text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'Let $(X_1,\dots,X_m)$ be self-adjoint non-commutative random variables distributed according to the free Gibbs law given by a sufficiently regular convex and semi-concave potential $V$, and let $(S_1,\dots,S_m)$ be a free semicircular family. We show that conditional expectations and conditional non-microstates free entropy given $X_1$, …, $X_k$ arise as the large $N$ limit of the corresponding conditional expectations and entropy for the random matrix models associated to $V$. Then by studying conditional transport of measure for the matrix models, we construct an isomorphism $\mathrm{W}^*(X_1,\dots,X_m) \to \mathrm{W}^*(S_1,\dots,S_m)$ which maps $\mathrm{W}^*(X_1,\dots,X_k)$ to $\mathrm{W}^*(S_1,\dots,S_k)$ for each $k = 1, \dots, m$, and which also witnesses the Talagrand inequality for the law of $(X_1,\dots,X_m)$ relative to the law of $(S_1,\dots,S_m)$.'
address: 'Department of Mathematics, UCLA, Los Angeles, CA 90095'
author:
- David Jekel
title: 'Conditional Expectation, Entropy, and Transport for Convex Gibbs Laws in Free Probability'
---
Introduction
============
Motivation
----------
Free probability initiated a fruitful exchange between random matrix theory and operator algebras. In many situations, tuples of $N \times N$ random matrices $(X_1^{(N)}, \dots, X_m^{(N)})$ can be described in the large $N$ limit by non-commutative random variables $X_1$, …, $X_m$ which are operators in a tracial $\mathrm{W}^*$-algebra. Conversely, many properties of non-commutative random variables (and the $\mathrm{W}^*$-algebras that they generate) are easier to understand when they can be simulated by finite-dimensional random matrix models. For instance, Voiculescu used free entropy, defined in terms of matricial microstates, to prove the absence of Cartan subalgebras for free group $\mathrm{W}^*$-algebras $L(\mathbb{F}_n)$ [@VoiculescuFE3]; similar techniques were used to give sufficient conditions for a von Neumann algebra to be non-prime and non-Gamma (a convenient list of results and references can be found in [@CN2019]). Further applications of random matrices to the properties of $\mathrm{C}^*$- and $\mathrm{W}^*$-algebras can be found for instance in [@HT2005] and [@GS2009 §4].
Free Gibbs laws are a prototypical example of the connection between random matrices and $\mathrm{W}^*$-algebras. Free Gibbs laws describe the large $N$ behavior of self-adjoint tuples of random matrices $X^{(N)} = (X_1^{(N)}, \dots, X_m^{(N)})$ given by a probability measure $\mu^{(N)}$ of the form $$d\mu^{(N)}(x) = \frac{1}{Z^{(N)}} e^{-N^2 V^{(N)}(x)}\,dx,$$ where $x \in M_N({\mathbb{C}})_{sa}^m$ is a self-adjoint tuple, $dx$ denotes Lebesgue measure, $V^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ is a function (known as a potential) chosen so that $e^{-N^2 V^{(N)}(x)}$ is integrable, and $Z^{(N)}$ is normalizing constant to make $\mu^{(N)}$ a probability measure. Here $V^{(N)}(x)$ could be given by $V^{(N)}(x) = \tau_N(p(x_1,\dots,x_m))$, where $\tau_N = (1/N) \operatorname{Tr}$ and $p$ is a non-commutative polynomial; for instance, taking $$V^{(N)}(x) = \frac{1}{2} \sum_{j=1}^m \tau_N(x_j^2)$$ produces the Gaussian unitary ensemble (GUE). Under certain assumptions on $V$ (e.g. convexity and good asymptotic behavior as $N \to \infty$), there will be non-commutative random variables $X_1$, …, $X_m$ in a tracial $\mathrm{W}^*$-algebra $(\mathcal{M},\tau)$ such that $$\tau_N(p(X_1^{(N)}, \dots, X_m^{(N)})) \to \tau(p(X_1,\dots,X_m)) \text{ in probability for every non-commutative polynomial } p;$$ see [@GMS2006 Theorems 3.3 and 3.4], [@DGS2016 Proposition 50 and Theorem 51], [@Jekel2018 Theorem 4.1]. The random matrix models satisfy the relation, derived from integration by parts, that $$E[\tau_N(D_{x_j} V^{(N)}(X^{(N)}) p(X^{(N)}))] = E[\tau_N \otimes \tau_N(\partial_{x_j}p(X^{(N)}))],$$ where $D_{x_j} V$ is a normalized gradient with respect to the coordinates of $x_j$ and $\partial_{x_j}$ denotes the free difference quotient, and hence the non-commutative tuple $X = (X_1,\dots,X_m)$ satisfies $$\tau(D_{x_j} V(X) p(X)) = \tau \otimes \tau(\partial_{x_j}p(X));$$ see [@GMS2006 §2.2 - 2.3]. The non-commutative law of a tuple $X$ satisfying such an equation is known as a *free Gibbs law* for the potential $V$.
Given sufficient assumptions on $V^{(N)}$ (for instance, Assumption \[ass:convexRMM\]), many of the classical quantities associated to $X^{(N)}$ will converge in the large $N$ limit to their free counterparts, besides obviously the convergence of the non-commutative moments $\tau_N(p(X^{(N)}))$. For instance, the normalized classical entropy will converge to the microstates free entropy (see [@VoiculescuFE1 §2], [@GS2009 Theorem 5.1], [@Jekel2018 §5.2]), and the normalized classical Fisher information will converge to the free Fisher information (see [@Jekel2018 §5.3]). The monotone transport maps of Guionnet and Shlyakhtenko are well-approximated by classical transport maps for the random matrix models [@GS2014 Theorem 4.7]. The solutions of classical SDE associated to the random matrix models approximate the solutions of free SDE; see for instance [@BCG2003], [@GS2009 §2], [@Dabrowski2017 §4].
Summary of Main Results
-----------------------
This paper will further develop the connection between classical and free probability for convex free Gibbs laws, by studying conditional expectation (§\[sec:conditionalexpectation\]), conditional entropy and Fisher information (§\[sec:entropy\]), and conditional transport (§\[sec:transport\]). This is an extension of our previous work [@Jekel2018].
We consider a sequence of random matrix tuples $(X^{(N)},Y^{(N)}) = (X_1^{(N)}, \dots, X_m^{(N)}, Y_1^{(N)}, \dots, Y_n^{(N)})$ given by a uniformly convex and semi-concave sequence of potentials $V^{(N)}$ such that the normalized gradient $DV^{(N)}$ is asymptotically approximable by trace polynomials (a notion of good asymptotic behavior as $N \to \infty$ defined in §\[subsec:AATP\]). Then the following results hold:
(1) The non-commutative moments $\tau_N(p(X^{(N)},Y^{(N)}))$ converge in probability to $\tau(p(X,Y))$ for some tuple $(X,Y)$ of non-commutative random variables in a tracial $\mathrm{W}^*$-algebra. See Theorem \[thm:freeGibbslaw\].
(2) The classical conditional expectation $E[f^{(N)}(X^{(N)},Y^{(N)})| Y^{(N)}]$ behaves asymptotically like the non-commutative conditional expectation $E_{\mathrm{W}^*(Y)}[f(X,Y)]$ where $f$ comes from an appropriate non-commutative function space and $f^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})$ is a sequence of uniformly Lipschitz functions that “behaves like $f$ in the large $N$ limit” in the sense of §\[sec:TP\]. See Theorem \[thm:conditionalexpectation\].
(3) The classical conditional entropy $N^{-2} h(X^{(N)} | Y^{(N)}) + (m/2) \log N$ converges to the conditional free entropy $\chi^*(X: \mathrm{W}^*(Y))$. This is a similar to a conditional version of $\chi = \chi^*$. See Theorem \[thm:convergenceofentropy\].
(4) There exists a function $f(X,Y)$ such that $(f(X,Y),Y) \sim (S,Y)$ in non-commutative law, where $S$ is a free semicircular $m$-tuple freely independent of $Y$, and this function also arises from functions $f^{(N)}$ such that $(f^{(N)}(X^{(N)},Y^{(N)}),Y^{(N)}) \sim (S^{(N)},Y^{(N)})$, where $S^{(N)}$ is an independent GUE $m$-tuple. This is the conditional version of transport to the Gaussian/semicircular law. See Theorems \[thm:transport3\].
(5) This transport map also witnesses the conditional entropy-cost inequality for the law of $X$ relative to semicircular conditioned on $Y$. See Theorem \[thm:transport3\].
(6) This transport map furnishes an isomorphism $\mathrm{W}^*(X,Y) \cong \mathrm{W}^*(S,Y) \cong \mathrm{W}^*(S) * \mathrm{W}^*(Y)$, which shows that $\mathrm{W}^*(Y)$ is freely complemented in $\mathrm{W}^*(X,Y)$.
(7) Actually, a second application of transport shows that $\mathrm{W}^*(Y)$ is isomorphic to the $\mathrm{W}^*$-algebra generated by a semicircular $n$-tuple, or in other words $L(\mathbb{F}_n)$. So altogether there is an isomorphism $\mathrm{W}^*(X,Y) \to L(\mathbb{F}_{m+n})$ that maps $\mathrm{W}^*(Y)$ to the canonical copy of $L(\mathbb{F}_n)$ inside $L(\mathbb{F}_{m+n})$.
Furthermore, the results about transport can be iterated to produce a “lower-triangular transport” as shown in Theorem \[thm:transport4\] and discussed further in §\[subsec:transportintro\]. This is analogous to the classical results on triangular transport of measure such as [@BKM2005].
In the rest of the introduction, we will review notation and then motivate and explain the main results in more detail. In the course of the paper, it will become clear that not only are our main results proved all by the similar techniques, but in fact their statements and proofs are tightly interrelated.
Notation and Background
-----------------------
We will continue to use the same notation and background as in [@Jekel2018]. The one major change is that we will write superscript $(N)$ rather than subscript $N$ for measures and functions defined on $N \times N$ matrices. Moreover, we will use the original notation $\partial$ for Voiculescu’s free difference quotient, even though [@Jekel2018] used $\mathcal{D}$.
We assume familiarity with the basic properties of tracial $\mathrm{W}^*$-algebras (or tracial von Neumann algebras); see for instance [@AP2017]. In particular, a tracial $\mathrm{W}^*$-algebra is a finite $\mathrm{W}^*$-algebra $\mathcal{M}$ with a specified trace $\tau: \mathcal{M} \to {\mathbb{C}}$. If $\mathcal{N} \subseteq \mathcal{M}$ is a $\mathrm{W}^*$-algebra, then there is a unique trace-preserving conditional expectation $E_{\mathcal{N}}: \mathcal{M} \to \mathcal{N}$. If $x = (x_1,\dots,x_m)$ is a tuple of operators in $\mathcal{M}$, then we denote by $\mathrm{W}^*(x)$ the $\mathrm{W}^*$-subalgebra which they generate.
There is an inner product on $\mathcal{M}$ defined by $\ip{x,y}_2 = \tau(x^*y)$, and the completion of $\mathcal{M}$ in this inner product is a Hilbert space known as $L^2(\mathcal{M},\tau)$. We denote the self-adjoint elements of $\mathcal{M}$ by $\mathcal{M}_{sa}$ and recall that if $x$ and $y$ are self-adjoint, then $\ip{x,y}_2$ is real. If $x = (x_1,\dots,x_m)$ and $y = (y_1,\dots,y_m)$ are tuples, we denote $\ip{x,y}_2 = \sum_{j=1}^m \ip{x_j,y_j}_2$. We define $\norm{x}_\infty = \max_j \norm{x_j}_\infty$, that is, the maximum of the operator norms of $x_j$.
We denote by $\operatorname{NCP}_m = {\mathbb{C}}\ip{X_1,\dots,X_m}$ the $*$-algebra of non-commutative polynomials in $m$ self-adjoint variables. A *non-commutative law* is a linear map $\lambda: {\mathbb{C}}\ip{X_1,\dots,X_m} \to {\mathbb{C}}$ satisfying
(A) $\lambda(1) = 1$.
(B) $\lambda(p^*p) \geq 0$ for all $p \in \operatorname{NCP}_m$.
(C) $\lambda(pq) = \lambda(qp)$ for all $p, q \in \operatorname{NCP}_m$.
(D) $|\lambda(X_{i_1} \dots X_{i_k})| \leq R^k$ for some constant $R$.
The set of non-commutative laws that satisfy (D) for a fixed value of $R$ is denoted $\Sigma_{m,R}$, and it is equipped with the topology of pointwise convergence on $\operatorname{NCP}_m$. Likewise, the space of all laws, equipped with the topology of pointwise convergence, will be denoted by $\Sigma_m$.
If $x = (x_1,\dots,x_m)$ is a tuple of self-adjoint elements of $(\mathcal{M},\tau)$, then we may define a non-commutative law $\lambda_x$ by $\lambda_x(p) = \tau(p(x))$. Conversely, every non-commutative law can be realized in this way through the GNS construction. In particular, a free Gibbs law can be realized by a tuple $(x_1,\dots,x_m)$ of self-adjoint operators, and thus the free Gibbs law has a corresponding $\mathrm{W}^*$-algebra $\mathrm{W}^*(x)$, that is unique up to isomorphism.
We always consider $M_N({\mathbb{C}})$ as a tracial $\mathrm{W}^*$-algebra with the normalized trace $\tau_N = (1/N) \operatorname{Tr}$, and in particular, we use the notation $\norm{x}_2$, $\norm{x}_\infty$, and $\lambda_x$ as defined above when $x$ is an $m$-tuple of matrices. The notation $\norm{\cdot}_2$ and $\norm{\cdot}_\infty$ will never be used for the $L^2$ or $L^\infty$ norms of *functions* on matrices, but if we write an $L^p$ norm it will be expressed $\norm{\cdot}_{L^p}$.
For a smooth function $u: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$, we denote by $Du$ and $Hu$ the gradient and Hessian with respect to the normalized inner product $\ip{\cdot,\cdot}_2$. In other words, $Du(x_0)$ is the vector in $M_N({\mathbb{C}})_{sa}^m$ and $Hu(x_0)$ is the ${\mathbb{R}}$-linear transformation of $M_N({\mathbb{C}})_{sa}^m$ satisfying $$u(x) = u(x_0) + \ip{Du(x_0), x - x_0}_2 + \frac{1}{2} \ip{Hu(x_0)(x - x_0), x - x_0}_2 + o(\norm*{x - x_0}_2^2).$$ For functions $f: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ or $M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$, we denote $\norm{f}_{\operatorname{Lip}}$ the Lipschitz (semi)norm with respect to using $\norm{\cdot}_2$ on $M_N({\mathbb{C}})_{sa}^m$ and $M_N({\mathbb{C}})$.
Note that $M_N({\mathbb{C}})_{sa}^m$ can also be equipped with the real inner product $\ip{x,y}_{\operatorname{Tr}} = \sum_{j=1}^m \operatorname{Tr}(x_jy_j) = N \ip{x,y}_2$. Being a real inner-product space, $M_N({\mathbb{C}})_{sa}^m$ may be identified with ${\mathbb{R}}^{mN^2}$ by choosing an orthonormal basis in $\ip{\cdot,\cdot}_{\operatorname{Tr}}$. Lebesgue measure on $M_N({\mathbb{C}})_{sa}^m$ should be understood with respect to this identification. Moreover, the gradient $\nabla$, Jacobian matrix $J$, divergence $\operatorname{Div}$, and Laplacian $\Delta$ for functions on $M_N({\mathbb{C}})_{sa}^m$ should also be understood with respect to this identification. Beware that this is *not* equivalent to using entrywise coordinates for $M_N({\mathbb{C}})_{sa}^m$ since the off-diagonal entries are complex and conjugate-symmetric, while the diagonal entries are real, and that the normalized gradient above satisfies $Df = N \nabla f$. For further discussion see [@Jekel2018 §2.1].
Main Results on Conditional Expectation
---------------------------------------
Consider a tuple $$(X^{(N)},Y^{(N)}) = (X_1^{(N)},\dots,X_m^{(N)},Y_1^{(N)},\dots,Y_n^{(N)})$$ of random self-adjoint matrices given by a probability density $(1/Z^{(N)}) e^{-N^2 V^{(N)}(x,y)}\,dx\,dy$. We assume that $V^{(N)}$ is uniformly convex and semi-concave and that the normalized gradient $DV^{(N)}$ is asymptotically approximable by trace polynomials (a certain notion of good asymptotic behavior as $N \to \infty$, explained below). The precise hypotheses are listed in Assumption \[ass:convexRMM\]. We showed in [@Jekel2018 Theorem 4.1] that in this case, there exists an $(m+n)$-tuple $(X,Y)$ of non-commutative random variables such that $\tau_N(p(X,Y)) \to \tau(p(X,Y))$ in probability.
Our first main result (Theorem \[thm:conditionalexpectation\]) says roughly that the classical conditional expectation given $Y^{(N)}$ well approximates the $\mathrm{W}^*$-algebraic conditional expectation $E_{\mathrm{W}^*(Y)}: \mathrm{W}^*(X,Y) \to \mathrm{W}^*(Y)$. This is motivated in general by the importance of conditional expectation in free probability, e.g. its relationship to free independence with amalgamation and to free score functions. See [@BCG2003 §4] for a study of the large $N$ limits of conditional expectations related to matrix SDE. The relationship between classical and free conditional expectation also has implications for the study of relative matricial microstate spaces, such as the “external averaging property” introduced in the upcoming work with Hayes, Nelson, and Sinclair [@HJNS2019]. Applications of conditional expectation within this paper include our results on free Fisher information and entropy (see Theorem \[thm:convergenceofentropy\] and Remark \[rem:simplifiedentropyproof\]), as well as our proof that Assumption \[ass:convexRMM\] is preserved under marginals (see Proposition \[prop:marginals\]).
The statement and proof of Theorem \[thm:conditionalexpectation\] rely on a notion of asymptotic approximation for functions on $M_N({\mathbb{C}})_{sa}^m$ explained in §\[sec:TP\]. We define a class of non-commutative functions $\overline{\operatorname{TrP}}_m^1$ as a certain Fr[é]{}chet space completion of trace polynomials, such that if $f \in \overline{\operatorname{TrP}}_m^1$ and $x_1$, …, $x_m$ are self-adjoint elements in an $\mathcal{R}^\omega$-embeddable tracial $\mathrm{W}^*$-algebra $(\mathcal{M},\tau)$, then $f(x_1,\dots,x_m)$ is a well-defined element of $L^2(\mathcal{M})$. In particular, every $f \in \overline{\operatorname{TrP}}_m^1$ can be evaluated on a tuple of self-adjoint matrices. Now if $f^{(N)}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$, we say that $f^{(N)} \rightsquigarrow f$ if for every $R > 0$, $$\lim_{N \to \infty} \sup_{\substack{x \in M_N({\mathbb{C}})_{sa}^m \\ \norm{x}_\infty \leq R}} \norm*{f^{(N)}(x) - f(x)}_2 = 0,$$ Moreover, if such an $f$ exists, then we say that $f^{(N)}$ is *asymptotically approximable by trace polynomials*.
Consider the random matrices $(X^{(N)},Y^{(N)})$ and non-commutative random variables $(X,Y)$ as above, and suppose that $f^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})$ is uniformly Lipschitz in $\norm{\cdot}_2$ and that $f^{(N)} \rightsquigarrow f \in \overline{\operatorname{TrP}}_{m+n}^1$. Then we show that $E[f^{(N)}(X^{(N)},Y^{(N)}) | Y^{(N)}]$ is given by a function $g^{(N)}(Y^{(N)})$ such that $g^{(N)} \rightsquigarrow g \in \overline{\operatorname{TrP}}_n^1$, and moreover $E_{\mathrm{W}^*(Y)}[f(X,Y)] = g(Y)$.
A curious feature of this result is that the function $g$ is defined for all self-adjoint $n$-tuples of non-commutative random variables, not only for the specific $n$-tuple $Y$ that we are concerned with. Similarly, the claim that $g^{(N)} \rightsquigarrow g$ describes the asymptotic behavior of $g^{(N)}(y)$ for all $y \in M_N({\mathbb{C}})_{sa}^n$, even though the distribution of the random matrix $Y^{(N)}$ is highly concentrated as $N \to \infty$ on much smaller sets, namely the “matricial microstate spaces” consisting of tuples $y \in M_N({\mathbb{C}})_{sa}^n$ with non-commutative moments close to those of $Y$. Thus, the statement we prove about the functions $g^{(N)}$ is stronger than an asymptotic result about $L^2$ approximation such as [@GS2014 Theorem 4.7].
Main Results on Entropy
-----------------------
Voiculescu defined two types of free entropy (see [@VoiculescuFE2], [@VoiculescuFE5], [@Voiculescu2002]). The first, called $\chi(X)$, is based on measuring the size of matricial microstate spaces, which is closely related to the classical entropy of the random matrix models (see [@Jekel2018 §5.2]). The second, called $\chi^*(X)$, is defined in terms of free Fisher information, which is based on classical Fisher information. Either one should heuristically be the large $N$ limit of the classical entropy of random matrix models, but there were many technical obstacles to proving this. The inequality $\chi \leq \chi^*$ is known in general thanks to [@BCG2003]. However, even for non-commutative laws as well-behaved and explicit as free Gibbs states given by convex potentials, the equality of $\chi$ and $\chi^*$ when $m > 1$ was not proved until Dabrowski’s paper [@Dabrowski2017], and the problem is still open for non-convex Gibbs states.
Our previous work [@Jekel2018] gave a proof of this equality in the convex case based on the asymptotic analysis of functions and PDE related to the random matrix models. Here we will use similar techniques for the conditional setting. We will show (Theorem \[thm:convergenceofentropy\]) that for a random tuple of matrices $(X^{(N)},Y^{(N)})$ given by a convex potential as above, the classical conditional entropy $N^{-2} h(X^{(N)} | Y^{(N)}) + (m/2) \log N$ converges to the conditional free entropy $\chi^*(X: \mathrm{W}^*(Y))$. Actually, the proof here is shorter than those of [@Dabrowski2017] and [@Jekel2018] (see Remark \[rem:simplifiedentropyproof\]), even considering the results we used from [@Jekel2018].
We focus here only on the non-microstates entropy (defined using Fisher information). It is not yet resolved in the literature what the correct definition of conditional microstates free entropy should be. In light of [@Jekel2018 §5.2], the conditional classical entropy for the random matrix models seems to be a reasonable substitute for microstates entropy, and in the convex setting we expect this to agree with any plausible definition of conditional microstates entropy due to the exponential concentration of measure.
Main Results on Transport {#subsec:transportintro}
-------------------------
A *transport map* from a probability measure $\mu$ and to another probability measure $\nu$ is a function $f$ such that $f_* \mu = \nu$. In probabilistic language, if $X \sim \mu$ and $Y \sim \nu$ are random variables, then $f_* \mu = \nu$ means that $f(X) \sim Y$ in distribution. The theory of transport (and in particular optimal transport) has numerous and significant applications in the classical setting. For instance, if we have a function $f$ such that $f(X) \sim Y$ and we can numerically simulate the random variable $X$, then we can also simulate $Y$.
In the non-commutative world, transport is even more significant. As remarked in [@GS2014 §1.1], there is no known analogue of a probability density in free probability. However, the existence of transport maps that would express our given random variables as functions of a free semicircular family (for instance) would serve a similar purpose to a density, namely to provide a fairly explicit and analytically tractable model for a large class of non-commutative laws.
Moreover, in contrast to the classical setting, the very existence of transport maps is a nontrivial condition. Being able to express a non-commutative tuple $Y$ as a function of another non-commutative tuple $X$ implies that $\mathrm{W}^*(Y)$ embeds into $\mathrm{W}^*(X)$, and having a transport map in the other direction as well implies that $\mathrm{W}^*(Y) \cong \mathrm{W}^*(X)$. In the classical setting, any two diffuse (non-atomic) standard Borel probability spaces are isomorphic. On the other hand, there are many non-isomorphic diffuse tracial $\mathrm{W}^*$-algebras, even after restricting our attention to factors (those which cannot be decomposed as direct sums); see [@McDuff1969]. Moreover, Ozawa [@Ozawa2004] showed that there is no separable tracial factor that contains an isomorphic copy of all the others. Thus, there are many instances where it is not even possible to transport one given non-commutative law to another.
The papers [@GS2009] and [@DGS2016] showed the existence of monotone transport maps between certain free Gibbs laws given by convex potentials and the law of a free semicircular family, and thus concluded that each of the corresponding $\mathrm{W}^*$-algebras was isomorphic to a free group factor $L(\mathbb{F}_n)$. In particular, this result applies to the $q$-Gaussian variables for sufficiently small $q$. These transport techniques have been extended to type III von Neumann algebras [@Nelson2015a], to planar algebras [@Nelson2015b], and to interpolated free group factors [@HN2018]. We will focus on “conditional transport” in the tracial setting.
Our first main result about transport is contained in Theorems \[thm:transport1\] and \[thm:transport2\]. Let $(X^{(N)},Y^{(N)})$ be an $(m+n)$-tuple of random matrices arising from a sequence of convex potentials satisfying Assumption \[ass:convexRMM\]. Let $(X,Y)$ be an $(m+n)$-tuple of non-commutative self-adjoint variables realizing the limiting free Gibbs law. Then we construct functions $F^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})_{sa}^m$ such that $(F^{(N)}(X^{(N)}, Y^{(N)}), Y^{(N)}) \sim (S^{(N)}, Y^{(N)})$ in distribution, where $S^{(N)}$ is a GUE $m$-tuple independent of $Y^{(N)}$. We think of this as a conditional transport, which transports the law of $X^{(N)}$ to the law of $S^{(N)}$ *conditioned on $Y^{(N)}$*.
Moreover, we show that the transport maps satisfy $F^{(N)} \rightsquigarrow F \in (\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$. In the large $N$ limit, we obtain $(F(X,Y), Y) \sim (S,Y)$ in non-commutative law, where $S$ is a free semicircular $m$-tuple freely independent of $Y$. In particular, this means that $\mathrm{W}^*(X,Y) \cong \mathrm{W}^*(S,Y) = \mathrm{W}^*(S) * \mathrm{W}^*(Y)$ (where $*$ denotes free product). In other words, $\mathrm{W}^*(Y)$ is freely complemented in $\mathrm{W}^*(X,Y)$.
By iterating this result, we can show that if $X = (X_1,\dots,X_m)$ is a tuple of non-commutative random variables given by a convex free Gibbs state as above, then there is an isomorphism $\mathrm{W}^*(X) \to \mathrm{W}^*(S)$ such that $\mathrm{W}^*(X_1,\dots,X_k)$ is mapped onto $\mathrm{W}^*(S_1,\dots,S_k)$ for each $k = 1$, …, $m$. In other words, there is a “lower-triangular transport.” See Theorem \[thm:transport4\]. This is a (partial) free analogue of [@BKM2005 Corollary 3.10].
This result implies in particular that $\mathrm{W}^*(X_1)$ is a maximal abelian subalgebra and in fact maximal amenable (since the subalgebra $\mathrm{W}^*(S_1)$ is known to be maximal amenable thanks to Popa [@Popa1983]), and the same holds for each $\mathrm{W}^*(X_j)$ by symmetry. For context on maximal amenable subalgebras, see for instance [@Popa1983] [@BC2015] [@BH2018]. More generally, *any von Neumann algebraic properties* of the sequence of inclusions $\mathrm{W}^*(X_1) \subseteq \mathrm{W}^*(X_1,X_2) \subseteq \dots \subseteq \mathrm{W}^*(X_1,\dots,X_m)$ are the same as for the case of free semicirculars, that is, for the standard inclusions $L({\mathbb{Z}}) \subseteq L({\mathbb{F}}_2) \subseteq \dots \subseteq L({\mathbb{F}}_m)$.
Denote by $F$ the transport map from the law of $X$ to the law of $S$ in our construction, so that $F(X) \sim S$. We can also arrange that $F$ witnesses the Talagrand entropy-cost inequality relative to the semicircular law, that is, $$\norm{F(X) - X}_2^2 \leq \norm{X}_2^2 + m \log 2\pi - 2 \chi^*(X),$$ where the left hand side is twice the entropy relative to semicircular (see §\[subsec:entropycost\]). This is not surprising because it was already known in the classical case that the Talagrand inequality can be witnessed by some triangular transport [@BKM2005 Corollary 3.10]. Moreover, our construction of the transport maps is a direct application of the same method that Otto and Villani used to prove the Talagrand entropy-cost inequality under the assumption of the log-Sobolev inequality [@OV2000 §4]. Thus, our main contribution is to study the large $N$ limit of the transport maps using asymptotic approximation by trace polynomials. We also show that $F$ is $\norm{\cdot}_2$-Lipschitz, and we estimate $\norm{F(X) - X}_\infty$ in terms of the constants $c$ and $C$ specifying the uniform convexity and semi-concavity of $V^{(N)}$. These estimates will in fact go to zero as $c, C \to 1$.
Unfortunately, the maps constructed here are not optimal triangular transport maps with respect to the $L^2$-Wasserstein distance, since Otto and Villani’s proof of [@OV2000 Theorem 1] uses a diffusion-semigroup interpolation between the two measures, not the displacement interpolation from optimal transport theory. In that sense, the results of this paper do not fully prove an analogue of [@BKM2005 Corollary 3.10]. Even in the work of Guionnet and Shlyakhtenko [@GS2009], which constructed monotone transport maps in the free setting, the question of whether these maps furnish an optimal coupling between $X$ and $S$ inside a tracial von Neumann algebra was left unresolved. Future research should study optimal transport in the free setting, and determine whether the classical optimal transport (or more generally optimal triangular transport) maps for the random matrix models converge in the large $N$ limit in the sense of this paper.
Outline
-------
The paper is organized as follows. We remark that §\[sec:RMbackground\] and §\[sec:diffeqtools\] are mostly technical background, and the reader may treat them like appendices if desired. In other words, it is feasible to read through the other sections in order and only refer to §\[sec:RMbackground\] and §\[sec:diffeqtools\] as needed to verify technical details of the main results.
§\[sec:RMbackground\] gives standard background on convex and semi-concave functions and on log-concave random matrix models.
§\[sec:TP\] sets up the algebra of trace polynomials, and the spaces $\overline{\operatorname{TrP}}_m^0$ and $\overline{\operatorname{TrP}}_m^1$ of functions that can be approximated by trace polynomials. These spaces provide a framework for functional calculus in multiple self-adjoint variables $X_1$, …, $X_m$ that can realize every element of $L^2(\mathrm{W}^*(X_1,\dots,X_m))$. They are a convenient tool to describe the large $N$ behavior of functions of several matrices, and thus will be used in the statements of our main theorems.
§\[sec:diffeqtools\] describes solving ODE’s and the heat equation over $\overline{\operatorname{TrP}}_m^1$. These are the technical lemmas used in the rest of the paper to show that the solutions of certain PDE’s have well-defined large $N$ limits.
§\[sec:conditionalexpectation\] explains the setup of our random matrix models given by convex potentials, and then proves our main result on conditional expectation (Theorem \[thm:conditionalexpectation\]).
§\[sec:entropy\] shows that the conditional entropy for random matrix models converges to the conditional non-microstates entropy (Theorem \[thm:convergenceofentropy\]).
§\[sec:transport\] proves the existence of transport maps from a free Gibbs law to the law of a free semicircular $m$-tuple which arise as the large $N$ limit of transport maps for the random matrix models (Theorem \[thm:transport1\] and \[thm:transport2\]).
§\[sec:applications\] discusses applications of our results. We show that our standard set of assumptions for log-concave random matrix models is preserved under marginals, independent joins, linear change of variables, and convolution (§\[subsec:operations\]). We show that the transport maps constructed above witness (the conditional version of) Talagrand’s entropy-cost inequality relative to Gaussian measure (Theorem \[thm:transport3\]). Then by iterating our conditional transport results, we show the existence of triangular transport (Theorem \[thm:transport4\]).
Multi-matrix Models from Convex Potentials {#sec:RMbackground}
==========================================
This section is a review and reference for basic results we will use throughout the paper.
We will be concerned with probability measures on $M_N({\mathbb{C}})_{sa}^m$ of the form $$d\mu(x) = \frac{1}{Z} e^{-N^2 V(x)}\,dx,$$ where $x = (x_1,\dots,x_m)$ is a tuple of self-adjoint matrices, $V: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ such that $e^{N^2 V}$ is integrable, and $Z = \int e^{-N^2 V(x)}\,dx$ is the normalizing constant. Here $dx$ denotes Lebesgue measure where we identify $M_N({\mathbb{C}})_{sa}^m$ with ${\mathbb{R}}^{mN^2}$ using the inner product associated to the trace (the normalization of Lebesgue measure is irrelevant here because if we multiply it by a constant, the normalizing constant $Z$ for $\mu$ will change to compensate). In this case, we will say that *$\mu$ is the measure given by the potential $V$*. We will often assume $V$ is convex. Note that $\mu$ only determines $V$ up to an additive constant, but we will still say that “$V$ is the potential corresponding to $\mu$” with a slight abuse of terminology.
A primary motivating example is $V(x) = \tau_N(f(x))$, where $\tau_N = (1/N) \operatorname{Tr}$ is the normalized trace and $f$ is a non-commutative polynomial in $x_1$, …, $x_m$. Unlike the notation in many random matrix papers, we prefer to write $N^2 \tau_N(f)$ rather than $N \operatorname{Tr}(f)$. This seems natural because $\tau_N(f)$ is a function with dimension-independent normalization and it would make sense for self-adjoint elements of a tracial $\mathrm{W}^*$-algebra. Meanwhile, $N^2$ is the dimension of $M_N({\mathbb{C}})_{sa}^N$ and also the scale (in the sense of large deviations) for the standard concentration estimates that hold when $V$ is uniformly convex (see for instance [@BCG2003] or §\[subsec:concentration\] below).
Semi-convex and Semi-concave Functions
--------------------------------------
\[def:convexityHnotation\] Let $A: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})_{sa}^m$ be a self-adjoint linear transformation and let $u: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$. We say that $Hu \leq A$ if $u(x) - (1/2) \ip{Ax,x}_2$ is concave. We say that $Hu \geq A$ if $u(x) - (1/2) \ip{Ax,x}_2$ is convex.
We will also regularly use the following observation:
\[lem:convexgradient\] Suppose that $u: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$, and let $A$ and $B$ be self-adjoint linear transformations. The following are equivalent:
(1) $A \leq Hu \leq B$.
(2) For each $x \in M_N({\mathbb{C}})_{sa}^m$, there exists $y \in M_N({\mathbb{C}})_{sa}^m$ such that $$\frac{1}{2} \ip{A(x' - x), x' - x} \leq u(x') - u(x) - \ip{y, x' - x}_2 \leq \frac{1}{2} \ip{B(x' - x), x' - x}_2$$ for all $x' \in M_N({\mathbb{C}})_{sa}^m$.
(3) $u$ is continuously differentiable and we have $$\ip{A(x'-x), x'-x}_2 \leq \ip{Du(x') - Du(x),x' - x}_2 \leq \ip{B(x'-x), x' - x}_2$$ for all $x, x' \in M_N({\mathbb{C}})_{sa}^m$.
Moreover, in this case, $Du$ is $\max(\norm{A}, \norm{B})$-Lipschitz with respect to $\norm{\cdot}_2$.
\(1) $\implies$ (3). Suppose (1) holds. If $C = \max(\norm{A},\norm{B})$, then for each $x$ there exists $y$ such that $$-\frac{C}{2} \norm{x' - x}_2^2 \leq u(x') - u(x) - \ip{y, x' - x}_2 \leq \frac{C}{2} \norm{x' - x}_2^2.$$ Hence, it follows from [@Jekel2018 Proposition 2.13] that $u$ must be continuously differentiable and $Du$ is $C$-Lipschitz (which proves the last claim of our lemma as well). To prove the inequality asserted by (3), we can reduce to the case when $u$ is smooth using a similar argument as in [@Jekel2018 Proposition 2.13]). But in the smooth case, the claim follows by estimating from above and below the formula $$\ip{Du(x') - Du(x), x' - x}_2 = \int_0^1 \ip{Hu(x + t(x' - x))(x' - x), x' - x}_2\,dt,$$ where $Hu$ is the Hessian defined in the standard pointwise sense.
\(3) $\implies$ (2). Recall the formula $$u(x') - u(x) = \int_0^1 \ip{Du(x + t(x' - x)), x' - x}\,dt.$$ This implies that $$\begin{aligned}
u(x') - u(x) - \ip{Du(x), x' - x}_2 &= \int_0^1 \ip{Du(x + t(x' - x)) - Du(x), x' - x}_2\,dt \\
&= \int_0^1 \frac{1}{t} \ip{Du(x + t(x' - x)) - Du(x), [x + t (x' - x)] - x}_2\,dt \\
&\leq \int_0^1 \frac{1}{t} \ip{B[t(x' - x)], t(x' - x)}_2\,dt \\
&= \frac{1}{2} \ip{B(x' - x), x' - x}_2.\end{aligned}$$ This proves the upper bound, and the lower bound is symmetrical.
\(2) $\implies$ (1). This follows from the characterization of convex functions by supporting hyperplanes. Indeed, $u(x) - (1/2) \ip{Ax,x}$ is convex if and only if for every $x$, there exists $y$ satisfying $$u(x') - u(x) + \frac{1}{2} \ip{Ax', x'}_2 - \frac{1}{2} \ip{Ax,x}_2 \geq \ip{y, x - x'}_2.$$ which is equivalent to the right inequality of (2), and the concavity of $u(x) - (1/2) \ip{Bx,x}$ follows similarly.
\[lem:convexgradientestimate\] Suppose that $0 \leq Hu \leq A$ for some linear transformation $A$. Then $u$ is differentiable and we have $$|\ip{Du(x) - Du(x'),y}_2| \leq \ip{A(x-x'),x-x'}_2^{1/2} \ip{Ay,Ay}_2^{1/2},$$ so that in particular, $\norm{Du(x) - Du(x')}_2 \leq \norm{A} \norm{x - x'}_2$.
As in [@Jekel2018 Proposition 2.13], we obtain differentiability; and moreover to prove the asserted estimate, it suffices to prove the claim for smooth functions $u$. In this case, $$\begin{aligned}
\ip{Du(x) - Du(x'),y} &= \int_0^1 \ip{Hu(tx + (1 - t) x')(x - x'),y}_2 \,dt \\
&\leq \int_0^1 \ip{Hu(tx + (1-t)x')(x-x'),x-x'}_2^{1/2} \ip{Hu(tx + (1-t)x')y,y}_2^{1/2}\,dt \\
&\leq \int_0^1 \ip{A(x-x'), x-x'}^{1/2} \ip{Ay,y}_2^{1/2}\,dt \\
&= \ip{A(x-x'),x-x'}_2^{1/2} \ip{Ay,y}_2^{1/2}. \qedhere\end{aligned}$$
Some Basic Lemmas
-----------------
Let $V: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ satisfy $HV \geq c$. Then one can check that $e^{-N^2 V(x)}$ is integrable; indeed, $V$ must achieve a minimum at some $x_0$ and we have $V(x) \geq V(x_0) + (c/2) \norm{x - x_0}_2^2$ and clearly $e^{-N^2 c \norm{x - x_0}_2^2}$ is integrable. Therefore, the probability measure $\mu$ given by $(1 / Z^{(N)}) e^{-N^2 V(x)}\,dx$ is well-defined.
\[lem:conjugatevariablebasics\] Let $V: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ satisfy $c \leq HV \leq C$ for some scalars $0 < c \leq C$. Let $\mu$ be the probability measure given by $d\mu(x) = (1/ Z^{(N)}) e^{-N^2 V(x)}\,dx$ and let $X$ be a random variables whose distribution is $\mu$. Then $$E[DV(X)] = 0$$ and $$\frac{m}{C} \leq E \norm*{X - E(X)}_2^2 \leq \frac{m}{c}.$$
We remark that $V$ is continuously differentiable by Lemma \[lem:convexgradient\] $V$ is differentiable and $DV$ is Lipschitz. It follows by some straightforward estimation that $\norm{DV}_2$ is integrable with respect to $\mu$, so that $E[DV(X)]$ is well-defined. Then $E[DV(X)] = 0$ follows from integration by parts (see §\[subsec:matrixentropy\] for further context on this integration by parts).
Next, let $D_j V$ denote the normalized gradient with respect to the matrix variable $x_j$. Using integration by parts again, we get $E \ip{D_j V(X), X_j - E(X_j)}_2 = 1$, so that $$E \ip{DV(X) - DV(E(X)), X - E(X)}_2 = E \ip{DV(X), X - E(X)}_2 = m.$$ On the other hand, by Lemma \[lem:convexgradientestimate\], $$c E \norm*{X - E(X)}_2^2 \leq E \ip{DV(X) - DV(E(X)), X - E(X)}_2 \leq C E \norm*{X - E(X)}_2^2.$$ Since the middle term evaluates to $m$, the proof is complete.
\[lem:RVboundedness\] Let $X$ be a random variable in $M_N({\mathbb{C}})_{sa}^m$ and let $G: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})_{sa}^n$ be Lipschitz with respect to $\norm{\cdot}_2$ in both the domain and target space, and let $\norm{G}_{\operatorname{Lip}}$ denote the corresponding Lipschitz (semi)norm. Then $$\norm*{G(x)}_2 \leq \norm*{E(G(X))}_2 + \norm*{G}_{\operatorname{Lip}} \left( \norm*{x - E(X)}_2 + (E\norm*{X - E(X)}_2^2)^{1/2} \right).$$
Note that $$\begin{aligned}
\norm*{G(x) - E(G(X))}_2 &\leq \norm*{G}_{\operatorname{Lip}} E \norm*{x - X}_2 \\
&\leq \norm*{G}_{\operatorname{Lip}} \left( \norm*{x - E(X)}_2 + E \norm*{X - E(X)}_2 \right) \\
&\leq \norm*{G}_{\operatorname{Lip}} \left( \norm*{x - E(X)}_2 + (E \norm*{X - E(X)}_2^2)^{1/2} \right). \qedhere\end{aligned}$$
\[cor:DVestimate\] Let $V: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ satisfies $c \leq HV \leq C$, let $\mu$ be the corresponding measure, and let $X \sim \mu$. Then $$\norm*{DV(x)}_2 \leq C \left( \norm*{x - E(X)}_2 + \frac{m^{1/2}}{c^{1/2}} \right).$$
We apply Lemma \[lem:RVboundedness\] to $DV(X)$. Also, $DV$ is $C$-Lipschitz by Lemma \[lem:convexgradient\]. By Lemma \[lem:conjugatevariablebasics\] $E(DV(X)) = 0$ and $E \norm*{X - E(X)}_2^2 \leq m / c$.
\[lem:limitoflogconcave\] Let $A$ and $B$ be positive definite linear transformations $M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})_{sa}^m$. Let $\{V_k\}_{k \in {\mathbb{N}}}$ be a sequence of functions such that $A \leq HV_k \leq B$. Let $d\mu_k(x) = (1/Z_k) e^{-N^2 V_k(x)}\,dx$ be the associated probability measure. Let $\mu$ be another measure with finite mean. Suppose $\mu_k$ converges weakly to $\mu$ and the mean of $\mu_k$ is bounded in $\norm*{\cdot}_2$ as $k \to \infty$. Then there exists $V$ such that $d\mu(x) = (1/Z) e^{-N^2 V(x)}\,dx$ and $A \leq HV \leq B$.
Since adding a constant to $V_k$ does not change $\mu_k$, we can assume without loss of generality that $V_k(0) = 0$. Now $DV_k$ is $C$-Lipschitz where $C = \max(\norm{A}, \norm{B})$, hence the sequence is equicontinuous. It is also pointwise bounded in light of the previous lemma, since we assumed the mean of $\mu_k$ is bounded as $k \to \infty$. Thus, by the Arzel[à]{}-Ascoli theorem, by passing to a subsequence, we can assume that $DV_k$ converges locally uniformly to some $F$ as $k \to \infty$. Since $V_k(0) = 0$, this also implies that $V_k$ converges locally uniformly to some $V$, which must satisfy $A \leq HV \leq B$ since the family of such functions is closed under pointwise limits (which follows from the family of convex functions being closed under pointwise limits; compare [@Jekel2018 Proposition 2.13(1)]). Moreover, $DV = F$.
Let $\nu$ be the probability measure given by $d\nu(x) = (1/Z) e^{-N^2 V(x)}\,dx$. Since $A$ is positive definite, we have $A \geq c$ for some scalar $c > 0$. Because $DV_k(0)$ is bounded in $\norm*{\cdot}_2$ as $k \to \infty$ and $V_k(x) \geq \ip{x, DV_k(0)}_2 + c \norm*{x}_2^2$, we can see using the dominated convergence theorem that $Z_k \to Z$ as $k \to \infty$. It follows again from dominated convergence that $\int \phi\,d\mu_k \to \int \phi\,d\nu$ for every continuous compactly supported $\phi$. Hence, $\nu = \mu$, so $\mu$ is given by the potential $V$.
Log-Sobolev Inequality and Concentration {#subsec:concentration}
----------------------------------------
Log-concave matrix models exhibit concentration of measure as $N \to \infty$ as a consequence of the following classical inequalities.
We say that a measure $\mu$ on ${\mathbb{R}}^m$ satisfies the *log-Sobolev inequality with constant $c$* if for all sufficiently smooth $f$, $$\label{eq:LSI}
\int f^2 \log \frac{f^2}{\int f^2 \,d\mu} \,d\mu \leq 2c \int |\nabla f|^2\,d\mu.$$
We say that a measure $\mu$ on ${\mathbb{R}}^m$ satisfies *Herbst’s concentration inequality with constant $c$* if for all Lipschitz functions $f: {\mathbb{R}}^m \to {\mathbb{R}}$ and $\delta > 0$, we have $E |f(X)| < +\infty$ and $$\label{eq:Herbst1}
P(f(X) - E[f(X)] \geq \delta) \leq e^{-c \delta^2 / 2 \norm{f}_{\operatorname{Lip}}^2}$$ where $X$ is a random variable distributed according to $\mu$. Note that by symmetry this implies $$\label{eq:Herbst2}
P(|f(X) - E[f(X)]| \geq \delta) \leq 2 e^{-c \delta^2 / 2 \norm{f}_{\operatorname{Lip}}^2}.$$
The following theorem is now standard. See for instance [@AGZ2009 §2.3.3 and 4.4.2] and [@BL2000]. To summarize the history, the log-Sobolev inequality was introduced by Gross [@Gross1975]. In the theorem below, (1) is due to Bakry and Emery and (2) is due to unpublished work of Herbst. The application to random matrices was introduced by Guionnet and Zeitouni [@GZ2000].
1. Suppose that $\mu$ is a probability measure on ${\mathbb{R}}^m$ satisfying $d\mu(x) = (1/Z)e^{-V(x)}\,dx$ and suppose that $V(x) - (c/2) |x|^2$ is convex. Then $\mu$ satisfies the log-Sobolev inequality with constant $1/c$.
2. If $\mu$ satisfies the log-Sobolev inequality with constant $1/c$, then it satisfies Herbst’s concentration inequality with constant $c$.
In particular, we have the following consequences for random matrices. Here we use the gradient $Df$ and Hessian $Hf$ with respect to the normalized inner product $\ip{\cdot,\cdot}_2$.
\[cor:matrixLSI\] Suppose that $V: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ satisfies $HV \geq c > 0$ and let $d\mu(x) = (1/Z) e^{-N^2 V(x)}\,dx$. Then $\mu$ satisfies the normalized log-Sobolev inequality $$\label{eq:normalizedLSI}
\int f^2 \log \frac{f^2}{\int f^2 \,d\mu} \,d\mu \leq \frac{2}{N^2 c} \int \norm{Df}_2^2\,d\mu,$$ and hence also satisfies the normalized Herbst concentration inequality $$\label{eq:normalizedHerbst}
P(f(X) - E[f(X)] \geq \delta) \leq e^{-cN^2 \delta^2 / 2 \norm{f}_{\operatorname{Lip}}^2},$$ where $f: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ is Lipschitz and $\norm{f}_{\operatorname{Lip}}$ denotes the Lipschitz norm with respect to $\norm{\cdot}_2$.
\[lem:epsilonnet\] Suppose that $\mu$ is a probability measure on $M_N({\mathbb{C}})_{sa}^m$ satisfying for some constant $c$. Let $f: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})_{sa}$ be Lipschitz with respect to $\norm*{\cdot}_2$. Then we have $$\label{eq:epsilonnet}
P\Bigl( \norm*{f(X) - E[f(X)]}_\infty \geq c^{-1/2} \norm*{f}_{\operatorname{Lip}} (\Theta + \delta) \Bigr) \leq e^{-N \delta^2 / 2}.$$ where $X \sim \mu$ and where $\Theta$ is a universal constant (independent of $N$ and $c$).
First, observe that $\norm{x}_\infty \leq N^{1/2} \norm{x}_2$ for $x \in M_N({\mathbb{C}})_{sa}^m$. In particular, $g(x) = \norm{f(x) - E[f(X)]}_\infty$ is $N^{1/2} \norm{f}_{\operatorname{Lip}}$-Lipschitz with respect to $\norm{\cdot}_2$, and thus $$P (g(X) \geq E[g(X)] + \delta) \leq e^{-cN \delta^2 / 2 \norm{f}_{\operatorname{Lip}}^2},$$ which implies after a change of variables for $\delta$ that $$P(g(X) \geq E[g(X)] + c^{-1/2} \norm{f}_{\operatorname{Lip}} \delta) \leq e^{-N \delta^2 / 2}.$$ Therefore, it suffices to show that for some constant $\Theta$, we have $$\label{eq:expectationofnorm}
E[g(X)] = E[ \norm{f(X) - E[f(X)]}_\infty] \leq \Theta c^{-1/2} \norm{f}_{\operatorname{Lip}}.$$
We may assume without loss of generality that $f$ is self-adjoint since in the general case, $f = (1/2)(f + f^*) + i(1/2i)(f - f^*)$, and each of the terms on the right hand side is Lipschitz. Thus, the self-adjoint case would imply the non-self-adjoint case at the cost of doubling the constant $\Theta$. Now to prove self-adjoint case, we use an “$\epsilon$-net argument” that is well-known in random matrix theory (see [@Tao2012 §2.3.1]). Fix $N$. Let $\{\eta_j\}_{j=1}^J$ be a maximal collection of unit vectors in ${\mathbb{C}}^N$ such that $|\eta_i - \eta_j| \geq 1/3$ for all $i \neq j$. Since this collection is maximal, for every unit vector $\eta$, there exists some $\eta_j$ with $|\eta - \eta_i| < 1/3$. Now if $a \in M_N({\mathbb{C}})_{sa}$, then there is a unit vector with $\norm*{a}_\infty = \ip{\eta, a \eta}$. We may then choose $\eta_j$ with $|\eta - \eta_j| < 1/3$ $$\begin{aligned}
\norm*{a}_\infty &= \ip{\eta, a \eta} \\
&= \ip{\eta_j, a \eta_j} + \ip{\eta_j, a(\eta - \eta_j)}_2 + \ip{\eta - \eta_j, a \eta} \\
&\leq \ip{\eta_j, a \eta_j} + \frac{1}{3} \norm*{a}_\infty + \frac{1}{3} \norm*{a}_\infty,\end{aligned}$$ so that $$\norm*{a}_\infty \leq 3 \max_j \ip{\eta_j, a \eta_j}.$$ Note that the balls $\{B(\eta_j,1/6)\}_{j=1}^J$ in ${\mathbb{C}}^N$ are disjoint and contained in $B(0,7/6)$. Hence, we can estimate the number of vectors by $$J \leq \frac{|B(0,7/6)|}{|B(0,1/6)|} = 7^{2N}.$$
Let $K = \norm{f}_{\operatorname{Lip}}$. For a matrix $a \in M_N({\mathbb{C}})_{sa}$, we have $$|\ip{\eta_i, a \eta_j}| \leq \norm*{a}_\infty \leq N^{1/2} \norm*{a}_2.$$ This implies that $x \mapsto \ip{\eta_j, f(x) \eta_j}$ is $KN^{1/2}$-Lipschitz with respect to $\norm*{\cdot}_2$ and hence $$P\Bigl( \ip{\eta_j, (f(X) - E[f(X)]) \eta_j} \geq \delta \Bigr) \leq e^{-cN \delta^2 / 2K^2}$$ Since $\norm*{a}_\infty \leq 3 \max_{j} \ip{\eta_j, a \eta_j}$, we have $$\begin{aligned}
P \Bigl(\norm*{f(X) - E[f(X)]}_\infty \geq 3 \delta \Bigr) &\leq J e^{-cN \delta^2 / 2} \\
&\leq 7^{2N} e^{-cN \delta^2 / 2K^2}.\end{aligned}$$ Thus, for any $t_0 > 0$, we have $$\begin{aligned}
E[\norm*{f(X) - E[f(X)]}_\infty] &= \int_0^\infty P(\norm{f(X) - E[f(X)]}_\infty \geq t)\,dt \\
&\leq \int_0^{t_0} 1 \,dt + \int_{t_0}^\infty 7^{2N} e^{-cNt^2 / 18 K^2}\,dt \\
&\leq t_0 + \int_{t_0}^\infty 7^{2N} \frac{t}{t_0} e^{-cNt^2 / 18 K^2}\,dt \\
&= t_0 + 7^{2N} \frac{9K^2}{cNt_0} e^{-cNt_0^2 / 18 K^2}.\end{aligned}$$ Now substitute $t_0 = 6c^{-1/2} K (\log 7)^{1/2}$ and obtain with $$\Theta = 6(\log 7)^{1/2} + \frac{9}{6 (\log 7)^{1/2}}.$$ (In fact, for a fixed $N$, we may use $\Theta_N = 6(\log 7)^{1/2} + 9 / 6N(\log 7)^{1/2}$ in the self-adjoint case.)
Functional Calculus and Asymptotic Approximation {#sec:TP}
================================================
In this section, we review the algebra $\operatorname{TrP}_m^1$ of trace polynomials in self-adjoint variables $X_1$, …, $X_m$, as well as a certain completed quotient $\overline{\operatorname{TrP}}_m^1$ of this algebra. The elements of $\overline{\operatorname{TrP}}_m^1$ represent functions that can be applied to *any* tuple of self-adjoint non-commutative random variables $(X_1,\dots,X_m)$ in an $\mathcal{R}^\omega$-embeddable tracial $\mathrm{W}^*$-algebra, and application of these functions will produce every element of $L^2(\mathrm{W}^*(X_1,\dots,X_m))$ (see Proposition \[prop:realizationofoperators\]). These functions are closed under certain algebraic and composition operations. Moreover, they are a natural tool to describe the large $N$ limit of functions on $M_N({\mathbb{C}})_{sa}^m$, which we will apply in the rest of the paper.
The Algebra of Trace Polynomials
--------------------------------
Trace polynomials have been used by several previous authors in the study of deterministic and random matrices; a brief list is [@Razmyslov1974], [@Razmyslov1987], [@Rains1997], [@Sengupta2008], [@Cebron2013], [@DHK2013] (which coined the term “trace polynomial”), [@Kemp2016], [@Kemp2017], [@DGS2016] but they are also used implicitly in many other works. We use the same notation as in our previous paper [@Jekel2018].
We denote by $\operatorname{NCP}_m = {\mathbb{C}}\ip{X_1,\dots,X_m}$ the $*$-algebra of polynomials in $m$ self-adjoint non-commuting variables $X_1$, …, $X_m$.
We denote by $\operatorname{TrP}_m^0$ the $*$-algebra of *scalar-valued trace polynomials*. A formal definition is given in [@Jekel2018]; in short, it is the tensor algebra of the vector space of non-commutative polynomials modulo cyclic symmetry. Informally, this is the commutative $*$-algebra generated by functions of the form $\tau(p(X_1,\dots,X_m))$, where $p$ is a non-commutative polynomial in $X = (X_1,\dots,X_m)$ and $\tau$ is a formal symbol (which stands in for a normalized trace on a von Neumann algebra), where $\tau(p(X))^* = \tau(p(X)^*)$, and where we identify $\tau(p(X) q(X))$ with $\tau(q(X) p(X))$ for all polynomials $p$ and $q$. Thus, $\operatorname{TrP}_m^0$ is spanned as a vector space by elements of the form $\tau(p_1(X)) \dots \tau(p_n(X))$ where $p_1$, …, $p_n \in \operatorname{NCP}_m$.
We denote by $\operatorname{TrP}_m^1$ the $*$-algebra of *operator-valued trace polynomials*. This is the $*$-algebra given formally as $\operatorname{TrP}_m^0 \otimes \operatorname{NCP}_m$. As a vector space, it is spanned by elements of the form $\tau(p_1(X)) \dots \tau(p_n(X)) q(X)$, where $p_1$, …, $p_n$ and $q$ are in $\operatorname{NCP}_m$. More generally, we would denote $\operatorname{TrP}_m^k = \operatorname{TrP}_m^0 \otimes (\operatorname{NCP}_m)^{\otimes k}$, but these spaces will not be needed in this paper.
The *degree* of a trace polynomial is defined as one would expect; see [@Jekel2018 §3.1] for precise explanation.
Suppose that $x_1$, …, $x_m$ are self-adjoint elements of a tracial von Neumann algebra $(\mathcal{M},\tau_0)$. Then elements of $\operatorname{NCP}_m$, $\operatorname{TrP}_m^0$, and $\operatorname{TrP}_m^1$ can be evaluated on $(x_1,\dots,x_m)$ and $\tau_0$ by substituting the operator $x_j$ and the trace $\tau_0$ in place of the formal symbols $X_j$ and $\tau$. More precisely, the evaluation map $\varepsilon_{(x_1,\dots,x_m)}: \operatorname{NCP}_m \to \mathcal{M}$ is the unique $*$-algebra homomorphism that sends $X_j$ to $x_j$. Similarly, the evaluation map $\varepsilon_{(x_1,\dots,x_m)}^0: \operatorname{TrP}_m^0 \to {\mathbb{C}}$ is the unique $*$-algebra homomorphism that sends $\tau(p(X))$ to $\tau_0(\varepsilon_{(x_1,\dots,x_m)}(p))$. Finally, the evaluation map $\varepsilon_{(x_1,\dots,x_m)}^1: \operatorname{TrP}_m^1 \to \mathcal{M}$ is $\varepsilon_{(x_1,\dots,x_m)}^0 \otimes \varepsilon_{(x_1,\dots,x_m)}$, that is, $$\tau(p_1(X)) \dots \tau(p_n(X)) q(X) \mapsto \tau_0(\varepsilon_{(x_1,\dots,x_m)}^0(p_1)) \dots \tau_0(\varepsilon_{(x_1,\dots,x_m)}^0(p_n)) \varepsilon_{(x_1,\dots,x_m)}(q).$$
For the most part, we will abuse notation and denote $f(x) = \varepsilon_{(x_1,\dots,x_m)}(f)$ when $f \in \operatorname{NCP}_m$, and similarly for $f \in \operatorname{TrP}_m^0$ or $f \in \operatorname{TrP}_m^1$. Note in particular that we can consider $(\mathcal{M},\tau_0) = (M_N({\mathbb{C}})_{sa}, \tau_N)$ and thus $f(x)$ is defined for $x \in M_N({\mathbb{C}})_{sa}^m$ and $f \in \operatorname{TrP}_m^0$ or $\operatorname{TrP}_m^1$.
These evaluation maps thus allow us to view $f \in \operatorname{TrP}_m^0$ as a function (or rather a family of functions) $\mathcal{M}_{sa}^m \to {\mathbb{C}}$ for every tracial $\mathrm{W}^*$-algebra $(\mathcal{M},\tau)$ and in particular $M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$ for every $m$. Similarly, every $f \in \operatorname{TrP}_m^1$ defines a function $\mathcal{M}_{sa}^m \to \mathcal{M}$ for every tracial $\mathrm{W}^*$-algebra and in particular a function $M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$ for every $N$.
Functions Approximable by Trace Polynomials
-------------------------------------------
From an analytic viewpoint, we prefer to work with certain separation-completions of $\operatorname{TrP}_m^0$ and $\operatorname{TrP}_m^1$. In [@Jekel2018 §8.1], we sketched several equivalent ways of defining these separation-completions. Here we emphasize their description as functions that can be evaluated on any self-adjoint tuple in $\mathcal{R}^\omega$ (or, as we will see, any $\mathcal{R}^\omega$-embeddable $\mathrm{W}^*$-algebra).
Let $\mathcal{R}$ denote the hyperfinite $\operatorname{II}_1$ factor (tracial $\mathrm{W}^*$-algebra with trivial center) and let $\mathcal{R}^\omega$ be its (tracial $\mathrm{W}^*$-algebra) ultrapower with respect to some fixed free ultrafilter $\omega \in \beta {\mathbb{N}}\setminus {\mathbb{N}}$.
Consider the case of $\operatorname{TrP}_m^0$ first. Let $\mathcal{F}_m^0$ denote the space of functions $(\mathcal{R}^\omega)_{sa}^m \to {\mathbb{C}}$ that are bounded on operator norm balls, equipped with the family of semi-norms $$\norm*{\phi}_{u,R} = \sup \{|\phi(x)|: \norm*{x}_\infty \leq R\}.$$ (Here “$u$” stands for uniform.) This is clearly a Fr[é]{}chet space since the topology is given by the countable family of semi-norms given by taking $R \in {\mathbb{N}}$ (for background on Fr[é]{}chet spaces, see e.g. [@Folland1999 §5.4]). Every $f \in \operatorname{TrP}_m^0$ defines a function $(\mathcal{R}^\omega)_{sa}^m \to {\mathbb{C}}$ that is a bounded an operator norm balls. In other words, evaluation produces a map $\operatorname{TrP}_m^0 \to \mathcal{F}_m^0$. We denote by $\overline{\operatorname{TrP}}_m^0$ the closure of the image of this map in $\mathcal{F}_m$. In other words, $\overline{\operatorname{TrP}}_m^0$ is the space of functions $(\mathcal{R}^\omega)_{sa}^m \to {\mathbb{C}}$ that can be approximated uniformly on operator-norm balls by trace polynomials.
This space was denoted as $\mathcal{T}_m^0$ in our earlier paper [@Jekel2018]. The notation $\overline{\operatorname{TrP}}_m^0$ is slightly abusive since we have not shown that the map $\operatorname{TrP}_m^0 \to \mathcal{F}_m^0$ is injective (and perhaps it is not). However, we will still use the notation $\overline{\operatorname{TrP}}_m^0$ since it indicates the connection with trace polynomials.
Earlier, we saw that it makes sense to evaluate a trace polynomial $f$ on any self-adjoint tuple $(x_1,\dots,x_m)$ in a tracial von Neumann algebra. In fact, $f(x_1,\dots,x_m)$ makes sense for every $f \in \overline{\operatorname{TrP}}_m^0$ when $x_1$, …, $x_m$ come from a tracial von Neumann algebra that embeds into $\mathcal{R}^\omega$. To see this, suppose $(\mathcal{M},\tau)$ admits a normal trace-preserving embedding $\iota: \mathcal{M} \to \mathcal{R}^\omega$. Then we define $f(x_1,\dots,x_m) = f(\iota(x_1),\dots,\iota(x_m))$. This is independent of the choice of trace-preserving embedding if $f$ is a trace polynomial, and hence it must also be independent of the choice of embedding when $f$ is in $\overline{\operatorname{TrP}}_m^0$.
A similar separation-completion can be defined for $\operatorname{TrP}_m^1$. Indeed, let $\mathcal{F}_m^1$ be the set of functions $\phi: (\mathcal{R}^\omega)_{sa}^m \to L^2(\mathcal{R}^\omega)$ such that $$\norm*{\phi}_{u,R} := \sup \{\norm*{\phi(x)}_2: \norm*{x}_\infty \leq R\}$$ is finite for each $R$. Again, this is a Fr[é]{}chet space. Through the evaluation map, every trace polynomial defines an element of $\mathcal{F}_m^1$ and hence there is a linear map $\operatorname{TrP}_m^1 \to \mathcal{F}_m^1$. We define $\overline{\operatorname{TrP}}_m^1$ to be the closure of the image of this map in $\mathcal{F}_m^1$.
Similar to the scalar-valued case, we can define evaluation of $f \in \overline{\operatorname{TrP}}_m^1$ for tuples in an $\mathcal{R}^\omega$-embeddable tracial $\mathrm{W}^*$-algebra $(\mathcal{M},\tau)$ by using any trace preserving embedding $\iota: \mathcal{M} \to \mathcal{R}^\omega$. Indeed, let $x_1,\dots,x_m \in \mathcal{M}_{sa}$. Clearly, for $f \in \operatorname{TrP}_m^1$, we have $f(\iota(x_1),\dots,\iota(x_m)) \in \iota(\mathcal{M}) \subseteq \iota(L^2(\mathcal{M}))$ where the latter is defined by extending $\iota$ to a map $L^2(\mathcal{M}) \to L^2(\mathcal{R}^\omega)$. Since this holds for $f \in \operatorname{TrP}_m^1$, then by taking limits, we have $f(\iota(x_1),\dots,\iota(x_m)) \in \iota(L^2(\mathcal{M}))$ for all $f \in \overline{\operatorname{TrP}}_m^1$. Therefore, we may define $f(x_1,\dots,x_m)$ by $\iota(f(x_1,\dots,x_m)) = f(\iota(x_1),\dots,\iota(x_m))$. Then one can check this is independent of the choice of embedding similarly as we did in the case of $\overline{\operatorname{TrP}}_m^0$.
Because the spaces $\overline{\operatorname{TrP}}_m^j$ used here are non-standard, let us briefly describe their relationship to other more familiar ideas. Recall that $\Sigma_{m,R}$ denotes the space of non-commutative laws of $m$-tuples with operator norms bounded by $R$. We denote by $\Sigma_{m,R}^{\operatorname{app}}$ the subspace of laws that can be realized by $m$-tuples in $\mathcal{R}^\omega$, and $\Sigma_m^{\operatorname{app}} = \bigcup_{R > 0} \Sigma_{m,R}^{\operatorname{app}}$. Then we showed in [@Jekel2018 Lemma 8.2] that $\overline{\operatorname{TrP}}_m^0$ consists of functions $\Sigma_m^{\operatorname{app}} \to {\mathbb{C}}$ such that the restriction to $\Sigma_{m,R}^{\operatorname{app}}$ is continuous for each $R$. One could think of this alternatively as an inverse limit of $C(\Sigma_{m,R}^{\operatorname{app}})$ over the directed system of restriction maps $C(\Sigma_{m,R'}^{\operatorname{app}}) \to C(\Sigma_{m,R}^{\operatorname{app}})$ for $R' > R$.
The spaces $\operatorname{TrP}_m^0$ and $\overline{\operatorname{TrP}}_m^0$ also arise naturally in the study of model theory of tracial von Neumann algebras introduced in [@FHS2013; @FHS2014; @FHS2014b]. To avoid some of the technical complexities of sorts, we follow the definitions in [@FHS2014] where the language has multiple domains of quantification for each sort (and thus we can get away with fewer sorts), and in which formulas are obtained by applying continuous functions ${\mathbb{R}}^n \to {\mathbb{R}}$ to atomic formulas (rather than functions defined on some compact set). For tracial von Neumann algebra $(M,\tau)$, the language includes (though this list is not exhaustive) a sort representing $M$ with domains of quantification for each operator norm ball of radius $n \in {\mathbb{N}}$, a special relation-like symbol $d(x,y)$ for the distance $\norm{x - y}_2$, a relation symbol for the trace $\tau(x)$, and function symbols for the adjoint, addition, and multiplication.
Now $\tau(p(x_1,\dots,x_m))$ is an example of a atomic formula (or strictly speaking, its real and imaginary parts are basic formulas). Similarly, $\tau(p(\operatorname{Re}(x_1),\dots, \operatorname{Re}(x_m))$ is an atomic formula, where $\operatorname{Re}(x_j) = (x_j + x_j^*) / 2$. Since the elements of $\operatorname{TrP}_m^0$ is obtained by multiplying formulas such as $\tau(p)$, we see that $f(\operatorname{Re}(x_1),\dots,\operatorname{Re}(x_m))$ is a quantifier-free formula for every $f \in \operatorname{TrP}_m^0$. Moreover, the supremum of $|f(\operatorname{Re}(x_1),\dots,\operatorname{Re}(x_m))$ over $\{x: \norm{x_j} \leq R_j\}$ is the same as the supremum of $f$ over $\{x: x_j = x_j^*, \norm{x_j} \leq R_j\}$. The limiting objects $\overline{\operatorname{TrP}}_m^0$ (evaluated on the real parts of operators) are thus uniform limits of quantifier-free formulas on each domain of quantification for every $\mathcal{R}^\omega$-embeddable tracial von Neumann algebra, that is, they are “quantifier-free definable predicates” relative to the theory of $\mathcal{R}^\omega$-embeddable tracial von Neumann algebras. Conversely, since $\overline{\operatorname{TrP}}_m^0$ is closed under the operation $(f_1,\dots,f_n) \mapsto \phi(f_1,\dots,f_n)$ for $\phi: {\mathbb{C}}^n \to {\mathbb{C}}$ continuous, every quantifier-free definable predicate $f$ satisfying $f(x_1,\dots,x_m) = f(\operatorname{Re}(x_1),\dots,\operatorname{Re}(x_m))$ is an element of $\overline{\operatorname{TrP}}_m^0$.
The elements of $\overline{\operatorname{TrP}}_m^1$, evaluated on the real parts of operators, may be viewed similarly as certain “quantifier-free definable functions” relative to the theory of $\mathcal{R}^\omega$-embeddable tracial von Neumann algebras, meaning that $\norm{f(x) - y}_2^2$ is a quantifier-free definable predicate | actually, for technical reasons a definable function is required to map an operator norm ball into an operator norm ball, so the last statement only applies if we assume our function $f \in \overline{\operatorname{TrP}}_m^1$ has this property (but it turns out that such functions exist in abundance in $\overline{\operatorname{TrP}}_m^1$; see Proposition \[prop:realizationofoperators\] and Proposition \[prop:operatornormestimate\]). Alternatively, in order to deal with functions with codomain $L^2$, we must first modify the language by adding another sort for $L^2(M)$, with domains of quantification corresponding to $L^2$-balls, which will act as the target space of the functions in $\overline{\operatorname{TrP}}_m^1$.
The quantifier-free nature of these formulas is a model-theoretic heuristic for why they behave well under limits in non-commutative law (hence describing the large $N$ limits of random matrix models). In fact,[@Jekel2018 Proposition 6.28] re-expresses a formula given by quantifiers in a quantifier-free way in order to get behavior under limits. There, we studied the inf-convolution $(Q_t V)(x) = \inf_y [V(y) - (1/2t) \norm{x - y}_2^2]$ for self-adjoint tuples $x$ and $y$. If $V \in \operatorname{TrP}_m^0$, then for each $ > 0$, $$W_{t,R}(\operatorname{Re}(x_1),\dots, \operatorname{Re}(x_m)) := \inf_{\norm{y} \leq R} \left[V(\operatorname{Re}(y_1),\dots,\operatorname{Re}(y_m)) + \frac{1}{2t} \norm{\operatorname{Re}(x) - \operatorname{Re}(y)}_2^2 \right]$$ is a formula in the language of tracial von Neumann algebras whose definition involves the quantifier $\inf$. But if $V$ is convex and semi-concave and $DV \in (\overline{\operatorname{TrP}}_m^1)_{sa}^m$, then the self-adjoint tuple $y$ where the infimum $$W_t(\operatorname{Re}(x_1),\dots, \operatorname{Re}(x_m)) := \inf_{\norm{y} \leq R} \left[V(\operatorname{Re}(y_1),\dots,\operatorname{Re}(y_m)) + \frac{1}{2t} \norm{\operatorname{Re}(x) - \operatorname{Re}(y)}_2^2 \right]$$ is achieved can be evaluated as the limit of a fixed-point iteration using functions from $(\overline{\operatorname{TrP}}_m^1)_{sa}^m$, and hence $y = \phi(\operatorname{Re}(x))$ for some $\phi \in (\overline{\operatorname{TrP}}_m^1)_{sa}^m$ (see [@Jekel2018 Proposition 6.28]). Moreover, it follows from the results in [@Jekel2018] that $\phi$ is Lipschitz in $\norm{\cdot}_2$, and thus in light of Proposition \[prop:operatornormestimate\] below, $\phi$ is bounded in operator norm on operator norm balls. So $\phi(\operatorname{Re}(x))$ is quantifier-free definable function. We can also conclude that $W_{t,R} \to W_t$ as $R \to \infty$ uniformly on operator norm balls, so $W_t$ is a definable formula (allowing quantifiers). But then because $$W_{t,R}(x) =V(\phi(x)) - \frac{1}{2t} \norm{\operatorname{Re}(x) - \phi(x)}_2^2,$$ we conclude that $W_t$ is in fact a *quantifier-free* definable predicate.
On the other hand, without the ability to eliminate the quantifier like this, we could not hope for $Q_t V$ to behave so well for the large $N$ limit of random matrix models. Indeed, for $Q_tV(x)$ to depend continuously on the non-commutative law $\lambda_x$ for $x$ in each operator norm ball, it must be in $\operatorname{TrP}_m^0$ by the last remark, and hence it is a quantifier-free definable predicate.
Many of the properties shown in the next section about operations on $\operatorname{TrP}_m^0$ and $\operatorname{TrP}_m^1$ are natural from the model theoretic viewpoint, but we sketch self-contained justifications nonetheless.
Asymptotic Approximation for Functions of Matrices {#subsec:AATP}
--------------------------------------------------
Our earlier work introduced asymptotic approximability by trace polynomials for a sequence of functions on $M_N({\mathbb{C}})_{sa}^m$, which is a precise description of good asymptotic behavior as $N \to \infty$ suitable for free probabilistic analysis in the limit.
\[def:AATP\] Let $\phi^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$. We say that $\{\phi^{(N)}\}$ is *asymptotically approximable by trace polynomials* if for every $R > 0$ and $\epsilon > 0$, there exists $f \in \operatorname{TrP}_m^0$ such that $$\limsup_{N \to \infty} \sup_{\substack{x \in M_N({\mathbb{C}})_{sa}^m \\ \norm*{x}_\infty \leq R}} |\phi^{(N)}(x) - f(x)| \leq \epsilon.$$ Similarly, for matrix-valued functions $\phi^{(N)}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$, we say that $\{\phi^{(N)}\}$ is *asymptotically approximable by trace polynomials* if for every $R > 0$ and $\epsilon > 0$, there exists $f \in \operatorname{TrP}_m^1$ such that $$\limsup_{N \to \infty} \sup_{\substack{x \in M_N({\mathbb{C}})_{sa}^m \\ \norm*{x}_\infty \leq R}} \norm*{\phi^{(N)}(x) - f(x)}_2 \leq \epsilon.$$
It will be convenient to denote $$\norm*{\phi^{(N)}}_{u,R}^{(N)} = \sup_{\substack{x \in M_N({\mathbb{C}})_{sa}^m \\ \norm*{x}_\infty \leq R}} |\phi(x)|$$ in the scalar-valued case and similarly for the matrix-valued case with $\norm*{\phi(x)}_2$ rather than $|\phi(x)|$. Thus, for instance, the preceding definition says that there exists a trace polynomial $f$ with $$\limsup_{N \to \infty} \norm*{\phi^{(N)} - f}_{u,R}^{(N)} < \epsilon.$$
Moreover, it is implicit from our discussion in [@Jekel2018 §8.1] that if $\phi^{(N)}$ is asymptotically approximable by trace polynomials, then it will be *asymptotic to* some $f \in \overline{\operatorname{TrP}}_m^0$ or $\overline{\operatorname{TrP}}_m^1$ in the following sense.
\[def:AA\] Let $\phi^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$ or $M_N({\mathbb{C}})$ respectively, and let $f \in \overline{\operatorname{TrP}}_m^0$ or $\overline{\operatorname{TrP}}_m^1$ respectively. Then we say that $\{\phi^{(N)}\}$ is *asymptotic to $f$*, or $\phi^{(N)} \rightsquigarrow f$ if for every $R > 0$, $$\lim_{N \to \infty} \norm*{\phi^{(N)} - f}_{u,R}^{(N)} = 0.$$ Similarly, if $\phi^{(N)}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$ and $f \in \overline{\operatorname{TrP}}_m^1$, we make the same definitions with $|\phi^{(N)}(x) - f(x)|$ replaced by $\norm*{\phi^{(N)}(x) - f(x)}_2$.
\[lem:AATP\] Let $\phi^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$ (respectively, $\to M_N({\mathbb{C}})$). Then $\phi^{(N)}$ is asymptotically approximable by trace polynomials if and only if there exists $f \in \overline{\operatorname{TrP}}_m^0$ (respectively, $f \in \overline{\operatorname{TrP}}_m^1$) such that $\phi^{(N)} \rightsquigarrow f$. Moreover, $\norm*{f}_{u,R} = \lim_{N \to \infty} \norm*{\phi^{(N)}}_{u,R}^{(N)}$ for each $R$.
We record the proof only for the case of scalar-valued functions, since the proof for operator-valued case is identical with minor changes of notation. Suppose that $\{\phi^{(N)}\}$ is asymptotically approximable by trace polynomials. Then there exists a sequence $\{f_k\}$ of trace polynomials such that for every $R > 0$, $$\lim_{k \to \infty} \limsup_{N \to \infty} \sup_{\substack{x \in M_N({\mathbb{C}})_{sa}^m \\ \norm*{x}_\infty \leq R}} |\phi^{(N)}(x) - f_k(x)| = 0.$$ As in [@Jekel2018 Lemma 8.1], if $g \in \operatorname{TrP}_m^0$, then $$\sup_{\substack{x \in (\mathcal{R}^\omega)_{sa}^m \\ \norm*{x}_\infty \leq R}} |g(x)| = \sup_{N \in {\mathbb{N}}} \sup_{\substack{x \in M_N({\mathbb{C}})_{sa}^m \\ \norm*{x}_\infty \leq R}} |g(x)| = \limsup_{N \to \infty} \sup_{\substack{x \in M_N({\mathbb{C}})_{sa}^m \\ \norm*{x}_\infty \leq R}} |g(x)|$$ which implies that $$\norm*{g}_{u,R} = \limsup_{N \to \infty} \norm*{g}_{u,R}^{(N)}.$$ Applying this to $g = f_j - f_k$, we obtain from the triangle inequality $$\norm*{f_j - f_k}_{u,R} \leq \limsup_{N \to \infty} \norm*{f_j(x) - \phi^{(N)}}_{u,R}^{(N)} + \limsup_{N \to \infty} \norm*{\phi^{(N)} - f_k}_{u,R}^{(N)},$$ and hence $f_k$ is Cauchy with respect to $\norm*{\cdot}_{u,R}$ for each $R > 0$. Hence, $f_k$ converges to some $f \in \overline{\operatorname{TrP}}_m^0$. By similar use of the triangle inequality, $$\limsup_{N \to \infty} \norm*{\phi^{(N)} - f}_{u,R}^{(N)} \leq \norm*{f - f_k}_{u,R} + \limsup_{N \to \infty} \norm*{\phi^{(N)} - f_k}_{u,R}^{(N)}.$$ Hence, $\phi^{(N)} \rightsquigarrow f$.
Conversely, suppose that $\phi^{(N)} \rightsquigarrow f \in \overline{\operatorname{TrP}}_m^0$. Choose $f_k \in \operatorname{TrP}_m^0$ such that $\norm*{f_k - f}_{u,R} \to 0$ for every $R$. Then $$\limsup_{N \to \infty} \norm*{\phi^{(N)}(x) - f_k}_{u,R}^{(N)} \leq \norm*{f - f_k}_{u,R} + \limsup_{N \to \infty} \norm*{\phi^{(N)} - f}_{u,R}^{(N)} = \norm*{f - f_k}_{u,R}.$$ Hence, it follows that $\{\phi^{(N)}\}$ is asymptotically approximable by trace polynomials, namely the polynomials $\{f_k\}$.
We leave the proof of the last claim that $\norm*{f}_{u,R} = \lim_{N \to \infty} \norm*{\phi^{(N)}}_{u,R}^{(N)}$ to the reader.
If $\phi^{(N)}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})_{sa}$ and $\{\phi^{(N)}\}$ is asymptotically approximable by trace polynomials, then we can asymptotically approximate it using *self-adjoint* trace polynomials. Indeed, if $$\limsup_{N \to \infty} \norm*{\phi^{(N)} - f}_{u,R}^{(N)} \leq \epsilon,$$ then the same holds with $f$ replaced by $(1/2)(f + f^*)$. Similarly, if $\phi^{(N)}(x)$ is self-adjoint and $\phi^{(N)} \rightsquigarrow f \in \overline{\operatorname{TrP}}_m^1$, then $f$ must be self-adjoint.
Definitions \[def:AATP\] and \[def:AA\] and Lemma \[lem:AATP\] extend naturally to tuples $f = (f_1,\dots,f_n) \in (\overline{\operatorname{TrP}}_m^1)^n$ and $\phi^{(N)} = (\phi_1^{(N)}, \dots, \phi_n^{(N)}): M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})^n$. We shall apply them to tuples without further comment in the rest of the paper.
Algebra, Composition, and Limits
--------------------------------
\[lem:algebra\] $\overline{\operatorname{TrP}}_m^0$ is an algebra and $\overline{\operatorname{TrP}}_m^1$ is a module over $\overline{\operatorname{TrP}}_m^0$. Also, if $f, g \in \overline{\operatorname{TrP}}_m^1$, then $\tau(fg) \in \overline{\operatorname{TrP}}_m^0$. Moreover, suppose that $\phi^{(N)}, \phi^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$ and $f^{(N)}, g^{(N)}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$ are asymptotically approximable, and $\phi^{(N)} \rightsquigarrow \phi$, $\psi^{(N)} \rightsquigarrow \psi$, $f^{(N)} \rightsquigarrow f$, and $g^{(N)} \rightsquigarrow g$. Then we have $$\begin{aligned}
\phi^{(N)} + \psi^{(N)} &\rightsquigarrow \phi + \psi \\
\phi^{(N)} \psi^{(N)} & \rightsquigarrow \phi \psi \\
f^{(N)} + g^{(N)} &\rightsquigarrow f + g \\
\phi^{(N)} f^{(N)} &\rightsquigarrow \phi f \\
\tau_N(f^{(N)} g^{(N)}) & \rightsquigarrow \tau(fg).\end{aligned}$$
Since the proofs of all the statements are straightforward and similar to each other, we will only explain how to show that if $\phi \in \overline{\operatorname{TrP}}_m^0$ and $f \in \overline{\operatorname{TrP}}_m^1$, then $\phi f \in \overline{\operatorname{TrP}}_m^1$ and that if $\phi^{(N)} \rightsquigarrow \phi$ and $f^{(N)} \rightsquigarrow f$, then $\phi^{(N)} f^{(N)} \rightsquigarrow \phi f$.
First, note that $\phi f$ is well-defined as a function on $(\mathcal{R}^\omega)_{sa}^m$ by multiplying the scalar $\phi(x)$ times the vector $f(x)$ for each $x \in (\mathcal{R}^\omega)_{sa}^m$, and also clearly $\norm*{\phi f}_{u,R} \leq \norm*{\phi}_{u,R} \norm*{f}_{u,R}$. To show that $\phi f \in \overline{\operatorname{TrP}}_m^1$, it suffices to show that for every $\epsilon > 0$ and $R > 0$, the function $\phi f$ can be approximated by an element of $\operatorname{TrP}_m^1$ with respect to $\norm*{\cdot}_{u,R}$ with error less than $\epsilon$. We first choose $h \in \operatorname{TrP}_m^1$ such that $$\norm*{h - f}_{u,R} \norm*{\phi}_{u,R} < \frac{\epsilon}{2}.$$ Then we choose $\theta \in \operatorname{TrP}_m^0$ such that $$\norm*{\theta - \phi}_{u,R} \norm*{h}_{u,R} < \frac{\epsilon}{2},$$ and we conclude with the routine observation that $$\begin{aligned}
\norm*{\theta h - \phi f}_{u,R} &\leq \norm*{(\theta - \phi) \cdot h}_{u,R} + \norm*{\phi \cdot (h - f)}_{u,R} \\
&\leq \norm*{\theta - \phi}_{u,R} \norm*{h}_{u,R} + \norm*{\phi}_{u,R} \norm*{h - f}_{u,R} < \epsilon.\end{aligned}$$
Next, to show $\phi^{(N)} f^{(N)} \rightsquigarrow \phi f$, first observe that $$M := \sup_N \norm*{\phi^{(N)}}_{u,R}^{(N)} < +\infty.$$ Then $$\begin{aligned}
\norm*{\phi^{(N)} f^{(N)} - \phi f}_{u,R}^{(N)} &\leq \norm*{\phi^{(N)} \cdot (f^{(N)} - f)}_{u,R}^{(N)} + \norm*{(\phi^{(N)} - \phi) \cdot f}_{u,R}^{(N)} \\
&\leq M \norm*{f^{(N)} - f}_{u,R}^{(N)} + \norm*{\phi^{(N)} - \phi}_{u,R}^{(N)} \norm*{f}_{u,R},\end{aligned}$$ which implies that $\norm*{\phi^{(N)} f^{(N)} - \phi f}_{u,R}^{(N)} \to 0$.
In addition to their algebraic structure, functions $(\mathcal{R}^\omega)_{sa}^m \to (\mathcal{R}^\omega)_{sa}^n$ given by trace polynomials are closed under composition. It turns out that self-adjoint tuples from $\overline{\operatorname{TrP}}_m^1$ are closed under composition under the assumption of $\norm*{\cdot}_2$-uniform continuity of the “outside” function (Lemma \[lem:composition\] below).
We say that $f \in \overline{\operatorname{TrP}}_m^1$ is *$\norm*{\cdot}_2$-uniformly continuous* if for every $\epsilon > 0$, there exists $\delta > 0$ such that $$\forall x, y \in (\mathcal{R}^\omega)_{sa}^m, \quad \norm*{x - y}_2 < \delta \implies \norm*{f(x) - f(y)}_2 < \epsilon.$$ Furthermore, we say $f \in \overline{\operatorname{TrP}}_m^1$ is *$\norm*{\cdot}_2$-Lipschitz* if $\norm*{f(x) - f(y)}_2 \leq K \norm*{x - y}$ for some constant $K$, which is an important special case of uniform continuity. We denote the minimum such constant by $\norm*{f}_{\operatorname{Lip}}$. We make the analogous definitions for $f \in \overline{\operatorname{TrP}}_m^0$.
\[obs:continuousextension\] If $f$ is a function from $(\mathcal{R}^\omega)_{sa}^m$ to $\mathcal{R}^\omega$ or ${\mathbb{C}}$ that is $\norm*{\cdot}_2$-uniformly continuous, then it has a unique continuous extension to $L^2(\mathcal{R}^\omega)_{sa}^m$, which is also $\norm*{\cdot}_2$-uniformly continuous. Similarly, if $f$ is Lipschitz on $(\mathcal{R}^\omega)_{sa}^m$, then the extension is also Lipschitz.
\[lem:limituniformlycontinuous\] Suppose that $f^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$ or $M_N({\mathbb{C}})$ and that $f^{(N)} \rightsquigarrow f$. If $f^{(N)}$ is $\norm{\cdot}_2$-uniformly continuous with respect to some modulus of continuity independent of $N$, then $f$ is $\norm{\cdot}_2$-uniformly continuous on $(\mathcal{R}^\omega)_{sa}^m$ with the same modulus of continuity.
Let us only explain the operator-valued case where $f^{(N)}$ is $M_N({\mathbb{C}})$-valued and $f \in \overline{\operatorname{TrP}}_m^1$, since the scalar-valued case is easier. We define scalar-valued functions of $2m$ variables by $F^{(N)}(x,y) = \norm{f^{(N)}(x) - g^{(N)}(y)}_2^2$ and $F(x,y) = \norm{f(x) - f(y)}_2^2$. By Lemma \[lem:algebra\], we have $F^{(N)} \rightsquigarrow F \in \overline{\operatorname{TrP}}_{2m}^0$.
Let $\epsilon(\delta)$ be a common modulus of continuity for $f^{(N)}$. Let $x$ and $y \in (\mathcal{R}^\omega)_{sa}^m$. Then we may embed $\mathrm{W}^*(x,y)$ into $(\mathcal{M},\tau) := \prod_{N \to \omega} (M_N({\mathbb{C}}),\tau_N)$, that is the tracial $\mathrm{W}^*$-ultraproduct of matrices. There exist tuples $x^{(N)}$ and $y^{(N)}$ of $N \times N$ matrices such that $x = \{x^{(N)}\}_{N \in {\mathbb{N}}}$ and $y = \{y^{(N)}\}_{N \in {\mathbb{N}}}$ in the ultraproduct and also $\norm{x^{(N)}}_\infty \leq \norm{x}_\infty$ and $\norm{y^{(N)}}_\infty \leq \norm{y}_\infty$. Observe that $$F(x,y) = \lim_{N \to \omega} F(x^{(N)},y^{(N)}).$$ (This equality holds for trace polynomials and hence holds for all functions in $\overline{\operatorname{TrP}}_{2m}^0$ by approximation.) On the other hand, we also have for $R > \max(\norm{x}_\infty, \norm{y}_\infty)$ that $$|F^{(N)}(x^{(N)}, y^{(N)}) - F(x^{(N)},y^{(N)})| \leq \norm{F - F^{(N)}}_{u,R}^{(N)} \to 0.$$ Therefore, $$\norm{f(x) - f(y)}_2 = \lim_{N \to \omega} \norm{f^{(N)}(x^{(N)}) - f^{(N)}(y^{(N)})}_2 \leq \lim_{N \to \omega} \epsilon( \norm{x^{(N)} - y^{(N)}}_2 ) \leq \epsilon (\norm{x - y}_2),$$ since $\norm{x^{(N)} - y^{(N)}}_2 \to \norm{x - y}_2$.
\[lem:composition\] Let $j = 0$ or $1$. Let $f \in \overline{\operatorname{TrP}}_m^j$ be $\norm*{\cdot}_2$-uniformly continuous and let $g = (g_1,\dots,g_m) \in (\overline{\operatorname{TrP}}_n^1)_{sa}^m$.
1. Then $f \circ g$ is a well-defined function on $(\mathcal{R}^\omega)_{sa}^n$, and it is in $\overline{\operatorname{TrP}}_n^j$.
2. If $g$ is also $\norm*{\cdot}_2$-uniformly continuous, then so is $f \circ g$.
3. Suppose $f^{(N)}$ is a function on $M_N({\mathbb{C}})_{sa}^m$ and $g^{(N)}: M_N({\mathbb{C}})_{sa}^n \to M_N({\mathbb{C}})_{sa}^m$ such that $f^{(N)} \rightsquigarrow f$ and $g^{(N)} \rightsquigarrow g$. Also, suppose that $f^{(N)}$ is $\norm*{\cdot}_2$-uniformly continuous with the modulus of continuity also uniform in $N$. Then $f^{(N)} \circ g^{(N)} \rightsquigarrow f \circ g$.
\(1) Because $f$ extends to a function on $L^2(\mathcal{R}^\omega)_{sa}^m$, we can define $f \circ g$. Now let us show $f \circ g \in \overline{\operatorname{TrP}}_m^j$. Choose $\epsilon > 0$ and $R > 0$. By uniform continuity of $f$, there exists a $\delta > 0$ such that $\norm*{x - y}_2 < \delta$ implies $|f(x) - f(y)|$ or $\norm*{f(x) - f(y)}_2 < \epsilon/2$ (for $j = 0$ or $1$ respectively). Now choose $\tilde{g} \in (\operatorname{TrP}_n^1)_{sa}^m$ such that $\norm*{\tilde{g} - g}_{u,R} < \delta$, and hence $$\norm*{f \circ g - f \circ \tilde{g}}_{u,R} < \frac{\epsilon}{2}.$$ Because $\tilde{g}$ is a trace polynomial, there is some $R'$ such that $\norm*{x}_\infty \leq R$ implies $\norm*{\tilde{g}}_\infty \leq R'$. Choose $\tilde{f} \in \overline{\operatorname{TrP}}_m^j$ with $\norm*{\tilde{f} - f}_{u,R'} < \epsilon / 2$, and hence $$\norm*{f \circ \tilde{g} - \tilde{f} \circ \tilde{g}}_{u,R} < \frac{\epsilon}{2}.$$ Then altogether we have $\norm*{f \circ g - \tilde{f} \circ \tilde{g}}_{u,R} < \epsilon$.
\(2) This is immediate.
\(3) This is similar to the proof of (1). Fix $R > 0$ and $\epsilon > 0$. Choose $\delta > 0$ such that $\norm*{x - y}_2 < \delta$ implies $|f(x) - f(y)|$ or $\norm*{f(x) - f(y)}_2 < \epsilon / 2$ and such that the same holds for $f^{(N)}$ as well. Let $\tilde{g} \in (\operatorname{TrP}_n^1)_{sa}^m$ such that $\norm*{\tilde{g} - g}_{u,R} < \delta$. Note that for sufficiently large $N$, we have $\norm*{g^{(N)} - \tilde{g}}_{u,R}^{(N)} < \delta$ and hence $$\norm*{f^{(N)} \circ g^{(N)} - f^{(N)} \circ \tilde{g}}_{u,R}^{(N)} < \epsilon /2.$$ Then let $R'$ and $\tilde{f}$ be as in (1). Then for sufficiently large $N$, we have $$\norm*{f^{(N)} \circ \tilde{g} - \tilde{f} \circ \tilde{g}}_{u,R}^{(N)} < \frac{\epsilon}{2},$$ so overall $$\norm*{f^{(N)} \circ g^{(N)} - \tilde{f} \circ \tilde{g}}_{u,R}^{(N)} < \epsilon, \qquad \norm*{f \circ g - \tilde{f} \circ \tilde{g}}_{u,R} < \epsilon,$$ so that$\norm*{f^{(N)} \circ g^{(N)} - f \circ g}_{u,R}^{(N)} < 2 \epsilon$ for large enough $N$.
Moreover, asymptotically approximable sequences are closed under limits in an appropriate sense.
\[lem:limits\] Let $f_k^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$ or to $M_N({\mathbb{C}})$ for $k$ and $N \in {\mathbb{N}}$. Suppose that $f_k^{(N)} \rightsquigarrow f_k$ in $\overline{\operatorname{TrP}}_m^j$ for each $k$, and that $$\label{eq:limithypothesis}
\lim_{k \to \infty} \limsup_{N \to \infty} \norm*{f_k^{(N)} - f^{(N)}}_{u,R} = 0 \text{ for all } R.$$ Then $f_k$ converges in $\overline{\operatorname{TrP}}_m^j$ to some $f$, and we have $f^{(N)} \rightsquigarrow f$.
Note that $$\norm*{f_k - f_j}_{u,R} = \lim_{N \to \infty} \norm*{f_k^{(N)} - f_j^{(N)}}_{u,R}^{(N)}$$ Then because of our assumption , we see that $\{f_k\}_{k \in {\mathbb{N}}}$ is Cauchy with respect to $\norm*{\cdot}_{u,R}$ for each $R$. Thus, $f_k$ converges to some $f$. Then to show that $f^{(N)} \rightsquigarrow f$ is a routine argument.
Functional Calculus and Operator Norm Bounds
--------------------------------------------
Now we will show that every element of $L^2(\mathrm{W}^*(x_1,\dots,x_m))$ can be expressed as $f(x_1,\dots,x_m)$ for some $f \in \overline{\operatorname{TrP}}_m^1$. In fact, we can arrange that $f$ can be approximated uniformly by Lipschitz functions. It will be convenient to define the uniform norm $$\norm*{f}_u = \sup_{R > 0} \norm*{f}_{u,R} = \sup_{x \in (\mathcal{R}^\omega)_{sa}^m} \norm*{f(x)}_2,$$ and we make the same definition for $\norm*{f}_u^{(N)}$ where the supremum is instead taken over $x \in M_N({\mathbb{C}})_{sa}^m$.
\[prop:realizationofoperators\] Let $x_1$, …, $x_m$ be self-adjoint variables which generate a tracial $\mathrm{W}^*$-algebra $(\mathcal{M},\tau)$ that is embeddable into $\mathcal{R}^\omega$. Let $z \in L^2(\mathcal{M},\tau)$.
1. There exists a $\norm*{\cdot}_2$-uniformly continuous $f \in \overline{\operatorname{TrP}}_m^1$ such that $Z = f(x_1, \dots, x_m)$.
2. The $f$ in (1) can be chosen so that there are $\norm*{\cdot}_2$-Lipschitz functions $f_k \in \overline{\operatorname{TrP}}_m^1$ such that $\norm*{f_k - f}_u \to 0$.
3. If $z \in {\mathbb{C}}\ip{x_1,\dots,x_m}$, then $f$ can be chosen to be $\norm*{\cdot}_2$-Lipschitz.
We use the following auxiliary observation. Here $\Sigma_{m,R}$ will denote the space of non-commutative laws for an $m$-tuple of operators with operator norm $\leq R$. We equip $\Sigma_{m,R}$ with the topology of convergence in moments. Recall that $\Sigma_{m,R}$ is compact, separable, and metrizable. In [@Jekel2018 Lemma 8.2], we noted the relationship between $\overline{\operatorname{TrP}}_m^0$ and continuous functions on $\Sigma_{m,R}$ for each $R$. This same idea motivates the proof of the next lemma.
Let $\mu \in \Sigma_{m,R}$ and let $\mathcal{U}$ be a neighborhood of $\mu$, and let $\epsilon > 0$. Then there exists a trace polynomial $f$ such that $$\mu(f) = 1 \qquad
0 \leq \nu(f) \leq \mathbf{1}_{\nu \in \mathcal{U}} + \epsilon \text{ for } \nu \in \Sigma_{m,R}.$$
By Urysohn’s lemma, there exists a continuous function $F: \Sigma_{m,R} \to [0,1]$ such that $F(\mu) = 1$ and $F(\nu) = 0$ for $\nu \not \in \mathcal{U}$. The functions $\Sigma_{m,R} \to {\mathbb{C}}$ of the form $\mu \mapsto \mu(f)$ for $f \in \operatorname{TrP}_m^0$ form a self-adjoint algebra in $C(\Sigma_{m,R})$, and they separate points because by definition two laws are the same if they agree on every non-commutative polynomial. So by the Stone-Weierstrass theorem, this algebra is dense in $C(\Sigma_{m,R})$. In particular, there exists a trace polynomial $g$ such that $|\nu(g) - F(\nu)| < \epsilon / 2$ for all $\nu \in \Sigma_{m,R}$. Then let $f = (g + \epsilon/2) / (g(\mu) + \epsilon / 2)$.
We will also use the following smooth cut-off trick.
Let $0 < R' \leq R$. Let $\phi \in C_c^\infty({\mathbb{R}};{\mathbb{R}})$ such that $\phi(t) = t$ for $t \leq R'$ and $|\phi(t)| \leq R$. For $y \in (\mathcal{R}^\omega)_{sa}$, define $\Phi(y) = \phi(y)$ where $\phi$ is applied through functional calculus. Then
1. $\Phi(y) = y$ if $\norm*{y}_\infty \leq R'$.
2. $\norm*{\Phi(y)}_\infty \leq R$ for all $y$.
3. $\Phi \in \overline{\operatorname{TrP}}_m^1$.
4. $\Phi$ is globally $\norm*{\cdot}_2$-Lipschitz.
\(1) and (2) follow from the properties of functional calculus. To prove (3), note by the Weierstrass approximation theorem that for every $r > 0$, there is a polynomial $p$ such that $|p(t) - \phi(t)| < \epsilon$ for $|t| \leq r$. This implies as with (1) that $|p(y) - \phi(y)| < \epsilon$ for all $y$ with $\norm*{y}_\infty \leq r$. Claim (4) follows from the results of [@Peller2006]; the argument is explained in [@Jekel2018 (8.9) and Proposition 8.8].
Let $\mu$ be the law of $x = (x_1,\dots,x_m)$, and let $R > \norm*{X}_\infty$. Since $z \in L^2(\mathcal{M},\tau)$, there exist non-commutative polynomials $\{p_k\}_{k=1}^\infty$ such that $\norm*{p_k(x) - z}_2 < 1 / 2^{k+1}$ and hence for $k \geq 1$, $$\norm*{p_{k+1}(x) - p_k(x)}_2 = \mu[(p_{k+1} - p_k)^2]^{1/2} \leq \frac{1}{2^{k+1}} + \frac{1}{2^{k+2}} < \frac{1}{2^k}.$$ By scaling, we may assume without loss of generality that $\norm*{z}_2 < 1$ and set $p_0 = 0$, and then the above statement also holds for $k = 0$. Now let $$\mathcal{U}_k = \{\nu \in \Sigma_{m,R}: \nu((p_{k+1} - p_k)^2)^{1/2} < 1/2^k\},$$ which is a neighborhood of $\mu$ in $\Sigma_{m,R}$. By the previous lemma, there exists a scalar-valued trace polynomial $u_k$ such that $\mu(u_k) = 1$ and $$0 \leq \nu(u_k) \leq \mathbf{1}_{\nu \in \mathcal{U}_k} + \frac{1}{2^k \norm*{p_{k+1} - p_k}_{u,R}}.$$ (We can assume without loss of generality that $\norm*{p_{k+1} - p_k}_{u,R} \neq 0$.) Now the function $u_k(p_{k+1} - p_k)$ will evaluate at the point $X$ to $p_{k+1}(x) - p_k(x)$. If $y \in (\mathcal{R}^\omega)_{sa}^m$ with $\norm*{y}_\infty \leq R$ and if the law of $y$ is in $\mathcal{U}_k$, then we will have $$\norm*{u_k(y)(p_{k+1}(Y) - p_k(y))}_2 \leq \norm*{p_{k+1}(y) - p_k(y)}_2 + \frac{1}{2^k \norm*{p_{k+1} - p_k}_{u,R}} \norm*{p_{k+1}(y) - p_k(y)}_2 \leq \frac{1}{2^k} + \frac{1}{2^k}.$$ On the other hand, if the law of $y$ is not in $\mathcal{U}_k$, then $\norm*{u_k(Y)(p_{k+1}(y) - p_k(y))}_2 \leq 1/2^k$. Overall, we have $$\norm*{u_k \cdot (p_{k+1} - p_k)}_{u,R} \leq \frac{2}{2^k}.$$ This implies that $\sum_{k=0}^\infty u_k \cdot (p_{k+1} - p_k)$ converges with respect to $\norm*{\cdot}_{u,R}$ for our given choice of $R$, and of course evaluating this function on $X$ it produces the desired operator $Z$ since $u_k(x) = 1$.
To extend the function to be be globally defined on $(\mathcal{R}^\omega)_{sa}^m$, we use the smooth cut-off trick. Let $\phi \in C_c^\infty({\mathbb{R}};{\mathbb{R}})$ such that $\phi(t) = t$ for $|t| \leq \norm*{X}_\infty$ and $|\phi| \leq R$. For $y = (y_1,\dots,y_m) \in (\mathcal{R}^\omega)_{sa}^m$, let $\Phi(y) = (\phi(y_1),\dots,\phi(y_m))$. Then $[u_k \cdot (p_{k+1} - p_k)] \circ \Phi \in \overline{\operatorname{TrP}}_m^1$ because it is the composition of a trace polynomial with a function $\Phi \in (\overline{\operatorname{TrP}}_m^1)_{sa}^m$ that is uniformly bounded in operator norm.
Also, since $\Phi$ is globally $\norm*{\cdot}_2$-Lipschitz and since $u_k \cdot (p_{k+1} - p_k)$ is $\norm*{\cdot}_2$-Lipschitz on the operator norm ball of radius $R$, we see that $[u_k \cdot (p_{k+1} - p_k)] \circ \Phi$ is globally Lipschitz in $\norm*{\cdot}_2$. For all $y \in (\mathcal{R}^\omega)_{sa}^m$, $$\norm*{u_k(\Phi(y))(p_{k+1}(\Phi(y)) - p_k(\Phi(y)))}_2 \leq \frac{2}{2^k}.$$ Therefore, $$f(y) := \sum_{k=0}^\infty u_k(\Phi(y))(p_{k+1}(\Phi(y)) - p_k(\Phi(y)))$$ converges, and clearly $f \in \overline{\operatorname{TrP}}_m^1$ since each of the individual terms is. Furthermore, $\norm*{\cdot}_2$-uniform continuity of each term and the uniform convergence of the series implies uniform continuity of $f$. Since $\norm*{x}_\infty \leq R$, we have $\Phi(x) = x$ and $u_k(x) = 1$, so that $$f(x) = \sum_{k=0}^\infty [p_{k+1}(x) - p_k(x)] = \lim_{k \to \infty} p_{k+1}(x) = z.$$ This concludes the proof of (1).
To verify (2), we take $f_n$ to be the $n$th partial sum of the series defining $f$; we have shown that the individual terms are $\norm*{\cdot}_2$-Lipschitz, hence so are the partial sums. Finally, to prove (3), note that if $z = p(x_1,\dots,x_m)$, then $z$ also equals $f(x_1,\dots,x_m)$ where $f = p \circ \Phi$, and by the same reasoning as above $p \circ \Phi$ is globally $\norm*{\cdot}_2$-Lipschitz.
We have shown that every element of $L^2(\mathrm{W}^*(x_1,\dots,x_m))$ has the form $f(x_1,\dots,x_m)$ for some $f \in \overline{\operatorname{TrP}}_m^1$. On the other hand, we will prove that if $f$ is Lipschitz, then $f(x)$ is actually bounded in operator norm. We state our estimate in terms of unitarily invariant random matrix models which satisfy concentration , but as explained in Remark \[rem:concentrationmodels\] such models exist whenever $L^2(\mathrm{W}^*(x_1,\dots,x_m))$ is embeddable into $\mathcal{R}^\omega$.
\[prop:operatornormestimate\] Let $x = (x_1, \dots, x_m)$ be a tuple of self-adjoint variables in a $\mathrm{W}^*$-algebra $(\mathcal{M},\tau)$ whose non-commutative law is $\lambda$. Suppose there is a sequence $\{\mu^{(N)}\}$ of probability measures on $M_N({\mathbb{C}})_{sa}^m$, invariant under unitary conjugation, that satisfies the concentration estimate for some constant $c$, and such that the corresponding random variables $X^{(N)} = (X_1^{(N)}, \dots, X_m^{(N)})$ satisfy $\lambda_{X^{(N)}} \to \lambda$ in probability. Then $\mathrm{W}^*(x)$ is embeddable into $\mathcal{R}^\omega$. Moreover, if $f \in \overline{\operatorname{TrP}}_m^1$ is $\norm{\cdot}_2$-Lipschitz, then $f(x)$ is a bounded operator and $$\norm*{f(x) - \tau(f(x))}_\infty \leq \Theta c^{-1/2} \norm*{f}_{\operatorname{Lip}},$$ where $\Theta$ is a universal constant.
In light of Lemma \[lem:epsilonnet\], $$P\left(\norm*{X^{(N)} - E(X^{(N)})}_\infty \leq c^{-1/2} (\Theta + N^{-1/3}) \right) \to 1$$ and $$P\left(\norm*{f(X^{(N)}) - E(f(X^{(N)}))}_\infty \leq c^{-1/2} \norm{f}_{\operatorname{Lip}} (\Theta + N^{-1/3}) \right) \to 1.$$ Also, the non-commutative law of $X^{(N)}$ converges in probability to that of $x$ and finally $\tau_N(f(X^{(N)})) - E[\tau_N(f(X^{(N)}))] \to 0$ in probability as a consequence of concentration. Therefore, we may choose a sequence of elements $y^{(N)} \in M_N({\mathbb{C}})_{sa}^m$ such that $$\begin{aligned}
\limsup_{N \to \infty} \norm*{y_j^{(N)} - E(X_j^{(N)})}_\infty &\leq c^{-1/2} \Theta, \\
\limsup_{N \to \infty} \norm*{f(y^{(N)}) - E(f(X^{(N)}))}_\infty &\leq c^{-1/2} \norm{f}_{\operatorname{Lip}} \Theta, \\
\left| \tau_N(f(y^{(N)})) - E[\tau_N(f(X^{(N)}))] \right| &\to 0, \\
\lambda_{y^{(N)}} &\to \lambda_x.\end{aligned}$$ Because $E(X_j^{(N)}) = E(\tau_N(X_j^{(N)}))$ by unitary invariance and because of concentration, $E(\tau_N(X_j^{(N)}))$ must converge to $\tau(x_j)$ since $\tau_N(X_j^{(N)})$ converges to the $\tau(x_j)$ in probability. So overall $E(X_j^{(N)}) - \tau_N(x_j) \to 0$ in operator norm. In particular, $$\limsup_{N \to \infty} \norm*{y_j^{(N)} - \tau(x_j)}_\infty \leq c^{-1/2} \Theta,$$ and hence $\norm{y^{(N)}}_\infty$ is bounded as $N \to \infty$. Moreover, our choice of $y^{(N)}$ also satisfies $$\limsup_{N \to \infty} \norm*{f(y^{(N)}) - \tau_N(f(y^{(N)}))}_\infty \leq c^{-1/2} \norm{f}_{\operatorname{Lip}} \Theta,$$ since $E[f(X^{(N)})] = E[\tau_N(f(X^{(N)})]$ again by unitary invariance.
Fix a free ultrafilter $\omega$ and let $(\mathcal{M},\tau) = \prod_{N \to \omega} (M_N({\mathbb{C}}), \tau_N)$ be the tracial $\mathrm{W}^*$-ultraproduct of the sequence of matrix algebras. Since $\{y^{(N)}\}$ is bounded in operator norm, $y = \{y^{(N)}\}_{N \in {\mathbb{N}}}$ defines an element of $(\mathcal{M},\tau)$. By definition of ultraproducts, $\tau(p(y)) = \lim_{N \to \omega} \tau_N(p(y^{(N)}))$ for every non-commutative polynomial $p$ and therefore the non-commutative law of $y$ is $\lambda$ (which is the same as that of $x$). In particular, $\mathrm{W}^*(x) \cong \mathrm{W}^*(y)$ embeds into $(\mathcal{M},\tau)$ and hence also into $\mathcal{R}^\omega$. (Compare [@GS2009 Theorem 4.4].)
Since $\mathrm{W}^*(x)$ is $\mathcal{R}^\omega$-embeddable, $f(x)$ is well-defined, and clearly $\norm{f(x) - \tau(f(x))}_\infty = \norm{f(y) - \tau(f(y))}_\infty$. Now we claim that $f(y)$ is given by the sequence $\{f(y^{(N)})\}_{N \in {\mathbb{N}}}$ as an element of $(\mathcal{M},\tau)$ (that is, application of $f$ commutes with ultralimits). It is easy to check that $g(y) = \{g(y^{(N)})\}_{N \in {\mathbb{N}}}$ when $g \in \operatorname{TrP}_m^1$. But for any $\epsilon > 0$, there exists $g \in \operatorname{TrP}_m^1$ with $\norm{f - g}_{c^{-1/2} \Theta + 1} < \epsilon$. Thus, $\norm{f(y) - g(y)}_2 < \epsilon$ and also $\norm{f(y^{(N)}) - g(y^{(N)})}_2 < \epsilon$ for sufficiently large $N$. This implies that $\norm{f(y) - \{f(y^{(N)})\}_{N \in {\mathbb{N}}}}_2 < 2 \epsilon$. Thus, $f(y) = \{f(y^{(N)})\}_{N \in {\mathbb{N}}}$ as claimed. The same holds with $f$ replaced by $f - \tau(f)$. This implies $$\norm{f(y) - \tau(f(y))}_\infty \leq \limsup_{N \to \infty} \norm{f(y^{(N)}) - \tau(f(y^{(N)}))}_\infty \leq c^{-1/2} \norm{f}_{\operatorname{Lip}} \Theta. \qedhere$$
\[rem:concentrationmodels\] Suppose that $\mathrm{W}^*(x_1,\dots,x_m)$ is embeddable into $\mathcal{R}^\omega$. Then there exist tuples $x^{(N)} = (x_1^{(N)},\dots,x_m^{(N)})$ in $M_N({\mathbb{C}})_{sa}^m$ such that $\norm{x^{(N)}}_\infty \leq \norm{x}_\infty$ and $\lambda_{x^{(N)}} \to \lambda_x$. Let $U^{(N)}$ be an $N \times N$ random Haar unitary matrix and let $X^{(N)} = U^{(N)} x^{(N)} (U^{(N)})^*$. Clearly, the probability distribution of $X^{(N)}$ is unitarily invariant and also $\lambda_{X^{(N)}} \to \lambda_x$ in probability.
To check concentration, observe that $u \mapsto u x^{(N)}u^*$ is a $2 m^{1/2} \norm{x}_\infty$-Lipschitz function from the unitary group to $M_N({\mathbb{C}})_{sa}^m$ with respect to $\norm{\cdot}_2$. Therefore, if $f: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ is Lipschitz, then $u \mapsto f(ux^{(N)}u^*)$ is also Lipschitz, with the Lipschitz constant $2 m^{1/2} \norm{x}_\infty \norm{f}_{\operatorname{Lip}}$. It was proved in [@Meckes2013 Theorem 15], [@Meckes2019 Theorem 5.16] that the Haar measure on the unitary group satisfies the (non-normalized) log-Sobolev inequality with constant $6/N$ and the corresponding concentration of measure for Lipschitz functions with respect to the Hilbert-Schmidt metric $N^{1/2} \norm{\cdot}_2$. After renormalization this implies that the Haar measure on the unitary group satisfies with $c = 1/6$. Hence, $X^{(N)}$ satisfies with $c = 1 / 12 m \norm{x}_\infty^2$.
Tools for Differential Equations in $\overline{\operatorname{TrP}}_m^j$ {#sec:diffeqtools}
=======================================================================
This section describes two analytic operations | solution of ODE and convolution with the Gaussian law | that can be performed on tuples in $\overline{\operatorname{TrP}}_m^1$ and on asymptotically approximable sequences of functions on $N \times N$ matrices. These operations were applied in [@Jekel2018], and will be applied in the remainder of this paper, to analyze the large $N$ limit of certain PDE associated to random matrix models, and hence to understand the behavior of convex matrix models in the large $N$ limit.
Flows Along Vector Fields {#subsec:vectorfields}
-------------------------
Several times in our study of partial differential equations, we will use flows along vector fields given by functions in $\overline{\operatorname{TrP}}_m^1$ and by asymptotically approximable sequences of functions on matrices. For instance, this idea was used in [@Jekel2018 Lemma 4.10], and in this paper, it will be used in the proof of Lemma \[lem:diffusionAATP\] and Theorem \[thm:transport1\].
The setup is roughly speaking as follows. Consider a time interval $[0,T] \subseteq {\mathbb{R}}$. Let $H: (\mathcal{R}^\omega)_{sa}^m \times [0,T] \to L^2(\mathcal{R}^\omega)_{sa}^m$ be a function such that $H(\cdot,t)$ is a tuple of functions in $\overline{\operatorname{TrP}}_m^1$ for each $t$ (satisfying certain uniform continuity assumptions). Also, let $F_0: (\mathcal{R}^\omega)_{sa}^m \to L^2(\mathcal{R}^\omega)_{sa}^m$. Then we would like to construct $F: (\mathcal{R}^\omega)_{sa}^m \times [0,T] \to (\mathcal{R}^\omega)_{sa}^m$ such that $$\begin{aligned}
F(x,0) &= F_0(x) \\
\partial_t F(x,t) &= H(F(x,t),t).\end{aligned}$$ Moreover, we would like to show that if $H^{(N)}$ is a function on $M_N({\mathbb{C}})_{sa}^m \times [0,T]$ that is asymptotic to $H$ and $F_0^{(N)} \rightsquigarrow F_0$, then the solutions $F^{(N)}$ are asymptotic to the solution $F$.
Such a proof was essentially carried out in [@Jekel2018 Lemma 4.10], but now we introduce the added complexity that $H$ will depend on $x$, $t$, and an auxiliary parameter $y \in (\mathcal{R}^\omega)_{sa}^m$, and we must solve the initial value problem $$\begin{aligned}
\label{eq:IVP}
F(x,y,0) &= F_0(x,y) \\
\partial_t F(x,y,t) &= H(F(x,y,t), y, t). \nonumber\end{aligned}$$ The added parameter $y$ arises naturally in our analysis of *conditional* expectation, entropy, and transport since it represents the variables we are conditioning upon (see for instance §\[subsec:conditionalexpectationstrategy\]).
For the sake of future reference, let us state the set of assumptions we make about the vector field $H(x,y,t)$. These assumptions are framed for a convenient and applicable level of generality rather than maximum generality.
\[ass:vectorfield\] We are given $T > 0$ and a function $H: (\mathcal{R}^\omega)_{sa}^m \times (\mathcal{R}^\omega)_{sa}^n \times [0,T] \to L^2(\mathcal{R}^\omega)_{sa}^m$ satisfying:
1. For each $t$, we have $H(\cdot,\cdot,t) \in (\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$.
2. $H$ is $\norm*{\cdot}_2$-Lipschitz in $(x,y)$, that is, for some constant $K$ independent of $t$, we have $$\norm*{H(x,y,t) - H(x',y',t)}_2 \leq K \norm*{(x,y) - (x',y')}_2.$$
3. The map $t \mapsto H(\cdot,\cdot,t)$ is a continuous function $[0,T] \to (\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$ with respect to the Fr[é]{}chet topology on $\overline{\operatorname{TrP}}_{m+n}^1$. This implies that for every $R > 0$ and for every $\epsilon > 0$, there exists $\delta > 0$, such that $$|t - t'| < \delta \implies \norm*{H(\cdot,\cdot,t) - H(\cdot,\cdot,t')}_{u,R} < \epsilon \text{ for all } t, t' \in [0,T].$$ (where we have upgraded from continuity to uniform continuity because of compactness of $[0,T]$).
Under this assumption, as in Observation \[obs:continuousextension\], we see that $H(\cdot,\cdot,t)$ has a unique continuous extension to $L^2(\mathcal{R}^\omega)_{sa}^{m+n}$. Furthermore, for each $(x,y) \in L^2(\mathcal{R}^\omega)_{sa}^{m+n}$, the function $t \mapsto H(x,y,t)$ is continuous (though the modulus of continuity cannot be chosen independent of $(x,y)$). Continuity follows because there exists a sequence $(x_n,y_n) \in (\mathcal{R}^\omega)_{sa}^{m+n}$ such that $(x_n,y_n) \to (x,y)$ in $\norm*{\cdot}_2$. Now $H(x_n,y_n,\cdot)$ is continuous by assumption (3), but assumption (2) implies that $H(x_n,y_n,\cdot) \to H(x,y,\cdot)$ uniformly on $[0,T]$.
Under these assumptions, can be solved by the standard method of Picard iteration. We first verify that Assumption \[ass:vectorfield\] is preserved under the composition and integration operations used to define Picard iterates.
\[lem:vectorfieldintegration\] Suppose that $H(x,y,t)$ satisfies Assumption \[ass:vectorfield\] and suppose that $G_0 \in (\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$ is globally $\norm*{\cdot}_2$-Lipschitz. Then the function $$G(x,y,t) = G_0(x,y) + \int_0^t H(x,y,s)\,ds$$ is well-defined by Riemann integration and it also satisfies Assumption \[ass:vectorfield\].
The Riemann integral is defined because $t \mapsto H(x,y,t)$ is continuous with respect to $\norm*{\cdot}_2$ for each $(x,y) \in (\mathcal{R}^\omega)_{sa}^{m+n}$ (and in fact, each $(x,y) \in L^2(\mathcal{R}^\omega)_{sa}^{m+n}$). Now let us check that $G$ satisfies Assumption \[ass:vectorfield\].
\(1) Fix $R > 0$ and $\epsilon > 0$. By assumption (2) for $H$, there exists $\delta > 0$ such that $$|t - t'| < \delta \implies \norm*{H(\cdot,\cdot,t) - H(\cdot,\cdot,t')}_{u,R} < \frac{\epsilon}{2T}.$$ Fix $t$, then choose a partition $0 = t_0$, …, $t_n = t$ of $[0,t]$ such that $|t_j - t_{j-1}| < \delta$. Then let $h_j \in (\operatorname{TrP}_{m+n}^1)_{sa}^m$ such that $$\norm*{h_j - H(\cdot,\cdot,t_j)}_{u,R} < \frac{\epsilon}{2T}.$$ Then $$\norm*{h_j - H(\cdot,\cdot,s)}_{u,R} < \frac{\epsilon}{T} \text{ for all } s \in [t_{j-1},t_j].$$ Therefore, $$\norm*{ \int_0^t H(\cdot,\cdot,s)\,ds - \sum_{j=1}^n (t_j - t_{j-1}) h_j }_{u,R} < \sum_{j=1}^n (t_j - t_{j-1}) \frac{\epsilon}{T} = \frac{\epsilon t}{T} \leq \epsilon.$$ This shows that $\int_0^t H(\cdot,\cdot,s)\,ds$ is in $(\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$. Because $G_0$ is in this space as well, this implies that $G(\cdot,\cdot,t)$ is in $(\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$ as desired.
\(2) If $H(\cdot,\cdot,t)$ is $K$-Lipschitz for all $t$, then $\norm*{G(\cdot,t)}_{\operatorname{Lip}} \leq \norm*{G_0}_{\operatorname{Lip}} + tK$.
\(3) Since $t \mapsto H(\cdot,\cdot,t)$ is continuous with respect to $\norm*{\cdot}_{u,R}$, we must have $\norm*{H(\cdot,\cdot,t)}_{u,R} \leq M$ for some constant $M$. Then $\norm*{G(\cdot,\cdot,t) - G(\cdot,\cdot,t')}_{u,R} \leq M |t - t'|$.
\[lem:vectorfieldcomposition\] Suppose that $H(x,y,t)$ and $G(x,y,t)$ satisfy Assumption \[ass:vectorfield\]. Then $H(G(x,y,t),y,t)$ also satisfies Assumption \[ass:vectorfield\].
The composition makes sense because $H(x,y,t)$ extends to be defined for $(x,y) \in L^2(\mathcal{R}^\omega)_{sa}^{m+n}$. It follows from Lemma \[lem:composition\] that $H(G(x,y,t),y,t)$ satisfies (1). The Lipschitz estimate (2) is straightforward and left to the reader. To prove (3), let $K$ be a Lipschitz constant for $H$ as a function of $(x,y)$ that works for all $t$. Fix $\epsilon > 0$. Proceeding as in the proof of Lemma \[lem:vectorfieldintegration\], we can choose a partition $\{t_0,\dots,t_n\}$ of $[0,T]$ and $g_j \in (\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$ such that $$\norm*{g_j - G(\cdot,\cdot,t)}_{u,R} < \frac{\epsilon}{4K} \text{ for } t \in [t_{j-1},t_j].$$ Then there exists some $R'$ such that $\norm*{(x,y)}_\infty \leq R$ implies $\norm*{(g_j(x,y),y)}_\infty \leq R'$ for all $j$. Then by applying assumption (3) to $H$, there exists $\delta$ such that $$|t - t'| < \delta \implies \norm*{H(\cdot,\cdot,t) - H(\cdot,\cdot,t')}_{u,R'} < \frac{\epsilon}{4}.$$ We also choose $\delta'$ such that $$|t - t'| < \delta' \implies \norm*{G(\cdot,\cdot,t) - G(\cdot,\cdot,t')}_{u,R} < \frac{\epsilon}{4K}.$$ Supposing that $\norm*{(x,y)}_\infty \leq R$ and $|t - t'| < \min(\delta,\delta')$, we have $$\begin{aligned}
&\norm*{H(G(x,y,t),y,t) - H(G(x,y,t'),y,t')}_2 \\
&\leq \norm*{H(G(x,y,t),y,t) - H(G(x,y,t'),y,t)}_2 + \norm*{H(G(x,y,t'),y,t) - H(G(x,y,t'),y,t')}_2 \\
&\leq K \norm*{G(x,y,t) - G(x,y,t')}_2 + \norm*{H(G(x,y,t'),y,t) - H(G(x,y,t'),y,t')}_2 \\
&\leq \frac{\epsilon}{4} + \norm*{H(G(x,y,t'),y,t) - H(G(x,y,t'),y,t')}_2.\end{aligned}$$ Meanwhile, after we pick $j$ such that $t' \in [t_{j-1},t_j]$, then $$\begin{aligned}
\norm*{H(G(x,y,t'),y,t) - H(G(x,y,t'),y,t')}_2 &\leq \norm*{H(G(x,y,t'),y,t) - H(g_j(x,y),y,t)}_2 \\
& \quad + \norm*{H(g_j(x,y),y,t) - H(g_j(x,y),y,t')}_2 \\
& \quad + \norm*{H(g_j(x,y),y,t) - H(G(x,y,t'),y,t')}_2.\end{aligned}$$ The middle term can be estimated by $\epsilon / 4$ because $\norm*{g_j(x,y),y)}_\infty \leq R'$. Meanwhile, the first and third terms can each be estimated by $K (\epsilon / 4K) = \epsilon / 4$ using the Lipschitz property of $H$ and our choice of $g_j$. Altogether, $|t - t'| < \min(\delta,\delta')$ implies that $\norm*{H(G(x,y,t),y,t) - H(G(x,y,t'),y,t')}_2 < \epsilon$ whenever $\norm*{(x,y)}_\infty \leq R$.
\[prop:ODE\] Let $H(x,y,t)$ satisfy Assumption \[ass:vectorfield\] and let $G_0 \in (\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$. Then there exists a unique continuous $F: L^2(\mathcal{R}^\omega)_{sa}^{m+n} \times [0,T] \to L^2(\mathcal{R}^\omega)_{sa}^m$ satisfying $$F(x,y,t) = G_0(x,y) + \int_0^t H(F(x,y,s),y,s)\,ds.$$ Moreover, $F(x,y,t)$ also satisfies Assumption \[ass:vectorfield\].
We define the Picard iterates $F_\ell$ inductively by $$\begin{aligned}
F_0(x,y,t) &= G_0(x,y) \\
F_{\ell+1}(x,y,t) &= G_0(x,y) + \int_0^t H(F_\ell(x,y,s),y,s)\,ds.\end{aligned}$$ The previous two lemmas imply that $F_k$ is well-defined and satisfies Assumption \[ass:vectorfield\]. Convergence of the Picard iterates follows from the standard proof of Picard-Lindel[ö]{}f. Briefly, given that $H$ is $K$-Lipschitz in $(x,y)$ with respect to $\norm*{\cdot}_2$, we have $$\norm*{F_{\ell+1}(x,y,t) - F_\ell(x,y,t)}_2 \leq K \int_0^t \norm*{F_\ell(x,y,s) - F_{\ell-1}(x,y,s)}_2\,ds.$$ Also, we have $$\norm*{F_1(x,y,t) - F_0(x,y,t)}_2 \leq t M(x,y).$$ where $M(x,y) = \sup_{s \in [0,T]} \norm*{H(G_0(x,y),y,s)}_2$, which is finite because of continuity of $H(G_0(x,y),y,t)$ in $t$. From here a straightforward induction on $\ell$ shows that for $\ell \geq 1$, $$\norm*{F_{\ell}(x,y,t) - F_{\ell-1}(x,y,t)}_2 \leq \frac{K^{\ell-1} t^\ell}{\ell !}$$ because $K \int_0^t K^{\ell-1} s^\ell / \ell!\,ds = K^\ell s^{\ell+1} / (\ell + 1)!$. Now because $\sum_{\ell=1}^\infty K^{\ell - 1} s^\ell / \ell!$ converges, we know that $$F(x,y,t) := \lim_{\ell \to \infty} F_\ell(x,y,t) \text{ exists,}$$ and $$\norm*{F_\ell(x,y,t) - F(x,y,t)}_2 \leq M(x,y) \sum_{j=\ell+1}^\infty \frac{K^{j-1} t^j}{j!}.$$ The fact that $F(x,y,t)$ satisfies the integral equation is straightforward, and the proof of the uniqueness of this $F$ is also standard.
It remains to show that $F$ satisfies Assumption \[ass:vectorfield\]. First, recall that $H(G_0(x,y),y,t)$ is Lipschitz in $(x,y)$ uniformly for all $t$. If $K'$ is a Lipschitz constant for this function, then $$M(x,y) \leq M(0,0) + K' \norm*{(x,y)}_2.$$ In particular, $$\norm*{F_\ell(x,y,t) - F(x,y,t)}_2 \leq (M(0,0) + K' \norm*{(x,y)}_2) \sum_{j=\ell+1}^\infty \frac{K^{j-1} t^j}{j!}.$$ This implies that the convergence of $F_\ell$ to $F$ occurs uniformly for $(x,y)$ with $\norm*{(x,y)}_\infty \leq R$ and all $t \in [0,T]$. Then because $F_\ell(\cdot,\cdot,t)$ can be approximated in $\norm*{\cdot}_{u,R}$ by trace polynomials, the same must be true for $F(\cdot,\cdot,t)$ for each $t$, which shows that $F$ satisfies (1). Similarly, because of the uniform convergence of $F_\ell$ to $F$ for $\norm*{(x,y)}_\infty \leq R$ and $t \in [0,T]$, the uniform continuity property (3) for $F$ follows from property (3) for $F_\ell$.
Finally, we must show (2) that $F$ is Lipschitz in $(x,y)$. More precisely, we claim that $$\norm*{F(x,y,t) - F(x',y',t)}_2 \leq e^{Kt} \norm*{G_0(x,y) - G_0(x',y')}_2 + (e^{Kt} - 1) \norm*{y - y'}_2.$$ Now it suffices to check that each Picard iterate $F_\ell$ satisfies this estimate. This can be verified by induction on $\ell$. The base case $F_0(x,y,t) = G_0(x,y)$ is immediate. For the induction step, we observe that $$\begin{aligned}
& \norm*{F_{\ell+1}(x,y,t) - F_\ell(x',y',t)}_2 \\
&\leq \norm*{G_0(x,y) - G_0(x',y')}_2 + \int \norm*{H(F_\ell(x,y,s),y,s) - H(F_\ell(x',y',s),y',s)}_2\,ds \\
&\leq \norm*{G_0(x,y) - G_0(x',y')}_2 + \int K \left( \norm*{F_\ell(x,y,s) - F_\ell(x',y',s)}_2 + \norm*{y - y'}_2 \right) \,ds,\end{aligned}$$ using the fact that $H$ is $K$-Lipschitz. Then we plug in our induction hypothesis that $\norm*{F_\ell(x,y,s) - F_\ell(x',y',s)}_2$ is bounded by $e^{Kt} \norm*{G_0(x,y) - G_0(x',y')}_2 + (e^{Kt} - 1) \norm*{y - y'}_2$, and then directly evaluate the integral to close the induction.
We have now shown that it makes sense to solve ODE for tuples in $(\overline{\operatorname{TrP}}_m^1)_{sa}$. There is a parallel list of results which instead deal with functions on $N \times N$ matrices that are asymptotically approximable as $N \to \infty$. We use the following assumptions.
\[ass:vectorfield2\] We are given $T > 0$ and for each $N \in {\mathbb{N}}$ a function $H^{(N)}: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \times [0,T] \to M_N({\mathbb{C}})_{sa}^m$ such that
1. For each $t$, there exists $H(\cdot,\cdot,t) \in (\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$ such that $H^{(N)}(\cdot,\cdot,t) \rightsquigarrow H(\cdot,\cdot,t)$.
2. $H^{(N)}$ is $\norm*{\cdot}_2$-Lipschitz in $(x,y)$ with some Lipschitz constant $K$ independent of $t$ and $N$.
3. For every $R > 0$ and for every $\epsilon > 0$, there exists $\delta > 0$, such that $$|t - t'| < \delta \implies \norm*{H^{(N)}(\cdot,\cdot,t) - H^{(N)}(\cdot,\cdot,t')}_{u,R}^{(N)} < \epsilon \text{ for all } t, t' \in [0,T] \text{ for all } N.$$
\[prop:ODE2\] Let $\{H^{(N)}\}$ satisfy Assumption \[ass:vectorfield2\], and let $G_0^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})_{sa}^m$ be asymptotically approximable such that $G_0^{(N)} \rightsquigarrow G_0$ and $G_0^{(N)}$ is $\norm*{\cdot}_2$ Lipschitz uniformly in $N$. Then for each $N$ there is a unique $F^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \times [0,T] \to M_N({\mathbb{C}})_{sa}^m$ satisfying $$F^{(N)}(x,y,t) = G_0^{(N)}(x,y) + \int_0^t H^{(N)}(F^{(N)}(x,y,s),y,s)\,ds.$$ Moreover, $\{F^{(N)}\}$ also satisfies Assumption \[ass:vectorfield\]. Furthermore, the vector field $H$ such that $H^{(N)}(\cdot, \cdot, t) \rightsquigarrow H(\cdot,\cdot,t)$ satisfies Assumption \[ass:vectorfield\], and we have $F^{(N)} \rightsquigarrow F$ where $F$ is the solution given by Proposition \[prop:ODE\].
The proof of existence and uniqueness of the solution is almost identical to that of Proposition \[prop:ODE\]. First, one shows that Assumption \[ass:vectorfield2\] is preserved under integration and composition (analogous to Lemma \[lem:vectorfieldintegration\] and \[lem:vectorfieldcomposition\]). Then exactly as in the proof of Proposition \[prop:ODE\], one defines Picard iterates, proves they converge, establishes Lipschitz bounds, and checks they satisfy Assumption \[ass:vectorfield2\]. The one additional feature in these proofs is to make all the estimates uniform in $N$. For instance, the quantity $M(x,y)$ in the proof of Proposition \[prop:ODE\] is replaced by $$M^{(N)}(x,y) = \sup_{s \in [0,T]} \norm*{H^{(N)}(G_0^{(N)}(x,y),y,s)}_2.$$ Then $H^{(N)}(G_0^{(N)}(x,y),y,t)$ has some Lipschitz constant $K'$ independent of $N$, and $$M^{(N)}(x,y) \leq M^{(N)}(0,0) + K' \norm*{(x,y)}_2.$$ But then we can show that $\sup_N M^{(N)}(0,0)$ is finite. This is because if $\Phi^{(N)}(x,y,t) = H^{(N)}(G_0^{(N)}(x,y),y,t)$, then $\sup_N \sup_t \norm*{\Phi^{(N)}(\cdot,\cdot,t)}_{u,R}^{(N)}$ is finite because of Assumption \[ass:vectorfield\] (3) and the fact that $\Phi^{(N)}(x,y,0)$ is asymptotically approximable and hence bounded in $\norm*{\cdot}_{u,R}^{(N)}$ as $N \to \infty$.
Now the fact that $H$ satisfies Assumption \[ass:vectorfield\] is a straightforward limiting argument. The key ingredient is that if $f^{(N)} \rightsquigarrow f$, then $\norm*{f}_{u,R} = \lim_{N \to \infty} \norm*{f^{(N)}}_{u,R}^{(N)}$.
Finally, to show that $F^{(N)} \rightsquigarrow F$, it suffices to show that for each of the Picard iterates $F_\ell^{(N)} \rightsquigarrow F_\ell$ because of the uniform convergence of $F_\ell^{(N)} \to F^{(N)}$ as $\ell \to \infty$ for $\norm*{(x,y)}_\infty \leq R$, where the rate of convergence is also independent of $N$. Furthermore, since the Picard iterates are defined inductively by composition and integration, it suffices to show that the asymptotic approximation relation $\rightsquigarrow$ is preserved by these operations. Preservation under integration follows because the integrals can be approximated by Riemann sums and this approximation is uniformly good for $\norm*{(x,y)}_\infty \leq R$ and for all $N$ because of the uniform continuity Assumption \[ass:vectorfield2\] (3). Preservation under composition follows from Lemma \[lem:composition\].
The Heat Semigroup {#subsec:heatsemigroup}
------------------
Recall that the solution to the classical heat equation is given by convolution the heat kernel (which is given by a Gaussian probability density). In particular, let $\sigma_{m,t}^{(N)}$ be the probability distribution of an $m$-tuple of independent GUE matrices $(S_1^{(N)}, \dots, S_m^{(N)})$ such that $E[\tau_N[(S_j^{(N)})^2]] = t$, which is given by density $(1/Z^{(N)}) e^{-\norm*{x}_2^2 / 2t}\,dx$. If $u_0: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$, then $u_t := u_0 * \sigma_t^{(N)}$ solves the normalized heat equation $$\partial_t u_t = \frac{1}{2N} \Delta u_t.$$ Here $u_0 * \sigma_{m,t}^{(N)}$ is meant in the sense of convolving a function with a measure, and this is the same as convolving of $u_0$ with the density function for $\sigma_{m,t}^{(N)}$. The meaning of $\Delta$ is to be interpreted using coordinates with respect to some orthonormal basis of $M_N({\mathbb{C}})_{sa}$ in the inner product $\ip{x,y} = \operatorname{Tr}(xy)$; this is *not* the same as differentiating entrywise since some of the entries are real and some are complex.
Our goal is to describe the large $N$ behavior of $u^{(N)} * \sigma_{m,t}^{(N)}$ when $\{u^{(N)}\}$ is asymptotically approximable by trace polynomials, and to define “$u \boxplus \sigma_{m,t}$” when $u \in \overline{\operatorname{TrP}}_m^j$.
In [@Jekel2018 §3.2 and 3.3], using similar methods to [@Cebron2013], we explained the computation of $(1/N) \Delta f$ as a function on $M_N({\mathbb{C}})_{sa}^m$ when $f \in \operatorname{TrP}_m^0$ or $\operatorname{TrP}_m^1$. More precisely, let $\Delta_j f(x_1,\dots,x_m)$ denote the Laplacian with respect to the coordinates of the matrix $x_j$. We found that for $j = 1, \dots, m$ there are linear maps $L_j^{(N)}, L_j: \operatorname{TrP}_m^0 \to \operatorname{TrP}_m^0$ defined purely algebraically, such that $(1/N) \Delta_j f = L_j^{(N)} f$ when $f$ is viewed as a function on $M_N({\mathbb{C}})_{sa}^m$, $L_j^{(N)}$ and $L_j$ do not increase the degree of a trace polynomial, and $\lim_{N \to \infty} L_j^{(N)} f = L_j f$ coefficient-wise.
A similar analysis holds for the Laplacian of $f \in \operatorname{TrP}_m^1$ viewed as a function $M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$. Here we follow the standard convention of using the same symbol $\Delta$ for the Laplacians of vector-valued functions as for the Laplacians of scalar-valued functions; thus, the reader must be careful to distinguish scalar-valued and vector-valued functions based on context. We saw that there were linear transformations $L_j^{(N)}, L_j: \operatorname{TrP}_m^1 \to \operatorname{TrP}_m^1$ such that $(1/N) \Delta_j f = L_j^{(N)}f$ as a function on matrices, $L_j^{(N)}$ and $L_j$ do not increase degree, and $L_j^{(N)} f \to L_j f$ coefficient-wise.
We deduced as a consequence that $e^{L^{(N)}t/2} f = f * \sigma_{m,t}^{(N)}$ has a well-defined large $N$ limit if $f$ is a trace polynomial [@Jekel2018 Lemma 3.21], and that if $\{u^{(N)}\}$ is asymptotically approximable by trace polynomials, then so is $\{u^{(N)} * \sigma_{m,t}^{(N)}\}$ [@Jekel2018 Lemma 3.28].
In order to establish “conditional versions” of our earlier results, we must consider trace polynomials $f(x_1,\dots,x_m, y_1, \dots, y_n)$ in $m + n$ variables and take the Laplacian with respect to $x = (x_1,\dots,x_m)$ while treating $y = (y_1,\dots,y_n)$ as an auxiliary parameter. We denote by $\Delta_x = \sum_{j=1}^m \Delta_{x_j}$, $L_x^{(N)} = \sum_{j=1}^m L_{x_j}^{(N)}$, and $L_x = \sum_{j=1}^m L_{x_j}$ the various Laplacian operators with respect to $x$.
Because $L_x^{(N)}$ and $L_x$ map the finite-dimensional vector space trace polynomials of degree $\leq d$ into itself, there are well-defined linear operators $e^{tL_x^{(N)} / 2}$ and $e^{tL_x / 2}$ on the space of trace polynomials in $\operatorname{TrP}_{m+n}^j$ of degree $\leq d$ for each $j = 0, 1$ each $d \in {\mathbb{N}}$, and each real $t \geq 0$. Since trace polynomials are the union of the subspaces of trace polynomials with degree $\leq d$, there are linear operators $e^{tL_x^{(N)}/2}, e^{tL_x / 2}: \operatorname{TrP}_{m+n}^j \to \operatorname{TrP}_{m+n}^j$. Moreover, these operators form a semigroup, and they satisfy the following property, which is an extension of [@Cebron2013 Theorem 2.4] to the spaces $\overline{\operatorname{TrP}}_m^j$.
Let $(X,Y)$ be a random variable in $M_N({\mathbb{C}})_{sa}^{m+n}$ with finite moments, and let $S \sim \sigma_{m,t}^{(N)}$ be an independent GUE random variable in $M_N({\mathbb{C}})_{sa}^m$. Then we have $$\label{eq:Gaussianconditionalexpectation}
E[f(X+S,Y) | (X,Y)] = [e^{tL_x^{(N)}/2} f](X,Y) \text{ for } f \in \operatorname{TrP}_{m+n}^0 \text{ or } f \in \operatorname{TrP}_{m+n}^1.$$ Similarly, suppose that $(X,Y)$ is a tuple of self-adjoint non-commutative random variables, and let $S$ be a freely independent tuple with non-commutative law $\sigma_{m,t}$. Then $$\label{eq:Gaussianconditionalexpectation2}
f(X+S,Y) = [e^{tL_x/2} f](X,Y) \text{ for } f \in \operatorname{TrP}_m^0,$$ and $$\label{eq:Gaussianconditionalexpectation3}
E_{\mathrm{W}^*(X,Y)}[f(X+S,Y)] = [e^{tL_x/2} f](X,Y) \text{ for } f \in \operatorname{TrP}_m^1,$$ where $E_{\mathrm{W}^*(X,Y)}: \mathrm{W}^*(X,Y,S) \to \mathrm{W}^*(X,Y)$ is the unique trace-preserving conditional expectation.
Since $S$ is independent and distributed according to $\sigma_{m,t}^{(N)}$, we have $$E[f(X+S,Y) | (X,Y)] = \int f(X + z,Y)\,d\sigma_{m,t}^{(N)}(z).$$ On the other hand, for $(x,y) \in M_N({\mathbb{C}})_{sa}^{m+n}$, $$\int f(x + z,y)\,d\sigma_{m,t}^{(N)}(z) = e^{tL_x^{(N)}/2} f(x,y),$$ because both sides are the solution to the heat equation on the space of coordinate-wise polynomials on $M_N({\mathbb{C}})_{sa}^{m+n}$ of degree $\leq d$. This shows .
To prove the free versions, we assume familiarity with the results of free probability (see e.g. [@VDN1992], [@NS2006], [@AGZ2009 Chapter 5]). Suppose that $(X,Y)$ are non-commutative random variables and $S_t$ is a freely independent free semicircular $m$-tuple with law $\sigma_{m,t}^{(N)}$. We may assume that $(S_t)_{t \geq 0}$ is a free Brownian motion, so that $S_t - S_s \sim S_{t - s}$ for $0 \leq s \leq t$ and $S_t \sim t^{1/2} S_1$. Note that $e^{-tL_x/2}$ is a well-defined operator on trace polynomials. To prove , it suffices to show that $[e^{tL_x/2}f](X+S_t,Y) = f(X,Y)$ for $f \in \operatorname{TrP}_m^0$. This will follow if we check that $$\frac{d}{dt} \left( [e^{-tL_x/2} f](X+S_t,Y) \right) = 0.$$ From a free probabilistic computation sketched in [@Jekel2018 Lemma 3.23], we have $$\frac{d}{dt} f(X+S_t,Y) = \frac{1}{2} [L_x f](X+S_t,Y),$$ and hence $$\begin{aligned}
\frac{d}{dt} \left( [e^{-tL_x/2} f](X+S_t,Y) \right) &= \frac{d}{dt_1} [e^{-t_1 L_x/2} f](X+S_{t_2},Y)|_{t_1=t_2=t} + \frac{d}{dt_2} [e^{-t_1 L_x/2} f](X+S_{t_2},Y)|_{t_1=t_2=t} \\
&= \left[ \frac{-L_x}{2} e^{-tL_x/2} f \right](X+S_t,Y) + \left[ \frac{L_x}{2} e^{-tL_x/2} f \right](X+S_t,Y) \\
&= 0.\end{aligned}$$
Next, to prove , it suffices to show that for $g \in \operatorname{TrP}_{m+n}^1$, we have $$\tau(f(X+S_t,Y) g(X,Y)) = \tau([e^{tL_x/2}f](X,Y) g(X,Y)),$$ since functions of the form $g(X,Y)$ for $g \in \operatorname{TrP}_m^1$ are dense in $L^2(\mathrm{W}^*(X,Y))$. Consider the function $F \in \operatorname{TrP}_{m+n+m}$ given by $F(x,y,x') = \tau(f(x,y) g(x',y))$. Notice that $$L_x F(x,y,x') = \tau(L_x[f(x,y) g(x',y)]) = \tau([L_x f(x,y)] g(x',y)).$$ Here the first equality is checked directly from the definition of the Laplacian [@Jekel2018 see Def. 3.13 and 3.16, proof of Lemma 3.18]. The equality $L_x[f(x,y) g(x',y)] = L_x[f(x,y)] g(x',y)$ again is checked from the definition of the Laplacian; this equality is intuitive since $g(x',y)$ is independent of $x$. Since the same reasoning may be applied to compute the Laplacian $L_x$ of $\tau([e^{tL_x/2}f](x,y) g(x,y))$, we have $$e^{tL_x/2} F(x,y,x') = \tau([e^{tL_x/2}f](x,y) g(x',y)).$$ We can view $F(x,y,x')$ as a function of the $m$-tuple $x$ and the $(n+m)$-tuple $(y,x')$, that is, an element of $\operatorname{TrP}_{m+(n+m)}^1$. We apply to $f$ and the pair $(X, (Y,X))$ and obtain $$F(X+S_t,Y,X) = e^{tL_x/2}F(X,Y,X)$$ which means precisely that $$\tau(f(X+S_t,Y) g(X,Y)) = \tau([e^{tL_x/2}f](X,Y) g(X,Y)),$$ which completes the proof of .
The free conditional expectation formulas and could also be proved using random matrices provided that $\mathrm{W}^*(X,Y)$ is $\mathcal{R}^\omega$-embeddable. Indeed, let $(X^{(N)},Y^{(N)})$ be (deterministic) tuples of matrices with non-commutative laws converging to the law of $(X,Y)$ and let $S^{(N)} \sim \sigma_{m,t}^{(N)}$. Then to prove for instance, we could use the fact that $E[f(X^{(N)}+S^{(N)},Y^{(N)}) = [e^{tL_x^{(N)}/2}f](X^{(N)},Y^{(N)})$ and take the limit as $N \to \infty$ using Voiculescu’s theorem on asymptotic freeness [@Voiculescu1998 Theorem 2.2]. A similar proof could be done for .
If $f \in \operatorname{TrP}_{m+n}^j$ for $j = 0, 1$, then we have $\norm*{e^{tL_x/2} f}_{u,R} \leq \norm*{f}_{u,R+2t^{1/2}}$ for $t \geq 0$. In particular, $f \mapsto e^{tL_x/2} f$ extends to a unique continuous linear operator $\overline{\operatorname{TrP}}_{m+n}^j \to \overline{\operatorname{TrP}}_{m+n}^j$.
Let $(X,Y) \in (\mathcal{R}^\omega)_{sa}^{m+n}$ with $\norm*{(X,Y)}_\infty \leq R$. Let $S \sim \sigma_{m,t}$ be a freely independent semicircular tuple. If $f \in \operatorname{TrP}_{m+n}^0$, then $$[e^{tL_x/2} f](X,Y) = f(X+S,Y).$$ Since $\norm*{S}_\infty = 2t^{1/2}$, we have $\norm*{(X+S,Y)}_\infty \leq R + 2t^{1/2}$. Therefore, $\norm*{e^{-tL_x/2} f}_{u,R} \leq \norm*{f}_{u,R+2t^{1/2}}$ as desired. Similarly, if $f \in \operatorname{TrP}_{m+n}^1$, then we check $\norm*{e^{-tL_x/2} f}_{u,R} \leq \norm*{f}_{u,R+2t^{1/2}}$ using the conditional expectation formula . Now the continuous extension of $e^{tL_x/2}$ to $\overline{\operatorname{TrP}}_m^j$ is immediate.
The semigroup $e^{tL_x/2}$ acting on $\overline{\operatorname{TrP}}_{m+n}^1$ describes the large $N$ limit of the Gaussian convolution semigroup on $M_N({\mathbb{C}})_{sa}^N$ defined as follows.
For $f: M_N({\mathbb{C}})_{sa}^{m+n} \to {\mathbb{C}}$ or $M_N({\mathbb{C}})$, we denote $$P_t^{(N)} f(x,y) = \int f(x+z,y)\,d\sigma_{m,t}^{(N)}(z).$$ Moreover, we denote by $P_t^{\operatorname{TrP}}: \overline{\operatorname{TrP}}_m^j \to \overline{\operatorname{TrP}}_m^j$ the continuous extension of $e^{tL_x/2}$.
\[lem:Gaussianconvolution\] Suppose that $f^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to {\mathbb{C}}$ is asymptotically approximable by trace polynomials and $f^{(N)} \rightsquigarrow f \in \overline{\operatorname{TrP}}_m^0$. Furthermore, assume that for some $A, B > 0$ and $k \in {\mathbb{N}}$, we have $$\label{eq:growthbound}
\norm*{f}_{u,R} \leq A + B R^k \\
\sup_{\substack{x \in M_N({\mathbb{C}})_{sa}^m \\ \norm*{x}_\infty \leq R}} |f^{(N)}(x)| \leq A + BR^k.$$ Then $P_t^{(N)} f^{(N)} \rightsquigarrow P_t^{\operatorname{TrP}} f$. The same holds for $f^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})$ and $f \in \overline{\operatorname{TrP}}_m^1$ with $|f^{(N)}(x)|$ replaced by $\norm*{f^{(N)}(x)}_2$.
The proof of this lemma is the same as in [@Jekel2018 Lemma 3.28].
\[rem:UCgrowthbound\] In both the scalar-valued and matrix-valued cases, the assumption holds automatically with $k = 1$ provided that $f^{(N)}$ and $f$ are $\norm*{\cdot}_2$-uniformly continuous (with modulus of continuity independent of $N$). Let us focus on the matrix-valued case of $\operatorname{TrP}_m^1$, there exists $\delta > 0$ such that $$\norm*{x-y}_2 < \delta \implies \norm*{f(x) - f(y)}_2 \leq 1.$$ In particular, given $x \in (\mathcal{R}^\omega)_{sa}^m$, we can choose an integer $j$ such that $j \delta < \norm*{x}_2 \leq 2 j \delta$. Then we have $$\norm*{f(x) - f(0)}_2 \leq \sum_{i=1}^{2j} \norm*{f(ix/2k) - f((i-1)x/2k)}_2 \leq 2j \leq 2 \norm*{x}_2 / \delta.$$ Thus, $$\norm*{f(x)}_2 \leq \norm*{f(0)}_2 + \frac{2}{\delta} \norm*{x}_2 \leq \norm*{f(0)}_2 + \frac{2m^{1/2}}{\delta} \norm*{x}_\infty,$$ which implies the first estimate of . The case for $f^{(N)}$ is handled similarly, and we note that $\norm*{f^{(N)}(0)}_2$ is bounded as $N \to \infty$ because of our assumption that $f^{(N)} \rightsquigarrow f$. The same argument works in the case of scalar-valued functions and $f \in \overline{\operatorname{TrP}}_m^0$.
Conditional Expectation for Free Gibbs States {#sec:conditionalexpectation}
=============================================
Free Gibbs States from Convex Potentials {#subsec:CEmotivation}
----------------------------------------
In [@Jekel2018] and in the present work, we focus on the following situation:
\[ass:convexRMM\] We are given $0 < c \leq C$ and $V^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ such that
1. $HV^{(N)} \geq c$, that is, $V^{(N)}(x) - \frac{1}{2} c \norm*{x}_2^2$ is convex.
2. $HV^{(N)} \leq C$, that is, $V^{(N)}(x) - \frac{1}{2} C \norm*{x}_2^2$ is concave.
3. $\{DV^{(N)}\}_{N \in {\mathbb{N}}}$ is asymptotically approximable by trace polynomials.
We denote by $\mu^{(N)}$ the probability measure on $M_N({\mathbb{C}})_{sa}^m$ given by $$d\mu^{(N)}(x) = e^{-N^2 V^{(N)}(x)}\,dx.$$ Furthermore, we assume that the mean $\int x_j \,d\mu^{(N)}(x)$ is a scalar multiple of the identity matrix.
The following was proved in [@Jekel2018 Theorem 4.1].
\[thm:freeGibbslaw\] Let $V^{(N)}$ and $\mu^{(N)}$ be as in Assumption \[ass:convexRMM\]. Then there exists a non-commutative law $\lambda$ such that for every non-commutative polynomial $p$, we have $$\lambda(p) = \lim_{N \to \infty} \int \tau_N(p(x))\,d\mu^{(N)}(x).$$ Moreover, we have for every $R > 0$ and $\epsilon > 0$ that $$\lim_{N \to \infty} \frac{1}{N^2} \log \mu_N\left(\left\{x \in M_N({\mathbb{C}})_{sa}^m: \norm*{x}_\infty \leq R, |\tau(p(x)) - \lambda(p)| > \epsilon \right\}\right) < 0$$
\[cor:convergenceofexpectation\] Let $\mu^{(N)}$ and $\lambda$ be as in Theorem \[thm:freeGibbslaw\]. Let $X^{(N)}$ be a random $m$-tuple of matrices distributed according to $\mu^{(N)}$ and let $X$ be a non-commutative random $m$-tuple distributed according to $\lambda$. Let $f^{(N)}, g^{(N)}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$. Suppose there are constants $A$ and $B > 0$ and $k \in {\mathbb{N}}$ such that $$\max(\norm*{f^{(N)}(x)}_2, \norm*{g^{(N)}(x)}_2) \leq A + B \norm*{x}_\infty^k$$ Suppose that $f^{(N)} \rightsquigarrow f$ and $g^{(N)} \rightsquigarrow g$ where $f, g \in \overline{\operatorname{TrP}}_m^1$. Then $$\lim_{N \to \infty} E[\tau_N(f^{(N)}(X^{(N)}) g^{(N)}(X^{(N)}))] = \tau(f(X) g(X)).$$
Let $a_j^{(N)} = E[X_j^{(N)}]$ which we assumed to be a scalar multiple of the identity, and which we know has a limit as $N \to \infty$. By Lemma \[lem:epsilonnet\], we have $$P(\norm*{X_j^{(N)} - a_j^{(N)}}_\infty \geq c^{-1/2} \Theta + \delta) \leq e^{-cN \delta^2 / 2}.$$ In particular, letting $R > \sup_{N,j} |a_j^{(N)}| + c^{-1/2} \Theta$, we have $$P(\norm*{X^{(N)}}_\infty \geq R) \to 0$$ and $$E[\mathbf{1}_{\norm*{X^{(N)}}_\infty \geq R} \tau_N(f^{(N)}(X^{(N)}) g^{(N)}(X^{(N)}))] \to 0.$$ Therefore, in order to prove convergence of the expectation, it suffices to check that $\tau_N(f^{(N)}(X^{(N)}) g^{(N)}(X^{(N)}))$ converges in probability to $\tau(f(X) g(X))$.
We already know that $\tau_N(p(X^{(N)}))$ converges to $\tau(p(X))$ in probability for every non-commutative polynomial $p$. It follows that if $u$ is a scalar-valued trace polynomial, then $u(X^{(N)}) \to u(X)$ in probability. This also holds for $u \in \overline{\operatorname{TrP}}_m^1$; indeed, we know that $\norm*{X^{(N)}}_\infty \leq R$ with probability tending to $1$ and $\norm*{X}_\infty \leq R$, whereas $u$ can be approximated in $\norm*{\cdot}_{u,R}$ by trace polynomials. Finally, if $u^{(N)}$ is a sequence of scalar-valued function such that $u^{(N)} \rightsquigarrow u \in \overline{\operatorname{TrP}}_m^0$, then $u^{(N)}(X^{(N)}) - u(X^{(N)})$ converges to $0$ in probability, and hence $u^{(N)}(X^{(N)})$ converges in probability to $u(X)$. By Lemma \[lem:algebra\], we can apply this statement to $u^{(N)} = \tau_N(f^{(N)} g^{(N)})$ and $u = \tau(fg)$, which completes the argument.
Let $V \in \overline{\operatorname{TrP}}_m^0$ and suppose $V$ extends to a function $L^2(\mathcal{R}^\omega)_{sa}^m \to {\mathbb{R}}$ such that $V(x) - (c/2) \norm*{x}_2^2$ is convex and $V(x) - (C/2) \norm*{x}_2^2$ is concave. In this case, $V$ is differentiable as a function on the real Hilbert space $L^2(\mathcal{R}^\omega)_{sa}^m$, as a consequence of the existence of supporting hyperplanes for convex functions on a Hilbert space. If we assume also that $DV \in \overline{\operatorname{TrP}}_m^1$, then we say that $V \in \mathcal{E}_m^{\operatorname{TrP}}(c,C)$.
We did not prove or assume that the trace polynomials which approximate $DV$ are the gradients of the *same* trace polynomials that approximate $V$. Thus, this definition is technically different from that of [@Jekel2018 §8.2].
If $V \in \mathcal{E}_m^{\operatorname{TrP}}(c,C)$, then we may define $V^{(N)} = V|_{M_N({\mathbb{C}})_{sa}^m}$, and in this case $DV^{(N)} = DV|_{M_N({\mathbb{C}})_{sa}^m}$. Clearly, $DV^{(N)}$ is asymptotically approximable by trace polynomials, and so by Theorem \[thm:freeGibbslaw\], there exists a non-commutative law $\lambda_V$ that arises as the large $N$ limit of the associated random matrix models. Furthermore, the limiting free Gibbs law $\lambda_V$ only depends on $V$, that is, every approximating sequence of functions $V^{(N)} \in \mathcal{E}_m^{(N)}(c,C)$ will produce the same free Gibbs law (see [@Jekel2018 §8.2]). We call $\lambda_V$ the *free Gibbs state given by potential $V$*.
One can check that if $V^{(N)}$ is as in Assumption \[ass:convexRMM\], then there exists a $V \in \mathcal{E}_m^{\operatorname{TrP}}(c,C)$ such that $V^{(N)} \rightsquigarrow V$ and $DV^{(N)} \rightsquigarrow DV$. Thus, the non-commutative laws that arise from these random matrix models are precisely $\lambda_V$ for $V \in \mathcal{E}_m^{\operatorname{TrP}}(c,C)$.
\[rem:unitaryinvariance\] Since $\lambda_V$ is independent of the choice of approximating sequence $V^{(N)}$, we can in particular take $V^{(N)} = V|_{M_N({\mathbb{C}})_{sa}^m}$, which produces a canonical unitarily invariant sequence of random matrices models.
Main Result on Conditional Expectation
--------------------------------------
Our main result in this section is in some sense a generalization of [@Jekel2018 Theorem 4.1], which deals with conditional expectations rather than expectations. The proof of the earlier theorem was reduced to the following statement: Suppose $V^{(N)}$ satisfies Assumption \[ass:convexRMM\] and that $u^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$ is $\norm*{\cdot}_2$-Lipschitz (uniformly in $N$) and asymptotically approximable by trace polynomials. Then $$\lim_{N \to \infty} \int u^{(N)}\,d\mu^{(N)} \text{ exists.}$$ Now, our goal is to prove the following.
\[thm:conditionalexpectation\] Consider functions $V^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to {\mathbb{R}}$, denoted as $V^{(N)}(x,y)$, which satisfy Assumption \[ass:convexRMM\] as functions of $(x,y)$. Let $\mu^{(N)}$ be the associated probability measure on $M_N({\mathbb{C}})_{sa}^{m+n}$. Let $(X^{(N)},Y^{(N)})$ be an $(m+n)$-tuple of random matrices distributed according to $\mu^{(N)}$, and let $(X,Y)$ be a $(m+n)$-tuple of non-commutative random variables distributed according to the limiting free Gibbs law $\lambda$ given by Theorem \[thm:freeGibbslaw\]
Let $f^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})$ be $\norm*{\cdot}_2$-Lipschitz (uniformly in $N$) and suppose $f^{(N)} \rightsquigarrow f \in \overline{\operatorname{TrP}}_{m+n}^1$. Let $g^{(N)}$ be the function given by $$g^{(N)}(Y^{(N)}) = E[f^{(N)}(X^{(N)}, Y^{(N)}) | Y^{(N)}],$$ which is well-defined function $M_N({\mathbb{C}})_{sa}^m$ because $\mu^{(N)}$ has positive density everywhere. Then $g^{(N)}$ is Lipschitz with $$\norm*{g^{(N)}}_{\operatorname{Lip}} \leq (1 + C/c) \norm*{f^{(N)}}_{\operatorname{Lip}}.$$ Moreover, there exists $g \in \overline{\operatorname{TrP}}_m^1$ such that $g^{(N)} \rightsquigarrow g$ and hence $$g(Y) = E_{\mathrm{W}^*(Y)}[f(X,Y)].$$
The gist of the theorem is that the conditional expectation $E[ \cdot | Y^{(N)}]$ behaves in the large $N$ limit like the $\mathrm{W}^*$-algebraic expectation $\mathrm{W}^*(X,Y) \to \mathrm{W}^*(Y)$. For instance, if $f \in \overline{\operatorname{TrP}}_{m+n}^1$ is globally Lipschitz in $\norm*{\cdot}_2$, then the $\mathrm{W}^*$-algebraic conditional expectation of $f(X,Y)$ can be approximated by the classical conditional expectation $E[f(X^{(N)}, Y^{(N)}) | Y^{(N)}]$.
In fact, we can approximate $E_{\mathrm{W}^*}(Z)$ for every $Z \in L^2(\mathrm{W}^*(X,Y))$ using classical conditional expectations in the same sense. Indeed, we showed in Proposition \[prop:realizationofoperators\] that every $Z$ can be expressed as $f(X,Y)$ where $f \in \overline{\operatorname{TrP}}_m^1$ is $\norm*{\cdot}_2$-uniformly continuous, and there exist $\norm*{\cdot}_2$-Lipschitz functions $f_k \in \overline{\operatorname{TrP}}_m^1$ such that $f_k \to f$ with respect to the uniform norm $\norm*{\cdot}_u$. Let $g_k^{(N)}$ and $g^{(N)}$ be given by $$g_k^{(N)}(Y^{(N)}) = E[f_k(X^{(N)}, Y^{(N)}) | Y^{(N)}],$$ and the analogous relation for $g^{(N)}$ and $f$. Because conditional expectation is a contraction in $L^\infty(\mu^{(N)})$ (for functions taking values in $M_N({\mathbb{C}})$ with $\norm*{\cdot}_2$), we have $$\norm*{g_k^{(N)} - g^{(N)}}_u^{(N)} = \norm*{g_k^{(N)} - g^{(N)}}_{L^\infty(\mu^{(N)})} \leq \norm*{f_k - f}_u$$ By the theorem, there exists $g_k \in \overline{\operatorname{TrP}}_m^1$ such that $g_k^{(N)} \rightsquigarrow g_k$. Given that $\norm*{g_k^{(N)} - g^{(N)}}_u^{(N)} \leq \norm*{f_k - f}_u \to 0$, a routine argument (“exchange of limits and uniform limits”) shows that there exists $g \in \overline{\operatorname{TrP}}_m^1$ such that $g^{(N)} \rightsquigarrow g$. In other words, the conclusion of Theorem \[thm:conditionalexpectation\] holds also for $f$ and thus $E_{\mathrm{W}^*(Y)}[Z] = E_{\mathrm{W}^*(Y)}[f(X,Y)]$ can be viewed as the large $N$ limit of $E[f(X^{(N)},Y^{(N)}) | Y^{(N)}]$.
Strategy {#subsec:conditionalexpectationstrategy}
--------
Our proof will follow the same strategy as the special case in [@Jekel2018 §4]. In that paper, we showed that if $V^{(N)}$ and $\mu^{(N)}$ on $M_N({\mathbb{C}})_{sa}^m$ are as in Assumption \[ass:convexRMM\] and if $u^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{C}}$ is uniformly Lipschitz and asymptotically approximable by trace polynomials, then $\lim_{N \to \infty} \int u^{(N)}\,d\mu^{(N)}$ exists.
We considered the diffusion semigroup $T_t^{(N)} = T_t^{V^{(N)}}$ that solves the equation $$\partial_t (T_t^{(N)} u^{(N)}) = \frac{1}{2N} \Delta (T_t^{(N)} u^{(N)}) - \frac{1}{2} \ip{DV^{(N)}, D(T_t^{(N)} u^{(N)})}_2.$$ As mentioned in [@Jekel2018 §4], this diffusion semigroup has an equivalent SDE formulation, and is a standard tool in proving the log-Sobolev inequality and concentration estimates (see for instance, [@Ledoux1992], [@AGZ2009 §4.4.2], [@DGS2016]). Now $\int T_t^{(N)} u^{(N)}\,d\mu^{(N)} = \int u^{(N)}\,d\mu^{(N)}$ and $\norm*{T_t^{(N)} u^{(N)}}_{\operatorname{Lip}} \leq e^{-ct/2} \norm*{u^{(N)}}_{\operatorname{Lip}}$. As $t \to \infty$, the function $T_t^{(N)} u^{(N)}$ converges to the constant function $\int u^{(N)}\,d\mu^{(N)}$ at a rate independent of $N$. On the other hand, we showed in [@Jekel2018 Lemma 4.10] that if $\{u^{(N)}\}_{N \in {\mathbb{N}}}$ and $\{DV^{(N)}\}_{N \in {\mathbb{N}}}$ are asymptotically approximable by trace polynomials, then so is $\{T_t^{(N)} u^{(N)}\}_{N \in {\mathbb{N}}}$. Hence, we concluded that the sequence of constant functions $\{ \int u^{(N)}\,d\mu^{(N)}\}$ is asymptotically approximable by trace polynomials, which means that the limit as $N \to \infty$ exists.
Now we apply the same method in the conditional setting to prove Theorem \[thm:conditionalexpectation\]. Let $V^{(N)}(x,y)$ be a function satisfying Assumption \[ass:convexRMM\]. If we fix $y$, then $V^{(N)}(\cdot,y)$ is uniformly convex and semi-concave function of $x$, so it defines a log-concave probability measure on $M_N({\mathbb{C}})_{sa}^m$. This produces a well-behaved conditional distribution of $X^{(N)}$ given $Y^{(N)}$, where $(X^{(N)},Y^{(N)}) \sim \mu^{(N)}$. Explicitly, for $f \in L^1(\mu^{(N)}, M_N({\mathbb{C}}))$, we have $$E[f(X^{(N)},Y^{(N)}) | Y^{(N)}] = \frac{\int f(x,Y^{(N)}) e^{-N^2 V^{(N)}(x,Y^{(N)})}\,dx}{\int e^{-N^2 V^{(N)}(x,Y^{(N)})}\,dx}.$$ We will evaluate this conditional expectation as the limit as $t \to \infty$ of $T_t^{(N)} f$, where $T_t^{(N)} = T_t^{V^{(N)}}$ is the semigroup, acting on Lipschitz functions of $(x,y)$, that solves $$\partial_t (T_t^{(N)} f) = \frac{1}{2N} \Delta_x (T_t^{(N)} f) - \frac{1}{2} J_x(T_t^{(N)} f)^* D_x V^{(N)},$$ where $J_x(T_t^{(N)} f)$ denotes the differential (Jacobian) of $T_t^{(N)} f$ as a function $x$ from $M_N({\mathbb{C}})_{sa}^m$ to $M_N({\mathbb{C}})$ and $*$ denotes the adjoint. In §\[subsec:relativediffusion\], we will analyze how $T_t^{(N)}$ affects the Lipschitz norms with respect to $x$ and $y$ separately and hence show that the conditional expectation is given by a Lipschitz function of $y$. In §\[subsec:relativediffusion2\], we will show that $T_t^{(N)}$ preserves asymptotic approximability by trace polynomials of $(x,y)$ and conclude our argument. The new aspect compared to [@Jekel2018] is that the functions are matrix-valued and depend on an extra parameter $y$.
Conditional Diffusion Semigroup {#subsec:relativediffusion}
-------------------------------
To simplify notation, let us fix $N$ and fix $V: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \to {\mathbb{R}}$ for the remainder of §\[subsec:relativediffusion\]. We will denote $$d\mu(x | y) = \frac{1}{\int e^{-N^2 V(x,y)}\,dx} e^{-N^2 V(x,y)}\,dx,$$ which is a measure on $M_N({\mathbb{C}})_{sa}^m$ depending on the parameter $y$. The associated semigroup $T_t$ will be approximated by alternating two other operators $P_t$ and $S_t$ on short time intervals. Let $P_t$ denote the semigroup of convolution with Gaussian with respect to $x$, that is, $$P_t f(x,y) = \int f(x+z,y)\,d\sigma_{t,m}^{(N)}(z).$$ The semigroup $S_t$ is given by $$S_t f(x,y) = f(W_t(x,y),y),$$ where $W_t: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \to M_N({\mathbb{C}})_{sa}^n$ is the solution to the initial value problem $$\begin{aligned}
W_0(x,y) &= x \\
\partial_t W_t(x,y) &= -\frac{1}{2} D_x V(W_t(x,y),y).\end{aligned}$$ This solution is defined for all $t \geq 0$ by the Picard-Lindelöf theorem because $D_x V(x,y)$ is globally Lipschitz in $x$ (compare §\[subsec:vectorfields\]).
\[prop:diffusionsemigroup\] There exists a semigroup $T_t$ acting on Lipschitz functions $M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n$ such that the following hold:
1. If $t = n / 2^\ell$ is a dyadic rational, let $T_{t,\ell} f = (P_{2^{-\ell}} S_{2^{-\ell}})^n$. Then $T_{t,\ell} f \to T_t f$ as $\ell \to \infty$ and more precisely $$\norm*{T_{t,\ell} f(\cdot,y) - T_t f(\cdot,y)}_{L^\infty} \leq \frac{Cm^{1/2}}{c(2 - 2^{1/2})} 2^{-\ell/2} \norm*{f(\cdot,y)}_{\operatorname{Lip}}.$$
2. If $0 \leq s \leq t$, we have $$\norm*{T_t f(x,y) - T_s f(x,y)}_2 \leq e^{-cs/2} \left( \frac{C}{c} (6 + 5 \sqrt{2})(t - s)^{1/2} + \norm*{D_x V(x,y)}_2 \right) \norm*{f(\cdot,y)}_{\operatorname{Lip}}.$$
3. $\norm*{T_t f(\cdot,y)}_{\operatorname{Lip}} \leq e^{-ct/2} \norm*{f(\cdot,y)}_{\operatorname{Lip}}$.
4. $\int T_t f(x,y)\,d\mu(x|y) = \int f(x,y)\,d\mu(x|y)$.
5. We have $T_t f(x,y) \to \int f(x',y)\,d\mu(x'|y)$ as $t \to \infty$ and specifically $$\norm*{T_t f(x,y) - \int f(x',y)\,d\mu(x'|y)}_2 \leq e^{-ct/2} \left( 4\frac{C}{c^2} (6 + 5 \sqrt{2}) t^{-1/2} + \frac{2}{c} \norm*{D_x V(x,y)}_2 \right) \norm*{f(\cdot,y)}_{\operatorname{Lip}}.$$
These results follow by freezing the variable $y$ and applying the results from our previous paper, specifically,
1. see [@Jekel2018 Lemma 4.5],
2. see [@Jekel2018 Lemma 4.6],
3. see [@Jekel2018 Lemma 4.6],
4. see [@Jekel2018 Lemma 4.8],
5. see [@Jekel2018 Lemma 4.9].
The results of [@Jekel2018 §4] were stated only for scalar-valued functions. However, the arguments hold for functions from $M_N({\mathbb{C}})_{sa}^m$ to any finite-dimensional normed vector space. The result (4) that $T_t$ is expectation-preserving follows immediately by applying the scalar-valued result to each coordinate of the vector-valued function in some basis. To verify the estimates, one simply replaces the “$|\cdot|$” in the arguments by the appropriate norm, which in our case would be $\norm*{\cdot}_2$ on $M_N({\mathbb{C}})$.
We will next show that $W_t(x,y)$ and $T_t f(x,y)$ depend in a Lipschitz manner upon $y$. Let us denote $$\begin{aligned}
\norm*{f}_{\operatorname{Lip}(dx)} &= \sup_y \norm*{f(\cdot,y)}_{\operatorname{Lip}} \\
\norm*{f}_{\operatorname{Lip}(dy)} &= \sup_x \norm*{f(x,\cdot)}_{\operatorname{Lip}}.\end{aligned}$$
With the setup above, we have for Lipschitz $f: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \to M_N({\mathbb{C}})$
1. $\norm*{W_t}_{\operatorname{Lip},dx} \leq e^{-ct/2}$ and $\norm*{W_t}_{\operatorname{Lip},dy} \leq (C/c)(1 - e^{-ct/2})$.
2. $\norm*{S_t f}_{\operatorname{Lip},dx} \leq e^{-ct/2} \norm*{f}_{\operatorname{Lip},dx}$.
3. $\norm*{S_t f}_{\operatorname{Lip},dy} \leq \norm*{f}_{\operatorname{Lip},dy} + (C/c)(1 - e^{-ct/2}) \norm*{f}_{\operatorname{Lip},dx}$.
4. $\norm*{P_t f}_{\operatorname{Lip},dy} \leq \norm*{f}_{\operatorname{Lip},dy}$ and $\norm*{P_t f}_{\operatorname{Lip},dx} \leq \norm*{f}_{\operatorname{Lip},dx}$.
5. $\norm*{T_t f}_{\operatorname{Lip},dx} \leq e^{-ct/2} \norm*{f}_{\operatorname{Lip},dx}$.
6. $\norm*{T_t f}_{\operatorname{Lip},dy} \leq \norm*{f}_{\operatorname{Lip},dy} + (C/c)(1 - e^{-ct/2}) \norm*{f}_{\operatorname{Lip},dx}$.
\(1) Fix $x, x' \in M_N({\mathbb{C}})_{sa}^m$ and $y, y' \in M_N({\mathbb{C}})_{sa}^n$. Define $$\phi(t) = \norm*{W_t(x,y) - W_t(x',y')}_2.$$ Note that $\phi$ is locally Lipschitz in $t$ and hence absolutely continuous. Moreover, $\phi(t)^2$ is $C^1$ with $$\begin{aligned}
\partial_t [\phi(t)^2] &= 2 \ip{\partial_t W_t(x,y) - \partial_t W_t(x',y'), W_t(x,y) - W_t(x',y')}_2 \\
&= -\ip{D_x V(W_t(x,y),y) - D_x V(W_t(x',y'),y'), W_t(x,y) - W_t(x',y')}_2 \\
&= -\ip{D_x V(W_t(x,y),y) - D_x V(W_t(x',y'),y), W_t(x,y) - W_t(x',y')}_2 \\
& \quad - \ip{D_x V(W_t(x',y'),y) - D_x V(W_t(x',y'),y'), W_t(x,y) - W_t(x',y')}_2 \\
&\leq -c \norm*{W_t(x,y) - W_t(x',y')}_2^2 \\
& \quad + \norm*{D_x V(W_t(x',y'),y) - D_x V(W_t(x',y'),y)}_2 \norm*{W_t(x,y) - W_t(x',y')}_2 \\
&\leq -c \norm*{W_t(x,y) - W_t(x',y')}_2^2 + C \norm*{y - y'}_2 \norm*{W_t(x,y) - W_t(x',y')}_2.\end{aligned}$$ Here we have employed the inequality $\ip{D_xV(z,w) - D_x V(z',w), z - z'}_2 \geq c \norm{z - z'}_2^2$ coming from the uniform convexity of $V$ as well as the Cauchy-Schwarz inequality. This implies that $$2 \phi'(t) \phi(t) = \partial_t [\phi(t)^2] \leq -c \phi(t)^2 + C \norm*{y - y'} \phi(t).$$ Thus, $\phi'(t) \leq -(c/2) \phi(t) + (C/2) \norm*{y - y'}$, so that $\partial_t[e^{ct/2} \phi(t)] \leq (C/2) e^{ct/2} \norm*{y - y'}_2$. This implies that $$e^{ct/2} \phi(t) - \phi(0) \leq \frac{C}{c} (e^{ct/2} - 1) \norm*{y - y'}_2.$$ But $\phi(t) = \norm*{W_t(x,y) - W_t(x',y')}_2$ and $\phi(0) = \norm*{x - x'}_2$. Hence, $$\norm*{W_t(x,y) - W_t(x',y')}_2 \leq e^{-ct/2} \norm*{x - x'}_2 + \frac{C}{c}(1 - e^{-ct/2}) \norm*{y - y'}_2.$$ This proves both estimates of (1).
\(2) This is immediate since $S_t f(x,y) = f(W_t(x,y),y)$, as in [@Jekel2018 Lemma 4.4 (5)].
\(3) Note that $$\begin{aligned}
\norm*{S_t f(x,y) - S_t f(x,y')}_2 &= \norm*{f(W_t(x,y),y) - f(W_t(x,y'),y')}_2 \\
&\leq \norm*{f(W_t(x,y),y) - f(W_t(x,y),y')}_2 + \norm*{f(W_t(x,y),y') - f(W_t(x,y'),y')}_2 \\
&\leq \norm*{f}_{\operatorname{Lip},dy} \norm*{y - y'}_2 + \norm*{f}_{\operatorname{Lip},dx} \norm*{W_t(x,y) - W_t(x,y')}_2 \\
&\leq \norm*{f}_{\operatorname{Lip},dy} \norm*{y - y'}_2 + \frac{C}{c}(1 - e^{-ct/2}) \norm*{f}_{\operatorname{Lip},dx} \norm*{y - y'}_2.\end{aligned}$$
\(4) This follows from basic properties of convolution of a function with a probability measure.
\(5) By iterating the estimates (2) and (4), we obtain $\norm*{T_{t,\ell} f}_{\operatorname{Lip},dx} \leq e^{-ct/2} \norm*{f}_{\operatorname{Lip},dx}$. Then by Proposition \[prop:diffusionsemigroup\] (2) and (3) we may take $\ell \to \infty$ and then extend to all real values of $t \geq 0$.
\(6) First, consider $T_{t,\ell}$ for a dyadic rational $t = n / 2^\ell$. Denote $\delta = 2^{-\ell}$. For $j = 0$, …, $n-1$, we have $$\begin{aligned}
\norm*{T_{(j+1)\delta,\ell}f}_{\operatorname{Lip},dy} &= \norm*{P_\delta S_\delta T_{j \delta,\ell} f}_{\operatorname{Lip},dy} \\
&\leq \norm*{S_\delta T_{j \delta,\ell} f}_{\operatorname{Lip},dy} \\
&\leq \norm*{T_{j \delta,\ell} f}_{\operatorname{Lip},dy} + \frac{C}{c}(1 - e^{-c\delta/2}) \norm*{T_{j \delta, \ell} f}_{\operatorname{Lip},dx},\end{aligned}$$ where the last inequality follows from (3). Therefore, by induction $$\begin{aligned}
\norm*{T_{t,\ell} f}_{\operatorname{Lip},dy} &= \norm*{T_{n\delta,\ell} f}_{\operatorname{Lip},dy} \\
&\leq \norm*{f}_{\operatorname{Lip},dy} + \frac{C}{c}(1 - e^{-c\delta/2}) \sum_{j=0}^{n-1} \norm*{T_{j\delta,\ell} f}_{\operatorname{Lip},dx} \\
&\leq \norm*{f}_{\operatorname{Lip},dy} + \frac{C}{c}(1 - e^{-c \delta/2}) \sum_{j=0}^{n-1} e^{-c\delta j / 2} \norm*{f}_{\operatorname{Lip},dx} \\
&= \norm*{f}_{\operatorname{Lip},dy} + \frac{C}{c}(1 - e^{-ct/2}) \norm*{f}_{\operatorname{Lip},dx}.\end{aligned}$$ In light of Proposition \[prop:diffusionsemigroup\] (1), we can take $\ell \to \infty$ and conclude that $\norm*{T_t f}_{\operatorname{Lip},dy} \leq \norm*{f}_{\operatorname{Lip},dy} + (C/c)(1 - e^{-ct/2}) \norm*{f}_{\operatorname{Lip},dx}$ for dyadic rational $t$. This inequality can then be extended to all real $t \geq 0$ by Proposition \[prop:diffusionsemigroup\] (2).
\[cor:conditionalexpectationLipschitz\] Let $f: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \to M_N({\mathbb{C}})$ be Lipschitz with respect to $\norm*{\cdot}_2$. Let $g(y) = \int f(x,y)\,d\mu(x|y)$. Then $g$ is Lipschitz with $$\norm*{g}_{\operatorname{Lip}} \leq (1 + C/c) \norm*{f}_{\operatorname{Lip}}.$$
By the previous lemma, $$\begin{aligned}
\norm*{T_t f}_{\operatorname{Lip},dy} &\leq \norm*{f}_{\operatorname{Lip},dy} + \frac{C}{c}(1 - e^{-ct/2}) \norm*{f}_{\operatorname{Lip},dx} \\
&\leq \norm*{f}_{\operatorname{Lip}} + \frac{C}{c} \norm*{f}_{\operatorname{Lip}}.\end{aligned}$$ As $t \to \infty$, we have $T_t f(x,y) \to g(y)$ by Proposition \[prop:diffusionsemigroup\] (5). Hence, $\norm*{g}_{\operatorname{Lip}} \leq (1 + C/c) \norm*{f}_{\operatorname{Lip}}$.
Asymptotic Approximation and Convergence {#subsec:relativediffusion2}
----------------------------------------
Let $V^{(N)}$ and $\mu^{(N)}$ be as in Theorem \[thm:conditionalexpectation\], let $(X^{(N)},Y^{(N)})$ be a random variable with distribution $\mu^{(N)}$. Let $\mu^{(N)}(x|y)$ denote the conditional distribution of $X^{(N)}$ given $Y^{(N)}$.
Let $P_t^{(N)}$, $S_t^{(N)}$, and $T_t^{(N)}$ be the semigroups acting on Lipschitz functions defined as in §\[subsec:relativediffusion\] with respect to the potential $V^{(N)}$.
\[lem:diffusionAATP\] With the notation above, suppose that $f^{(N)}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})$, that $f^{(N)}$ is $K$-Lipschitz for every $N$, and that $f^{(N)}$ is asymptotically approximable by trace polynomials. Then
1. $\{P_t^{(N)} f^{(N)}\}$ is asymptotically approximable by trace polynomials.
2. $\{S_t^{(N)} f^{(N)}\}$ is asymptotically approximable by trace polynomials.
3. $\{T_t^{(N)} f^{(N)}\}$ is asymptotically approximable by trace polynomials.
\(1) We proved in Lemma \[lem:Gaussianconvolution\] that $P_t^{(N)}$ preserves asymptotic approximability by trace polynomials.
\(2) Recall that $S_t^{(N)} f^{(N)}(x,y) = f^{(N)}(W_t^{(N)}(x,y),y)$, where $$\begin{aligned}
W_0^{(N)}(x,y) &= x \\
\partial_t W_t^{(N)}(x,y) &= -\frac{1}{2} D_x V^{(N)}(W_t(x,y),y).\end{aligned}$$ Now $D_xV^{(N)}(x,y)$ is $C$-Lipschitz in $(x,y)$, asymptotically approximable by trace polynomials, and independent of $t$, and thus it satisfies Assumption \[ass:vectorfield2\], so by Proposition \[prop:ODE2\], $W_t^{(N)}(x,y)$ is asymptotically approximable by trace polynomials (here we rely on Lemma \[lem:AATP\] that asymptotic approximability is equivalent to being asymptotic to some element of $\overline{\operatorname{TrP}}_{m+n}^1$). Then because $f^{(N)}$ is $K$-Lipschitz in $(x,y)$, Lemma \[lem:composition\] implies asymptotic approximability of $f^{(N)}(W_t^{(N)}(x,y),y)$.
\(3) Let $T_{t,\ell}^{(N)} = (P_{2^{-\ell}}^{(N)} S_{2^{-\ell}}^{(N)})^n$ whenever $t = n 2^{-\ell}$. From (1) and (2), it follows that $T_{t,\ell}^{(N)} f^{(N)}$ is asymptotically approximable by trace polynomials. Now for each dyadic $t$, Proposition \[prop:diffusionsemigroup\] (1) shows that $T_{t,\ell}^{(N)} f^{(N)} \to T_t^{(N)} f^{(N)}$ uniformly on $\norm*{\cdot}_2$-balls (and hence on $\norm*{\cdot}_\infty$). Therefore, by Lemma \[lem:limits\], $T_t^{(N)} f^{(N)}$ is asymptotically approximable by trace polynomials. Then we extend this property from dyadic $t$ to all real $t$ using Proposition \[prop:diffusionsemigroup\] (2) and Lemma \[lem:limits\].
Let $f^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})$ be $K$-Lipschitz and asymptotically approximable by trace polynomials. Let $$g^{(N)}(y) = \int f^{(N)}(x,y) \,d\mu^{(N)}(x|y)$$ We showed in Corollary \[cor:conditionalexpectationLipschitz\] that $g^{(N)}$ is Lipschitz with $\norm*{g^{(N)}}_{\operatorname{Lip}} \leq (1 + C/c) \norm*{f^{(N)}}_{\operatorname{Lip}}$. We know that $T_t^{(N)} f^{(N)}$ is asymptotically approximable by trace polynomials in $(x,y)$. By Proposition \[prop:diffusionsemigroup\] (5), we have $T_t^{(N)} f^{(N)}(x,y) \to g^{(N)}(x,y)$ as $t \to \infty$, with the error bounded by $$e^{-ct/2} \left( 4\frac{C}{c^2} (6 + 5 \sqrt{2}) t^{-1/2} + \frac{2}{c} \norm*{D_x V(x,y)}_2 \right) \norm{f^{(N)}}_{\operatorname{Lip}}.$$ Given that $\{DV^{(N)}\}$ is asymptotically approximable by trace polynomials, $\norm*{D_x V}_{u,R}^{(N)}$ is bounded as $N \to \infty$. This implies that the rate of convergence of $T_t^{(N)} f^{(N)}(x,y) \to g^{(N)}(x,y)$ as $t \to \infty$ is uniform on $\norm*{(x,y)}_\infty \leq R$ and independent of $N$. So by Lemma \[lem:limits\], $g^{(N)}$ is asymptotically approximable by trace polynomials of $(x,y)$. Yet $g^{(N)}$ is independent of $x$, and so we may approximate $g^{(N)}(y)$ by evaluating these trace polynomials at $(0,y)$, which reduces them to trace polynomials of $y$.
Since $g^{(N)}$ is asymptotically approximable by trace polynomials, let $g \in \overline{\operatorname{TrP}}_m^1$ such that $g^{(N)} \rightsquigarrow g$. Then it remains to show that $g(Y) = E_{\mathrm{W}^*(Y)}[f(X,Y)]$, where $(X,Y)$ are non-commutative random variables for the free Gibbs law $\lambda$ as in the theorem statement. It suffices to check that $$\tau(\phi(Y) g(Y)) = \tau(\phi(Y) f(X,Y))$$ whenever $\phi$ is a non-commutative polynomial. But using Corollary \[cor:convergenceofexpectation\], $$\begin{aligned}
\tau(\phi(Y) g(Y)) &= \lim_{N \to \infty} E[\tau_N(\phi(Y^{(N)}) g^{(N)}(Y))] \\
&= \lim_{N \to \infty} E[\tau_N(\phi(Y^{(N)}) f^{(N)}(X^{(N)}Y^{(N)}))] \\
&= \tau(\phi(Y) f(X,Y)). \qedhere\end{aligned}$$
We showed in §\[subsec:heatsemigroup\] that $P_t^{(N)}$ has a large $N$ limit $P_t^{\operatorname{TrP}}$ acting on $\overline{\operatorname{TrP}}_{m+n}^1$. Similarly, the results of §\[subsec:vectorfields\] imply that $S_t^{(N)}$ has a large $N$ limit $S_t^{\operatorname{TrP}}$ acting on $\overline{\operatorname{TrP}}_{m+n}^1$. This implies that the semigroup $T_t^{(N)}$ also has a large $N$ limit $T_t^{\operatorname{TrP}}$ in light of Proposition \[prop:diffusionsemigroup\] (1) and (2) and Lemma \[lem:limits\]. Future research should investigate in what sense $F(x,t) = T_t^{\operatorname{TrP}}f(x)$ would solve the differential equation $$\partial_t F = \frac{1}{2} L_x F - \frac{1}{2} (J_xF)^* (D_x V),$$ where $V$ is the large $N$ limit of $\{V^{(N)}\}$ and $J_xF$ is the Jacobian matrix of $F$ with respect to the variable $x$.
Conditional Entropy and Fisher’s Information {#sec:entropy}
============================================
In this section, we show that for random matrix models satisfying Assumption \[ass:convexRMM\], the conditional (classical) entropy $h(X^{(N)} | Y^{(N)})$ converges to the conditional non-microstates free entropy $\chi^*(X | Y)$ (also known as $\chi^*(X: \mathrm{W}^*(Y))$).
Conditional Entropy and Fisher’s Information in the Classical Setting
---------------------------------------------------------------------
We refer to [@Voiculescu2002 §3] and [@Jekel2018 §5] for background on classical entropy and Fisher’s information and motivation for the free case. The conditional setting is more technical, and we will state several standard results without proof, since the proofs in the non-conditional case were repeated in some detail in [@Jekel2018].
Recall that the classical entropy of a random variable $X$ in ${\mathbb{R}}^m$ with probability density $\rho$ is $h(X) = -\int \rho \log \rho$. Similarly, if $(X,Y)$ is a random variable in ${\mathbb{R}}^m \times {\mathbb{R}}^n$ with density $\rho_{X,Y}(x,y)$, then the *conditional entropy* $h(X|Y)$ is defined by $$\label{eq:classicalentropy}
h(X | Y) = \int_{{\mathbb{R}}^n} \int_{{\mathbb{R}}^m} \rho_{X|Y}(x|y) \log \rho_{X|Y}(x|y)\,dx \rho_Y(y)\,dy = \int_{{\mathbb{R}}^n \times {\mathbb{R}}^m} (\log \rho_{X|Y}(x|y)) \rho(x,y)\,dx\,dy,$$ where $\rho_Y$ is the marginal density $$\rho_Y(y) = \int_{{\mathbb{R}}^m} \rho_{X,Y}(x,y)\,dx$$ and $\rho_{X|Y}$ is the conditional density $$\rho_{X|Y}(x|y) = \frac{\rho_{X,Y}(x,y)}{\rho_Y(y)} \text{ defined when } \rho_Y(y) > 0.$$ It is a standard fact that if $X$ has finite variance, then $h(X|Y)$ is well-defined. The proof for the non-conditional entropy $h(X)$ was reviewed in [@Jekel2018 Lemma 5.1], and the conditional case can be handled similarly.
The *conditional Fisher information* given by $$\label{eq:classicalFisherinfo}
\mathcal{I}(X|Y) = \int_{{\mathbb{R}}^m \times {\mathbb{R}}^n} \left| \frac{\nabla_x \rho_{X|Y}(x|y)}{\rho_{X|Y}(x|y)} \right|^2 \rho_{X,Y}(x,y)\,dx\,dy,$$ whenever the right hand side makes sense and $\infty$ otherwise. It describes the rate of change of $h(X+t^{1/2}S | Y)$, where $S$ is a Gaussian random variable in ${\mathbb{R}}^m$ with covariance matrix $I$ independent of $(X,Y)$. Knowing that the density $\rho_{X+t^{1/2}S,Y}$ satisfies the heat equation $$\partial_t \rho_{X+t^{1/2}S,Y} = \frac{1}{2} \Delta_x \rho_{X+t^{1/2}S,Y},$$ one can show that $\mathcal{I}(X + t^{1/2}S | Y)$ is well-defined and finite for $t > 0$ and that $$\label{eq:classicalentropyrateofchange}
\frac{d}{dt} h(X + t^{1/2} S | Y) = \frac{1}{2} \mathcal{I}(X + t^{1/2} S | Y).$$
The Fisher information is the $L^2$ norm of the (${\mathbb{R}}^m$-valued) random variable $\Xi$ given by evaluating the *score function* $-\nabla_x \rho_{X|Y} / \rho_{X|Y}$ on the random variable $(X,Y)$, provided that this random variable is in $L^2$. In this case, the random variable $\Xi$ is known as the *score function* for $X$ given $Y$, and it is the unique element of $L^2$ satisfying the integration-by-parts relation $$\label{eq:integrationbyparts}
E[\Xi f(X,Y)] = E[ \nabla_x f(X,Y) ] \text{ for all } f \in C_c^\infty({\mathbb{R}}^m \times {\mathbb{R}}^n).$$ More generally, if there exists a random variable $\Xi$ in $L^2$ satisfying this integration-by-parts formula, then we define the conditional Fisher information to be $\mathcal{I}(X|Y) = E |\Xi|^2$ (and this extends our previous definition of $\mathcal{I}(X|Y)$). Otherwise, $\mathcal{I}(X|Y)$ is defined to be $\infty$.
In light of the integration-by-parts characterization, score functions behave well under conditionally independent sums. The following lemma is proved in the same way as the non-conditional case (see [@Jekel2018 Lemma 5.6]) and the free case (see [@VoiculescuFE5 Proposition 3.7]).
\[lem:xiconditionalexpectation\] Let $Y$ be a random variable in ${\mathbb{R}}^n$ and let $X_1$ and $X_2$ be random variables in ${\mathbb{R}}^m$ that are conditionally independent given $Y$. Suppose that $\Xi$ is a score function for $X_1$ given $Y$. Then $E[\Xi| X_1 + X_2, Y]$ is a score function for $X_1 + X_2$ given $Y$. Hence, $$\mathcal{I}(X_1+X_2|Y) \leq \mathcal{I}(X_1 | Y).$$ In particular, this holds if $X_2$ is independent from $(X_1,Y)$ or $X_1$ is independent of $(X_2,Y)$.
Score functions also scale in the following way. The proof is straightforward from the integration-by-parts relation.
\[lem:xiscaling\] If $\Xi$ is a score function for $X$ given $Y$ and $t > 0$, then $(1/t) \Xi$ is a score function for $tX$ given $Y$, and hence $\mathcal{I}(tX | Y) = t^{-2} \mathcal{I}(X|Y)$.
Random Matrix Renormalization {#subsec:matrixentropy}
-----------------------------
Suppose that $(X,Y)$ is a random variable in $M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n$ with density $\rho_{X,Y}$. The trace on $M_N({\mathbb{C}})_{sa}$ produces a real inner product. But to study the large $N$ limit, we use the normalized trace $\tau_N = (1/N) \operatorname{Tr}$. The corresponding normalized Gaussian is the GUE ensemble $S = (S_1,\dots,S_m)$ where $S_j$ has variance $1$ with respect to $\tau_N$. We use the following renormalized entropy, which is motivated by computation of the Gaussian case and by below, $$h^{(N)}(X|Y) = \frac{1}{N^2} h(X|Y) + \frac{m}{2} \log N.$$ Due to the normalization of Gaussian, the evolution of the density for $(X + t^{1/2} S, Y)$ is given by the renormalized heat equation $$\partial_t \rho_{X+t^{1/2}S, Y} = \frac{1}{2N} \Delta \rho_{X+t^{1/2}S,Y}.$$ This results in $$\label{eq:normalizedentropyrateofchange}
\partial_t h^{(N)}(X + t^{1/2} S | Y) = \frac{1}{2N^3} \mathcal{I}(X + t^{1/2}S | Y) =: \frac{1}{2} \mathcal{I}^{(N)}(X+t^{1/2}S | Y),$$ where $\mathcal{I}^{(N)}(X|Y) := N^{-3} \mathcal{I}(X|Y)$, assuming that $X$ has finite variance and $t > 0$.
Another heuristic for the normalization $\mathcal{I}^{(N)} = N^{-3} \mathcal{I}$ comes from analyzing the case where $(X,Y)$ have density $(1/Z) e^{-N^2 V(x,y)}\,dx\,dy$ where $V$ is uniformly convex and semi-concave. Indeed, in this case, the classical score function for $X$ given $Y$ is $-N^2 \nabla_x V(X,Y)$. Recall that $D_x V = N \nabla_x V$ is the gradient of $V$ with respect to the normalized inner product $\ip{\cdot,\cdot}_2$. Thus, $$\frac{1}{N^3} \mathcal{I}(X|Y) = \frac{1}{N} E \norm{N \nabla_x V(X,Y)}_{\operatorname{Tr}}^2 = E \norm{D_x V(X,Y)}_2^2$$ is a dimension-independent normalization. Furthermore, the normalized score function $\xi = (1/N) \Xi$ (which would be $D_x V(X,Y)$ in the case where the law is given by a potential $V$) satisfies the integration-by-parts relation $$E \ip{\xi_j, f(X,Y)}_2 = E \frac{1}{N^2} \operatorname{Div}_{x_j} f(X,Y),$$ where $\xi = (\xi_1,\dots,\xi_m)$ and where $\operatorname{Div}$ is the divergence with respect to the classical coordinates (not normalized). But if $f$ is a non-commutative polynomial, then $$\frac{1}{N^2} \operatorname{Div}_{x_j} f(x,y) = \frac{1}{N^2} \operatorname{Tr}\otimes \operatorname{Tr}(\partial_{x_j} f(x,y)) = \tau_N \otimes \tau_N(\partial_{x_j} f(x,y)),$$ where $\partial_{x_j}$ denotes the non-commutative derivative or free difference quotient with respect to $x_j$. Thus, applying the integration-by-parts relation to non-commutative polynomials results in the dimension-independent relation $$E \ip{\xi_j, f(X,Y)}_2 = \sum_{j=1}^m E[\tau_N \otimes \tau_N(\partial_{x_j} f(x,y))]$$ that characterizes the normalized score function.
As a consequence of , $h^{(N)}(X|Y)$ can be recovered by integrating $\mathcal{I}^{(N)}(X+t^{1/2}S | Y)$ and modifying the integral to converge at $\infty$. This results in $$\label{eq:normalizedentropyformula}
h^{(N)}(X|Y) = \frac{1}{2} \int_0^\infty \left( \frac{m}{1 + t} - \mathcal{I}^{(N)}(X + t^{1/2}S | Y) \right)\,dt + \frac{m}{2} \log 2\pi e$$ provided that $(X,Y)$ has a density $\rho_{X,Y}$ and that $X$ has finite variance. The proof is similar to [@Jekel2018 Lemma 5.7]. Convergence of the integral at $\infty$ can be deduced from the following estimate, and it also shows convergence of the integral at $0$ if $\mathcal{I}(X|Y)$ is finite. Compare [@VoiculescuFE5 Corollary 6.14 and Remark 6.15] and [@Jekel2018 Lemma 5.7].
\[lem:Fisherestimates\] Let $(X,Y)$ be a random variable in $M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n$ such that $a = (1/m) \sum_{j=1}^m E[\tau_N(X_j^2)] < \infty$, and let $S$ be an independent GUE $m$-tuple. Then $$\frac{m}{a + t} \leq \mathcal{I}^{(N)}(X + t^{1/2} S | Y) \leq \min \left( \frac{m}{t}, \mathcal{I}^{(N)}(X|Y) \right).$$
We observe that $\xi_t = E[t^{-1/2} S | X + t^{1/2} S, Y]$ is a normalized score function for $X + t^{1/2} S$ given $Y$ by Lemma \[lem:xiconditionalexpectation\]. This yields $\mathcal{I}^{(N)}(X + t^{1/2} S | Y) \leq m / t$. On the other hand, if $\xi$ is a normalized score function for $X$ given $Y$, we also have $\xi_t = E[\xi | X + t^{1/2} S, Y]$, which yields the upper bound $\mathcal{I}^{(N)}(X|Y)$. The lower bound follows from observing $(E \norm{\xi_t}_2^2)^{1/2} (E \norm{X + t^{1/2}S}_2^2)^{1/2} \geq E \ip{\xi_t, X + t^{1/2} S}_2$ and evaluating the right hand side using integration by parts.
Convergence to Conditional Free Entropy
---------------------------------------
Motivated by the normalized entropy and Fisher’s information in the previous section, Voiculescu defined the free versions as follows. Let $(X,Y)$ be an $(m+n)$-tuple of self-adjoint non-commutative random variables in a tracial $\mathrm{W}^*$-algebra $(\mathcal{M},\tau)$. We say that $\xi = (\xi_1,\dots,\xi_m) \in L^2(\mathcal{M},\tau)_{sa}^m$ is a *free score function for $X$ given $Y$* (also known as a *conjugate variable*) if for every non-commutative polynomial $f(X,Y)$, we have $$\tau(\xi_j f(X,Y)) = \tau \otimes \tau(\partial_{x_j} f(X,Y)).$$ The *free Fisher information* $\Phi^*(X|Y)$ is defined to be $\norm{\xi}_2^2$ if such a $\xi$ exists, and $\infty$ otherwise. The *non-microstates free entropy* $\chi^*(X|Y)$ is defined to be $$\label{eq:definechi*}
\chi^*(X|Y) = \frac{1}{2} \int_0^\infty \left( \frac{m}{1 + t} - \Phi^*(X + t^{1/2}S | Y) \right)\,dt + \frac{m}{2} \log 2\pi e.$$ Convergence of the integral at $\infty$ follows from the free analogue of Lemma \[lem:Fisherestimates\], so that $\chi^*(X | Y)$ is well-defined in $[-\infty,\infty)$ whenever $X$ has finite variance.
Voiculescu’s original notation in [@VoiculescuFE5 §7] was $\chi^*(X: \mathrm{W}^*(Y))$ rather than $\chi^*(X | Y)$, since the definition of the free score function can be rephrased so as to depend only on $\mathrm{W}^*(Y)$ rather than $Y$. However, we prefer to write $\chi^*(X | Y)$ instead by analogy with the classical case, using the vertical bar to denote “conditioning.” This avoids potential confusion with the notation $\chi(X:Y)$ for microstates entropy of $X$ in the presence of $Y$ used in [@VoiculescuFE3 §1].
The following lemma gives sufficient conditions for classical Fisher information for random matrix models to converge to free Fisher information. The main hypotheses are that the non-commutative laws converge, the score functions $D_x V^{(N)}$ for the $N \times N$ matrix models are asymptotically approximable by trace polynomials, and some mild growth conditions on score functions and probability measures as $\norm{(x,y)}_\infty \to \infty$. We omit the proof since it is a direct adaptation of the proof of [@Jekel2018 Proposition 5.10].
\[lem:convergenceofFisherinfo\] Let $V^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to {\mathbb{R}}$ be a potential with $\int_{M_N({\mathbb{C}})^{m+n}} e^{-N^2 V^{(N)}(x,y)}\,dx\,dy < +\infty$, let $\mu^{(N)}$ be the associated probability density, and let $(X^{(N)}, Y^{(N)})$ be a random variable distributed according to $\mu^{(N)}$. Let $(X,Y)$ be an $(m+n)$-tuple of self-adjoint non-commutative random variables in the tracial $\mathrm{W}^*$-algebra $(\mathcal{M},\tau)$. Assume that:
(A) The non-commutative law of $(X^{(N)}, Y^{(N)})$ with respect to $\tau_N$ converges in probability to the non-commutative law of $(X,Y)$.
(B) $D_x V^{(N)}$ is defined and continuous, and the sequence $\{D_x V^{(N)}\}$ is asymptotically approximable by trace polynomials, and hence $D_x V^{(N)} \rightsquigarrow g \in (\overline{\operatorname{TrP}}_{m+n}^1)_{sa}^m$.
(C) For some $k \geq 0$ and $a, b > 0$, we have $$\norm{D_x V^{(N)}(x,y)}_2^2 \leq a + b \norm{(x,y)}_\infty^k$$
(D) There exists $R_0 > 0$ such that $$\lim_{N \to \infty} E \left[ \mathbf{1}_{\norm{(X^{(N)},Y^{(N)})}_\infty \geq R_0} \left(1 + \norm{(X^{(N)},Y^{(N)})}_\infty^k \right) \right] = 0.$$
Then $\mathcal{I}^{(N)}(X|Y)$ is finite. Moreover, $g(X,Y)$ is in $L^2(\mathcal{M},\tau)$ and it is the free score function for $X$ given $Y$, and we have $$\mathcal{I}^{(N)}(X|Y) = E[\norm{D_x V^{(N)}(X^{(N)},Y^{(N)})}_2^2] \to \norm{g(X,Y)}_2^2 = \Phi^*(X|Y).$$
\[thm:convergenceofentropy\] Let $V^{(N)}: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \to {\mathbb{R}}$ satisfy Assumption \[ass:convexRMM\] for some $0 < c \leq C$. Let $\mu^{(N)}$ be the corresponding measure, let $X^{(N)}, Y^{(N)}$ be random variables chosen according to $\mu^{(N)}$, and let $S^{(N)}$ be an independent $m$-tuple of GUE matrices.
Let $X = (X_1,\dots,X_m)$ and $Y = (Y_1,\dots,Y_n)$ be non-commutative random variables with non-commutative law $\mu = \mu_V$, and let $S$ be a freely independent free semicircular $m$-tuple. Then for every $t \geq 0$, we have $$\label{eq:convergenceofFisherinfo}
\Phi^*(X + t^{1/2} S | Y) = \lim_{N \to \infty} \mathcal{I}^{(N)}(X^{(N)} + t^{1/2} S^{(N)} | Y^{(N)})$$ and $$\label{eq:convergenceofentropy}
\chi^*(X + t^{1/2} S | Y) = \lim_{N \to \infty} h^{(N)}(X^{(N)} + t^{1/2} S^{(N)} | Y^{(N)}).$$
We want to show that the law of $(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)})$ satisfies the assumptions of Lemma \[lem:convergenceofFisherinfo\] for each $t \geq 0$. The joint law of $(X^{(N)},Y^{(N)},S^{(N)})$ is given by the convex potential $U^{(N)}(x,y,s) = V(x,y) + (1/2) \norm{s}_2^2$. Now $U^{(N)}$ satisfies $\min(c,1) \leq HU^{(N)} \leq \max(C,1)$ and $DU^{(N)}$ is asymptotically approximable by trace polynomials. Thus, the law of $(X^{(N)},Y^{(N)},S^{(N)})$ has a large $N$ limit given by Theorem \[thm:freeGibbslaw\]. In fact, the large $N$ limit must be non-commutative law of $(X,Y,S)$ because of Voiculescu’s asymptotic freeness theorem [@Voiculescu1998] and because the non-commutative law of $S^{(N)}$ converges to the non-commutative law of $S$. (Alternatively, this could be proved the same way as [@Jekel2018 Lemma 7.4].)
Since the non-commutative law of $(X^{(N)},Y^{(N)},S^{(N)})$ converges in probability to that of $(X,Y,S)$, the non-commutative law of $(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)})$ converges in probability to that of $(X + t^{1/2}S,Y)$, and thus (A) of Lemma \[lem:convergenceofFisherinfo\] holds. Moreover, Lemma \[lem:epsilonnet\] shows that $$P(\norm{(X^{(N)} - E(X^{(N)}),Y^{(N)} - E(Y^{(N)}),S^{(N)})}_\infty \geq \min(1,c)^{-1/2} (\Theta + \delta)) \leq e^{-N \delta^2/2}.$$ From this it is not hard to show that $(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)})$ satisfies (D).
It remains to check (B) and (C). The potential for $(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)})$ is given by $$W_t^{(N)}(\tilde{x},y,s) = U^{(N)}(\tilde{x} - t^{1/2} s, y, s),$$ which follows by applying the change of variables formula for the density. Here we write $\tilde{x}$ to emphasize that this variable corresponds to $X^{(N)} + t^{1/2} S^{(N)}$ rather than $X^{(N)}$. Note that $W_t^{(N)}$ is uniformly convex and semi-concave since it is the composition of $U^{(N)}$ with an invertible linear transformation. Also, $$DW_t^{(N)}(\tilde{x},y,s) = (DV^{(N)}(\tilde{x}-t^{1/2}s,y), s - t^{1/2} D_xV^{(N)}(\tilde{x}-t^{1/2}s,y))$$ is asymptotically approximable by trace polynomials. The potential corresponding to $(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)})$ is $$V_t^{(N)}(\tilde{x},y) = -\frac{1}{N^2} \log \int e^{-N^2 W_t^{(N)}(\tilde{x},y,s)}\,ds.$$ Since $W_t^{(N)}$ is uniformly convex, the integrand vanishes rapidly at $\infty$, and thus it is straightforward to differentiate under the integral by dominated convergence, and deduce that $V_t^{(N)}$ is continuously differentiable. Furthermore, $$DV_t^{(N)}(\tilde{x},y) e^{-N^2 V_t^{(N)}(\tilde{x},y)} = \int DW_t^{(N)}(\tilde{x},y,s) e^{-N^2W_t^{(N)}(\tilde{x},y,s)} \,ds,$$ so that $$DV_t^{(N)}(\tilde{x},y) = \frac{\int DW_t^{(N)}(\tilde{x},y,s) e^{-N^2W_t^{(N)}(\tilde{x},y,s)} \,ds}{\int e^{-N^2W_t^{(N)}(\tilde{x},y,s)} \,ds},$$ or in other words $DV_t^{(N)}$ is given by the conditional expectation $$\begin{aligned}
\label{eq:xiconditionalexpectation2}
DV_t^{(N)}(X^{(N)} + t^{1/2} S^{(N)},Y^{(N)}) &= E\left[D_{(x,y)} W_t^{(N)}(X^{(N)} + t^{1/2} S^{(N)},Y^{(N)},S^{(N)}) \bigl| X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)} \right] \\
&= E\left[DV^{(N)}(X^{(N)}, Y^{(N)}) \bigl| X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)} \right]. \nonumber\end{aligned}$$ Now we apply Theorem \[thm:conditionalexpectation\] using the potential $W_t^{(N)}$ and conditioning on $(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)})$ to conclude conclude that $DV_t^{(N)}(\tilde{x},y)$ is asymptotically approximable by trace polynomials, which establishes (B).
Furthermore, Theorem \[thm:conditionalexpectation\] implies that $$\norm{DV_t^{(N)}}_{\operatorname{Lip}} \leq (1 + C/c) \norm{D_{(x,y)}W_t^{(N)}}_{\operatorname{Lip}} \leq (1 + C/c)(1 + t^{1/2}) \norm{DV^{(N)}}_{\operatorname{Lip}} \leq (1 + C/c)(1 + t^{1/2}) C.$$ This implies that (C) of Lemma \[lem:convergenceofFisherinfo\] holds with $k = 1$, using Remark \[rem:UCgrowthbound\].
Therefore, we may apply Lemma \[lem:convergenceofFisherinfo\] to $(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)})$ to obtain that holds for every $t \geq 0$, that is, $$\lim_{N \to \infty} \mathcal{I}^{(N)}(X^{(N)} + t^{1/2} S^{(N)} | Y^{(N)}) \to \Phi^*(X+t^{1/2}S | Y).$$ For the second claim regarding $h^{(N)}$ and $\chi^*$, it remains to show that $$\lim_{N \to \infty} \frac{1}{2} \int_0^\infty \left( \frac{m}{1 + t} - \mathcal{I}^{(N)}(X + t^{1/2}S | Y) \right)\,dt = \frac{1}{2} \int_0^\infty \left( \frac{m}{1 + t} - \Phi^*(X + t^{1/2}S | Y) \right)\,dt.$$ We just showed the integrand converges pointwise. But we can take the limit inside the integral by the dominated convergence theorem, because by Lemma \[lem:Fisherestimates\], we have $$\frac{m}{a + t} \leq \mathcal{I}^{(N)}(X^{(N)} + t^{1/2} S^{(N)} | Y^{(N)}) \leq \min \left( \frac{m}{t}, \mathcal{I}^{(N)}(X^{(N)}|Y^{(N)}) \right),$$ and we also know that $\mathcal{I}^{(N)}(X^{(N)} | Y^{(N)})$ is bounded as $N \to \infty$ because it converges to $\Phi^*(X|Y)$.
Of course, leads to the same conclusion as Lemma \[lem:xiconditionalexpectation\]. Indeed, $\xi_t = D_x V_t^{(N)}(X^{(N)} + t^{1/2}S^{(N)},Y^{(N)})$ is the score function for $X^{(N)} + t^{1/2} S^{(N)}$, and Lemma \[lem:xiconditionalexpectation\] says that $\xi_t$ is the conditional expectation of $\xi_0 = D_x V^{(N)}(X^{(N)}, Y^{(N)})$ given $X^{(N)} + t^{1/2} S^{(N)}$ and $Y^{(N)}$.
\[rem:simplifiedentropyproof\] In [@Jekel2018 §7], we did not use the conditional expectation method to prove $DV_t^{(N)}$ is asymptotically approximable by trace polynomials, but rather we analyzed the evolution of $DV_t^{(N)}$ directly using PDE semigroups. The proof given here for convergence of entropy is thus considerably shorter. However, our results on the evolution of $DV_t^{(N)}$ will come in handy for our construction of transport in the next section.
Conditional Transport to Gaussian {#sec:transport}
=================================
In this section, we prove our main results about transport (Theorems \[thm:transport1\] and \[thm:transport2\]). Suppose that $V^{(N)}(x,y)$ is a potential as in Assumption \[ass:convexRMM\], $\mu^{(N)}$ is the corresponding probability distribution and that $(X^{(N)},Y^{(N)})$ is a random variable with this law. Let $S^{(N)}$ be an independent $m$-tuple of GUE matrices. Let $\mu_t^{(N)}$ be the law of $(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)})$.
The evolution of the potential $V_t^{(N)}$ corresponding to $\mu_t^{(N)}$ was studied in [@Jekel2018], and in particular, we established a dimension-independent way to obtain $DV_t^{(N)}$ from $DV^{(N)}$ using operations that preserve asymptotic approximability by trace polynomials. By solving an ODE in terms of $DV_t^{(N)}$, we will obtain transport maps $F_{s,t}^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})_{sa}^m$ such that $$(F_{s,t}^{(N)}(X^{(N)} + t^{1/2} S^{(N)}, Y^{(N)}), Y^{(N)}) \sim (X^{(N)} + s^{1/2} S^{(N)}, Y^{(N)}).$$ Upon renormalizing and taking the limit as $s$ or $t$ goes to $\infty$, we obtain transport to the law of $(S^{(N)}, Y^{(N)})$.
To make each part of the proof more computationally tractable, we proceed in stages. Up until §\[subsec:largeNtransport\], we fix $N$ (and thus suppress it in the notation). First, in §\[subsec:basictransport\], we describe the basic construction of transport for functions of $x$ alone (imagining that we have frozen the variable $y$). In §\[subsec:conditionalHJB\], we describe the properties of $V_t^{(N)}(x,y)$. Next, §\[subsec:Lipschitztransport\] proves Lipschitz estimates for the transport maps $F_{s,t}^{(N)}(x,y)$.
In §\[subsec:largeTtransport\], we introduce renormalized transport maps $\tilde{F}_{s,t}^{(N)}$ that transport $\tilde{\mu}_t$ to $\tilde{\mu}_s$, where $\tilde{\mu}_t$ is the law of $(e^{-t/2} X^{(N)} + e^{-t/2}(e^t - 1)^{1/2} S^{(N)}, Y^{(N)})$. The renormalized transport map $\tilde{F}_{s,t}$ is the same one used by Otto and Villani in their proof of the Talagrand transportation-entropy inequality [@OV2000 §4, proof of Lemma 2], in the special case where the target measure is Gaussian (and generalized to the conditional setting). We will explain this inequality further in §\[subsec:entropycost\].
The new element in our paper is the analysis of the large $t$ and large $N$ limits of the transport maps. In §\[subsec:largeTtransport\], we show that the limit as $s$ or $t$ tends to $\infty$ exists. Then in §\[subsec:largeNtransport\], we use the machinery of asymptotic approximation by trace polynomials to study the large $N$ limit of $\tilde{F}_{s,t}^{(N)}$. In order to get dimension-independent estimates for convergence as $s$ or $t$ tends to $\infty$, we conduct a finer analysis of convexity properties of $V_t$ and Lipschitz properties of $\tilde{F}_{s,t}$. It is convenient to carry out the earlier stages of this analysis in §\[subsec:conditionalHJB\] and §\[subsec:Lipschitztransport\] for $F_{s,t}$ rather than $\tilde{F}_{s,t}$.
Basic Construction of Transport {#subsec:basictransport}
-------------------------------
In this section, we will fix $N$ and fix a function $V: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ in $\mathcal{E}_{c,C}$ for some $0 < c < C$. Later, we will allow $V$ to depend on $N$ and to depend on another self-adjoint tuple $y$, but we prefer to simplify notation for the sake of carrying out the basic computation.
Let $\mu$ be the probability measure with density $(1/Z) e^{-N^2 V}$ where $Z = \int_{M_N({\mathbb{C}})_{sa}^m} e^{-N^2 V}$. We showed in [@Jekel2018] that the density of $\mu_t := \mu * \sigma_{t,N}$ is $(1/Z) e^{-N^2 V_t}$, where $V_t$ solves the equation $$\label{eq:simpleHJB}
\partial_t V_t = \frac{1}{2N} \Delta V_t - \norm*{DV_t}_2^2.$$ Because $(1/Z) e^{-N^2 V_t}$ solves the heat equation, we know that $V_t$ is a smooth function of $(x,t)$ for $t > 0$ and a continuous function of $(x,t)$ for $t \geq 0$. Moreover, $V_t \in \mathcal{E}(c(1+ct)^{-1},C(1+Ct)^{-1})$ for each $t$ as proved in Theorem 6.1 (1) of [@Jekel2018].
Now we can describe explicit transport functions $F_{s,t}$ such that $(F_{s,t})_* \mu_s = \mu_t$ for all $s, t \in [0,+\infty)$.
\[prop:basictransport1\] Let $V$, $\mu$, $V_t$, and $\mu_t$ be as above.
1. There exists a unique family of functions $F_{s,t}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})_{sa}^m$ for $0 \leq s \leq t < +\infty$ such that $$\begin{aligned}
F_{t,t}(x) &= x \text{ for } t \in [0,+\infty) \label{eq:transportODE} \\
\partial_s F_{s,t}(x) &= \frac{1}{2} DV_s(F_{s,t}(x)) \text{ for } s, t \in [0,+\infty). \nonumber
\end{aligned}$$
2. $F_{t_1,t_2} \circ F_{t_2,t_3} = F_{t_1,t_3}$ and in particular $F_{t,s} = F_{s,t}^{-1}$.
3. $(F_{s,t})_* \mu_t = \mu_s$.
\(1) Because $V_s \in \mathcal{E}(c(1+cs)^{-1},C(t+Cs)^{-1})$, we know that $DV_s(x)$ is $C$-Lipschitz with respect to $\norm*{\cdot}_2$. Hence, given $t \in [0,+\infty)$, by the Picard-Lindelöf theorem, the initial value problem has a solution defined for all $s \in [0,+\infty)$.
\(2) Fix $t_1$, $t_2$, and $t_3$ and fix $x \in M_N({\mathbb{C}})_{sa}^m$. Let $G(t)$ be the function defined by $G(t_3) = x$ and $\partial_t G(t) = DV_t(G(t))$. By definition of the functions $F_{s,t}$, we have $G(t_1) = F_{t_1,t_3}(x)$ and $G(t_2) = F_{t_2,t_3}(x)$. So $G$ also satisfies the initial value problem $\partial_t G(t) = DV_t(G(t))$ and $G(t_2) = F_{t_2,t_3}(x)$. Therefore, $G(t_1) = F_{t_1,t_2}(F_{t_2,t_3}(x))$, so that $F_{t_1,t_2}(F_{t_2,t_3}(x)) = F_{t_1,t_3}(x)$.
\(3) We first prove the claim for $s, t > 0$. Because $V_s$ is smooth, it follows that $F_{s,t}$ is smooth for $s, t > 0$ by standard theory of smooth dependence for ODE. Let $JF_{s,t}$ denote the Jacobian linear transformation (differential) of $F_{s,t}$. Let $\rho_t = (1/Z)e^{-V_t}$ is the density of $\mu_t$. As a consequence of the change-of-variables formula for multivariable integrals, we see that the density of $(F_{s,t})_* \mu_t = (F_{t,s}^{-1})_* \mu_t$ is $$(\rho_t \circ F_{t,s}) |\det JF_{t,s}| = \frac{1}{Z} \exp\left[ -N^2\left(V_t \circ F_{t,s} - \frac{1}{N^2} \log |\det JF_{t,s}| \right) \right].$$ Fix $s$. If $t = s$, then clearly this reduces to $\rho_s$. Therefore, it suffices to show that $(\rho_t \circ F_{t,s}) |\det JF_{t,s}|$ is a constant function of $t$, or equivalently that $$\partial_t \left[ V_t \circ F_{t,s} - \frac{1}{N^2} \log |\det JF_{t,s}| \right] = 0.$$ Recalling smoothness $V_t$ and $F_{t,s}$ for $s, t > 0$ and using the differential equations for $V_t$ and for $F_{t,s}$, we obtain $$\begin{aligned}
\partial_t[V_t \circ F_{t,s}] &= (\partial_t V_t) \circ F_{t,s} + \ip{DV_t(F_{t,s}), \partial_t F_{t,s}}_2 \\
&= \frac{1}{2N} \Delta V_t \circ F_{t,s} - \frac{1}{2} \norm*{DV_t(F_{t,s})}_2^2 + \frac{1}{2} \ip{DV_t(F_{t,s}), DV_t(F_{t,s})}_2 \\
&= \frac{1}{2N} \Delta V_t.\end{aligned}$$ Meanwhile, to compute $\partial_t \log |\det JF_{t,s}|$, note that for small $\epsilon \in {\mathbb{R}}$, $$JF_{t+\epsilon,s} = J(F_{t+\epsilon,t} \circ F_{t,s}) = JF_{t+\epsilon,t}(F_{t,s}) JF_{t,s},$$ so that $$\log |\det JF_{t+\epsilon,s}| = \log |\det JF_{t,s}| + \log |\det JF_{t+\epsilon,t}(F_{t,s})|.$$ Using smoothness, $$\begin{aligned}
\frac{d}{d\epsilon} \bigr|_{\epsilon = 0} JF_{t+\epsilon,t} &= J \left( \frac{d}{d\epsilon} \bigr|_{\epsilon = 0} F_{t+\epsilon,t} \right) \\
&= \frac{1}{2} J(DV_t) = \frac{1}{2} N J(\nabla V_t),\end{aligned}$$ Since $JF_{t+\epsilon,t}$ becomes the identity when $\epsilon = 0$, we know that for small enough $\epsilon$, the linear transformation $JF_{t+\epsilon,t}$ has positive determinant and $\log JF_{t+\epsilon,t}$ is well-defined by power series, so that $$\begin{aligned}
\frac{d}{d\epsilon} \Bigr|_{\epsilon = 0} \log |\det JF_{t+\epsilon,t} &= \frac{d}{d\epsilon} \bigr|_{\epsilon = 0} \operatorname{Tr}\log JF_{t+\epsilon,t} \\
&= \operatorname{Tr}\left( \frac{d}{d\epsilon} \Bigr|_{\epsilon = 0} JF_{t+\epsilon,t} \right) \\
&= \operatorname{Tr}( N(J \nabla V_t)) = \frac{1}{2} N \Delta V_t.\end{aligned}$$ Hence, $\partial_t \log |\det JF_{t,s}| = \frac{N}{2} \Delta V_t \circ F_{t,s}$. This implies that $$\partial_t \left[ V_t \circ F_{t,s} - \frac{1}{N^2} \log |\det JF_{t,s}| \right] = \frac{1}{2N} \Delta V_t - \frac{1}{N^2} \cdot \frac{N}{2} \Delta V_t \circ F_{t,s} = 0,$$ completing the proof of the claim for $s, t > 0$. The equality $(F_{s,t})_* \mu_t = \mu_s$ extends to the case where $s$ or $t$ is zero because both sides depend continuously on $s$ and $t$ with respect to the weak topology on measures.
In particular, the map $F_{0,t}$ transports $\mu_t = \mu * \sigma_{t,N}$ to the original law $\mu$. In other words, if $X \sim \mu$ and $S \sim \sigma_{1,N}$, then $F_{0,t}(X + t^{1/2}S) \sim X$ and $F_{t,0}(X) \sim X + t^{1/2} S$. This implies that $(1 + t)^{-1/2} F_{t,0}(X) \sim (1 + t)^{-1/2}(X + t^{1/2} S)$. This suggests that we can find a transport map from the law of $X$ to the law of $S$ as the large $t$ limit of $(1 + t)^{-1/2} F_{t,0}$. In the interest of efficiency, we postpone the details of this argument until after we introduce the dependence on the other set of parameters $y$.
Conditional Hamilton-Jacobi-Bellman Equation and Semigroups {#subsec:conditionalHJB}
-----------------------------------------------------------
Let us now fix $N$ and fix a potential $V: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \to {\mathbb{R}}$ in $\mathcal{E}_{m+n}^{(N)}(c,C)$ for some $0 \leq c \leq C$. Let $\mu$ be the corresponding law and let $(X,Y)$ be a random variable in $M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n$ distributed according to $\mu$. Let $\mu_t$ be the law of $(X + t^{1/2}S,Y)$, where $S$ is an independent tuple of independent GUE.
Our goal is to transport the law $\mu_s$ to the law $\mu_t$. Upon freezing the variable $y$, the methods of the previous section will produce a transport map $F_{s,t}(x,y)$ such that $F_{s,t}(\cdot,y)$ pushes forward the conditional distribution of $X + s^{1/2} S$ given $Y$ to the conditional distribution of $X + t^{1/2} S$ given $Y$. Specifically, $F_{s,t}: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \to M_N({\mathbb{C}})_{sa}^m$ is the solution to the initial value problem $$\begin{aligned}
F_{t,t}(x,y) &= x, \\
\partial_s F_{s,t}(x,y) &= D_x V_s(F_{s,t}(x,y),y).\end{aligned}$$ Then $(F_{s,t}(X + t^{1/2} S, Y),Y) \sim (X+s^{1/2} S, Y)$.
We seek to understand the large $t$ and large $N$ behavior of $F_{s,t}(x,y)$ as a function of $(x,y)$ rather than simply as a function of $x$ for a fixed $y$. To achieve this, we must understand the behavior of $V(x,y)$ and $D_x V(x,y)$ as a functions of $(x,y)$. We will first import the results of [@Jekel2018 §6] regarding $D_x V(x,y)$ as a function of $x$, then we will extend them to handle the dependence on $y$.
The potential $V_t$ satisfies $$\label{eq:conditionalHJB}
\partial_t V_t = \frac{1}{2N} \Delta_x V_t - \frac{1}{2} \norm*{D_x V_t}_2^2.$$ We express $V_t = R_t V$, where $R_t$ is a semigroup acting on convex and semi-concave functions defined as follows. Let $$\begin{aligned}
P_t u(x,y) &= \int_{M_N({\mathbb{C}})_{sa}^m} u(x+z,y)\,d\sigma_{m,t}^{(N)}(z) \\
Q_t u(x,y) &= \inf_{z} \left[u(z,y) + \frac{1}{2t} \norm*{z - x}_2^2 \right].\end{aligned}$$ Then as suggested by Trotter’s formula, we want to express $R_t u = \lim_{n \to \infty} (P_{t/n} Q_{t/n})^n u$, but for technical convenience we only apply this to dyadic rationals $t$ and values of $n$ that are powers of $2$. The following is a direct application of [@Jekel2018 Theorems 6.1 and 6.17] to $V(\cdot,y)$.
\[thm:Rtsemigroup\] There exists a semigroup of nonlinear operators $R_t: \bigcup_{C > 0} \mathcal{E}_{m+n}^{(N)}(0,C) \to \bigcup_{C > 0} \mathcal{E}_{m+n}^{(N)}(0,C)$ with the following properties:
(1) [**Change in Convexity:**]{} If $u(\cdot,y) \in \mathcal{E}_m^{(N)}(c,C)$, then $R_t u(\cdot,y) \in \mathcal{E}_m^{(N)}(c(1+ct)^{-1}, C(1+Ct)^{-1})$.
(2) [**Approximation by Iteration:**]{} For $\ell \in {\mathbb{Z}}$ and $t \in 2^{-\ell} {\mathbb{N}}_0$, denote $R_{t,\ell}u = (P_{2^{-\ell}} Q_{2^{-\ell}})^{2^\ell t} u$. Suppose $t \in {\mathbb{Q}}_2^+$ and $u \in \mathcal{E}_{m+n}^{(N)}(0,C)$.
(a) If $2^{-\ell-1} C \leq 1$, then $$|R_t u - R_{t,\ell} u| \leq \left( \frac{3}{2} \frac{C^2mt}{1+Ct} + \log(1 + Ct) (m + Cm + \norm*{D_x u}_2^2) \right) 2^{-\ell}.$$
(b) $\displaystyle \norm*{D_x(R_{t,\ell} u) - D_x(R_t u)}_{L^\infty} \leq [t/2 + C(t/2)^2] C^2 m^{1/2}(2 \cdot 2^{-\ell/2} + 2^{-3\ell/2}C)$.
(3) [**Continuity in Time:**]{} Suppose $s \leq t \in {\mathbb{R}}_+$ and $u \in \mathcal{E}_{m+n}^{(N)}(0,C)$. Then
(a) $R_t u \leq R_s u + \frac{m}{2} [\log(1 + Ct) - \log(1 + Cs)]$.
(b) $R_t u \geq R_s u - \frac{1}{2} (t - s)(Cm + \norm*{D_x u}_2^2)$.
(c) If $C(t - s) \leq 1$, then $\norm*{D_x(R_t u) - D_x(R_s u)}_2 \leq 5Cm^{1/2} 2^{1/2}(t - s)^{1/2} + C(t - s) \norm*{D_x u}_2$.
(4) [**Differential Equation:**]{} $R_t u(x)$ is continuous as a function of $(x,t)$ on $M_N({\mathbb{C}})_{sa}^m \times [0,+\infty)$ and smooth on $M_N(|C)_{sa}^m \times (0,+\infty)$, and it satisfies , and we have $P_t[\exp(-N^2u)] = \exp(-N^2 R_t u)$.
Result (1) regarding convexity and semi-concavity only applies to $R_t u$ as a function of $x$ for a fixed $y$. We now extend this result to control the dependence on $y$, using the same techniques as in [@Jekel2018 Lemma 6.6]. As remarked in that paper, this type of analysis of $Q_t$ is standard in the PDE literature on viscosity solutions.
We use the following notation, as in Definition \[def:convexityHnotation\]: Consider a function $u(x,y)$ on $M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n$. Let us write $Hu \geq c I_m \oplus c' I_n$ to mean that $$u(x,y) - \frac{c}{2} \norm*{x}_2^2 - \frac{c'}{2} \norm*{y}_2^2 \text{ is convex}$$ and similarly let us write $Hu \leq C I_m \oplus C' I_n$ to mean that $$u(x,y) - \frac{C}{2} \norm*{x}_2^2 - \frac{C'}{2} \norm*{y}_2^2 \text{ is concave.}$$
\[lem:Qt\] Suppose that $u: M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n \to {\mathbb{R}}$ and that $$c I_m \oplus c' I_n \leq Hu \leq C I_m \oplus C' I_n.$$ Then
1. $c I_m \oplus c' I_n \leq H(P_t u) \leq C I_m \oplus C' I_n$.
2. $D(Q_t u)(x,y) = Du(x - t D_x(Q_t u)(x,y), y)$.
3. $c(1 + ct)^{-1} I_m \oplus c' I_n \leq H(Q_t u) \leq C(1 + Ct)^{-1} I_m \oplus C' I_n$.
4. $c(1 + ct)^{-1} I_m \oplus c' I_n \leq H(R_t u) \leq C(1 + Ct)^{-1} I_m \oplus C' I_n$.
\(1) This is left as an exercise.
\(2) The proof is a modification of that of [@Jekel2018 Lemma 6.6], which proves an analogous result in the simpler case of functions of $x$ without the extra variable $y$. Fix $x_0$ and $y_0$. Because the function $u(z,y_0) + \frac{1}{2t} \norm*{z - x_0}_2^2$ is uniformly convex with respect to $z$, it has a unique minimizer $z_0$. This minimizer must be a critical point with respect to the first variable, and hence $$D_x u(z_0,y_0) + \frac{1}{t}(z_0 - x_0),$$ that is, $$z_0 = x_0 - t D_x u(z_0,y_0).$$ Let $p = D_x u(z_0,y_0)$ and $q = D_y u(z_0,y_0)$, so that $Du(z_0,y_0) = (p,q)$. Our assumption $c I_m \oplus c' I_n\leq Hu \leq C I_m \oplus C' I_n$ implies that $$\underline{v}(x,y) \leq u(x,y) \leq \overline{v}(x,y),$$ where $$\begin{aligned}
\underline{v}(x,y) &= u(z_0,y_0) + \ip{p, x - z_0} + \ip{q, y - y_0} + \frac{c}{2} \norm*{x - z_0}_2^2 + \frac{c'}{2} \norm*{y - y_0}_2^2 \\
\overline{v}(x,y) &= u(z_0,y_0) + \ip{p, x - z_0} + \ip{q, y - y_0} + \frac{C}{2} \norm*{x - z_0}_2^2 + \frac{C'}{2} \norm*{y - y_0}_2^2.\end{aligned}$$ Note that $\underline{v} \leq u \leq v$ implies $Q_t \underline{v} \leq Q_t u \leq Q_t \overline{v}$ since monotonicity of $Q_t$ is immediate from the definition. One can compute $Q_t \underline{v}$ and $Q_t \overline{v}$ directly as in Lemma 6.4 (2) and the proof of Lemma 6.6 in [@Jekel2018] and obtain $$\begin{aligned}
Q_t \underline{v}(x,y) &= u(z_0,y_0) - \frac{t}{2} \norm*{p}_2^2 + \ip{p, x - z_0} + \ip{q, y - y_0} + \frac{c}{2(1 + ct)} \norm*{x - tp - z_0}_2^2 + \frac{c'}{2} \norm*{y - y_0}_2^2 \\
&= u(z_0,y_0) + \frac{1}{2t} \norm*{z_0 - x_0}_2^2 + \ip{p, x - x_0} + \ip{q, y - y_0} + \frac{c}{2(1 + ct)} \norm*{x - tp - z_0}_2^2 + \frac{c'}{2} \norm*{y - y_0}_2^2 \\
&= Q_t u(x_0,y_0) + \ip{p, x - x_0} + \ip{q, y - y_0} + \frac{c}{2(1 + ct)} \norm*{x - x_0}_2^2 + \frac{c'}{2} \norm*{y - y_0}_2^2,\end{aligned}$$ where the last two lines following from substituting $z_0 = x_0 - tp$ and that the infimum defining $Q_t u$ is achieved at $z_0$. The analogous formula for $Q_t \overline{v}(x,t)$ holds as well. The functions $Q_t \underline{v}$ and $Q_t \overline{v}$ thus provide second-order Taylor expansions from above and below for the function $Q_t u$ with respect to $(x,y)$ at the point $(x_0,y_0)$. Looking at the first-order terms in the expansions shows that $Q_t u$ is differentiable at $(x_0,y_0)$ with $$D(Q_t u)(x_0,y_0) = (p,q) = Du(z_0,x_0) = Du(x_0 - tp, y_0) = Du(x_0 - t D_x(Q_t u)(x_0,y_0),y_0),$$ which proves (2).
\(3) We examine the second-order terms of upper and lower Taylor expansions $Q_t \underline{v}$ and $Q_t \overline{v}$ and apply the claim (2) $\implies$ (1) from Lemma \[lem:convexgradient\]. This is the same argument as in the proof of [@Jekel2018 Proposition 2.13 (2)].
\(4) Recall that if $A \leq Hu \leq B$, then $A \leq H(P_tu) \leq B$. Using this fact together with (3) iteratively, we see that if $t$ is a dyadic rational and $t = 2^{-\ell} k$, then $$c(1 + ct)^{-1} I_m \oplus c' I_n \leq H(R_{t,\ell} u) \leq C(1 + Ct)^{-1} I_m \oplus C' I_n.$$ In light of Theorem \[thm:Rtsemigroup\] (2), this will also hold in the limit as $\ell \to \infty$, since for any two self-adjoint matrices $A$ and $B$, the family of functions with $A \leq Hu \leq B$ is closed under pointwise limits. Similarly, using Theorem \[thm:Rtsemigroup\] (3), we extend this to all real $t \geq 0$.
The convexity conditions of Lemma \[lem:Qt\] (4) can alternatively be deduced from [@BL1976 Theorem 4.3]. However, it is convenient for us to use Theorem \[thm:Rtsemigroup\] here because we want the dimension-independent time-continuity estimates Theorem \[thm:Rtsemigroup\] (3) in the proof of Theorem \[thm:transport1\] below.
Lipschitz Estimates for Conditional Transport {#subsec:Lipschitztransport}
---------------------------------------------
This subsection proves the technical estimate Lemma \[lem:transportLipschitzestimate\] on the Lipschitz seminorm of $F_{s,t}$. This depends crucially on the convexity properties of $V_t(x,y)$.
\[lem:twotermgradientestimate\] $$\begin{aligned}
\ip{D_x V_t(x,y) - D_x V_t(x',y'), x - x'}_2 &\leq \frac{C}{1+Ct} \norm*{x - x'}_2^2 + \frac{C-c}{(1+Ct)^{1/2}(1+ct)^{1/2}} \norm*{x - x'}_2 \norm*{y - y'}_2 \\
\ip{D_x V_t(x,y) - D_x V_t(x',y'), x - x'}_2 &\geq \frac{c}{1+ct} \norm*{x - x'}_2^2 - \frac{C-c}{(1+Ct)^{1/2}(1+ct)^{1/2}} \norm*{x - x'}_2 \norm*{y - y'}_2.\end{aligned}$$
First, note that $$\begin{gathered}
\label{eq:twoterms}
\ip{D_x V_t(x,y) - D_x V_t(x',y'), x - x'}_2 \\
= \ip{D_x V_t(x,y) - D_x V_t(x',y), x - x'}_2 + \ip{D_x V_t(x',y) - D_x V_t(x',y'), x - x'}_2\end{gathered}$$ By Lemma \[lem:convexgradient\], the first term on the right hand side of can be estimated by $$\frac{c}{1+ct} \norm*{x - x'}_2^2 \leq \ip{D_x V_t(x,y) - D_x V_t(x',y), x - x'}_2 \leq \frac{C}{1+Ct} \norm*{x - x'}_2^2.$$ To handle the second term on the right hand side of , define $$\begin{aligned}
\overline{V}_t(x,y) &= V_t(x,y) - \frac{c}{2(1+ct)} \norm*{x}_2^2 - \frac{c}{2} \norm*{y}_2^2 \\
\underline{V}_t(x,y) &= V_t(x,y) - \frac{C}{2(1+Ct)} \norm*{x}_2^2 - \frac{C}{2} \norm*{y}_2^2\end{aligned}$$ and recall that $\overline{V}_t$ is convex and $\underline{V}_t$ is concave and in particular $$\begin{aligned}
0 \leq H\overline{V}_t &\leq \left( \frac{C}{1+Ct} - \frac{c}{1+ct} \right) I_m \oplus (C - c) I_n \\
&= \frac{C-c}{(1+Ct)(1+ct)} I_m \oplus (C - c) I_n.\end{aligned}$$ Note that $$\begin{aligned}
D_x V_t(x',y) - D_x V_t(x',y') &= \left(D_x V_t(x',y) - \frac{c}{1+ct} x' \right) - \left( D_x V_t(x',y) - \frac{c}{1+ct} x' \right) \\
&= D_x \overline{V}_t(x',y) - D_x \overline{V}_t(x',y').\end{aligned}$$ Therefore, $$\begin{aligned}
\ip{D_x V_t(x',y) - D_x V_t(x',y'), x - x'}_2 &= \ip{D_x \overline{V}_t(x',y) - D_x \overline{V}_t(x',y'), x - x'}_2 \\
&= \ip{D \overline{V}_t(x',y) - D\overline{V}_t(x',y'), (x - x', 0)}_2.\end{aligned}$$ Now we apply Lemma \[lem:convexgradientestimate\] to $\overline{V}_t$ with the matrix $A = \frac{C-c}{(1+Ct)(1+ct)} I_m \oplus (C - c) I_n$ and conclude that $$\begin{aligned}
|\ip{D \overline{V}_t(x',y) - D\overline{V}_t(x',y'), (x - x', 0)}_2| &\leq \left( (C-c) \norm*{y-y'}_2^2 \right)^{1/2} \left( \frac{C - c}{(1+Ct)(1+ct)} \norm*{x - x'}_2^2 \right)^{1/2} \\
&\leq \frac{C-c}{(1+Ct)^{1/2}(1+ct)^{1/2}} \norm*{x - x'} \norm*{y - y'}.\end{aligned}$$ Combining this estimate for the second term of with our earlier estimate for the first term completes the proof.
\[lem:transportLipschitzestimate\] We have $$\label{eq:transportLipdx}
\norm*{F_{s,t}}_{\operatorname{Lip},dx} \leq \begin{cases}
\frac{(1+Cs)^{1/2}}{(1+Ct)^{1/2}}, & s \geq t \\
\frac{(1+cs)^{1/2}}{(1 + ct)^{1/2}} & s \leq t. \end{cases}$$ and $$\label{eq:transportLipdy}
\norm*{F_{s,t}}_{\operatorname{Lip},dy} \leq \begin{cases}
(C/c - 1)(1 + Cs)^{1/2}\left( \frac{1}{(1+Ct)^{1/2}} - \frac{1}{(1+Cs)^{1/2}} \right), & s \geq t, \\
(C/c - 1)(1 + cs)^{1/2} \left( \frac{1}{(1+Cs)^{1/2}} - \frac{1}{(1+Ct)^{1/2}} \right) & s \leq t.
\end{cases}$$
Fix $t \geq 0$ and $(x,y)$ and $(x',y')$ in $M_N({\mathbb{C}})_{sa}^m \times M_N({\mathbb{C}})_{sa}^n$ and define $$\phi(s) = \norm*{F_{s,t}(x,y) - F_{s,t}(x',y')}_2.$$ Note that $\phi$ is locally Lipschitz, hence absolutely continuous. Also, $$\begin{aligned}
2 \phi(s) \phi'(s) &= \partial_s [\phi(s)^2] \\
&= 2 \ip{\partial_s F_{s,t}(x,y) - \partial_s F_{s,t}(x',y'), F_{s,t}(x,y) - F_{s,t}(x',y')}_2 \\
&= \ip{DV_s(F_{s,t}(x,y),y) - DV_s(F_{s,t}(x',y'),y'), F_{s,t}(x,y) - F_{s,t}(x',y')}_2 \\
&\leq \frac{C}{1+Ct} \norm*{F_{s,t}(x,y) - F_{s,t}(x',y')}_2^2 \\
& \quad + \frac{C-c}{(1+Ct)^{1/2}(1+ct)^{1/2}} \norm*{F_{s,t}(x,y) - F_{s,t}(x',y')}_2 \norm*{y - y'}_2 \\
&= \frac{C}{1+Cs} \phi(s)^2 + \frac{C-c}{(1+Cs)^{1/2}(1+cs)^{1/2}} \phi(s) \norm*{y - y'}_2,\end{aligned}$$ where we have applied Lemma \[lem:twotermgradientestimate\]. It follows that whenever $\phi(s) > 0$, $$\phi'(s) \leq \frac{C}{2(1+Cs)} \phi(s) + \frac{C-c}{2(1+Cs)^{1/2}(1+cs)^{1/2}} \norm*{y - y'}_2.$$ On the other hand, since $\phi(s) \geq 0$, any point where $\phi$ is zero and $\phi$ is differentiable must be a critical point, so when $\phi(s) = 0$ the estimate is vacuously true. This inequality implies $$\begin{aligned}
\frac{d}{ds} \left[ \frac{1}{(1 + Cs)^{1/2}} \phi(s) \right] &\leq \frac{C-c}{2(1+Cs)(1+cs)^{1/2}} \norm*{y - y'}_2 \\
&\leq \frac{C(C-c)}{2c(1+Cs)^{3/2}} \norm*{y - y'}_2,\end{aligned}$$ where in the last line we have observed that $(1 + cs)^{1/2} \geq (c/C)^{1/2} (1 + Cs)^{1/2} \geq (c/C)(1 + Cs)^{1/2}$. Hence for $s \geq t$ $$\frac{1}{(1+Cs)^{1/2}} \phi(s) - \frac{1}{(1+Ct)^{1/2}} \phi(t) \leq \frac{C-c}{c} \left( \frac{1}{(1+Ct)^{1/2}} - \frac{1}{(1+Cs)^{1/2}} \right) \norm*{y - y'}_2$$ Now we substitute $\phi(s) = \norm*{F_{s,t}(x,y) - F_{s,t}(x',y')}_2$ and $\phi(t) = \norm*{x - x'}_2$ and rearrange to obtain $$\frac{1}{(1+Cs)^{1/2}} \norm*{F_{s,t}(x,y) - F_{s,t}(x',y')}_2 \leq \frac{1}{(1+Ct)^{1/2}} \norm*{x - x'}_2 + \frac{C-c}{c} \left( \frac{1}{(1+Ct)^{1/2}} - \frac{1}{(1+Cs)^{1/2}} \right) \norm*{y - y'}_2.$$ This proves the asserted estimates in the case where $s \geq t$. The argument for the case $s \leq t$ is similar. Here we use the lower bound rather than the upper bound in Lemma \[lem:twotermgradientestimate\] and get $$\phi'(s) \geq \frac{c}{2(1+cs)} \phi(s) - \frac{C-c}{2(1+Cs)^{1/2}(1+cs)^{1/2}} \norm*{y - y'}_2$$ so that $$\begin{aligned}
\frac{d}{ds} \left[ \frac{1}{(1+cs)^{1/2}} \phi(s) \right] &\geq -\frac{C-c}{2(1+Cs)^{1/2}(1+cs)} \norm*{y - y'}_2 \\
&\geq -\frac{C(C-c)}{2c(1+Cs)^{3/2}} \norm*{y - y'}_2.\end{aligned}$$ Now we take $s \leq t$ and obtain $$\begin{gathered}
\frac{1}{(1+ct)^{1/2}} \norm*{x - x'}_2 - \frac{1}{(1+cs)^{1/2}} \norm*{F_{s,t}(x,y) - F_{s,t}(x',y')}_2 \\
\geq - \frac{C-c}{c} \left( \frac{1}{(1+Cs)^{1/2}} - \frac{1}{(1+Ct)^{1/2}} \right) \norm*{y - y'}_2,\end{gathered}$$ which yields the desired estimates.
Transport in the Large $t$ Limit {#subsec:largeTtransport}
--------------------------------
We remind the reader here that we are still working in the finite-dimensional setting for a fixed value of $N$ which is suppressed in the notation. To understand the large $t$ limit of our transport maps, consider the renormalized law $$\label{eq:renormalizedlaw}
\tilde{\mu}_t := \text{law of } (\tilde{X}_t, Y) := \left( e^{-t/2} X + (1 - e^{-t})^{1/2} S, Y \right).$$ A brief computation shows that the corresponding potential is $$\label{eq:renormalizedpotential}
\tilde{V}_t(x,y) := V_{e^t - 1}(e^{t/2} x, y),$$ (here the potential is only well-defined up to an additive constant because the probability measure $\tilde{\mu}_t$ includes a normalizing constant $1/\tilde{Z}_t$ anyway, so we made a convenient choice of the additive constant). This potential satisfies the equation $$\partial_t \tilde{V}_t = \frac{1}{2N} \Delta_x \tilde{V}_t - \frac{1}{2} \norm*{D_x \tilde{V}_t}_2^2 + \frac{1}{2} \ip{D_x \tilde{V}_t, x}_2.$$ We remark that if $\tilde{\rho}_t = (1/Z_t) e^{-N^2 \tilde{V}_t}$ is the density at time $t$ and $r(x,y) = \text{const} e^{-\norm{(x,y)}_2^2/2}$ is the Gaussian density, then $$\begin{aligned}
\partial_t \tilde{\rho}_t &= \frac{1}{2N} \Delta_x \tilde{\rho}_t + \frac{1}{2} \ip{\nabla_x \tilde{\rho}_t, x}_{\operatorname{Tr}} + \frac{Nm}{2} \tilde{\rho}_t \\
&= \frac{1}{2N} \operatorname{Div}_x \left( \tilde{\rho}_t \nabla_x \left( \log \frac{\tilde{\rho}_t}{r} \right) \right).\end{aligned}$$ In other words, $\tilde{\rho}_t$ evolves according to the diffusion semigroup with respect to Gaussian measure (compare equation (33) of [@OV2000]), while the heat equation represents diffusion with respect to Lebesgue measure.
The transport functions are renormalized as follows. Because $(F_{s,t}(x,y),y)$ pushes forward $\mu_t$ to $\mu_s$, we may compute that $(\tilde{F}_{s,t}(x,y),y)$ pushes forward $\tilde{\mu}_t = \tilde{\mu}_s$, where $$\label{eq:renormalizedtransport}
\tilde{F}_{s,t}(x,y) := e^{-s/2} F_{e^s-1,e^t-1}(e^{t/2}x,y).$$ Moreover, from the differential equation , we deduce that $$\label{eq:renormalizedtransportODE}
\partial_s \tilde{F}_{s,t}(x,y) = \frac{1}{2} \left( D_x \tilde{V}_s(\tilde{F}_{s,t}(x,y),y) - \tilde{F}_{s,t}(x,y) \right).$$
As $t \to \infty$, the law $\tilde{\mu}_t$ converges to the law of $(S,Y)$, which we denote $\tilde{\mu}_\infty$. Thus, if we show that $\tilde{F}_{s,t}$ has a limit as $s \to +\infty$ or $t \to +\infty$, we will be able to transport our given law $\mu = \tilde{\mu}_0$ of $(X,Y)$ to the law of $(S,Y)$. As the first step, we deduce from Lemma \[lem:transportLipschitzestimate\] the following Lipschitz estimates on $\tilde{F}_{s,t}$ which are uniform in $s$ and $t$. Note also that the coefficient of $\norm*{y - y'}_2$ goes to zero as $s, t \to \infty$.
\[lem:transportLipschitzestimate2\] We have $$\label{eq:transportLipdx2}
\norm*{\tilde{F}_{s,t}}_{\operatorname{Lip},dx} \leq \max(C,1/c)^{1/2}$$ and $$\label{eq:transportLipdy2}
\norm*{\tilde{F}_{s,t}}_{\operatorname{Lip},dy} \leq (C/c - 1) \max(C,1/C)^{3/2} |e^{-s/2} - e^{-t/2}|.$$ In particular, $$\label{eq:transportLip2}
\norm*{\tilde{F}_{s,t}}_{\operatorname{Lip}} \leq \max(C,1/c)^{7/2}.$$
For the first estimate, for the case where $s \geq t$, direct substitution of into of Lemma \[lem:transportLipschitzestimate\] shows that $$\norm*{\tilde{F}_{s,t}}_{\operatorname{Lip},dx} \leq e^{-s/2} \frac{(1 + C(e^s - 1))^{1/2}}{(1 + C(e^t - 1))^{1/2}} e^{t/2} = \frac{(C + (1 - C)e^{-s})^{1/2}}{(C + (1 - C)e^{-t})^{1/2}}.$$ The function $C + (1 - C) e^{-s}$ is clearly monotone on $[0,+\infty)$ and achieves the values $1$ and $C$ at $0$ and $+\infty$ respectively, and hence is between $\min(1,C)$ and $\max(1,C)$. Hence, $$\norm*{\tilde{F}_{s,t}}_{\operatorname{Lip},dx} \leq \frac{\max(1,C)^{1/2}}{\min(1,C)^{1/2}} = \max(C,1/C)^{1/2} \leq \max(C,1/c)^{1/2}.$$ The case where $s \leq t$ follows by the same argument, where the bound this time is $\max(c,1/c)^{1/2} \leq \max(C,1/c)^{1/2}$.
For the second estimate, we apply . Note in , in the case $s \leq t$, we may use $(1 + cs)^{1/2} \leq (1 + Cs)^{1/2}$ and thus in both cases $s \geq t$ or $s \leq t$, $$\begin{aligned}
\norm{F_{s,t}}_{\operatorname{Lip},dy} &\leq (C/c - 1) (1 + Cs)^{1/2} \left| \frac{1}{(1 + Cs)^{1/2}} - \frac{1}{(1 + Ct)^{1/2}} \right| \\
&= (C/c - 1)(1 + Cs)^{1/2} \left| \int_s^t \frac{C}{2(1 + Cu)^{3/2}}\,du \right|\end{aligned}$$ This implies that $$\begin{aligned}
\norm*{\tilde{F}_{s,t}}_{\operatorname{Lip},dy} &\leq e^{-s/2} (C/c - 1)(1 + C(e^s - 1))^{1/2} \left| \int_{e^s-1}^{e^t-1} \frac{C}{2(1 + Cu)^{3/2}}\,du \right| \\
&= (C/c - 1)e^{-s/2}(1 + C(e^s - 1))^{1/2} \left| \int_s^t \frac{Ce^w}{2(1 + C(e^w - 1))^{3/2}}\,dw \right| \\
&\leq (C/c - 1) \max(1,C)^{1/2} \left| \int_s^t \frac{Ce^w}{2 \min(1,C)^{3/2} e^{3w/2}}\,dw \right| \\
&\leq (C/c - 1) \frac{\max(1,C)^{1/2} C}{\min(1,C)^{3/2}} |e^{-t/2} - e^{-s/2}| \\
&\leq (C/c - 1) \max(C,1/C)^{3/2} |e^{-t/2} - e^{-s/2}|.\end{aligned}$$ where we have again applied $\min(1,C) e^s \leq 1 + C(e^s - 1) \leq \max(1,C) e^s$.
For the last estimate , observe that $$\begin{aligned}
\norm*{\tilde{F}_{s,t}}_{\operatorname{Lip}} &\leq \norm*{\tilde{F}_{s,t}}_{\operatorname{Lip},dx} + \norm*{\tilde{F}_{s,t}}_{\operatorname{Lip},dy} \\
&\leq \max(C,1/c)^{1/2} + (C/c - 1) \max(C,1/C)^{3/2} |e^{-s/2} - e^{-t/2}| \\
&\leq \max(C,1/c)^{3/2} + (\max(C,1/c)^2 - 1) \max(C,1/c)^{3/2} \\
&= \max(C,1/c)^{7/2}. \qedhere\end{aligned}$$
\[lem:transportLipschitzestimate3\] Let $\pi_1$ denote the function $\pi_1(x,y) = x$. Then $$\label{eq:transportLipdx3}
\norm*{\tilde{F}_{s,t} - \pi_1}_{\operatorname{Lip},dx} \leq \frac{1}{2} (\max(C,1/c) - 1) \max(C,1/C)^{1/2} |e^{-s} - e^{-t}|$$ and $$\label{eq:transportLip3}
\norm*{\tilde{F}_{s,t} - \pi_1}_{\operatorname{Lip}} \leq (\max(C,1/c)^3 - 1) \max(C,1/c)^{1/2} |e^{-s/2} - e^{-t/2}|.$$
Let $U_s(x,y) = \tilde{V}_s(x,y) - (1/2) \norm*{x}_2^2$. Then says that $$\partial_s \tilde{F}_{s,t}(x,y) = \frac{1}{2} D_x U_s(\tilde{F}_{s,t}(x,y),y).$$ Moreover, we have $$\frac{ce^s}{1 + c(e^s - 1)} \leq H_x \tilde{V}_s \leq \frac{Ce^s}{1 + C(e^s - 1)}.$$ We can bound $H_x U_s$ above and below by subtracting $1$ from both sides, which after some computation reduces to $$\frac{c - 1}{1 + c(e^s - 1)} \leq H_x U_s \leq \frac{C - 1}{1 + C(e^s - 1)}.$$ Therefore, we have $-L \leq H_x U_s \leq L$ where $$L = \max \left( -\frac{c - 1}{1 + c(e^s - 1)}, \frac{C - 1}{1 + C(e^s - 1)} \right).$$ We claim that $L \leq L' := (\max(C,1/c) - 1) e^{-s}$. If the first term $(1 - c) / (1 + c(e^s - 1))$ is negative, then it is $\leq L'$ automatically, but if it is positive, then $c \leq 1$ and hence $$\frac{1 - c}{1 + c(e^s - 1)} \leq \frac{1 - c}{c + c(e^s - 1)} = (1/c - 1)e^{-s} \leq (\max(C,1/c) - 1)e^{-s}.$$ Similarly, if $(C - 1) / (1 + C(e^s - 1))$ is negative, there is nothing to prove, but otherwise $C \geq 1$, and hence $$\frac{C - 1}{1 + C(e^s - 1)} \leq \frac{C - 1}{1 + (e^s - 1)} = (C - 1) e^{-s} \leq (\max(C,1/c) - 1) e^{-s}.$$ But $-L' \leq H_x U_s \leq L'$ implies that $D_x U_s$ is $L'$-Lipschitz in $x$. Therefore, $$\begin{aligned}
\norm*{\partial_s \tilde{F}_{s,t}(x,y) - \partial_s \tilde{F}_{s,t}(x',y)}_2 &= \frac{1}{2} \norm*{DU_s(\tilde{F}_{s,t}(x,y),y) - DU_s(\tilde{F}_{s,t}(x',y))}_2 \\
&\leq \frac{1}{2}(\max(C,1/c) - 1) e^{-s} \norm*{\tilde{F}_{s,t}(x,y) - \tilde{F}_{s,t}(x',y)}_2.\end{aligned}$$ Applying in the case where $s \geq t$, we get $$\norm*{\partial_s \tilde{F}_{s,t}(x,y) - \partial_s \tilde{F}_{s,t}(x',y)}_2 \leq \frac{1}{2} (\max(C,1/c) - 1) \max(C,1/c)^{1/2} e^{-s} \norm*{x - x'}_2.$$ Hence, $$\begin{aligned}
\norm*{\tilde{F}_{s,t}(x,y) - \tilde{F}_{s,t}(x',y) - (x - x')}_2 &\leq \left| \int_t^s \norm*{\partial_u \tilde{F}_{u,t}(x,y) - \partial_u \tilde{F}_{u,t}(x',y)}_2\,du \right| \\
&\leq \frac{1}{2} (\max(C,1/c) - 1) \max(C,1/c)^{1/2} |e^{-s} - e^{-t}| \norm*{x - x'}_2.\end{aligned}$$ which proves the desired estimate .
To check the second estimate , first observe $$\frac{1}{2} |e^{-s} - e^{-t}| = \int_{\min(s,t)}^{\max(s,t)} \frac{1}{2} e^{-u}\,du \leq \int_{\min(s,t)}^{\max(s,t)} \frac{1}{2} e^{-u/2}\,du = |e^{-s/2} - e^{-t/2}|,$$ Moreover, $\norm{\tilde{F}_{s,t} - \pi_1}_{\operatorname{Lip},dy} = \norm{\tilde{F}_{s,t}}_{\operatorname{Lip},dy}$. Therefore, using and , $$\begin{aligned}
\norm{\tilde{F}_{s,t} - \pi_1}_{\operatorname{Lip}} &\leq \norm{\tilde{F}_{s,t} - \pi_1}_{\operatorname{Lip},dx} + \norm{\tilde{F}_{s,t} - \pi_1}_{\operatorname{Lip},dy} \\
&\leq (\max(C,1/c) - 1) \max(C,1/C)^{1/2} \frac{1}{2} |e^{-s} - e^{-t}| + (C/c - 1) \max(C,1/C)^{3/2} |e^{-s/2} - e^{-t/2}| \\
&\leq [(\max(C,1/c) - 1) \max(C,1/c)^{1/2} + (\max(C,1/c)^2 - 1) \max(C,1/c)^{3/2}] |e^{-s/2} - e^{-t/2}| \\
&= (\max(C,1/c)^3 - 1) \max(C,1/c)^{1/2} |e^{-s/2} - e^{-t/2}|. \qedhere.\end{aligned}$$
\[prop:largeTtransport\] The limits $\tilde{F}_{s,\infty} := \lim_{t \to \infty} \tilde{F}_{s,t}$ and $\tilde{F}_{\infty,t} = \lim_{s \to \infty} \tilde{F}_{s,t}$ exist for $s, t \geq 0$. More precisely, let $(X,Y)$ and $(\tilde{X}_t,Y)$ be a pair of random variables with the laws $\mu$ and $\tilde{\mu}_t$ as above. Then $$\begin{gathered}
\label{eq:transportconvergenceestimate}
\norm*{\tilde{F}_{s,\infty}(x,y) - \tilde{F}_{s,t}(x,y)}_2 \leq \max(C,1/c)^{1/2} e^{-t/2} \norm*{E(X)}_2 \\ + e^{-t/2}(\max(C,1/c)^3 - 1) \max(C,1/c) \left( \norm{(x, y - E(Y))}_2 + (m + \operatorname{Var}(Y))^{1/2} \right)\end{gathered}$$ and $$\begin{gathered}
\norm*{\tilde{F}_{s,t}(x,y) - \tilde{F}_{\infty,t}(x,y)}_2 \leq \frac{1}{2} (\max(C,1/c) - 1) \max(C,1/c)^{1/2} e^{-s} \biggl( e^{-s/2} \norm{E(X)}_2 \\ + \max(C,1/c)^{7/2} \left( \norm{(x - e^{-t/2}E(X), y - E(Y))}_2 + (e^{-t} \operatorname{Var}(X) + (1 - e^{-t})m + \operatorname{Var}(Y))^{1/2} \right) \biggr)\end{gathered}$$ The estimates of Lemmas \[lem:transportLipschitzestimate2\] and \[lem:transportLipschitzestimate3\] extend to the cases where $s$ or $t$ is infinite, where we define $\tilde{F}_{\infty,\infty}(x,y) = x$. Moreover, if $(\tilde{X}_t,Y) \sim \tilde{\mu}_t$, then we have the relation $(\tilde{F}_{s,t}(\tilde{X}_t,Y),Y) \sim (\tilde{X}_s,Y)$ when $s, t \in [0,\infty]$.
We have written the explicit form of the estimates here in order to emphasize that the bounds are dimension-independent; they only depend on the parameters $m$, $n$, $c$, $C$, $\norm{E(X)}_2$, $\norm{E(Y)}_2$, $\operatorname{Var}(X)$, and $\operatorname{Var}(Y)$. The estimates also become sharper when $c$ and $C$ are close to $1$, which would include the situation where $V(x,y)$ is a perturbation of the quadratic potential $(1/2)[\norm*{x}_2^2 + \norm*{y}_2^2]$. This perturbative setting was studied first in the literature, for instance in [@GMS2006] and [@GS2014]; see [@Jekel2018 §8.3] for further discussion.
We first consider the case where $s$ is fixed and $t \to +\infty$. Note that by , $$\begin{aligned}
\label{eq:convergenceproof1}
\norm*{\tilde{F}_{s,t'}(x,y) - \tilde{F}_{s,t}(x,y)}_2 &= \norm*{\tilde{F}_{s,t'}(\tilde{F}_{t',t}(x,y),y) - \tilde{F}_{s,t'}(x,y)}_2 \\
&\leq \norm*{\tilde{F}_{s,t'}}_{\operatorname{Lip},dx} \norm*{\tilde{F}_{t,t'}(x,y) - x}_2 \nonumber \\
&\leq \max(C,1/c)^{1/2} \norm*{\tilde{F}_{t,t'}(x,y) - x}_2. \nonumber\end{aligned}$$ By Lemma \[lem:transportLipschitzestimate3\], $$\norm*{\tilde{F}_{t,t'} - \pi_1}_{\operatorname{Lip}} \leq L |e^{-t/2} - e^{t'/2}|,$$ where $L = \max(C,1/c)^3 - 1) \max(C,1/c)^{1/2}$. Then we apply Lemma \[lem:RVboundedness\] to $G(x,y) = \tilde{F}_{t,t'}(x,y) - x$ with the random variable $(\tilde{X}_{t'},Y)$. Note that $(\tilde{X}_{t'},Y)$ has mean $(e^{-t'/2} E(X), E(Y))$ and variance $e^{-t'} \operatorname{Var}(X) + (1 - e^{-t'})m + \operatorname{Var}(Y)$. Moreover, $$E[G(\tilde{X}_{t'},Y)] = E[\tilde{X}_t] - E[\tilde{X}_{t'}] = (e^{-t/2} - e^{-t'/2}) E(X).$$ Thus, by Lemma \[lem:RVboundedness\], $$\begin{gathered}
\label{eq:transportCauchyestimate1}
\norm{\tilde{F}_{t,t'}(x,y) - x}_2 \leq |e^{-t/2} - e^{-t'/2}| \norm{E(X)}_2 \\ + L |e^{-t/2} - e^{t'/2}| \left( \norm{(x - e^{-t'/2} E(X), y - E(Y))}_2 + (e^{-t'} \operatorname{Var}(X) + (1 - e^{-t'})m + \operatorname{Var}(Y))^{1/2} \right).\end{gathered}$$ Plugging this into , we see that $\tilde{F}_{s,t}$ is Cauchy in $t$ as $t \to +\infty$. Moreover, we obtain the estimate by taking $t' \to \infty$ in and multiplying by $\norm*{\tilde{F}_{s,t}}_{\operatorname{Lip},dx} \leq \max(c,1/c)^{1/2}$.
Now let us fix $t$ and consider when $s'$ and $s$ approach $\infty$. The argument for this case is similar but antisymmetrical. We estimate $$\begin{aligned}
\norm*{\tilde{F}_{s',t}(x,y) - \tilde{F}_{s,t}(x,y)}_2 &= \norm*{\tilde{F}_{s',s}(\tilde{F}_{s,t}(x,y),y) - \tilde{F}_{s,t}(x,y)}_2 \\
&\leq \norm*{\tilde{F}_{s',s} - \pi_1}_{\operatorname{Lip},dx} \norm*{\tilde{F}_{s,t}(x,y)}_2 \\
&\leq \frac{1}{2} (\max(C,1/c) - 1) \max(C,1/c)^{1/2} |e^{-s} - e^{-s'}| \norm*{\tilde{F}_{s,t}(x,y)}_2,\end{aligned}$$ where the last line follows from . Then by applying Lemma \[lem:RVboundedness\] to the function $\tilde{F}_{s,t}(x,y)$ and the random variable $(\tilde{X}_t,Y)$, together with , we obtain $$\begin{gathered}
\norm*{\tilde{F}_{s,t}(x,y)}_2 \leq e^{-s/2} \norm{E(X)}_2 \\
+ \max(C,1/c)^{7/2} \left( \norm{(x - e^{-t/2}E(X), y - E(Y))}_2 + (e^{-t} \operatorname{Var}(X) + (1 - e^{-t})m + \operatorname{Var}(Y))^{1/2} \right)\end{gathered}$$ This produces an estimate on $\norm*{\tilde{F}_{s',t} - \tilde{F}_{s,t}}_2$ which shows that $\tilde{F}_{s,t}$ is Cauchy as $s \to \infty$, so that $\tilde{F}_{\infty,t}$ is well-defined. The explicit bound on the rate of convergence follows fixing $s$ and $t$, combining the above estimates, and taking $s' \to \infty$.
Finally, since we have established convergence of $\tilde{F}_{s,t}$ as $s$ or $t$ approaches $\infty$, a routine argument with limits will extend the estimates of Lemmas \[lem:transportLipschitzestimate2\] and \[lem:transportLipschitzestimate3\], and the transport relations, to the cases where $s$ or $t$ is $+\infty$.
Transport in the Large $N$ Limit {#subsec:largeNtransport}
--------------------------------
If $V^{(N)} \in \mathcal{E}_{m+n}^{(N)}(c,C)$ and $\{DV^{(N)}\}$ is asymptotically approximable by trace polynomials, then we must show that the associated sequence of transport maps is asymptotically approximable by trace polynomials, and hence conclude that they define transport for the non-commutative random variables in the large $N$ limit.
\[thm:transport1\] For $N \in {\mathbb{N}}$, let $V^{(N)}(x,y)$ be a potential on $M_N({\mathbb{C}})_{sa}^{m+n}$ satisfying Assumption \[ass:convexRMM\] for some $0 < c \leq C$, and let $\mu^{(N)}$ be the corresponding probability measures on $M_N({\mathbb{C}})_{sa}^{m+n}$. Let $(X^{(N)}, Y^{(N)})$ be a random variable given by $\mu^{(N)}$ and let $S^{(N)}$ be an independent GUE $m$-tuple. Let $$\tilde{\mu}_t^{(N)} \text{ be the law of } (\tilde{X}_t, Y) = (e^{-t/2} X^{(N)} + (1 - e^{-t})^{1/2} S^{(N)}, Y^{(N)}),$$ and let $\tilde{V}_t^{(N)}(x,y) = R_{e^t-1}^{(N)} V^{(N)}(e^{t/2} x, y)$ be the corresponding potential. Similarly, let $\mu_\infty^{(N)}$ be the law of $(S^{(N)}, Y^{(N)})$. For $s, t \in [0,\infty)$, let $\tilde{F}_{s,t}^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})_{sa}^m$ be the solution of the initial value problem $$\begin{aligned}
\tilde{F}_{t,t}^{(N)}(x,y) &= x \\
\partial_s \tilde{F}_{s,t}^{(N)}(x,y) &= \frac{1}{2} \left( D_x \tilde{V}_s(\tilde{F}_{s,t}(x,y), y) - \tilde{F}_{s,t}(x,y) \right).\end{aligned}$$ Then
1. The family $\tilde{F}_{s,t}^{(N)}$ extends continuously to $(s,t) \in [0,\infty]^2$.
2. $\tilde{F}_{s,t}^{(N)} \circ \tilde{F}_{t,u}^{(N)} = \tilde{F}_{s,u}^{(N)}$.
3. $(\tilde{F}_{s,t}^{(N)}(\tilde{X}_t^{(N)},Y^{(N)}), Y^{(N)}) \sim (\tilde{X}_s^{(N)}, Y^{(N)})$.
4. For $s, t \in [0,\infty]$, the sequence $\{\tilde{F}_{s,t}^{(N)}\}_{N \in {\mathbb{N}}}$ is $(C/c) \max(C,1/c)^{1/2}$-Lipschitz for all $s$, $t$, and $N$, and it is asymptotically approximable by trace polynomials as $N \to \infty$.
Recall in §\[subsec:largeTtransport\] we defined $\tilde{F}_{s,t}^{(N)}$ by renormalizing $F_{s,t}^{(N)}$. However, that definition is equivalent to the definition of $\tilde{F}_{s,t}$ given in this theorem because both definitions produce a solution to the ODE . Of course, global uniqueness of the solution holds because the vector field $D_x \tilde{V}_t^{(N)}(x,y) - x$ is uniformly Lipschitz in $(x,y)$ on any compact time interval (as we discuss in more detail below).
So claims (1), (2), and (3) follow immediately from Proposition \[prop:largeTtransport\]. The estimate for the Lipschitz norm of $\tilde{F}_{s,t}^{(N)}$ was shown in .
We finish by showing asymptotic approximability using the results of §\[subsec:vectorfields\]. Let $V_t^{(N)} = R_t^{(N)} V^{(N)}$. By Theorem \[thm:Rtsemigroup\] (3c), $D_x V_t^{(N)}$ is uniformly continuous in $t$ on $[0,\infty)$. Since $D_x \tilde{V}_t^{(N)}(x,y) = e^{t/2} D_x V_{e^t - 1}^{(N)}(e^{t/2}x,y)$, it follows that $D_x \tilde{V}_t^{(N)}$ is uniformly continuous in $t$ on $[0,T]$ for every $T > 0$, with modulus of continuity independent of $N$, and recall it is also uniformly Lipschitz in $(x,y)$, since $0 \leq H\tilde{V}_t^{(N)} \leq \max(C, Ce^t / (1 + C(e^t - 1))$.
Consequently, $(1/2)(D_x \tilde{V}_t^{(N)}(x,y) - x)$ is uniformly continuous in $t$ on $[0,T]$ and uniformly Lipschitz in $(x,y)$. Also, we showed that $D_x V_t^{(N)}$ is asymptotically approximable by trace polynomials in the proof of Theorem \[thm:convergenceofentropy\], and hence so is $D_x \tilde{V}_t^{(N)}$. Thus, $(1/2)(D_x \tilde{V}_t^{(N)}(x,y) - x)$ satisfies Assumption \[ass:vectorfield2\], so we may apply Proposition \[prop:ODE2\] to deduce that $\tilde{F}_{s,t}^{(N)}$ is asymptotically approximable by trace polynomials for $s, t \in [0,\infty)$. This property extends to the case where $s$ or $t$ is infinite using Lemma \[lem:limits\] and Proposition \[prop:largeTtransport\].
Rather than citing the proof of Theorem \[thm:convergenceofentropy\], one could also argue that $D_x V_t^{(N)}$ is asymtotically approximable directly from the construction of the semigroup $R_t^{(N)}$ using the same reasoning as [@Jekel2018 Proposition 6.8]. Moreover, this method would also show that $D(R_t^{(N)} V^{(N)})$ is asymptotically approximable by trace polynomials provided we can prove analogues of Theorem \[thm:Rtsemigroup\] (2) and (3) for $D(R_t^{(N)} V^{(N)})$ rather than only $D_x(R_t^{(N)} V^{(N)})$. However, all this is unnecessary work for our present purpose.
\[thm:transport2\] With all the notation of the previous theorem, let $(X,Y)$ be a non-commutative random variable distributed according to the limiting free Gibbs law $\lambda$, let $S$ be a freely independent free semicircular $m$-tuple, and let $\tilde{X}_t = e^{-t/2} X + (1 - e^{-t})^{1/2} S$. Define $\tilde{F}_{s,t}$ by $\tilde{F}_{s,t}^{(N)} \rightsquigarrow \tilde{F}_{s,t}$. For $s,t,u \in [0,+\infty]$, we have
1. $\tilde{F}_{s,t}$ is $(C/c) \max(C,1/c)^{1/2}$-Lipschitz with respect to $\norm{\cdot}_2$.
2. $\tilde{F}_{s,t} \circ \tilde{F}_{t,u} = \tilde{F}_{s,u}$.
3. $(\tilde{F}_{s,t}(\tilde{X}_t,Y), Y) \sim (\tilde{X}_s,Y)$ in non-commutative law.
4. We have $$\begin{gathered}
\norm*{\tilde{F}_{s,t}(\tilde{X}_t,Y) - \tilde{F}_{s',t}(\tilde{X}_t,Y) - (e^{-s/2} - e^{-s'/2})\tau(X)}_\infty \\
\leq (\max(C,1/c)^3 - 1) \max(C,1/c) |e^{-s/2} - e^{-s'/2}| \Theta.
\end{gathered}$$ where $\Theta$ is the universal constant from Proposition \[prop:operatornormestimate\].
In particular, $\mathrm{W}^*(X,Y)$ is isomorphic to $\mathrm{W}^*(S,Y)$, which is the free product $\mathrm{W}^*(S) * \mathrm{W}^*(Y)$.
We know that there exists $\tilde{F}_{s,t}$ such that $\tilde{F}_{s,t}^{(N)} \rightsquigarrow \tilde{F}_{s,t}$ because of Lemma \[lem:AATP\]. Then (1) and (2) follow from the corresponding properties of $\tilde{F}_{s,t}^{(N)}$ by straightforward limit arguments.
As remarked in the last proof $D\tilde{V}_t^{(N)}$ is asymptotically approximable by trace polynomials. We also know $D\tilde{V}_t^{(N)}$ is uniformly convex and semi-concave, and thus by Theorem \[thm:freeGibbslaw\], the non-commutative law of $(\tilde{X}_t^{(N)}, Y^{(N)})$ converges in probability to some non-commutative law. Of course, the limiting non-commutative law must be the non-commutative law of $(\tilde{X}_t,Y)$ because the joint non-commutative law of $(X^{(N)},Y^{(N)},S^{(N)})$ converges in probability to that $(X,Y,S)$ (as in the proof of Theorem \[thm:convergenceofentropy\]).
With this relation between the laws of $(\tilde{X}_t^{(N)},Y^{(N)})$ and $(\tilde{X}_t,Y)$ in hand, we can prove (3) by taking the large $N$ limit using Corollary \[cor:convergenceofexpectation\]. Indeed, if $f \in \overline{\operatorname{TrP}}_{m+n}^1$ is $\norm{\cdot}_2$-uniformly continuous, then $f(\tilde{F}_{s,t}^{(N)}(x,y),y)$ is also $\norm{\cdot}_2$-uniformly continuous and asymptotically approximable by trace polynomials by Lemma \[lem:composition\]. Thus, applying Corollary \[cor:convergenceofexpectation\] to this function and the function $1$, we get $$\begin{aligned}
\tau\left( f(\tilde{F}_{s,t}(\tilde{X}_t,Y), Y) \right) &= \lim_{N \to \infty} E[\tau_N (f(\tilde{F}_{s,t}(\tilde{X}_t^{(N)},Y^{(N)}), Y^{(N)}))] \\
&= \lim_{N \to \infty} E[\tau_N(f(\tilde{X}_s^{(N)},Y^{(N)}))] \\
&= \tau(f(\tilde{X}_s,Y)).\end{aligned}$$ Hence, $\tau\left( f(\tilde{F}_{s,t}(\tilde{X}_t,Y), Y) \right) = \tau(f(\tilde{X}_s,Y))$ for all $f \in \overline{\operatorname{TrP}}_m^1$ that are uniformly continuous in $\norm{\cdot}_2$. But by Proposition \[prop:realizationofoperators\] such functions $f$ can realize every element in the $\mathrm{W}^*$-algebra generated by $(\tilde{X}_s,Y)$, and in particular all the non-commutative polynomials in $(\tilde{X}_s,Y)$. Hence, $(\tilde{F}_{s,t}(\tilde{X}_t,Y), Y) \sim (\tilde{X}_s,Y)$ in non-commutative law as desired.
\(4) Note that $$\tilde{F}_{s,t}(\tilde{X}_t,Y) - \tilde{F}_{s',t}(\tilde{X}_t,Y) = (\pi_1 - \tilde{F}_{s,s'}) \circ (\tilde{F}_{s',t}(\tilde{X}_t,Y),Y),$$ but $(\tilde{F}_{s',t}(\tilde{X}_t,Y),Y) \sim (\tilde{X}_{s'}, Y)$ in non-commutative law. Hence, it suffices to prove the desired estimate for $\tilde{F}_{s,s'}(\tilde{X}_{s'},Y) - \tilde{X}_{s'}$ rather than $\tilde{F}_{s,t}(\tilde{X}_t,Y) - \tilde{F}_{s',t}(\tilde{X}_t,Y)$. Now $(\tilde{X}_{s'}, Y)$ arises as the large $N$ limit of the matrix models given by potential $\tilde{V}_{s'}^{(N)}$. By Lemma \[lem:Qt\] (4), we have $HV_t^{(N)} \geq c(1+ct)^{-1} I_m \oplus c I_n$, so that $$H \tilde{V}_{s'}^{(N)} \geq \frac{ce^{s'}}{1 + c(e^{s'} - 1)} I_m \oplus c I_n \geq \min(1,c) I_{m+n} \geq \frac{1}{\max(C,1/c)} I_{m+n}.$$ By Remark \[rem:unitaryinvariance\], there exists a sequence of random matrix models for $(\tilde{X}_{s'},Y)$ given by uniformly convex potentials which are also unitarily invariant (even if this is not true of our original model), with the same lower bound $1 / \max(C,1/c)$ for the Hessian of the potential. Therefore, by Proposition \[prop:operatornormestimate\], $$\norm*{\tilde{F}_{s,s'}(\tilde{X}_{s'},Y) - \tilde{X}_{s'} - \tau[\tilde{F}_{s,s'}(\tilde{X}_{s'},Y) - \tilde{X}_{s'}]}_\infty \leq \max(C,1/c)^{1/2} \Theta \norm*{\tilde{F}_{s,s'} - \pi_1}_{\operatorname{Lip}}.$$ We finish by substituting the estimate $$\norm{\tilde{F}_{s,s'} - \pi_1}_{\operatorname{Lip}} \leq (\max(C,1/c)^3 - 1) \max(C,1/c)^{1/2} |e^{-s/2} - e^{-s'/2}|$$ which follows from and Lemma \[lem:limituniformlycontinuous\] (the latter lemma is needed since the original statement of is for the finite-dimensional setting for a fixed $N$).
The last claim regarding $\mathrm{W}^*$-algebras follows from (3) by examining the case with $s = 0$ and $t = \infty$ or vice versa.
Applications {#sec:applications}
============
We show that Assumption \[ass:convexRMM\] is preserved under independent joins, marginals, convolution, and linear changes of variables. We conclude that for the convex free Gibbs laws considered here, $\chi^*$ satisfies additivity under conditioning. Moreover, by iterating our conditional transport results, we obtain “lower-triangular” transport maps from a convex free Gibbs law to the law of a free semicircular family, which also satisfy the entropy-cost inequality relative to the semicircular law, analogous to the triangular transport achieved in the classical case by [@BKM2005 Corollary 3.10].
Operations on Convex Gibbs Laws {#subsec:operations}
-------------------------------
Recall that Assumption \[ass:convexRMM\] for a sequence $\{V^{(N)}\}$ of potentials $M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ states that $c \leq HV^{(N)} \leq C$ for some constants $c$ and $C$, the sequence $\{DV^{(N)}\}$ is asymptotically approximable by trace polynomials, and $\int x_j \,d\mu^{(N)}(x)$ is a scalar matrix for each $j$, where $\mu^{(N)}$ is the measure associated to $DV^{(N)}$.
\[prop:independentjoins\] Suppose that $V_1^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ and $V_2^{(N)}: M_N({\mathbb{C}})_{sa}^n \to {\mathbb{R}}$ satisfy Assumption \[ass:convexRMM\] for some $0 < c \leq C$. Then $V^{(N)}(x,y) := V_1^{(N)}(x) + V_2^{(N)}(y)$ also satisfies Assumption \[ass:convexRMM\] for the same $c$ and $C$.
Moreover, let $\mu_1^{(N)}$, $\mu_2^{(N)}$, and $\mu^{(N)}$ be the measures associated to $V_1^{(N)}$, $V_2^{(N)}$, and $V^{(N)}$ respectively, and let $\lambda_1$, $\lambda_2$, and $\lambda$ be the respective limiting free Gibbs laws given by Theorem \[thm:freeGibbslaw\]. Then $\mu^{(N)}$ is the independent join of $\mu_1^{(N)}$ and $\mu_2^{(N)}$ and $\lambda$ is the freely independent join of $\lambda_1$ and $\lambda_2$.
The claim $c \leq HV^{(N)} \leq C$ follows because $HV^{(N)}(x,y) = HV_1^{(N)}(x) \oplus HV_2^{(N)}(y)$. The claim about asymptotic approximation by trace polynomials follows because $DV^{(N)}(x,y) = (DV_1^{(N)}(x), DV_2^{(N)}(y))$ and each component is asymptotically approximable by trace polynomials.
The probability density for $\mu^{(N)}$ is the tensor product of the probability densities for $\mu_1^{(N)}$ and $\mu_2^{(N)}$ and hence $\mu^{(N)}$ is the independent join of these two marginal laws. It follows that $\int x_j \,d\mu^{(N)}(x)$ and $\int y_j \,d\mu^{(N)}(y)$ are scalar matrices, hence Assumption \[ass:convexRMM\] holds for $V^{(N)}$.
Let $(X^{(N)},Y^{(N)}) \sim \mu^{(N)}$ be random variables and let $(X,Y) \sim \lambda$ be non-commutative random variables. Then by Theorem \[thm:convergenceofentropy\], $$\begin{aligned}
\Phi^*(X,Y) &= \lim_{N \to \infty} E\left[ \norm{DV^{(N)}(X^{(N)},Y^{(N)})}_2^2 \right] \\
&= \lim_{N \to \infty} E\left[ \norm{DV_1^{(N)}(X^{(N)})}_2^2 \right] + \lim_{N \to \infty} E\left[ \norm{DV_2^{(N)}(Y^{(N)})}_2^2 \right] \\
&= \Phi^*(X) + \Phi^*(Y).\end{aligned}$$ It was shown in [@VoiculescuFE6 Proposition 5.18(c)] that $\Phi^*(X,Y) = \Phi^*(X) + \Phi^*(Y)$ implies that $X$ and $Y$ are freely independent.
\[prop:marginals\] Suppose that $V^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to {\mathbb{R}}$ satisfies Assumption \[ass:convexRMM\]. Let $\mu^{(N)}$ be the corresponding law, let $(X^{(N)},Y^{(N)}) \sim \mu^{(N)}$ and let $\mu_1^{(N)}$ and $\mu_2^{(N)}$ be the laws of $X^{(N)}$ and $Y^{(N)}$. Then $\mu_1^{(N)}$ and $\mu_2^{(N)}$ are given by a potentials $W_1^{(N)}$ and $W_2^{(N)}$ that also satisfy Assumption \[ass:convexRMM\] for the same values of $c$ and $C$.
By symmetry, it suffices to prove the claims for $\mu_2^{(N)}$. First, it is immediate that the mean of $y_j$ under $\mu_2^{(N)}$ is a scalar, since it is $E[Y_j^{(N)}]$. Moreover, if we define $$V_2^{(N)}(x) = -\frac{1}{N^2} \int e^{-N^2 V^{(N)}(x,y)}\,dx,$$ then (as in the proof of Theorem \[thm:convergenceofentropy\]) we may compute $DV_2^{(N)}$ by differentiating under the integral and obtain $$DV_2^{(N)}(Y^{(N)}) = E[D_y V(X^{(N)}, Y^{(N)}) | Y^{(N)}].$$ It follows by Theorem \[thm:conditionalexpectation\] that $\{DV_2^{(N)}\}$ is asymptotically approximable by trace polynomials.
Finally, the fact that $c \leq HV_2^{(N)} \leq C$ follows from [@BL1976 Theorem 4.3], or alternatively by the following reasoning. Let $\mu_t^{(N)}$ be the law of $(e^{-t/2} X^{(N)} + (1 - e^{-t})^{1/2} S^{(N)}, Y^{(N)})$, where $S^{(N)}$ is an independent GUE tuple. The corresponding potential $\tilde{V}_t^{(N)}$ is given by and it satisfies $$\frac{ce^t}{1 + c(e^t - 1)} I_m \oplus c I_n \leq H\tilde{V}_t^{(N)} \leq \frac{Ce^t}{1 + C(e^t - 1)} I_m \oplus C I_n$$ by direct substitution of into Lemma \[lem:Qt\] (4) and hence $$\min(c,1) I_m \oplus c I_n \leq H\tilde{V}_t^{(N)} \leq \max(C,1) I_m \oplus C I_n$$ Now as $t \to \infty$, the law $\tilde{\mu}_t$ converges to the law $\tilde{\mu}_\infty$ of $(S^{(N)},Y^{(N)})$. By applying Lemma \[lem:limitoflogconcave\], $\tilde{\mu}_\infty$ is given by some potential $W_2^{(N)}(x,y)$ satisfying $$\min(c,1) I_m \oplus c I_n \leq HW_2^{(N)} \leq \max(C,1) I_m \oplus C I_n.$$ However, we know that $W_2^{(N)}(x,y) = (1/2) \norm{x}_2^2 + V_2^{(N)}(y) + \text{constant}$ because the potential corresponding to a law is unique up to an additive constant. This implies that $c \leq HV_2^{(N)} \leq C$ as desired.
\[prop:lineartransformations\] Let $V^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ satisfy Assumption \[ass:convexRMM\] for some $0 < c \leq C$, and let $X^{(N)}$ be the corresponding random variable. Let $A$ be an invertible $m \times m$ matrix with real entries and let $A^{(N)}$ denote the linear transformation $M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})_{sa}^m$ given by $$(A^{(N)}x)_i = \sum_{j=1}^m A_{i,j}x.$$ Then $\widehat{V}^{(N)} = V^{(N)}((A^{-1})^{(N)})$ is the potential corresponding to $A^{(N)} X^{(N)}$, and $\widehat{V}^{(N)}$ satisfies Assumption \[ass:convexRMM\] with constants $c / \norm{A}$ and $C \norm{A^{-1}}$.
The fact that $\widehat{V}^{(N)}$ is the potential corresponding to $A^{(N)} X^{(N)}$ follows from change of variables. Now it is immediate that the expectation of $(A^{(N)} X^{(N)})_i$ is a scalar multiple of identity for each $i$. Next, by the chain rule $$D\widehat{V}^{(N)}(x) = ((A^{-1})^{(N)})^T DV^{(N)}((A^{-1})^{(N)}x) = ((A^{-1})^T)^{(N)} DV^{(N)}((A^{-1})^{(N)} x),$$ and from this it follows that $\{D\widehat{V}^{(N)}\}$ is asymptotically approximable by trace polynomials. Similarly, by the chain rule, $$H\widehat{V}^{(N)}(x) = ((A^{-1})^T)^{(N)} HV^{(N)}((A^{-1})^{(N)}x) (A^{-1})^{(N)}.$$ The maximum and minimum singular values of $(A^{-1})^{(N)}$ are the same as those of $A^{-1}$, which are $\norm{A^{-1}}$ and $1 / \norm{A}$ respectively. By a basic linear algebra argument, it follows that $c / \norm{A} \leq H\widehat{V}^{(N)} \leq C \norm{A^{-1}}$.
\[prop:convolutions\] Let $V_1^{(N)}$ and $V_2^{(N)}$ be two potentials $M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ satisfying Assumption \[ass:convexRMM\] with constants $c$ and $C$. Let $X^{(N)}$ and $Y^{(N)}$ be the corresponding random tuples of matrices. Then the law of $X^{(N)} + Y^{(N)}$ is given by another potential $\widehat{V}^{(N)}$ satisfying Assumption \[ass:convexRMM\] with constants $\sqrt{2} c$ and $\sqrt{2} C$. Moreover, the free Gibbs state corresponding to $\{\widehat{V}^{(N)}\}$ is the free convolution of those corresponding to $\{V_1^{(N)}\}$ and $\{V_2^{(N)}\}$.
Let $V^{(N)}(x,y) = V_1^{(N)}(x) + V_2^{(N)}(y)$, which satisfies Assumption \[ass:convexRMM\] (with the same constants) by Proposition \[prop:independentjoins\]. Now let $A$ be the $2m \times 2m$ matrix $$A = \begin{pmatrix} I & I \\ -I & I \end{pmatrix}.$$ Since $A / \sqrt{2}$ is an isometry, we have $\norm{A} = \sqrt{2}$ and $\norm{A^{-1}} = 1/\sqrt{2}$. Therefore, by Proposition \[prop:lineartransformations\], the law of $(X^{(N)} + Y^{(N)}, -X^{(N)} + Y^{(N)})$ is given by a potential satisfying Assumption \[ass:convexRMM\] with constants $\sqrt{2} c$ and $\sqrt{2} C$. Then by Proposition \[prop:marginals\], the law of $X^{(N)} + Y^{(N)}$ is given by such a potential with the same constants $\sqrt{2} c$ and $\sqrt{2} C$.
We showed in Proposition \[prop:independentjoins\] that the large $N$ limit of the law of $(X^{(N)},Y^{(N)})$ given a freely independent join of the corresponding marginals. Hence, the large $N$ limit of the law of $X^{(N)} + Y^{(N)}$ is given by the free convolution.
As a consequence, we have additivity of entropy under conditioning.
Let $V^{(N)}(x,y)$ be a potential satisfying Assumption \[ass:convexRMM\] as in the setup of Theorem \[thm:convergenceofentropy\]. Let $(X,Y)$ be a tuple of non-commutative random variables distributed according to the limiting free Gibbs law associated to $V^{(N)}$. Then $$\chi^*(X,Y) = \chi^*(X | Y) + \chi^*(Y).$$
From standard classical results, we have $$h(X^{(N)}, Y^{(N)}) = h(X^{(N)} | Y^{(N)}) + h(Y^{(N)}).$$ Dividing by $N^2$ and adding $\frac{1}{2} (m+n) \log N$ to both sides, we obtain the normalized version $$h^{(N)}(X^{(N)}, Y^{(N)}) = h^{(N)}(X^{(N)} | Y^{(N)}) + h^{(N)}(Y^{(N)}).$$ By the previous theorem, we obtain the desired relation for $\chi^*$ in the limit as $N \to \infty$. More precisely, we apply the theorem as stated to $h^{(N)}(X^{(N)} | Y^{(N)})$. Meanwhile, for $h^{(N)}(X^{(N)},Y^{(N)})$ and $h^{(N)}(Y^{(N)})$ we apply the special case of the theorem where we condition on $0$ variables.
Entropy and Fisher Information Relative to Gaussian
---------------------------------------------------
As background for our discussion of the entropy-cost inequality in §\[subsec:entropycost\], we review the entropy of one probability measure relative to another. If $\nu$ is a measure on ${\mathbb{R}}^m$, then the *entropy of $\mu$ relative to $\nu$* is $$h(\mu | \nu) := -\int \rho \log \rho\,d\mu, \text{ where } \rho = \frac{d\mu}{d\nu}$$ whenever the integral is well-defined. The standard entropy $h(\mu) = -\int \rho \log \rho$ corresponds to the choice of Lebesgue measure for $\nu$.
The reader should be careful to distinguish between the relative entropy $h(\mu | \nu)$ and the conditional entropy $h(X | Y)$. The first changes the ambient measure while the second describes conditioning on $Y$.
If $\mu$ and $\nu$ are both probability measures, then $h(\mu | \nu) \leq 0$. For this reason, many authors choose to change the sign. We will keep the sign convention given above to be consistent with our convention for $h(\mu)$ relative to Lebesgue measure, but we will write absolute value signs around relative entropy when it is natural to use the positive version.
For probability measures on $M_N({\mathbb{C}})_{sa}^m$, we may study entropy relative to the Gaussian measure $\sigma_{m,t}^{(N)}$ on $M_N({\mathbb{C}})_{sa}^m$. A direct computation shows that if $X \sim \mu$ is a random variable in $M_N({\mathbb{C}})_{sa}^m$, then we have $$h(\mu | \sigma_{m,t}^{(N)}) = h(\mu) - \frac{N^2}{2} E \norm{X}_2^2 + \frac{mN^2}{2} \log \frac{N}{2\pi} .$$ We denote the normalized version by $$h_g^{(N)}(\mu) := h_g^{(N)}(X) := \frac{1}{N^2} h(\mu | \sigma_{m,t}^{(N)}) = h^{(N)}(X) - \frac{1}{2} E \norm{X}_2^2 - \frac{m}{2} \log 2\pi.$$ Similarly, if $\mu$ is a measure on $M_N({\mathbb{C}})_{sa}^{m+n}$ which absolutely continuous with respect to Lebesgue measure and $(X,Y)$ is the corresponding random variable, we define $$h_g^{(N)}(X | Y) = h^{(N)}(X | Y) - \frac{1}{2} E \norm{X}_2^2 - \frac{m}{2} \log 2\pi,$$ which is equivalent to $$h_g^{(N)}(X | Y) = \int h_g^{(N)}(\mu_{X|Y=y})\,d\mu_Y(y),$$ where $\mu_{X|Y=y}$ is the conditional distribution of $X$ given $Y = y$, and $\mu_Y$ is the marginal law of $Y$. Similarly, if $(X,Y)$ is an $(m+n)$-tuple of non-commutative random variables, we define the free entropy $\chi^*$ relative to Gaussian by $$\chi_g^*(X | Y) = \chi^*(X | Y) - \frac{1}{2} \norm{X}_2^2 - \frac{m}{2} \log 2\pi.$$
We define the *normalized conditional Fisher information relative to Gaussian* by $$\label{eq:definenormalizedFisher}
\mathcal{I}_g^{(N)}(X | Y) = \mathcal{I}^{(N)}(X | Y) - 2m + E \norm{X}_2^2.$$ Note that if this Fisher information is finite and if $\xi$ is the normalized score function for $X$ given $Y$ as in §\[subsec:matrixentropy\], then $$\mathcal{I}_g^{(N)}(X | Y) = E \norm{\xi - X}_2^2$$ because $$\begin{aligned}
E \norm{\xi - X}_2^2 &= E \norm{\xi}_2^2 - 2 E \ip{\xi,X}_2 + E \norm{X}_2^2 \\
&= \mathcal{I}^{(N)}(X | Y) - 2m + E \norm{X}_2^2,\end{aligned}$$ where we have evaluated the middle term on the right hand side using integration by parts. Similarly, for an $(m+n)$-tuple $(X,Y)$ of non-commutative random variables, we define $$\label{eq:definenormalizedFisher2}
\Phi_g^*(X | Y) = \Phi^*(X | Y) - 2m + \norm{X}_2^2 = \norm{\xi - X}_2^2,$$ where the second equality holds provided that $\Phi^*$ is finite and $\xi$ is the free score function. We have the following version of and Lemma \[lem:Fisherestimates\] for entropy and Fisher information relative to Gaussian.
\[lem:GaussianFisherinfo\] Let $(X,Y)$ be a random variable in $M_N({\mathbb{C}})_{sa}^{m+n}$ with a density and with finite variance and let $S$ be an independent GUE $m$-tuple. Then $$|h_g^{(N)}(X | Y)| = \frac{1}{2} \int_0^\infty \mathcal{I}_g^{(N)}(e^{-t/2}X + (1 - e^{-t})^{1/2}S | Y)\,dt.$$ Similarly, suppose that $(X,Y)$ is an $(m+n)$-tuple of non-commutative random variables and let $S$ be a freely independent free semicircular $m$-tuple. Then $$|\chi_g^*(X | Y)| = \frac{1}{2} \int_0^\infty \Phi_g^*(e^{-t/2} X + (1 - e^{-t})^{1/2} S | Y)\,dt.$$
The first formula follows from [@OV2000 §4, Lemma 1] after renormalization. However, we will give an argument by a change of variables in that will apply to both $h_g^{(N)}$ and $\chi_g^*$. Note that by $$\begin{aligned}
h_g^{(N)}(X|Y) &= h^{(N)}(X|Y) - \frac{1}{2} E \norm{X}_2^2 - \frac{m}{2} \log 2\pi \\
&= \frac{1}{2} \int_0^\infty \left( \frac{m}{1 + t} - \mathcal{I}^{(N)}(X + t^{1/2}S | Y) \right)\,dt + \frac{m}{2} - \frac{1}{2} E \norm{X}_2^2,\end{aligned}$$ and in particular, we know that the integral is well-defined in $[-\infty,+\infty)$. Now we do a change of variables in the integral $t = e^u - 1$, $dt = e^u \,du$ and obtain $$\begin{aligned}
\frac{1}{2} \int_0^\infty \left( \frac{m}{1 + t} - \mathcal{I}^{(N)}(X + t^{1/2}S | Y) \right)\,dt &= \frac{1}{2} \int_0^\infty \left( \frac{m}{e^u} - \mathcal{I}^{(N)}(X + (e^u - 1)^{1/2}S| Y) \right) e^u\,du \\
&= \frac{1}{2} \int_0^\infty \left( m - \mathcal{I}^{(N)}(e^{-u/2} X + (1 - e^{-u})^{1/2}S| Y) \right) \,du,\end{aligned}$$ where we have applied the scaling relation Lemma \[lem:xiscaling\] for Fisher information. On the other hand, $$\begin{aligned}
\frac{m}{2} - \frac{1}{2} E \norm{X}_2^2 &= \frac{1}{2} \left( E \norm{S}_2^2 - E \norm{X}_2^2 \right) \\
&= \frac{1}{2} \int_0^\infty e^{-u} \left( E \norm{S}_2^2 - E \norm{X}_2^2 \right)\,du \\
&= \frac{1}{2} \int_0^\infty \left( m - E \norm{e^{-u/2} X + (1 - e^{-u})^{1/2} S}_2^2 \right)\,du.\end{aligned}$$ Therefore, altogether $$\begin{aligned}
h_g^{(N)}(X|Y) &= \frac{1}{2} \int_0^\infty \left( 2m - \mathcal{I}^{(N)}(e^{-u/2} X + (1 - e^{-u})^{1/2}S| Y) - E \norm{e^{-u/2} X + (1 - e^{-u})^{1/2} S}_2^2 \right) \,du \\
&= -\frac{1}{2} \int_0^\infty \mathcal{I}_g^{(N)}(e^{-u/2}X + (1 - e^{-u})^{1/2}S | Y)\,du,\end{aligned}$$ which is the desired formula. The statement for $\chi^*$ can be proved by exactly the same computation, since the definition of $\chi^*$ in is completely analogous to .
Furthermore, the log-Sobolev inequality for the Gaussian measure has the following interpretation for entropy and Fisher’s information. This in fact generalizes to entropy and Fisher’s information relative to any measure $\nu$ satisfying LSI, see [@OV2000 Definition 1], but we only use the case where $\nu$ is Gaussian and $\mu$ is sufficiently regular.
\[lem:informationLSI\] Let $X$ be a random variable in $M_N({\mathbb{C}})_{sa}^m$ that has a $C^1$ density with respect to Lebesgue measure. Then $$|h_g^{(N)}(X | Y)| \leq \frac{1}{2} \mathcal{I}_g^{(N)}(X | Y).$$
First, it suffices to check the non-conditional version $h_g^{(N)}(X) \leq \frac{1}{2} \mathcal{I}_g^{(N)}(X)$. Indeed, in the conditional case, the left hand side is $\int h_g^{(N)}(\mu_{X|Y=y})\,d\mu_Y(y)$ and the right hand side is $\int \mathcal{I}_g^{(N)}(\mu_{X|Y=y})\,d\mu_Y(y)$, and solving the non-conditional case would allow us to compare the integrands pointwise.
Now suppose that $X$ has density $\rho$ with respect to Lebesgue measure and let $\tilde{\rho}$ be the density with respect to Gaussian, so that $$\rho(x) = \tilde{\rho}(x) \frac{1}{(2\pi / N)^{N^2/2}} e^{-N^2 \norm{x}_2^2 / 2}\,dx.$$ By Corollary \[cor:matrixLSI\], the measure $\sigma_{m,t}^{(N)}$ satisfies the normalized log-Sobolev inequality with $c = 1$, so that $$\frac{1}{N^2} \int f^2 \log \frac{f^2}{\int f^2 \,d\sigma_{m,t}^{(N)}} \,d\sigma_{m,t}^{(N)} \leq \frac{2}{N^4} \int \norm{Df}_2^2\,d\sigma_{m,t}^{(N)}.$$ Let $f = \tilde{\rho}^{1/2}$. Then $\int f^2\,d\sigma_{m,t}^{(N)}$ reduces to $1$, so the right hand side is $|h_g^{(N)}(X)|$. On the other hand, letting $V(x) = -(1/N^2) \log \rho$, we get $$f(x) = \tilde{\rho}(x)^{1/2} = (\text{constant}) e^{-N^2V(x)/2 - N^2 \norm{x}_2^2 / 4},$$ and hence on the support of $f$, we have $$Df(x) = -\frac{N^2}{2} (DV(x) - x) \tilde{\rho}(x)^{1/2}.$$ Thus, $$\frac{2}{N^4} \int \norm{Df}_2^2\,d\sigma_{m,t}^{(N)} = \frac{1}{2} \int \norm{DV(x) - x}_2^2\,d\mu(x) = \frac{1}{2} \mathcal{I}_g^{(N)}(X).$$ Hence, the log-Sobolev inequality implies the desired inequality.
Conditional Transport and the Entropy-Cost Inequality {#subsec:entropycost}
-----------------------------------------------------
Now we will show that the transport maps constructed in §\[subsec:largeNtransport\] satisfy the Talagrand entropy-cost inequality. It was shown in [@OV2000 Theorem 1] that if a measure $\nu$ satisfies the log-Sobolev inequality with some constant $c$ (and some regularity conditions), then it satisfies the Talagrand inequality $$W_2(\mu,\nu)^2 \leq \frac{2 h(\mu|\nu)}{\rho} \text{ for all } \mu,$$ where $W_2$ is the $L^2$-Wasserstein distance, which is equivalent to the infimum of $\norm{X - Y}_{L^2}$ over all coupled random variables $X$ and $Y$ with $X \sim \mu$ and $Y \sim \nu$.
Adapting Otto and Villani’s argument, we will show that the transport maps constructed in §\[subsec:largeNtransport\] witness the (conditional) entropy-cost inequality relative to the GUE law for the $N \times N$ matrix models and the corresponding free entropy-cost inequality for the non-commutative random variables. This is claim (5) below, while the other claims in Theorem \[thm:transport3\] summarize the results of our earlier construction.
We remark that the free Talagrand inequality for self-adjoint tuples was studied in greater generality in [@HU2006] and [@Dabrowski2010 §3.3]. Although we restricted ourselves to the case where the target measure is Gaussian/semicircular, our goal in this paper was not merely to estimate the Wasserstein distance using some coupling, but rather to exhibit a coupling that arises from a transport map, and to show Lipschitzness of this transport map.
\[thm:transport3\] As in Theorem \[thm:transport1\], let $V^{(N)}(x,y)$ be a potential on $M_N({\mathbb{C}})_{sa}^{m+n}$ satisfying Assumption \[ass:convexRMM\] for some $0 < c \leq C$, and let $\mu^{(N)}$ and $(X^{(N)}, Y^{(N)})$ be the corresponding probability measures and random variables. Let $S^{(N)}$ be an independent GUE $m$-tuple. Let $(X,Y)$ be a tuple of non-commutative random variables given by the limiting free Gibbs law $\lambda$ and let $S$ be a freely independent free semicircular $m$-tuple. Let $\pi_1(x,y) = x$ and $\pi_2(x,y) = y$. Then there exist functions $F^{(N)}$, $G^{(N)}: M_N({\mathbb{C}})_{sa}^{m+n} \to M_N({\mathbb{C}})_{sa}^m$ and $F, G \in (\overline{\operatorname{TrP}}_{m+n}^1)^m$ such that
1. We have $(F^{(N)}(X^{(N)},Y^{(N)}),Y^{(N)}) \sim (S^{(N)},Y^{(N)})$ and $(G^{(N)}(S^{(N)},Y^{(N)}),Y^{(N)}) \sim (X^{(N)},Y^{(N)})$ in law, and $(F(X,Y),Y) \sim (S,Y)$ and $(G(S,Y),Y) \sim (X,Y)$ in non-commutative law.
2. $(F^{(N)},\pi_2) \circ (G^{(N)},\pi_2) = \operatorname{id}= (G^{(N)},\pi_2) \circ (F^{(N)},\pi_2)$ and the same holds for $F$ and $G$.
3. $F^{(N)} \rightsquigarrow F$ and $G^{(N)} \rightsquigarrow G$.
4. We have $\norm{F^{(N)} - \pi_1}_{\operatorname{Lip}}$ and $\norm{G^{(N)} - \pi_1}_{\operatorname{Lip}} \leq (\max(C,1/c)^3 - 1) \max(C,1/c)^{1/2}$, and the same holds for $F$ and $G$.
5. We have $$\norm{F^{(N)}(X^{(N)}, Y^{(N)}) - X^{(N)}}_{L^2}^2 = \norm{G^{(N)}(S^{(N)}, Y^{(N)}) - S^{(N)}}_{L^2}^2 \leq 2 h_g^{(N)}(X^{(N)} | Y^{(N)}).$$ and $$\norm{F(X,Y) - X}_2^2 = \norm{G(S,Y) - S}_2^2 \leq 2 \chi_g^*(X | Y).$$
Let $\tilde{F}_{s,t}^{(N)}$ and $\tilde{F}_{s,t}$ be as in Theorems \[thm:transport1\] and \[thm:transport2\]. Then let $$\begin{aligned}
F^{(N)} &= \tilde{F}_{\infty,0}^{(N)} \\
G^{(N)} &= \tilde{F}_{0,\infty}^{(N)} \\
F &= \tilde{F}_{\infty,0} \\
G &= \tilde{F}_{0,\infty}.\end{aligned}$$ The only property that was not shown in the earlier theorems is (5). First, note that as a consequence of (1), $$\norm{F^{(N)}(X^{(N)}, Y^{(N)}) - X^{(N)}}_{L^2} = \norm{S^{(N)} - G^{(N)}(S^{(N)}, Y^{(N)})}_{L^2}.$$ The rest of the proof of (5) proceeds as in [@OV2000 §4]. As in §\[subsec:largeNtransport\], let $\tilde{V}_t^{(N)}$ denote the potential corresponding to $(\tilde{X}_t^{(N)},Y^{(N)}) = (e^{-t/2}X^{(N)} + (1 - e^{-t})^{1/2} S^{(N)}, Y^{(N)})$ and recall that $$\partial_s \tilde{F}_{s,0}^{(N)}(x,y) = \frac{1}{2} \left( D_x \tilde{V}_s^{(N)}(\tilde{F}_{s,t}^{(N)}(x,y),y) - \tilde{F}_{s,t}^{(N)}(x,y) \right),$$ and hence $$\norm*{\tilde{F}_{\infty,0}^{(N)}(x,y) - x}_2 \leq \frac{1}{2} \int_0^\infty \norm*{D_x \tilde{V}_s(\tilde{F}_{s,t}^{(N)}(x,y),y) - \tilde{F}_{s,t}^{(N)}(x,y)}_2\,ds.$$ Then we apply Minkowski’s inequality with respect to integration $d\mu^{(N)}(x,y)$ to obtain $$\left( \int \norm*{\tilde{F}_{\infty,0}^{(N)}(x,y) - x}_2^2 \,d\mu^{(N)}(x,y) \right)^{1/2} \leq \frac{1}{2} \int_0^\infty \left( \int \norm*{D_x \tilde{V}_s(\tilde{F}_{s,t}^{(N)}(x,y),y) - \tilde{F}_{s,t}^{(N)}(x,y)}_2^2 \,d\mu^{(N)}(x,y) \right)^{1/2} \,ds$$ which can be rewritten as $$\begin{aligned}
\norm*{F^{(N)}(X^{(N)}, Y^{(N)}) - X^{(N)}}_{L^2} &\leq \frac{1}{2} \int_0^\infty \norm*{D_x \tilde{V}_s(\tilde{F}_{s,t}^{(N)}(X^{(N)},Y^{(N)}),Y^{(N)}) - \tilde{F}_{s,t}^{(N)}(X^{(N)},Y^{(N)})}_{L^2}\,ds \\
&= \frac{1}{2} \int_0^\infty \norm*{D_x \tilde{V}_s^{(N)}(\tilde{X}_s^{(N)},Y^{(N)}) - \tilde{X}_s^{(N)}}_{L^2}\,ds \\
&= \frac{1}{2} \int_0^\infty \mathcal{I}_g^{(N)}(\tilde{X}_s^{(N)} | Y^{(N)})^{1/2}\,ds,\end{aligned}$$ where we have applied the fact that $(\tilde{F}_{s,0}^{(N)}(X^{(N)},Y^{(N)}), Y^{(N)}) \sim (\tilde{X}_s^{(N)}, Y^{(N)})$. It follows from Lemma \[lem:GaussianFisherinfo\] and a change of variables that $$\label{eq:integrateentropy}
|h_g^{(N)}(\tilde{X}_t^{(N)} | Y^{(N)})| = \int_t^\infty \mathcal{I}_g^{(N)}(\tilde{X}_s^{(N)} | Y^{(N)})\,ds.$$ It is easy to see that $s \mapsto \mathcal{I}_g^{(N)}(\tilde{X}_s^{(N)} | Y^{(N)})$ is bounded on compact sets because of Lemma \[lem:Fisherestimates\] and . Therefore, we have for almost every $t$, $$\frac{d}{dt} |h_g^{(N)}(\tilde{X}_t^{(N)} | Y^{(N)})| = \frac{1}{2} \mathcal{I}_g^{(N)}(\tilde{X}_t^{(N)} | Y^{(N)}).$$ Hence, for almost every $t$, $$\begin{aligned}
\frac{d}{dt} |h_g^{(N)}(\tilde{X}_t^{(N)} | Y^{(N)})|^{1/2} &= \frac{1}{4} \mathcal{I}_g^{(N)}(\tilde{X}_t^{(N)} | Y^{(N)}) |h_g^{(N)}(\tilde{X}_t^{(N)} | Y^{(N)})|^{-1/2} \\
&\geq \frac{1}{2 \sqrt{2}} \mathcal{I}_g^{(N)}(\tilde{X}_t^{(N)} | Y_t^{(N)})^{1/2}, \end{aligned}$$ where the last line follows from Lemma \[lem:informationLSI\]. Therefore, $$\begin{aligned}
\norm*{F^{(N)}(X^{(N)}, Y^{(N)}) - X^{(N)}}_{L^2} &\leq \sqrt{2} \int_0^\infty \frac{d}{dt} |h_g^{(N)}(\tilde{X}_t^{(N)} | Y^{(N)})|^{1/2} \,dt \\
&= \sqrt{2} |h_g^{(N)}(X^{(N)} | Y^{(N)})|^{1/2},\end{aligned}$$ where we have employed the fact that $\lim_{t \to \infty} |h_g^{(N)}(\tilde{X}_t^{(N)} | Y^{(N)})| = 0$ by . This establishes the first claim of (5).
The second claim of (5) follows by taking the large $N$ limit using Corollary \[cor:convergenceofexpectation\] and Theorem \[thm:convergenceofentropy\]. More precisely, for the left hand side, we take the limit using Corollary \[cor:convergenceofexpectation\]. Meanwhile, for the right hand side, note that $h_g^{(N)}(X^{(N)} | Y^{(N)}) \to \chi_g^*(X | Y)$ because $h^{(N)}(X^{(N)} | Y^{(N)}) \to \chi^*(X | Y)$ and $E \norm{X^{(N)}}_2^2 \to \norm{X}_2^2$ by Corollary \[cor:convergenceofexpectation\].
Construction of Triangular Transport
------------------------------------
Finally, by iterating Theorem \[thm:transport3\], we obtain the following result concerning “lower-triangular transport.” This is analogous to the classical result [@BKM2005 Corollary 3.10]. Of course, the challenge in our situation was to understand the large $N$ behavior of the transport maps in a dimension-independent way. Unfortunately, the transport constructed here is not optimal among triangular mappings, since indeed Otto and Villani’s construction does not produce the optimal transport map.
\[thm:transport4\] Let $V^{(N)}: M_N({\mathbb{C}})_{sa}^m \to {\mathbb{R}}$ be a potential satisfying Assumption \[ass:convexRMM\]. Let $\mu^{(N)}$ and $X^{(N)}$ be the corresponding law and random variable. Let $\lambda$ be the limiting free Gibbs law, and let $X \sim \mu$ be an $m$-tuple of non-commutative random variables. Let $S^{(N)}$ be an independent GUE $m$-tuple and let $S$ be a freely independent free semicircular family. Then there exist functions $\Phi^{(N)}$, $\Psi^{(N)}: M_N({\mathbb{C}})_{sa}^m \to M_N({\mathbb{C}})_{sa}^m$ and $\Phi, \Psi \in (\overline{\operatorname{TrP}}_m^1)^m$ such that
1. $\Phi^{(N)}(X^{(N)}) \sim S^{(N)}$ and $\Psi^{(N)}(S^{(N)}) \sim X^{(N)}$ in law, and similarly, $\Phi(X) \sim S$ and $\Psi(S) \sim X$ in non-commutative law.
2. $\Phi^{(N)}$ and $\Psi^{(N)}$ are inverse functions of each other, and the same holds for $\Phi$ and $\Psi$.
3. $\Phi^{(N)} \rightsquigarrow \Phi$ and $\Psi^{(N)} \rightsquigarrow \Psi$.
4. $\Phi^{(N)}$ is upper triangular in the sense that $$\Phi^{(N)}(x_1,\dots,x_m) = (\Phi_1^{(N)}(x_1), \Phi_2^{(N)}(x_1,x_2), \dots, \Phi_m^{(N)}(x_1,\dots,x_m))$$ and the same holds for $\Psi^{(N)}$, $\Phi$, and $\Psi$. In particular, the isomorphism $\mathrm{W}^*(X) \to \mathrm{W}^*(S)$ induced by $\Phi$ maps $\mathrm{W}^*(X_1,\dots,X_k)$ onto $\mathrm{W}^*(S_1,\dots,S_k)$ for each $k = 1$, …, $m$.
5. We have $\norm{\Phi^{(N)} - \operatorname{id}}_{\operatorname{Lip}} \leq m^{1/2} (\max(C,1/c)^3 - 1) \max(C,1/c)^{1/2}$ and $\norm{\Psi^{(N)} - \operatorname{id}}_{\operatorname{Lip}}$ is bounded by some constant $L(c,C,m)$ which goes to zero as $c, C \to 1$.
6. We have $$\norm{\Phi^{(N)}(X^{(N)}) - X^{(N)}}_{L^2}^2 = \norm{\Psi^{(N)}(S^{(N)}) - S^{(N)}}_{L^2}^2 \leq 2 h_g^{(N)}(X^{(N)}).$$ and $$\norm{\Phi(X) - X}_2^2 = \norm{\Psi(S) - S}_2^2 \leq 2 \chi_g^*(X).$$
7. We have $$\begin{aligned}
\norm{\Phi_j(X_1,\dots,X_j) - (X_j - \tau(X_j))}_\infty &= \norm{(\Psi_j(S_1,\dots,S_j) - \tau(\Psi_j(S_1,\dots,S_j))) - S_j}_\infty \\
&\leq (\max(C,1/c)^3 - 1) \max(C,1/c) \Theta,
\end{aligned}$$ where $\Theta$ is the universal constant from Proposition \[prop:operatornormestimate\].
First, by Proposition \[prop:marginals\], the marginal law of $(X_1^{(N)}, \dots, X_j^{(N)})$ is given by a convex potential satisfying the same assumptions.
For each $j$, we apply Theorem \[thm:transport3\] with $X_j^{(N)}$ as the first variable and $(X_1^{(N)}, \dots, X_{j-1}^{(N)})$ as the second variable. We thus obtain maps $\Phi_j^{(N)}: M_N({\mathbb{C}})_{sa}^j \to M_N({\mathbb{C}})_{sa}$ such that $$(X_1^{(N)}, \dots, X_{j-1}^{(N)}, \Phi_j^{(N)}(X_1^{(N)}, \dots, X_j^{(N)})) \sim (X_1^{(N)}, \dots, X_{j-1}^{(N)}, S_j^{(N)}).$$ Let $$\Phi^{(N)}(x_1,\dots,x_m) = (\Phi_1^{(N)}(x_1), \Phi_2^{(N)}(x_1,x_2), \dots \Phi_m^{(N)}(x_1,\dots,x_m)).$$ Let $Y^{(N)} = \Phi^{(N)}(X^{(N)})$. Then we can check by backwards induction on $j$ that $$(X_1^{(N)}, \dots, X_j^{(N)}, Y_{j+1}^{(N)}, \dots, Y_m^{(N)}) \sim (X_1^{(N)}, \dots, X_j^{(N)}, S_{j+1}^{(N)}, \dots, S_m^{(N)}).$$ Indeed, the base case $j = m$ is trivial. For the induction step, suppose the claim holds for $j$. Since $Y_{j+1}^{(N)}$ is a function of $X_1^{(N)}$, …, $X_j^{(N)}$, then the induction hypothesis implies that $$\begin{aligned}
(X_1^{(N)}, \dots, X_{j-1}^{(N)}, Y_j^{(N)}, Y_{j+1}^{(N)}, \dots, Y_m^{(N)}) &\sim (X_1^{(N)}, \dots, X_{j-1}^{(N)}, Y_j^{(N)}, S_{j+1}^{(N)}, \dots, S_m^{(N)}) \\
&\sim(X_1^{(N)}, \dots, X_{j-1}^{(N)}, S_j^{(N)}, S_{j+1}^{(N)}, \dots, S_m^{(N)}),\end{aligned}$$ where the last line follows because $(X_1^{(N)}, \dots, X_{j-1}^{(N)}, Y_j^{(N)}) \sim (X_1^{(N)}, \dots, X_{j-1}^{(N)}, S_j^{(N)})$ and because $S_{j+1}^{(N)}$, …, $S_m^{(N)}$ are independent of the other variables. By Theorem \[thm:transport3\], $\Phi_j^{(N)}$ is asymptotic to some $\Phi_j \in (\overline{\operatorname{TrP}}_j^1)_{sa}$, and the objects $\Phi$, $X$, and $S$ satisfy the analogous transport relations in the non-commutative setting. Now because each $\Phi_j^{(N)} - \pi_{x_j}$ is $(\max(C,1/c)^3 - 1) \max(C,1/c)^{1/2}$-Lipschitz, we see that $\Phi^{(N)} - \operatorname{id}$ is $m^{1/2} (\max(C,1/c)^3 - 1) \max(C,1/c)^{1/2}$-Lipschitz.
By Theorem \[thm:transport3\], there is a map $G_j^{(N)}: M_N({\mathbb{C}})_{sa}^{m-j+1} \to M_N({\mathbb{C}})_{sa}$ such that $(x_1,\dots,x_{j-1},G_j^{(N)}(x_1,\dots,x_j))$ is the inverse of $(x_1,\dots,x_{j-1},\Phi_j(x_1,\dots,x_j))$. Define $\Psi_j^{(N)}$ by induction by $$\Psi_j^{(N)}(x_j,\dots,x_m) = G_j^{(N)}(\Psi_1(x_1),\dots,\Psi_{j-1}(x_1,\dots,x_{j-1}),x_j).$$ Then $\Psi^{(N)} = (\Psi_1^{(N)},\dots,\Psi_m^{(N)})$ is the inverse of $\Phi^{(N)}$. Since $G_j^{(N)} - \operatorname{id}$ is $(\max(C,1/c)^3 - 1) \max(C,1/c)^{1/2}$-Lipschitz, we can show by induction that $\norm{\Psi_j^{(N)}}_{\operatorname{Lip}}$ is bounded by a constant depending only on $c$, $C$, and $m - j$, and which goes to zero as $c, C \to 1$. Moreover, by Lemma \[lem:composition\], $\Psi^{(N)}$ is asymptotic to some $\Psi \in (\overline{\operatorname{TrP}}_m^1)_{sa}^m$.
This concludes the verification of (1) - (5). Now to prove (6), we apply Theorem \[thm:transport3\] (5) and get $$\begin{aligned}
\norm{\Phi^{(N)}(X^{(N)}) - X^{(N)}}_{L^2}^2 &= \sum_{j=1}^m \norm*{\Phi_j^{(N)}(X_j^{(N)},\dots,X_m^{(N)}) - X_j^{(N)}}_{L^2}^2 \\
&\leq \sum_{j=1}^m 2 h_g^{(N)}(X_j^{(N)} | X_1^{(N)}, \dots, X_{j-1}^{(N)}) \\
&= 2 \sum_{j=1}^m \left( h^{(N)}(X_j^{(N)} | X_1^{(N)}, \dots, X_{j-1}^{(N)}) - \frac{1}{2} E \norm{X_j^{(N)}}_2^2 - \frac{1}{2} \log 2 \pi \right) \\
&= 2 \left( h^{(N)}(X^{(N)}) - \frac{1}{2} E \norm{X^{(N)}}_2^2 - \frac{m}{2} \log 2 \pi \right) \\
&= 2 h_g^{(N)}(X^{(N)}),\end{aligned}$$ where we have applied the definition of $h_g^{(N)}$ and the classical fact that $h^{(N)}$ is additive under conditioning. As before, because $\Phi^{(N)}(X^{(N)}) \sim S^{(N)}$, we see that $\norm{S^{(N)} - \Psi^{(N)}(S^{(N)})}_{L^2} = \norm{\Phi^{(N)}(X^{(N)}) - X^{(N)}}_{L^2}$. Finally, the second claim of (6) regarding the free case follows by taking the limit as $N \to \infty$.
Finally, to prove (7), recall that the map $\Phi_j$ is a special case of the map $\tilde{F}_{0,\infty}$ in Theorem \[thm:transport2\]. Thus, by applying Theorem \[thm:transport2\] (4) in the case where $s = \infty$ and $s' = t = 0$, we obtain $\norm{\Phi_j(X_1,\dots,X_j) - (X_j - \tau(X_j))}_\infty \leq (\max(C,1/c)^3 - 1) \max(C,1/c) \Theta$. Moreover, the middle quantity in claim (7) equals the left hand side because $\Phi(X) \sim S$.
Funding {#funding .unnumbered}
-------
This work was supported by the National Science Foundation \[grant DMS-1762360\] and the UCLA graduate division.
Acknowledgements {#acknowledgements .unnumbered}
----------------
I thank Dima Shlyakhtenko, Ben Hayes, Brent Nelson, Yoann Dabrowski, Yoshimichi Ueda, and Todd Kemp for various useful conversations and comments on drafts of this paper. The results of this paper were motivated in part by discussions with Ben Hayes regarding free entropy and maximal amenable subalgebras. Dima Shlyakhtenko suggested the name “triangular transport.” The anonymous referees suggested several references and improvements to the exposition, including the connection with model theory.
[99]{}
, [*An Introduction to Random Matrices*]{}, Cambridge Studies in Advanced Mathematics, Cambridge University Press, 2009.
, [*An Introduction to [II1]{} Factors*]{}, 2017. preprint available at http://www.math.ucla.edu/ popa/Books/IIun-v10.pdf.
, [*Large deviation bounds for matrix [B]{}rownian motion*]{}, Invent. Math., 152 (2003), pp. 433–459.
, [*Large deviations via transference plans*]{}, Advances in Mathematics Research, 2 (2003), pp. 151-175.
, [*From [B]{}runn-[M]{}inkowski to [B]{}raskamp-[L]{}ieb and to logarithmic [S]{}obolev inequalities*]{}, Geom. Funct. Anal., 10 (2000), pp. 1028–1052.
, [*Triangular transformations of measures*]{}, Sb. Math., 196.3 (2005), pp. 309–335. , [*Maximal amenable von neumann subalgebras arising from maximal amenable subgroups*]{}, Geometric and Functional Analysis, 25 (2015), pp. 1688–1705.
, [*Amenable absorption in amalgamated free product von neumann algebras*]{}, Kyoto J. Math., 58 (2018), pp. 583–593.
, [*On extensions of the [B]{}runn-[M]{}inkowski and [P]{}r[é]{}kopa-[L]{}eindler theorems, including inequalities for log concave functions, and with an application to the diffusion equation*]{}, Journal of Functional Analysis, 22 (1976), pp. 366–389.
, [*Free convolution operators and free Hall transform*]{}, Journal of Functional Analysis, 265, pp. 2645–2708 (2013).
, [*Free Stein Irregularity and Dimension*]{}, preprint at arXiv:1902.02379 (2019).
, [*A [L]{}aplace principle for [H]{}ermitian [B]{}rownian motion and free entropy [I]{}: the convex functional case*]{}, preprint at arXiv:1604.06420 (2017).
height 2pt depth -1.6pt width 23pt, [*A non-commutative path space approach to stationary free stochastic differential equations*]{}, preprint at arXiv:1006.4351 (2010).
, [*Free transport for convex potentials*]{}, preprint at arXiv:1701.00132 (2017).
, [*The large-$N$ limit of the Segal-Bargmann transform on $U_N$*]{}, Journal of Functional Analysis 265 (2013), pp. 2585–2644.
, [*Model theory of operator algebras I: stability*]{}, Bulletin of the London Mathematical Society 45 (2013), pp. 825-838.
, [*Model theory of operator algebras II: model theory*]{}, Israel Journal of Mathematics 201.1 (2014), pp. 477 - 505.
, [*Model theory of operator algebras III: elementary equivalence and $\operatorname{II}_1$ factors*]{}, Bulletin of the London Mathematical Society 46 (2014), pp. 609-628.
, [*Real Analysis: Modern Techniques and their Applications*]{}, Pure and Applied Mathematics, John Wiley & Sons, New York, 2 ed., 1999.
, [*Logarithmic Sobolev Inequalities*]{}, American Journal of Mathematics, 97:4 (1975), pp. 1061–1083.
, [*Combinatorial aspects of random matrix models*]{}, Alea, 1 (2006), pp. 241–279.
, [*Free diffusions and matrix models with strictly convex interaction*]{}, Geometric and Functional Analysis, 18 (2009), pp. 1875–1916.
height 2pt depth -1.6pt width 23pt, [*Free monotone transport*]{}, Inventiones mathematicae, 197 (2014), pp. 613–661.
A. Guionnet and O. Zeitouni. [*Concentration of the spectral measure for large matrices*]{}, Electron. Commun. Probab. 5 (2000), pp. 199-136.
, [*A new application of random matrices: [$\text{Ext}(C_{\text{red}}^*(F_2))$]{} is not a group*]{}, Ann. Math., 162 (2005), pp. 711–775.
, [*Free transport for interpolated free group factors*]{}, Journal of Functional Analysis, 274 (2018), pp. 222–251.
, [*A random matrix approach to absorption theorems for free products*]{}, preprint at arXiv:1912.11569 (2019).
, [*Free transportation cost inequalities for noncommutative multi-variables*]{}, Infinite Dimensional Analysis, Quantum Probability and Related Topics, 09, pp. 391–412 (2006).
, [*An elementary approach to free entropy theory for convex potentials*]{}, preprint at arXiv:1805.08814 (2018), to appear in Analysis and PDE Journal.
, [*The large-$N$ limits of Brownian motions on $GL(N)$*]{}, Int. Math. Res. Not. IMRN, no. 13 (2016), pp. 4012–4057.
height 2pt depth -1.6pt width 23pt, [*Heat kernel empirical laws on $U(N)$ and $GL(N)$*]{}, J. Theoret. Probab. 30, no. 2, pp. 397–451 (2017)
, [*A heat semigroup approach to concentration on the sphere and on a compact [R]{}iemannian manifold*]{}, Geometric [&]{} Functional Analysis, 2 (1992), pp. 221–224.
, [*Uncountably many $\operatorname{II}_1$ factors*]{}, Ann. Math. (2), 90.2 (1969), pp. 372-377.
, [*The Random Matrix Theory of the Classical Compact Groups*]{}, Cambridge Tracts in Mathematics, Cambridge University Press, 2019.
, [*Spectral powers of random matrices*]{}, Electron. Comm. Probab., 18 (2013).
, [*Free monotone transport without a trace*]{}, Communications in Mathematical Physics, 334 (2015), pp. 1245–1298.
height 2pt depth -1.6pt width 23pt, [*Free transport for finite depth subfactor planar algebras*]{}, Journal of Functional Analysis, 268 (2015), pp. 2586 – 2620.
, [*Lectures on the Combinatorics of Free Probability*]{}, vol. 335 of London Mathematical Society Lecture Note Series, Cambridge University Press, 2006.
, [*Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality*]{}, Journal of Functional Analysis, 173 (2000), pp. 361 – 400.
, [*There is no separable universal $\operatorname{II}_1$ factor*]{}, Proc. Amer. Math. Soc., 132 (2004), pp. 487–490.
, [*Multiple operator integrals and higher operator derivatives*]{}, Journal of Functional Analysis, 233 (2006), pp. 515–544.
, [*Maximal injective subalgebras in factors associated with free groups*]{}, Advances in Mathematics, 50 (1983), pp. 27 – 48.
, [*Combinatorial properties of Brownian motion on the compact classical groups*]{}, Journal of Theoretical Probability, 10 (1997), pp. 659-679.
, [*Trace identities of full matrix algebras over a field of characteristic zero*]{}, Mathematics of the USSR-Izvestiya, 8 (1974), p. 727.
height 2pt depth -1.6pt width 23pt, [*Trace identities and central polynomials in the matrix superalgebras $M_{n,k}$*]{}, Mathematics of the USSR-Sbornik, 56 (1987), p. 187.
, [*Traces in two-dimensional QCD: the large-N limit*]{}, Traces in number theory, geometry and quantumfields, pp. 193–212, Vieweg Aspects of Mathematics, vol. 38. , [*An Introduction to Random Matrix Theory*]{}, vol. 132 of Graduate Texts in Mathematics, American Mathematical Society, Providence, 2012.
, [*The analogues of entropy and of [F]{}isher’s information measure in free probability, [I]{}*]{}, Comm. Math. Phys., 155 (1993), pp. 71–92.
height 2pt depth -1.6pt width 23pt, [*The analogues of entropy and of [F]{}isher’s information measure in free probability, [II]{}*]{}, Invent. Math., 118 (1994), pp. 411–440.
height 2pt depth -1.6pt width 23pt, [*The analogues of entropy and of [F]{}isher’s information measure in free probability, [III]{}: Absence of [C]{}artan subalgebras*]{}, Geometric and Functional Analysis, 6 (1996), pp. 172–199.
height 2pt depth -1.6pt width 23pt, [*The analogues of entropy and of [F]{}isher’s information measure in free probability [V]{}*]{}, Invent. Math., 132 (1998), pp. 189–227.
height 2pt depth -1.6pt width 23pt, [*A strengthened asymptotic freeness result for random matrices with applications to free entropy*]{}, International Mathematics Research Notices, 1998 (1998), pp. 41–63.
height 2pt depth -1.6pt width 23pt, [*The analogues of entropy and of Fisher’s information measure in free probability theory, [VI]{}: Liberation and mutual free information*]{}, Advances in Mathematics, 146 (1999), pp. 101–166.
height 2pt depth -1.6pt width 23pt, [*Free entropy*]{}, Bulletin of the London Mathematical Society, 34 (2002), pp. 257–278.
, [*Free Random Variables*]{}, vol. 1 of CRM Monograph Series, American Mathematical Society, 1992.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Requiring the existence of a unitary, causal and local UV-completion places a set of positivity bounds on the corresponding effective field theories (EFTs). We apply this positivity argument to the EFT of cosmological perturbations. Taking a $c_T=1$ beyond-Horndeski EFT as an illustrative example, we drive such bounds, which naturally incorporate the cosmological correction of order $H^2/\Lambda^2$, in which $\Lambda$ is the cutoff scale. Applications of our results are briefly discussed.'
author:
- 'Gen Ye$^{1}$[^1]'
- 'Yun-Song Piao$^{1,2}$[^2]'
title: Positivity in the effective field theory of cosmological perturbations
---
Introduction
============
The effective field theory (EFT) of cosmological perturbations is a powerful tool to study perturbations around a given cosmological background. Since the EFT of inflation [@Cheung:2007st], the relevant idea has been also applied to other cosmological fields, such as the EFT of dark energy [@Gubitosi:2012hu; @Bloomfield:2012ff; @Gleyzes:2013ooa; @Langlois:2017mxy] which captures the physics of scalar-tensor theories [@Horndeski:1974wa; @Deffayet:2011gz; @Kobayashi:2011nu; @Gleyzes:2014dya; @Langlois:2015cwa] (for a review, see [@Langlois:2018dxi; @Kobayashi:2019hrl]) at the cosmological scale. As another example, based on the EFT approach, it has been found that fully stable nonsingular cosmologies exist in theories beyond Horndeski [@Cai:2016thi; @Creminelli:2016zwa; @Cai:2017dyi; @Cai:2017tku; @Kolevatov:2017voe; @Mironov:2018oec; @Ye:2019frg; @Ye:2019sth].
The full UV-complete theory of gravity is yet unknown. Instead of starting top-down from a UV theory, usually one works directly with the EFT, which captures the physics of the underlying theory at certain scales. However, not all low-energy EFTs have a consistent UV theory, see e.g.[@Ooguri:2006in; @Obied:2018sgi]. Assuming the UV-complete theory is causal, unitary and local Lorentz-invariant, one can derive dispersion relations relating the IR limit of a scattering amplitude with its UV behavior, which place a set of bounds on the properties of the corresponding low-energy EFTs, the so-called positivity bound [@Adams:2006sv; @Nicolis:2009qm; @Bellazzini:2014waa; @Bellazzini:2016xrt; @deRham:2017avq; @deRham:2017zjm; @Chandrasekaran:2018qmx; @deRham:2018qqo; @Tokuda:2019nqb]. The positivity bounds have been applied to various EFTs [@Bellazzini:2015cra; @Cheung:2016yqr; @Cheung:2016wjt; @deRham:2017imi; @Bellazzini:2017fep; @deRham:2017xox; @Bellazzini:2019bzh; @Bellazzini:2019xts].
Recently, Ref.[@Melville:2019wyy] derived positivity bounds on a covariant shift-symmetric Horndeski theory (which might explain the current accelerated expansion), and paired these bounds with a cosmological parameter estimation analysis. It is natural to ask what will happen if we incorporate the cosmological background evolution. It is convenient for our purpose to work with the EFT of cosmological perturbations. However, the background evolution is itself the biggest obstruction in applying the positivity arguments because it breaks time-translation symmetry and makes Lorentz-invariant scattering ill-defined. In the limit of a Lorentz-invariant background, the positivity bounds saturated by the UV completions of single-field inflation have been investigated in [@Baumann:2015nta].
Theoretically, an EFT is obtained after one integrates out physics above the cutoff scale $\Lambda$. On the other hand, the curvature of a homogeneous FRW universe is proportional to $H^2, \dot{H}$, so one can use $H^2, \dot{H}$ as additional dimensionful parameters in constructing the EFT, e.g.[@Kaloper:2002uj]. In this paper, based on similar observation, we will regard the cosmological background as $\mathcal{O}(H^{2n}/\Lambda^{2n},\dot{H}^n/\Lambda^{2n})$ correction in the amplitude, as pointed out also in Ref.[@Baumann:2015nta], and derive the corresponding positivity bounds incorporating the cosmological background.
This paper is structured as follows. In section-\[sec:bound\], we briefly review the dispersion relation that relates the UV and IR in relativistic quantum field theories. The effective Goldstone Lagrangian is derived in section-\[subsec:goldstone lagrangian\]. Then in section-\[subsec:apply bound\], under some assumptions, we derive positivity bounds with leading cosmological correction. Applications of our bounds to the cosmolgical scenarios of interest are discussed in section-\[subsec:example\].
Positivity bounds {#sec:bound}
=================
We briefly review the positivity argument, see e.g.[@Adams:2006sv] for details. In a Lorentz-invariant UV theory with locality and causality, the $2\to2$ scattering amplitude is expected to be an analytic function of the Mandelstam variables $(s,t,u)$ with poles and branch cuts. Unitarity is encoded as the polynomial boundedness of the amplitude [@Froissart:1961ux; @Martin:1962rt].
![The analytic structure of ${\mathcal{A}}(s)$ in the forward limit and the integration contour $\mathcal{C}$. The branch cut starts from the biparticle production threshold and extends all the way to infinity.[]{data-label="analytic structure"}](contour.pdf){width="5in"}
Consider a massive scalar field with mass $m$. The Mandelstam variables are not independent $s+t+u=4m^2$. We denote the $2\to2$ amplitude by ${\mathcal{A}}(s,t)$. For fixed $t$, it is an analytic function ${\mathcal{A}}_t(s)$ of $s$. We assume $|t|<m^2$ so no poles are encountered when pushed to the forward limit $t\to0$. ${\mathcal{A}}_t(s)$ can be extended to the complex s-plane by crossing symmetry ${\mathcal{A}}_t(s)={\mathcal{A}}_t(4m^2-t-s)$ and analytic continuation ${\mathcal{A}}_t(s)={\mathcal{A}}_t^*(s^*)$. Consider the following Cauchy integral at fixed $t$ $$\label{Cauchy int}
\sum Res\left(\frac{{\mathcal{A}}_t(s)}{(s-M^2)^3}\right)=\frac{1}{2\pi i}\oint_\mathcal{C} \frac{{\mathcal{A}}_t(s)}{(s-M^2)^3} ds.$$ The analytic structure of $\lim_{t\to0}{\mathcal{A}}_t(s)$ and the integration contour $\mathcal{C}$ are depicted in Fig.\[analytic structure\]. By the Froissart-Martin bound $|{\mathcal{A}}_t(s)|\sim \mathcal{O}(s\ln^2s)$ as $s\to\infty$ [@Froissart:1961ux; @Martin:1962rt], only integration along the branch cuts is not equal to 0 when the contour is pushed to infinity. To extract the positivity bound, we now pass to the forward limit $t\to0$. Using the optical theorem $$Im[{\mathcal{A}}(s)]=\sqrt{s(s-4m^2)}\sigma^{2\to any}(s),$$ we can write the integration along the cuts as $$\label{branch int}
\frac{1}{\pi}\int_{-\infty}^{m^2}+\int_{4m^2}^{+\infty}\frac{Im[{\mathcal{A}}(s)]}{(s-M^2)^3}ds=\frac{1}{\pi}\int_{-\infty}^{m^2}+\int_{4m^2}^{+\infty}\frac{\sqrt{s(s-4m^2)}\sigma^{2\to any}(s)}{(s-M^2)^3}ds.$$ There are three poles included. When the EFT has only derivative interactions (the Goldstone action we are to consider later falls into this category), residues at $s=m^2$ and $s=3m^2$, associated with propagators in the exchange diagrams, are proportional to powers of $m^2$ and thus vanish when $m^2\to0$. Then the final bound in the massless limit is $$\label{the bound}
{\mathcal{A}}''(s=M^2)\simeq\frac{4}{\pi}\int_0^\infty \frac{\sigma(s)}{s^2}ds>0,$$ up to $\mathcal{O}(M^2/\Lambda^2)$ correction. Besides, combine the optical theorem with the partial wave expansion ${\mathcal{A}}(s,t)=16\pi\sqrt{s/(s-4m^2)}\sum_l
(2l+1)P_l(\cos\theta)a_l(s)$, one obtains another bound for $s\ge4m^2$ [@Martin:1965jj] (for recent exploitation of this bound, see e.g.[@deRham:2017avq]) $$\label{the t bound}
\frac{\partial^n}{\partial t^n}Im[{\mathcal{A}}(s+i\epsilon,0)]\Big|_{t=0}\ge0.$$ Apply this bound to Eq. and Eq., one reaches the positivity of $\partial_t^n{\mathcal{A}}''(s=M^2,t)\Big|_{t=0}$ with $\mathcal{O}(M^2/\Lambda^2)$ correction.
Positivity bounds in cosmology {#sec:bound with background}
==============================
We will apply the positivity argument to the EFT of cosmological perturbations. We want to focus on the Goldstone EFT, so we will work in the decoupling limit [@Cheung:2007st]. Such limit, however, becomes subtle in a $c_T\ne1$ theory [@Creminelli:2014wna]. It is also noticed that the EFTs of modified gravity at cosmological scales have been strictly constrained by GW170817 to $c_T=1$ [@Abbott:2018lct; @Creminelli:2017sry; @Langlois:2017dyl]. Also, as mentioned, the stable nonsingular cosmological models can be implemented only in theories beyond Horndeski. Thus we are well-motivated to consider a $c_T=1$ beyond-Horndeski EFT as an illustrative example.
The shift-symmetric $c_T=1$ beyond-Horndeski theory can be written as [@Gleyzes:2014dya; @Creminelli:2017sry] $$\label{L_bH}
L=M_p^2\Lambda^2\left[B(X)\frac{R}{\Lambda^2}+G_2(X)+G_{3}(X)\frac{\Box\phi}{M_p\Lambda^2}-\frac{4}{X}B_X\frac{\phi^\mu\phi^\nu\phi_{\mu\nu}\Box\phi-\phi^\mu\phi_{\mu\nu}\phi_\lambda\phi^{\lambda\nu}}{M_p^4\Lambda^6}\right],$$ where $M_p$ is the reduced Planck mass and $\Lambda$ is the EFT cutoff. All coefficients $B(X)$, $G_{2,3}(X)$ and fields are dedimensionalised by $\phi\to \phi/M_p, \ \partial\to
\partial/\Lambda$ and $X\equiv\phi^\mu\phi_\mu/(M_p^2\Lambda^2)$. Subscript $X$ denotes partial derivatives with respect to $X$, for instance $B_X\equiv \frac{\partial}{\partial X}B$. The positivity argument in section-\[sec:bound\] only holds for a massive scalar field. In (\[L\_bH\]), we can add a small but nonvanishing $\phi$-dependent potential, which provides a small effective Goldstone mass (well outside the regime of validity of the EFT), and work in the vanishing-mass limit.
In the unitary gauge ($\delta \phi=0$), the Lagrangian is equivalent to [@Creminelli:2017sry] $$\label{unitary L}
L=M_p^2\Lambda^2\left[G_2(X)+Q(X)\frac{K}{\Lambda}+B(X)\frac{R^{(3)}+K^\mu_{\
\nu}K^{\nu}_{\ \mu}-K^2}{\Lambda^2}\right]$$ where $K^{\nu}_{ \ \mu}$ and $R^{(3)}$ are the extrinsic curvature tensor and Ricci scalar of the spacelike uniform-$\phi$ hypersurface, respectively, and $Q(X)\equiv
-\int \sqrt{-X}G_{3X}(X)dX$.
The effective Goldstone Lagrangian {#subsec:goldstone lagrangian}
----------------------------------
An arbitrary time slicing $(t,\bm{x})$ is related to the unitary gauge time $(\tilde{t},\tilde{\bm{x}})$ by $\tilde{t}=t+\pi(t,\bm{x})$ and $\bm{x}=\tilde{\bm{x}}$, then $$\label{goldstone-metric}
\tilde{g}^{00}(\tilde{t},\tilde{\bm{x}})=\frac{\partial \tilde{t}}{\partial x^\mu}\frac{\partial \tilde{t}}{\partial x^\nu}g^{\mu\nu}(t,\bm{x})=(1+\dot{\pi})^2g^{00}+2(1+\dot{\pi})\partial_i \pi g^{0i}+g^{ij}\partial_i \pi \partial_j \pi,$$ where $\pi$ is a Goldstone field. Quantities in the unitary gauge are labeled with tildes. Due to the derivative coupling, the scattering process $\pi\pi\to\pi\pi$ is dominated by sub-Hubble contribution, where the Goldstone mode decouples from gravity [@Cheung:2007st]. We are thus allowed to choose the standard FRW metric $$\label{FRW}
ds^2=g_{\mu\nu}dx^\mu dx^\nu=-dt^2+a^2(t)\delta_{ij}dx^idx^j$$ as the metric in the new coordinate system $(t,\bm{x})$. Relevant Stuckelberg tricks are given in Appendix-\[apdx:Stuekelberg\].
Conventional EFTs (as well as the underlying UV theories) are Poincare invariant. The EFT of cosmological perturbations, on the other hand, are obtained by breaking the time-diffeomorphism invariance with a gauge choice (unitary gauge). When the time-translation symmetry is broken because of the evolving background, the notion of Mandelstam variables becomes ill-defined. However, since the Goldstone theory only contains derivative interactions, the scattering process are dominated by contribution near the cutoff scale $\Lambda$. Symmetry breaking correction from the cosmological background is expected to be small. We assume that *the cosmological evolution only manifests as $\mathcal{O}(H/\Lambda,\dot{H}/\Lambda^2)$ correction to the coefficients and does NOT affect perturbative calculation of the scattering amplitudes.*
It is convenient to define the dimensionless parameters $$\label{epsilon}
\epsilon\equiv\frac{\dot{\phi}^2/2}{M_p^2H^2},\qquad\epsilon_H\equiv-\frac{\dot{H}}{H^2},$$ where $\epsilon_H$ describes evolution of the universe. We will assume $|\ddot{\phi}/(H\dot{\phi})|\ll1$, so that we can safely neglect $\ddot{\phi}$. Here, $\epsilon<1$ is not necessary[^3]. In a $\phi$-dominating universe, $\epsilon$ is related to $\epsilon_H$ by the Friedman equation and the equation of motion of $\phi$, in particular, $\epsilon=\epsilon_H$ if $\phi$ is canonical. We thus have the time derivative of a function $f$ as $$\label{int by part}
\frac{df}{dt}=\left(-\epsilon_H H^2\frac{\partial}{\partial H}+\epsilon^{(1)}_H\frac{\partial}{\partial \epsilon_H}\right)f,$$ where $\dot{\epsilon}_H\equiv H\epsilon^{(1)}_H\simeq2H\epsilon_H^2$ if $\phi$ dominates.
At the tree level, it is sufficient to consider $S^{(2)}$, $S^{(3)}$ and $S^{(4)}$. To quadratic order, after some integration by parts, one has $$\label{L2}
S^{(2)}=(M_p\Lambda)^2\int dx^4a^3\left[U\dot{\pi}^2-V\frac{(\partial_i\pi)^2}{a^2}\right],$$ where
$$U=\epsilon {\left(\frac{H}{\Lambda}\right)}^2\left[-2G_{2X}+(12B_X+12\sqrt{2}\epsilon^{1/2}G_{3X}+8\epsilon G_{2XX}){\left(\frac{H}{\Lambda}\right)}^2+\dots\right],$$
$$V=\epsilon {\left(\frac{H}{\Lambda}\right)}^2\left[-2G_{2X}+(4(7-4\epsilon_H)B_X+8\sqrt{2}\epsilon^{1/2}G_{3X}){\left(\frac{H}{\Lambda}\right)}^2+\dots\right].$$
Here, the ellipsis refers to the parts proportional to higher powers of $H/\Lambda$ and the coefficients are evaluated on the background $\bar{X}=-\dot{\phi}^2/(M_p^2\Lambda^2)=2(H/\Lambda)^2\epsilon$. Here and in the rest of this paper, the results are presented using $\epsilon$ and $\epsilon_H$ instead of $\dot{\phi}^2$ and $\dot{H}$. Truncation to leading order in $H/\Lambda$ is valid if the correction is $<{\cal O}(1)$. For unity $B_X,\epsilon
G_{2XX},\epsilon^{1/2}G_{3X}$, it is sufficient to have $H\lesssim0.1\Lambda$. The sound speed squared is $$\label{sound speed}
c_s^2\equiv
V/U=1+\frac{2}{G_{2X}}\left(4(-1+\epsilon_H)B_X+\sqrt{2}
\epsilon^{1/2} G_{3X}+2 \epsilon G_{2XX}\right){\left(\frac{H}{\Lambda}\right)}^2+\dots,$$ which is approximately constant since $\frac{d}{dt}c_s^2/\Lambda\sim\epsilon_H H^3/\Lambda^3$ is of higher order in $H/\Lambda$. Rescale the spacial coordinates $x\to
c_s x$ and define the canonically normalised field $\pi_c=\sqrt{2c_s^3U}M_p \Lambda\pi$, we have $$\label{quadratic Lagragian}
S^{(2)}=\int dx^4a^3\left[\frac{1}{2}(\partial\pi_c)^2\right],$$ where $(\partial\pi_c)^2=\dot{\pi}_c^2-(\partial_i\pi_c)^2$.
The final Lagrangian for the UV scattering is (the subscript $c$ is omitted) $$\label{scattering Lagrangian}
\begin{aligned}
L=&\frac{1}{2}(\partial\pi)^2+\frac{1}{M_p\Lambda}\left[\alpha_1\dot{\pi}^3+\alpha_2\dot{\pi}(\partial\pi)^2\right]+\frac{1}{M_p^2\Lambda^2}\Big[\beta_1\dot{\pi}^4+\beta_2\dot{\pi}^2(\partial\pi)^2+\beta_3[(\partial\pi)^2]^2\Big]\\&+\frac{\beta_4}{M_p^2\Lambda^4}
\left[\dot{\pi} \partial_i\partial_j\pi \partial_i\dot{\pi}
\partial_j \pi-\dot{\pi}\partial_i\dot{\pi}\partial_i \pi
\nabla^2\pi \right],
\end{aligned}$$ see Appendix-\[apdx:L calculation\] for the details of derivation. The explicit expressions of $\alpha,\beta$ in terms of $G_2(X)$, $G_3(X)$ and $B(X)$ in are given in Appendix-\[apdx:coefficient\]. The coefficients of the final Lagrangian , as expected, do not contain $B(X)$ or $Q(X)$, since in the covariant Lagrangian the Ricci curvature $R$ is invariant and only the $X$ dependent part of $G_3(X)$ is relevant, though they appear in coefficients before individual vertices (for example, eq. contains $B(X)$). This can serve as a quick consistency check of our result.
Positivity bounds {#subsec:apply bound}
-----------------
The tree level $2\to2$ amplitude corresponding to in the center of mass frame is $$\label{amplitude}
{\mathcal{A}}(s,t)=\left(-\frac{9}{4}\alpha_1^2-6\alpha_1\alpha_2-4\alpha_2^2+\frac{3}{2}\beta_1+2\beta_2\right)\frac{s^2}{M_p^2\Lambda^2}+2\beta_3\frac{s^2+t^2+u^2}{M_p^2\Lambda^2}+\frac{1}{2}\beta_4\frac{stu}{M_p^2\Lambda^4}.$$ Detailed calculation of but with $\beta_4=0$ has been presented in Ref.[@Baumann:2015nta]. The Lagrangian is actually non-relativistic, in that uncontracted time derivatives are present. However, if it indeed captures the physics of some UV-complete theory below the high energy cutoff $\Lambda$ and above the decoupling scale $E_{mix}\lesssim \epsilon_H^{1/2}H$ [@Cheung:2007st], positivity derived from locality and unitarity should be inherited. To the leading order in $H/\Lambda$, the positivity bound reads (without loss of generality we set $G_{2X}=-1/2$ hereafter) $$\label{tree level bound}
\begin{aligned}
&(1+\epsilon_H)B_X+\epsilon G_{2XX}
\\&+\bigg[4B_X^2(-61-34\epsilon_H+34\epsilon^2_H-2\epsilon_H^{(1)})+4\epsilon (1+33\epsilon_H)B_XG_{2XX}\\&+2\sqrt{2}\epsilon^{1/2}(-23+5\epsilon_H)B_XG_{3X}+\epsilon(15+\epsilon_H)B_{XX}+4\epsilon^2(2G_{2XX}^2-G_{2XXX})\\&-2\epsilon G_{3X}^2+\sqrt{2}\epsilon^{3/2}(4G_{2XX}G_{3X}-3G_{3XX})\bigg]{\left(\frac{H}{\Lambda}\right)}^2\ge0.
\end{aligned}$$ Applying bound yields ($n=1$) $$\label{tree level t bound}
\begin{aligned}
-B_X+4\bigg[\epsilon B_{XX}+(34-28\epsilon_H)B_X^2-10\epsilon B_X G_{2XX}-\sqrt{2}\epsilon^{1/2}B_X G_{3X}\bigg]{\left(\frac{H}{\Lambda}\right)}^2\ge0.
\end{aligned}$$ Recall in section-\[sec:bound\], we commented that both bounds have $\mathcal{O}(M^2/\Lambda^2)$ uncertainty with $M$ being the energy scale at which the bounds are evaluated. The EFT considered here has both UV cutoff $\Lambda$ and IR cutoff $E_{mix}$. Usually, the Hubble scale satisfies $\Lambda>H>E_{mix}$. If one evaluates the bounds at $M^2\simeq H^2$, the uncertainty $\mathcal{O}(M^2/\Lambda^2)$ of the bound itself is comparable to the background correction, since $M^2/\Lambda^2\simeq
H^2/\Lambda^2$ is of the same order as the leading order cosmological correction. Another interesting case is when $\epsilon_H\ll1$, so we can have $M^2\simeq E^2_{mix}\ll H^2$. In such scenarios, the cosmological correction is much larger than the uncertainty $\mathcal{O}(M^2/\Lambda^2)$, so one is allowed to consider the full corrected bounds.
In the calculation, we did not assume the correction is even powers of $H/\Lambda$. But it turns out the leading order correction is indeed $H^2/\Lambda^2$, as expected in the introduction section, so the bounds do not distinguish between contraction and expansion. As commented before, the cosmological evolution breaks time-translation symmetry of the EFT, so it is not surprising that the bound in Minkowski spacetime (or the limit $H/\Lambda\rightarrow 0$) may be violated for a relatively large $H^2/\Lambda^2$, even if $H^2/\Lambda^2<1$.
Discussion {#subsec:example}
----------
We can ignore the $H^2/\Lambda^2$ correction. Without the corrections, the bounds and are $$\label{0 bound}
(1+\epsilon_H)B_X+\epsilon G_{2XX}\ge0,\qquad B_X\le0.$$ For $1+\epsilon_H>0$, we have the positivity constraint on $B_X$ as $$\label{0 constraint}
\epsilon G_{2XX}\ge -(1+\epsilon_H)B_X\ge0.$$ In comparison, the Minkowski bound only gives $G_{2XX}\ge0$ (see Appendix-\[apdx:minkowski bound\]). Ref.[@Melville:2019wyy] reports $G_{4X}\le0$ for a shift-symmetric Horndeski Lagrangian with only $G_2(X)$ and $G_4(X)$. The covariant theory ($c_T=1$) studied here is essentially different from that in [@Melville:2019wyy] ($c_T\neq 1$), in which the bound $G_{4X}\le0$ is actually equivalent to the subluminal condition $c_T\le1$. In fact, the $G_{4X}\le0$ bound in [@Melville:2019wyy] is derived from $Y^{(2,1)}\ge0$ in [@deRham:2017avq]. The second bound in Eq. has similar origin to $Y^{(2,1)}\ge0$ and it also implies $B_X\le0$. However, as $B_X$ also contributes in the tree level $\pi\pi\to\pi\pi$ scattering, we have an additional bound in Eq. which further constrains $B_X$. As the last comment, one must be careful when applying the results of Ref.[@deRham:2017avq] where the bounds are evaluated at $s\sim \mathcal{O}(m^2)$, which is much smaller than the low energy cutoff $E_{mix}$ of the EFT considered here.
In addition, the subluminality of $c_s^2$ brings another bound $$\label{subluminal}
\sqrt{2}\epsilon^{1/2}G_{3X}\ge 4(1-\epsilon_H)B_X-2\epsilon G_{2XX}.$$ This relation is automatically satisfied if $G_{3X}\ge0$ and $\epsilon_H<1$. However, if $G_{3X}=0$ and $\epsilon_H>1$, it requires $\epsilon G_{2XX}\ge -2(\epsilon_H-1)B_X$, which is stronger than (\[0 constraint\]) for $\epsilon_H>3$.
An interesting example is the slow-roll inflation, where $0<\epsilon\simeq\epsilon_H\ll1$, so the bounds (\[tree level bound\]) and (\[tree level t bound\]) with $H^2/\Lambda^2$ correction can be simplified. To see this, setting $\epsilon=\epsilon_H=0$ in and , we have $$\label{infl bound}
B_X-244 B_X^2{\left(\frac{H}{\Lambda}\right)}^2\ge0,\qquad -B_X+136 B_X^2 {\left(\frac{H}{\Lambda}\right)}^2\ge0.$$ This suggests $B_X=0$ unless we consider the slow-roll suppressed parts. This implies that for the potential-driving dS inflation, the $c_T=1$ beyond-Horndeski EFT reduces to GR. However, for the $\dot \phi$-driving inflation, such as k-inflation [@ArmendarizPicon:1999rj; @Garriga:1999vw] and G-inflation [@Kobayashi:2010cm], since $\epsilon\ll1$ may be violated, the constraint on $B_X$ will be released.
![The shaded region plots parameter space where $G_{2XX}\ge0$ and $c_s^2\le1$ are satisfied. The $H^2/\Lambda^2$ correction in is positive in the darker triangle.[]{data-label="phase"}](G2G3.pdf){width="4in"}
Inspired by the previous example, we look at the GR limit, in which $\phi$ is not coupled to the Ricci scalar ($B\equiv1/2$). The corresponding EFT is called Galileon [@Nicolis:2009qm]. Assuming $G_{2XXX}=G_{3XX}=0$ for simplicity, we have $$\label{GR bound}
\epsilon G_{2XX}+\left[8(\epsilon G_{2XX})^2+4(\epsilon G_{2XX})(\sqrt{2}\epsilon^{1/2}G_{3X})-(\sqrt{2}\epsilon^{1/2}G_{3X})^2\right]{\left(\frac{H}{\Lambda}\right)}^2\ge0.$$ Combining it with $c_s^2\le 1$, we plot the constrained parameter space for $\epsilon
G_{2XX},\sqrt{2}\epsilon^{1/2}G_{3X}$ in Fig.\[phase\]. It is seen that the positivity bound at $H/\Lambda\rightarrow 0$ may be violated (the region shaded with light gray) when the cosmological background $H^2/\Lambda^2$ becomes important.
Conclusion {#sec:conclusion}
==========
We have investigated the application of positivity arguments in the EFT of cosmological perturbations. As an illustrative example, we considered a $c_T=1$ beyond-Horndeski EFT . We explicitly showed by calculation that the leading cosmological correction to positivity bounds indeed comes at $H^2/\Lambda^2$ and $\dot{H}/\Lambda^2$ order, consistent with our observation in the introduction. It is also observed that the positivity bounds found in the limit $H/\Lambda\rightarrow 0$ (or in Minkowski spacetime) might be violated when $H$ is not far smaller than the cutoff scale $\Lambda$, since the coefficient of $H^2/\Lambda^2$ correction is at $10^2$ order. Depending on the sign of coefficient, the bound with cosmological corrections is either weaker or stronger. We also discussed the applications of our bound. It is found that positivity favors a suppressed $B_X$ (comparable in size with $\epsilon G_{2XX}$ or $\sqrt{2}\epsilon^{1/2}G_{3X}$) for slow-roll inflation.
Lagrangian can be used to implement fully stable cosmological bounce. Nonpathological bouncing models built in Ref.[@Cai:2017tku; @Kolevatov:2017voe; @Mironov:2018oec; @Ye:2019frg; @Ye:2019sth] all have $G_{2XX}/G_{2X}<0$ somewhere, which seems to be inconsistent with the bound (\[0 bound\]) at first sight. There is no tension for now. Typically, to violate the null energy condition (NEC, see [@Rubakov:2014jja] for a review), one requires that the operator $X^2$ is not negligible (depending on the value of $\phi$) at the NEC-violating regime, while the beyond-Horndeski operator takes effect as a stabiliser that controls gradient stability. Usually, such models display obvious $\phi$-dependence and nonnegligible $\ddot{\phi}$ around the bounce point, which invalidates the assumptions we used to derive the bounds here. Thus our bounds cannot be directly applied to the bouncing models. It is possible to relax these assumptions in more complete study. We might come back to relevant issues in future works.
It is also interesting to integrate out the IR part of the RHS of within the regime of validity of the EFT to give a more precise bound [@Bellazzini:2017fep]. Recently, the positivity bounds with heavy spinning intermediate states have been studied in inflation but from a covariant point of view [@Kim:2019wjo], and also Ref.[@Herrero-Valea:2019hde] has explored positivity in the Higgs-Dilaton inflation model. It is also well-motivated to go beyond the decoupling limit and study the EFT with graviton and high-spin particles included [@Arkani-Hamed:2015bza; @Lee:2016vti].
**Acknowledgments**
We thank Yong Cai for helpful discussions. This work is supported by by NSFC, Nos.11575188, 11690021.
The Stuekelberg trick {#apdx:Stuekelberg}
=====================
In the unitary guage $$X\equiv \partial_\mu \phi\partial^\mu\phi=\dot{\phi}^2 \tilde{g}^{00},$$ where $\tilde{g}^{00}$ is given in Eq.. To expand Lagrangian , we still need the expression of the extrinsic tensor. The normal of the uniform-$\phi$ hypersurface is $$\label{normal}
n_\mu=\frac{\partial_\mu \tilde{t}}{\sqrt{-\partial_\mu \tilde{t} g^{\mu\nu} \partial_\nu \tilde{t} }}=\frac{\delta^0_\mu+\partial_\mu\pi}{\sqrt{-\tilde{g}^{00}}}.$$ Recall that we are allowed to raise and lower indices with the unperturbed FRW metric in the decoupling limit $$n^\mu=\frac{-\delta^\mu_0+\partial^\mu \pi}{\sqrt{-\tilde{g}^{00}}}.$$ The extrinsic curvature $K$ is $$K=-\nabla_\mu n^\mu, \qquad K^\mu_{\ \nu} K^\nu_{\ \mu}=\nabla_\mu n^\nu \nabla_\nu n^\mu,$$ with the relevant expressions as follows
\[goldstone-extrinsic curvature\] $$\nabla_0 n^0=\partial_0\left(\frac{-1-\dot{\pi}}{\sqrt{-\tilde{g}^{00}}}\right),$$ $$\nabla_0 n^i=\partial_0\left(\frac{\partial_i\pi/a^2}{\sqrt{-\tilde{g}^{00}}}\right)+\frac{H\partial_i\pi/a^2}{\sqrt{-\tilde{g}^{00}}},$$ $$\nabla_i n^0=\partial_i\left(\frac{-1-\dot{\pi}}{\sqrt{-\tilde{g}^{00}}}\right)+\frac{H\partial_i\pi}{\sqrt{-\tilde{g}^{00}}},$$ $$\nabla_i n^j=\partial_i\left(\frac{\partial_j\pi/a^2}{\sqrt{-\tilde{g}^{00}}}\right)-\frac{(1+\dot{\pi})H\delta_i^j}{\sqrt{-\tilde{g}^{00}}}.$$
According to the Gauss-Codazzi formula, the 3d Ricci scalar $R^{(3)}$ is $$R^{(3)}=R+2R_{\mu\nu}n^\mu n^\nu-K^2+K^\mu_{\ \nu} K^\nu_{\ \mu}.$$ On the RHS, the Ricci scalar $R$ is invariant and $R_{\mu\nu}$ can be easily calculated using the spatially flat FRW metric $$R_{00}=-3\frac{\ddot{a}}{a},\qquad R_{ij}=(a\ddot{a}+2\dot{a}^2)\delta_{ij}.$$Now we are well-equipped to expand Lagrangian in powers of $\pi$ and its derivatives.
Derivation of Goldstone Lagrangian {#apdx:L calculation}
===================================
The equation of motion (EoM) of the free Goldstone field is $$\label{pi EoM}
\ddot{\pi}+3H\dot{\pi}-c_s^2\frac{\nabla^2\pi}{a^2}=0.$$ Each $\pi$ field can have at most two derivatives. In fact, only $K_{\mu\nu}$ contributes second derivatives of $\pi$. Thus the n-th order Lagrangian $L^{(n)}$ contains at most $(n+2)$ derivatives. We first consider the part of $L^{(3)}$ with up to four derivatives $$\label{general L3} g_1\dot{\pi}^3+g_2\dot{\pi}\frac{(\partial_i\pi)^2}{a^2}+g_3\dot{\pi}\frac{\partial_i\dot{\pi}\partial_i\pi}{a^2}+g_4\ddot{\pi}\frac{(\partial_i\pi)^2}{a^2}+g_5\dot{\pi}^2\frac{\nabla^2\pi}{a^2}+g_6\ddot{\pi}\frac{\dot{\pi}^2}{a^2},$$ which includes all possible three point interactions with at most four derivatives that may yield nonvanishing amplitudes in the center of mass (CM) frame. Note that in the CM frame, vertices with no time-derivatives (e.g: $\partial_i\partial_j\pi\partial_i\pi\partial_j\pi$) do not contribute in exchange diagrams since the exchanged virtue particle has vanishing 3-momentum. After some integration by parts, we have $$S^{(3)}=M_p^2\Lambda^2\int dx^4 a^3\left\{g_1\dot{\pi}^3+(g_2-{\mathcal{D}_1}g_4)\dot{\pi}\frac{(\partial_i \pi)^2}{a^2}+\dot{\pi}^2\left[g_6\ddot{\pi}-(g_3/2-g_4-g_5)\frac{ \nabla^2\pi }{a^2}\right]\right\},$$ where $\mathcal{D}_n\equiv nH+d/dt$ is defined, and $g_n, \
n=1,2,\dots,6$ are only dependent on time. Insert the EoM and switch to the rescaled coordinates and normalised field, we get $$\label{L3}
\begin{aligned}
S^{(3)}=M_p^2\Lambda^2\int dx^4\sqrt{-g}&\left\{\left[g_1+\frac{g_2}{{c_s^2}}-\frac{{\mathcal{D}_1}g_4}{{c_s^2}}-\frac{1}{3}{\mathcal{D}_3}g_6+(3H-{\mathcal{D}_3}/3)\frac{-g_3/2+g_4+g_5}{{c_s^2}}\right]\dot{\pi}_c^3\right.\\&\left.+\left(-\frac{g_2}{{c_s^2}}+\frac{{\mathcal{D}_1}g_4}{{c_s^2}}\right)\dot{\pi}_c\frac{(\partial\pi_c)^2}{a^2}\right\}\frac{1}{(\sqrt{2c_s^3U}M_p\Lambda)^3}\\\equiv\int dx^4\sqrt{-g}&\left[\alpha_1\dot{\pi}_c^3+\alpha_2\dot{\pi}_c\frac{(\partial\pi_c)^2}{a^2}\right]\frac{1}{M_p\Lambda}.
\end{aligned}$$ Now consider the part with higher-order derivatives $$\frac{1}{a^4}\left[g_7\left(\dot{\pi} (\nabla^2\pi)^2-\dot{\pi}
(\partial_i\partial_j\pi)^2 \right)+g_8(\partial_i
\dot{\pi}\partial_i \pi \nabla^2\pi -\partial_i \dot{\pi}
\partial_j \pi \partial_i\partial_j\pi )\right].$$ After some integration by parts, we find that the higher derivatives cancel out and $$-\frac{1}{a^4}(-H+\frac{d}{dt})\frac{g_8-g_7}{2} \nabla^2\pi ( \partial_j \pi)^2$$ remains, which can be safely neglected in the CM frame. The calculation of $S^{(4)}$ is similar but more involved. In particular, higher-order derivative operators may contribute in contact diagrams. We first consider operators with at most five derivatives and then look at the higher derivative part. The part of $S^{(4)}$ with at most five derivatives is $$\begin{aligned}
&h_1\dot{\pi}^4+h_2\dot{\pi}^2\frac{(\partial_i \pi)^2}{a^2}+h_3\frac{(\partial_i \pi)^4}{a^4}+h_4\frac{ \partial_i\dot{\pi} \partial_i \pi( \partial_j \pi)^2}{a^4}+h_5\dot{\pi}\frac{ \partial_i\partial_j\pi \partial_i \pi \partial_j \pi}{a^4}\\&+h_6\dot{\pi}\frac{ \nabla^2\pi ( \partial_j \pi)^2}{a^4}+h_7\dot{\pi}\ddot{\pi}\frac{(\partial_i \pi)^2}{a^2}+h_8\dot{\pi}^2\frac{ \partial_i\dot{\pi} \partial_i \pi}{a^2}+h_9\dot{\pi}^3\frac{ \nabla^2\pi }{a^2}+h_{10}\ddot{\pi}(\dot{\pi})^3\\
\to&\left[h_1+(3H-{\mathcal{D}_3}/4)\left(\frac{h_9}{{c_s^2}}-\frac{h_8}{3{c_s^2}}+\frac{h_7}{3{c_s^2}}+\frac{h_6}{3c_s^4}-\frac{h_5}{6c_s^4}\right)-\frac{{\mathcal{D}_3}}{4}h_{10}\right]\dot{\pi}^4\\
&+\left[h_2+({\mathcal{D}_1}/4-3H/2)\left(\frac{h_5}{c_s^2}-\frac{2h_6}{c_s^2}\right)-\frac{{\mathcal{D}_1}}{2}h_7\right]\dot{\pi}^2\frac{(\partial_i \pi)^2}{a^2}\\
&+\left[h_3-\frac{\mathcal{D}_{-1}}{8}(2h_4-h_5)\right]\frac{(\partial_i \pi)^4}{a^4}\\\equiv&A\dot{\pi}^4+B\dot{\pi}^2(\partial_i \pi)^2+C(\partial_i \pi)^4\\\to&\left[\left(A+\frac{B}{{c_s^2}}+\frac{C}{c_s^4}\right)\dot{\pi}_c^4-\left(\frac{B}{{c_s^2}}+\frac{2C}{c_s^4}\right)\dot{\pi}_c^2(\partial\pi_c)^2+\frac{C}{c_s^4}(\partial\pi_c)^4\right]\frac{1}{(\sqrt{2c_s^3U}M_p\Lambda)^4}\\
\equiv&\frac{1}{M_p^4\Lambda^4}\left[\beta_1\dot{\pi}_c^4+\beta_2\dot{\pi}_c^2(\partial\pi_c)^2+\beta_3(\partial\pi_c)^4\right].
\end{aligned}$$
The highest derivative part of $S^{(4)}$ divides into to two sectors. The first sector consists solely of spacial derivatives $$\begin{aligned}
\sim &h_{11}\left[(\partial_k\pi)^2( \partial_i\partial_j\pi
)^2-(\partial_k\pi)^2( \nabla^2\pi )^2\right]+h_{12}(\partial_i
\pi \partial_i\partial_j\pi \partial_j\partial_k\pi
\partial_k\pi-\partial_i\partial_j\pi\partial_i \pi \partial_j \pi
\nabla^2\pi ).
\end{aligned}$$ It is easy to check that both the $h_{11}$ and $h_{12}$ vertices cancel out in the CM frame. Vertices in the other sector are $$\begin{aligned}
&h_{13}\left(\dot{\pi}^2 (\partial_i\partial_j\pi)^2-\dot{\pi}^2 (\nabla^2\pi) ^2\right)+h_{14}(\dot{\pi} \partial_i\partial_j\pi \partial_i\dot{\pi} \partial_j \pi-\dot{\pi} \partial_i\dot{\pi} \partial_i \pi \nabla^2\pi )\\
&+h_{15}\left( (\partial_i\dot{\pi} )^2 (\partial_j \pi)^2-\ddot{\pi}(\partial_i \pi)^2 \nabla^2\pi \right)+h_{16}\left( (\partial_i\dot{\pi} \partial_i \pi)^2 -\ddot{\pi} \partial_i\partial_j\pi \partial_i \pi \partial_j \pi\right)\\
\to&(h_{14}-2h_{13})(\dot{\pi} \partial_i\partial_j\pi \partial_i\dot{\pi} \partial_j \pi-\dot{\pi} \partial_i\dot{\pi} \partial_i \pi \nabla^2\pi )+h_{15}\left( (\partial_i\dot{\pi})^2 (\partial_j \pi)^2-\ddot{\pi}(\partial_i\pi)^2 \nabla^2\pi \right)\\&+h_{16}\left( (\partial_i\dot{\pi} \partial_i \pi)^2-\ddot{\pi} \partial_i\partial_j\pi \partial_i \pi \partial_j \pi\right).
\end{aligned}$$ with the coefficients $$\begin{aligned}
h_{13}&=6B+6{\left(\frac{\dot{\phi}}{M_p\Lambda}\right)}^2B_X+4{\left(\frac{\dot{\phi}}{M_p\Lambda}\right)}^4B_{XX},\\
h_{14}&=24B+16{\left(\frac{\dot{\phi}}{M_p\Lambda}\right)}^2B_X,\qquad h_{15}=-h_{16}=4B.
\end{aligned}$$ Performing integration by part again to the terms proportional to $B(X)$, we have $$\begin{aligned}
&\left[4{\left(\frac{\dot{\phi}}{M_p\Lambda}\right)}^2B_X-8{\left(\frac{\dot{\phi}}{M_p\Lambda}\right)}^4B_{XX}\right](\dot{\pi} \partial_i\partial_j\pi \partial_i\dot{\pi} \partial_j \pi-\dot{\pi} \partial_i\dot{\pi} \partial_i \pi \nabla^2\pi )\\&+\mathcal{D}_{-1}(6B) \dot{\pi} \nabla^2\pi \partial_j \pi^2-\frac{B}{2}\frac{d^2}{dt^2}[(\partial_i \pi)^4].
\end{aligned}$$ So in conclusion, the higher derivative operators introduce one new vertex $(\dot{\pi} \partial_i\partial_j\pi \partial_i\dot{\pi} \partial_j \pi-\dot{\pi} \partial_i\dot{\pi} \partial_i \pi \nabla^2\pi )$ and modify the coefficients $$\label{S4 modification}
h_3\to h_3-\frac{B}{2}(1+\epsilon_H)H^2,\qquad h_6\to h_6-6BH.$$ Putting together all the results in this Appendix, we get Lagrangian .
Coefficients in {#apdx:coefficient}
================
We list here the explicit expressions of the coefficients in , calculated to the leading order in $H/\Lambda$,
$$\alpha_1=\frac{1}{\epsilon^{1/2}(-G_{2X})^{3/2}}{\left(\frac{H}{\Lambda}\right)}\left[(2+8\epsilon_H/3)B_X+\sqrt{2}\epsilon^{1/2}G_{3X}\right],$$
$$\alpha_2=\frac{1}{\epsilon^{1/2}(-G_{2X})^{3/2}}{\left(\frac{H}{\Lambda}\right)}\left[2(-1+\epsilon_H)B_X+\frac{1}{2}\sqrt{2}\epsilon^{1/2}G_{3X}+\epsilon G_{2XX}\right],$$
$$\begin{aligned}
\beta_1=\frac{1}{\epsilon G_{2X}^2}&\left\{\frac{2 B_X (-3+\epsilon_H)}{3 }\right.\\&\left.+\frac{1}{3G_{2X}}\bigg[-12 B_X^2 \left(5 \epsilon_H^2-36 \epsilon_H+\epsilon^{(1)}_H+35\right)-14 B_X \epsilon (\epsilon_H-9) G_{2XX}\right.\\&\left.+3 \sqrt{2} B_X \epsilon^{1/2} (3 \epsilon_H+5) G_{3X}+B_{XX} \epsilon (\epsilon_H-27) G_{2X}\bigg]{\left(\frac{H}{\Lambda}\right)}^2\right\},
\end{aligned}$$
$$\begin{aligned}
\beta_2=\frac{1}{\epsilon G_{2X}^2}&\bigg\{B_X (3-\epsilon_H)\\&+\frac{1}{G_{2X}}\bigg[4 B_X^2 \left(10 \epsilon_H^2-53 \epsilon_H+\epsilon^{(1)}_H+49\right)+4 B_X \epsilon (4 \epsilon_H-15) G_{2XX}\\&-2 \sqrt{2} B_X \epsilon^{1/2} (\epsilon_H+2) G_{3X}\\&+\epsilon \left(12 B_{XX} G_{2X}-\epsilon G_{2X} G_{2XXXX}+2 \epsilon G_{2XX}^2+\sqrt{2} \epsilon^{1/2} G_{2XX} G_{3X}\right)\bigg]{\left(\frac{H}{\Lambda}\right)}^2\bigg\},
\end{aligned}$$
$$\begin{aligned}
\beta_3=\frac{1}{\epsilon G_{2X}^2} &\left\{\frac{B_X (3 \epsilon_H-5)+\epsilon G_{2XX}}{8}\right.\\&\left.-\frac{1}{8G_{2X}}\bigg[4 B_X^2 \left(42 \epsilon_H^2-121 \epsilon_H+85\right)+4 B_X \epsilon (29 \epsilon_H-42) G_{2XX}\right.\\&\left.+2 \sqrt{2} B_X \epsilon^{1/2} (3 \epsilon_H-5) G_{3X}+6 B_{XX} \epsilon G_{2X}+\sqrt{2} \epsilon^{3/2} (3 G_{2X} G_{3XX}+2 G_{2XX} G_{3X})\right.\\&\left.+20 \epsilon^2 G_{2XX}^2\bigg]{\left(\frac{H}{\Lambda}\right)}^2\right\}.
\end{aligned}$$
$$\begin{aligned}
\beta_4=&\frac{1}{\epsilon G_{2X}^2}\left[\frac{B_X}{2}{\left(\frac{H}{\Lambda}\right)}^{-2}+\frac{ (34-28\epsilon_H)B_X^2-B_X \left(10 \epsilon G_{2XX}+ \sqrt{2} \epsilon^{1/2} G_{3X}\right)-2 \epsilon B_{XX}G_{2X}}{ G_{2X}}\right]
\end{aligned}$$
Positivity in the Minkowski spacetime {#apdx:minkowski bound}
=====================================
To derive positivity bounds in the Minkowski spacetime for , one should switch back to $\dot{\phi}^2$ and $\dot{H}$ by inserting , and pass to the $H,\dot{H}\to0$ limit. The extrinsic curvature $K$ vanishes in the Minkowski spacetime. Thus positivity only constrains $G_2(X)$ and its derivatives. The Minkowski bound as well as the sound speed are ($G_{2X}=-1/2$) $$\label{minkowski bound}
\begin{aligned}
&G_{2XX}+\left(4 G_{2XX}^2-2 G_{2XXX}\right){\left(\frac{\dot{\phi}}{M_p\Lambda}\right)}^2\\&+ \left(40 G_{2XX}^3-4 G_{2XX} G_{2XXX}+\frac{G_{2XXXX}}{2}\right){\left(\frac{\dot{\phi}}{M_p\Lambda}\right)}^4+\dots\ge0,
\end{aligned}$$ $$c_s^2=\frac{1}{1+4G_{2XX}{\left(\frac{\dot{\phi}}{M_p\Lambda}\right)}^2}.$$ Absence of superluminality places the bound $G_{2XX}\ge0$. An interesting case is that if $G_2(X)\sim X^n$, in which $X^n$ is the first nonnegligible higher-order derivative operator in the $G_2(X)$ polynomial, the positivity bound implies $G_{2XXX}\le0$ if $G_{2XX}=0$ and $G_{2XXXX}\ge0$ if $G_{2XX}=G_{2XXX}=0$, consistent with the result of Ref.[@Chandrasekaran:2018qmx].
[99]{} C. Cheung, P. Creminelli, A. L. Fitzpatrick, J. Kaplan and L. Senatore, JHEP [**0803**]{}, 014 (2008) doi:10.1088/1126-6708/2008/03/014 \[arXiv:0709.0293 \[hep-th\]\]. G. Gubitosi, F. Piazza and F. Vernizzi, JCAP [**1302**]{}, 032 (2013) \[JCAP [**1302**]{}, 032 (2013)\] doi:10.1088/1475-7516/2013/02/032 \[arXiv:1210.0201 \[hep-th\]\]. J. K. Bloomfield, É. É. Flanagan, M. Park and S. Watson, JCAP [**1308**]{}, 010 (2013) doi:10.1088/1475-7516/2013/08/010 \[arXiv:1211.7054 \[astro-ph.CO\]\]. J. Gleyzes, D. Langlois, F. Piazza and F. Vernizzi, JCAP [**1308**]{}, 025 (2013) doi:10.1088/1475-7516/2013/08/025 \[arXiv:1304.4840 \[hep-th\]\]. D. Langlois, M. Mancarella, K. Noui and F. Vernizzi, JCAP [**1705**]{}, no. 05, 033 (2017) doi:10.1088/1475-7516/2017/05/033 \[arXiv:1703.03797 \[hep-th\]\]. G. W. Horndeski, Int. J. Theor. Phys. [**10**]{}, 363 (1974). doi:10.1007/BF01807638 C. Deffayet, X. Gao, D. A. Steer and G. Zahariade, Phys. Rev. D [**84**]{}, 064039 (2011) doi:10.1103/PhysRevD.84.064039 \[arXiv:1103.3260 \[hep-th\]\]. T. Kobayashi, M. Yamaguchi and J. Yokoyama, Prog. Theor. Phys. [**126**]{}, 511 (2011) doi:10.1143/PTP.126.511 \[arXiv:1105.5723 \[hep-th\]\].
J. Gleyzes, D. Langlois, F. Piazza and F. Vernizzi, Phys. Rev. Lett. [**114**]{}, no. 21, 211101 (2015) doi:10.1103/PhysRevLett.114.211101 \[arXiv:1404.6495 \[hep-th\]\]. D. Langlois and K. Noui, JCAP [**1602**]{}, no. 02, 034 (2016) doi:10.1088/1475-7516/2016/02/034 \[arXiv:1510.06930 \[gr-qc\]\]. D. Langlois, Int. J. Mod. Phys. D [**28**]{}, no. 05, 1942006 (2019) doi:10.1142/S0218271819420069 \[arXiv:1811.06271 \[gr-qc\]\]. T. Kobayashi, Rept. Prog. Phys. [**82**]{}, no. 8, 086901 (2019) \[arXiv:1901.07183 \[gr-qc\]\]. Y. Cai, Y. Wan, H. G. Li, T. Qiu and Y. S. Piao, JHEP [**1701**]{}, 090 (2017) doi:10.1007/JHEP01(2017)090 \[arXiv:1610.03400 \[gr-qc\]\]. P. Creminelli, D. Pirtskhalava, L. Santoni and E. Trincherini, JCAP [**1611**]{}, no. 11, 047 (2016) doi:10.1088/1475-7516/2016/11/047 \[arXiv:1610.04207 \[hep-th\]\]. Y. Cai, H. G. Li, T. Qiu and Y. S. Piao, Eur. Phys. J. C [**77**]{}, no. 6, 369 (2017) doi:10.1140/epjc/s10052-017-4938-y \[arXiv:1701.04330 \[gr-qc\]\]. Y. Cai and Y. S. Piao, JHEP [**1709**]{}, 027 (2017) doi:10.1007/JHEP09(2017)027 \[arXiv:1705.03401 \[gr-qc\]\]. R. Kolevatov, S. Mironov, N. Sukhov and V. Volkova, JCAP [**1708**]{}, no. 08, 038 (2017) doi:10.1088/1475-7516/2017/08/038 \[arXiv:1705.06626 \[hep-th\]\]. S. Mironov, V. Rubakov and V. Volkova, JCAP [**1810**]{}, no. 10, 050 (2018) doi:10.1088/1475-7516/2018/10/050 \[arXiv:1807.08361 \[hep-th\]\]. G. Ye and Y. S. Piao, Commun. Theor. Phys. [**71**]{}, no. 4, 427 (2019) doi:10.1088/0253-6102/71/4/427 \[arXiv:1901.02202 \[gr-qc\]\]. G. Ye and Y. S. Piao, Phys. Rev. D [**99**]{}, no. 8, 084019 (2019) doi:10.1103/PhysRevD.99.084019 \[arXiv:1901.08283 \[gr-qc\]\]. G. Obied, H. Ooguri, L. Spodyneiko and C. Vafa, arXiv:1806.08362 \[hep-th\]. H. Ooguri and C. Vafa, Nucl. Phys. B [**766**]{}, 21 (2007) \[hep-th/0605264\]. A. Adams, N. Arkani-Hamed, S. Dubovsky, A. Nicolis and R. Rattazzi, JHEP [**0610**]{}, 014 (2006) doi:10.1088/1126-6708/2006/10/014 \[hep-th/0602178\]. A. Nicolis, R. Rattazzi and E. Trincherini, JHEP [**1005**]{}, 095 (2010) Erratum: \[JHEP [**1111**]{}, 128 (2011)\] doi:10.1007/JHEP05(2010)095, 10.1007/JHEP11(2011)128 \[arXiv:0912.4258 \[hep-th\]\]. B. Bellazzini, L. Martucci and R. Torre, JHEP [**1409**]{}, 100 (2014) doi:10.1007/JHEP09(2014)100 \[arXiv:1405.2960 \[hep-th\]\]. B. Bellazzini, JHEP [**1702**]{}, 034 (2017) doi:10.1007/JHEP02(2017)034 \[arXiv:1605.06111 \[hep-th\]\]. C. de Rham, S. Melville, A. J. Tolley and S. Y. Zhou, Phys. Rev. D [**96**]{}, no. 8, 081702 (2017) doi:10.1103/PhysRevD.96.081702 \[arXiv:1702.06134 \[hep-th\]\]. C. de Rham, S. Melville, A. J. Tolley and S. Y. Zhou, JHEP [**1803**]{}, 011 (2018) doi:10.1007/JHEP03(2018)011 \[arXiv:1706.02712 \[hep-th\]\]. V. Chandrasekaran, G. N. Remmen and A. Shahbazi-Moghaddam, JHEP [**1811**]{}, 015 (2018) doi:10.1007/JHEP11(2018)015 \[arXiv:1804.03153 \[hep-th\]\]. C. de Rham, S. Melville, A. J. Tolley and S. Y. Zhou, JHEP [**1903**]{}, 182 (2019) doi:10.1007/JHEP03(2019)182 \[arXiv:1804.10624 \[hep-th\]\]. J. Tokuda, JHEP [**1905**]{}, 216 (2019) doi:10.1007/JHEP05(2019)216 \[arXiv:1902.10039 \[hep-th\]\]. B. Bellazzini, C. Cheung and G. N. Remmen, Phys. Rev. D [**93**]{}, no. 6, 064076 (2016) doi:10.1103/PhysRevD.93.064076 \[arXiv:1509.00851 \[hep-th\]\]. C. Cheung and G. N. Remmen, JHEP [**1604**]{}, 002 (2016) doi:10.1007/JHEP04(2016)002 \[arXiv:1601.04068 \[hep-th\]\]. C. Cheung and G. N. Remmen, Phys. Rev. Lett. [**118**]{}, no. 5, 051601 (2017) doi:10.1103/PhysRevLett.118.051601 \[arXiv:1608.02942 \[hep-th\]\]. C. de Rham, S. Melville, A. J. Tolley and S. Y. Zhou, JHEP [**1709**]{}, 072 (2017) doi:10.1007/JHEP09(2017)072 \[arXiv:1702.08577 \[hep-th\]\]. B. Bellazzini, F. Riva, J. Serra and F. Sgarlata, Phys. Rev. Lett. [**120**]{}, no. 16, 161101 (2018) doi:10.1103/PhysRevLett.120.161101 \[arXiv:1710.02539 \[hep-th\]\]. C. de Rham, S. Melville and A. J. Tolley, JHEP [**1804**]{}, 083 (2018) doi:10.1007/JHEP04(2018)083 \[arXiv:1710.09611 \[hep-th\]\]. B. Bellazzini, F. Riva, J. Serra and F. Sgarlata, arXiv:1903.08664 \[hep-th\]. B. Bellazzini, M. Lewandowski and J. Serra, arXiv:1902.03250 \[hep-th\]. S. Melville and J. Noller, arXiv:1904.05874 \[astro-ph.CO\]. D. Baumann, D. Green, H. Lee and R. A. Porto, Phys. Rev. D [**93**]{}, no. 2, 023523 (2016) doi:10.1103/PhysRevD.93.023523 \[arXiv:1502.07304 \[hep-th\]\]. N. Kaloper, M. Kleban, A. E. Lawrence and S. Shenker, Phys. Rev. D [**66**]{}, 123510 (2002) doi:10.1103/PhysRevD.66.123510 \[hep-th/0201158\]. M. Froissart, Phys. Rev. [**123**]{}, 1053 (1961). doi:10.1103/PhysRev.123.1053 A. Martin, Phys. Rev. [**129**]{}, 1432 (1963). doi:10.1103/PhysRev.129.1432 A. Martin, Nuovo Cim. A [**42**]{}, 930 (1965). doi:10.1007/BF02720568 P. Creminelli, J. Gleyzes, J. Noreña and F. Vernizzi, Phys. Rev. Lett. [**113**]{}, no. 23, 231301 (2014) doi:10.1103/PhysRevLett.113.231301 \[arXiv:1407.8439 \[astro-ph.CO\]\].
B. P. Abbott [*et al.*]{} \[LIGO Scientific and Virgo Collaborations\], Phys. Rev. Lett. [**123**]{}, no. 1, 011102 (2019) doi:10.1103/PhysRevLett.123.011102 \[arXiv:1811.00364 \[gr-qc\]\]. P. Creminelli and F. Vernizzi, Phys. Rev. Lett. [**119**]{}, no. 25, 251302 (2017) doi:10.1103/PhysRevLett.119.251302 \[arXiv:1710.05877 \[astro-ph.CO\]\]. D. Langlois, R. Saito, D. Yamauchi and K. Noui, Phys. Rev. D [**97**]{}, no. 6, 061501 (2018) doi:10.1103/PhysRevD.97.061501 \[arXiv:1711.07403 \[gr-qc\]\].
P. Agrawal, G. Obied, P. J. Steinhardt and C. Vafa, Phys. Lett. B [**784**]{}, 271 (2018) doi:10.1016/j.physletb.2018.07.040 \[arXiv:1806.09718 \[hep-th\]\]. C. Armendariz-Picon, T. Damour and V. F. Mukhanov, Phys. Lett. B [**458**]{}, 209 (1999) \[hep-th/9904075\]. J. Garriga and V. F. Mukhanov, Phys. Lett. B [**458**]{}, 219(1999) \[hep-th/9904176\].
T. Kobayashi, M. Yamaguchi and J. Yokoyama, Phys. Rev. Lett. [**105**]{}, 231302 (2010) \[arXiv:1008.0603 \[hep-th\]\].
V. A. Rubakov, Phys. Usp. [**57**]{}, 128 (2014) \[Usp. Fiz. Nauk [**184**]{}, no. 2, 137 (2014)\] doi:10.3367/UFNe.0184.201402b.0137 \[arXiv:1401.4024 \[hep-th\]\]. S. Kim, T. Noumi, K. Takeuchi and S. Zhou, arXiv:1906.11840 \[hep-th\]. M. Herrero-Valea, I. Timiryasov and A. Tokareva, arXiv:1905.08816 \[hep-ph\]. N. Arkani-Hamed and J. Maldacena, arXiv:1503.08043 \[hep-th\]. H. Lee, D. Baumann and G. L. Pimentel, JHEP [**1612**]{}, 040 (2016) doi:10.1007/JHEP12(2016)040 \[arXiv:1607.03735 \[hep-th\]\].
[^1]: [email protected]
[^2]: [email protected]
[^3]: Since $|\ddot{\phi}/(H\dot{\phi})|\ll1$ here, the scalar field rolls at nearly constant speed. In one Hubble time, $\Delta\phi\sim
\sqrt{2\epsilon}M_p$. The swampland conjecture [@Obied:2018sgi; @Agrawal:2018own] implies $\epsilon\lesssim
\mathcal{O}(1)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work we compare two different random dilutions on a mean field ferromagnet: the first model is built on a Bernoulli-diluted graph while the second lives on a Poisson-diluted one. While it is known that the two models have, in the thermodynamic limit, the same free energy, we investigate on the structural constraints that the two models must fulfill. We rigorously derive for each model the set of identities for the multi-overlaps distribution, using different methods for the two dilutions: constraints in the former model are obtained by studying the consequences of the self-averaging of the internal energy density, while in the latter are obtained by a stochastic-stability technique. Finally we prove that the identities emerging in the two models are the same, showing [*robustness*]{} of the ferromagnetic properties of diluted graphs with respect to the details of dilution.'
---
**[Adriano Barra$^{*\dag}$, Federico Camboni$^{*}$, Pierluigi Contucci$^{\dag}$]{}**
$^*$ Dipartimento di Fisica, Sapienza Università di Roma
$^\dag$ Dipartimento di Matematica, Università di Bologna
\[section\]
\[Theorem\][Definition]{}
\[Theorem\][Corollary]{}
\[Theorem\][Lemma]{}
\[Theorem\][Example]{}
[**Keywords:**]{} diluted graphs, spin glasses, polynomial identities.
Introduction
============
In the past decade an increasing interest has been shown in statistical mechanics built on diluted graphs (see i.e. [@barabasi][@barrat][@bucca][@barabba3][@ton3][@smallworld]). For diluted spin glasses [@GTVB][@BarDes] this interest is, at least, double motivated: despite their mean field nature, they share with finite-dimensional models the fact that each spin interact with a finite number of other spins. Moreover they are mathematically equivalent to some random optimization problems (i.e. K-SAT or X-OR-SAT depending on the size of the instantaneous interaction [@Ksat1][@Ksat2]).
Although the cases of simpler ferromagnetic models [@barra0][@barraP][@ellis] are not interesting from the point of view of the hard satisfiability interpretation, they are still interesting for their finite connectivity nature and for testing different variations on the topology of the graph they are based on.
With this aim we consider two different ways of diluting the graph [@fan]: the first ferromagnet has its links distributed according to a Bernoulli probability distribution, the second according to a Poisson.
For these models we compared the properties of a family of linear constraints for the order parameters (known as Aizenman-Contucci polynomials [@aizcon][@barra1] in the case of spin glasses).
These relations were investigated earlier [@aizcon][@guerra2] in the spin glass framework, where they were obtained as a result of the stability of the quenched measure with respect to random perturbation, or equivalently through the bound on the fluctuation of the internal energy. Here we propose them as a test for robustness under dilution.
The methods to approach the identities for the two models are structurally different. For the Poisson case, in fact, the additive law of Poissonian random variables makes possible the direct exploitation of the stochastic stability property. The same strategy is not available for the Bernoulli random variables but, for those, we derive the set of identities from the general bound on the quenched fluctuations (even though, for the sake of completeness, in the Appendix we derive the constraints within this general framework for the Poisson case too). The methods we use are generalizations of those appearing in [@barraguerra][@BCK][@CL][@CDGG][@limterm][@pastur][@CGN][@gg][@barra1][@ABC].
Our main result is a rigorous proof of the identities and, especially, the fact that they coincide for the two dilutions.
The mean field diluted ferromagnet
==================================
We introduce a large number $N$ of sites, labeled by the index $i=1,...,N$, and associate to each of them an Ising variable $\sigma_i=\pm1$. We introduce furthermore two families of discrete independent random variables $\{i_{\nu}\}$,$\{j_{\nu}\}$, uniformly distributed on $1,2,...,N$. The Hamiltonian $H_N(\sigma)$ of the diluted ferromagnet is expressed trough $$H_{N}(\sigma) = - \sum_{\nu=1}^{x} \sigma_{i_\nu}\sigma_{j_\nu}$$ where $x$ does depend on the dilution probability distribution. For the Bernoullian dilution case the variable $x$ is called $k$, and defined by $$E_B[\cdot] = \sum_{k=0}^{M}\frac{M!}{(M-k)!k!}
(\frac{\alpha}{N})^k(1-\frac{\alpha}{N})^{M-k}[\cdot],$$ $M=N(N-1)/2$ being the maximum amount of couples $\sigma_i\sigma_j$ existing in the model and $\alpha/N$ the probability that two spins interact. $\alpha >0$ plays the role of the connectivity. The mean and the variance of $k$ are obtained as $$\begin{aligned}
E_B[k] &=&
\frac{M\alpha}{N}, \\
E_B[k^2] - E_B^2[k] &=& \frac{M\alpha}{N}(1 - \frac{\alpha}{N}).\end{aligned}$$ We will make use in the sequel of following properties of a Bernoulli distribution: $$\begin{aligned}
E_B[kg(k)] &=& \frac{M\alpha}{N}E_B[g(k+1)], \label{B1}\\
E_B[k^2g(k)] &=& \frac{M(M-1)\alpha^2}{N^2}E_B[g(k+2)] -
\frac{M\alpha}{N}E_B[g(k+1)],\label{B2} \\
\frac{d}{d\alpha}E_B[g(k)] &=& \frac{M}{N}E_B[g(k+1) -
g(k)].\label{B3}\end{aligned}$$
For the Poissonian dilution case, we denote $x$ as $\xi_{\alpha
N}$, which is a random variable of mean $\alpha N$, for some $\alpha > 0$ (again defining the connectivity of the model), such that P(\_[N]{} = k) = (k,N)= ( -N ) , k = 0,1,2,...Furthermore, we stress that the Poisson distribution obeys the following properties $$\begin{aligned}
k \pi(k,\lambda) &=& \lambda \pi(k-1,\lambda), \\
\frac{d}{d\lambda}\pi(k,\lambda) &=&
-\pi(k,\lambda)+\pi(k-1,\lambda)(1-\delta_{k,0}).\end{aligned}$$ As for the Bernoulli case, the average with respect to the Poisson measure will be denoted by the index $$E_P =
\sum_{k=0}^{\infty} \frac{e^{\alpha N} (\alpha N)^k}{k!}.$$ The expectation with respect to all the quenched variables will be denoted by **E** and represents the product of the expectation over the dilution distribution and the expectation over the uniformly distributed variables $$\textbf{E} = E_{B,P} \cdot \frac{1}{N^2}\sum_{i,j}^{1,N}.$$ The thermodynamic objects we deal with are the partition function $$Z_{N}(\alpha,\beta) = \sum_{\{\sigma\}} e^{-\beta H_N(\alpha)},$$ the quenched intensive free energy $$A_{N}(\alpha,\beta) = \frac{1}{N} \textbf{E}\ln
Z_{N}(\alpha,\beta),$$ the Boltzmann state $$\omega(g(\sigma)) = \frac{1}{Z_{N}(\alpha,\beta)}
\sum_{\{\sigma_N\}} g(\sigma) e^{-\beta H_{N}(\alpha)},$$ the replicated Boltzmann state $$\Omega(g(\sigma)) = \prod_s \omega^{(s)}(g(\sigma^{(s)}))$$ and the global average $\langle g(\sigma)\rangle$ defined as $$\langle g(\sigma)\rangle = \textbf{E}[\Omega(g(\sigma))].$$ The functional order parameter of the theory is the infinite family of multi-overlaps, defined as $$q_{1\cdots
n}=\frac1N\sum_{i=1}^{N}\sigma^{(1)}_{i}\cdots\sigma^{(n)}_{i} ,$$ where particular emphasis is due to the magnetization $m = q_1 =
(1/N)\sum_{i=1}^{N}\sigma_{i}$ and to the two replica overlap $q_{12} = (1/N)\sum_{i=1}^{N}\sigma_{i}^1\sigma_i^2$.
Bernoullian diluted case
========================
Identities in the Bernoullian model will be obtained as a consequence of the internal energy self-average [@contgia][@gg]; before focusing on this procedure, let us recall that
A quantity $A(\sigma)$ is called [*self-averaging*]{} if \_[N ]{}(A() - A())\^2 = \_[N ]{} ( A\^2() - A() \^2 ) = 0,
by which we can remind the
Given two regular functions $A(\sigma)$ and $B(\sigma)$, if at least one of them is self-averaging, then the following relation holds \[idea\] \_[N]{}A()B() = \_[N]{} A() \_[N]{}B()
**Proof** The proof is straightforward. Let us suppose the self-averaging quantity is $B(\sigma)$ and use $A(\sigma)$ as a trial function. Then we have $$\begin{aligned}
0 &\leq& |\langle A(\sigma)B(\sigma)\rangle-\langle A(\sigma)
\rangle \langle B(\sigma) \rangle| \\ \nonumber &=& |\langle
A(\sigma)B(\sigma) - A(\sigma)\langle B(\sigma) \rangle + \langle
A(\sigma) \rangle B(\sigma)\rangle - \langle A(\sigma) \rangle
\langle B(\sigma) \rangle|
\\ &\leq& \sqrt{\langle
A^2(\sigma)\rangle}\sqrt{\langle (B(\sigma) - \langle B(\sigma)
\rangle)^2 \rangle}, \nonumber\end{aligned}$$ where, in the last passage, we used Cauchy-Schwartz relation. In the thermodynamic the proof becomes completed. $\Box$ The scheme to follow is then clear: using the above proposition as the underlying backbone in the derivation of the constraints in this section, we must, at first, show that the internal energy density of the model self-averages. Then we use as trial functions suitably chosen quantities of the order parameters. The identities follow by evaluating explicitly both the terms of eq. (\[idea\]): this operation produces several contributions, all involving the order parameters, among which cancelations occur and the remaining part gives the identities.
Self-averaging of the internal energy density
---------------------------------------------
Once defined $h_l = H(\sigma^{(l)})/N$ as the density of the Hamiltonian evaluated on the generic $l^{th}$ replica and $$\begin{aligned}
\theta &=& \tanh(\beta) \\
\alpha'&=& M\alpha / N^2
\quad \stackrel{N\rightarrow\infty}{\longrightarrow}
\quad \alpha / 2\end{aligned}$$ for simplicity, we start with the following
In the thermodynamic limit, and in $\beta$-average, the internal energy density self-averages \_[N ]{} \_[\_1]{}\^[\_2]{}( (h\^2) - (h)\^2 ) d= 0.
This Theorem has already been proved in [@dembo] by using essentially the existence of the thermodynamic limit for the free energy density. To make the paper self contained we provide an alternative proof in the appendix.
We can now introduce the following lemma.
Let us consider for simplicity the quantity $$\label{delta1G}
\Delta G = \sum_{l=1}^{s} \Big[E \big(\Omega(h_lG) -
\Omega(h_l)\Omega(G)\big) \Big].$$ For every regular, smooth function $G$, in $\beta$-average, we have $$\lim_{N\rightarrow\infty} \int_{\beta_1}^{\beta_2}|\Delta G|d\beta
= 0.$$
**Proof** $$\begin{aligned}
\int_{\beta_1}^{\beta_2}|\Delta G|d\beta &\leq&
\int_{\beta_1}^{\beta_2}\sum_{l=1}^{s}|
\:\textbf{E}[\Omega(h_lG) - \Omega(h_l)\Omega(G)] \:|\:d\beta \label{G1}\\
&\leq& \int_{\beta_1}^{\beta_2}\sum_{l=1}^{s}
\sqrt{\textbf{E}[(\Omega(h_lG) - \Omega(h_l)\Omega(G))^2] }d\beta \label{G2} \\
&\leq& s \int_{\beta_1}^{\beta_2}
\sqrt{\textbf{E}[\Omega(h^2) - \Omega^2(h)] }d\beta \label{G3} \\
&\leq& s \sqrt{\beta_2 - \beta_1}\sqrt{\int_{\beta_1}^{\beta_2}
\textbf{E}[\Omega(h^2) - \Omega^2(h)] d\beta} \;
\stackrel{N\rightarrow\infty}{\longrightarrow}0 \label{G4}\end{aligned}$$ where (\[G1\]) comes from triangular inequality; (\[G2\]) is obtained via the Jensen inequality applied to the measure $\textbf{E}[\cdot]$. In the same way (\[G3\]) comes from Schwarz inequality applied on the measure $\Omega(\cdot)$ (being $G$ bounded), while (\[G4\]) is obtained via Jensen inequality applied on the measure $(\beta_2-\beta_1)^{-1}\int_{\beta_1}^{\beta_2}(\cdot)d\beta$. $\Box$ Now we can state the main theorem for the linear constraints. We are going to introduce a specific trial functions that we call $f_G(\alpha,\beta)$.
\[T1\] Let us consider the following series of multi-overlap functions $G$ acting, in complete generality, on $s$ replicas $$\begin{aligned}
\label{F1}
f_G(\alpha,\beta) &=& \alpha' \Big[ \Big(\sum_{l=1}^{s}\langle G
m_l^2 \rangle - s\langle G m_{s+1}^2 \rangle\Big)
\Big(1 - \theta^2\Big) + \nonumber \\
&& + 2\theta \Big(\sum_{a<l}^{1,s}\langle G q_{al}^2 \rangle -
s\sum_{l}^{1,s}\langle G q_{l,s+1}^2 \rangle +
\frac{s(s+1)}{2}\langle G q_{s+1,s+2}^2 \rangle \Big) + \nonumber \\
&& + 3\theta^2\Big(\sum_{l<a<b}^{1,s}\langle G q_{l,a,b}^2 \rangle
- s\sum_{l<a}^{1,s}\langle G q_{l,a,s+1}^2 \rangle +
\frac{s(s+1)}{2}\sum_{l}^{1,s}\langle G q_{l,s+1,s+2}^2 \rangle + \nonumber \\
&& \quad - \frac{s(s+1)(s+2)}{3!}\langle G q_{s+1,s+2,s+3}^2
\rangle \Big)
+ O(\theta^3)\Big],\end{aligned}$$ in the thermodynamic limit the following generator of linear constraints holds: $$\label{f1}
\lim_{N\rightarrow\infty}\int_{\beta_1}^{\beta_2} d\beta
|f_G(\alpha,\beta)| = 0.$$
**Proof** Let us consider explicitly the quantities encoded in (\[delta1G\]). For the sake of clearness all the calculations are reported in appendix, here we present just the results. $$\begin{aligned}
\textbf{E}[\Omega(h_lG)] &=& -\alpha' \Big[\langle G m_l^2 \rangle
+ \theta\Big(\sum_{a=1}^{s}\langle G q_{a,l}^2 \rangle -
s\langle G q_{l,s+1}^2 \rangle\Big) + \nonumber \\
&& + \theta^2\Big(\sum_{a<b}^{1,s}\langle Gq_{l,a,b}^2 \rangle -
s\sum_{a}^{1,s}\langle Gq_{l,a,s+1}^2 \rangle +
\frac{s(s+1)}{2}\langle Gq_{l,s+1,s+2}^2 \rangle\Big) \nonumber \\
&& + O(\theta^2)\Big],\end{aligned}$$ $$\begin{aligned}
\textbf{E}[\Omega(h_l)\Omega(G)] &=& -\alpha' \Big[\langle G
m_{l}^2 \rangle + \theta\Big(\sum_{a=1}^{s+1}\langle G q_{a,l}^2
\rangle -
(s+1)\langle G q_{l,s+1}^2 \rangle\Big) + \nonumber \\
&& + \theta^2\Big(\sum_{a}^{1,s}\langle Gm_{a}^2 \rangle -
(s+1)\langle Gm_{l}^2 \rangle +
\sum_{a<b}^{1,s}\langle Gq_{l,a,b}^2 \rangle + \nonumber \\
&& - (s+1)\sum_{a}^{1,s}\langle Gq_{l,a,s+1}^2 \rangle +
\frac{(s+1)(s+2)}{2}\langle Gq_{l,s+1,s+2}^2 \rangle\Big) \nonumber \\
&& + O(\theta^2)\Big].\end{aligned}$$ Subtracting the last equation from the former, immediately we conclude that $$\label{delta1Gesp}
\Delta G = - f_G(\alpha,\beta),$$ from which theorem thesis follows. $\Box$
Linear constraints for multi-overlaps
-------------------------------------
We outline here the first order identities as it is customary to do in the spin-glasses counterpart [@barra1] or in neural network [@BG2].
The first class of multi-overlap constraints is obtained by choosing $G=m^2$.
In fact, if we set $G = q_1^2 = m^2$, the function $f_G(\alpha,\beta)$ becomes $$\begin{aligned}
\nonumber
f_{m^2}(\alpha,\beta) &=& \alpha'\Big[\Big(\langle m_1^4\rangle -
\langle m_1^2m_2^2\rangle\Big)\Big(1 - \theta^2\Big) + \\
\nonumber && - 2\theta\Big(\langle m_1^2q_{12}^2\rangle - \langle
m_1^2q_{23}^2\rangle\Big) + \\ \nonumber && + 3\theta^2
\Big(\langle m_1^2q_{123}^2\rangle - \langle m_1^2q_{234}^2\rangle
\Big) + O(\theta^3)\Big],\end{aligned}$$ from which, changing the Jacobian $d\theta = (1 -
\theta^2)d\beta$, we get $$\begin{aligned}
\label{aAC1}
\lim_{N\rightarrow\infty}\int_{\beta_1}^{\beta_2}
|f_{m^2}(\alpha,\beta)|d\beta &=&
\frac{\alpha}{2}\int_{\theta_1}^{\theta_2}d\frac{\theta}{(1 -
\theta^2)} \Big[|\; \Big(\langle m_1^4\rangle - \langle m_1^2m_2^2\rangle\Big) + \\
\nonumber && - 2\theta \Big(\langle
m_1^2q_{12}^2\rangle - \langle m_1^2q_{23}^2\rangle\Big) + \\
\nonumber && + 3\theta^2 \Big(\langle
m_1^2q_{123}^2\rangle - \langle m_1^2q_{234}^2\rangle \\
\nonumber && + O(\theta^3)\;|\Big] = 0,\end{aligned}$$ where, the (not interesting) breakdown at $\theta=1$, of the expression above, reflects the lack of convergence of the harmonic series we used in eq. (\[breakdown\]).
The second class of multi-overlap constraints is obtained by choosing $G=q_{12}^2$.
In fact, if we set $G = q_{12}^2$, the function $f_G(\alpha,\beta)$ becomes $$\begin{aligned}
\nonumber
f_{q_{12}^2}(\alpha,\beta) &=& \alpha'\Big[ \Big( 2\langle
m_1^2q_{12}^2\rangle - 2\langle m_3^2q_{12}^2\rangle\Big)\Big(1 -
\theta^2\Big) + \\ \nonumber && + 2\theta\Big( \langle
q_{12}^4\rangle - 4\langle q_{12}^2q_{23}^2\rangle + 3\langle
q_{12}^2q_{34}^2\rangle\Big) + \\ \nonumber && - 6\theta^2\Big(
\langle q_{12}^2q_{123}^2\rangle - 3\langle
q_{12}^2q_{234}^2\rangle + 2\langle q_{12}^2q_{345}^2\rangle\Big)
+ O(\theta^3)\Big].\end{aligned}$$ Again $$\begin{aligned}
\label{aAC2}
\lim_{N\rightarrow\infty}\int_{\beta_1}^{\beta_2}
|f_{q^2}(\alpha,\beta)|d\beta &=&
\frac{\alpha}{2}\int_{\theta_1}^{\theta_2}d\frac{\theta}{(1 -
\theta^2)} \Big[|\; \Big(\langle
m_1^2q_{12}^2\rangle - \langle m_3^2q_{12}^2\rangle\Big) + \\
\nonumber && + \theta \Big( \langle q_{12}^4\rangle - 4\langle
q_{12}^2q_{23}^2\rangle + 3\langle q_{12}^2q_{34}^2\rangle\Big) +
\\ \nonumber && - 3\theta^2 \Big( \langle
q_{12}^2q_{123}^2\rangle - 3\langle q_{12}^2q_{234}^2\rangle +
2\langle q_{12}^2q_{345}^2\rangle\Big) \nonumber \\ \nonumber && +
O(\theta^3)\;|\Big] = 0,\end{aligned}$$ from which the constraints are obtained as the r.h.s. of (\[aAC1\],\[aAC2\]) set to zero. $\Box$
Poissonian diluted case
=======================
To tackle the Poisson diluted ferromagnet, we are going to use the cavity field approach. The idea beyond this technique is that, calling $F(\beta)$ the extensive free energy and $f(\beta)$ the intensive one, a bridge among the two, in the large $N$ limit, is offered simply by the relation (- F\_[N+1]{}()- F\_N())= f() + O(N\^[-1]{}). As our system has a topologically quenched disordered the $N+1$ spin can be seen as an ”external random cavity field” for the former system of $N$ particles. The identities derived by tuning that field are called of ”stochastic stability”. The simplest way to find them is to consider monomials which are left invariant by the random field: the derivative with respect to it, being zero, will produce the desired polynomial.
Cavity field decompositions for the pressure density
----------------------------------------------------
To start applying the sketched plan let us decompose (in distribution) a Poissonian random Hamiltonian of $N+1$ spins in two Hamiltonians [@ABC]: The former of the “inner” $N$ interacting spins, the latter as the remaining spin interacting with the internal $N$ spins of the cavity.
Up to negligible corrections that go to zero in the thermodynamic limit, we can write in distribution $$H_{N+1}(\alpha) = -\sum_{\nu=1}^{P_{\alpha (N+1)}}
\sigma_{i_\nu}\sigma_{j_\nu} \quad \sim \quad
-\sum_{\nu=1}^{P_{\tilde{\alpha} N}} \sigma_{i_\nu}\sigma_{j_\nu}
- \sum_{\nu=1}^{P_{2\tilde{\alpha }}} \sigma_{i_\nu}\sigma_{N+1},$$ or simply for compactness $$\label{hsplit}
H_{N+1}(\alpha) \sim H_{N}(\tilde{\alpha}) +
\hat{H}_{N}(\tilde{\alpha})\sigma_{N+1}$$ where $$\label{decompo}
\tilde{\alpha} = \frac{N}{N+1}\alpha \stackrel{N\rightarrow
\infty}{\longrightarrow} \alpha, \qquad
\hat{H}_{N}(\tilde{\alpha}) = - \sum_{\nu=1}^{P_{2\tilde{\alpha}}}
\sigma_{i_\nu}.$$ It is useful now to introduce an interpolating parameter $t \in
[0,1]$ on the term encoding the linear connectivity shift so to menage the derivative with respect to the random field by differentiating with respect to this parameter.
We define the $t$-dependent Boltzmann state $\tilde{\omega}_t$ as $$\label{dante}
\tilde{\omega}_t(g(\sigma)) = \frac{1}{Z_{N,t}(\alpha,\beta)}
\sum_{\{\sigma\}}g(\sigma) e^{\beta\sum_{\nu=1}^{P_{\tilde{\alpha}
N}} \sigma_{i_\nu}\sigma_{j_\nu} + \beta
\sum_{\nu=1}^{P_{2\tilde{\alpha }t}} \sigma_{i_\nu}}.$$
We stress the simplicity by which the $t$ parameter switches among the system of $N+1$ spins and the one built just by the former $N$, in the large $N$ limit: In fact, being the two body Hamiltonian left invariant by the gauge symmetry $\sigma_i \to
\epsilon\sigma_i$ for all $i\in(1,...,N)$ with $\epsilon = \pm 1$, by choosing $\epsilon = \sigma_{N+1}$ we have $$\begin{aligned}
Z_{N,t=1}(\tilde{\alpha},\beta) &=& Z_{N+1}(\alpha,\beta), \\
Z_{N,t=0}(\tilde{\alpha},\beta) &=& Z_{N}(\tilde{\alpha},\beta).\end{aligned}$$ Note that $Z_{N,t}(\tilde{\alpha},\beta)$ is defined accordingly to (\[dante\]) and coherently, dealing with the perturbed Boltzmann measure, we introduce an index $t$ also to the global averages $\langle . \rangle \to \langle . \rangle_t$.
Stochastic stability via cavity fields
--------------------------------------
We are now ready to attack the problem. We divide the ensemble of overlap monomials in two large categories: stochastically stable monomials and (as a side results) not stochastically stable ones. Then we find explicitly the family of the stochastically stable monomials, and by putting their $t$-derivative equal to zero we obtain the identities. To follow the plan let us introduce the
We define as stochastically stable monomials those multi-overlap monomials where each replica appears an even number of times.
We are ready to introduce the main theorem, which offers, as a straightforward consequence, a useful corollary, stated immediately after.
\[Tcwdiluito1\] *At $t=1$ the Boltzmannfaktor of the perturbed measure is comparable with the canonical Boltzmannfaktor, and, in the thermodynamic limit, we get* $$\lim_{N \to \infty}
\mathbb{E}\tilde{\Omega}_{N,t=1}(\sigma_{i_1}\sigma_{i_2}...\sigma_{i_n})
=
\lim_{N\to\infty}\mathbb{E}\Omega_{N+1}(\sigma_{i_1}\sigma_{i_2}...\sigma_{i_n}\sigma_{N+1}^n).$$
\[Tcwdiluito3\] *In the thermodynamic limit, the averages $\langle\cdot\rangle_t$ of the stochastically stable monomials become t-independent in $\beta$-average.*
**Proof** Let us focus on the proof of Theorem \[Tcwdiluito1\]. Corollary \[Tcwdiluito3\] will be produced as a straightforward application of Theorem \[Tcwdiluito1\] on stochastically stable monomials. Let us start the proof. Let us assume for a generic multi-overlap monomial the following representation $$Q = \sum_{i_l^1}...\sum_{i_l^s}\prod_{l=1}^{n^a}\sigma_{i_l^a}^a
I(\{i_l^a \})$$ where *a* labels replicas, the inner product accounts for the spins depicted by the index *l* which belong to the Boltzmann state *a* of the product state $\Omega$ for the multi-overlap $q_{a,a'}$ and runs over the integers from $1$ to the amount of times the replica *a* appears into the expression. The external product multiplies all the terms coming from the internal one. The factor $I$ fixes replica-bond constraints. For example the monomial $Q = q_{12}q_{23}$ has $s=3, n^1=n^3=1,
n^2=2$ and $I=N^{-2}\delta_{i_1^1,i_1^2}\delta_{i_1^2,i_2^3}$, there the $\delta$-functions give the correlations $1,2
\rightarrow q_{1,2}$ and $2,3 \rightarrow q_{2,3}$. By applying the Boltzmann and quenched-disordered expectations we get $$\langle Q \rangle_t = \textbf{E}\sum_{i_l^a}I(\{i_l^a
\})\prod_{a=1}^s \omega_{t}(\prod_{l=1}^{n^a}\sigma_{i_l^a}^a).$$ Let us suppose now that $Q$ is not stochastically stable (for otherwise the proof will be simply ended) and let us decompose it by factorizing the Boltzmann state $\omega$ and splitting the terms involving replicas appearing an even number of times from the ones involving replicas appearing an odd number of times. Then, evaluate the whole receipt at $t=1$. $$\langle Q \rangle_t = \textbf{E}\sum_{i_l^a,i_l^b}I(\{i_l^a \},
\{i_l^b \})\prod_{a=1}^u \omega_a (
\prod_{l=1}^{n^a}\sigma_{i_l^a}^a) \prod_{b=u+1}^s \omega_b (
\prod_{l=1}^{n^b}\sigma_{i_l^b}^b),$$ where *u* stands for the amount of replicas which appear an odd number of times inside $Q$. In this way we split the measure $\Omega$ in two ensembles $\omega_a$ and $\omega_b$. Replicas belonging to $\omega_b$ are an even number while the ones in $\omega_a$ an odd number. At this point, as the Hamiltonian has two body interaction and consequently is left unchanged by the symmetry $\sigma_i^a
\rightarrow\sigma_i^a\sigma_{N+1}^a, \forall i \in (1,N)$ (as $\sigma_{N+1}^2 \equiv 1$), we apply such a symmetry globally to the whole set of $N$ spins. The even measure is left unchanged by this symmetry while the odd one takes a multiplying term $\sigma_{N+1}$ $$\langle Q \rangle = \sum_{i_l^a,i_l^b}I(\{i_l^a \}, \{i_l^b \})
\prod_{a=1}^u \omega ( \sigma_{N+1}^a
\prod_{l=1}^{n^a}\sigma_{i_l^a}^a) \prod_{b=u+1}^s \omega
(\sigma_{N+1}^b\prod_{l=1}^{n^b}\sigma_{i_l^b}^b).$$ The last trick is that, by noticing the arbitrariness of the $N+1$ label in $\sigma_{N+1}$, we can change it to a generic label $k$ for each $k \neq \{ i_l^a \}$ and multiply by $1=N^{-1}\sum_{k=1}^N$. At finite $N$ the thesis is recovered forgetting terms $O(1/N)$ and becomes exact in the thermodynamic limit. $\Box$ It is straightforward to check that the effect of Theorem \[Tcwdiluito1\] is not felt by stochastically stable multi-overlap monomials (Corollary \[Tcwdiluito3\]) thanks to the dichotomy of the Ising spins ($\sigma_{N+1}^{2n}\equiv1
\forall n \in\mathbf{N} $). $\Box$ The last point missing to obtain the identities is finding a streaming equation to work out the derivatives with respect to the random field of the stochastically stable monomials. To this task we introduce the following
\[Tcwdfluw\] *Given $F_s$ as a generic function of the spins of $s$ replicas, the following streaming equation holds* $$\begin{aligned}
\nonumber
\frac{\partial\langle F_s \rangle_{t,\tilde{\alpha}}}{\partial t}
&=& 2\tilde{\alpha}\theta [\sum_{a=1}^s\langle F_s \sigma_{i_0}^a
\rangle_{t,\tilde{\alpha}} - s \langle F_s \sigma_{i_0}^{s+1}
\rangle_{t,\tilde{\alpha}}] \quad +
\\ \nonumber
&+& \quad 2\tilde{\alpha}\theta^2 [ \sum_{a<b}^{1,s}\langle F_s
\sigma_{i_0}^a\sigma_{i_0}^b \rangle_{t,\tilde{\alpha}} - s
\sum_{a=1}^s\langle F_s \sigma_{i_0}^a\sigma_{i_0}^{s+1}
\rangle_{t,\tilde{\alpha}} + \\ \nonumber &+&
\frac{s(s+1)}{2!}\langle F_s \sigma_{i_0}^{s+1}\sigma_{i_0}^{s+2}
\rangle_{t,\tilde{\alpha}}] \quad +
\\ \nonumber
&+& \quad 2\tilde{\alpha}\theta^3 [\sum_{a<b<c}^{1,s}\langle F_s
\sigma_{i_0}^a\sigma_{i_0}^b\sigma_{i_0}^c
\rangle_{t,\tilde{\alpha}} - s \sum_{a<b}^{1,s}\langle F_s
\sigma_{i_0}^a\sigma_{i_0}^b\sigma_{i_0}^{s+1}
\rangle_{t,\tilde{\alpha}} +
\\ \nonumber
&+& \frac{s(s+1)}{2!}\sum_{a=1}^{s}\langle F_s \sigma_{i_0}^a
\sigma_{i_0}^{s+1}\sigma_{i_0}^{s+2} \rangle_{t,\tilde{\alpha}} +
\quad \\ \label{cwdfluw} &+& \frac{s(s+1)(s+2)}{3!} \langle F_s
\sigma_{i_0}^{s+1}\sigma_{i_0}^{s+2}\sigma_{i_0}^{s+3}
\rangle_{t,\tilde{\alpha}}] \, + \, O(\theta^3)\end{aligned}$$
**Proof** The proof follows by direct calculations $$\begin{aligned}
\frac{\partial\langle F_s \rangle_{t,\tilde{\alpha}}}{\partial t}
&=& \frac{\partial}{\partial t} \textbf{E} [
\frac{\sum_{\{\sigma\}}F_s
e^{\sum_{a=1}^s(\beta\sum_{\nu=1}^{P_{\tilde{\alpha} N}}
\sigma_{i_\nu}^a\sigma_{j_\nu}^a + \beta
\sum_{\nu=1}^{P_{2\tilde{\alpha }t}} \sigma_{i_\nu}^a)}}
{\sum_{\{\sigma\}}
e^{\sum_{a=1}^s(\beta\sum_{\nu=1}^{P_{\tilde{\alpha} N}}
\sigma_{i_\nu}^a\sigma_{j_\nu}^a + \beta
\sum_{\nu=1}^{P_{2\tilde{\alpha }t}} \sigma_{i_\nu}^a)}}] =
\\ \nonumber
&=& 2\tilde{\alpha} \textbf{E} [ \frac{\sum_{\{\sigma\}}F_s
e^{\sum_{a=1}^s(\beta\sigma_{i_0}^a +
\beta\sum_{\nu=1}^{P_{\tilde{\alpha} N}}
\sigma_{i_\nu}^a\sigma_{j_\nu}^a + \beta
\sum_{\nu=1}^{P_{2\tilde{\alpha }t}} \sigma_{i_\nu}^a)}}
{\sum_{\{\sigma\}} e^{\sum_{a=1}^s(\beta\sigma_{i_0}^a +
\beta\sum_{\nu=1}^{P_{\tilde{\alpha} N}}
\sigma_{i_\nu}^a\sigma_{j_\nu}^a + \beta
\sum_{\nu=1}^{P_{2\tilde{\alpha }t}} \sigma_{i_\nu}^a)}}] -
2\tilde{\alpha}\langle F_s \rangle_{t,\tilde{\alpha}} =
\\ \nonumber
&=& 2\tilde{\alpha} \textbf{E}[ \frac{\tilde{\Omega}_t (F_s
e^{\sum_{a=1}^s\beta\sigma_{i_0}^a})} {\tilde{\Omega}_t
(e^{\sum_{a=1}^s\beta\sigma_{i_0}^a})}] - 2\tilde{\alpha}\langle
F_s \rangle_{t,\tilde{\alpha}} =
\\ \nonumber
&=& 2\tilde{\alpha} \textbf{E}[ \frac{\tilde{\Omega}_t (F_s
\Pi_{a=1}^{s}(\cosh\beta + \sigma_{i_0}^a\sinh\beta))}
{\tilde{\Omega}_t (\Pi_{a=1}^{s}(\cosh\beta +
\sigma_{i_0}^a\sinh\beta))}] - 2\tilde{\alpha}\langle F_s
\rangle_{t,\tilde{\alpha}} =
\\ \nonumber
&=& 2\tilde{\alpha} \textbf{E}[ \frac{\tilde{\Omega}_t (F_s
\Pi_{a=1}^{s}(1 + \sigma_{i_0}^a\theta))} {(1 +
\tilde{\omega}_t(\sigma_{i_0}^a)\theta)^s}] -
2\tilde{\alpha}\langle F_s \rangle_{t,\tilde{\alpha}},\end{aligned}$$ then, by noting that $$\Pi_{a=1}^{s}(1 + \sigma_{i_0}^a\theta) = 1 +
\sum_{a=1}^{s}\sigma_{i_0}^a\theta +
\sum_{a<b}^{1,s}\sigma_{i_0}^a\sigma_{i_0}^b\theta^2 +
\sum_{a<b<c}^{1,s}\sigma_{i_0}^a\sigma_{i_0}^b\sigma_{i_0}^c\theta^3
+ ...$$ $$\frac{1}{(1 + \tilde{\omega}_t \theta)^s} = 1 - s\tilde{\omega}_t
\theta + \frac{s(s+1)}{2!}\tilde{\omega}_t^2 \theta^2 -
\frac{s(s+1)(s+2)}{3!}\tilde{\omega}_t^3 \theta^3 + ...$$ we get $$\frac{\partial\langle F_s \rangle_{t,\tilde{\alpha}}}{\partial t}
= 2\tilde{\alpha} \textbf{E}[ \tilde{\Omega}_t (F_s(1 +
\sum_{a=1}^{s}\sigma_{i_0}^a\theta +
\sum_{a<b}^{1,s}\sigma_{i_0}^a\sigma_{i_0}^b\theta^2 +
\sum_{a<b<c}^{1,s}\sigma_{i_0}^a\sigma_{i_0}^b\sigma_{i_0}^c\theta^3
+ ...)) \times$$ $$\times (1 - s\tilde{\omega}_t \theta +
\frac{s(s+1)}{2!}\tilde{\omega}_t^2 \theta^2 -
\frac{s(s+1)(s+2)}{3!}\tilde{\omega}_t^3 \theta^3 + ...)] -
2\tilde{\alpha}\langle F_s \rangle_{t,\tilde{\alpha}},$$ from which the thesis follows. $\Box$
Linear constraints for multi-overlaps
-------------------------------------
We saw that the stochastically stable multi-overlap monomials become asymptotically independent by the $t$ parameter upon increasing the size of the system. Calling for simplicity $G_N(q)$ a stochastically stable multi-overlap monomial, identities follow as a consequence of Corollary \[Tcwdiluito3\] and are encoded in the following relation $$\lim_{N\to \infty} \partial_t \langle G_{N}(q) \rangle_t = 0.$$ As we did when we investigated the Bernoullian model, we analyze the stability of $\langle m^2 \rangle$ and $\langle
q_{12}^2\rangle$, up to the third order in $\theta$, so to compare the results at the end. $$\begin{aligned}
\nonumber
\partial_{t}\langle m_1^2 \rangle_t &=&
2\tilde{\alpha}\theta\Big (\langle m_1^3 \rangle_t - \langle
m_1^2m_2 \rangle_t\Big) - 2\tilde{\alpha}\theta^2 \Big(\langle
m_1^2q_{12} \rangle_t - \langle m_1^2q_{23} \rangle_t\Big) \\
\nonumber &+& \; 2\tilde{\alpha}\theta^3 \Big(\langle m_1^2q_{123}
\rangle_t - \langle m_1^2q_{234} \rangle_t\Big) + O(\theta^3)\end{aligned}$$ $$\begin{aligned}
\label{comp3}
\Rightarrow \quad && \Big[2\alpha\theta\Big (\langle m_1^4 \rangle
- \langle m_1^2m_2^2 \rangle\Big) - 2\alpha\theta^2
\Big( \langle m_1^2q_{12}^2 \rangle - \langle m_1^2q_{23}^2 \rangle\Big) \nonumber \\
&& + \; 2\alpha\theta^3 \Big(\langle m_1^2q_{123}^2 \rangle -
\langle m_1^2q_{234}^2 \rangle\Big) + O(\theta^3)\Big] = 0\end{aligned}$$ $$\begin{aligned}
\partial_{t}\langle q_{12}^2 \rangle_t &=& 4\tilde{\alpha}\theta
\Big(\langle m_1q_{12}^2 \rangle_t - \langle m_3q_{12}^2
\rangle_t\Big) + 2\tilde{\alpha}\theta^2 \Big(\langle q_{12}^3
\rangle_t - 4\langle
q_{12}^2q_{13}\rangle_t + 3\langle q_{12}^2q_{34} \rangle_t\Big) + \nonumber \\
&-& \quad 4\tilde{\alpha}\theta^3 \Big(\langle q_{12}^2q_{123}
\rangle_t - 3\langle q_{12}^2q_{134} \rangle_t + 4\langle
q_{12}^2q_{345} \rangle_t\Big) + O(\theta^3) \nonumber\end{aligned}$$ $$\begin{aligned}
\label{comp4}
\Rightarrow && \Big[4\alpha\theta \Big(\langle m_1^2q_{12}^2
\rangle - \langle m_3^2q_{12}^2 \rangle\Big) + 2\alpha\theta^2
\Big(\langle q_{12}^4 \rangle - 4\langle
q_{12}^2q_{13}^2\rangle + 3\langle q_{12}^2q_{34}^2 \rangle\Big) + \nonumber \\
&& - \quad 4\alpha\theta^3 \Big(\langle q_{12}^2q_{123}^2 \rangle
- 3\langle q_{12}^2q_{134}^2 \rangle + 2\langle q_{12}^2q_{345}^2
\rangle\Big) + O(\theta^3)\Big] = 0\end{aligned}$$
Discussion and outlook
======================
Let us start this section by comparing the results we get from the two models. We have to compare respectively eq.s (\[aAC1\]) versus (\[comp3\]) and (\[aAC2\]) versus (\[comp4\]). We see that the details of the dilution do not affect the constraints: the series show the same set of identities. In fact, despite we are not generally allowed to set to zero each term in the expressions (\[comp3\],\[comp4\],\[aAC1\],\[aAC2\]) (as we do to obtain alone the following identities (\[64\]-\[69\])), at least close to the critical line, where different multi-overlaps have different scaling laws [@ABC], i.e. $q_{n}^2 \propto (\alpha\theta - 1)^n$, such a spreading is possible and we can forget each single $(\alpha,\beta)$-coefficient as it does not affect the identities (it is never involved into the averages $\langle . \rangle$). We get $$\begin{aligned}
\label{64}
0 &=& \langle m_1^4 \rangle - \langle m_1^2 m_2^2 \rangle, \\
\label{65} 0 &=& \langle m_1^2 q_{12}^2 \rangle - \langle
m_1^2q_{23}^2 \rangle, \\ \label{66} 0 &=& \langle m_1^2 q_{123}^2
\rangle - \langle m_1^2q_{234}^2 \rangle,
\end{aligned}$$ when investigating the magnetization as a trial function and $$\begin{aligned}
\label{67}
0 &=& \langle m_1^2 q_{12}^2 \rangle - \langle m_3^2 q_{12}^2 \rangle, \\
\label{68} 0 &=& \langle q_{12}^4 \rangle - 4 \langle
q_{12}^2q_{23}^2 \rangle + 3 \langle q_{12}^2 q_{34}^2 \rangle, \\
\label{69} 0 &=& \langle q_{12}^2 q_{123}^2 \rangle - 3\langle
q_{12}^2 q_{134}^2 \rangle +2 \langle q_{12}^2 q_{345}^2 \rangle,
\end{aligned}$$ when investigating the two replica overlap. Even if a minor point, we stress that the differences among the global $(\alpha,\beta)$-coefficients (not shown here) clearly are related to the differences in the two method involved (in the former the constraints appear under the integral over the temperature, while in the latter this $\beta$-average is already worked out), furthermore the limiting connectivity in the Bernoulli dilution is $\alpha/2$, while is $2\alpha$ in the Poisson model; so there is an overall factor $4$ of difference among the results (for the sake of clearness we worked out in the appendix also the constraints in the Poisson diluted case via the first method, to check explicitly the coherence among the two model coefficients). Then, by looking explicitly at the constraints, several physical features can be recognized, in fact every term is well known: the first class (Eq.s \[64\],\[65\],\[66\]) is the standard first momentum self-averaging on replica symmetric systems; In fact, by assuming replica equivalence, eq. (\[64\]) turns out to be the standard internal energy self-averaging of the Curie-Weiss model. Eq.(\[65\]) and (\[66\]) contribute as higher order internal energy self-average by taking into account the dilution (in fact, they go to zero whenever $\alpha \to \infty$ because $\theta$-powers higher than one go to zero and only the Curie-Weiss self-averaging for the internal energy survives as it should). With a glance at the identities coming from the second constraint series (Eq.s \[67\],\[68\],\[69\]) we recognize immediately the replica symmetry ansatz for the magnetization in the first identity, followed by the first and the second Aizenman-Contucci relation for systems with quenched disorder [@aizcon][@barra0]. Interestingly these series are in agreement even with other models, apparently quite far away, as spin-glasses with Gaussian coupling $\mathcal{N}[1,1]$ instead of $\mathcal{N}[0,1]$ [@CGN]. A very interesting conjecture may be that these constraints hold for systems whose interaction has on average positive strength and are affected by quenched disorder, independently if the disorder affects the strength of the interaction or the topology of the interaction. We plan to report soon on this topics.
Acknowledgment {#acknowledgment .unnumbered}
==============
The author are pleased to thank an anonymous referee for useful suggestions. AB work in this paper is partially supported by the Technological Vaucher Contract n.11606 of Calabria Region and partially by the CULTAPTATION Project (European Commission contract FP6 - 2004-NEST-PATH-043434). PC acknowledge Strategic Grant of University of Bologna.
Appendix
========
Alternative proof of energy self-averaging
------------------------------------------
Starting from the thermodynamic relation $$\label{varh}
\textbf{E}[\Omega(h^2) - \Omega^2(h)] =
-\frac{1}{N}\frac{d}{d\beta}\textbf{E}[\Omega(h)]$$ we evaluate explicitly the term $E[\Omega(h)]$ as $$\begin{aligned}
\textbf{E}[\Omega(h)] &=& - \frac{1}{N}\textbf{E}\Big[
\frac{\sum_{\{\sigma\}}\sum_{\nu=1}^{k}
\sigma_{i_\nu}\sigma_{j_\nu}e^{-\beta H} }
{Z_N(\alpha,\beta)}\Big] \label{h1} =\\
&=& - \frac{1}{N}\textbf{E}\Big[ \frac{\sum_{\{\sigma\}}
k\sigma_{i_0}\sigma_{j_0}e^{-\beta H} }
{Z_N(\alpha,\beta)}\Big] \label{h2} =\\
&=& - \alpha'\textbf{E}\Big[ \frac{\sum_{\{\sigma\}}
\sigma_{i_0}\sigma_{j_0}e^{\beta\sigma_{i_0}\sigma_{j_0}}e^{-\beta
H} } {\sum_{\{\sigma\}}
e^{\beta\sigma_{k_0}\sigma_{l_0}}e^{-\beta H}}\Big] \label{h3} =\\
&=& - \alpha'\textbf{E}\Big[ \frac{\sum_{\{\sigma\}}
\sigma_{i_0}\sigma_{j_0}(\cosh\beta +
\sigma_{i_0}\sigma_{j_0}\sinh\beta)e^{-\beta H} }
{\sum_{\{\sigma\}}
(\cosh\beta + \sigma_{k_0}\sigma_{l_0}\sinh\beta)e^{-\beta H}}\Big] \label{h4} =\\
&=& - \alpha'\textbf{E}\Big[ \frac{\sum_{\{\sigma\}}
\sigma_{i_0}\sigma_{j_0}(1 +
\sigma_{i_0}\sigma_{j_0}\theta)e^{-\beta H} } {\sum_{\{\sigma\}}
(1 + \sigma_{k_0}\sigma_{l_0}\theta)e^{-\beta H}}\Big] \label{h5} =\\
&=& - \alpha'\textbf{E}\Big[
\frac{\omega(\sigma_{i_0}\sigma_{j_0}) + \theta} {1 +
\omega(\sigma_{k_0}\sigma_{l_0})\theta}\Big], \label{h6}\end{aligned}$$ where in (\[h2\]) we fixed the index $\nu$, in (\[h3\]) we used the property (\[B1\]) of the Bernoulli distribution and we introduced two further families of random variables $\{k_{\nu}\}$,$\{l_{\nu}\}$, and in (\[h4\]) we used $e^{\beta\sigma_{i_0}\sigma_{j_0}} = \cosh\beta +
\sigma_{i_0}\sigma_{j_0}\sinh\beta$. Let us now expand the denominator of (\[h6\]) keeping in mind the relation $$\frac{1}{(1 + \tilde{\omega}_t \theta)^p} = 1 - p\tilde{\omega}_t
\theta + \frac{p(p+1)}{2!}\tilde{\omega}_t^2 \theta^2 -
\frac{p(p+1)(p+2)}{3!}\tilde{\omega}_t^3 \theta^3 + ...$$ such that, by posing $p=1$, we obtain $$\textbf{E}[\Omega(h)] = - \alpha'\textbf{E}\Big[ \theta +
\sum_{n=1}^{\infty}(-1)^n\theta^n(1-\theta^2)\langle
q_{1...n}^2\rangle \Big].$$ By applying the modulus function to the equation above we can proceed further with the following bound $$\label{breakdown}
|\textbf{E}[\Omega(h)]| \leq \alpha'\textbf{E}\Big[ |\theta| +
\sum_{n=1}^{\infty}|\theta^n(1-\theta^2)\langle
q_{1...n}^2\rangle| \Big].$$ Both $|\theta|$ and $|\langle q_{1...n}^2\rangle|$ belong to $[0,1]$ so we get $$|E[\Omega(h)]| \leq \alpha'\Big[ 1 +
(1-\theta^2)\sum_{n=1}^{\infty}\theta^n \Big],$$ whose harmonic series converges to $1/(1-\theta)$, $|\theta|<1$; The fact that the convergence is not guaranteed at zero temperature with this technique is not a problem because, first the identities we are looking for hold in $\beta$-average, secondly the zero temperature has been intensively investigated elsewhere [@GDS]. For each finite $\beta$, then, we can write $$\begin{aligned}
|\textbf{E}[\Omega(h)]| &\leq& \alpha'\Big[ 1 +
\frac{(1-\theta^2)}{1-\theta} \Big] = \\
&=& \alpha'\Big[ 1 +
\frac{(1-\theta)(1+\theta)}{1-\theta} \Big] = \\
&=&\alpha'\Big[ 1 +
(1+\theta) \Big] \leq \\
&\leq& 3\alpha', \label{minor}\end{aligned}$$ and consequently $$\begin{aligned}
\int_{\beta_1}^{\beta_2}\textbf{E}[\Omega(h^2) -
\Omega^2(h)]d\beta &\leq&
\int_{\beta_1}^{\beta_2}|\;\textbf{E}[\Omega(h^2) - \Omega^2(h)]\;|d\beta = \nonumber \\
&=& \frac{1}{N}\int_{\beta_1}^{\beta_2}|\;\frac{d}{d\beta}\textbf{E}[\Omega(h)]\;|d\beta \quad\leq \nonumber \\
&\leq& 3\frac{\alpha'}{N}\end{aligned}$$ $$\Rightarrow \qquad \lim_{N\rightarrow\infty}
\int_{\beta_1}^{\beta_2}\textbf{E}[\Omega(h^2) -
\Omega^2(h)]d\beta = 0$$ and the proof is closed. $\Box$
Details in Bernoulli dilution calculations
------------------------------------------
Here some technical calculations concerning the self-averaging technique on the Bernouilli diluted graph are reported. $$\begin{aligned}
\textbf{E}[\Omega(h_lG)] &=& -\frac{1}{N}
\textbf{E}\Big[\frac{\sum_{\{\sigma\}}
\sum_{\nu=1}^{k}\sigma_{i_\nu}^l\sigma_{j_\nu}^lG
\:e^{(\beta\sum_{a=1}^{s}\sum_{\nu=1}^{k}\sigma_{i_\nu}^a\sigma_{j_\nu}^a)}}
{\sum_{\{\sigma\}}
e^{(\beta\sum_{a=1}^{s}\sum_{\nu=1}^{k}\sigma_{i_\nu}^a\sigma_{j_\nu}^a)}}] = \nonumber \\
&=& -\frac{1}{N} \textbf{E}\Big[\frac{\sum_{\{\sigma\}} k
\sigma_{i_0}^l\sigma_{j_0}^lG
\:e^{(\beta\sum_{a=1}^{s}\sum_{\nu=1}^{k}\sigma_{i_\nu}^a\sigma_{j_\nu}^a)}}
{\sum_{\{\sigma\}}
e^{(\beta\sum_{a=1}^{s}\sum_{\nu=1}^{k}\sigma_{i_\nu}^a\sigma_{j_\nu}^a)}}],
\nonumber\end{aligned}$$ and remembering the properties of the Bernoullian distribution (\[B1\]) we can write $$\begin{aligned}
&=& -\frac{\alpha M}{N^2} \textbf{E}\Big[\frac{\sum_{\{\sigma\}}
\sigma_{i_0}^l\sigma_{j_0}^lG
\prod_{a=1}^{s}e^{\beta\sigma_{i_0}^a\sigma_{j_0}^a} \:e^{-\beta
H_s}} {\sum_{\{\sigma\}}
\prod_{a=1}^{s}e^{\beta\sigma_{i_0}^a\sigma_{j_0}^a} e^{-\beta
H_s}}\Big] = \nonumber \\ \nonumber &=& -\frac{\alpha M}{N^2}
\textbf{E}\Big[\frac{\sum_{\{\sigma\}}
\sigma_{i_0}^l\sigma_{j_0}^lG \prod_{a=1}^{s}[\cosh\beta +
\sigma_{i_0}^a\sigma_{j_0}^a\sinh\beta] \:e^{-\beta H_s}}
{\sum_{\{\sigma\}} \prod_{a=1}^{s}[\cosh\beta +
\sigma_{i_0}^a\sigma_{j_0}^a\sinh\beta] \:e^{-\beta H_s}}\Big] =\end{aligned}$$ $$\begin{aligned}
\nonumber &=& -\frac{\alpha M}{N^2}
\textbf{E}\Big[\frac{\sum_{\{\sigma\}}
\sigma_{i_0}^l\sigma_{j_0}^lG \prod_{a=1}^{s}[1 +
\sigma_{i_0}^a\sigma_{j_0}^a\theta] \:e^{-\beta H_s}}
{\sum_{\{\sigma\}} \prod_{a=1}^{s}[1 +
\sigma_{i_0}^a\sigma_{j_0}^a\theta]
\:e^{-\beta H_s}}\Big] = \nonumber \\
&=& -\frac{\alpha M}{N^2}
\textbf{E}\Big[\frac{\Omega\Big(\sigma_{i_0}^l\sigma_{j_0}^lG
\prod_{a=1}^{s}[1 + \sigma_{i_0}^a\sigma_{j_0}^a\theta]\Big)}
{\Big(1 + \omega(\sigma_{i_0}^a\sigma_{j_0}^a)\theta\Big)^s}\Big]
= \nonumber\end{aligned}$$ let us expand both the numerator and the denominator up to the second order in $\theta$ $$\begin{aligned}
&&= -\frac{\alpha M}{N^2}
\textbf{E}\Big[\Omega\Big((\sigma_{i_0}^l\sigma_{j_0}^lG) (1 +
\sum_{a}^{1,s}\sigma_{i_0}^a\sigma_{j_0}^a\theta +
\sum_{a<b}^{1,s}\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^b\sigma_{j_0}^b\theta^2)\Big) \times \nonumber \\
&& \times \; \Big(1 - s\omega(\sigma_{i_0}\sigma_{j_0})\theta +
\frac{s(s+1)}{2}\omega^2(\sigma_{i_0}\sigma_{j_0})\theta^2\Big)\Big] = \nonumber \\
&&= -\frac{\alpha M}{N^2}
\textbf{E}\Big[\Omega\Big(G\sigma_{i_0}^l\sigma_{j_0}^l +
G\sum_{a}^{1,s}\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^l\sigma_{j_0}^l\theta
+
G\sum_{a<b}^{1,s}\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^b\sigma_{j_0}^b\sigma_{i_0}^l\sigma_{j_0}^l\theta^2\Big)
\times \nonumber\end{aligned}$$ $$\begin{aligned}
&& \times \; \Big(1 - s\omega(\sigma_{i_0}^a\sigma_{j_0}^a)\theta +
\frac{s(s+1)}{2}\omega^2(\sigma_{i_0}\sigma_{j_0})\theta^2\Big)\Big] = \nonumber \\
&&= -\frac{\alpha M}{N^2}
\textbf{E}\Big[\Omega(G\sigma_{i_0}^l\sigma_{j_0}^l) +
\theta\Big(\sum_{a}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^l\sigma_{j_0}^l)
-
s\Omega(G\sigma_{i_0}^l\sigma_{j_0}^l)\omega(\sigma_{i_0}\sigma_{j_0})\Big)
+ \nonumber \\
&& +
\;\theta^2\Big(\sum_{a<b}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a
\sigma_{i_0}^b\sigma_{j_0}^b\sigma_{i_0}^l\sigma_{j_0}^l) -
s\sum_{a}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^l\sigma_{j_0}^l)
\omega(\sigma_{i_0}\sigma_{j_0}) + \nonumber\end{aligned}$$ $$\begin{aligned}
&& \qquad + \quad \frac{s(s+1)}{2}\Omega(G\sigma_{i_0}^l\sigma_{j_0}^l)\omega^2(\sigma_{i_0}^a\sigma_{j_0}^a)\Big)\Big] = \nonumber \\
&&= -\frac{\alpha M}{N^2}\Big[ \langle Gm_l^2 \rangle +
\theta\Big(\sum_{a=1}^{s}\langle G q_{a,l}^2 \rangle -
s\langle G q_{l,s+1}^2 \rangle\Big) + \nonumber \\
&&+ \theta^2\Big(\sum_{a<b}^{1,s}\langle Gq_{l,a,b}^2 \rangle -
s\sum_{a}^{1,s}\langle Gq_{l,a,s+1}^2 \rangle +
\frac{s(s+1)}{2}\langle Gq_{l,s+1,s+2}^2 \rangle\Big)\Big].
\nonumber\end{aligned}$$ While the other term $\textbf{E}[\Omega(h_l)\Omega(G)]$ can be worked out as follows: **E**\[(h\_l)(G)\] = $$\begin{aligned}
&&= -\frac{1}{N} \textbf{E}\Big[\frac{\sum_{\{\sigma\}}
\sum_{\nu=1}^{k}\sigma_{i_\nu}^l\sigma_{j_\nu}^lG
\:e^{\beta\sum_{\nu=1}^{k}\sigma_{i_\nu}^l\sigma_{j_\nu}^l}
\:e^{(\beta\sum_{a=1}^{s}\sum_{\nu=1}^{k}\sigma_{i_\nu}^a\sigma_{j_\nu}^a)}}
{\sum_{\{\sigma\}}
e^{\beta\sum_{\nu=1}^{k}\sigma_{i_\nu}^l\sigma_{j_\nu}^l}
e^{(\beta\sum_{a=1}^{s}\sum_{\nu=1}^{k}\sigma_{i_\nu}^a\sigma_{j_\nu}^a)}}] = \nonumber \\
&&= -\frac{1}{N} \textbf{E}\Big[\frac{\sum_{\{\sigma\}} k
\sigma_{i_0}^l\sigma_{j_0}^lG
\:e^{\beta\sum_{\nu=1}^{k}\sigma_{i_\nu}^l\sigma_{j_\nu}^l}
\:e^{(\beta\sum_{a=1}^{s}\sum_{\nu=1}^{k}\sigma_{i_\nu}^a\sigma_{j_\nu}^a)}}
{\sum_{\{\sigma\}}
e^{\beta\sum_{\nu=1}^{k}\sigma_{i_\nu}^l\sigma_{j_\nu}^l}
e^{(\beta\sum_{a=1}^{s}\sum_{\nu=1}^{k}\sigma_{i_\nu}^a\sigma_{j_\nu}^a)}}]
= \nonumber\end{aligned}$$ $$\begin{aligned}
&&= -\frac{\alpha M}{N^2} \textbf{E}\Big[\frac{\sum_{\{\sigma\}}
\sigma_{i_0}^l\sigma_{j_0}^lG
\:e^{\beta\sigma_{i_0}^l\sigma_{j_0}^l}
[\prod_{a=1}^{s}e^{\beta\sigma_{i_0}^a\sigma_{j_0}^a}] \:e^{-\beta
H_{s+1}}} {\sum_{\{\sigma\}} e^{\beta\sigma_{i_0}^l\sigma_{j_0}^l}
[\prod_{a=1}^{s}e^{\beta\sigma_{i_0}^a\sigma_{j_0}^a}]
e^{-\beta H_{s+1}}}\Big] = \nonumber \\
&&= -\frac{\alpha M}{N^2}
\textbf{E}\Big[\frac{\Omega\Big(\sigma_{i_0}^l\sigma_{j_0}^lG (1 +
\sigma_{i_0}^l\sigma_{j_0}^l\theta) [\prod_{a=1}^{s}(1 +
\sigma_{i_0}^a\sigma_{j_0}^a\theta)]\Big)}
{\Big(1 + \omega(\sigma_{i_0}\sigma_{j_0})\theta\Big)^{s+1}}\Big] = \nonumber \\
&&= -\frac{\alpha M}{N^2}
\textbf{E}\Big[\Omega\Big((\sigma_{i_0}^l\sigma_{j_0}^lG +
G\theta) (1 + \sum_{a}^{1,s}\sigma_{i_0}^a\sigma_{j_0}^a\theta +
\sum_{a<b}^{1,s}\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^b\sigma_{j_0}^b\theta^2)\Big) \times \nonumber \\
&& \times \; \Big(1 - (s+1)\omega(\sigma_{i_0}\sigma_{j_0})\theta
+
\frac{(s+1)(s+2)}{2}\omega^2(\sigma_{i_0}\sigma_{j_0})\theta^2\Big)\Big]
= \nonumber\end{aligned}$$ $$\begin{aligned}
&&= -\frac{\alpha M}{N^2}
\textbf{E}\Big[\Big(\Omega(G\sigma_{i_0}^l\sigma_{j_0}^l) +
\theta\Big(\Omega(G) + \sum_{a}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^l\sigma_{j_0}^l)\Big) + \nonumber \\
&& + \:\theta^2\Big(\sum_{a}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a) + \sum_{a<b}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^b\sigma_{j_0}^b\sigma_{i_0}^l\sigma_{j_0}^l)\Big) \times \nonumber \\
&& \times \; \Big(1 - (s+1)\omega(\sigma_{i_0}\sigma_{j_0})\theta
+
\frac{(s+1)(s+2)}{2}\omega^2(\sigma_{i_0}\sigma_{j_0})\theta^2\Big)\Big]
= \nonumber \end{aligned}$$ $$\begin{aligned}
&&= -\frac{\alpha M}{N^2}
\textbf{E}\Big[\Omega(G\sigma_{i_0}^l\sigma_{j_0}^l) +
\theta\Big(\Omega(G) +
\sum_{a}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^l\sigma_{j_0}^l)
\nonumber
\\
&& -
(s+1)\Omega(G\sigma_{i_0}^l\sigma_{j_0}^l)\omega(\sigma_{i_0}\sigma_{j_0})\Big) + \nonumber \\\end{aligned}$$ $$\begin{aligned}
&& +
\;\theta^2\Big(\sum_{a}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a)
+ \sum_{a<b}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a
\sigma_{i_0}^b\sigma_{j_0}^b\sigma_{i_0}^l\sigma_{j_0}^l)
\nonumber \\
&&-(s+1)\Omega(G)\omega(\sigma_{i_0}\sigma_{j_0}) + \nonumber \\
&& \quad - (s+1)\sum_{a}^{1,s}\Omega(G\sigma_{i_0}^a\sigma_{j_0}^a\sigma_{i_0}^l\sigma_{j_0}^l)
\omega(\sigma_{i_0}\sigma_{j_0})
\nonumber \\ && + \frac{(s+1)(s+2)}{2}\Omega(G\sigma_{i_0}^l\sigma_{j_0}^l)\omega^2(\sigma_{i_0}^a\sigma_{j_0}^a)\Big)\Big] = \nonumber\end{aligned}$$ $$\begin{aligned}
\nonumber
&&= -\frac{\alpha M}{N^2} \Big[\langle G m_{l}^2 \rangle +
\theta\Big(\langle G \rangle + \sum_{a=1}^{s}\langle G q_{a,l}^2
\rangle -
(s+1)\langle G q_{l,s+1}^2 \rangle\Big) + \\
&&\qquad + \quad \theta^2\Big(\sum_{a}^{1,s}\langle Gm_{a}^2 \rangle -
(s+1)\langle Gm_{l}^2 \rangle +
\sum_{a<b}^{1,s}\langle Gq_{l,a,b}^2 \rangle + \nonumber \\
&&\qquad - \quad (s+1)\sum_{a}^{1,s}\langle Gq_{l,a,s+1}^2 \rangle +
\frac{(s+1)(s+2)}{2}\langle Gq_{l,s+1,s+2}^2 \rangle\Big) \Big] \nonumber\end{aligned}$$
Poisson identities via the self-averaging technique
---------------------------------------------------
For the sake of completeness we report also the constraints in the Poisson diluted model obtained by using the first method: $$\begin{aligned}
\label{F1}
f_G(\alpha,\beta) &=& \alpha \Big[ \Big(\sum_{l=1}^{s}\langle G
m_l^2 \rangle - s\langle G m_{s+1}^2 \rangle\Big)
\Big(1 - \theta^2\Big) + \nonumber \\
&+& 2\theta \Big(\sum_{a<l}^{1,s}\langle G q_{al}^2 \rangle -
s\sum_{l}^{1,s}\langle G q_{l,s+1}^2 \rangle +
\frac{s(s+1)}{2}\langle G q_{s+1,s+2}^2 \rangle \Big) + \nonumber \\
&+& 3\theta^2\Big(\sum_{l<a<b}^{1,s}\langle G q_{l,a,b}^2 \rangle
- s\sum_{l<a}^{1,s}\langle G q_{l,a,s+1}^2 \rangle +
\frac{s(s+1)}{2}\sum_{l}^{1,s}\langle G q_{l,s+1,s+2}^2 \rangle + \nonumber \\
&-& \frac{s(s+1)(s+2)}{3!}\langle G q_{s+1,s+2,s+3}^2 \rangle
\Big)
+ O(\theta^3)\Big],\end{aligned}$$ by which, choosing as a trial function $m^2$ we have $$\begin{aligned}
\nonumber
f_{m^2}^P(\alpha,\beta) &=& \alpha\Big[\Big(\langle m_1^4\rangle -
\langle m_1^2m_2^2\rangle\Big)\Big(1 - \theta^2\Big) + \\
\nonumber && - 2\theta\Big(\langle m_1^2q_{12}^2\rangle - \langle
m_1^2q_{23}^2\rangle\Big) + \\ \nonumber && + 3\theta^2
\Big(\langle m_1^2q_{123}^2\rangle - \langle m_1^2q_{234}^2\rangle
\Big) + O(\theta^3)\Big],\end{aligned}$$ from which, changing the Jacobian $d\theta = (1 -
\theta^2)d\beta$, we get $$\begin{aligned}
\label{AC1}
\lim_{N\rightarrow\infty}\int_{\beta_1}^{\beta_2}
|f_{m^2}^P(\alpha,\beta)|d\beta &=&
\alpha\int_{\theta_1}^{\theta_2}d\theta \Big[|\;
\Big(\langle m_1^4\rangle - \langle m_1^2m_2^2\rangle\Big) + \\
\nonumber &-& 2\frac{\theta}{(1 - \theta^2)} \Big(\langle
m_1^2q_{12}^2\rangle - \langle m_1^2q_{23}^2\rangle\Big) + \\
\nonumber &+& 3\frac{\theta^2}{(1 - \theta^2)} \Big(\langle
m_1^2q_{123}^2\rangle - \langle m_1^2q_{234}^2\rangle \\
\nonumber &+& O(\theta^3)\;|\Big] = 0.\end{aligned}$$ If we set $G = q_{12}^2$ as the trial function $f_G^P(\alpha,\beta)$ becomes $$\begin{aligned}
\nonumber
f_{q^2}^P(\alpha,\beta) &=& \alpha\Big[ \Big( 2\langle
m_1^2q_{12}^2\rangle - 2\langle m_3^2q_{12}^2\rangle\Big)\Big(1 -
\theta^2\Big) + \\ \nonumber &+& 2\theta\Big( \langle
q_{12}^4\rangle - 4\langle q_{12}^2q_{23}^2\rangle + 3\langle
q_{12}^2q_{34}^2\rangle\Big) + \\ \nonumber &-& 6\theta^2\Big(
\langle q_{12}^2q_{123}^2\rangle - 3\langle
q_{12}^2q_{234}^2\rangle + 2\langle q_{12}^2q_{345}^2\rangle\Big)
+ O(\theta^3)\Big].\end{aligned}$$ Again $$\begin{aligned}
\label{AC2}
\lim_{N\rightarrow\infty}\int_{\beta_1}^{\beta_2}
|f_{q^2}^P(\alpha,\beta)|d\beta &=&
2\alpha\int_{\theta_1}^{\theta_2}d\theta \Big[|\; \Big(\langle
m_1^2q_{12}^2\rangle - \langle m_3^2q_{12}^2\rangle\Big) + \\
\nonumber &+& \frac{\theta}{(1 - \theta^2)}\Big( \langle
q_{12}^4\rangle - 4\langle q_{12}^2q_{23}^2\rangle + 3\langle
q_{12}^2q_{34}^2\rangle\Big) + \\ \nonumber &-&
3\frac{\theta^2}{(1 - \theta^2)}\Big( \langle
q_{12}^2q_{123}^2\rangle - 3\langle q_{12}^2q_{234}^2\rangle +
2\langle q_{12}^2q_{345}^2\rangle\Big) \nonumber \\ \nonumber &+&
O(\theta^3)\;|\Big] = 0,\end{aligned}$$
[9]{}
E. Agliari, A. Barra, F. Camboni, *Criticality in diluted ferromagnets*, J. Stat. Mech. P1003 (2008).
M. Aizenman, P. Contucci, [*On the stability of the quenched state in mean field spin glass models*]{}, J. Stat. Phys. [**92**]{}, 765-783 (1998).
R. Albert, A. L. Barabasi *Statistical mechanics of complex networks*, Reviews of Modern Physics **74**, 47-97 (2002).
A. Barra, *The mean field Ising model trhought interpolating techniques*, J. Stat. Phys. **145**, 234-261, (2008).
A. Barra *Notes of ferromagnetic Pspin and REM*, Math. Meth. Appl. Sci. $10.1002$/mma.$1065$, Wiley (2008).
A. Barra, *Irreducible free energy expansion and overlap locking in mean field spin glasses*, J. Stat. Phys. **123**, 601-614 (2006).
A. Barra, L. DeSanctis *Stability properties and probability distribution of multi-overlaps in dilute spin glasses*, J. Stat. Mech. P08025 (2007).
A. Barra, F. Guerra, *About the ergodicity in Hopfield analogical neural networks*, J. Math. Phys. **49** $125217$ (2008).
A. Barra, F. Guerra, *Order parameters and their locking in analogical neural networks*, ”Percorsi Incrociati", Dedicated Volume, Collana d’Ateneo, University of Salerno (2008).
A. Barrat, M. Weight, *On the properties of small world network models*, Eur. Phys. Journ. B **13**-3 (2000).
A. Bianchi, P. Contucci, A. Knauf, *Stochastically Stable Quenched Measures*, Journal of Statistical Physics, **117**, 831-844, (2004).
M. Buchanan, *Nexus: Small Worlds and the Groundbreaking Theory of Networks*. Norton, W. W. Company, Inc. (2003).
F. Chung, L. Lu, *Complex Graphs and Networks*, Americ. Math. Soc. Publ. (2006).
P. Contucci, C. Giardiná, *The Ghirlanda-Guerra identities*, J. Stat. Phys., **126**, N. 4/5, 917-931, (2007).
P. Contucci, C. Giardiná, I. Nishimori, *Spin Glass Identities and the Nishimori Line*, to appear in “Progress in Probability”, (2008).
P. Contucci, J. Lebowitz, *Correlation Inequalities for Spin Glasses*, Annales Henri Poincare, **8**, 1461-1467, (2007).
P. Contucci, M. Degli Esposti, C. Giardiná, S. Graffi, *Thermodynamical Limit for Correlated Gaussian Random Energy Models*, Commun. in Math. Phys. **236**, 55-63, (2003).
A. Dembo, A. Montanari, *Ising models on locally tree-like graphs*, arXiv:0804.4726 (2008).
R.S. Ellis, *Large deviations and statistical mechanics*, Springer, New York (1985).
S. Ghirlanda, F. Guerra, [*General properties of overlap distributions in disordered spin systems. Towards Parisi ultrametricity*]{}, J. Phys. A, [**31**]{}, 9149-9155, (1998).
F. Guerra, [*About the overlap distribution in mean field spin glass models*]{}, Int. Jou. Mod. Phys. B [**10**]{}, 1675-1684, (1996).
F. Guerra, L. De Sanctis, [*Mean field dilute ferromagnet I. High temperature and zero temperature behavior*]{}, J. Stat. Phys. **129**, 231, (2008).
F. Guerra, F. L. Toninelli, [*The Thermodynamic Limit in Mean Field Spin Glass Models*]{}, Comm. Math. Phys. [**230**]{}, 71-79, (2002).
F. Guerra, F. L. Toninelli, [*The high temperature region of the Viana-Bray diluted spin glass model*]{}, J. Stat. Phys. **115**, (2004).
M.Mezard, G.Parisi, R. Zecchina, *Analytic and Algorithmic Solution of Random Satisfiability Problems*, Science **297**, 812 (2002).
R. Monasson, S.Kirkpatrick, B.Selman, L.Troyansky, R. Zecchina, *Determining computational complexity from characteristic phase transitions*, Nature **400**, 133 (1999).
M. Newman, D. Watts, A.-L. Barabasi *The Structure and Dynamics of Networks*, Princeton University Press, (2006).
T. Nikoletoupolous, A.C.C. Coolen, I. Perez Castillo, N.S.Skantos, J.P.L. Hatchett, B. Wemmenthove, *Replicated transfer matrix analysis of Ising spin models on small world lattices*, J. Phys. A, 6455-6475 (2004).
L. Pastur, M. Shcherbina, [*The absence of self-averaging of the order parameter in the Sherrington-Kirkpatrick model*]{}, J. Stat. Phys. [**62**]{}, 1-19, (1991).
D.J. Watts, S.H. Strogatz, *Collective dynamics of small world networks*, Nature **393**, 6684 (1998)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the conformal symmetry and the energy-momentum conservation of scalar field interacting with a curved background at $D=2$. We avoid to incorporate the metric determinant into the measure of the scalar field to explain the conformal anomaly and the consequent energy-momentum conservation. Contrarily, we split the scalar field in two other fields, in such a way that just one of them can be quantized. We show that the same usual geometric quantities of the anomaly are obtained, which are accompanied by terms containing the new field of the theory.'
author:
- 'M. Alves'
- 'J. Barcelos-Neto'
title: 'On the trace anomaly and the energy-momentum conservation of quantum fields at D=2 in classical curved backgrounds'
---
Introduction
============
The interaction between quantum fields and classical gravity has been intensively studied for a long time [@Birrell]. One of the main motivations is that this procedure is considered to be the first initial step to understand the full quantum theory including the gravitatonal field itself. It resembles the old formalism of dealing with quantum fields interacting with the classical external electromagnetic background before the advent of the quantum electrodynamics.
However, even though these are aparently similar procedures, the former is much more involved. For example, the concept of particle is not well defined. Consequently, the definition of $\vert in\rangle$ and $\vert out\rangle$ states cannot be done in a clear way and the definition $S$-matrix becomes meaningless. So, instead os particles, it is the energy-momentum tensor that plays the fundamental role, due to its local nature and being the source of the curvature in the general relativity theory. There is an important property that the energy-momentum tensor has to satisfy, the (covariant) divergenceless, which corresponds to the conservation of energy and momentum of the theory. This is a kind of symmetry that cannot be modified by quantum corrections.
On the other hand, we have an interesting symmetry related to massless theories, called conformal symmetry, which means, in a broader sense, the absence of scales. This symmetry is manifested in the traceless of the energy-momentum tensor. Contrarily to previous case, it is not necessarily kept in the quantum scenario (trace anomaly). This occurs because the quantum formalism naturally introduces scale parameters in order to deal with infinities during the regularization procedures.
In this paper, we consider a quantum massless scalar field interacting with a classical curved background. We mention that this theory exhibits the conformal symmetry, where for $D>2$ it is necessary to couple the scalar field to the classical curvature (nominimal coupling) [@Birrell]. The particular case of $D=2$, that is the subject of the present paper, has an interesting feature. The conformal symmetry is verified without necessity of coupling the scalar field to the curvature, because there is no conformal transformation for it. This result may lead to a wrong conclusion that there is no trace anomaly for scalar fields at $D=2$ because the absence of conformal transformation for them leads to an invariance of the corresponding measure in the path integral formalism.
The above reasoning and conclusion, which there is no trace anomaly for $D=2$, cannot be true because it is accompanied by an unpleasant absence of energy and momentum conservation (after quantum corrections). A more carefull study of the conformal symmetry shows that the conformal anomaly does actually exist and the energy and momentum are actually conserved, as they should be [@Fujikawa; @Fujikawa2]. We mention that this problem can be circumvented by splitting the $\sqrt{-g}$ of the action as $\sqrt{-g}=
(-g)^{-\frac{1}{4}}(-g) ^{-\frac{1}{4}}$ and incorporating each one of these factors to the scalar field [@Fujikawa]. In this way, the new scalar field acquires a convenient conformal transformation [@Alves], whose noninvariance of the measure renders the expected trace anomaly and the energy-momentum conservation.
The purpose of the present paper is to display a different alternative of dealing with this problem. We avoid to incorporate any factor involving the metric tensor to the scalar field because, since we intend to work with the path integral formalism, this would be inconsistent with the initial assumption that the gravitational field is classical. Our proposal consists in splitting the scalar field in a product of two fields with different conformal transformations. The classical conformal symmetry is not modified, but the measure of one of them is. We show that this leads to a trace anomaly, that has the same geometrical terms of the usual case, plus other ones related to the new field. We also show that, eventhough more involved, the energy momentum conservation is also achieved quantically.
Our paper is organized as follows. In Sec. II we brief discuss how this this problem can be solved by incorporating a factor inovolving the metric tensor into the scalar field. Eventhough this is a section review, we follow the lines that we shall used into the next section, where our formalism is presented. We left Sec. IV for some concluding remarks.
Trace anomaly and energy-momentum conservation
==============================================
Let us start from the action
$$S=\frac{1}{2}\int d^2x\,\sqrt{-g}\,g^{\mu\nu}
\partial_\mu\phi\,\partial_\nu\phi
\label{2.1}$$
The classical energy-momentum tensor is
$$\begin{aligned}
T_{\mu\nu}&=&\frac{2}{\sqrt{-g}}\,\frac{\delta S}{\delta g^{\mu\nu}}
\nonumber\\
&=&\partial_\mu\phi\,\partial_\nu\phi
-\frac{1}{2}\,g_{\mu\nu}\,\partial^\rho\phi\,\partial_\rho\phi
\label{2.2}\end{aligned}$$
which has the following properties
$$\begin{aligned}
&&g^{\mu\nu}T_{\mu\nu}=0
\label{2.3}\\
&&\nabla^\mu T_{\mu\nu}=\partial_\nu\phi\,\Box\,\phi=0
\label{2.4}\end{aligned}$$
where $\nabla^\mu$ is the covariante derivative and $\Box$ is the Laplace-Beltrami operator, $\Box\phi=\frac{1}{\sqrt{-g}}
\partial_\mu (\sqrt{-g}\partial^\mu\phi)
=g^{\mu\nu}\nabla_\mu\nabla_\nu\phi$.
Expression (\[2.3\]) tell us that the theory exhibits the conformal symmetry. The conformal transformation for the metric tensor is $\tilde g_{\mu\nu}=e^{2\alpha}g_{\mu\nu}$ and, consequently, $\sqrt{-g}\,g^{\mu\nu}$ is conformally invariant. This means that the conformal symmetry is verified at $D=2$ without any transformation for the scalar field $\phi$. The meaning of expression (\[2.4\]) is that the energy and momentum are conserved. It is opportune to mention that this expression was obtained by using the equation of motion for $\phi$.
In the quantum scenario (with a classical background metric) the energy momentum tensor can be obtained by means of the vacuum functional as
$$\langle T_{\mu\nu}\rangle
=\frac{2}{\sqrt{-g}}\,\frac{\delta Z}{\delta g^{\mu\nu}}
\label{2.6}$$
where
$$Z=\,\int[d\phi]\,\exp\,\Bigl(\frac{i}{2}\int d^2x\,
\sqrt{-g}\,g^{\mu\nu}\partial_\mu\phi\partial_\nu\phi\Bigr)
\label{2.5}$$
Since the scalar field does not change under conformal transformation, we have that the measure $[d\phi]$ also remains invariant. Consequently, the semiclassical expression for the energy-momentum tensor reads
$$\langle T_{\mu\nu}\rangle=
\langle\partial_\mu\phi\,\partial_\nu\phi
-\frac{1}{2}\,g_{\mu\nu}\,\partial^\rho\phi\,\partial_\rho\phi\rangle
\label{2.7}$$
The curved background is considered to be classical, so we may have
$$\begin{aligned}
&&g^{\mu\nu}\langle T_{\mu\nu}\rangle
=\langle g^{\mu\nu}T_{\mu\nu}\rangle=0
\label{2.8}\\
&&\nabla^\mu\langle T_{\mu\nu}\rangle
=\langle\nabla^\mu T_{\mu\nu}\rangle
=\langle\partial_\nu\,\phi\Box\,\phi\rangle
\label{2.9}\end{aligned}$$
It is not possible to conclude that expression (\[2.9\]) is zero because the equation of motion cannot be used in the quantum scenario. Of course, the results above do not merit confidence because there is no reason to believe that energy and momentum are not conserved after quantum effects are taken into account.
To circumvent this problem, the action (\[2.1\]) can be rewriten as
$$S=\frac{1}{2}\int d^2x\,g^{\mu\nu}\partial_\mu\Phi\,\partial_\nu\Phi
\label{2.10}$$
where
$$\Phi=(-g)^\frac{1}{4}\phi
\label{2.11}$$
Of course, the classical $T_{\mu\nu}$ is precisely the previous one given by (\[2.2\]) (which can be rewritten in terms of $\Phi$) and the action, $S$ given by (\[2.10\]), is still conformally invariant. The conformal transformation for the new scalar field $\Phi$ is
$$\tilde\Phi=e^\alpha\Phi
\label{2.12}$$
However, for the vacuum functional, the measure $[d\Phi]$ is not invariant under conformal transformation. In a general way, we have [@Fujikawa3]
$$[d\tilde\Phi]=\exp\Bigl(i\int d^2x\sqrt{-g}\,
\alpha(x)\,A(x)\Bigr)\,[d\Phi]
\label{2.13}$$
where $A(x)$ is a badly divergent quantity that can be regularized by means of the zeta function technique leading to [@Hawking]
$$\begin{aligned}
A(x)&=&\lim_{s\rightarrow0}\,{\rm tr}\,\zeta(x,s)
\nonumber\\
&=&\frac{[a_1(x)]}{4\pi}
\label{2.14}\end{aligned}$$
The last step of the expression above is restrict to $D=2$, and the coefficient $a_1(x,x^\prime)$ is related to heat kernel expansion. The notation $[a_1(x)]$ means $[a_1(x)]= \lim_{x^\prime\rightarrow x}
a_1(x,x^\prime)$. These coefficients can be obtained by means of a recursion relation that depends on the kind of operator that acts on the field [@DeWitt]. For the present case, $a_1(x)=-\frac{1}{6}
\,R$, where $R$ is the Ricci scalar curvature
Now, the corresponding energy-momentum tensor obtained by means of expression (\[2.5\]), which we shall denote by $\tilde T_{\mu\nu}$, is not traceless. The trace $\langle T^\mu\,_\mu\rangle$ can be directly obtained by
$$\begin{aligned}
\langle\tilde T^\mu\,_\mu\rangle
&=&-\frac{i}{\sqrt{-g}}\,\frac{\delta Z}{\delta\alpha}
\nonumber\\
&=&A(x)
\nonumber\\
&=&-\frac{1}{24\pi}\,R
\label{2.14a}\end{aligned}$$
This result embodies the trace anomaly. Since in two spacetime dimensions we have the identity $R_{\mu\nu}= \frac{1}{2}g_{\mu\nu}R$, one may say that the full expression for the energy-momentum tensor $\langle\tilde T_{\mu\nu}\rangle$ should be
$$\langle\tilde T_{\mu\nu}\rangle
=\langle T_{\mu\nu}\rangle-\frac{1}{48\pi}g_{\mu\nu}R
\label{2.14b}$$
where $\langle T_{\mu\nu}\rangle$ is the one given by (\[2.7\]) (it is indifferent to write it in terms of $\phi$ or $\Phi$). Acting the covariant derivative in both sides of the expression above, we get
$$\nabla^\mu\langle\tilde T_{\mu\nu}\rangle
=\langle\partial_\nu\phi\,\Box\,\phi\rangle
-\frac{1}{48\pi}\partial_\nu R
\label{2.15}$$
Expanding the field $\phi$ in terms of eigenfunctions of the operator $\Box$, one can show that [@Fujikawa]
$$\langle\partial_\nu\phi\,\Box\phi\rangle
=\frac{1}{2}\partial_\nu\langle\phi\Box\phi\rangle
\label{2.16}$$
and the quantity $\langle\phi\,\Box\,\phi\rangle$ can be regularized and leads to [@Fujikawa]
$$\langle\phi\,\Box\,\phi\rangle=\frac{1}{24\pi}\,R
\label{2.18}$$
So,
$$\nabla^\mu\langle\tilde T_{\mu\nu}\rangle=0
\label{2.19}$$
as it should be.
Alternative procedure
=====================
Now, instead of incorporating the factor $(-g)^{-\frac{1}{4}}$ to the scalar field, we go in a opposite direction by splitting the field $\phi$ as
$$\phi=e^\theta\,\varphi
\label{3.1}$$
where $\theta$ and $\varphi$ are considered to be two independent quantities with the following conformal tranformations
$$\begin{aligned}
&&\tilde\varphi=e^{-\alpha}\varphi
\label{3.2}\\
&&\tilde\theta=\theta+\alpha
\label{3.3}\end{aligned}$$
The field $\varphi$ remains quantum, but $\theta$ can be quantum or not. It is important the field $\theta$ appears in a exponential term and, consequently, with a conformal transformation like (\[3.3\]). This is so because, in the hypothesis that $\theta$ is also quantum, its corresponding measure, $[d\theta]$, remains unchanged (the jacobian is trivial). In the developments which follow, we shall consider $\theta$ classical. At the end, we briefly talk on the possibility of $\theta$ being quantum.
Replacing $\phi$ given by (\[3.1\]) into the initial expression for $S$, (\[2.1\]), we have
$$S=-\frac{1}{2}\int d^2x\,\sqrt{-g}\,
\varphi\,\Bigl[e^{2\theta}
\bigl(\Box -\partial_\mu\theta\partial^\mu\theta)\Bigr]\varphi
\label{3.4}$$
We have just done a change of variables and, consequently, there is no changing into the classical case. But, in the path integral, the measure $[d\varphi]$ is not invariant under conformal transformation. Considering $\theta$ classical, se have the vacuum functional
$$Z=\int[d\varphi]\exp\biggl\{-\frac{i}{2}
\int d^2x\sqrt{-g}\varphi\Bigl[e^{2\theta}\bigl(\Box
-\partial_\mu\theta\partial^\mu\theta)\Bigr]\varphi\biggr\}
\label{3.5}$$
Now, the coefficient $[a_1]$, related to the operator that is acting on $\varphi$, is
$$[a_1]=-e^{2\theta}\,\Bigl(\frac{1}{6}R
+\partial_\mu\theta\partial^\mu\theta\Bigr)
\label{3.6}$$
So, the trace anomaly reads
$$\langle T^\mu\,_\mu\rangle=-\frac{e^{2\theta}}{4\pi}\,
\Bigl(\frac{1}{6}R+\partial_\mu\theta\partial^\mu\theta\Bigr)
\label{3.7}$$
Since the field $\theta$ is considered to be classical, one may infer that the expression for the energy momentum tensor $\langle\tilde T_{\mu\nu}\rangle$ is given by
$$\langle\tilde T_{\mu\nu}\rangle=\langle T_{\mu\nu}\rangle
-\frac{e^{2\theta}}{4\pi}\,
\Bigl(\frac{1}{12}\,g_{\mu\nu}R
+\partial_\mu\theta\partial_\nu\theta\Bigr)
\label{3.8}$$
where $\langle T_{\mu\nu}\rangle$ is the same one as given by (\[2.7\]), with $\phi$ replaced by $e^\theta\varphi$, i.e.
$$\begin{aligned}
&&\langle T_{\mu\nu}\rangle=e^{2\theta}\,
\Bigl[\Bigl(\partial_\mu\theta\partial_\nu\theta
-\frac{1}{2}g_{\mu\nu}\,\partial_\rho\theta\partial^\rho\theta
\Bigr)\langle\varphi^2\rangle
\nonumber\\
&&\phantom{\langle\tilde T_{\mu\nu}\rangle=e^{2\theta}\,}
+\frac{1}{2}\,\partial_\mu\theta\langle\partial_\nu\varphi^2\rangle
+\frac{1}{2}\,\partial_\nu\theta\langle\partial_\mu\varphi^2\rangle
\nonumber\\
&&\phantom{\langle\tilde T_{\mu\nu}\rangle=e^{2\theta}\,}
-\frac{1}{2}\,g_{\mu\nu}\,
\partial_\rho\theta\langle\partial^\rho\varphi^2\rangle
\nonumber\\
&&\phantom{\langle\tilde T_{\mu\nu}\rangle=e^{2\theta}\,}
+\langle\partial_\mu\varphi\partial_\nu\varphi
-\frac{1}{2}g_{\mu\nu}\,
\partial_\rho\varphi\partial^\rho\varphi\rangle\Bigr]
\label{3.9}\end{aligned}$$
Now, let us verify the consistency with respect the energy and momentum conservation. First, we consider $\nabla^\mu\langle T_{\mu\nu}\rangle$. This can be done by directly acting the operator $\nabla^\mu$ on $\langle\tilde T_{\mu\nu}\rangle$, given by the expression (\[3.9\]), or by using (\[2.9\]), where we should replace $\phi$ by $e^\theta\varphi$ \[notice, however, that this replacement cannot be done into (\[2.16\]), because just part of $\phi$ is quantum\]. We thus have
$$\begin{aligned}
&&\nabla^\mu\langle T_{\mu\nu}\rangle=e^{2\theta}\Bigl[
\bigl(\partial_\rho\theta\partial^\rho\theta+\Box\,\theta\bigr)\,
\partial_\nu\theta\langle\varphi^2\rangle
\nonumber\\
&&\phantom{\nabla^\mu\langle\tilde T_{\mu\nu}\rangle
=e^{2\theta}\Bigl[}
+\partial_\nu\theta\partial^\rho\theta\,
\langle\partial_\rho\varphi^2\rangle
+\partial_\nu\theta\langle\varphi\,\Box\,\varphi\rangle
\nonumber\\
&&\phantom{\nabla^\mu\langle\tilde T_{\mu\nu}\rangle
=e^{2\theta}\Bigl[}
+\frac{1}{2}\bigl(\partial_\rho\theta\partial^\rho\theta
+\Box\,\theta\bigr)\,\langle\partial_\nu\varphi^2\rangle
\nonumber\\
&&\phantom{\nabla^\mu\langle\tilde T_{\mu\nu}\rangle
=e^{2\theta}\Bigl[}
+2\partial^\rho\theta
\langle\partial_\rho\varphi\partial_\nu\varphi\rangle
+\langle\partial_\nu\varphi\,\Box\,\varphi\rangle\Bigr]
\label{3.10}\end{aligned}$$
For the second term of (\[3.8\]) we have
$$\begin{aligned}
&&-\frac{1}{4\pi}\,\nabla^\mu\,
\Bigl[e^{2\theta}\,\Bigl(\frac{1}{12}\,g_{\mu\nu}\,R
+\partial_\mu\theta\partial_\nu\theta\Bigr)\Bigr]
\nonumber\\
&&\phantom{-\frac{1}{4\pi}\,\nabla^\mu\,}
=-\frac{1}{4\pi}\,e^{2\theta}\,
\Bigl(\frac{1}{6}\,R\,\partial_\nu\theta
+\frac{1}{12}\,\partial_\nu R
\nonumber\\
&&\phantom{-\frac{1}{4\pi}\,\nabla^\mu\,=}
+2\partial_\nu\theta\partial_\rho\theta\partial^\rho\theta
\nonumber\\
&&\phantom{-\frac{1}{4\pi}\,\nabla^\mu\,=}
+\Box\,\theta\partial_\nu\theta
+\partial^\rho\theta\partial_\rho\partial_\nu\theta\Bigr)
\label{3.11}\end{aligned}$$
Now, since just $\varphi$ is quantum, we can write
$$\langle\partial_\nu\varphi\,\Box\,\varphi\rangle
=\frac{1}{2}\partial_\nu\langle\varphi\,\Box\,\varphi\rangle
\label{3.12}$$
Considering the action involving $\varphi$ and with the term $e^{2\theta}$ factorized, as well as the renormalization factor we are using into (\[2.14\]), we can obtain the regularized quantities
$$\begin{aligned}
&&\langle\varphi\,\Box\,\varphi\rangle
=\frac{1}{4\pi}\Bigl(\frac{1}{6}\,R
+\partial_\rho\theta\partial^\rho\theta\Bigr)
\label{3.13}\\
&&\langle\varphi^2\rangle=\frac{1}{4\pi}
\label{3.14}\end{aligned}$$
Finally, since $\varphi$ is a scalar quantity and $\langle\partial_\nu\varphi^2\rangle$ is an average involving all directions aleatory, and the same occurs with $\langle\partial_\rho\varphi\partial_\nu\varphi\rangle$, we have that these quantities are null. Replacing all these results into (\[3.10\]) and (\[3.11\]), we get
$$\nabla^\mu\langle\tilde T_{\mu\nu}\rangle=0
\label{3.15}$$
which express the consistency with the energy-momentum conservation.
In the case that $\theta$ is also quantum, the measure $[d\theta]$ does not change under conformal transformation, but the problem is much more involved and difficult to solve. Just formally, one may write that the trace anomaly reads
$$\langle\tilde T^\mu\,_\mu\rangle=-\frac{1}{4\pi}\,
\langle e^{2\theta}\Bigl(\frac{1}{6}R
+\partial_\mu\theta\partial^\mu\theta\Bigr)\rangle
\label{3.16}$$
From which one cannot either infer an expression similar as the second term of (\[3.8\]) or try to regularize it because the bad divergencies occuring in the exponential $e^{2\theta}$.
Conclusion
==========
In this paper we have study the problem of conformal anomaly and the energy-momentum conservation for a quantum scalar field interacting with a classical curved background in the spacetime dimension D=2. We have considered the scalar field as split in two other scalar fields, and kept just of them quantum. This procedure is in opposite direction to what is done in literature, where a factor containing the metric determinant is absorbed by the scalar field, leading to a new field with a convenient conformal transformation. We have shown that our procedure is consistent with the geometric terms of the usual treatment of the conformal anomaly and also consistent with the expected result that energy and momentum should be conserved after quantum effects are taken into account.
This work is supported in part by Conselho Nacional de Desenvolvimento Científico e Tecnológico - CNPq and Fundação Universitária José Bonifácio - FUJB (Brazilian Research Agencies) One of us, J.B.-N. has also the support of PRONEX 66.2002/1998-9.
[30]{} For a general review see N.D. Birrell and P.C.W. Davies, [*Quantum fields in curved space*]{}, Cambridge (1982), and references therein. K. Fujikawa, U. Lindström, N.K. Nielsen, M. Roček, and P van Nieuwenhuizen, Phys. Rev. [**D37**]{} (1988) 391. Fujikawa K [*Quantum Gravity and Cosmology*]{} ed H Sato and I Inami (Singapore: World Scientific)(1982). M. Alves and C. Farina, Class Quantum Grav. [**9**]{} (1992) 1841. K. Fujikawa, Phys. Rev. Lett. [**44**]{} (1980) 1733. See, for example, S. Hawking, Commun. Math. Phys. [**55**]{} (1977) 133 and references therein. B.S. DeWitt, [*Dynamical theory of groups and fields*]{}, Les Houches: Relativiry, groups and topology (1963). Editors: B.S. DeWitt and C. DeWitt; B.S. DeWitt, [*The spacetime approach to quantum field theory*]{}, Les Houches: Relativity, groups and topology (1983). Editors: B.S. DeWitt and R. Stora; S.M. Christensen, Phys. Rev. [**D14**]{} (1976) 2490. W.Kummer, H.Liebl and D.V.Vassilevich, Mod. Phys. Lett [**A 12** ]{} (1997) 2683; S. Hawking and R. Boussos, hep-th/9705236; J. S. Dowker, hep-th/9802029; S. Ichinoise and S.Odintsov, hep-th/9802043.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We explain some phenomena existing generally in the timing residuals: amplitude and sign of the second derivative of a pulsar’s spin-frequency ($\ddot\nu$), some sophisticated residual patterns, which also change with the time span of data segments. The sample is taken from Hobbs et al.(2010), in which the pulsar’s spin-frequency and its first derivative have been subtracted from the timing solution fitting. We first classify the timing residual patterns into different types based on the sign of $\ddot\nu$. Then we use the magnetic field oscillation model developed in our group [@zhang12a] to fit successfully the different kinds of timing residuals with the Markov Chain Monte Carlo method. Finally, we simulate the spin evolution over 20 years for a pulsar with typical parameters and analyze the data with the conventional timing solution fitting. By choosing different segments of the simulated data, we find that most of the observed residual patterns can be reproduced successfully. This is the first time that the observed residual patterns are fitted by a model and reproduced by simulations with very few parameters. From the distribution of the different residual patterns in the $P-\dot P$ diagram, we argue that (1) a single magnetic field oscillation mode exists commonly in all pulsars throughout their lifetimes; (2) there may be a transition period over the lifetimes of pulsars, in which multiple magnetic field oscillation modes exist.'
author:
- |
Xu-Dong Gao$^{1,2}$ [^1], Shuang-Nan Zhang$^{2,3}$ [^2], Shu-Xu Yi$^{3,4}$, Yi Xie$^{2,4}$, and Jian-Ning Fu$^{1}$\
$^{1}$ Department of Astronomy, Beijing Normal University, Beijing 100875, China\
$^{2}$ National Astronomical Observatories, Chinese Academy Of Sciences, Beijing 100012, China\
$^{3}$ Key Laboratory of Particle Astrophysics, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China\
$^{4}$ University of Chinese Academy of Sciences, Beijing 100049, China
title: Understanding the residual patterns of timing solutions of radio pulsars with a model of magnetic field oscillation
---
\[firstpage\]
stars: magnetic field, stars: neutron, (stars): pulsars: general
Introduction {#sec-introduction}
============
Radio pulsars are renowned for unique high-precision clocks in the universe driven by their stable rotation. The pulsar timing method provides a model of the pulsar’s astrometric, orbital and rotational parameters to compare with the observed pulse times of arrival (TOAs) [@manchester77; @lyne04; @lorimer05; @edwards06]. Timing residuals of a pulsar are defined as the differences between the predicted TOAs and the actual TOAs. However, radio pulsars exhibit two main timing irregularities, namely ‘glitches’ that are sudden increases in spin rate followed by a period of relaxation, and ‘timing noise’ that shows long-term and stochastic deviations from a regular spin-down model and mainly consists of low-frequency structures.
Timing noise can be characterized as a random walk in the rotational phase, angular velocity or torque and describes the un-modelled red noise in the observed TOAs [@cordes85; @shannon10], implying a process autocorrelated on a time-scale of hours to years [@cordes80]. Previously the analyses of timing noise have been limited by the relatively short data spans. Only a few papers have analysed a small number of pulsars for long data spans [@baykal99; @shabanova95; @stairs00; @shabanova01]. Most of the analyses are mainly about obtaining high-quality power spectral estimates of the timing residuals or fitting the timing residuals of a single pulsar with a simple model. Recently Hobbs et al.(2010) (hereafter H10) carried out so far the most comprehensive survey of timing irregularities in 366 pulsars over time-scales longer than a decade, which have characteristic ages larger than $10^{4}$ years. Based on the simple spin-down model, the second frequency derivative can be modelled by the cubic terms of a Taylor series as follows, $$\label{Phi0}
\Phi(t)= \Phi_{0}+\nu(t-t_{0})+\frac{1}{2}\dot\nu(t-t_{0})^{2}+\frac{1}{6}\ddot\nu(t-t_{0})^{3}+...,$$ where $\Phi_{0}$ is the phase at time $t_{0}$. However, values of $\ddot\nu$ obtained from timing fits are typically orders of magnitude larger than the prediction of the vacuum dipole model and usually have different signs. In the meantime, the braking indices determined by these derivatives have anomalous values range from $-2.6\times10^{8}$ to $+2.5\times10^{8}$. Therefore, Hobbs et al. argued that these cubic terms are not dominated by the intrinsic dipole braking of pulsars and only fit the pulse frequency and its first derivative and consider the remaining features as timing noise. Subsequently, They tried to categorize the pulsars based on the structures existing in their timing residuals. Nevertheless, they argued that two problems exist in their categorization, namely, the TOA precision achievable and the data spans available.
However, the physical processes behind timing noise have not been well explained. The phenomenon is attributed to various mechanisms. It has been suggested that the timing noise is dominated by the recovery from unseen glitches [@johnston99]. As the pulsar ages, glitch activity decreases and the timing noise is dominated by changes in the magnetosphere. Lyne et al.(2010) demonstrated that the observed timing noise correlates with the observed pulse shape, which indicates that some timing noise may be caused by magnetospheric processes. Some authors argued that timing noise may be driven by variability in the coupling between the crust and superfluid interior [@alpar86; @jones90] and fluctuations in the external spin-down torque [@cheng87a; @cheng87b; @urama06]. A more complete understanding of timing noise will provide us an insight into the interior structure of neutron stars.
An additional possibility is that the magnetic fields of pulsars decay with age. Baym et al. (1969) gave the first estimate of the characteristic decay time of their magnetic fields. Sang $\&$ Chanmugam (1987) argued that the field does not decay exponentially during the lifetime of the neutron stars. However, a complete theoretical model that could explain all observations does not exist yet, as the evolution of the magnetic field depend on its configuration that is still unknown and even the origin of the field is not clear until now. In general, the magnetic field is often assumed to occupy a substantial fraction of a neutron star’s volume and pass through its core or is confined to relatively not very deep layers. Recently, three physical processes were proposed to explain the evolution of the magnetic field inside a neutron star: (1) Hall drift, which is the advection of the magnetic field because of the motion of the free electrons; (2) Ohmic dissipation, which converts magnetic energy to heat because of the finite conductivity of the crust; (3) ambipolar diffusion, which is the interaction of the electric currents with the neutrons deeper inside a neutron star [@goldreich92]. On the other hand, population synthesis studies suggested that old pulsars show no significant magnetic decay over their lifetimes [@regimbau01; @faucher06], although the opposite conclusion has also been claimed [@gonthier04; @popov10]. These contradictions may be resolved by the assumption that the neutron star magnetic field is maintained by two current systems. Long living currents in the superconducting core support the large-scale dipolar field and are responsible for the spin down of old pulsars, but currents in the crust support the short-lived part of the field [@pons07].
The estimates of the magnetic field strength of neutron stars usually come from radio pulsars with measured spin-down rates. As the real ages of these pulsars are usually unknown, one can determine their spin-down age, $\tau_{c}=P/2\dot P$ , where $P$ is the spin period of the pulsar, as the indicators of the true ages of pulsars. However, Zhang $\&$ Xie (2011) showed that spin-down ages are normally significantly larger than the ages of the supernova remnants physically associated with them, which in principle should be the unbiased age indicators of the pulsars [@lyne75; @geppert99]. It may be evidence of the magnetic field decay over their lifetimes, as the decay can alter the spin-down rate of a pulsar significantly. To explain the observed periodic or quasi-periodic evolution of the spin of pulsars, it is natural to introduce some oscillation parameters. Zhang $\&$ Xie (2012a) proposed a magnetic field evolution model, which consists a long-term decay modulated by short-term oscillations and can explain the observed statistical properties of $\ddot\nu$ well.
In this paper, we will mainly explain the reason for amplitude and sign of second frequency derivative and the structure of timing residuals of a given pulsar varying with the time span of data segments. By contrasting with Hobbs’s opinion, we argue that timing residuals, which have subtracted pulsar’s spin frequency and its first derivative and are dominated by $\ddot\nu$ (i.e. Fig.3 in H10), are still due to the evolution of magnetic field of pulsars. The structure of the paper is as follows. In the next section, we introduce the sample selection of the pulsars, and classify the timing residuals based on the the sign of $\ddot\nu$ and the detailed structure of timing residuals. In section 3, we review magnetic field evolution model we developed previously and use Markov Chain Monte Carlo (MCMC) method to fit the timing residuals with a single oscillation component and two oscillation components of the model. Then we reclassify the timing residuals based on the number of oscillation components. In section 4, we discuss the evolution of pulsars through their distribution of different kinds of timing residuals in the $P-\dot P$ diagram and physical implication of the oscillation term. The summary of our results is given in section 5.
Classify the Timing Residuals {#sec-classify-residuals}
=============================
As we mentioned above, Hobbs et al.(2010) attempted to categorize pulsars based on the structures existing in their timing residuals: (1) 37% of the timing residuals are dominated by the measurement errors and show no features; (2) 20% have residuals that can be modelled by a significant positive $\ddot\nu$ value; (3) 16% have cubic terms that correspond to a negative $\ddot\nu$ value; and (4) 27% show more complicated structures. In the meantime, they showed that their simple categorization has two problems: (1) The TOAs of some pulsars are measured more precisely than the TOAs of others (see Fig.4 in H10); (2) The structure of timing residuals of any pulsar is likely to change as the data span varies (see Fig.5 in H10).
To better classify the timing residuals, we try to separate the pulsars that their timing residuals have detailed structures and those that have no structures in the residuals from the H10 sample. Figure 1 shows the distributions of $\sigma_{3}$ and $\sigma_{1}$ of all 366 pulsars taken from H10, where $\sigma_{1}$ represents the un-weighted rms of the residuals after fitting for $\nu$ and $\dot\nu$, and $\sigma_{3}$ represents the rms after whitening the data set by fitting and removing harmonically related sinusoids [@hobbs04; @hobbs10]. From these distributions alone we can not identify those pulsars with obvious structures in their timing residuals. In Figure 2 we plot the distribution of $\sigma_{3}/\sigma_{1}$ of all 366 pulsars we take from H10. We find that there are two distinctly different peaks in Figure 2; the pulsars in one peak that the values of $\sigma_{3}/\sigma_{1}$ are approximately 0 have distinct structures of timing residuals as shown in Figures 3 and 4, but in the other peak no distinct structure exists, which means that the residuals after subtraction of the pulsar’s $\nu$ and $\dot\nu$ are approximately white noise. Therefore, the pulsars in our sample are selected with $\sigma_{3}/\sigma_{1} < 0.4$, which have periodic or quasi-periodic structures in their timing residuals; the number of pulsars with $\sigma_{3}/\sigma_{1} > 0.4$ or $\sigma_{3}/\sigma_{1} < 0.4$ are 166 and 200, respectively. In Figure 5 we plot the distributions of $\sigma_{1}$ and $\sigma_{3}$ of the two samples; we can not distinguish the two samples based on $\sigma_{1}$ or $\sigma_{3}$ individually, since their main peaks overlap in the distributions.
![Distributions of $\sigma_{1}$ and $\sigma_{3}$ of all 366 pulsars, where $\sigma_{1}$ represents the unweighted rms of the residuals after fitting for $\nu$ and $\dot\nu$, and $\sigma_{3}$ represents the rms after whitening the data set by fitting and removing harmonically related sinusoids.[]{data-label="dis"}](Figure1.pdf){width="\columnwidth"}
![Distribution of $\sigma_{3}/\sigma_{1}$ of all pulsars. The number of $\sigma_{3}/\sigma_{1} < 0.4$ or $\sigma_{3}/\sigma_{1} >0.4$ are 200 and 166, respectively.[]{data-label="dis"}](Figure2.pdf){width="\columnwidth"}
![$W$ mode timing residuals. Each sub-panel represents the different subclass, for which the pulsar’s spin-frequency and its first derivative have been fitted and removed.[]{data-label="W-mode"}](Figure3.pdf){width="\columnwidth"}
![$M$ mode timing residuals. Each sub-panel represents the different subclass, for which the pulsar’s spin-frequency and its first derivative have been fitted and removed.[]{data-label="M-mode"}](Figure4.pdf){width="\columnwidth"}
![Distribution of $\sigma_{1}$ and $\sigma_{3}$ of two samples, respectively.[]{data-label="dis"}](Figure5.pdf){width="\columnwidth"}
We thus select the sample that have detailed structures to study the evolution of $\ddot\nu$ with their timing residuals. In the first step, we classify the timing residuals based on the sign of $\ddot\nu$. In addition, based on the structural tendency of timing residuals, we define the residuals of $\ddot\nu >0$ as $W$ mode (see Fig. \[W-mode\]). Conversely, we define the residuals of $\ddot\nu <0$ as $M$ mode (see Fig. \[M-mode\]). According to equation (\[Phi0\]), to the timing residuals after substraction of the pulsar’s spin-frequency and its first derivative, the overall structural tendency is dominated by the sign of $\ddot\nu$. That is to say, the overall shapes of these two modes (see Fig.3a and Fig.4a respectively) are the basic shape of a cubic polynomial after removal of $\nu$ and $\dot\nu$. Therefore, these features are inevitable results of mathematics. However, there is still no self-consistent theory to explain the reason for the sign of $\ddot\nu$. Besides, there exist some residuals in both modes that show more complicated structures beyond the simple mathematics. Therefore, based on the numbers of local extrema, the $W$ mode residuals are further divided into $Wa$, $Wb$, $Wc$ and $Wd$ subclasses corresponding to the residual patterns with two, three, four and more local extreme points (see Fig.3a, 3b, 3c and 3d which have the typical characteristic structures in the timing residuals for the different subclasses among the 200 pulsars in the sample of $\sigma_{3}/\sigma_{1} < 0.4$), respectively. Similarly, the $M$ mode residuals are further divided into $Ma$, $Mb$ and $Mc$ subclasses corresponding to the residual patterns with two, three and four local extreme points (see Fig.4a, 4b and 4c).
In Table 1 we present the numbers of the different subclasses in our sample. In Figure 6 we plot the $P-\dot P$ diagram for our sample and all the observed pulsars taken from the ATNF Pulsar Catalogue [@manchester05]. From the distribution of our sample we see that it covers radio pulsars, millisecond pulsars and high energy pulsars. In the following section, we will attempt to use phenomenological model to reproduce these timing residuals, find the physical process that leads to the phenomena, i.e., timing residuals generally includes positive and negative values of $\ddot\nu$ and some of them have sophisticated structures. Then we will further explain the reason for the structure of timing residuals of a given pulsar varying with the time span of the data analyzed.
$Type$ $Wa$ $Wb$ $Wc$ $Wd$
---------- ------ ------ ------ ------
$Number$ 60 13 22 6
$Type$ $Ma$ $Mb$ $Mc$
$Number$ 63 19 7
: Number of the corresponding types in our sample[]{data-label="typenum"}
![The top panel shows the distribution of pulsars in our sample and all other pulsars taken from the ATNF Pulsar Catalog [@manchester05] in the $P-\dot P$ diagram. The triangle represents the $W$ mode pulsars, the $M$ mode pulsars are marked with star, and different colors represent the different subclasses. The lower panel is the enlarged view of the distribution of pulsars in our sample; the dotted and dashed lines represent constant characteristic age and constant magnetic field, respectively.[]{data-label="p-pdot"}](Figure6.pdf){width="\columnwidth"}
Magnetic Field Decay and Oscillation Model {#sec-model}
==========================================
For the case of a neutron star in vacuum, due to the asymmetry of magnetic field about its rotation axis, it radiates energy and losses angular momentum. Assuming pure magnetic dipole radiation dominates over the gravitational quadrupole radiation as the braking mechanism for a pulsar’s spin down, we have the magnetic-dipole model [@pacini67; @pacini68; @ostriker69] $$\label{dipole}
\dot E_{\rm rot}=I\Omega \dot \Omega=-\frac{2}{3c^{3}}{|\ddot m|^{2}}=-\frac{2(BR^3)^2{sin^2\alpha}}{3c^3}{\Omega ^4},$$ or $$\label{other}
\dot\nu = -AB^2\nu^3,$$ where $A=\frac{8\pi^{2}R^{6}{\sin^2\alpha}}{3c^3I}$, $m$ is the magnetic dipole moment, $B$ is the strength of the dipole magnetic field at its equatorial surface, $R$ denotes the radius of neutron star, $\alpha$ is the inclination of the magnetic axis with respect to the rotation axis, and $I$ is its moment of inertia. Deriving $\ddot{\nu}$ from equation (\[other\]), we have $$\label{nu_ddot1}
\ddot \nu = 3\dot{\nu}^2/\nu +2\dot{\nu}\dot{B}/B.$$ For pulsars with characteristic ages larger than $10^{5}$yr in H10 sample, almost equal numbers of pulsars have each sign of $\ddot\nu$. According to equation (\[nu\_ddot1\]), this phenomenon can be caused by the evolution of magnetic field of second term. However, by considering only monotonic change of magnetic field or torque cannot explain this phenomenon. That is also the reason why other authors suspect that the cubic terms are not due to the intrinsic dipole braking of pulsars [@hobbs04].
However, we have constructed a phenomenological model for the evolution of magnetic field $B$, which contains a long-term power-law decay modulated by short-term oscillations [@zhang12a; @zhang12b], $$\label{b_decay}
B(t)= B_{\rm L}(t)(1+\sum k_{i}\sin(\phi_{i}+2\pi\frac{t}{T_{i}})),$$ where $t$ represents the pulsar’s age in arbitrary units, $k_{i}\ll1$, $\phi_{i}$, $T_{i}$ are the amplitude, phase and period of the $i-th$ oscillation component, respectively. Assuming $B_{\rm L}=B_0 ((t+t_{0})/t_{0})^{-\alpha}$, in which $B_0$ is the field strength at age $t_0$, i.e. at the epoch $t=0$, and $\alpha$ is the index of the power law decay. From equation (\[b\_decay\]), we can obtain the analytic approximation for $\ddot\nu$ (see Zhang $\&$ Xie 2012a Eq. 21), $$\label{ddot_p2}
\ddot{\nu}(t)\simeq 2\dot{\nu}(t)(-\frac{\alpha}{t}+\sum f_i\cos(\phi_i+2\pi\frac{t}{T_i})),$$ where $f_i={2\pi k_i/T_i}$, $t$ is the real age of the pulsar. We can see that the second term of equation (\[ddot\_p2\]) dominates the sign of $\ddot\nu$, if $|\frac{\alpha}{t}|<f_i$. Here we assume that the decay index $\alpha =0.5$ is constant, the value of $k$ and $T$ range from $10^{-5}$ to $10^{-2}$ and from 2 yr to 100 yr, and $f_{\rm max}=1.0\times 10^{-9}$ can be derived [@zhang12a]. Therefore, only to the pulsars younger than about 1000 yr, the first term of equation (\[ddot\_p2\]) can dominate the sign of $\ddot\nu$. Since the ages of the pulsars in our sample are all more than $10^{4}$ yr, the first term of equation (\[ddot\_p2\]) can be neglected, especially over the short-term evolution. In our phenomenological model, therefore, the sign and magnitude of $\ddot\nu$ can be well described and the structure of timing residuals reflects the oscillation of its magnetic field. Furthermore, for constant $k$ and $T$ (i.e. periodic oscillation), the key factor that influences the sign of $\ddot\nu$ is the variation of the phase of oscillation component.
To further verify our magnetic oscillation model, we try to use it to simulate the observed timing residuals. Since $(t/t_0)$ and $k$ are small quantities, $$\begin{aligned}
\centering
B^2(t)&\simeq& B_0^2(1-2\alpha\frac{t}{t_0})(1+\sum2k_i\sin(\phi_i+\omega_i t)) \nonumber\\
&\simeq&B_0^2\left(1-2\alpha\frac{t}{t_0}+\sum2k_i\sin(\phi_i+\omega_i t)\right).\nonumber\end{aligned}$$
Then, integrating equation (\[other\]) and combining with equation (\[b\_decay\]), we can obtain the expression, $$\label{second}
\frac{1}{\nu^2}=\frac{1}{\nu_0^2}+2AB_0^2\left(t-\frac{\alpha}{t_0}t^2-\sum\frac{2k_i}{\omega_i}\cos(\phi_i+\omega_i t)\right),$$ where $\nu_0=\nu(t=t_0)$ and $\omega_i = 2\pi/T_i$.
From equation (\[other\]) we can obtain $-AB^2\nu^2=\dot{\nu}/\nu$. $-AB^2\nu^2$ is a small quantity, since $\Delta\nu/\nu\ll1$ during an observation span. Therefore from equation (\[second\]), we get, $$\label{third}
\nu=\nu_0-AB_0^2\nu_0^3\left(t-\frac{\alpha}{t_0}t^2-\sum\frac{2k_i}{\omega_i}\cos(\phi_i+\omega_i t)\right).$$
Integrating equation (\[third\]) gives the expected pulse phase, $$\label{phifan}
\Phi_{m}(t)=\Phi_0+\nu_0t+AB_0^2\nu_0^3(\frac{\alpha}{3t_0}t^3-\frac{t^2}{2}+\sum\frac{2k_i}{\omega_i^2}\sin(\phi_i+\omega_i t)).$$
If we define $\dot{\nu}_0\equiv-AB_0^2\nu_0^3$, $\ddot{\nu}_0\equiv2AB_0^2\nu_0^3\alpha/t_0$, $F_i\equiv2k_iAB_0^2\nu_0^3/\omega_i^2$, then equation (\[phifan\]) can be rewritten as, $$\label{six}
\Phi_{m}(t)=\Phi_0+\nu_0t+\frac{\dot{\nu}_0t^2}{2}+\frac{\ddot{\nu}_0t^3}{6}+\sum F_i\sin(\phi_i+\omega_i t).$$ From equation (\[six\]) we can find that the pulse phase can be written as a polynomial to the third order, plus several sinusoidal waves. One thing we would like to emphasize here is that $\ddot{\nu}_0$ we defined in equation (\[six\]) is not equivalent to our derived instantaneous $\ddot{\nu}(t)$ in equation (\[ddot\_p2\]), neither to the averaged $\bar{\ddot{\nu}}$ from conventional fitting observed data span with the third order polynomial of Taylor expansion. The power-law decay of magnetic field always results in $\ddot{\nu}_0>0$, since the decay index $\alpha >0$. In Section 2, the classification of timing residuals of our sample is based on the value of practical fitting averaged $\bar{\ddot{\nu}}$.
In practical procedure of pulsar timing, fitting a polynomial to the third order without sinusoidal waves is a common treatment:
$$\Phi_{\rm{fit}}(t)=\Phi_0+\bar{\nu}t+\frac{\bar{\dot{\nu}}}{2}t^2+\frac{\bar{\ddot{\nu}}}{6}t^3.
\label{taylor}$$
By comparing between equations (\[six\]) and (\[taylor\]), we can obtain, $$\begin{aligned}
\bar{\nu}&=\nu_0+\sum_{i}\xi_{1i}(\phi_{i},\omega_{i}\tau)\omega_{i}F_{i},
\\
\bar{\dot{\nu}}&=\dot{\nu}_0+\sum_{i}\xi_{2i}(\phi_{i},\omega_{i}\tau)\omega_{i}^{2}F_{i},
\\
\bar{\ddot{\nu}}&=\ddot{\nu}_0+\sum_{i}\xi_{3i}(\phi_{i},\omega_{i}\tau)\omega_{i}^{3}F_{i},
\\
\end{aligned}\label{threeterms}$$ where $\xi_{1i}$, $\xi_{2i}$ and $\xi_{3i}$ correspond to the coefficients of the first, second and third order terms of polynomial fitting of $\sin(\phi_{i}+\omega_{i}t)$, respectively, which are all functions of time span $\tau$, since: $$\sin(\phi_{i}+\omega_{i}t)\rightarrow \xi_{1i}(\phi_{i},\omega_{i}\tau)\omega t+\frac{\xi_{2i}(\phi_{i},\omega_{i}\tau)}{2}\omega^{2}t^2+\frac{\xi_{3i}(\phi_{i},\omega_{i}\tau)}{6}\omega^{3}t^3\cdots$$ Therefore, from equation (\[threeterms\]) we can see that $\bar{\ddot{\nu}}$ is consisted of $\ddot{\nu}_0$ and the third order terms of polynomial fitting of oscillation components. But instantaneous $\ddot{\nu}(t)$ includes the contributions of power-law decay and all the oscillation components at every moment. Only if there is no oscillation in the magnetic field, i.e., $F_i=0$, then $\ddot{\nu}(t)\simeq \bar{\ddot{\nu}}=\ddot{\nu}_0$
Simulations and Fitting Observations {#sec-model}
====================================
Therefore, the timing series of phases $\Phi_{s}(t)$ of the simulated TOAs can be obtained based on equation (\[six\]). Then, we fit the simulated $\Phi_{s}(t)$ by the second order of Taylor expansion over the observation time span $\tau$: $$\label{Phi}
\Phi_{s}(t)= \Phi_{s0}+\bar{\nu_{0}}(t)+\frac{1}{2}\bar{\dot\nu}t^{2}.$$ The simulated timing residuals can be obtained by, $$\label{res}
R_{\rm sim}(t_{i})=\frac{\Phi_{s}(t_{i})-\Phi_{m}(t_{i})}{\nu_0}.$$
To extract the best model parameters by comparing with observations, we use [emcee]{} code from [@foreman13], which is an affine invariant ensemble sampler for MCMC and designed for Bayesian parameter estimation, to sample the full parameter space. Based on Bayes theorem, assuming $\bmath{p}$ and $\bmath{d}$ represent the model parameters and the data(t, residuals, errors.), respectively. Therefore, the posterior probability function can be written as, $$\label{posterior}
\mathrm{P}(\bmath{p} \, | \, \bmath{d}) \propto
\mathrm{P}(\bmath{d} \, | \, \bmath{p}) \, \mathrm{P}(\bmath{p}),$$ where $\mathrm{P}(\bmath{d} \, | \, \bmath{p})$ is the likelihood function, $\mathrm{P}(\bmath{p})$ is the prior function.
Here we define the likelihood function as a simple Gaussian, $$\mathrm{P}(\bmath{d} \, | \, \bmath{p}) = \exp\left(-\frac{\chi^2}{2}\right),$$ with $$\chi^2 = \sum_{i=1}^{N_{\rm{obs}}} \left(\frac{R_{\rm re}(t_{i})-R_{\rm sim}(t_{i})}{ \sigma_{3} }\right)^2,$$ where $N_{\rm{obs}}$ denotes the number of the reported timing residuals, $R_{\rm re}(t_{i})$ and $R_{\rm sim}(t_{i})$ represent the reported timing residuals and the simulated timing residuals at $t_{i}$, respectively. The error of each point of timing residual corresponds to the value of $\sigma_{3}$.
In prior function($\mathrm{P}(\bmath{p})$), we need to give physically acceptable ranges of our parameters. Here we set priors within given bounds:
\(i) $\omega_{\rm{min}}<\omega_{i}<\omega_{\rm{max}}$, where $\omega_{\rm{min}}=2.0\times10^{-9}\,\rm{s^{-1}}$, $\omega_{\rm{max}}=1.0\times10^{-7}\,\rm{s^{-1}}$;
\(ii) $2k_{\rm{min}}AB_0^2\nu_0^3/\omega_{\rm{max}}^2<F_{i}<2k_{\rm{max}}AB_0^2\nu_0^3/\omega_{\rm{min}}^2$, where $k_{\rm{min}}=10^{-5}$, $k_{\rm{max}}=10^{-2}$ and $\nu_0$ corresponds to the initial value of a pulsar;
\(iii) $\phi_{i}$ is between 0 and $2\pi$.
After all these have been prepared, we obtain the best fitting parameters of our model with the following steps. (1)We fit the $R_{\rm sim}(t_{i})$ to the $R_{\rm re}(t_{i})$ preliminarily to find the maximum likelihood. To each parameter, we start by initializing 100 walkers in a tiny Gaussian ball around the maximum likelihood result; (2)We sample the parameter space according to the $\mathrm{P}(\bmath{p} \, | \, \bmath{d})$ and regard the set of parameters for which the $\mathrm{P}(\bmath{p} \, | \, \bmath{d})$ is maximized as our best model, since the parameter set can reproduce the $R_{\rm re}(t_{i})$ most closely; (3)We can get marginalized distribution for each parameter independently in one-dimension histograms and two dimensional projections of posterior probability distributions(In Appendix A, we present the two-dimensional marginalized distributions of all the parameters of the seven typical pulsars); (4)We choose the medians of the one-dimension posterior distributions as the best fitting parameters, and the parameter uncertainties are calculated with their 68 percent confidence intervals.
Specifically, the number of parameters in our model is determined by the number of oscillation parameters of magnetic field. Figures 7 and 8 show the corresponding power spectra of the typical pulsars in Figures 3 and 4, respectively. The number of main peaks in the power spectrum equals the number of oscillation components in our magnetic field evolution model that we assume initially. Then we give a set of initial values of the oscillation parameters $k_{i}$, $T_{i}$ and $\phi_{i}$ in equation (\[b\_decay\]) to obtain the timing residuals to be fitted, as well as the initial values of $\nu_0$, $\dot{\nu}_{0}$ and $\ddot{\nu}_0$ and the observation time span $t_s$. The values of $\nu_0$ and $\dot{\nu}_{0}$ are selected from H10 to each pulsar. We keep them constant, since they can hardly change during the short-term evolution. Here we set the power-law index $\alpha =0.5$ and $t_{0}$ as characteristic age to determine the value of $\ddot{\nu}_0$ in equation (\[six\]), because the short observation time span can not really constrain the monotonic evolution of the magnetic fields of these pulsars. (We will further discuss if the long-term monotonic evolution of magnetic field would influence the patterns of timing residuals in such short-term observations in section 5.1.) Based on equation (\[res\]), to $Wa$, $Wb$, $Ma$ and $Mb$ types, three parameters are needed to be sampled by MCMC($F$, $\omega$ and $\phi$). To $Wc$, $Wd$ and $Mc$ types, which include two oscillation component of magnetic field, six parameters should be sampled by MCMC($F_{1}$, $\omega_{1}$, $\phi_{1}$, $F_{2}$, $\omega_{2}$ and $\phi_{2}$). After getting the best fitting parameters, we convert them to the parameters we need, i.e., $k_{i}$, $T_{i}$, $\phi_{i}$, and $f_{i}$.
![Power spectra of B1855+02, B1829-08, B1736-31 and B1826-17.[]{data-label="W-pwd"}](Figure7.pdf){width="\columnwidth"}
![Power spectra of B0114+58, B1845-01 and B2227+61.[]{data-label="M-pwd"}](Figure8.pdf){width="1\columnwidth"}
Figures 9 and 10 show the comparisons between the reported and fitting timing residuals for those several typical pulsars. In the framework of magnetic field oscillation model, we can reproduce almost all types of timing residuals except the $Wd$ type. The $Wd$ type timing residuals of B1826-17 contain three main peaks in its power spectrum; however, we only use two oscillation components to fit the timing residuals to simplify the problem.
![Comparisons between our model fitting timing residuals and the corresponding observed timing residuals.[]{data-label="W-fit"}](Figure9.pdf){width="\columnwidth"}
![Comparisons between our model fitting timing residuals and the corresponding observed timing residuals.[]{data-label="M-fit"}](Figure10.pdf){width="\columnwidth"}
In Tables 2 and 3 we present the best fitting parameters which are converted from $F_{i}$, $\omega_{i}$ and $\phi_{i}$, corresponding to the different kinds of timing residuals. Combining equation (\[ddot\_p2\]) and equation (\[nu\_ddot1\]), $\sum{f_{i}}\simeq \ddot{\nu}(t)/2\dot{\nu}(t)=\dot{B}/B$, which can reflect the required fractional variation of the magnetic field in the short-term evolution. Figures 11 and 12 show the comparisons between the power spectra of the fitting timing residuals and the reported timing residuals of typical pulsars.
$Name$ $Type$ $k\times10^{-4}$ $T$(yr) $\phi$ $f\times10^{-12}$
---------- -------- --------------------------- ------------------------- ----------------------------- --------------------------- --
B1855+02 $Wa$ $5.83_{-0.18}^{+0.20}$ $64.48_{-0.43}^{+0.42}$ ${4.943}_{-0.004}^{+0.004}$ $1.803_{-0.069}^{+0.072}$
B1829-08 $Wb$ $0.673_{-0.003}^{+0.004}$ $13.22_{-0.02}^{+0.02}$ ${0.379}_{-0.005}^{+0.005}$ $1.015_{-0.007}^{+0.006}$
B0114+58 $Ma$ $1.61_{-0.07}^{+0.07}$ $38.32_{-0.30}^{+0.29}$ ${5.30}_{-0.02}^{+0.02}$ $0.84_{-0.04}^{+0.04}$
B1845-01 $Mb$ $23.9_{-3.5}^{+4.2}$ $31.93_{-0.85}^{+0.95}$ ${6.12}_{-0.05}^{+0.05}$ $14.89_{-2.54}^{+3.10}$
$Name$ B1736-31 B1826-17 B2227+61
--------------------- ----------------------------- ----------------------------- -----------------------------
$Type$ $Wc$ $Wd$ $Mc$
$k_1\times10^{-3}$ $1.03_{-0.01}^{+0.01}$ $0.68_{-0.01}^{+0.01}$ $6.57_{-2.75}^{+4.73}$
$T_1$(yr) $16.47_{-0.04}^{+0.04}$ $2.829_{-0.003}^{+0.003}$ $64.69_{-6.43}^{+8.50}$
$\phi_1$ $5.043_{-0.007}^{+0.006}$ $0.27_{-0.02}^{+0.02}$ $5.77_{-0.21}^{+0.23}$
$f_1\times10^{-11}$ ${1.249}_{-0.018}^{+0.016}$ ${4.808}_{-0.075}^{+0.076}$ ${2.023}_{-0.967}^{+1.841}$
$k_2\times10^{-5}$ $263.4_{-0.5}^{+0.3}$ $9.81_{-0.76}^{+0.82}$ $179.4_{-90.0}^{+158.1}$
$T_2$(yr) $7.210_{-0.002}^{+0.003}$ $21.70_{-0.32}^{+0.39}$ $18.98_{-2.10}^{+1.77}$
$\phi_2$ $5.076_{-0.002}^{+0.002}$ $5.86_{-0.03}^{+0.04}$ $6.19_{-0.40}^{+0.31}$
$f_2\times10^{-11}$ $7.278_{-0.014}^{+0.013}$ $0.090_{-0.008}^{+0.009}$ $1.884_{-1.026}^{+2.098}$
: Best fitting parameters of the double oscillation component of magnetic field[]{data-label="parameters2"}
![Comparisons between the power spectra of our model fitting timing residuals and the corresponding observed timing residuals.[]{data-label="W-fit-pwd"}](Figure11.pdf){width="\columnwidth"}
![Comparisons between the power spectra of our model fitting timing residuals and the corresponding observed timing residuals.[]{data-label="M-fit-pwd"}](Figure12.pdf){width="\columnwidth"}
Reclassify the Timing Residuals {#sec-reclassify}
===============================
As we pointed out in the previous section, to each oscillation component, the sign of $\bar{\ddot\nu}$ is dominated by the variation of phase, which is an oscillation parameter in our magnetic field oscillation model; namely, the sign of $\bar{\ddot\nu}$ is mainly determined by the observation time. In the meantime, our magnetic oscillation model have reproduced all the typical timing residuals well. Subsequently, we will further attempt to explain the sign of $\bar{\ddot\nu}$ varying with the time span of data analyzed so as to influence the structure of timing residuals. The different types of timing residuals we defined may be caused by the different stages of magnetic field oscillation.
Long-term Monotonic Evolution of Magnetic Field
-----------------------------------------------
Here we only discuss two cases of long-term monotonic evolution of magnetic field for simplicity. We first explore if the exponential decay can influence the patterns of timing residuals, although Zhang & Xie(2012a) have ruled out the exponential decay as the long-term monotonic evolution of the magnetic fields for the sample of H10. We first assume an exponential field decay $B_{\rm D}{(t)}=B_{\rm 0}e^{-t/\tau_{\rm D}}$, where the timescale of decay $\tau_{\rm D}=10\,\rm {Myr}$ and $B_{\rm 0}$ is the strength of the surface dipole magnetic field at time $t_{\rm 0}=0$, the initial parameters $\bar{\nu}=9.858\,\rm s^{-1}$, $\bar{\dot\nu}=-5.96\times 10^{-13}\,\rm s^{-2}$, and the characteristic age is $2.62\times 10^{5}\,\rm{yr}$; the observation time span we simulate is $\tau=20\,\rm {yr}$. We plot the timing residuals by extracting the first $10$ yr of the simulated observational data segment that is obtained by numerical computation just like in the previous section. Then we extract the data segment from the $1\rm st$ year to the $11\rm th$ year of the simulated observation to plot the timing residuals; the time step is one year and other data segments are taken in a similar fashion, until we extract the data segment from the $11\rm th$ year to the $20\rm th$ year. The upper panel of Figure 13 shows the timing residuals of all the data segments extracted. To the short term observation of the pulsar, we can see that the structures and the magnitude of all the timing residuals are exactly the same during the same observation time interval. The sign and magnitude of $\bar{\ddot\nu}$ do not change, because the observation time span and the structure of timing residuals are all the same.
{width="70.00000%"}
We then assume a power-law field decay $B_{\rm L}{(t)}=B_0 (t+t_{0}/t_{0})^{-\alpha}$, where the decay index $\alpha=0.5$, and $B_{\rm 0}$ is the strength of surface dipole magnetic field at time $t_{\rm 0}=0$. The initial parameters and observation time span are kept the same as the conditions of the exponential decay. The timing residuals are obtained also by the same operation steps as in the previous subsection. The lower panel of Figure 13 shows the timing residuals of all the data segments extracted. The result is nearly the same as for the exponential decay, namely, the structure of timing residuals and the sign of $\bar{\ddot\nu}$ are all independent from the different observation stages in the short term observation. Therefore, the long-term monotonic evolution of magnetic field can not influence the patterns of timing residuals during short-term observations.
Periodic Oscillations
---------------------
We use our magnetic field oscillation model in equation (\[b\_decay\]) and adopt the same initial parameters and observation time as before, and assume a single oscillation component, corresponding to the oscillation parameters $k=10^{-4}$, $T=10$ yr and $\phi=0$, respectively. The timing residuals are obtained also by the same steps above. Figure 14 shows the timing residuals of all the data segments extracted. These timing residuals include $Wa$, $Wb$, $Ma$ and $Mb$ types defined in section 2. The correlation between the sign of $\bar{\ddot\nu}$ and different types of timing residuals is the same as that in observations. From the simulation we also find that, the pulsar keeps $Wa$ and $Ma$ types for most of the time during the periodic oscillation. This result agrees with the fact that $Wa$ and $Ma$ types occupy the majority of all the timing residuals in our sample. This indicates that these four types of timing residuals should be the different oscillation stages of the single oscillation component of magnetic field.
As a consequence, the long-term monotonic decay of magnetic field has no impact to the structure of the timing residuals, but short-term oscillations can explain the observations well. Therefore, the different structures of timing residuals of observations and the corresponding sign of $\bar{\ddot\nu}$ are determined by the observation time, namely, the variation of the oscillation phase $\phi$ of magnetic field. Obviously, our previous classification of timing residuals only according to the tendency of structure does not have physical significance. Hereafter, we reclassify the previously defined timing residuals of $Wa$, $Wb$, $Ma$ and $Mb$ types, which are all reproduced by a single oscillation component of magnetic field, as Single-component Residuals ($SR$) mode; the other types which include double or multiple oscillation components are defined as Multi-component Residuals ($MR$) mode. Since the timing residuals that include double or multiple oscillation components are complex, we do not simulate these situations in this paper.
{width="70.00000%"}
Physical Implications of $MR$ mode {#sec-discuss}
==================================
Figure 15 shows the distribution of the pulsars of the reclassified timing residuals in the $P-\dot P$ diagram. For the $SR$ mode pulsars, their distribution covers almost all the evolution stages of pulsars, including normal radio pulsars, millisecond pulsars and high energy pulsars. However, the distribution of $MR$ mode pulsars are mainly concentrated in the region of normal radio pulsars.
![The distribution of the reclassified timing residuals in the $P-\dot P$ diagram. $SR$ mode pulsars are marked with triangles, and the circles represent $MR$ mode pulsars. The dashed lines and dotted lines represent constant magnetic field and constant characteristic age, respectively. The Transition Region is confined by the the pink dotted lines and pink dashed lines.[]{data-label="p-pdot-new"}](Figure15.pdf){width="\columnwidth"}
To examine if these two samples are derived from the same distribution, we make the two-dimensional Kolmogorov-Smirnov (K-S) test; the returned p-value is 0.13, which means these two samples unlikely originate from the same distribution. Therefore, we speculate that the physical processes that show a single oscillation form exist over the whole evolution of pulsars all along. However, as the neutron stars evolve to normal pulsars, there may exist a transition region ($TR$), where double or multiple oscillation components could appear in some pulsars. Afterwards the single oscillation mode dominates again, accompanied with the evolution of pulsars. $TR$ is shown in Figure 15 with the pink dashed lines and dotted lines, where the magnetic field is from $3\times 10^{11}\ \rm G$ to $3\times 10^{12}\ \rm G$ and the characteristic age is from $10^{5}$ yr to $10^{7}$ yr. The $TR$ phenomenon may be indirect evidence of long term magnetic field evolution of pulsars.
In both isolated and accreting neutron stars, the Hall effect and Ohmic decay are both very important to the long term magnetic field evolution [@cumming14]. Since the Hall drift conserves magnetic energy, it cannot be a direct cause of magnetic field decay. However, Jones (1988) first demonstrated that Hall drift could transport magnetic field from the inner crust to the outer crust, where Ohmic dissipation is much more effective than in the inner crust. Hall drift can also pump energy from an internal strong toroidal field to the poloidal component, resulting in increase of magnetic field [@pons12; @gourgouliatos14]. It indicates that Hall drift may produce damped oscillation with long period between toroidal and poloidal field components [@marchant14]. Cumming et al. (2004) proposed that the Hall effect dominates over ohmic decay after the time $t_{\rm switch}\simeq10^{4}B_{12}^{-3}$ yr, when the magnetization parameter, $R_{B}=\frac{t_{\rm Ohm}}{t_{\rm Hall}}>1$, where $t_{\rm Ohm}$ and $t_{\rm Hall}$ represent the timescales of Ohmic decay and Hall effect, respectively. The Hall timescale at lower densities of the crust is [@cumming14] $$\label{Hallout}
\tau_{\rm Hall,~Outer}=\frac{5.7\times 10^4~{\rm
yr}}{B_{12}}\rho_{12}^{5/3}(\frac{Y_e}{0.25})^{11/3}(\frac{g_{14}}{2.45})^{-2},$$ where $\rho_{12}=\rho/10^{12}~{\rm g~cm^{-3}}$, $Y_e$ is the number fraction of electrons, and $g_{14}$ is the local gravity, assumed constant.
Gourgouliatos $\&$ Cumming (2014) investigated the evolution due to Hall drift using different initial conditions, and found that eventually the field evolves towards a similar configuration defined as ‘attractor’, which consists of a dominant dipolar poloidal field coupled to a octopole through a weak toroidal quadrupole. Recently Marchant et al. (2014) did a similar numerical simulation about the effect of Hall drift to magnetic field evolution in neutron stars. The initial condition $E_{p}/E=0.9$ represents the ratio of poloidal to total energy, namely the poloidal component contains a large amount of energy, and $R_{B}=100$ means that Hall drift dominates the evolution. From the result of the numerical simulation, we find that the current associated to the toroidal field drags the poloidal field lines closer to one of the poles, since the bending of the poloidal field lines changes the orientation of the toroidal field; therefore, the poloidal field lines are dragged to the opposite pole. This stable oscillation is repeated until the evolution is dominated by Ohmic dissipation (see Fig.1 in Marchant et al 2014). From Fig. 1 in Marchant et al. (2014) we estimate the beginning of the oscillation is nearly $\tau_{\rm Hall}$; after approximately 100 $\tau_{\rm Hall}$, the oscillation vanishes when it reaches the attractor state. According to equation (\[Hallout\]), we estimate that the evolution time is from $0.6\times 10^{5}$ yr to $0.6\times 10^{7}$ yr, consistent with the $TR$ defined above. It is reasonable to speculate that this stage of oscillations between the poloidal and toroidal components is the main stage to produce the $MR$ mode timing residuals. We thus further argue that the evolution of the modes of timing residuals may have a relation with the long term evolution of magnetic field configurations.
Summary and Conclusion {#sec-summary}
======================
In this paper, we separate the pulsars that have more detailed structure in the timing residuals than others from the H10 sample as our study sample ($\sigma_{3}/\sigma_{1} < 0.4$). According to the sign of $\bar{\ddot\nu}$, we have classified the timing residuals as $W$ mode and $M$ mode, which are further subdivided into several kinds of subclasses based on the structure of timing residuals (Fig.3 and Fig.4). Then we used our phenomenological model to fit the timing residuals of the typical pulsars by the MCMC method respectively (Fig.9 and Fig.10) and obtained the best fitting oscillation parameters to all the types of timing residuals. After that, we simulated the short term observation of an individual pulsar with different magnetic field evolution models. Furthermore, we reclassified the timing residuals as $SR$ mode and $MR$ mode based on the number of oscillation components of magnetic field. We also analyzed the physical implications from the distribution of timing residuals in the $P-\dot P$ diagram and the physical mechanisms causing the observational behaviors of timing residuals. Our main results and conclusions are summarized as follows.
1\. Our magnetic field oscillation model can reproduce the general behaviors of timing residuals well.
2\. The variation of the observed $\bar{\ddot\nu}$ and the detailed structure of timing residuals reflect the oscillation of the magnetic field of pulsars.
3\. We rule out long-term monotonic evolution of magnetic field as the underlying mechanism for the observational timing residual patterns.
4\. The different structures of timing residuals of observation and the corresponding sign of $\bar{\ddot\nu}$ are determined by the observation time, namely, the variation of oscillation phase $\phi$ of magnetic field.
5\. A single magnetic field oscillation mode exists generally over the whole evolution of pulsars, while $MR$ mode only exists in the region of normal radio pulsars.
6\. The $SR$ mode and $MR$ mode may transit from each other with the evolution stage of magnetic field configurations.
In this work we only fit the timing residual patterns of seven typical pulsars in the sample of 200 pulsars with obvious timing residual patterns, in order to understand the underlying mechanism of the observational residual patterns and their variations with observation time. In our next work, we will fit the data for all other pulsars to understand the statistical properties of the parameters in our model.
If we can understand better about the behavior of timing residuals, we may be able to improve the sensitivity of using pulsars to detect gravitational waves and explore the physical processes in the interiors of neutron stars. We suggest that our simulation results could be confirmed by using the existing data and predict that the structure of timing residuals should vary along with different segments of data. In doing so, we could further understand the behavior of magnetic field oscillation of pulsars.
Acknowledgements {#acknowledgements .unnumbered}
================
SNZ acknowledges partial funding support by 973 Program of China under grant 2014CB845802, by the National Natural Science Foundation of China (NSFC) under grant Nos. 11133002 and 11373036, by the Qianren start-up grant 292012312D1117210, and by the Strategic Priority Research Program “The Emergence of Cosmological Structures” of the Chinese Academy of Sciences (CAS) under grant No. XDB09000000. JNF acknowledges partial funding support by the Joint Fund of Astronomy of NSFC and CAS through grant U1231202, and 973 Program under grants 2014CB845700 and 2013CB834900.
Alpar, M. A., Nandkumar, R., & Pines, D., 1986, ApJ, 311, 197
Baykal, A., Ali Alpar, M., Boynton, P., Deeter, J., 1999, MNRAS, 306, 207
Baym, G., Pethick, C. J., & Pines, D., 1969, Nature, 224, 674
Cheng, K. S., 1987a, ApJ, 321, 799
Cheng, K. S., 1987a, ApJ, 321, 805
Cordes, J. M., & Helfand, D. J., 1980, ApJ, 239, 640
Cordes, J. M., & Downs, G. S., 1985, ApJS, 59, 343
Cumming, A., Arras, P., & Zweibel, E., 2004, ApJ, 609, 999
Edwards, R. T., Hobbs, G. B., & Manchester, R. N., 2006, MNRAS, 372, 1549
Faucher-Giguere, C. -A., Kaspi, V. M., 2006, ApJ, 643, 332
Foreman-Mackey D., Hogg D. W., Lang D., Goodman J., 2013, PASP, 125, 306
Geppert, U., Page, D., & Zannias, T., 1999, A&A, 345, 847
Goldreich, P., & Reisenegger, A., 1992, ApJ, 395, 250
Gonthier, P., van Guilder, R., & Harding, A., 2004, ApJ, 604, 775
Gourgouliatos, K. N., & Cumming, A., 2014, MNRAS, 438, 1618
Hobbs, G., Lyne, A. G., Kramer, M., Marin C. E., & Jordan, C., 2004, MNRAS, 353, 1311
Hobbs, G., Lyne, A. G., & Kramer, M., 2010, MNRAS, 402, 1027
Johnston, S., & Galloway, D., 1999, MNRAS, 306, L50
Jones, P. B., 1988, MNRAS, 233, 875
Jones, P. B., 1990, MNRAS, 246, 364
Lorimer, D. R., & Kramer, M., 2005, Handbook of Pulsar Astronomy. Cambridge Univ. Press, Cambridge
Lyne, A. G., Ritchings, R. T., & Smith, F. G., 1975, MNRAS, 171, 579
Lyne, A. G., & Smith, F. G., 2004, Pulsar Astronomy, 3rd edn. Cambridge Univ. Press, Cambridge
Lyne, A. et al., 2010, Science, 329, 408
Manchester, R. N., & Taylor, J. H., 1997, Pulsars. Freeman, San Francisco
Manchester, R. N., Hobbs, G. B., Teoh A., & Hobbs, M., 2005, AJ, 129, 1993
Marchant, P., Reisenegger, A., Alejandro Valdivia, J., & Hoyos, J. H., 2014, ApJ, 796, 94
Ostriker, J. P., & Gunn, J. E., 1969, ApJ, 157, 1395
Pacini, F., 1967, Nature, 216, 567
Pacini, F., 1968, Nature, 219, 145
Pons, J. A., & Geppert, U., 2007, A&A, 470, 303
Pons, J. A., Vigano, D., & Geppert, U., 2012, A&A, 547, A9
Popov, S. B., Pons, J. A., Miralles, J. A., Boldin, P. A., & Posselt, B., 2010, MNRAS, 401, 2675
Regimbau, T., & de Freitas Pacheco, J., 2001, A&A, 374, 182
Sang, Y., & Chanmugam, G., 1987, ApJL, 323, L61
Shabanova, T. V., 1995, ApJ, 453, 779
Shabanova, T. V., Lyne, A. G., & Urama, J. O., 2001, ApJ, 552, 321
Shannon, R. M., & Cordes, J. M., 2010, ApJ, 725, 1607
Stairs, I. H., Lyne, A. G., & Shemar, S., 2000, Nat, 406, 484
Urama, J. O., Link, B., & Weisberg, J. M., 2006, MNRAS, 370, L76
Zhang, S. N., & Xie, Y., 2011, Astron. Soc. Pac. Conf. Ser., 451, 231
Zhang, S. N., & Xie, Y., 2012, ApJ, 757, 153 (Paper I)
Zhang, S. N., & Xie, Y., 2012, ApJ, 761, 102 (Paper II)
2D marginalized distributions of all the parameters {#sec:2D}
===================================================
We obtain two dimensional marginalized distributions of all the parameters of our model with the following steps.
(1)To sample the posterior distribution of parameters, we use [emcee]{} code from [@foreman13], which provides a fast and stable implementation of an affine-invariant ensemble sampler for MCMC. We can get the $R_{\rm sim}(t_{i})$ through the analytic expression of pulse phases(see equation (\[six\]) and equation (\[res\])).
(2)We fit the $R_{\rm sim}(t_{i})$ to the $R_{\rm re}(t_{i})$ preliminarily to find the maximum likelihood. To each parameter, we start by initializing 100 walkers in a tiny Gaussian ball around the maximum likelihood result.
(3)We sample the parameter space according to the $\mathrm{P}(\bmath{p} \, | \, \bmath{d})$ regarding the set of parameters for which the $\mathrm{P}(\bmath{p} \, | \, \bmath{d})$ is maximized as our best model, since the parameter set can reproduce the $R_{\rm re}(t_{i})$ most closely. In this process, we also need to output the time series of each parameter in the chain, which shows the parameter values for each walker at each step in the corresponding chain. The walkers start in small distributions around the maximum likelihood values and then they quickly wander and start to explore the full posterior distribution. Therefore, we can check if all the parameters are convergent through this.
(4)After making sure all the parameters are convergent, we can get marginalized distribution for each parameter independently in one-dimension histograms and two dimensional projections of posterior probability distributions(see Fig.\[Wa-corner\], Fig.\[Wb-corner\], Fig.\[Wc-corner\], Fig.\[Wd-corner\], Fig.\[Ma-corner\], Fig.\[Mb-corner\] and Fig.\[Mc-corner\]).
(5)We choose the medians of the one-dimension posterior distributions as the best fitting parameters, and the parameter uncertainties are calculated with their 68 percent confidence intervals.
(6)We need to make the projection of our result into the corresponding observed timing residuals.
To further make sure that our best fitting parameters are credible, after getting the final posterior distribution, we also chose 500 samples from the time series of parameters in the chain randomly to produce 500 simulated timing residuals. The result shows that they all overlap well with the ‘best’ simulated timing residuals, which are produced by our best fitting parameters.
{width="80.00000%"}
{width="80.00000%"}
{width="80.00000%"}
{width="80.00000%"}
{width="80.00000%"}
{width="80.00000%"}
{width="80.00000%"}
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'With the ubiquitous diffusion of social networks, images are becoming a dominant and powerful communication channel. Not surprisingly, they are also increasingly subject to manipulations aimed at distorting information and spreading fake news. In recent years, the scientific community has devoted major efforts to contrast this menace, and many image forgery detectors have been proposed. Currently, due to the success of deep learning in many multimedia processing tasks, there is high interest towards CNN-based detectors, and early results are already very promising. Recent studies in computer vision, however, have shown CNNs to be highly vulnerable to adversarial attacks, small perturbations of the input data which drive the network towards erroneous classification. In this paper we analyze the vulnerability of CNN-based image forensics methods to adversarial attacks, considering several detectors and several types of attack, and testing performance on a wide range of common manipulations, both easily and hardly detectable.'
author:
-
bibliography:
- 'refs.bib'
title: |
Analysis of adversarial attacks against\
CNN-based image forgery detectors\
[^1]
---
Image counterforensics, convolutional neural networks, generative adversarial networks.
Introduction
============
In the era of social networks, images have become a dominant communication vehicle. They convey information with higher immediacy and depth than text, and have the potential to elicit strong responses in the observers. Unfortunately, with modern media editing tools, tampering with images has become very easy. The manipulated images can be used to discredit people, direct public opinion, even change the course of political events, and pass easily unnoticed from ordinary people.
A number of multimedia forensic tools have been proposed in the last years to detect image manipulations [@Korus2017]. In particular, methods based on high-order statistics of image residuals have drawn great attention since long time [@Farid2003; @Bayram2006]. Indeed, when a pristine image is modified, by inserting or removing objects, or modifying global characteristics, several low-level operations are usually involved, like linear or non-linear filtering, resizing, or compression. All these operations leave subtle but distinctive traces in the image micro-patterns, which can be discovered by means of suitable image descriptors extracted from the high-pass image residual. To this end, the SPAM (subtractive pixel adjacency matrix) features [@Pevny2010] and the SRM (spatial rich models) [@Fridrich2012] have shown great potential for many image forensics tasks [@Cozzolino2014a; @Boroumand2017; @Li2018]. In particular, excellent results [@Cozzolino2014a; @Cozzolino2015b; @Li2018] can be obtained even by considering one specific single model from [@Fridrich2012], the one computing 4-pixel co-occurrences on the residuals of 3rd order linear filter. Given their similarity wirh SPAM features, here for the sake of brevity we will refer to them as S3SPAM or simply SPAM features.
Nonetheless, the current trend in forensics, and in multimedia processing in general, is to abandon handcrafted features in favor of deep learning. Given a sufficiently large training set, deep nets, typically convolutional neural networks (CNN), learn from the data which features best address the given task, reaching usually impressive performance. The first CNN-based detector of image manipulation was proposed in [@Bayar2016], inspired to previous work in steganalysis. Its main peculiarity is an [*ad hoc*]{} first layer, comprising a bank of filters constrained to extract high-pass features. Since the most relevant information for discrimination is hidden in the high-pass image content, such filters speed up convergence to a satisfactory solution. In [@Cozzolino2017], instead, it was proven that S3SPAM features can be extracted by a simple shallow CNN. Besides reproducing the very good results of the original detector, the resulting net can be further improved by fine tuning on a specific dataset, providing a very good performance even with a small training set. Very recently, another deep learning solution has been proposed, aimed at detecting the processing history of JPEG images [@Boroumand2018].
All the above networks, though very effective, are relatively shallow. Very deep architectures can be expected to provide a further performance boost. Tellingly, in a recent competition on camera model identification organized by the IEEE Signal Processing society on the Kaggle platform[^2], all top-ranking teams proposed solutions based on an ensemble of very deep networks. Likewise, very deep networks have shown top performance and higher robustness [@Marra2018] in detecting images manipulated by generative adversarial networks (GAN).
{width="0.9\linewidth"}
Although deep learning holds great potential for multimedia forensics, one should not rely on a safe environment, counting on the attacker’s naivete. On the contrary, the risks incurred by counter-forensic actions, aimed at neutralizing forensic tools (see Fig.1), must be taken into serious account and analyzed in depth. Some recent papers [@Marra2015; @Chen2017], for example, propose to attack SPAM-based detectors by means of iterative gradient descent algorithms, which prove very effective, although definitely slow. Attacking CNNs, however, has proven to be much simpler [@Goodfellow2015]. By exploiting the intrinsic differentiability of the loss function, a suitable adversarial noise can be easily generated and added to the input image to modify the network decision, without visible image impairments. Following this seminal paper, many more attacks based on adversarial noise have been devised. In addition, deep learning can be used itself for counter-forensics. In [@Kim2018], a GAN-based architecture was proposed to conceal the traces of 3$\times$3 median filtering. Such a study, though limited to a very special case, opens the way to interesting developments.
Here, we investigate on the effectiveness of adversarial attacks to CNN-based detectors. We consider a large set of manipulations, both easily detectable and more challenging, and several CNN-based detectors. Specific adversarial noise is generated for each detector, and the effects are assessed both on the target detector and on non-target ones. The performance of GAN-based restoration is also assessed, with reference to the especially challenging case of median filtering. To the best of our knowledge, this is the first study on this topic.
In the rest of the paper, we describe the detectors (Section 2), the attacks (Section 3), and the experimental analysis (Section 4), before drawing conclusions (Section 5).
CNN-based detectors of image manipulation
=========================================
In this Section we briefly recall some relevant CNN-based detectors, with their main features. However we also consider a baseline conventional detector, using handcrafted features [@Fridrich2012] and support vector machine (SVM) classification.
SPAM+SVM
--------
To extract the residual features proposed in [@Fridrich2012] the original image is high-pass filtered and quantized with a small number of bins. Then, co-occurrences are computed, encoded, and collected in a linear histogram feature, normalized to unit energy. Depending on the specific parameters of this process, different features are obtained, collectively called rich models [@Fridrich2012]. As said before, we consider one single model here, with third-order linear filter, 5-bin quantization, and 4-lag co-occurrences ([*s3\_spam14hv*]{}) and will refer to it as SPAM features from now on.
Bayar2016
---------
In [@Bayar2016] a relatively small CNN is proposed for image manipulation detection, referred to as Bayar2016 from now on, comprising three convolutional layers, two max-pooling layers, and three fully-connected layers. In order to immediately extract residual-based features, as suggested by the literature, filters of the first layer, with 5$\times$5 receptive field, are constrained to respect the following rule $$\left\{ \begin{array}{l}
w(0,0) = -1 \\
\sum_{l,m \neq 0} w(l,m) = 1
\end{array} \right.$$ Therefore, the sum of all weights is 0, enforcing the high-pass nature of the filters. In particular, the off-center pixels are combined to compute a prediction of the center pixel, so the output of the filter can be regarded as a prediction error.
Cozzolino2017
-------------
The main result of [@Cozzolino2017] is that a large class of conventional features can be computed exactly by suitable convolutional networks. Although the result is quite general, the work focuses on the SPAM feature described before. Exact SPAM feature extraction requires only two convolutional layers, followed by hardmax and average pooling. The extracted features could then be used to train an external SVM. However, a full-fledged CNN-based detector is also built in [@Cozzolino2017], by complementing the feature extractor subnet with a fully connected layer which replaces the external SVM classifier. Then, the hardmax is also replaced by softmax to ensure differentiability, allowing further training by backpropagation. Besides the theoretical result, the CNN proposed in [@Cozzolino2017] can faithfully replicate the SPAM-SVM suite, and improve upon it by means of quick fine tuning over a very small training set. This latter version, referred to as Cozzolino2017, is considered here.
Very deep nets: Xception
------------------------
In recent years, a large experimental evidence has accumulated showing that network depth plays a fundamental role for generalization ability. State-of-the-art architectures in computer vision and related fields, such as ResNet, DenseNet, InceptionNet, XceptionNet, all comprise from several dozens to hundreds of layers. Our own experience in forensic applications [@Marra2018] confirms the superior robustness of deep nets to challenging and off-training conditions. On the down side, deep nets require very large datasets for correct training, a condition not always met in practice.
To include a deeper net in our comparative assessment we selected Xception [@Chollet2017], comprising a total of 42 layers, 36 convolutional, 5 pooling, and one fully connected. Its main architectural innovation is the use of separable filters. That is, 3D convolutions are obtained by the cascade of 2D spatial and 1D cross-map convolutions. Thanks to this constraint, the number of free parameters drops significantly w.r.t. competing nets or, under a different point of view, a deeper architecture can be adopted for the same level of complexity, allowing the use of such a deep net even with a relatively small training set.
Attacking forensic detectors
============================
In this Section, we describe some possible strategies to attack image forensic detectors, in particular
- gradient descent algorithms for SPAM+SVM;
- generation of adversarial noise for CNN-based detectors;
- GAN-based restoration of manipulated images.
Although we focus on targeted attacks, designed against a specific detector, universal counter-forensic methods are also studied, [*e.g.*]{}, [@Barni2013b]
Attacking a SPAM-based detector by gradient descent
---------------------------------------------------
![Attacks in the feature space. Left: restoring the feature of the pristine image. Right: crossing the decision boundary of the target detector. The second attack is simpler but may fail with non-target detectors.[]{data-label="fig:Marra_attacks"}](figure/Marra_Attacks){width="1\linewidth"}
Let $X^{{prist}}$ and $X^0$ be the pristine and manipulated images, with $f^{{prist}}$ and $f^0$ the corresponding feature vectors, SPAM in our case. Lacking perfect knowledge on the classifier, the attacker wants to modify $X^0$ into a new image, $\hat{X}$, similar to $X^0$, to limit distortion, but whose feature, $\hat{f}$, is so close to $f^{{prist}}$ to fool the detector, see Fig.\[fig:Marra\_attacks\](left). In formulas, the problem can be cast as $$\hat{X} = \arg\min_X \psi(f(X),f^{{prist}}), \hspace{4mm} {\rm s.t.}\;\; \phi(X,X^0)<T
\label{eq:formulation3}$$ where $\phi(\cdot,\cdot)$ and $\psi(\cdot,\cdot)$ are image and feature space distances, and $T$ a suitable threshold on distortion.
In [@Marra2015] an iterative algorithm is proposed, where the objective function is minimized through local changes on the image, like in the iterated conditional modes method. This approach is effective but quite slow, because the feature must be recomputed at each new step, due to its complex nonlinear relationship with the image. Note that, if the classifier is perfectly known, one can target $f^{close}$, the feature closest to $f^0$ across the decision boundary, rather than $f^{{prist}}$, as shown in Fig.\[fig:Marra\_attacks\](right). This speeds up convergence considerably, but reduces robustness with respect to off-target detectors, as also depicted in Fig.\[fig:Marra\_attacks\](right).
Attacking CNNs by adversarial noise
-----------------------------------
------------------------------ ------- -------- ------- ------ ------- ------- ------ -------- ------- ------- ------- -------
Manipulation/parameter FPR TPR ACC FPR TPR ACC FPR TPR ACC FPR TPR ACC
Blurring, $\sigma$=1.10 0.06 100.00 99.97 0.02 99.98 99.98 0.07 100.00 99.96 2.26 99.87 98.81
JPEG compression, $Q$=70 0.02 100.00 99.99 0.83 99.69 99.43 0.00 99.98 99.99 0.63 98.54 98.95
Median filtering, 7$\times$7 0.54 99.90 99.68 0.56 99.93 99.69 1.26 100.00 99.37 0.85 99.96 99.56
Resizing, scale=1.500 0.26 100.00 99.87 0.50 99.94 99.72 0.00 99.96 99.98 6.56 99.11 96.28
Blurring, $\sigma$=0.50 24.02 99.12 87.55 8.74 97.93 94.59 7.69 99.06 95.69 13.76 88.06 87.15
JPEG compression, $Q$=90 6.46 88.26 90.90 0.72 90.17 79.72 5.83 94.81 94.49 2.93 90.37 93.72
Median filtering, 3$\times$3 0.06 99.80 99.87 0.31 99.91 99.80 0.43 99.91 99.74 7.46 99.54 96.04
Resizing, scale=1.010 9.07 99.29 95.11 2.72 99.59 98.44 3.00 99.67 98.33 8.26 98.22 94.98
------------------------------ ------- -------- ------- ------ ------- ------- ------ -------- ------- ------- ------- -------
\[tab:no\_attack\]
Experiments in computer vision [@Goodfellow2015; @Papernot2016JSMA] have clearly established the vulnerability of CNN-based detectors to adversarial attacks. A suitable adversarial noise pattern can be added to the input image to mislead the classifier, see Fig.1, without impairing its visual quality. With CNNs, some simple algorithms for the generation of adversarial noise are available.
The Fast Gradient Sign Method (FGSM), proposed in [@Goodfellow2015], exploits the differentiability of the loss function. The gradient of the loss with respect to each pixel of the input image is first computed by backpropagation. Then, each pixel is modified by a small quantity, $\pm\epsilon$, taking the sign of the local gradient. Neglecting higher order effects, all perturbations increase the loss, and hence a large change in output can be obtained with very low-variance adversarial noise.
Following this early, and simple, method, more sophisticated solutions have been proposed. DeepFool [@Moosavi2016] is based on a local linearization of the classifier under attack, which allows one to project the input image on the approximate decision boundary, and to introduce the minimum perturbation necessary to cross it. The Jacobian-based Saliency Map Attack (JSMA) [@Papernot2016JSMA] relies on a greedy iterative procedure. Unlike FGSM, it attacks only the pixels that contribute most to the correct classification, identified by a suitable saliency map. In [@Madry2017], adversarial noise generation is formulated as a min-max optimization, with the double aim of generating effective adversarial examples and training robust classifiers. The resulting algorithm, projected gradient descent (PGD), provides the optimum adversarial examples when the network is perfectly known. Noteworthy, FGSM can be regarded as a single-step scheme to solve the maximization step of PGD.
In the experiments, we will consider only the FGSM algorithm, because of its low complexity (JSMA and PGD are orders of magnitude slower) and easy interpretation. Note that, in a realistic setting, images must be rounded to integer values to be stored or transmitted, so, unlike in theoretical analyses, we consider only integer values for the $\epsilon$ parameter.
Manipulation SPAM Bayar2016 Cozz.2017 Xception
------------------- ------- -------------- ----------- ----------
Blurring 1.10 0.19 [**0.00**]{} 0.00 3.51
JPEG 70 0.06 [**0.00**]{} 0.08 10.22
Median 7$\times$7 99.16 [**0.00**]{} 99.96 26.56
Resizing 1.50 0.02 [**0.00**]{} 0.00 3.28
Blurring 0.50 1.43 [**0.00**]{} 7.67 12.17
JPEG 90 0.26 [**0.02**]{} 0.13 18.01
Median 3$\times$3 86.65 [**0.00**]{} 45.43 10.49
Resizing 1.01 0.95 [**0.00**]{} 5.92 10.18
: TPR with adversarial noise (FGSM, $\epsilon=1$). Target: Bayar2016.
\[tab:target\_Bayar\]
Manipulation SPAM Bayar2016 Cozz.2017 Xception
------------------- ------- ----------- --------------- ----------
Blurring 1.10 32.63 32.68 [**32.19**]{} 32.75
JPEG 70 0.20 13.51 [**0.00**]{} 17.16
Median 7$\times$7 98.78 93.34 [**88.20**]{} 96.71
Resizing 1.50 18.76 18.92 [**18.91**]{} 18.58
Blurring 0.50 1.05 1.66 [**15.67**]{} 4.90
JPEG 90 0.02 27.94 [**0.00**]{} 16.42
Median 3$\times$3 92.29 99.95 [**6.43**]{} 35.22
Resizing 1.01 0.71 4.49 [**13.89**]{} 9.02
: TPR with adversarial noise (FGSM, $\epsilon=1$). Target: Cozz.2017.
\[tab:target\_Cozzolino\]
Manipulation SPAM Bayar2016 Cozz.2017 Xception
------------------- ------- ----------- ----------- ---------------
Blurring 1.10 8.33 12.50 0.63 [**0.00**]{}
JPEG 70 3.00 38.85 0.81 [**0.41**]{}
Median 7$\times$7 99.30 24.74 100.00 [**0.00**]{}
Resizing 1.50 11.56 13.22 1.06 [**0.00**]{}
Blurring 0.50 26.31 30.41 17.20 [**0.00**]{}
JPEG 90 2.24 33.46 0.33 [**17.98**]{}
Median 3$\times$3 99.83 43.93 100.00 [**7.50**]{}
Resizing 1.01 14.24 27.91 8.70 [**0.00**]{}
: TPR with adversarial noise (FGSM, $\epsilon=1$). Target: Xception.
\[tab:target\_Xception\]
Attacks based on GANs
---------------------
Starting from the 2015 seminal work of Goodfellow [*et al.*]{} [@Goodfellow2015] generative adversarial networks have gained a major role in deep learning, providing remarkable results in a large number of tasks involving image synthesis and/or manipulation. The basic idea is to train in parallel two competing nets, a generator, which tries to synthesize images with a natural appearance, and a discriminator, which tries to tell apart natural from synthetic images. This competition gradually improves the performance of both nets. Ideally, at convergence, the generator should be able to produce images that are indistinguishable from natural ones. Recently, a GAN-based method has been proposed [@Kim2018] for the restoration of median filtered images. For this application, the generator does not start from a random noise vector to synthesize the output, as usual with GANs, but takes in input the manipulated image and restores its natural features. Accordingly, the generator loss includes not only an adversarial term, which measures its ability to fool the detector, but also two image quality terms. These measure objective quality (distance from the original) and perceptual quality of the generated image. We refer the reader to the original paper for all details of the method, underlining only that the generator relies heavily on residual connections to improve stability and speed up convergence.
Experimental analysis
=====================
To carry out our experimental analysis we generate a dataset taking 200 images from each of 9 different devices, and 192 partially overlapping $128\times 128$ patches from each image. We consider 4 types of image manipulation: Gaussian blurring, JPEG compression, median filtering, and resizing, with two different settings for each case corresponding to “easy” and “challenging” tasks. For example, a Gaussian filter with $\sigma=1.10$ causes easily detectable blurring, unlike with $\sigma=0.50$. In each binary classification task, patches from 6 devices chosen at random are used for training, the others for testing. Overall, each training set comprises more than 200k+200k patches, still relatively small for deep learning applications.
In Tab.\[tab:no\_attack\] we report, for the considered detectors, false positive rate (FPR), true positive rate (TPR), and overall accuracy (ACC), in the absence of counter-forensic attacks. For easy cases (top), accuracies are always close to 100%, only Xception shows a somewhat worse performance, very likely due to the limited training set. For more challenging manipulations (bottom), larger differences are observed, with some poorer results on JPEG compression (Bayar2016, SPAM) and Gaussian blurring (SPAM, Xception). Nonetheless, a very good detection performance is still observed, in general.
In Tables [\[tab:target\_Bayar\] through \[tab:target\_Xception\]]{} we study the case in which adversarial noise is added to the manipulated images, using FGSM with $\epsilon=1$, namely, the weakest adversarial noise which survives the image rounding. Since neither the pristine images nor the detectors change, we report only the TPR for the attacked manipulated images. Note also that the PSNR is always 48.13 dB (MSE=1), hence no visual impairment can be appreciated. The attack is very effective when the same net is used to generate the adversarial noise and to detect the manipulation (boldface entries). Only Cozzolino2017, and only for the 7$\times$7 median filtering, keeps providing a good TPR. In the absence of alignment, however, the attack is much less effective, especially for median filtering, both 3$\times$3 and 7$\times$7, for which both SPAM and Cozzolino2017 provide a TPR close to 100%. These results suggest that, at least in such cases, the adversarial noise is not restoring the features of pristine images disrupted by the manipulation, but only exploiting some detector weaknesses.
This latter consideration further motivates us to explore the GAN-based attack, which has the very goal of restoring manipulated images. In Tab.\[tab:GAN\_attack\] we report results only for the critical median filtering cases. They seem to confirm a better ability of the GAN-based method to attack uniformly all detectors. Actually, the original architecture proposed in [@Kim2018] works well only in the 3$\times$3 case, and never fools Xception. However, if we replace the original discriminator with a VGG net [@Simonyan2014], the attack becomes more effective for all detectors, and none of them reaches a 50% TPR.
Manipulation SPAM Bayar2016 Cozz.2017 Xception
-------------------- ------- ----------- ----------- ----------
M-7$\times$7 (Kim) 70.44 97.78 89.39 89.19
M-3$\times$3 (Kim) 19.30 49.20 15.69 91.70
M-7$\times$7 (VGG) 3.74 49.15 18.48 35.56
M-3$\times$3 (VGG) 1.17 0.74 0.78 23.76
: TPR for median filtering after GAN-based restoration. Top: Kim2018 discriminator. Bottom VGG discriminator.
\[tab:GAN\_attack\]
Conclusions
===========
We have presented an investigation on adversarial attacks to CNN-based image manipulation detectors. Even a rather simple attack can completely mislead the target detector and largely reduce the detection performance of off-target detectors. As only exception, the adversarial noise attack was not able to conceal 7$\times$7 median filtering, which deeply modifies the image fine structures. However, a suitable GAN-based attack proves to work well even in this challenging case.
Obviously, these early results represent only a proof of concept, and more thorough analyses are necessary to gather a solid understanding of the relevant issues. More sophisticated attacks must be considered, and more detectors tested, on a wider range of manipulations. In particular, realistic applications over social networks, involving resizing and compression, should be considered.
[^1]: This material is based on research sponsored by the Air Force Research Laboratory and the Defense Advanced Research Projects Agency under agreement number FA8750-16-2-0204. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the Air Force Research Laboratory and the Defense Advanced Research Projects Agency or the U.S. Government.
[^2]: https://www.kaggle.com/c/sp-society-camera-model-identification
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The restoration of particle number within Energy Density Functional theory is analyzed. It is shown that the standard method based on configuration mixing leads to a functional of both the projected and non-projected densities. As an alternative that might be advantageous for mass models, nuclear dynamics and thermodynamics, we propose to formulate the functional in terms directly of the one-body and two-body density matrices of the state with good particle number. Our approach does not contain the pathologies recently observed when restoring the particle number in an Energy Density Functional framework based on transition density matrices and can eventually be applied with functionals having arbitrary density dependencies.'
author:
- Guillaume Hupin
- Denis Lacroix
- Michael Bender
bibliography:
- 'PAV-func.bib'
date: 28 April 2011
title: On the formulation of functional theory for pairing with particle number restoration
---
Introduction {#sec:intro}
============
Energy Density Functional (EDF) methods provide a universal framework to describe nuclear structure, nuclear dynamics or thermodynamics. Tremendous advances have been made in the last few decades on the practical application of EDF methods to nuclei [@Ben03]. Still, despite their long success, some of the fundamental assumptions made to justify the usual strategies how the EDF techniques are constructed and used for nuclear systems have not yet been satisfactorily clarified. Most, if not all, EDF approaches break as many symmetries of the nuclear Hamiltonian as possible: translational, rotational and $U(1)$ symmetry in gauge space, among the most important ones. In fact, the exploitation of symmetry breaking in nuclei is strongly motivated by experimental observations. For instance, the appearance of highly collective rotational bands in spectroscopic data clearly points to the existence of deformed intrinsic states in many nuclei [@Rin80]. Similarly, there is evidence that pairing can be often treated by explicitly breaking the $U(1)$ gauge symmetry of eigenstates of the particle-number operator, like for instance in a Bardeen-Cooper-Schrieffer (BCS) or Hartree-Fock-Bogoliubov (HFB) approach [@Bla86; @Rin80]. Nuclei are, however, finite systems and methods like BCS or HFB do not properly treat quantum fluctuations of the order parameter associated with the broken symmetry [@Bla86]. These fluctuations can be incorporated either by a statistical treatment of the order parameter, or by the restoration of the relevant symmetry [@Bla86]. The concept of symmetry breaking and restoration stands out as the tool of choice within the EDF framework.
It has, however, been recently shown that restoration of symmetries has to be handled with great care in an EDF framework [@Ang01; @Dob07; @Lac09a; @Dug10]. In particular, the configuration mixing within a Multi-Reference (MR) EDF approach might lead to serious practical difficulties that can, however, eventually be cured [@Lac09a; @Ben09]. Besides compromising applications, these difficulties have clearly pointed out the necessity to clarify the theoretical framework on which the theory can be build.
The discussion in the present paper is restricted to ground-state properties and to particle-number projection, for which detailed analyses have been recently made. This case is the simplest situation where pathologies of the MR-EDF approach have been observed [@Ang01], analyzed and regularized [@Lac09a; @Ben09]. The first goal of the present work is to provide an alternative analysis of the EDF theory using configuration mixing to restore symmetries without and with the regularization. It will be shown that neither the non-regularized nor the regularized functionals can straightforwardly be interpreted in terms of the densities of projected or non-projected states. Starting from this analysis, the second intent of this work is to propose an alternative way to introduce a functional theory that is $U(1)$ symmetry conserving, and that without making use of the Multi-Reference technique. Our approach remains close to the Hohenberg-Kohn [@Hoh64] and Kohn-Sham [@Koh65] framework and uses a projected state as an intermediate trial state to construct the components of the functional. This approach avoids the difficulties recently encountered in MR-EDF approaches and can be applied also with functionals that cannot be safely employed within the standard MR-EDF approach, as for example functionals with non-analytical density dependences.
Particle number restoration within EDF theory: standard approach {#sec:edfstandard}
================================================================
The strategy to obtain a functional for pairing including particle-number restoration has been extensively analyzed recently [@Dob07; @Lac09a; @Ben09; @Dug09] and we only give here the main steps necessary for our discussion. Following these references, in this section we will consider a specific class of functional form that will be sufficient for the present discussion.[^1] At the so-called Single-Reference (SR) level, a quasi-particle (QP) vacuum of Bogoliubov type $| \Phi_0 \rangle$ is used to construct the normal and anomalous density matrices, denoted by $\rho$ and $\kappa$, that serve to construct the functional. The energy is then written as $$\begin{aligned}
{\cal E}_{SR}[\Phi_0] =&& \mathcal{E}_{SR} \left[ \rho , \kappa, \kappa^* \right] \nonumber \\
=&& \sum_{i} t_{ii} \rho_{ii}+ \frac{1}{2} \sum_{i,j} \overline{v}_{ijij}^{\rho \rho} \rho_{ii}\rho_{jj} \nonumber \\
&+& \frac{1}{4}
\sum_{i,j} \overline{v}_{i\bar\imath j\bar\jmath }^{\kappa \kappa} \kappa_{i \bar\imath }^* \kappa_{j \bar\jmath } \, ,
\label{eq:denssr}\end{aligned}$$ where $\overline{v}^{\rho \rho}$ and $\overline{v}^{\kappa \kappa}$ denote the effective vertices in the particle-hole and particle-particle channels. Standard SR-EDF can be schematically seen as the sequence $$\label{eq:seqSR}
\Phi_0 ~\Longrightarrow~ ( \rho , \kappa, \kappa^*) ~\Longrightarrow ~ {\cal E}_{SR} \, .$$ The price to be paid for incorporating pairing with a rather simple functional is to use an intermediate state $| \Phi_0 \rangle$ that is not an eigenstate of particle number. In a second step, the symmetry can be restored projecting out the component with $N$ particles $$\begin{aligned}
| \Psi_N \rangle = P^N | \Phi_0 \rangle \, ,\end{aligned}$$ where $P^N$ denotes the particle number projection operator defined through [@Rin80; @Bla86] $$\label{eq:Pop}
{P}^N
= \frac{1}{2\pi} \int_{0}^{2\pi} \! d{\varphi} \; \,e^{i\varphi (\hat{N}-N)}
\, .$$ The expectation value of any operator $O$ that conserves particle number can then be expressed as $$\label{eq:expproj}
\frac{\langle \Psi_{N} | \, O \, | \Psi_{N} \rangle}{\langle \Psi_{N} | \Psi_{N} \rangle}
= \int_{0}^{2\pi} d\varphi \,
\frac{ \langle \Phi _0 | O | \Phi_\varphi \rangle } {\langle \Phi _0 | \Phi_\varphi \rangle} {\cal N}_N ({0, \varphi})
\, ,$$ where the shorthand $$\begin{aligned}
{\cal N}_N ({0, \varphi}) \equiv \frac{e^{-i\varphi N}}{2\pi}
\frac{\langle \Phi_0 | \Phi _{\varphi} \rangle}{\langle \Psi_{N} | \Psi_{N} \rangle}\, ,\end{aligned}$$ has been introduced. Here $\varphi$ denotes the gauge angle, whereas $| \Phi_{\varphi} \rangle = e^{i \varphi \hat N} \, | \Phi_0 \rangle$ refers to the state $| \Phi_0 \rangle$ rotated in gauge space. The kernel entering in the integral of Eq. (\[eq:expproj\]) corresponds to the transition matrix element of an operator between two quasi-particle vacua. One can then take advantage of the Generalized Wick Theorem (GWT) [@Bal69; @Rin80] to express the kernel in terms of the transition density matrices $$\begin{aligned}
\label{contractph}
\rho^{0\varphi}_{i j }
& \equiv & \frac{\langle \Phi_{0} | a^{\dagger}_{j} a_{i} | \Phi_{\varphi} \rangle}
{\langle \Phi_{0} | \Phi_{\varphi} \rangle} , \\
\label{contracthh}
\kappa^{0\varphi}_{i j }
& \equiv & \frac{\langle \Phi_{0} | a_{j} a_{i} | \Phi_{\varphi} \rangle}
{\langle \Phi_{0} | \Phi_{\varphi} \rangle},
\\
\label{contractpp}
{\kappa^{\varphi 0 }_{j i}}^\star
& \equiv & \frac{\langle \Phi_{0} | a^{\dagger}_{i } a^{\dagger}_{j} | \Phi_{\varphi} \rangle}
{\langle \Phi_{0} | \Phi_{\varphi} \rangle}.\end{aligned}$$ For instance, when $O$ is a two-body Hamiltonian, the two-body interaction $\overline{v}$ entering in (\[eq:expproj\]) takes a form similar to Eq. (\[eq:denssr\]) with $\overline{v}^{\rho\rho} =
\overline{v}^{\kappa \kappa}= \overline{v}$ and where the densities, $\rho$ and $\kappa$ are replaced by the corresponding transition densities, Eqns. (\[contractph\]-\[contractpp\]). Guided by the Hamiltonian case, the energy functional associated with particle-number restoration is usually defined through $$\begin{aligned}
\label{eq:ekernel}
{\cal E}_N [\Psi_{N}] \equiv \int_{0}^{2\pi} d\varphi \, \mathcal{E}_{SR} \left[ \rho^{0\varphi} , \kappa^{0\varphi},
{\kappa^{\varphi 0 }}^\star \right]
{\cal N}_N ({0, \varphi}) \, . \end{aligned}$$ This energy functional is a special case of a so-called Multi-Reference EDF (MR-EDF). The present strategy to restore symmetries in an EDF framework deserves some further remarks:
- The expression (\[eq:ekernel\]) is postulated having in mind the Hamiltonian case. However, the MR-EDF theory should not be confounded with the expectation value of a Hamilton operator. In particular, an energy functional has much more flexibility regarding the functional form of the energy kernels in Eq. (\[eq:ekernel\]), which can be used for the efficient modeling of in-medium correlations.
- The construction of the MR-EDF, Eq. (\[eq:ekernel\]), from the SR-EDF by simply replacing the normal and anomalous density matrices in the SR EDF by the corresponding transition density matrices is postulated by analogy to the GWT. While it appears rather natural, it was shown recently that this strategy to construct the MR-EDF might lead to an ill-defined functional that exhibits divergencies and jumps in practical applications [@Dob07; @Lac09a; @Ben09]. While a solution to this problem has been proposed and applied in Refs. [@Lac09a; @Ben09], a consistent framework for MR-EDF approaches is still missing. A clear illustration of this is the ongoing debate about which densities should enter in the functional [@Rob10], as well as the recently recognized impossibility to use non-integer powers of the transition density in MR energy functionals [@Dug09].
- The very notion of symmetry restoration within an EDF framework remains to be clarified. For instance, it has been shown recently [@Dug10] that also regularized MR energy functionals may in general not transform as an irreducible representation of the restored symmetry, unless additional constraints are introduced.
In the present section, we will further analyze the way the MR-EDF is constructed and the possible sources of difficulties. For simplicity, we restrict ourselves to a case where the two-body effective interaction kernels entering Eq. (\[eq:denssr\]) are [*independent of the densities*]{}.
A peculiarity of particle-number projection is that the canonical basis of the original state $| \Phi_0 \rangle$ and of the rotated states $| \Phi_\varphi \rangle$ are the same when making a suitable choice of the Bogoliubov transformation between quasi-particle operators of these states. Accordingly, the canonical base of the projected state $| \Psi_N \rangle$ is also the same as the one of the original reference state $| \Phi_0 \rangle$. In the following, we will implicitly assume that densities are written in this canonical basis in which we have $$\label{eq:canodens}
\rho^{0\varphi}_{i j } = \delta_{ij} n^{0\varphi}_{i}, ~~~\kappa^{0\varphi}_{i j } = \delta_{j \bar\imath } \kappa^{0\varphi}_{i \bar\imath },
~~~{\kappa^{\varphi 0 }_{i j }}^\star = \delta_{j \bar\imath } {\kappa^{\varphi 0 }_{i \bar\imath }}^\star \,,$$ whereas the energy $ {\cal E}_{N} $ takes the form $$\begin{aligned}
{\cal E}_{N} [\Psi_{N}] &=& \sum_{i} t_{ii} \int_{0}^{2\pi} \!\!\! d\varphi ~n^{0\varphi}_i {\cal N}_N(0,\varphi)
\nonumber \\
&+& \frac{1}{2} \sum_{i,j} \overline{v}_{ijij}^{\rho \rho} \int_{0}^{2\pi} \!\!\! d\varphi ~n^{0\varphi}_i ~n^{0\varphi}_j {\cal N}_N({0,\varphi}) \nonumber \\
&+&\frac{1}{4} \sum_{i, j} \overline{v}_{i\bar\imath j\bar\jmath }^{\kappa \kappa}
\int_{0}^{2\pi} \!\!\! d\varphi ~{\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{j \bar\jmath }} \, {\cal N}_N({0,\varphi}) \, .\end{aligned}$$ After a lengthy, but straightforward calculation, the energy functional can be expressed as $$\begin{aligned}
{\cal E}_{N} [\Psi_{N}] &=& \sum_{i} t_{ii} n^N_i \nonumber \\
%\nonumber \\&+&
&+& \frac{1}{2} \sum_{i,j, j\neq \bar\imath } \overline{v}_{ijij}^{\rho \rho} R^N_{ijij} \nonumber \\
&+& \frac{1}{4}
\sum_{i \neq j, i \neq \bar \jmath } \overline{v}_{i\bar\imath j\bar\jmath }^{\kappa \kappa} R^N_{j \bar\jmath i\bar\imath } \nonumber \\
&+& \frac{1}{2} \sum_{i} \overline{v}_{i\bar\imath i \bar\imath }^{\rho \rho}
\int_{0}^{2\pi} \!\!\! d\varphi ~n^{0\varphi}_i ~n^{0\varphi}_i {\cal N}_N({0,\varphi}) \nonumber \\
&+&\frac{1}{2} \sum_{i}\overline{v}_{i\bar\imath i\bar\imath }^{\kappa \kappa}
\int_{0}^{2\pi} \!\!\! d\varphi {\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{i \bar\imath }} \, {\cal N}_N({0,\varphi}) \, ,
\label{eq:edftotproj}\end{aligned}$$ where $n^N_i$ are the occupation numbers: $$\begin{aligned}
n^N_i &\equiv&
\frac{ \langle \Psi_{N} | a^\dagger_i a_i | \Psi_{N} \rangle }{\langle \Psi_{N} | \Psi_{N} \rangle}
%\nonumber \\
%&=& \int_{0}^{2\pi} d\varphi \, ~ n^{0\varphi}_i {\cal N}_N({0,\varphi}) \, ,\end{aligned}$$ and $R^N_{ijkl}$ corresponds to the two-body density matrix $$\begin{aligned}
R^N_{ijkl} &\equiv& \frac{ \langle \Psi_{N} | a^\dagger_k a^\dagger_l a_j a_i | \Psi_{N} \rangle }{\langle \Psi_{N} | \Psi_{N} \rangle} %\nonumber \\
%&=& (\delta_{ik} \delta_{jl} - \delta_{il} \delta_{jk}) \int_{0}^{2\pi} \!\!\! d\varphi ~ n^{0\varphi}_i n^{0\varphi}_j {\cal N}_N({0,\varphi}) \nonumber \\
%&+& \delta_{j \bar\imath } \delta_{l \bar k} \int_{0}^{2\pi} \!\!\! d\varphi {\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{k \bar k}} \, {\cal N}_N({0,\varphi}) \, ,\end{aligned}$$ of the projected state. They can be expressed in terms of the gauge angle integrals as $$\begin{aligned}
n^N_i
&=& \int_{0}^{2\pi} d\varphi \, ~ n^{0\varphi}_i {\cal N}_N({0,\varphi}) \, ,\end{aligned}$$ and $$\begin{aligned}
R^N_{ijkl}
%&\equiv& \frac{ \langle \Psi_{N} | a^\dagger_k a^\dagger_l a_j a_i | \Psi_{N} \rangle }{\langle \Psi_{N} | \Psi_{N} \rangle} \nonumber \\
&=& (\delta_{ik} \delta_{jl} - \delta_{il} \delta_{jk}) \int_{0}^{2\pi} \!\!\! d\varphi ~ n^{0\varphi}_i n^{0\varphi}_j {\cal N}_N({0,\varphi}) \nonumber \\
&+& \delta_{j \bar\imath } \delta_{l \bar k} \int_{0}^{2\pi} \!\!\! d\varphi {\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{k \bar k}} \, {\cal N}_N({0,\varphi}) \, .\end{aligned}$$ Equation (\[eq:edftotproj\]) is rather enlightening with respect to the physical content of present MR-EDF calculations. Indeed, if one neglects the last two terms in Eq. (\[eq:edftotproj\]), one sees that the functional associated with the projected state can be regarded as a functional of the one- and two-body components of this very state. Similarly, if one uses the same effective interaction $\overline{v}^{\rho\rho} =
\overline{v}^{\kappa \kappa}$, then the last two terms of Eq. (\[eq:edftotproj\]) recombine and the two-body component $ R^N_{i\bar\imath i \bar\imath }$ can be recognized, thanks to the relation $$\begin{aligned}
R^N_{i\bar\imath i \bar\imath } &=& n_i^N \nonumber \\
&=& \int_{0}^{2\pi} \!\!\! d\varphi ~(n^{0\varphi}_i ~n^{0\varphi}_i + {\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{i \bar\imath }}) {\cal N}_N({0,\varphi}) \, . \label{eq:sumrule}\end{aligned}$$ However, when using different effective vertices $\overline{v}^{\rho\rho} \neq
\overline{v}^{\kappa \kappa}$ in the particle-hole and particle-particle channels, or when using vertices $\overline{v}^{\rho\rho}$ or $\overline{v}^{\kappa \kappa}$ that cannot be written as an antisymmetrized matrix elements of the two-body force, then the identification of the energy as a functional of one- and two-body density matrices of the projected state cannot be made anymore. Instead, it can only be written as a functional of the transition density matrices.[^2] This subtlety is intimately connected to the presence of pathologies encountered in MR-EDF calculations. Indeed, the last two terms in Eq. (\[eq:edftotproj\]) are nothing but the ones at the heart of the difficulties to construct a well-defined MR-EDF theory. As discussed in Refs. [@Dob07; @Ben09], for near-orthogonal states $\langle \Phi _0 | \Phi_\varphi \rangle \simeq 0$ there is at least one $n_i^{0\varphi}$ and the corresponding ${\kappa^{\varphi 0}_{i \bar\imath }}^\star$ and ${\kappa^{0\varphi}_{i \bar\imath }}$ that all go to infinity. As a consequence, the two terms can separately become larger than any physical scale in the nucleus. They do, however, recombine to a well-behaved expression when a Hamiltonian is used, i.e. when $\overline{v}^{\rho\rho} = \overline{v}^{\kappa \kappa}$. Without taking specific care of these terms in the restoration of symmetry within the functional framework, there is a spurious contribution that leads to discontinuities and divergences when plotting the particle-number projected energy as a function of a collective coordinate.
MR-EDF theory with regularization
---------------------------------
A strategy to construct a well-behaved MR-EDF theory proposed in Refs. [@Lac09a; @Ben09] is to remove terms that might not properly recombine in the MR-EDF approach in such a way that the spurious contamination is removed without touching the physical content of the functional. The resulting functional then takes the form (technical details are given in appendix \[app:correctedmr\]) $$\begin{aligned}
{\cal E}_{N} [\Psi_{N}] &=& \sum_{i} t_{ii} n^N_i \nonumber \\
&+& \frac{1}{2} \sum_{i,j, j\neq \bar\imath } \overline{v}_{ijij}^{\rho \rho} R^N_{ijij} \nonumber \\
&+& \frac{1}{4}
\sum_{i \neq j, j\neq \bar\imath } \overline{v}_{i\bar\imath j\bar\jmath }^{\kappa \kappa} R^N_{j \bar\jmath i\bar\imath } \nonumber \\
&+& \frac{1}{2} \sum_{i} \overline{v}_{i\bar\imath i \bar\imath }^{\rho \rho} (n^N_i n^N_i - \delta n_i \delta n_i) \nonumber \\
&+&\frac{1}{2} \sum_{i}\overline{v}_{i\bar\imath i\bar\imath }^{\kappa \kappa} \left[ n^N_i (1 - n^N_i) + \delta n_i \delta n_i \right] \, ,
\label{eq:edftotprojcor}\end{aligned}$$ where $\delta n_i = n^N_i -n^0_i$ is the difference between the occupation number of the level $i$ in the projected and the non-projected state.
Expression (\[eq:edftotprojcor\]) is of particular interest for the following discussion regarding the construction of energy functional theory. First, let us remark that, compared to the previous form (\[eq:edftotproj\]), the gauge space integrals are now hidden in the components of the one- and two-body density matrices of the projected state. In addition, the last two lines of Eq. (\[eq:edftotprojcor\]) are also functionals of the occupation numbers $n^0_i$ in the original non-projected state. The analysis of the regularization procedure to remove spurious contribution to the MR-EDF method [@Lac09a; @Ben09; @Dug09] suggests that these terms will always be well-behaved.
An example for a deformation energy curve obtained from a particle-number projected MR-EDF calculation with the Skyrme interaction SIII and a pairing functional of volume type is shown in Fig. \[fig1:pav\] (dashed line). The MR-EDF is numerically calculated using expression (\[eq:ekernel\]) and the Fomenko discretization procedure of the gauge-space integrals described, for instance, in Ref. [@Ben09]. Here, 199 discretization points have been used. This large number is necessary to resolve the discontinuities that stem from the spurious contribution to the non-regularized MR-EDF [@Ben09]. As in Ref. [@Ben09], the Lipkin-Nogami procedure is used in the minimization of the energy of the state $| \Phi_0 \rangle$. The solid line corresponds to the MR-EDF method with the regularization proposed in [@Lac09a]. In this Figure, we also show the results (filled circles) obtained using directly the expression (\[eq:edftotprojcor\]) that has been proven above to be analytically equivalent to the regularized MR-EDF functional. Note that, in the latter case, we have used a method called hereafter “recurrence method” to compute the projected occupation numbers and components of the projected two-body densities. This method is described in detail in appendix \[app:rec\]. Although the use of gauge angle integration would have given exactly the same results, this method has the advantage to be very simple, numerically efficient and to not make use of transition density matrices. As expected, the energy obtained with expression (\[eq:edftotprojcor\]) exactly matches the one obtained using the regularized MR-EDF functional. This formulation provides a new and alternative insight into the content of particle-number restored energy functionals.
![\[fig1:pav\] (Color online) Particle-number restored deformation energy curve of $^{18}$O as a function of quadrupole deformation $\beta_2$ and calculated within standard MR-EDF technique using Projection After Variation (PAV) with SIII and a delta pairing interaction before (blue dashed curve) and after regularization (black solid curve). The red filled circles correspond to the result obtained using directly Eq. (\[eq:edftotprojcor\]) (see text). To compare with previous work [@Ben09], the Coulomb exchange contribution has been subtracted from the energy. ](fig1pav.eps){width="8.cm"}
Critical analyses of standard method
------------------------------------
As discussed above, specific regularizations in MR-EDF functionals are needed to avoid discontinuities such as the jumps appearing in Fig. \[fig1:pav\]. At this point, even with the regularization, two important problems remain:
- Terms that have a non-analytical density dependence, for example a non-integer power such as often used in parameterizations of the Skyrme and Gogny interactions, cannot be regularized with the procedure proposed in Ref. [@Lac09a]. Indeed, the functional itself becomes in that case multivalued in the complex plane and cannot be properly defined [@Dob07; @Dug09].
- A second issue illustrated in Eq. (\[eq:edftotprojcor\]) is that the last two terms are not only a functional of the occupation numbers of the projected state, but also of the occupation probabilities of the original reference state $| \Phi_0 \rangle$. Accordingly, the energy remains a functional of the density of the quasi-particle vacuum that is not an eigenstate of particle number. This raises the question which density, i.e. projected, transition, or non-projected can be used to construct a functional for MR calculations.
In the following, we show that both (i) and (ii) can eventually be avoided by changing the strategy to construct the functional for pairing that accounts for particle number restoration.
Discussion on EDF theory for pairing with particle number restoration
=====================================================================
Let us now discuss the critique (ii) made above concerning the components of the projected energy functional. In most functional approaches, an intermediate state is introduced to construct densities that are used to minimize the energy. This is the case in usual DFT or at the SR-EDF level where the trial state is a Slater determinant or a quasi-particle vacuum. When restoring the symmetry in a MR-EDF framework, then, according to Eq. (\[eq:edftotprojcor\]), the projected state can be [*almost*]{} regarded as an intermediate many-body state from which the one- and two-body density matrices used to define the functional are obtained.
However, due to the presence of $n^0_i$ in the energy, this functional happens to depend on components not only of the projected state, but also of the original reference state. A slight modification, however, can easily restore the unique dependence of the functional on the projected state. If, for instance, the following replacements $$\begin{aligned}
\label{eq:nidnini}
\begin{array}{lcl}
(n^N_i n^N_i - \delta n_i \delta n_i) & \Longrightarrow & n^N_i n^N_i \\
\\
\left[ n^N_i (1 - n^N_i) + \delta n_i \delta n_i \right] & \Longrightarrow & n^N_i (1 - n^N_i) \, ,
\end{array}\end{aligned}$$ are made in Eq. (\[eq:edftotprojcor\]), then the strategy of standard DFT to construct the EDF as a functional of a density of an auxiliary state, the projected state here, is recovered.[^3]
The use of a projected product state a auxiliary state has the advantage that it allows to treat pairing in a particle-number conserving framework. An illustration of a result obtained taking into account this modification in Eq. (\[eq:edftotprojcor\]) is shown in Fig. \[fig2:pav\] and compared to the original curve. This figure illustrates that the small change in the functional does not significantly modify the energy landscape. This is indeed not unexpected since the difference $\delta n_i$ (resp. $\delta n_i \delta n_i$) is likely to be much smaller than $n^N_i$ (resp. $n^N_i n^N_i$).
![\[fig2:pav\] (Color online) Particle-number restored deformation energy surface of $^{18}$O calculated using Eq. (\[eq:edftotprojcor\]). The dotted curve is obtained by making the replacement (\[eq:nidnini\]) in Eq. (\[eq:edftotprojcor\]). ](fig2pav.eps){width="8.cm"}
By making the simple modification (\[eq:nidnini\]), the EDF framework can now be interpreted as a functional of the projected-state densities. Indeed, the state with good particle number can now be regarded as the auxiliary many-body state that provides the quantities used to construct the EDF. Similarly to Eq. (\[eq:seqSR\]), the corresponding theory can be regarded as a new sequence $$\label{eq:seqN}
\Psi_N~\Longrightarrow~ ( \rho^N , R^N) ~\Longrightarrow ~ {\cal E}_{N}.$$ We note in passing that the slight modification (\[eq:nidnini\]) does not break the shift invariance of the energy functional discussed in Refs. [@Dob07; @Ben09].
At this point, let us make a few further important remarks:
- The functional form (\[eq:edftotprojcor\]) is not completely surprising. Indeed, our starting point, Eq. (\[eq:denssr\]), is very close to a form one would have obtained by taking the expectation value of a two-body Hamiltonian. In the case of an energy functional calculated as the expectation value of a genuine Hamiltonian operator, the energy can be written as a functional of one-body transition densities, but also as a functional of projected one- and two-body (and eventually higher-order) densities. This freedom is lost in the functional framework, where a choice has to be made for either one or the other. MR-EDF follows the former strategy, whereas the latter has not been explored yet. For a regularized bilinear functional, the differences between both formulations remain very small, see Fig. \[fig2:pav\].
- [ Expression (\[eq:edftotprojcor\]) contains not only one-body but also two-body matrix elements and might appear out of the scope of a density functional approach aiming at replacing the original $N$-body problem by a functional of the one-body density matrix only. Indeed, in the Hamiltonian case, the expectation value of any two-body Hamiltonian for any state can directly be regarded as a functional of the one- and two-body densities of this state. Density functional theories are introduced to avoid the explicit use of two-body and higher density matrices. Therefore, by itself, the use of a functional of the two-body density might appear useless. The important simplification here stems from the fact that these densities are constructed from a very specific class of states, namely projected product states. For instance, we have shown recently that the two-body density matrix elements can eventually be written as an explicit functional of the one-body density under some approximation [@Lac10]. Accordingly, while two-body density matrix elements are used to get a compact expression in Eq. (\[eq:edftotprojcor\]), this functional can truly be considered as a functional of the projected state one-body density consistently with density matrix functional theory, such that the sequence become $$\label{eq:seqNone}
\Psi_N~\Longrightarrow~ \rho^N ~\Longrightarrow ~ {\cal E}_{N} \, .$$ ]{}
- When making the replacement (20) in Eq. (19), then the functional directly incorporates symmetry breaking and its restoration in a single step, contrary to standard approaches in EDF theory. From that point of view, it could be seen as a “Symmetry-Conserving” EDF theory.[^4]
- It is quite interesting to note that the MR-EDF can already almost be regarded as a functional of the components of the projected state. While this was hidden in formula (\[eq:ekernel\]), it becomes evident in Eq. (\[eq:edftotprojcor\]). In particular, as noted in Refs. [@Lac09a; @Ben09; @Dug09], there exists some flexibility in the regularization of the pathologies of the MR-EDF. It is possible to slightly modify the original prescription (\[eq:cor\_rr\]-\[eq:cor\_kk\]), such that the regularization automatically leads to (\[eq:nidnini\]). In that case, the method based on the use of MR-EDF and “symmetry conserved” EDF framework are strictly equivalent. As an important consequence, while the use of techniques inspired from configuration mixing was unclear within a functional framework, we give here evidence that it can be formulated consistently in a functional framework. It is, however, worth mentioning that while this connection can be made only in the simple functional form given in Eq. (\[eq:denssr\]), most functionals currently used do not allow their controlled usage in an MR EDF framework.
- Finally, it is important to mention that this equivalence holds only true for the schematic bilinear functional given by Eq. (\[eq:denssr\]) with two-body vertices independent on the density. If density dependent terms are present in the functional, like in all currently used parameterizations of the EDF, such an equivalence does not exist anymore. Note, however, that, in this case, a safely usable MR-EDF cannot be constructed anymore due to the absence of a regularization scheme. In Eq. (\[eq:edftotprojcor\]), one then obtains a functional that remains closer to the spirit of DFT based on the Hohenberg-Kohn theorem than the usual MR-EDF approach. Indeed, in the HK-theorem-based DFT, the functional is constructed from the density matrices of the correlated (i.e. in our case projected) state. As we will illustrate below, on the contrary, the alternative formulation proposed here that treats both symmetry breaking and restoration simultaneously can still be applied for functionals that cannot be regularized in a MR-EDF framework.
Constraints on the symmetry-conserving functional
-------------------------------------------------
If the standard projection method is used as guidance to construct the functional, then the form of the functional is almost entirely constrained. Indeed, this corresponds to use Eq. (\[eq:edftotproj\]) or eventually Eq. (\[eq:edftotprojcor\]) as a starting point. Eq. (\[eq:nidnini\]) corresponds to a specific choice. Here, we discuss whether alternative choices can be made for the last two lines of Eq. (\[eq:edftotprojcor\]). At present, it is not clear if, within the functional framework, a unique prescription of the functional form exists. Nevertheless, one can propose a few rules to better constrain its form. Let us assume a more general prescription than Eq. (\[eq:nidnini\]) $$\begin{aligned}
\label{eq:nidninigen}
\begin{array}{lcl}
(n^N_i n^N_i - \delta n_i \delta n_i) & \Longrightarrow & F^N_{i \bar\imath } , \\
\\
(n^N_i (1 - n^N_i) + \delta n_i \delta n_i) & \Longrightarrow & G^N_{i \bar\imath } \, ,
\end{array}\end{aligned}$$ where $F^N$ and $G^N$ are the unknown quantities.
Let us specify some rules to constrain them:
- [**Sum-rule:**]{} When $\overline{v}^{\rho\rho} = \overline{v}^{\kappa \kappa}$, then the last two terms in Eq. (\[eq:edftotproj\]) should recombine to give $R^N_{i \bar \imath i \bar \imath} =n^N_i$. Accordingly, it seems reasonable to impose $$\label{eq:const1}
F^N_{i \bar\imath } + G^N_{i \bar\imath } = n^N_i\, .$$
- [**No-pairing limit:**]{} Slater determinants belong to the Hilbert space spanned by projected states. Consequently, one can interpret the functional for particle-number projected wave functions as a generalization of the SR-EDF theory expressed for Slater determinant, i.e. $$\label{eq:srlimit}
{\cal E}_{N} [\Psi_{N}] \Longrightarrow {\cal E}_{SR} [\Phi_{\rm SD}]\, ,$$ as $\Phi_{N} \longrightarrow \Phi_{\rm SD}$. $\Phi_{\rm SD}$ denotes any Slater determinant. As a consequence, in this limit, we should have $$\label{eq:cons2}
F^N_{i \bar\imath } \Longrightarrow n^0_i n^0_i, \hspace*{.5cm} G^N_{i \bar\imath } \Longrightarrow 0.$$
- [**Large $N$ limit:**]{} In the limit of infinite particle number, the projected state and the reference state should become identical (for instance $\delta n^0_i \Longrightarrow 0$). Accordingly, we do expect $$\begin{aligned}
\begin{array}{ccc}
\lim_{N \rightarrow + \infty} F^N_{i \bar\imath } & =& n^N_i n^N_i , \\
\\
\lim_{N \rightarrow + \infty} G^N_{i \bar\imath } & =& n^N_i (1 - n^N_i) \,.
\end{array}\end{aligned}$$
These three constraints significantly reduce the freedom of choosing the form of the functional that can be used. The prescription (\[eq:nidnini\]) naturally fulfills all of them.
Can we use terms with non integer power of the density?
-------------------------------------------------------
When the effective two-body vertex depends explicitly on the density, then the energy cannot be directly mapped on Eq. (\[eq:edftotproj\]). If the density dependence is in integer powers of the density, then one could eventually generalize the derivation of Eq. (\[eq:edftotproj\]) to three-body or even higher-order effective interactions. For all other forms of the density dependence, such as the widely used non-integer powers of the density, there is no way to deduce an equivalent expression because the integration over gauge angles cannot be uniquely defined from a mathematical point of view [@Dob07; @Dug09]. It is worth to mention that the same difficulty appears when the Coulomb exchange term is approximated using the Slater prescription. Above, we have shown that, with a slight change in the functional used in the standard MR-EDF method, one obtain an EDF that can can be interpreted consistently within the usual functional approach where the projected state becomes a trial wave function to construct the ingredients of the functional.
Guided by the setup of functional (\[eq:edftotproj\]), the most natural and simple way to extend the SR-EDF functional using density dependent two-body effective vertices with non-integer powers of the density is to directly replace the density entering in the effective vertex by the density of the projected state, i.e. $$\label{eq:twoext}
\overline{v}^{\rho \rho}[\rho] \Longrightarrow
\overline{v}^{\rho \rho}[\rho^N] \, , ~~~~~
\overline{v}^{\kappa \kappa}[\rho] \Longrightarrow
\overline{v}^{\kappa \kappa}[\rho^N] \, .$$ Again, by doing this, we ensure that the functional used for the projected state is consistent with the one used in the no-pairing case (Eq. (\[eq:srlimit\])) and in the large-$N$ limit. In Fig. \[fig3:pav\], the deformation energy curve obtained by using Eq. (\[eq:twoext\]) is compared to the result deduced from the standard non-regularized MR-EDF procedure using Eq. (\[eq:ekernel\]). The SLy4 effective interaction used here contains density dependent terms with non-integer powers i.e. $\rho^{2+1/6}$. Note that in this case, the MR-EDF cannot be regularized. The new alternative method we propose here, however, does lead to a perfectly well behaved energy curve.
![\[fig3:pav\] (Color online) Same as Figure \[fig1:pav\] when the SLy4 effective interaction is used in the particle-hole channel. The dashed line corresponds to the non-regularized MR-EDF result directly obtained by gauge angle integration using 199 points in the discretization. The filled circles correspond to the result obtained using the Symmetry-Conserved EDF using Eq. (\[eq:edftotprojcor\]) and the prescription (\[eq:nidnini\]). ](fig3pav.eps){width="8.cm"}
In our approach, the main difference between effective interactions that depend on non-integer powers of the density and those depending only on integer powers of the density, is that while in the latter case one might eventually recognize terms coming from three-body or four-body forces and so on, this is impossible in the former case. It should, however, be kept in mind that the use of effective interactions should be regarded more as a guidance for the set-up of the actual form of the functional, and not as a prerequisite for the functional approaches as such.
The example of non-integer powers of the density shows that functional theory including pairing and particle number restoration and extending the usual SR-EDF approach, but without using the MR-EDF framework, can eventually be defined for rather general class of functionals if the strategy to construct the functional proposed here is followed. Let us add a few remarks:
- It is important to realize that for particle-number projection the present strategy becomes equivalent to the MR-EDF one when the regularization is slightly modified compared to the one originally proposed in Ref. [@Lac09a], i. e. the present strategy and the modified regularized MR EDF calculation will give the same energy for regularizable functionals. For those, it should therefore be more regarded as an alternative way of implementing MR-EDF approach to particle-number projection than as a new framework.
- With the present strategy, one will never have practical difficulties in applying the method to rather general and complex forms of functionals. However, some effort has been made recently to outline the constraints that a functional should fulfill to be truly regarded as a symmetry-conserving functional [@Dug10]. While these constraints are even partially unknown, one might anticipate that they will significantly restrict the functional form that might be used. We are therefore facing the following dilemma: from condensed matter physics, we know that the powerfulness of DFT comes from the large flexibility in choosing the functional form. Putting too many formal constraints will make it increasingly difficult to model the relevant physics with a tractable functional. In particular, one can already see from Ref. [@Dug10] that a functional that fulfills the constraints elaborated there will be very close to the energy functional one obtains from an Hamiltonian.
- It should be mentioned that the use of the projected density entering effective density-dependent vertices $\bar v[\rho^N]$ has already been proposed and used in Refs. [@Egi91; @Rod10; @Rod07]. However, in those references, a hybrid approach is set up where transition densities are used in other parts of the functional, and for the restoration of spatial symmetries.[^5] It has been pointed out in Ref. [@Rob10] that such hybrid approach may lead to unphysical results when set up for the restoration of spatial symmetries. Here, the theory is completely formulated in terms of the projected one- and two-body density matrices only. An open question that has to be addressed in the future is if and how the strategy to set up the functional we propose here can be generalized to the restoration of spatial symmetries, and perhaps even more general configuration mixing without becoming numerically intractable.
Discussion and Conclusion
=========================
In this work, projection made by MR-EDF techniques including the recently proposed regularization [@Lac09a; @Ben09; @Dug09] is further analyzed for the case of particle-number restoration of quasi-particle vacua of Bogoliubov type. Starting from a simple toy functional where the two-body effective interaction is not explicitly density dependent, we show that the regularized energy can [*almost*]{} be regarded as a functional of the one- and two-body densities of the projected state. To follow the density functional methodology, we propose to slightly modify the functional such that it becomes a function of projected state densities [*only*]{}, and that the projected state becomes the intermediate trial state from which the functional and other observables are constructed. For particle-number projection, such a modification could for instance be achieved within standard MR-EDF by slightly modifying the regularization proposed in Ref. [@Lac09a] while still removing the pathologies. Such an alternative interpretation may eventually serve as a justification of MR-EDF framework within a functional approach for particle number restoration when the effective kernels are not density dependent.
As a matter of fact, most of the functional forms used nowadays do not enter into the class of functionals which can be safely used in MR-EDF. We show, however, that such functional can still be used in a symmetry restoring framework that does not make use of the MR-EDF technique, but directly formulates the theory in terms of the one- and two-body density matrices of projected product states.
This theory can be seen as a direct extension of the SR-EDF level that we proposed and is called here Symmetry-Conserving EDF approach. An illustration of the resulting projected energy is given, showing that the method could be a valuable tool for the description of the ground state of a system with pairing including the restoration of particle number even when density dependence with non-integer powers is used in the functional.
The analysis of similarities and differences between the MR-EDF theory and symmetry-conserving approaches was greatly simplified here because the original quasi-particle state and the projected state share the same canonical basis. For instance, expression (\[eq:edftotprojcor\]) only holds in the canonical basis. In the present article, the applications are restricted to projection after variation for which this equation is perfectly suited. The next the step will be the extension approach to perform variation after projection (VAP). VAP is usually solved using MR-EDF techniques by making variations with respect to the components of the original quasi-particle vacuum and not the projected state itself [@Egido1982; @Egido1982a; @She00; @Sheikh2001]. In the symmetry conserving approach, one could follow the same strategy as in the standard MR-EDF approach, i.e. perform variations of the reference state. Work in that direction is currently underway.
Last, we would like to mention that the present article only discusses the case of particle-number projection and the possibility to determine the ground-state energy. The MR-EDF technique is frequently used to restore other symmetries and to calculate excited states in a Generator-Coordinate framework. What these other configuration mixings have in common, is the fact that there does not exist a common canonical basis in which the one-body density matrices of the original and of the correlated states are simultaneously diagonal. An important point to be clarified is if and how the formalism developed here can be generalized to those more general configuration mixings, and that without becoming numerically intractable. Finally, it has to be stressed that the method proposed and explored here is not meant to replace the MR EDF framework for the description of excited states and transition moments in complex nuclei. Instead, it might provide a numerically much more efficient alternative to the MR EDF scheme when one is interested just in the ground state and its evolution, either in dynamics or thermodynamics.
We thank Thomas Duguet for stimulating discussions.
Proof of Eq. (\[eq:edftotprojcor\]) {#app:correctedmr}
===================================
To prove Eq. (\[eq:edftotprojcor\]), we have to explicitly remove terms that cause pathologies from the energy calculation as proposed in Ref. [@Lac09a]. Starting from Eq. (50) of Ref. [@Lac09a], the transition matrix elements can be expressed as $$\begin{aligned}
\label{eq:rho01rho0}
n^{0 \varphi}_i
& \equiv & n^0_i + \delta n_i \left[ \varphi \right], \nonumber \\
\kappa^{0 \varphi}_{i \bar\imath }
& \equiv & \kappa^0_{i \bar\imath } + \delta \kappa_{i \bar\imath } \left[ \varphi \right] , \nonumber \\
{\kappa^{\varphi 0 }_{i \bar\imath }}^\star & \equiv & \kappa^{0*}_{i \bar\imath } + \delta \kappa_{i \bar\imath }^{ \star}\left[ \varphi \right]\, \, \, ,\end{aligned}$$ where $n^0_i$ and $\kappa^0_{i \bar\imath }$ refer to the occupation probabilities and anomalous densities of the state $\Phi_0$. Following Ref. [@Lac09a], we decompose the energy kernels entering into the integral of Eq. ([eq:ekernel]{}) into three terms ${\cal E}^\rho$, ${\cal E}^{\rho \rho}$ and ${\cal E}^{\kappa \kappa}$ corresponding to the kinetic, mean-field and pairing terms respectively. Then, ${\cal E}^{\rho \rho}$ and ${\cal E}^{\kappa \kappa}$ can be expressed as $$\begin{aligned}
{\cal E}^{\rho \rho} & = & \frac{1}{2} \sum_{ij} \bar{v}_{ijij}^{\rho \rho} n^0_i n^0_j \nonumber \\
& +& \frac{1}{2} \sum_{ij} \bar{v}_{ijij}^{\rho \rho} \left(n^0_i \delta n_j \left[ \varphi \right] + n^0_j \delta n_i \left[ \varphi \right] \right)\nonumber \\
&+& \frac{1}{2} \sum_{ij} \bar{v}_{ijij}^{\rho \rho} \delta n_i \left[ \varphi \right] \delta n_j \left[ \varphi \right] \, ,
\label{eq:drhodrho}\end{aligned}$$ whereas $$\begin{aligned}
{\cal E}^{\kappa \kappa} & = &
\frac{1}{4} \sum_{ij} \bar{v}_{i{\bar\imath }j{\bar\jmath }}^{\kappa \kappa} \, \kappa^{0 *}_{i{\bar\imath }} \, \kappa^{0}_{j{\bar\jmath }} \nonumber \\
&+& \frac{1}{4} \sum_{ij} \bar{v}_{i{\bar\imath }j{\bar\jmath }}^{\kappa \kappa} \,
\left( \kappa^{0 *}_{i{\bar\imath }} \, \delta \kappa_{j \bar\jmath } \left[ \varphi \right] + \kappa^{0}_{j{\bar\jmath }} \delta \kappa^{\star}_{i \bar\imath } \left[ \varphi \right] \right) \nonumber \\
&+& \frac{1}{4} \sum_{ij} \bar{v}_{i{\bar\imath }j{\bar\jmath }}^{\kappa \kappa} \,
\delta \kappa^{\star}_{i \bar\imath } \left[ \varphi \right] \delta \kappa_{j \bar\jmath } \left[ \varphi \right] \, .
\label{eq:dkappadkappa}\end{aligned}$$ These expressions are the strict equivalent of the ones given in Eqns. (51-54) in Ref. [@Lac09a]. For instance, regularizations have been obtained by removing terms with $j = \bar\imath $ in the last line of Eqs. (\[eq:drhodrho\]) and Eq. (\[eq:dkappadkappa\]). Accordingly, the spurious contribution to be removed from the functional is $$\begin{aligned}
\label{eq:cor_rr} {\cal E}^{\rho \rho}_{CG} & = & \frac{1}{2} \sum_{i} \bar{v}_{ijij}^{\rho \rho} \,
\int \delta n_{i}\left[ \varphi \right] \delta n_{i}\left[ \varphi \right] {\cal N}_N({0,\varphi})d\varphi
\, ,\nonumber \\
\\
\label{eq:cor_kk} {\cal E}^{\kappa \kappa} _{CG}& = & \frac{1}{2} \sum_{i}
\bar{v}_{i{\bar\imath }i{\bar\imath }}^{\kappa \kappa} \,\int
\delta \kappa^{\star}_{i \bar\imath } \left[ \varphi \right] \delta \kappa_{i \bar\imath } \left[ \varphi \right]
{\cal N}_N({0,\varphi})d\varphi \, .\nonumber \\\end{aligned}$$ Therefore, when the regularization is included, this is equivalent to make the replacements $$\begin{aligned}
&& \int_{0}^{2\pi} \!\!\! d\varphi ~n^{0\varphi}_i ~n^{0\varphi}_i {\cal N}_N({0,\varphi}) \nonumber \\
&& \Longrightarrow \int_{0}^{2\pi} \!\!\! d\varphi ~\left(
n^{0\varphi}_i ~n^{0\varphi}_i - \delta n_{i}\left[ \varphi \right] \delta n_{i}\left[ \varphi \right] \right) {\cal N}_N({0,\varphi})
\nonumber \end{aligned}$$ and $$\begin{aligned}
&& \int_{0}^{2\pi} \!\!\! d\varphi {\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{i \bar\imath }} {\cal N}_N({0,\varphi}) \nonumber \\
&& \Longrightarrow \int_{0}^{2\pi} \!\!\! d\varphi \left(
{\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{i \bar\imath }}
- \delta \kappa^{\star}_{i \bar\imath } \left[ \varphi \right] \delta \kappa_{i \bar\imath } \left[ \varphi \right] \right) {\cal N}_N({0,\varphi}) \, , \nonumber \end{aligned}$$ in the last two terms of Eq. (\[eq:edftotproj\]).
From the equalities (\[eq:rho01rho0\]), one can deduce new interesting relationships between the projected observables. For instance, performing the gauge integration of the first equation, we obtain $$n^N_i = \int_{0}^{2\pi} \!\!\! d\varphi ~n^{0\varphi}_i {\cal N}_N({0,\varphi}) =n^0_i + \delta n_i \, ,$$ with $$\delta n_i = n^N_i -n^0_i = \int d\varphi \, \delta n_{i}\left[ \varphi \right] {\cal N}_N({0,\varphi}) \, .$$ From this, let us now re-express the different quantities entering in Eq. (\[eq:edftotproj\]) $$\begin{aligned}
\int_{0}^{2\pi} \!\!\! d\varphi ~n^{0\varphi}_i ~n^{0\varphi}_j {\cal N}_N({0,\varphi}) & = & n^0_i n^0_j \nonumber \\
&+& n^0_i \delta n_j + \delta n_i n^0_j \nonumber \\
&+&
\int_{0}^{2\pi} \!\!\! d\varphi ~\delta n_i\left[ \varphi \right] ~ \delta n_j \left[ \varphi \right] {\cal N}_N({0,\varphi}) \, ,\nonumber \end{aligned}$$ where, in the specific case $i=j$, we recognize the term that enters in the regularization to be the last one. Therefore, the term entering into the regularization of ${\cal E}^{\rho\rho}$ can be expressed as $$\begin{aligned}
&& \int_{0}^{2\pi} \!\!\! d\varphi ~\left(
n^{0\varphi}_i ~n^{0\varphi}_i - \delta n_{i}\left[ \varphi \right] \delta n_{i}\left[ \varphi \right] \right) {\cal N}_N({0,\varphi})
\nonumber \\
&& \hspace*{3.cm} = n^0_i n^0_i + 2 n^0_i \delta n_i \nonumber \\
&& \hspace*{3.cm} =n^N_i n^N_i - \delta n_i \delta n_i \, .\label{eq:funnini0}\end{aligned}$$
To derive an expression of the term entering in the regularization of ${\cal E}^{\kappa\kappa}$, one can proceed in a similar way. We first define $\delta \kappa_{i \bar\imath }^{ *}$ and $\delta \kappa_{i \bar\imath }$ through $$\begin{aligned}
\int_{0}^{2\pi} \!\!\! d\varphi ~ {\kappa^{\varphi 0 }_{i \bar\imath }}^\star {\cal N}_N({0,\varphi})& = &
\int_{0}^{2\pi} \!\!\! d\varphi \left(\kappa^{0 *}_{i \bar\imath } + \delta \kappa_{i \bar\imath }^{\star}\left[ \varphi \right] \right) {\cal N}_N({0,\varphi})\nonumber \\
&\equiv& \kappa^{0 *}_{i \bar\imath } + \delta \kappa_{i \bar\imath }^{\star} \nonumber \\
\int_{0}^{2\pi} \!\!\! d\varphi \kappa^{0 \varphi}_{i \bar\imath }{\cal N}_N({0,\varphi})
& = & \int_{0}^{2\pi} \!\!\! d\varphi \left(\kappa^0_{i \bar\imath } + \delta \kappa_{i \bar\imath } \left[ \varphi \right] \right) {\cal N}_N({0,\varphi})\nonumber \\
&\equiv& \kappa^0_{i \bar\imath } + \delta \kappa_{i \bar\imath } \, . \nonumber \end{aligned}$$ Therefore the term entering in the regularized functional is given by $$\begin{aligned}
&& \int_{0}^{2\pi} \!\!\! d\varphi ~\left(
{\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{i \bar\imath }}
-\delta \kappa_{i \bar\imath }^{ \star} \left[ \varphi \right] \delta \kappa_{i \bar\imath } \left[ \varphi \right]
\right) {\cal N}_N({0,\varphi})
\nonumber \\
&& \hspace*{1.cm} = \kappa^{0 *}_{i \bar\imath } \kappa^{0}_{i \bar\imath } + \delta \kappa^{\star}_{i \bar\imath } \kappa^0_{i \bar\imath } + \kappa^{0 *}_{i \bar\imath } \delta \kappa_{i \bar\imath } \, .\end{aligned}$$ One can then take advantage of the fact that $$\begin{aligned}
n^N_i &=& n^0_i n^0_i + 2 n^0_i \delta n_i \nonumber \\
&+&
\int_{0}^{2\pi} \!\!\! d\varphi ~\delta n_i\left[ \varphi \right] ~ \delta n_j \left[ \varphi \right] {\cal N}_N({0,\varphi}) \nonumber \\
&
+ & \kappa^{0 *}_{i \bar\imath } \kappa^{0}_{i \bar\imath } + \delta \kappa^{\star}_{i \bar\imath } \kappa^0_{i \bar\imath } + \kappa^{0 *}_{i \bar\imath } \delta \kappa_{i \bar\imath }
\nonumber \\
&+& \int_{0}^{2\pi} \!\!\! d\varphi
\delta \kappa_{i \bar\imath }^{ \star} \left[ \varphi \right] \delta \kappa_{i \bar\imath } \left[ \varphi \right] {\cal N}_N({0,\varphi}) \nonumber \end{aligned}$$ and that $$\begin{aligned}
\delta n_i\left[ \varphi \right]\delta n_i\left[ \varphi \right] = - \delta \kappa_{i \bar\imath } \left[ \varphi \right]
\delta \kappa_{i \bar\imath }^{ \star} \left[ \varphi \right].\end{aligned}$$ The first equality is nothing but Eq. (\[eq:sumrule\]), whereas the second equality can be proved by expressing $\delta n_i\left[ \varphi \right]$, $\delta \kappa_{i \bar\imath } \left[ \varphi \right] $ and $\delta \kappa_{i \bar\imath }^{ *} \left[ \varphi \right]$ directly in terms of the $u_i$ and $v_i$ of the SR-EDF theory and the gauge angle $\varphi$ starting from Eq. (72-74) of Ref. [@Lac09a]. Altogether, we obtain: $$\begin{aligned}
&& \int_{0}^{2\pi} \!\!\! d\varphi ~\left(
{\kappa^{\varphi 0}_{i \bar\imath }}^\star {\kappa^{0\varphi}_{i \bar\imath }}
-\delta \kappa_{i \bar\imath }^{ \star} \left[ \varphi \right] \delta \kappa_{i \bar\imath } \left[ \varphi \right]
\right) {\cal N}_N({0,\varphi}) = n^N_i
\nonumber \\
&& -\int_{0}^{2\pi} \!\!\! d\varphi ~\left(
n^{0\varphi}_i ~n^{0\varphi}_i - \delta n_{i}\left[ \varphi \right] \delta n_{i}\left[ \varphi \right] \right) {\cal N}_N({0,\varphi}) \nonumber \\
&& = (n^N_i (1 - n^N_i) + \delta n_i \delta n_i)\, . \label{eq:corkk}\end{aligned}$$ Combining this expression with Eq. (\[eq:funnini0\]), we finally deduce the expression (\[eq:edftotprojcor\]) for the regularized functional.
Particle number restoration with recurrence relation {#app:rec}
====================================================
A method, alternative to the the gauge-integration method is presented here to calculate the one- and two-body density matrix components of a projected product state. This method turns out to be very fast and efficient numerically.
Let us start from a quasi-particle state written in its canonical basis as $$\begin{aligned}
| \Phi_{0} \rangle & = & \prod_{i>0} \left( 1 + x_i a^\dagger_i a^\dagger_{\bar\imath } \right) | 0 \rangle \, ,
\label{eq:bcsstate}\end{aligned}$$ where $|x_i|^2 = n_i^0/ (1-n_i^0)$. The associated projected state with $N$ particles can be expressed as $$\begin{aligned}
| \Psi_N \rangle & \propto & \left( \sum_{i>0} x_i a^\dagger_i a^\dagger_{\bar\imath } \right)^N | 0 \rangle \, .
\label{eq:pbcsstate}\end{aligned}$$ Starting from these expressions, it has recently been shown [@Lac10] that the elements of the one- and two-body density matrix are given by $$\begin{aligned}
n^N_i &=& \displaystyle N |x_i|^2 \frac{I_{N-1}(i)}{I_N} \, , \label{eq:nin}\\
R^N_{i\bar\imath j \bar\jmath } &=& \displaystyle N x^*_i x_j \frac{I_{N-1}(i,j)}{I_N} ~~{\rm for} ~~ (i \neq j) \, , \\
R^N_{ijij} &=& N(N-1) |x_i|^2 |x_j|^2 \frac{I_{N-2}(i,j)}{I_N} \, ,
\label{eq:nicipbcs}\end{aligned}$$ while as already mentioned $R^N_{i\bar\imath i \bar\imath } = n^N_i$. The different coefficients entering in $n^N$ and $R^N$ are given by: $$\begin{aligned}
\left\{
\begin{array}{ccc}
I_K &=& \sum^{\neq}_{(i_1, \cdots ,i_{K})} |x_{i_1}|^2 \cdots |x_{i_{K}}|^2 \\
\\
I_K(i) &=& \sum^{\neq}_{(i_1, \cdots ,i_{K}) \neq i} |x_{i_1}|^2 \cdots |x_{i_{K}}|^2 \\
\\
I_K(i,j) &=& \sum^{\neq}_{(i_1, \cdots ,i_{K}) \neq (i,j)} |x_{i_1}|^2 \cdots |x_{i_{K}}|^2\nonumber \\
\nonumber \\
& \cdots & \nonumber
\label{eq:ikijdef}
\end{array}
\right. \end{aligned}$$ Direct use of these expressions for $K=N$ is rather difficult numerically. However, these coefficients verify simple recurrence relations that are straightforward to implement on a computer. These recurrence relations have been recently used to solve numerically the Variation After Projection (VAP) [@San08; @San09] and to set up a new functional for pairing accounting for particle-number conservation [@Lac10].
In the present work, we use the recurrence method to perform PAV within the symmetry-conserving EDF framework. In that case, a preliminary SR-EDF calculation is performed leading to a quasi-particle state given by (\[eq:bcsstate\]) with a set of $\{ x_i \}$ values. Here, we have used the ev8 code [@Bon05]. From this set, the quantities $I_{N-1}(i)$ and $I_{N}$ are evaluated via the recurrence relations $$\begin{aligned}
I_{K}(i) &=& I_{K} - (K-1) |x_i|^2 I_{K-1}(i) \nonumber\\
I_{K} &=& \sum_i |x_i|^2 I_{K-1} - (K-2) \sum_i |x_i|^4 I_{K-2}(i) \, . \nonumber\\\end{aligned}$$ with the condition $I_0 = I_0 (i) = 1 $, $I_1 = \sum_k |x_k|^2 $ and $I_1(i) = I_1 - |x_i|^2$. The occupation numbers of the projected state can then be calculated as well as the correlation components using the relation [@Lac10; @Hup10]: $$\begin{aligned}
R^N_{i\bar\imath j \bar\jmath } &=& \displaystyle x^*_i x_j \frac{n_j^N - n_i^N}{|x_j|^2 - |x_i|^2} ~~{\rm for} ~~ (i \neq j) \, , \\
R^N_{ijij} &=& \frac{|x_j|^2 n_i^N - |x_i|^2 n_j^N}{|x_j|^2 - |x_i|^2} \, ,
\label{eq:pratikpbcs}\end{aligned}$$ where for $i=j$, we have $R^N_{i\bar\imath j \bar\jmath } = n_i^N$ and $R^N_{ijij} = 0$. This method is referred to as “recurrence method” in the text.
[^1]: Note that, none of the currently used SR-EDF functionals belongs to this class as they have non-analytical density dependences. The form (\[eq:denssr\]) is the only one (restricting ourselves here to bilinear functionals) for which the recently proposed regularization applies.
[^2]: We recall that the expectation value of the two-body operator in a projected state can be written as a functional of the two-body density of this state, or, fully equivalently, as a functional of the one-body density matrices. This property does not hold for general functionals that are constructed without reference to an underlying Hamiltonian.
[^3]: This does not mean, however, that we recover a theory that is equivalent to DFT. Indeed, at this stage, the functional (\[eq:edftotprojcor\]) is still a functional of the two-body density matrix. However, as will be discussed below, for the specific case of particle-number projection, the two-body density matrix is itself a functional of the one-body density matrix.
[^4]: It should be, however, kept in mind that the present functional only takes care of the restoration of $U(1)$ gauge symmetry while others still remain broken.
[^5]: For the special case of a pure particle-number projected MR-EDF calculation, the functional used in [@Val97; @Val00; @Rod07] could be mapped on a functional of the one- and two-body density matrices of the projected state.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Static rankings of papers play a key role in the academic search setting. Many features are commonly used in the literature to produce such rankings, some examples are citation-based metrics, distinct applications of PageRank, among others. More recently, learning to rank techniques have been successfully applied to combine sets of features producing effective results. In this work, we propose the metric S-RCR, which is a simplified version of a metric called Relative Citation Ratio — both based on the idea of a co-citation network. When compared to the classical version, our simplification S-RCR leads to improved efficiency with a reasonable effectiveness. We use S-RCR to rank over 120 million papers in the Microsoft Academic Graph dataset. By using this single feature, which has no parameters and does not need to be tuned, our team was able to reach the 3rd position in the first phase of the WSDM Cup 2016.'
author:
- |
Sabir Ribas\
\
\
Alberto Ueda\
\
\
Rodrygo L. T. Santos\
\
\
- |
Berthier Ribeiro-Neto\
\
\
Nivio Ziviani\
\
\
bibliography:
- 'ribas2016wsdmcup.bib'
subtitle: 'UFMG/LATIN at WSDM Cup 2016 '
title: |
Simplified Relative Citation Ratio\
for Static Paper Ranking
---
ACKNOWLEDGEMENTS {#acknowledgements .unnumbered}
================
This work was partially sponsored by the Brazilian National Institute of Science and Technology for the Web (MCT/ CNPq 573871/2008-6) and the authors’ individual grants and scholarships from CNPq and FAPEMIG.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a quantum circuit representation consisting entirely of qubit initialisations (I), a network of controlled-NOT gates (C) and measurements with respect to different bases (M). The ICM representation is useful for optimisation of quantum circuits that include teleportation, which is required for fault-tolerant, error corrected quantum computation. The non-deterministic nature of teleportation necessitates the conditional introduction of corrective quantum gates and additional ancillae during circuit execution. Therefore, the standard optimisation objectives, gate count and number of wires, are not well-defined for general teleportation-based circuits. The transformation of a circuit into the ICM representation provides a canonical form for an exact fault-tolerant, error corrected circuit needed for optimisation prior to the final implementation in a realistic hardware model.'
author:
- 'Alexandru Paler$^*$'
- 'Ilia Polian$^*$'
- 'Kae Nemoto$^{**}$'
- 'Simon J. Devitt$^{+,-,**}$'
bibliography:
- 'bib.bib'
title: 'A Fully Fault-Tolerant Representation of Quantum Circuits'
---
Introduction {#sec:intro}
============
Quantum computing promises speed-ups for a number of relevant computational problems. Building a scalable and reliable quantum computer is one of the challenges of modern science. As the size of quantum computers increases, the focus of interest shifts from their basic physical principles to structured design methodologies that will allow us to realise large-scale systems.
In general, quantum circuit optimisation methods are used to minimise the implementation costs like the number of gates or the number of wires [@wille2010towards]. Classical circuit optimisation assumes fixed gate lists even in the presence of gate errors, but classical circuits are more robust towards errors, whereas quantum information is fragile [@NC00 Ch. 8]. Classical gate failures are usually solved either by hardening the circuit (e.g. modifying transistor sizes), or by introducing various types of information redundancies that mitigate the failures. Gate hardening is not considered realistic in quantum computing architectures, and a feasible solution requires quantum error-correcting codes (QECC) [@devitt2013quantum]. The structure and design of QECC allows encoded quantum gates to be applied directly to the encoded quantum data.
In contrast to the classical case, the most practical implementations of QECC and fault-tolerant quantum circuits are composed of gates which are *nondeterministic* even in the absence of errors [@FMM13]. They either work correctly or require a correction, which is only determined during the execution of the circuit. Most such correction gates do not need to be dynamically included into the executing circuit, because their effect can be classically tracked through the subsequent gates [@paler2014software]. This is not true for all possible corrections occurring during the execution of a quantum circuit and some need to be actively applied to the quantum data[@FMM13]. This means that the overall circuit is dynamic, because its gate list needs to be modified during its execution based on certain measurement results. Reducing the incidence of such gates is difficult because when a fully error-corrected, fault-tolerant circuit is examined, it is exactly these measurement based corrections that appear to give quantum computing its power [@fowler2012time]. In general, fault-tolerant quantum circuits are constructed from Clifford and $T$ (Section \[sec:quant\]) gates, and the $T$ gate is the main source of the complications [@amy2013meet] for which dynamic corrections cannot be avoided.
The separation of circuit gates into Clifford and $T$ gates is generally performed at the higher level circuit design layer in order to make fault-tolerant error constructions more amenable to practical implementation. The physical mapping of these circuits to an actual error corrected architecture is then done with a specific QECC and hardware architecture in mind, preserving fault-tolerance. Fault-tolerance is understood as the set of procedures by which the cascade of quantum errors (bit and phase flips) caused by the circuit [@devitt2013quantum] is restricted allowing the underlying QECC to be effective when mapped to actual operations in a hardware model. In standard fault-tolerant constructions (those that are widely used in state-of-the-art hardware models [@devitt2009architectural; @Y13; @N14; @J13]), the only dynamic corrections needed are when we implement logical layer corrections for $T$ gates. These correctional gates are constructed using ancillae initialised into high-fidelity states (see Section \[sec:igtele\]) and gate teleportation protocols [@FMM13]. Our results are quite similar to those present in Ref. [@danos2007measurement], however this work focuses on producing a representation that is compatible with fault-tolerant error correction protocols.
The solution to having all the required corrections into the logical layer of the computation is to translate circuits into a regular representation that replaces correctional gate dynamics with the dynamics of reading and interpreting the circuit outputs. Such an approach is similar to the model of measurement based quantum computing (MBQC) [@briegel2009measurement], where a computation is solely described by the interpretation of the measurements performed on a specifically initialised quantum state. A circuit is described in this work as an $ICM$ sequence, where the $I$ part contains qubit initialisations, the $C$ part is a sub-circuit consisting entirely of CNOT gates, and the qubits are measured in the $M$ part. This work represents a separate and distinct approach from the work of [@miller2014mapping], where $NCV$ (reversible) circuits were mapped into Clifford and $T$ gate circuits, because the ICM representation is regular and consists entirely of ancillae, CNOTs and measurements.
The ICM representation is the extension of the methods presented in [@fowler2012time] to fit into the measurement based paradigm [@briegel2009measurement]. The presented algorithmic formulation will output the ICM representation for arbitrary quantum and reversible circuits. Such a formulation, although it requires an increased number of ancillae, allows us to directly synthesise fully fault-tolerant error corrected circuits for an underlying higher level circuit (including all required ancillary protocols), represents the realistic resource requirements of fault-tolerant quantum computations for state-of-the-art quantum architectures [@devitt2013requirements; @gottesman2013overhead] and provides an elegant form for further circuit optimisation techniques for QECC models such as topological codes [@PF13; @mequanics].
The paper is organised as follows: Section \[sec:quant\] offers a short introduction to quantum computing, illustrates the concepts of controlled and rotational gates, discusses the reversibility aspects of computing and the applications of information and gate teleportations. Section \[sec:materials\] details the non-deterministic resource requirements of arbitrary quantum circuits, introduces the ICM representation and presents the algorithm used for achieving it. The algorithm is benchmarked using circuits from the RevLib library and the results are discussed in Section \[sec:discussion\]. Finally, conclusions and future work are formulated.
Quantum and reversible computing {#sec:quant}
--------------------------------
*Quantum circuits* represent and manipulate information in *qubits* (quantum bits). The *quantum state* of a qubit is the vector $\ket{\psi}= (\alpha_0, \alpha_1)^T = \alpha_0\ket{0} + \alpha_1\ket{1}$. Here, $\ket0 = (1, 0)^T$ and $\ket1 = (0, 1)^T$ are quantum analogues of classical logic values 0 and 1, respectively. $\alpha_0$ and $\alpha_1$ are complex numbers called *amplitudes* with $|\alpha_0|^2 +|\alpha_1|^2 = 1$.
A state may be modified by applying single-qubit *quantum gates*. Each quantum gate corresponds to a complex unitary matrix, and gate function is given by multiplying that matrix with the quantum state. The application of $X$ gate to a state results in a *bit flip*: $X(\alpha_0, \alpha_1)^T = (\alpha_1, \alpha_0)^T$. The application of the $Z$ gate results in a *phase flip*: $Z(\alpha_0, \alpha_1)^T= (\alpha_0, -\alpha_1)^T$. The matrices of the Pauli gates $I,X,Y,Z$ are: $$\begin{aligned}
\small{
\begin{array}{cccccccccccc}
I & = &
\left( \begin{array}{cc}
1 & 0\\
0 & 1
\end{array} \right)
&
Y & = &
\left( \begin{array}{cc}
0 & -i\\
i & 0
\end{array} \right)
&
X & = &
\left( \begin{array}{cc}
0 & 1\\
1 & 0
\end{array} \right)
&
Z & = &
\left( \begin{array}{cc}
1 & 0\\
0 & -1
\end{array} \right)
\end{array}}\end{aligned}$$
Further important single-qubit quantum gates in the context of this work are $H,P,T$, where $T^2 = P$ and $P^2 = Z$. $$\begin{aligned}
\small
H = \frac1{\sqrt2}\begin{pmatrix}
1 & 1 \\
1 & -1
\end{pmatrix}\,\,
P = \begin{pmatrix}
1 & 0 \\
0 & i
\end{pmatrix} \,\,
T = \begin{pmatrix}
1 & 0 \\
0 & e^{i\frac{\pi}{4}}
\end{pmatrix}\end{aligned}$$
*Quantum measurement* is defined with respect to a basis and yields one of the basis vectors with a probability related to the amplitudes of the quantum state. Of importance in this work are $Z$- and $X$-measurements. $Z$-measurement is defined with respect to basis $(\ket0, \ket1)$. Applying a $Z$-measurement to a qubit in state $\ket{\psi}= \alpha_0\ket{0} + \alpha_1\ket{1}$ yields $\ket0$ with probability $|\alpha_0|^2$ and $\ket1$ with probability $|\alpha_1|^2$. Moreover, the state $\ket\psi$ *collapses* into the measured state (i.e. only the components of $\ket{\psi}$ consistent with the measurement result remains). $X$-measurement is defined with respect to the basis $(\ket+, \ket-)$, where $\ket{+} = \frac{1}{\sqrt2}(\ket{0} + \ket{1})$ and $\ket{-}=\frac{1}{\sqrt2}(\ket{0}-\ket{1})$.
Rotational gates
----------------
The exponentiation of the Pauli matrices results in the rotational gates $R_x$, $R_y$, $R_z$ parametrised by the angle of the rotation [@NC00 Ch. 4]. Hence the bit flip is a rotation by $\pi$ around the $X$-axis, implying that $X=R_x(\pi)$, and the phase-flip is a rotation by $\pi$ around the $Z$-axis, such that $Z=R_z(\pi)$. Furthermore, $P=R_z(\pi/2)$ and $T=R_z(\pi/4)$. The $V$ and $V^\dagger$ gates are parametrised $X$-rotations, $V=R_x(\pi/2)$. The Hadamard gate is $H=R_z(\pi/2)R_x(\pi/2)R_z(\pi/2)=PVP$.
$$\begin{aligned}
\small
R_x(\theta) =
\left[\begin{array}{c c}
\cos\frac{\theta}{2} & -i\sin\frac{\theta}{2} \\
-i\sin\frac{\theta}{2} & \cos\frac{\theta}{2}
\end{array}\right]\,\,
R_y(\theta) =
\left[\begin{array}{c c}
\cos\frac{\theta}{2} & -\sin\frac{\theta}{2} \\
\sin\frac{\theta}{2} & \cos\frac{\theta}{2}
\end{array}\right]\\
R_z(\theta) =
\left[\begin{array}{c c}
e^{-i\theta/2} & 0 \\
0 & e^{i\theta/2}
\end{array}\right]\,\,
CNOT =
\left[\begin{array}{c c c c}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 \\
0 & 0 & 1 & 0
\end{array}\right]\end{aligned}$$
Controlled gates {#sec:cgates}
----------------
An $n$-qubit circuit processes states represented by $2^n$ amplitudes, $\alpha_y$, with $y \in \{0, 1\}^n$ and $\sum_y|\alpha_y|^2 = 1$. Measuring all qubits of the circuit results in one basis vector with the probability given by the corresponding amplitude, $|\alpha_y|^2$. Quantum gates may act on several qubits simultaneously. A gate operating on $n$ qubits is represented by a $2^n \times 2^n$ complex unitary matrix. One important two-qubit gate is the *controlled-not* CNOT$(c,t)$ gate, where the $c$ qubit conditionally flips the state of the $t$ qubit when set to $\ket{1}$. In general, any quantum gate can be used in a controlled manner, and other versions are controlled-$Z$ (CPHASE), controlled-$V$ (C-$V$) and controlled-$V^\dagger$ (C-$V^\dagger$), where $V^2 = X$.
Similarly to how arbitrary classical Boolean functions can be constructed entirely from NAND gates, universal quantum computations can be constructed using a discrete set of gates. The universal gate set has to contain at least one coupling operation, and the most often used one is CNOT. A commonly used gate set in fault-tolerant quantum computing is $UGS_{ft}=\{CNOT, H, T\}$ [@NC00 Ch. 4]. There are gate sets that are not universal, an example is the Clifford gate group, generated by the gates $\{CNOT, H, P\}$. Circuits comprised of gates exclusively from the Clifford group can be efficiently simulated on a classical computer [@GKtheorem], but the Clifford group together with the $T$ gate is quantum universal. The $T$ gate is one of the most expensive quantum gates to implement when QECC and fault-tolerant computation is taken into account [@devitt2013requirements; @gottesman2013overhead]. Thus, there is ongoing research into reducing the $T$ gate count of synthesised quantum circuits [@jones2012novel; @amy2013meet; @miller2014mapping].
Reversibility
-------------
The linearity of quantum mechanics has the effect that information can not be erased, therefore, for an arbitrary computation, the number of input qubits equals the number of output qubits. *Reversible circuits*, as presented in [@wille2010towards; @saeedi2013synthesis], are the result of enforcing this requirement on classical Boolean circuits. The interest in classical reversible computing was initially motivated by Landauer’s principle, which states that the erasure of information is dissipating energy [@moore2012computing]. The hope was that computers might become more energy-efficient if classical computations would be reversible. Therefore, FANINs and FANOUTs are not allowed into the circuits. The majority of the classical gates are not linear maps. For example the inputs $a$ and $b$ of the $AND(a,b)=c$ gate are impossible to infer from the output $c$. However, the $NOT$ gate is reversible because its output is the negation of the input, and no information is erased.
The reversibility of classical circuits is achieved by the *Toffoli* gate (Fig. \[circ:toffoli\]), operating on three bits, where two of them control the bit-flip of the third: *toffoli*$(a,b,c)=(a,b,c \oplus ab)$. Arbitrary classical circuits can be completely constructed using Toffoli gates [@NC00 Ch. 3]. While a quantum Toffoli performs effectively the same transformation on qubits, the key difference between quantum and reversible circuits is that the Toffoli gate is not universal for quantum computations because universality also require at least the $H$ gate [@aharonov2003simple]. Reversible circuits can be considered restricted quantum circuits operating only on computational basis states. However, it is possible to decompose the Toffoli gate into quantum gates (Fig. \[circ:toffoli\] and Fig. \[circ:toffoli2a\]). One decomposition (the quantum version) uses the gate set $\{CNOT, H, T\}$ ($T^\dagger=T^7$), while a second decomposition uses the gates $\{CNOT,V,V^\dagger\}$. The second representation will be called *the reversible version* (although the $V$ gate is quantum), because its lower gate cost makes it widely used in the designs of reversible circuits [@saeedi2013synthesis; @wille2010towards], although these costs generally don’t account for the true nature of error corrected quantum circuits.
Information and Gate Teleportation {#sec:igtele}
----------------------------------
Quantum information (qubit states) cannot be copied [@WZ82], but there are ways to *move* information from one qubit to another through quantum state *teleportation* (Fig. \[circ:teleport2\]) [@teleport]. The most general teleportation technique [@NC00 Ch. 4] is implemented using a slightly different mechanism, but quantum computing models and architectures like [@devitt2009architectural; @N14; @J13; @Y13; @FMM13] use the two circuits presented herein.
Each of the circuits requires an ancilla initialised into either $\ket{0}$ or $\ket{+}$. For the first circuit, after applying the CNOT on the states $\ket{\psi}=a\ket{0}+b\ket{1}$ and $\ket{0}$, the two-qubit state will be $a\ket{00}+b\ket{11}$. The measurement of the input qubit, in the $X$-basis is probabilistic, and depending on its result the final state of the ancilla will be either $\ket{\psi_1}=a\ket{0}+b\ket{1}$ if $\ket{+}$ is measured, or $\ket{\psi_2}=a\ket{0}-b\ket{1}$ for $\ket{-}$. The execution of the second circuit, where instead a $Z$-basis measurement is used, will result in the state of the ancilla being $\ket{\psi_3}=a\ket{0}+b\ket{1}$ after measuring $\ket{0}$, or $\ket{\psi_4}=a\ket{1}+b\ket{0}$ after measuring $\ket{1}$. For both teleportations the final state is the desired one with $50\%$ probability ($\ket{\psi_1}$ and $\ket{\psi_2}$), while otherwise correctional gates are required, because $\ket{\psi_1}=Z\ket{\psi_2}$ and $\ket{\psi_3}=X\ket{\psi_4}$. The corrections are a direct result of the measurements being probabilistic. The correction mechanism is illustrated in the circuit diagrams by the double vertical lines connecting the measurements to the $X/Z$ gates, indicating a classically controlled gate of either $X$ or $Z$.
Information teleportation is a linear transformation of the destination qubit, such that its state is exactly the state of the source, but quantum gates are linear transformations, too. It follows that it is possible to construct teleported versions for single-qubit quantum gates. Such constructs are commonly used in the fault-tolerant implementation of quantum gates. The teleportation-based gate circuits for the $V$, $T$ and $P$ gates are shown in Fig. \[circ:ftcircs\]. The teleportations are again probabilistic and the output state requires corrections (derived in [@paler2014software]). Gate teleportations are based on *magic states* [@bravyi2005universal] like $\ket{Y}=\frac{1}{\sqrt{2}}(\ket{0}+i\ket{1})$ and $\ket{A}=\frac{1}{\sqrt{2}}(\ket{0}+e^{i\frac{\pi}{4}}\ket{1})$. The utilization of magic states and the above teleportation circuits is that they can be implemented using fault-tolerant QECC through a process known as state distillation [@bravyi2005universal; @FMM13], which accounts for the majority of resources necessary for a large-scale error corrected algorithm [@devitt2013quantum].
The $R_z(\pi/4)^\dagger=T^\dagger = R_z(-\pi/4)$ rotation is implemented using the same circuit as the gate $T$, the only difference being the interpretation of the measurement result in terms of any subsequent correction. Because the $T$,$T^\dagger$,$V$ and $V^\dagger$ gates can be implemented by teleportations, it follows that the Toffoli gate (in both its quantum and reversible versions) can be decomposed into teleportation sub-circuits.
The magic states in the construction of fault-tolerant gates are assumed to be high-fidelity (As high as the fidelity of the underlying quantum information protected by the QECC). Otherwise, high-fidelity instances are obtained after *distilling* multiple low-fidelity states using circuits consisting entirely of CNOTs and measurements [@bravyi2005universal]. For example, the distillation of a single $\ket{Y}$ state from low-fidelity $\ket{Y}$ ancillae is reported in [@bravyi2005universal], reducing the infidelity, $p$, of the output from $O(p)$, $p<1$, of the seven inputs to $O(p^3)$ on the output.
The ICM Representation {#sec:materials}
======================
In state-of-the-art fault-tolerant quantum circuits, two sources of non-determinism can be distinguished. First, errors can occur during calculation due to undesired interaction with the environment. The errors are handled by quantum error-correcting codes [@devitt2013quantum]. Second, as mentioned above, the realisation of gates by teleportation is inherently probabilistic. The outcome of the gate application is correct with 50% probability and requires a correction with 50% probability even in absence of errors.
Circuit gate dynamics, as presented in Section \[sec:intro\], is the consequence of applying specific quantum gates (e.g. $T$) by teleportation. Correctional gates may or may not be required, depending on the outcome of a measurement that is only available when the circuit is being executed. A further source of non-determinism is error-correction, which is not considered herein and is handled at a lower level in the overall design stack of a quantum circuit [@FMM13].
A circuit with a dynamic gate list is difficult to execute on a quantum computer, and is furthermore difficult to optimise. This section introduces the ICM representation, which replaces the non-deterministic gate dynamics with an exact gate list. The resulting circuit still contains correctional mechanisms, but these are controlled by measurement results of introduced ancillae and active feedforward determining subsequent measurement choices. We essentially fan-out using extra ancillae to remove the complication of dynamic circuit construction with fault-tolerant and reversible quantum circuits.
Non-deterministic Resource Requirements {#sec:nondet}
---------------------------------------
Gate corrections may or may not be required after each teleportation. They consist in applying $X$, $Y$, $Z$ or $P$ gates to the calculated result. Therefore, the total number of gates in the circuit depends on the number of corrections, and this number is not known *a priori* because the need for corrections is determined only during circuit execution (each individual teleportation has a 50:50 chance of each ancilla measurement result, so the possibilities grow exponentially in the number of teleported gates). Moreover, corrections require an introduction of additional ancillae qubits, thus making the computation total number of qubits unpredictable as well.
It can be shown that $X$, $Y$ and $Z$ corrections (Pauli corrections) do not have to be addressed immediately in a quantum way after an unsuccessful gate application. Instead they can be postponed to the end of calculation using *Pauli tracking* [@paler2014software] and instead of applying an active quantum gate to the data, we simply reinterpret the meaning of the classical measurement results. However, this technique does not apply to $P$ corrections necessary for implementing the $T$ gate (Section \[sec:igtele\]). This is because the $P$ correction does not commute through either the $H$ gate of the target of a $CNOT$ gate in a straightforward manner and changes the probability distribution of subsequent $X$-basis measurements.
For example, in the teleported $T$ gate (Fig. \[circ:ftt\]), applying a CNOT on two qubits $\ket{t}=\frac{1}{\sqrt{2}}(\ket{0}+r\ket{1})$ (where $r=e^{\frac{i\cdot\pi}{4}}$) and $\ket{q}=a\ket{0}+b\ket{1}$ results in $\ket{qt}=(a\ket{00}+ar\ket{11}+b\ket{10}+br\ket{01})/\sqrt{2}$. The $\ket{0}$ result of the first qubit’s $Z$-measurement will result in the second qubit’s state as if it were directly rotated by $T$: $a\ket{0}+br\ket{1}$. If the measurement result is $\ket{1}$, the state is $ar\ket{1} + b\ket{0}$, which after a $PX$ correction is required [@paler2014software], and it can be applied using the circuit from Fig. \[circ:ftp\].
The $P$ correction requires us to dynamically change the circuit being executed as this correction cannot be classically tracked. A second ancilla is introduced in the $\ket{Y} = \ket{0}+i\ket{1}$ state, a CNOT applied between the ancilla and the state to be corrected, and the input is measured according to Fig. \[circ:ftp\]. For an $n$-qubit circuit $C$ with a gate list $GL(C)$, each probabilistic $P$ correction increments the number of qubits by one, and inserts a $P$ gate into the gate list.
The problem of applying the $P$ gate dynamically is solved by introducing into the circuit the possibility to operate both a teleported identity gate, used when no correction is needed, and a teleported $P$ gate. Similarly to a classical demultiplexer the measurement result of the teleported $T$ gate is used to decide, at run-time, whether $I$ or $P$ gate is applied. Finally, after performing either the $I$ or $P$ correction, the corresponding state has to be routed to a single qubit. This is realised by classically controlled teleportations in a manner similar to a classical multiplexer with the select signal being the measurement result of the teleported $T$ gate. Classically controlled teleportations were described in [@fowler2012time], and a circuit using these mechanisms will have a fixed number of qubits and a determined gate list. Compared to a dynamically changing circuit, these are larger, but the predictability of these parameters is useful for circuit optimisation.
For *selective destination teleportation* (Fig. \[circ:selective\]) the first group of measurements ($Z_1X_2$ where the subscripts indicate the qubit’s number) will teleport $\ket{\psi}$ on the third qubit where it will be corrected by $P$. The second group of measurements ($X_1Z_2$) will teleport the state to the fourth qubit where the trivial correction $I$ is applied, thus leaving the state unchanged. In the *selective source teleportation* the $X_1Z_2$ measurements will select $\ket{\psi_1}$ for teleportation on the third qubit, while the second measurement group ($Z_1X_2$) will teleport $\ket{\psi_2}$ [@fowler2012time]. The selective teleportation circuits require only Pauli corrections, which are not shown in the diagrams, because their application can be postponed to the end of the computation and classically tracked.
As a consequence, Pauli tracking can reduce but not completely eliminate the non-determinism of fault-tolerant circuits. This implies that standard synthesis methods which optimise gate count and/or number of qubits are not applicable to teleportation-based quantum circuits because these numbers are not well-defined. It is possible to circumvent the non-determinism by using “conditional-identity construction” which results in the maximal possible number of gates. The initial gate dynamics of a circuit, with all the classically controlled corrections replaced by classically controlled teleportations, is interpreted as the dynamics of the measurements .
@C=.5em @R=.4em [ & & & & & &&&&&&&&& &\
& & & & & & &&&&&&&& &\
& & & & && &&&&&&&&&\
& && & & & &&&&&\
\
&&(a)&&&&&&&(b) ]{}
ICM Correctness and Construction
--------------------------------
The role played by the structured representation of circuits was recognised in [@aaronson2004improved], where stabiliser circuits were decomposed into a canonical sequence of sub-circuits constructed from a single type of gates. In the context of fault-tolerant quantum computing, the systematic derivations of the circuits [@zhou2000methodology] uses teleportation sub-circuits, too. However, the combination of the fault-tolerant constructions with the regular gate decompositions [@shende2006synthesis], required for efficient synthesis algorithms, is limited by the realistic requirements of future quantum computing architectures. Nevertheless, structured mapping techniques between various architectures were investigated in [@hirata2011efficient; @cheung2007translation]. These approaches were targeted at specific quantum hardware properties, such as nearest-neighbour interaction between qubits, but fault-tolerant constructions were not specifically addressed.
ICM is a structured representation, which consists in the regular representation of arbitrary quantum and reversible circuits using the $UGS_{ft}$ gate set, where the single-qubit rotational gates are teleportation-based. Circuits are transformed into the ICM representation after decomposing all non-$UGS_{ft}$ gates into $UGS_{ft}=\{CNOT, H, T\}$ component gates, and simultaneously introducing, where necessary, selective source and destination teleportation circuits into the resulting circuit.
The correctness of the ICM representation is based on the observation that the teleported gate circuits (Figs. \[circ:ftp\],\[circ:ftt\] and \[circ:ftv\]) and the selective teleportation circuits (Fig. \[circ:selective\]) consist entirely of qubit initialisations, CNOT gates and qubit measurements. Thus, decomposing an arbitrary circuit into elements that can be expressed entirely using the above mentioned sub-circuits, will consist only of initialisations, CNOTs and measurements. The circuit from Fig. \[circ:selective\]a can be rewritten, such that the $P$ gate will not be directly applied: in general, $R_z$ rotations (e.g. the $P$ gate) commute with the control of CNOT gates [@NC00 Ch. 4]. As a result, the $P$ gate can be moved on the left side of the CNOT, and $P\ket{+}=\ket{Y}$. The third qubit from Fig. \[circ:selective\]a will be initialised into $\ket{Y}$ instead of $\ket{+}$.
The ICM representation of an arbitrary quantum circuit is the result of applying algorithm presented in this paper. The algorithm is taking a circuit composed of gates from the set $\{$*Toffoli*, $CNOT, C\mbox{-}V$ and $C\mbox{-}V^\dagger, H, P, T\}$, and performs pattern replacements resulting in the circuit $CICM$ (Line 1) consisting of gates from $UGS_{ft}$. The Toffoli gates are decomposed into single qubit rotations (either $\{V,V^\dagger\}$ or $\{H,T,T^\dagger\}$) and CNOT gates. The Hadamard gates are replaced with the series of $Z$- and $X$-axis rotational gates ($P$ and $V$ gates). Afterwards, each $P$ and $V$ gate is replaced using the corresponding teleportation-based gate implementations from Figs. \[circ:ftp\],\[circ:ftt\],\[circ:ftv\]. The effect of replacing a gate $G$ acting on qubit $i$ is that an ancilla is introduced on the position $i+1$. Thus, all the gates following the initial application of $G$ on $i$ are moved to $i+1$ (Line 20).
The ICM representation is obtained by *moving* all the single-qubit measurements to the end of the circuit, and all the ancillae initialisations to the beginning of the circuit. The middle part of the resulting circuit consists entirely of CNOT gates. The single qubit measurements are then temporally staggered (e.g. Fig. \[fig:papert1\]), such that the results of previous measurements determine the basis choices for subsequent measurements to teleport data to pre-prepared ancillae. In the case of the teleported $T$ gate, this procedure dictates to either apply $P$ gate corrections or not, as required.
Resource Analysis
-----------------
Transforming arbitrary quantum and reversible circuits into the ICM representation requires the introduction of supplemental ancillae, CNOT gates and measurements. The obtained representation is an augmented version of the initial circuit, and there is a constant resource overhead associated with each gate transformation. In the following the gate cost of implementing a sub-circuit (gate) $S$ is represented by $gc(S)$, and the ancilla cost is denoted $ac(S)$.
### Theorem:
The ICM representation of a quantum circuit $C$ with $n_T$ $T$ gates, $n_P$ $P$ gates, $n_V$ $V$ gates, $n_H$ Hadamard gates and $n_{Tf}$ Toffoli gates requires $ac(C)=5n_T + n_P + n_V + 3n_H + 42n_{Tf}$ ancillae and $gc(C)=6N_T + n_P +n_V + 3n_H + 55n_{Tf}$ additional gates.
### Proof:
The central quantum gate is $T$, which requires $ac(T)=5$ ancillae and $gc(T)=6$ CNOTs. One of the ancillae is the one initialised into $\ket{A}$, three other ancillae are used for the selective destination teleportation sub-circuit, and, finally, the fifth ancilla is introduced for the selective source teleportation and represents the output of the teleported $T$ gate.
The $P$ and the $V$ gates introduce a single ancilla $ac(P)=ac(V)=1$ initialised into the $\ket{Y}$ state, and because the teleportation circuits require a single CNOT $gc(P)=gc(V)=1$. The Hadamard gate being implemented as a sequence of $P$ and $V$ gates generates a gate cost of $gc(H)=3gc(P)=3$, and an ancilla cost of $ac(H)=3ac(P)=3$.
The quantum version of the Toffoli gate (denoted *Toffoli*$_q$) decomposition contains $6$ CNOTs, $7$ $T$ gates, one $P$ and two $H$ gates (Fig. \[circ:toffoli\]), and thus $gc($*Toffoli*$_q)=6+7gc(T)+(1+2\times 3)gc(P)=55$ and $ac($*Toffoli*$_q)=7ac(T)+(1+2\times 3)ac(P)=42$.
### Note:
The Theorem was formulated for the ICM decomposition of quantum Toffoli gates, but can easily be updated to include the reversible version of these gates (in the following denoted *Toffoli*$_2$). These gates are decomposed into quantum gates, and the initial version contains $2$ CNOT gates and $3$ controlled-$V$ gates (denoted by $CV$), which are further decomposed (Fig. \[circ:toffoli2b\]) into 2 Hadamard gates, 3 $T$ and 2 CNOTs. Therefore, because $gc(CV)=2gc(H)+3gc(T)+2=26$ and $ac(CV)=2ac(H)+3ac(T)=21$, the gate cost of the reversible Toffoli is $gc($*Toffoli*$_2)=3gc(CV)+2=80$ and the ancilla cost $ac($*Toffoli*$_2)=3ac(CV)=63$.
$$\begin{aligned}
\scriptsize
\begin{aligned}
& \textbf{Require: } \text{Circuit } C \text{ composed from } \{Toffoli, CNOT, H, P, T\}\\
&\text{1: } \text{Circuit }CICM \gets C\\
&\text{2: } \text{Replace in } CICM \text{ the Toffoli gates with their decomposition (Figure~\ref{circ:toffoli} or Figure~\ref{circ:toffoli2a}})\\
&\text{3: } \text{Replace in } CICM \text{ the } H \text{ gates with }PVP\\
&\text{4: } \textbf{forall } P \text{ gates in }CICM\\
&\text{5: } \quad\text{Introduce the ancilla } a_p \text{ below the qubit having }P\\
&\text{6: } \quad\text{Construct the circuit for the teleported $P$ gate}\\&\text{7: } \quad\text{Move all the gates following the initial } P \text{ onto }a_p\\
&\text{8: } \textbf{endfor}\\
&\text{9: } \textbf{forall } V \text{ gates in } CICM\\
&\text{10: } \quad\text{Introduce the ancilla }a_v \text{ below the qubit having }V\\
&\text{11: } \quad\text{Construct the circuit for the teleported $V$ gate}\\&\text{12: } \quad\text{Move all the gates following the initial }V \text{ onto }a_v\\
&\text{13: } \textbf{endfor}\\
&\text{14: } \textbf{forall } T \text{ gates in }CICM\\
&\text{15: } \quad\text{Introduce the ancilla } a_c \text{ below the qubit having }T\\
&\text{16: } \quad\text{Construct the circuit for the teleported $T$ gate}\\&\text{17: } \quad\text{Introduce }4 \text{ ancillae below the previous ancilla}\\
&\text{18: } \quad\text{Construct the selective destination circuit where } a_c \text{ corresponds to the first qubit, }\\
&\text{and } s_3 \text{ and } s_4 \text{ are the third and fourth qubits respectively}\\
&\text{19: } \quad\text{Construct the selective source circuit where } s_3 \text{ corresponds to the first qubit, }\\
& s_4 \text{ to the second qubit, and } a_{out} \text{ is the third qubit}\\
&\text{20: } \quad\text{Move all the gates following the initial } T \text{ onto the ancilla } a_{out}\\
&\text{21: } \textbf{endfor}\\
&\text{22: } \textbf{return } CICM\\
\end{aligned}
\label{alg:1}\end{aligned}$$
State distillation (see Section \[sec:igtele\]) is not analysed here, as it is an intrinsic requirement for any type of computation where magic states are required. An exhaustive and complete analysis of the distillation circuits overhead is presented in [@devitt2013requirements] and, as a consequence, the present ICM resource analysis is a continuation of that work.
Discussion {#sec:discussion}
==========
The ICM representation of an arbitrary circuit prepared into a fault-tolerant manner will not affect its properties. Therefore, fault-tolerance statistics will not be discussed. The results of executing the implementation of Algorithm \[alg:1\] on circuits from the RevLib benchmark are presented in Table \[tbl:res\]. The *EQ* circuits consisted of gates from the set $\{CNOT, C\mbox{-}V, C\mbox{-}V^\dagger\}$ and the *NCT* circuits from the set $\{$*Toffoli*$,CNOT, X\}$. The best-case non-ICM representation consists of the teleportation-based gate construction where no $P$ corrections is required for the $T$ gate. The worst-case non-ICM scenario assumed that all the $T$ gates require the $P$ correction. For other types of gates the corrections can be tracked through the circuit [@paler2014software], but tracking is not possible for the probabilistic $P$-correction (see Section \[sec:nondet\]). In order to illustrate the benefit of the ICM representation the time required for executing the critical path of the decomposed circuits was computed. The model presumed a time cost of $10$ for initialisations, and a cost of $1$ for the CNOTs and the measurements.
It can be seen that the time required by ICM circuits is predictable and better than the worst-case time of circuits before transformation. Note that longer time translates to higher decoherence and more stringent requirements on quantum error-correction.
Example
-------
The systematic transformations of the $T$ gate and of the controlled-$V$ gate decomposition from Fig. \[circ:toffoli2b\] are presented after applying Algorithm \[alg:1\] and obtaining a circuit composed from $UGS_{ft}$ (see Section \[sec:cgates\]). The ICM representation of the $T$ gate (Fig. \[fig:papert1\]) takes the $\ket{in_0}$ qubit, and after performing the CNOT with the $\ket{A}$ ancilla, selectively teleports (the leftmost group of gates) the intermediary state to either the fourth or the fifth qubit.
The measurement of the first qubit ($Z_1$) is followed by either the measurement pattern $Z_2X_3$ if the result of the teleported $T$ needs a $P$ correction, or the measurement pattern $X_2Z_3$ if the result was correct up to Pauli corrections. The correctness of the teleported gate application is indicated by the measurement result. Applying the $Z_2X_3$ pattern teleports the intermediary state on the output qubit marked by $\ket{out_0}$, and the fourth and fifth qubits are measured using $X_4Z_5$. Otherwise, the measurement $Z_4X_5$ will result in teleporting the state of the fifth qubit on the sixth qubit. The measurement of specific qubit groups depends on the results of previous measurements.
The controlled-$V$ gate ICM representation (Fig. \[fig:papert2\] after applying Algorithm \[alg:1\]) has the input states $c_{in}$ (control) and $t_{in}$ (target) and outputs $c_{out}$ and $t_{out}$. The individual decomposition of the single-qubit gates from Fig. \[circ:toffoli2b\] is highlighted by the dashed bounding boxes. The boxes containing three CNOTs are implementations of the Hadamard gate where for each constituent sub-gate a CNOT and a $\ket{Y}$-qubit are used. The ancillae introduced by the ICM transformation are affecting the distance between the control and the target of the initial CNOTs (not marked by bounding boxes). The order of the measurements is dictated by the temporal order of the bounding boxes, meaning that the measurements implementing the leftmost $T$ and $H$ can be applied in parallel. Afterwards, the measurements associated to the middle bounding boxes can be again executed in parallel. Finally, the last Hadamard gate from the initial circuit is applied by measuring the last three qubits.
![The ICM representation of the controlled-$V$ gate. There are three ICM $T$-gate applications (see Fig. \[fig:papert1\]) and two ICM Hadamard applications (marked by bounding boxes in which three ancillae are measured using the $ZXZ$ pattern)](papert2){width=".6\textwidth"}
. \[fig:papert2\]
Conclusion {#sec:conclusion}
==========
The usual assumptions made for quantum optimisation techniques do not necessarily hold for the fault-tolerant circuits because of their inherent dynamicity. A regular representation of quantum and reversible circuits was presented starting from the fault-tolerant implementation of quantum circuits. The ICM representation is a consequence of the results presented in [@fowler2012time; @paler2014software] and has the potential, when combined with the synthesis method from [@amy2013meet; @amy2014polynomial], to be used for future circuit optimisation techniques.
The results indicate that, while making a quantum circuit fault-tolerant significantly increases its gate count and the number of required ancilla qubits, the ICM representation outperforms direct mapping without enforcing the ICM condition with respect to both predictability and worst-case execution time. The major advantage of this representation is that it produces a deterministic circuit description for a higher level circuit. A deterministic description is essential to allow for more global circuit optimisations in various error corrected implementations. Future work will investigate quantum circuit synthesis, optimisation and validation techniques based on the ICM representation.
Acknowledgements {#acknowledgements .unnumbered}
================
SJD acknowledges support from the JSPS Grant-in-aid for Challenging Exploratory Research, NICT, Japan and JSPS KAKENHI Kiban B 25280034.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a search for new members of the 300 [kms$^{-1}$]{} stream (300S) near the dwarf galaxy Segue 1 using wide-field survey data. We identify 11 previously unknown bright stream members in the APOGEE-2 and SEGUE-1 and 2 spectroscopic surveys. Based on the spatial distribution of the high-velocity stars, we confirm for the first time that this kinematic structure is associated with a 24-long stream seen in SDSS and Pan-STARRS imaging data. The 300S stars display a metallicity range of $-2.17 < {\rm [Fe/H]} < -1.24$, with an intrinsic dispersion of 0.21$_{-0.09}^{+0.12}$ dex. They also have chemical abundance patterns similar to those of Local Group dwarf galaxies, as well as that of the Milky Way halo. Using the open-source code *galpy* to model the orbit of the stream, we find that the progenitor of the stream passed perigalacticon about 70 Myr ago, with a closest approach to the Galactic Center of about 4.5 kpc. Thus, 300S may be the most recent Milky Way satellite to be tidally disrupted. Using Pan-STARRS DR1 data, we obtain an integrated stream luminosity of $4 \times 10^3$ L$_{\odot}$. We conclude that the progenitor of the stream was a dwarf galaxy that is probably similar to the satellites that were accreted to build the present-day Milky Way halo.'
author:
- Sal Wanying Fu
- 'Joshua D. Simon'
- Matthew Shetrone
- Jo Bovy
- 'Timothy C. Beers'
- 'J. G. Fern[á]{}ndez-Trincado'
- 'Vinicius M. Placco'
- Olga Zamora
- Carlos Allende Prieto
- 'D. A. Garc[í]{}a-Hern[á]{}ndez'
- Paul Harding
- Inese Ivans
- Richard Lane
- Christian Nitschelm
- 'Alexandre Roman-Lopes'
- Jennifer Sobeck
bibliography:
- 'bibliography.bib'
title: 'The Origin of the 300 km s$^{-1}$ Stream Near Segue 1'
---
Introduction
============
The $\Lambda$CDM cosmological model posits that larger galaxies form via the hierarchical mergers of smaller galaxies [e.g., @searle1978compositions; @white1978]. Simulations using $\Lambda$CDM predict that hierarchical merging processes form stellar substructures, remnants of individually tidally disrupted dwarf satellite galaxies, in the halos of larger galaxies [e.g., @bullock2001hierarchical; @bullock2005; @cooper2010]. The Milky Way halo, which we can study in great detail, encodes the history of tidal accretion events in the form of stellar streams and other substructures. Studying these streams can provide clues to the Milky Way’s recent formation history, such as the timescale of accretion events, as well as the type of objects that were accreted to form the Galaxy’s halo [e.g., @helmi1999; @zolotov2010; @bonaca2012cold; @tissera2014a; @koposov2014discovery; @pillepich2015; @guglielmo2017origin; @kupper2017exploding].
Multiple spectroscopic studies of the dwarf galaxy Segue 1 have uncovered a distinct population of stars with a heliocentric recessional velocity of $\sim300$ [kms$^{-1}$]{} in the same field (@geha2009least [henceforth G09], @norris2010chemical [henceforth N10], @simon2011complete [henceforth S11]). S11 suggested that these stars belong to a stellar stream (henceforth 300S) rather than a compact object due to the stars’ diffuse positions over their $\sim0.25\degr$ survey area. @frebel2013300 [henceforth F13] studied the chemical abundances of the brightest star in the stellar stream and suggested that it is similar to halo stars; this means that the stream’s only distinct signature from the halo is its 300 [kms$^{-1}$]{} heliocentric velocity. If that is the case, then 300S may be representative of a large number of similar objects that were accreted at earlier times. However, the nature of the progenitor of 300S is currently unknown.
From SDSS photometric data, [@niederste2009origin] used a matched-filter technique to discover an elongated structure in the vicinity of Segue 1, extending about 4$\degree$ east-west. S11 suggested that this feature could be the photometric counterpart of 300S. @bernard2016synoptic [henceforth B16] used Pan-STARRS photometry to trace the same feature over a wider area of the sky, showing that it extends spatially over the range $144\degr < \mbox{RA} < 168\degr$ (also see @grillmair2014). Follow-up spectroscopic observations over this patch of the sky are crucial for determining whether the photometric structure and the kinematic structure are linked.
All-sky surveys are optimal for studying stellar streams because of (1) their ability to detect and map out the full extent of the stream, and (2) their ability to provide photometric and spectroscopic data that allows for further characterization of stellar streams. In particular, radial velocities and proper motions allow us to model the orbit of the stream, providing insight into the stream’s tidal disruption history. All-sky, high-resolution spectroscopic surveys such as APOGEE-2 can also yield detailed insights into the chemical evolution of the stream progenitor, and discern whether the progenitor is a globular cluster or a dwarf galaxy.
In this study, we present an analysis of members of 300S found in APOGEE-2 and SEGUE data. Section 2 describes the member selection process. In Section 3, we compare the chemical abundances of 300S to those of other Milky Way populations. In Section 4, we model the orbit of the stream and discuss its tidal disruption history. In Section 5, we discuss the nature of the 300S progenitor and infall scenarios. In Section 6, we summarize our results and present our conclusions.
Stream Member Selection {#sec:new_members}
=======================
Stream Members in APOGEE-2 {#sec:apogee_members}
--------------------------
The Apache Point Observatory Galactic Evolution Experiment [APOGEE; @majewski2015apache] survey obtained high signal-to-noise ratio ($\mbox{S/N}~\gtrsim 100$), high-resolution ($R \approx 22,500$) near-infrared ($1.51-1.70~\mu$m) spectra of 146,000 stars in and around the Milky Way with the Sloan Digital Sky Survey 2.5 m telescope [@gunn2006]. After processing with the APOGEE data reduction pipeline [@nidever2015], stellar atmospheric parameters such as temperature, metallicity, and surface gravity are determined by the APOGEE Stellar Parameter and Chemical Abundances Pipeline [ASPCAP; @garciaperez2016aspcap]. The APOGEE-2 project, part of the fourth-generation Sloan Digital Sky Survey [SDSS-IV; @blanton2017sloan], is extending APOGEE observations to a larger sample of stars (270,000 in DR14; @dr14). This extended sample includes 9 dwarf galaxies out to distances of $\sim100$ kpc, and covers the southern hemisphere [@majewski2016; @zasowski2017]. Several APOGEE-2 plug-plates span the location of the photometric features possibly associated with 300S, and provide the potential to study the chemical evolution history of the stream progenitor.
To search for potential 300S members, we begin with the APOGEE-2 allstars file, which was released in SDSS DR14 [@dr14]. This file includes the spectra of the original APOGEE survey, re-reduced and re-analyzed with the same software used on APOGEE-2 data. To isolate potential stream stars, we apply a velocity cut of $275~\mbox{{km\,s$^{-1}$}} < V_{\rm helio} < 325$ [kms$^{-1}$]{}, and a spatial cut of $145\degree < \alpha_{2000} < 170\degree$, $10\degree < \delta_{2000} < 22\degree$. Given the stream velocity dispersion measured by S11, this criterion is quite generous ($3.5\sigma$), but it allows for a possible velocity gradient along the stream. As it turns out, none of the member stars we identify are near the edge of the velocity selection window, so the exact limits chosen do not affect our results. Our spatial cut is based on the general region of the stream trace from B16. The stars resulting from this cut are shown in Figures \[fig:mempos\], \[fig:vel\_ra\] and \[fig:memdistr\] as open blue circles.
In order to obtain a reference isochrone against which to compare potential 300S members, we fit the known stream population from S11 and N10. For these stars, as for the rest of the stars in this study, we obtain their Pan-STARRS DR1 [PS1; @chambers2016pan] photometry[^1], and apply extinction corrections using the Bayestar17 3D dust map and extinction laws of @green2018galactic. A PARSEC isochrone [@marigo2017new] with $\mbox{[Fe/H]} = -1.5$ and an age of 12 Gyr at a distance of 18 kpc is a good match to the known stream population, in agreement with the results from S11 and F13.
For the APOGEE-2 stars, we do an initial selection by examining their position on the color-magnitude diagram (CMD) shown in Figure \[fig:memdistr\]a. B16 reported that the stream distance ranges from $14-19$ kpc. Because the stream distance near Segue 1 is 18 kpc, we posit that the stream distance increases toward smaller right ascension, and make a qualitative color selection based on identifying stars whose position on the CMD is close to, or in between, the best-fit isochrone shifted to distances of 14 kpc and 18 kpc.
The star 2M10231170+1608483 lacks PS1 and SDSS photometry, so we could not confirm its membership using its position on the CMD. The stars 2M10310815+1550149, 2M10555533+1409147, and 2M10532191+1638441 lie well away from the stream isochrone; 2M10532191+1638441 also lies too far north of the stream trace to be considered a member. The star 2M11081786+1018043 has colors consistent with it being a member of 300S, but falls several degrees beyond the trace of the stream from B16. It also falls outside of the track of the modeled orbit (see Section \[sec:orbit\]). For that reason, we exclude it from 300S membership.
The stars that remain from this qualitative selection all lie along the trace of the stream from B16 (Figure \[fig:mempos\]), confirming for the first time that 300S is the kinematic component of the photometric substructure. We then proceed to a more quantitative selection for photometric stream members. We approximate the distance gradient along the stream as varying linearly with right ascension ($\alpha$), where $(d/1~\mbox{kpc}) = 48.9952 - 0.2083\alpha$ and $\alpha$ is in degrees. The $y$-intercept and slope were determined by using the RA and distance measurements from B16, where we assume $d_{\rm 300S} = 14$ kpc at $\alpha = 168\degr$, and $d_{\rm 300S} = 19$ kpc at $\alpha = 144\degr$. Within the APOGEE-2 stars, we select photometric members of the stream by requiring that members be within 0.12 mag of the theoretical isochrone at its corresponding distance along the stream. Although the photometric uncertainties for these bright stars are quite small, we allow for a relatively wide selection window around the isochrone in order to account for uncertainties in our distance model and the stellar populations of the stream.
Table \[tab:apogeemem\] provides the list of stars in APOGEE-2 that passed the spatial and velocity selection criteria. The stars that we consider members of 300S are marked with a “1" under the “MEM" field. Out of the 27 stars from APOGEE-2 in this patch of sky with $V_{\rm helio} > 250$ [kms$^{-1}$]{}, 6 are ultimately members of 300S.
Stream Members in SEGUE-1, SEGUE-2 {#sec:seguemembers}
----------------------------------
The SEGUE (Sloan Extension for Galactic Understanding and Exploration) 1 & 2 surveys collected $R \sim 1800$ spectra in the $3900-9000$ Å wavelength range for $\sim 240,000$ stars with $14 < g < 20.3$ across a wide range of spectral types [@yanny2009segue]. The SEGUE data releases provide the radial velocity for every star, and, if the spectrum is of sufficient S/N, stellar atmospheric parameters such as metallicity, surface gravity, and effective temperature from the SEGUE Stellar Parameter Pipeline (SSPP) (@lee2008segue, @lee2008segue2, @allendeprieto2008segue, @smolinski2011segue, @lee2011alpha).
To search for 300S members from the SEGUE surveys, we begin by selecting stars in the region displayed in Figure 3 from B16 with velocities between 275 [kms$^{-1}$]{} and 325 [kms$^{-1}$]{}, and with uncertainties of less than 10 [kms$^{-1}$]{}. The spatial distribution of SEGUE stars with velocities near 300 [kms$^{-1}$]{} does not obviously reveal the presence of the stream, so we make another spatial cut by selecting stars that are less than 0.75 away from the center of the stream trace shown in Figure \[fig:mempos\]. There are 14 entries from the SEGUE survey that pass this stricter criterion, but two of them correspond to the same star, PSO J105127.367+144705.996. In reporting its values, we select the entry with the lower uncertainty in velocity. Thus, a total of 13 stars pass the velocity and position cuts, shown in Figures \[fig:mempos\], \[fig:vel\_ra\] and \[fig:memdistr\] as pink squares.
In selecting for photometric members, we apply the same criterion described in Section \[sec:apogee\_members\], leaving 6 candidates. One of the stars, PSO J104717.120+145503.948, should lie on the RGB of the stream at an extinction-corrected $r$-band magnitude of 16.65. However, its atmospheric parameters determined from SEGUE spectroscopy, $T_{\rm eff} = 5353$ K and $\log {g} = 4.4$, suggest that it is a main sequence star. Thus, we exclude this star from the member sample.
The results of 300S member selection are shown in Figures \[fig:mempos\], \[fig:vel\_ra\], and \[fig:memdistr\]. Table \[tab:segmem\] provides the list of stars in the SEGUE survey data that passed the spatial and velocity selection criteria. Table \[tab:segabund\] provides the atmospheric parameters of the same stars, obtained from the publicly available SEGUE data. It also contains \[Fe/H\], \[$\alpha$/Fe\], and \[C/Fe\] abundance ratios derived from the updated SSPP described in @lee2011alpha [$\alpha$-elements] and @lee2013carbon [carbon]. For these stars, the \[Fe/H\] measurements from the publicly available SEGUE data and from the updated SSPP pipeline are consistent with each other. In the rest of our analysis, we use the \[Fe/H\] measurements presented in Table 3. In both tables, the 5 stars that we consider members of 300S are marked with a “1" under the “MEM" field.
[lcccccccccccc]{} \[tab:apogeemem\] 2M10203784+1514471 & 155.15769 & 15.24642 & 292.54 & 0.01 & $-$1.38 & 0.10 & $-$0.57 & $-$0.02 & 14.99 & 14.02 & 13.61 & 1\
2M10204415+1555327 & 155.18399 & 15.92577 & 297.74 & 0.06 & $-$1.24 & 0.10 & $-$0.69 & $-$0.45 & 16.02 & 15.36 & 15.03 & 1\
2M10235791+1530589 & 155.99132 & 15.51638 & 292.62 & 0.13 & ... & ... & ... & ... & 17.40 & 16.83 & 16.57 & 1\
2M10243358+1531009 & 156.13992 & 15.51694 & 298.52 & 0.04 & $-$1.28 & 0.10 & $-$0.50 & $-$0.33 & 16.25 & 15.60 & 15.31 & 1\
2M10292189+1520453 & 157.34122 & 15.34594 & 299.18 & 0.07 & $-$1.32 & 0.10 & $-$0.21 & $-$0.18 & 16.22 & 15.67 & 15.41 & 1\
2M10494291+1500530 & 162.42883 & 15.01473 & 289.68 & 0.01 & $-$1.28 & 0.10 & $-$0.65 & $-$0.10 & 14.21 & 13.39 & 12.89 & 1\
2M10231170+1608483 & 155.79876 & 16.14675 & 303.77 & 0.02 & $-$1.20 & 0.10 & $-$0.38 & ... & ... & 14.34 & 13.77 & 0\
2M10310815+1550149 & 157.78400 & 15.83748 & 319.03 & 0.05 & $-$1.00 & 0.10 & $-$0.08 & ... & 14.71 & 14.11 & 13.87 & 0\
2M10532191+1638441 & 163.34129 & 16.64560 & 293.58 & 0.06 & $-$1.14 & 0.10 & $-$0.23 & ... & 13.64 & 13.33 & 13.06 & 0\
2M10555533+1409147 & 163.98056 & 14.15408 & 296.72 & 0.09 & $-$1.39 & 0.10 & $-$0.72 & ... & 14.43 & 13.97 & 13.77 & 0\
2M11081786+1018043 & 167.07442 & 10.30121 & 291.28 & 0.07 & $-$1.07 & 0.10 & $-$0.25 & ... & 15.67 & 15.06 & 14.77 & 0\
[lcccccccccc]{} \[tab:segmem\] PSO J101213.614+162336.187 & 153.05674 & 16.39340 & 297.5 & 6.0 & $-$2.07 & 0.08 & 18.25 & 17.76 & 17.51 & 1\
PSO J104236.584+150006.746 & 160.65240 & 15.00186 & 297.0 & 5.6 & $-$1.45 & 0.08 & 19.24 & 18.82 & 18.64 & 1\
PSO J104241.253+151913.286 & 160.67186 & 15.32034 & 297.2 & 3.4 & $-$1.76 & 0.08 & 18.42 & 17.95 & 17.73 & 1\
PSO J104552.952+144850.129 & 161.47066 & 14.81393 & 300.8 & 10.0 & $-$1.43 & 0.04 & 19.45 & 19.23 & 19.17 & 1\
PSO J105146.576+142850.068 & 162.94404 & 14.48055 & 287.3 & 1.7 & $-$1.41 & 0.05 & 16.79 & 16.52 & 16.42 & 1\
PSO J102747.821+151112.482 & 156.94925 & 15.18678 & 311.7 & 5.5 & $-$1.44 & 0.24 & 17.48 & 17.55 & 17.64 & 0\
PSO J103009.937+152241.212 & 157.54140 & 15.37809 & 312.4 & 5.3 & $-$1.57 & 0.14 & 17.64 & 17.69 & 17.78 & 0\
PSO J104309.200+152210.855 & 160.78830 & 15.36964 & 281.2 & 4.7 & $-$1.33 & 0.04 & 18.83 & 18.52 & 18.40 & 0\
PSO J104416.759+145459.165 & 161.06981 & 14.91639 & 279.0 & 6.1 & $-$1.23 & 0.02 & 18.60 & 18.34 & 18.24 & 0\
PSO J104717.120+145503.948 & 161.82127 & 14.91769 & 283.9 & 1.5 & $-$1.43 & 0.06 & 17.21 & 16.73 & 16.52 & 0\
PSO J105016.344+144644.466 & 162.56806 & 14.77903 & 302.9 & 2.8 & $-$1.26 & 0.01 & 16.96 & 16.30 & 16.34 & 0\
PSO J105127.367+144705.996 & 162.86398 & 14.78494 & 290.7 & 2.1 & $-$1.83 & 0.02 & 16.23 & 16.01 & 15.95 & 0\
PSO J105510.494+141100.053 & 163.79369 & 14.18334 & 275.1 & 6.4 & $-$2.48 & 0.01 & 18.54 & 18.27 & 18.17 & 0\
Properties of Newly Identified Stream Members
---------------------------------------------
As illustrated in Figure \[fig:memdistr\]a, most of the 300S members appear to lie on the red giant branch. However, there is one star from the SEGUE survey that lies on the horizontal branch, which may be advantageous for determining a more robust distance measurement of the stream at a position away from Segue 1. The mean metallicity of the stream from these measurements is $\mbox{[Fe/H]} = -1.48$. For the SEGUE stars, we adopt an external metallicity uncertainty of 0.2 dex. For the APOGEE-2 stars, we adopt an external metallicity uncertainty of 0.1 dex. Using those values, we calculate an intrinsic metallicity dispersion of 0.21$^{+0.12}_{-0.09}$ dex. @holtzman2015abundances note that the external uncertainty on ASPCAP \[Fe/H\] measurements ranges from 0.1 dex to 0.2 dex. The exact value that we adopt for the APOGEE-2 stars does not significantly affect our results because the metallicity dispersion is driven largely by the more metal-poor SEGUE stars. However, the choice of metallicity uncertainty for the SEGUE members does matter; a larger value can substantially reduce the derived intrinsic metallicity dispersion for 300S. Figure \[fig:vel\_ra\] shows the heliocentric velocity of the stream members as a function of RA. There may be a slight velocity gradient along the stream, with velocity increasing with negative RA, but it is subtle. A larger sample size would be needed to verify the existence of the gradient.
For the 300S members in both the APOGEE-2 and SEGUE samples, we correct for the depletion of \[C/Fe\] along the red giant branch by applying the methods used in @placco2014. For the APOGEE-2 members, the respective carbon corrections for stars with available abundances (shown in the order in Table \[tab:apogeemem\]) are 0.55, 0.24, 0.17, 0.03, and 0.55 dex. We present the corrected values in Table \[tab:apogeemem\]. For the SEGUE stars, the corrected carbon abundances are included in Table \[tab:segabund\]. For these stars, the corrections are less than 0.1 dex. One 300S star, PSO J104236.584+150006.746, can be classified as carbon-enhanced ($\mbox{[C/Fe]} = +0.90$). Given its metallicity, PSO J104236.584+150006.746 is likely a CEMP-$s$ star enriched by a binary companion. Follow-up spectroscopy to measure its neutron-capture element abundances and velocity variability may be interesting to confirm this possibility.
[ccccccccccccc]{} \[tab:segabund\] PSO J101213.614+162336.187 & 5057 & 83 & 2.3 & 0.1 & $-$2.17 & 0.13 & $+$0.85 & 0.17 & $-$0.26 & 0.12 & $-$0.26 & 1\
PSO J104236.584+150006.746 & 5367 & 47 & 2.6 & 0.1 & $-$1.65 & 0.11 & $+$0.28 & 0.22 & $+$0.88 & 0.11 & $+$0.90 & 1\
PSO J104241.253+151913.286 & 5216 & 63 & 2.5 & 0.1 & $-$1.75 & 0.13 & $+$0.39 & 0.15 & $+$0.04 & 0.10 & $+$0.06 & 1\
PSO J104552.952+144850.129 & 6557 & 53 & 4.0 & 0.6 & $-$1.26 & 0.20 & $+$0.21 & 0.20 & $+$0.57 & 0.42 & $+$0.57 & 1\
PSO J105146.576+142850.068 & 6131 & 62 & 2.3 & 0.3 & $-$1.45 & 0.07 & $+$0.43 & 0.08 & $-$0.12 & 0.22 & $-$0.10 & 1\
PSO J102747.821+151112.482 & 8275 & 22 & 4.1 & 0.2 & $-$1.56 & 0.11 & ... & ... & ... & ... & ... & 0\
PSO J103009.937+152241.212 & 7970 & 76 & 4.0 & 0.1 & $-$1.49 & 0.08 & ... & ... & $+$2.19 & 0.5 & $+$2.20 & 0\
PSO J104309.200+152210.855 & 6047 & 55 & 4.2 & 0.1 & $-$1.30 & 0.02 & $+$0.31 & 0.16 & $+$0.21 & 0.14 & $+$0.21 & 0\
PSO J104416.759+145459.165 & 6313 & 53 & 3.3 & 0.2 & $-$1.27 & 0.11 & $+$0.26 & 0.20 & $+$0.58 & 0.28 & $+$0.59 & 0\
PSO J104717.120+145503.948 & 5353 & 17 & 4.4 & 0.0 & $-$1.44 & 0.04 & $+$0.46 & 0.04 & $+$0.04 & 0.02 & $+$0.04 & 0\
PSO J105016.344+144644.466 & 6590 & 105 & 3.2 & 0.2 & $-$1.46 & 0.07 & $+$0.67 & 0.10 & $+$0.46 & 0.16 & $+$0.48 & 0\
PSO J105127.367+144705.996 & 6453 & 32 & 3.7 & 0.1 & $-$1.87 & 0.06 & $+$0.54 & 0.14 & $<$0.01 & ... & ... & 0\
PSO J105510.494+141100.053 & 6329 & 56 & 3.8 & 0.2 & $-$2.75 & 0.12 & $+$0.64 & 0.15 & $+$1.66 & 0.42 & $+$1.66 & 0\
Chemical Abundance Analysis {#sec:abundances}
===========================
In this section we examine the chemical abundances of the stream members identified in Section \[sec:new\_members\]. We analyze detailed abundances for the six APOGEE-2 stream members identified in Section \[sec:apogee\_members\]. For one of these stars, 2M10235791+1530589, the ASPCAP pipeline was unable to determine any abundances. However, a line-by-line analysis of its spectrum suggests similar abundances to the other five stars. We also examine the \[$\alpha$/Fe\] abundance ratios for the SEGUE stream members.
Comparison Sample Selection
---------------------------
From the APOGEE-2 dataset, we select various other sets of stars with which to compare chemical abundances. For all stars, we verify that none have the ‘STAR\_BAD’ ASPCAP flag, which encodes any unreliable ASPCAP measurements, set [@holtzman2015abundances]. We also ensure that these stars have internal uncertainties less of than 0.2 dex for each element considered.
For a dwarf galaxy comparison sample, we select members of the Sagittarius (Sgr) dwarf galaxy from APOGEE (see @hasselquist2017apogee for a detailed study of the chemical abundances of Sagittarius). We begin by selecting stars with the ‘APOGEE\_SGR\_DSPH’ flag set in the APOGEE\_TARGET1 column. From that sample, we apply a velocity cut, selecting stars with $120~\mbox{{km\,s$^{-1}$}} < V_{\rm helio} < 160$ [kms$^{-1}$]{}, which isolates most stars in the core and removes some potential contaminants. For data on other dwarf galaxies, we use the measurements of @shetrone2003vlt for Sculptor, Leo I, Carina and Fornax, @cohen2009chemical for Draco, and @cohen2010chemical for Ursa Minor.
We also select APOGEE stars in the globular clusters M13 and M92 for comparison. We chose M13 because its metallicity is similar to that of the stream, and M92 as a metal-poor reference. We take cluster membership information from @meszaros2015exploring. However, we use chemical abundances from ASPCAP to control for potential systematics due to different analysis methods.
We construct our halo sample by obtaining distances from the APOGEE DR14-Based Distance Estimations value added catalog, which was constructed using the isochrone matching technique [NICE; @schultheis2014extinction]. To determine the height above the Galactic plane where 90% of the stars are from the halo as a function of Galactocentric radius, we use the Trilegal model [@vanhollebeke2009stellar]. We select all APOGEE-2 stars above this height for our halo sample.
We also supplement our position-selected halo sample with stars that have 3D space motions consistent with halo membership, which we base on Gaia DR1 proper motions (see @gaia2016gaia, @brown2016gaia, @lindegren2016astrometry, and @arenou2017gaia, among others) and APOGEE DR14 radial velocities. We select our stars as those that have rotational and $UW$ velocities that differ from those of thick- and thin-disk stars by more than 2$\sigma$. Due to the magnitude limits in Gaia DR1, there was very little overlap between these two differently selected samples.
Light-Element Abundance Correlations
------------------------------------
We begin our investigation into the nature of the stream progenitor by comparing its chemical abundance pattern to those of globular clusters. Globular cluster stars obey well-known correlations between abundances of light elements including carbon, nitrogen, oxygen, sodium, aluminum, and magnesium [see, e.g., @gratton2004 and references therein]. In Figure \[fig:chem\_corr\] we examine the light-element abundances of 300S in comparison to M92 and M13, as well as to Local Group dwarf galaxies. Because Na is too weak to be measured in the $H$-band at low metallicity ($[\mathrm{Na/H}] \lesssim -1$), we could not test for a Na-O anti-correlation among the 300S stars.
It is apparent that members of 300S do not display the chemical abundance correlations of globular clusters, either in the direction of the correlation or the shape of the distribution. From comparison with Figure 9 of @meszaros2015exploring, which shows the Mg-Al correlations of the globular clusters from that study, we note in particular that 300S does not resemble either the first-generation or second-generation stars in M13. The chemical abundances of 300S are also much more similar to those of the Sagittarius dwarf galaxy than either of the globular clusters. This comparison strongly suggests that the progenitor of 300S is more likely to be a dwarf galaxy than a globular cluster.
\[C/N\] Ratio
-------------
To first order, the \[C/N\] ratio serves as an indicator of age. In Figure \[fig:cnlogg\] we compare the \[C/N\] ratios of 300S to those of Sagittarius and the Milky Way halo. To control for evolutionary stage and possible deviations from local thermodynamic equilibrium (LTE), we select stars from our comparison sample that have similar $\log {g}$ and \[Fe/H\] to 300S. In the regime of $-1.5 < \mbox{[Fe/H]} < -1.0$, the 300S members appear to be approximately as old as Sgr and the halo.
Chemical Abundance Patterns
---------------------------
To control for non-LTE effects, we originally only selected stars that are similar to the 300S sample in $T_{\rm eff}$ and $\log{g}$. However, the conclusions we drew from considering only such $T_{\rm eff}$ and $\log{g}$ “twins” were the same as those from considering the full halo sample. Therefore, we present our full sample of halo stars in the following figures in order to better illustrate the halo chemical abundance distribution, and highlight the stars that are $T_{\rm eff}$ and $\log{g}$ twins of 300S.
Figure \[fig:alphageneral\] compares the \[$\alpha$/Fe\] abundance ratios of 300S to those of Sgr, the MW halo, and Ursa Minor. To ensure a consistent comparison to the \[$\alpha$/Fe\] ratios for the SEGUE stars, we calculate \[$\alpha$/Fe\] for the other populations by taking a weighted average abundance of Mg, Ca, Ti, and Si, where the weights for these four elements are given in @lee2011alpha. For Ti, we give the same weight to the abundances of its different ionization states. Ursa Minor was the only dSph from the literature included in this sample, because it alone has published abundances for all four of the above-mentioned elements. The “knee" of the \[$\alpha$/Fe\] ratio as a function of metallicity [@tinsley1979stellar] for 300S appears to occur at around \[Fe/H\] = $-1.3$. 300S reaches solar \[$\alpha$/Fe\] at a similar metallicity to the classical dSphs.
Figure \[fig:alpha\] compares individual $\alpha$-elements as a function of metallicity in 300S and Sgr. The $\alpha$-element abundances of 300S generally match those of the classical dSphs. 300S may be slightly enriched in calcium compared to Sgr and the other dwarf spheroidals. Compared to the MW halo, 300S also seems to have similar $\alpha$-element abundances.
Figure \[fig:fe\_peak\] compares the abundances of Fe-peak elements in 300S to those of the other Milky Way systems. 300S appears to be deficient in Cr relative to the other dwarf spheroidals. Within the uncertainties, 300S has similar Mn and Ni abundances to the dwarf galaxies. Overall, 300S has similar Fe-peak abundance patterns relative to the MW halo. Compared to the reference globular clusters, 300S is deficient in Fe-peak elements.
That 300S has similar chemical abundance patterns to the Milky Way halo, as well as dSphs, and different ones from those of globular clusters, further suggests that the 300S progenitor is a dwarf galaxy.
Tidal Disruption History {#sec:orbit}
========================
Proper Motion Modeling
----------------------
In Section \[sec:new\_members\], we determined the path of 300S along the sky and its distance and velocity as a function of position. In order to calculate the orbit of the stream around the Milky Way, we also require proper motions.
We initially attempted to employ proper motion catalogs such as UCAC5 [@zacharias2017ucac5], and the recently released GPS-1 [@tian2017gaia], to constrain the proper motions of the stream members. However, we found that the published proper motions of the APOGEE-2 and SEGUE stars in the stream exhibit a large scatter, and the UCAC5 and GPS-1 proper motions are not in very good agreement for the stars with measurements in both catalogs. Table \[tab:pm\] presents the catalog proper motion values for the members of 300S identified in Section \[sec:new\_members\].
[cccccccccc]{} 2M10203784+1514471 & -5.7 & 3.0 & -2.7 & 2.5 & ... & ... & ... & ...\
2M10204415+1555327 & -5.7 & 1.7 & -1.4 & 1.3 & ... & ... & ... & ...\
2M10235791+1530589 & -0.4 & 1.4 & 0.0 & 1.2 & ... & ... & ... & ...\
2M10243358+1531009 & -2.4 & 1.3 & -7.7 & 1.2 & -6.5 & 3.8 & -3.4 & 2.7\
2M10292189+1520453 & 0.0 & 1.5 & 0.2 & 1.2 & ... & ... & ... & ...\
2M10494291+1500530 & 0.0 & 3.0 & -8.3 & 2.6 & -4.8 & 1.1 & -4.5 & 1.1\
PSO J101213.614+162336.187 & -1.9 & 1.7 & -1.8 & 1.2 & ... & ... & ... & ...\
PSO J104236.584+150006.746 & -3.5 & 1.9 & -2.7 & 1.5 & ... & ... & ... & ...\
PSO J104241.253+151913.286 & -6.4 & 1.8 & -6.1 & 1.8 & ... & ... & ... & ...\
PSO J104552.952+144850.129 & -3.4 & 2.1 & -1.6 & 1.7 & ... & ... & ... & ...\
PSO J105146.576+142850.068 & -1.5 & 1.4 & -4.4 & 1.4 & -8.9 & 14.9 & 1.0 & 9.7\
\[tab:pm\]
Thus, we instead infer the proper motion of the stream by considering many possible orbits subject to the constraints of its known properties. We test a grid covering the full range of plausible proper motions to see which of them produce a stream with (1) the observed path, constrained by B16, (2) the distance along the stream, which we were able to obtain for APOGEE-2 stars from [@queiroz2018starhorse], and (3) the velocity along the stream, constrained by the APOGEE-2 and SEGUE 300S members.
We model 300S using the open-source code *galpy* [@bovy2015galpy], and approximate the orbit of the stream using point-particle integration. We integrate the orbits in the `MWPotential2014` potential, which is the standard Milky Way model in *galpy*, and adopt the solar motion from @schoenrich2010. In order to initialize an orbit, *galpy* requires 6D phase space information about the point where the orbit is initialized. We initialize the orbit of 300S at the location of the S11 stars (i.e., Segue 1) because that region has the best-constrained values as a result of the large sample of confirmed main sequence member stars. At that location, $\alpha = 151.8\degr$, $\delta = 16.1\degr$, $V_{\rm helio} = 298.8$ [kms$^{-1}$]{}, and $D_{\odot} = 18$ kpc (S11).
![Reduced $\chi^{2}$ value of the best-fitting orbit. The best-fit $\mu_{\alpha}\cos{\delta}$, $\mu_{\delta}$ values are $-2.35$ [masyr$^{-1}$]{} and $-2.2$ [masyr$^{-1}$]{}, respectively. The orbit from those proper motions has a reduced $\chi^{2}$ value close to 1, suggesting that the orbit model is a good fit to the data.[]{data-label="fig:pmfit"}](mura_mude_convergence.png)
For every orbit corresponding to a proper motion, we compute its $\chi^{2}$ value based on its fit to the stream observables. Figure \[fig:pmfit\] shows the results of the $\chi^{2}$ fits for a region around the best fit proper motion. The proper motion of the orbit corresponding to the lowest $\chi^{2}$ fit is $\mu_{\alpha}\cos{\delta} = -2.35$ [masyr$^{-1}$]{}, and $\mu_{\delta} = -2.2$ [masyr$^{-1}$]{}. The reduced $\chi^{2}$ value of the orbit fit at that point is close to 1, suggesting that the model is a good fit. This proper motion is also close to the weighted mean of the GPS-1 proper motions of the member stars ($\mu_{\alpha}\cos{\delta} = -2.5 \pm 0.5$ [masyr$^{-1}$]{}, $\mu_{\delta} = -2.8 \pm 0.4$ [masyr$^{-1}$]{}), despite the large uncertainties of many of the individual measurements.
For completeness, we fit an orbit with a positive $\mu_{\alpha}$, where the stream would travel in the opposite direction, and obtain proper motions of $\mu_{\alpha}\cos{\delta} = 1.48$ [masyr$^{-1}$]{}, and $\mu_{\delta} = -2.63$ [masyr$^{-1}$]{}. However, this orbit has a corresponding reduced $\chi^2$ value of 2.2, suggesting that the positive $\mu_{\alpha}$ orbit is a poorer fit to the data. This orbit also has a larger perigalactic distance of 17 kpc.
Thus, we prefer the negative $\mu_{\alpha}$ solution due to its better agreement with the GPS-1 proper motions, and because its much smaller perigalactic distance is more consistent with the observed disruption of the stream progenitor. The forthcoming Gaia DR2 measurements will eliminate this degeneracy.
Properties of the Modeled Orbit
-------------------------------
![(a) Modeled orbit of the stream compared to its observed position on the sky. The orbit passes through the position of the known members, but deviates slightly from the trace of the stream as seen in B16. (b) Modeled heliocentric velocity along the stream track. The modeled orbit is in good agreement with the data. (c) Modeled distance along the stream. Distances plotted are to the APOGEE-2 stream members according to @queiroz2018starhorse. The modeled orbit is consistent with all of the distance measurements to within $2\sigma$.[]{data-label="fig:galpyfits"}](galpy_fits.png)
The panels in Figure \[fig:galpyfits\] compare the modeled orbit with the observed properties of the stream. In Figure \[fig:galpyfits\]a, the position of the modeled orbit is consistent with the location of the 300S members. The orbit also does not pass close to the position of 2M11081786+1018043, corroborating our decision in Section \[sec:apogee\_members\] that the star is not a likely 300S member. In Figure \[fig:galpyfits\]b and \[fig:galpyfits\]c, the modeled orbit is consistent with the heliocentric velocity and distance along the stream track. Although the orbit deviates slightly from the trace of the stream as seen in B16, it is still reasonable to use it to infer general features of the stream’s kinematic history.
![Galactocentric distance of the stream orbit as a function of time. The stream passed perigalacticon about 70 Myr ago, approaching within $\sim4.5$ kpc of the Galactic Center.[]{data-label="fig:time_dist"}](time_streamdist.png)
Figure \[fig:time\_dist\] shows the distance of the stream away from the Galactic Center as a function of time, suggesting that the progenitor of 300S passed perigalacticon only $\sim$ 70 Myr ago. At its closest approach, the progenitor was $\sim4.5$ kpc from the Galactic Center, and thus must have experienced powerful tidal forces. If this interpretation is correct, then 300S may be the most recently disrupted system currently known.
Figure \[fig:galactocentric\] shows the modeled orbit projected onto various Cartesian planes centered on the Galactic Center. The orbit of 300S does not resemble that of the Sagittarius dwarf spheroidal galaxy [@law2010sagittarius], supporting previous suggestions that 300S is unlikely to be kinematically associated with the Sagittarius stream (G09, S11, B16). If the two are related, the 300S progenitor must have been stripped from Sgr long ago. The orbit of the stream is perpendicular to the proper motion of Segue 1 [@fritz2017orbit], ruling out any association with that galaxy as well.
Integrated Luminosity of the Stream
-----------------------------------
In order to place a lower limit on the original mass of the core of the progenitor, we calculate the integrated luminosity of the stream using PS1 photometry [@chambers2016pan]. We follow the same procedures outlined in B16 to select stellar-like objects. In particular, we select sources with $| r_{psf} - r_{aperture}| \leq 0.2$ mag. To ensure the quality of our photometry, we also reject sources with uncertainties larger than 0.2 mag in $g$, $r$, and $i$ bands. We correct for reddening effects using the dust maps of @sfd1998 and using the extinction laws of @schlafly2011.
We create a Hess diagram of 300S by subtracting a matching Hess diagram of the Milky Way foreground population from the Hess diagram of the region within the stream. The width of the stream is $\sim1\degr$ (see Section \[sec:origin\]). Thus, to construct our foreground distribution, we consider two regions. At a distance of 1 north of the center of the stream, we obtain the foreground distribution by using the stars within a region of sky that is 1 wide and runs parallel to the stream track. At a distance of 1 south of the stream center, we construct another foreground distribution in an analogous fashion. We average the distributions from the regions above and below the stream to produce the foreground Hess diagram, and subtract that from the Hess diagram of the region within the stream.
From the foreground-subtracted Hess diagram, we find a total of 637 stars within the stream down to a magnitude of $r=22$. For comparison, @martin08 measured 65 stars in Segue 1 ($L = 335$ L$_{\odot}$) down to the same magnitude limit in SDSS. In each Hess diagram bin along the stream CMD sequence we calculate a luminosity using the isochrone described in Section \[sec:apogee\_members\] and shown in Figure \[fig:memdistr\]a. After correcting for the contribution of sources below the PS1 magnitude limit following @martin08, we obtain an integrated luminosity for 300S of $4 \times 10^3$ L$_{\odot}$.
Since the chemical abundances of the stream stars suggest that the progenitor of 300S is a dwarf galaxy, we can invoke the mass-metallicity relationship for dwarf galaxies [@kirby2013universal] to infer a progenitor stellar mass of $10^{6.9 \pm 0.6}$ M$_{\odot}$, which is on the order of a classical dwarf galaxy (e.g., Leo I, Sculptor). The implied luminosity of this stellar mass is far greater than the observed luminosity of the stream. This suggests that the 300S progenitor was either abnormally metal-rich for its luminosity, or that most of its mass was stripped earlier and is not present in the currently-known part of the stream.
The Origin of 300S {#sec:origin}
==================
Until the discovery of the ultra-faint dwarf galaxies in 2005 [@willman05a; @willman05b], globular clusters and dwarf galaxies exhibited clearly distinct structural properties, with all dwarf galaxies having radii of more than 100 pc and all globular clusters having radii less than 30 pc [e.g., @belokurov07]. Over the last decade the size distributions of the two populations have begun to encroach upon one another at faint luminosities ($L \lesssim
10^{4}$ L$_{\odot}$), such that size alone is no longer sufficient to classify compact stellar systems. This convergence has led to the adoption of alternative classification criteria, namely dynamical mass-to-light ratio and metallicity dispersion [@willman12]. While the interpretation of the stellar kinematics is rendered more difficult in the case of a tidally disrupted object, the chemical properties of the stars are preserved during the disruption process.
To ascertain the nature of the 300S progenitor, we first consider the chemical abundance measurements presented in Sections \[sec:new\_members\] and \[sec:abundances\]. The metallicites of the APOGEE-2 member stars are all quite similar to each other (and to 300S-1 from F13). However, the SEGUE members span a range in metallicity from $\mbox{[Fe/H]} = -1.3$ to $-2.2$, such that the intrinsic dispersion of the metallicity distribution for our full sample of new members is 0.21$_{-0.09}^{+0.12}$ dex. While we cannot distinguish this value from zero with high statistical significance, the data do suggest that 300S is not a mono-metallic system. Although the sample size is small, we also do not see any sign of the characteristic globular cluster light-element abundance correlations in 300S. With detailed abundance patterns for only five stars, we cannot exclude the possibility that all of the APOGEE-2 stars happen to belong to a single stellar generation and that high-resolution spectroscopy of additional stars would reveal multiple populations with correlated abundances. While each individual piece of evidence is not a strong indicator of the nature of the progenitor, taken together, the best explanation for the data is that the progenitor of 300S was a dwarf galaxy.
To constrain the size of the stream progenitor, we estimate the width of the stream. The profile of the stream perpendicular to its length has a full-width at half maximum of 0.94, corresponding to a physical extent of $\sim260$ pc at an average distance of 16 kpc. This size is consistent with the hypothesis that the progenitor was a relatively compact dwarf galaxy.
S11 measured a velocity dispersion for the stream at the position of Segue 1 of $7.0 \pm 1.4$ [kms$^{-1}$]{} with all candidate member stars included, and $5.6 \pm 1.2$ [kms$^{-1}$]{} if several possible foreground stars are excluded. Because the best-fit stream model we computed in Section \[sec:new\_members\] indicates that there may be a velocity gradient of $\sim10$ [kms$^{-1}$]{} across the full region spanned by the spectroscopic data, we cannot simply use the entire sample of member stars to calculate the velocity dispersion. Instead, we use the fact that the five of the six APOGEE-2 stars and four of the five SEGUE stars are clustered together (near $\alpha_{2000} = 156\degr$ and $\alpha_{2000} =
161.5\degr$, respectively) to determine local velocity dispersions at these two positions. The dispersion of the APOGEE-2 stars near $\alpha_{2000} = 156\degr$ is $\sigma = 3.3^{+1.8}_{-1.1}$ [kms$^{-1}$]{}, while that of the SEGUE stars near $\alpha_{2000} = 161.5\degr$ is $\sigma =
5.1^{+4.8}_{-2.8}$ [kms$^{-1}$]{}. Within the uncertainties, we therefore conclude that the data are consistent with the stream having a constant velocity dispersion of $\sim4-5$ [kms$^{-1}$]{} over the RA range $\alpha_{2000} = 151.8\degr-161.5\degr$. Recognizing that the disruption of the progenitor may have resulted in a stream velocity dispersion that is higher than that of the progenitor system, we note that this dispersion is significantly larger than that of the prototypical globular cluster stream Pal 5 [@odenkirchen2009; @kuzma2015].
Finally, we consider the stream orbit derived in Section \[sec:orbit\]. The path of the stream indicates that it is moving on a highly elliptical orbit. Such an orbit would be quite unusual for a globular cluster [e.g., @dinescu1999; @allen2006]. More eccentric orbits are expected for dwarf galaxies, which may have fallen into the Milky Way from large distances, as opposed to forming in situ like many globular clusters. Orbits approaching within a few kpc of the Galactic Center are likely not common for dwarfs either, but very small perigalacticon distances are necessary in order to completely disrupt a dark matter-dominated system. As an example, S11 calculated that a galaxy with the mass and size of Segue 1 would need to pass within $\sim4$ kpc of the Galactic Center to be disrupted.
We conclude that the observed properties of 300S favor a dwarf galaxy, rather than a globular cluster, progenitor. Next, we examine some of its characteristics in the context of other dwarf galaxies. If we invoke the mass-metallicity relation for dwarf galaxies [@kirby2013universal], the metallicity of the stream corresponds to a progenitor stellar mass of $10^{6.9 \pm 0.6}$ M$_{\odot}$, which is comparable to a classical dwarf spheroidal galaxy such as Leo I (${\rm [Fe/H]} = -1.43$). That 300S reaches solar levels of \[$\alpha$/Fe\] at a metallicity between that of Ursa Minor and Sgr also suggests that the stellar mass of its progenitor is between $2\times10^5$ M$_{\odot}$ (UMi) and at least $2\times10^7$ M$_{\odot}$ (Sgr, core). Using Equation 28 from @erkal2016number, which calculates the mass of the stream progenitor from the width of the stream on the sky and the enclosed mass at the stream distance,[^2] we obtain a progenitor mass of $10^{5.3}$ M$_{\odot}$. However, the integrated luminosity of the stream over its observed extent is only $4000$ L$_{\odot}$. Reconciling these numbers requires either that the progenitor dwarf was unusually metal-rich (and perhaps unusually extended) for its luminosity, that the progenitor was strongly dark-matter dominated, or that nearly all of the stars belonging to the progenitor lie outside the known stream. Since the stream’s orbit extends out to nearly 60 kpc, where most of its stars would be too faint to be detected by current surveys, the latter scenario may be plausible. In addition, it is possible that the progenitor made previous close passages to the Galactic Center during which most of its mass was lost. Given the derived orbital period of $\sim1$ Gyr (see Fig. \[fig:time\_dist\]), such stars could now be located quite far away from 300S.
Although the radial velocity of 300S is not consistent with the known portions of the Sgr stream (G09), the spatial overlap between the two and their chemical similarity suggests the possibility that the stream progenitor might once have been a dwarf satellite of Sagittarius. Cosmological simulations suggest that $\sim30-60$% of dwarf satellites around MW, M31-like halos were accreted as members of galaxy groups [@wetzel2015satellite]; thus, the phenomenon of group infall more generally is not unlikely. Previous wraps of Sgr debris around the Galaxy are not well-constrained by existing data, so an association with Sgr cannot currently be ruled out observationally. Improved proper motion measurements for both Sgr and 300S and more detailed modeling of the early history of the interaction between Sgr and the Milky Way would be needed to test this idea.
Conclusions
===========
In this study, we present 11 new members of 300S identified in the APOGEE-2 and SEGUE spectroscopic surveys. From the position of these stars on the sky, we show that 300S is the kinematic counterpart of the elongated photometric substructure found in the same region.
We find that the 300S members from APOGEE-2 are chemically similar to Local Group dwarf galaxies and the Milky Way halo, and do not display the characteristic light-element abundance correlations of globular clusters. The new known members also display a metallicity dispersion of 0.21$_{-0.09}^{+0.12}$ dex, exceeding an intrinsic dispersion of 0 by 2$\sigma$. This suggests that the progenitor may have had an extended period of star formation and a potential well sufficiently deep to retain supernovae ejecta. The relatively large width and velocity dispersion of the stream also point to a massive progenitor. Thus, we conclude that 300S is likely the remnant of a tidally disrupted dwarf galaxy.
We infer the proper motion of the stream by fitting the observed properties of the stream to orbits generated from a grid of possible proper motions. The best-fit orbit is highly eccentric, with an apogalacticon distance of 60 kpc and perigalacticon distance of 4.5 kpc away from the Galactic Center. The orbital period of 300S is $\sim1$ Gyr, with its most recent perigalacticon passage 70 Myr ago, suggesting that 300S may be the most recent tidally disrupted system known to date.
Invoking the mass-metallicity relationship for dwarf galaxies, we find that the progenitor of 300S should have a stellar mass of $10^{6.9 \pm 0.6}$ M$_{\odot}$, which is comparable to classical dwarf spheroidal galaxies such as Leo I, Sculptor, and Fornax. We also calculate the integrated luminosity of the stream to be $4 \times 10^3$ L$_{\odot}$, which is much lower than the luminosity implied by the stellar mass from the previous relation. However, at a perigalacticon distance of 4.5 kpc, the tidal field of the Milky Way is sufficiently strong for even a dark matter-dominated system to undergo tidal disruption. With an orbital period of 1 Gyr, it is quite possible that the progenitor of 300S lost most of its stars over multiple close passages to the Milky Way. This is consistent with matched-filter maps of the stream, which show a system that is completely tidally disrupted.
At an observed distance of 20 kpc away from the Milky Way center, 300S may be a valuable probe of the Milky Way potential interior to that distance. With a perigalacticon passage of 4.5 kpc, the orbit of stars in 300S may be affected by time-dependent effects of the Galactic Bar. For reference, the Pal 5 stream, with a perigalacticon distance of of 8 kpc, displays gaps that may have resulted from the bar rotation [@pearson2017bar]. Thus, the modeling of 300S in tandem with other stellar streams should provide a more complete picture of the Milky Way potential within its inner tens of kpc.
The upcoming Gaia Data Release 2 should provide strong constraints on the proper motion along the stream track, as the brightest members of 300S are predicted to have proper motion uncertainties of just $\sim0.06$ [masyr$^{-1}$]{} [@gaiadr2]. The Gaia data should also aid in selecting additional 300S targets for spectroscopic followup to determine the velocity dispersion and gradient along the stream, as well as for determining detailed chemical abundances of more stream members.
[^1]: We use PS1 photometry because some APOGEE-2 candidates are bright enough to be saturated in SDSS images.
[^2]: The width used in this calculation is the $\sigma$ of a Gaussian fit to the stream profile (D. Erkal 2018, private communication), which we measure to be 0.4. It is also important to note that @erkal2016number derived this relation for the case of a single-component (i.e., purely stellar) progenitor. For a dwarf galaxy progenitor containing both stars and dark matter, the mass determined with this method should correspond to the dynamical mass within a radius comparable to the width of the stream rather than the stellar mass (D. Erkal 2018, private communication). This value of course may be much smaller than the mass with which the progenitor formed if stripping has been ongoing for a long time.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Combining the mutual information criterion with a forward feature selection strategy offers a good trade-off between optimality of the selected feature subset and computation time. However, it requires to set the parameter(s) of the mutual information estimator and to determine when to halt the forward procedure. These two choices are difficult to make because, as the dimensionality of the subset increases, the estimation of the mutual information becomes less and less reliable. This paper proposes to use resampling methods, a K-fold cross-validation and the permutation test, to address both issues. The resampling methods bring information about the variance of the estimator, information which can then be used to automatically set the parameter and to calculate a threshold to stop the forward procedure. The procedure is illustrated on a synthetic dataset as well as on real-world examples.'
address:
- |
Université catholique de Louvain, Machine Learning Group, CESAME\
Av. Georges Lemaitre, 4, B-1348 Louvain-la-Neuve, Belgium
- |
Projet AxIS, INRIA\
Domaine de Voluceau, Rocquencourt, B.P. 105, 78153 Le Chesnay Cedex, France
- |
Université catholique de Louvain, Machine Learning Group, DICE\
Place du Levant, 3, B-1348 Louvain-la-Neuve, Belgium
author:
- '[D. François]{}'
- 'F. Rossi'
- 'V. Wertz'
- 'M. Verleysen'
bibliography:
- 'PermutationTest.bib'
title: 'Resampling methods for parameter-free and robust feature selection with mutual information'
---
,
,
and
Mutual information ,Permutation test ,Feature selection
Introduction {#sec:introduction}
============
Feature selection consists in choosing, among a set of input features, or variables, the subset of features that has maximum prediction power for the output. More formally, let us consider ${\bf X} = (X_1, \cdots, X_d)$ a random input vector and $Y$ a continuous random output variable that has to be predicted from ${\bf X}$. The task of feature selection consists in finding the features $X_i$ that are most relevant to predict the value of $Y$ [@Guyon:2003].
Selecting features is important in practice, especially when distance-based methods like k-nearest neighbors (k-NN), Radial Basis Function Networks (RBFN) and Support Vector Machines (SVM) (depending on the kernel) are considered. These methods are indeed quite sensitive to irrelevant inputs: their performances tend to decrease when useless variables are added to the data.
When the data are high-dimensional (i.e. the initial number of variables is large) the exhaustive search of an optimal feature set is of course intractable. In such cases, furthermore, most methods that ‘work backwards’ by eliminating useless features perform badly. The backward elimination procedure for instance, or pruning methods for the MultiLayer Perceptron [@Verikas:2002], SVM-based feature selection [@Fung:2004], or weighting methods like the Generalized Relevance Learning Vector Quantization algorithm [@Hammer:2002] require building a model with all initial features. With high-dimensional data, this will often lead to large computation times, overfitting, convergence problems, and, more generally, issues related to the curse of dimensionality. These approaches are furthermore bound to a specific prediction model.
By contrast, a forward feature selection procedure can be applied using any model and begins with small feature subsets. Such procedure is furthermore simple and often efficient. Nevertheless, when data are high-dimensional, it becomes difficult to perform the forward search using the prediction model directly. This is because, for every candidate feature subset, a prediction model must be fit, involving resampling techniques and grid searching for optimal structural parameters. A cheaper alternative is to estimate the relevance of each candidate subset with a statistical or information-theoretic measure, without using the prediction model itself.
The combined use of a forward feature search and an information-theoretic-based relevance criterion is generally considered to be a good option, when nonlinear effects prevent from using the correlation coefficient [@guyon:2006]. In this context, the mutual information estimated using a nearest neighbour-based approach has been shown to be effective [@VanDijk:2006; @CILS2006]. Nevertheless, this approach, just like most feature selection methodologies, faces two difficulties.
The first one, which is generic for all feature selection methods, lies in the optimal choice of the number of features to select. Most of the time, the number of features to select is chosen a priori or so as to maximize the relevance criterion. The former approach leaves no room for optimization, while the latter may be very sensitive to the estimation of the relevance criterion.
The second difficulty concerns the choice of parameter(s) in the estimation of the relevance criterion. Indeed, most of these criteria, except maybe for the correlation coefficient, have at least one structural parameter, like a number of units or a kernel width in a prediction model, a number of neighbours or a number of bins in a nonparametric relevance estimator, etc. Often, the result of the selection highly depends on the value of that (those) parameter(s).
The aim of this paper is to provide an automatic procedure to choose the two above-mentioned important parameters, i.e. the number of features to select in the forward search and the structural parameter(s) in the relevance criterion estimation. This procedure will be detailed in a situation where the mutual information is used as relevance criterion, and is estimated through nearest neighbours. Resampling methods will be used to obtain this automatic choice. Those methods increase the computational cost of the forward search, but provide meaningful information about the quality of the estimations and the setting of parameters: it will be shown that a permutation test can be used to automatically stop the forward procedure, and that a combination of permutation and K-fold resampling allows choosing the optimal number of neighbors in the estimation of the mutual information.
The remaining of this paper is organized as follows. Section \[sec:priorart\] introduces the mutual information, the permutation test and the K-fold resampling, and briefly reviews how they can be used together. Section \[sec:problems\] illustrates the challenges in choosing the number of neighbours in the mutual information estimation and the number of features to select in a forward search. Section \[sec:proposedapproach\] then presents the proposed approach. The performances of the method on real-world data are reported in Section \[sec:experiments\].
Prior art {#sec:priorart}
=========
Mutual information-based forward feature selection
--------------------------------------------------
The mutual information is a nonparametric, nonlinear, measure of relevance derived from information theory. Unlike correlation that only considers linear relationships between variables, the mutual information is theoretically able to identify relations of any type. It furthermore makes no assumption about the distribution of the data.
The mutual information of two random variables $Z_1$ and $Z_2$ is a measure of how $Z_1$ depends on $Z_2$ and *vice versa*. It can be defined from the entropy $H(.)$: $$\label{eq:mientropy}
{M\!I\left(Z_1;Z_2\right)} = H(Z_1) + H(Z_2) - H(Z_1,Z_2) = H(Z_1) - H(Z_2|Z_1),$$ where $H(Z_2|Z_1)$ is the *conditional* entropy of $Z_2$ given $Z_1$. In that sense, it measures the loss of entropy (i.e. reduction of uncertainty) of $Z_2$ when $Z_1$ is known. If $Z_1$ and $Z_2$ are independent, $H(Z_1,Z_2) = H(Z_1) + H(Z_2)$, and $H(Z_2|Z_1) = H(Z_2)$. In consequence, the mutual information of two independent variables is zero.
For a continuous random variable $Z_1$, the entropy is defined as $$H(Z_1) = - \int p_{Z_1}(\zeta_1) \, \log\,p_{Z_1}(\zeta_1) \, {\rm d}\zeta_1,$$ where $p_{Z_1}$ is the probability distribution of $Z_1$. Consequently, the mutual information can be rewritten, for continuous $Z_1$ and $Z_2$, as $$\label{eq:midivergency}
{M\!I\left(Z_1;Z_2\right)} = \int\!\!\!\int p_{Z_1,Z_2}(\zeta_1, \zeta_2) \, \log \, \frac{p_{Z_1,Z_2}(\zeta_1, \zeta_2)}{p_{Z_1}(\zeta_1)\cdot p_{Z_2}(\zeta_2)}\,{\rm d}\zeta_1{\rm d}\zeta_2.$$ It corresponds to the Kullback-Leibler distance between $p_{Z_1,Z_2}(\zeta_1, \zeta_2)$, the joint probability density of $Z_1$ and $Z_2$, and the product of their respective marginal distributions. In the discrete case, the integral is replaced by a finite sum.
In practice, the mutual information has to be estimated from the dataset, as the exact probability density functions in the above equations are not known. The most sensitive part of the estimation of the mutual information is the estimation of the joint probability density function $p_{Z_1,Z_2}(\zeta_1,
\zeta_2)$. Several methods have been developed in the literature to estimate such joint densities: histograms, kernel-based methods and splines [@Scott:1992]. All those estimators depend on at least one parameter that has to be chosen appropriately.
In the context of a forward procedure, the mutual information is estimated between a **set** of inputs $X_i$ (instead of a single variable $X_i$) and the output $Y$. The above definitions of entropy and mutual information remain valid, provided that $Z_1$ is replaced by a multi-dimensional variable. The dimension of the latter grows at each iteration of the forward procedure. Therefore the estimations of the $p_{Z_1}$ and $p_{Z_1,Z_2}$ densities must also be performed in spaces of increasing dimension.
Unfortunately, most of the density estimation methods require a sample whose size grows exponentially with both the dimension of $Z_1$ and the dimension of $Z_2$ to provide an accurate estimation. This is sometimes referred to as one instance of the curse of dimensionality [@Bellman:1961]. In practice, one seldom has the required number of points for an accurate estimation when the dimension is above 10. For dimensions below or close to that value, the estimation of the multi-dimensional mutual information can be performed with classical multivariate density estimators [@Bonnlander:1994; @Kwak:2002b]. With more than 10 dimensions the estimation becomes quite unreliable with those estimators. However, nearest neighbor-based density estimators have been reported to be less sensitive to dimensionality than many others [@Kraskov:2004; @Rossi:2005] and are therefore more suitable for the forward search strategy.
The forward search is incremental and “greedy” in the sense that the method makes final decisions about features at each iteration: once a feature is chosen, its relevance is never questioned again. The forward search will therefore perform at most $O(d^2)$ estimations of the criterion (rather than $2^d$ for the exhaustive search). The forward search begins with an empty set of features and adds at each iteration the feature that has the most positive influence on the criterion. The procedure is halted either when the *a priori* chosen number of features has been selected or when adding one more feature does not improve the relevance criterion.
Combining a forward search procedure with a mutual information estimator for the relevance criterion is an idea dating back to 1994 [@Battiti:1994]. Before the nearest neighbor estimator was popularized by Kraskov et al. [@Kraskov:2004], the multivariate mutual information measures were most often approximated using combinations of bi-variate [@Battiti:1994; @Kwak:2002a] or tri-variate [@Fleuret:2004] mutual information estimations. Those approximations, however, do not estimate the true value of the mutual information between the set of $X_i$ and $Y$, and make strong independence assumptions between the input features. The forward strategy with the mutual information estimated using nearest neighbors was shown to be successful [@Rossi:2005] and is used as the foundation method in the present paper. It however requires manual tuning of the number of neighbors and comparisons between the respective mutual informations between sets of features of different sizes and the output, which is not always advisable in practice, as detailed in Section \[sec:problems\].
Resampling methods
------------------
Additional information is needed to select a priori sound values (i) for the structural parameter of the estimator and (ii) for the number of selected features in the subset, without optimizing these numbers with respect to the prediction performances of the model. This additional information, namely an estimation of the variance of the estimator, is brought by two resampling methods: the permutation resampling and the K-fold resampling.
Resampling methods have heavy computational requirements that increase the time needed to perform the forward selection procedure. However, the running time of the scheme proposed in Section \[sec:proposedapproach\] remains acceptable compared to the computational burden of alternate solutions that could be used to choose the number of features and the parameter of the estimator (e.g. optimizing those elements based on the performances of a prediction model).
It should be noted that bootstrap resampling, while generally advisable for exploring the behavior of an estimator, is not adapted to the $k$ nearest neighbors estimator [@Kraskov:2004] used in this paper. When a bootstrap sample is generated from the original dataset, it contains duplicates of many of the observations. As a consequence, the $k$ nearest neighbors of each observation may contain this observation itself (sometimes even repeated), which leads to a strong overestimation of the mutual information.
### K-fold resampling
The K-fold resampling is very similar to the K-fold cross-validation scheme used for validating prediction models, except that it is used in an unsupervised manner. Given $z_1$ and $z_2$ respectively realizations of $Z_1$ and $Z_2$, and some statistic $\theta$, it consists in computing the $K$ estimates $\hat\theta_k$ of $\theta$ where one (or several) data element(s) has(ve) been removed from the analysis. Typically, the sample is partitioned into $K$ clusters of roughly equal size, and the statistic is estimated $K$ times on the sample from which the $K$th cluster was excluded. The average of those estimations is often found to be a more robust estimator of $\theta$, while the variance of the estimations gives an idea of the sensitivity of the estimator to the particular sample.
### The permutation test or randomized resampling
The permutation test [@good] is a nonparametric hypothesis test over some estimated statistic $\hat\theta$ involving $z_1$ and $z_2$. The statistic $\hat\theta$ can be a difference of means in a classification context, or a correlation, or, as in this paper, a mutual information. Let $\hat\theta$ be the estimation of the statistics for the given $z_1$ and $z_2$, both vectors of size $n$ drawn from $p_{Z_1}$ and $p_{Z_2}$ respectively. The aim of the test is to answer the following question : how likely is the value $\hat\theta$ given the vectors $z_1$ and $z_2$ if we suppose that $Z_1$ and $Z_2$ are independent? In particular, the value of the mutual information under such hypothesis should be zero.
The permutation test considers the empirical distribution of $z_1$ and $z_2$ to be fixed, as well as the sample size. The random variable of interest is the value of the statistic $\theta$. In such a framework, the distribution of $\hat\theta$ is the set of all values of $\hat\theta_k$ for all $n!$ possible permutations of the elements of the vector $z_1$, or, equivalently, all permutations of the elements of the vector $z_2$. The P-value $\alpha$ associated to the test is the proportion of $\hat\theta_k$ that are larger than the value of $\hat\theta$ estimated with $z_1$ and $z_2$ without permutation.
In practice, it is not necessary to perform all $n!$ permutations. Several tens or hundreds of them are randomly performed. In this case, the exact P-value cannot be known but a 95% confidence interval around the observed P-value can be estimated [@Opdyke:2003].
Combined uses
-------------
The permutation test has been extensively used in conjunction with the mutual information to perform a nonparametric statistical test of independence of variables or signals. It has been of much use in identifying nonlinear relationships between pairs of variables in exploratory analysis [@Craddock:2006; @Hahn:2005; @Hummel:2005; @Purushothaman:2005; @Hoffman:2003], and to test serial independence in time series [@Diks:2005].
The permutation test has also been used specifically to filter out features, by measuring independence via mean differences, student statistics, or chi-squared measures. The test is used, for instance, to discard features for which the independence hypothesis cannot be statistically rejected [@Conrad:2004], or to rank features according to the p-value estimated by the permutation test [@Radivojac:2004]. The permutation test can also be used in the process of building a decision tree, to choose the features that should be used at a split-point [@Frank:1998].
Feature filtering with the mutual information and the permutation test was also recently proposed [@VanDijk:2006; @Francois:2006; @Radivojac:2004], in a pure feature ranking approach where the permutation test is used to automatically set a threshold on the value of the mutual information.
Resampling approaches similar to the K-fold resampling (Jackknife, bootstrap, etc.) have also been used to get better estimates of the mutual information [@Zhou:2004] and to choose among several estimators (nearest neighbor-based, histogram-based, spline-based, etc) to estimate the mutual information between EEG signals [@Nicolaou:2005]. The estimator that is chosen is the one that is most robust with respect to resampling, i.e. that has the lowest variance around the estimated value.
Mutual information with permutation testing has thus been used for automatic feature filtering, that is for discarding features that are statistically non-relevant for the prediction. This approach however selects many features, more than necessary since redundancy in the features is not considered. That is why automatic forward selection is preferable to actually select features rather than discarding them. Furthermore, in choosing the value of the estimator structural parameter and the number of variables to consider in the forward search, we should not only consider the variance of the estimator but also, and more importantly, how well it discriminates dependent features (with $M\!I>0$) from independent ones (with $M\!I=0$). The methodology described in the next section answers these questions.
The sensitivity to parameter values {#sec:problems}
===================================
The mutual information, with a nearest neighbor-based estimator, and the forward search combined together present a good compromise between computation time and performances. As already discussed, two issues must be addressed however, namely the number of features to select and the choice of the parameter in the estimation of the mutual information to discriminate at best relevant features from useless ones. The results of the feature selection process highly depend on those two parameters, especially when the mutual information must be estimated from a few samples. This section illustrates those difficulties in a simple case.
The problem discussed here is a synthetic prediction problem, derived from Friedman’s [@friedman]. We consider 10 input variables $X_i$ and one output variable $Y$ given by $$\label{eq:friedman}
Y = 10 \sin\left(X_1 \cdot X_2\right) + 20 \left( X_3-0.5\right)^2 + 10 X_4 + 5 X_5 + \epsilon.$$ All $X_i, 1\leq i\leq 10$, are uniformly distributed over $[0,1]$, and $\epsilon$ is a centered Gaussian noise with unit variance. Variables $X_6$ to $X_{10}$ are just noise and have no predictive power. The sample size is 100.
Parameter sensitivity
---------------------
The number $k$ of neighbors taken into account in the estimation of the mutual information must be chosen carefully, especially in the case of a small sample and noisy data. If the number of neighbors is too small, the estimation will have a large variance; if the number of neighbors is chosen too large, the variance of the estimator will be small, but all estimations will converge to zero, even for highly-dependent variables.
![- Mutual information between the 10 variables of the synthetic example and the output, for several values of the estimator number of neighbors. All relevant features have a higher mutual information than non-relevant ones only for well-chosen values.[]{data-label="fig:illustratechoiceofk"}](illustratechoiceofk2){width="100.00000%"}
In practice, a bad choice of $k$ can modify the ranking between variables and lead to false conclusions. As an illustration, Figure \[fig:illustratechoiceofk\] displays the mutual information between each $X_i$ and $Y$, using the nearest neighbor-based estimator for a single dataset generated from Equation \[eq:friedman\]. The number of neighbors used in the estimation of the mutual information is shown at the top of the graphs. Although only features $X_1$ to $X_5$ are informative, they do not always have a mutual information larger than the other features. Furthermore, a significant, large, difference can be observed between $X_1$ and $X_2$ while they have the same influence on the output.
This simple experiment shows that the number of neighbors must be chosen correctly to avoid artefacts from the estimator, even in simple cases.
Stopping criterion instability
------------------------------
The stopping criterion of the forward search will determine how many features are selected. When nested subsets of features are considered, as in the forward search, the mutual information is theoretically a non-decreasing function of the subset size; it can only increase or remain constant as more features are added. Maximizing the mutual information therefore does not make sense: the whole feature subset will always, in theory, have the largest mutual information with the value to predict.
In practice however, as illustrated in Figure \[fig:Forward\], the evaluation of the mutual information tends to decrease when useless variables are added, especially with an estimator based on the distances between observations. It is therefore tempting to look for the maximum value of the mutual information. But again, as shown in Figure \[fig:Forward\], this will frequently lead to sub-optimal feature subsets. On this example, stopping the forward procedure at the first peak selects a wrong number of features in almost all cases. Moreover, searching for the global maximum does not improve a lot the situation: the optimal set of features is selected only in three cases (for $k$ equal to 1, 3, and 6).
![Result of the forward procedure on the artificial example with different values of the number of neighbors in the estimation of the mutual information. Only for well-chosen values of the number of neighbours the correct features ($X_1$ to $X_5$) are selected.[]{data-label="fig:Forward"}](friedman-forward.eps){width="100.00000%"}
In fact, there is no particular reason for this strategy (maximization of the mutual information) to give optimal results when the mutual information is estimated via a distance-based method. Indeed, the forward procedure tends to add features in their relevance order. Moreover, when a feature is included in the current subset, it has the same individual importance in the distance calculations as each previously selected feature. As a consequence, the influence of the previous features, which might be more relevant than the last one, on the mutual information estimator tends to decrease. As shown in Figure \[fig:Forward\], there are many cases in which the first five features are the optimal ones and yet the mutual information is not maximal for the five feature set. In fact, the forward procedure only fails for $k$ equal to 5, 7 and 9, when it selects the irrelevant feature 8 before the relevant feature 3. While an optimal choice of $k$ should in theory prevent estimator problems to lead to bad estimations of the mutual information, and therefore rule out values 5, 7, and 9 for $k$, we cannot guarantee that the optimal feature subset will correspond to the highest value of the estimated mutual information. This is in fact more an intrinsic limitation of the chosen estimator than a problem of its tuning; it is in a sense the price we have to pay for an estimator that is able to handle higher-dimensional data.
There is thus a need for a sound stopping criterion of the forward search based on the mutual information, in addition to the optimal choice of $k$.
Proposed methodology {#sec:proposedapproach}
====================
The number of neighbours {#sec:neighborChoice}
------------------------
In order to determine the optimal number of neighbors in the estimation of the mutual information, the notion of optimality must be explicitly defined since there is no obvious criterion that we could maximize or minimize. As already discussed, we do not want to optimize the number of neighbors with respect to the performances of a prediction model built with the variables chosen by the procedure, because this would render the search procedure too time-consuming.
The goal is to discriminate between features that are relevant for the problem and features that are useless. We therefore consider the optimal value of $k$ to be the value for which the separation between the relevant features and an independent feature is maximum. Since the estimator of the mutual information has some variance, it is important to take this variance into account in measuring the separability. If we had access to the distribution of the mutual information estimate over the data, we could calculate a separation between ${M\!I\left(X;Y\right)}$ and ${M\!I\left(U;Y\right)}$ (considered as random variables) for an important feature $X$ and an useless feature $U$.
To show the behavior of those variables on a simple example, 100 datasets are randomly generated from Equation \[eq:friedman\]. From those datasets, 100 realizations of the random variables ${M\!I\left(X_4;Y\right)}$ and ${M\!I\left(X_{10};Y\right)}$ are produced, for different values of $k$. Figure \[fig:BiasVariance\] represents the means of ${M\!I\left(X_4;Y\right)}$ and of ${M\!I\left(X_{10};Y\right)}$ over the 100 datasets, as well as the $0.01$ and $0.99$ percentiles of the same realizations. Those values are reliable estimates of the theoretical values of the considered quantities.
![Mutual information estimator distribution for datasets generated from Equation \[eq:friedman\]. Solid lines correspond to variable $X_4$ and dashed lines to variable $X_{10}$. See text for details.[]{data-label="fig:BiasVariance"}](friedman-mi-v4v10-quantile.eps){width="100.00000%"}
As expected, the variability of the estimator reduces with the number of neighbors. However, the mutual information ${M\!I\left(X_4;Y\right)}$ also decreases, whereas there is a strong relationship between $X_4$ and $Y$. For a low number of neighbors (1 and 2), the variability of the estimator is important enough to blur the distinction between $X_4$ and $X_{10}$ in term of potential predictive power: for some of the datasets, ${M\!I\left(X_{10};Y\right)}$ is larger than ${M\!I\left(X_4;Y\right)}$. When $k$ increases, the estimator becomes stable enough to show that $Y$ depends more on $X_4$ than on $X_{10}$ (for $k\geq 3$). However, after a first growing phase, the separation between the distributions of ${M\!I\left(X_4;Y\right)}$ and ${M\!I\left(X_{10};Y\right)}$ decreases with $k$: the reduction of the mean estimated value of ${M\!I\left(X_4;Y\right)}$ tends to negate the positive effect of the reduction of variability. The lowest values of ${M\!I\left(X_4;Y\right)}$ are getting closer and closer to the highest values of ${M\!I\left(X_{10};Y\right)}$. It seems therefore important to choose $k$ so as to ensure a good separation between relevant and irrelevant variables.
In practice however, the true distribution of ${M\!I\left(X;Y\right)}$ is unknown. We therefore rely on a combined K-fold/permutation test to estimate the bias and the variance of the estimator for relevant features and for independent ones. The idea is the following. Consider $X_i$ a feature that is supposed to be relevant to predict $Y$. Two resampling distributions are built for both ${M\!I\left(X_i;Y\right)}$ and ${M\!I\left(X_i^\pi;Y\right)}$ where $X_i^\pi$ denotes a randomized $X_i$ that is made independent from $Y$ through permutations. This is done by performing several estimations of (i) the mutual information between $X_i$ and $Y$ and (ii) the mutual information between a randomized version of $X_i$ and $Y$, using several non-overlapping subsets of the original sample, in a K-fold resampling scheme. A good value for $K$ is around 20 or 30. Less than 20 renders the estimation of mean and variance hazardous, while the estimations with more than 30 are often very close to those with K=30. The procedure results in two samples of estimates of ${M\!I\left(X_i;Y\right)}$ and ${M\!I\left(X_i^\pi;Y\right)}$.
The optimal value of $k$ is the one that best separates those two distributions, for instance according to a Student-like measure: $$t_{i,k} = \frac{\mu-\mu_\pi}{\sqrt{\sigma^2 + \sigma^2_\pi}},$$ where $\mu$ and $\sigma^2$ represent the mean and variance of the cross-validated distribution of ${M\!I\left(X_i;Y\right)}$, and $\mu_\pi$ and $\sigma^2_\pi$ are those of the cross-validated distribution of ${M\!I\left(X_i^\pi;Y\right)}$ (illustrated on Figure \[fig:illustratechoiceofkDistrib\]).
The optimal $k$ for all features is chosen as the one corresponding to the largest value of $t_{i,k}$ over all values of $k$ over all features. This way, features that are useless do not participate in the choice of the optimal value. Using useless features to choose the value that best separates the resampling of the mutual information from the permuted sample would indeed make no sense if they are independent from the output value. It should be noted that other solutions could be thought of, like, for instance, to optimize the mean value of $t_{i,k}$ over features for which $t_{i,k}$ is above a pre-specified significance threshold, but at the cost of an additional parameter.
![Distribution of mutual information for a relevant feature. On the left, the distribution of the mutual information of the features with permuted values, on the right, the distribution of the mutual information of the relevant feature; as given by the K-fold method. The value of $k$ is chosen so as to best separate those two distributions.[]{data-label="fig:illustratechoiceofkDistrib"}](illustratechoiceofk){width=".65\textwidth"}
The stopping criterion {#subsection:stoppingcriterion}
----------------------
As choosing the maximum or the peak of the mutual information is nor sound neither efficient, a more promising approach consists in trying to avoid adding useless features to the current subset by comparing the value of the mutual information with the added feature to the one without that feature in a way that incorporates the variability of the estimator.
Let us consider $S$, the subset of already selected features, and $X^*$, the best candidate among all remaining features. We consider the distribution of $MI\left( S \cup \{X^*\};Y \right)$ under the hypothesis that $X^*$ is independent from $Y$ and $S$, that is all values of $MI\left( S \cup
\{X^{*\pi}\};Y \right)$ where $X^{*\pi}$ is a random permutation of $X^*$. If the P-value of $MI\left( S \cup \{X^*\} ;Y\right)$ is small and the hypothesis is rejected, it means that $X^*$ brings sufficient new information about $Y$ to be added to the feature subset.
Note that this way, the increment in mutual information between $MI\left( S
\cup \{X^*\};Y \right)$ and $MI\left( S ;Y\right)$ is estimated without comparing estimations of mutual information on subsets of different sizes. In theory this should not be an issue; in practice however, it is important. Indeed, as we observed before, adding an informative variable should, in theory, strictly increase the mutual information, but the contrary is frequently observed (see for example Figure \[fig:Forward\].)
![Mutual information in a forward feature subset search on the toy example. Thresholds (horizontal lines) are computed as the 95% percentiles of the permutation distribution; the actual mutual information is represented with circles. The number of neighbors is $k=19$ (selected according to the criterion proposed in section \[sec:neighborChoice\]). []{data-label="fig:forwardtoy"}](forwardtoy){width=".65\textwidth"}
Figure \[fig:forwardtoy\] summarizes the results of the proposed stopping criterion applied to the synthetic dataset introduced above. The procedure selects the right features ($X_1$ to $X_5$) and finds that the sixth added feature does not improve the mutual information significantly. As already shown on Figure \[fig:Forward\], the mutual information decreases when the third feature is added, which can wrongfully be taken as a clue that the procedure should be halted. The permutation test is able to cope with the instabilities of the estimator and to detect the relevance of the added feature even if it makes the mutual information decrease.
Computational burden
--------------------
In most traditional resampling schemes, the overall computation time is simply multiplied by the number of resamplings performed. In this case however, a more detailed analysis is needed to grasp the overhead cost brought by the proposed method.
The number of mutual information estimations to perform at iteration $t$ in the forward search, is equal to the number $d-t+1$ of features that are candidate for entering the optimal feature subset plus the number $P$ of permutations performed to evaluate the threshold of the stopping criterion. The cost of each iteration, in terms of mutual information estimation, thus amounts to $ d-t+1+P$. As the number of permutations is often limited to 20 or 30, the additional cost at each iteration needed to estimate the threshold is small compared with the cost needed to find the feature that should be added to the optimal feature subset. For instance, on a 100-dimensional dataset (like the Delve census dataset presented in Section \[sec:delve\]), 955 estimations of the mutual information are needed to find the optimal subset of size 10 while 200 estimations, that is a bit more than 15% were used to set the threshold. Of course, when the number of original features is small, permutations tend to represent a more important part of the total computational burden.
The cost of the choice of the optimal number of neighbors is roughly equal to the cost of the first step of the forward search multiplied by $K$, the number of folds in the cross-validation scheme used in the proposed method. In practice, $K$ is chosen between 20 and 30. If the expected number of optimal features has the same order of magnitude, the total cost of the forward procedure will also be of the same order of magnitude than the cost of the cross-validation, which means that the overall cost is roughly doubled. However, this is much less than if the number of neighbors was optimized using the performances of the prediction model, as this would imply performing as many forward searches as the number of values that are tested.
The total cost of the automatic determination of the parameters, in the case of high-dimensional data, is thus a bit more than the double of the cost when the number of neighbors is chosen arbitrarily and the mutual information is maximized. This additional cost brings in better and more stable results, as shown in the next section.
Experiments {#sec:experiments}
===========
This section presents further experiments on the synthetic example and on three real-world datasets.
A simulation study
------------------
To further validate the interest of the proposed approach, the forward procedure is applied to 100 datasets randomly generated from Equation \[eq:friedman\]. For each dataset, the optimal value of $k$ is selected between 1 and 20, then the forward procedure is conducted. The feature set that maximizes the mutual information and the best feature set according to the stopping criterion presented in the previous section are retained for comparison. Results are summarized in Tables \[table:featuresubsetsize:friedman\], \[table:informativefeaturesubsetsize:friedman\] and \[table:uselessfeaturesubsetsize:friedman\].
Number of features 1 2 3 4 5 6
---------------------------- --- ---- ---- ---- ---- ---
Maximal mutual information 7 45 33 14 1 0
Stopping criterion 0 1 12 52 29 6
: Number of feature subsets of a given size obtained by both criteria[]{data-label="table:featuresubsetsize:friedman"}
It appears clearly from Table \[table:featuresubsetsize:friedman\] that maximizing the mutual information does not provide good results: this leads to the selection of an optimal set of features (5 variables) only in one case out of one hundred. The stopping criterion defined in Section \[subsection:stoppingcriterion\] tends to select more features: in fact, the feature sets obtained by this methods have strictly more features that the ones selected by maximizing the mutual information in 84 % of the cases (and equal sizes in other situations).
Number of informative features 1 2 3 4 5
-------------------------------- --- ---- ---- ---- ---- --
Maximal mutual information 7 45 33 14 1
Stopping criterion 0 1 16 66 17
: Number of feature subsets that contain the specified number of relevant features obtained by both criteria[]{data-label="table:informativefeaturesubsetsize:friedman"}
Moreover, the additional features are generally informative ones, as illustrated by Table \[table:informativefeaturesubsetsize:friedman\]. The positive aspect of maximizing the mutual information is that it leads, on those experiments, only to the selection of relevant features. The stopping criterion proposed in Section \[subsection:stoppingcriterion\] selects sometimes irrelevant features (see Table \[table:uselessfeaturesubsetsize:friedman\]), but it also selects always at least as much relevant features as the former method. Moreover, in 79% of the experiments, it selects strictly more relevant features than the maximizing strategy. In 5 % of the experiments, the feature set selected by the significance stopping criterion consists in the set that maximizes the mutual information with an additional uninformative variable: this corresponds to the error level expected as the forward procedure was controlled by using the 95% percentile of the permutation distribution.
Number of uninformative features 0 1 2
---------------------------------- ---- ---- ---
Stopping criterion 75 22 3
: Number of feature subsets that contain the specified number of irrelevant features obtained by the stopping criterion of Section \[subsection:stoppingcriterion\][]{data-label="table:uselessfeaturesubsetsize:friedman"}
This simulation study shows that while the proposed stopping criterion is not perfect, it provides significant improvements over the standard practice of maximizing the mutual information. Moreover, it does not lead to the selection of too large feature sets that would reduce its practical benefit.
The utility of the method is further illustrated below on a well-known dataset from the UCI Machine Learning Repository (Housing), on a high-dimensional nitrogen spectra data and on a high-dimensional data set from the Delve repository.
The HOUSING dataset
-------------------
The goal with the Housing dataset is to predict the value of houses (in k\$) described by 13 attributes representing demographic statistics of the area around each house. The dataset contains 506 instances split into 338 learning examples and 169 test ones.
The optimal value (on the learning set) of $k$, searched between 1 and 20, is found to be 18.
The forward search procedure described in the previous section is run with 50 permutations on the learning examples. The threshold p-value is set to 0.05. When the mutual information is below the 95% percentile of the permutation distribution, the procedure is halted.
Figure \[fig:forwardhousing\] displays the mutual information as a function of the forward search iterations. The horizontal lines correspond to the critical values (i.e. the 95% percentile of the permutation distribution) while the circles represent the mutual information between the selected subset and the value to predict. Four features are selected ($X_6$, $X_{13}$, $X_1$ and $X_4$). Interestingly, the procedure does not stop when the peak in mutual information is observed.
A RBFN model was built using the selected features and optimized by 5-fold cross-validation on the learning set, according to the method described in [@Benoudjit:2003]. The root mean squared error (RMSE) on the test set is 9.48. By comparison, the RMSE on the test set with the all set of features is 18.97, while the RMSE with the first two features, corresponding to the peak in mutual information, is 19.39.
![The evolution of the mutual information in a forward feature subset search on the Boston Housing dataset. Thresholds (horizontal lines) are computed as the 95% percentile of the permutation distribution; the actual values of the mutual information are represented with circles. The procedure stops after four features have been selected (dashed line).[]{data-label="fig:forwardhousing"}](forwardhousing){width=".65\textwidth"}
The nitrogen dataset
--------------------
The nitrogen dataset originates from a software contest organized at the International Diffuse Reflectance Conference[^1] held in 1998 in Chambersburg, Pennsylvania, USA. It consists of scans and chemistry gathered from fescue grass (*Festuca elatior*). The data set contains 141 spectra discretized to 1050 different wavelengths, from 400 to 2 498 nm. The goal is to determine the nitrogen content of the grass samples (ranging from 0.8 to 1.7 approximately). The data can be obtained from the Analytical Spectroscopy Research Group of the University of Kentucky[^2].
The dataset is split into a test set containing 36 spectra and a training set with the remaining 105 spectra. We apply moreover a functional preprocessing, as proposed in [@RossiEtAl05Neurocomputing]: this consists in replacing each spectrum by its coordinates on a B-spline basis, which is itself selected by minimizing a leave-one-out criterion (see [@RossiEtAl05Neurocomputing] for details). The purpose of this functional preprocessing is to reduce the huge number of original features (1050) to a more reasonable number: the optimal B-spline basis consists indeed in 166 B-splines of order four.
Figure \[fig:forwardnitro\] illustrates the behavior of the forward feature selection with resampling on this dataset. The optimal number of neighbors is 12. It leads to the selection of 25 variables (among the 166 B-spline coordinates). The RMSE on the test set, using a RBFN model built on those features, is 0.6649.
Maximizing the mutual information leads to a smaller feature set with 6 features. The RMSE on the test, using a RBFN model built on those features, is 0.7753. As a reference, the RMSE on the test set when all features are used is 3.1197.
![The evolution of the mutual information in a forward feature subset search on the nitrogen dataset. Thresholds (horizontal lines) are computed as the 95% percentile of the permutation distribution; the actual values of the mutual information are represented with circles. Twenty-five features are selected.[]{data-label="fig:forwardnitro"}](forwardnitro){width=".65\textwidth"}
The Delve-Census dataest {#sec:delve}
------------------------
The Delve Census dataset, available directly from the University of Toronto[^3], comprises data collected during the 1990 US Census. Each of 22,784 the data elements concerns a small survey region and is described by 139 features measuring demographic information like the total person count in the region, the proportion of males, the percentage of people aged between 25 and 64, etc. The aim is to predict the median price of the houses in each survey region. This problem can be considered as a large scale version of the Housing dataset.
For the sake of this analysis, we used only 104 of the 139 original features. We indeed discarded the features that are too much correlated with the value to predict like the average price, the first and third percentiles, etc. In the dataset, 52 regions were found to have a median house price of zero; they were considered to be erroneous and removed from the analysis[^4].
Of the 22,732 remaining observations, 14,540 are used for the test set. The 8192 remaining observations are split into 8 subsets and used for training. This corresponds to the classical splitting for this dataset; it allows one to study the variability of the feature selection procedure while retaining enough data both for learning and testing. For each observation subset, the optimal feature subset is determined using the proposed approach and a RBFN model is built using a 3-fold cross validation procedure. The RBFN model is then applied on the test set and the results are compared with those obtained using the peak in mutual information and using all features.
![The evolution of the mutual information in a forward feature subset search on the first subset of the Delve dataset. Thresholds (horizontal lines) are computed as the 95% percentile of the permutation distribution; the actual values of the mutual information are represented with circles. Eight features are selected.[]{data-label="fig:forwarddelve"}](forward-delve-1){width=".65\textwidth"}
Figure \[fig:forwarddelve\] displays the evolution of the mutual information and of the thresholds found by permutation over each iteration of the forward search procedure. Figure \[fig:forwarddelve\] shows the results of the first of the eight learning sets. The number of selected features is eight, while the maximum of mutual information is observed for six features.
Table \[table:resultsdelve\] shows the Root Mean Square Error of the model on the test set, for each learning subset. The permutation approach always selects either 8 or 9 features, while stopping the forward procedure at the peak of mutual information gives from 2 to 6 features. Except for Subset number 2, the results obtained with the permutation are either equivalent either far better than those obtained with features selected by taking the peak of mutual information.
--------------- ------------- --------------------------- -------------- --------------------------- ---------------------------
All features
Subset number \# features test RMSE ($\times 10^4$) \# features test RMSE ($\times 10^4$) test RMSE ($\times 10^4$)
1 8 1.3286 6 1.3223 1.4304
2 9 1.0748 6 0.9472 1.5393
3 8 1.2883 3 2.5643 1.4338
4 8 1.2214 2 2.3125 1.419
5 9 1.2575 3 1.1799 1.4628
6 8 0.9504 5 2.363 1.4146
7 8 1.1987 2 2.2381 2.1855
8 9 1.1929 3 1.19 1.5314
--------------- ------------- --------------------------- -------------- --------------------------- ---------------------------
: Root mean square error on the test set obtained by the RBFN built on each learning subset. []{data-label="table:resultsdelve"}
Discussion
----------
The three real-world examples illustrate the gain in prediction performances that can be obtained when using a well-chosen subset of features. Simulations show the significant improvements obtained when using the proposed method for selecting the subset, rather than using as traditionally the peak of the mutual information, or the full set of features.
It appears therefore that the proposed strategy allows the automatic selection of good subsets of the original feature set. Moreover, it could easily be combined with a simple wrapper approach that compares the feature set that maximizes the mutual information with the one obtained by the proposed method. This would further increase the robustness of the feature selection process without leading to the enormous computation time that would be required by a full wrapper forward selection process.
Conclusions {#sec:conclusions}
===========
Combining the use of the mutual information and a forward procedure is a good option for feature selection. It is indeed faster than a wrapper approach (that uses the prediction model itself for all evaluations) and still make very few assumptions about the data as it is nonlinear and nonparametric. The major drawback of this approach is that the estimation of the mutual information is often difficult in high-dimensional spaces, i.e. when several features have already been selected.
Nearest neighbor-based mutual information estimators are one of the few sustainable options for such estimation. However, two issues must be addressed. The first one is the choice of the parameter of the estimator, namely the number of neighbors. This number must be chosen carefully, especially with high-dimensional subsets. The second one is the number of features to select, or, equivalently, when to halt the forward procedure.
These two parameters of the approach could be optimized with respect to the performances of the prediction model, but this would require a large amount of computations. Rather, resampling methods can be used.
In this paper, the K-fold and permutation resamplings are used in a combined way to obtain an estimate of the variance of the estimator both in the case of relevant features and of independent ones. The optimal number of neighbors is then chosen so as to maximize the separation between the two cases.
Once the number of neighbors is chosen, the forward procedure may begin. It is halted when the added feature does not significantly increase the mutual information compared with the estimation of the mutual information if the same feature was independent from the value to predict. This is done using the permutation test.
Combining the forward feature selection procedure, the mutual information to estimate the relevance of the input subsets and resampling methods to estimate the reliability of the estimation thus brings a feature selection methodology that is faster than a wrapper approach and only requires the user to choose a significance level; all other parameters are set in an automated way.
The method is illustrated on a synthetic dataset, as well as on three real-world examples. The method is shown to perform better than choosing the peak in mutual information. The test error of a Radial Basis Function Network built with the features selected by the method is always much lower than if the whole set of features is used and significantly lower than if the features up to the peak in mutual information are used.
Although the procedure described here uses a forward feature selection, it could be used as well with other incremental search methods like backward feature elimination, or add-$r$ remove-$s$ methods that remove and/or add several features at each step. Adaptive methods could be used also to detect when performing more permutations is not necessary (for instance the variance in the permuted data gets to a stable value). Furthermore, this paper focusses on mutual information because it has been shown to be well adapted to forward feature selection, but the methodology could be applied to quadratic mutual information [@Principe] or to the Gamma test [@Jones] as well.
Acknowledgments {#acknowledgments .unnumbered}
===============
D. François is funded by a grant from the Belgian F.R.I.A. Parts of this research result from the Belgian Program on Interuniversity Attraction Poles, initiated by the Belgian Federal Science Policy Office. The scientific responsibility rests with its authors.
[^1]: <http://www.idrc-chambersburg.org/index.htm>
[^2]: <http://kerouac.pharm.uky.edu/asrg/cnirs/shoot_out_1998/>
[^3]: <http://www.cs.toronto.edu/~delve/data/census-house/desc.html>
[^4]: The preprocessed data can be downloaded from the UCL Machine Learning Group website: <http://www.ucl.ac.be/mlg>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is commonly believed that in the absence of disorder or an external magnetic field, there are three possible types of superconducting excitation gaps: the gap is nodeless, it has point nodes, or it has line nodes. Here, we show that for an even-parity nodal superconducting state which spontaneously breaks time-reversal symmetry, the low-energy excitation spectrum generally does not belong to any of these categories; instead it has extended Bogoliubov Fermi surfaces. These Fermi surfaces can be visualized as two-dimensional surfaces generated by “inflating” point or line nodes into spheroids or tori, respectively. These inflated nodes are topologically protected from being gapped by a $\mathbb{Z}_2$ invariant, which we give in terms of a Pfaffian. We also show that superconducting states possessing these Fermi surfaces can be energetically stable. A crucial ingredient in our theory is that more than one band is involved in the pairing; since all candidate materials for even-parity superconductivity with broken time-reversal symmetry are multiband systems, we expect these $\mathbb{Z}_2$-protected Fermi surfaces to be ubiquitous.'
author:
- 'D. F. Agterberg'
- 'P. M. R. Brydon'
- 'C. Timm'
title: 'Bogoliubov Fermi Surfaces in Superconductors with Broken Time-Reversal Symmetry'
---
*Introduction*.—The theory of superconductivity is conventionally formulated in terms of pairing of spin-$1/2$ fermions [@vol85; @sig91]. The complications introduced by additional electronic degrees of freedom, e.g., orbitals, are not usually thought to qualitatively alter the physics. This picture has recently been challenged for a number of materials. For iron-based superconductors, the role of interorbital pairing is attracting increased attention [@GSZ10; @NGR12; @NKT16; @ong16]. Another example is the nematic superconductivity of Cu$_x$Bi$_2$Se$_3$ [@CuBi2Se3], where the odd parity of the gap is encoded in the orbital degrees of freedom. Furthermore, theories of pairing in YPtBi and UPt$_3$ based on $j=3/2$ and $j=5/2$ fermions, respectively, have greatly enriched the allowed superconducting states [@bry16; @NoI16].
In this Letter, we show that the presence of multiple bands qualitatively changes the nodal structure of a time-reversal-symmetry-breaking (TRSB) superconductor. Specifically, the expected line or point nodes of an even-parity superconducting gap [@vol85; @sig91] are replaced by two-dimensional Fermi surfaces of Bogoliubov quasiparticles, which are topologically protected by a $\mathbb{Z}_2$ invariant. We further interpret these Fermi surfaces in terms of a pseudomagnetic field arising from interband Cooper pairs, here referred to as “interband pairing”.
Our conclusions are relevant for a wide range of candidate TRSB superconductors, such as UPt$_3$ [@luk93; @joy02; @sch14; @NoI16], Th-doped UBe$_{13}$ [@hef90; @zie04], PrOs$_4$Sb$_{12}$ [@aok03], Sr$_2$RuO$_4$ [@luk98; @xia06], URu$_2$Si$_2$ [@kas07; @sch15], SrPtAs [@bis13], and Bi/Ni bilayers [@gon16]. Remarkably, signatures of these Fermi surfaces may have already been observed in Th-doped UBe$_{13}$ [@zie04] (and possibly in UPt$_3$ [@joy02]), where there is evidence for a non-zero density of states at zero temperature, which appears not to be due to impurities. In addition to these known superconductors, theory has predicted TRSB superconductivity in graphene [@nan12; @abs14], the half-Heusler compound YPtBi [@bry16], water-intercalated sodium cobaltate Na$_x$CoO$_{2}\cdot y$H$_2$O [@sar04; @kie13], Cu-doped TiSe$_2$ [@gan14], and monolayer transition metal dichalcogenides [@hsu16]. A common feature of all these materials is that the electronic structure involves multiple bands, and so we expect our theory to apply.
*Model*.—Our starting point is a general Hamiltonian with time-reversal and inversion symmetries containing four electronic degrees of freedom at each momentum. These four degrees of freedom can arise from either the combination of spin $1/2$ and two orbitals of equal parity, or from fermions with effective angular momentum $j=3/2$. We work with a $j=3/2$ generalized Luttinger-Kohn Hamiltonian [@lut55] in order to make our arguments most transparent. Although these two descriptions have different symmetry properties, we show in the Supplemental Material (SM) [@suppinfo] that they can be unitarily transformed into each other and that our model represents a generic two-band theory including all symmetry-allowed crystal-field and spin-orbit-coupling terms. The general form of the normal-state Hamiltonian is $$\begin{aligned}
H_N &= c_0 1_4
+c_{yz}\frac{J_y J_z+J_z J_y}{\sqrt{3}}
+c_{xz}\frac{J_x J_z+ J_z J_x}{\sqrt{3}} \nonumber \\
&+c_{xy}\frac{J_x J_y+J_y J_x}{\sqrt{3}}
+c_{3z^2-r^2}\frac{2J_z^2-J_x^2-J_y^2}{3} \nonumber \\
&+c_{x^2-y^2}\frac{J_x^2-J_y^2}{\sqrt{3}} ,
\label{Ham0}\end{aligned}$$ where $1_4$ is the $4\times4$ unit matrix, and the $J_i$ are spin-$3/2$ matrices given in the SM [@suppinfo]. The coefficients $c_i=c_i({\bf k})$ of the matrices in [Eq. (\[Ham0\])]{} are real, satisfy $c_i({\bf k})=c_i(-{\bf k})$, and transform in the same way as the corresponding matrix under spatial symmetries. $H_N$ has twofold degenerate eigenvalues $\epsilon_\pm = c_0
\pm (c_{yz}^2+c_{xz}^2+c_{xy}^2+c_{3z^2-r^2}^2+c_{x^2-y^2}^2)^{1/2}$. For definiteness, we discuss the case of only one of the two bands crossing the chemical potential so that there is only one normal-state Fermi surface; if there are two, identical arguments pertain to both. While our conclusions are general, numerical results are given for the spherically symmetric Hamiltonian $H_N = \alpha k^2 + \beta\, (\mathbf{k}\cdot\mathbf{J})^2 - \mu$ [@lut55], where $\alpha$ and $\beta$ are constants ($\beta$ is the spin-orbit coupling) and $\mu$ is the chemical potential, which leads to $c_0=(\alpha+5\beta/4)k^2-\mu$, $c_{3z^2-r^2}=\beta[k_z^2-(k_x^2+k_y^2)/2]$, $c_{x^2-y^2}=\sqrt{3}\beta(k_x^2-k_y^2)/2$, $c_{yz}=\sqrt{3}\beta k_yk_z$, $c_{xz}=\sqrt{3}\beta k_xk_z$, and $c_{xy}=\sqrt{3}\beta k_xk_y$.
The superconducting state is taken to have even parity. Fermionic antisymmetry permits six possible gap matrices $\eta_i$ in the spin-$3/2$ space: $\eta_s=U_T$, $\eta_{yz}=(J_y J_z+J_z J_y) U_T/\sqrt{3}$, $\eta_{xz}=(J_x J_z+J_z J_x)U_T/\sqrt{3}$, $\eta_{xy}=(J_x J_y+J_y J_x) U_T/\sqrt{3}$, $\eta_{3z^2-r^2}=(2J_z^2-J_x^2-J_y^2) U_T/3$, and $\eta_{x^2-y^2}=(J_x^2-J_y^2) U_T/\sqrt{3}$, where $$U_T = \left(
\begin{array}{cccc}
0 & 0 & 0 & 1 \\
0 & 0 & -1 & 0 \\
0 & 1 & 0 & 0 \\
-1 & 0 & 0 & 0
\end{array}
\right)$$is the unitary part of the time-reversal operator $T=U_T\mathcal{K}$, with $\mathcal{K}$ denoting complex conjugation. The $\eta_s$ gap is a spin-singlet state and represents pure intraband pairing. The other gaps, however, describe spin-quintet ($J=2$) pairs and involve both intra- and interband pairing [@bry16]. Since we consider zero-momentum Cooper pairs, this implies that quintet pairing involves states away from the Fermi energy. A general superconducting state is a linear combination of these gap matrices with symmetry-compliant $\mathbf{k}$-dependent coefficients and is described by the Bogoliubov-de Gennes Hamiltonian $$H(\mathbf{k}) =\left(
\begin{array}{cc}
H_N({\bf k}) & \Delta({\bf k}) \\
\Delta^{\dagger}({\bf k}) & -H_N^T(-\mathbf{k})
\end{array}
\right) .
\label{Ham}$$ While our results apply to all TRSB even-parity superconducting states, for concreteness we consider the gap function $$\Delta({\bf k})=\Delta_{1}\,\psi({\bf k})\,\eta_s
+\Delta_0\,(\eta_{xz}+i\eta_{yz}) ,
\label{gap}$$ where $\Delta_1$ and $\Delta_0$ are real constants. Although the latter term describes purely on-site pairing, the gap matrix $\eta_{xz}+i\eta_{yz}$ transforms under rotations as the spherical harmonic $Y_{2,1}(\hat{{\bf k}})$, i.e., it is *chiral*. It generically accompanies a spin-singlet term with a form factor $\psi({\bf k})$ of the same symmetry. Being chiral, this pairing state, which we call the $k_z(k_x+ ik_y)$ state, breaks time-reversal symmetry. It has the same symmetry as proposed for URu$_2$Si$_2$ [@kas07], YPtBi [@bry16], and UPt$_3$ [@joy02]. For pure singlet pairing (i.e., $\Delta_0=0$), the gap has line nodes in the $k_z=0$ plane and point nodes on the $k_z$ axis ($k_x=k_y=0$). Mixing in a quintet component has a dramatic effect on the excitation spectrum: *the expected point and line nodes are replaced by Fermi surfaces*. In Fig. \[figure1\], we plot these Fermi surfaces for the $k_z(k_x+ik_y)$ state. We find them to be a generic feature of all the TRSB even-parity states classified in [@sig91]. As we will see below, these Fermi surfaces bear some resemblance to those found in the presence of an exchange field [@gub05], although they have a completely different origin.
![Bogoliubov Fermi surfaces of the superconducting $k_z(k_x+ik_y)$ state, shown here for the case where only one band has a Fermi surface. The normal-state Fermi surface, shown as the semitransparent sphere, is gapped out by the superconductivity. The point and line nodes of the single-band theory (red dots and line, respectively), however, are “inflated” into spheroidal and toroidal $\mathbb{Z}_2$-protected Fermi surfaces (orange surfaces).[]{data-label="figure1"}](fig1.pdf){width="0.8\columnwidth"}
*Existence of Fermi surfaces and $\mathbb{Z}_2$ invariant*.—We now show that Bogoliubov Fermi surfaces are a generic feature of the Hamiltonian in [Eq. (\[Ham\])]{} and construct their topological invariant. The first step is to show that $H(\mathbf{k})$ can be unitarily transformed into an antisymmetric matrix, i.e., there exists a unitary $\Omega$ such that $\tilde{H}^T(\mathbf{k})=-\tilde{H}(\mathbf{k})$ for $\tilde{H}(\mathbf{k})\equiv \Omega H(\mathbf{k}) \Omega^{\dagger}$. The main ideas of the proof are explained in the following, details and a representative $\Omega$ are given in the SM [@suppinfo]. The Hamiltonian $H(\mathbf{k})$ possesses charge-conjugation symmetry $C$ and parity symmetry $P$. $C$ acts as $U_CH^T(-{\bf k})U^{\dagger}_C=-H({\bf k})$ with $U_C=\hat{\tau}_x \otimes 1_4$, where $\hat{\tau}_i$ are the Pauli matrices in particle-hole space, while $P$ acts as $U_PH(-{\bf k})U_P^{\dagger}=H({\bf k})$ with $U_P=\hat{\tau}_0 \otimes
1_4$. Hence, $CP$ symmetry reads $$U_{CP}\, H^T(\mathbf{k})\, U_{CP}^\dagger = -H(\mathbf{k}) ,
\label{CPsymmetry}$$ with $U_{CP}=U_CU_P^*=\hat{\tau}_x \otimes 1_4$. This implies that $(CP)^2=U_{CP}U_{CP}^*=+1$ and thus $U_{CP}=U_{CP}^T$. Any symmetric matrix can be diagonalized by a unitary congruence, i.e., there exist a unitary $Q$ and a diagonal $\Lambda$ such that $U_{CP}=Q\Lambda Q^T$ with the *transposed* matrix $Q^T$. Insertion into [Eq. (\[CPsymmetry\])]{} gives $Q \Lambda Q^T\, H^T(\mathbf{k})\, Q^* \Lambda^\dagger Q^\dagger
= -H(\mathbf{k})$. Since $\Lambda$ is diagonal, we can define a square root $\sqrt{\Lambda}$. The sign of the root for each diagonal component of $\Lambda$ can be chosen arbitrarily and is then held fixed. This allows us to split the unitary matrices in the $CP$ symmetry relation, which gives $(\sqrt{\Lambda}{}^\dagger Q^\dagger H(\mathbf{k}) Q \sqrt{\Lambda})^T
= - \sqrt{\Lambda}{}^\dagger Q^\dagger H(\mathbf{k}) Q \sqrt{\Lambda}$. This can be written as $(\Omega H(\mathbf{k}) \Omega^\dagger)^T
= - \Omega H(\mathbf{k}) \Omega^\dagger$ with $\Omega \equiv \sqrt{\Lambda}{}^\dagger Q^\dagger$. Hence, $\tilde{H}(\mathbf{k}) = \Omega H(\mathbf{k})
\Omega^\dagger$ is antisymmetric, as we wanted to show. We can then define the Pfaffian $P({\bf k})\equiv\text{Pf}\, \tilde{H}({\bf k})$. Since $\det H({\bf k})=\det \tilde{H}(\mathbf{k})=P^2({\bf k})$, the zeros of $P({\bf k})$ give the zero-energy states of $H(\mathbf{k})$.
It has recently been shown that Fermi surfaces of Hamiltonians with $CP$ symmetry squaring to $+1$ can possess a non-trivial $\mathbb{Z}_2$ charge, making them topologically stable against $CP$-preserving perturbations [@kob14; @zha16]. We now express the invariant in terms of the Pfaffian $P(\mathbf{k})$. The Pfaffian is real since it is a polynomial of even degree of the components of the matrix $\tilde{H}(\mathbf{k})$, which is hermitian and antisymmetric and thus purely imaginary. Regions in momentum space in which $P(\mathbf{k})$ has opposite signs are thus necessarily separated by a two-dimensional (Fermi) surface on which $P(\mathbf{k})=0$. The existence of such Fermi surfaces is thus guaranteed if $P(\mathbf{k})$ changes sign and hence we can identify $(-1)^l=\text{sgn}[P({\bf k}_{-})P({\bf k}_{+})]$ as the $\mathbb{Z}_2$ invariant, where ${\bf k}_{+}$ (${\bf k}_{-}$) refers to momenta inside (outside) of the Fermi surface.
Under what conditions do protected Fermi surfaces exist? In the normal state, the Pfaffian $P(\mathbf{k}) = \epsilon_+^2
\epsilon_-^2$ is always non-negative and has second-order zeros on the Fermi surface. Hence, the normal-state Fermi surface is not protected by the $\mathbb{Z}_2$ charge. Furthermore, for *time-reversal-symmetric* superconductivity, $P(\mathbf{k})$ can also be chosen non-negative for all $\mathbf{k}$ so that there is no non-trivial $\mathbb{Z}_2$ invariant [@kob14; @zha16]. Our proof in the SM [@suppinfo] simplifies the earlier proof by Kobayashi *et al.* [@kob14]. For TRSB pairing, the second-order zeros of $P(\mathbf{k})$ are generically (i.e., in the absence of additional symmetries or fine tuning) lifted, leading to $P(\mathbf{k}) > 0$ in a neighborhood of the former zero, or split into first-order zeros, in which case there is a region with $P(\mathbf{k}) < 0$. Such a region is bounded by a two-dimensional Fermi surface; as this shrinks to a point or line node in the limit of infinitesimal pairing, we call it an “inflated” node.
The existence of protected Bogoliubov Fermi surfaces is now illustrated for the $k_z(k_x+ik_y)$ state of [Eq. (\[gap\])]{}. We restrict ourselves to pure quintet pairing but our results hold for any gap with nonzero quintet component, see the SM [@suppinfo]. The Pfaffian is $$P({\bf k}) = \epsilon_+^2\epsilon_{-}^2 + 4 \Delta_0^2\,
(\epsilon_+\epsilon_{-}+c_{xz}^2+c_{yz}^2) ,
\label{Pf}$$ which is negative for all $\mathbf{k}$ such that $s_-<\epsilon_{+}\epsilon_{-}<s_+$, where $s_\pm= -2\Delta_0^2 \pm 2\Delta_0\,(\Delta_0^2-c_{xz}^2-c_{yz}^2)^{1/2}$. $s_+$ and $s_-$ exist and are distinct if the radicand is positive. In the plane $k_z=0$ and along the line $k_x=k_y=0$, this holds for *any* $\Delta_0>0$ since symmetry dictates that $c_{xz}^2+c_{yz}^2$ vanishes there. We then find $s_-=-4\Delta_0^2<0$ and $s_+=0$, i.e., $s_+$ vanishes on the normal-state Fermi surface. Since $\epsilon_{+}\epsilon_{-}$ changes sign across the normal-state Fermi surface, there is always a region with $s_-<\epsilon_{+}\epsilon_{-}<s_+$ in this plane and along this line and, due to the continuity of the $c_i(\mathbf{k})$, also in their neighborhood. The resulting region with $P(\mathbf{k})<0$ is bounded by a Fermi surface, as illustrated in Fig. \[figure1\].
*Stability of Fermi surfaces*.—Even-parity TRSB superconductors are usually argued to be energetically favored over time-reversal-symmetric states because they maximize the gap in momentum space [@sig91]. The existence of extended Bogoliubov Fermi surfaces invalidates this argument. To show that TRSB states can nevertheless be stable, we consider a model with an on-site pairing interaction of strength $V$ in both the quintet $\eta_{xz}$ and $\eta_{yz}$ channels. In this case, the TRSB state $\Delta_0\,(\eta_{xz}+i\eta_{yz})$ \[i.e., the $k_z(k_x+ik_y)$ state introduced above\] and the time-reversal symmetric state $\sqrt{2}\,\Delta_0\,\eta_{xz}$ have the same critical temperature $T_c$. To decide which is energetically stable at temperatures $T$ near $T_c$, we perform a standard expansion of the free energy $F$ in the gap $\Delta$ [@ho99; @min98], $$\begin{aligned}
F &=\frac{1}{2V}\, \text{Tr}\,\Delta^{\dagger}\Delta \notag \\
& + \frac{k_B T}{2} \sum_{\mathbf{k},\omega_n} \sum_{l=1}^{\infty}
\frac{1}{l}\, \text{Tr}\, \big[\Delta\tilde{G}(\mathbf{k},\omega_n)
\Delta^\dagger G(\mathbf{k},\omega_n)\big]^l ,
\label{eq:freeE}\end{aligned}$$ where $G(\mathbf{k},\omega_n)$ and $\tilde{G}(\mathbf{k},\omega_n)$ are the normal-state electron and hole Matsubara Green’s functions, respectively. For vanishing spin-orbit coupling $\beta$, introduced below [Eq. (\[Ham0\])]{}, the normal bands are fourfold degenerate. As was previously shown in the context of $j=3/2$ pairing in cold atoms [@ho99], the TRSB state is *unstable* towards the time-reversal symmetric state in this limit, as it leaves two of these bands ungapped. Nonzero spin-orbit coupling partially lifts this degeneracy and allows the TRSB state to open a gap on all Fermi surfaces, reducing its energy. An analysis of the fourth-order term in [Eq. (\[eq:freeE\])]{} predicts that the TRSB $\eta_{xz}+i\eta_{yz}$ state is energetically favored for $|\beta|k_F^2/k_BT_c \gtrsim 9.324$. Details are given in the SM [@suppinfo]. Hence, the presence of the Bogoliubov Fermi surfaces does not necessarily compromise the stability of TRSB states.
*Pairing-induced pseudomagnetic field*.—To gain additional insight into the Bogoliubov Fermi surfaces, it is useful to rewrite the Hamiltonian in a basis for which the normal-state Hamiltonian is diagonal. We denote the eigenvectors of $H_N$ to the twofold degenerate eigenvalues $\epsilon_\pm$ by $|\pm,i{=}1,2\rangle$, which we choose to form a pseudospin basis, i.e., $|\pm,2\rangle = PT\,|\pm,1\rangle$. In this basis, the superconducting state is described by the Hamiltonian $$\begin{aligned}
\bar{H}&=\left(\begin{array}{cccc}
H_{N,+} & \Delta_{++} & 0 & \Delta_{+-} \\
\Delta_{++}^{\dagger} & -H_{N,+} & -\Delta_{+-}^* & 0 \\
0 & -\Delta_{+-}^T & H_{N,-} & \Delta_{--} \\
\Delta_{+-}^{\dagger} & 0 & \Delta_{--}^{\dagger} & -H_{N,-}
\end{array}\right) .
\label{eq:bandHam}\end{aligned}$$ Here, $H_{N,\pm}=\epsilon_{\pm}\hat{\sigma}_0$ and $\Delta_{\pm\pm}$ are antisymmetric matrices with $\Delta_{\pm\pm}=\psi_{\pm}({\bf k})\,i\hat{\sigma}_y$, where $\hat{\sigma}_\mu$ are the Pauli matrices in pseudospin space. $\Delta_{+-}$ is the interband pairing potential, the explicit form of which depends on the choice of bases in the two-dimensional eigenspaces of $\epsilon_\pm$ and is hence not illuminating. The intraband gap functions $\psi_{\pm}({\bf k})$ are obtained by transforming [Eq. (\[gap\])]{} into the pseudospin basis and are given by $$\psi_\pm({\bf k})=\Delta_1\psi({\bf k})\pm2\Delta_0\frac{c_{xz}({\bf
k})+ic_{yz}({\bf k})}{\epsilon_+({\bf k}) -\epsilon_-({\bf k})} .$$ In the absence of $\Delta_{+-}$, $\bar{H}$ would describe two decoupled pseudospin-$1/2$ singlet superconductors with, at most, line or point nodes. Hence, the interband pairing is responsible for the appearance of the extended Fermi surfaces. This can be shown by treating the off-diagonal interband blocks of [Eq. (\[eq:bandHam\])]{} as a perturbation to the intraband Hamiltonians: focusing on the $+$ states (analogous results can be found for the $-$ states if they have a normal-state Fermi surface), the second-order corrections due to the interband pairing appear only in the normal-state components, which become $$H^\prime_{N,+}=[\epsilon_++\gamma({\bf k})]\, \hat{\sigma}_0
+ {\bf h}({\bf k})\cdot \hat{\mbox{\boldmath$\sigma$}},$$ where $$\begin{aligned}
\gamma({\bf k}) & = \frac{2|\Delta_0|^2}{(\epsilon_+-\epsilon_-)^3}\,
\left[(\epsilon_+-\epsilon_-)^2-2c_{yz}^2-2c_{xz}^2\right] ,\\
|{\bf h}({\bf k})| & =\frac{4|\Delta_0|^2}{(\epsilon_+-\epsilon_-)^2}\,
\sqrt{c_{3z^2-r^2}^2+c_{x^2-y^2}^2+c_{xy}^2} .
\label{field}\end{aligned}$$ The direction of $\mathbf{h}(\mathbf{k})$ is basis-dependent and is thus not physically meaningful. The correction $\gamma({\bf k})$ is always present and results in a small modification of the normal-state dispersion $\epsilon_+$, whereas the second term ${\bf h}({\bf k})\cdot\hat{\mbox{\boldmath$\sigma$}}$ only appears for TRSB gaps. This reveals that in the TRSB state the *interband pairing manifests itself as a pseudomagnetic field in the normal-state Hamiltonian*.
The origin of the extended Fermi surfaces becomes clear from the excitation spectrum of the low-energy pairing Hamiltonian, $$E_{{\bf k},\pm,\nu}= \nu|{\bf h}({\bf k})|\pm\sqrt{[\epsilon_+({\bf k})
+ \gamma({\bf k})]^2+|\psi_+({\bf k})|^2} ,
\label{Epseudo}$$ where $\nu=\pm1$. The pseudomagnetic field $|\mathbf{h}(\mathbf{k})|$ evidently splits the dispersion. The square root in [Eq. (\[Epseudo\])]{} goes to zero at the intersection of the nodes of the intraband gap $\psi_+({\bf k})$ with the surface $\epsilon_++\gamma({\bf k})=0$, and the pseudomagnetic field gives rise to the Bogoliubov Fermi surfaces by shifting the nodes to finite energies $\pm|\mathbf{h}(\mathbf{k})|$. The generation of ${\bf h}({\bf k})$ by the superconducting state itself exhibits the *intrinsic* nature of the Bogoliubov Fermi surfaces. This distinguishes our results from the “breached-pairing” state of population-imbalanced cold atoms or exchange-split superconductors, where the required breaking of time-reversal symmetry is *extrinsic* to the pairing [@liu03; @gub05].
The low-energy effective model also allows us to estimate the dimensions of the Fermi surfaces. Perpendicular to the normal-state Fermi surface, the Bogoliubov Fermi surfaces have a width $\delta k_\perp/k_F \sim \Delta^2_0/[\mu(\epsilon_+-\epsilon_-)]$, where $k_F$ is the normal-state Fermi momentum. Their width in the direction parallel to the normal-state Fermi surface is $\delta
k_\parallel/k_F \sim \Delta_0/(\epsilon_+-\epsilon_-)$. Since we typically expect $\Delta_0\ll \mu$, we find that $\delta k_\perp \ll \delta k_\parallel$, which implies oblate spheroidal Fermi surfaces near the original point nodes and flattened toroidal Fermi surfaces near the original line nodes, as seen in Fig. \[figure1\].
Pseudomagnetic fields appear in any TRSB phase of our model with interband (spin-quintet) pairs. Our analysis generalizes to other systems with multiband pairing: we expect non-vanishing contributions to the pseudomagnetic field from all the interband potentials, implying that Bogoliubov Fermi surfaces are a generic feature of such systems. Equation (\[field\]) shows that these Fermi surfaces will be largest in strong-coupling materials in which the different bands lie close to each other. These conditions are likely satisfied in heavy-fermion superconductors such as UPt$_3$, Th-doped UBe$_{13}$, PrOs$_4$Sb$_{12}$, and URu$_2$Si$_2$, which makes them ideal systems in which to search for Bogoliubov Fermi surfaces. Indeed, as mentioned in the Introduction, Th-doped UBe$_{13}$ shows a large (and so far unexplained) residual density of states [@zie04] which is consistent with our theory.
*Conclusions*.—We have established that broken-time-reversal even-parity superconductors generically support two-dimensional Fermi surfaces. The states at these Fermi surfaces are charge-neutral Bogoliubov quasiparticles. The Fermi surfaces are protected by a topological $\mathbb{Z}_2$ invariant, which can be written in terms of a Pfaffian, and thus cannot be removed by any perturbation that is $CP$ invariant. They are also energetically stable for plausible parameters, i.e., the corresponding state has lower free energy than an associated time-reversal-symmetric state. The $\mathbb{Z}_2$-protected Fermi surfaces appear in multiband systems, where interband pairing produces an effective pseudomagnetic field, which “inflates” the expected point or line nodes. Since all candidate materials for TRSB even-parity superconductivity have multiple relevant bands, we expect that these $\mathbb{Z}_2$-protected Fermi surfaces are ubiquitous. The existence of Bogoliubov Fermi surfaces in the superconducting state should lead to characteristic experimental consequences: a nonzero density of states at the Fermi energy, visible for example in tunneling and photoemission experiments, would coexist with ideal conductivity and flux expulsion. The low-temperature thermodynamic response would show a linear temperature dependence of the specific heat, and heat conduction should also be unconventional [@gub05]. Such anomalies may have already been observed in heavy-fermion superconductors [@zie04; @joy02]. Our work raises many interesting new questions, e.g., is a superconductor with Bogoliubov Fermi surfaces a Fermi liquid when residual interactions are taken into account? The search for and study of such systems thus constitutes a very promising task for future research.
The authors thank A.P. Schnyder and M. Punk for useful discussions. D.F.A. was supported by the National Science Foundation grant DMR-1335215. C.T. acknowledges financial support by the Deutsche Forschungsgemeinschaft, in part through Research Training Group GRK 1621 and Collaborative Research Center SFB 1143.
[99]{}
G. E. Volovik and L. P. Gor’kov, Superconducting classes in heavy-fermion systems, Zh. Eksp. Teor. Fiz. [**88**]{}, 1412 (1985).
M. Sigrist and K. Ueda, Phenomenological theory of unconventional superconductivity, Rev. Mod. Phys. [**63**]{}, 239 (1991).
Y. Gao, W.-P. Su, and J.-X. Zhu, Interorbital pairing and its physical consequences for iron pnictide superconductors, Phys. Rev. B **81**, 104504 (2010).
A. Nicholson, W. Ge, J. Riera, M. Daghofer, A. Moreo, and E. Dagotto, Pairing symmetries of a hole-doped extended two-orbital model for the pnictides, Phys. Rev. B **85**, 024532 (2012).
R. Nourafkan, G. Kotliar, A. S. Tremblay, Correlation-Enhanced Odd-Parity Interorbital Singlet Pairing in the Iron-Pnictide Superconductor LiFeAs, Phys. Rev. Lett. **117**, 137001 (2016).
T. Ong, P. Coleman, and J. Schmalian, Concealed $d$-wave pairs in the $s^{\pm}$ condensate of iron-based superconductors, Proc. Natl. Acad. Sci. U.S.A. [**113**]{}, 5486 (2016).
L. Fu, Odd-parity topological superconductor with nematic order: Application to Cu$_x$Bi$_2$Se$_3$, Phys. Rev. B [**90**]{}, 100509(R) (2014); S. Yonezawa, K. Tajiri, S. Nakata, Y. Nagai, Z. Wang, K. Segawa, Y. Ando, and Y. Maeno, Thermodynamic evidence for nematic superconductivity in Cu$_x$Bi$_2$Se$_3$, Nature Phys. [**13**]{}, 123 (2017).
P. M. R. Brydon, L. M. Wang, M. Weinert, and D. F. Agterberg, Pairing of $j=3/2$ Fermions in Half-Heusler Superconductors, Phys. Rev. Lett. [**116**]{}, 177001 (2016); H. Kim, K. Wang, Y. Nakajima, R. Hu, S. Ziemak, P. Syers, L. Wang, H. Hodovanets, J. D. Denlinger, P. M. R. Brydon, D. F. Agterberg, M. A. Tanatar, R. Prozorov, amd J. Paglione, Beyond Triplet: Unconventional Superconductivity in a Spin-3/2 Topological Semimetal, arXiv:1603.03375 (2016).
T. Nomoto and H. Ikeda, Exotic Multigap Structure in UPt$_3$ Unveiled by a First-Principles Analysis, Phys. Rev. Lett. **117**, 217002 (2016); T. Nomoto, K. Hattori, and H. Ikeda, Classification of “multipole” superconductivity in multiorbital systems and its implications, Phys. Rev. B **94**, 174513 (2016).
G. M. Luke, A. Keren, L. P. Le, W. D. Wu, Y. J. Uemura, D. A. Bonn, L. Taillefer, and J. D. Garrett, Muon spin relaxation in UPt$_3$, Phys. Rev. Lett. [**71**]{}, 1466 (1993).
R. Joynt and L. Taillefer, The superconducting phases of UPt$_3$, Rev. Mod. Phys. [**74**]{}, 235 (2002).
E. R. Schemm, W. J. Gannon, C. M. Wishne, W. P. Halperin, and A. Kapitulnik, Observation of broken time-reversal symmetry in the heavy-fermion superconductor UPt$_3$, Science [**345**]{}, 190 (2014).
R. H. Heffner, J. L. Smith, J. O. Willis, P. Birrer, C. Baines, F. N. Gygax, B. Hitti, E. Lippelt, H. R. Ott, A. Schenck, E. A. Knetsch, J. A. Mydosh, and D. E. MacLaughlin, New phase diagram for (U,Th)Be$_{13}$: A muon-spin-resonance and $H_{c1}$ study, Phys. Rev. Lett. [**65**]{}, 2816 (1990).
R. J. Zieve, R. Duke, and J. L. Smith, Pressure and linear heat capacity in the superconducting state of thoriated UBe$_{13}$, Phys. Rev. B [**69**]{}, 144503 (2004).
Y. Aoki, A. Tsuchiya, T. Kanayama, S. R. Saha, H. Sugawara, H. Sato, W. Higemoto, A. Koda, K. Ohishi, K. Nishiyama, and R. Kadono, Time-Reversal Symmetry-Breaking Superconductivity in Heavy-Fermion PrOs$_4$Sb$_{12}$ Detected by Muon-Spin Relaxation, Phys. Rev. Lett [**91**]{}, 067003 (2003).
G. M. Luke, Y. Fudamoto, K. M. Kojima, M. I. Larkin, J. Merrin, B. Nachumi, Y. J. Uemura, Y. Maeno, Z. Q. Mao, Y. Mori, H. Nakamura, and M. Sigrist, Time-Reversal Symmetry Breaking Superconductivity in Sr$_2$RuO$_4$, Nature [**394**]{}, 558 (1998).
J. Xia, Y. Maeno, P. T. Beyersdorf, M. M. Fejer, and A. Kapitulnik, High Resolution Polar Kerr Effect Measurements of Sr$_2$RuO$_4$: Evidence for Broken Time-Reversal Symmetry in the Superconducting State, Phys. Rev. Lett. [**97**]{}, 167002 (2006).
Y. Kasahara, T. Iwasawa, H. Shishido, T. Shibauchi, K. Behnia, Y. Haga, T. D. Matsuda, Y. Onuki, M. Sigrist, and Y. Matsuda, Exotic Superconducting Properties in the Electron-Hole-Compensated Heavy-Fermion “Semimetal” URu$_2$Si$_2$, Phys. Rev. Lett. [**99**]{}, 116402 (2007).
E. R. Schemm, R. E. Baumbach, P. H. Tobash, F. Ronning, E. D. Bauer, and A. Kapitulnik, Evidence for broken time-reversal symmetry in the superconducting phase of URu$_2$Si$_2$, Phys. Rev. B [**91**]{}, 140506(R) (2015).
P. K. Biswas, H. Luetkens, T. Neupert, T. Stürzer, C. Baines, G. Pascua, A. P. Schnyder, M. H. Fischer, J. Goryo, M. R. Lees, H. Maeter, F. Brückner, H.-H. Klauss, M. Nicklas, P. J. Baker, A. D. Hillier, M. Sigrist, A. Amato, and D. Johrendt, Evidence for superconductivity with broken time-reversal symmetry in locally noncentrosymmetric SrPtAs, Phys. Rev. B [**87**]{}, 180503(R) (2013).
X. Gong, M. Kargarian, A. Stern, D. Yue, H. Zhou, X. Jin, V. M. Galitski, V. M. Yakovenko, and J. Xia, Time-Reversal-Symmetry-Breaking Superconductivity in Epitaxial Bismuth/Nickel Bilayers, arXiv:1609.08538 (2016).
R. Nandkishore, L. S. Levitov, and A. V. Chubukov, Chiral superconductivity from repulsive interactions in doped graphene, Nature Phys. [**8**]{}, 158 (2012).
A. Black-Schaffer and C. Honerkamp, Chiral *d*-wave superconductivity in doped graphene, J. Phys. Condens. Matter [**26**]{}, 423201 (2014).
D. Sa, M. Sardar, and G. Baskaran, Superconductivity in Na$_x$CoO$_2\cdot y$H$_2$O: Protection of a $d_1+id_2$ state by spin-charge separation, Phys. Rev. B [**70**]{}, 104505 (2004).
M. L. Kiesel, C. Platt, W. Hanke, and R. Thomale, Model Evidence of an Anisotropic Chiral $d+id$-Wave Pairing State for the Water-Intercalated Na$_x$CoO$_2\cdot y$H$_2$O Superconductor, Phys. Rev. Lett. [**111**]{}, 097001 (2013).
R. Ganesh, G. Baskaran, J. van den Brink, and D. V. Efremov, Theoretical Prediction of a Time-Reversal Broken Chiral Superconducting Phase Driven by Electronic Correlations in a Single TiSe$_2$ Layer, Phys. Rev. Lett. [**113**]{}, 177001 (2014).
Y.-T. Hsu, A. Vaezi, M. H. Fischer, and E.-A. Kim, Topological superconductivity in monolayer transition metal dichalcogenides, arXiv:1606.00857 (2016).
J. M. Luttinger and W. Kohn, Motion of Electrons and Holes in Perturbed Periodic Fields, Phys. Rev. [**97**]{}, 869 (1955).
See Supplemental Material at http://link.aps.org/ supplemental/10.1103/XXX for details on the mapping between orbital and spin $j=3/2$ representations, the properties of the Pfaffian, and the stability analysis. This includes Refs. [@S:AKM98; @mer74].
D. Alekseevsky, A. Kriegl, P. W. Michor, and M. Losik, Choosing roots of polynomials smoothly, Israel J. Math. **105**, 203 (1998).
N. D. Mermin, *d*-wave pairing near the transition temperature, Phys. Rev. A [**9**]{}, 868 (1974).
S. Kobayashi, K. Shiozaki, Y. Tanaka, and M. Sato, Topological Blount’s theorem of odd-parity superconductors, Phys. Rev. B [**90**]{}, 024516 (2014).
Y. X. Zhao, A. P. Schnyder, and Z. D. Wang, Unified Theory of $PT$ and $CP$ Invariant Topological Metals and Nodal Superconductors, Phys. Rev. Lett. [**116**]{}, 156402 (2016).
T. L. Ho and S.-K. Yip, Pairing of Fermions with Arbitrary Spin, Phys. Rev. Lett. [**82**]{}, 247 (1999).
V. P. Mineev and K. V. Samokhin, *Introduction to Unconventional Superconductivity* (Gordon and Breach Science Publishers, New York, 1999).
E. Gubankova, E. G. Mishchenko, and F. Wilczek, Breached Superfluidity via *p*-Wave Coupling, Phys. Rev. Lett. [**94**]{}, 110402 (2005); Gapless surfaces in anisotropic superfluids, Phys. Rev. B [**74**]{}, 184516 (2006).
W. V. Liu and F. Wilczek, Interior Gap Superfluidity, Phys. Rev. Lett. [**90**]{}, 047002 (2003).
**Supplemental Material for\
Bogoliubov Fermi Surfaces in Superconductors with Broken Time-Reversal Symmetry**\
D. F. Agterberg, P. M. R. Brydon, and C. Timm
Mapping to a general two-orbital model
======================================
In this section, we show that the most general model for a normal-state system with a two-valued orbital degree of freedom where the orbitals have the same parity and are chosen to be real can be unitarily transformed into the $j=3/2$ Hamiltonian given in the main text. We expand the single-particle Hamiltonian in terms of Kronecker products in orbital-spin space, $$H_N({\bf k}) = \sum_{\mu,\nu=0,x,y,z} c_{\mu,\nu}({\bf k})\,
\hat{s}_{\mu} \otimes \hat{\sigma}_{\nu} ,
\label{eq:suppH}$$ where $\hat{s}_{\mu}$ ($\hat{\sigma}_{\nu}$) are the Pauli matrices in orbital (spin) space. Hermiticity implies $c_{\mu,\nu}({\bf k})=c_{\mu,\nu}^*({\bf k})$. Inversion maps $H_N(\mathbf{k})$ onto $U_P H_N(-\mathbf{k}) U_P^\dagger$, where $U_P = 1_4$ ($U_P = -1_4$) if the orbitals are both even (odd). The extra sign for the odd case obviously drops out and can be disregarded. Hence, for the Hamiltonian to be symmetric under inversion, we simply require $c_{\mu,\nu}({\bf k})=c_{\mu,\nu}(-{\bf k})$. Time reversal maps $H_N(\mathbf{k})$ onto $U_T H_N^*(-\mathbf{k}) U_T^\dagger$. $c_{\mu,\nu}({\bf k})$ is thus mapped onto $c_{\mu,\nu}^*(-{\bf k})=c_{\mu,\nu}({\bf k})$. Under the assumption of real-valued orbitals, $\hat{s}_y$ and the spin components $\hat{\sigma}_x$, $\hat{\sigma}_y$, $\hat{\sigma}_z$ are odd under time reversal, whereas the remaining matrices are even. This is achieved by $U_T = \hat{s}_0 \otimes i\hat{\sigma}_y$.
In order for the Hamiltonian $H_N(\mathbf{k})$ to be invariant under time reversal, only those Kronecker products can appear in [Eq. (\[eq:suppH\])]{} that are even, specifically the six products with $\{\mu,\nu\} = \{0,0\}, \{x,0\}, \{z,0\}, \{y,x\}, \{y,y\}, \{y,z\}$. The corresponding $c_{\mu,\nu}({\bf k})$ are real and even functions of ${\bf k}$. Note that the five nontrivial Kronecker products are mutually anticommuting and thus form a representation of the five Dirac matrices.
We now show that the generalized Luttinger-Kohn model [@S:lut55] adopted in the main text is equivalent to [Eq. (\[eq:suppH\])]{}. It is expressed in terms of the angular-momentum $j=3/2$ matrices $$\begin{aligned}
J_x &= \frac{1}{2}\left(
\begin{array}{cccc}
0 & \sqrt{3} & 0 & 0 \\
\sqrt{3} & 0 & 2 & 0 \\
0 & 2 & 0 & \sqrt{3} \\
0 & 0 & \sqrt{3} & 0
\end{array}
\right) ,\\
J_y &= \frac{i}{2}\left(
\begin{array}{cccc}
0 & -\sqrt{3} & 0 & 0 \\
\sqrt{3} & 0 & -2 & 0 \\
0 & 2 & 0 & -\sqrt{3} \\
0 & 0 & \sqrt{3} & 0
\end{array}
\right) ,\\
J_z &= \frac{1}{2}\left(
\begin{array}{cccc}
3 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & -1 & 0 \\
0 & 0 & 0 & -3
\end{array}
\right) .\end{aligned}$$ The generalized Luttinger-Kohn model involves five sets of products of the $j=3/2$ matrices, specifically $(2 J_z^2 - J_x^2 - J_y^2/6$, $(J_x^2 - J_y^2)/(2\sqrt{3})$, $(J_x J_z + J_z J_x)/(2\sqrt{3})$, $(J_x J_y + J_y J_x)/(2\sqrt{3})$, and $(J_y J_z + J_z J_y)/(2\sqrt{3})$. These matrices also form a set of five Dirac matrices and there is a unitary transformation between them and the five allowed nontrivial Kronecker products discussed above. Specifically, we find $$\begin{aligned}
U^\dagger\, \frac{2 J_z^2 - J_x^2 - J_y^2}{6}\, U
&= \hat{s}_z\otimes \hat{\sigma}_0 , \\
U^\dagger\, \frac{J_x^2 - J_y^2}{2\sqrt{3}}\, U
&= \hat{s}_x\otimes \hat{\sigma}_0 , \\
U^\dagger\, \frac{J_x J_z + J_z J_x}{2\sqrt{3}}\, U
&= \hat{s}_y\otimes \hat{\sigma}_y , \\
U^\dagger\, \frac{J_x J_y + J_y J_x}{2\sqrt{3}}\, U
&= \hat{s}_y\otimes \hat{\sigma}_z , \\
U^\dagger\, \frac{J_y J_z + J_z J_y}{2\sqrt{3}}\, U
&= \hat{s}_y\otimes \hat{\sigma}_x ,\end{aligned}$$ with the unitary matrix $$U=\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 \\
0 & 0 & 1 & 0 \\
0 & 1 & 0 & 0
\end{array}
\right) .$$ Note that one can pairwise swap the orbital-spin matrices $\hat{s}_\mu\otimes\hat{\sigma}_\nu$ by an appropriate unitary transformation so that this mapping is entirely general. We hence see that the generalized Luttinger-Kohn model is equivalent to the most general single-particle Hamiltonian of a time- and inversion-symmetric material with two identical-parity orbitals and spin. The generalized Luttinger-Kohn representation is convenient, however, as it allows us to exploit the symmetry properties of the angular-momentum matrices $J_i$ under rotations to gain insight into the underlying physics.
We can obtain six on-site pairing states in the orbital-spin representation by multiplying the six allowed Kronecker products by the unitary part of the time-reversal operator, $U_T=\hat{s}_0\otimes i\hat{\sigma}_y$. We hence obtain one orbitally trivial spin-singlet gap proportional to $\hat{s}_0\otimes \hat{\sigma}_y$ and five “anomalous” gaps proportional to $\hat{s}_x\otimes \hat{\sigma}_y$, $\hat{s}_z\otimes \hat{\sigma}_y$, $\hat{s}_y\otimes\hat{\sigma}_z$, $\hat{s}_y\otimes \hat{\sigma}_0$, and $\hat{s}_y\otimes
\hat{\sigma}_x$, which are either orbital-triplet spin-singlet or orbital-singlet spin-triplet pairing states. It is straightforward to map these states onto the gap matrices in the equivalent spin-$3/2$ formulation: the orbitally trivial spin-singlet state maps onto the singlet gap matrix $\eta_s$, while the anomalous gaps map onto the quintet gap matrices. Hence, we can describe any even-parity pairing state, without loss of generality, in terms of a linear combination of the six spin-$3/2$ gap functions with even-parity coefficients $\psi_i({\bf k})$.
Existence and properties of the Pfaffian
========================================
In this section, we provide additional details on constructing the topological invariant protecting the Fermi surfaces. Our starting point is the Hamiltonian in the superconducting state, $$H(\mathbf{k}) = \left(
\begin{array}{cc}
H_N({\bf k}) & \Delta({\bf k}) \\
\Delta^{\dagger}({\bf k}) & -H_N^T({\bf k})
\end{array}
\right) .
\label{eq:Ham3}$$ Recall that $H_N(\mathbf{k})$ is even in $\mathbf{k}$. We employ the spin-$3/2$ basis used in the main text. We first show that a $\mathbf{k}$-independent unitary matrix $\Omega$ exists such that $\tilde{H}^T(\mathbf{k})=-\tilde{H}(\mathbf{k})$ for $\tilde{H}({\bf k})=\Omega H({\bf k}) \Omega^{\dagger}$, i.e., the Hamiltonian can be transformed into antisymmetric form.
The proof proceeds as follows: 1. The Hamiltonian $H(\mathbf{k})$ satisfies parity and charge-conjugation symmetries and thus also their product, i.e., $$U_{CP}\, H^T(\mathbf{k})\, U_{CP}^\dagger = -H(\mathbf{k}) ,
\label{1.sup.1}$$ where $U_{CP} \equiv U_C U_P^* = (\hat{\tau}_x \otimes 1_4)
(\hat{\tau}_0 \otimes 1_4) = \hat{\tau}_x \otimes 1_4$. The $\hat{\tau}_i$ denote Pauli matrices in particle-hole (Nambu) space. We find that the $CP$ symmetry squares to $+1$ since $(CP)^2 = (U_{CP} \mathcal{K})^2
= U_{CP} U_{CP}^* = \hat{\tau}_0 \otimes 1_4 = +1_8$. In the presence of such a symmetry, two-dimensional Fermi surfaces are characterized by a $\mathbb{Z}_2$ invariant [@S:kob14; @S:zha16].
2. For *any* $CP$ symmetry that squares to unity we have $U_{CP}^* = U^{-1}_{CP} = U^\dagger_{CP}$ and thus $U_{CP} = U^T_{CP}$, hence $U_{CP}$ is symmetric. Any (complex) symmetric matrix can be diagonalized by a unitary *congruence*, i.e., there exist a unitary matrix $Q$ and a diagonal matrix $\Lambda$ such that (note the transpose) $$U_{CP} = Q \Lambda Q^T .
\label{1.UCP.5}$$ Inserting this equation into [Eq. (\[1.sup.1\])]{} yields $$Q \Lambda Q^T\, H^T(\mathbf{k})\, Q^* \Lambda^\dagger Q^\dagger
= -H(\mathbf{k}) .
\label{1.CPQ.2}$$ Since $\Lambda = Q^\dagger U_{CP} Q^*$ is unitary and diagonal, it can be written as $\Lambda = \mathrm{diag}(\lambda_1,\lambda_2,\ldots)$ with $|\lambda_i|=1$. Now let $$\sqrt{\Lambda} \equiv
\mathrm{diag}\left(\sqrt{\lambda_1},\sqrt{\lambda_2},\ldots\right) ,
\label{1.Lambda.4}$$ where for each $i$, $\sqrt{\lambda_i}$ is the complex root with arbitrary but fixed sign. Also let $$\sqrt{\Lambda}{}^\dagger \equiv \mathrm{diag}
\left(\sqrt{\lambda_1}^{\,*},\sqrt{\lambda_2}^{\,*},\ldots\right) ,
\label{1.Lambda.5}$$ with the same choice of signs as in [Eq. (\[1.Lambda.4\])]{}. It is trivial to show that this is a root of $\Lambda^\dagger$. Furthermore, $\sqrt{\Lambda}$ and $\sqrt{\Lambda}{}^\dagger$ are obviously diagonal, and thus symmetric, and also unitary. We can thus rewrite [Eq. (\[1.CPQ.2\])]{} as $Q \sqrt{\Lambda} \sqrt{\Lambda} Q^T H^T(\mathbf{k})
Q^* \sqrt{\Lambda}{}^\dagger \sqrt{\Lambda}{}^\dagger Q^\dagger
= -H(\mathbf{k})$ and find that $$(\sqrt{\Lambda}{}^\dagger Q^\dagger H(\mathbf{k}) Q \sqrt{\Lambda})^T
= - \sqrt{\Lambda}{}^\dagger Q^\dagger H(\mathbf{k}) Q \sqrt{\Lambda} .
\label{1.CPQ.4}$$ With the unitary matrix $\Omega \equiv \sqrt{\Lambda}{}^\dagger Q^\dagger$, [Eq. (\[1.CPQ.4\])]{} can be written as $(\Omega H(\mathbf{k}) \Omega^\dagger)^T
= - \Omega H(\mathbf{k}) \Omega^\dagger$. Hence, $\tilde{H}(\mathbf{k}) \equiv \Omega H(\mathbf{k})
\Omega^\dagger$ is indeed antisymmetric.
For the Hamiltonian $H(\mathbf{k})$ given in [Eq. (\[eq:Ham3\])]{} above together with Eq. ([1]{}) in the main text, a possible choice is $$\Omega = \frac{1}{\sqrt{2}} \left(\begin{array}{cc}
1 & 1 \\ i & -i
\end{array}\right) \otimes 1_4 .$$ This yields $$\tilde{H}(\mathbf{k}) = \frac{1}{2} \left(\begin{array}{cc}
H_N - H_N^T + \Delta + \Delta^\dagger &
-i\,(H_N+H_N^T) + i\,(\Delta-\Delta^\dagger) \\[1ex]
i\,(H_N+H_N^T) + i\,(\Delta-\Delta^\dagger) &
H_N - H_N^T - \Delta - \Delta^\dagger \end{array}\right) ,$$ where we have suppressed the argument $\mathbf{k}$. The terms involving the normal-state Hamiltonian $H_N$ are obviously antisymmetric. For this, it is crucial that $H_N$ is even in $\mathbf{k}$. From [Eq. (\[1.sup.1\])]{}, one easily finds $\Delta^T = -\Delta$, which implies that also the superconducting terms in $\tilde{H}(\mathbf{k})$ are antisymmetric. Hence, we do find $\tilde{H}^T(\mathbf{k}) = -\tilde{H}(\mathbf{k})$.
Since $\tilde{H}(\mathbf{k})$ is antisymmetric, its Pfaffian $P({\bf k}) \equiv \text{Pf}\, \tilde{H}({\bf k})$ exists. The Pfaffian is real for any spinful system since (a) the dimension of the Hamiltonian is a multiple of four ($2\times2$ for Nambu and spin space), thus the Pfaffian is a polynomial of even degree of the components of $\tilde{H}(\mathbf{k})$, and (b) $\tilde{H}(\mathbf{k})$ is hermitian and antisymmetric and thus these components are purely imaginary.
Note further that the ambiguity in the signs of the roots in $\sqrt{\Lambda}$ means that $\Omega$ can be multiplied on the left by a diagonal matrix $D$ with arbitrary components $\pm 1$ on the diagonal. With $\Omega \to D\Omega$ we get $\Omega^\dagger \to \Omega^\dagger D$ and $\tilde{H}(\mathbf{k}) \to D
\tilde{H}(\mathbf{k}) D$. This leads to $\text{Pf}\,\tilde{H}(\mathbf{k}) \to
\det D\: \text{Pf}\,\tilde{H}(\mathbf{k})$. Hence, $P(\mathbf{k})$ is only determined up to an overall sign. But this sign is selected by fixing the root $\sqrt{\Lambda}$ once for all $\mathbf{k}$. Thus sign *changes* in $P(\mathbf{k})$ are meaningful.
Since $\det H(\mathbf{k})=\det \tilde{H}(\mathbf{k})=P^2(\mathbf{k})$, the zeros of $P(\mathbf{k})$ give the nodes of the superconducting state. Thus if the Pfaffian changes sign as a function of $\mathbf{k}$, the surface separating regions with $P(\mathbf{k})\gtrless 0$ is a two-dimensional Fermi surface. However, $P(\mathbf{k})$ can in addition have zeros of even multiplicity, which do not separate regions with different sign. These *accidental* zeros can form two-dimensional surfaces or line or point nodes.
For the general $k_z(k_x+ik_y)$ state of Eq. ([4]{}) in the main text, the Pfaffian is $$P({\bf k}) = (\epsilon_+\epsilon_-)^2
+ 4 \Delta_0^2\, (\epsilon_+\epsilon_-+c_{xz}^2+c_{yz}^2)
+ \Delta_1^2\, |\psi|^2\, (\epsilon_+^2+\epsilon_-^2)
+ 8\Delta_0\Delta_1\, c_0\, |\psi|^2
+ \Delta_1^4\, |\psi|^4 ,
\label{Pf}$$ where $\epsilon_\pm(\mathbf{k})$ are the normal-state eigenenergies. We have chosen the signs in the root $\sqrt{\Lambda}$ in such a way that $P(\mathbf{k})>0$ in the limit of large $\mathbf{k}$. Equation (\[Pf\]) is correct for arbitrary real and even coefficients $c_i(\mathbf{k})$. Note that when $\Delta_0=0$, the equation $P(\mathbf{k})=0$ implies the usual relationship for zero-energy states in a spin-singlet superconductor, i.e., $(\epsilon_+^2+\Delta_1^2\,|\psi|^2)(\epsilon_-^2+\Delta_1^2\,|\psi|^2)=0$, implying nodes when $\epsilon_{\pm}=0$ and $\psi=0$; since $\psi(\mathbf{k})\propto c_{xz}(\mathbf{k})+ic_{yz}(\mathbf{k})$ to ensure that the full gap function $\Delta(\mathbf{k})$ transforms under rotations like $Y_{2,1}(\hat{\mathbf{k}})$, we have line nodes in the $k_z=0$ plane and point nodes along the line $k_x=k_y=0$. With both $\Delta_0\ne 0$ and $\Delta_1\ne 0$, we see that along the nodal directions, for which $\psi \propto c_{xz}+ic_{yz}=0$, we have zero-energy excitations for $\epsilon_+\epsilon_- (4\Delta_0^2+\epsilon_+\epsilon_-)=0$. The two solutions for $\epsilon_+\epsilon_-$ that follow from the latter equation give the two points of the Bogoliubov Fermi surface along these nodal directions. Away from the nodal directions, the other terms in [Eq. (\[Pf\])]{} no longer vanish. However, they are continuous functions of the direction $\hat{\mathbf{k}}$, implying that the Fermi surface still exists at least for a finite range away from the nodal directions. In the main text, we consider the special case of a pure quintet gap, $\Delta_1=0$.
Finally, we show that the Pfaffian can be chosen non-negative if the system is also time-reversal symmetric and the combined inversion and time-reversal symmetry squares to $-1$, i.e., if there exists a unitary matrix $U_{PT}$ so that $U_{PT} H^T(\mathbf{k}) U_{PT}^\dagger = H(\mathbf{k})$ and $U_{PT} U^*_{PT} = -1$. Hence, if both symmetries are present, the $\mathbb{Z}_2$ invariant exists but is necessarily trivial. This has essentially been shown using the method of Clifford algebra extensions in [@S:kob14]. In the following, we give a more elementary proof.
1. Kramers’ theorem shows that under the last two conditions all eigenvalues of $H(\mathbf{k})$ have even degeneracy. Furthermore, condition (\[1.sup.1\]) above implies that the eigenvalues come in pairs $\{E_i(\mathbf{k}), -E_i(\mathbf{k})\}$. Since the dimension of $H(\mathbf{k})$ is a multiple of four, the spectrum thus consists of quadruplets $\{E_i(\mathbf{k}), E_i(\mathbf{k}), -E_i(\mathbf{k}),
-E_i(\mathbf{k})\}$.
2. $P(\mathbf{k})$ is a polynomial of the components of $\tilde{H}(\mathbf{k})$, which are linear combinations of the components of $H(\mathbf{k})$. Hence, $P(\mathbf{k})$ is a polynomial of the components of $\mathcal{H}(\mathbf{k})$. The coefficients are independent of $\mathbf{k}$.
3. We have $P^2(\mathbf{k}) = \det \tilde{H}(\mathbf{k})
= \det H(\mathbf{k}) = \prod_i E_i^4(\mathbf{k})$ so that $P(\mathbf{k}) = \pi(\mathbf{k}) \prod_i E_i^2(\mathbf{k})$ with $\pi(\mathbf{k}) = \pm 1$.
4. Consider the eigenenergies $E_i(\mathbf{k})$ as functions of a real parameter $h$, which can be any real or imaginary part of a component of the hermitian, finite-dimensional matrix $H(\mathbf{k})$. Theorem 7.6 of Alekseevsky *et al*. [@S:AKM98] shows under the weak additional condition that no two eigenvalues meet of infinite order for any real $h$ unless they are equal for all $h$ that the $E_i(\mathbf{k})$ can be chosen as smooth functions of $h$. Then $\prod_i E_i^2(\mathbf{k})$ is a smooth function of $h$.
5. We have already shown that $P(\mathbf{k})$ is a polynomial in $h$, that $\prod_i E_i^2(\mathbf{k}) = \pi(\mathbf{k})\, P(\mathbf{k})$ is a smooth function of $h$, and that $\pi(\mathbf{k})=\pm 1$. Then $\pi(\mathbf{k})$ is constant.
6. Noting the ambiguity in choosing $\Omega$ and thus the overall sign of the Pfaffian, we can choose $\pi(\mathbf{k})=+1$ at some $\mathbf{k}$ and consequently at all $\mathbf{k}$. Thus we indeed obtain $P(\mathbf{k}) = \prod_i E_i^2(\mathbf{k}) \ge 0$.
Stability of states with broken time-reversal symmetry
======================================================
In this section, we provide additional details on the free-energy expansion. We show that states with broken time-reversal symmetry featuring Bogoliubov Fermi surfaces can be energetically stable. The relative stability of spin-singlet broken-time-reversal states over time-reversal-invariant states with the same transition temperature has been attributed to the gapping of nodes that appear in the time-reversal-symmetric state by breaking this symmetry [@S:sig91]. A common example is the chiral *d*-wave state with line nodes, for which the usual spin-singlet gap function takes the form $\psi({\bf k})=\psi_0\, k_z(\nu_x k_x+\nu_y k_y)$. In the broken-time-reversal state, we have $(\nu_x,\nu_y)=(1,i)/\sqrt{2}$, which leads to line nodes for $k_z=0$ and point nodes at $k_x=k_y=0$. In the nodal time-reversal-invariant state, we instead have $(\nu_x,\nu_y)=(1,0)$, which also leads to line nodes for $k_z=0$ and additional line nodes for $k_x=0$. In this case, the broken-time-reversal state is believed to be stable because it gaps the $k_x=0$ line node so that only two point nodes remain, gaining condensation energy. However, we have found that these point nodes become inflated $\mathbb{Z}_2$-protected Fermi surfaces, and it is reasonable to ask if the broken-time-reversal state is still stable.
We consider the spherically symmetric normal-state Hamiltonian $H_N(\mathbf{k})=\alpha k^2+\beta\,({\bf k}\cdot {\bf J})^2-\mu$ and an on-site gap function that is a linear combination of $\eta_{xz}$ and $\eta_{yz}$. For this gap function, it is known that there are two essentially different possible ground states [@S:sig91]: a time-reversal-invariant state $\Delta_r=(\bar{\Delta}/2)\,\eta_{xz}$ with $k_zk_x$ symmetry and a broken-time-reversal state $\Delta_a=(\bar{\Delta}/2\sqrt{2})\,(\eta_{xz}+i\eta_{yz})$ with $k_z(k_x+ik_y)$ symmetry. The time-reversal-invariant state $(\bar{\Delta}/2)\,\eta_{yz}$ is degenerate with $\Delta_r$. For these two states, the off-diagonal block in the Hamiltonian $H(\mathbf{k})$ reads $$\Delta_{r}=\frac{\bar{\Delta}}{2}\left(
\begin{array}{cccc}
0 & 0 & -1 & 0 \\
0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 \\
0 & -1 & 0 & 0
\end{array}\right)$$ and $$\Delta_{a}=\frac{\bar{\Delta}}{\sqrt{2}}\left(
\begin{array}{cccc}
0 & 0 & -1 & 0 \\
0 & 0 & 0 & 0 \\
\,1\, & 0 & 0 & 0 \\
0 & 0 & 0 & 0
\end{array}\right),$$ respectively. These two matrices have been normalized such that $\text{Tr}\,\Delta^\dagger \Delta=\bar{\Delta}^2$ ($\bar{\Delta}$ is assumed real).
Near $T_c$, the mean-field free energy can be expressed as a power series in $\Delta$. The expression is standard [@S:ho99; @S:min98], $$F =\frac{1}{2V}\, \text{Tr}\,\Delta^{\dagger}\Delta
+ \frac{k_B T}{2} \sum_{\mathbf{k},\omega_n} \sum_{l=1}^{\infty}
\frac{1}{l}\, \text{Tr}\, \big[\Delta\tilde{G}(\mathbf{k},\omega_n)
\Delta^\dagger G(\mathbf{k},\omega_n)\big]^l ,$$ where the constant $V$ is the BCS pairing interaction (which is the same for $\Delta_a$ and $\Delta_r$), $\omega_n$ are the Matsubara frequencies, $T$ is the temperature, and the normal-state Green’s functions $G$ and $\tilde{G}$ satisfy $[i\omega_n-H_N(\mathbf{k})]\, G(\mathbf{k},\omega_n)=1$ and $[i\omega_n+H_N^T(\mathbf{k})]\, \tilde{G}(\mathbf{k},\omega_n)=1$. To find $G$ and $\tilde{G}$, we note that $H_N^T(k_x,k_y,k_z)=H_N(k_x,-k_y,k_z)$ so that $\tilde{G}$ is given once we know $G$. Some algebra then gives $$\begin{aligned}
G(\mathbf{k},\omega_n)
&= G_+(\mathbf{k},\omega_n)
+ \left[\big(\hat{\mathbf{k}}\cdot {\bf J}\big)^2-\frac{5}{4}\right]
G_-(\mathbf{k},\omega_n) , \\
\tilde{G}(\mathbf{k},\omega_n)
& = \tilde{G}_+(\mathbf{k},\omega_n)
+ \left[\big(\hat{\mathbf{k}}\cdot {\bf J}^T\big)^2-\frac{5}{4}\right]
\tilde{G}_-(\mathbf{k},\omega_n) ,\end{aligned}$$ with $$\begin{aligned}
G_{\pm}(\mathbf{k},\omega_n)
&= \frac{1}{2}\left(\frac{1}{i\omega_n-\epsilon_1}
\pm\frac{1}{i\omega_n-\epsilon_2}\right), \\
\tilde{G}_{\pm}(\mathbf{k},\omega_n)
& = \frac{1}{2}\left(\frac{1}{i\omega_n+\epsilon_1}
\pm\frac{1}{i\omega_n+\epsilon_2}\right),\end{aligned}$$ $\epsilon_1=(\alpha+9\beta/4)\,k^2-\mu$, and $\epsilon_2=(\alpha+\beta/4)\,k^2-\mu$. Denoting the free energies for $\Delta_r$ and $\Delta_a$ by $F_r$ and $F_a$, respectively, we find, to fourth order in $\Delta$, $$F_{a}-F_r \cong \frac{\bar{\Delta}^4k_B T}{16} \sum_{{\bf k},\omega_n}
\Big\{\tilde{G}_+^2G_+^2
+ (1-2l_1l_{-1})\left(\tilde{G}_-^2G_+^2+\tilde{G}_+^2G_-^2\right)
+ 4(l_1l_{-1}-1)\,\tilde{G}_-G_-\tilde{G}_+G_+
+ (1-l_1^2l_{-1}^2)\, \tilde{G}_-^2G_-^2\Big \} ,
\label{free}$$ where $l_{\pm 1}=\sqrt{3}\,\cos\theta\sin\theta\: e^{\pm i \phi}$ and $\theta$ and $\phi$ are the spherical angles denoting the direction of ${\bf k}$. Note that the term of order $\Delta^2$ drops out of the free-energy difference under the integral over $\phi$. The analysis of the above expression reveals that either $\Delta_r$ or $\Delta_a$ can be stable, depending on the parameters. In particular, first consider the case of vanishing spin-orbit coupling, $\beta=0$. In this limit, we find $\tilde{G}_-=G_-=0$ so that only the first term in the sum in [Eq. (\[free\])]{} survives, and we have $F_{a}>F_{r}$ so that $\Delta_r$ has lower free energy. This limit has also been considered in the context of $j=3/2$ pairing in cold atoms [@S:ho99; @S:mer74] and the results agree with what we find. Consequently, the broken-time-reversal state is not stable for $\beta=0$.
Now consider the single-band limit, which can be reached, for example, by taking a large $|\beta|$, such that $\alpha+9\beta/4$ and $\alpha+\beta/4$ have opposite sign, and a large chemical potential. If only $\epsilon_1$ crosses the Fermi surface, for sufficiently small $k_BT/\mu\ll 1$ we can safely take the limit $|\epsilon_2|/k_BT\rightarrow \infty$ to find the asymptotic expression $$F_a - F_r
\cong \frac{\bar{\Delta}^4}{1024 k_BT} \sum_{{\bf k}} l_1^2l_{-1}^2
\left( \frac{1-k_BT\sinh(\epsilon_1/k_BT)/\epsilon_1}{\epsilon_1^2(1+\cosh(\epsilon_1/k_BT))}
\right) < 0\,.$$ This is the result expected from single-band weak-coupling theory, and demonstrates that the broken time-reversal state is stable in this limit.
The above analysis implies that for fixed $\alpha$ and $\mu$ there will be a transition from $\Delta_a$ to $\Delta_r$ as a function of spin-orbit coupling $\beta$. It is straightforward to show that the momentum and Matsubara sum in [Eq. (\[free\])]{} only depends on the dimensionless ratios $\beta/\alpha$ and $\mu/k_BT$, apart from prefactors that do not affect its sign. Numerical evaluation indicates that transitions exist for any value of $\mu/k_BT$. We consider explicitly the case of $|\beta|\ll \alpha$, in which the two spherical Fermi surfaces have nearly identical $k_F$. We then find by expanding in $\beta/\alpha$ that this transition takes place at $x \equiv |\beta|\mu/\alpha k_BT = |\beta|k_F^2/k_BT \approx 9.324$. For $x \lesssim 9.324$, we find that the time-reversal-invariant state $\Delta_r$ is the ground state, while for $x \gtrsim 9.324$, $\Delta_r$ is unstable towards $\Delta_a$. This indicates that the broken-time-reversal state is stabilized by modest spin-orbit coupling.
[9]{}
J. M. Luttinger and W. Kohn, Motion of Electrons and Holes in Perturbed Periodic Fields, Phys. Rev. [**97**]{}, 869 (1955).
S. Kobayashi, K. Shiozaki, Y. Tanaka, and M. Sato, Topological Blount’s theorem of odd-parity superconductors, Phys. Rev. B [**90**]{}, 024516 (2014).
Y. X. Zhao, A. P. Schnyder, and Z. D. Wang, Unified Theory of $PT$ and $CP$ Invariant Topological Metals and Nodal Superconductors, Phys. Rev. Lett. [**116**]{}, 156402 (2016).
D. Alekseevsky, A. Kriegl, P. W. Michor, and M. Losik, Choosing roots of polynomials smoothly, Israel J. Math. **105**, 203 (1998).
M. Sigrist and K. Ueda, Phenomenological theory of unconventional superconductivity, Rev. Mod. Phys. [**63**]{}, 239 (1991).
T. L. Ho and S.-K. Yip, Pairing of Fermions with Arbitrary Spin, Phys. Rev. Lett. [**82**]{}, 247 (1999).
N. D. Mermin, *d*-wave pairing near the transition temperature, Phys. Rev. A [**9**]{}, 868 (1974).
V. P. Mineev and K. V. Samokhin, *Introduction to Unconventional Superconductivity* (Gordon and Breach Science Publishers, New York, 1999).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We observe a strong peak in the capacitive photocurrent of a MDMO-PPV / PCBM bulk heterojunction solar cell for excitation below the absorbance threshold energy. Illumination at the peak energy blocks charge capture at other wavelengths, and causes the photovoltage to drop dramatically. These results suggest that the new peak is due to a charge transfer state, which provides a pathway for charge separation and photocurrent generation in the solar cell.'
author:
- 'H.M. Shah'
- 'A.D. Mohite'
- 'T. Bansal'
- 'B.W. Alphenaar'
bibliography:
- 'OPVtext2.bib'
title: |
Photovoltage Bleaching in Bulk Heterojunction Solar Cells through\
Occupation of the Charge Transfer State
---
Organic photovoltaic (OPV) cells are cheaper and more easily fabricated than silicon solar cells and other inorganic photovoltaics. However, they also suffer from relatively low efficiency and time dependent output degradation. Solving these problems requires an improved understanding of the unique charge generation mechanism in OPVs. The highest OPV efficiencies have been observed in bulk heterojunction (BHJ) solar cells. BHJs consist of a mixture of two different organic materials a $\pi$ conjugated polymer (which acts as an electron donor) and a fullerene based polymer (which acts as an electron acceptor)[@yu1; @yu2; @shah]. Under illumination, excitons (or bound electron-hole pairs) are formed[@nunzi]. For photocurrent to be produced, the excitons must dissociate across the donor/acceptor interface and the resulting free carriers diffuse to the contacts. Recent experimental[@goris; @jjbs; @hall; @trivngs] and theoretical[@aryan] work suggests that the excitons dissociate via an intermediate charge transfer state (or exciplex[@morteani]) consisting of an electron hole pair bound across the donor / acceptor interface. There is still considerable debate however as to the nature of this state, and what its role is in the charge dissociation process.
Here, we describe characterization of BHJ solar cells using capacitive photocurrent spectroscopy, (CPS) a novel spectroscopy technique developed in our laboratory[@mohite1] that is particularly sensitive to the exciton dissociation process. CPS has been successfully used to distinguish between excitonic and free carrier states in individual carbon nanotubes[@mohite2; @mohite3] and is able to detect exciton dissociation for carriers not captured by the electrical contacts. Using CPS we are able to identify a photo-absorption state lying at an energy below the main exciton peak in an MDMO-PPV / PCBM solar cell. This peak has low absorbance, but very high dissociation efficiency, and its energy correlates well with previous observations of the charge transfer state[@hall]. Illumination at the peak energy results in a decrease in the photovoltage signal by more than 70%, while no decrease is observed under lower or higher energy illumination. This strongly suggests that this state has a significant role in the charge dissociation process.
A comparison of standard photovoltage and capacitive photocurrent measurement techniques is shown in Figs. 1(a) and (b), respectively. In both cases, the active region consists of a 1:4 mixture of MDMO-PPV:PCBM. In the standard photovoltage measurement, electrical contacts are made to the top and bottom of the sample, using Al/LiF and ITO/PEDOT:PSS, respectively. Light incident on the polymer (through the transparent ITO contact) excites electron-hole pairs into excitonic states. Some fraction of the excitons dissociate into free carriers, and some fraction of these then diffuse to the contacts to be detected as a photovoltage.
In the capacitive photocurrent measurement, the ITO contact is separated from the polymer by an insulating quartz layer. This blocks the flow of dc current to the contact; however, the probe remains sensitive to charge that dissociates and separates (due to the built-in potential) to form a net dipole moment in the polymer layer. For a modulated light source this produces an out-of-phase ac voltage which can be measured with respect to the isolated ITO contact. An advantage of this technique is that it is sensitive to charge that dissociates, but is unable to diffuse to the contacts. By contrast, in the standard photovoltage measurement the out-of-phase signal is dominated by the in-phase (or dc) signal due to the diffusion of carriers to the contacts. Carriers which dissociate, but do not make it to the contact remain undetected.
Figure 2(a) shows the absorbance spectrum of a 100 nm thick layer of 1:4 MDMO-PPV:PCBM, spin coated onto a glass slide. The measurements were performed using a Perkin Elmer Lambda 950 UV-VIS spectrometer under atmospheric conditions. A peak in the absorbance is observed at 2.6 eV (477 nm), and drops-off at lower energy. This peak has been attributed to the ground excitonic state in the MDMO-PPV . A second smaller peak associated with the ground state exciton of the PCBM lies at 3.28 eV (378 nm)[@cook]. Fig. 2(b) shows the open circuit photovoltage of the BHJ measured using the two-contact set-up shown in Fig. 1(a) with the sample in vacuum. The sample is illuminated by a tungsten halogen white light source (Newport Q-T-Halogen, 1 kW ) resolved by a monochromator (Acton Research, SpectraPro 500i) over the wavelength range of 380 - 800 nm. The incident light is chopped at low frequency (13 Hz) and the photovoltage detected with a lock-in amplifier. The photovoltage roughly correlates with the absorbance (showing a peak at 2.6 eV), and is similar to photocurrent and photovoltage measurements of the MDMO-PPV:PCBM system described in the literature[@hoppe; @shah].
Figure 2(c) shows the results of the capacitive photocurrent measurement of the BHJ, using the set-up shown in Fig. 1(b). The capacitor structure is made by spin coating a 100 nm film of 1:4 (MDMO: PPV-PCBM) onto a quartz slide with an ITO contact on the opposite side. An aluminum contact is then deposited onto the polymer film. The polymer and ITO function as the positive and negative electrodes of the capacitor, with the quartz functioning as the capacitor dielectric. The sample is anchored to a copper block within an optical access flow cryostat and kept at vacuum. Illumination is done with an optical parametric amplifier (OPA) excited by a pulsed Ti-Sapphire regenerative amplifier. This produces tunable excitation between 0.4 and 2.4 eV. The pulse width is 120 fs with a repetition rate of 1 kHz, and the output power is kept constant at 14 mW. The capacitive photocurrent is detected by passing the output into a current amplifier and then measuring the out-of-phase signal using a lock-in amplifier.
The capacitive photocurrent spectrum shows a peak at 2.6 eV, similar to that observed in the absorbance and standard photovoltage. However, a second peak is also observed at 1.77 eV ( 699 nm) whose magnitude is even larger than the ground state exciton peak. Comparison to Figs. 2(a) and 2(b) shows that evidence for the low energy feature can also be observed in the absorbance and standard photovoltage. In both cases, a small deflection is observed near 1.77 eV. However, the feature is much stronger in the capacitive photocurrent spectrum. Since evidence for the new feature is observed in all three measurements, its appearance is clearly not dependent on the use of a femtosecond pulsed laser (such as a multiple photon transition) or to an anomaly of the capacitive photocurrent technique. To further confirm that the low energy CPS peak is not due to two-photon absorption, the magnitude of the peak was measured as a function of laser power. As shown in the inset to Fig. 2(c), the magnitude of the peak increases approximately linearly with increasing power, indicating that it is due to a single photon absorption process.
The observation of a sub-absorption threshold feature in the CPS signal follows reports from a number of groups who have observed sub-gap features in the photoexcitation spectrum of BHJ solar cells using a variety of techniques (including absorption[@goris], photothermal deflection[@jjbs], photoluminescence[@hall], electroluminescence[@trivngs] and electroabsorption[@holt]). Photoluminescence measurements of MDMO:PPV-PCBM show a broad sub-gap peak centered at 1.65 eV[@hall], close in energy to the CPS peak that we observe. A question that remains is to what extent the observed sub-gap states contribute to the charge dissociation process. From Fig. 2(a), it is clear that the absorbance is relatively weak at the low energy CPS peak. The large magnitude of the CPS signal must mean that charge carriers photoexcited at this energy have a very high dissociation efficiency. However, most of the dissociated carriers do not diffuse far enough to be captured by the contacts, (as indicated by relatively weak signal in the standard photovoltage measurement). It is still possible, though, that the low energy state provides a pathway through which higher energy excitons can dissociate into free carriers. If so, it is expected that the occupation of the low energy state would reduce its availability for the dissociation of higher energy excitons, and that this would then lead to a reduction in the photovoltage.
To test this hypothesis, we performed a series of standard photovoltage measurements while exposing the sample to fixed wavelength light. The photovoltage is measured as a function of wavelength using light from the monochrometer (probe beam). For each measurement, the sample is also exposed to fixed wavelength light from the OPA laser (pump beam). The results are plotted in Fig. 3 for four different pump beam wavelengths (solid lines) and for the pump beam blocked (dashed line). The magnitude of the standard photovoltage decreases under exposure to the OPA light, by an amount that is strongly dependent on the wavelength of the pump beam. The maximum change is observed for a pump wavelength of 699 nm (1.77 eV), while almost no change is observed for a pump wavelength of 474 nm (2.6 eV).
Figure 4(a) plots the percent change in the photovoltage signal (integrated over the 380-800 nm probe beam wavelength range) as a function of the pump beam wavelength. As stated above, the greatest change in the photovoltage is centered at an excitation energy of 1.77 eV (699 nm) where a reduction of 71% is observed. For pump beam energies on either side of this value, the bleaching is reduced, forming a pronounced minimum in the photovoltage signal. A comparison to the capacitive photocurrent spectrum (reproduced from Fig 2(c) in Fig. 4(b)) shows that the wavelength dependence of the bleaching directly correlates with the longer wavelength feature in the CPS. To prove that this result is not specific to a particular device, Figs. 4(c) and 4(d) show results of the same measurement performed on a second device structure. Although there is some variation in the detailed spectrum, the main result is reproduced a large, sub-absorbance threshold peak is observed in the CP spectrum, and a decrease in the photovoltage is observed when the sample is exposed to light at the peak energy. It is interesting to note that in the second device, a pair of sub-threshold peaks are observed, however, only one is reproduced in the bleaching measurements. This suggests that lowest energy peak is not involved in the charge dissociation process, and is perhaps due to a sample specific defect state.
It is striking that such large photovoltage bleaching occurs under exposure to long wavelength light even though the absorbance in this regime is relatively weak (see Fig. 2(a)). In contrast, little or no bleaching is observed under exposure to short wavelength light corresponding to the main absorbance peak. This demonstrates that the effect is not simply due to heating of the sample through absorption of the pump beam energy. Instead, the bleaching must somehow be caused by the occupation of the 1.77 eV state. One possibility is that filling of the 1.77 eV state and the subsequent charge dissociation creates an electric field which blocks the flow of additional charge to the contacts. To test this, we measured the voltage directly by replacing the current amplifier used in the capacitive photocurrent measurement with a voltage amplifier. We observe that the photocurrent peak corresponds to a voltage difference of only 30 $\mu$V, which is far less than the change of 10 mV observed in the standard photovoltage, making this explanation unlikely. In addition, no bleaching is observed under exposure to high energy light even though a similar or larger blocking voltage is created.
These results can be understood in terms of exciton dissociation through an interfacial charge transfer state, as has been described in the literature[@goris; @jjbs; @hall; @trivngs; @aryan; @morteani]. Here, the long wavelength peak corresponds to an intermediate state needed for the efficient dissociation of photo generated excitons, while the main absorbance peak (at 2.6 eV) corresponds to the ground exciton state of the MDMO-PPV. The majority of absorbance occurs into the MDMO-PPV and PCBM exciton states. These have long diffusion lengths, but also long dissociation times[@muller]. At the interface between the MDMO-PPV and PCBM, charge transfer excitons are formed with fast dissociation times[@sariciftci]. However, these states also have short diffusion lengths and a small absorption cross section. Charge pairs in the bulk exciton states diffuse to the interface, where dissociation occurs. If the interface is sufficiently close to the contact, a photovoltage is produced. In our bleaching experiment, we populate the interfacial states, blocking the dissociation of the main excitonic states. This reduces the photovoltage by removing the pathway for efficient charge dissociation.
In conclusion, using the capacitive photocurrent technique we are able to resolve a low energy feature in the photo-excitation spectrum of BHJ solar cells that is much more weakly observed in standard absorbance and photovoltage spectroscopy. The state has a large dissociation rate and a low absorbance cross-section compared to states at higher energy. Bleaching of the photovoltage signal is observed for illumination at the low energy feature suggesting that filling this state impedes the charge dissociation process. The experimental results counterintuitively demonstrate that increasing the amount of light on a BHJ solar cell can actually cause the output to go down. This implies that the solar cell efficiency could be improved by filtering out the light over a narrow band of wavelengths corresponding to the interfacial state energies.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Little attention has so far been paid to the division of the observed population between pulsars of the two spin directions that are possible. Almost all pulsars with positive corotational charge density at the polar caps are expected to satisfy space-charge limited flow boundary conditions. Charge separation by blackbody photo-electric transitions in moving ions limits the acceleration potential, analogously with the more usually considered pair creation. But the limitation is more severe so that proton and ion energies can be relativistic but not ultra-relativistic, and these allow the growth of Langmuir-mode induced turbulence that couples directly with the radiation field, as shown by Asseo, Pelletier & Sol. The consequences of this, and of the several possible physical states of the polar cap, are described, qualitatively, as possible explanations for the complex phenomena of nulls, subpulse drift and mode-switching observed in sub-sets of pulsars.'
author:
- |
P. B. Jones[^1]\
Department of Physics, University of Oxford, Denys Wilkinson Building,\
Keble Road, Oxford OX1 3RH, England
title: 'Nulls subpulse drift and mode-switching in pulsars: the polar-cap surface'
---
\[firstpage\]
pulsars: general - stars: neutron - plasma - instabilities
Introduction
============
The past twenty years have seen a number of developments that are likely to be relevant to pulsar radio and X-ray emission. Apart from the many improvements in the quality and scope of observations, there has been the recognition of the role of the Lense-Thirring effect in acceleration at the polar cap (Muslimov & Tsygan 1992) and of the significance of plasma turbulence (see the review of Melrose 2000). Most of the very many theoretical papers published have assumed a relative orientation of rotation spin ${\bf \Omega}$ and polar-cap magnetic flux density ${\bf B}$ such that ${\bf %%@
\Omega}\cdot{{\bf B}} > 0$ with Goldreich-Julian charge density $\rho_{GJ} < 0$ and electron acceleration. In this case, there can be no doubt that the electron work function is so small that the space-charge limited flow boundary condition on the corotating-frame electric field, ${\bf E}_{\parallel} = 0$, is satisfied on the polar-cap surface at all instants of time. (The subscripts $\parallel$ and $\perp$ refer to directions locally parallel with and perpendicular to ${\bf B}$.) There appears to be no reason why neutron stars with ${\bf \Omega}\cdot{\bf B} < 0$ and $\rho_{GJ} > 0$ should not be present in the pulsar population, but in this case, the best existing calculations of ion separation energies (see Medin & Lai 2006) combined with estimates of the polar-cap temperatures can be used to obtain values of the minimum magnetic field for which the boundary condition ${\bf E}_{\parallel} \neq 0$ is possible. These are typically of the order of $10^{14}$ G (Jones 2011) and are indirect evidence that all except possibly a very small number of pulsars with spin direction such that ${\bf \Omega}\cdot{\bf B} < 0$ satisfy the space-charge limited flow boundary condition rather than the ${\bf E}_{\parallel} \neq 0$ condition assumed in the classic polar-cap model of Ruderman & Sutherland (1975).
The concept of localized regions of electron-positron pair production, referred to as sparks, moving over the neutron-star surface within the polar cap in an organized way, has been adapted from the Ruderman-Sutherland model and the ${\bf E}_{\parallel} \neq %%@
0$ boundary condition, but is now widely used phenomenologically, with no reference to boundary conditions or to the sign of ${\bf \Omega}\cdot{\bf B}$, in describing observations such as nulls and subpulse drift. In the ${\bf \Omega}\cdot{\bf B} > 0$ case, with space-charge limited flow, no physical basis for such a phenomenological description has so far been published, the reason being that electrons are the only possible negatively-charged particles for outward acceleration and the space-charge limited flow boundary condition ${\bf E}_{\parallel} %%@
= 0$ can be satisfied at any conceivable polar-cap surface temperature. The state of the accelerated plasma is then simply a problem in electromagnetism with defined boundary conditions. However, for ${\bf \Omega}\cdot{\bf B} < 0$, it has been shown that subpulse formation and movement, either organized or disorganized, over the polar-cap surface is quite natural and should be observable in most radio pulsars with this spin direction (Jones 2010a, 2011; hereafter Papers I and II). The basis for this conclusion is an examination of the physical processes occurring in electromagnetic showers formed at the polar-cap surface by photo-electrons accelerated inward. The problem of finding the state of the accelerated plasma is not restricted to electromagnetism but also depends on proton production by the formation and decay of the giant dipole resonance in the electromagnetic showers.
Subpulse modulation (the wide variations of intensity at a fixed longitude in a sequence of observed pulses) has been recognized as an almost universal property of pulsars ever since their discovery. Weltevrede, Edwards & Stappers (2006) found it present in 170 out of a sample of 187 pulsars selected only by signal-to-noise ratio, evidence that its presence is very likely to be independent of the sign of ${\bf %%@
\Omega}\cdot{\bf B}$. Strong plasma turbulence is now viewed as the most probable source of radio emission (see Melrose 2000). Its development from a quasi-longitudinal Langmuir mode and the possible formation of a random array of stable Langmuir solitons has been investigated by Asseo, Pelletier & Sol 1990, and by Asseo & Porzio 2006. It has been shown recently (Jones 2012; hereafter Paper III) that the formation of this mode is also possible in a two-component beam consisting of protons and relativistic but not ultra-relativistic ions that can be formed as a consequence of the reverse flow of photo-electrons to the polar cap. Thus it is not surprising that subpulse modulation is almost universal. That a secondary low-energy electron-positron plasma is not the only possible source of plasma turbulence is also consistent with the fact that common characteristics of coherent radio emission are observed in the pulsar population even though the inferred polar-cap magnetic flux densities vary by five orders of magnitude. Pulsars with either sign of ${\bf \Omega}\cdot{\bf B}$ are therefore expected to produce individual pulses of similar form although if there is no pair production, there may be differences that are difficult to detect such as in the distributions of spectral index caused by the different plasma frequencies and Lorentz factors in the two cases. But the present paper (see also Paper II) proposes that phenomena such as nulls, mode-changes and organized sub-pulse motion are restricted to the ${\bf %%@
\Omega}\cdot{\bf B} < 0$ case. These involve time-scales that have no obvious explanation in the ${\bf \Omega}\cdot{\bf B} > 0$ direction of spin.
Existing observations show that subpulse drift is a frequent though not universal phenomenon. Weltevrede, Edwards & Stappers found that 68 of their sample of 187 pulsars show observable subpulse drift. There appears to be no published survey either of nulls or of mode-changes based on a carefully defined sample as in the above work. But the most recent and largest tabulation of data (Wang, Manchester & Johnston 2007) shows that nulls must be a moderately frequent phenomenon. Systematic information about mode-changing is much more sparse (see Kramer et al 2006).
Models of subpulse formation and drift recently developed (see, for example, Gil & Sendyk 2000) are based on the classic paper of Ruderman & Sutherland (1975) and assume the ${\bf \Omega}\cdot{\bf B} < 0$ case with the polar-cap surface boundary condition ${\bf E}_{\parallel}\neq 0$. These have all the attributes of a good physical model, particularly simplicity and predictive capability in that it is possible to write down an expression for the ${\bf E}\times{\bf B}$ subpulse drift velocity. But the boundary condition requires surface magnetic flux densities approximately two orders of magnitude larger than the dipole fields inferred from the spin-down rates. Although the presence of highly complex field geometries formed from higher-multipole components cannot be excluded, it would be strange if they were present with the frequency indicated by the survey of Weltevrede, Edwards & Stappers. Consequently, Papers I and II investigated physical processes at the polar-cap surface and have shown that the composition of the plasma accelerated under the space-charge limited flow (SCLF) boundary condition is usually neither time-independent nor uniform over its whole area. Localized areas in which the accelerated plasma is suitable for the growth of the quasi-longitudinal Langmuir mode exist naturally and can move in either an organized or disorganized way, so acting as sources for subpulses with the same properties. Thus plasma systems analogous with Ruderman-Sutherland sparks are present under the ${\bf E}_{\parallel} = 0$ boundary condition. But we emphasize that their motion is unrelated with ${\bf E}\times{\bf B}$ drift but is determined by the time $\tau _{p}$ in which electromagnetic shower protons diffuse to the surface.
This paper attempts to present a description of the early stages of the coherent emission mechanism that is very different from the canonical. In particular, the unexpected conclusion of Paper III that relatively low-energy ion and proton beams can be formed means that it is necessary to re-assess the model of sub-pulse drift discussed in Paper II. Polar caps of pulsars with spin ${\bf \Omega}\cdot{\bf B} < 0$ are seen to be very complex systems compared with those of the electron acceleration ${\bf \Omega}\cdot{\bf B} > 0$ case. But we must presume that both are present in the neutron-star population and so it is worth examining how the properties of the observed pulsar population may be related with spin direction.
Before proceeding further, it must be admitted that the work has at least one drawback. Processes of the kind considered in Papers I and II, occurring at a continuously changing real condensed-matter surface, are not necessarily susceptible to the formulation of simple physical theories. The behaviour of such systems can be varied and complex. But this is not inconsistent with the observed properties of many individual pulsars.
Section 2 summarizes the assumptions about the properties of the open region of the magnetosphere that we are obliged to make in both qualitative and quantitative work. Strong plasma turbulence is possible only for limited intervals of particle energy at altitudes $z$ above the polar-cap surface smaller than $z \sim 10R$, where $R$ is the neutron-star radius. This is described in more detail in Section 3. Section 4 draws on the work of Paper III to obtain values of the parameter $K$, which is the number of protons produced by photo-electrons per unit nuclear charge accelerated. We then give in Section 5 a description of the polar cap showing how subpulses are formed and that either chaotic or organized motion is possible. An attempt is made to show that this model of the polar cap provides a physical basis for phenomena such as nulls, subpulse drift, and mode changes.
The open magnetosphere at the polar cap
=======================================
At the base of the magnetosphere there is an atmosphere of ions, assumed to be in local thermodynamic equilibrium (LTE), having a scale height of the order of $10^{-1}$ cm. Its total mass is temperature-dependent but is broadly equivalent to $10^{-1} - 10^{1}$ radiation lengths. Thus it contains only the earlier part of the electromagnetic showers formed by the reverse-electron flux. There is fractionation of the ionic charge to mass ratio. Ions with the highest value are concentrated at the top of the atmosphere and so preferentially enter the acceleration region. The proton number density is many orders of magnitude smaller than that of the ions and consequently has no effect on the structure of the atmosphere. Protons are not in static equilibrium within the LTE ion atmosphere but are subject to an outward-directed net gravitational and electrostatic force. Their motion is therefore a combination of diffusion and drift with a characteristic time $\tau _{p}$ elapsing between formation and arrival at the top of the atmosphere. If the rate of formation is such that the Goldreich-Julian flux would be exceeded, the excess protons form a thin atmosphere in equilibrium at the top of the LTE ion atmosphere. We refer to Sections 2, 3 and 5 of Paper II for a more complete account of these topics. Values $\tau _{p}\sim 10^{-1} - %%@
10^{0}$ s are estimated.
The polar-cap radius adopted here and in Paper III is based on dipole-field geometry and is that given by Harding & Muslimov (2001), $$\begin{aligned}
u_{0}(0) = \left(\frac{2\pi R^{3}}{cPf(1)}\right)^{1/2},\end{aligned}$$ for rotation period $P$, in which $f(1) = 1.368$ for the neutron star of mass $1.4M_{\odot}$ and radius $R = 1.2 %%@
\times 10^{6}$ cm. As in Paper III, we use the fact that the open magnetic flux lines are contained within a narrow tube, assumed to be of circular cross-section whose radius, $u_{0}(z) = u_{0}(0)\eta^{3/2}$ for a dipole field, increases only very slowly as a function of altitude on the magnetic axis $z = (\eta - 1)R$ at radii $\eta %%@
\sim 10$ or smaller. Therefore, we can adequately represent a small section of the tube at altitude $z$ as a right cylinder of locally-defined radius ${\bf u} = {\bf u}_{0}(z)$. Then the electrostatic potential $\Phi$ at $z\gg u_{0}(0)$ and under the SCLF boundary conditions is approximately $$\begin{aligned}
\Phi(u,z) = \pi\left(u_{0}^{2}(z) - u^{2}\right)\left(\rho(z) - \rho _{GJ}(z)\right),\end{aligned}$$ in terms of the difference between the charge density $\rho$ and the Goldreich-Julian charge density $\rho _{GJ}$ which is almost constant over the cross-sectional area of the tube. We refer again to Harding & Muslimov (2001) for these quantities at radial coordinate $\eta$, $$\begin{aligned}
\rho _{GJ} = -\frac{Bf(\eta)}{cP\alpha \eta^{3}f(1)}\left(1 - \frac{\kappa}{\eta^{3}}
\right)\cos\psi\end{aligned}$$ and, $$\begin{aligned}
\rho = - \frac{Bf(\eta)}{cP\alpha \eta^{3}f(1)}\left(1 - \kappa \right)\cos\psi,\end{aligned}$$ in which $B$ is the surface magnetic flux density. Equation (4) gives the SCLF charge density undisturbed by photo-electric transitions or electron-positron pair creation. In these expressions, $\kappa$ is the dimensionless Lense-Thirring factor which we assume to be $\kappa = 0.15$, and $\psi$ is the angle between ${\bf \Omega}$ and ${\bf %%@
B}$. Equation (4) also assumes that particle velocities, including those of protons and ions, differ negligibly from the velocity of light. The function $f(\eta)/\alpha f(1)$ including the red-shift factor $\alpha$ is a slowly-varying function of $\eta$ and is approximately unity at low altitudes. Thus $\rho _{GJ}(0) = (1 - \kappa)\rho _{GJ}(\infty)$ where $\rho _{GJ}(\infty)$ is the flat space-time Goldreich-Julian charge density. Equations (3) and (4) exclude terms that can become significant at higher altitudes for which our geometrical model also fails, but nevertheless equation (2) is adequate for the estimation of proton and ion energies and of photo-electric transition rates. It is convenient to use the energy units of particle physics for polar-cap processes and therefore we define a potential energy difference $V({\bf u},z) = -e\Phi$. It has the value $V_{max}({\bf u},z)$ on a flux line with polar-cap coordinate ${\bf %%@
u}$ for the undisturbed SCLF charge density of equation (4). It is also useful to have a simple expression for $V_{max}$ on the magnetic axis, $$\begin{aligned}
V_{max}(0,\infty) \approx \frac{2\pi^{2}R^{3}\kappa eB}{c^{2}f(1)P^{2}}
= 1.25\times 10^{3}\frac{B_{12}}{P^{2}} \hspace{5mm}{\rm GeV},\end{aligned}$$ for acceleration above the polar cap, where $B_{12}$ is the surface magnetic flux density in units of $10^{12}$ G.
The above statement is subject to a serious qualification concerning the definition of the open magnetosphere. If many secondary low-energy electron-positron pairs were created per primary particle accelerated, there would be no difficulty in assuming the unconstrained outward flow of primary particles and secondary pairs on any magnetic flux line intersecting the light cylinder. The separate velocity distributions of secondary electrons and positrons would have the capacity to adjust so as to maintain the condition ${\bf E}_{\parallel} = 0$ at all points in the open region. But there must be serious doubts about the universal existence, in the pulsar population, of the necessary density of secondary-pair plasma. Apart from ${\bf \Omega}\cdot{\bf B} < 0$ pulsars in which there may be no pair creation, calculations of the secondary pair density produced by the inverse Compton scattering (ICS) of polar-cap photons by outward accelerated electrons in the ${\bf \Omega}\cdot{{\bf B}} > 0$ case (Hibschman & Arons 2001; Harding & Muslimov 2002) predict in many pulsars pair densities that are small compared with unity, per primary electron. It is true that more complex field geometries, such as the off-set dipole, are possible and may enhance pair production (see Harding & Muslimov 2011) but there remain at least some hints of a problem in the pair density. In the absence of a pair plasma with the required properties, motion in the magnetosphere beyond the $\eta\sim 10$ region depends on the sense of curvature of the open flux line: there is further acceleration for $|\cos\psi|$ increasing, but deceleration in the opposite case which is capable of cancelling the acceleration at lower altitudes. A section of the polar cap is then a dead zone from which there is no outward particle flux. Whilst the change in active polar-cap shape is not very significant, the potential given by equation (2) is reduced, very roughly, in proportion to the decrease in area. The component ${\bf E}_{\parallel}$ in the dead volume is very small and is that required to maintain a charge-separated equilibrium with gravity and the centrifugal force.
A further source of uncertainty is the consistency with which the total electric charge of the neutron star remains constant. The rate of loss of charge at the polar caps is $\dot{q} = 2\pi u^{2}_{0}(0)\rho _{GJ}(0)c$ whose compensation is likely to be easy if there is strong secondary pair creation, but is otherwise obscure. An order of magnitude estimate of the relevant time-scale can be found by dividing a typical light-cylinder field strength by the rate of change of the light-cylinder radial electric field given by $\dot{q}$. This is, $$\begin{aligned}
\left(\frac{BR^{3}}{R^{3}_{LC}}\right)\left(\frac{4\pi R^{2}_{LC}}{\dot{q}}\right)
\sim 2P \hspace{2mm} {\rm s},\end{aligned}$$ where $R_{LC}$ is the light-cylinder radius. This indicates that if there is to be no substantial change in the structure of the magnetosphere, cessation of the processes maintaining charge constancy is possible only for time intervals no more than of the order of $P$.
But the great uncertainty affecting ${\bf \Omega}\cdot{\bf B} < 0$ pulsars is in the atomic number of matter at the neutron-star surface. If this were small $(Z\sim 6)$ there would be a negligible reverse-electron flux and, as we shall see in Section 3, no coherent radio emission unless curvature radiation electron-positron pair production is possible. There is some indirect observational evidence for $Z = 6$ (Ho & Heinke 2009; see also Bogdanov & Grindlay 2009, Yakovlev et al 2011, and Zhu et al 2009) but it does not appear strong. Theoretical work does indicate the possibility of atomic numbers other than the canonical $Z = 26$. Calculations by Hoffman & Heyl (2009) for isolated neutron stars with no mass accretion show the presence of $Z = 14$, though with small total mass, but Pearson, Goriely & Chamel (2011) find a range of values $Z %%@
> 26$. The mass per unit area removed by a Goldreich-Julian current density of nuclei with mass number $A$ in an ${\bf \Omega}\cdot{\bf B} < 0$ pulsar in a life-time $t$ is, $$\begin{aligned}
\frac{Am_{p}}{Z}\int^{t}_{0}\frac{-B\cos\psi}{eP(t^{\prime})}dt^{\prime} \approx %%@
\frac{-2Am_{p}Bt\cos\psi}{ZeP(t)},\end{aligned}$$ assuming that $B$ and $\psi$ are time-independent. This is $\sim 4\times %%@
10^{11}B_{12}P^{-1}$ g cm$^{-2}$ at $t = 10^{6}$ yr, and over the whole surface of the star amounts to a layer of no more than $4\times 10^{-9}B_{12}P^{-1}$ $M_{\odot}$. It is a quantity that could easily be the result of unknown details of the formation process such as fall-back. In view of this, direct experimental evidence of the nuclei that are actually present at the polar-cap surface is clearly needed.
The uncertainty in the atomic number of polar-cap surface nuclei has an obvious effect on photo-electric transitions, and therefore on the reverse-electron flux and the acceleration potential difference for particles moving on any given magnetic flux line. The dependence of $\Phi$ on $\rho(z)$ in equation (2) means that the production of oppositely charged particles in a charge-separated region at altitude $z$ under the SCLF boundary condition tends to cancel the local electric field ${\bf E}_{\parallel}$. This is well-known in the case of electron-positron pairs but is, of course, also true for photo-electrons (see Paper III).
Equation (2) is not valid at low altitudes $z \sim u_{0}(0)$. Near the polar-cap surface, $z \ll u_{0}(0)$, ${\bf E}_{\parallel}$ is the inertial acceleration field first described by Michel (1974). The problem of finding the electric field is one-dimensional in $z$ and the inertial ${\bf E}_{\parallel}$ is not large. Its contribution to particle acceleration is small compared with that of the Muslimov-Tsygan effect. We refer to Paper III for an approximate model for this region.
The polar-cap and whole-surface neutron-star temperatures are treated as distinct in Paper III and in the present work. Calculated photo-electric transition rates for polar-cap blackbody photons are negligibly small above an altitude $z = h \approx 4u_{0}(0)$, principally because the photon momentum component perpendicular to ${\bf B}$ becomes small at higher altitudes. This is unfavourable both for the Lorentz transformation to the rest-frame of the accelerated ion and for the cross-section to the lowest Landau state of the emitted electron (see Paper III).
The SCLF boundary conditions lead to an electric field having a complex form in the $z %%@
< h$ region which is unlikely to be well represented by equation (2). There also exist the complications, just discussed, of the inertial ${\bf E}_{\parallel}$ and of photo-electric transitions in this interval. But the boundary conditions also limit the extent to which $\rho(z)$ can change if the condition $E_{\parallel} > 0$ is to be satisfied. Thus the difference between the mean ion charges $Z_{h}$ at $z = h$ and $\tilde{Z}$ in the assumed local thermodynamic equilibrium of the atmosphere at $z = 0$ cannot be large. Adjustments in $\rho(z)$ and in ${\bf E}_{\parallel}$ maintain this state by limiting ion Lorentz factors and hence photo-electric transition rates. The ratio, $$\begin{aligned}
\frac{\rho(h)}{\rho(0)} = \frac{KZ_{s} + 2Z_{h} - Z_{\infty}}
{KZ_{s} + 2\tilde{Z} - Z_{\infty}}, \nonumber\end{aligned}$$ including ion, proton and electron charge densities is always very nearly unity at $z < h$. Here $K$ is the mean number of protons produced in electromagnetic showers per unit nuclear charge accelerated. The mean charge of nuclei reaching the surface after moving through the electromagnetic shower region is $Z_{s}$, and $Z_{\infty}$ is the ion charge at $z_{max}$, where photo-electric transition rates become negligibly small. Because the acceleration in this interval is minor compared with that at $z > h$ we can assume that the effect of this region is to inject ions of mean charge $Z_{h}$ and Lorentz factor $\gamma _{h}$ into the main acceleration region $z > h$ in which $\rho(z)$ is modified by photo-electric transitions and possibly by electron-positron pair creation. Although the mean value of $Z_{h} - \tilde{Z}$ is small, it is undoubtedly finite, and the associated reverse-electron flux gives a modest energy input, $\epsilon _{h}$ , to the polar-cap surface even in the absence of any whole-surface contribution to photo-ionization at $z > h$. We refer to Paper III for its approximate size, $\epsilon _{h} \sim 20$ GeV.
For a relatively small fraction of pulsars, pair creation by the conversion of curvature radiation (CR) photons is possible. But for most ${\bf \Omega}\cdot{\bf B} > %%@
0$ pulsars, it is believed that the source of pairs is the conversion of outward-directed ICS polar-cap photons (see Hibschman & Arons 2001, Harding & Muslimov 2002). However, in the case of the ${\bf \Omega}\cdot{\bf B} < 0$ pulsars considered here, there is an essential difference in that the electron-photon momentum in the frame of the rotating neutron star is directed inwards. There is likely to be some pair creation at low values of $z$ from $\gamma$-rays emitted by the capture of shower-generated neutrons at the polar-cap surface, but the rate is difficult to estimate. Thus pair creation rates in most ${\bf \Omega}\cdot{\bf B} < 0$ pulsars are likely to be one or two orders of magnitude smaller than for those with spin direction ${\bf \Omega}\cdot{\bf B} > 0$. But even for the latter spin direction, the number of ICS pairs formed per primary electron can be much smaller than unity (see Harding & Muslimov 2002, also Fig. 8 of Hibschman & Arons). Thus for the ${\bf \Omega}\cdot{\bf B} < 0$ spin direction, there is a very real problem in understanding how a density of low-energy secondary electron-positron pairs adequate for coherent radio emission can be formed.
Conditions for strong turbulence
================================
There is now a consensus that, in almost all observable radio pulsars, the source of the coherent emission lies within radii $\eta \sim 10^{1} - 10^{2}$ and its energy is derived from particle acceleration at the polar cap. There is also a growing consensus that the formation of strong plasma turbulence is involved (Melrose 2000). The unstable mode that develops into strong plasma turbulence and transfers energy from particle beams to the electromagnetic field must therefore have a high growth rate. This severely constrains the energy of the beam particles. The equations of motion for the system are the Maxwell equations and those for a relativistic particle fluid. In a strong magnetic field, the mass-to-charge ratio present in the equations is $m_{i}\gamma^{3} _{i}q^{-1}_{i}$ for particles of species $i$, where $q_{i}$ is the charge and $\gamma _{i}$ the Lorentz factor. The longitudinal effective mass is $m_{i}\gamma^{3}_{i}$ and this factor necessarily determines the growth rate of any unstable mode.
We consider in this paper, specifically, the growth of a quasi-longitudinal Langmuir mode (Asseo, Pelletier & Sol 1990) from which a random array of stable Langmuir solitons may develop (Asseo & Porzio 2006). A different longitudinal mode has been studied by Weatherall (1997, 1998) and may be the source of the nanosecond pulses observed in certain pulsars (Hankins et al 2003; Soglasnov et al 2004). (We refer to the paper of Asseo & Porzio for an account of the history of plasma turbulence in relation to pulsar physics.) The Langmuir mode and solitons have components ${\bf %%@
E_{\parallel}} \neq 0$, ${\bf E_{\perp}} \neq 0$ and ${\bf B}_{t}$ perpendicular to ${\bf E}_{\perp}$ and so can transfer energy to the radiation field directly (see Asseo, Pelletier & Sol 1990). Hence the direct transfer of energy to the radiation field is determined only by the longitudinal effective mass and by the departure of the mode wave-vector from the purely longitudinal state. Examination of growth rates shows at once that primary electrons or positrons of $\sim %%@
10^{3}$ GeV energy have longitudinal effective masses so large that they are unable to participate in the mode. The beams that interact in ${\bf \Omega}\cdot{\bf B} > 0$ pulsars can be only secondary electrons and positrons whose velocity distributions satisfy the relativistic Penrose condition (Buschauer & Benford 1977). But ion and proton beams that have longitudinal effective masses of the same order as secondary electrons and positrons can be produced in ${\bf \Omega}\cdot{\bf B} < 0$ neutron stars. Thus it is anticipated that pulsars of both directions exist and produce very similar coherent emission.
Beams of cold ions and protons (or, hypothetically, cold electrons and positrons) permit growth of the quasi-longitudinal Langmuir mode for which the dispersion relation is, $$\begin{aligned}
\left(\omega^{2} - k^{2}_{\parallel}\right)\left(1 - \frac{\omega^{*2}_{1}}
{(\omega - k_{\parallel}\beta _{1})^{2}} - \frac{\omega^{*2}_{2}}
{(\omega - k_{\parallel}\beta _{2})^{2}}\right) = k^{2}_{\perp},\end{aligned}$$ for angular frequency $\omega$, wave-vector ${\bf k}$, and ion and proton (or, electron and positron) velocities $\beta _{1}$ and $\beta _{2}$, respectively. The remaining quantities are $\omega^{*} _{i} = \gamma^{-3/2}_{i}\omega _{i}$ in which, $$\begin{aligned}
\omega^{2}_{i} = \frac{4\pi N_{i}q^{2}_{i}}{m_{i}},\end{aligned}$$ where $q_{i}$, $m_{i}$ and $N_{i}$ are respectively the charge, mass and neutron-star frame number density of each component $i = 1,2$. The transverse field components are, $$\begin{aligned}
E_{\perp} & = & \frac{k_{\perp}E_{\parallel}}{2(k_{\parallel} - \omega)},\nonumber %%@
\\
B_{t} & = & \frac{k_{\perp}E_{\parallel}}{\omega} - %%@
\frac{k_{\perp}k^{2}_{\parallel}E_{\parallel}}{\omega(k^{2}_{\parallel} - \omega^{2})}.\end{aligned}$$ The plasma frequency in the rest frame of a component is $\gamma^{-1/2}_{i}\omega _{i}$ and so, following Asseo et al, a natural wavenumber for which to seek a solution is defined in terms of the slower-moving component, and is $k_{\parallel} = %%@
2\omega^{*}_{1}\gamma^{2}_{1}$. With the definition of a new variable $s$ such that, $$\begin{aligned}
\omega - k_{\parallel}\beta _{1} = \omega^{*}_{1}(1 + s),\end{aligned}$$ and, $$\begin{aligned}
\omega - k_{\parallel} = \omega^{*}_{1}(1 + s) - \frac{k_{\parallel}}{2\gamma^{2}_{1}},\end{aligned}$$ valid for $\gamma _{1} \gg 1$ equation (8) can be expressed as, $$\begin{aligned}
1 - \frac{1}{(s + 1)^{2}} - \frac{C}{(s + \mu)^{2}} = %%@
\frac{k^{2}_{\perp}\gamma^{2}_{1}}{k^{2}_{\parallel}s},\end{aligned}$$ in which $\mu = \gamma^{2}_{1}/\gamma^{2}_{2}$ and $C = \omega^{*2}_{2}/\omega^{*2}_{1}$. In the limit $k_{\perp} \rightarrow 0$, equation (13) is a quartic with two real and two complex roots, whose properties were investigated in Paper III. The amplitude growth rate was found to be of the order of ${\rm Im}\omega \sim 0.2\omega^{*}_{1}$ and is a slowly varying function of the ratio $\gamma _{1}/\gamma _{2}$ and, for ions, the proton-ion number density ratio. But we have to consider the quasi-longitudinal case of $k_{\perp} \neq 0$ in which the transverse field components are non-zero. Equation (13) is then less transparent, being a quintic with three real and two complex roots, but we can use perturbation theory to estimate the extent to which a non-zero $k_{\perp}$ changes a given complex root. Let $s = s_{0} + \delta s$, where $s_{0}$ is a root of the quartic. Then, $$\begin{aligned}
\delta s\left(\frac{1}{(s_{0} + 1)^{3}} + \frac{C}{(s_{0} + \mu)^{3}}\right) = %%@
\frac{k^{2}_{\perp}\gamma^{2}_{1}}{k^{2}_{\parallel}s_{0}}.\end{aligned}$$ From equations (10), we see that $|E_{\perp}| = |E_{\parallel}$ for $k_{\perp} = %%@
k_{\parallel}|s|/\gamma^{2}_{1}$. Then for typical values of $C$, $\mu$ and $s_{0}$, we have, $$\begin{aligned}
|\delta s| \approx \frac{k^{2}_{\perp}\gamma^{2}_{1}}{k^{2}_{\parallel}|s_{0}|}
= \frac{|s_{0}|}{\gamma^{2}_{1}},\end{aligned}$$ indicating that for a significant range of $k_{\perp}$, the growth rate does not differ much from the $k_{\perp} = 0$ value.
The growth rate decreases with distance above the polar cap and so it is appropriate to obtain the amplitude growth factor $\exp \Lambda$ at a given radius $r = \eta R$ by integration. It is, $$\begin{aligned}
\Lambda = {\rm Im}s\left(\frac{-16\pi \alpha _{1}q_{1}R^{2}B\cos\psi}
{Pm_{1}c^{3}\gamma^{3}_{1}}\right)^{1/2} \left(1 - \eta^{-1/2}\right),\end{aligned}$$ assuming a dipole field and constant $\gamma _{1}$. Here, the ion (or electron) component number density in the frame of the rotating neutron star is a fraction $\alpha _{1}$ of the Goldreich-Julian number density. For ${\rm Im}s = 0.2$ and $\cos\psi = -1$, the factor is, $$\begin{aligned}
\Lambda = 2.4\times 10^{5}
\left(\frac{B_{12}q_{1}m\alpha _{1}}{Pem_{1}\gamma^{3}_{1}}\right)^{1/2}
\left(1 - \eta^{-1/2}\right),\end{aligned}$$ where $m$ is the electron rest mass. We can see that the growth factor increases most rapidly at small $\eta$ and is substantially developed at $\eta = 2$. The only particle property that it depends on is the charge to longitudinal effective mass ratio. A minimum value necessary for the development of nonlinearity is presumably given by the size of particle number fluctuations within a cubic-wavelength volume at the Goldreich-Julian number density. This indicates a minimum $\Lambda \sim 20$. We assume $\Lambda = 30$. For a typical pulsar with $PB^{-1}_{12} = 1$, equation (17) then requires $\gamma _{1} %%@
< 176\alpha^{1/3}_{1}$ for electrons or $\gamma _{1} < 11\alpha^{1/3}_{1}$ for ion beams at $\eta = 2$ increasing to $\gamma _{1} < 311\alpha^{1/3}_{1}$ and $\gamma _{1} < 20\alpha^{1/3}_{1}$, respectively, at $\eta = 10$. Development of turbulence or of stable solitons is then possible for either electrons or ions although for electrons, the cold beam hypothesis is clearly unrealistic and the distributions of electron and positron velocities must satisfy the Penrose condition.
Quasi-longitudinal modes transfer energy directly to the radiation field at angular frequencies of the order of $\omega _{0} = 2\gamma^{1/2}_{1}\omega _{1}$. The actual maximum in the intensity distribution presumably depends on the extent to which developed turbulence moves energy toward higher wavenumbers. This is unknown except that some examples of soliton formation considered by Asseo & Porzio have maxima at $\omega \approx 1.5\omega _{0}$. As a function of $\eta$ and for $\cos\psi = -1$, the frequencies $\nu _{0} = \omega _{0}/2\pi$ are, $$\begin{aligned}
\nu _{0} & = & 4.8\times 10^{9}\left(\frac{B_{12}\alpha _{1}}{P}\right)^{1/2}
\left(\frac{\gamma _{1}}{\eta^{3}}\right)^{1/2} \nonumber \\
\nu _{0} & = & 1.1\times 10^{8}
\left(\frac{B_{12}Z_{\infty}\alpha _{1}}{AP}\right)^{1/2}
\left(\frac{\gamma _{1}}{\eta^{3}}\right)^{1/2}
\hspace{3mm}{\rm Hz} \end{aligned}$$ respectively, for electrons and for ions of mass number $A$ and charge $Z_{\infty}$. These values differ by two orders of magnitude and suggest the existence of two classes of pulsar with very different radiation spectra though with individual pulses of similar form. There have been relatively few measurements of flux density below 400 MHz but those of Deshpande & Radhakrishnan (1992) and of Malofeev et al (1994) both confirm that low-frequency spectral cut-offs are usually below $100$ MHz and in some cases, below $50$ MHz.
There have been many calculations of pair formation densities above polar caps, but of particular interest is the paper of Harding & Muslimov (2011) who have studied the effect of a specific type of field, that of an off-set dipole. Pair formation densities, assuming outward moving ICS photons to be the source, are rapidly increasing functions of the degree of off-set. The most extreme case in which pair formation might be thought doubtful is that of PSR J2144-3933 which has a period $P = 8.51$ s and whose observers (Young, Manchester & Johnston 1999) commented specifically on the problem. The value of its parameter $BP^{-2} = 2.87\times 10^{10}$ G s$^{-2}$ is almost an order of magnitude smaller than the apparent cut-off value of $2.2\times 10^{11}$ G s$^{-2}$ in the distribution of that quantity obtained from the ATNF pulsar catalogue (Manchester et al 2005) representing a death-line in the $P$-$\dot{P}$ distribution below which pair creation is not possible. It is old ($2.7\times 10^{8}$ yr) with a relatively low mean flux density at 400 MHz, and may be observable only as a consequence of close proximity (170 pc). Although the maximum acceleration potential given by equation (5) is very small, $V_{max} = 36$ GeV, ICS pair formation is possible in principle although a very high degree of off-set would be required (see Figs. 7 - 9 of Harding & Muslimov 2011). However, given a spin direction ${\bf \Omega}\cdot{\bf B} < 0$, accelerated ions would have Lorentz factors appropriate for growth of the quasi-longitudinal mode for any field geometry. But the value of $\nu _{0}$ given by equations (18) is rather small in this case, even assuming that the radiation field is largely decoupled from the magnetosphere at quite low altitudes $\eta \sim 2$. (For this spin direction, the pair multiplicities produced by ICS photons would probably be one or two orders of magnitude smaller than for ${\bf \Omega}\cdot{\bf B} > 0$ because the momentum of the electron-photon system is inward directed.) But more extensive observations at low frequencies may reveal whether or not J2144-3933 is unique or merely a nearby example of an otherwise unobserved population.
Instability and proton production
=================================
Owing to our late recognition that photo-electric transitions have a very significant effect on particle acceleration under SCLF boundary conditions, it is necessary to revise the estimates of the reverse-electron energy flux that were used in Papers I and II. Photo-electric transitions and the consequent reverse-electron flux produce changes in mean nuclear and ion charge that we can summarize here as follows. The initial nuclear charge is $Z$ (canonical value $Z = 26$) but formation and decay of the giant dipole state in electromagnetic showers reduces it to $Z_{s}$ at the top of the neutron-star atmosphere. We assume local thermodynamic equilibrium (LTE) in this region with ion charge $\tilde{Z}$. Then ions with charge $Z_{h}$ are injected into the main acceleration region at $z = h$. Typically, $Z_{h}\approx \tilde{Z} + 1$ owing to interaction with polar-cap blackbody photons. Then further acceleration produces more extensive ionization to $Z_{\infty} \leq Z_{s}$ by the whole-surface blackbody field with Z(z) as the altitude-dependent ion charge during this process.
In papers I-III we defined $KZ_{s}$ to be the number of protons produced per ion of nuclear charge $Z_{s}$ accelerated. The initial problem treated here is of acceleration given by equations (2) - (4) with $\rho(z)$ independent of time and of position ${\bf u}$. Then, $$\begin{aligned}
\frac{\rho(z)}{\rho(h)} = \frac{KZ_{s} + 2Z(z) - Z_{\infty}}{KZ_{s} + 2Z_{h} - %%@
Z_{\infty}}, \nonumber\end{aligned}$$ and the total photo-ionization at altitude $z$ is related to the potential at that point by, $$\begin{aligned}
\frac{2Z(z) - 2Z_{h}}{KZ_{s} + 2Z_{h} - Z_{\infty}} = \frac{\kappa(\eta^{-3}_{h} - %%@
\eta^{-3})}
{1 - \kappa \eta^{-3}_{h}}\left(1 - \frac{V({\bf u},z)}{V_{max}({\bf u},z)}\right),\end{aligned}$$ where $\eta _{h} = 1 + h/R$. This result takes into account protons, ions and the reverse flux of photo-electrons, but not CR electron-positron pair creation or the conversion of ICS photons.
Photo-electric transition rates are rapidly increasing functions of the ion Lorentz factor $\gamma$. The reverse-electron flux they produce leads to a charge density that increases with altitude, so reducing the potential given by equation (2). Thus we might expect a self-consistent set of functions $V(z,{\bf u})$, $\rho(z,{\bf u})$ and $\gamma(z,{\bf u})$ that are time-independent. In order to investigate this possibility, it is necessary to calculate the reverse-electron energy flux taking into account the fact that photo-electric transitions reduce the potential energy $V = - %%@
e\Phi$ derived from equation (2) to levels below $V_{max}$. Broadly, we expect that $\gamma(z)$ will increase from its initial value $\gamma _{h}$ but that $V({\bf u},\infty) < V_{max}({\bf u},\infty)$, and our initial attempt to model this is to assume that $V({\bf u},z) = V_{max}({\bf u},z)$ until a cut-off $V_{c}$ is reached.
With the procedures and assumptions of Paper III, we calculate the mean reverse-electron energy $\epsilon _{s}$ per ion resulting from interaction with whole-surface blackbody photons, and the mean ion charge $Z_{\infty}$. Transition rates are obtained to an altitude $z_{max} = 3R$, beyond which they become small and equation (2) predicts negligible further acceleration. The polar-cap temperature is treated as distinct from the general surface temperature of the star and its value is equal to that generated by the reverse-electron flux which produces protons at the Goldreich-Julian current density. A flux exceeding this produces an accumulation of protons at the top of the neutron-star atmosphere which, until exhausted, reduces the ion component of the current density to zero. We adopt equation (33) of Paper I, with $\cos\psi = -1$ and rotation period $P = 1$ s, and a typical proton production rate per unit shower energy of $W_{p}mc^{2} = 2\times 10^{-4}$, that is, one proton per $2.5$ GeV shower energy. This quantity is a slowly varying function of $B$ and $Z$ and has been estimated in Papers I and II. Then $(\epsilon _{h} + \epsilon _{s})W_{p} = KZ_{s}$ with $\epsilon %%@
_{h} \sim 20$ GeV being the contribution of polar-cap blackbody photons. The temperatures are $T_{pc} = 0.74\times 10^{6}$ K, $0.82\times 10^{6}$ K and $1.10\times 10^{6}$ K respectively for $B = 1.0\times 10^{12}$, $3.0\times 10^{12}$ and $10^{13}$ G. These have negligible effect on the acceleration but determine the initial ion charge $\tilde{Z}$. Actual values of $Z_{s}$ are unknown [*ab initio*]{} but are given by $Z_{s} = Z/(1 + %%@
\bar{K})$, where $\bar{K}$ is the time-average of $K$ at a specific position ${\bf u}$ on the polar cap and $Z$ is here the pre-shower nuclear charge. Hence we assume arbitrary values $Z_{s} = 10$ or $20$ with mass numbers $A = 20$ or $40$ and note that ions with very small $Z_{s}$ tend to be completely ionized by the polar-cap or whole-surface temperature at the start of acceleration and so produce no reverse-electron flux. The initial ion Lorentz factor is fixed as $\gamma _{h} = 5$ at an altitude $z = h = 0.05R$. The initial ion charges at $z = h$ are: for $Z_{s} = 10$; $Z_{h} = 8,6$ or $5$ and for $Z_{s} = 20$; $Z_{h} = 15,12$ or $9$ respectively at $B = %%@
1.0\times 10^{12}$, $3.0\times 10^{12}$ G or $10^{13}$ G..
Values of $\epsilon _{s}$ and $Z_{\infty}$ obtained on this basis are given in Table 1. With the approximations and methods of Paper III in mind, it is obvious that the absolute values of $\epsilon _{s}$ do not have the accuracy that the number of figures given might indicate. But comparison of columns for different values of $T_{s}$ shows that whole-surface photons can generate very high transition rates not only at large $V_{c}$ for $T_{s} = 10^{5}$ K, but for progressively lower $V_{c}$ as whole-surface temperature increases. By combining the results of Table 1, which are estimates of $\epsilon _{s}(V)$ and $Z_{\infty}(V)$, with equation (19) it should be possible, given $Z_{h}$, to solve for $V({\bf u},\infty)$. It can be seen by inspection of the equation, with reference to values of $K(V)$ found from the Table, that there is always a solution, the value of $V$ depending principally on $T_{s}$ and to a lesser extent on $B$. Values $V \ll V_{max}$ occur at high $T_{s}$ and small $B$.
Basically, incomplete ionization at $z = h$ means that there will be further ionization at $z > h$ unless $T_{s}$ or $V_{max}$, or both, are too small for it to be possible. In this latter case, relevant for old pulsars, there remains the polar-cap contribution $\epsilon _{h}$ giving a small factor $K < 1$ and a solution $V({\bf u},z) = %%@
V_{max}({\bf u},z)$. Such a solution would be time-independent and of uniform $\rho$ over the polar cap, with both ion and proton components. But it would not lead to radio emission unless $V_{max}$ were small enough to give an adequate growth rate for the Langmuir mode, as described in Section 3. The general cut-off $BP^{-2} = 2.2\times %%@
10^{11}$ G s$^{-2}$ obtained from the ATNF pulsar catalogue gives $V_{max}({\bf %%@
u},\infty) = 275(1 - u^{2}/u^{2}_{0})$ GeV. Ion Lorentz factors $\gamma \sim 20$ and proton energies $\sim 60$ GeV are possible on flux lines originating near the edge of the polar cap. However, in the more general case, large values of $\epsilon _{s}$ in the uniform time-independent model lead to $K \gg 1$ and hence to instability, as shown in Section 4.3 of Paper I. The actual state of the polar-cap surface is therefore time-dependent.
Qualitative modelling of the polar cap
======================================
The basic principle is that the state of any element $\delta {\bf u}$ of surface at any instant is a function only of its past time-distribution of reverse-electron energy flux. The flux of accelerated protons $J^{p}({\bf u},t)$ is given by, $$\begin{aligned}
J^{p}({\bf u},t) + \tilde{J}^{p}({\bf u},t) = \int^{t}_{-\infty}f_{p}(t - %%@
t^{\prime})K({\bf u},t^{\prime})J^{z}({\bf u},t^{\prime}),\end{aligned}$$ in terms of the ion flux $J^{z}$. The quantity $\tilde{J}^{p} = 0$ within intervals for which $J^{p}$ does not exceed the Goldreich-Julian current density $\rho _{GJ}(0)c$. Most protons are produced in the vicinity of the shower maximum whose depth is only a slowly varying function of the primary electron energy. Their motion toward the surface is better described by a drift velocity, rather than diffusion, in most of the LTE atmosphere so that the distribution of arrival times can be modelled as $f_{p}(t - t^{\prime}) = \delta(t - t^{\prime} - \tau _{p})$. We refer to Section 5 of Paper II for further details. Following Paper III, equation (20) recognizes that $K$ is a function of ${\bf u}$ and $t$ owing to its $V-$dependence, and can vary on time-scales of the order of $\tau _{p}$. The protons are preferentially accelerated, but if $J^{p}$ reaches the Goldreich-Julian flux, as can be the case if $K \gg 1$, a thin proton atmosphere forms at the top of the ion atmosphere at a rate given by $\tilde{J}^{p}$ and the local ion flux falls to zero until it is exhausted.
As a first approximation, equation (20) shows that the condition of an element of polar-cap area can be represented by two possible states: ion emission with duration $\tau _{p}$ or proton emission of duration $\tau _{gap} \approx \bar{K}\tau _{p}$, where $\bar{K}$ is here the time-average over the interval $\tau _{p}$. Values $\bar{K} \gg 1$ ensure that proton production is so large that an atmosphere of protons forms at the top of the LTE atmosphere of ions. Then the ion phase ends and the proton phase is maintained until the proton component of the atmosphere is exhausted. Areas in the ion phase necessarily move with time and so can be seen as moving subpulses within the time-averaged radio-pulse profile. The motion could be chaotic or, as proposed in Paper II, organized so as to give regular subpulse drift.
Pair creation by curvature radiation (CR) photons was referred to briefly at the end of Section 2 but has been otherwise ignored here. The necessary conditions are thought to be present in a relatively small fraction of the observed population (Harding & Muslimov 2002) but we shall review briefly its effects in the ${\bf \Omega}\cdot {\bf B} < 0$ case with SCLF boundary conditions. Self-sustaining pair creation requires adequate conversion probabilities for both outward and inward-directed CR photons at $z < z_{max}$. Thus pair creation is not uniform over the whole cross-sectional area $\pi u^{2}_{0}(z)$ of the open flux tube, but depends on a combination of $V$ and of flux-line curvature. For the suitable values of ${\bf %%@
u}$, the equation analogous to (19), obtained by comparison of current densities at $z %%@
= 0$ and $z = z_{max}$, can be most compactly expressed in terms of the particle flux $\phi$ on the relevant flux lines, $$\begin{aligned}
\frac{\phi _{p} + |\phi _{e}|}{\phi _{p} - |\phi _{e}|} = 1 + \frac{\kappa}{1 - \kappa}
\left(1 - \frac{V({\bf u},\infty)}{V_{max}({\bf u},\infty)}\right),\end{aligned}$$ in which $\phi _{e}$ is the flux either of electrons or positrons, and $\phi _{p}$ is the proton flux. This assumes, as is almost certainly the case, that the reverse-electron energy flux produces an excess of protons at the top of the LTE atmosphere. There is then no ion flow except on flux lines that do not satisfy the CR criteria such as those near $u_{0}$ on which $V_{max}$ is too small.
---------- --------- --------- ----------------- -------------- ------------ -------------- ------------ --------------
$B_{12}$ $Z_{s}$ $V_{c}$ $\epsilon _{s}$ $Z_{\infty}$ $\epsilon$ $Z_{\infty}$ $\epsilon$ $Z_{\infty}$
GeV $T_{s5}$ $= 1$ $2$ $4$
1.0 10 1000 1389 9.9 650 10.0 255 10.0
500 253 8.5 650 10.0 255 10.0
250 4 8.0 392 9.6 255 10.0
125 38 8.3 228 10.0
80 12 8.1 128 9.6
50 2 8.0 67 9.2
20 1000 2196 18.0 2044 20.0 803 20.0
500 328 15.7 1783 19.7 803 20.0
250 5 15.0 640 17.5 802 20.0
125 56 15.4 395 18.4
80 14 15.2 241 17.9
50 2 15.0 136 17.3
3.0 10 2000 6438 9.3 4264 10.0 1607 10.0
1000 138 6.1 3767 10.0 1607 10.0
500 36 6.1 581 7.2 1524 10.0
250 7 6.0 244 7.0 734 9.3
125 64 6.5 373 8.9
80 16 6.2 225 8.5
50 2 6.0 119 7.9
20 2000 10580 17.4 9196 20.0 3760 20.0
1000 104 12.1 6677 18.9 3760 20.0
500 27 12.1 647 13.3 3346 19.8
250 3 12.0 225 12.9 1348 17.7
125 35 12.3 604 16.5
80 5 12.1 319 15.4
50 124 14.0
10.0 10 2000 85 5.0 3114 6.6 8521 10.0
1000 47 5.0 631 5.7 3578 8.9
500 12 5.0 345 5.7 1724 8.6
250 1 5.0 98 5.4 814 8.1
125 10 5.1 314 7.3
80 1 5.0 144 6.5
50 47 5.7
20 2000 137 9.1 3061 10.6 19778 20.0
1000 58 9.1 1055 10.1 6523 16.1
500 10 9.0 456 9.9 3269 15.6
250 1 9.0 90 9.3 1388 14.3
125 5 9.0 440 12.1
80 155 10.6
50 31 9.5
---------- --------- --------- ----------------- -------------- ------------ -------------- ------------ --------------
: Values of the mean reverse-electron energy per ion ($\epsilon _{s}$) at the polar-cap surface in units of GeV, and of the mean ion charge at the end of photo-ionization ($Z_{\infty}$) have been calculated. The magnetic flux density is in units of $10^{12}$ G and the acceleration potential cut-off $V_{c}$ is in units of GeV. Arbitrary values $Z_{s} = 10$ or $20$ have been assumed for the mean surface nuclear charge. We refer to Section 4 for discussion of $Z_{s}$ and other parameters on which the calculated values are dependent. The final three pairs of columns give the values of $\epsilon _{s}$ and of $Z_{\infty}$ for whole surface temperatures $T_{s} = 1,2$ and $4 \times 10^{5}$ K. The rotation period is $P = 1$ s. Blank spaces denote energies that are negligibly small.
Quantitative model construction is not attempted here. Owing to the uncertainties in significant parameters that are listed in Section 6, quantitative testing of a model against observed data does not appear immediately profitable. The purpose of this paper is less ambitious. It is an attempt to show that the condition of the polar-cap surface provides a physical basis for nulls, mode-switches and subpulse drift. To do this, we describe qualitatively some of the possible polar-cap states.
Assume, initially, that CR pair production as described above is possible over some limited area of the polar cap. Its boundary can be defined as that within which proton production by the reverse-electrons of CR pairs exceeds the Goldreich-Julian rate so that a proton atmosphere forms and grows in density. Between this boundary and ${\bf %%@
u_{0}}$ there is an annular strip within which ion acceleration can occur and proton formation is governed by equation (20). (Any contribution from CR reverse-electrons is neglected here.) Solutions of equation (20) then exist in which zones of ion phase circulate around the CR region. Equation (20) is local in ${\bf u}$ and so cannot itself determine the circulation sense or velocity. Suppose that there are $n$ ion zones. Then in the usual notation by which subpulse drift is described, the circulation time is $\hat{P}_{3} = nP_{3}$, where $P_{3}$ is the band separation and in the model is given by $P_{3} = \tau _{gap} + \tau _{p} = (K + 1)\tau _{p}$. Association with subpulse drift would require either secondary pair creation on the ion phase flux lines or the presence of proton and ion components in the current density near to ${\bf u_{0}}$ so that energies are low enough for Langmuir-mode growth. A further consideration is that excess proton production within the CR area must be accompanied by lateral diffusion on the polar-cap surface, perpendicular to ${\bf B}$. We have chosen the example of a clearly defined annular region but there is no reason why organized motion of this kind should not occur more generally and in the absence of a CR component.
A further possibility is that the active CR region could be switched off by changes in the potential inside it which is also dependent on the state of the remaining parts of the polar cap. In general, in a less-organized state, there is no reason why either the whole polar cap or large parts of it should not be in the proton phase for intervals of time of the order of $\tau _{gap}$ and, in the absence of CR pair production, there would be no possibility of Langmuir mode growth. Thus nulls are a natural phenomenon within the model. Obviously, there are many possible polar-cap states, their relevance to widely-observed phenomena being limited only by the requirement that they should not be too dependent on neutron-star parameters having critical or particular values.
Mode-changes have been studied extensively in a small number of pulsars, some of which also exhibit subpulse drift. The paper of Bartel et al (1982) summarizes the main characteristics of this phenomenon. These authors emphasize that the spectral and polarization changes that occur on mode-switching are consistent with a change in the optical depth of the source region within the magnetosphere and that the sources of the two or more modes would be on different sets of flux lines. (They also suggest that the explanation of mode-switching lies in the polar-cap surface, though in the context of the Ruderman & Sutherland model.) Switching of emission between different sets of flux lines occurs naturally in our qualitative model which is sufficiently diverse to accommodate the characteristics of mode-switching that are observed, including changes in subpulse drift rates.
The qualitative model is applicable to subpulse drift and to short-duration nulls or mode changes and there is no difficulty in understanding how non-randomness or quasi-periodicity, such as that reported by Rankin & Wright (2008) and Redman & Rankin (2009) might occur naturally in such a system. The time-scales are necessarily short, perhaps of the order of $\tau _{gap}$, but there is also the possibility of a medium time-scale instability in the atomic number $Z_{s}$ of surface nuclei, described in Section 4 of Paper II. This can result in values of $Z_{s}$ that are too small to produce any reverse-electron flux because the atoms are completely ionized in the LTE atmosphere. Having the largest charge-to-mass ratio, apart from protons, they are in equilibrium at the top of the atmosphere and so are preferentially accelerated through the full potential difference $V({\bf u},\infty)$ with no possibility of Langmuir-mode growth if CR pair creation is not possible. A broad measure of the time-scale for this instability is provided by the time in which a depth of surface nuclei equivalent to one radiation length is removed from the polar cap at the Goldreich-Julian current density. From equations (4) and (5) of Paper I, this is given by, $$\begin{aligned}
t_{rl} = 2.1\times %%@
10^{5}\left(\frac{-P\sec\psi}{ZB_{12}\ln(12Z^{1/2}B_{12}^{-1/2})}\right) \hspace{2mm}
{\rm s}.\end{aligned}$$ The distribution of $Z_{s}$ over the polar cap at any instant need not, and probably will not, be uniform. Thus complex distributions of potential and of accelerated beam composition are to be expected.
Conclusions and uncertainties
=============================
Both neutron-star spin directions, defined by the sign of ${\bf \Omega}\cdot{\bf B}$ at the polar caps, presumably exist and it is one of the purposes of this paper to question how, if at all, each sign contributes to the observed radio pulsar population. Previous papers (I - III) have been directed toward the ${\bf \Omega}\cdot{\bf B} < 0$ case because it has positive polar-cap corotational charge density so that the flux of accelerated particles can have ionic, proton and positron components. There appears to have been some reluctance in the published literature to consider in any detail the physics of the polar cap for this spin direction. But Papers I - III have attempted to show that such considerations are essential and that the radio emission characteristics are not solely determined by electrodynamics, as may be so for ${\bf \Omega}\cdot{\bf %%@
B} > 0$.
Weltevrede, Edwards & Stappers (2006) detected subpulse drift in almost one half of a sample of $187$ pulsars selected only by signal-to-noise ratio. Nulls and mode changes are less frequently observed. We propose that results obtained in Papers I - III and in the present paper lead to the idea that these phenomena are all related to the physics of the polar cap in ${\bf \Omega}\cdot{\bf B} < 0$ pulsars and that the opposite spin direction, with only electrons as primary accelerated particles, is then observed as the population showing subpulse modulation, perhaps as a consequence of plasma turbulence, but none of the above phenomena. The problem with polar-cap physics is that a real condensed-matter system can be extremely complex and, as we noted in Section 1, does not always lend itself to the construction of simple models (see also Section 2 of Paper II). A further difficulty is that some of the parameters concerned are either unknown, such as the atomic number of surface nuclei, or can vary over several orders of magnitude.
Calculated values in Table 1 are limited to $B = 10^{12} - 10^{13}$ G which is representative of pulsars in the nulls tabulation of Wang et al (2007) and, for example, of those pulsars with drifting subpulses investigated by Gil et al (2008). Photo-electric cross-sections obtained in Paper III are not valid at $B < 10^{12}$ G in which region interpolation between zero-field cross-sections and those at $10^{12}$ G would be necessary. At higher fields, $B > B_{c} = 4.41 \times 10^{13}$ G, approximate cross-sections for bremsstrahlung and pair creation have been found (Jones 2010b) and electromagnetic shower development investigated. It was confirmed that the value of the proton production parameter $W_{p}$ decreases only slowly at $B > B_{c}$ but also that the Landau-Pomeranchuk-Migdal effect, known to exist in zero field, is significant in the neutron-star atmosphere, and has the effect of reducing high-energy bremsstrahlung and pair creation cross-sections and therefore of increasing shower depth and hence the proton diffusion-drift time $\tau _{p}$.
The whole-surface temperature $T_{s}$ is a further important parameter that is not well known. Paper III did not incorporate general-relativistic corrections that should strictly have been made in transforming the blackbody radiation field to the ion rest frame. But for most transitions, the radiation field is better approximated by that of the neutron-star surface proper frame rather than by $T_{s}^{\infty}$, the temperature inferred by a distant observer. Since $T^{\infty}_{s}\approx 0.8T_{s}$ for the neutron-star mass and radius assumed here in Section 2, the temperatures used in Table 1 are well below those that can presently be observed. Yakovlev & Pethick (2004) list only a small number of young pulsars whose blackbody temperatures have been measured. These are $T^{\infty}_{s} > 5\times 10^{5}$ K. All model predictions reviewed by these authors show $T^{\infty}_{s}$ decreasing very rapidly at age $> 10^{6}$ yr. But on the other hand, temperatures below $T_{s} = 10^{5}$ K, the lowest used in Table 1, could be maintained by a very small and obscure level of dissipation within the star. Isolated neutron stars (INS) certainly have higher temperatures $T^{\infty}_{s}\sim 10^{6}$ K (see, for example, Kaplan & van Kerkwijk 2009). But they are few in number and their radio beams (if they exist) would be narrow compared with the $4\pi$ observability of unpulsed blackbody radiation. Given this uncertainty, we do not regard them as part of the population considered here.
The energy of the coherent radio emission must presumably be transferred from particle beams in high magnetic field regions near the polar caps. Whatever the coherent emission mechanism is eventually established to be, provided it is not curvature radiation, the criteria in Section 3 for growth of Langmuir modes must be relevant. This paper does not attempt to judge whether low-energy secondary electrons or small Lorentz factor ions form the particle beams producing radio emission in any particular group, but asserts that both possibilities exist for ${\bf \Omega}\cdot{\bf B} < 0$ pulsars. It follows that evolution of a pulsar to old age may include long intervals, perhaps of the order of $10^{6}$ yr, in which pair creation is not possible because $V_{max}$ is too small but ion Lorentz factors remain too large to give adequate Langmuir-mode growth rates. Because $V_{max}({\bf u},\infty) \propto u^{2}_{0} - u^{2}$, at least approximately, the highest growth rates will be on flux lines with $u$ near $u_{0}$. This is entirely consistent with the conal pattern of emission in the pulsar morphology developed by Rankin (1983) .
Pulsars exhibit nulls with mean duration as short as several rotation periods or, for those described as intermittent pulsars, times of the order of $10$ d. PSR 1931+24 is an example of the latter (Kramer et al 2006), but long-duration nulls have now also been observed in J1832+0029 (see Lyne 2009) and in J1841-0500 (Camilo et al 2012). Such null durations allow separate measurement of the spin-down rate in both on and off-states of emission. In each case, the off-state spin-down rate is about half that of the on-state, indicating some substantial difference in the condition of the magnetosphere. Obviously, it is not known if this is also true for short-duration nulls but for ${\bf \Omega}\cdot{\bf B} < 0$ pulsars in our model of the polar cap, it occurs naturally owing to the change in the extent to which the electromagnetic fields near and beyond the light cylinder are loaded and the total momentum density modified by particle emission. If the whole polar cap were in the proton phase, as discussed in Section 5, there would be no significant pair creation (assuming no CR pair production) and, of course, no possibility of Langmuir-mode growth. However, if creation of low-energy secondary pairs were possible, the flux of particles to be accelerated at or beyond the light cylinder could greatly exceed the Goldreich-Julian value. Even though we lack a comprehensive understanding of acceleration at $R_{LC}$ and beyond, it is entirely plausible that the changes in particle number density inherent in our model would lead to the observed changes in spin-down torque.
Interesting reviews of pulsars with large nulling fractions and of the Rotating Radio Transients (RRAT) have been given recently by Burke-Spolaor & Bailes (2010), Keane et al (2011) and by Keane & McLaughlin (2011). In particular, pulsars are shown in the $g - P$ plane, where $g$ is here the fraction of periods in which a pulse is detected. The distribution appears roughly uniform in $\log g$ within the interval $10^{-4} < g < %%@
10^{-1}$. The existence of these objects is not inconsistent with the complexity of possible polar-cap states of the ${\bf \Omega}\cdot{\bf B} < 0$ pulsars described here given the time-scale $t_{rl}$ defined by equation (22). But Keane & McLaughlin note that there is a scarcity of objects whose behaviour is governed by time-scales that are longer than the above but shorter than the very long time-scales ($\sim 10^{6} - 10^{7}$ s) associated with the intermittent pulsars such as PSR 1931+24. Very long time-scales much exceeding those given by equation (22) still remain a problem whose explanation may be more subtle than those attempted in this paper or may lie in quite different considerations.
A simple explanation for the RRAT has been proposed by Weltevrede et al (2006). These authors compared RRAT pulses with the infrequent very large-amplitude pulses seen in PSR 0656+14. They note that if 0656+14 were as distant in the galaxy as most of the RRAT, its observed emission would appear to be that of the typical RRAT. But this may be inconsistent with the periodicities in individual RRAT pulse arrival times found recently by Palliyaguru et al (2011). These are mostly of the order of $10^{4}$ s but a minority are much longer, of the order of $10^{6} - 10^{7}$ s. If these complex sets of periodicities are established and are not loose quasi-periodicities, they represent a very real problem for the model of the polar cap described here. It is difficult to see how they can be present in a truly isolated neutron star, with either spin direction, not interacting with any other periodic system. Quasi-periodicities might be understandable in terms of natural phenomena at the condensed-matter surface, possibly connected with the diffusion of low-$Z_{s}$ nuclei perpendicular to ${\bf B}$ in the vicinity of the polar cap or with the two sections of the open magnetosphere described in Section 2. But well-defined periodicities would remain a problem.
[99]{}
Asseo E., Pelletier G., Sol H., 1990, MNRAS, 247, 529 Asseo E., Porzio A., 2006, MNRAS, 369, 1469 Bartel N., Morris D., Sieber W., Hankins T. H., 1982, ApJ, 258, 776 Bogdanov S., Grindlay J. E., 2009, ApJ, 703, 1557 Burke-Spolaor S., Bailes M., 2010, MNRAS, 402, 855 Buschauer R., Benford B., 1977, MNRAS, 179, 99 Camilo F., Ransom S. M., Chatterjee S., Johnston S., Demorest P., 2012, ApJ, 746:63 Deshpande A. A., Radhakrishnan V., 1992, J. Ap. Astr., 13, 151 Gil J. A., Sendyk M., 2000, ApJ, 541, 351 Gil J. A., Haberl F., Melikidze G., Geppert U., Zhang B., Melikidze G. Jr., 2008, ApJ, 686, 497 Hankins T. H., Kern J. S., Weatherall J. C., Eilek J. A., 2003, Nat., 422, 141 Harding A. K., Muslimov A. G., 2001, ApJ, 556, 987 Harding A. K., Muslimov A. G., 2002, ApJ, 568, 862 Harding A. K., Muslimov A. G., 2011, ApJ, 743, 181 Hibschman J. A., Arons J., 2001, ApJ, 554, 624 Hoffman K., Heyl J., 2009, MNRAS, 400, 1986 Ho W. C. G., Heinke C. O., 2009, Nat., 462, 71 Jones P. B., 2010a, MNRAS, 401, 513 (Paper I) Jones P. B., 2010b, MNRAS, 409, 1719 Jones P. B., 2011, MNRAS, 414, 759 (Paper II) Jones P. B., 2012, MNRAS, 419, 1682 (Paper III) Kaplan D. L., van Kerkwijk M. H., 2009, ApJ, 705, 798 Keane E. F., McLaughlin M. A., 2011, Bull. Astr. Soc. India, 39, 333 Keane E. F., Kramer M., Lyne A. G., Stappers B. W., McLaughlin M. A., 2011, MNRAS, 415, 3065 Kramer M., Lyne A. G., O’Brien J. T., Jordan C. A., Lorimer D. R., 2006, Sci., 312, 549 Lyne A. G., 2009, in Becker W., ed., Astrophysics and Space Science Library, Vol. 357, Neutron Stars and Pulsars. Springer, Berlin, p.67 Malofeev V. M., Gil J. A., Jessner A., Malov I. F., Seiradakis J. H., Sieber W., Wielebinski R., 1994, A&A, 285, 201 Manchester R. N., Hobbs G. B., Teoh A., Hobbs M., 2005, Astron J., 129, 1993 Medin Z., Lai D., 2006, Phys. Rev. A, 74, 062508 Melrose D. B., 2000, in Kramer M., Wex N., Wielebinski N., eds, ASP Conf. Ser. Vol. 202, Pulsar Astronomy - 2000 and Beyond. Astron. Soc. Pac., San Francisco, p.721 Michel F. C., 1974, ApJ, 192, 713 Muslimov A. G., Tsygan A. I., 1992, MNRAS, 255, 61 Palliyaguru N. T., et al, 2011, MNRAS, 417, 1871 Pearson J. M., Goriely S., Chamel N., 2011, Phys. Rev. C, 83, 065810 Rankin J. M., 1983, ApJ, 274, 333 Rankin J. M., Wright G. A. E., 2008, MNRAS, 385, 1923 Redman S. L., Rankin J. M., 2009, MNRAS, 395, 1529 Ruderman M. A., Sutherland P. G., 1975, ApJ, 196, 51 Soglasnov V. A., Popov M. V., Bartel N., Cannon W., Novikov A. Yu., Kondratiev V. I., Altunin V. I., 2004, ApJ, 616, 439 Wang N., Manchester R. N., Johnston S., 2007, MNRAS, 377, 1383 Weatherall J. C., 1997, ApJ, 483, 402 Weatherall J. C., 1998, ApJ, 506, 341 Weltevrede P., Edwards R. T., Stappers B. W., 2006, A&A, 445, 243 Weltevrede P., Stappers B. W., Rankin J. M., Wright G. A. E., 2006, ApJ, 645, L149 Yakovlev D. G., Ho W. C. G., Shternin P. S., Heinke C. O., Potekhin A. Y., 2011, MNRAS, 411, 1977 Yakovlev D. G., Pethick C. J., 2004, Ann. Rev. Astron. Ap., 42, 169 Young M. D., Manchester R. N., Johnston S., 1999, Nat., 400, 848 Zhu W., Kaspi V. M., Gonzalez M. E., Lyne A. G., 2009, ApJ, 704, 1321
\[lastpage\]
[^1]: E-mail:[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the power-law tails in the evolution of massless fields around a fixed background geometry corresponding to a black hole. We give analytical arguments for their existence at $scri_{+}$, at the future horizon and at future timelike infinity. We confirm their existence with numerical integrations of the curved spacetime wave equation on the background of a Schwarzschild and a Reissner-Nordström black hole. These results are relevant to studies of mass inflation and the instability of Cauchy horizons. The analytic arguments also suggest the behavior of the full nonlinear dynamics, which we study numerically in a companion paper.'
address: 'Department of Physics, University of Utah, Salt Lake City, UT 84112-1195'
author:
- 'Carsten Gundlach, Richard H. Price and Jorge Pullin'
date: '9 July, 1993'
title: |
Late-time behavior of stellar collapse and explosions:\
I. Linearized perturbations
---
Introduction
============
In the study of nonspherical gravitational collapse, the late stages of black hole formation and nonspherical stellar dynamics, certain simplifications have been traditional. Typically, linearized perturbation theory has been used on a fixed background, and the results have been taken to be representative of nonperturbative collapse.
One of the most basic results [@DeBr; @RoWi; @Ro] about the evolution of fields on a curved background is that “sandwich waves” are not usually possible. At late times waves do not cut off sharply but die off in “tails.”
In the context of perturbations of spherical objects, like stars or black holes, arguments have been given [@Pr] leading to the conclusion that the tails have a specific power-law form. The intention of this paper is to analyze in detail in which regions of the spacetime this form for the tails holds and under what conditions they develop.
In section II we start by giving an analytical outline of the development of tails in spherically symmetric fixed backgrounds. We present a somewhat more general and more pedagogical derivation of the results of the appendix of reference [@Pr]. Moreover, we give two new results:
1\) We show that for perturbations around a black hole, power-law tails develop not only at timelike infinity as was proven in [@Pr] but also at $scri_{+}$ and at the black-hole horizon. This result is of relevance since the development of tails in these regions is crucial for the physics of mass inflation [@HeKo] and the stability of Cauchy horizons. At least twice [@CuMoPr; @Le] in the literature it has been stated that the power-law tails are “nonradiative,” suggesting that the tails made no appearance at $scri_{+}$ or at the horizon.
2\) Generalizing the arguments for the Schwarzschild background, we show that power-law tails develop even when no horizon is present in the background. This would mean, among other things, that power-law tails should be present in perturbations of stars, or after the implosion and subsequent explosion of a massless field which does not result in black hole formation.
In sections III and IV we confirm the first of these results for the Schwarzschild background by performing numerical integrations of the perturbation equations for different initial shapes and different multipole moments of the field. We also confirm that the results extend to the case of Reissner-Nordström black holes; this is the first clear evidence that power-law tails actually develop for the background of direct relevance to mass inflation and stability of Cauchy horizons. Section III contains a brief discussion of our numerical method; section IV contains our numerical results.
In section V we make some final remarks, especially about possible implications, both for the behavior of test fields evolving on on a time-dependent stellar collapse or explosion background, and for the behavior of a (spherically symmetric) self-gravitating massless field in a collapse or explosion situation. We develop this subject in a subsequent paper.
Tails have also been found important in the detailed calculation of gravitational waveforms from the inspiral collapse of binary systems. The tail back reaction appear as a correction to the 3/2 Post Newtonian equations of motions [@AKOP].
Schwarzschild background and beyond
===================================
In this section we examine the evolution of massless perturbation fields outside a star collapsing to form a black hole. We make the idealization that the mass loss in the transition from star to black hole can be neglected. (If there are massless perturbation fields outside the star at all, they will carry some energy away to null infinity, but we neglect this compared to the mass of the star.) If the star is spherically symmetric (again neglecting the perturbation fields, which need not be spherically symmetric), spacetime outside the star is Schwarzschild spacetime. We are therefore dealing with the (massless) wave equation on a region of Schwarzschild spacetime given by $$ds^{2}= -\left(1-{2 M \over r}\right) dt^{2}+\left(1-{2 M \over
r}\right)^{-1} dr^{2} +r^{2}(d \theta^{2}+\sin^{2}\theta d\phi^{2})\ ,$$ where we have used $c=G=1$ units. We introduce a “tortoise” radial coordinate $r_{*}\equiv r+2M\ln(r-2M)$, advanced time $v\equiv t+r_*$, and retarded time $u\equiv t-r_*$, in terms of which the line element becomes $$ds^{2}= -(1-{2 M \over r}) dudv
+r^{2}(d \theta^{2}+\sin^{2}\theta d\phi^{2})\ .$$ We let the star begin its collapse at retarded time $u=u_0$. The characteristic $u=u_0$ will be one boundary for our problem.
After the surface has started collapsing, it rapidly approaches the speed of light. The wordline of the stellar surface is therefore asymptotic to an ingoing null geodesic $v=v_0$. It is convenient to consider this $v=v_0$ null geodesic, rather than the stellar surface, as the left boundary of the problem, and the data on this line for $u>u_0$ to complete the specification of the problem. The variation of $\phi$ on the stellar surface is asymptotically infinitely redshifted. (See [@Pr] for details.) This means that $\phi,_{u}$ will be small (more precisely, exponentially small) at late times and we make the specific assumption that, after some retarded time $u=u_{1}$, variations in $\phi$ on $v=v_0$ can be ignored. This black hole formation scenario is illustrated in Fig. 1.
Since the background is spherically symmetric, each multipole of a perturbation field evolves separately. For a scalar field $\phi(t,r,\theta,\varphi)$, for example, we can write $\phi=
\sum_{l,m}
\Psi_{m}^{l}(u,r) Y_l^m(\theta,\varphi)/r $ and for each multipole moment we have $$\label{wave}
-4\Psi_{,uv}=V_l(r)\Psi,\qquad\qquad V_l\equiv\left(1-
{2M\over
r}\right)
\left({l(l+1)\over r^2}+ R(r)\right),$$ where $R(r)$ falls off as $r^{-3}$ for large $r$, and where we have suppressed the indices $l,m$ on $\Psi$. The evolutions of scalar, electromagnetic, and gravitational perturbations are all governed by an equation with the form of (\[wave\]). Only the details of $R(r)$ differ from one type of field to another.
The general solution to (\[wave\]) can be written as a series depending on two arbitrary functions $F$ and $G$: $$\label{full}
\Psi=\sum_{p=0}^l A_l^p r^{-p} \left[(G^{(l-p)}(u)+(-1)^p F^{(l-
p)}(v)\
\right]
+\sum_{p=0}^\infty B_l^p(r) \left[(G^{(l-p-1)}(u)+(-1)^p F^{(l-p-
1)}(v)\
\right],$$ where the superindices on $F$ and $G$ indicate the number of times the function is differentiated; negative superindices are to be interpreted as integrals. The coefficients $A_l^p$ are dimensionless fixed numbers which have the same values as in flat spacetime. The coefficient functions $B_l^p(r)$, however, vanish for $M\rightarrow0$ and have the form $$\label{Beq}
B_l^p(r)=a_l^p
r^{-(p+2)}\left[ 1+O(M/r)\right]\ ,$$ where the coefficients $a_l^p$ are proportional to $M$. (See [@Pr] for additional details.)
The most interesting characteristic data on $u=u_0$ correspond to two different cases: a) an initially static perturbation field for $u\le
u_0$, and b) a vanishing perturbation field for $u\le u_0$. The first case corresponds to the collapse of an initially static star with an initially static perturbations. The second case corresponds to the collapse of an initially static star in which the collapse itself generates the perturbations.
One may also consider a more general (nonstatic) initial perturbation on $u=u_0$, but its generic effects can be modeled by the data on $v=v_0$, after reflection through the center. We shall see, on the other hand, that there is a qualitative difference in the tails if there is a static perturbation present initially.
We break the problem of evolution of the fields into two steps. In the first step we find $\phi$ on $u=u_1$. This step requires the data on $u=u_0$, but only the early (small $u$) data on the left boundary. In this step, then, the type of the left boundary (whether stellar center, surface of completely collapsing star, etc.) is unimportant. From the data on $u=u_1$ we find $\phi$ at late times, and at $scri_{+}$, in a second step. [*To leading order in $M$*]{}, the evolution depends on the spacetime curvature in the first step, but not in the second step. The validity of this approximation is ultimately justified by numerical experiment.
We begin with the first step, the scattering in the region $u_0<u<u_1$. We have taken the variation in $\phi$ on $v=v_0$ to be negligible after $u_{1}$. This means that, aside from backscattering, there is no outgoing radiation for $u>u_{1 } $. In (\[full\]) the first series represents the non-backscattered waves, so our assumptions about no late outgoing waves are equivalent to taking $G(u_1)=0.$ For large $r$ at $u=u_1$ the dominant term in (\[full\]) is then $$\Psi(u=u_1,r) =a_l^l r^{-l-2} G^{(-1)}(u_1) \left[ 1+O(M/r)\right]\ ,$$ where $G^{(-1)}(u_1)\equiv\int_{u_0}^{u_1} G(u) du$. The characteristic data, then, is proportional to the integrated initial burst, as well as to $M$. In the case of an initial static field, with the form $$\label{statsoln}
\Psi_{\rm static}=\mu r^{-l}\left[ 1+O(M/r)\right]\ ,$$ the prediction is more definitive. In this case, at $u=u_1$, we have $$\label{case2}
\Psi=M\mu\left[(2 l+1)/(l+1)\right] r^{-(l+1)}\left[
1+O(M/r)\right]\ .$$ (See again [@Pr] for details.)
With this specification of characteristic data on $u=u_1$ we can go on to consider the subsequent evolution of fields. We will show that the late time behavior of the fields at constant $r$, and the late-time behavior at $scri_{+}$, are independent of the details of the background at small $r$. The evolution is now the same in the spacetime geometry exterior to a star, a black hole, or to, e. g., an imploding-exploding shell of scalar field.
We confine our attention to the region $u>u_1, r_*\gg M$. The leading order effect on the propagation of $\Psi$ is now that of the “centrifugal term" $l(l+1)/r^2$ in $V_l$. In this sense we now approximate spacetime as flat. The solution for $\Psi$ is then that of (\[full\]) with $M=0$, $$\label{flat}
\Psi(u,v)=\sum_{p=0}^l A_l^p r_*^{-p} \left[(g^{(l-p)}(u)+(-1)^p
f^{(l-p)}(v)\
\right] \ .$$ (To the approximation we are using here, for $r_*\gg M$, we have also $r=r_*$.) By matching this form to the initial data on $u=u_1$ we find that $$\label{ftail}
f(v)=F_0/v^P$$ where $$F_0=(-1)^l 2MG^{(-1)}(u_1), \quad P=2 ,$$ if there is no initial static field,and $$F_0=(-2)^l 2M\mu\left[ l!/(2 l)! \right], \quad P=1,$$ if there is an initial static field, with the form of (\[statsoln\]). (These expressions involve a sum over the $A_l^p$, which are given in [@Pr].)
We next take $t\gg r_*$ and we expand $g^{(k)}(u)= g^{(k)}(t) -
g^{(k+1)}(t)
r_* +\cdots ,$ and similarly for $f(v)$. By reordering terms we arrive at $$\Psi=\sum_{n=-l}^\infty K_l^n r_*^n
\Bigl[f^{(l+n)}(t)+(-1)^n g^{(l+n)}(t)\Bigr].$$ The coefficients $K_l^n$ are constructed from the $A_l^p$ and vanish for $l-n$ even, $-l\le
n\le
l$. We define $h\equiv f+(-1)^l g$, and we note that for $t\gg r^*$ $\Psi$ has the form $$\begin{aligned}
\label{Taylor}
\Psi=&&K_l^{-l}h(t)r_*^{-l}+K_l^{-l+2}h''(t)r_*^{-l+2}+...\\
&&+K_l^lh^{(2l)}(t)r_*^l+K_l^{2l+1}\Big[2f^{(2l+1)}(t)-
h^{(2l+1)}(t)\Big]\,
r_*^{l+1}+...\nonumber\end{aligned}$$ We know that at late times $f(v)$ has the power-law form given in (\[ftail\]); we make the [*ansatz*]{} that $g(u)$ also falls off as a power law, so that we may write $h(t)=H_0/t^N$, where $H_0$ is a constant.
We note that as $t\rightarrow\infty$ the term $h(t)r_*^{-l}$ will dominate if $N<P+2l+1$ and the term $2f^{(2l+1)}(t)r_*^{l+1}$ will dominate if $N>P+2l+1$. In either case only one of the constants $H_0$ or $F_0$ has an influence on the solution, and in fact the constant is simply an overall scaling of the solution. If either $N<P+2l+1$ or $N>P+2l+1$, therefore, the form of the solution (aside from overall scaling) is fixed for $r_*\gg M$, its continuation to smaller values of $r_*$ is fixed, and we cannot satisfy boundary conditions for small $r$ (e.g., regularity at $r=0$ for a stellar model or, as $r_*\to-\infty$, at the horizon of a black hole). In order to satisfy a small-$r$ boundary condition we must have $N=P+2l+1$ and hence $$\label{same}
f(t)\simeq F_0 t^{-P}, \quad g(t)\sim (-1)^{(l+1)} F_0 t^{-P},$$$$h(t)\equiv f(t)+(-1)^l g(t)\sim 1/t^{P+2l+1}.$$
For $M\ll r_*\ll t$ the form of the fields is particularly simple. In that case we have $$\label{larger}
\Psi=K_l^{2l+l} 2 f^{(2l+1)}(t) r_*^{l+1}
=-2 K_l^{2l+l} F_0 (P+2l)! t^{-(P+2l+1)} r_*^{l+1} \ ,$$ and $\Psi$ therefore falls off as $1/t^{2l+2}$ (initial static field) or $1/t^{2l+3}$ (no initial static field) at timelike infinity $i_+$. It follows from (\[same\]) and (\[flat\]) that at $scri_{+}$ (i. e., at $v\rightarrow\infty$) we have $$\Psi(v\to\infty,u)\simeq A_l^0 g^{(l)}(u)\simeq -A_l^0 F_0(l+P-1)! u^{-
P-1}\ .$$ At null infinity $scri_{+}$ the fields therefore fall off as $u^{-l-1}$ (static initial field) or as $u^{-l-2}$ (no static initial field).
We keep in mind that the above analysis did not depend on the small-$r$ details of the problem, and we go on to consider the specific case of a black hole. (It does not matter if it is eternal or formed in collapse, only that there is an internal “infinity", i. e. an event horizon.) As $r_*\rightarrow- \infty$ the curvature potential $V_l$ in (\[wave\]) is negligibly small and the solution to (\[wave\]) can be written as $\Psi=\alpha(u)+\gamma(v)$. The nature of the characteristic data at $v=v_0$ requires that $\alpha(u)$ be a constant (aside from exponentially small variation) and with no loss of generality we choose it to be zero. For $|r_*|\ll t$ we can then expand the solution, for large $u$ and $r_*\ll-M$ as $$\label{lateH}
\Psi=\gamma(v)=\gamma(t)+\gamma'(t)r_*+\frac{1}{2} \gamma''(t)
r_*^2+\cdots\ .$$
To join this solution, at $r_*\ll-M$ to our previous solution in (\[larger\]) for $r_*\gg M$, we make one further assumption. We assume that when $t\gg|r_*|$ then, for the whole range of $r_*$ (from $r_*\ll-M$ to $r_*\gg M$), the solution has the form $\Psi\approx\Psi_{\rm
finstat}(r)/t^{P+2l+1}$, where $\Psi_{\rm finstat}(r)$ is a $t$- independent solution of (\[wave\]). This is clearly the case in the region $t\gg r_*\gg M$. With this assumption we can match the solution in (\[lateH\]) at $r_*\ll-M$ and that in (\[larger\]) for $r_*\gg M$, and conclude that $\gamma(t)=\Gamma_0 t^{-P-2l-
1}$, and therefore that the late time behavior at the horizon is $$\label{utoinf}
\Psi(u\to\infty,v)=\gamma(v)=\Gamma_0 v^{-P-2l-1}\ .$$ The constant $\Gamma_0$ is determined by the condition that there is a static solution $\Psi_{\rm finstat}$ such that $$\label{horlimit}
\lim_{r_*\rightarrow -\infty}\Psi_{\rm finstat}(r)
=\Gamma_0,$$$$\label{infrlim}
\lim_{r_*\rightarrow \infty}\Psi_{\rm finstat}(r)
=-2K_l^{2l+1} F_0(P+2l)! r^{l+1}.$$ The coefficient $\Gamma_0$, and therefore the behavior at the horizon, like all other aspects of the late time behavior, is fixed by the initial backscattering.
In the above analysis, we have seen that the backscattering of the initially outgoing waves, and the subsequent evolution in time, produces the power-law tails at $scri_{+}$ and at future null infinity $i_ +$. The small-$r$ nature of the background does not enter (except, of course, in the discussion of the tails at the horizon). This means that the same power-law tails should develop at $scri_{+}$ and at $i_+$ in models other than black hole models. We might consider as backgrounds: incompletely collapsing stars, static stars, imploding and exploding shells, and so forth. These different models would have different small-$r$ boundary conditions. All that is important to tail formation is that the source of the perturbations is sharply cut off as happens, due to the infinite redshift, in the black hole collapse case. We might consider, as an example, a quadrupole deformation of a nonrotating neutron star. A dynamical process might change the quadrupole perturbation from one static value to another (nonzero) static value. In this case, clearly, there cannot be a power-law fall off of the perturbation in time; the perturbation does not fall off at late times. On the other hand there might be a phase change in which the stellar crust loses the ability to support shear stresses responsible for the quadrupole moment, and the star might quickly become spherical. In this case, our analysis predicts that at late times the exterior quadrupole perturbation will fall off as $1/t^{6}$.
The only specific detail of our analysis that must be modified for non-hole models is the nature of the small-$r$ solution. In particular, the horizon condition in (\[lateH\]), (\[utoinf\]), and (\[horlimit\]) must be replaced by the appropriate small-$r$ condition, and the static solution $\Psi_{\rm finstat}(r)$ is no longer the solution well behaved at $r\rightarrow 2M$, but rather the solution satisfying the appropriate small-$r$ condition (e.g., regularity at $r=0$). The result in (\[infrlim\]) remains unchanged.
The analysis of perturbations on a Reissner-Nordström (RN) background, of mass $M$ and charge $Q$, is very similar to that given above for the Schwarzschild background. For scalar perturbations or for all radiative degrees of freedom of the mixed electromagnetic- gravitational perturbations, the field equations can be put in the form of (\[wave\]), but with a potential of the form $$\label{RN}
V_l(r)=\frac{(r-r_+)(r-r_-)}{r^2}\left[
\frac{l(l+1)}{r^2}+R(r)\right]\ ,$$ where $r_+, r_-$ are the radii of the outer and inner horizons $$r_\pm=2M\left(1\pm\sqrt{1-(Q/M)^2}\right)\ ,$$ and where $R(r)$ falls off as $r^{-3}$. The tortoise coordinate $r_*$ for the RN background is the solution of $$\frac{dr_*}{dr} =\frac{r^2}{(r-r_+)(r-r_-)}$$ for $r_+<r<\infty$.
A review of the analysis of late time behavior for the Schwarzschild background confirms that almost all of the arguments depend on the general form of $V_l(r)$ at large $r$ and the exponentially sharp fall off of $V_l(r)$ at the horizon. In these features there seems to be no difference between the Schwarzschild and the RN problems. The difference in the relationship of $r$ and $r_*$ in the two cases must be carefully considered, however; it is this relationship \[not the details of $R(r)$\] that determines the initial backscattering and therefore the behavior of the late time tails. It turns out that there is no difference (for the tails) between the Schwarzschild and the RN cases. When the expression in (4) is substituted in (3) (with the RN form of $u,v$ and $V_l$) the result in (5) again emerges, and it is this result that determines the initial backscattering. The differences between the RN and Schwarzschild cases enter into the determination of $B_l^p(r)$ only to higher order in $r^{-1}$.
Numerical method
================
It is straightforward to integrate equation (\[wave\]) on a $uv$ null grid. The two-dimensional wave equation $-4\Psi_{,uv}=V_{l}(r) \Psi$ is most simply discretized as $$\Psi_N=\Psi_E+\Psi_W-\Psi_S-\Delta u\,\Delta v\,
V_l\left({v_N+v_W-u_N-u_E\over4}\right){\Psi_W+\Psi_E\over8} +O(h^4).$$ Here the points N, S, E and W form a null rectangle with relative positions indicated by the points of the compass, and $h$ is an overall grid scale factor, so that $$\Delta u=u_N-u_E=u_W-u_S=O(h),\qquad \Delta v=v_N-v_W=v_E-v_S=O(h).$$ Starting from null data on $u=u_0$ and $v=v_0$, integration proceeds to the northeast (increasing $v$) on one $u=const.$ line after another.
The error estimate for $\Psi_N$ follows directly if one assumes that the exact solution $\Psi(u,v)$ has a Taylor expansion in the given null rectangle. On a grid of fixed total size in $u$ and $v$ there are $O(h^{-2})$ grid points, so that the total error on the far end of the grid from the null data should scale as $h^2$ when the grid size is scaled by $h$. We have done numerical tests which have confirmed this $O(h^2)$ convergence.
The only difficulty in developing a code for this problem was the calculation of $r$ from $r_*=(v-u)/2$. We have done it by iteration of the defining equation, in the form $r=r_*-\ln(r-1)$ for large $r$, and in the form $r_1=\exp(r_*-1-r_1)$ for $r$ close to $1$, that is, close to the horizon, where $r_1\equiv r-1$. We have used a grid of constant $\Delta u=\Delta v$. Because of the scale-invariance of the problem we have set $2M=1$.
There is a small difference between the null data on $u=u_0$ and $v=v_0$ which we posed in the previous, analytical, section and our numerical work. For the numerical work, we posed [*constant*]{} data on $v=v_0$, and either generic (here Gaussian) or static data on $u=u_0$. The justification for this is the following. As we consider a linear wave equation, we can examine the evolution of data on $v=v_0$ and of data on $u=u_0$ separately. If we want to consider generic data on $v=v_0$, i. e. data which are localized, we can reflect them back at the scattering potential to get generic (although different data) on some $u=\tilde
u_0$. So if we are not very interested in the exact shape of our data, putting them on $u=u_0$ is sufficient.
Numerical results
=================
We now report on the results of our numerical simulations of test fields on a Schwarzschild or RN background. Without loss of generality we have set $2M=1$. As we are dealing with a linear wave equation, the overall amplitude of our initial data can also be chosen arbitrarily and is physically irrelevant. As our initial data null surfaces we chose $v=0$ for $u\ge0$ and $u=0$ for $v\ge 0$, which meet at the point $t=0$, $r_*=0$. (The background is of course $t$-independent.)
In a first series of numerical experiments, we chose $V_{l}(r)$ appropriate for a massless minimally coupled scalar field on a Schwarzschild background for different multipole indices $l$. It is of the form (\[wave\]), with $R(r)=2M/r^3$. As null data we used a Gaussian of width $3$ centered at $v=10$ on $u=0$. $\Psi$ is chosen to be constant on the null boundary $v=0$, with the constant determined by $\Psi(u=0,v=0)$. This is a simple approximation of the idea that the field is anchored in a star whose surface is rapidly redshifting, as explained in section II. We extended the grid from $v=0$ and $u=0$ to $u=400$ and $v=400$, with a typical value of $\Delta u=\Delta v=0.1$.
We examined the value of $\Psi$ as a function of $t$ on three lines, namely $r_*=10$, $u=400$ and $v=400$. We took these lines as finite approximations of the future timelike infinity $i_+$, of the future horizon ${\cal H}_+$ and future null infinity $scri_+$. Log-log plots of these three “tails" for $l=0$ are shown in Fig. 2. The predicted power-law behaviors are apparent for the case of an initial static field. On ${\cal H}_+$ and $i_{+}$ the field falls off as $t^{-3}$; on $scri_+$ the fall-off is as $t^{-2}$. The bend at the end of the $scri_+$ line is not surprising. It represents null infinity only for $v\gg u$, which is no longer the case at that end of the line. The large wiggles at the left are a remnant of quasinormal ringing.
Quasinormal ringing is shown more clearly in the linear plot, for $l=1$ scalar perturbations, shown in Fig. 3. The agreement with the theoretically calculated frequency [@Le] is good. From the plot we read off values of $0.56$ and $-0.19$ for the real and imaginary parts of the frequency, for a scalar field with $l=1$, for $2M=1$. The predicted values [@Le] are $0.58587$ and $-0.19532$ for the least damped modes.
In Fig. 4 we show the effect of the multipole index $l$ on the power law of the tail at $i_+$, again for a scalar field and Gaussian null data. The agreement of the power laws with the prediction is striking.
In a second series of numerical experiments we looked at the potential $V_l$ appropriate for an electromagnetic field on a Schwarzschild background, which is (\[wave\]) with $R(r)=0$. We took null data corresponding to the initial presence of a static $l$-pole moment. (These are given in power-series form in [@CuMoPr], and we numerically evaluated the first few lowest orders in $1/r$.) The tails at constant $r$ for $l=1$ and $l=2$ are shown in figure 5. (There is no electromagnetic monopole field.) It was shown in [@Pr] how electromagnetic and spin-2 fields can be encoded in a scalar field. The potential $V_l$ depends on the spin as well as on $l$, but the terms by which it differs for different spins are essentially Riemann tensor components and therefore of order $2M/r^3$. Terms of this order have been neglected in our analytic derivation, and the surprising prediction is therefore that the power law of the tails is independent of the spin of the testfield.
In a third series of experiments we examined a scalar test field on Reissner-Nordström backgrounds. The potential for a scalar field is now of the form (\[RN\]), with $R(r)=2Mr^{-3}-2Q^2r^{-4}$ As explained in section III one expects the tails to be independent of the charge of the black hole. Figure 6 shows that this is in fact so, with the example of the tails at the (outer) horizon of an $l=0$ test field on Reissner-Nordström backgrounds with charge/mass ratios of $Q^2/M^2=0.1$ and $0.9$ respectively. This figure constitutes also direct evidence for the existence of power-law tails on the (outer) horizon for a generic perturbations, thus underpinning a central condition for the mass inflation scenario.
Conclusions
===========
When the dust of an approximately spherical collapse has settled, and spacetime inside the future lightcone of the collapse has approached Schwarzschild, Minkowski, or a stellar interior spacetime, any massless fields that were present in the collapse still show “tails" that linger, decaying only as a power of time. In particular the $l$-th multipole moment of a massless test field decays like $t^{-(2l+P+1)}$ at fixed radius at large times, with $P=1$ if there is an $l$-pole moment present before the collapse and $P=2$ otherwise. But power-law tails are also present on $scri_{+}$, where they decay like $u^{-(l+P)}$, and if a black hole has formed, on the horizon, where they decay like $v^{-(2l+P+1)}$. The amplitude of the tails can be calculated as a function of the initial multipole moment, and in its absence, as an integral over the radiation going out to infinity during the collapse.
Neither this amplitude, nor the exponent of the power law, depend upon the spin of the massless field in question, nor, if the black hole is charged, on its charge. We have shown the origin of these features in an analysis which is based on [@Pr]. The spin- and charge-dependent parts of the effective radial potential of a massless test field on a Reissner-Nordström background are of an order that can be neglected in this analysis.
We have checked numerically that the tails are indeed present on null infinity and the horizon, and are independent of the spin of the field and the black hole charge. The fact that power-law tails are present on the outer horizon of a Reissner-Nordström black hole after a generic collapse situation is of crucial importance to the mass inflation scenario. It has, to our knowledge, never been demonstrated explicitly.
Finally, one decisive step of our analysis was a regularity condition, either on the horizon when a black hole formed in the collapse, or else at the center of spherical symmetry. This argument generalizes to any kind of boundary condition posed at small radius, and strongly suggests that perturbations of massless fields outside any spherical background object should also have power-law tails with the powers given above. In a subsequent paper we report results for a closely related [*nonlinear*]{} problem: the implosion of a shell of scalar field.
We wish to thank Josh Goldberg and Fritz Rohrlich for pointing out early references. This work was supported in part by grant NSF PHY92-07225 and by research funds of the University of Utah. J.P. acknowledges hospitality and support from the Institute for Theoretical Physics at UCSB and the National Science Foundation grant PHY89-04035.
B. S. DeWitt and R. W. Brehme, Ann. Phys. (NY) [**9**]{}, 220 (1960). F. Rohrlich and J. Winicour, in [*Perspectives in Geometry and Relativity, Essays in Honor of Václav Hlavatý*]{} (University of Indiana Press, 1966) (see correction in ref. [@Ro]). F. Rohrlich, [*Classical charged particles*]{} (Addison Wesley, Reading, 1990). R. H. Price, Phys. Rev. D [**5**]{}, 2419 (1972). T. M. Helliwell and D. A. Konkowski, Phys. Rev. D [**47**]{}, 4322 (1993) and references therein. C. T. Cunningham, V. Moncrief, and R. H. Price, Astrophys. J. [**224**]{}, 643 (1978). E. Leaver, Ph. D. Thesis, University of Utah (1985) (unpublished). T. Apostolatos, D. Kennefick, A. Ori, and E. Poisson, Phys. Rev. D [**47**]{}, 5376 (1993). K. Thorne, in [*Magic without Magic*]{}, J. R. Klauder, ed. (W. H. Freeman, San Francisco, 1972)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Lijun Zhang
- 'Jun-Wei Luo'
- 'A. L. Saraiva'
- Belita Koiller
- Alex Zunger
bibliography:
- 'ref\_vs.bib'
title: Genetic Design of Enhanced Valley Splitting towards a Spin Qubit in Silicon
---
[**Electronic spins in Silicon (Si) are rising contenders for qubits – the logical unit of quantum computation [@nielsen_quantum_2000] – owing to its outstanding spin coherence properties and compatibility to standard electronics [@morton_embracing_2011; @zwanenburg_silicon_2012]. A remarkable limitation for spin quantum computing in Si hosts is the orbital degeneracy of this material’s conduction band, preventing the spin-$\frac{1}{2}$ states from being an isolated two-level system [@koiller_exchange_2001; @zwanenburg_silicon_2012]. So far available samples of Si quantum wells cladded by Ge-Si alloy barriers provide relatively small valley splitting (VS), with the order of 1 meV or less [@nicholas_investigation_1980; @lai_valley_2006; @goswami_controllable_2007; @boykin_valley_2004; @friesen_valley_2007; @saraiva_physical_2009], degrading the fidelity of qubits encoded in spin “up” and “down” states in Si. Here, based on an atomically resolved pseudopotential theory, we demonstrate that ordered Ge-Si layered barriers confining a Si slab can be harnessed to enhance the VS in the active Si region by up to one order of magnitude compared to the random alloy barriers adopted so far. A biologically inspired genetic-algorithm search [@franceschetti_the_1999; @piquini_band_2008] is employed to identify magic Ge/Si layer sequences of the superlattice barriers that isolate the electron ground state in a single valley composition with VS as large as $\sim$9 meV. The enhanced VS is preserved with the reasonable inter-layer mixing between different species, and is interestingly “protected” even if some larger mixing occurs. Implementation of the optimized layer sequences of barriers, within reach of modern superlattice growth techniques [@menczigar_enhanced_1993], overcomes in a practical systematic way the main current limitations related to the orbital degeneracy, thus providing a roadmap for reliable spin-only quantum computing in Si.** ]{}
The qubits for quantum information processing are encoded in two-level quantum systems [$\{|0\rangle,|1\rangle\}$]{}, and can be realized, for example, by two spin states [$\{|\uparrow\rangle, |\downarrow\rangle\}$]{} of an electron at the conduction band edge of a semiconductor [@loss_quantum_1998; @kane_silicon_1998]. While Si enjoys a number of advantages over III-V semiconductors in this respect, including long spin coherence lifetime (associated with its weak spin-orbit coupling and small content of nonzero-nuclear-spin isotopes), as well as advanced fabrication know-how, its major drawback is the (six-fold) orbital degeneracy of its lowest conduction band (Fig. \[intro\]a) located close to the $X$ point in the Brillouin zone. This is no longer a two-level system determined solely by its spin, leading to considerable leakage and decoherence driven by the energetic proximity among the degenerate orbitals [@koiller_exchange_2001]. Whereas this six-fold valley degeneracy in the O$_h$-symmetric bulk Si can be partially removed by application of tensile biaxial strain [@schaffler_high_1997], thus, isolating the two lowest $|+z\rangle$ and $|-z\rangle$ components from the rest (Fig. \[intro\]b), the creation of a sufficiently large energy splitting within this $Z$-valley subspace (hereby called valley splitting (VS), see Fig. \[intro\]c) has proven to be a challenge for the experimental realization of Si-based spin qubits [@zwanenburg_silicon_2012].
The geometry of the basic physical system explored (Fig. \[intro\]d) includes a Si slab (“Well”) interfaced by a material with higher conduction band (“barrier”). The VS of this system depends on a multitude of degrees of freedom present in the realistic device growth. The Si well of thickness $d$ cladded by barrier materials of composition $X_b$ is coherently strained on a substrate with the planar lattice constant $a_s$ (determined by its composition $X_s$). We focus on the substrate and barrier composed of Ge-Si based materials, which provide better-quality interfaces than oxides. The barrier can be a Ge-Si random alloy of composition $X_b$ or any corresponding atomistically ordered structure. We incorporate monolithically the full system containing up to 10$^{5}$ atoms/computational-cell, via an atomistic pseudopotential Hamiltonian [@zunger_first_1996; @wang_local_1995], solved in a plane-wave basis for each relaxed atomic configuration, which gives directly the energies {$\varepsilon_i$} and wavefunctions {$\Psi_i$} of the valley states.
**Macroscopic degrees of freedom:** We start by exploring the continuum *configuration-averaged* degrees of freedom in this system, as common in the literature [@boykin_valley_2004; @nestoklon_spin_2006; @friesen_valley_2007; @valavanis_intervalley_2007; @boykin_valley_2008; @saraiva_physical_2009], finding that whereas they do not provide a clear avenue to major VS enhancement, their exploration hints at the importance of another length scale. We consider a fixed-thickness Si well embedded in the Ge-Si alloy barriers with varied composition $X_b$, on three substrates with different composition $X_s$. For each alloy composition $X_b$ of barriers, we calculated 20 randomly realized atomic configurations and the averaged VS is evaluated. The solid red line in Figs \[alloy\_sl\]a-c shows the calculated configuration-averaged VS as a function of composition $X_b$. Generally, one observes an uneventful monotonic increase of the averaged VS as the barrier becomes richer in Ge (see also Supplementary Fig. S1b, which shows the VS for a few distinct $X_b$). Such continuum-like effect of the configuration-averaged alloy barriers can be understood by the gradual change of the barrier height, *i.e.*, the band offset between the valley states of Si well and barrier (“b-Valley” in Fig. \[intro\]c and Supplementary Fig. S1a) [@saraiva_physical_2009; @saraiva_intervalley_2011]. Although the averaged VS (red lines in Figs \[alloy\_sl\]a-c) shows substantial dependence on the epitaxial strain (also see Supplementary Fig. S1b), the variation of the macroscopic barrier composition $X_b$ and substrate composition $X_s$ provides limited tuning of VS.
**Atomically resolved length scale:** Important clues emerge as to the significance of the atomically resolved length scale and symmetry, as indicated in a recent work on the intervalley splittings of PbSe [@poddubny_anomalous_2012]. In principle, the splitting within the $Z$ valleys is closely related to the interface-induced deviation from the O$_h$ symmetry of bulk Si (or the D$_{4h}$ symmetry of biaxially strained Si). For a Si quantum well (Fig. \[intro\]d), the interfacial perturbation potential $\Delta V$ with the D$_{2h}$/D$_{2d}$ symmetry provides a coupling channel between two $Z$-valley states, giving a VS magnitude in perturbation theory of $2|\langle+z|\Delta V|-z\rangle|$. To tune VS, we can engineer the magnitude and profile of the perturbation potential $\Delta V$ by varying the *atomic-scale structure and symmetry* for the well and barriers. The importance of the atomic scale is revealed, for example, in Fig. \[alloy\_sl\]a-c, where the blue circles represent the VS obtained by resolving distinct random realizations of site occupations in alloy barriers. The VS ranging from 0 to an upper bound of $\sim$1 meV is in reasonable agreement with experiments [@nicholas_investigation_1980; @lai_valley_2006; @goswami_controllable_2007; @zwanenburg_silicon_2012]. We can see that the VS of Si can vary significantly for different atomic configurations of barriers at the same composition $X_b$. This is consistent with the recent calculation showing that specific atomic arrangements at the interface region can result in distinct VS (however the assumed Si$_{3}$Ge luzonite structure is difficult for experimental realization) [@jiang_effects_2012]. Also, the critical role of atomic resolution and symmetry is apparent by considering a system of short-period Si-Ge superlattices located directly on a substrate (i.e, no active Si layer in Fig. \[intro\]d), where our calculated VS reaches values as large as several tens of meV, although the Si-Ge superlattice system is not the case of interest here (but may relate to different qubit proposals [@vrijen_electron_2000]).
Inspired by these basic insights from the atomic length scale, we next explore in a systematic way whether and how atomic degrees of freedom in the Si well, barrier composition and structure, and epitaxial substrate could raise the VS. By varying the above degrees of freedom, we aim at identifying the rule of how the relevant physical factors govern VS, and use it to seek an optimized VS.
**Effect of Si well thickness:** The thickness $d$ of the Si well is the first obvious parameter to tune the perturbation potential $\Delta V$, and thus manipulate VS. Fig. \[o\_e\]a shows the dependence of VS on the thickness $d$ in monolayers (MLs) for fixed pure Ge barrier from the pseudopotential calculations. We observe an overall decay in the magnitude of VS as the thickness $d$ increases, while the VS for $d$ with an odd (blue circles) and even (red squares) number of MLs appears to oscillate independently, with a common period $\sim$14 ML and a phase shift of $\pi$/2. This intriguing oscillatory behavior has been reported previously, and was attributed to the symmetry change of the Si well of $d$ MLs: D$_{2d} \leftrightarrow$ D$_{2h}$ for $d$ odd $\leftrightarrow$ even [@boykin_valley_2004; @nestoklon_spin_2006; @srinivasan_valley_2008].
In Fig. \[o\_e\]b, we show the calculated VS within the effective mass approximation (EMA) as a function of the continuum thickness $d_{con}$ (solid black line), as well as the data sampled at odd (blue circles) and even (red squares) atomic MLs. We find that while the EMA results with continuum $d_{con}$ show a much faster oscillation, clearly they reproduce well the existence of “independent” oscillation for discrete $d$ of odd and even MLs. Thus, we attribute this atomic-scale odd-even independent oscillation to a manifestation of the [*aliasing effect*]{} (introduced by sampling a function at a rate which is not fine enough to capture each oscillation), rather than to a symmetry change (see Supplementary Information for detailed description). This understanding underlines that to gain an optimized VS of Si well, a well-controlled growth of monolayer precision is required to reach the thickness $d$ at the peak of the oscillation.
**Atomically ordered superlattices barriers:** The substantial effect of specific atomic realization for the disordered alloy barriers (open circles in Figs \[alloy\_sl\]a-c) stimulates us to investigate the situation where the barriers are composed of ordered superlattices, *i.e.*, a repeated sequence of Si and Ge layers of arbitrarily assigned widths. We explore the system composed of a 40 MLs Si well (located at an even peak of Fig. \[o\_e\]a) embedded in the superlattice barriers with a period of 80 MLs (with the minimum stacking unit of bilayer to comply with current experimental growth conditions). This gives an astronomical number (2$^{40}$) of candidate layer-stacking configurations of barriers, so a direct calculation for enumeration of all the candidates is not practical. We perform an *inverse-band-structure search* calculation [@franceschetti_the_1999; @piquini_band_2008] where the best fitness is defined by the maximum VS, and favorable structures are selected within a genetic algorithm approach. Fig. \[ibs\]a shows the evolution of fitness (VS) with generation (evolution step). One clearly observes that the VS can be effectively tuned within a wide energy range, from negligibly small up to $\sim$9 meV, by varying Ge/Si stacking sequence of superlattice barriers. Less than 100 generations already identify the best individuals, which remain superior for the following hundreds of generations, while new individuals still emerge with intermediate VS values.
Figs \[alloy\_sl\]d-f show the achieved VS of all the atomic configurations visited by the inverse-band-structure search, sorted in terms of the Ge content in the barriers on three varied substrates. It is demonstrated that a remarkable VS enhancement by a factor of 5-10 is achievable with ordered superlattice barriers as compared to disordered alloy barriers (Figs \[alloy\_sl\]a-c) for all substrates. Particularly, comparing with the maximum VS for the disordered alloy barriers – $\sim$1.0 meV on all the substrates, the maximum VS (accompanied by the optimum configuration) for the ordered superlattice barriers reaches:
$\%$ 0 Ge substrate: 5.7 meV (Ge$_4$Si$_4$Ge$_2$Si$_6$Ge$_4$Si$_4$Ge$_4$Si$_2\ldots$);\
$\%$20 Ge substrate: 7.4 meV (Ge$_4$Si$_4$Ge$_4$Si$_2$Ge$_4$Si$_6$Ge$_4$Si$_2\ldots$);\
$\%$40 Ge substrate: 8.7 meV (Ge$_4$Si$_6$Ge$_2$Si$_6$Ge$_4$Si$_4$Ge$_4$Si$_4\ldots$).\
We find that the multilayer superlattice barriers show larger VS around the central region, *i.e.*, at 40-60$\%$ Ge content in the barrier, different from higher Ge content leading to larger VS for random alloy barriers. The same Ge content in the superlattice barriers can lead to both high and low VS extremes, again emphasizing the key role of atomistic scale ordering in controlling VS.
**The Si/Ge$_4$ motif:** Interestingly, all the optimum configurations identified start the barrier sequence by a Ge$_{4}$ sub-layer. This same“magic” thickness for the first Ge sub-layer is also identified in the exhaustive enumeration calculations for the superlattice barriers with a shorter period of 16 MLs (see Supplementary Fig. S2a-c). Similar results are obtained for a Si well with the thickness of 47 MLs (located at an odd peak of Fig. \[o\_e\]a, see Supplementary Fig. S2d-f). In order to better understand this, we explore a simple case – the fixed 40 MLs Si well embedded in Ge$_n$Si$_n$ superlattice barriers with $n$ = 1, 2, 4, 8, 16, as shown in Fig. \[wav\]a. We see that the barrier of Ge$_4$Si$_4$ superlattice indeed exhibits the largest VS ($>$7 meV), whereas all other barriers (including pure Ge) show typically low VS ($< 2$ meV). This indicates that the starting sub-layer thinner or thicker than Ge$_4$ seem to equally suppress VS. We unravel the underlying origin within the EMA context. Briefly, the VS induced by a Si/Ge (ascending offset) interface has opposite sign to the Ge/Si (descending offset) interface with the same wavefunction. Choosing the interface positions to match the maxima/minima of the VS at the ascending/descending interfaces would maximize the total VS. It is impossible to match the interface positions perfectly to the incommensurate oscillations of well-thickness dependent VS (Fig. \[o\_e\]), but the Ge$_4$ sub-layer is the closest we can get to this matching. Conversely, starting with a Ge$_2$ sub-layer cladding the Si well, we find a destructive interference, in agreement with the suppressed VS for Si$_2$Ge$_2$ superlattice barrier in Fig. \[wav\]a. This engineering is analogous to that of a *distributed Bragg reflector* (see Supplementary Information for detailed description). But the fact that the oscillations are incommensurate with the lattice and the strong dependence of VS on atomic ordering makes it impossible to analytically predict the optimal structure. For this reason, the genetic selection of candidate structures is an essential ingredient of this work.
Previous studies correlate the VS with the electronic wave-function magnitude of the $Z$ valley at the interface (interfacial $|\Psi|^2$) [@nestoklon_spin_2006; @friesen_valley_2007; @saraiva_physical_2009; @saraiva_intervalley_2011] and the wave-function penetration into the barrier region (penetration of $|\Psi|^2$) [@saraiva_physical_2009; @saraiva_intervalley_2011]. In Fig. \[wav\]b and \[wav\]c, we probed the VS as a function of interfacial $|\Psi|^2$ and penetration of $|\Psi|^2$, for the 40 MLs Si well cladded by alloys barriers (blue crosses) and superlattice barriers (red pluses). Compared with the alloy barriers, the stronger confinement power of superlattice barriers give much narrower distribution of both interfacial $|\Psi|^2$ and penetration of $|\Psi|^2$. The optimum VS values for the superlattice barriers emerge in the region of the narrowest distribution of these two quantities. This is related to the sharp well/barrier interface for superlattice barriers, which could in principle enhance the VS [@saraiva_physical_2009; @culcer_interface_2010].
**Effect of Ge-Si intermixing in barriers:** Since it is still a challenge to grow perfectly pure sub-layer of Si or Ge in superlattices due to atomic inter-diffusion [@menczigar_enhanced_1993], we examine how much VS is affected by the *interfacial mixing* between Si and Ge. In particular, the inter-layer mixing is modeled by mapping pure Si into Si$_{1-\eta}$Ge$_\eta$, and pure Ge into Ge$_{1-\eta}$Si$_\eta$ at the interfacial first few layers, determined by a mixing length. The parameter $\eta$ quantifies the degree of inter-layer mixing, with $\eta$ = 0 corresponding to no mixing and $\eta$ = 0.5 meaning maximum mixing, *i.e.*, complete destruction of Si-rich or Ge-rich pattern within this layer. Fig. \[ibs\]b shows the calculated VS as a function of $\eta$ for the above optimized superlattice barrier on $\%$40 Ge substrate (Ge$_4$Si$_6$Ge$_2$Si$_6$Ge$_4$Si$_4$Ge$_4$Si$_4\ldots$, see Supplementary Fig. S3 for more ordered superlattice barriers), when two cases of mixing lengths \[2 MLs (green) and 4 MLs (red)\] are explored. Note that the favorable Ge$_4$ starting sub-layer is only partially damaged if the mixing length is 2 ML (1 ML towards each side of the interface), while for the 4ML mixing length the Ge-pure layer is totally destroyed. The non-trivial, non-monotonic behavior indicates that the intermixing may lead to the formation of a more complex geometry which tunes VS by affecting the interference pattern discussed before. This is reflected in a surprisingly steeper suppression of VS in the shorter mixing length of 2 MLs compared to the longer mixing length of 4 MLs for small $\eta$. Similarly, at very large $\eta$, the structure becomes a complex layering of alloys, pure Si and pure Ge, which might keep suppressing (the case of 4ML) or invert the symmetry and enhance the VS (the case of 2ML). In both mixing lengths, for a reasonable degree of mixing ($\eta <$ 0.1), the rather high VS of $>$6 meV is preserved.
**Advantage of atomically ordered barriers:** We anticipate that the choice of ordered superlattice barriers instead of random alloy barriers might mitigate many problems of real samples. For instance, the intrinsic non-deterministic nature of alloys induces disorder ranging from the geometry of the interface plane to the inhomogeneous strain fields [@evans_nanoscale_2012]. The leakage of electrons tunneling through the superlattice barrier should also be reduced since the electronic density inside the barrier is much reduced. Our atomically-resolved pseudopotential calculations of the Si well strained on various substrates and interfaced with different barriers can be used to explore the effects on VS of both global macroscopic quantities (strain, alloy compositions, geometric well thickness), as well as atomic scale effects (ordering, inter-layer mixing, even-odd independent oscillation). We identify the critical Si well thicknesses as well as an emerging “magic” motif of $Ge_4$ starting in the ordered superlattice barrier that causes strong coupling between $|+z\rangle$ and $|-z\rangle$ valley states, leading to significantly enhanced VS as large as $\sim$9 meV. The predicted structure is accessible within current experimental fabrication capabilities. This opens the way to fundamental understanding of the hitherto rather intangible $Z$-valley splitting in indirect-gap semiconductors such as Si with the possible benefit of isolating single electron valley state for spin-based quantum computing.
**Methods**
The structures employed to optimize VS in this work (Fig. \[intro\]d) involve an active Si well with the thickness of $d$ MLs, cladded on both sides by a barrier consisting of Si-Ge based materials, including homogeneous random alloy and layer-by-layer superlattice structures. The whole system is coherently strained on a substrate, via minimization of atomically-resolved strain with a generalized valence force field method (see Supplementary Methods A) [@bernard_strain_1991]. To comply with what is currently accessible in experimental growth we used two restrictions: (i) Since too high Ge content in substrate is known to cause dislocations in thick Si active layers to relieve excessive strain, up to 40$\%$ Ge content is considered in substrate. (ii) a bilayer is used as the minimum stacking unit of each specie (Si/Ge) for the superlattice barrier.
The energies and wave-functions of conduction valley states for candidate structures are calculated “on the fly” with the atomistic pseudopotential method, described in detail in Refs. [@wang_local_1995; @zunger_first_1996]. The atomistic pseudopotential method (overcoming the well-known Density-Functional-Theory limitations on electronic structure calculations), accompanied with a plane-wave basis set and folded-spectrum diagonalization, allow us to accurately calculate energy splitting of $Z$-valley states (at the order of meV or lower) for numerous candidate structures with economic efficiency, as described in Supplementary Methods B.
Effective mass approach calculations were performed to accompany the interpretation of the pseudopotential results. The effective mass calculations follow essentially the model presented in Refs. [@saraiva_physical_2009; @saraiva_intervalley_2011], adapted to describe quantum wells in first order approximation, as described briefly in Supplementary Methods C.
For Si wells embedded in layer-by-layer superlattice barriers, since the search space shows a combinatorial burst of degrees of freedom, we employ the developed inverse-band-structure approach [@franceschetti_the_1999; @piquini_band_2008; @davezac_genetic_2012], *i.e.*, a biologically inspired (Darwinian) genetic algorithm to guide the electronic structure calculations, with the aim at finding the optimum configuration that gives the maximum VS (Supplementary Methods D).
**Acknowledgements**
We thank M. A. Eriksson and M d’Avezac for helpful discussions. This work is supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Contract No. DE-AC36-08GO28308 to NREL. The “Center for Inverse Design” is a DOE Energy Frontier Research Center. A.S. and B.K.’s work is part of the Brazilian National Institute for Science and Technology on Quantum Information. A.S. and B.K. acknowledge partial support from FAPERJ, CNPq and CAPES.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Jahn-Teller systems and the Jahn-Teller effect are discussed in terms of cavity QED models. By expressing the field modes in a quadrature representation, it is shown that certain setups of a two-level system interacting with a bimodal cavity are described by the Jahn-Teller $E\!\times\!\varepsilon$ Hamiltonian. We identify the corresponding adiabatic potential surfaces and the conical intersection. The effects of a non-zero geometrical Berry phase, governed by encircling the conical intersection, are studied in detail both theoretically and numerically. The numerical analysis is carried out by applying a wave packet propagation method, more commonly used in molecul or chemical physics, and analytic expressions for the characteristic time scales are presented. It is found that the collapse-revival structure is greatly influenced by the geometrical phase and as a consequence, the field intensities contain direct information about this phase. We also mention the link between the Jahn-Teller effect and the Dicke phase transition in cavity QED.'
author:
- Jonas Larson
title: 'Jahn-Teller systems from a cavity QED perspective'
---
Introduction
============
The [*Jahn-Teller*]{} (JT) effect, due to Hermann Jahn and Edward Teller [@JTorig], states that a symmetry-breaking is likely (only exceptions are linear molecules or molecules possessing [*Kramers degeneracy points*]{} [@JTeff2]) to occur whenever there is an isolated degeneracy of electronic states in a molecule, a so called [*conical intersection*]{} (CI) [@ci; @cibook]. Over the years, the JT effect has gained enormous attention, mainly in molecular and condensed matter physics [@JTeff2; @JTeff1; @JTcrystal]. A simple model system Hamiltonian possessing a CI, later termed $E\!\times\!\varepsilon$, was presented by H. C. Longuet-Higgens [*et al.*]{} [@longuet-higgens]. The main result of this work states that the angular momentum quantum number is half integer valued rather than an integer. This phenomenon arises from a [*geometric phase*]{}, on top of the dynamical one, obtained while encircling the CI. The additional phase must be introduced in order to have a single valued total (electronic and vibrational) wave function. This was further analyzed in [@mt], where it was shown that the double valuedness of the electronic wave function can indeed be removed by introducing a “vector potential" term in the Hamiltonian. For CI models, this resembles the [*Aharanov-Bohm effect*]{} [@ab], and gave rise to the [*molecular Aharanov-Bohm effect*]{} and [*molecular gauge theory*]{} [@berrymol1; @berrymol2]. A deeper understanding of this phase effect was gained with the seminal paper of M. V. Berry [@berry], which presents a general formalism for the geometrical phase factors that an adiabatic change in the Hamiltonian brings about. In the spring from this work came several papers on the geometrical phase related to CIs, see Refs. [@berrymol1; @berrymol2].
The effect of the geometrical phase on physical observables has since then been discussed and experimentally verified in several reports [@faseffects; @singlesurf]. The modulation, caused by the geometrical phase, of the wave function has been addressed in Ref. [@koppel], where a dynamical wave packet approach, like the one used in this article, is applied to the $E\!\times\!\varepsilon$ JT model. Inclucion of spin-orbit couplings has been investigated in terms of the JT effect [@JTso] and of wave packets [@spinorbit2]. Recently, other properties of JT models, not only the $E\!\times\!\varepsilon$, have been considered, for example, quantum chaos [@chaos] and ground state entanglement [@ent; @magnetic2].
Although the JT effect has not, to the best of my knowledge, been discussed within [*cavity quantum electrodynamics*]{} (QED), the geometrical Berry phase has been analyzed in the framework of cavity QED [@berrycav; @vedralham; @berrycav2]. These references study the effects induced by the vacuum field on the geometrical phase. In other words, the treatment of the two-level particle in the time-varying field is then considered on a fully quantum mechanical footing. The degeneracy point is not, however, identified as a CI in these works, and the [*rotating wave approximation*]{} (RWA) has been applied which is not the case in the present article. In addition, in the situation studied here, the geometrical phase is said to be of [*dynamical character*]{} as it originates from the intrinsic evolution of the system [@berrymol2] rather than from “external" changing of the Hamiltonian [@berrycav; @vedralham; @berrycav2]. Thus, the circumstances and approaches are notably different between this work and the ones of Refs. [@berrycav; @vedralham; @berrycav2]. Among others, we examine the cavity QED system in a [*conjugate representation*]{} in which the intracavity fields are expressed in their quadrature operators rather than the standard used creation and annihilation ladder operators. In this picture, the link to JT systems and to CIs is revealed and, in fact, the [*Dicke normal-superradiant phase transition*]{} in cavity QED is seen to be related to the JT effect.
The model system is a two-level quantum-dot embedded in a cavity and interacting with two degenerate field modes. The numerical analysis is carried out using wave packet propagation methods; an initial state of the system is let to evolve under the corresponding Hamiltonian. The full Hamiltonian dynamics is considered and compared to results obtained from a second Hamiltonian which shares the same [*adibatic potential surfaces*]{} (APS) but lacks a geometrical phase. We give analytical expressions for the characteristic time scales, collapse and revivals, for both systems and it is found that the revival time is in a sense twice as long in the case where the geometrical phase is excluded. A consequence of this is reflected in the intracavity field intensities. Thus, measurements of the field intensities of the two modes give a direct indication of the geometrical phase.
The outline of the paper is as follows. In Sec. \[sec2\] we review some basics of JT models and CIs. Especially, in Subsec. \[subsec2a\] we introduce the $E\!\times\!\varepsilon$ Hamiltonian and discuss its APSs, while Subsec. \[subsec2b\] derives the geometrical phase accumulated by encircling the CI, and Subsec. \[subsec2c\] considers the JT effect in general and in the $E\!\times\!\varepsilon$ Hamiltonian in particular. The next Sec. \[sec3\] is devoted to our cavity QED model and it is shown how the $E\!\times\!\varepsilon$ model occurs for a two-level system interacting with degenerate bimodal fields. A discussion of the corresponding JT effect in cavity QED is outlined in Subsec. \[subsec3b\], where a parallel with the Dicke normal-superradiant phase transition is drawn. Our numerical results of the cavity JT system are presented in Sec. \[sec4\], both analyzing the short and long term behavior and how the geometrical phase comes into play. Finally we summarize in Sec. \[sec5\].
The Jahn-Teller model {#sec2}
=====================
Jahn-Teller systems are characterized by a degeneracy point of coupled potential surfaces, a CI. In one dimension, the simplest example is the $E\!\times\!\beta$ model, also called [*Rabi*]{} or [*spin-boson*]{} model [@oldwine]. It describs a spin 1/2 particle coupled to a single boson mode [@eb]. In certain parameter regimes, a RWA can be applied in which this model relaxes to the one of Jaynes and Cummings [@jc]. In one dimension, the wave packet (state of the system) cannot encircle the CI without passing through it, and therefore there is no corresponding dynamical geometric phase [@berry]. In two dimensions, generalization of the $E\!\times\!\beta$ model leads in certain situations to the $E\!\times\!\varepsilon$ model which will be the subject of this section.
The $E\!\times\!\varepsilon$ Hamiltonian {#subsec2a}
----------------------------------------
The simplest Jahn-Teller Hamiltonian with two vibrational degrees of freedom is the so called $E\!\times\!\varepsilon$ one, given by [@longuet-higgens] $$\label{jtham}
\begin{array}{lll}
H_{JT} & = &
\displaystyle{\frac{\hat{p}_x^2}{2m}+\frac{\hat{p}_y^2}{2m}+\frac{m\omega^2}{2}(\hat{x}^2+\hat{y}^2)}
\\ \\
& &
\displaystyle{+\lambda\hat{x}\hat{\sigma}_x+\lambda\hat{y}\hat{\sigma}_y}.
\end{array}$$ Here $\hat{p}_i$ and $\hat{x}_i$ are momentum and position in the $i$ direction of the “particle" with mass $m$. The $\hat{\sigma}$-operators are the standard Pauli matrices obeying the commutation relations $$[\hat{\sigma}_i,\hat{\sigma}_j]=2\varepsilon_{ijk}\hat{\sigma}_k$$ and with the $z$-eigenstates, $\hat{\sigma}_z|\pm\rangle=\pm|\pm\rangle$ and $\lambda$ the coupling constant. Clearly, at the origin $\hat{x}=\hat{y}=0$, the two potential surfaces are degenerate. In the presence of either spin-orbit coupling [@spinorbit1; @spinorbit2] or an external magnetic field [@magnetic1; @magnetic2], an additional [*detuning*]{} term $\Delta\hat{\sigma}_z/2$ is added to the Hamiltonian, where $\Delta$ is the spin-orbit splitting or the magnetic strength. With this term present, the degeneracy is lifted and the intersection becomes avoided.
The form of the Hamiltonian (\[jtham\]) defines the [*diabatic basis*]{} and [*diabatic potentials*]{}, namely; a diabatic state is written as $\Psi(x,y)=f_\pm(x,y)|\pm\rangle$ for some normalized function $f_\pm(x,y)$, and the diabatic potentials, once the detuning $\Delta$ is included, are $m\omega^2(x^2+y^2)/2\pm\Delta/2$. Before defining the APSs, we express the Hamiltonian in polar coordinates $$x\pm iy=\rho\mathrm{e}^{\pm i\varphi}$$ giving [@koppel; @spinorbit2; @spinorbit1] $$\label{jtpolar}
\begin{array}{lll}
H_{JT} & = &
\displaystyle{-\frac{\hbar^2}{2m}\left(\frac{\partial^2}{\partial\hat{\rho}^2}+
\frac{1}{\hat{\rho}}\frac{\partial}{\partial\hat{\rho}}+
\frac{1}{\hat{\rho}^2}\frac{\partial^2}{\partial\hat{\varphi}^2}\right)}
\\ \\
& &
\displaystyle{+\frac{m\omega^2}{2}\hat{\rho}^2+\left[\begin{array}{cc}\displaystyle{\frac{\Delta}{2}}
& \lambda\hat{\rho}\mathrm{e}^{i\hat{\varphi}} \\
\lambda\hat{\rho}\mathrm{e}^{-i\hat{\varphi}} &
-\displaystyle{\frac{\Delta}{2}}\end{array}\right]}
\end{array}$$
Let us introduce the unitary operator [@koppel; @mt] $$\label{adstate}
U=\left[\begin{array}{cc} \sin(\nu)
& \cos(\nu) \\
-\cos(\nu)\mathrm{e}^{i\mu} & \sin(\nu)\mathrm{e}^{i\mu}
\end{array}\right],$$ where $$\tan(2\nu)=\frac{2\lambda\hat{\rho}}{\Delta}$$ and $\mu=\hat{\varphi}$, which diagonalizes the last term of Eq. (\[jtpolar\]). However, $U$ does not commute with the kinetic term in Eq. (\[jtpolar\]) and consequently, the transformed Hamiltonian, $\tilde{H}_{JT}=U^{-1}H_{JT}U$, is non-diagonal. The off-diagonal terms are the non-adiabatic couplings, which usually are small far from the crossing. Omitting these terms defines the [*adiabatic Hamiltonian*]{} $$\label{jtpolarad}
H_{JT}^{ad}=T+V_\pm^{ad}+V_{cent}+V_{gauge},$$ where [@koppel; @mt] $$\label{jtpolaradterms}
\begin{array}{l}
T=\displaystyle{-\frac{\hbar^2}{2m}\left(\frac{\partial^2}{\partial\hat{\rho}^2}+
\frac{1}{\hat{\rho}}\frac{\partial}{\partial\hat{\rho}}+\frac{1}{\hat{\rho}^2}\frac{\partial^2}{\partial\hat{\varphi}^2}\right),}
\\ \\
V_\pm^{ad}=\displaystyle{+\frac{m\omega^2}{2}\hat{\rho}^2+\hat{\sigma}_z\sqrt{\left(\!\frac{\Delta}{2}\right)^2+\lambda^2\hat{\rho}^2},}
\\ \\ V_{cent}=\displaystyle{\frac{\lambda^2\Delta^2\omega}{2\left(\Delta^2+4\lambda^2\hat{\rho}^2\right)^2},}
\\ \\
V_{gauge}=\displaystyle{\frac{\omega}{2\hat{\rho}^2}\left(1+\frac{\Delta}{\sqrt{\Delta^2+4\lambda^2\hat{\rho}^2}}\right)\left(\frac{1}{2}-\frac{\partial}{\partial\hat{\varphi}}\right)}.
\end{array}$$ We will assume that the evolution takes place mainly upon the lower APS, [*i.e.*]{}, $\hat{\sigma}_z=-1$, but in the numerical simulations we consider full dynamics without any approximations. Close to the crossing, the non-adiabatic couplings may have a significant impact on the dynamics [@cibook; @bobreak]. The term $V_\pm^{ad}$ defines the APSs, while $V_{cent}$ and $V_{gauge}$, arising from the commutator between $U$ and the kinetic energy operator, are [*centrifugal corrections*]{} [@centcorr]. The last term has been labeled [*gauge*]{} for reasons that will become clear later on. We display two examples, $\Delta=0$ and $\Delta\neq0$, of the APSs $V_\pm^{ad}$ in Fig. \[fig1\]. For a non-zero detuning, as pointed out, the crossing at the CI becomes avoided, with splitting amplitude $\Delta$. The lower surface has the familiar sombrero shape, while the upper possesses a single global minimum at $x=y=0$. For large detunings $\Delta$, the Mexican hat structure is lost, and the minimum of the lower APS is at the origin. Especially, the radius giving the potential minima is given by [@spinorbit2] $$\label{potmin}
\rho_{min}=\left\{\begin{array}{lll}
\displaystyle{\sqrt{\left(\frac{\lambda}{\omega}\right)^2-\left(\frac{\Delta}{\lambda}\right)^2}},
& \hspace{1cm} & \omega|\Delta|<\lambda^2 \\ \\
0, & \hspace{1cm} & \omega|\Delta|\ge\lambda^2
\end{array}\right..$$
![Two examples of the adiabatic surfaces in the $E\!\times\!\varepsilon$ Jahn-Teller model with detuning $\Delta=0$ (a) and $\Delta\neq0$ (b). []{data-label="fig1"}](fig1.eps){width="8cm"}
Berrys geometrical phase in the $E\!\times\!\varepsilon$ model {#subsec2b}
--------------------------------------------------------------
The [*adiabatic states*]{}, defined by $U$, are arbitrary up to an overall phase. The phase choice in Eq. (\[adstate\]) is chosen such that the states are singled valued as $\varphi$ is varied by $2\pi$. For example, the alternative obtained by multiplying $U$ by $\exp\left(-i\varphi/2\right)$ implies double-valued adiabatic states. Unitary transformation of the Hamiltonian in this second case leads to an adiabatic Hamiltonian lacking the term proportional to $i\partial/\partial\hat{\varphi}$. The very last term of (\[jtpolarad\]), containing the differential operator $\partial/\partial\hat{\varphi}$, can be viewed as a vector potential. Indeed, this term can be combined with the canonical momenta to define a kinetic momenta. Thus, the two options of overall phase of the adiabatic states given above result in either single-valued states with a vector potential present in the Hamiltonian or no such vector potential term but double-valued states [@vector]. The source of a vector potential term is the cause for having a non-zero geometric phase as the system encircles the CI in analogous to the Aharanov-Bohm effect [@spinorbit2].
For the system evolving along a closed loop $C$ in parameter space, the geometrical phase can be calculated according to [@berry] $$\gamma_n(C)=\oint_C\langle n({\bf R})|\nabla_{{\bf R}}n({\bf
R})\rangle\cdot d{\bf R}.$$ Here, $|n({\bf R})\rangle$ is the $n$’th adiabatic eigenstate and ${\bf
R}$ the set of parameters. In particular, in our case $|n({\bf
R})\rangle=\left(\sin(\theta),-\cos(\theta)\mathrm{e}^{i\varphi}\right)$ and as we consider a time-independent dynamical problem, the varying parameters ${\bf R}$ are the coordinates $\rho$ and $\varphi$. Especially, we consider a wave packet located at the minima of the sombrero potential, such that $\langle\hat{\rho}\rangle\approx\rho_{min}$, and $\varphi$ is changed from 0 to $2\pi$. For a general radius $R$ we find the $E\!\times\!\varepsilon$ geometric phase [@spinorbit2] $$\gamma_{JT}(R)=-\pi\left(1+\frac{\Delta}{\sqrt{\Delta^2+\lambda^2R^2}}\right),$$ which for $R=\rho_{min}$ becomes $$\label{berryfas}
\gamma_{JT}(\rho_{min})=-\pi\left(1+\frac{\omega\Delta}{\lambda^2}\right).$$ For $\Delta=0$ we obtain the well known sign change of the wave function when encircling a CI, causing half integer angular momentum quantum numbers.
The Jahn-Teller effect {#subsec2c}
----------------------
Using group theoretical arguments, Jahn and Teller proved that for almost any degeneracy (CIs) among electronic states in a molecule, a symmetry-breaking is “allowed" which removes the degeneracy and lowers the total energy of the model system ground state [@JTorig]. It turns out that this symmetry-breaking indeed takes place in the majority of cases with some exceptions [@JTeff2; @JTeff1]. Hence, the molecule favors a distortion of its most symmetrical state. This effect is quenched when spin-orbit coupling is taken into account, but may still exist [@JTso]. Returning to the APSs of the $E\!\times\!\varepsilon$ Hamiltonian (\[jtpolarad\]), the state with highest symmetrical is a wave packet centered around the origin ($x=y=0$). Loosely speaking, for vanishing detuning $\Delta$, is it intuitive to expect the wave packet to slide down the potential surfaces towards the minima of the sombrero. Semi-classically, the wave packet experiences a non-zero force $F=-\nabla V(x,y)$. For large detunings however, we saw from Eq. (\[potmin\]) that the potential surfaces may possess a single global minimum which will prevent the JT-distortion. On the other hand, for small but non-zero detuning, the sombrero structure is present for the lower APS and here quantum fluctuations will permit a symmetry breaking. Naturally, the above arguments are semi-classical and the full evolution is quantum mechanical and described by the coupled system. Nonetheless, it gives some insight and intuition of the dynamics.
Jahn-Teller models in cavity QED {#sec3}
================================
Spin boson models naturally occur in cavity QED. Here the boson subsystem represents a single or several quantized modes of an intracavity field, while the spin degrees of freedom describes either two-level atom/atoms [@haroche] or solid state quantum-dot/dots [@qdot]. Contrary to standard formulations of Jahn-Teller models, here the bosons are the photons of the field rather than vibrational phonons, and the internal structure corresponds to two discrete energy levels of the atom or quantum-dot. In the single mode case, a microscopic derivation gives the $E\!\times\!\beta$ Hamiltonian [@scully] in the assumption of [*dipole approximation*]{} and neglecting the [*self-energy*]{} (see below). In most cavity QED experiments involving atoms, the application of the RWA is justified, in which the Hamiltonian identifies the analytically solvable Jaynes-Cummings one [@jc]. The APSs (or rather adiabatic potential curves) of the Jaynes-Cummings model contain the differential momentum operator [@oldwine], and they are therefore said to be of [*non-potential form*]{}. Nonetheless, even if the picture of potential surfaces is less intuitive due to the momentum dependence, the JC model renders a sort of generalized CI (curve crossing). To go beyond the RWA regime, the coupling to the field must be substantially increased compared to atomic cavity QED setups. This is indeed the case for solid state quantum-dots coupled to a cavity. In fact, the crucial parameter, coupling divided by the two-level transition frequency, can be made several orders of magnitude larger in condensed matter systems [@qdot] compared to atom-cavity ones. Another possibility to achieve ultrastrong atom-field couplings is to consider Bose-Einstein condensates coupled to an intracavity field [@bec].
The model system Hamiltonian {#subsec3a}
----------------------------
To obtain CIs rather than curve crossings, multi-mode cavities must be considered [@twomode]. For simplicity we will assume two degenerate cavity modes such that they share the same frequency $\tilde{\omega}$ and also same coupling amplitude $\tilde{\lambda}$ to the quantum-dot. The Hamiltonian in the dipole approximation reads [@hamref; @hamref2] $$\label{hamcav1}
\begin{array}{lll}
H_{cav} & = &
\displaystyle{\hbar\tilde{\omega}\left(\hat{a}^\dagger\hat{a}+\hat{b}^\dagger\hat{b}\right)+\hbar\frac{\tilde{\Omega}}{2}\hat{\sigma}_z}\\
\\
& & \displaystyle{+\hbar
\frac{\tilde{\lambda}}{\sqrt{2}}\Big[\left(\hat{a}^\dagger+\hat{a}\right)\left(\hat{\sigma}^+\mathrm{e}^{-i\phi}+\hat{\sigma}^-\mathrm{e}^{i\phi}\right)}\\
\\
& & \displaystyle{+
\left(\hat{b}^\dagger+\hat{b}\right)\left(\hat{\sigma}^+\mathrm{e}^{-i\theta}+\hat{\sigma}^-\mathrm{e}^{i\theta}\right)\Big]}.
\end{array}$$ Here $\hat{a}^\dagger$ and $\hat{b}^\dagger$ ($\hat{a}$ and $\hat{b}$) are creation (annihilation) operators for the two field modes, $\phi$ and $\theta$ field phases, $\tilde{\Omega}$ the quantum-dot transition frequency, and $2\hat{\sigma}^\pm=\hat{\sigma}_x\pm
i\hat{\sigma}_y$. In the following we will label the two cavity modes by $a$ and $b$. Before proceeding, for brevity we introduce a characteristic energy $\hbar\tilde{\omega}$ and time scale $\tilde{\omega}^{-1}$, such that we consider dimensionless variables $$\begin{array}{lllll}
\displaystyle{\lambda=\frac{\tilde{\lambda}}{\tilde{\omega}}}, & &
\displaystyle{\Omega=\frac{\tilde{\Omega}}{\tilde{\omega}}}, & &
\tau=\tilde{\omega}t,
\end{array}$$ where $t$ is the unscaled time. In a [*conjugate variable representation*]{} defined by the operator relations $$\label{convar}
\begin{array}{c}
\displaystyle{\hat{p}_x=i\frac{1}{\sqrt{2}}\left(\hat{a}^\dagger-\hat{a}\right)},\hspace{1cm}
\displaystyle{\hat{x}=\frac{1}{\sqrt{2}}\left(\hat{a}^\dagger+\hat{a}\right)},\\
\\
\displaystyle{\hat{p}_y=i\frac{1}{\sqrt{2}}\left(\hat{b}^\dagger-\hat{b}\right)},\hspace{1cm}
\displaystyle{\hat{y}=\frac{1}{\sqrt{2}}\left(\hat{b}^\dagger+\hat{b}\right)},
\end{array}$$ where $[\hat{x},\hat{p}_x]=[\hat{y},\hat{p}_y]=i$, the Hamiltonian (\[hamcav1\]) takes the form $$\label{hamcav2}
\begin{array}{lll}
H_{cav} & = &
\displaystyle{\frac{\hat{p}_x^2}{2}+\frac{\hat{p}_y^2}{2}+\frac{\hat{x}^2}{2}+\frac{\hat{y}^2}{2}+\frac{\Omega}{2}\hat{\sigma}_z}
\\ \\
& &
\displaystyle{+2\lambda\hat{x}\big[\cos(\phi)\hat{\sigma}_x+\sin(\phi)\hat{\sigma}_y\big]}
\\ \\
& &
\displaystyle{+2\lambda\hat{y}\big[\cos(\theta)\hat{\sigma}_x+\sin(\theta)\hat{\sigma}_y\big]}.
\end{array}$$ For the simple example of $\phi=0$ and $\theta=\pi/2$ we recover the cylindically symmetric $E\!\times\!\varepsilon$ Hamiltonian (\[jtham\]). In fact, for $$\label{eecon}
|\phi-\theta|=(j+1/2)\pi,\hspace{1.2cm} j
\hspace{0.3cm}\mathrm{integer},$$ $H_{cav}$ is unitarilly equivalent with the $E\!\times\!\varepsilon$ Hamiltonian $H_{JT}$ by identifying $\Omega$ with $\Delta$. In some special cases of the phases, the last two terms of (\[hamcav2\]) can be written as $$\begin{array}{l}
2\lambda(\hat{x}+\hat{y})\left[\begin{array}{cc}
0 & \mathrm{e}^{-i\phi}\\
\mathrm{e}^{i\phi} & 0\end{array}\right],\hspace{0.7cm}\mathrm{for}\hspace{0.2cm}\theta-\phi=2j\pi,\\ \\
2\lambda(\hat{x}-\hat{y})\left[\begin{array}{cc}
0 & \mathrm{e}^{-i\phi}\\
\mathrm{e}^{i\phi} & 0\end{array}\right],\hspace{0.7cm}\mathrm{for}\hspace{0.2cm}\theta-\phi=(2j+1)\pi,
\end{array}$$ for some integer $j$. For these situations the CI is replaced by an intersecting curve in the directions of $\varphi=3\pi/4,\,7\pi/4$ or $\varphi=\pi/4,\,5\pi/4$ respectively. Here, it is clear that by a unitary rotation, the adiabatic states can be made real, indicating that the geometrical phase becomes identically zero as the wave packet is encircling the CI. This is indeed seen in Fig. \[fig2\] displaying the geometric phases of the Hamiltonian (\[hamcav2\]). The general form of the APSs, in polar coordinates, reads $$V_\pm^{ad}(\rho,\varphi)=\!\frac{\rho^2}{2}\!\pm\!\sqrt{\!\left(\frac{\Omega}{2}\right)^{\!2}\!\!+4\lambda^2\rho^2\!\left[1\!+\!\cos(\phi\!-\!\theta)\sin(2\varphi)\right]}.$$ The lower APS has two minima for angels $\varphi=\pi/4,\,3\pi/4$. It is known that the “quadratic" $E\!\times\!\varepsilon$ Hamiltonian has three local minima in the sombrero shaped potential [@koppel; @spinorbit2]. This derives from a term of the form $\sin(3\varphi)$ in the APSs. Here we have a $\sin(2\varphi)$-dependence instead and hence the double minima structure.
The single valued adiabatic states can again be written $$\label{adstate2}
\Psi_u(\rho,\varphi)=\!\left[\!\begin{array}{c}\sin(\nu) \\ -\cos(\nu)\mathrm{e}^{i\mu}\end{array}\!\right]\!,\hspace{0.5cm}\Psi_l(\rho,\varphi)=\!\left[\!\begin{array}{c}\cos(\nu) \\ \sin(\nu)\mathrm{e}^{i\mu}\end{array}\!\right],$$ but with $$\begin{array}{l}
\displaystyle{\tan(2\nu)=\frac{4\lambda\rho}{\Omega}\sqrt{1+\sin(2\varphi)\cos(\phi-\theta)}},\\
\\
\displaystyle{\mu=\tan\left(\frac{\cos(\varphi)\sin(\phi)+\sin(\varphi)\sin(\theta)}{\cos(\varphi)\cos(\phi)+\sin(\varphi)\cos(\theta)}\right)}.
\end{array}$$ Encircling the CI at a radius $R$ we find the geometric phase (\[berryfas\]) $$\label{berryfas2}
\gamma_{cav}(R)=-\left.\int_0^{2\pi}\cos^2(\alpha)\frac{\partial\psi}{\partial\varphi}\,d\varphi\right|_{\rho=R}.$$ The phase (\[berryfas2\]) is depicted in Fig. \[fig2\] as a function of $\theta$ and $\lambda$ for fixed $\phi=0$ and $\Omega=1$ (a) and $\phi=1/2$ and $\Omega=1$ (b). The asymptotic value for large couplings $\lambda$ is either $-\pi$ or $0$ (modulo $2\pi$). The radius $R$ is taken to be the minimum of the cylindrically symmetric case ($\theta=\pi/2$). For $\phi=0$, the greatest effect generated by the geometrical phase is seen to be in the symmetric case of $\theta=\pi/2$, while for $\phi\neq0$ the situation becomes more complex. Note that, according to Eq. (\[potmin\]), and identifying $\Omega$ with $\Delta$, $\lambda\neq0$ in order to have a sombrero structure of the lower APS, which is the reason why $\lambda$ is not approaching 0.
![The geometric phase (\[berryfas2\]) as function of coupling $\lambda$ and field phase $\theta$. The radius $R=\rho_{min}$, where $\rho_{min}=\sqrt{4\lambda^2-(\Omega/4\lambda)^2}$ is the adiabatic potential minima of the sombrero when $\theta-\phi=\pi/2$ (the cylindrically symmetric case). The other dimensionless parameters are $\phi=0$ and $\Omega=1$ (a) and $\phi=1/2$ and $\Omega=1$ (b). []{data-label="fig2"}](fig2a.eps "fig:"){width="8cm"} ![The geometric phase (\[berryfas2\]) as function of coupling $\lambda$ and field phase $\theta$. The radius $R=\rho_{min}$, where $\rho_{min}=\sqrt{4\lambda^2-(\Omega/4\lambda)^2}$ is the adiabatic potential minima of the sombrero when $\theta-\phi=\pi/2$ (the cylindrically symmetric case). The other dimensionless parameters are $\phi=0$ and $\Omega=1$ (a) and $\phi=1/2$ and $\Omega=1$ (b). []{data-label="fig2"}](fig2b.eps "fig:"){width="8cm"}
The Jahn-Teller effect in cavity QED {#subsec3b}
------------------------------------
The Jahn-Teller effect states that, in the presence of a CI, lowering the symmetry may be energetically favorable in various systems, typically for molecules and crystals. In terms of the $E\!\times\!\varepsilon$ Hamiltonian this is intuitive, since a wave packet centered at the origin (the CI) will have a larger “potential energy" than a wave packet located at $\rho_{min}$. The corresponding symmetry breaking in cavity QED implies that the system ground state consists of non-zero fields in the two modes. At the same time, the quantum-dot is not entirely in its lower state but in a superposition of its two internal states. This is related to a well known phenomenon in quantum optics, namely [*superradiance*]{} [@dicke]. This, of course, comes about due to the strong interaction between the quantum-dot and the cavity fields. For multi quantum-dot systems and in the thermodynamic limit, where the number of two-level quantum-dots and the volume tend to infinity while the density is kept fixed, this results in a second order quantum phase transition between a [*normal*]{} (the field in its vacuum) and a [*superradiant phase*]{} (a macroscopic non-zero field) [@dicke2]. The critical coupling of this phase transition is given by $\lambda_c=\sqrt{|\Delta|\omega}$ [@dicke2], which indeed follows from Eq. (\[potmin\]).
However, it can be shown that for a quantum-dot in which the lower state is its ground state, the normal-superradiant phase transition is an artifact from neglecting the self-energy term from the Hamiltonian [@a2]. For a single mode, and in unscaled units, this term is given by $$H_{se}=\frac{e^2}{2m}\frac{\pi\hbar}{V\omega}\hat{x}^2,$$ where $e$ is the electron charge, $m$ its mass and $V$ the effective mode volume. This should be compared with the matter-field coupling $$\lambda=\Omega d\sqrt{\frac{2\pi\hbar}{V\omega}},$$ where $d$ is the dipole moment of the transition of interest in the quantum-dot. It is clear that the self-energy term $H_{se}$ tends to quench the Mexican hat structure, and further that $\lambda$ and $H_{sc}$ are not fully independent. Indeed, in Ref. [@a2] it is demonstrated, either using the Thomas-Reiche-Kuhn sum rule or simple thermodynamical and gauge invariance arguments, that the sombrero shape cannot be obtained for any set of physical parameters.
To circumvent this obstacle one may use a two-photon Raman type of interaction, where three levels of the quantum-dot are coupled through the cavity mode and an external classical laser field [@ramandicke; @jonasdicke]. In the large detuning limit of one of the internal levels it can be adiabatically eliminated [@adbose] and one arrives at an effective model very similar to the one above. In such procedure one introduces an additional independent parameter, the detuning $\delta$ of the eliminated level [@cond]. As the detuning enters in the effective matter-field coupling parameter, but not in the self energy term, these two become in principle independent, and in particular $\lambda$ can be made large in comparison with $H_{se}$. The effective model, once the detuned level has been eliminated, contains some [*Stark shift*]{} terms that will modify the potential surfaces, but the CI and sombrero structure are still present. The external laser fields have the advantage of being easily controllable in terms of system parameters such as amplitude and phase. A system Hamiltonian suitable for realizing the Jahn-Teller model can be found in Ref. [@vedralham]. It should be noted though, that the effective Hamiltonian is in general time dependent, which is prevailed by imposing a RWA. This, however, induces a “momentum-dependent" potential surface, but nonetheless, the JT symmetry breaking is still present in this approximation [@dicke2].
Another possibility to surmount the problem with the self-energy, which is assumed in this paper, is to use the fact that the two internal levels that couple to the cavity modes are normally highly excited meta stable (Rydberg) states. For these states, neither the Thomas-Reiche-Kuhn sum rule nor the thermodynamical arguments apply, and the symmetry breaking may still occur. We therefore discard the self-energy terms as they would only modify the frequencies of the harmonic potentials. In general, also for the Raman coupled model, the states involved are highly excited metastable states and the sum rule cannot be applied in those cases either. A benefit of the Raman model, compared to a one photon model, is the higher controllability of the system parameters; especially the diagonal element is a detuning parameter (and not a transition frequency $\Omega$) that can be made small compared to the matter-field coupling. The drawback of an effective Raman model is that the analysis is considerably less intuitive due to the RWA. Here, we therefore choose the simpler model of the two as the physical phenomena may be more easily extracted from it.
Numerical results {#sec4}
=================
Contrary to molecular or solid state systems, properties of the cavity fields are in comparison easily measured, for example, phoson distribution [@photonmeas], field quadratures [@squezmeas] and, in fact, the whole phase space distribution using quantum tomography [@tomo]. Using wave packet propagation methods, the dynamics of such quantities will be studied in this section with emphasizes on the effects emerging from the geometrical phase. As an initial state we take a disentangled one, given in cartesian coordinates by $$\label{initialstate2}
\Psi(x,y,0)=\psi(x,y,0)\frac{1}{\sqrt{2}}\left[\begin{array}{c} 1 \\ -1
\end{array}\right],$$ where $$\label{initialstate}
\psi(x,y,0)=\frac{1}{\sqrt{\pi}}\mathrm{e}^{-(\mathrm{Im}\,x_0)^2-(\mathrm{Im}\,y_0)^2}\mathrm{e}^{-\frac{(x-x_0)^2}{2}-\frac{(y-y_0)^2}{2}}.$$ The initial quantum-dot is a linear combination of its two internal states with equal amplitudes, and the two field modes are in Gaussian states corresponding to coherent field states; $|x_0/\sqrt{2}\rangle$ and $|y_0/\sqrt{2}\rangle$ respectively [@mandel]. Such initial states are readily prepared experimentally. We will further pick $y_0=0$ and $x_0=2\lambda$ such that the initial wave packet is approximately centered at the minima of the sombrero. Note that the initial average momentum is zero, and that $\Psi(x,y,0)$ is different from the adiabatic states (\[adstate2\]). A consequence of this is that the wave packet evolution will not be restricted to a single APS. However, the upper adiabatic state is only marginally populated for our particular choice of initial state and the main phenomena studied here, the effects of the geometrical phase on the field properties, is indeed seen even though slight interference between the two adiabatic states occurs. Hence, we emphasize that the dynamics take place mainly on the lower adiabatic surface. Properties of the upper APS have been studied in Ref. [@uaps].
We restrict the analysis to the cylindrically symmetric case, where the time evolved state $$\label{timedepstate}
\Psi(x,y,\tau)=\frac{1}{\sqrt{2}}\left(\psi_e(x,y,\tau)\left[\begin{array}{c} 1 \\ 0\end{array}\right]+\psi_g(x,y,\tau)\left[\begin{array}{c} 0 \\ 1\end{array}\right]\right),$$ will predominantly spread along the minima of the sombrero potential. As the wave packet broadens it will, after a certain time, start to self-interfere. We may estimate the characteristic time for this process by approximate the inherent spreading by free evolution along the minima of the sombrero potential to get $$T_{in}\approx\sqrt{4\pi^2\lambda^2-1}\approx2\pi\lambda.$$ Within this time, the wave packet width has expanded over a distance $2\pi\rho_{min}$.
From the full system state (\[timedepstate\]), we can derive the reduced density operators for the separate constitutes $$\rho_i(\tau)=\mathrm{Tr}_{j,k}\Big[\rho(\tau)\Big],$$ where the subscripts represent, either the two modes $a$ and $b$ or the quantum-dot, and $\rho(\tau)=\Psi^*(x,y,\tau)\Psi(x,y,\tau)$. Using the reduced density operators we will especially study the photon statistics and the Husimi $Q$-distribution [@mandel; @schleich] $$\label{quant}
\begin{array}{l}
\displaystyle{P_i(n)=\langle n|\rho_i(\tau)|n\rangle}, \\ \\
\displaystyle{\langle \hat{n}_i\rangle=\sum_nnP_i(n)}, \\ \\
\displaystyle{Q_i(\alpha)=\frac{1}{\pi}\langle\alpha|\rho_i(\tau)|\alpha\rangle}.
\end{array}$$ Here, $|n\rangle$ is the $n$’th-photon Fock state, $|\alpha\rangle$ a coherent state with amplitude $\alpha$ and the subscript $i=a,b$ for the respective modes.
Dynamics on the $T_{in}$ time scale {#subsec4a}
-----------------------------------
Discussed in Sec. \[sec2\], it is the term $V_{gauge}$ that gives rise to a geometric phase. To correctly describe the adiabatic evolution each term of $H_{JT}^{ad}$ in Eq. (\[jtpolarad\]) must be taken into account, and it is not enough to study dynamics upon the potentials $V_\pm^{ad}$. In this subsection we study the full dynamics using Hamiltonian (\[hamcav2\]), and we hence go beyond any adiabatic approximation. However, in order to identify the effects of the geometrical phase we compare the results with the ones obtained by propagating the same initial state using the “semi-adiabatic" Hamiltonian defined as $$\label{semiadh}
\tilde{H}_{JT}^{ad}=T+V_-^{ad},$$ where $T$ and $V_-^{ad}$ are both given in Eq. (\[jtpolaradterms\]). Accordingly, a wave packet evolving via the Hamiltonian $\tilde{H}_{JT}^{ad}$ around the origin will not accumulate any geometrical phase.
The characteristic time scale $T_{in}$ determines how long it takes for the particular initial state (\[initialstate\]) to inherently spread out across the CI and start to self-interfere. It is therefore a measure of the collapse time. The effect of the geometrical phase on the probability wave functions $|\Psi(x,y,\tau)|^2=|\psi_e(x,y,\tau)|^2/2+|\psi_g(x,y,\tau)|^2/2$ has been discussed in Refs. [@koppel; @spinorbit2]. Initially we choose $x_0>0$ while $y_0=0$ such that interference of the evolved wave packet sets off at $-x_0$ where the two tails of the packet first join. Destructive and constructive interference cause nodes (vanishing probability distribution) and anti-nodes (non-vanishing probability distribution) in $|\Psi(x,y,t)|^2$, and the ring-shaped wave packet splits up in localized blobs. In the case of $\tilde{H}_{JT}^{ad}$, in which the geometric phase is zero, an anti-node builds up at $x=-x_0$, while for $\gamma_{JT}(\rho_{min})=-\pi$ (as is the case of zero detuning in the $E\!\times\!\varepsilon$ model) a node is formed at $x=-x_0$. The location of the corresponding node or anti-node depends on the value of $\gamma_{JT}(\rho_{min})$, and in all our examples $\Omega<\lambda$ such that $\gamma_{JT}(\rho_{min})\sim-\pi$ giving a node at $x\approx -x_0$. These features are visible in Fig. \[fig3\] showing the numerical results of the propagated distributions $|\Psi(x,y,\tau)|^2$ for three different times $\tau$. Full dynamics governed by Hamiltonian (\[hamcav2\]), with $\phi=0$ and $\theta=\pi/2$ (cylindrically symmetric case) are shown in the left plots, while the right ones reproduce the results from propagation using the semi-adiabatic Hamiltonian (\[semiadh\]). The effect of the geometrical phase becomes clear once the wave packet starts to self-interfere. The number of localized blobs depends on time $\tau$ and system parameters and then especially $\rho_{min}$. Note that very similar results where presented in Refs. [@koppel; @spinorbit2]. However, the Hamiltonians used for the simulations in Refs. [@koppel; @spinorbit2] are in general different from the one $\tilde{H}_{JT}^{ad}$ utalized here; a [*single surface approximation*]{} [@singlesurf; @singlesurf2] is applied in most examples of Refs. [@koppel; @spinorbit2] while here it is only considered for the non-geometrical phase case.
![Snapshots of the wave packet distributions $|\Psi(x,y,\tau)|^2$ at times $\tau=T_{in}/4$ (a) and (b), $\tau=T_{in}/2$ (c) and (d), and $\tau=T_{in}$ (e) and (f) for the cases with (left) and without geometrical phase (right). In the last two plots, the difference between the interference structures is clearly visible. See Refs. [@koppel; @spinorbit2] for similar results. The dimensionless parameters are $\Omega=0.5$ and $\lambda=3$. []{data-label="fig3"}](fig3a.eps){width="8cm"}
As the initial wave packet starts to spread, a non-zero field will begin to build up in the vacuum $b$ mode, on the cost of decreasing field intensity of mode $a$. However, without the RWA, the total number of excitations is not conserved. In Fig. \[fig4\] we display the individual photon distributions $P_i(n)$ at a quarter of the interference time $T_{in}$. Already at this instant has the initially empty mode a non-zero field intensity, and its photon distribution consists mostly of even number of photon states. This is a typical characteristic of Schrödinger cat states [@haroche], and in the next subsection we will indeed show that such a state is created in the system at certain times. The small but non-zero population of odd photon numbers in the $b$ mode is caused by non-adiabaticity; for the semi-adiabatic Hamiltonian (\[semiadh\]) the odd photon numbers are never populated for the given initial state (\[initialstate2\]) with $y_0=0$.
![The photon distributions $P_a(n)$ (white) and $P_b(n)$ (black) after a time $\tau=T_{in}/4$. The dimensionless parameters are the same as in Fig. \[fig3\]. []{data-label="fig4"}](fig4.eps){width="6cm"}
Dynamics beyond the $T_{in}$ time scale {#subsec4b}
---------------------------------------
Seen in the previous subsection, for an initial localized state mainly located at the minima of the sombrero potential and with zero average momentum, the time scale $T_{in}$ determines the collapse time; the time it takes for the localized wave packet to spread out over its accessible phase space region. Over longer periods, a revival structure in physical quantities is expected [@robinett], where localized bumps are formed in phase space signalizing fractional or full revivals [@fracrev]. It has been pointed out however, that the collapse-revival characteristics are rather different in models where the RWA has been applied [@oldwine; @dong]. Typically, in the RWA regime the various time scales become long. In this work we are outside such a regime, and we will in particular find that the revival time is given by a multiple of $\lambda T_{in}$ and that phase space evolution is significantly different for the two Hamiltonians (\[hamcav2\]) and (\[semiadh\]) due to the geometrical phase.
From Fig. \[fig3\] we see that after a time $T_{in}$, the initial wave packet is spread throughout the minima of the sombrero potential and the self-interference causes nodes in the probability distribution. The number of localized bumps depends on $\rho_{min}$ (\[potmin\]), but also on the time $\tau$; at first, when the self-interference sets off, the number of bumps increases to a maximum value and then the number begins to decrease and eventually form a single localized wave packet. Full revival occurs when a single localized bump is formed at the same position as the initial wave packet. To study the field dynamics we use the $Q$-function for the two modes $a$ and $b$ of Eq. (\[quant\]). We will present the two functions $Q_a$ and $Q_b$ in the same plots for brevity, but mark them with letters $a$ and $b$ respectively. At the initial time $\tau=0$, $Q_a$ and $Q_b$ are Gaussians centered at ${\bf \alpha}=(\sqrt{2}\lambda,0)$ and ${\bf \alpha}=(0,0)$ respectively. As time evolves, the $b$ mode builds up its intensity and the $Q$-function moves away from the origin, while $Q_a$ at first decreases its intensity by tending towards the origin. However, over longer time scales, $\tau>T_{in}$, a swapping of energy between the two modes will take place. This phenomenon has been discussed in our model, but only when the RWA has been imposed [@hamref2]. As our analysis concerns a regime far from the RWA one, this exchange of energy between the modes occurs at very different time scales than in Ref. [@hamref2], similar to what was found for the inversion in the JC model [@oldwine]. Namely, the characteristic time scales in the parameter regimes of the RWA and without the RWA in the JC model can differ by orders of magnitude.
From our numerical simulations we have found that localization of the phase space distributions comes about at multiples of time $T_{frac}=\lambda T_{in}$, which hence are the characteristic scales for fractional revivals [@robinett; @fracrev]. The larger the radius $\rho_{min}$, the better resolved wave packet localizations. In Fig. \[fig5\] we display examples of the $Q$-functions $Q_a$ and $Q_b$ (indicated in the figures by $a$ and $b$) obtained either from the full system Hamiltonian (\[hamcav2\]) (left) or from the semi-adiabatic Hamiltonian (\[semiadh\]) (right). The times are here, $\tau=\lambda T_{in},\,2\lambda T_{in},\,3\lambda T_{in},\,4\lambda T_{in}$. A clear discrepancy is seen between the two models. For example, at $\tau=2\lambda T_{in}$ (c) and (d), mode $a$ in the left plot (with geometrical phase) is approximately in vacuum, while for the semi-adiabatic system (without geometrical phase), mode $b$ is roughly empty. At this instant, the non-empty mode is in a Schrödinger cat state. For $\tau=4\lambda T_{in}$ the full system has revived; the $Q$-functions have evolved into approximate replicas of their initial states. This is true up to an overall phase for the semi-adiabatic case, which is typical for a half-revival [@robinett]. From this figure we find the revival time for a wave packet encircling the CI in the $E\!\times\!\varepsilon$ model to be $$\label{revtim}
T_{rev}=4\lambda T_{in}\approx8\pi\lambda^2.$$ For the semi-adiabatic model, exact revivals (in terms of restoring also the overall phase) occur at twice this time. It should be pointed out that formula (\[revtim\]) has been verified for a large set of different parameters.
![Snapshots of the Husime $Q$-functions at times $\tau=\lambda T_{in}$ (a) and (b), $\tau=2\lambda T_{in}$ (c) and (d), $\tau=3\lambda T_{in}$ (e) and (f) and $\tau=4\lambda T_{in}$ (g) and (h). The left plots are the results with geometric phase, using Hamiltonian (\[hamcav2\]), and the non-geometric phase results, obtained by the Hamiltonian in Eq. (\[semiadh\]), are displayed in the right figures. The $Q$-function of the $a$-mode is labeled by $a$ in the plots, while $b$ labels the second mode $Q$-function. The dimensionless parameters are $\lambda=6$ and $\Omega=0.5$. []{data-label="fig5"}](fig5.eps){width="9cm"}
Even though the phase space distribution of a cavity mode is in principle measureable [@tomo], the field intensity is directly regained from the cavity output field using a photon-counter detector. Already Fig. \[fig5\] indicates that the average number of photons $\langle n_i\rangle$ differ considerably between the full model and the semi-adiabatic one. This is verified in Fig. \[fig6\] showing the time evolution of $\langle n_a\rangle$ and $\langle n_b\rangle$ for both models. Judging from the field intensities in this figure, the revival time of the semi-adiabatic model seems to be half the one of the full model, but here the $a$ mode is indeed not in a coherent state but in a Schrödinger cat.
![The average photon numbers $\langle \hat{n}_i\rangle$ for both modes $a$ (black) and $b$ (gray) as a function of scaled time $\tau/T_{in}$ . The upper plot presents the results from using the full Hamiltonian (\[hamcav2\]), which includes a geometrical phase, while the lower plot shows the results from using Hamiltonian (\[semiadh\]). The effect of the geometrical phase is remarkably reflected in the two field intensities. The dimensionless parameters are the same as in Fig. \[fig5\]. []{data-label="fig6"}](fig6.eps){width="6cm"}
Conclusions {#sec5}
===========
In this paper we have shown how a system of a two-level “particle" interacting with the fields of a bimodal cavity may fall in the category of JT models. By representing the model Hamiltonian in terms of field quadrature operators, rather than boson ladder operators, we identified its APSs and a CI. In this nomeclature, and in particular for the multi-particle analogue (Dicke model), the JT effect of cavity QED was identified with the normal-superradiant phase transition. The system studied here was described by the well known $E\!\times\!\varepsilon$ Hamiltonian. Knowledge from earlier research on this model, almost exclusively in molecular or chemical and condensed matter physics, has been applied on this cavity QED counterpart. Our main interest concerned the geometrical Berry phase reign from encircling the CI. The effect of the geometrical phase was studied by comparing physical quantities, such as the field phase space distributions and the field intensities, obtained from the evolution of either the $E\!\times\!\varepsilon$ Hamiltonian or the semi-adiabatic one in which no geometrical phase occurs. Clear distinctions between the two models were found when the system is let to evolve over longer time periods. Energy is swapped between the two field modes, and this exchange is highly affected by the geometrical phase. From our numerical results we could as well present analytical expressions for the collapse-revival times for a wave packet encircling the CI in the $E\!\times\!\varepsilon$ model.
In addition, by introducing the notion of a wave packet evolving on two coupled potential surfaces, a deeper understanding of cavity QED models is obtained. This work, analyzing the geometrical phase, serves as an alternative viewpoint of the phenomenon in comparison to previous studies such as Refs. [@vedralham; @berrycav]. It is indeed believed that the wave packet method used here will give even more thorough insight into cavity QED problems, or even trapped ion systems where related CI models are expect to occur [@dong]. We plan to study, using the current approach, the dynamics of the Dicke normal-superradiant phase transition. Another project underway is to investigate the “ molecular Aharanov-Bohm effect" and “molecular gauge theory" in terms of cavity QED models.
The author would like to thank Dr. Åsa Larson for inspiring discussions.
[999]{}
H. A. Jahn and E. Teller, Proc. R. Soc. London Ser. A [**161**]{}, 220 (1937).
M. C. M. O’Brien and C. C. Chancey, Am. J. Phys. [**61**]{}, 688 (1993); I. Bersuker, [*The Jahn-Teller Effect*]{}, (Cambridge University Press, Cambridge, 2006).
C. A. Mead, J. Chem. Phys. [**78**]{}, 807 (1983); D. R. Yarkony, Rev. Mod. Phys. [**68**]{}, 985 (1996); D. R. Yarkony, J. Phys. Chem. A [**103**]{}, 8579 (1999).
M. Baer, [*Beyond Born-Oppenheimer*]{}, (Wiley, New York, 2006).
R. Englman, [*The Jahn-Teller Effects in Molecules and Crystals*]{}, (Wiley, New York, 1972).
C. A. Bates, Phys. Rep. [**35**]{}, 178 (1978).
H. C. Longuet-Higgins, U. Öpik, M. H. L. Pryce, and R. A. Sack, Proc. R. Soc. London Ser. A [**244**]{}, 1 (1958).
C. A. Mead and D. G. Truhlar, J. Chem. Phys. [**70**]{}, 2284 (1979).
Y. Aharanov and D. Bohm, Phys. Rev. [**115**]{}, 485 (1959).
A. Bohm, A. Mostafazadeh, H. Kuizumi, Q. Niu, and J. Zwanziger, [*The Geometric Phase in Quantum Systems*]{}, (Springer, Berlin, 2003).
C. A. Mead, Rev. Mod. Phys. [**64**]{}, 51 (1992).
M. V. Berry, Proc. R. Soc. London Ser. A [**392**]{}, 45 (1984).
G. Delacrétaz, E. R. Grant, R. L. Whetten, L. Wöste, and J. W. Zwanziger, Phys. Rev. Lett. [**56**]{}, 2598 (1986); F. S. Ham, [*ibid.*]{} [**58**]{}, 725 (1987); J. Schön and H. Köppel, Chem. Phys. Lett. [**231**]{}, 55 (1994).
R. Baer, D. M. Charutz, R. Kosloff, and M. Baer, J. Chem. Phys. [**105**]{}, 9141 (1991).
J. Schön and H. Köppel, J. Chem. Phys. [**103**]{}, 9292 (1995).
H. A. Jahn, Proc. R. Soc. London Ser. A [**164**]{} 117 (1938).
J. Schön and H. Köppel, J. Chem. Phys. [**1008**]{}, 1503 (1998).
A. P. Hines, C. M. Dawson, R. H. McKenzie, and G. J. Milburn, Phys. Rev. A [**70**]{}, 022303 (2004).
H. Yamasaki, Y. Natsume, A. Terai, and K. Nakamura, Phys. Rev. E [**68**]{}, 046201 (2003); E. Majernikova, and S. Shpyrko, Phys. Rev. A [**73**]{}, 066215 (2006).
G. Liberti, R. L. Zaffino, F. Piperno, and F. Plastina, Phys. Rev. A [**76**]{}, 042332 (2007).
I. Fuentes-Guridi, A. Carollo, S. Bose, and V. Vedral, Phys. Rev. Lett. [**89**]{}, 220404 (2002); S. Bose, A. Carollo, I. Fuentes-Guridi, M. F. Santos, and V. Vedral, J. Mod. Opt. [**50**]{}, 1175 (2003); A. Carollo, I. Fuentes-Guridi, M. F. Santos, and V. Vedral, Phys. Rev. Lett. [**92**]{}, 020402 (2004).
A. Carollo, M. F. Santos, and V. Vedral, Phys. Rev. A. [**67**]{}, 063804 (2003).
X. Z. Yuan and K. D. Zhu, Phys. Rev. B [**74**]{}, 073309 (2006); S. Siddiqui, and J. Gea-Banacloche, Phys. Rev. A [**74**]{}, 052337 (2006).
J. Larson, Physica. Scr. [**76**]{}, 146 (2007); J. Larson, J. Phys: Conf. Ser. [**99**]{}, 012011 (2008).
G. Levine and V. N. Muthukumar, Phys. Rev. B [**69**]{}, 113203 (2004).
E. T. Jaynes and F. W. Cummings, Proc. IEEE [**51**]{}, 89 (1963); B. W. Shore and P. L. Knight, J. Mod. Opt. [**40**]{}, 1195 (1993).
C. A. Mead, Chem. Phys. [**49**]{}, 33 (1980); H. Koizumi, and S. Sugano, J. Chem. Phys. [**102**]{}, 4472 (1995); L. V. Poluyanov, S. Mishra, and W. Domcke, Mol. Phys. [**105**]{}, 1471 (2007).
G. Bevilacqua, L. Martinelli, and G. P. Parravicini, Phys. Rev. B [**63**]{}, 132403 (2001).
H. Köppel, W. Domcke, and L. S. Cederbaum, Adv. Chem. Phys. [**57**]{}, 59 (1984); M. Baer, Phys. Rep. [**358**]{}, 75 (2002); G. A. Worth and L. S. Cederbaum, Ann. Rev. Phys. Chem. [**55**]{}, 127 (2004).
A. S. Davydov, [*Quantum Mechanics*]{}, 2nd ed. (Pergamon, Oxford, 1976).
C. A. Mead, and D. G. Truhlar, J. Chem. Phys. [**70**]{}, 2284 (1979); C. A. Mead, Chem. Phys. [**49**]{}, 23 (1980).
S. Haroche and J. M. Raimond, [*Exploring the Quantum*]{}, (Oxford University Press, Oxford, 2006).
A. Wallraf, D. I. Schuster, A. Blais, L. Frunzio, J. Majer, M. H. Devoret, S. M. Girvin, and R. J. Schoelkopf, Nature [**431**]{}, 162 (2004); I. Chiorescu, P. Bertet, K. Semba, Y. Nakamura, C. J. P. M. Harmans, and J. E. Mooil, Nature [**431**]{}, 159 (2004); N. Hatakenaka, and S. Kurihara, Phys. Rev. A [**54**]{}, 1729 (1996); A. T. Sornborger, A. N. Cleland, and M. R. Geller, Phys. Rev. A [**70**]{}, 052315 (2004).
S. Slama, G. Krentz, S. Bux, C. Zimmermann, and P. W. Courteille, Phys. Rev. A [**75**]{}, 063620 (2007); F. Brennecke, T. Donner, S. Ritter, T. Bourdel, M. Köhn, and T. Esslinger, Nature [**450**]{}, 268 (2007); P. Truetlein, D. Hunger, S. Camerer, T. W. Hänsch, and J. Reichel, Phys. Rev. Lett. [**99**]{}, 140403 (2007); Y. Colombe, T. Steinmetz, G. Dubois, F. Linke, D. Hunger, and J. Reichel, Nature [**450**]{}, 272 (2007); H. T. Ng, Phys. Rev. A [**77**]{}, 033617 (2008).
M. O. Scully and M. S. Zubairy, [*Quantum Optics*]{}, ( Cambridge University Press, Cambridge, 1997).
A. Messina, S. Maniscalco, and A. Napoli, J. Mod. Opt. [**50**]{}, 1 (2003).
G. J. Papadopoulos, Phys. Rev. A [**37**]{}, 2482 (1988); C. Wildfeur and D. H. Schiller, Phys. Rev. A [**67**]{}, 053801 (2003).
G. Benivegna and A. Messina, J. Mod. Opt. [**41**]{}, 907 (1994).
R. H. Dicke, Phys. Rev. [**93**]{}, 99 (1954).
Y. K. Wang and F. T. Hioe, Phys. Rev. A [**7**]{}, 831 (1973); K. Hepp and E. H. Lieb, Ann. Phys. [**76**]{}, 360 (1973); [*ibid.*]{}, Phys. Rev. A [**8**]{}, 2517 (1973); N. Lambert, C. Emary, and T. Brandes, Phys. Rev. Lett. [**92**]{}, 073602 (2004).
K. Rzazewski, K. Wodkiewicz, and W. Zakowicz, Phys. Rev. Lett. [**35**]{}, 432 (1975); I. Bialynicki-Birula and K. Rzazewski, Phys. Rev. A [**19**]{}, 301 (1979).
F. Dimer, B. Estienne, A. S. Parkins, and H. J. Carmichael , Phys. Rev. A 75, 013804 (2007); S. Morris and A. S. Parkins, [*ibid.*]{} [**77**]{}, 043810 (2008).
J. Larson, K. Rzazewski, and M. Lewenstein, (unpublished).
M. Alexanian and S. K. Bose, Phys. Rev. A [**52**]{}, 2218 (1995).
The detuning is independent as a control parameter, but it should be remembered that it can only be varied within the validity regime of the adiabatic elimination approximation.
M. J. Holland, D. F. Walls, and P. Zoller, Phys. Rev. Lett. [**67**]{}, 1716 (1991); D. I. Schuster, A. A. Houck, J. A. Schreier, A. Wallraff, J. M. Gambetta, A. Blais, L. Frunzio, J. Majer, B. Johnson, M. H. Devoret, S, M Girvin, R. J. Schoelkopf, Nature [**445**]{}, 515 (2007).
H. M. Wiseman and G. J. Milburn, Phys. Rev. A [**47**]{}, 642 (1993); Q. A. Turchette, N. P. Georgiades, C. J. Hood, H. J. Kimble, and A. S. Parkins, Phys. Rev. A [**58**]{}, 4056 (1998).
K. Vogel and H. Risken, Phys. Rev. A [**40**]{}, 2847 (1989); D. T. Smithey, M. Beck, M. G. Raymer, and F. Faridani, Phys. Rev. Lett. [**70**]{}, 1244 (1993); L. G. Lutterbach, and L. Davidovich, Phys. rev. Lett. [**78**]{}, 2547 (1997); P. Bertet, A. Auffeves, P. Maioli, S. Osnaghi, T. Meunier, M. Brune, J. M. Raimond, and S. Haroche, Phys. Rev. Lett. [**89**]{}, 200402 (2002); XuBo Zou, K. Pahlke, and W. Mathis, Phys. Rev. A [**69**]{}, 015802 (2004); A. A. Houck, D. I. Schuster, J. M. Gambetta, J. A. Schreier, B. R. Johnson, J. M. Chow, L. Frunzio, J. Majer, M. H. Devoret, S. M. Girvin, and R. J. Schoelkopf, Nature [**449**]{}, 328 (2007). S. Deleglise, I. Dotsenko, C. Sayin, J. Bernu, M. Brune, J. -M. Raimond, and S. Haroche, arXiv:08091064.
L. V. Poluyanov, S. Mishra, and W. Domcke. Chem. Phys. [**332**]{}, 243 (2007).
L. Mandel, and E. Wolf, [*Optical Coherence and Quantum Optics*]{}, (Cambridge University Press, Cambridge, 1995).
W. P. Schleich, [*Quantum Optics in Phase Space*]{} (Wiley, New York, 2001).
S. Adhikari and G. D. Billing, J. Chem. Phys. [**111**]{}, 40 (1999).
R. W. Robinett, Phys. Rep. [**392**]{}, 1 (2004).
E. Romera and F. de los Santos, Phys. Rev. Lett. [**99**]{} 263601 (2007).
D. Wang, T. Hansson, Å. Larson, H. O. Karlsson, and J. Larson, Phys. Rev. A, [**77**]{}, 053808 (2008); H. Moya-Cessa, A. Vidiella-Barranco, J. A. Roversi, D. S. Freitas, and S. M. Dutra, [*ibid.*]{} [**59**]{}, 2518 (1999).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this study, main parameters of Super proton-proton Collider (SppC) based lepton-proton colliders are estimated. For electron beam parameters, highest energy International Linear Collider (ILC) and Plasma Wake Field Accelerator-Linear Collider (PWFA-LC) options are taken into account. For muon beams, 1.5 TeV and 3 TeV center of mass energy Muon Collider parameters are used. In addition, ultimate $\mu$p collider which assumes construction of additional 50 TeV muon ring in the SppC tunnel is considered as well. It is shown that luminosity values exceeding $10^{32}$ $cm^{-2}s^{-1}$ can be achieved with moderate upgrade of the SppC proton beam parameters. Physics search potential of proposed lepton-proton colliders is illustrated by considering small Bj[ö]{}rken x region as an example of SM physics and resonant production of color octet leptons as an example of BSM physics.'
address: |
$^{1}$TOBB University of Economics and Technology, Ankara, Turkey\
$^{2}$Ankara University, Department of Physics, Ankara, Turkey\
$^{3}$Ankara University, Department of Engineering Physics, Ankara, Turkey\
$^{4}$ANAS Institute of Physics, Baku, Azerbaijan
author:
- 'Ali Can Canbay$^{1,2}$, Umit Kaya$^{1, 2, *}$, Bora Ketenoglu$^{3}$, Bilgehan Baris Oner$^{1}$, Saleh Sultansoy$^{1,4}$'
title: 'SppC based energy frontier lepton-proton colliders: luminosity and physics'
---
$^{*}$Correspondence: [email protected]
April 2017
[***Keywords:*** SppC, lepton-proton colliders, luminosity, beam-beam effects, small Bj[ö]{}rken $x$, color octet leptons]{}
Introduction
============
It is known that lepton-hadron scattering had played crucial role in our understanding of deep inside of matter. For example, electron scattering on atomic nuclei reveals structure of nucleons in Hofstadter experiment \[1\]. Moreover, quark parton model was originated from lepton-hadron collisions at SLAC \[2\]. Extending the kinematic region by two orders of magnitude both in high $Q^{2}$ and small $x$, HERA (the first and still unique lepton-hadron collider) with $\sqrt{s}=0.32$ TeV has shown its superiority compared to the fixed target experiments and provided parton distribution functions (PDF) for LHC and Tevatron experiments (for review of HERA results see \[3, 4\]). Unfortunately, the region of sufficiently small $x$ ($<10^{-5}$) and high $Q^{2}$ ($\geq10\,GeV^{2}$) simultaneously, where saturation of parton densities should manifest itself, has not been reached yet. Hopefully, LHeC \[5\] with $\sqrt{s}=1.3$ TeV will give opportunity to touch this region.
Construction of linear $e^{+}e^{-}$colliders (or dedicated linac) and muon colliders (or dedicated muon ring) tangential to the future circular pp colliders, FCC or SppC, as shown in Fig. 1, will give opportunity to use highest energy proton beams in order to obtain highest center of mass energy in lepton-hadron and photon-hadron collisions. (For earlier studies on linac-ring type ep, $\gamma$p, eA and $\gamma$A colliders, see reviews \[6, 7\] and papers \[8-14\].)
FCC is the future 100 TeV center-of-mass energy pp collider studied at CERN and supported by European Union within the Horizon 2020 Framework Programme for Research and Innovation \[15\]. SppC is the Chinese analog of the FCC. Main parameters of the SppC proton beam \[16, 17\] are presented in Table \[tab:tablo1\]. The FCC based ep and $\mu$p colliders have been considered recently (see \[18\] and references therein).
Beam Energy (TeV) 35.6 68.0
------------------------------------------------ ------- -------
Circumference (km) 54.7 100.0
Peak Luminosity ($10^{34}\,cm^{-2}s^{-1}$) 11 102
Particle per Bunch ($10^{10}$) 20 20
Norm. Transverse Emittance ($\mu m$) 4.10 3.05
[$\beta$]{}[\*]{} amplitude function at IP (m) 0.75 0.24
IP beam size ($\mu m$) 9.0 3.04
Bunches per Beam 5835 10667
Bunch Spacing (ns) 25 25
Bunch length (mm) 75.5 15.8
Beam-beam parameter, $\xi_{pp}$ 0.006 0.008
: Main parameters of proton beams in SppC.[]{data-label="tab:tablo1"}
In this paper we consider SppC based ep and $\mu$p colliders. In Section 2, main parameters of proposed colliders, namely center of mass energy and luminosity, are estimated taken into account beam-beam tune shift and disruption effects. Physics search potential of the SppC based lp colliders have been evaluated in Section 3, where small Bj[ö]{}rken-x region is considered as an example of the SM physics and resonant production of color octet leptons is considered as an example of the BSM physics. Our conclusions and recommendations are presented in Section 4.\
Main Parameters of the SppC Based ep and $\mu$p Colliders
=========================================================
General expression for luminosity of SppC based $lp$ colliders is given by ($l$ denotes electron or muon):
$$\begin{aligned}
L_{lp} & = & \frac{N_{l}N_{p}}{4\pi max[\sigma_{x_{p}},\sigma_{x_{l}}]max[\sigma_{y_{p}},\sigma_{y_{l}}]}min[f_{c_{p},}\,f_{c_{l}}]\label{eq:Denklem1}\end{aligned}$$
\
where $N_{l}$ and $N_{p}$ are numbers of leptons and protons per bunch, respectively; $\sigma_{x_{p}}$ ($\sigma_{x_{l}}$) and $\sigma_{y_{p}}$ ($\sigma_{y_{l}}$) are the horizontal and vertical proton (lepton) beam sizes at interaction point (IP); $f_{c_{l}}$ and $f_{c_{p}}$ are LC/$\mu$C and SppC bunch frequencies. $f_{c}$ is expressed by $f_{c}=N_{b}f_{rep}$, where $N_{b}$ denotes number of bunches, $f_{rep}$ means revolution frequency for SppC/$\mu$C and pulse frequency for LC. In order to determine collision frequency of lp collider, minimum value should be chosen among lepton and hadron bunch frequencies. Some of these parameters can be rearranged in order to maximize $L_{lp}$ but one should note that there are main limitations due to beam-beam effects that should be kept in mind. While beam-beam tune shift affects proton and muon beams, disruption has influence on electron beams.
Disruption parameter for electron beam is given by:
$$\begin{aligned}
D_{x_{e}} & = & \frac{2\,N_{p}r_{e}\sigma_{z_{p}}}{\gamma_{e}\sigma_{x_{p}}(\sigma_{x_{p}}+\sigma_{y_{p}})}\label{eq:Denklem2}\end{aligned}$$
$\,$
$$D_{y_{e}}=\frac{2\,N_{p}r_{e}\sigma_{z_{p}}}{\gamma_{e}\sigma_{y_{p}}(\sigma_{y_{p}}+\sigma_{x_{p}})}$$
where, $r_{e}=2.82\times10^{-15}$ $m$ is classical radius for electron, $\gamma_{e}$ is the Lorentz factor of electron beam, $\sigma_{x_{p}}$ and $\sigma_{y_{p}}$ are horizontal and vertical proton beam sizes at IP, respectively. $\sigma_{z_{p}}$ is bunch length of proton beam. Beam-beam parameter for proton beam is given by:\
$$\xi_{x_{p}}=\frac{N_{l}r_{p}\beta_{p}^{*}}{2\pi\gamma_{p}\sigma_{x_{l}}(\sigma_{x_{l}}+\sigma_{y_{l}})}\label{eq:Denklem3}$$
$ $
$$\xi_{y_{p}}=\frac{N_{l}r_{p}\beta_{p}^{*}}{2\pi\gamma_{p}\sigma_{y_{l}}(\sigma_{y_{l}}+\sigma_{x_{l}})}$$
where $r_{p}$ is classical radius for proton, $r_{p}=1.54\times10^{-18}$ $m$, $\beta_{p}^{*}$ is beta function of proton beam at IP, $\gamma_{p}$ is the Lorentz factor of proton beam. $\sigma_{x_{l}}$ and $\sigma_{y_{l}}$ are horizontal and vertical sizes of lepton beam at IP, respectively.\
Beam-beam parameter for muon beam is given by:\
$$\xi_{x_{\mu}}=\frac{N_{p}r_{\mu}\beta_{\mu}^{*}}{2\pi\gamma_{\mu}\sigma_{x_{p}}(\sigma_{x_{p}}+\sigma_{y_{p}})}\label{eq:Denklem3}$$
$ $
$$\xi_{y_{\mu}}=\frac{N_{p}r_{\mu}\beta_{\mu}^{*}}{2\pi\gamma_{\mu}\sigma_{y_{p}}(\sigma_{y_{p}}+\sigma_{x_{p}})}$$
where $r_{\mu}=1.37\times10^{-17}$ $m$ is classical muon radius, $\beta_{\mu}^{*}$ is beta function of muon beam at IP, $\gamma_{\mu}$ is the Lorentz factor of muon beam. $\sigma_{x_{p}}$ and $\sigma_{y_{p}}$ are horizontal and vertical sizes of proton beam at IP, respectively.
ep option
---------
Preliminary study of CepC-SppC based e-p collider with $\sqrt{s}=4.1$ TeV and $L_{ep}=10^{33}$ $cm^{-2}s^{-1}$ has been performed in \[19\]. In this subsection, we consider ILC (International Linear Collider) \[20\] and PWFA-LC (Plasma Wake Field Accelerator - Linear Collider) \[21\] as a source of electron/positron beam for SppC based energy frontier ep colliders. Main parameters of ILC and PWFA-LC electron beams are given Table \[tab:tablo2\].
Beam Energy (GeV) $500$ $5000$
-------------------------------------------------------- --------- ----------------
Peak Luminosity ($10^{34}\,cm^{-2}s^{-1}$) $4.90$ $6.27$
Particle per Bunch ($10^{10}$) $1.74$ $1.00$
Norm. Horiz. Emittance ($\mu m$) $10.0$ $10.0$
Norm. Vert. Emittance (nm) $30.0$ $35.0$
Horiz. [$\beta$]{}[\*]{} amplitude function at IP (mm) $11.0$ $11.0$
Vert. [$\beta$]{}[\*]{} amplitude function at IP (mm) $0.23$ $0.099$
Horiz. IP beam size (nm) $335$ $106$
Vert. IP beam size (nm) $2.70$ $59.8$
Bunches per Beam $2450$ $1$
Repetition Rate (Hz) $4.00$ $5000$
Beam Power at IP (MW) $27.2$ $40$
Bunch Spacing (ns) $366$ $20$x$10^{4} $
Bunch length (mm) $0.225$ $0.02$
: Main parameters of the ILC (second column) and PWFA-LC (third column) electron beams.[]{data-label="tab:tablo2"}
It is seen that bunch spacings of ILC and PWFA-LC are much greater than SppC bunch spacing. On the other hand, transverse size of proton beam is much greater than transverse sizes of electron beam. Therefore, Eq. (1) for luminosity turns into:
$$L_{ep}=\frac{N_{e}N_{p}}{4\pi\sigma_{p}^{2}}f_{c_{e}}\label{eq:Denklem4}$$
For transversely matched electron and proton beams at IP, equations for electron beam disruption and proton beam tune shift become:
$$D_{e}=\frac{N_{p}r_{e}\sigma_{z_{p}}}{\gamma_{e}\sigma_{p}^{2}}\label{eq:Denklem5}$$
$$\xi_{p}=\frac{N_{e}r_{p}\beta_{p}^{*}}{4\pi\gamma_{p}\sigma_{p}^{2}}=\frac{N_{e}r_{p}}{4\pi\epsilon_{np}}\label{eq:Denklem6}$$
where $\epsilon_{np}$ is normalized transverse emittance of proton beam.
Using nominal parameters of ILC, PWFA-LC and SppC, we obtain values of L$_{ep}$, D$_e$ and ${\xi}_{p}$ parameters for LC$\otimes$SppC based ep colliders, which are given in Table \[tab:tablo3\]. The values for luminosity given in parantheses represent results of beam-beam simulations by ALOHEP software \[22\], which is being developed for linac-ring type ep colliders.
E$_{e}$, TeV E$_{p}$, TeV $\sqrt{s}$, TeV L$_{ep}$, $cm^{-2}s^{-1}$ D$_{e}$ $\xi_{p}$, $10^{-3}$
-------------- -------------- ----------------- --------------------------- --------- ----------------------
0.5 35.6 8.44 3.35 (6.64) x $10^{30}$ 0.537 0.5
0.5 68 11.66 2.69 (5.33) x $10^{31}$ 0.902 0.7
5 35.6 26.68 0.98 (1.94) x $10^{30}$ 0.054 0.3
5 68 36.88 0.78 (1.56) x $10^{31}$ 0.090 0.4
: Main parameters of LC$\otimes$SppC based ep colliders.[]{data-label="tab:tablo3"}
In order to increase luminosity of ep collisions LHeC-like upgrade of the SppC proton beam parameters have been used. Namely, $\beta$ function of proton beam at IP is arranged to be 7.5/2.4 times lower (0.1 m instead of 0.75/0.24 m) which corresponds to LHeC \[5\] and THERA \[23\] designs. This leads to increase of luminosity and D$_{e}$ by factor 7.5 and 2.4 for SppC with 35.6 TeV and 68 TeV proton beam, respectively. Results are shown in Table \[tab:tablo4\].
E$_{e}$, TeV E$_{p}$, TeV $\sqrt{s}$, TeV L$_{ep}$, $cm^{-2}s^{-1}$ D$_{e}$ $\xi_{p}$, $10^{-3}$
-------------- -------------- ----------------- --------------------------- --------- ----------------------
0.5 35.6 8.44 2.51 (4.41) x $10^{31}$ 4.03 0.5
0.5 68 11.66 6.45 (10.8) x $10^{31}$ 2.16 0.7
5 35.6 26.68 7.37 (13.3) x $10^{30}$ 0.403 0.3
5 68 36.88 1.89 (3.75) x $10^{31}$ 0.216 0.4
: Main parameters of LC$\otimes$SppC based ep colliders with upgraded $\beta$\*.[]{data-label="tab:tablo4"}
In principle “dynamic focusing scheme” \[24\] which was proposed for THERA, could provide additional factor of 3-4. Therefore, luminosity values exceeding $10^{32}$ $cm^{-2}s^{-1} $ can be achieved for all options. Concerning ILC$\otimes$SppC based ep colliders, a new scheme for energy recovery proposed for higher-energy LHeC (see Section 7.1.5 in \[5\]) may give an opportunity to increase luminosity by an additional order, resulting in L$_{ep}$ exceeding $10^{33}$ $cm^{-2}s^{-1} $. Unfortunately, this scheme can not be applied at PWFA-LC$\otimes $SppC.
$\mu$p option
-------------
Muon-proton colliders were proposed almost two decades ago: construction of additional proton ring in $\sqrt{s}$ = 4 TeV muon collider tunnel was suggested in \[25\], construction of additional 200 GeV energy muon ring in the Tevatron tunnel was considered in \[26\] and ultimate $\mu$p collider with 50 TeV proton ring in $\sqrt{s}$ = 100 TeV muon collider tunnel was suggested in \[27\]. Here, we consider construction of TeV energy muon colliders ($\mu$C) \[28\] tangential to the SppC. Parameters of $\mu$C are given in Table \[tab:tablo5\].
Keeping in mind that both SppC and $\mu$C have round beams, luminosity Eq. (1) turns to:
$$\begin{aligned}
L_{pp} & = & f_{pp}\frac{N_{p}^{2}}{4\pi\sigma_{p}^{2}}\label{eq:Denklem7}\end{aligned}$$
$$\begin{aligned}
L_{\mu\mu} & = & f_{\mu\mu}\frac{N_{\mu}^{2}}{4\pi\sigma_{\mu}^{2}}\label{eq:Denklem8}\end{aligned}$$
for SppC-$pp$ and $\mu$C, respectively. Concerning muon-proton collisions one should use larger transverse beam sizes and smaller collision frequency values. Keeping in mind that $f_{\mu\mu}$ is smaller than $f_{pp}$ by more than two orders, following correlation between $\mu p$ and $\mu\mu$ luminosities take place:
$$\begin{aligned}
L_{\mu p} & = & (\frac{N_{p}}{N_{\mu}})(\frac{\sigma_{\mu}}{max[\sigma_{p},\,\sigma_{\mu}]})^{2}L_{\mu\mu}\label{eq:Denklem9}\end{aligned}$$
Using nominal parameters of $\mu\mu$ colliders given in Table 5, parameters of the SppC based $\mu p$ colliders are calculated according to Eq. (\[eq:Denklem9\]) and presented in Table \[tab:tablo6\]. Concerning beam beam tune shifts, for round and matched beams Eqs. (4,5) and Eqs. (6,7) turns to:
$$\begin{aligned}
\xi_{p} = \frac{N_{\mu}r_{p}\beta_{p}^{*}}{4\pi\gamma_{p}\sigma_{\mu}^{2}} = \frac{N_{\mu}r_{p}}{4\pi\epsilon_{np}}\label{eq:Denklem10}\end{aligned}$$
and
$$\begin{aligned}
\xi_{\mu} = \frac{N_{p}r_{\mu}\beta_{\mu}^{*}}{4\pi\gamma_{\mu}\sigma_{p}^{2}} = \frac{N_{p}r_{\mu}}{4\pi\epsilon_{n\mu}},\label{eq:Denklem11}\end{aligned}$$
respectively.
As one can see from Table \[tab:tablo6\], where nominal parameters of SppC proton beam are used, $\xi_{p}$ is unacceptably high and should be decreased to 0.02 which seems acceptable for $\mu$p colliders \[26\]. According to Eq. (14), $\xi_{p}$ can be decreased, for example, by decrement of N$_{\mu}$ which leads to corresponding reduction of luminosity (three times and four times for $\mu$p 35.6 TeV and 68 TeV, respectively). Alternatively, crab crossing \[29\] can be used for decreasing of $\xi_{p}$ without change of the luminosity.
Beam Energy (GeV) $750$ $1500$
------------------------------------------------- ------------- ---------------
Circumference (km) $2.5$ $4.5$
Average Luminosity ($10^{34}\,cm^{-2}s^{-1}$) $1.25$ $4.4$
Particle per Bunch ($10^{12}$) $2$ $2$
Norm. Trans. Emitt. (mm-rad) $0.025$ $0.025$
[$\beta$]{}[\*]{} amplitude function at IP (cm) $1 (0.5-2)$ $0.5 (0.3-3)$
IP beam size ($\mu$m) $6$ $3$
Bunches per Beam $1$ $1$
Repetition Rate (Hz) $15$ $12$
Bunch Spacing (ns) $8300$ $15000$
Bunch length (cm) $1$ $0.5$
: Main parameters of the muon beams.[]{data-label="tab:tablo5"}
$E_{\mu}$, TeV $E_{p}$, TeV $\surd$S, TeV $L_{{\mu}p} $, $cm^{-2}s^{-1}$ $\xi_{\mu}$ $\xi_{p}$
---------------- -------------- --------------- -------------------------------- ----------------- -----------------
0.75 35.6 10.33 5.5 x $10^{32}$ 8.7 x $10^{-3}$ 6.0 x $10^{-2}$
0.75 68 14.28 12.5 x $10^{32}$ 8.7 x $10^{-3}$ 8.0 x $10^{-2}$
1.5 35.6 14.61 4.9 x $10^{32}$ 8.7 x $10^{-3}$ 6.0 x $10^{-2}$
1.5 68 20.2 42.8 x $10^{32}$ 8.7 x $10^{-3}$ 8.0 x $10^{-2}$
: Main parameters of SppC based $\mu$p colliders.[]{data-label="tab:tablo6"}
Ultimate $\mu$p option
----------------------
This option can be realized if an additional muon ring is constructed in the SppC tunnel. In order to estimate CM energy and luminosity of $\mu$p collisions we use muon beam parameters from \[30\], where 100 TeV center of mass energy muon collider with 100 km ring circumference have been proposed. These parameters are presented in Table \[tab:tablo7\].
CM energy, luminosity and tune shifts for ultimate $\mu$p collider are given in Table \[tab:tablo8\]. Again $\xi_{\mu}$ and $\xi_{p}$ can be decreased by lowering of $N_{p}$ and $N_{\mu}$ respectively (which lead to corresponding decrease of luminosity) or crab crossing can be used without change of the luminosity.
Beam Energy (TeV) $50$
------------------------------------------------- --------
Circumference (km) $100$
Average Luminosity ($10^{34}\,cm^{-2}s^{-1}$) $100$
Particle per Bunch ($10^{12}$) $0.80$
Norm. Trans. Emitt. (mm-mrad) $8.7$
[$\beta$]{}[\*]{} amplitude function at IP (mm) $2.5$
IP beam size ($\mu$m) $0.21$
Bunches per Beam $1$
Repetition Rate (Hz) $7.9$
Bunch Spacing ($\mu$s) $333$
Bunch length (mm) $2.5$
: Main parameters of the ultimate muon beam.[]{data-label="tab:tablo7"}
$E_{\mu}$, TeV $E_{p}$, TeV $\surd$S, TeV $L_{{\mu}p} $, $cm^{-2}s^{-1}$ $\xi_{\mu}$ $\xi_{p}$
---------------- -------------- --------------- -------------------------------- ----------------- -----------------
50 68 116.6 1.2 x $10^{33}$ 2.6 x $10^{-2}$ 3.5 x $10^{-2}$
: Main parameters of the ultimate SppC based ${\mu}$p collider.[]{data-label="tab:tablo8"}
Physics
=======
In order to evaluate physics search potential of the SppC based lp colliders we consider two phenomena, namely, small Bj[ö]{}rken $x$ region is considered as an example of the SM physics and resonant production of color octet electron and muon is considered as an example of the BSM physics.
Small Bj[ö]{}rken $x$
---------------------
As mentioned above, investigation of extremely small $x$ region ($x$ $<$ $10^{-5}$) at sufficiently large $Q^{2}$ ($>$ 10 $GeV^{2}$), where saturation of parton density should manifest itself, is crucial for understanding of QCD basics. Smallest achievable $x$ at lp colliders is given by $Q^{2}$/S. For LHeC with $\sqrt{s}=1.3$ TeV minimal acvievable value is $x$ = 6 x $10^{-6}$. In Table \[tab:tablo9\], we present smallest $x$ values for different SppC based lepton-proton colliders (E$_{p}$ is chosen as 68 TeV). It is seen that proposed machines has great potential for enligthening of QCD basics.
E$_{l}$ (TeV) 0.5 5 1.5 50
--------------- ----------------- ----------------- ----------------- ------------------
$x$ $7$ x $10^{-8}$ $7$ x $10^{-9}$ $2$ x $10^{-8}$ $7$ x $10^{-10}$
: Attainable Bj[ö]{}rken $x$ values at $Q^{2}=10$ $GeV^{2}$.[]{data-label="tab:tablo9"}
Color octet leptons
-------------------
Color octet leptons ($l_{8}$) are predicted in preonic models with colored preons \[31\]. There are various phenomenological studies on $l_{8}$ at TeV energy scale colliders \[32-39\]. Resonant production of color octet electron ($e_{8}$) and muon ($\mu_{8}$) at the FCC based lp colliders have been considered in \[40\] and \[41\] respectively. Performing similar analyses for SppC based lp colliders we obtain mass discovedynamicry limits for $e_{8}$ and $\mu_{8}$ in $\Lambda = M_{l_{8}}$ case (where $\Lambda$ is compositeness scale) which are presented in Figs 2 and 3, respectively. Discovery mass limit value for LHC and SppC are obtained by rescaling ATLAS/CMS second generation LQ results \[42, 43\] using the method developed by G. Salam and A. Weiler \[44\]. For lepton colliders, it is obvious that discovery mass limit for pair production of $l_{8}$ are approximately half of CM energies. It is seen that $l_{8}$ search potential of SppC based lp colliders overwhelmingly exceeds that of LHC and lepton colliders. Moreover lp colliders will give an opportunity to determine compositeness scale (for details see \[40, 41\]).
It should be noted that FCC/SppC based lp colliders has great potential for search of a lot of BSM phenomena, such as excited leptons (see \[45\] for ${\mu}^*$), contact interactions, R-parity violating SUSY etc.
Conclusion
==========
It is shown that construction of linear $e^{+}e^{-}$colliders (or dedicated linac) and muon colliders (or dedicated muon ring) tangential to the SppC will give opportunity to handle lepton-proton collisions with multi-TeV CM energies and sufficiently high luminosities. Concerning SM physics, these machines will certainly shed light on QCD basics. BSM search potential of lp colliders essentially exceeds that of corresponding lepton colliders. Also these type of colliders exceed the search potential of the SppC itself for a lot of BSM phenomena.
Acceleration of ion beams at the SppC will give opportunity to provide multi-TeV center of mass energy in eA and $\mu$A collisions. In addition, electron beam can be converted to high energy photon beam using Compton backdynamic-scattering of laser photons which will give opportunity to construct LC$\bigotimes$SppC based $\gamma$p and $\gamma$A colliders. Studies on these topics are ongoing.
In conclusion, systematic study of accelerator, detector and physics search potential issues of the SppC based ep, eA, $\gamma$p, $\gamma$A, $\mu$p and $\mu$A colliders are essential to foreseen the future of particle physics. Certainly, realization of these machines depend on the future results from the LHC as well as FCC and/or SppC.
Acknowledgments {#acknowledgments .unnumbered}
===============
This study is supported by TUBITAK under the grant no 114F337.
References {#references .unnumbered}
==========
[1]{}
R. Hofstadter and R. W. McAllister, Electron Scattering from the Proton,“ Physical Reviev, vol. 98, no. 1, pp. 217, 1955. J. I. Friedman and H. W. Kendall, Deep inelastic electron scattering,” Annual Review of Nuclear Science, vol. 22, no. 1, pp. 203-254, 1972. H. Abramowicz and C. C. Allen, HERA collider physics,“ Reviews of Modern Physics, vol. 71, no. 5, pp. 1275-1409, 1999. M. Klein and R. Yoshida, Collider physics at HERA,” Progress in Particle and Nuclear Physics, vol. 61, no. 2, pp. 43-393, 2008. J. L. Abelleira Fernandez, C. Adolphsen, A. N. Akay et al., A Large Hadron Electron Collider at CERNReport on tA beam driven plasma-wakefield linear collider from Higgs factory to multi-TeV,he Physics and Design Concepts for Machine and Detector,“ Journal of Physics G: Nuclear and Particle Physics, vol. 39, no. 7, Article ID 075001, 2012. S. Sultansoy, Linac ring type colliders: Second way to TeV scale,” The European Physical Journal C-Particles and Fields, vol. 33, pp. 1064-1066, 2004. A. N. Akay, H. Karadeniz and S. Sultansoy, Review of Linac-Ring Type Collider Proposals,“ International Journal of Modern Physics A, vol. 25, no. 25, pp. 4589-4602, 2010. Z. Z. Aydin, A. K. Ciftci and S. Sultansoy, Linac-ring type ep machines and $\gamma$p colliders based on them,” Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 351, no. 2-3, pp. 261-265, 1994. A. K. Ciftci, S. Sultansoy, S. Turkoz and O. Yavas, Main parameters of TeV energy $\gamma$p colliders,“ Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 365, no. 2-3, pp. 317-328, 1995. H. Aktas, N. Buget, A. K. Ciftci, N. Meric, S. Sultansoy and O. Yavas, New tool for “old” nuclear physics: FEL $\gamma$-nucleus colliders,” Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 428, no. 2, pp. 271-275, 1999. A. K. Ciftci, S. Sultansoy and O. Yavas, TESLA\*HERA based $\gamma$p and $\gamma$A colliders,“ Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 472, no. 1-2, pp. 72-78, 2001. H. Braun, R. Corsini, E. Guliyev et al., CLIC-LHC-based FEL-nucleus collider: Feasibility and physics search potential,” Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 552, no.3, pp. 440-448, 2005. H. Aksakal, A. K. Ciftci, Z. Nergiz et al., Conversion efficiency and luminosity for gamma-proton colliders based on LHC-CLIC or LHC-ILC QCD explorer scheme,“ Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 576, no. 2-3, pp. 287-293, 2007. E. Arikan and H. Aksakal, Positron source investigation by using CLIC drive beam for Linac-LHC based $e^{+}$p collider,” Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 683, pp. 63-70, 2012. FCC, https://fcc.web.cern.ch. F. Su, J. Go, M. Xiao and D. Wong, Method study of parameter choice for a circular proton-proton collider,“ Chinese Physics C, vol. 40, no. 1, pp. 103-109, 2016. F. Su, S. Bai, T. Bian et al., SPPC Parameter Choice and Lattice Design,” 7th International Particle Accelerator Conference (IPAC’16), pp. 1400-1402, Busan, Korea, May 2016. Y.C. Acar, A. N. Akay, S. Beser et al., FCC Based Lepton-Hadron and Photon-Hadron Colliders: Luminosity and Physics,“ arXiv preprint arXiv:1608.02190, 2016. Y. Zhang and Y. Peng, A High Energy e-p/A Collider Based on CepC-SppC,” 6th International Particle Accelerator Conference (IPAC’15), Richmond, USA, May 2015. C. Adolphsen, M. Barone, B. Barish et al., The International Linear Collider Technical Design Report - Volume 3.II: Accelerator Baseline Design,“ https://arxiv.org/abs/1306.6328. J. P. Delahaye, E. Adli, S. J. Gessner et al., A beam driven plasma-wakefield linear collider from Higgs factory to multi-TeV,” in Proceedings of the 5th International Particle Accelerator Conference (IPAC ’14), pp. 3791-3793, Dresden, Germany, June 2014. ALOHEP, http://alohep.hepforge.org. U. Katz, M. Klein, A. Levy, and S. Schlenstedt, The THERA Book,“ DESY-LC-REV-2001-062 2001. R. Brinkmann and M. Dohlus, A method to overcome the bunch length limitations on $\beta_{p}^{*}$ for electron-proton colliders,” DESY-M-95-11 1995. I. F. Ginzburg, Physics at future ep, $\gamma$p (linac-ring) and $\mu$p colliders,“ Turkish Journal of Physics, vol. 22, no. 7, pp. 607-610, 1998. V. D. Shiltsev, An asymmetric muon proton collider: luminosity consideration,” Particle Accelerator Conference Proceedings, vol. 1, pp. 420-421, Vancouver, British Columbia, Canada, 1997. S. Sultansoy, The post-HERA era: brief review of future lepton-hadron and photon-hadron colliders,“ arXiv preprint arXiv:9911417v2, 1999. J. P. Delahaye, C. Ankenbrandt, A. Bogacz et al., Enabling intensity and energy frontier science with a muon accelerator facility in the U.S.,” arXiv preprint arXiv:1308.0494v2, 2013. S. Verdu-Andres, S. Belomenstnykh, I. Ben-Zvi et al., Crab cavities for colliders: past, present and future,“ Nuclear and Particle Physics Proceedings, vol. 273, pp. 193-197 , 2016. B. J. King, Parameter Sets for 10 TeV and 100 TeV Muon Colliders, and their Study at the HEMC’99 Workshop,” arXiv preprint arXiv:0005008, 2000. I. A. D’Souza and C. S. Kalman, PREONS: Models of Leptons, Quarks and Gauge Bosons as Composite Objects,“ World Scientific, Singapore, 1992. A. Celikel, M. Kantar and S. Sultansoy, A Search for sextet quarks and leptogluons at the LHC,” Physics Letters B, vol. 443, no. 1, pp. 359-364, 1998. T. Mandal and S. Mitra, Probing color octet electrons at the LHC,“ Physical Review D, vol. 87, no. 9, Article ID 095008, 2013. D. Gonçalves-Netto, D. Lopez-Val, K. Mawatari et al., Looking for leptogluons,” Physical Review D, vol. 87, no. 9, Article ID 094023, 2013. T. Jelinski and D. Zuridov, Leptogluons in dilepton production at LHC,“ Acta Phys. Pol. B, vol. 46, no. 11, pp. 2185, 2015. T. Mandal, S. Mitra and S. Seth, Probing compositeness with the CMS eejj$\&$ejj data,” Physics Letters B, vol. 758, pp. 219-25, 2016. M. Sahin, S. Sultansoy and S. Turkoz, Resonant production of color octet electron at the LHeC,“ Phyics Letters B, vol. 689, no. 4, pp. 172-176, 2010. M. Sahin, Resonant production of spin 3/2 color octet electron at the LHeC,” Acta Phys. Pol. B, vol. 45, pp. 1811, 2014. A. N. Akay, H. Karadeniz, M. Sahin and S. Sultansoy, Indirect search for color octet electron at next-generation linear collider,“ Europhysics Letters, vol. 95, no. 3, Article ID 31001, 2011. Y. C. Acar, U. Kaya, B. B. Oner, and S. Sultansoy, Color octet electron search potential of the FCC based e-p colliders,” Journal of Physics G: Nuclear and Particle Physics, vol. 44, no. 4, Article ID 045005, 2017. Y. C. Acar, U. Kaya and B. B. Oner, Resonant Production of Color Octet Muons at the Future Circular Collider Based Muon-Proton Colliders,“ arxiv preprint arXiv:1703.04030v2, 2017. The ATLAS Collaboration, Search for scalar leptoquarks in pp collisions at $\surd$S=13 TeV with the ATLAS experiment,” New J. Phys., vol. 18, Article ID 093016, 2016. The CMS Collaboration, Search for pair production of first and second generation leptoquarks in proton-proton collisions at $\surd$S= 8 TeV,“ Physical Review D, vol. 93, no. 3, Article ID 032004, 2016. G. Salam and A. Weiler, The Collider Reach project,” http://collider-reach.web.cern.ch/collider-reach.
A. Caliskan, S. O. Kara, A. Ozansoy, Excited Muon Searches at the FCC-Based Muon-Hadron Colliders," Advances in High Energy Physics, vol. 2017, Article ID 1540243, 2017.
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'VPipe.bib'
---
=1
Acknowledgement {#acknowledgement .unnumbered}
===============
We thank the numerous team members who have contributed to the success of the SDO mission and particularly to the HMI instrument. This work was supported by NASA Contract NAS5-02139 (HMI) to Stanford University. Some of the research described here was carried out by staff of the Jet Propulsion Laboratory, California Institute of Technology. HAO/NCAR is supported by the National Science Foundation. Efforts at NWRA were also supported through NASA Contract NNH09CF22C and by PO\# NNG12PP28D/C\# GS-23F-0197P from NASA/Goddard Space Flight Center. J. Schou acknowledges support from EU FP7 Collaborative Project [*Exploitation of Space Data for Innovative Helio- and Asteroseismology*]{} (SPACEINN).
| {
"pile_set_name": "ArXiv"
} |
2.1cm
**Stress Tensors in p-adic String Theory and Truncated OSFT**
Haitang Yang
*Center for Theoretical Physics*
*Massachussetts Institute of Technology,*
*Cambridge, MA 02139, USA*
E-mail: [email protected]
**Abstract**
We construct the stress tensors for the p-adic string model and for the pure tachyonic sector of open string field theory by naive metric covariantization of the action. Then we give the concrete energy density of a lump solution of the p-adic model. In the cubic open bosonic string field theory, we also give the energy density of a lump solution and pressure evolution of a rolling tachyon solution.
Introduction
============
Much work has been devoted to looking for solutions in string field theory (SFT). Generally speaking, physicists are concerned with two kinds of solutions with different properties. One kind of solutions are the time independent ones which represent the tachyon vacuum or lower dimensional D-branes [@9902105]-[@0008252]. Initiated by Sen [@0203211], time dependent rolling tachyon solutions have recently attracted much attention [@0203265]-[@0209122]. Studying rolling tachyon solutions can give us information about how the tachyon approaches the tachyon vacuum. At the same time, the p-adic model [@BFOW], which exhibits a lot of properties of string field theory, is also of interest. In this model, the potential has a stable vacuum and a tachyon. Studying the dynamics of the tachyon may suggest to us what happens in the same situation for the SFT. Furthermore, one also has lump solutions in the p-adic theory which are identified as lower dimensional D-branes [@0003278].
In [@0207107], Moeller and Zwiebach discussed how to construct the stress tensor for the rolling tachyon solution in the p-adic model. They obtained an unambiguous expression for the energy through a generalized Noether procedure. This analysis could not be extended to the pressure calculation, however, as there are ambiguities in that case. Instead, they included the metric in the action and used the definition of stress tensor in general relativity to calculate the pressure. Then they constructed the rolling tachyon solutions for both the p-adic model and open string field theory (OSFT) in the form of series expansions. After that, they calculated the pressure evolution in the p-adic string case.
It is of interest to consider the stress tensor in the case when the scalar field in the p-adic model depends on all the coordinates. Especially, for a lump solution, what is the profile of the energy distribution along the spatial coordinate? Is it the same as what we expect intuitively? Furthermore, in OSFT, it is important to know if the profile of the energy density has the same properties as that in p-adic string theory. Moeller and Zwiebach showed in [@0207107] that the pressure of the rolling solution in p-adic model does not vanish at large times. For the rolling solution in OSFT, it is of interest to test if one gets vanishing pressure asymptotically or not.
In this paper, we first give the stress tensor in a general form for the p-adic model. When our results are specialized to the time dependent solution in p-adic model, they reproduce the results in [@0207107]. A nontrivial lump solution in p-adic model was given in [@BFOW], [@0003278]. We construct the energy density of this solution and compare it with that of the lump solution of ordinary $\phi^3$ field theory. We find that these two energy densities have similar spatial profiles. Section $3$ is devoted to the case of the pure tachyon field in OSFT. We again construct the stress tensor in a general form. The energy density of a solitonic solution [@0005036] is then constructed in subsection $3.1$. Finally we calculate the pressure evolution of a rolling tachyon solution [@0207107].
p-adic String Theory Case
=========================
In this section, we first construct the stress tensor of the p-adic string theory by varying the metric. We will find that the expression is exactly the same as the one obtained in [@0207107] if we constrain scalar field to only depend on time. We will also consider the case where the tachyon scalar only depends one spatial coordinate. In that situation, one nontrivial solitonic solution was already given [@BFOW], [@0003278]. We then calculate the energy density of that solution. The results show that the total energy, integrated over all space, perfectly agrees with the D24 brane tension as expected. The spatial profile of this energy density looks very like the one of the solitonic solution of ordinary $\phi^3$ field theory.
Stress Tensor for p-adic model
------------------------------
The p-adic string theory is defined by the action: $$\label{pmodel} S = \int {d^d x{{\cal L}}= \frac{1}{{g_p ^2 }}} \int {d^d
x} \left[ { - \frac{1}{2}\phi p^{ - \frac{1}{2}\Box } \phi +
\frac{1}{{p + 1}}\phi ^{p + 1} } \right],\hspace{3mm}
\frac{1}{g_p^2}=\frac{1}{g^2}\frac{p^2}{p-1},$$ where $\phi(x)$ is a scalar field, $p$ is a prime integer and $g$ is the open string coupling constant. Though the theory makes sense even as $p\rightarrow 1$, in most cases, we will consider $p\geq 2$ in this paper. In this action, there is an infinite number of both time and spatial derivatives. One defines: $$p^{ - \frac{1}{2}\Box } \equiv \exp \left( { - \frac{1}{2}\ln
p\Box } \right) = \sum\limits_{n = 0}^\infty {\left( { -
\frac{1}{2}\ln p} \right)} ^n \frac{1}{{n!}}\Box
^{n},$$ and $$\square = - \frac{{\partial ^2 }} {{\partial t^2 }} + \nabla
^2.$$ Now we include the metric in the action [@0207107]: $$\begin{aligned}
\label{gpmodel} S =S_1+S_2&=& \frac{1}{{g_p ^2 }}\int {d^d x} \sqrt { -
g} \left[ { - \frac{1}{2}\phi ^2 + \frac{1}{{p + 1}}\phi ^{p + 1}
} \right]\nonumber\\
&&- \frac{1}{{2g_p ^2 }}\sum\limits_{l = 1}^\infty {\left( { -
\frac{1}{2}\ln p} \right)^l \frac{1}{{l!}}} \int {d^d x} \sqrt { -
g} \phi \Box ^{l} \phi,\end{aligned}$$ where we have split the action into two parts: $S_1$ represents the potential and $S_2$ represents the kinetic term. After introduction of the metric $\square$ becomes the covariant D’Alembertian.
$$\begin{aligned}
\label{covd} B_l\equiv\int d^dx \sqrt{-g} \phi\, \Box^l \,\phi &=&
\int d^dx\,\, \phi \,\, \partial_{\mu_1}
\sqrt{-g}g^{\mu_1\nu_1}\partial_{\nu_1} \,\,
\,{1\over \sqrt{-g}} \partial_{\mu_2}
\sqrt{-g}g^{\mu_2\nu_2}\partial_{\nu_2} \,\, \nonumber\\
&& \cdots {1\over \sqrt{-g}} \partial_{\mu_l}
\sqrt{-g}g^{\mu_l\nu_l}\partial_{\nu_l} \, \phi \,.\end{aligned}$$
The stress tensor is given by: $$T_{\alpha \beta } = \frac{2} {{\sqrt { - g} }}\frac{{\delta
S}} {{\delta g^{\alpha \beta } }}.$$ The variation of the potential $S_1$ in (\[gpmodel\]) contributes: $$\label{vs1} \frac{2} {{\sqrt { - g} }}\frac{{\delta S_1}} {{\delta
g^{\alpha \beta } }}=- \frac{1} {{g_p ^2 }}\left( { - \frac{1}
{2}\phi ^2 + \frac{1} {{p + 1}}\phi ^{p + 1} } \right)g_{\alpha
\beta },$$ where we have set the metric to be flat with signature $(-,+,+\cdots +)$ after the variation and we will use the same convention in the rest of this paper. As for the variation of the kinetic term $S_2$ in (\[gpmodel\]), from (\[covd\]), we need to vary both factors of $\sqrt{-g}$ and $g^{\mu_i\nu_i}$ with respect to $g^{\alpha\beta}$. First consider varying factors of $\sqrt{-g}$ in (\[covd\]) with respect to $g^{\alpha\beta}$: $$\begin{aligned}
\frac{\delta B_l}{\delta \sqrt{-g}} \frac{\delta\sqrt{-g}}{\delta
g^{\alpha\beta}}&=& g^{\mu _1 \nu _1 } g^{\mu _2 \nu _2 } \cdots
g^{\mu _l \nu _l } ( \phi _{\mu _1 } \phi _{\nu _1 \mu _2 \nu _2
\cdots \mu _l \nu _l } + \phi _{\mu _1 \nu _1 } \phi _{\mu _2 \nu
_2 \cdots \mu _l \nu _l } + \cdots \nonumber\\
&&\cdots+ \phi _{\mu _1 \nu _1 \mu _2 \nu _2 \cdots \mu _l } \phi
_{\nu _l } )g_{\alpha \beta },\end{aligned}$$ with the definition: $$\phi _{\mu _1 \nu _1 \mu _2 \nu _2 \cdots \mu _l \nu _l } \equiv
\partial_{\mu _1 }\partial_{\nu
_1 }\partial_{ \mu _2 }\partial_{\nu _2 }\cdots \partial_{\mu _l
}\partial_{\nu _l }\phi (x).$$ The variation of the factors of $g^{\mu_i\nu_i}$ in (\[covd\]) with respect to $g^{\alpha\beta}$ contributes: $$\begin{aligned}
\frac{\delta B_l}{\delta g^{\mu_i\nu_i}}\frac{\delta
g^{\mu_i\nu_i}} {\delta g^{\alpha\beta}}&=&- 2g^{\mu _1 \nu _1 }
g^{\mu _2 \nu _2 } \cdots g^{\mu _{l - 1} \nu _{l - 1} } \Big(\phi
_\alpha \phi _{\beta \mu _1 \nu _1 \mu _2 \nu _2 \cdots \mu _{l -
1} \nu _{l - 1} } \nonumber \\ &&+ \phi _{\alpha \mu _1 \nu _1 }
\phi _{\beta \mu _2 \nu _2 \cdots \mu _{l - 1} \nu _{l - 1} } +
\cdots + \phi _{\alpha \mu _1 \nu _1 \mu _2 \nu _2 \cdots \mu _l
} \phi _\beta \Big).\end{aligned}$$ So, we can calculate $\delta S_2$. Finally, the stress tensor is: $$\begin{aligned}
\label{padicT}T_{\alpha\beta}&=&- \frac{1} {{g_p ^2 }}\left(
{ - \frac{1} {2}\phi ^2 + \frac{1} {{p + 1}}\phi ^{p + 1} }
\right)g_{\alpha
\beta }\nonumber\\
&&-\frac{1}{{2g_p ^2 }}\sum\limits_{l = 1}^\infty {\left( { -
\frac{1}{2}\ln p} \right)^l } \frac{1}{{l!}}\bigg\{g^{\mu _1 \nu
_1 } g^{\mu _2 \nu _2 } \cdots g^{\mu _l \nu _l }\Big( \phi _{\mu _1
} \phi _{\nu _1 \mu _2 \nu _2 \cdots \mu _l \nu _l }\nonumber\\
&& + \phi_{\mu _1 \nu _1 } \phi _{\mu _2 \nu _2
\cdots \mu _l \nu _l } +
\cdots+ \phi _{\mu _1 \nu _1 \mu _2 \nu _2 \cdots \mu _l } \phi
_{\nu _l } \Big )g_{\alpha \beta }\nonumber\\
&& - 2g^{\mu _1 \nu _1 } g^{\mu _2 \nu
_2 } \cdots g^{\mu _{l - 1} \nu _{l - 1}}\Big(\phi
_\alpha \phi _{\beta \mu _1 \nu _1 \mu _2 \nu _2 \cdots \mu _{l -1}
\nu _{l - 1} }\nonumber\\
&& + \phi _{\alpha \mu _1 \nu _1 } \phi _{\beta \mu _2 \nu _2
\cdots \mu _{l - 1} \nu _{l - 1} }+ \cdots + \phi _{\alpha \mu _1
\nu _1 \mu _2 \nu _2 \cdots \mu _l } \phi _\beta \Big
)\bigg\}.\end{aligned}$$ If $\phi(x)$ is only time dependent, in (\[padicT\]), each $g^{\mu_i\nu_i}$ contributes one ‘$-$’ sign and the second term in the sum survives only for the component $T_{00}$. This gives the same results as in [@0207107].
One can also use the following identity [^1] $$\delta e^A=\int_0^1 dt\, e^{tA}(\delta A) e^{(1-t)A}$$ to get an alternative “closed” form of the stress tensor, compared with the series expression (\[padicT\]): $$\begin{aligned}
\label{padicT2} T_{\alpha \beta } &=&
\frac{{g_{\alpha \beta } }} {{2g_p ^2 }}\bigg\{\phi e^{ - k\square
} \phi - \frac{2} {{p + 1}}\phi ^{p + 1} + k\int\limits_0^1 {dt}
(e^{ - kt\square } \phi )(\square e^{ - k(1 - t)\square } \phi
)\nonumber\\
&&+ k\int\limits_0^1 {dt} (\partial _\mu e^{ - kt\square } \phi )
(\partial ^\mu e^{ - k(1 - t)\square } \phi )\bigg\}\nonumber\\
&&- \frac{k} {{g_p ^2 }}\int\limits_0^1 {dt} (\partial _\alpha e^{
- kt\square } \phi )(\partial _\beta e^{ - k(1 - t)\square }
\phi),\end{aligned}$$ where $k\equiv \frac{1}{2}\ln p$.
In the case that $\phi(x)$ only depends on one spatial coordinate, say $x\equiv x^{25}$, the last term in the right hand side of (\[padicT2\]) vanishes for all the components except for $T_{25,25}$. The energy density is
$$\begin{aligned}
\label{peng} E(x)&=&T^0_0=
\frac{1} {{2g_p ^2 }}\bigg\{\phi e^{ - k\partial^2 } \phi -
\frac{2} {{p + 1}}\phi ^{p + 1} + k\int\limits_0^1 {dt} (e^{ -
kt\partial^2 } \phi )(\partial^2 e^{ - k(1 - t)\partial^2 } \phi
)\nonumber\\
&&+ k\int\limits_0^1 {dt} (\partial e^{ - kt\partial^2 } \phi )
(\partial e^{ - k(1 - t)\partial^2 } \phi )\bigg\},\end{aligned}$$
where $\partial^2\equiv\frac{\partial^2}{\partial x^2}$.
Energy of The Lump Solution
---------------------------
There are some previously known solutions for the p-adic model [@BFOW], [@0003278]. One of them is the lump solution: $$\label{plump} \phi (x) = p^{\frac{1}{{2(p - 1)}}} \exp \left( { -
\frac{1}{2}\frac{{p - 1}}{{p\ln p}}x^2 }\right).$$ This solution is interpreted as a D24-brane, where $x$ is the coordinate transverse to the brane. This solution can be generalized to lower dimensional branes [@0003278]. The D-brane tension of this solution is: $$\begin{aligned}
\label{tension} {{\cal T}}_{24}&=&-\int dx {{\cal L}}(\phi(x))
=-\int dx\frac{1}{2g_p^2}\frac{1-p}{1+p}\phi^{(p+1)}(x)\nonumber\\
&=&\frac{1}{g_p^2}\frac{p-1}{2(p+1)}p^{\frac{p}{p-1}}\sqrt\frac{2\pi\ln
p}{p^2-1}.\end{aligned}$$ Using the identity $$\exp\left(-a\frac{d^2}{dx^2}\right)\exp(-bx^2)=\frac{1}
{\sqrt{1-4ab}}\exp\left(-\frac{bx^2}{1-4ab}\right),$$ from (\[peng\]), we can write down the energy density: $$\label{pengs} E(x)=\frac{p-1}{p+1}\sqrt{\frac{2\pi}{(p^2-1)\ln
p}}\, p^{\frac{p}{p-1}}|x| \,{\rm
Erf}\left[\frac{p-1}{p+1}\sqrt{\frac{p^2-1}{2p\ln p}}\,|x|\right]
e^{-\frac{2(p-1)x^2}{(p+1)\ln p}},$$ where ${\rm Erf}[x]\equiv\frac{2}{\sqrt\pi}\int_0^x
dt\,\exp(-t^2)$ is the error function. In Figure \[padiclumpE\], we plot this energy density (the solid line) for $p=2$. At $x=0$ and $x\rightarrow\pm\infty$, this energy density vanishes. By solving $\frac{d}{dx}E(x)=0$ numerically as $p=2$, one can see the energy reaches its maxima at $x\approx\pm0.9997$. From (\[pmodel\]), the potential is $$\frac{1}{2}\phi^2-\frac{1}{p+1}\phi^{p+1},$$ so, the D-brane vacuum is at $\phi=1$. Moreover, from (\[plump\]), one gets $\phi=1$ at $x=\pm\sqrt 2 \ln2\approx \pm
0.9803$, which are close to the locations where the energy gets its maxima.
The lump solution (\[plump\]) we are considering here, as we mentioned at the beginning of this subsection, is interpreted as a D-24 brane sharply localized on the hyperplane $x=0$. Therefore, intuitively one may expect the energy to be sharply localized around $x=0$. But from figure \[padiclumpE\], one can see that the energy is somewhat localised around $x\approx\pm 0.9997$ and reaches a local minimum at $x=0$.
The total energy is: $$\label{PTE} \int\limits_{-\infty}^{\infty} dx\,
E(x)=\frac{1}{g_p^2}\frac{p-1}{2(p+1)}p^{\frac{p}{p-1}}\sqrt\frac{2\pi\ln
p}{p^2-1}$$ which is exactly the same as (\[tension\]). In the limit $p\rightarrow 1$, $E(x)$ becomes: $$\label{limite}\mathop {\lim }\limits_{p \to 1} E(x) =
\frac{1}{{2g^2 }}x^2 \exp (1 - x^2 ).$$ On the other hand, from (\[pmodel\]), as $p\rightarrow 1$, the action becomes: $$S=\frac{1}{2g^2}\int d^d x \bigg( \frac{1}{2}\phi\square\phi
-\frac{1}{2}\phi^2+\phi^2 \ln\phi\bigg).$$ This action has a lump solution: $$\phi(x)=\exp\Big(\frac{1}{2}(1-x^2)\Big),$$ whose energy density is exactly the same as (\[limite\]).
This energy density looks very similar to that of the ordinary $\phi^3$ field theory with coupling constant $g_0$ and unit mass [@0008227]: $$S=\frac{1}{g_0^2}\int d^d x\bigg\{\frac{1}{2}(\partial\phi)^2-\frac{1}{2}
\phi^2+\frac{1}{3}\phi^3\bigg\},$$ which has the lump solution: $$\label{phi3lump}\phi (x)=\frac{3}{2}(1-{\rm tanh}^2\frac{x}{2}),$$ with energy density $$E(x)=\frac{1}{g_0^2}\frac{9}{4}{\rm sech}^4\frac{x}{2} {\rm
tanh}^2\frac{x}{2},$$ which is plotted in Figure 1 (dashed line).
The Pure Tachyon Field of String Field Theory Case
==================================================
When we expand the string field in the Hilbert space of the first quantized string theory, we can read off the action of the pure tachyonic cubic string field theory. As in the last section, we include the metric in the action and convert all the ordinary derivatives to covariant ones. Variations of the metric again give the stress tensor. Then we calculate the energy density of the lump solution given in [@0005036] and the pressure of the rolling tachyon solution given in [@0207107].
Stress Tensor for the Tachyon field in SFT
------------------------------------------
Firstly, we write down the pure tachyonic action of the cubic SFT. From Sen’s conjecture [@9902105], we should add the D-brane tension into the SFT action to cancel the negative energy due to the tachyon. We know that after adding the D-brane tension term to the potential of the cubic SFT, the local minimum of the new potential vanishes [@9912249]. In the same spirit, here we should add a term $\frac{1}{6} K^{-6}$ to the potential to set the local minimum of the potential to zero. $$\label{sftaction} S = \frac{1}{{g_0 ^2 }}\int {d^d x} \left(
{\frac{1}{2}\phi ^2 - \frac{1}{2}(\partial \phi )^2 -
\frac{1}{3}K^3 \widetilde\phi ^3 -\frac{1}{6} K^{-6}} \right),$$ where $$\label{dphitild} \widetilde\phi = \exp \left( {\ln K\square }
\right)\phi (x) = K^\square \phi (x).$$ $g_0$ is the open bosonic string coupling constant and $K = 3\sqrt
3 /4$. $\square$ is defined as in the last section. The equation of motion from this action is: $$K^{ - 2\square } (1 + \square )\widetilde\phi = K^3
\widetilde\phi ^2.$$ In order to separate the term without derivatives from $\widetilde\phi (x)$, we define: $$\begin{aligned}
\label{dpsi}\psi (x) &=& \widetilde\phi (x) - \phi (x) =
\sum\limits_{l = 1}^\infty {\frac{{\left( {\ln K} \right)^l }}
{{l!}}} \square ^l \phi (x)\nonumber\\
&=& \sum\limits_{l = 1}^\infty {\frac{{\left( {\ln K} \right)^l
}} {{l!}}} \frac{1} {{\sqrt { - g} }}\partial _{\mu _1 } \sqrt { -
g} g^{\mu _1 \nu _1 } \partial _{\nu _1 } \frac{1} {{\sqrt { - g}
}}\partial _{\mu _2 } \sqrt { - g} g^{\mu _2 \nu _2 } \partial
_{\nu _2 } \cdots\nonumber\\
&&\cdots \frac{1} {{\sqrt { - g} }}\partial _{\mu _l } \sqrt { -
g} g^{\mu _l \nu _l } \partial _{\nu _l } \phi (x),\end{aligned}$$ where in the last step, we have written the expression in the covariant form. For an arbitrary differentiable function $f(x)$, $$\label{general}\int {d^d xf(x)} \frac{\delta\psi (x)}{\delta
g^{\alpha \beta } }= \frac{1} {2}f\psi g_{\alpha \beta
}+A_{\alpha\beta}(f)$$ where $$\begin{aligned}
\label{defineA} A_{\alpha\beta}(f)&=& \frac{1}
{2}g_{\alpha \beta } \sum\limits_{l = 1}^\infty {\frac{{\left(
{\ln K} \right)^l }} {{l!}}} g^{\mu _1 \nu _1 } \cdots g^{\mu _l
\nu _l }
\nonumber\\
&&\cdot\Big( {f_{\mu _1 } } \phi _{\nu _1 \mu _2 \nu _2 \cdots
\mu _l \nu _l }
+ f_{\mu _1 \nu _1 } \phi _{\mu _2 \nu _2 \cdots \mu _l \nu _l
} + \cdots + {f_{\mu _1 \nu _1 \cdots \mu _l } \phi _{\nu
_l } } \Big )\nonumber\\
&& - \sum\limits_{l = 1}^\infty {\frac{{\left( {\ln K} \right)^l
}} {{l!}}} g^{\mu _1 \nu _1 } \cdots g^{\mu _{l - 1} \nu _{l - 1}
} \Big( {f_\alpha } \phi _{\beta \mu _1 \nu _1 \cdots
\mu _{l - 1} \nu _{l - 1} } +\nonumber\\&&
f_{\alpha \mu _1 \nu_1 }
\phi _{\beta \mu _2 \nu _2 \cdots \mu _{l - 1} \nu _{l - 1} } +
\cdots + {f_{\alpha \mu _1 \nu _1 \cdots \mu _{l - 1} \nu _{l -
1} } \phi _\beta } \Big).\end{aligned}$$ Again, we set the metric to be flat with signature $(-1,1,1\cdots
1)$ after the variation. Replace $\widetilde\phi$ by $\phi+\psi$ in (\[sftaction\]), expanding and coupling to the metric: $$\begin{aligned}
\label{SFTag}S &=& \frac{1} {{g_0 ^2 }}\int {d^d x} \sqrt {
- g} \left\{ {\left( {\frac{1} {2}\phi ^2 - \frac{1}
{2}g^{\mu\nu}\partial_\mu \phi\partial_\nu\phi - \frac{1}
{3}K^3 \phi ^3-\frac{1}{6} K^{-6} } \right)} \right.\nonumber\\
&& \left. { - K^3 \left( {\phi ^2 \psi + \phi \psi ^2 + \frac{1}
{3}\psi ^3 } \right)} \right\}.\end{aligned}$$ Varying the first term in the last right hand side of (\[SFTag\]) with respect to $\delta g^{\alpha\beta}$ gives $$\begin{aligned}
\label{svs1}&&\frac{1} {{g_0 ^2
}}\delta_{g^{\alpha\beta}}\int {d^d x} \sqrt { - g} {\left(
{\frac{1} {2}\phi ^2 - \frac{1} {2}g^{\mu\nu}\partial_\mu
\phi\partial_\nu\phi - \frac{1} {3}K^3
\phi ^3-\frac{1}{6} K^{-6} } \right)} \nonumber\\
&=&- \frac{{g_{_{\alpha \beta } } }}
{{2g_0 ^2 }}\left( {\frac{1} {2}\phi ^2 - \frac{1} {2}(\partial
\phi )^2 - \frac{1} {3}K^3 \phi ^3 }-\frac{1}{6} K^{-6} \right) -
\frac{1} {2g_0^2}\partial _\alpha \phi \partial _\beta \phi\nonumber\\
&\equiv&-\frac{1}{g_0^2}C_{\alpha\beta},\end{aligned}$$ where we have defined $C_{\alpha\beta}$ to simplify our notation. As for the second term in the last right hand side of (\[SFTag\]), note $$\begin{aligned}
&&-K^3\delta _{g^{\alpha \beta } } \int {d^d x} \sqrt { - g} \left(\phi ^2 \psi
+\phi\psi^2+\frac{1}{3}\psi^3 \right)\\
&&= \frac{1} {2}K^3 \Big(\phi ^2 \psi+\phi\psi^2+\frac{1}{3}\psi^3\Big)
g_{_{\alpha \beta } } - K^3 \int {d^d x} \sqrt { - g}\widetilde\phi^2
\delta _{g^{\alpha \beta } } \psi.\end{aligned}$$ So, from (\[general\]) and (\[defineA\]) the variation of the second term in the last step of (\[SFTag\]) contributes: $$\begin{aligned}
\label{svs2} && \left( \frac{{ - K^3 }}{g_0^2} \right)\delta
_{g^{\alpha \beta } }\int {d^d x} \sqrt { - g} \left( {\phi ^2
\psi + \phi \psi ^2 + \frac{1} {3}\psi ^3 } \right)\nonumber\\
&&=\frac{-K^3}{g_0^2}\left\{ A_{\alpha \beta }(\widetilde\phi ^2)
+ \frac{1} {2}\left( {\phi \psi ^2 + \frac{2} {3}\psi ^3 }
\right)g_{\alpha \beta }\right\},\end{aligned}$$ Finally, from (\[defineA\]), (\[svs1\]) and (\[svs2\]), the stress tensor is: $$\label{SFTstress} T_{\alpha \beta } = \frac{2} {{\sqrt { -
g}}}\frac{{\delta S}} {{\delta g^{\alpha \beta } }} =-
\frac{2K^{-3}} {{g_0 ^2 }}A_{\alpha \beta
}(\widetilde\phi^2)-\frac{2}{g_0^2}C_{\alpha\beta} .$$ In the case that $\phi (x)$ only depends on one spatial coordinate, say $x^{25}$, from (\[defineA\]), $$\label{lumpA} A_{\alpha \beta }(\widetilde\phi^2)=\sum\limits_{l =
1}^\infty {\frac{{\left( {\ln K} \right)^l }} {{l!}}} \left\{
{\frac{1} {2}g_{\alpha \beta } \sum\limits_{m = 1}^{2l - 1}
{\widetilde\phi ^2 _m } \phi _{2l - m} - \delta _{\alpha,25}
\delta _{\beta,25} \sum\limits_{m = 1}^l {\widetilde\phi ^2 _{2m -
1} \phi _{2l - 2m + 1} } } \right\}.$$ Plug it into (\[SFTstress\]), we obtain the stress tensor for lump solutions. Similarly, if $\phi (x)$ only depends on time, we can write: $$\label{rollA} A_{\alpha \beta }(\widetilde\phi^2) = - K^3
\sum\limits_{l = 1}^\infty {\frac{{\left( { - \ln K} \right)^l }}
{{l!}}} \left\{ {\frac{1} {2}g_{\alpha \beta } \sum\limits_{m =
1}^{2l - 1} {\widetilde\phi ^2 _m } \phi _{2l - m} + \delta
_{\alpha,0} \delta _{\beta,0} \sum\limits_{m = 1}^l
{\widetilde\phi ^2 _{2m - 1} \phi _{2l - 2m + 1} } }
\right\}.$$ Plug it into (\[SFTstress\]), we obtain the stress tensor for rolling solutions
Energy distribution of the SFT lump solution
--------------------------------------------
In [@0005036], a lump solution of OSFT has been given in the form of an expansion in terms of cosines. We are only concerned with the pure tachyonic mode here, so drop the higher modes: $$\label{SFTlumpsolution} \phi (x) = t_0 + t_1 \cos \left(
{\frac{x} {R}} \right) + t_2 \cos \left( {\frac{{2x}} {R}} \right)
+ \cdots,$$ where $R$ is the radius of the circle on which the coordinate $x$ is compactified. We can calculate the energy distribution of this solution, from (\[dphitild\]), (\[dpsi\]), (\[SFTstress\]) and (\[lumpA\]): $$\widetilde\phi (x) = K^{\partial _x ^2 } \phi (x) = t_0 + t_1 K^{
- \frac{1} {{R^2 }}} \cos \left( {\frac{x} {R}} \right) + t_2 K^{
- \frac{4} {{R^2 }}} \cos \left( {\frac{{2x}} {R}} \right) +
\cdots,$$ $$\psi (x) = \widetilde\phi (x) - \phi (x) = t_1 \left( {K^{ -
\frac{1} {{R^2 }}} - 1} \right)\cos \left( {\frac{x} {R}} \right)
+ t_2 \left( {K^{ - \frac{4} {{R^2 }}} - 1} \right)\cos \left(
{\frac{{2x}} {R}} \right) + \cdots,$$
$$\begin{aligned}
\label{SFTE}E(x) &=& T^0 _0 = - T_{00}\nonumber\\
&=& - \frac{1} {{g_0
^2 }}\left( {\frac{1} {2}\phi ^2 - \frac{1} {3}K^3 \phi ^3 -
\frac{1} {2}\left( {\partial _x \phi } \right)^2 - \frac{1}
{6}K^{ - 6}+K^3 \phi \psi ^2 + \frac{2}{3}K^3 \psi ^3}\right)\nonumber\\
&& - \frac{{K^3 }} {{g_0 ^2 }}\sum\limits_{l = 1}^\infty
{\frac{{\left( {\ln K} \right)^l }} {{l!}}} \sum\limits_{m =
1}^{2l - 1} {\left( {\widetilde\phi ^2 } \right)_m } \phi _{2l -
m}.\end{aligned}$$
In $R=\sqrt 3$ case, using the method introduced in [@0005036], one can obtain: $$t_0=0.216046,\hspace{5mm}t_1=-0.343268,\hspace{5mm}t_2=-0.0978441,$$ when we plug these values into (\[SFTE\]), we find: $$\begin{aligned}
E(x)&=&\frac{1}{g_0^2}\Big(0.0206937
+0.0242345 \cos \frac{x}{R}\\
&&-0.00780954 \cos \frac{2x}{R} -0.0204855 \cos
\frac{3x}{R}\\
&&-0.0111187\cos \frac{4x}{R}-0.00218278\cos
\frac{5x}{R}-0.000177055\cos \frac{6x}{R}\Big).\end{aligned}$$ This lump solution has the interpretation of D24 brane, the tension is: $${\cal T}_{24}=\int\limits_{-\pi R}^{\pi R} dx\, E(x)\simeq 0.225206 \frac{1}{g_0^2}.$$ On the other hand, $\phi=0$ is supposed to represent the D25 brane. We have ${\cal
T}_{25}=-V(\phi=0)=\frac{1}{6}\frac{K^{-6}}{g_0^2}\simeq
0.0346831\frac{1}{g_0^2}$. Therefore, $$\frac{1}{2\pi}\frac{{\cal T}_{24}}{{\cal T}_{25}}\simeq 1.03343$$ a ratio that is unity in string theory.
Figure 2 shows the energy density $E(x)$. As the lump solutions in the p-adic string theory, the energy density is not localised around the hyperplane $x=0$. Instead, $E(x=0)$ is a local minimum. A difference from the p-adic model is that $E(0)$ does not vanish here.
Pressure evolution of the SFT rolling tachyon solution
------------------------------------------------------
In [@0207107], a rolling tachyon solution of OSFT is expressed as a series expansion in cosh($nt$): $$\phi(t)=t_0+t_1{\rm cosh}t+t_2{\rm cosh}2t+\cdots.$$ From (\[dphitild\]), (\[dpsi\]), (\[SFTstress\]) and (\[rollA\]): $$\widetilde\phi (t) = K^{ - \partial _t ^2 } \phi (t) = t_0 + t_1
K^{ - 1} \cosh t + t_2
K^{ - 4} \cosh 2t +
\cdots,$$ $$\psi (t) = \widetilde\phi (t) - \phi (t) = t_1 \left( K^{ -
1} - 1 \right)\cosh t
+ t_2 \left (K^{ - 4} - 1\right)\cosh
2t + \cdots,$$
$$\begin{aligned}
p(t) &=& - T_{11}\nonumber\\
&=& \frac{1} {{g_0 ^2 }}\left( {\frac{1}
{2}\phi ^2 - \frac{1} {3}K^3 \phi ^3 + \frac{1} {2}\left(
{\partial _t \phi } \right)^2 - \frac{1} {6}K^{ - 6} + K^3 \phi
\psi ^2 + \frac{2} {3}K^3 \psi ^3 } \right)\nonumber\\
&& +\frac{{K^3 }} {{g_0 ^2 }}\sum\limits_{l = 1}^\infty
{\frac{{\left( { - \ln K} \right)^l }} {{l!}}} \sum\limits_{m =
1}^{2l - 1} {\left( {\widetilde\phi ^2 } \right)_m } \phi _{2l -
m}.\end{aligned}$$
From section 7 in [@0207107], $$t_0=0.00162997,\hspace{5mm}t_1=0.05,\hspace{5mm}t_2=-0.000189714,$$ and therefore,
$$\begin{aligned}
p(t)&=&\frac{1}{g_0^2}\Big(-0.0346844+0.0000416895\cosh t
+0.00124462\cosh 2t\\
&&-0.0000416042\cosh 3t
+2.59666\times 10^{-7}\cosh 4t\\
&&-3.97466\times
10^{-10}\cosh 5t +2.09045\times 10^{-13}
\cosh 6t\Big).\end{aligned}$$
Figure \[0207SFTrollP\] shows the pressure evolution. It has the same property as the pressure in p-adic theory (Figure 10 in [@0207107]).The pressure starts from negative value at time $t=0$ to force the tachyon roll to the vacuum. But instead of decreasing to zero as $t\rightarrow
\infty$, it oscillates without bound at large times. So, this solution does not seem to represent tachyon matter.
Conclusion
==========
By introducing the metric, we have obtained general expressions for the stress tensors both for the p-adic model and for the pure tachyonic sector of open bosonic string field theory [@9902105], [@9912249], [@0002237], [@0005036].
Furthermore, we considered some available solutions and wrote down the corresponding energy densities for space dependent ones and pressure evolutions for time dependent ones. In conformal field theory, D-branes are boundary conditions and one could expect the energy to be sharply localized at the D-brane position. It was not clear whether or not the lumps of the padic string theory would have this property. Our results show that they do not. The energy density vanishes at $x=0,\pm\infty$. It has two maxima. These two maxima are symmetrically localized with respect to $x=0$. In the pure tachyonic sector of OSFT, the energy density for the lump solution reaches a local minimum at $x=0$. For the rolling tachyon solution, the pressure oscillates with growing amplitude instead of asymptotically vanishing. Therefore, as in the p-adic model, the rolling solution we considered in this paper does not seem to represent tachyon matter.
There are two shortcomings of the calculations in OSFT. The first is not including the massive fields. The second is that the coupling of open strings to the metric could have additional terms that vanish in the flat space limit but contribute to the stress tensor. Such phenomena happens in noncommutative field theory [@0012218]. Open-closed string field theory [@9705241] might be needed to calculate the stress tensor with complete confidence. I thank M. Schnabl for bringing this point to my attention.
[**Acknowledgements.**]{} The author is especially grateful to B. Zwiebach and M. Schnabl for their critical help in finishing this work. Also I thank Jessie Shelton and Nikhil Mittal for useful discussions. This work was supported by DOE contract \#DE-FC02-94ER40818.
[99]{}
A. Sen, “Descent relations among bosonic D-branes”, Int. J. Mod. Phys. A14, 4061 (1999) \[hep-th/9902105\]. A. Sen and B. Zwiebach, “Tachyon condensation in string field theory”, JHEP 9912 (1999) 027, \[hep-th/9912249\]. Jeffrey A. Harvey and Per Kraus, “D-Branes as Unstable Lumps in Bosonic Open String Field Theory”, JHEP 0004 (2000) 012, \[hep-th/0002117\]. N. Moeller and W. Taylor, “Level truncation and the tachyon in open bosonic string field theory”, Nucl.Phys. B583 (2000) 105-144 , \[htp-th/0002237\]. R. de Mello Koch, A. Jevicki, M. Mihailescu and R. Tatar, “Lumps and P-branes in Open String Field Theory”, Phys.Lett. B482 (2000) 249-254, \[hep-th/0003031\]. N. Moeller, A. Sen and B. Zwiebach, “D-branes as Tachyon Lumps in String Field Thoery”, JHEP 0008 (2000) 039 , \[hep-th/0005036\]. L. Rastelli and B. Zwiebach, “Tachyon potentials, star products and universality”, JHEP 0109 (2001) 038 , \[hep-th/0006240\]. V. A. Kostelecky and R. Potting, “Analytical construction of a nonperturbative vacuum for the open bosonic string”, Phys.Rev. D63 (2001) 046007, \[hep-th/0008252\]. A. Sen, “Rolling Tachyon“, JHEP 0204 (2002) 048 , \[hep-th/0203211\]. A. Sen “Tachyon Matter”, JHEP 0207 (2002) 065, \[hep-th/0203265\]. M. Gutperle and A. Strominger, “Spacelike Branes”, JHEP 0204 (2002) 018, \[hep-th/0202210\]. K. Hashimoto, “Dynamical Decay of Brane-Antibrane and Dielectric Brane”, JHEP 0207 (2002) 035, \[hep-th/0204203\]. S. Sugimoto and S. Terashima, “Tachyon matter in boundary string field theory”, JHEP 0207 (2002) 025 , \[hep-th/0205085\]. J. A. Minahan, “Rolling the tachyon in super BSFT”, JHEP 0207 (2002) 030, \[hep-th/0205098\]. A. Ishida and S. Uehara, “Gauge Fields on Tachyon Matter”, \[hep-th/0206102\]. K. Ohta and T. Yokono,“Gravitational Approach to Tachyon Matter”, \[hep-th/0207004\]. A. Sen, ”Time evolution in open string theory", \[hep-th/0207105\]. N. Moeller and B. Zwiebach, “Dynamics with Infinitely Many Time Derivatives and Rolling Tachyons” \[hep-th/0207107\]. J. Kluson, “Time Dependent Solution in Open Bosonic String Field Theory”, \[hep-th/0208028\]. P. Mukhopadhyay and A. Sen, “Decay of Unstable D-branes with Electric Field”, \[hep-th/0208142\]. A. Sen, “Time and Tachyon”, \[hep-th/0209122\]. L. Brekke, P.O. Freund, M. Olson and E. Witten, “Non-archimedean string dynamics”, Nucl. Phys. B302 (1988) 365. D. Ghoshal and A. Sen, “Tachyon condensation and Brane Descent Relations in p-adic String Theory”, Nucl.Phys. B584 (2000) 300-312, \[hep-th/0003278\]. B. Zwiebach, “A Solvable Toy Model for Tachyon Condensation in String Field Theory”, JHEP 0009 (2000) 028, \[hep-th/0008227\] Y. Okawa and H. Ooguri, “How Noncommutative Gauge Theories Couple to Gravity”, Nucl.Phys. B599 (2001) 55-82, \[hep-th/0012218\] B. Zwiebach, “Oriented Open-Closed String Theory Revisited”, Annals Phys. 267 (1998) 193-248, \[hep-th/9705241\]
[^1]: I thank M. Schnabl for suggesting the use of this identity.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study a simple left-right symmetric (LRS) extension of the Zee model for neutrino mass generation. An extra $SU(2)_{L/R}$ singlet charged scalar helps in generating a loop-induced Majorana mass for neutrinos in this model. The right-handed neutrinos in this case are very light of the order of a few eV to a few MeV which makes this scenario quite different from other LRS models. We have analyzed the scalar potential and Higgs spectrum in detail, which also play an important role for the neutrino phenomenology. We identified the parameter regions in the model which satisfy the experimentally observed neutrino masses and mixings along with other experimental constraints. We have then studied the collider signatures of the charged scalar at $e^+e^-$ colliders with different benchmark points. It is possible to get a huge enhancement in the production cross-section of the charged scalar at lepton collider compared to the hadron colliders, resulting in a much stronger signal which can be easily observed at the upcoming ILC or CLIC experiments.'
author:
- Sarif Khan
- Manimala Mitra
- Ayon Patra
title: 'Neutrino and Collider Implications of a Left-Right Extended Zee Model'
---
Introduction {#intro}
============
The observation of neutrino oscillation leading to the realization that neutrinos are massive, is one of the biggest motivation for physics beyond the Standard Model (SM). A large number of models have been suggested to explain neutrino masses and mixings either by the seesaw mechanism [@csaw] or through loop induced processes [@rad]. The Zee model [@zee] is one of the simplest such scenarios where neutrino masses are generated at one-loop by extending the SM scalar sector with an extra doublet and a charged singlet scalar field. The charged singlet scalar can mix with other charged scalars while also having non-zero flavor violating couplings with leptons, giving rise to neutrino masses at one-loop. Unfortunately the simplest form of the Zee model was shown to be ruled out by experimental neutrino data [@zee-out]. However its extensions might still be viable. In this work we study an extended Zee model in a left-right symmetric (LRS) framework [@lr]. The model was proposed and studied in context of the LHC in [@pavel1] and the low energy flavor violating processes were discussed in [@pavel2]. In this work we examine its viability from neutrino oscillation data, study the scalar potential in detail and derive charged Higgs spectrum, as well as analyze the possible electron-positron collider implications for the charged singlet Higgs boson.
Left-right symmetric (LRS) models are attractive extensions of the SM with the gauge group being extended to $SU(3)_C \times SU(2)_L \times SU(2)_R \times U(1)_{B-L}$. The parity symmetry is fundamentally conserved in these models which provides a natural solution to the strong CP problem [@scp] without introducing a global Peccei-Quinn symmetry. The parity symmetry is broken once the $SU(2)_R \times U(1)_{B-L}$ is spontaneously broken into $U(1)_Y$ at a scale $v_R$ much above the electroweak scale. Thus the observed parity violation in the SM can be easily understood. The gauge structure of LRS framework naturally requires the existence of right-handed neutrinos which can help generate light neutrino masses through seesaw mechanism. This usually requires the presence of an $SU(2)_R$ triplet scalar whose neutral component acquires a non-zero vacuum expectation value, leading to the right-handed symmetry breaking and the generation of Majorana masses for the right-handed neutrinos. The simplest LRS scenario, on the other hand, requires only an $SU(2)_R$ doublet scalar to achieve a consistent right-handed symmetry breaking but cannot generate light neutrino masses[^1]. A simple LRS framework consisting of two doublets and a bidoublet scalar field, as will be considered here, can only generate a Dirac mass term for the neutrinos and the introduction of an extra charged singlet scalar is a very economical way to generate neutrino Majorana masses in such a scenario. Hence it is quite natural to extend the Zee model in a simple LRS framework to generate the neutrino masses and mixings.
There are several other advantages of LRS extended Zee model. Firstly, since the neutrino Majorana masses are generated at one-loop, the right-handed neutrino masses also remain quite light ranging from a few MeV to a few eV. This is quite different from other LRS scenarios, where right-handed neutrinos are very heavy with masses proportional to the right-handed symmetry breaking scale (typically more than a few TeV). The presence of lighter right-handed neutrino states is a unique feature of this model. The recent results from the LSND [@Athanassopoulos:1995iw; @Aguilar:2001ty] and MiniBooNE experiments [@AguilarArevalo:2007it; @AguilarArevalo:2010wv; @Aguilar-Arevalo:2013pmq] hint at the existence of a light sterile neutrino with mass around a few eV. The LRS Zee model would be a prime candidate for explaining such a particle if these experimental results were to persist. Another important consequence of light right-handed neutrinos is the enhanced cross-section for the production of the $SU(2)_{L/R}$ singlet charged Higgs boson in this model, especially in the context of electron-positron colliders. The singlet charged Higgs bosons can be pair produced via a $t$-channel process. This process can either be mediated by a left-handed or a right-handed neutrino. The left-handed neutrino mediated processes suffer from extremely small couplings while the right-handed neutrino mediated processes (for models with heavy right-handed neutrinos) are suppressed by the large right-handed neutrino masses. The $t$-channel mediated charged Higgs pair-production cross-section thus remains extremely small for both these processes. Our scenario, with light MeV scale right-handed neutrinos, can alleviate this shortcoming and deliver large pair-production cross-section for the charged Higgs boson. Owing to the large couplings with the leptons, the charged singlet Higgs bosons can be copiously produced at lepton colliders, and thus give rise to rich collider phenomenology. Since the singlet charged Higgs does not interact with the quarks of the SM, it has a limited discovery prospect in the hadronic colliders including the Large Hadron Collider (LHC). A lepton collider, instead, is a perfect setup to test the singlet charged Higgs of this model. In this work, we pursue a detailed study of LRS extended Zee model by analyzing the neutrino mass and mixing constraints on the model parameters, taking into account three generations of light neutrinos. We explicitly show the hierarchical structure of Dirac mass matrix. We also analyze the potential and evaluate the Higgs spectrum in detail. Furthermore, with the set of model parameters that satisfy neutrino oscillation measurements, we carry out an in-depth analysis of the pair-production and decay of these charged scalars in the upcoming International Linear Collider (ILC) and Compact Linear Collider (CLIC) experiments. The final state of two opposite sign leptons and missing energy can be measured quite significantly over the SM background resulting in a possibility to observe such a process even with a very low luminosity $\mathcal{L} \sim 1-3 \, \rm{fb}^{-1}$ at these experiments. Therefore, even an early run of ILC/CLIC can detect the presence of such a gauge singlet charged Higgs state. The rest of this paper is organized as follows. We discuss the model and the particle spectrum in Section. \[model\]. Following that, the pair-production of the charged Higgs and its detailed collider phenomenology is discussed in Section. \[col\]. We present our conclusions in Section. \[conc\].
Model and Spectrum {#model}
==================
LRS models are simple gauge extensions of the SM with the gauge group being $SU(3)_C \times SU(2)_L \times SU(2)_R \times U(1)_{B-L}$. The charge of a particle in this model is defined as $$\mathcal{Q}=I_{3L}+I_{3R}+\frac{B-L}{2},$$ where $I_{3L/3R}$ is the third component of isospin under $SU(2)_{L/R}$ symmetry. The quarks and leptons consist of three generation of left-handed and right-handed doublet fields: $$\begin{aligned}
\!\!Q_L\left (3,2, 1, \frac13 \right )\!\!&=&\!\!\left (\begin{array}{c}
u\\ d \end{array} \right )_L,~~
Q_R \left( 3,1, 2, \frac13 \right )\!=\!\left (\begin{array}{c}
u\\d \end{array} \right )_R ,\nonumber \\
l_L \left ( 1,2, 1, -1 \right )&=&\left (\begin{array}{c}
\nu\\ e\end{array}\right )_L,~~
l_R\left ( 1,1, 2, -1 \right )=\left (\begin{array}{c}
\nu \\ e \end{array}\right )_R,~~~\end{aligned}$$ where the numbers in the brackets denote the quantum numbers under $SU(3)_C$, $SU(2)_L$, $SU(2)_R$, $U(1)_{B-L}$ gauge groups respectively. Here we see that the right-handed neutrinos are naturally present due to the gauge symmetry of the models.
The minimal Higgs sector, required for a consistent symmetry breaking mechanism and generation of quark and lepton masses and mixing angles, consists of $$\begin{aligned}
H_R(1,1,2,1)&=&\left (\begin{array}{c}
H_R^+ \\H_R^0 \end{array} \right ),~~
H_L(1,2,1,1)=\left (\begin{array}{c}
H_L^+ \\H_L^0 \end{array} \right ),~~
\Phi(1,2,2,0)={\left (\begin{array}{cc}
\phi^{0}_1 & \phi^{+}_{2} \\ \phi^{-}_{1} & \phi^{0}_{2} \end{array} \right )},~~ \delta(1,1,1,2) = \delta^+ .~~~~~\end{aligned}$$ The right-handed doublet field $H_R$ is required for breaking the $SU(2)_R \times U(1)_{B-L}$ into $U(1)_Y$ at some high scale to obtain the SM gauge symmetry at the electroweak (EW) scale. The $H_L$ doublet is required for preservation of the left-right symmetry. The bidoublet field $\Phi$ is responsible for generation of quark and charged lepton masses and Cabibbo-Kobayashi-Maskawa (CKM) mixing angles. The charged singlet field $\delta^{\pm}$ is needed for generation of neutrino masses through one-loop diagrams as will be discussed later in this section.
The Yukawa Lagrangian is given as: $$\begin{aligned}
\mathcal{L}_Y&=&Y_{ij}^{q1}\overline{Q}_{Li}\Phi Q_{Rj}+Y_{ij}^{q2}\overline{Q}_{Li}\widetilde{\Phi}Q_{Rj}+Y_{ij}^{l1}\overline{l}_{Li}\Phi l_{Rj} + Y_{ij}^{l2}\overline{l}_{Li}\widetilde{\Phi}l_{Rj} + \l_{L_{ij}} l^T_{Li} i \tau_2 l_{Lj} \delta^+ + \l_{R_{ij}} l^T_{Ri} i \tau_2 l_{Rj} \delta^+ + H.C.~,~~
\label{eq:yuk}\end{aligned}$$ where $Y$ and $\l$ are the Yukawa couplings and $$\widetilde{\Phi}=\tau_2\Phi^\ast\tau_2,$$ The structure of $\l_{L/R_{ij}}$ term is such that the only terms that will survive are the ones with $i \neq j$. This is exactly the same as in the Zee mechanism of neutrino mass generation. If we expand out any one of the terms involving $\d^+$ in the Yukawa Lagrangian we will get: $$\mathcal{L} \supset \sum_{i \neq j} \nu_i e_j(\l_{ij}-\l_{ji}),$$ where $\nu_i$ and $e_j$ are both in the flavor basis. Thus if we redefine the $\l$ matrix to $\l'_{ij}=\l_{ij}-\l_{ji}$, then this new $\l'$ matrix is completely anti-symmetric and the Lagrangian terms can now be written as: $$\mathcal{L} \supset \sum_{i,j} \nu_i e_j \l'_{ij}.$$
The Vacuum expectation values (VEVs) of the Higgs fields are given as: $$\left<\phi_1^0\right>=v_1,~\left<\phi_2^0\right>=v_2,~\left<H_R^0\right>=v_R,~\left<H_L^0\right>=v_L,$$ with the effective EW VEV given as $v_{EW}=\sqrt{v_1^2+v_2^2+v_L^2}$. Without loss of generality, one of the bidoublet VEVs can be chosen to be small. Also since $v_L$ does not contribute to the top mass, a large $v_L$ would automatically require a large top Yukawa coupling resulting in the theory being non-perturbative at quite low scales. The hierarchy in the VEVs thus has been chosen such that $$v_R>>v_1>v_2,v_L.$$
The gauge sector of the model consist of two charged $W_R^\pm$ and $W^\pm$ gauge bosons and three neutral bosons including the $Z_R$, $Z$ and the photon. The $W_R^\pm$ and the $Z_R$ bosons get their masses at the right-handed symmetry breaking scale and remain heavy while the others are the same as in the SM. The heavy gauge boson masses in this model are given as: $$\begin{aligned}
M^2_{W_R^\pm} &&\simeq \frac{1}{2} g_R^2 (v_R^2+v_1^2+v_2^2), ~~~~M^2_{Z_R} \simeq \frac{1}{2} \left[(g_R^2+g_V^2) v_R^2+ \frac{g_R^4(v_1^2+v_2^2)+g_V^4 v_L^2}{g_R^2+g_V^2} \right],~~~~~~
\label{eq:ZR}\end{aligned}$$ where $g_R$ and $g_V$ are the $SU(2)_R$ and $U(1)_{B-L}$ gauge couplings respectively. The left-handed (SM-like) gauge boson masses are given by their usual expressions with the effective $U(1)_Y$ gauge coupling $g_Y$ given as $$g_Y=\frac{g_R g_V}{\sqrt{g_R^2+g_V^2}}.$$
The scalar potential of this model is given as: $$\begin{aligned}
V(\Delta,\Phi)&=& -\mu_{1}^2{\text{Tr}}\left(\Phi^\dagger \Phi \right)-\mu_2^2{\text{Tr}}\left[\widetilde{\Phi}\Phi^\dagger+\widetilde{\Phi}^\dagger\Phi\right]-\mu_3^2 H_R^\dagger H_R - \mu_4^2 H_L^\dagger H_L- \mu_5^2 \delta^+ \delta^-+ \left(M_1 H_L^\dagger \Phi H_R \right. \notag \\
&+& \left. M_2 H_L^\dagger \widetilde\Phi H_R + H.C.\right) + \l_1\left[{\text{Tr}}(\Phi^\dagger\Phi)\right]^2 + \l_2\left[\left\{{\text{Tr}}(\widetilde\Phi\Phi^\dagger)\right\}^2+\left\{{\text{Tr}}(\widetilde\Phi^\dagger\Phi)\right\}^2\right]+\l_3{\text{Tr}}(\widetilde\Phi\Phi^\dagger){\text{Tr}}(\widetilde\Phi^\dagger\Phi)\notag\\
&+&\l_4{\text{Tr}}(\Phi\Phi^\dagger)\left[{\text{Tr}}(\widetilde\Phi\Phi^\dagger)+{\text{Tr}}(\widetilde\Phi^\dagger\Phi)\right] + \left[ i \a_1 {\text{Tr}}(H_L^T \tau_2 \Phi H_R \delta^-) + i \a_2 {\text{Tr}}(H_L^T \tau_2 \widetilde\Phi H_R \delta^-) + H.C.\right] \notag \notag\\
&+& \a_3(H_L^\dagger \Phi \Phi^\dagger H_L) + \a_4\left[{\text{Tr}}(\widetilde\Phi\Phi^\dagger+\widetilde{\Phi}^\dagger\Phi)H_L^\dagger H_L\right]+\a_5{\text{Tr}}(\Phi\Phi^\dagger)H_L^\dagger H_L +\a_6 (H_R^\dagger \Phi^\dagger \Phi H_R) \notag\\
&+&\a_7\left[{\text{Tr}}(\widetilde\Phi\Phi^\dagger+\widetilde{\Phi}^\dagger\Phi)H_R^\dagger H_R\right] + \a_8{\text{Tr}}(\Phi\Phi^\dagger)H_R^\dagger H_R + \b_1 (H_L^\dagger H_L)^2 + \b_2 (H_R^\dagger H_R)^2 +\b_3 (H_R^\dagger H_R)(H_L^\dagger H_L) \notag \\
&+&\g_1 {\text{Tr}}(\Phi^\dagger \Phi) \delta^+ \delta^- + \g_2 {\text{Tr}}\left[\widetilde{\Phi}\Phi^\dagger+\widetilde{\Phi}^\dagger\Phi\right]\delta^+ \delta^- + \g_3 H_L^\dagger H_L \delta^+ \delta^- +\g_4 H_R^\dagger H_R \delta^+ \delta^- + \g_5 (\delta{^+} \delta{^-})^2.
\label{eq:spot}\end{aligned}$$
This gives four CP-even, two CP-odd and three charged Higgs boson states. Two CP-odd and two charged states are eaten up to give mass to the $Z_R,~Z,~W_R,W$ gauge bosons respectively. We will mainly focus our discussion on the charged Higgs sector, as that is the most important for the neutrino masses and the collider analysis which will be studied in this paper.
Minimizing the scalar potential of Eq. \[eq:spot\] we get four minimization conditions given as $$\begin{aligned}
&&2 (\l_1 + 4 \l_2 + 2 \l_3) v_1 v_2^2 + 2 \l_4 v_2^3 + v_1 (\a_5 v_L^2 + \a_8 v_R^2 + 2 \l_1 v_1^2 - \mu_1^2) + 2 v_2 (\a_4 v_L^2 + \a_7 v_R^2 + 3 \l_4 v_1^2 - \mu_2^2) + M_2 v_L v_R=0, \notag \\
&&2 ( \l_1 v_2 + 3 \l_4 v_1) v_2^2+ v_2 \left\{ (\a_3 + \a_5) v_L^2 + (\a_6 + \a_8) v_R^2 + 2 (\l_1 + 4 \l_2 + 2 \l_3) v_1^2 - \mu_1^2 \right\} + M_1 v_L v_R \notag \\
&& + 2 v_1 ( \a_4 v_L^2+ \a_7 v_R^2 + \l_4 v_1^2 - \mu_2^2) =0, \notag \\
&& \left\{ 4 \a_4 v_1 v_2 + \a_5 v_1^2 + (\a_3 + \a_5) v_2^2 + 2 \b_1 v_L^2+ \b_3 v_R^2 - \mu_4^2\right\}v_L + M_2 v_1 v_R + M_1 v_2 v_R = 0, \notag \\
&& \left\{ 4 \a_7 v_1 v_2 + \a_8 v_1^2 + (\a_6 + \a_8) v_2^2 + 2 \b_2 v_R^2+ \b_3 v_L^2 - \mu_3^2 \right\} v_R + M_2 v_1 v_L + M_1 v_2 v_L =0.\end{aligned}$$
Using these conditions along with the scalar potential, the charged Higgs mass-squared matrix in the gauge basis $({\phi_1^-}^*, \phi_2^+, H_R^+,H_L^+,\delta^+)$ is given as $$M_{H^\pm}^2 =
\begin{pmatrix}
M_{11} & M_{12} & \a_3 v_1 v_L - M_2 v_R & \a_6 v_2 v_R + M_1 v_L & -\a_2 v_L v_R \\
M_{12} & M_{22} & \a_3 v_2 v_L + M_1 v_R & \a_6 v_1 v_R - M_2 v_L & \a_1 v_L v_R \\
\a_3 v_1 v_L - M_2 v_R & \a_3 v_2 v_L + M_1 v_R & M_{33} & M_1 v_1 + M_2 v_2 & -(\a_1 v_2 + \a_2 v_1) v_R \\
\a_6 v_2 v_R + M_1 v_L & \a_6 v_1 v_R - M_2 v_L & M_1 v_1 + M_2 v_2 & M_{44} & (\a_1 v_1 + \a_2 v_2) v_L \\
-\a_2 v_L v_R & \a_1 v_L v_R & -(\a_1 v_2 + \a_2 v_1) v_R & (\a_1 v_1 + \a_2 v_2) v_L & M_{55}
\end{pmatrix},$$ where $$\begin{aligned}
M_{11} &=& \left(-M_2 v_1 v_L v_R + M_1 v_2 v_L v_R + \a_3 v_1^2 v_L^2 + \a_6 v_2^2 v_R^2\right)/ (v_1^2 - v_2^2), \notag \\
M_{12} &=& (M_1 v_1 v_L v_R - M_2 v_2 v_L v_R + \a_3 v_1 v_2 v_L^2 + \a_6 v_1 v_2 v_R^2)/(v_1^2 - v_2^2),\notag \\
M_{22} &=& \left(-M_2 v_1 v_L v_R + M_1 v_2 v_L v_R + \a_3 v_2^2 v_L^2 + \a_6 v_1^2 v_R^2\right)/ (v_1^2 - v_2^2), \notag \\
M_{33} &=& -\frac{1}{v_L} (M_2 v_1 v_R + M_1 v_2 v_R) + \a_3 (v_2^2 - v_1^2), \notag \\
M_{44} &=& -\frac{1}{v_R} (M_2 v_1 v_L + M_1 v_2 v_L) + \a_6 (v_2^2 - v_1^2), \notag \\
M_{55} &=& \gamma_1 (v_1^2 + v_2^2)+ 4 \gamma_2 v_1 v_2 + \gamma_3 v_L^2 + \gamma_4 v_R^2 - \mu_5^2.\end{aligned}$$ This $5\times 5$ charged Higgs mass-squared matrix can be diagonalized to obtain their mass eigenvalues as $$M^2_{Diag} = V^{\dagger} M_{H^\pm}^2 V,
\label{eq:hdiag}$$ where $M^2_{Diag}$ is the diagonalized charged Higgs boson mass-squared matrix and $V$ is the corresponding diagonalizing matrix. There are two zero eigenvalues corresponding to the two Goldstone bosons absorbed by the $W_R^\pm$ and $W^\pm$ bosons to give them mass. The Goldstone bosons primarily consist of $H_R^\pm$ and $\phi_1^\pm$ states respectively as their corresponding doublet neutral fields get the large non-zero VEVs. The other three eigenstates give the three physical charged Higgses and are linear combinations of $\phi_2^\pm$, $H_L^\pm$ and $\d^\pm$. Flavor constraints, such as, $K^0-\bar{K}^0$ and $B^0-\bar{B}^0$ mixings require the neutral component of the bidoublet field $\phi_2^0$ mass to be heavier than 15 TeV [@Zhang:2007da], forcing its charged counterpart to be very massive as well. So $\d^\pm$ can primarily mix only with $H_L^\pm$ as $\phi_2^\pm$ is effectively decoupled owing to its large mass. We will consider two scenarios for our analysis. One where the lightest charged Higgs consists almost entirely of the charged singlet field $\d^\pm$ and another where the lightest physical state is almost equal admixture of $\d^\pm$ and $H_L^\pm$. In Tab. \[tab:chig\] we provide four benchmark points for the lightest charged Higgs boson $H_1^\pm$, two for the minimal mixing and two for the maximal mixing scenarios respectively.
[C[1.5cm]{}|C[6.5cm]{}]{} Mass & Composition\
473.32 & $0.002 \phi_2^+ + 0.999 \d^+$\
1000.7 & $0.002 \phi_2^+ + 0.999 \d^+$\
432.58 & 0.03 ${\phi_1{^-}}^\ast-0.006 \phi_2^+ +0.72H_L^+ + 0.69 \d^+$\
1000.9 & 0.03 ${\phi_1{^-}}^\ast-0.006 \phi_2^+ +0.76H_L^+ + 0.65 \d^+$\
We also cross-check the corresponding scalar and pseudo-scalar neutral Higgs bosons for the set of parameters that we use to generate the above charged Higgs masses, given in Table. \[tab:chig\]. We ensure that the lightest scalar Higgs boson mass is 125 GeV and the pseudo-scalar sector has two massless Goldstone bosons, required to give masses to the $Z_R$ and $Z$ bosons.
The quark and lepton masses can be obtained from Eq. \[eq:yuk\] as: $$\begin{aligned}
M_{u} &=& Y^{q1} v_1+Y^{q2} v_2,~~M_d=Y^{q1} v_2+Y^{q2} v_1,~~M_l = Y^{l1} v_2+Y^{l2}v_1,~~M_\nu^D = Y^{l1} v_1+Y^{l2} v_2.\end{aligned}$$ One can perform a simple rotation of the neutral bidoublet fields to obtain two new scalar fields $$h_1^0 = \frac{v_1 \phi_1^0 + v_2 \phi_2^0}{\sqrt{v_1^2 + v_2^2}},~~~~ h_2^0 = \frac{v_2 \phi_1^0 - v_1 \phi_2^0}{\sqrt{v_1^2 + v_2^2}}.~~~~$$ In this rotated basis, only one of these new fields ($h_1^0$) gets a non-zero VEV. This along with a redefinition of the couplings gives $$\hspace*{-0.3cm}
{M_{u} = Y^q v'_1,~~M_d = \wt{Y}^q v'_1,~~M_l = \wt{Y}^l v'_1,~~M_\nu^D = Y^l v'_1,}
\label{eq:fmass}$$ where $\left< h_1^0 \right> = v'_1 $ is the VEV in the redefined basis and $$\begin{aligned}
Y^q &=& \frac{1}{v'_1}\left( Y^{q1} v_1+Y^{q2} v_2 \right),~~\wt Y^q = \frac{1}{v'_1}\left( Y^{q1} v_2+Y^{q2} v_1 \right),\notag \\ \wt{Y}^l &=& \frac{1}{v'_1}\left(Y^{l1} v_2+Y^{l2}v_1\right),~~Y^l = \frac{1}{v'_1}\left(Y^{l1} v_1+Y^{l2}v_2\right).~~\end{aligned}$$ The $\delta^+$ field is responsible for producing the Majorana mass terms in the neutrino mass matrix which are given as [@pavel1]: $$\begin{aligned}
\displaystyle {(M_\nu^L)}^{\alpha \gamma}&&=\frac{1}{4\pi^2}\lambda'_L{^{\alpha \beta}}m_{e_\beta}\sum_{i=1}^3 \text{Log}\left(\frac{M_{h_i}^2}{m_{e_\beta}^2}\right) \times V_{5i}\left [ (Y_l^\dagger)^{\beta \gamma}V_{2i}^*-(\wt Y_l^\dagger)^{\beta \gamma}V_{1i}^* \right ] \ + \ \alpha \leftrightarrow \gamma
\,,\notag \\
\displaystyle {(M^R_\nu)}^{\alpha \gamma}&&=\frac{1}{4\pi^2}\lambda'_R{^{\alpha \beta}}m_{e_\beta}\sum_{i=1}^3\text{Log}\left(\frac{M_{h_i}^2}{m_{e_\beta}^2}\right) \times V_{5i}\left[(Y_l)^{\beta \gamma}V_{1i}^*-(\wt Y_l)^{\beta \gamma}V_{2i}^*\right] \ + \ \alpha \leftrightarrow \gamma \,.
\label{RHM}\end{aligned}$$ Here $\alpha$, $\beta$ and $\gamma$ each run from $1-3$, $V_{ij}$ corresponds to the $ij$-th element of the charged Higgs boson mixing matrix $V$ defined in Eq. \[eq:hdiag\], $M_{h_i}(i=1-3)$ is the mass of the charged Higgs boson eigenstates and $m_{e_{\alpha}}$ is the charged lepton mass with $\alpha = 1,2~\text{and}~3$ representing the electron, muon and tau respectively. The neutrino mass matrix would thus be a $6\times6$ matrix in the $(\nu_{L_i},\nu_{R_j})$ $(i,j=1-3)$ basis given as: $$M_\nu = \begin{bmatrix}
M^L_\nu&M^D_\nu\\(M^D_\nu)^T&M^R_\nu
\end{bmatrix},$$ where $M_{\nu}^L$ and $M^R_{\nu}$ are generated at one-loop while $M^D_\nu$ is the neutrino Dirac mass term. With the seesaw approximation, the light neutrino mass matrix appears as a combination of Type-I and Type-II seesaw: $$M_{\nu}=M^L_{\nu}-{M^D_{\nu}}^T {M^R_{\nu}}^{-1} {M^D_{\nu}}.
\label{t2t1}$$
The redefined coupling $\wt Y_l$, which we have chosen to be diagonal, is entirely determined from the charged lepton masses as can be seen from Eq. \[eq:fmass\]. Similarly $Y^q$ (chosen to be diagonal) and $\wt Y^q$ can be determined from the up and down sector quark masses and CKM mixings. For the neutrino sector we first chose $Y_l$ to be zero to get the light neutrino masses and mixings from $M^L_\nu$ alone. This approach does not work as there are too few free parameters to fit the experimental neutrino data ($\lambda^{\prime}_L$ is anti-symmetric). We then considered the case with non-zero $Y_l$ while $\l'_L$ was chosen to be zero. The light neutrino masses in this case arises entirely from $M_\nu^D$ and $M_R$ similar to type-I seesaw mechanism: $$M_{\nu}=-{M^D_{\nu}}^T {M^R_{\nu}}^{-1} {M^D_{\nu}}.
\label{t1}$$
This gave us the correct experimentally observed masses and mixings for the light neutrino and hence this is the approach we have chosen for the neutrino sector [^2].
![Right-handed neutrino masses as a function of $\lambda'_R$ for minimal charged Higgs mixing.[]{data-label="fig:rhneumass"}](rh_neutrino_mass2.png){height="3.0in" width="3.50in"}
The right-handed neutrino masses in this scenario are generated at one loop and proportional to the square of the charged lepton Yukawa coupling $\tilde{Y}_l$. Therefore, right handed neutrino masses are quite small. As the other Yukawa coupling $Y_l$ is responsible for generating Dirac masses for the neutrinos, it is orders of magnitude smaller than $\tilde{Y}_l$ and hence does not have any impact on right-handed neutrino masses. We show the variation of the three right-handed neutrino masses $M_{N_{1,2,3}}$ with the Yukawa coupling $\lambda^{\prime}_R$ in Fig. \[fig:rhneumass\]. As can be seen, that for $\lambda^{\prime}_R \sim 0.1-1$, the lightest right handed neutrino mass $M_{N_1}$ varies from $3 \, \rm{eV}-30$ eV, while $M_{N_{2,3}}$ are in the sub-MeV scale. In deriving this, we utilize Eq. \[RHM\], where we diagonalize the right-handed Majorana mass matrix $M_\nu^R$. The charged Higgs boson masses and mixings used to obtain the neutrino Majorana masses are the ones corresponding to the first benchmark point in Tab. \[tab:chig\]. We have also provided these charged Higgs boson masses and mixings in details in Appendix A. This is quite different from other left-right symmetric models where the right-handed neutrino is naturally heavy as its mass is proportional to the right-handed symmetry breaking scale.
[||C[9.0cm]{}|]{}
7.03$\times 10^{-5}~\text{eV}^2$ $<\Delta m_{21}^2<$ 8.09$\times 10^{-5}~\text{eV}^2$\
2.407$\times 10^{-3}~\text{eV}^2$ $<\Delta m_{31}^2<$ 2.643$\times 10^{-3}~\text{eV}^2$\
$0.271<\sin^2{\theta_{12}}<0.345$\
$0.385<\sin^2{\theta_{23}}<0.635$\
$0.01934<\sin^2{\theta_{13}}<0.02392$\
$U_{PMNS}$ $$\begin{pmatrix}
0.800 \rightarrow 0.844 & 0.515 \rightarrow 0.581 & 0.139 \rightarrow 0.155 \\
0.229 \rightarrow 0.516 & 0.438 \rightarrow 0.699 & 0.614 \rightarrow 0.790 \\
0.249 \rightarrow 0.528 & 0.462 \rightarrow 0.715 & 0.595 \rightarrow 0.776\end{pmatrix}$$\
![Scatter plot of neutrino Dirac mass matrix elements $M^D_{\nu_{ij}}$ (denoted by $M_{\nu_{ij}}$ in the figure) satisfying the neutrino oscillation data in Table \[tab:neut\].[]{data-label="fig:nyuk"}](Yukd.pdf){width="4.5in"}
For our subsequent analysis, we choose $\lambda'_R \sim \mathcal{O}(1)$. Since we use a type-I seesaw-like structure for the neutrino mass, the Dirac Yukawa couplings $Y^l$ in Eq. \[eq:fmass\] are chosen accordingly to satisfy the correct neutrino oscillation parameters, given in Table \[tab:neut\]. As an illustrative example, we consider a normal hierarchy spectrum in light neutrino sector. The allowed values for the elements of Dirac mass matrix $M^D_\nu$ are obtained by scanning over the allowed parameter space. We have varied the elements of $\l^{\prime}_R$ matrix between 0.5 to 1 keeping them very close to each other by allowing a spread of only 10%. To generate $M^{R}_{\nu}$, we set the charged Higgs masses and mixings as given in Appendix A. Fig. \[fig:nyuk\] gives a scatter plot of the allowed neutrino Dirac masses (directly proportional to the Dirac Yukawa coupling $Y_l$) satisfying the experimental 3$\sigma$ ranges for the light neutrino parameters given in Tab. \[tab:neut\]. Here we plot the neutrino Dirac masses along the Y-axis with $M^D_{\nu_{11}}$ along the X-axis. This gives us an clear indication of the allowed values of the various terms in the $M^D_\nu$ matrix relative to each other. Note that the hierarchy between $(M^D_{\nu})_{11}, (M^D_{\nu})_{12},$ and $(M^D_{\nu})_{13}$ is clearly visible from the figure, with $(13)$ element of the Dirac mass matrix allowed to take highest values. The hierarchy between $(13)$ and $(12)$ element is largest for lower $(M^D_{\nu})_{11}$ mass $(M^D_{\nu})_{11} \sim 0.1 $ eV.
As it is clear from the preceding discussion, in the present model we have an eV scale right-handed neutrino. Hence it may give the contribution to the relativistic degree of freedom (d.o.f) of the universe if they equilibrate with the cosmic soup through their mixing with the active neutrinos. Recently from Planck data there is a strong bound on the sum of the light degrees of freedom ([*d.o.f*]{}) which at $2\,\sigma$ gives $N_{\nu} < 3.2$ and comes when we combine the $D/H$ ratio with the cosmic microwave background (CMB) baryon density [@Ade:2015xua; @Cyburt:2015mya]. However, the recent LSND [@Athanassopoulos:1995iw; @Aguilar:2001ty] and MiniBooNE [@AguilarArevalo:2007it; @AguilarArevalo:2010wv; @Aguilar-Arevalo:2013pmq] data of electron excess in the antineutrino mode requiress an eV scale sterile neutrino [@Abazajian:2012ys]. The Reactor anomalies [@Mention:2011rk; @Mueller:2011nm; @Huber:2011wv; @Abdurashitov:2005tb] and the gallium experiments calibration data [@Bahcall:1994bq; @Giunti:2006bj; @Giunti:2010wz; @Giunti:2010zu] also hinted the presence of eV scale sterile neutrinos. Therefore, to go around the bound on the light relativistic [*d.o.f*]{}, a number of mechanisms have been suggested to overcome it. Among them, the popular ones are as follows. In [@Dasgupta:2013zpn; @Hannestad:2013ana; @Cherry:2016jol; @Chu:2018gxk], authors have used secret interactions where the sterile neutrinos are charged under some hidden symmetry mediated by the light gauge boson, resulting in the mixing between active and sterile neutrinos being suppressed due to the large thermal potential experienced by the sterile neutrinos. In [@Yaguna:2007wi], they have shown that relativistic [*d.o.f*]{} can be alleviated if the sterile neutrino is produced in a scenario where the reheating temperature ($T_R$) is low, $T_{R} < 7$ MeV. The authors of [@Saviano:2013ktj] have shown how to reduce $N_{\nu}$ by studying the active-sterile flavor conversion. In [@Ho:2012br], they have used MeV dark matter to reduce $N_{\nu}$ with the help of $p$-wave annihilations. Ref.[@Giovannini:2002qw] discuss about the fact that without violating cosmology we can increase the relativistic [*d.o.f*]{} by reducing the neutron to proton ratio ($n/p$). A number of these possible resolutions can be applied for our model. For example, we can consider the existence of secret interactions with some hidden sector particles which would help lower the neutrino mixing between the left-handed and right-handed neutrinos. The effect of these interactions though would have ceased to exist at a much earlier time in the universe and today we will not be able to observe them anymore. Hence our current study would not be sensitive to them. Again, the other two right handed neutrinos are in MeV mass range and have a warm spectrum [*i.e.*]{} they are neither relativistic (which makes the problem with the cosmological structure formation [@Abazajian:2004zh]) nor non-relativistic. Extensive studies in the context of structure formation for such sub-MeV RH neutrinos are there in the literature [@Bode:2000gq; @Hansen:2001zv; @Boyarsky:2008xj; @Boyanovsky:2010pw; @Lovell; @Boyanovsky:2010sv; @Villaescusa; @Merle:2013ibc; @Abazajian:2012ys; @Adhikari:2016bei]. Our model thus can be made consistent with the cosmological constraints but we have not considered them here as it is beyond the scope of this work.
Experimental limits and Collider signature {#col}
==========================================
In Table \[tab:chig\] we present a list of the various charged Higgs eigenstates that we consider in this study. We consider two cases with minimal mixings (thus consisting entirely of $\d^+$) and two with maximal mixing of $\d^+$ with $H_L^+$. For these benchmark points we study the pair production of charged Higgs states and their decay to a final state of two opposite sign charged leptons and two neutrinos. The most recent experimental bound on this process is from the ALTAS search [@Aaboud:2018jiw] of two opposite sign leptons and missing energy. They have put a bound of 500 GeV if the final state is coming from pair production of two sleptons. The production cross-section of the charged Higgs at LHC is however much lower for our model and even a 430 GeV charged Higgs is safe from the LHC bounds [^3]. Therefore, the benchmark points we have considered are allowed by the experimental observations.
{width="4.5in"}
. \[fig:feyn\]
The pair-production of the charged Higgs at LHC is through the s-channel process mediated by $\gamma$, $Z$ and $Z_R$ bosons which gives small production cross-section. In a lepton collider, on the other hand, there is an additional t-channel process mediated by the neutrinos as shown in Fig. \[fig:feyn\]. Owing to the large couplings of the charged singlet with the right-handed leptons and the small masses of the right-handed neutrinos in this model, this t-channel process will be the major pair-production channel. The masses of the right-handed neutrinos used in our analysis were taken as $$M_{N_1} = 17~\text{eV},~M_{N_2} = 6.8~\text{KeV},~ M_{N_3} = 8.2~\text{KeV},$$ for which the values of the Yukawa coupling $\lambda'_R \sim \mathcal{O}(1)$. We thus study the pair production of the charged Higgs at 1 TeV run of the International Linear Collider (ILC) [@ilc] and 3 TeV run by Compact Linear Collider (CLIC) [@clic].
![Production cross section of $H^{+}_1 H^{-}_1$ at $e^{+}e^{-}$ collider for different center-of-mass energies.[]{data-label="fig:prod"}](hc1_hc1_prod.pdf){width="3.3in" height="3.0in"}
We include the relevant vertices in FeynRules [@Alloul:2013bka], and use MadGraph[@Alwall:2014hca] for event generation, Pythia [@Sjostrand:2006za] for hadronization, and DelPhes [@deFavereau:2013fsa] for detector simulation. Fig. \[fig:prod\] shows the pair-production cross-section of the charged singlet Higgs as a function of its mass for four different center-of-mass (c.m.) energies at the lepton colliders. Here we consider the scenario of minimal mixing of the singlet charged scalar for this figure. The charged Higgs decays to a charged lepton and a right-handed neutrino and gives rise to a final state of dileptons with opposite charge ($l^{+}$ and $l^{-}$) and missing energy. Even the case where the charged Higgs is a mixture of $\d^\pm$ and $H_L^\pm$, this is the only kinematically allowed 2-body decay channel with its branching into 3-body decays being almost negligible. This is because $H_L$ does not couple to the quarks or leptons and its other physical states (the charged state with $H_L^\pm$ and $\d^\pm$ orthogonal to the one considered here and the CP-odd and CP-even neutral states coming from $H_L^0$) are much heavier. Schematically, the signal looks like $$\begin{aligned}
e^+ e^{-} \rightarrow H^{+} H^{-} \rightarrow l^{+} l^{-} \cancel{E}_T+ X,\end{aligned}$$ where $l^{\pm}$ is either one of $e^{\pm}$, $\mu^{\pm}$ and $\tau^{\pm}$ or combination of them. Inside the detector $\tau$ lepton will decay leptonically or hadronically and a small portion of it will give opposite sign dilepton and will increase the signal strength. As $\tau$ decays, eventually we get a final state signal which consist of opposite sign electron ($e^{\pm}$) or muon ($\mu^{\pm}$) or di-jet. For simulation, we consider Yukawa couplings $\lambda^{\prime}_R$, that are allowed by neutrino oscillation data.
Since we are interested in the opposite sign dilepton ($l^{+} l^{-}$) and missing energy ($\cancel{E}_{T}$) signal in the final state, the corresponding SM dominant backgrounds are as follows,
1. At the time of electron positron collision, opposite sign dilepton and missing energy can be produced as $e^+ e^- \rightarrow l^+ l^- Z\, (\rightarrow \nu_{l} \bar{\nu_{l}})$. This includes both the $ZZ$ and the virtual photon contribution.
2. Another dominant background is the $W^+ W^-$ pair production and its further leptonic decay. This can mimic the signal as $ e^+ e^- \rightarrow W^{+} W^{-} \, \rightarrow l^+ l^- \nu_{l}\, \bar{\nu_{l}}$.
3. Moreover, $t \bar{t}$ final state production and its subsequent decay will also affect the signal as background in the following manner: $e^{+} e^{-} \rightarrow t(\rightarrow b\, l^{+} \nu_{l})\, \bar{t} (\rightarrow \bar{b}\, l^{-} \bar{\nu_{l}})$.
We do not put any veto on the light jet in our analysis. Additionally, this is to note that the signal does not comprises of any $b$ jet. Therefore, a $b$-veto will reduce the backgrounds, such as $t\bar{t}$ production. Depending on the various kinematical variables, there is a clear distinction between the signal and the backgrounds as can be seen clearly in Fig.\[sig-bkg2\]. The leftmost plot in Fig. \[sig-bkg2\] shows the distribution of the transverse momentum of the hardest lepton $(p_T^{l_1})$, the one in the middle is its pseudo-rapidity distribution $(\eta_{l_1})$ and the rightmost plot shows the missing energy distribution of the signal and background events. Following these, we can select appropriate cuts on different kinematical variables to minimize the background, while protecting the signal as much as possible. The details of the cuts, that we use in our analysis are as follows:
1. We consider a signal in which final state contains two opposite sign dilepton with missing energy [*i.e.,* ]{} $l^{+} l^{-} \cancel{E}_{T}$. We implement a minimum cut on the $p_T$ of the leptons which is $p_{T,\,\, l}^{min} \geq 10$ GeV. We also implement an upper limit on the pseudo-rapidity which is $|\eta^{l}| < 2.5$. These cuts have been implemented at the time of generating the partonic event samples.
2. We select our events which contains two opposite sign dilepton.
3. From the left panel of Fig.\[sig-bkg2\] one can see that if we use cut on the hardest lepton around 130 GeV then background can be reduced. We therefore use cut on the $p_T$ of the hardest lepton, which is equal or greater than 130 GeV, $p_{T}^{l_1} \geq 130$ GeV and relatively softer cut on the second lepton which is $p_{T}^{l_2} \geq 60$ GeV.
4. The background from $ZZ$ pair production can be safely removed by applying $Z$-veto. We put a small window on the dilepton invariant mass ($m_{ll}$) which is $|m_{ll} - 91.2| \leq 10$ GeV, and reject the events, that falls within this window.
5. One of the background ($t\,\bar{t}$) contains $b$-jets in the final state. However, the signal of our interest doesn’t have any $b$-jets. Therefore we have used $b$-veto in the final state to reduce the background without affecting the signal.
6. From the middle panel of Fig.\[sig-bkg2\], it is evident that signal and backgrounds peak at different value of the pseudo-rapidity of the leading lepton. We use tighter cut on $\eta^{l_1}$. We reject events which have $|\eta^{l_1}| \geq 1 $.
7. The RH neutrinos in our scenario are very light, as they have $\sim$ eV to MeV scale masses. The decay of RH neutrinos can not happen inside the detector. Hence, they will be undetected and will give missing energy. We show the distribution of $\slashed{E}_T$ in the right panel of Fig.\[sig-bkg2\]. To reduce the background we also use cut on the missing energy, which is $\slashed{E}_{T} > 80$ GeV. This further enhances the signal to background ratio.
Using these above mentioned cuts, we can reduce the background significantly while keeping the signal at a significant level. In Table \[tab:1tev\] and Table \[tab:3tev\], we show the background cross-sections at 1 Tev ILC and 3 TeV CLIC experiments respectively, after implementing all the above-mentioned cuts. The dominant background is $W^{\pm} W^{\mp}$ production, that has a cross-section $126.88$ fb (for 1 TeV ILC) at partonic level. It is evident that the backgrounds become quite small $\sigma \sim 7, 1$ fb for ILC and CLIC respectively after the cuts.
![Distribution of signal events and background processes for different kinematical variables. The plot on the left is the transverse momentum of the hardest lepton, the middle plot is for the pseudo-rapidity of the hardest lepton and the right plot is the missing energy distribution.[]{data-label="sig-bkg2"}](lep1_pt_new3.pdf "fig:"){height="5.5cm" width="5.7cm"} ![Distribution of signal events and background processes for different kinematical variables. The plot on the left is the transverse momentum of the hardest lepton, the middle plot is for the pseudo-rapidity of the hardest lepton and the right plot is the missing energy distribution.[]{data-label="sig-bkg2"}](rap_lep1_new3.pdf "fig:"){height="5.5cm" width="5.7cm"} ![Distribution of signal events and background processes for different kinematical variables. The plot on the left is the transverse momentum of the hardest lepton, the middle plot is for the pseudo-rapidity of the hardest lepton and the right plot is the missing energy distribution.[]{data-label="sig-bkg2"}](met_new3.pdf "fig:"){height="5.5cm" width="5.7cm"}
[||c|c||]{}
[c|c]{}\
Channels & Cross-section (fb)\
$l^{+} l^{-} Z \,(\rightarrow \nu_{l}\bar{\nu_{l}}) $ & $18.68$\
$W^{+}(\rightarrow l^{+} \nu_{l})\, W^{-} (\rightarrow l^{-} \bar{\nu_{l}})$ & $126.88$\
$t (\rightarrow b l^{+} \nu_{l})\,\bar{t} (\rightarrow \bar{b} l^{-} \bar{\nu_{l}})$ & $13.96$\
Total Backgrounds & \
&
[c|c|c|c|c|c]{}\
A0+A1 & A2 & A3 & A4 & A5 & A6 \
$10.79$ & $5.99$ & $5.54$ & $5.54$ & $2.30$ & $1.67$\
$52.72$ & $32.15$ & $32.15$ & $32.15$ & $12.44$ & $7.05$\
$3.10$ & $0.78$ & $0.78$ & $0.1$ & $0.08$ & $0.05$\
& & & & & 8.77\
\
[||c|c||]{}
[c|c]{}\
Channels & Cross-section (fb)\
$l^{+} l^{-} Z \,(\rightarrow \nu_{l}\bar{\nu_{l}}) $ & $6.33$\
$W^{+}(\rightarrow l^{+} \nu_{l})\, W^{-} (\rightarrow l^{-} \bar{\nu_{l}})$ & $13.85$\
$t (\rightarrow b l^{+} \nu_{l})\,\bar{t} (\rightarrow \bar{b} l^{-} \bar{\nu_{l}})$ & $1.76$\
Total Backgrounds & \
&
[c|c|c|c|c|c]{}\
A0+A1 & A2 & A3 & A4 & A5 & A6 \
$3.0$ & $2.89$ & $2.86$ & $2.86$ & $0.54$ & $0.44$\
$5.45$ & $5.1$ & $5.1$ & $5.1$ & $1.34$ & $1.13$\
$0.05$ & $0.02$ & $0.02$ & 0.005 & 0.002 & 0.002\
& & & & & 1.57\
\
The signal cross-sections and their statistical significance over the background are given in Tab. \[tab:signal\] for the chosen benchmark points. Clearly the case with no mixing in the Higgs state gives a much larger cross-section. This is because the $\d^\pm l_R^\mp \nu_R$ vertex is primarily responsible for the charged Higgs pair-production. The mixing of $\d^\pm$ with $H_L^\pm$ will only introduce an extra factor of $\cos^4 \theta$ in the pair-production cross-section, where $\theta$ is the mixing angle, resulting in a decrease in the cross-section. As evident the cross-section is enormous in the lepton collider. As an illustrative example, for a 1 TeV charged Higgs $H^{\pm}_1$, the partonic cross-section is $\sigma \sim 100$ fb. After the cuts, the cross-section reduces to $\sigma \sim 27$ fb. This is order of magnitude larger than the after-cut background cross-section. We compute the statistical significance ($\mathcal{S}$) of signal over background using the following expression, $$\mathcal{S} = \sqrt{2 \times \left[ (s+b) {\rm ln}(1 + \frac{s}{b}) - s \right]}.$$ In the above, $s$ and $b$ denote the signal and background events. The significance has been shown in Tab. \[tab:signal\]. As expected the case with zero mixing has a much better significance of signal over background boosting its chances to be discovered even in the early run of the upcoming lepton colliders. In particular, we show that only $\mathcal{L}=$1 $\rm{fb}^{-1}$ luminosity is required in the zero-mixing scenario to discover charged Higgs $H^{\pm}_1$ with mass range $473$ GeV - 1 TeV. For the relatively less optimistic scenario of half-mixing, 3 $\rm{fb}^{-1}$ will be required to claim discovery.
[||c|c|c||]{}
[C[0.7cm]{}|C[1.2cm]{}|C[1.3cm]{}|C[1.1cm]{}|C[1.2cm]{}]{}\
& COM Energy & Mass (GeV) & Mixing & CS (fb)\
BP1 & 1 TeV & 473.32 & Zero & 192.67\
BP2 & 3 TeV & 1000.70 & Zero & 100.31\
BP3 & 1 TeV & 432.58 & Half & 49.50\
BP4 & 3 TeV & 1000.92 & Half & 17.86\
&
[C[1.2cm]{}|C[0.9cm]{}|C[0.9cm]{}|C[0.9cm]{}|C[0.9cm]{}|C[0.9cm]{}|C[0.9cm]{}]{}\
& & & & &\
& & & & &\
79.75 & 62.13 & 62.02 & 62.02 & 57.78 & 53.63\
38.21 & 35.57 & 35.56 & 35.55 & 28.08 & 27.07\
19.19 & 14.62 & 14.59 & 14.59 & 13.54 & 12.51\
6.83 & 6.33 & 6.33 & 6.33 & 5.08 & 4.99\
&
[C[1.7cm]{}|C[1.7cm]{}]{}\
&\
&\
11.73 & 20.32\
10.78 & 18.67\
3.56 & 6.174\
2.96 & 5.13\
\
Conclusion {#conc}
==========
In this work, we have studied left-right symmetric extension of Zee model. The model has very different characteristics features as compared to the minimal left-right symmetric model. It is well known that the basic Zee model is ruled out from light neutrino mass and mixing constraints. Going to the left-right symmetric framework, it is possible to evade the tension with the neutrino oscillation data. The model consists of three lighter right-handed neutrino states, that can have masses from MeV down to eV scale. Additionally, the model also contains an additional charged scalar $\delta^{\pm}$. The charged scalar, due to its additional interaction with charged leptons and right-handed neutrinos, can be copiously produced at a lepton collider via the $t$-channel processes.
We discuss light neutrino mass generation in this model and fit the observed data. The light neutrino mass matrix is a combination of both the Type-I and Type-II seesaw matrices. The Type-II contribution and the right-handed neutrino mass matrix, that participates in Type-I seesaw, are however generated through one loop process with the charged leptons and charged Higgs fields as mediators. We fit the observed light neutrino mass square differences and the PMNS mixing in this model, and derive constraints on model parameters. With the set of parameters, that satisfy the neutrino mass constrains, we extensively analyze the charged Higgs phenomenology at 1 TeV ILC and 3 TeV CLIC. Owing to the extra interaction of the charged Higgs with the right-handed neutrinos and for moderately large Yukawa couplings, the cross-section at $e^+e^-$ collider is enormous, as compared to the LHC. We find that in the most optimistic scenario, where the lighter charged Higgs state $H^{\pm}_1$ is a pure charged scalar state $\delta^{\pm}$, the cross-section for pair-production of charged Higgs can be $\sigma \sim \mathcal{O}(1)$ pb for $M_{H^{\pm}} \sim 473$ GeV, and c.m.energy $\sqrt{s}=1$ TeV. For CLIC, that can operate with c.m.energy $\sqrt{s}=3$ TeV, the charged Higgs of mass 1 TeV is also accessible ($\sigma \sim 100 $ fb for pair-production).
We consider the subsequent decay of the charged Higgs into a lepton and a neutrino, that is the only possible channel for this model. This leads to the final states $l^{+}l^{-} + \slashed{E}_T$, that we analyze in detail, taking into account detector simulation. We show that a discovery of the charged Higgs of mass in between 473-1000 GeV in the di-lepton + $\slashed{E}_T$ will require only 1-3 $\rm{fb}^{-1}$ integrated luminosity at an $e^+e^-$ collider, operating with c.m.energy $\sqrt{s}=1, 3$ TeV. Therefore, this model can most economically be tested at the very early run of ILC or CLIC.
M.M. would like to acknowledge the DST-INSPIRE research grant IFA14-PH-99 and hospitality of CHEP, IISc, Bengaluru, where part of the discussion has been carried out. A.P. is supported by the SERB National Postdoctoral Fellowship \[PDF/2016/000202\]. S.K. thanks Prof. Sandhya Choubey for discussions. S.K. also acknowledges the cluster computing facility at HRI (http://cluster.hri.res.in). S.K. would also like to thank the Department of Atomic Energy (DAE) Neutrino Project of Harish-Chandra Research Institute. The authors would like to thank Dr. Arnab Dasgupta for very useful discussions at the early stage of this work.
Charged Higgs boson eigenstates used for neutrino phenomenology
===============================================================
Here we list the charged Higgs boson masses and mixings that have been used for the neutrino phenomenology in our study. We consider that the lightest charged Higgs boson $H^{\pm}_1$ has a mass around 473 GeV and is almost entirely consisting of the singlet charged Higgs field $\delta^{\pm}$. The charged Higgs boson states, after diagonalization, consist of two Goldstone bosons $G_1^\pm$ and $G_2^\pm$ and three physical charged Higgs bosons with $$M_{H_1^\pm} = 473.32~\text{GeV},~~M_{H_2^\pm} = 2534.94~\text{GeV},~~M_{H_3^\pm} = 15.95~\text{TeV}.$$ The corresponding eigenstates can be identified as $$\begin{pmatrix}
H_1^\pm\\H_2^\pm\\H_3^\pm\\G_1^\pm\\G_2^\pm
\end{pmatrix}
=
\begin{pmatrix}
0.0000127106 & 0.00225768 & 0.000109819 & 0.0000433475 & 0.999997 \\ -0.114973 & 0.000916758 & -0.993368 & 0.0000176018 & 0.000108482 \\ -0.000105399 & -0.999813 & -0.000910601 & -0.0191964 & 0.0022582 \\ -0.00574986 & -0.0191961 & 0.000665494 & 0.999799 & 4.70382 \times 10^{-16} \\ 0.993352 & -0.00011112 & -0.114971 & 0.00578718 &
2.76194 \times 10^{-18}
\end{pmatrix}
\begin{pmatrix}
\phi_1^\pm \\ \phi_2^\pm \\ H_L^\pm \\H_R^\pm \\ \delta^\pm
\end{pmatrix}.$$ The $5\times5$ matrix in the above is the charged Higgs boson rotation matrix $V$.
\[
P. Minkowski, Phys. Lett. B [**67**]{} (1977) 421; T. Yanagida, proceedings of the [*Workshop on Unified Theories and Baryon Number in the Universe*]{}, Tsukuba, 1979, eds; A. Sawada, A. Sugamoto, KEK Report No. 79-18, Tsukuba; S. Glashow, in [*Quarks and Leptons, Cargèse 1979*]{}, eds; M. L' evy. et al., (Plenum, 1980, New York); M. Gell-Mann, P. Ramond, R. Slansky, proceedings of the [*Supergravity Stony Brook Workshop*]{}, New York, 1979, eds. P. Van Niewenhuizen, D. Freeman (North-Holland, Amsterdam). R. Mohapatra, G. Senjanovi' c, Phys.Rev.Lett. [**44**]{} (1980) 912. E. Ma, Phys. Rev. Lett. [**81**]{}, 1171 (1998) doi:10.1103/PhysRevLett.81.1171 \[hep-ph/9805219\]; F. Bonnet, M. Hirsch, T. Ota and W. Winter, JHEP [**1207**]{}, 153 (2012) doi:10.1007/JHEP07(2012)153 \[arXiv:1204.5862 \[hep-ph\]\]; D. Aristizabal Sierra, A. Degee, L. Dorame and M. Hirsch, JHEP [**1503**]{}, 040 (2015) doi:10.1007/JHEP03(2015)040 \[arXiv:1411.7038 \[hep-ph\]\]. A. Zee, Phys. Lett. [**93B**]{}, 389 (1980) Erratum: \[Phys. Lett. [**95B**]{}, 461 (1980)\]. doi:10.1016/0370-2693(80)90349-4, 10.1016/0370-2693(80)90193-8 P. H. Frampton, M. C. Oh and T. Yoshikawa, Phys. Rev. D [**65**]{}, 073014 (2002) doi:10.1103/PhysRevD.65.073014 \[hep-ph/0110300\]. Y. Koide, Nucl. Phys. Proc. Suppl. [**111**]{}, 294 (2002) doi:10.1016/S0920-5632(02)01726-7 \[hep-ph/0201250\]; X. G. He, Eur. Phys. J. C [**34**]{}, 371 (2004) doi:10.1140/epjc/s2004-01669-8 \[hep-ph/0307172\]; J. Herrero-García, T. Ohlsson, S. Riad and J. Wirén, JHEP [**1704**]{}, 130 (2017) doi:10.1007/JHEP04(2017)130 \[arXiv:1701.05345 \[hep-ph\]\].
R. N. Mohapatra and J. C. Pati, Phys. Rev. D [**11**]{}, 566 (1975); G. Senjanovic and R. N. Mohapatra, Phys. Rev. D [**12**]{}, 1502 (1975). P. Fileviez Perez, C. Murgui and S. Ohmer, Phys. Rev. D [**94**]{}, no. 5, 051701 (2016) doi:10.1103/PhysRevD.94.051701 \[arXiv:1607.00246 \[hep-ph\]\]. P. Fileviez Perez and C. Murgui, Phys. Rev. D [**95**]{}, no. 7, 075010 (2017) doi:10.1103/PhysRevD.95.075010 \[arXiv:1701.06801 \[hep-ph\]\]. M. A. B. Beg and H. -S. Tsao, Phys. Rev. Lett. [**41**]{}, 278 (1978); R. N. Mohapatra and G. Senjanovic, Phys. Lett. B [**79**]{}, 283 (1978); K. S. Babu and R. N. Mohapatra, Phys. Rev. D [**41**]{}, 1286 (1990); S. M. Barr, D. Chang and G. Senjanovic, Phys. Rev. Lett. [**67**]{}, 2765 (1991); R. N. Mohapatra and A. Rasin, Phys. Rev. Lett. [**76**]{}, 3490 (1996); R. Kuchimanchi, Phys. Rev. Lett. [**76**]{}, 3486 (1996); R. N. Mohapatra, A. Rasin and G. Senjanovic, Phys. Rev. Lett. [**79**]{}, 4744 (1997); K. S. Babu, B. Dutta and R. N. Mohapatra, Phys. Rev. D [**65**]{}, 016005 (2001); R. Kuchimanchi, Phys. Rev. D [**82**]{}, 116008 (2010).
C. Athanassopoulos [*et al.*]{} \[LSND Collaboration\], Phys. Rev. Lett. [**75**]{}, 2650 (1995) \[nucl-ex/9504002\]. A. Aguilar-Arevalo [*et al.*]{} \[LSND Collaboration\], Phys. Rev. D [**64**]{}, 112007 (2001) \[hep-ex/0104049\]. A. A. Aguilar-Arevalo [*et al.*]{} \[MiniBooNE Collaboration\], Phys. Rev. Lett. [**98**]{}, 231801 (2007) \[arXiv:0704.1500 \[hep-ex\]\]. A. A. Aguilar-Arevalo [*et al.*]{} \[MiniBooNE Collaboration\], Phys. Rev. Lett. [**105**]{}, 181801 (2010) \[arXiv:1007.1150 \[hep-ex\]\].
A. A. Aguilar-Arevalo [*et al.*]{} \[MiniBooNE Collaboration\], Phys. Rev. Lett. [**110**]{}, 161801 (2013) \[arXiv:1303.2588 \[hep-ex\]\].
G. Ecker, W. Grimus and H. Neufeld, Phys. Lett. B [**127**]{} , 365 (1983) \[Erratum-ibid. B [**132**]{} , 467 (1983)\]; R. N. Mohapatra, G. Senjanovic and M. D. Tran, Phys. Rev. D [**28**]{}, 546 (1983); M. E. Pospelov, Phys. Rev. D [**56**]{}, 259 (1997); Y. Zhang, H. An, X. Ji and R. N. Mohapatra, Nucl. Phys. B [**802**]{}, 247 (2008) \[arXiv:0712.4218 \[hep-ph\]\]; A. Maiezza, M. Nemevsek, F. Nesti and G. Senjanovic, Phys. Rev. D [**82**]{}, 055022 (2010) \[arXiv:1005.5160 \[hep-ph\]\]; For bounds on $W_R$ mass from radiative correction effects, see J. Chakrabortty, J. Gluza, R. Sevillano and R. Szafron, JHEP [**1207**]{}, 038 (2012).
F. Capozzi, E. Di Valentino, E. Lisi, A. Marrone, A. Melchiorri and A. Palazzo, Phys. Rev. D [**95**]{}, no. 9, 096014 (2017) doi:10.1103/PhysRevD.95.096014 \[arXiv:1703.04471 \[hep-ph\]\]. I. Esteban, M. C. Gonzalez-Garcia, M. Maltoni, I. Martinez-Soler and T. Schwetz, JHEP [**1701**]{}, 087 (2017) doi:10.1007/JHEP01(2017)087 \[arXiv:1611.01514 \[hep-ph\]\]; NuFIT 3.2 (2018), www.nu-fit.org.
P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], Astron. Astrophys. [**594**]{}, A13 (2016) \[arXiv:1502.01589 \[astro-ph.CO\]\]. R. H. Cyburt, B. D. Fields, K. A. Olive and T. H. Yeh, Rev. Mod. Phys. [**88**]{}, 015004 (2016) \[arXiv:1505.01076 \[astro-ph.CO\]\].
K. N. Abazajian [*et al.*]{}, arXiv:1204.5379 \[hep-ph\]. G. Mention, M. Fechner, T. Lasserre, T. A. Mueller, D. Lhuillier, M. Cribier and A. Letourneau, Phys. Rev. D [**83**]{}, 073006 (2011) \[arXiv:1101.2755 \[hep-ex\]\]. T. A. Mueller [*et al.*]{}, Phys. Rev. C [**83**]{}, 054615 (2011) \[arXiv:1101.2663 \[hep-ex\]\]. P. Huber, Phys. Rev. C [**84**]{}, 024617 (2011) Erratum: \[Phys. Rev. C [**85**]{}, 029901 (2012)\] \[arXiv:1106.0687 \[hep-ph\]\]. J. N. Abdurashitov [*et al.*]{}, Phys. Rev. C [**73**]{}, 045805 (2006) \[nucl-ex/0512041\]. J. N. Bahcall, P. I. Krastev and E. Lisi, Phys. Lett. B [**348**]{}, 121 (1995) \[hep-ph/9411414\]. C. Giunti and M. Laveder, Mod. Phys. Lett. A [**22**]{}, 2499 (2007) \[hep-ph/0610352\]. C. Giunti and M. Laveder, Phys. Rev. D [**82**]{}, 053005 (2010) \[arXiv:1005.4599 \[hep-ph\]\]. C. Giunti and M. Laveder, Phys. Rev. C [**83**]{}, 065504 (2011) doi:10.1103/PhysRevC.83.065504 \[arXiv:1006.3244 \[hep-ph\]\].
B. Dasgupta and J. Kopp, Phys. Rev. Lett. [**112**]{}, no. 3, 031803 (2014) \[arXiv:1310.6337 \[hep-ph\]\]. S. Hannestad, R. S. Hansen and T. Tram, Phys. Rev. Lett. [**112**]{}, no. 3, 031802 (2014) \[arXiv:1310.5926 \[astro-ph.CO\]\]. J. F. Cherry, A. Friedland and I. M. Shoemaker, arXiv:1605.06506 \[hep-ph\].
X. Chu, B. Dasgupta, M. Dentler, J. Kopp and N. Saviano, arXiv:1806.10629 \[hep-ph\]. C. E. Yaguna, JHEP [**0706**]{}, 002 (2007) \[arXiv:0706.0178 \[hep-ph\]\]. N. Saviano, A. Mirizzi, O. Pisanti, P. D. Serpico, G. Mangano and G. Miele, Phys. Rev. D [**87**]{}, 073006 (2013) \[arXiv:1302.1200 \[astro-ph.CO\]\]. C. M. Ho and R. J. Scherrer, Phys. Rev. D [**87**]{}, no. 6, 065016 (2013) \[arXiv:1212.1689 \[hep-ph\]\]. M. Giovannini, H. Kurki-Suonio and E. Sihvola, Phys. Rev. D [**66**]{}, 043504 (2002) \[astro-ph/0203430\]. K. Abazajian, E. R. Switzer, S. Dodelson, K. Heitmann and S. Habib, Phys. Rev. D [**71**]{}, 043507 (2005) \[astro-ph/0411552\]; R. de Putter, O. Mena, E. Giusarma, S. Ho, A. Cuesta, et al., Astrophys. J. 761,12 (2012), 1201.1909.
P. Bode, J. P. Ostriker and N. Turok, Astrophys. J. [**556**]{}, 93 (2001) \[astro-ph/0010389\]. S. H. Hansen, J. Lesgourgues, S. Pastor and J. Silk, Mon. Not. Roy. Astron. Soc. [**333**]{}, 544 (2002) \[astro-ph/0106108\]. A. Boyarsky, J. Lesgourgues, O. Ruchayskiy and M. Viel, JCAP [**0905**]{}, 012 (2009) \[arXiv:0812.0010 \[astro-ph\]\]. M. R. Lovell, V. Eke, C. S. Frenk, L. Gao, A. Jenkins, et al., Mon. Not. Roy. Astron. Soc.420, 2318 (2012),1104.2929.
D. Boyanovsky and J. Wu, Phys. Rev. D [**83**]{}, 043524 (2011) \[arXiv:1008.0992 \[astro-ph.CO\]\]. D. Boyanovsky, Phys. Rev. D [**83**]{}, 103504 (2011) \[arXiv:1011.2217 \[astro-ph.CO\]\]. F. Villaescusa-Navarro and N. Dalal, JCAP 1103, 024 (2011), 1010.3008.
A. Merle and V. Niro, Phys. Rev. D [**88**]{}, no. 11, 113004 (2013) \[arXiv:1302.2032 \[hep-ph\]\]. M. Drewes [*et al.*]{}, JCAP [**1701**]{}, no. 01, 025 (2017) \[arXiv:1602.04816 \[hep-ph\]\]. M. Aaboud [*et al.*]{} \[ATLAS Collaboration\], arXiv:1803.02762 \[hep-ex\]. H. Baer [*et al.*]{}, arXiv:1306.6352 \[hep-ph\]; C. Adolphsen [*et al.*]{}, arXiv:1306.6353 \[physics.acc-ph\]; C. Adolphsen [*et al.*]{}, arXiv:1306.6328 \[physics.acc-ph\]; T. Behnke [*et al.*]{}, arXiv:1306.6329 \[physics.ins-det\]; T. Behnke [*et al.*]{}, arXiv:1306.6327 \[physics.acc-ph\]. H. Abramowicz [*et al.*]{} \[CLIC Detector and Physics Study Collaboration\], arXiv:1307.5288 \[hep-ex\]; D. Dannheim [*et al.*]{}, arXiv:1208.1402 \[hep-ex\]; E. Accomando [*et al.*]{} \[CLIC Physics Working Group\], hep-ph/0412251; L. Linssen [*et al.*]{}, arXiv:1202.5940; N. Alipour Tehrani, J.-J. Blaising, B. Cure, D. Dannheim, F. Duarte Ramos, K. Elsener et al., CLICdet: The post-CDR CLIC detector model, CLICdp-Note-2017-001 (Mar, 2017).
A. Alloul, N. D. Christensen, C. Degrande, C. Duhr and B. Fuks, Comput. Phys. Commun. [**185**]{}, 2250 (2014) doi:10.1016/j.cpc.2014.04.012 \[arXiv:1310.1921 \[hep-ph\]\]. J. Alwall [*et al.*]{}, JHEP [**1407**]{}, 079 (2014) \[arXiv:1405.0301 \[hep-ph\]\]; J. Alwall, M. Herquet, F. Maltoni, O. Mattelaer and T. Stelzer, JHEP [**1106**]{}, 128 (2011) \[arXiv:1106.0522 \[hep-ph\]\].
T. Sjostrand, S. Mrenna and P. Z. Skands, JHEP [**0605**]{}, 026 (2006) \[hep-ph/0603175\].
J. de Favereau [*et al.*]{} \[DELPHES 3 Collaboration\], JHEP [**1402**]{}, 057 (2014) \[arXiv:1307.6346 \[hep-ex\]\]; M. Selvaggi, J. Phys. Conf. Ser. [**523**]{}, 012033 (2014); A. Mertens, J. Phys. Conf. Ser. [**608**]{}, no. 1, 012045 (2015).
[^1]: These models need extra singlet fermions for neutrino mass generation.
[^2]: Even if we keep both $Y_l$ and $\lambda'_L$ to be non-zero, for which $M^L_{\nu} \neq 0$, the values of the elements of $\lambda'_L$ matrix satisfying the neutrino constraints turn out to be very small to have any observable consequences for our study.
[^3]: For a set of loose cuts denoted by SF1 in [@Aaboud:2018jiw], a production cross-section for $l^+ l^- \cancel{E}_T$ greater than 2 fb is ruled out while we only get 0.23 fb for $M_{H^\pm}$ = 450 GeV with similar cuts.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss a novel use of the Geant4 simulation toolkit to model molecular transport in a vacuum environment, in the molecular flow regime. The Geant4 toolkit was originally developed by the high energy physics community to simulate the interactions of elementary particles within complex detector systems. Here its capabilities are utilized to model molecular vacuum transport in geometries where other techniques are impractical. The techniques are verified with an application representing a simple vacuum geometry that has been studied previously both analytically and by basic Monte Carlo simulation. We discuss the use of an application with a very complicated geometry, that of the Large Synoptic Survey Telescope camera cryostat, to determine probabilities of transport of contaminant molecules to optical surfaces where control of contamination is crucial.'
author:
- JACK SINGAL
- 'J. BRIAN LANGTON, RAFE SCHINDLER'
title: |
Geant4 Applications for Modeling Molecular Transport\
in Complex Vacuum Geometries
---
Introduction {#intro}
============
The cryostat of the Large Synoptic Survey Telescope camera (LSSTCam)[@LSSTover; @SPIE1] represents a significant challenge for vacuum design.[@RSI] With a volume of 2.9 m$^3$, it will contain nearly 1000 kg of material, and have more than 40 m$^2$ of exposed surface area within. The geometry of the LSSTCam cryostat presents an environment too complicated for vacuum transport to be modeled by analytical calculations, or by techniques such as thermal conductance calculations in conjunction with finite element analysis.[@AnsysP] Instead, a fully three dimensional model of the geometry is required, with a full Monte Carlo treatment of the initial conditions and subsequent propagation of particles through their interactions with surfaces. Monte Carlo calculations of vacuum transport have been performed extensively via stand-alone custom code, but for relatively simpler geometries.[@Davis; @Vaktrak] The complexity of the LSSTCam cryostat would render the specification of the geometry by similar stand-alone code impossible. Some complex vacuum geometries have be modeled in the MolFlow+ simulation suite.[@Molflow] However, for even more complicated geometries such as in the LSSTCam cryostat, the Geant4 simulation toolkit[@Geant1; @Geant2] presents itself as an ideal platform for developing an application to model vacuum transport.
![CAD rendering of an LSST camera cryostat design in cross section. A brief summary of the cryostat is given in §\[intro\] and its geometry is discussed in §\[setup\]. Shown here is one science raft tower module (RTM) containing CCD detectors and electronics boards and other structures, the grid (rendered in blue), the cryoplate (pink and dark orange), the cold shroud (red), the lower shroud (orange), some of the pumping plenum and chimneys (dark green and white), the feed-thru flange (transparent light green), and the pump plate (blue). The cryostat body and the L3 lens are rendered in grey. This particular scenario is known as “option 3”, however the differences between the various scenarios are subtle enough not to dominate a large-scale view. Some structures are not shown. []{data-label="cryo"}](fig1.eps){width="3.0in"}
[cccc]{} $L$/$R$ & $P_e$ & $P_m$ & $P_G$\
0.5 & 0.80 & 0.80 & 0.80\
1.0 & 0.67 & 0.68 & 0.67\
1.5 & 0.58 & 0.57 & 0.57\
2.0 & 0.51 & 0.53 & 0.51\
3.0 & 0.42 & 0.43 & 0.44\
4.0 & 0.36 & 0.36 & 0.35\
5.0 & 0.32 & 0.32 & 0.29\
10.0 & 0.20 & — & 0.18\
20.0 & 0.11 & — & 0.10\
\[resu\]
Geant4 is a Monte Carlo simulation toolkit used extensively in particle physics, nuclear physics, medical physics, and aerospace applications. It is a flexible, self-contained software package that allows the definition of a volume containing real materials arranged into complex geometries and then the transport of elementary or composite particles within the volume. Geant4 follows any fundamental interactions of the particles with the materials, as well as any products of those interactions. The toolkit consists of a large number of C++ classes which the application developer uses in order to create a specific application that models a physical system.
In this letter we describe Geant4 applications that were developed to model vacuum transport in the molecular flow regime. The general strategy is summarized in §\[ao\]. In §\[confirm\] an application with a simple geometry to verify the techniques used is discussed. In §\[setup\] a specific application to the LSSTCam cryostat is described. Several cases of the cryostat geometry and initial conditions of the molecular distribution are evaluated to demonstrate the power and flexibility of this design tool. These general techniques could of course be applied to any vacuum geometry, and this is explored in §\[disc\].[^1]
Application Strategy {#ao}
====================
To develop a Geant4 application, particles, their interactions, a detector geometry, and relevant cataloging of outcomes (known as “scoring”) must be implemented. In Geant4 particles are specified by properties including mass, and charge, and, reflecting its origin in the particle physics community, quantities such as spin, quark contents, and lifetime and decay modes. There are many pre-defined particles (e.g. electrons, positrons, photons, etc), and application developers can create their own by specifying the relevant properties. Particles can interact with other particles and surfaces, as well as leave the area of interest or decay into other particles if possible, as specified by the application developer. The possible interaction outcomes and their probabilities for any particle type with any other particle type or surface can be specified by implementations of included classes. In Geant4, evolution of a system moves the particles through tiny steps, at the end of which it is evaluated whether the particle has encountered any other particles or any surfaces, and if so, an interaction is carried out. Any original and subsequently created particles are tracked until they reach a defined outcome, such as encountering a given surface or exiting the detector, or when a defined amount of time has elapsed. Scoring involves counting the number of times a certain event has happened, for example a type of particle decaying or touching a given region of a surface. Recent years have seen applications using the capabilities of Geant4 extend beyond the original fields, such as in modeling phonon propagation in a silicon crystal for dark matter direct detection[@CDMS] and modeling the damage to DNA from radiation in various environments[@Geant4DNA].
To model molecular transport in vacuum with the Geant4 toolkit, the vacuum space of interest is defined as the detector geometry. All structures are specified according to their material (stainless steel, copper, G10, silicon, and so on) in the standard Geant4 way taking advantage of the Geant4 material database, which uses National Institutes of Standards and Technology (NIST) standard values. They are also assigned their predicted operating temperatures. Custom particles are specified with implementations of G4ParticleDefinition, to represent the desired molecular species (“Waterons” and “Nitroginos” for example, representing water and nitrogen molecules, respectively) that lack internal structure and charge but have the proper molecular weight, allowing, for example, the mean velocities and residence times to be affected by contact with surfaces of different temperatures.
There are no particle-particle interactions, as the physical situations being modeled are high vacuum systems where transport is in the molecular flow regime, so that the “ballistic” approximation is valid. For example, the LSSTCam cryostat will have pressures of $\sim$10$^{-6}$ torr upon pump-down and around two orders of magnitude lower when surfaces are cooled to their operating temperatures, so that the molecular mean free path is significantly larger than the dimensions involved. Particles thus only interact with material surfaces, and then in only a particular way that represents how surfaces direct particle transport in the molecular flow regime.
When a particle encounters a surface in the simulation, it is adsorbed at the point of contact and then desorbed with the probability of a given direction proportional to the cosine of the angle with the normal vector to the surface at the point. We specify these interactions in an implementation of G4VDiscreteProcess. This model of surface interaction is used in other Monte Carlo implementations of molecular flow vacuum transport.[@Davis; @Vaktrak; @Molflow] It is based on the considerations that the probability of a particle successfully desorbing from a surface onto which it is adsorbed is proportional to the normal component of its velocity away from the surface, which is the cosine of the angle of the velocity vector to the surface normal. Thus the angular distribution of desorbed molecules is described by a cosine function from the surface normal. This is discussed from a theoretical perspective in surface chemistry texts[@PMC], and some experimental verifications have been carried out.[@Rettner89; @Steinruck84] Like many assumptions in simulations, the cosine to the surface normal desorption probability distribution is a model which represents real-world data in a variety of circumstances, but may not hold on a microscopic level in every single case.[@Steinruck84] However, we believe that following this approach, which has a sound theoretical basis and is the one used in the simulation literature, is the most appropriate.
An implementation of G4GeneralParticleSource is used to originate particles with initial momentum directions and positions chosen randomly according to distributions appropriate to the application, and from the regions of interest within the volume of the vacuum. Particles are tagged with the name of the first surface they encounter, which can be considered the originating item, and this is done in an implementation of G4UserSteppingAction with the originating surface name and any other desired information attached with an implementation of G4UserTrackInformation.
Scoring is implemented such that particles propagate until they reach one of several particular specified terminating outcomes, for example exiting through a pump port, or contacting a particular surface. These terminating outcomes are implemented through the use of regions specified in the detector construction, which are checked at each step in our implementation of G4UserSteppingAction. Each molecule’s originating item and terminating outcome is cataloged in the output data file, which consists of columns representing the different outcomes with a “1” entered into the proper outcome column for each molecule. In this manner the total number of molecules resulting in each outcome can be easily summed.
![ [View of an OpenGL rendering of the geometry used in the LSST camera cryostat Geant4 application. For clarity, certain structures are rendered in wireframe and others are rendered invisible. This particular geometrical scenario shown is known as “option 3”, however the differences between the various scenarios are subtle enough not to dominate such a large-scale view. ]{}[]{data-label="geo1"}](fig2.eps){width="3.0in"}
[cccc]{} & & &\
Geometry & Pump & Getter & CCD Front Surf.\
Baseline & 12% & 32% & 56%\
Option 2 & 15% & 32% & 54%\
Option 3 & 8% & 36% & 56%\
Option 3 + 2xI & 23% & 29% & 47%\
Option 3 + 4xI & 50% & 20% & 30%\
& & &\
Geometry & Pump & Getter & CCD Front Surf.\
Baseline & 13% & 31% & 57%\
Option 2 & 13% & 35% & 52%\
Option 3 & 8% & 37% & 55%\
Option 3 + 2xI & 24% & 24% & 52%\
Option 3 + 4xI & 50% & 17% & 33%\
\[contribs\]
Confirmation of techniques with a simple geometry {#confirm}
=================================================
To test the validity of using the Geant4 toolkit to model molecular transport in vacuum with the techniques discussed here, we have developed a Geant4 application with a very simple geometry for which transport results can be compared to those in the literature. Following Davis [@Davis], the geometry modeled is a right circular cylinder (i.e. a tube), where molecules enter at one end with randomly distributed initial radial and azimuthal positions in the plane of the opening and momentum vectors randomly distributed proportional to the cosine of their angle with the surface normal of the opening plane. Molecules are propagated through the cylinder with wall interactions as discussed in §\[ao\]. The quantity of interest is whether a molecule first reaches the opposite end of the cylinder or returns to the opening through which it entered. Thus this is just the problem of molecular transport probability through a pipe of a given length and radius.
The probabilities vary with the ratio of the length ($L$) to the radius ($R$) of the cylinder, and have been calculated analytically by Clausing[@Clausing] in 1930 and subsequently by Davis[@Davis] using simple Monte Carlo methods. The Davis code tracks a particle and, if it is coincident with the coordinates of a surface, resets the trajectory to be in another direction with a likelihood probabilty proportional to the cosine of the angle with the normal vector to the surface at the point in question. A Geant4 application for this geometry is implemented and the results for various $L/R$ values compared. The application has regions specified at the two ends of the cylinder that terminate a given molecule’s trajectory when they are entered, and the terminating region is recorded in the output file as described in §\[ao\]. Table 1 shows the probabilities for a molecule to reach the opposite opening for different values of $L$/$R$, as calculated analytically by Clausing, via Monte Carlo methods by Davis, and by the Geant4 application. The statistical uncertainty on the probabilities we have obtained can be estimated from standard theory as $\sigma=[\sqrt{n}/M]/M$, for a situation where there are $n$ randomized trials with $M$ possible discrete outcomes. In this case, $\sigma$=0.005. The results from the Geant4 application show very good agreement with the probabilities obtained by the other methods.
LSSTCam Cryostat application {#setup}
============================
Figure \[cryo\] shows a CAD rendering of an LSSTCam cryostat layout. The strict requirements on LSST’s photon throughput necessitate a detailed knowledge of the effect of potential molecular contaminants that may deposit on the cold CCD focal plane surface within the cryostat. One crucial input to any model that addresses this question is the transport of molecules within the cryostat. In particular, how likely is a contaminant molecule from the bulk of the vacuum space where most of the electronics and other structures are located to reach the focal plane region rather than being removed from the system elsewhere?
The LSSTCam cryostat geometry consists of a significant number of internal structures contained within the cryostat body, which forms the vacuum enclosure. The body is a truncated conical section with small interior dimension (ID) 0.95 m and large ID 1.06 m. It is sealed at the smaller “front” end (through which light from the telescope enters) by a glass lens designated “L3”. Directly behind L3 is the focal plane consisting of the 3.2 billion 10 $\mu$m CCD pixels arranged in 21 raft tower modules (RTMs). Each RTM contains nine 4kx4k pixel CCD detectors, and is itself a self-contained camera with the CCD packages, a Silicon Carbide “raft” on which they are mounted and aligned, conductance barriers, electronics boards, thermal and wall structures, and connectors and cabling. Four triangular shaped RTMs carrying guide and wavefront sensors are located at the corners of the focal plane. The RTMs pass through a Silicon Carbide ceramic honeycomb structure known as the grid, which kinematically supports the raft. Behind the grid, the cryoplate is a mechanical structure that carries the balance of the load of the RTM. The cryoplate contains cryogenic refrigerant channels that provide cooling to the RTM electronics and other structures. Behind the cryoplate, a separate cold circuit removes heat from the rearmost electronics at a higher temperature. There are also various shrouds, chimneys, and plenums that provide thermal radiation shields and direct molecular transport. The rear of the cryostat is an annular feed-through flange containing hermetic signal and cryogenic feed-throughs, and an octagonal plate containing ports for turbomolecular and ion pumps and gauges. In combination, they seal the rear of the vacuum space. Two large molecular sieve getter pump structures reside just behind L3, facing the focal plane.
![ [View of an OpenGL rendering of the geometry used in the LSST camera cryostat Geant4 application, along with an example trajectory of a simulated molecule. This particular molecule has undergone a far lower than average number of surface interactions and ended up contacting the front surface of a CCD. The geometrical scenario shown is known as the “baseline”, however the differences between the various scenarios are subtle enough not to dominate such a large-scale view. ]{}[]{data-label="geo2"}](fig3.eps){width="3.0in"}
Three different primary cryostat layouts have been under consideration for LSSTCam. All three preserve the same basic structure, and differ only in details. The first, or “baseline,” design features RTMs with distinct front and rear electronics boards connected by 2m long cables. Inside the RTM, the frontmost boards are interspersed with copper thermal planes and connected to the cryoplate. An additional warmer coldplate structure is provided to support and cool the rearmost boards. A modified design, “Option 2” contains the electronics on a single longer board with built-in copper thermal bars, replaces the coldplate with a simpler tubular structure, and in addition makes several other modifications. “Option 3” modifies the design even further by reconfiguring the radiation shrouds and pumping plenum geometry. To explore the design parameters, further modifications to Option 3 have been examined, for example increasing the diameter of the ion pump ports by either two or four times. These cases are designated “Option 3 + 2xI” and “Option 3 + 4xI” respectively. Figures \[geo1\] and \[geo2\] show examples of the detector geometry developed for this application as rendered in the OpenGL viewer.
A number of features within the cryostat have not yet been rendered for the application detector geometry, because they would be too cumbersome to implement and/or are not yet specified, nor are they likely have a significant effect on the bulk molecular transport issues under investigation. These features include fasteners, cables, thermal straps, fluid lines, hermetic feed-throughs, CCD package kinematic mounts, raft hold downs, and perimeter cutouts in the grid. Additionally some features have necessarily been simplified to allow for their specification with Geant4 detector construction. It is unlikely that these simplifications and omissions would have a significant effect on the bulk vacuum molecular transport issues being modeled, because they represent a small portion of the total surface area and are mostly perturbations on top of other surfaces.
For this application scoring has been implemented such that each particle interacts with surfaces until it reaches one of three fates — either contacting the front surface of the CCDs in the focal plane, contacting the surface of a getter pump located near the focal plane, or exiting the cryostat through one of two ion pump ports. These three fates, any of which ends any given molecule’s trajectory, are determined through the use of regions specified in the detector construction as discussed in §\[ao\]. As an indication of the complexity of the LSSTCam geometry, it is found that molecules can have anywhere from less than one hundred to hundreds of thousands of surface encounters before encountering one of these three terminating regions.
The probabilities for each of these outcomes for several alternate geometries under consideration are presented in Table 2; each is based on simulations of 1000 particles. The statistical uncertainties on these values can be obtained as outlined in §\[confirm\] and are $\sigma$=1%. These results show that in the first three geometries most molecules ($>$85%) eventually reach the focal plane region (i.e. touch the getter pumpus or the front surface of the CCD detectors) rather than exiting through the rear ion pump ports. Once in that region, they then can encounter either the getter pump or the front surface of the CCD detectors first. From this one concludes that the getter pumps will need to have a large enough capacity to handle this fraction of the flow. It also suggests these pump surfaces should be made larger or be better positioned. Alternately, modifications to the geometry to make molecular transport into the ion pumps more likely is warranted. The simplest modification would be to increase the diameter of the ion pump ports, and as shown in Table II this by itself greatly increases the likelihood of a molecule exiting through an ion pump port before reaching the focal plane region. The ability to readily test the effects of such modifications makes a strong case for this Geant4 modeling and simulation approach.
Discussion {#disc}
==========
![[View of an OpenGL rendering of the geometry of a Geant4 application representing a section of accelerator pipe, where regularly spaced irises form cavities, as discussed in §\[disc\]. A trajectory of a molecule created with initial position on the wall of the middle section and with random initial momentum is shown. ]{}[]{data-label="beamcavity"}](fig4.eps){width="3.0in"}
Geant4 provides a flexible and well-documented framework to simulate particle transport and interactions across complex geometries. These features, in concert with the techniques described herein, provide a unique method to model molecular transport in complex vacuum environments, where analytical calculations or other techniques are impractical. Such is the case in many real applications like the LSSTCam cryostat. As demonstrated in §\[confirm\], these general techniques reproduce results from analytical calculations and other Monte Carlo methods that are available for comparison.
The techniques we discuss are clearly not unique to a particular vacuum geometry and could be used to model molecular transport in any geometry. Indeed such modeling can also provide additional information as well. As an example, Figure \[beamcavity\] shows a rendering of the geometry of another Geant4 application we have developed, that of an electron accelerator beam pipe with regularly spaced irises that form cavities, with pumps located at either end. This application has been used to determine mean residence lengths of molecules in a given cavity section, and therefore the partial pressures within, given an outgassing rate for the interior surfaces of the pipe. To sum the path lengths within a given volume, one simply carries the net length through in an implementation of G4UserTrackInformation and adds to it in an implementation of G4UserSteppingAction, where the step length is added if the step terminates within the volume of interest.
As discussed in §\[setup\], for the LSSTCam cryostat application, we have so far implemented a simple scoring (i.e.: fraction of particles reaching a focal plane structure versus reaching a pump first). As already indicated, the efficacy of simple changes to the model such as increasing the diameter of the ion pump ports can be readily evaluated. In further studies, the effects of adding, resizing, and moving pump ports and getters will be evaluated, as well as adding large area cold activated-charcoal getter pumps which may be placed in the transport regions between the front and the rear of the cryostat, but are not yet included in the current models.. More extensive changes to the geometry and design (e.g.: plenum shape, charcoal getters, molecular flow barriers, etc.) will also be studied.
Thus far, interest has been in the steady state or “worst-case” scenario estimates. However these applications can be easily extended to include temperature-dependent vacuum sticking coefficients representing the mean duration for a molecular species to be adsorbed on a particular surface. This time dependency can be included in evaluating, for example, whether a particle will actually contribute to building monolayers of contaminants on the front surface of the CCDs within a given interval of time. Results from this application are crucial for evaluating the transport, and therefore focal plane contamination consequences, of different layouts and features under consideration. They will also provide a necessary component of the mapping from outgassing of components (that are measured in a test stand[@RSI]) to the resulting effect on the contaminant levels adsorbed on the focal plane, and ultimately to the impact on CCD performance via modeling.[@SPIE2]
We wish to thank Dennis Wright, Makoto Asai, and Daniel Brandt for their assistance and expertise in developing Geant4 applications, and Gordon Bowden for useful discussions and critique. LSST project activities are supported in part by the National Science Foundation through Governing Cooperative Agreement 0809409 managed by the Association of Universities for Research in Astronomy (AURA), and the Department of Energy under contract DE-AC02-76-SFO0515 with the SLAC National Accelerator Laboratory. Additional LSST funding comes from private donations, grants to universities, and in-kind support from LSSTC Institutional Members.
[0]{} Z. Ivezic [*et al*]{}., LSST: from Science Drivers to Reference Design and Anticipated Data Products, 2008. (arXiv:0805.2366, 2008) S. Kahn [*et al*]{}., Design and Development of the 3.2 Gigapixel Camera for the Large Synoptic Survey Telescope, [*Proc. SPIE — Ground-based and Airborne Instrumentation for Astronomy III*]{}, 2010. J. Singal, R. Schindler, C. Chang, C. Czodrowski, and P. Kim, A Multi-Chamber System for Analyzing the Outgassing, Deposition, and Associated Optical Degradation Properties of Materials in a Vacuum, [*Rev. Sci. Instrum.*]{}, [**81**]{}:025101–025108, 2010. H. Howell, J. Wehrle, and J. Jostlein, Calculation of pressure distribution in vacuum systems using a commercial finite element program, [*Proceedings of 1991 IEEE particle accelerator conference (PAC)*]{}, pp. 2295–2297, 1991. D.H. Davis, Monte Carlo Calculation of Molecular Flow Rates through a Cylindrical Elbow and Pipes of Other Shapes, [*Journal of Applied Physics*]{}, [**31**]{}:1169–1175, 1960. Ziemann V. Vacuum Tracking, [*Proceedings of 1993 IEEE particle accelerator conference (PAC)*]{}, pp. 3909–3911, 1993. R.Kersevan, & J-L.Pons, Introduction to MOLFLOW+, [*J. Vac. Sci. Technol. A*]{}, [**27**]{}:1017–1023, 2009. S. Agostinelli et al., Geant4 — A Simulation Toolkit, [*Nuclear Instruments and Methods A*]{}, [**506**]{}:250–303, 2003. J. Allison [*et al*]{}., Geant4 Developments and Applications, [*Int. J. Model. Simul. Sci. Comput.*]{}, [**1**]{}:157–178, 2010. R. Agnese [*et al*]{}., Dark Matter Search Results Using the Silicon Detectors of CDMS II, [*PRL*]{}, submitted, 2013. (arXiv:1304.4279) S. Incerti [*et al*]{}., The Geant4-DNA project, [*PRL*]{}, submitted, 2013. (arXiv:0910.5684) S. Ceyer, D. Gladstone, M. McGonigal, & M. Schulberg, Molecular Beams: Probes of the Dynamics of Reactions on Surfaces, in Physical Methods of Chemistry: Investigations of Surfaces and Interfaces, eds. B. Rossiter & R. Baetzold, New York:Wiley, pp. 383–449, 1992. C. Rettner, E. Schweizer, & C. Mullins, Desorption and trapping of argon at a 2H–W(100) surface and a test of the applicability of detailed balance to a nonequilibrium system, [*Journal of Chemical Physics*]{}, [**90**]{}:3800–3814, 1989. H. Steinruck, A. Winkler, & D. Rendulik, Angle-resolved thermal desorption spectra for CO and H2 from Ni(111), Ni(110) and polycrystalline nickel, [*Journal of Physics C*]{}, [**17**]{}:L311–L316, 1984. P. Clausing, The Flow of Highly Rarefied Gases through Tubes of Arbitrary Length, [*Ann Physik*]{}, [**12**]{}:961–972, 1932. K. Gilmore [*et al*]{}., LSST Camera Instrument Modeling, [*Proc. SPIE — Modeling, Systems Engineering, and Project Management for Astronomy V*]{}, 2012.
[^1]: All applications discussed in this work are available at http://sourceforge.net/projects/geant4vacmod/.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Limit sets of ${\mathrm{AdS}}$-quasi-Fuchsian groups of ${\mathrm{PO}}(n,2)$ are always Lipschitz submanifolds. The aim of this article is to show that they are never $\mathcal{C}^1$, except for the case of Fuchsian groups. As a byproduct we show that ${\mathrm{AdS}}$-quasi-Fuchsian groups that are not Fuchsian are Zariski dense in ${\mathrm{PO}}(n,2)$.'
author:
- 'Olivier Glorieux, Daniel Monclair'
title: ' Regularity of limit sets of ${\mathrm{AdS}}$ quasi-Fuchsian groups'
---
Introduction
============
The study of various notions of convex cocompact groups in semi-simple Lie groups has gain considerable interest the last decade, thanks to its relation with Anosov representations. A particularly nice setting is for subgroups of ${\mathrm{PO}}(p,q)$ where the quadratic form helps to construct invariant domains of dicontinuity, see [@DGK].
In a previous paper, we studied the metric properties of limit sets for such representations [@glorieux2017hausdorff] and proved a rigidity result for quasi-Fuchsian representations in ${\mathrm{PO}}(2,2)$. Recently Zimmer [@Zimmer] showed a $\mathcal{C}^2$ rigidity result for Hitchin representations in ${\mathrm{PSL}}_n({\mathbb{R}})$ (${\mathcal{C}}^\infty$ rigidity was known from the work of Potrie-Sambarino [@potrie2014]).
In this paper, we study the $\mathcal{C}^1$ regularity of such a limit set and prove a rigidity result for quasi-Fuchsian subgroups ${\mathrm{PO}}(n,2)$. They are examples of ${\mathrm{AdS}}$-convex cocompact groups, as defined by [@DGK].\
Given the standard quadratic form $q_{n,2}$ of signature $(n,2)$ on ${\mathbb{R}}^{n+2}$, we define ${\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ as the subset of ${\mathbb{R}}{\mathbb{P}}^{n+1}$ consisting of negative lines for $q_{n,2}$. Its boundary $\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ is the set of $q_{n,2}$-isotropic lines.
[@DGK] A discrete subgroup ${\Gamma}$ of $G={\mathrm{PO}}(n,2) $ is ${\mathrm{AdS}}$-convex cocompact if it acts properly discontinously and cocompactly on some properly convex closed subset ${\mathcal{C}}$ of ${\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ with nonempty interior whose ideal boundary $\partial_i {\mathcal{C}}:= \overline{{\mathcal{C}}}\setminus {\mathcal{C}}$ does not contain any nontrivial projective segment.
Any infinite convex-cocompact group contains proximal elements, ie. elements that have a unique attractive fixed point in $\partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$. For ${\Gamma}$ a discrete subgroup of ${\mathrm{PO}}(n,2)$, the *proximal limit set of ${\Gamma}$* is the closure $\Lambda_{\Gamma}\subset {\mathbb{RP}}^{n,2}$ of the set of attracting fixed points of proximal elements of ${\Gamma}$. Since ${\Gamma}$ acts properly discontinuously on a convex set ${\mathcal{C}}$, the proximal limit set coincides with the ideal boundary of ${\mathcal{C}}.$ It is shown in [@DGK] that this notion of limit set coincides with the closure of orbits in the boundary.
A discrete group of ${\mathrm{PO}}(n,2)$ is ${\mathrm{AdS}}$-quasi-Fuchsian if it is ${\mathrm{AdS}}$-convex cocompact and its proximal limit set is homeomorphic to a $n-1$ dimensional sphere.
If moreover, the group preserves a totally geodesic copy of ${\mathbb{H}}^n$, it is called ${\mathrm{AdS}}$-Fuchsian.
The limit set of an ${\mathrm{AdS}}$-Fuchsian group is a geometric sphere, hence a ${\mathcal{C}}^1$-submanifold of $\partial {\mathbb{A}\mathrm{d}\mathbb{S}}$. The principal aim of this article is to show that the converse holds:
\[theorem-main\] Let ${\Gamma}\subset {\mathrm{PO}}(n,2)$ be ${\mathrm{AdS}}$ quasi-Fuchsian. If $\Lambda_{\Gamma}$ is a $\mathcal C^1$ submanifold of $\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$, then ${\Gamma}$ is Fuchsian.
The proof is based on the following result which is interesting on its own:
\[prop - fuchsien ou zariski dense intro\] Let ${\Gamma}\subset {\mathrm{PO}}(n,2)$ be ${\mathrm{AdS}}$ quasi-Fuchsian. If ${\Gamma}$ is not ${\mathrm{AdS}}$-Fuchsian, then it is Zariski dense in ${\mathrm{PO}}(n,2)$.
Remark that this proposition and Zimmer’s result [@Zimmer Corollary 1.48] imply that the limit set is not $\mathcal{C}^2$.
Background on ${\mathrm{AdS}}$-quasi-Fuchsian groups.
======================================================
We introduce the results needed for the proofs of Theorem \[theorem-main\] and Proposition \[prop - fuchsien ou zariski dense intro\]. Most of this section follows directly from the work of [@merigot2012anosov] and [@DGK], except maybe the characterization of Fuchsian groups as subgroups of ${\mathrm{O}}(n,1)$ in Proposition \[prop-fuchsian est equivalent sous groupe de O(n,1)\].
First, let us define the anti-de Sitter space. We denote by ${\langle \cdot\, |\, \cdot \rangle_{n,2} }$
The anti-de Sitter space is defined by $${\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1} := \{ [x] \in {\mathbb{RP}}^{n+1} \, |\, \langle x | x \rangle_{n,2} < 0\}.$$ Its boundary is $$\partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1} := \{ [x] \in {\mathbb{RP}}^{n+1} \, |\, \langle x | x \rangle_{n,2} = 0\}.$$ Two points $[x],[y]\in\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ are called transverse if ${\langle x\, |\, y \rangle_{n,2} }\neq 0$.
We now give a brief review of the proximal limit set:
Given ${\gamma}\in {\mathrm{PO}}(n,2)$, we denote by $\lambda_1({\gamma})\geq\lambda_2({\gamma})\geq\cdots\geq\lambda_{n+2}({\gamma})$ the logarithms of the moduli of the eigenvalues of any of its representants in ${\mathrm{O}}(n,2)$.\
We say that ${\gamma}$ is proximal if $\lambda_1({\gamma})>\lambda_2({\gamma})$.
Remark that an element of ${\mathrm{PO}}(n,2)$ has not always a lift in ${\mathrm{SO}}(n,2)$. However since it is the quotient of ${\mathrm{O}}(n,2)$ by $\pm \operatorname{Id}$, the set of moduli of eigenvalues of a lift is well defined. If ${\gamma}\in{\mathrm{PO}}(n,2)$ is proximal, it has a unique lift $\hat{\gamma}\in{\mathrm{O}}(n,2)$ which has $e^{\lambda_1({\gamma})}$ as an eigenvalue.\
Notice that we always have $\lambda_3({\gamma})=\cdots=\lambda_n({\gamma})=0$, as well as $\lambda_1({\gamma})+\lambda_{n+2}({\gamma})=\lambda_2({\gamma})+\lambda_{n+1}({\gamma})=0$.\
If ${\gamma}\in {\mathrm{PO}}(n,2)$ is proximal, we denote by ${\gamma}_+\in {\mathbb{R}}{\mathbb{P}}^{n+1}$ its attractive fixed point, i.e. the eigendirection for the eigenvalue of modulus $e^{\lambda_1({\gamma})}$ of a lift of ${\gamma}$ to ${\mathrm{O}}(n,2)$. We also set ${\gamma}_-=({\gamma}^{-1})_+$.
Note that ${\gamma}_+$ is necessarily isotropic, i.e. ${\gamma}_+\in \partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$.
If ${\gamma}\in {\mathrm{PO}}(n,2)$ is proximal, then $\lim_{n\to+\infty}{\gamma}^n(\xi)={\gamma}_+$ for all $\xi\in\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ which is transverse to ${\gamma}_-$ (i.e. such that ${\langle \xi\, |\, {\gamma}_- \rangle_{n,2} }\neq 0$).
Recall that the proximal limit set of a discrete subgroup ${\Gamma}\subset {\mathrm{PO}}(n,2)$ is the closure $\Lambda_{\Gamma}$ in ${\mathbb{RP}}^{n+1}$ of the set of all attractive fixed points of proximal elements of ${\Gamma}$, it is therefore a subset of $\partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$.\
If additionally ${\Gamma}$ is ${\mathrm{AdS}}$-convex cocompact, then it is word-hyperbolic and the action of ${\Gamma}$ on its proximal limit set is conjugated to the action on its Gromov boundary [@DGK]. As a consequence, we have:
\[prop-action minimale sur l ensemble limite \][@ghysdelaHarpe] If ${\Gamma}\subset {\mathrm{PO}}(n,2)$ is ${\mathrm{AdS}}$-convex cocompact, the action of ${\Gamma}$ on the limit set $\Lambda_{\Gamma}$ is minimal, ie. all orbits are dense.
The group ${\mathrm{O}}(n,1)$ can be embedded in ${\mathrm{PO}}(n,2)$ by the following map: $$A {\rightarrow}\left[\begin{array}{cc}
A&0 \\
0 & 1\\
\end{array} \right].$$ We will say that an element (respectively a subgroup) of ${\mathrm{PO}}(n,2)$ is conjugate to an element (respectively to a subgroup) of ${\mathrm{O}}(n,1)$ if it has a conjugate in the image of this embedding.\
Note that if ${\gamma}\in{\mathrm{PO}}(n,2)$ is proximal, we have $\lambda_2({\gamma})= 0$ if and only if ${\gamma}$ is conjugate to an element of ${\mathrm{O}}(n,1)$.\
A subgroup of ${\mathrm{PO}}(n,2)$ which is conjugate to a cocompact lattice of ${\mathrm{O}}(n,1)$ is ${\mathrm{AdS}}$-Fuchsian, as it fixes a totally geodesic copy of ${\mathbb{H}}^{n}$ on which it acts properly discontinuously and cocompactly. These are the only ${\mathrm{AdS}}$-Fuchsian groups:
\[prop-fuchsian est equivalent sous groupe de O(n,1)\] A discrete group of ${\mathrm{PO}}(n,2)$ is ${\mathrm{AdS}}$-Fuchsian if and only if it is conjugate to a cocompact lattice of ${\mathrm{O}}(n,1)$.
Let ${\Gamma}\subset {\mathrm{PO}}(n,2)$ be an ${\mathrm{AdS}}$-Fuchsian group. Let $H$ be a totally geodesic copy of ${\mathbb{H}}^n$ in ${\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ preserved by ${\Gamma}$. Since the stabilizer $L\subset {\mathrm{PO}}(n,2)$ of $H$ is conjugate to ${\mathrm{O}}(n,1)$, we only have to show that ${\Gamma}$ is a cocompact lattice of $L$. This will be a consequence of the fact that ${\Gamma}$ acts properly discontinuously and cocompactly on $H$.\
Let ${\gamma}$ be a proximal element of ${\Gamma}$. Let $\xi \in \partial H$ be transverse to the repelling fixed point ${\gamma}_-$. The sequence ${\gamma}^n \xi $ lies in $H$ and converges to ${\gamma}^+$. Therefore, $\partial H$ contains the attracting point of ${\gamma}$, and it follows that $\Lambda_{\Gamma}\subset \partial H$. Since $\Lambda_{\Gamma}$ and $\partial H$ are homeomorphic to ${\mathbb{S}}^{n-1}$, we have $\Lambda_{\Gamma}=\partial H$.
Finally since, ${\Gamma}$ is convex-cocompact, ${\Gamma}$ acts properly discontinuously and cocompactly on the convex hull of $\Lambda_{\Gamma}$ that is $H$ (see [@DGK]).
The boundary $\partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ is naturally equipped with a conformal Lorentzian structure. It is conformally equivalent to the quotient of ${\mathbb{S}}^{n-1} \times {\mathbb{S}}^1 $ endowed with the Lorentzian conformal metric $[g_{{\mathbb{S}}^{n-1}} - d\theta^2]$ (where $g_{{\mathbb{S}}^{n-1}} $ is the round metric of curvature $1$ on ${{\mathbb{S}}^{n-1}} $, and $d\theta^2$ is the round metric on the circle of radius one) by the antipodal map $(x,\theta)\mapsto(-x,-\theta)$. See [@merigot2012anosov paragraph 2.3] for more details.\
Using the absence of segments in the limit sets of ${\mathrm{AdS}}$-quasi-Fuchsian groups we have:
\[prop-l’ensemble limite est le graph d’une fonction lip\] The limit set $\Lambda_{\Gamma}\subset\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ of an ${\mathrm{AdS}}$-quasi-Fuchsian group ${\Gamma}\subset{\mathrm{PO}}(n,2)$ is the quotient by the antipodal map of the graph of a distance-decreasing[^1] map $f \, :\, {\mathbb{S}}^{n-1} {\rightarrow}{\mathbb{S}}^{1}$ where ${\mathbb{S}}^{n-1}$ and ${\mathbb{S}}^{1}$ are endowed with the round metrics.
Barbot-Mérigot showed in [@merigot2012anosov] that the limit set of a quasi-Fuchsian group lifts to the graph of a 1-Lipschitz map. Since the limit set does not contain any non trivial segment of $\partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ the map strictly decreases the distance.
Finally we will need the following proposition, which in the Lorentzain vocabulary translates as the fact that the limit set is a Cauchy hypersurface:
\[prop-les geoedeisc de type lumiere intersect l ensemble limite\] If ${\Gamma}\subset {\mathrm{PO}}(n,2)$ is ${\mathrm{AdS}}$-quasi-Fuchsian, then every isotropic geodesic of $\partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1} $ intersects $\Lambda_{\Gamma}$ at exactly one point.
Let $f \, :\, {\mathbb{S}}^{n-1} {\rightarrow}{\mathbb{S}}^{1}$ be a distance-decreasing map such that the quotient by the antipodal map of its graph is $\Lambda_{\Gamma}$. An isotropic geodesic can be parametrized by $(c(\theta), \theta)$, where $c : \theta {\rightarrow}c(\theta) $ is a unit speed geodesic on ${\mathbb{S}}^{n-1}$. Then the proposition is equivalent to the existence and uniqueness of a fixed point for the map $f\circ c :{\mathbb{S}}^1\to{\mathbb{S}}^1$.\
It is a simple exercise to show that a distance-decreasing map of a compact metric space to itself has a unique fixed point.
The Zariski closure of ${\mathrm{AdS}}$ quasi-Fuchsian groups
=============================================================
We prove in this section the Zariski density of ${\mathrm{AdS}}$-quasi-Fuchsian subgroups of ${\mathrm{PO}}(n,2)$ which are not ${\mathrm{AdS}}$-Fuchsian. This result, which happens to be interesting in itself, will considerably simplify the proof of Theorem \[theorem-main\] when we will use Benoist’s Theorem [@benoist1997asymptotiques] about Jordan projections for discrete subgroups of semi-simple Lie groups in the last section.
\[lem - fuchsien ou irreductible\] Let ${\Gamma}\subset {\mathrm{PO}}(n,2)$ be ${\mathrm{AdS}}$ quasi-Fuchsian. If ${\Gamma}$ is reducible, then it is Fuchsian.
Assume that ${\Gamma}$ is not Fuchsian, and let $V\subset {\mathbb{R}}^{n+2}$ be a ${\Gamma}$-invariant subspace with $0<\dim(V)<n+2$.\
First, lets us show that the restriction of ${\langle \cdot\, |\, \cdot \rangle_{n,2} }$ to $V$ is non degenerate.\
Assume that it is not the case. Then ${\Gamma}$ preserves the totally isotropic space $V\cap V^\perp$. It has dimension $1$ or $2$. If $\dim(V\cap V^\perp)=1$, then ${\mathbb{P}}(V\cap V^\perp)$ is a global fixed point for the action on $\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$, which cannot exist. The case $\dim(V\cap V^\perp)=2$ is impossible because it also implies the existence of a global fixed point on $\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ (the intersection of the null geodesic ${\mathbb{P}}(V\cap V^\perp)$ of $\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ with $\Lambda_{\Gamma}$).\
We can now assume that the restriction of ${\langle \cdot\, |\, \cdot \rangle_{n,2} }$ to $V$ is non degenerate. It can have signature $(k,2)$, $(k,1)$ or $(k,0)$ (where $k\geq 0$ is the number of positive signs).\
In the first case, ${\Gamma}$ acts on some totally geodesic copy $X$ of ${\mathbb{A}\mathrm{d}\mathbb{S}}^{k+1}$ (with $k<n$) in ${\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$. Then $\partial X\cap \Lambda_{\Gamma}$ is a non empty closed invariant subset of $\Lambda_{\Gamma}$, hence $\Lambda_{\Gamma}\subset \partial X$ and $C(\Lambda_{\Gamma})\subset X$. Since $C(\Lambda_{\Gamma})$ has non empty interior in ${\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ (Lemma 3.13 in [@merigot2012anosov]), we see that $X={\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$, i.e. $V={\mathbb{R}}^{n+2}$, which is absurd.\
Now assume that $V$ has Lorentzian signature $(k,1)$. Then ${\Gamma}$ preserves $X={\mathbb{P}}(V)\cap {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ which is a totally geodesic copy of ${\mathbb{H}}^k$. It also acts on $X'={\mathbb{P}}(V^\perp)\cap {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ which is a totally geodesic copy of ${\mathbb{H}}^{k'}$ (with $k+k'=n$). Considering a proximal element ${\gamma}\in {\Gamma}$, there is a point in $\partial X\cup \partial X'$ which is transverse to the repelling fixed point ${\gamma}_-$ of ${\gamma}$ (otherwise ${\gamma}_-$ would be in $V\cap V^\perp$). This implies that ${\gamma}_+\in \partial X\cup\partial X'$, hence $\Lambda_{\Gamma}\cap \partial X\neq \emptyset$ or $\Lambda_{\Gamma}\cap \partial X'\neq \emptyset$. The action of ${\Gamma}$ on $\Lambda_{\Gamma}$ being minimal, we find that $\Lambda_{\Gamma}\subset \partial X$ or $\Lambda_{\Gamma}\subset \partial X'$. This is impossible because $\Lambda_{\Gamma}$ is homeomorphic to ${\mathbb{S}}^{n-1}$ and $\partial X$ (resp. $\partial X'$) is homeomorphic to ${\mathbb{S}}^{k-1}$ (resp. ${\mathbb{S}}^{k'-1}$).\
Finally, if $V$ is positive definite, then $V^\perp$ has signature $(n-k,2)$, this case has already been ruled out.
\[coro - composante neutre irreductible\] If ${\Gamma}\subset {\mathrm{PO}}(n,2)$ is ${\mathrm{AdS}}$-quasi-Fuchsian but not ${\mathrm{AdS}}$-Fuchsian, then the identity component of the Zariski closure of ${\Gamma}$ acts irreducibly on ${\mathbb{R}}^{n+2}$.
Let ${\Gamma}_\circ\subset G$ be a finite index subgroup. Since $\Lambda_{{\Gamma}_\circ}=\Lambda_{\Gamma}$, it cannot be Fuchsian, so it acts irreducibly on ${\mathbb{R}}^{n+2}$ by Lemma \[lem - fuchsien ou irreductible\].
\[prop - fuchsien ou zariski dense\] Let ${\Gamma}\subset {\mathrm{PO}}(n,2)$ be ${\mathrm{AdS}}$-quasi-Fuchsian. If ${\Gamma}$ is not ${\mathrm{AdS}}$-Fuchsian, then it is Zariski dense in ${\mathrm{PO}}(n,2)$.
Let $G\subset {\mathrm{SO}}_0(n,2)$ be the pre-image by the quotient map ${\mathrm{SO}}_0(n,2)\to{\mathrm{PO}}(n,2)$ of the identity component of the Zariski closure of ${\Gamma}$, and assume that ${\Gamma}$ is not Fuchsian.\
By Corollary \[coro - composante neutre irreductible\], we know that $G$ acts irreducibly on ${\mathbb{R}}^{n+2}$.\
According to [@discala] the only connected irreducible subgroups of ${\mathrm{SO}}(n,2)$ other than ${\mathrm{SO}}_0(n,2)$ are $\mathrm{U}(\frac{n}{2},1)$, $\mathrm{SU}(\frac{n}{2},1)$, ${\mathbb{S}}^1.{\mathrm{SO}}_0(\frac{n}{2},1)$ (when $n$ is even) and ${\mathrm{SO}}_0(2,1)$ (when $n=3$).\
The first three cases are subgroups of $\mathrm{U}(\frac{n}{2},1)$, which only contains elements ${\gamma}\in {\mathrm{SO}}(n,2)$ satisfying $\lambda_1({\gamma})=\lambda_2({\gamma})$ so $G$ cannot be one of them (otherwise ${\Gamma}$ would not contain any proximal element and $\Lambda_{\Gamma}=\emptyset$).\
The irreducible copy of ${\mathrm{SO}}_0(2,1)$ in ${\mathrm{SO}}(3,2)$ can also be ruled out because a quasi-Fuchsian subgroup of ${\mathrm{PO}}(3,2)$ has cohomological dimension $3$, so it cannot be isomorphic to a discrete subgroup of ${\mathrm{SO}}_0(2,1)\approx {\mathrm{PSL}}(2,{\mathbb{R}})$.\
The only possibility left is that ${\Gamma}$ is Zariski dense in ${\mathrm{PO}}(n,2)$.
Non differentiability of limit sets
===================================
We finally prove the main result, Theorem \[theorem-main\]. The proof goes as follows: first, we prove that the tangent spaces of the limit set are space like (i.e. positive definite for the natural Lorentzian conformal structure on $\partial {\mathrm{AdS}}^{n+1}$). Then by an algebraic argument, this shows that all proximal elements of ${\Gamma}$ are conjugate (by an *a priori* different element of ${\mathrm{PO}}(n,2)$) to an element of ${\mathrm{O}}(n,1)$. Finally, using a famous theorem of Benoist, this implies that ${\Gamma}$ is not Zariski-dense, and therefore by Proposition \[prop - fuchsien ou zariski dense intro\] that the group is Fuchsian.
Spacelike points
----------------
\[lem - point spacelike\] If ${\Gamma}\subset {\mathrm{PO}}(n,2)$ is ${\mathrm{AdS}}$ quasi-Fuchsian and $\Lambda_{\Gamma}$ is a $\mathcal C^1$ submanifold of $\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$, then there is $\xi\in \Lambda_{\Gamma}$ such that $T_\xi\Lambda_{\Gamma}$ is spacelike.
Let $f \, :\, {\mathbb{S}}^{n-1} {\rightarrow}{\mathbb{S}}^{1}$ be a distance-decreasing map such that the quotient by the antipodal map of its graph is $\Lambda_{\Gamma}$.\
Knowing that the graph of $f$ is a ${\mathcal{C}}^1$-submanifold, we first want to show that $f$ is ${\mathcal{C}}^1$. Using the Implicit Function Theorem, it is enough to know that the tangent space of the graph projects non trivially to the tangent space of ${\mathbb{S}}^{n-1}$. This is true because $\Lambda_{\Gamma}$ is acausal.\
Since $f$ satisfies $d(f(x),f(y))<d(x,y)$ for $x\neq y$ [@merigot2012anosov], it cannot be onto, so it can be seen as a function $f:{\mathbb{S}}^{n-1}\to {\mathbb{R}}$. At a point $x\in{\mathbb{S}}^{n-1}$ where it reaches its maximum, it satisfies $df_x=0$, so the tangent space to $\Lambda_{\Gamma}$ at $(x,f(x))$ is $T_x{\mathbb{S}}^{n-1}\times \{0\}$, which is spacelike.
\[coro - partout spacelike\]If ${\Gamma}\subset {\mathrm{PO}}(n,2)$ is ${\mathrm{AdS}}$ quasi-Fuchsian and $\Lambda_{\Gamma}$ is a $\mathcal C^1$ submanifold of $\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$, then for all $\xi\in \Lambda_{\Gamma}$, the tangent space $T_\xi\Lambda_{\Gamma}$ is spacelike.
Let $E=\{\xi\in\Lambda_{\Gamma}: T_\xi\Lambda_{\Gamma}\textrm{ is spacelike} \}$. Then $E$ is open and ${\Gamma}$-invariant. Since the action of ${\Gamma}$ on $\Lambda_{\Gamma}$ is conjugate to the action on its Gromov boundary, it is minimal (i.e. all orbits are dense). It follows that $E$ is either empty or equal to $\Lambda_{\Gamma}$ and by Lemma \[lem - point spacelike\], it is not empty.
#### Remark:
Lemma \[lem - point spacelike\] fails in general in higher rank pseudo-Riemannian symmetric spaces, i.e. for ${\mathbb{H}}^{p,q}$-quasi-Fuchsian groups. Indeed, Hitchin representations in $PO(3,2)$ provide ${\mathbb{H}}^{2,2}$-quasi-Fuchsian groups which are not ${\mathbb{H}}^{2,2}$-Fuchsian, yet have a ${\mathcal{C}}^1$ limit set (which is isotropic for the natural Lorentzian conformal structure on $\partial{\mathbb{H}}^{2,2}$).
Fixed points and Benoist’s asymptotic cone
------------------------------------------
\[lem - conjugaison elements\] Let ${\Gamma}\subset {\mathrm{PO}}(n,2)$ be ${\mathrm{AdS}}$-quasi-Fuchsian. If the limit set $\Lambda_{\Gamma}\subset \partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ is a ${\mathcal{C}}^1$ submanifold, then every proximal element ${\gamma}\in{\Gamma}$ is conjugate in ${\mathrm{PO}}(n,2)$ to an element of $\mathrm O(n,1)$.
Let ${\gamma}\in{\Gamma}$ be proximal, and let $\hat{\gamma}\in {\mathrm{O}}(n,2)$ be the lift with eigenvalue $e^{\lambda_1({\gamma})}$. Let ${\gamma}_+\in\Lambda_{\Gamma}$ be the attractive fixed point. Then the differential of ${\gamma}$ acting on $\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ at ${\gamma}_+$ preserves $T_{{\gamma}_+}\Lambda_{\Gamma}$. It also preserves $(T_{{\gamma}_+}\Lambda_{\Gamma})^\perp$, which is a timelike line because of Corollary \[coro - partout spacelike\].\
Lifting everything to ${\mathbb{R}}^{n+2}$ and using the identification of $T_{{\gamma}_+}\partial{\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ with ${\gamma}_+^\perp/{\gamma}_+$, we see that $\hat{\gamma}$ preserves a two-dimensional plane $V\subset {\gamma}_+^\perp$ which contains ${\gamma}_+$ and a negative direction. Let $(u,v)$ be a basis of $V$, where $u\in{\gamma}_+$ and ${\langle v\, |\, v \rangle_{n,2} }=-1$.\
By writing $\hat\gamma v = au+bv$, we find that $b^2=-{\langle \hat\gamma v\, |\, \hat\gamma v \rangle_{n,2} }=-{\langle v\, |\, v \rangle_{n,2} }=1$. So the matrix of the restriction of $\hat{\gamma}$ to $P$ in the basis $(u,v)$ has the form $$\begin{pmatrix}
e^{\lambda_1({\gamma})} & a\\
0 & \pm 1
\end{pmatrix}$$ It has $\pm 1$ as an eigenvalue, and the eigendirection is in $V$ but is not ${\gamma}_+$ (because $\lambda_1({\gamma})>0$), so it is negative for ${\langle \cdot\, |\, \cdot \rangle_{n,2} }$. This eigendirection is a point of ${\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ fixed by ${\gamma}$.
Let ${\Gamma}\subset {\mathrm{PO}}(n,2)$ be ${\mathrm{AdS}}$ quasi-Fuchsian. If the limit set $\Lambda_{\Gamma}\subset \partial {\mathbb{A}\mathrm{d}\mathbb{S}}^{n+1}$ is a ${\mathcal{C}}^1$ submanifold, then ${\Gamma}$ is Fuchsian.
By Lemma \[lem - conjugaison elements\], the Jordan projections of proximal elements of ${\Gamma}$ all lie in a half line in a Weyl chamber $\mathfrak a^+$ of ${\mathrm{PO}}(n,2)$, therefore its asymptotic cone has empty interior in $\mathfrak a^+$. Benoist’s Theorem [@benoist1997asymptotiques] implies ${\Gamma}$ is not Zariski dense. Proposition \[prop - fuchsien ou zariski dense intro\] implies that ${\Gamma}$ is Fuchsian.
[Glo15b]{}
Thierry Barbot. Deformations of fuchsian AdS representations are quasi-fuchsian. , 101(1):1–46, 2015.
Yves Benoist. Propriétés asymptotiques des groupes linéaires. 7 (1997), no. 1, 1–47.
Thierry Barbot, Fran[ç]{}ois B[é]{}guin, and Abdelghani Zeghib. Constant mean curvature foliations of globally hyperbolic spacetimes locally modelled on ads 3. , 126(1):71–129, 2007.
Thierry Barbot and Quentin M[é]{}rigot Anosov AdS representations are quasi-fuchsian. , 6(3):441–483, 2012.
Jeffrey Dancinger, Fanny Kassel, Francois Guéritaud.
Jeffrey Dancinger, Fanny Kassel, Francois Guéritaud. 192 (2018), p. 87-126.
Antonio J. Di Scala and Thomas Leistner. Charles Frances. Lorentzian Kleinian groups.
Etienne Ghys, Pierre de la Harpe. Sur les groupes Hyperboliques d’après Mikhael Gromov ,Springer, 1990 Olivier Glorieux, Daniel Monclair Critical exponent and Hausdorff dimension in pseudo-Riemannian hyperbolic geometry .
Fanny Kassel and Toshiyuki Kobayashi. Poincaré series for non-riemannian locally symmetric spaces. In [*Advances in Mathematics*]{}, 287, p.123-236, 2016 Fran[ç]{}ois Labourie. Anosov flows, surface groups and curves in projective space. , 165(1):51–114, 2006.
Geoffrey Mess. Lorentz spacetimes of constant curvature. , 126(1):3–45, 2007. Raphael Potrie and Andres Sambarino. Eigenvalues and Entropy of a Hitchin representation , 209(3):885–925, 2017
Andrew Zimmer. Projective Anosov representations, convex cocompact actions, and rigidity.
\
<span style="font-variant:small-caps;"></span>\
*E-mail address:* `[email protected]` \
<span style="font-variant:small-caps;">Université du Luxembourg, Campus Kirchberg, 6, rue Richard Coudenhove-Kalergi, L-1359 Luxembourg</span>\
*E-mail address:* `[email protected]` \
<span style="font-variant:small-caps;">Institut de Mathématique d’Orsay, Bâtiment 307, Université Paris-Sud, F-91405 Orsay Cedex, France</span>\
[^1]: That is $\forall x\neq y, \, d(f(x),f(y)) < d(x,y)$ .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we examine a variant of the voter model on a dynamically changing network where agents have the option of changing their friends rather than changing their opinions. We analyse, in the context of dense random graphs, two models considered in Durrett et. al. [@Durrett12]. When an edge with two agents holding different opinion is updated, with probability $\frac{\beta}{n}$, one agent performs a voter model step and changes its opinion to copy the other, and with probability $1-\frac{\beta}{n}$, the edge between them is broken and reconnected to a new agent chosen randomly from (i) the whole network (rewire-to-random model) or, (ii) the agents having the same opinion (rewire-to-same model). We rigorously establish in both the models, the time for this dynamics to terminate exhibits a phase transition in the model parameter $\beta$. For $\beta$ sufficiently small, with high probability the network rapidly splits into two disconnected communities with opposing opinions, whereas for $\beta$ large enough the dynamics runs for longer and the density of opinion changes significantly before the process stops. In the rewire-to-random model, we show that a positive fraction of both opinions survive with high probability.'
author:
- 'Riddhipratim Basu[^1]'
- 'Allan Sly[^2]'
bibliography:
- 'evm.bib'
title: Evolving Voter Model on Dense Random Graphs
---
Introduction {#s:intro}
============
In recent years, a significant research effort in various fields, including biology, ecology, economics, sociolgy among others, has been concentrated on studying and modelling behaviour of large complex networks with many interacting agents. Different dynamics on large networks has been studied focussing on the structural impact of these dynamics on different models of networks. Some of the problems which received attention are consensus of opinion and polarisation, spread of epidemics, information cascades etc. (see [@N10; @Dur06]). In many real world networks the evolution of the links in the network depend upon the states of the connecting agents and vice versa. The general class of network models that model this dependence are called adaptive or coevolutionary networks (see [@GB08; @SP00]). As in the case of static networks, the problems of spread of information and epidemic, evolution of opinion and polarization into communities have been studied numerically and also using a variety of rigorous and partly non-rigorous methods ([@Volz07; @Volz09; @Gil06; @KB08; @Henry11], see also [@Durrett12; @Vas13] and references therein for more background). The problems we consider in this paper belong to this general class.
The voter model has classically been studied in the probability literature as an interacting particle system mainly on lattices [@HL1975; @Lig85]. More recently voter models have been studied in the context of general networks as a model for spread of opinion [@PRL05; @PRE05]. In the classical voter model on a fixed graph, each vertex has one of the two prevalent opinions, neighbours interact at some fixed rate, and one of the neighbours adopts the opinion of the other after the interaction. A simplified model of coevolution of network and opinion was introduced and studied using non-rigorous methods by Holme and Newman [@HN06] where they try to model the property that an agent is less likely to interact (remain connected with) another agent if there opinions do not match. Their model is similar to the classical voter model (but with number of opinions proportional to the size of the network) but with the added feature that, whenever there is an interaction between two vertices (agents) with different opinion, with probability $\alpha\in (0,1)$, one of the vertices breaks the link and connects to a different vertex of the same opinion, i.e., the network connections evolve with time as well. Using finite size scaling, Holme and Newman conjectured a phase transition in $\alpha$, where in the supercritical phase all the opinions will have small number of followers, but in the subcritical phase a giant community holding the same opinion will emerge. This model and its further extensions were investigated in [@PRE08KH; @PRL08].
Durrett et al. [@Durrett12], studied two variants of this model using numerical simulations and formulated some conjectures about the behaviour of the model. They use certain non-rigorous and numerical methods to formulate some conjectures about the asymptotic behaviour of the models on sparse random graphs as network size becomes large. They take the initial network to be a sparse Erdős-Rényi graph $G$ with average degree $>1$ and the two initial opinions distributed as product measure. In each step a uniformly chosen disagreeing edge is selected and a voter model step is performed with probability $1-\alpha$ and a rewiring step with probability $\alpha$. They consider two variants of reconnecting edge (i) rewire-to-random where the edge is connected to a randomly chosen neighbour and (ii) rewire-to-same where the edge is connected to a random neighbour of the same opinion. Based on numerical evidence and heuristics Durrett et. al. conjecture in [@Durrett12] that
1. Supercritical phase: In both variants, there exist $\alpha_c(u)\in (0,1)$ independent of initial density of $1$’s such that for $\alpha >\alpha_c$, the process reaches an absorbing state in time $O(n)$ and the final fraction $\rho$ of minority opinion is $\approx u$.
2. Subcritical phase: In the [**rewire-to-random**]{} model, for $\alpha<\alpha_c(u)$, the time to absorption is order $n^2$ on average and at the absorption time, the density of the minority opinion $\rho$ is bounded away from $0$ and independent of $u$. In contrast, for the [**rewire-to-same**]{} model, $\rho\approx 0$, so one of the opinions takes over almost the whole network at the time of absorption.
We establish analogous results in the dense case but without establishing a sharp transition. We also prove that both opinions survive in [**rewire-to-random**]{} model, however we cannot prove the contrasting result for the [**rewire-to-same**]{} model as is conjectured in [@Durrett12]. Durrett et al. also formulates conjectures about finer behaviours of the evolving voter model along the path to absorption (see Conjecture 1 and Conjecture 2 in [@Durrett12]). Further extensions of these models with different social dynamics and multiple possible opinions were considered in [@PRE13; @RG13PRE; @MM13; @Dur13PRE].
Main Results
------------
In this paper, we study the dense version of the model from [@Durrett12] where the initial graph is $G(n,1/2)$. It is easy to see that to obtain in non-trivial transition, we must renormalize the opinion update rate to $1-\alpha= \beta/n$ (this is due to the average degree being linear in $n$). Let $\tau$ denote the time to reach an absorbing state, i.e., $\tau$ is the first time when there are no disagreeing edges in the graph (for the rewire-to-same model absorbing states are slightly different, see § \[s:def\] below). For $\frac{1}{2}>\varepsilon>0$, let $\tau_{*}(\varepsilon)$ be the first time that the fraction of the minority opinion reaches $\varepsilon$, i.e., $\tau_*(\varepsilon)=\min\{t:N_*(t)\leq \varepsilon n\}$, where $N_*(t)$ is the number of vertices holding the minority opinion at time $t$. Now we state our main theorems.
\[t:rewiretorandom\] Let $\frac{1}{2}>\varepsilon '>0$ be given. For both variants of the model, there exist $0<\beta_0<\beta_{*}(\varepsilon')<\infty$ such that each of the following hold.
1. For all $\beta<\beta_0$ and any $\eta>0$, we have $\{\tau<10n^2, N_*(\tau)\geq \frac{1}{2}-\eta\}$ holds with high probability.
2. For all $\beta>\beta_*(\varepsilon')$, we have that $\tau_{*}(\varepsilon')\leq \tau$ with high probability and $$\lim_{c \downarrow 0} \liminf_n {\mathbb{P}}[\tau >cn^3] = 1.$$
The following theorem addresses the issue of fraction of minority opinion in the process at the absorption time for the rewire-to-random model.
\[t:rtoresplit\] Let $\beta>0$ be fixed. For the rewire-to-random model there exists $\varepsilon_{*}=\varepsilon_{*}(\beta)>0$ such that $\tau<\tau_*(\varepsilon_{*})$ with high probability.
Formal Model Definitions {#s:def}
------------------------
Now we describe formally the models we consider in this paper. Let $n$ be a fixed positive integer. Let $V$ be a fixed set with $|V|=n$. Let $V^{(2)}$ denote the set of all unordered pairs in $V$, we shall call elements of $V^{(2)}$ as *bonds*. Also let $\mathcal{E}$ be a fixed set. We consider discrete time Markov chains $\{G(t)\}_{t\geq 0}$ taking values in $$\left\{\{0,1\}^{V}, \left(V^{(2)}\right)^{\mathcal{E}}\right\},$$ i.e., for each $t$, $G(t)$ is a multi-graph on the vertex set $V$ with labelled edges coming from the set $\mathcal{E}$ (each edge in $\mathcal{E}$ is placed at one of the bonds); each vertex has one of the two opinions $0$ and $1$. The following notations will be used throughout this paper.
The opinion of a vertex $v$ at time $t$ shall be denoted by $v(t)$. The vector of opinions of vertices in $G(t)$ shall be denoted by $V(t)$. We shall denote by $N_0(t)$ and $N_1(t)$ the number of $0$s and $1$s in $V(t)$ respectively. Let $N_*(t)=\min\{N_0(t),N_1(t)\}$. For $v\in V$, let $C_v(t)$ denote the set of all vertices in $V$ which have the same opinion as $v$ in $G(t)$. By $\tilde{G}(t)=(V,E(t))$, we denote the underlying graph of $G(t)$. Often, when there is no scope of confusion we shall use $G(t)$ instead of $\tilde{G}(t)$ to denote the same. Notice that we are allowing multi-edges but not self loops, i.e., at a time $t$, a bond $(u,v)\in V^{(2)}$ may be connected by more than one edge, but there are no edges connecting $v$ to itself. For an edge connecting the bond $(u,v)$ in $G(t)$ we shall call it [**disagreeing**]{} if $u(t)\neq v(t)$ and [**agreeing**]{} otherwise.
[**Initial condition:**]{} To simplify matters we only consider the following initial condition. We take $\tilde{G}(0)$ is distributed as $G(n,\frac{1}{2})$, i.e., each bond contains $0$ edge with probability $\frac{1}{2}$ and $1$ edge with probability $\frac{1}{2}$ independent of each other. Also, let $\{v(0)\}_{v\in V}$ be $i.i.d.$ $\mbox{Ber}(\frac{1}{2})$. Also denote the set of the labelled edges $\mathcal{E}=\{e_0,e_1,\ldots, e_N\}$.
[**Transition Probabilities:**]{} We describe the one step evolution of the two variants of the Markov chains as follows.
Let $G(t)$ be the state of the chain at time $t$. Let $Z(t)=\mbox{Ber}(\frac{\beta}{n})$ be independent of $G(t)$. If $Z(t)=1$ we obtain $G(t+1)$ from $G(t)$ by taking a *voter model step*, and if $Z(t)=0$ then we obtain $G(t+1)$ from $G(t)$ by taking a *rewiring step*. We call $\beta>0$ the *relabelling rate*, this is a parameter of the model.
Let $\mathcal{E}^{\times}(t)\subseteq \mathcal{E}$ denote the set of edges that are disagreeing in $G(t)$. Choose one edge $e$ from $\mathcal{E}^{\times}(t)$ uniformly at random. Let $(u,v)$ be the bond which this edge connects in $G(t)$. Choose one vertex randomly among $u$ and $v$, say $u$. The vertex $u$ as above will be called the [**root**]{} of a rewiring move.
[**The voter model step (relabelling step):**]{} If $Z(t)=1$, then $u$ adopts the opinion of $v$, i.e., $G(t+1)$ is obtained from $G(t)$ by taking $\tilde{G}(t+1)=\tilde{G(t)}$, $v'(t+1)=v'(t)$ for all $v'\in V\setminus \{u\}$ and $u(t+1)=v(t+1)$.
[**The rewiring step:**]{} If $Z(t)=0$, the two chains we consider evolve differently.
[**Rewire-to-random model:**]{} In this model, we choose a vertex $v'$ uniformly at random from $V \setminus \{u\}$. We obtain $G(t+1)$ from $G(t)$ by taking $V(t+1)=V(t)$ and $E(t+1)$ is obtained from $E(t)$ by removing the edge $e$ from the bond $(u,v)$ and adding it to the bond $(u,v')$.
[**Rewire-to-same model:**]{} In this model, we choose a vertex $v'$ uniformly at random from $C_u(t)\setminus \{u\}$. We obtain $G(t+1)$ from $G(t)$ by taking $V(t+1)=V(t)$ and $E(t+1)$ is obtained from $E(t)$ by removing the edge $e$ from the bond $(u,v)$ and adding it to the bond $(u,v')$.
We make the following basic observation characterising the absorbing states.
\[o:absorbing\] Notice that, on finite networks both the chains are absorbing. For the [*rewire-to-random*]{} model the only absorbing states are those which corresponds to the graph having no disagreeing edges, i.e., either one opinion has taken over all the vertices, or the graph is split into disconnected communities, where all the vertices in a community has the same opinion. For the [*rewire-to-same*]{} model the absorbing states are those that either have no disagreeing edges, or those in which one of the opinions are held by only one vertex.
Notice that the number of edges is conserved in each step of the chain, i.e., we have that $|E(t)|=|\mathcal{E}|$ for all $t$. Also observe that even though we have labelled edges, this fact does not affect the behaviour of the model at all. The edges are labelled, simply because it will be convenient while constructing a coupling of this chain with another process which we shall use.
One of the main questions we are interested in for both the models described above is the asymptotics of absorption time as a function of $\beta$ as $n\rightarrow \infty$, and whether it exhibits a phase transition in $\beta$ or not. Notice that if $\beta=0$, then we have only rewiring moves the absorption time is $\Theta(n^2)$, i.e., the graph splits immediately into two communities having different opinions. We investigate whether similar phenomenon occurs if $\beta>0$ is sufficiently small. In the other extreme, if the rewiring moves are much rarer compared to the relabelling moves (i.e., $\beta>>1$) one might expect the model to behave similarly as the voter model on a static graph, where the minority opinion density will become very small before reaching an absorbing state, and the absorption time will be at least $\Theta(n^3)$. This is established in Theorem \[t:rewiretorandom\]. A related quantity of interest is the fraction of the minority opinion vertices when the process reaches an absorbing state. For $\beta$ sufficiently large does the minority opinion persist with a positive fraction? Theorem \[t:rtoresplit\] provide the answers for the rewire-to-random model.
Outline of the proof
--------------------
We prove parts $(i)$ and $(ii)$ of Theorem \[t:rewiretorandom\] separately. The arguments are similar for the rewire-to-random model and the rewire-to-same model. We provide details only for the rewire-to-random model while pointing out the differences for the rewire-to-same model. To prove part $(i)$, we essentially show that before the density of either opinion changes, a rewiring move is likely to decrease the number of disagreeing edges. Using a martingale argument we show that, by time $\Theta(n^2)$ (by which time the opinion densities cannot change significantly), the number of disagreeing edges decay to 0.
Most of the work goes into proving part $(ii)$ of Theorem \[t:rewiretorandom\]. We show that for $\beta=\beta(\varepsilon')$ sufficiently large, the graph $G(t)$ remains close enough to an Erdős-Rényi graph, in a sense to be made precise, as long as the minority opinion density does not drop below $\varepsilon'$. To this end, we define a number of stopping times detecting when $G(t)$ deviates too much from an Erdős-Rényi graph for the first time with respect to certain different properties, and roughly show that all those stopping times are with high probability at least as large as than $\tau_{*}(\varepsilon')$. The properties we need to consider are vertex degrees, the Cheeger constant and edge-multiplicities.
Corresponding to each of the properties we consider, we define two stopping times, one with a stronger threshold and the other with a weaker threshold. We show that provided none of the weaker thresholds have been reached, the opinions quickly mix to an approximate product measure which guarantees that the properties of interest are sufficiently mean reverting for our purposes.
For the proof of Theorem \[t:rtoresplit\], we show that for a fixed $\beta$, there exist sufficiently small but positive $\varepsilon_{*}$, such that once the minority opinion reaches $\varepsilon_{*}$, the typical vertices having minority opinions start losing disagreeing edges at a higher rate than it gains them, and eventually all the disagreeing edges are lost before the minority opinion density can change substantially.
[**Organisation of the paper:**]{} The rest of the paper is organised as follows. In § \[s:smallbeta\], we prove part $(i)$ of Theorem \[t:rewiretorandom\] for the rewire-to-random model. Most of the work in this paper goes towards the proof of part $(ii)$ of Theorem \[t:rewiretorandom\] for the rewire-to-random model, which spans § \[s:stoptime\], § \[s:weakbound\] and § \[s:strongbound\]. In § \[s:stoptime\] we define all the stopping times that we need to use. In § \[s:weakbound\] we show that, if by time $t$ the graph does not reach any of the stronger stopping times, then the graph does not reach any of the weaker stopping times by time $t+t'$ with high probability, where $t'$ is of order $n^2$. That the the graph is also unlikely to reach any of the strong stopping times by time $t+t'$ as long as the minority opinion density does not become too small is shown in § \[s:strongbound\]. Together these complete the proof of Theorem \[t:rewiretorandom\], part $(ii)$. Theorem \[t:rtoresplit\] is proved in § \[s:esplit\]. In § \[s:samemod\], we point out the significant adaptations to the argument that are necessary to prove Theorem \[t:rewiretorandom\] for the rewire-to-same model. We finish with the discussion of some open problems in § \[s:conclusion\].
Fast polarization for small $\beta$ {#s:smallbeta}
===================================
In this section we prove part $(i)$ of Theorem \[t:rewiretorandom\] for the *rewire-to-random model* with relabelling rate $\beta$. First we make the following definitions. Let $D_{\max}(t)$ denote the maximum degree of a vertex in $G(t)$. The degree of a vertex is defined as the number of edges incident to it, and not the number of bonds containing edges. Also, let $X_t=|\mathcal{E}^{\times}(t)|$. Consider the following stopping times. Let $\tau_1=\min\{t: D_{\max}(t)\geq 8n\}$ and let $\tau_{2}=\tau_*(\frac{1}{3})$, i.e., $\tau_{2}=\min\{t: N_{*}(t)=\min\{N_0(t),N_1(t)\}\leq \frac{n}{3}\}$. Define $\tau_0=\tau \wedge \tau_1 \wedge \tau_2$. We have the following lemmas.
\[l:betasmallmg\] There exists $\beta_0>0$, such that for all $\beta<\beta_0$, we have $\tau_0\leq 6n^2$ with high probability.
Let ${\mathcal{F}}_{t}$ denote the filtration generated by the process up to time $t$. Observe that whenever an edge is rewired, $X_t$ either remains the same or decreases by $1$. Conditional on ${\mathcal{F}}_t$, the chance that $X_t$ is decreased by a rewiring move is at least $\frac{N_{*}(t)-1}{n-1}$. Also notice that a relabelling move, i.e., a voter model step can increase $X_t$ by at most $D_{\max}(t)$. Hence we have for $\lambda>0$, $$\label{e:supmart}
{\mathbb{E}}\left(e^{\frac{\lambda X_{t+1}}{n}} \mid {\mathcal{F}}_t\right) \leq e^{\frac{\lambda X_{t}}{n}}\left((1-\frac{\beta}{n})\left(1+\frac{N_{*}(t)-1}{n-1}(e^{-\frac{\lambda}{n}}-1)\right)+\frac{\beta}{n} e^{\frac{\lambda D_{\max}(t)}{n}}\right).$$ Now for large $n$ on the event that $\{t<\tau_0\}$ we have $\frac{N_*(t)-1}{n-1}\geq \frac{1}{4}$. Taking $\lambda>0$ sufficiently small so that $e^{8\lambda}\leq 1+9\lambda$ and $e^{-\frac{\lambda}{n}}-1\leq -\frac{\lambda}{2n}$. Then on $\{t< \tau_0\}$, we have for $\beta<\frac{1}{400}$, $$\begin{aligned}
\label{e:supmart2}
{\mathbb{E}}\left(e^{\frac{\lambda X_{t+1}}{n}} \mid {\mathcal{F}}_t\right) &\leq
& e^{\frac{\lambda X_{t}}{n}}\left((1-\frac{\beta}{n})\left(1-\frac{1}{4}(e^{-\frac{\lambda}{n}}-1)\right)+\frac{\beta}{n} e^{8\lambda}\right)\nonumber\\
&\leq & e^{\frac{\lambda X_{t}}{n}}\left((1-\frac{\beta}{n})(1-\frac{\lambda}{8n}) +\frac{\beta}{n}(1+9\lambda)\right)\nonumber\\
&\leq & e^{\frac{\lambda X_{t}}{n}}(1-\frac{\lambda}{10n}).\end{aligned}$$ It follows from above that $$\label{e:markov}
{\mathbb{P}}[\tau_0 > t\mid {\mathcal{F}}_0]\leq {\mathbb{E}}\left[e^{\frac{\lambda X_t}{n}}1_{\{\tau_0 >t\}}\mid {\mathcal{F}}_{0}\right] \leq
e^{\frac{\lambda X_0}{n}}e^{-\frac{\lambda t}{10n}}\leq e^{\frac{\lambda n}{2}}e^{-\frac{\lambda t}{10 n}}$$ since $X_0\leq \frac{n^2}{2}$. Hence we have $$\label{e:tau}
{\mathbb{P}}[\tau_0 > 6n^2]\leq e^{-\frac{\lambda n}{10}}.$$
\[l:betasmallincoming\] We have $\tau_1>6n^2$ with high probability.
It is easy to see that the *rewire-to-random* dynamics can be implemented in the following way. Without loss of generality let $V=[n]$. And let $\mathbb{W}=\{W_{i}\}_{i\geq 1}$ be a sequence of i.i.d. random variables with each $W_i$ being uniformly distributed over $\{1,2,\ldots, n\}$. Let us define $L_0=0$, and we define $L_i$ recursively as follows. Let, $v_i$ be the *root* of the $i$-th rewiring move. Then we define $L_{i}=\min\{j>L_{i-1}:W_{j}\neq v_i\}$. Then in the $i$-th rewiring move, we add an edge to the bond $(v_i,W_{L_i})$. The algorithm can be described as follows. For each rewiring move, start inspecting the list $\mathbb{W}$ from the first previously uninspected element upto the first time you find a vertex which is not equal to the root of the current rewiring. Rewire the edge to this vertex. Clearly, in this way the chosen vertex is uniform among all vertices other than $v_i$, and hence this is indeed an implementation of the *rewire-to-random* dynamics.
Now observe that $L_{i+1}-L_{i}$ are i.i.d. $\mbox{Geom}(\frac{n-1}{n})$ variables. It follows by a large deviation estimate that $L_{6n^2}<\frac{13n^2}{2}$ with exponentially high probability. Now for $v\in V$, let $$N(v)=\#\{i\leq \frac{13n^2}{2}: W_{i}=v\}.$$ Clearly, $N(v)$ is distributed as $\mbox{Bin}(\frac{13n^2}{2},\frac{1}{n})$, and a Chernoff bound implies $${\mathbb{P}}[N(v)\geq 7n] \leq e^{-n/78}.$$ Hence, noting that $D_{\max}(0)\leq n$, we have using a union bound over all the vertices
$$\label{e:tau1}
{\mathbb{P}}[\tau_1 \leq 6n^2]\leq ne^{-n/78}+{\mathbb{P}}[L_{6n^2}>\frac{13n^2}{2}].$$
This completes the proof of the lemma.
\[l:betasmalldensity\] There exists $\beta_0>0$, such that for all $\beta<\beta_0$, we have $\tau_2\geq 6n^2\wedge \tau$ with high probability.
For $t\geq 1$, it is easy to see that $RL(t)$, the number of relabelling moves upto time $t$, is stochastically dominated by a $\mbox{Bin}(6n^2,\frac{\beta}{n})$ variable. On $\{t<\tau\}$, we have that $N_0(t)-N_0(0)$ is distributed as $Z_{RL(t)}$ where $\{Z_i\}_{i\geq 0}$ is a simple symmetric random walk on ${\mathbb{Z}}$ started from $0$. Using a union bound it follows that
$${\mathbb{P}}[\tau _2 < 6n^2\wedge \tau]\leq {\mathbb{P}}[N_{*}(0)\leq \frac{2n}{5}]+ {\mathbb{P}}[RL(6n^2)>12\beta n]+{\mathbb{P}}[\max_{i\leq 12\beta n}|Z_i| \geq \frac{3n}{20}].$$ By choosing $\beta_0$ sufficiently small, the last term in the above inequality is $0$ for all $\beta<\beta_0$. Noticing that ${\mathbb{P}}[N_{*}(0)<\frac{2n}{5}]=2{\mathbb{P}}[\mbox{Bin}(n,\frac{1}{2})<\frac{2n}{5}]$ and using Hoeffding inequality to bound the first term and a Chernoff bound on the second term yields
$$\label{e:tau2}
{\mathbb{P}}[\tau _2 < 6n^2\wedge \tau]\leq 2e^{-2n/25} + e^{-2\beta n}.$$
This completes the proof of the lemma.
Now we are ready to prove Theorem \[t:rewiretorandom\]$(i)$.
From Lemma \[l:betasmallmg\], Lemma \[l:betasmallincoming\] and Lemma \[l:betasmalldensity\] it follows that with high probability we have $\{\tau_0\leq 6n^2, \tau_1\wedge \tau_2\geq 6n^2\wedge \tau\}$. It follows that, $\tau\leq 6n^2$ with high probability. The second part of the theorem follows from noting that using a random walk estimate as in Lemma \[l:betasmalldensity\], we see that for each $\eta>0$, the probability that the density of the minority opinion drops below $\frac{1}{2}-\eta$, within $6n^2$ steps tends to $0$ as $n\rightarrow\infty$. This completes the proof of the theorem.
High relabelling rate case: Stopping times {#s:stoptime}
==========================================
A time change: *Rewire-to-random-\** dynamics
---------------------------------------------
For the proof of Theorem \[t:rewiretorandom\]$(ii)$, we shall consider a time changed variant of *rewire-to-random* dynamics, which we call *rewire-to-random-\** model. This model is same as the *rewire-to-random* model, except that now at time $(t+1)$, instead of choosing a disagreeing edge at random, we choose an edge at random from $G(t)$. If the edge is not disagreeing, then we do nothing. It is clear that *rewite-to-random-\** model is a slowed down version of *rewire-to-random* model. It is also clear that if we prove Theorem \[t:rewiretorandom\]$(ii)$ for the *rewire-to-random-\** dynamics, then it will imply the same theorem for the *rewire-to-random* dynamics.
[**Assumption on the initial condition:**]{} For this section and the next two, we shall always assume that $G(0)$ satisfies the following conditions.
1. $|E(0)|$, the number of edges in $G(0)$ is in $[\frac{n^2}{4}-n^{3/2}, \frac{n^2}{4}+n^{3/2}]$.
2. $\#\{v\in V: v(0)=0\} \in [\frac{n}{2}-n^{3/4}, \frac{n}{2}+n^{3/4}]$.
Since both the events hold with probability $1-o(1)$, this assumption does not affect any of our results.
Now we move towards proving Theorem \[t:rewiretorandom\]$(ii)$. Let us fix $\varepsilon'>\varepsilon>0$ for the rest of this paper. Let $\tau_{*}=\tau_{*}(\varepsilon)$, i.e., $$\tau_*=\min\{t:\min(N_{0}(t),N_{1}(t))< \varepsilon n\}.$$
**Parameters:** Now we define the following stopping times. In the definition of these stopping times and the proofs that follow we use a number of parameters that need to satisfy the following relationships. For a fixed $
\varepsilon$ our parameters satisfy the following inequalities. $\varepsilon_2<\varepsilon^2/1000$, $\varepsilon_3< \varepsilon_2^{2}/1000$. $\varepsilon_7<\varepsilon_3^{2}/1000$. Fixing these parameters we choose $C_2=2$, $\varepsilon_4< \frac{1}{4\log 10}$, $\delta < \frac{\varepsilon_3}{10000C_2}$. We choose $\varepsilon_{14}< \varepsilon^2/100$. Fixing all these $C_1$ is chosen sufficiently large depending on $\delta$ and $\varepsilon$, here the exact functional dependence is not of interest to us. After fixing all these parameters, $C$ is chosen sufficiently large depending on these. There are also many other parameters used in the proofs which are chosen either sufficiently small or large depending on other parameters, again where the functional dependence is not of importance to us. Finally $\beta$ is taken sufficiently large depending on all the parameters used. Also we shall always take $n$ sufficiently large depending on everything else.
$\bullet$ [**Stopping times for large Cuts**]{}:
Let $S$ and $T$ be two disjoints subsets of $V$ with $S\cup T=V$. We denote by $N_{ST}(t)$ the number of edges in $G(t)$ with one endpoint in $S$ and another endpoint in $T$. Define $N_{SS}(t)$ similarly. Also let $N(t)=N$ denote the total number of edges in $G(t)$. Let $$K_{ST}(t)=(\dfrac{N_{SS}(t)-\frac{1}{4}|S|^2}{N(t)})^2+ (\dfrac{N_{TT}(t)-\frac{1}{4}|T|^2}{N(t)})^2.$$ Let $$L(t)=\max_{S,T:\min(|S|,|T|)\geq \varepsilon_2 n}K_{ST}(t).$$ Also let $$L'(t)=\max_{S,T:\min(|S|,|T|)\geq \varepsilon_2 n} \mid \dfrac{N_{ST}(t)-\frac{1}{2}|S||T|}{N(t)} \mid \vee \mid \dfrac{N_{SS}(t)-\frac{1}{4}|S|^2}{N(t)} \mid.$$
Now the two stopping times are defined as follows:
- The stronger stopping time: $\tau_2=\min\{t: L(t)\geq \varepsilon_3^2\}$.
- The weaker stopping time: $\tau'_2=\min\{t: L'(t)\geq 2\varepsilon_3\}$.
$\bullet$ [**Stopping times for individual edge multiplicities**]{}:
For $u,v\in V$, let $M_{uv}(t)$ denote the number of edges in the bond $(u,v)$ in $G(t)$. Let $M(t)=\max_{u\neq v}M_{uv}(t)$. Now the two stopping times are defined as follows:
- The stronger stopping time: $\tau_3=\min\{t: M(t)\geq \varepsilon_4\log n\}$.
- The weaker stopping time: $\tau'_3=\min\{t: M(t)\geq 2\varepsilon_4\log n\}$.
$\bullet$ [**Stopping times for balanced vertices**]{}:
Let us call a vertex $v$ $\epsilon$-balanced if for all $k$, $\#\{u\in V: M_{uv}=k\}\leq \epsilon 10^{-k}n$. We define the two stopping times as follows.
- The stronger stopping time: $\tau_4=\min\{t:\exists v\in V ~\text{not}~C_1\text{-balanced in}~G(t)\}$.
- The weaker stopping time: $\tau'_4=\min\{t:\exists v\in V ~\text{not}~2C_1\text{-balanced in}~G(t)\}$.
$\bullet$ [**Stopping times for maximum and minimum degrees**]{}:
Let $D_{\max}(t)$ and $D_{\min}(t)$ denote the maximum and minimum degree in $G(t)$ respectively. The stopping times are defined as follows.
- The stronger stopping time: $\tau_5=\min\{t: D_{\max}(G(t))> (1-\frac{\varepsilon}{2})n~\text{or}~D_{\min}(t)<\frac{\varepsilon n}{2} \}$.
- The weaker stopping time: $\tau'_5=\min\{t: D_{\max}(t)>C_2n~\text{or}~D_{\min}(G(t))<\frac{\varepsilon n}{4}\}$.
It is easy to see that for each $i=2,3,4,5$, we have $\tau'_i\leq \tau_i$.\
Finally we define $\tau_0=\tau_{*}\wedge \tau_2\wedge \tau_3\wedge \tau_4\wedge \tau_5$ and $\tau'_0=\tau_{*}\wedge \tau'_2\wedge \tau'_3\wedge \tau'_4\wedge \tau'_5$.\
Part $(ii)$ of Theorem \[t:rewiretorandom\] follows from the next theorem.
\[t:stoptime\] There exist $\beta_{*}=\beta_{*}(\varepsilon)$ such that for all $\beta>\beta_{*}$, we have for the rewire-to-random-\* model $\tau_0\geq \tau_*-n^2$ w.h.p.
We shall prove Theorem \[t:stoptime\] over the next two sections. Before that we show how this implies part $(ii)$ of Theorem \[t:rewiretorandom\] for *rewire-to-random-\** model.
Notice that, it follows from a random walk estimate that $\tau_{*}\geq \tau_*(\varepsilon')+n^2$ with high probability. On $\{\tau_0\geq \tau_{*}(\varepsilon')\}$, we have that $\tau_{*}(\varepsilon')-1<\tau_0$. And hence, in particular, $\tau_{*}(\varepsilon')-1<\tau_2$. Let $S$ be the set of all vertices with the minority opinion at time $\tau_{*}-1$. Since $\varepsilon_2< \varepsilon<\varepsilon'$, we have that $N_{ST}(\tau_{*}-1)\geq \frac{1}{2}|S||T|-2\varepsilon_{3}N(t)>0$, since $2\varepsilon_{3}<\varepsilon_2(1-\varepsilon_{2})$. It then follows that $\tau\geq \tau_{*}(\varepsilon')$ for the *rewire-to-random-\** dynamics. since the *rewire-to-random-\** dynamics is merely a time changed version of the *rewire-to-random* dynamics, the first statement in Theorem \[t:rewiretorandom\],(ii) follows. The second statement is an obvious consequence after observing that on $\{t<\tau\}$ the number of vertices of one opinion does a simple random walk which takes one step with roughly $\frac{n}{\beta}$ steps of the rewire-to-random dynamics.
Before starting with the proof of Theorem \[t:stoptime\] we prove the following lemma which establishes the connection between the evolving voter model dynamics and our stopping times that we shall exploit extensively.
\[l:mixing\] Consider the following continuous time random walk on $G(t)$. Let each directed edge ring at rate $\frac{\beta}{2n}$. Whenever an edge rings a walker at the starting point of the edge moves along the edge. Let $\lambda(G(t))$ denote the spectral gap of this Markov chain. Then there exists $\varepsilon_{14}>0$, such that we have, on $\{t<\tau'_2\wedge \tau'_3\wedge \tau'_4\wedge \tau'_5\}$, $\lambda(G(t))\geq \beta\varepsilon_{14}$.
Let $h(G(t))$ denote the Cheeger constant of the corresponding random walk. We have $$h(G(t)):=\min_{S,T S\cup T=V, S\cap T=\emptyset, |S|\leq n/2} \frac{N_{ST}(t)\beta}{2|S|n}.$$ Now on $\{t\leq \tau'_2\}$, if $|S|\geq \varepsilon_2 n$, $N_{ST}(t)\geq \frac{1}{2}|S||T|-2\varepsilon_3 N(t)\geq \frac{1}{2}|S||T|-\varepsilon_3 n^2$ and hence $$\frac{N_{ST}(t)\beta}{2|S|n}\geq \beta (\frac{1}{8}-\frac{\varepsilon_3}{2\varepsilon_2})\geq 2\beta \sqrt{C_2} \sqrt{\varepsilon_{14}}$$ provided $$\varepsilon_{14}\leq \dfrac{(\frac{1}{8}-\frac{\varepsilon_3}{2\varepsilon_2})^2}{4C_2}.$$ On $\{t\leq \tau'_4 \wedge \tau'_5\}$, if $|S|\leq \varepsilon_2 n$, then $$N_{SS}(t)+N_{ST}(t)\geq \frac{\varepsilon|S|n}{4}$$ and $$N_{SS}(t)\leq 2|S|n\varepsilon_2\log \frac{1}{\varepsilon_2}.$$
It follows that
$$\frac{N_{ST}(t)\beta}{2|S|n}\geq \beta (\frac{\varepsilon}{8}-2\varepsilon_2\log \frac{1}{\varepsilon_2})\geq 2\beta \sqrt{C_2} \sqrt{\varepsilon_{14}}$$ provided $$\varepsilon_{14}\leq \dfrac{(\frac{\varepsilon}{8}-2\varepsilon_2\log \frac{1}{\varepsilon_2})^2}{4C_2}.$$
Now it follows that on $\{t\leq \tau'_2\wedge \tau'_4\wedge \tau'_5\}$,
$$h(G(t))\geq 2\beta \sqrt{C_2} \sqrt{\varepsilon_{14}}.$$
Now, using Cheeger inequality (Theorem 13.14 of [@LPW09], see [@FN02] for the variant used here) we get,
$$\lambda(G(t))\geq \frac{h(G(t))^2}{2\frac{\beta D_{max}(t)}{n}}\geq \frac{h(G(t))^2}{2\beta C_2}\geq \beta \varepsilon_{14}$$ which completes the proof of the lemma.
Estimates for the weak stopping times {#s:weakbound}
=====================================
In this section we show that provided the process has not reached any of the stronger stopping times by time $t$, i.e., $t<\tau$, then within a small number ($\delta n^2$) steps, the process is unlikely to reach any of the weaker stopping times, i.e., $t+\delta n^2<\tau'_0$ with high probability. The general idea is that by time $\delta n^2$, there are not enough rewiring steps to change the graph substantially. We start with the following lemma which controls the fraction of minority opinion vertices in the time interval $\{t+1,t+2,\ldots, t+\delta n^2\}$.
\[lbw0\] We have that ${\mathbb{P}}(\tau_{*}(\frac{4\varepsilon}{5})\leq t+\delta n^2\mid {\mathcal{F}}_{t},t<\tau_0)\leq e^{-cn}$ where $c=c(\delta,\beta,\varepsilon)>0$.
It follows from a Chernoff bound that the probability that there are more than $2\delta\beta n$ many relabelling steps in $[t+1,t+\delta n^2]$ is exponentially small in $n$. Note that the number of vertices of a certain opinion does a simple symmetric random walk in absorbed at $0$ or $n$ in the *rewire-to-random* dynamics. The lemma now follows from a random walk estimate by observing that *rewire-to-random-\** dynamics is slower than rewire-to-random dynamics.
The next lemma considers the weaker stopping times for the large cuts.
\[l:lbw2\] On the event $\{t<\tau_0\}$, $t+\delta n^2<\tau'_2$.
Clearly for all $S,T$ that make a partition of $V$ and any $t'$ with $t\leq t'\leq t+\delta n^2$, $|N_{ST}(t')-N_{ST}(t)|\vee |N_{SS}(t')-N_{SS}(t')|\leq \delta n^2$. It follows from definitions that if $\delta <\frac{\varepsilon_3}{100}$ and $1000\varepsilon_7<\varepsilon$, then $t+\delta n^2\leq \tau'_2$.
The next lemma shows the weaker degree estimates continue to hold till time $t+\delta n^2$ with high probability if $t<\tau_0$.
\[l:lbw5\] We have ${\mathbb{P}}(t+\delta n^2\geq \tau'_5 \mid {\mathcal{F}}_t, t<\tau_0)\leq e^{-cn}$ for some constant $c>0$.
Condition on ${\mathcal{F}}_{t}$. For any fixed vertex $v$, the number of times in $[t+1,t+\delta n^2]$ an edge is rewired to $v$ is stochastically dominated by a $\mbox{Bin}(\delta n^2, \frac{1}{n-1})$ variable. By Chernoff’s inequality and a union bound it follows that the probability that any vertex gets more than $2\delta n$ edges is exponentially small in $n$. It follows that with exponentially high probability $\max_{t'\in [t+1,t+\delta n^2]}D_{\max}(t')<C_2n$ provided $C_2> (1-\frac{\varepsilon}{2})+2\delta$.
For the lower bound notice that the probability that a vertex $v$ with degree at least $\varepsilon n/2$ at time $t$ becomes of degree less that $\varepsilon n/4$ in time $[t+1,t+\delta n^2]$ is at most $${\mathbb{P}}\left(\mbox{Bin}\biggl(\delta n^2, \frac{5\varepsilon}{4n}\biggr) \geq \varepsilon n/4\right).$$ The above probability is exponentially small in $n$ by a Chernoff bound provided $6\delta < 1$. Taking a union bound over all the vertices completes the proof of the lemma.
Now we prove a similar statement for individual edge-multiplicities.
\[l:lbw3\] We have ${\mathbb{P}}(t+\delta n^2 \geq \tau'_3\mid {\mathcal{F}}_{t}, t<\tau_0) \leq \frac{1}{n^{20}}$.
For every bond $(u,v)$ in $V^{(2)}$, it follows from Lemma \[l:lbw5\] that with exponentially high probability $D_u(t')+D_v(t')\leq 2C_2 n$ for all $t'\in [t+1,t+\delta n^2]$ where $D_w(t')$ denotes the degree of the vertex $w$ in $G(t')$. Let $A_{uv}$ denote that event. Then, on $A_{uv}$, the number of edges added to the bond $(u,v)$ is stochastically dominated by $$\mbox{Bin}(\delta n^2, \frac{10C_2}{n(n-1)})\preceq \mbox{Bin}(\delta n^2, \frac{12C_2}{n^2})$$ variables provided $\varepsilon>16\varepsilon_3$ and $n$ sufficiently large. By a Chernoff bound again, provided $n$ is sufficiently large we get that $${\mathbb{P}}\left( \mbox{Bin}(\delta n^2, \frac{12C_2}{n^2})\geq \varepsilon_4\log n \right)\leq \frac{1}{n^{23}}.$$ Taking a union bound over all bonds completes the proof of the lemma.
Finally we work out the estimates for number of multi-edges incident to a given vertex.
\[l:lbw4\] We have ${\mathbb{P}}(t+\delta n^2\geq \tau'_4\mid {\mathcal{F}}_t, t< \tau_0)\leq \frac{1}{n^{18}}$.
Condition on ${\mathcal{F}}_t$. As in the previous lemma we shall ignore without loss of generality the event with exponentially small probability that for some $t'\in [t+1,t+\delta n^2]$ there exists $u,v\in V$ such that $D_u(t')+D_v(t')>2C_2n$. Fix $v\in V$ and $0<k<2\varepsilon_4\log n$. Let $N(\ell\rightarrow k)$ denote the number of bonds containing $v$ that gained at least $(k-\ell)$ edges during time $[t+1,t+\delta n^2]$. Clearly, $$\max_{t'\in [t+1,t+\delta n^2]}\#\{u:M_{uv}(t')=k\}\leq C_1 10^{-k}n+\sum_{\ell=0}^{k-1} N(\ell \rightarrow k)+ \#\{u:M_{uv}(t)>k\}.$$
Clearly the third term in the above sum is bounded by $\frac{1}{9}C_1 10^{-k}n$. To bound the second term fix $\ell$ with $0\leq \ell \leq k-1$. Let $u_1,u_2,\ldots, u_D$ be the set of vertices in $V$ such that $M_{vu_i}(t)=\ell$. Without loss of generality we can assume $D=C_110^{-\ell}n$. Let $T_i$ denote the number of edges gained by the bond $(v,u_i)$ in $[t+1,t+\delta n^2]$. We construct a family of random variables $(Y_1,Y_2,\ldots ,Y_D)$ which jointly stochastically dominates $(T_1,T_2,\ldots, T_D)$ as follows.
Consider $D$ urns all of which are empty to begin with. For $i=\{1,2,\ldots, \delta n^2\}$, at the ith step with probability $\frac{12C_2D}{n^2}$ we choose one of urns uniformly and put a ball in it. Let $(Y_1,Y_2,\ldots, Y_D)$ denote the vector of the number of balls in the urns after $\delta n^2$ steps. Notice that since at each step in the voter model dynamics only one of the bonds $(v,u_i)$ can gain an edge, and since for each of the bonds, the chance to gain an edge is at most $\frac{12C_2}{n^2}$ as shown in the previous lemma, it follows that $(Y_1,Y_2,\ldots, Y_D)$ indeed jointly stochastically dominates $(T_1,T_2,\ldots, T_D)$. Observe further that conditional on $\sum_{i=1}^{D}Y_i\leq 24\delta C_2 D$, the joint law of $(Y_1,Y_2,\ldots, Y_D)$ is stochastically dominated by the conditional joint law of $(Y'_1,Y'_2,\ldots, Y'_D)$ conditional on $\sum_{i=1}^{D}Y'_i\geq 24\delta C_2 D$ where $\{Y'_i\}$ is an i.i.d. sequence of $\mbox{Bin}(24\delta C_2 D, \frac{2}{D})$ variables. Let $Z_i$ (resp. $Z'_i$) denote the indicator of $Y_i\geq (k-\ell)$ (resp. $Y'_i\geq (k-\ell)$). Let $\mathcal{C}$ (resp. $\mathcal{C}'$) denote the event that $\sum_{i=1}^{D}Y_i\leq 24\delta C_2 D$ (resp. $\sum_{i=1}^{D}Y'_i\geq 24\delta C_2 D$). It follows therefore that $$\begin{aligned}
{\mathbb{P}}[\sum_{i=1}^D Z_i\geq 25^{-(k-\ell)}D]& \leq & {\mathbb{P}}[\mathcal{C}^c]+{\mathbb{P}}[\sum_{i=1}^D Z_i\geq 25^{-(k-\ell)}D\mid \mathcal{C}\\
&\leq & e^{-4\delta C_2 D}+{\mathbb{P}}[\sum_{i=1}^D Z'_i\geq 25^{-(k-\ell)}D\mid \mathcal{C}']\\
&\leq & e^{-4\delta C_2 D}+ 2\times {\mathbb{P}}[\mbox{Bin}(D,q(k-\ell))\geq 25^{-(k-\ell)}D]\end{aligned}$$ where $q(k-\ell)=\mbox{Bin}(24\delta C_2 D, \frac{2}{D})\geq (k-\ell))$. In the above equation we have used Chernoff bounds to deduce ${\mathbb{P}}[\mathcal{C}^c]\leq e^{-4\delta C_2 D}$ and ${\mathbb{P}}[\mathcal{C}_1]
\geq \frac{1}{2}$. Using another Chernoff bound we get $$q(k-\ell)\leq e^{-\frac{1}{2}(k-\ell)\log (\frac{1}{48C_2\delta})(k-\ell)}\leq 30^{-(k-\ell)}$$ provided $1>48000C_2\delta$. Since $N(\ell\rightarrow k) \preceq \sum Z_i$ it follows using Chernoff’s inequality yet again that
$${\mathbb{P}}(N(\ell\rightarrow k)\geq (2.5)^{-(k-\ell)}C_1 10^{-k}n)\leq e^{-\frac{C_1 10^{-k}n}{75}}\leq 2e^{-\frac{C_1\sqrt{n}}{75}}+e^{-4\delta C_1C_2\sqrt{n}/100}$$ since $k<2\varepsilon_4\log n$ provided $2\varepsilon_4 \log 10 <\frac{1}{2}$. Taking a union bound over all $\ell\in [0,k-1]$ gives that ${\mathbb{P}}[\max_{t'\in [t+1,t+\delta n^2]}\#\{u:M_{uv}(t')=k\}>2C_110^{-k}n]\leq \frac{1}{n^{25}}$. A union bound over all $k\in [1,2\varepsilon_4\log n]$, Lemma \[l:lbw3\] and another union bound over all vertices $v$ completes the proof of the lemma.
All the lemmas in this section together imply the following Theorem.
\[t:weakbound\] For the rewire-to-random-\* model, we have ${\mathbb{P}}(t+\delta n^2<\tau_{*},t+\delta n^2 \geq \tau'_0\mid {\mathcal{F}}_t, t<\tau_0)\leq \frac{1}{n^{17}}$.
Estimates for Strong Stopping Times {#s:strongbound}
===================================
Our goal in this section is to prove that if by time $t$ the process does not reach any of the strong stopping times, then it is also unlikely that any of the strong stopping times will be hit by time $t+\delta n^2$ unless the minority opinion density drops below $\epsilon$. We shall prove this by separate analysis of each of the stopping times. In the heart of the analysis, in each case, is some estimates on how the opinions of the vertices get mixed in a short (compared to $\delta n^2$) time, which we prove by constructing a coupling of a random walk on the graph $G(t)$ with the evolving voter model dynamics.
The Coupling Construction {#s:coupling}
-------------------------
The dual relationship between the random walk and the voter model on a fixed graph $H$ is well known. The distribution of the voter model $X(t)$ started from $X(0)$ can be constructed by running coalescing random walks for time $t$ from each vertex and setting $X_u(t)$ to be the value of $X(0)$ at the location of the walker started from $u$ (c.f. §1.7, [@Dur06]). We prove that an analogue of that result holds in our setup. We want to say that if $t$ is small so that that graph has not changed sufficiently in the meantime, then the two distributions are not far apart. Now we formally describe the coupling.\
[**An equivalent implementation of the *rewire-to-random-\** dynamics starting at time $t+1$:**]{} Let us condition on ${\mathcal{F}}_{t}$. Let $N=N(t)$ be the number of edges in $G(t)$. Let the set of all labelled edges be $\{e_1,e_2,\ldots, e_N\}$. For the purpose of this subsection we shall assume that these edges are directed, i.e., the edges have two identifiable ends $e_{i}^{+}$ and $e_{i}^{-}$. When an edge is placed in a bond, we think of it as $e_i^{+}$ being placed at one vertex of the bond and $e_{i}^{-}$ being placed at the other. Suppose $e_{i}$ is placed in the bond $(u,v)$ and $e_i^{+}$ is placed at $u$ and $e_{i}^{-}$ placed at $v$. Then if a rewiring move rewires $e$ with root $u$ to the vertex $w$, then $e_i^{-}$ is placed at $w$. Consider the two independent sequences $\mathbb{RW}=\{RW_i\}_{i\geq 1}$ and $\mathbb{RL}=\{RL_i\}_{i\in {\mathbb{Z}}}$ where each $RL_i$ are chosen independently and uniformly from the set $E^*=\{e_1^+,e_1^-,e_2^+,e_2^-,\ldots , e_N^{+}, e_N^{-}\}$. The sequence $\mathbb{RW}$ is also an i.i.d. sequence where each, $RW_i=(e^*,v)$, where $e^*$ is picked uniformly from $E^*=\{e_1^+,e_1^-,e_2^+,e_2^-,\ldots , e_N^{+}, e_N^{-}\}$, and $v$ is a uniform random vertex $v$ picked from $V$ independently of $e^*$. Also let $Z_{i}$ be a sequence of i.i.d. $\mbox{Ber}\left({\frac{\beta}{n}}\right)$ variables.
We construct the equivalent formulation of the process as follows. At time $t+i$, if $Z_i=0$, then choose the first uninspected element from $\mathbb{RW}$, let that element be $(e_j^{+},v)$. If the edge $e_j$ is not disagreeing in $G(t+i-1)$ then do nothing. Otherwise, we try to rewire the edge $e_j$ to $v$, with the root as the vertex having $e_j^{+}$. If this rewiring is not legal, (i.e., $e_j^{+}$ is already placed at $v$ in $G(t+i-1)$), then we choose the next element from $\mathbb{RW}$ and repeat the process. If $Z_i=1$, then we choose the first unused element from $\mathbb{RL}$, let that element be $e_j^{+}$. If the edge $e_j$ is not disagreeing in $G(t+i-1)$ then do nothing. Otherwise, we change the opinion of the vertex containing $e_j^{+}$. Notice that $\mathbb{RL}$ is a bi-infinite sequence but we start inspecting the elements starting from $RL_1$. It is clear from our construction that this indeed is an equivalent implementation of the *rewire-to-random-\** dynamics. Let us run this dynamics starting with $G(t)$ upto $\frac{Cn^2}{\beta}$ steps. Let $\sigma$ and $\omega$ be the number of elements of $\mathbb{RL}$ and $\mathbb{RW}$ that gets inspected in the process. Notice that $\sigma$ is independent of the sequences $\mathbb{RL}$ and $\mathbb{RW}$.
[**Coupling with continuous time random walks:**]{} Now consider the following continuous time random walk on $G(t)$. Each directed edge rings at rate $\frac{\beta}{2n}$. When a directed edge rings a walker at the starting point of the edge moves along the edge. Consider the process where we start with one walker at each vertex of some arbitrary subset $W\subseteq V$, and each walker independently performs the random walk described above. We consider the following coupling between this process and the evolving voter model process described above. To start with each of the random walks are of type $A$. Now choose $T_i$ i.i.d. $exp(2N)$. At the $i$-th step, wait time $T_i$. If $Z_{Cn^2/\beta+1-i}=0$, then do nothing. If $Z_{Cn^2/\beta+1-i}=1$, then look at $RL_{\sigma+1-k(i)}$ where $k(i)=\#\{j\in [Cn^2/\beta+1-i, Cn^2/\beta]: Z_j=1\}$. If there is any walker of type $A$ at the starting point of that edge, then that walker takes a step along that edge. If any walker of type $A$ takes a step to a vertex where there is already one or more walkers, then all the walkers become of type $B$. Type $B$ walkers do a random walk having the same waiting time distributions but using independent randomness. It is clear that the random walks are independent. Also since $\sigma$ is independent of $\mathbb{RL}$, the random walks also have correct marginals. So this is indeed a coupling as we claimed.
Let $T=\sum_{i=1}^{Cn^2/\beta}T_i$. Let us make the following definitions. For $v\in V$, let $$E^*(v)=\{e^*\in E^*: (e^*,v)=RW_i~\text{for some}~i<\omega\}.$$
\[d:happy\] Consider the random walks described above. At time $s$, we call the walker starting at $v_0$ [**happy**]{} if the following conditions are satisfied.
1. It is of type $A$ at time $s$.
2. None of the edges the walker has traversed have been rewired in the voter model process.
3. Let the path traversed by the walker be $\{v_0,v_1,\ldots, v_k\}$, with the times of jump being $0=T_0^*<T_1^*<T_2^*<\ldots <T_k^*\leq s$. Then for each $\ell$, and each $e^*\in E^*(v_{\ell})$ there was no ring in $e^*$ in $[T_{\ell}^*, T_{\ell+1}^*]$.
The following lemma records the most basic useful fact about this coupling construction.
\[l:couplingbasic1\] Consider the coupling described above. At time $T$, if a walker starting from $v$ is [**happy**]{}, then the opinion of the position of that walker at time $T$ is the same as the opinion of $v$ in the *rewire-to-random-\** dynamics at time $t+\frac{Cn^2}{\beta}$.
This follows from the definition of the coupling.
\[l:couplingbasic2\] For the coupling described as above, let $T=\sum_{i=1}^{\frac{Cn^2}{\beta}} T_i$. Then, for any $\kappa_3>0$, with exponentially high probability $\frac{2C-\kappa_3}{\beta} \leq T \leq \frac{2C+\kappa_3}{\beta}$.
The result follows from a large deviation estimate for sum of independent exponential variables.
\[l:couplingtv\] Consider the continuous time random walk on $G(t)$ described above starting from an arbitrary vertex $v$. Let $Y(t+t')$ denote the position of the walk at time $t'$. Also let $U$ be a uniformly chosen vertex from $V$. Then for sufficiently large $C$, we have that on $\{t<\tau'_0\}$, for all $t'\geq \frac{C}{\beta}$, $$||Y(t+t')-U||_{TV}\leq e^{-\frac{\sqrt{C}}{1000}}.$$
We know from Lemma \[l:mixing\] that on $\{t<\tau'_0\}$, $\lambda:=\lambda(G(t))\geq \beta \varepsilon_{14}$. Now let $T_1$ be the time of the first jump of the walker initially at $v$. Then we know by $L^2$ Contraction Lemma (Lemma 3.26, [@AF02]) $$||Y(t+T_1+t^*)-U||_2 \leq e^{-t^*\lambda}||Y(t+T_1)-U||_2.$$
Now we know that on $\{t<\tau'_0\}$,
$$\begin{aligned}
||Y(t+T_1)-U||_2^{2} &\leq & \sum _{k=1}^{\infty} \frac{9k^2}{n^2}2C_110^{-k}n^2 \\ &\leq & 18C_1\sum_{k=1}^{\infty} k^210^{-k}\\
&\leq & 50C_1.\end{aligned}$$
Now observe that for $t'\geq \frac{C}{\beta}$,
$$\begin{aligned}
||Y(t+t')-U||_1 &\leq & 2||Y(t+T+t'/2)-U||_1 +{\mathbb{P}}(T_1\geq t'/2)\\
&\leq & 2 e^{-\frac{t'\lambda}{2}}||Y(t+T+t'/2)-U||_2+ {\mathbb{P}}(T_1\geq t'/2)\\
&\leq & 10\sqrt{2C_1}e^{-t'\beta \varepsilon_{14}/2}+ e^{-\frac{C\varepsilon}{16}}\\
&\leq & e^{-C\varepsilon_{14}/20}+e^{-\frac{3C\varepsilon}{40}}.\end{aligned}$$
The lemma follows by taking $C$ sufficiently large.
Now we need the following properties of the coupling described above.
\[l:nonintersection\] Let $\kappa>0$ be fixed. Let $v_1,v_2,\ldots, v_{\varepsilon_{13}n}$ be given vertices in $V$. From each of the vertices we run independent discrete time simple random walks in $G(t)$ upto $20C$ steps. Then, on $\{t<\tau'_0\}$, we have for $\varepsilon_{13}$ small enough, with exponentially high probability there exists at least $(1-\kappa)\varepsilon_{13}n$ vertices among these such that the paths of random walks started from these vertices do not intersect.
Condition on $\{{\mathcal{F}}_t, t<\tau'_0\}$. For $i=1,2,\ldots, \varepsilon_{13}n$, and $j=1,2,\ldots,4C$, let $Z_{ij}$ denote the indicator of the event that the random walk started from $i$ hits the set $V^*=\{v_1,v_2, \ldots, v_{\varepsilon_{13}n}\}$ in step $j$. Let ${\mathcal{F}}_{i,j}$ denote the filtration generated by the random walk paths of the walks started from $v_1,v_2,\ldots, v_{i-1}$ and the first $(j-1)$ steps of the random walk started from $v_i$. Now notice that on $\{t<\tau'_0\}$, for any vertex $v$, the number of edges from $v$ to $V^*$ in $G(t)$ is at most $$\sum_{j: \sum_{k\geq j}{2C_110^{-k}\leq \varepsilon_{13}}} 2C_1j10^{-j}n\leq \sum_{j\geq \log(\frac{4C_1}{\varepsilon_{13}})} 2C_1j10^{-j}n\leq 25\varepsilon_{13}\log(\frac{4C_1}{\varepsilon_{13}})n \leq \frac{\varepsilon \kappa n}{400C}$$ for $\varepsilon_{13}$ sufficiently small. Since the degree of each vertex is at least $\frac{\varepsilon n}{4}$, it follows that $$E[Z_{ij}\mid {\mathcal{F}}_{i,j-1}]\leq \frac{\kappa}{100C}.$$ It follows from Azuma’s inequality that
$$\label{e:azuma3}
{\mathbb{P}}\left(\sum_{i=1}^{\varepsilon_{13}n}\sum_{j=1}^{20C}\left[Z_{ij} -E[Z_{ij}\mid {\mathcal{F}}_{i,j-1}]\right]\geq \frac{\kappa\varepsilon_{13}}{4}n\right) \leq e^{-\kappa^2\varepsilon_{13}n/640C}.$$
Hence $$\label{e:azuma4}
{\mathbb{P}}\left(\sum_{i=1}^{\varepsilon_{13}n}\sum_{j=1}^{20C} Z_{ij} \geq \frac{\kappa\varepsilon_{13}}{2}n\right) \leq e^{-\kappa^2\varepsilon_{13}n/640C}.$$
Now let $B_i$ denote the event that the random walk starting from $v_i$ intersects a random walk started from $v_j$ for some $j<i$ at a point other than $v_i$. Let $Y_i=1_{B_i}$. Let $\mathcal{G}_i$ denote the filtration generated by the paths of random walks started from $v_1,v_2,\ldots ,v_i$. Let $C_i$ be the set of vertices visited by the first $(i-1)$ random walks except possibly $v_i$. Clearly $|C_i|\leq 25C\varepsilon_{13}n$. Arguing as before, the number of edges from any vertex $v$ to $C_i$ is at most $625C\varepsilon_{13}\log(\frac{4C_1}{25C\varepsilon_{13}})n\leq \frac{\kappa \varepsilon n}{400C}$ for $\varepsilon_{13}$ sufficiently small. By a union bound over the steps of the random walk started from $v_i$ it follows that ${\mathbb{E}}(Y_i\mid \mathcal{G}_{i-1})\leq \frac{\kappa}{4}$. Using Azuma’s inequality as before we get
$$\label{e:azuma5}
{\mathbb{P}}\left(\sum_{i=1}^{\varepsilon_{13}n}\left[Y_i -E[Y_i\mid \mathcal{G}_{i-1}]\right]\geq \frac{\delta\varepsilon_{13}}{4}n\right) \leq e^{-\kappa^2\varepsilon_{13}n/32}.$$
Hence $$\label{e:azuma6}
{\mathbb{P}}\left(\sum_{i=1}^{\varepsilon_{13}n} Y_i \geq \frac{\kappa\varepsilon_{13}}{2}n\right) \leq e^{-\kappa^2\varepsilon_{13}n/32}.$$
Now let $D_1$ be the set of vertices $v_i$ such that $v_i$ is hit by the random walk started from some $v_j$, $j\neq i$. Clearly $$|D_1|\leq \sum_{i=1}^{\varepsilon_{13}n}\sum_{j=1}^{20C}Z_{ij}.$$ Also let $D_2$ be the set of vertices $v_i$ such that the random walk started from $v_i$ intersects a random walk started from $v_j$ for some $j<i$. Clearly $$|D_2|\leq \sum_{i=1}^{\varepsilon_{13}n} Y_i.$$
It is also clear that for $i,j\in V^*\setminus (D_1\cup D_2)$, random walks started from $v_i$ and $v_j$ do not intersect. The lemma now follows from (\[e:azuma4\]) and (\[e:azuma6\]).
\[l:notmanysteps\] Let $v_1,v_2,\ldots, v_{\varepsilon_{13}n}$ be given vertices in $V$. From each of the vertices we run independent continuous time random walks in $G(t)$ as described in the coupling upto time $\frac{10C}{\beta}$. Let $Z_1$ denote the number of walks that take more than $20C$ steps in this time. Then, ${\mathbb{P}}[Z_1\geq \kappa \varepsilon_{13}n\mid {\mathcal{F}}_t, t<\tau'_0]$ is exponentially small in $n$ for $C$ sufficiently large .
Condition on $\{{\mathcal{F}}_t, t<\tau'_0\}$. Since the number of steps taken by each random walk is independent and are stochastically dominated by a $\mbox{Poi}(5CC_2)$ variable, it follows from a large deviation estimate and $C_2\leq 2$, thar ${\mathbb{P}}(Z_1>\kappa\varepsilon_{13}n)$ is exponentially small in $n$ by taking $C$ sufficiently large.
\[l:notrewired\] Assume the hypothesis of Lemma \[l:notmanysteps\]. Consider the coupling of the random walks with the evolving voter model as described above. Let $Z_2$ denote the number of walkers which traverse by time $\frac{10C}{\beta}$, some edge that is rewired during the voter model process. Then ${\mathbb{P}}(Z_2\geq 3\kappa \varepsilon_{13}n\mid {\mathcal{F}}_t, t<\tau'_0)\leq e^{-\gamma n}$ for $C$ sufficiently large, $\varepsilon_{13}=\varepsilon_{13}(C)$ sufficiently small and $\beta=\beta(C)$ sufficiently large, and for some $\gamma>0$ that does not depend on $\beta$.
Condition on $\{{\mathcal{F}}_t, t<\tau'_0\}$. In the coupling construction, let $\omega$ be the number of entries in $\mathbb{RW}$ that were inspected. We consider the case $\omega\leq \frac{2Cn^2}{\beta}$ since the complement of this event has probability that is exponentially small in $n^2$ and can be ignored. Notice that $\mathbb{RW}$ is independent of $\mathbb{RL}$ and hence is also independent of the random walks. Let $\mathcal{H}$ denote the filtration generated by the random walk paths. Let $D_i$ denote the set of edges traversed by the random walk started at $v_i$. Let $\mathcal{D}$ denote the event that there is $J\subseteq [\varepsilon_{13}n]$ with $|J|\geq (1-2\kappa)\varepsilon_{13}n$ such that for all $j_1,j_2\in J$, $D_{j_1}\cap D_{j_2}=\emptyset$ and $|D_{j_1}|\leq 20C$. It follows from Lemma \[l:notmanysteps\] and Lemma \[l:nonintersection\] that ${\mathbb{P}}(\mathcal{D}^{c}\mid {\mathcal{F}}_t, t<\tau'_0)$ is exponentially small in $n$ for $C$ sufficiently large and $\varepsilon_{13}=\varepsilon_{13}(C)$ sufficiently small. Now let us condition on $\mathcal{H}$ and $\mathcal{D}$. Since $\mathbb{RW}$ is independent of $\mathcal{H}$, for $j\in J$, we have $${\mathbb{P}}(\text{the edge in}~RW_{k}\in D_j~\text{for some}~k\leq \frac{2Cn^2}{\beta})\leq \frac{200C^2}{\beta}.$$
By a large deviation estimate, it follows that if $\beta$ is sufficiently large so that $\beta>\frac{2000C^2}{\kappa}$ then $${\mathbb{P}}(Z_2\geq 3\kappa\varepsilon_{13}n\mid {\mathcal{F}}_t, t<\tau'_0, \mathcal{H},\mathcal{D})$$ is exponentially small in $n$. This finishes the proof of the lemma.
\[l:happycondition\] Assume the hypothesis of Lemma \[l:notmanysteps\]. Consider the coupling of the random walks with the evolving voter model as described above. Let $Z_3$ be the number of walks that take less than $20C$ steps up to time $\frac{10C}{\beta}$ but violates condition $3$ in Definition \[d:happy\] at time $T$. Then ${\mathbb{P}}(Z_3>\kappa\varepsilon_{13}n)$ is exponentially small in $n$ for $\beta$ sufficiently large where the exponent does not depend on $\beta$.
Let us fix a function $q(\cdot)$ such that $q(\beta)<<\beta << q(\beta)\log q(\beta)$ as $\beta \rightarrow \infty$. Since the random walks are independent of $\mathbb{RW}$, we condition on $\mathbb{RW}$ and the following event $$\mathcal{G}=\{\omega < \frac{2Cn^2}{\beta}, |E^*(v)|\leq \frac{4Cnq(\beta)}{\beta}\forall v\in V\}.$$ By a Chernoff bound ${\mathbb{P}}(\mathcal{G})\geq 1-e^{-\gamma n}$ where $\gamma$ does not depend on $\beta$. Conditional on $\mathcal{G}$, for each $i$, the chance that the random walk started from $v_i$ violates condition $3$ in Definition \[d:happy\] before making $20C$ many jumps is at most $\frac{200C^2 q(\beta)}{\varepsilon \beta}$. Since the events are independent for different $i$ conditional on $\mathcal{G}$, the lemma follows for $\beta$ sufficiently large.
\[l:couplemany\] Let $v_1, v_2,\ldots , v_{\varepsilon_{13}n}$ be fixed vertices in $V$. Consider the coupling between the evolving voter model starting with $G(t)$ at time $t$, with independent continuous time random walks on $G(t)$ starting with one walker at each $v_i$ as described above. Let us denote the position of the random walk started at $v_i$ at time $s$ by $Y_i(t+s)$. Let $\Pi$ denote the event that there exists a time $\sigma_0\in \{C/\beta +\frac{1}{n^3}, C/\beta +\frac{2}{n^3}, \ldots, 3C/\beta\}$ and $J\subseteq [\varepsilon_{13}n]$ with $|J|\geq (1-7\kappa)\varepsilon_{13}n$, such that for each $i\in J$, opinion of $Y_i(t+\sigma_0)$ in $G(t)$ is same as $v_i(t+\frac{Cn^2}{\beta})$. Then for $C$ sufficiently large, $\varepsilon_{13}=\varepsilon_{13}(C)$ sufficiently small and $\beta=\beta(C)$ sufficiently large, we have that ${\mathbb{P}}(\Pi\mid {\mathcal{F}}_t, t<\tau'_0)\geq 1-e^{-\gamma n/2}$ where $\gamma>0$ does not depend on $\beta$.
In this proof the value of the constant $\gamma$ may change from line to line but $\gamma$ is always a positive constant independent of $\beta$. Condition on $\{{\mathcal{F}}_t, t<\tau'_0\}$. Let $Z_2$ denote the number of walkers at time $\frac{3C}{\beta}$ of type $B$. Then it follows from Lemma \[l:notmanysteps\] and Lemma \[l:nonintersection\] that ${\mathbb{P}}(Z_2>\kappa\varepsilon_{13}n)\leq e^{-\gamma n}$ for some $\gamma>0$ not depending on $\beta$. Let $Q$ denote the event that there exists $J\subseteq [\varepsilon_{13}n]$ with $|J|\geq (1-5\kappa)\varepsilon_{13}n$ such that for all $j\in J$, the opinion of $Y_j(t+T)$ in $G(t)$ is same as $v_j(t+\frac{Cn^2}{\beta})$. It now follows from Lemma \[l:couplingbasic1\], Lemma \[l:notrewired\] and Lemma \[l:happycondition\] that for appropriate choices of $C,\varepsilon_{13}$ and $\beta$, ${\mathbb{P}}(Q^c, T<\frac{3C}{\beta}\mid {\mathcal{F}}_t, t<\tau'_0)\leq e^{-\gamma n}$.
Now let $A$ denote the event there exist $k\in \{1,2,\ldots ,\frac{2n^3C}{\beta}\}$ such that there are more than $\kappa\varepsilon_{13}n$ of the random walks take a step within time $[\frac{C}{\beta}+\frac{i}{n^3}, \frac{C}{\beta}+\frac{i+1}{n^3}]$. By a union bound it follows that $P(A\mid {\mathcal{F}}_t, t<\tau'_0)\leq e^{-\gamma n}$. It follows now that, $${\mathbb{P}}(\Pi\mid {\mathcal{F}}_t, t<\tau'_0)\geq 1-{\mathbb{P}}(T>\frac{3C}{\beta}\mid {\mathcal{F}}_t, t<\tau'_0)-{\mathbb{P}}(A\mid {\mathcal{F}}_t, t<\tau'_0)-{\mathbb{P}}(Q^c, T<\frac{3C}{\beta}\mid {\mathcal{F}}_t, t<\tau'_0).$$ The proof of the lemma is completed using Lemma \[l:couplingbasic2\].
Bound for Large Cuts
--------------------
Let us fix $S,T\subseteq V$, such that $S\cap T=\emptyset$ and $S\cup T=V$ with $\varepsilon_2 n\leq |S|\leq |T|$.
\[p:proportionofbadedges\] For the *rewire-to-random-\** dynamics, let us condition on $\{{\mathcal{F}}_t, t<\tau'_0, N_1(t)=pn\}$. For $t'>t$, let $X_{ST}(t')$ denote the number of disagreeing edges at time $t'$. Then there exists a constant $C$ sufficiently large, and $\beta=\beta(C)$ sufficiently large such that, $${\mathbb{P}}\left(X_{ST}(t+\frac{Cn^2}{\beta})\notin ((2p(1-p)-\varepsilon_7)N_{ST}(t+\frac{Cn^2}{\beta}), (2p(1-p)+\varepsilon_7)N_{ST}(t+\frac{Cn^2}{\beta})\mid {\mathcal{F}}_{t}, t< \tau'_0\right)\leq e^{-\gamma n}$$ for some $\gamma>0$ that does not depend on $\beta$.
We shall need the following lemma in order to prove Proposition \[p:proportionofbadedges\].
\[l:vdisjoint\] Let $\mathcal{E}_{S,T}(t)$ denote the set of edges that have one endpoint in $S$ and another endpoint $T$ at time $t$. On $\{t<\tau'_{0}\}$, we have $|\mathcal{E}_{ST}(t)|\geq \frac{\varepsilon_2 n^2}{5}$ provided $1000\varepsilon_3< \varepsilon_2$. Let $\kappa>0$ be fixed. Let $e_1,e_2,\ldots e_{\varepsilon_{12}n}$ be uniformly chosen edges from $\mathcal{E}_{ST}(t)$. For $\varepsilon_{12}$ sufficiently small, with exponentially high probability there exists at least $(1-\kappa)\varepsilon_{12}n$ many edges among the sample that are vertex disjoint at time $t$.
For $i=1,2,\ldots, \varepsilon_{12}n$, let $A_i$ denote the event that $e_i$ is not vertex disjoint with $e_1,e_2,\ldots , e_{i-1}$. Let $Z_i=1_{A_i}$. Let ${\mathcal{F}}_{i-1}$ denote the filtration generated by $e_1,e_2,\ldots, e_{i-1}$. Then it is clear from the assumption on the graph that $${\mathbb{E}}(Z_i\mid {\mathcal{F}}_{i-1})\leq \frac{10C_2n+10\varepsilon_{3}\log n}{\varepsilon_2 n^2} \varepsilon_{12}n\leq \frac{12C_2 \varepsilon_{12}}{\varepsilon_2}.$$ Also notice that $|Z_i-{\mathbb{E}}(Z_i\mid {\mathcal{F}}_{i-1})|\leq 1$ and hence Azuma’s inequality yields
$$\label{e:azuma1}
{\mathbb{P}}\left(\sum_{i=1}^{\varepsilon_{12}n}\left[Z_i -E[Z\mid {\mathcal{F}}_{i-1}]\right]\geq \frac{\kappa\varepsilon_{12}}{2}n\right) \leq e^{-\kappa^2\varepsilon_{12}n/8}.$$
It follows that
$$\label{e:azuma2}
{\mathbb{P}}\left(\sum_{i=1}^{\varepsilon_{12}n}Z_i \geq \frac{12C_2\varepsilon_{12}^2n}{\varepsilon_2}+ \frac{\kappa\varepsilon_{12}}{2}n\right) \leq e^{-\kappa^2\varepsilon_{12}n/8}.$$
By choosing $\varepsilon_{12}$ sufficiently small such that $\frac{12C_2\varepsilon_{12}}{\varepsilon_2}\leq \frac{\kappa}{2}$ completes the proof.
Let us condition on $\{{\mathcal{F}}_{t}, t<\tau'_0, N_1(t)=pn\}$. Let $N_{ST}(t)$ be the number of edges in $G(t)$ with one endpoint in $S$ and another endpoint in $T$. Let $\{e_1,e_2,\ldots , e_{N_{ST}(t)}\}$ denote the set of those edges. Let $X_i$ be the indicator that endpoints of $e_i$ in $G(t)$ are disagreeing in $G(t+\frac{Cn^2}{\beta})$. Since in $Cn^2/\beta$ steps, at most $Cn^2/\beta$ edges can be rewired it follows that for $\beta$ sufficiently large, it suffices to prove that
$${\mathbb{P}}(\frac{1}{N_{ST}(t)}\sum_{i=1}^{N_{ST}(t)}X_i \notin ((2p(1-p)-\varepsilon_7/2), (2p(1-p)+\varepsilon_7/2)\mid {\mathcal{F}}_{t}, t< \tau'_0)\leq e^{-\gamma n}.$$
Let $J$ be a set of size $\varepsilon_{12}n$ where each element is an independent uniform sample from $[N_{ST}(t)]$. Clearly by Hoeffding’s inequality,
$${\mathbb{P}}\left(\mid \frac{1}{\varepsilon_{12}n}\sum_{j\in J} X_j - \frac{1}{N_{ST}(t)}\sum_{i=1}^{N}X_i \mid \geq \varepsilon_7/4 \right)\leq e^{-\frac{\varepsilon_{7}^2\varepsilon_{12}n}{32}}.$$
So it suffices for us to prove that with probability at least $1-e^{-2\gamma n}$,
$$\frac{1}{\varepsilon_{12}n}\sum_{j\in J} X_j\in (2p(1-p)-\varepsilon_7/4, 2p(1-p)+\varepsilon_7/4).$$
Choose $\varepsilon_{12}$ sufficiently small, and set $\varepsilon_{13}=2\varepsilon_{12}(1-\kappa)$ so that the conclusions of Lemma \[l:vdisjoint\] and Lemma \[l:couplemany\] are satisfied. Let $\mathcal{H}_1$ denote the event that there is a subset $J^*\subseteq J$ with $|J^*|=(1-\kappa)\varepsilon_{12}n$ such that endpoints of $e_j$ are disjoint for all $j\in J^*$. It follows from Lemma \[l:vdisjoint\] that ${\mathbb{P}}(\mathcal{H}\mid {\mathcal{F}}_t, t<\tau'_0)\geq 1-e^{-100\gamma n}$. Condition on $\mathcal{H}_1$ and $J^*$. Let $v_1,v_2,\ldots, v_{\varepsilon_{13}n}$ be endpoints of edges of $J^*$. By choosing $100\kappa < \varepsilon \varepsilon_{7}$ it follows that it suffices to prove $$\label{e:reducedlc}
{\mathbb{P}}[\frac{2}{\varepsilon_{13}n}\sum_{j\in J^*} X_j\notin (2p(1-p)-\varepsilon_7/8, 2p(1-p)+\varepsilon_7/8)\mid {\mathcal{F}}_t, t<\tau'_0, N_1(t)=pn, \mathcal{H}_1, J]\leq e^{-10\gamma n}.$$
Consider the coupling described in § \[s:coupling\]. Fix $j\in J^*$, let $v_{j_1}$ and $v_{j_2}$ be endpoints of $e_j$ in $G(t)$. For $\tilde{\sigma}\in \Sigma=\{C/\beta +\frac{1}{n^3}, C/\beta +\frac{2}{n^3}, \ldots, 3C/\beta\}$ let $U_{j,\tilde{\sigma}}$ denote the indicator that the position of the coupled random walks started from $v_{j_1}$ and $v_{j_2}$ at time $\tilde{\sigma}$ have different opinions in $G(t)$. Clearly, for a fixed $\tilde{\sigma}$, for all $j\in J^*$, $U_{j,\tilde{\sigma}}$ are conditionally independent. Also, it follows from Lemma \[l:couplingtv\] that ${\mathbb{E}}(U_{j,\tilde{\sigma}}\mid {\mathcal{F}}_t, t<\tau'_0, N_1(t)=pn)\in [2p(1-p)-\varepsilon_7/32,2p(1-p)+\varepsilon_7/32]$.
A standard Hoeffding bound now shows that conditional on $\mathcal{G}=\{{\mathcal{F}}_t, t<\tau'_0, N_1(t)=pn, \mathcal{H}_1,J^*\}$, with probability at least $1-e^{-20\gamma n}$, $$\label{e:reducedlc3}
\frac{2}{\varepsilon_{13}n}\sum_{j\in J^*} U_{j,\tilde{\sigma}}\in (2p(1-p)-\varepsilon_7/16, 2p(1-p)+\varepsilon_7/16).$$ By taking a union bound over all possible values of $\tilde{\sigma}$, it follows that that above holds for all $\tilde{\sigma}$ in $\Sigma$ with conditional probability at least $1-e^{-15\gamma n}$.
By observing that by Lemma \[l:couplemany\], we have, conditional on $\mathcal{G}$ there exists $\tilde{\sigma}\in \Sigma$ $$\frac{2}{\varepsilon_{13}n}\sum_{j\in J^*} |U_{j,\tilde{\sigma}}-X_j|\leq 12\kappa \leq \varepsilon_7/32$$ with probability at least $1-e^{-15\gamma n}$. This and the previous observation implies (\[e:reducedlc\]) and the proof of the proposition is complete.
The following proposition follows along the same lines as Proposition \[p:proportionofbadedges\] and we shall omit the proof.
\[p:proportionofbadedgesss\] Fix $S\subseteq V$ with $|S|\geq \varepsilon_2 n$. For the *rewire-to-random-\** dynamics, let us condition on $\{{\mathcal{F}}_t, t<\tau'_0, N_1(t)=pn\}$. For $t'>t$, let $X_{SS}(t')$ denote the number of disagreeing edges at time $t'$. Then there exists a constant $C$ sufficiently large, and $\beta=\beta(C)$ sufficiently large such that, $${\mathbb{P}}\left(X_{SS}(t+\frac{Cn^2}{\beta})\notin ((2p(1-p)-\varepsilon_7)N_{SS}(t+\frac{Cn^2}{\beta}), (2p(1-p)+\varepsilon_7)N_{SS}(t+\frac{Cn^2}{\beta})\mid {\mathcal{F}}_{t}, t< \tau'_0\right)\leq e^{-\gamma n}$$ for some $\gamma>0$ that does not depend on $\beta$.
Condition on $\{{\mathcal{F}}_{t}, t<\tau_0, N_1(t)=pn\}$. Let $D^*(t')$ denote the event that $N_1(t'')\in ((p-\varepsilon_7)n, (p+\varepsilon_7)n)$ for all $t''\in [t+1, t']$. Fix $S$ and $T$ as in Proposition \[p:proportionofbadedges\]. For $t'\in [t+1,t+\delta n^2]$, let us define events $\mathcal{A}^{t'}_{ST}$, $\mathcal{A}_{SS}^{t'}$ and $\mathcal{A}_{T,T}^{t'}$ as follows. $$\mathcal{A}^{t'}_{ST}=\left\{X_{ST}(t')\in (2p(1-p)-2\varepsilon_7, 2p(1-p)+2\varepsilon_7)N_{ST}(t')\right\}.$$ $$\mathcal{A}^{t'}_{SS}=\left\{X_{SS}(t')\in (2p(1-p)-2\varepsilon_7, 2p(1-p)+2\varepsilon_7)N_{SS}(t')\right\}.$$ $$\mathcal{A}_{TT}^{t'}=\left\{X_{TT}(t')\in (2p(1-p)-2\varepsilon_7, 2p(1-p)+2\varepsilon_7)N_{TT}(t')\right\}.$$ Finally, let us define $$\mathcal{A}^{t'}= \mathcal{A}_{SS}^{t'}\cap \mathcal{A}_{ST}^{t'}\cap \mathcal{A}_{TT}^{t'}.$$ Let $Z_{t'}$ be the indicator of $\overline{\mathcal{A}^{t'}}$.
\[l:propcount\] Let $\mathcal{G}$ denote the conditioning $\mathcal{G}=\{{\mathcal{F}}_t, t<\tau_0, N_1(t)=pn\}$. Then we have
$${\mathbb{P}}\left(\frac{1}{\delta n^2}\sum_{t'=t+1}^{t+\delta n^2} Z_{t'}\geq \varepsilon_{15}, t+\delta n^2<\tau_0', D^*(t+\delta n^2) \mid \mathcal{G}\right)\leq e^{-h(\beta)n}$$
where $h(\beta)$ can be made arbitrarily large by taking $\beta$ sufficiently large.
For $i=1,2,\ldots ,\frac{Cn^2}{\beta}$, and for $k=1,2, \ldots \frac{\delta \beta}{C}-1$ let $t_{i,k}=t+i+k\frac{Cn^2}{\beta}$. Let
$$W_i=\sum_{k=1}^{\frac{\delta \beta}{C}-1}Z_{t_{i,k}}.$$
From Proposition \[p:proportionofbadedges\] and Proposition \[p:proportionofbadedgesss\] it follows that for a fixed $i$, ${\mathbb{P}}[Z_{t_{i,k+1}}=1\mid {\mathcal{F}}_{t_{i,k}}, t_{i,k}<\tau'_0, D^*(t_{i,k})]\leq e^{-\gamma n/5}$. Using a Chernoff bound it follows that
$$\begin{aligned}
\label{e:timeprop1}
{\mathbb{P}}\left[\frac{C}{\delta\beta} W_{i} \geq \varepsilon_{15}/2, t+\delta n^2\leq \tau'_0, D^{*}(t+\delta n^2) \mid \mathcal{G}\right] &\leq & \exp(-\varepsilon_{15}\delta\beta\log({\frac{\varepsilon_{15}e^{-\gamma n/5}}{2}})/4C)\nonumber \\
&\leq & (\frac{2}{\varepsilon_{15}})^{1/4C}\exp(-\varepsilon_{15}\delta\beta\eta n/20C).\end{aligned}$$
Taking a union bound over all $i$ and choosing $\beta$ sufficiently large so that $\frac{C}{\delta\beta}\leq \varepsilon_{15}/2$ it follows that
$${\mathbb{P}}\left(\frac{1}{\delta n^2}\sum_{t'=t+1}^{t+\delta n^2} Z_{t'}\geq \varepsilon_{15}, t+\delta n^2<\tau' \mid {\mathcal{F}}_t \right)\leq \frac{Cn^{2}}{\beta}(\frac{2}{\varepsilon_{15}})^{1/4C}\exp(-\varepsilon_{15}\delta\beta\gamma n/20C).$$
Taking $\beta$ sufficiently large completes the proof of the lemma.
\[p:bigcut\] Let $\mathcal{G}$ denote the conditioning $\mathcal{G}=\{{\mathcal{F}}_t, t<\tau_0,N_1(t)=pn\}$. Let $S,T$ be as above. Then we have, $${\mathbb{P}}(K_{ST}(t+\delta n^2)> \varepsilon_3^2, t+\delta n^2<\tau'_0, D^*(t+\delta n^2) \mid \mathcal{G})\leq \exp (-h(\beta)n)$$
where $h(\beta)$ can be made arbitrarily large by taking $\beta$ sufficiently large.
For $s\in [t,t+\delta n^2-1]$, let ${\mathcal{F}}_{s}$ denote the filtration generated by the process up to time $s$. Conditioned on ${\mathcal{F}}_{s}$ the transition rule for the evolution of $(N_{SS}(s), N_{TT}(s))$ is given by the following.
$$(N_{SS}(s+1), N_{TT}(s+1))=
\begin{cases}
(N_{SS}(s)+1, N_{TT}(s))~~w.p. ~~ \frac{X_{ST}(s)}{N(s)}\frac{1}{2}\frac{|S|-1}{n-1}(1-\frac{\beta}{n}) \\
(N_{SS}(s), N_{TT}(s)+1)~~ w.p. ~~ \frac{X_{ST}(s)}{N(s)}\frac{1}{2}\frac{|T|-1}{n-1}(1-\frac{\beta}{n}) \\
(N_{SS}(s)-1, N_{TT}(s))~~ w.p.~~ \frac{X_{SS}(s)}{N(s)}\frac{|T|}{n-1}(1-\frac{\beta}{n}) \\
(N_{SS}(s), N_{TT}(s)-1)~~ w.p.~~ \frac{X_{TT}(s)}{N(s)}\frac{|S|}{n-1}(1-\frac{\beta}{n}) \\
(N_{SS}(s), N_{TT}(s)) ~~~~~~ \text{otherwise}.
\end{cases}$$
For this proof, let us write $\Delta=2p(1-p)(1-\frac{\beta}{n})$ and $\varepsilon_{8}=\varepsilon_{7}/2p(1-p)$. For $n$ sufficiently large and on $Z_{s}=0$, we have
$$(N_{SS}(s+1), N_{TT}(s+1))=
\begin{cases}
(N_{SS}(s)+1, N_{TT}(s))~~ w.p. ~~ \in \Delta\frac{N_{ST}(s)}{N(s)}\frac{1}{2}\frac{|S|}{n}(1\pm 3\varepsilon_8) \\
(N_{SS}(s), N_{TT}(s)+1)~~ w.p. ~~ \in \Delta\frac{N_{ST}(s)}{N(s)}\frac{1}{2}\frac{|T|}{n}(1\pm 3\varepsilon_8) \\
(N_{SS}(s)-1, N_{TT}(s)) ~~ w.p. ~~ \in \Delta\frac{N_{SS}(s)}{N(s)}\frac{|T|}{n}(1\pm 3\varepsilon_8) \\
(N_{SS}(s), N_{TT}(s)-1)~~ w.p. ~~ \in \Delta\frac{N_{TT}(s)}{N(s)}\frac{|S|}{n}(1\pm 3\varepsilon_8) \\
\end{cases}$$
Doing a change of variable $W_{S}(s)=\frac{N_{SS}(s)-|S|^2/4}{N}$ and $W_{T}(s)=\frac{N_{TT}(s)-|T|^2/4}{N}$ it follows that on $\{Z_{s}=0\}$, we have $$(W_{S}(s+1), W_{T}(s+1))=
\begin{cases}
(W_{S}(s)+1/N, W_{T}(s))~~ w.p. ~~ \in\Delta\left(1-W_S(s)-W_{T}(s)-\frac{|S|^2+|T|^2}{4N}\right)\frac{|S|}{2n}\pm 3\varepsilon_7 \\
(W_{S}(s), W_{T}(s)+1/N)~~ w.p. ~~ \in\Delta\left(1-W_S(s)-W_{T}(s)-\frac{|S|^2+|T|^2}{4N}\right)\frac{|T|}{2n}\pm 3\varepsilon_7 \\
(W_{S}(s)-1/N, W_{T}(s)) ~~ w.p. ~~ \in \Delta\left(W_{S}(s)+\frac{|S|^2}{4N}\right)\frac{|T|}{n}\pm 3\varepsilon_7 \\
(W_{S}(s), W_{T}(s)-1/N)~~ w.p. ~~ \in \Delta\left(W_T(s)+\frac{|T|^2}{4N}\right)\frac{|S|}{n} \pm 3\varepsilon_7 \\
\end{cases}$$ It follows that on $\{Z_s=0\}$,
$$\begin{aligned}
{\mathbb{E}}[W_S(s+1)^2+W_{T}(s+1)^2\mid {\mathcal{F}}_s] &\leq & W_S(s)^2+W_{T}(s)^2 \\
&+&\frac{2W_S(s)\Delta}{N}\left[\frac{|S|}{2n}-\frac{|S|n}{8N}-W_S(s)(\frac{|S|+2|T|}{2n})-W_T(s)\frac{|S|}{2n}\right]\\
&+& \frac{2W_T(s)\Delta}{N}\left[\frac{|T|}{2n}-\frac{|T|n}{8N}-W_T(s)(\frac{2|S|+|T|}{2n})-W_T(s)\frac{|T|}{2n}\right]\\
&+&\frac{25\varepsilon_7}{n^2}+o(\frac{1}{n^2})\\
&\leq & W_S(s)^2+W_{T}(s)^2\\
&-& \frac{4\Delta}{n^2}\left[\frac{1}{4}(W_S(s)^2+W_{T}(s)^2)+\frac{1}{4}(W_s(s)+W_T(s))^2\right]\\
&-& \frac{4\Delta}{n^2}\left[\frac{|T|}{2n}W_S(s)^2+\frac{|S|}{2n}W_T(s)^2\right] +\frac{32\varepsilon_7}{n^2}\\
&\leq & W_S(s)^2+W_{T}(s)^2-\frac{\Delta}{n^2}\left[(W_S(s)^2+W_{T}(s)^2)-32\varepsilon_8\right]\end{aligned}$$
Hence, we have
$${\mathbb{E}}[K_{ST}(s+1)-K_{ST}(s)\mid {\mathcal{F}}_s, Z_{s}=0]\leq -\frac{\Delta}{n^2}(K_{ST}(s)-32\varepsilon_{8}).$$
In particular, on $\{Z_{s}=0\}\cap \{K_{ST}(s)\geq \varepsilon_3^2/2\}$, we have $${\mathbb{E}}[K_{ST}(s+1)-K_{ST}(s)\mid {\mathcal{F}}_s]\leq -\frac{\Delta}{4n^2}\varepsilon_3^2$$by choosing $\varepsilon_3^2\geq 128\varepsilon_8$.
Now notice that $K_{ST}(s+1)-K_{ST}(s)\leq \frac{16}{n^2}$. Let $\mathcal{C}$ denote the event $$\mathcal{C}=\{\min_{s\in [t,t+\delta n^2]}K_{ST}\leq \varepsilon_{3}^2/2\}.$$ Let $\mathcal{D}$ denote the event $$\mathcal{D}=\left\{\sum_{s=t}^{t+\delta n^2-1}{\mathbb{E}}[K_{ST}(s+1)-K_{ST}(s)\mid {\mathcal{F}}_s]>\delta(16\varepsilon_{15}-(1-\varepsilon_{15})\Delta\varepsilon_{3}^2/4)\right\}.$$ It follows from Lemma \[l:propcount\] that $${\mathbb{P}}[\mathcal{C}^c, \mathcal{D}, D^*(t+\delta n^2), t+\delta n^2 <\tau'_0\mid \mathcal{G}]\leq \exp (-h(\beta)n)$$
where $h(\beta)$ can be made arbitrarily large by choosing $\beta$ sufficiently large.
Now an application of Azuma-Hoeffding inequality gives $$\label{e:azuma7}
{\mathbb{P}}\left[\sum_{s=t}^{t+\delta n^2-1} K_{ST}(s+1)-K_{ST}(s)-{\mathbb{E}}(K_{ST}(s+1)-K_{ST}(s)\mid {\mathcal{F}}_s)\geq \frac{\delta\varepsilon_{3}^2}{10}\mid \mathcal{G}\right]\leq \exp (-\frac{\delta\varepsilon_{3}^4n^2}{204800}).$$
It follows that if $\varepsilon_{15}$ is chosen sufficiently small so that $\varepsilon \varepsilon_3^{2}> \frac{64\varepsilon_{15}}{1-\varepsilon_{15}}$, then we have,
$${\mathbb{P}}[K_{ST}(t+\delta n^2)> K_{ST}(t), \mathcal{C}^c, D^*(t+\delta n^2) t+\delta n^2< \tau'_0\mid \mathcal{G}]\leq \exp (-h(\beta)n).$$
Observe that by choosing $\delta$ sufficiently small ($16 \delta < \varepsilon_3^2$), it follows that on $\mathcal{C}$, $K_{ST}(t+\delta n^2)< \varepsilon_3^2$. Hence
$${\mathbb{P}}[K_{ST}(t+\delta n^2)> \varepsilon_{3}^2, D^*(t+\delta n^2), t+\delta n^2< \tau'_0\mid \mathcal{G}]\leq \exp (-h(\beta)n).$$
This completes the proof of the proposition.
Now we are ready to prove the main result of this subsection.
\[t:bigcut\] We have for all $t\geq \delta n^2$, $${\mathbb{P}}[t+\delta n^2\geq \tau_2, t+\delta n^2<\tau_*, t<\tau_0]\leq \frac{1}{n^{14}}.$$
Let $\tilde{D}(t+\delta n^2)$ be defined as follows.
$$\tilde{D}(t+\delta n^2)=\{\forall t'\in [t+1, t+\delta n^2]: |N_1(t')-N_1(t)|\leq \varepsilon_{7}n\}.$$
By taking a union bound over all cuts $S,T$ such that $\varepsilon_2 n\leq |S|\leq |T|$ and using Proposition \[p:bigcut\] we get that
$${\mathbb{P}}[L(t+\delta n^2)\geq \varepsilon_3^{2}, \tilde{D}(t+\delta n^2), t+\delta n^2 <\tau_0' \mid {\mathcal{F}}_t, t<\tau_0]\leq 2^n\exp(-h(\beta)n)\leq \frac{1}{n^{18}}$$ by taking $\beta$ sufficiently large. It follows by a random walk estimate that ${\mathbb{P}}[\tilde{D}(t+\delta n^2)\mid {\mathcal{F}}_t, t<\tau_0]$ is exponentially close to $1$ and hence we have,
$${\mathbb{P}}[L(t+\delta n^2)\geq \varepsilon_{3}^2, t+\delta n^2<\tau'_0\mid t<\tau, {\mathcal{F}}_t]\leq \frac{2}{n^{18}}.$$
By Theorem \[t:weakbound\], we know that ${\mathbb{P}}[t+\delta n^2\geq \tau'_0, t+\delta n^2< \tau_{*}\mid {\mathcal{F}}_t, t<\tau_0]\leq \frac{1}{n^{17}}$ and hence
${\mathbb{P}}[L(t+\delta n^2)\geq \varepsilon_3^2, t+\delta n^2< \tau_*\mid {\mathcal{F}}_t , t<\tau_0]\leq \frac{2}{n^{17}}$.
Now as $\{t<\tau\}\subseteq \{t-s<\tau\}$ and $\{t+\delta n^2 <\tau_{*}\}\subseteq \{t+\delta n^2-s <\tau_{*}\}$ for each $s\geq 0$, by taking a union over $s\in [0,\delta n^2 -1]$ we get that
$${\mathbb{P}}[\tau_2\leq t+\delta n^2, \tau_{*}< t+\delta n^2, t<\tau_0]\leq \frac{1}{n^{14}}.$$ This completes the proof of the theorem.
Edge Multiplicity
-----------------
In this subsection we consider the strong stopping time associated with edge multiplicities, i.e., $\tau_3$. We need the following lemma.
\[l:probagreeing\] Let $u,v$ be two fixed vertices in $V$. Let $X_{uv}(t')$ be the indicator that $u$ and $v$ are disagreeing in $G(t')$. Then we have, ${\mathbb{E}}(X_{uv}(t+\frac{Cn^2}{\beta})\mid {\mathcal{F}}_{t}, t<\tau'_0)\geq \frac{\varepsilon}{2}$.
To prove this consider the coupling of the evolving voter model with independent continuous time random walk started from $u$ and $v$ as described in § \[s:coupling\]. Notice that the chance that the random walks intersect upto time $\frac{3C}{\beta}$ is $o(1)$ as $n\rightarrow \infty$. Also notice that the chance that either of the walk traverses any edge that was rewired can be made less than $\frac{\varepsilon}{100}$ by choosing $\beta$ sufficiently large. Let $Y_{uv}$ denote the indicator that the positions of the random walks started from $u$ and $v$ after time $T$ are disagreeing in $G(t)$. Clearly $P(X_{uv}(t+\frac{Cn^2}{\beta})\neq Y_{uv})\leq \varepsilon/4$ using Lemma \[l:couplingbasic1\]. Also let $Y^*_{uv}$ be the indicator that the position of random walks started from $u$ and $v$ after time $\frac{C-\varepsilon/100}{\beta}$ are disagreeing in $G(t)$. Using Lemma \[l:couplingbasic2\] it follows that ${\mathbb{P}}(Y_{uv}\neq Y^*_{uv})\leq \varepsilon/4$. The result follows by noticing that Lemma \[l:couplingtv\] implies that for $C$ sufficiently large ${\mathbb{E}}(Y^*_{uv})\geq \varepsilon$.
\[l:multbadprounion\] Let $u,v\in V$ be two vertices in $V$. Fix $t>\varepsilon_{16}n^2\log n$. For $\varepsilon_{16}n^2\log n<t'<t$, let $A_{t'}$ denote the event that there exists $\mathcal{T}\in \{1,2,\ldots, t-t'\}$ such that $$\#\{s\in \{\mathcal{T}, \mathcal{T}+1, \ldots , \mathcal{T}+t'\}:u(s)\neq v(s)\}\leq t'\varepsilon/4.$$ Then we have $${\mathbb{P}}\left[\cup A_{t'}, t<n^4\wedge \tau'_0\right]\leq \frac{1}{n^{r(\beta)}}$$ where $r(\beta)$ can be made arbitrarily large by taking $\beta$ sufficiently large.
Fix $\varepsilon_{16}n^2\log n<t'<t\wedge n^4$ and $\mathcal{T}\in \{1,2,\ldots t-t'\}$. For $t''\in \{\mathcal{T}, \mathcal{T}+1,\ldots, \mathcal{T}+t'\}$ it follows from Lemma \[l:probagreeing\] that on $\{t''<\tau'_0\}$, ${\mathbb{P}}[u(t''+Cn^2/\beta)\neq v(t''+Cn^2/\beta)\mid {\mathcal{F}}_{t''}]\geq \varepsilon/2$ for $\beta$ sufficiently large. It follows using a Chernoff’s bound that for each $i=1,2,\ldots, Cn^2/\beta$, $${\mathbb{P}}\left[\#\{t''\in \{\mathcal{T}+i+kCn^2/\beta: k\in [\beta t'/n^2C]\}: u(t'')\neq v(t'')\}\leq \frac{\beta t'\varepsilon}{4Cn^2}, t<\tau'_0\right]\leq \exp (-\frac{\beta t' \varepsilon}{12Cn^2}).$$ For all $t'>\varepsilon_{16}n^2\log n$, we have the right hand side of the above inequality is at most $(\frac{1}{n})^{\beta \varepsilon_{16}\varepsilon/12C}$ and it follows by taking a union bound over all $i\in [Cn^2/\beta]$ and all $\mathcal{T}\in \{1,2,\ldots t-t'\}$ that $${\mathbb{P}}[A_{t'},t<n^4\wedge \tau'_0]\leq \frac{1}{n^{r'(\beta)}}$$ where $r'(\beta)$ can be made sufficiently large by choosing $\beta$ to be sufficiently large. The lemma now follows by taking union bound over $\varepsilon_{16}n^2\log n<t'<t\wedge n^4$.
Now we define the following family of random walks which we couple with the *rewire-to-random-\** dynamics as follows, $X^{s}(\cdot)$ indexed by $s\in \{1,2,\ldots, t\}$ with each starting from $K>0$ (i.e., $X^{s}(0)=K~ \forall s$) with transition probabilities as described below.
$$X^s(h+1)=
\begin{cases}
X^{s}(h)+1 ~~ w.p. ~~ \frac{9C_2}{n^2} \\
X^{s}(h)-1 ~~ w.p. ~~ \frac{K}{n^2}~\text{if}~O_u(s+h)\neq O_v(s+h) \\
X^{s}(h) ~~~ ~otherwise.
\end{cases}$$
The following lemma is immediate by comparing one step transition probabilities of $M_{uv}(t)$ and $X^{s}(t-s)$.
\[l:multstocdom\] Let $M_{uv}^*(t)=\max_{t'\in[1,t]}M_{uv}(t')$ and $X^{*}(t)=\max_{s,h: s+h\leq t} X^{s}(t)$. Then we have, on $\{t<\tau'_0\}$, $M_{uv}^{*}(t)\preceq X^{*}(t)$ where $\preceq$ denotes stochastic domination.
From the previous lemma, we deduce the following.
\[l:multmart\] We have ${\mathbb{P}}[M_{uv}^*(t)>\varepsilon_4 \log n , t<\tau'_0\wedge n^4]\leq \frac{1}{n^{10}}$.
By Lemma \[l:multstocdom\] it suffices to prove the inequality in the statement with $M_{uv}^*(t)$ replaced by $X^{*}(t)$. Let $\mathcal{C}$ denote the following event. $$\mathcal{C}=\{\forall T\in [1,t], t'> \varepsilon_{16}n^2\log n~ \#\{s\in [T,T+t']:
u(s)\neq v(s)\}\geq t'\varepsilon/4\}.$$
Then we have that for all $t-s>t'>\varepsilon_{16}n^2\log n$,
$${\mathbb{E}}(e^{\lambda X^{s}(t')}1_{\mathcal{C}\cap \{t<\tau'_0\}})\leq e^{\lambda K}\left(1+\frac{9C_2}{n^2}(e^{\lambda}-1)\right)^{t'(1-\varepsilon/4)}\left(1+\frac{9C_2}{n^2}(e^{\lambda}-1)+\frac{K}{n^2}(1-e^{-\lambda})\right)^{t'\varepsilon/4}.$$
Fix $\lambda$ large enough such that $\lambda\varepsilon_{4}>20$. Choosing $K$ sufficiently large depending on $\lambda$ and $\varepsilon$, it follows that
$$(1+\frac{9C_2}{n^2}(e^{\lambda}-1))^{1-\varepsilon/4}(1+\frac{9C_2}{n^2}(e^{\lambda}-1)+\frac{K}{n^2}(1-e^{-\lambda}))^{\varepsilon/4}< 1$$ and hence
$${\mathbb{E}}(e^{\lambda X^{s}(t')} 1_{\mathcal{C}\cap \{t<\tau'_0\}})\leq e^{\lambda K}.$$
By Markov’s inequality it follows that
$${\mathbb{P}}(\{X^{s}_{t'}>\varepsilon_4\log n\}\cap \mathcal{C}\cap \{t<\tau'_0\})\leq e^{-\lambda(K-\varepsilon_4\log n)}\leq (\frac{1}{n})^{19}.$$
For $t'<\varepsilon_{16}n^2\log n$, $X^{s}(t')-K$ is stochastically dominated by a $\mbox{Bin}(\varepsilon_{16}n^2\log n, \frac{9C_2}{n^2})$ variable. Using a Chernoff bound in this case, we get for $n$ sufficiently large $${\mathbb{P}}[X^s(t')\geq \varepsilon_4\log n]\leq e^{-\varepsilon_4\log n\log(\varepsilon_{4}/9C_2\varepsilon_{16})/2} \leq (\frac{1}{n})^{19}$$ by choosing $\varepsilon_{16}$ sufficiently small such that $\varepsilon_4\log(\varepsilon_{4}/9C_2\varepsilon_{16})>38.$
By taking a union bound over all $s,t'$ it follows that
$${\mathbb{P}}[X^{*}(t)>\varepsilon_4 \log n, \mathcal{C}, \{t<\tau'_0\}]\leq (\frac{1}{n})^{11}.$$
The result now follows from Lemma \[l:multbadprounion\].
\[t:indmultbound\] Let $t<n^4$. Then ${\mathbb{P}}[t\geq \tau_3, t< \tau'_0]\leq \frac{1}{n^4}$.
The theorem follows from using Lemma \[l:multmart\], taking a union bound over all $(u,v) \in V^{(2)}$.
Degree Estimate
---------------
In this section we prove the following theorem.
\[t:degsbound\] We have for all $t\geq \delta n^2$, ${\mathbb{P}}[t+\delta n^2\geq \tau_5, t+\delta n^2<\tau_{*}, t<\tau]\leq \frac{1}{n^{14}}$.
We start with the following lemma.
\[l:degcoupling\] Let $\kappa_2>0$ be fixed. Let us condition on $\{{\mathcal{F}}_{t_1}, t_1<\tau', N_{*}(t)=pn\}$. Let $v$ be a fixed vertex in $V$. Let $X_v(t')$ denote the number of disagreeing edges incident to $v$ at time $t'$. Then for sufficiently large $C$ and sufficiently large $\beta=\beta(C)$, at time $t_2=t_1+\frac{Cn^2}{\beta}$ we have ${\mathbb{P}}[X_v(t_2)\notin (p(1-\kappa_2), 1-p(1-\kappa_2))D_v(t_2)\mid {\mathcal{F}}_{t_1}, \tau'_0>t_1]\leq e^{-\varepsilon_{20}n}$ for some constant $\varepsilon_{20}>0$.
The proof of this lemma goes along the same lines as that of Proposition \[p:proportionofbadedges\]. We shall therefore only give the sketch of the steps. Let the edges incident to $v$ at time $t_1$ be $\{e_1,e_2,\ldots, e_{D_v(t_1)}\}$. Let $Y_i$ denote the indicator that the endpoints of $e_i$ are disagreeing in $G(t_2)$. It follows by taking $\beta$ sufficiently large that it suffices to prove that
$$\frac{1}{D_v(t_1)}\sum_{i=1}^{D_v(t)} Y_i \in (p(1-\kappa_2/2),1-p(1-\kappa_2/2))$$ with exponentially high probability. To this end, we choose a subset of these edges of size $\varepsilon_{12}n$. Condition on a subset $J$ of these edges of size at least $\varepsilon_{13}n=(1-\frac{\kappa_2}{100})\varepsilon_{12}n$ which correspond to distinct bonds in $V^{(2)}$. Arguing as in the proof of Proposition \[p:proportionofbadedges\], we see that it suffices to show that, conditionally, $$\frac{1}{\varepsilon_{13}n}\sum_{j\in J} Y_j \in (p(1-\kappa_2/50),1-p(1-\kappa_2/50))$$ with exponentially high probability.
Consider the coupling described in § \[s:coupling\] of the *rewire-to-random-\** dynamics started with $G(t_1)$ with independent random walks started from $v_j$ where $e_j$ is placed in the bond $(v,v_j)$ is $G(t_1)$. Let for $j\in J$ and $\tilde{\sigma}>\frac{C}{\beta}$, $U^0_{j,\tilde{\sigma}}$ and $U^1_{j,\tilde{\sigma}}$ denote the indicators that the position of the random walk started from $v_j$ has the opinion $0$ and opinion $1$ respectively at time $\tilde{\sigma}$. It follows by taking $C$ sufficiently large that ${\mathbb{E}}(U^0_{j,\tilde{\sigma}}), {\mathbb{E}}(U^1_{j,\tilde{\sigma}})\in (p(1-\kappa_2/100),1-p(1-\kappa_2/100))$. Now the proof is completed arguing as in the proof of Proposition \[p:proportionofbadedges\].
\[l:deginterval\] With the notation as in Lemma \[l:degcoupling\], let $A_t$ denote the event that for some $t'$ with $t'\in [t+\frac{Cn^2}{\beta}, t+\delta n^2]$, $X_v(t') \notin (p-\varepsilon/8, 1-(p-\varepsilon/8))D_v(t')$. Then we have, $${\mathbb{P}}(A_t, t+\delta n^2< \tau'_0\mid {\mathcal{F}}_t, t<\tau'_0, N_*(t)=pn)\leq e^{-\varepsilon_{20}n/2}.$$
This follows from a union bound and the previous lemma.
\[l:degmartingale\] We have ${\mathbb{P}}[D_v(t+\delta n^2)\notin (\varepsilon/2, 1-\varepsilon/2)n, t+\delta n^2<\tau'_0\mid {\mathcal{F}}_t, t<\tau_0]\leq e^{-\varepsilon_{21}n}$.
Let $C_t$ denote the event that for all $t'\in [t+1,t+\delta n^2]$, $|N_*(t')-N_*(t)|\leq \varepsilon n/16$. Let $H_t=A_t^{c}\cap C_t \cap \{t+\delta n^2 <\tau'_0\}$. Notice that on $H_t$, we have at time $t'\in [t,t+\delta n^2]$, the number of disagreeing edge at time $t$, $Z({t'})\geq (p(1-p)-\varepsilon/8)n^2$. Also notice that on $A_t$, the number of disagreeing edges incident to $v$ is at time $t'$ is in $[p(1-\varepsilon/8),1-p(1-\varepsilon/8)]D_v(t)$. Set $X(t')=D_v(t')-(1-3\varepsilon/4)n$, also set $\Delta^*=(1-\frac{\beta}{n})$. It follows therefore that for $\lambda>0$ $$\begin{aligned}
\label{e:degmartcalc}
{\mathbb{E}}(e^{\lambda X_{t'+1}}1_{H_t}\mid {\mathcal{F}}_{t'})&\leq & e^{\lambda X_{t'}}\left(1+(e^{\lambda}-1)\frac{\Delta^* Z(t')}{N(n-1)}+(e^{-\lambda}-1)\frac{\Delta^*(p-\varepsilon/8)(X_{t'}+(1-3\varepsilon/4)n)}{2N}\right)\nonumber
$$ Now take $\lambda$ so small such that $e^{\lambda}-1\leq (1+\varepsilon/100)\lambda$ and $e^{-\lambda}-1\leq -(1-\varepsilon/100)\lambda$. Then we have
$$\begin{aligned}
\label{e:degmartcalc}
{\mathbb{E}}(e^{\lambda X_{t'+1}}1_{H_t}\mid {\mathcal{F}}_{t'})&\leq & e^{\lambda X_{t'}}\left(1+ \lambda\Delta^*\left(\frac{(1+\varepsilon/50)Z_{t'}}{nN}-\frac{(p-\varepsilon/6) X_{t'}}{2N}-\frac{(p-\varepsilon/6)(1-3\varepsilon/4)n}{2N}\right) \right)\nonumber\\
& \leq & e^{\lambda X_{t'}}\left(1-\lambda\frac{\varepsilon X_{t'}}{8N}\right)\nonumber\\
&\leq & e^{\lambda_{*}X_{t'}}\end{aligned}$$
where $0<\lambda_{*}<\lambda(1-\frac{\varepsilon}{10N})$ and since $p>\varepsilon$ implies that on $H_t$ we have $\frac{(1+\varepsilon/50)Z_{t'}}{nN}<\frac{(p-\varepsilon/6)(1-3\varepsilon/4)n}{2N}$. It follows that there exist $\lambda_0$ bounded away from $\lambda$ such that
$${\mathbb{E}}(e^{\lambda X_{t+\delta n^2}}1_{H_t}\mid {\mathcal{F}}_{t+Cn^2/\beta})\leq e^{\lambda_0 X_{t+Cn^2/\beta}}1_{H_t}\leq e^{\lambda_0\varepsilon n/4},$$
$$\text{i.e.},~{\mathbb{E}}(e^{\lambda X_{t+\delta n^2}}1_{H_t}\mid {\mathcal{F}}_{t}, t<\tau_0)\leq e^{\lambda_0\varepsilon n/4}.$$ By Markov’s inequality it now follows that
$${\mathbb{P}}[D_v(t+\delta n^2)\geq (1-\varepsilon/2)n, H_t\mid {\mathcal{F}}_t, t<\tau_0]\leq e^{-(\lambda-\lambda_0)\varepsilon n/4}.$$ It follows from Lemma \[l:deginterval\] and another random walk estimate that ${\mathbb{P}}(H_t^c, t+\delta n^2<\tau'_0)$ is exponentially small in $n$, which completes the proof of one side of the bound in this lemma.
The other side of the bound can be proved similarly by starting with $\lambda$ negative and $X(t)=D_v(t)-3\varepsilon/4$. This completes the proof of the lemma.
This theorem follows from Lemma \[l:degmartingale\] by taking a union bound over all vertices $v$, and all times $t'\in [t-\delta n^2, t]$ as in the proof of Theorem \[t:bigcut\], and using Theorem \[t:weakbound\].
Multiple-Edge Estimates
-----------------------
\[t:multedgesbound\] We have for all $t\geq \delta n^2$, ${\mathbb{P}}[t+\delta n^2\geq \tau_4, t+\delta n^2< \tau_*, t<\tau]\leq \frac{1}{n^{4}}$.
The above theorem follows from the following lemma.
\[l:kedge\] Condition on $\{{\mathcal{F}}_t, t<\tau_0\}$. Let $v$ be a vertex in $V$. Let $X_{v,k}(t)$ denote the number of vertices $u\in G$ such that $M_{uv}(t)=k$. Then for each $1\leq k\leq 2\varepsilon_{4}\log n$, we have $${\mathbb{P}}[X_{v,k}(t+\delta n^2)> C_110^{-k}n, t+\delta n^2<\tau'_0\mid {\mathcal{F}}_t, t<\tau]\leq \frac{1}{n^{10}}.$$
The proof of the above lemma depends upon the next two lemmas.
\[l:kedge1\] Fix $v$ and $k$ as in the above lemma. Let $Y_{k,r}$ denote the number of $r$-edges that became $k$ edges in time $[t,t+\delta n^2]$. Fix an integer $L_0\geq 1000C_2\delta \vee 1$. Then we have for each $k>3L_0$ and each $r<k-L_0$ we have ${\mathbb{P}}(Y_{k,r}\geq 3^{-(k-r)}C10^{-k}n, t+\delta n^2<\tau'_0\mid {\mathcal{F}}_t, t<\tau_0)\leq \frac{1}{n^{20}}$.
This proof is similar to the proof of Lemma \[l:lbw4\] and we give only a sketch. Condition on $\{{\mathcal{F}}_t, t<\tau_0\}$. Let $u_1,u_2,\ldots, u_{D}$ be the vertices in $V$ such that $\{M_{vu_i}(t)=r\}$. Now without loss of generality we assume $D=C10^{-r}n$. Let $T_i$ be the number of times a rewired edges rewires along $vu_i$. It is clear that on $\{t+\delta n^2<\tau'_0\}$, $T_i$ is stochastically dominated by a $\mbox{Bin}(\delta n^2, \frac{10C_2}{n^2})$. Let $Z_i$ denote the indicator that $T_i\geq (k-r)$. Using a joint stochastic domination argument as in the proof of Lemma \[l:lbw4\] it is easy to show that upto terms much smaller than $n^{-100}$, ${\mathbb{P}}[\sum_{i}Z_{i}\geq (30)^{-(k-r)}D]$ can be approximated by ${\mathbb{P}}[\sum_{i}Z'_{i}\geq (30)^{-(k-r)}D]$ where $Z'_i$ are i.i.d. $\mbox{Bin}(\delta n^2, \frac{20C_2}{n^2})$. For our choice of $L_0$ it follows that for all $\ell\geq L_0$, ${\mathbb{P}}(\mbox{Bin}(\delta n^2, \frac{10C_2}{n^2})\geq \ell)\leq 40^{-\ell}$. Now using another Chernoff bound gives us that on $\{t+\delta n^2<\tau'_0\}$ $$P[Y_{k,r}\geq (30)^{-(k-r)}C_110^{-r}n]\leq \frac{1}{n^{20}}.$$ This completes the proof of the lemma.
\[l:kedge2\] Let $L_0$ be chosen as in Lemma \[l:kedge1\] and $k>\frac{10^6L_0}{\varepsilon\delta}$. Let $k-L_0\leq r\leq k$ be fixed. Let $u_1,u_2,\ldots, u_{D}$ be the vertices in $G$ such that $\{M_{vu_i}(t)=r\}$. Let $R_i$ be the number of edges lost by $vu_i$ in time $[t,t+\delta n^2]$. Let $Z'_i=1_{\{R_i\leq L_0\}}$. Then $${\mathbb{P}}[\sum_{i=1}^{D}Z'_i\geq 20^{-L_0}C10^{-r}n]\leq \frac{1}{n^{20}}.$$
For the proof of this lemma we shall consider the continuous time *rewire-to-random-\** dynamics. We need the following proposition.
\[p:kedge2continuousdisagreeing\] Let us condition on $\{{\mathcal{F}}_t, t<\tau_0\}$. Let $v$ be a fixed vertex in $V$. Let $u_1,u_2,\ldots , u_{D}$ be the vertices in $v$ such that we have $M_{vu_i}(t)=r$, where $1<r<2\varepsilon_4\log n$. Set $G(t)=H(0)$ and Run the following continuous time *rewire-to-random-\** process $H(\cdot)$ from time $0$ to $\delta/2$. Each directed edge rings at rate $1$. If the endpoints of the edge are agreeing in the current graph, no change occurs. If they are diasgreeing then we do a voter model step with probability $\frac{\beta}{n}$ and a rewire-to-random step with probability $(1-\frac{\beta}{n})$. Let $Z_i$ be the indicator that $(v,u_i)$ is disagreeing for less that $\frac{\varepsilon\delta}{200}$ time in $H(\cdot)$. Then we have $P(\sum_{i=1}^{D} Z_i > 25^{-L_0}D, \tau'_0>\delta/2 )\leq e^{-\gamma\sqrt{n}}$ for some $\gamma>0$.
Without loss of generality, we assume $D=C_110^{-r}n >> \sqrt{n}$. Let us choose a random subset $D^*\subseteq {1,2,\ldots, D}$ with $|D^*|=\sqrt{n}$. It therefore suffices to prove that ${\mathbb{P}}(\sum_{i\in D^*} Z_i> 30^{-L_0}\sqrt{n}\mid D^*)\leq e^{-\gamma\sqrt{n}}$. This fact is established by Lemma \[l:continuousdisagreeingkedgesample\] and Lemma \[l:kedgecontdependence\] below which completes the proof.
\[l:continuousdisagreeingkedgesample\] Condition on $\{{\mathcal{F}}_t, t<\tau_0\}$. Let us set $G(t)=H(0)$, and consider running the continuous time *rewire-to-random-\** dynamics $H(\cdot)$ from time $0$ to $\delta/2$. Let $v, v_1,v_2,\ldots , v_{\sqrt{n}}$ be fixed vertices in $V$. Let us consider the independent continuous time random walks described in § \[s:coupling\]. Let us $X_i^{j}(\cdot)$ be the random walk started from $v_i$ on $H(2j\frac{C}{\beta})$ run for time $\frac{2C}{\beta}$. Let for $\frac{C}{\beta}\leq s \leq \frac{2C}{\beta}$, $Y_i(2j\frac{C}{\beta}+s)$ is the indicator of the event that the opinion of $X_i^{j}(s)$ in $H(0)$ is different from the opinion of $v$ in $H(s)$. Let $Y_i^*=\int_{0}^{\frac{\delta}{2}}Y_i(s)~ds$. Further, let $Z_i^*$ denote the indicator that $Y_i^*< \frac{\varepsilon\delta}{64}$. Then we have, $${\mathbb{P}}\left[\sum_{i} Z_i^* \geq 40^{-L_0}\sqrt{n}\right]\leq e^{-c\sqrt{n}}.$$
For $k=0,2,\ldots , \frac{C}{\theta}-1$, $j=1,2,\ldots \frac{\delta \beta}{4C}$, let $\chi_v^{j,k}=1$ if $v$ spends majority of its time in the interval $[(2j-1)\frac{C}{\beta}+\frac{k\theta}{\beta},(2j-1)\frac{C}{\beta}+\frac{(k+1)\theta}{\beta}]$ with opinion $1$ and $0$ otherwise. Let $\chi_i^{j,k}=1$ if the opinion of $X_i^{j}(s)=1$ for all $s\in [(2j-1)\frac{C}{\beta}+\frac{k\theta}{\beta},(2j-1)\frac{C}{\beta}+\frac{(k+1)\theta}{\beta}]$, $\chi_i^{j,k}=0$ if the opinion of $X_i^{j}(s)=0$ for all $s\in [(2j-1)\frac{C}{\beta}+\frac{k\theta}{\beta},(2j-1)\frac{C}{\beta}+\frac{(k+1)\theta}{\beta}]$, and $\chi_i^{j,k}=\star$ otherwise. Now let $U_i^{j,k}=1_{\{\chi_i^{j,k}=1, \chi_v^{j,k}=0\}}+1_{\{\chi_i^{j,k}=1, \chi_v^{j,k}=0\}}$. Let us fix $k$. Now choose $\theta$ sufficiently small so that the chance that the random walk takes a step in time $\theta/\beta$ is at most $\frac{\varepsilon}{4}$. Clearly, for a fixed realisation of the sequence $\chi_v^{j,k}$, and on $\{2(j-1)C/\beta<\tau'_0\}$, we have by Lemma \[l:couplingtv\], that ${\mathbb{E}}[U_i^{j,k}\mid {\mathcal{F}}_{2(j-1)C/\beta}]\geq \varepsilon/4$. Since the random walks are independent, it follows that
$${\mathbb{P}}\left[\#\{i: \sum_j U_i^{j,k}\leq \frac{\delta \beta \varepsilon}{8C}\}\geq e^{-\gamma'\beta}\sqrt{n}, \chi_v^{j,k}, \frac{\delta}{2}<\tau'_0\right]\leq e^{-c\sqrt{n}}.$$
Taking a union bound over $2^{\delta\beta/4C}$ possible realisations of the sequence $\chi_v^{k,j}$ (for a fixed $k$), we get that,
$${\mathbb{P}}\left[\#\{i: \sum_j U_i^{j,k}\leq \frac{\delta \beta \varepsilon}{32C}\}\geq e^{-\gamma'\beta}\sqrt{n}, \frac{\delta}{2}<\tau'_0\right]\leq e^{-c\sqrt{n}}$$ for some constant $\gamma'$ and $c>0$.
Now taking a union bound over $k$ we get,
$${\mathbb{P}}\left[\#\{i: \sum_j\sum_{k} U_i^{j,k}\leq \frac{\delta \beta \varepsilon}{32\theta}\}\geq \frac{C}{\theta}e^{-\gamma'\beta}\sqrt{n}, \frac{\delta}{2}<\tau'_0\right]\leq \frac{C}{\theta}e^{-c\sqrt{n}}.$$ Now notice that on $\{\sum_j\sum_{k} U_i^{j,k}>\frac{\delta \beta \varepsilon}{32\theta}\}$, we have $Y_i^*>\frac{\varepsilon\delta}{64}$, and the proof of the lemma is completed by taking $\beta$ sufficiently large.
\[l:kedgecontdependence\] Assume the setting of Lemma \[l:continuousdisagreeingkedgesample\]. Also let $\tilde{Y}_i$ denote the amount of time the bond $(v,v_i)$ is disagreeing in $[0,\frac{\delta}{2}]$. Then there is a coupling of the continuous time evolving voter model with the continuous time random walks started at $v_i$ as described in Lemma \[l:continuousdisagreeingkedgesample\] such that $${\mathbb{P}}[\#\{i: \tilde{Y}_{i}\leq Y_i^*- \frac{\varepsilon\delta}{128}\}\geq 40^{-L_0}\sqrt{n}, \frac{\delta}{2}<\tau'_0]\leq e^{-c\sqrt{n}}$$ for some constant $c$.
Consider the coupling described in § \[s:coupling\], with the obvious modification for the continuous time dynamics. Define $Y_{i,j}=0$ if the opinion of $X_i^j(s)$ is the same as the opinion of $v_i$ in $H(\frac{2jC}{\beta}+s)$ for all $s\in [0, \frac{2C}{\beta}]$ and $Y_{i,j}=1$ otherwise. It follows by arguments similar to those in the proof of Lemma \[l:couplemany\] that $${\mathbb{P}}\left[\sum_{j=1}^{\delta\beta/4C}\sum_{i=1}^{\sqrt{n}}Y_{i,j} \geq 10C\delta\sqrt{\beta}\sqrt{n}\right]\leq e^{-c\sqrt{n}}$$.
Again notice that, $\tilde{Y}_i-Y_i^* \geq (\sum_{j}Y_{i,j})\frac{2C}{\beta}$. It follows that,
$${\mathbb{P}}[\#\{i: \tilde{Y}_{i}\leq Y_i^*- \frac{\varepsilon\delta}{64}\}\geq \frac{2560C^2}{\sqrt{\beta}\varepsilon\delta}\sqrt{n}]\leq e^{-c\sqrt{n}}$$ The proof of the lemma is completed by taking $\beta$ sufficiently large.
\[l:kedge2continuous\] Condition on $\{{\mathcal{F}}_t, t<\tau_0\}$. Assume the setting of Proposition \[p:kedge2continuousdisagreeing\]. Then, Let $W_i^*$ be the indicator that the bond $vu_i$ loses at least $L_0$ edges by time $\frac{\delta}{2}$. For $r> \frac{10^6L_0}{\varepsilon\delta}$, we have $${\mathbb{P}}\left[\sum_{i=1}^{D}W^*_i\geq 20^{-L_0}C_110^{-r}n, \tau'_0>\frac{\delta}{2}\right]\leq \frac{1}{n^{20}}.$$
Without loss of generality assume $D=C_110^{-r}n>> \sqrt{n}$. In the continuous time model, the rate at which a bond with $k$-disagreeing edges lose an edge is $k$, and the rings of the different edges are independent. Let $S_1, S_2,\ldots, S_{D}$ be the number of edges lost by the bonds $vu_1, vu_2,\ldots vu_D$ respectively. Let $S'_i$ be independent $\mbox{Poi}(1000L_0)$ variables. It follows from Proposition \[p:kedge2continuousdisagreeing\] that there exist a coupling such that, $${\mathbb{P}}[\#\{i: S_i\leq S'_i\wedge L_0 \}\geq 25^{-L_0}C_1 10^{-k}n, \tau'_0>\frac{\delta}{2}]\leq e^{-c\sqrt{n}}$$ for some constant $c>0$. Also notice that
$P[\#\{i: S'_i\leq L_0\}\geq 40^{-L_0}D]\leq e^{-c\sqrt{n}}$. The lemma follows.
Now we are ready to prove Lemma \[l:kedge2\].
The proof follows from the obvious coupling of the continuous time [rewire-to-random-\*]{} dynamics with the discrete time [rewire-to-random-\*]{} dynamics and observing that with exponentially high probability, the number of step taken in the discrete time process upto time $\frac{\delta}{2}$ in the continuous time process is less that $\delta n^2$.
We now complete the proof of Lemma \[l:kedge\].
Choose $C_1$ large enough such that the condition is satisfied for all $k<\frac{10^6L_0}{\varepsilon\delta}$ by the degree estimate. Now notice that $$X_{v,k}(t+\delta n^2)\leq \sum_{r\geq k+1} X_{v,r}(t)+\sum_{r=0}^{k-L_0}Y_{r,k}+\sum_{r=k-L_0}^{k} Y_{r,k}.$$ On $\{t<\tau_0\}$ the first sum is $\frac{C_110^{-k}n}{9}$. By Lemma \[l:kedge1\] we have that $${\mathbb{P}}\left[\sum_{r=0}^{k-L_0}Y_{r,k}\geq 3^{-L_0}C10^{-k}n, t+\delta n^2<\tau'_0\right]\leq \frac{1}{n^{19}}.$$ Also using Lemma \[l:kedge1\] and Lemma \[l:kedge2\] it follows that $${\mathbb{P}}\left[\sum_{r=k-L_0}^{k}Y_{r,k}\geq 2^{-L_0}C10^{-k}n, t+\delta n^2<\tau'_0\right]\leq \frac{1}{n^{19}}.$$ Putting together all these gives us the statement of the lemma.
For each fixed $v$, taking a union bound over all $k$, and then taking a union bound over all $v$, and then taking a union bound over times in $[t-\delta n^2, t]$ yields the theorem from Lemma \[l:kedge\].
Completing the proof of Theorem \[t:stoptime\]
----------------------------------------------
Now we are ready to prove Theorem \[t:stoptime\].
Using a random walk estimate it is clear that ${\mathbb{P}}[\tau_0>n^4]=o(1)$. Also it is clear from the properties of an Erdős-Rényi graph that $P[\tau_0<\delta n^2]=o(1)$. Now for $k\geq 0$, and $i=2,3,4,5$, let $A_{k,i}$ denote the event $\{k\delta n^2<\tau_0, (k+1)\delta n^2<\tau_*, (k+1)\delta n^2\geq \tau_i\}$. Using Theorem \[t:bigcut\], Theorem \[t:indmultbound\], Theorem \[t:multedgesbound\] and Theorem \[t:degsbound\] and taking a union over $0\leq k \leq n^2/\delta$, it follows that $${\mathbb{P}}[\tau_0<\tau_*-\delta n^2]\leq o(1)+\sum_{i,k} {\mathbb{P}}(A_{k,i}) =o(1).$$ This completes the proof.
Rewire-to-random Eventually Splits {#s:esplit}
==================================
In this section we prove Theorem \[t:rtoresplit\]. For this section we shall consider running the *rewire-to-random* model with a different initial condition. For $0<p<1$, let $\mathcal{G}^{*}(p)$ be the subset of the state space of our markov chain, i.e., let $\mathcal{G}^*(p)$ is a set of multi-graphs of $n$ vertices with labelled edges where each vertex has either of the two opinions $0$ and $1$, such that $N_1(G)=pn$ and the number of edges in $G$ is in $[\frac{12n^2}{50},\frac{13n^2}{50}]$. Theorem \[t:rtoresplit\] will follow from the the next theorem.
\[t:es\] Let $\beta>0$ be fixed. Consider running the *rewire-to-random* model with relabelling rate $\beta$ starting with the state $G(0)$. Consider the stopping times $\tau=\min\{t:\mathcal{E}^{\times}(t)=\emptyset\}$ and $\tau_*=\tau_{*}(p/2)=\min\{t:N_*(t)\leq \frac{pn}{2}\}$. Then there exists $p=p(\beta)$ sufficiently small such that for all $G(0)\in \mathcal{G}^{*}(p)$, we have $\tau<\tau_{*}$ with high probability.
Before starting with the proof of Theorem \[t:es\] we make the following definitions. Let us fix $G(0)\in \mathcal{G}^{*}(p)$. Let $S$ be the set of vertices in $G(0)$ with degree at most $10n$ and let $T$ be the set of vertices with degree more than $10n$. Clearly $|S|\geq \frac{24n}{25}$. Let us run the process till $10n^2$ steps.
Let $W_{SS}$ denotes the total number of rewirings of edges with both endpoints in $S$. $W_{ST}$ and $W_{TT}$ are defined similarly. Let $Y_{SS}$ denote the number of edges with both endpoints in $S$ at the end of the process (i.e., after running $10n^2$ steps). $Y_{ST}$, $Y_{TT}$ are defined similarly.
We next describe an equivalent way of constructing the *rewire-to-random* dynamics.
An Equivalent Construction of the Dynamics {#an-equivalent-construction-of-the-dynamics .unnumbered}
------------------------------------------
Let $\{X_i\}$ and $\{X'_i\}$ be two sequences of i.i.d. $\mbox{Geom}(\frac{\beta}{n})$ variables (taking values in $\{0,1, \ldots\}$). Let $\{Z_i\}$ be a sequence of i.i.d. $\mbox{Ber}(\frac{1}{2})$ variables. Let $\{W_i\}$ be a sequence of vertices of $G$ where each $W_i$ is uniformly chosen vertex of $G$. All these sequences are distributed independently of each other.
We now describe how to run the process starting with $G(0)$ using only the randomness in the above sequences and the randomness used to choose a disagreeing edge uniformly at each step. Having chosen a disagreeing edge the variables $Z_i$ will be used to designate one of the endpoints of the edge uniformly as the root of the current (relabelling or rewiring) update. Also for each vertex $v$ in $V$ we shall define a sequence $K_i(v)$. To start with, list the vertices in $V$ in some order, say $\{v_1,v_2,\ldots, v_n\}$. Define $K_0(v_j)=X'_{j}$ for all $j$. This encodes the number of updates at the vertex $v_j$ (i.e., the number of moves with $v_j$ being the root) before it changes its opinion for the first time which clearly has a $\mbox{Geom}(\frac{\beta}{n})$ distribution. Roughly speaking, for each vertex $v$ the sequence $\{K_i(v)\}$ will be a counter which shall denote how many more rewiring updates one needs to make at $v$ before the next relabelling update. Once the counter runs to $0$, the next update at that vertex is a relabelling one, and a new value from either the sequence $\{X_i\}$ or the sequence $\{X'_i\}$ will be assigned to the counter. We describe the process formally below.
We shall define the sequences $L_{i}, L'_{i}, T_{i}$ recursively, these will be indices of different elements chosen from $\{X_i\}$, $\{X'_i\}$ and $\{W_i\}$ respectively. Let $L_0=T_0=0$ and $L'_{0}=n$. At step $i$, pick a disagreeing edge $e$ uniformly at random, if such an edge exists. If $Z_i=1$, then choose the vertex with opinion $1$ to be the root of the rewiring or relabelling step, if $Z_i=0$, choose the other one. Let $v$ be the chosen vertex. If $v$ is in $S$ and the opinion of $v$ is $0$, do the following. Set $L'_{i}=L'_{i-1}$. If $K_{i-1}(v)$ is positive, then define $T_{i}=\min\{k>T_{i-1}:W_{k}\neq v\}$, that is $T_i$ is the index of the first hitherto uninspected element in $\{W_j\}$ which allows a legal rewiring move. Rewire the edge to $W_k$ and reduce $K_{i}(v)$ by 1, and set $L_{i}=L_{i-1}$. If $K_{i-1}(v)=0$, then relabel $v$ and set $L_i=L_{i-1}+1$ and $K_i(v)=X_{L{i}}$, in this case, also set $T_{i}=T_{i-1}$. If $v$ is not in $S$, or the opinion of $v$ is $1$, then do the same as in the previous case except use elements from the sequence $X'$ and $L'$ in stead of the elements from sequence $X$ and $L$, and change the values in the sequence $L'$ instead of the sequence $L$.
It is easy to see that this is indeed an implementation of the *rewire-to-random* dynamics.
We need the following lemmas. The first follows immediately from the fact that the number of vertices with opinion $1$ does a random walk.
\[l:es1\] For a fixed $p>0$ and $G(0)\in \mathcal{G}^*(p)$, the number of vertices of opinion $1$ remains between $pn/2$ and $3pn/2$ throughout the first $10n^2$ steps w.h.p..
We call an element of the sequence $X$ *stubborn* if it is at least $25n$.
\[l:es2\] Let $Y=\#\{i\leq L_{10n^2}: X_{i} > 25 n\}$ denote the number of stubborn elements $X$ which are used in first $10n^2$ steps. Then with high probability, $N_1(10n^2)\geq Y$, i.e., the number of vertices with label $1$ after $10n^2$ steps is at least the number of used *stubborn* elements of the sequence $X$.
Let $\mathcal{S}$ denote the following event. $$\mathcal{S}=\Bigl\{\forall v\in G: \#\{i\leq T_{10n^2}:W_i=v\}\leq 14n\Bigr\}.$$ We show that on $\mathcal{S}$, the vertices in $S$ that each stubborn element gets assigned to (i.e., those $v$ such that $X_{\ell}=K_i(v)$ for some $i$ and some stubborn $X_{\ell}$) are distinct and each of them has label $1$ after $10n^2$ steps. Consider a specific stubborn element, suppose it was used and assigned to the vertex $v$. By definition, at that point the opinion of $v$ was $1$. Now by definition of *stubbornness*, before it changes its opinion again, $v$ needs to be the root of at least $25n$ rewiring moves. Notice now that the number of rewirings rooted at $v$ is at most the sum of the initial degree of $v$ (which is at most $10n$ since $v\in S$) and the number of rewirings to $v$ (which is at most $14n$ on $\mathcal{S}$). Hence the vertex $v$ never changes its opinion again and in particular is never associated with any other stubborn element. Hence corresponding to each used stubborn element, there are distinct vertices in $V$ which have opinion $1$ after $10n^2$ steps.
It remains to show that $\mathcal{S}$ occurs with high probability. First notice that using an argument similar to that used in the proof of Lemma \[l:betasmallincoming\], it follows that ${\mathbb{P}}(T_{10n^2}>11n^2)$ is exponentially small in $n$. Also, we note that for each $v\in V$, the chance that $v$ occurs more than $14n$ times in the first $11n^2$ elements of the list $W$ is exponentially small in $n$ using a Chernoff bound. Taking a union bound over all the vertices completes the proof of the lemma.
\[l:es3\] Let $RL_{SS}$ denote the number of times a relabelling occurs when an edge with both endpoints in $S$ was chosen. Then, for $p$ sufficiently small, $RL_{SS}\leq \frac{\beta n}{20}$ w.h.p..
Let $RL_{SS}^+$ denote the number of times we have a relabelling changing an opinion to $1$ after an edge with both endpoints in $S$ was chosen. Notice that each time we choose an edge with both endpoints in $S$, and do a relabelling update changing an opinion from $0$ to $1$, and element from the sequence $X$ gets used. Hence it follows from Lemma \[l:es1\] and Lemma \[l:es2\] it follows that w.h.p. $$RL_{SS}^{+}\leq \min_i\Bigl\{\#\{j\leq i: X_j\geq 25n\}>\frac{3pn}{2} \Bigr\}.$$ Since each $X_j$ is a $\mbox{Geom}(\frac{\beta}{n})$ variable it follows that ${\mathbb{P}}(X_j\geq 25n)=(1-\frac{\beta}{n})^{25n}\geq e^{-50\beta}$. Since $X_j$’s are i.i.d., it follows that for $p$ sufficiently small (depending on $\beta$), within first $\frac{\beta n}{50}$ elements of $X$, there are more than $\frac{3pn}{2}$ *stubborn* elements with high probability. It follows that with high probability $RL_{SS}^{+}\leq \frac{\beta n}{50}$.
Now notice that each time a relabelling occurs when an edge with both endpoints in $S$ is chosen, with probability $\frac{1}{2}$ the relabelling changes an opinion to $1$ and these events are independent of one another. It follows that ${\mathbb{P}}(RL_{SS}>\frac{\beta n}{20}, RL_{SS}^{+}\leq \frac{\beta n}{50})$ is exponentially small in $n$. It now follows that for $p$ sufficiently small, $RL_{SS}\leq \frac{\beta n}{20}$ w.h.p..
\[l:es4\] Let $R_{SS}$ be the number of times an edge with both endpoints on $S$ was picked. For $p$ sufficiently small, $R_{SS}\leq \frac{n^2}{10}$ w.h.p..
Each time an edge is picked, it leads to a relabelling with probability $\frac{\beta}{n}$. It follows using a Chernoff bound that $${\mathbb{P}}(R_{SS} > \frac{n^2}{10}, RL_{SS}\leq \frac{\beta n}{20})\leq {\mathbb{P}}(\mbox{Bin}(\frac{n^2}{10}, \frac{\beta}{n})\leq \frac{\beta n}{20})\leq e^{-\frac{\beta n}{80}}.$$ The lemma now follows from Lemma \[l:es3\].
\[l:es5\] Let $R_{ST}$ be the number of times a disagreeing edge is picked with one endpoint in $S$ and another in $T$. For $p$ sufficiently small $R_{ST}\leq 3n^2$ w.h.p..
Let $W_{ST}$ denote the number of rewiring moves where a disagreeing edge with one endpoint in $S$ and another endpoint in $T$ is rewired. Each time an edge with one endpoint in $S$ and the other in $T$ is rewired, with probability $\frac{1}{2}$ it is rewired with the root at the vertex in $S$ and independent of that with probability at least $\frac{|S|-1}{n-1}$ the edge is rewired to a vertex in $S$. That is, after rewiring an edge with one endpoint in $S$ and another in $T$, the chance that it becomes an edge with both endpoints in $S$ is at least $\frac{|S|-1}{2(n-1)}\geq \frac{23}{50}$ for $n$ sufficiently large. Let $W_{ST\rightarrow SS}$ denote the number of such rewirings. It follows that ${\mathbb{P}}(W_{ST\rightarrow SS}\leq \frac{11n^2}{10}, W_{ST}> \frac{5n^2}{2})$ is exponentially small in $n^2$. Now notice that, $$n^2\geq Y_{SS}\geq -R_{SS}+W_{ST\rightarrow SS}$$ and it follows from Lemma \[l:es4\] that for $p$ sufficiently small $W_{ST\rightarrow SS}\leq \frac{11n^2}{10}$ w.h.p. and hence $W_{ST}\leq \frac{5n^2}{2}$ w.h.p.. Since each time a disagreeing edge is picked, with probability $1-\frac{\beta}{n}$, it leads to a rewiring, it follows that ${\mathbb{P}}(R_{ST}>3n^2, W_{ST}\leq \frac{5n^2}{2})$ is exponentially small in $n^2$ and hence for $p$ sufficiently small $R_{ST}\leq 3n^2$ w.h.p..
\[l:es6\] Let $R_{TT}$ be the number of times a disagreeing edge is picked with both endpoints in $T$. For $p$ sufficiently small $R_{TT}\leq 6n^2$ w.h.p..
Arguing as in the proof of Lemma \[l:es5\] we have that after rewiring an edge with both endpoints in $T$, the chance that it becomes an edge with one endpoint in $S$ and another in $T$ is $\frac{|S|}{n}\geq \frac{24}{25}$. Let $W_{TT\rightarrow ST}$ denote the number of such rewirings. It follows that ${\mathbb{P}}(W_{TT\rightarrow ST}\leq 4n^2, W_{TT}> 5n^2)$ is exponentially small in $n^2$. Now notice that $$n^2\geq X_{ST}\geq -R_{ST}+W_{TT\rightarrow ST}$$ and it follows from Lemma \[l:es5\] that for $p$ sufficiently small $W_{TT\rightarrow ST}\leq 4n^2$ w.h.p.. It follows that $W_{TT}\leq 5n^2$ w.h.p.. Arguing as in the proof of Lemma \[l:es5\] we conclude that $R_{TT}\leq 6n^2$ w.h.p..
We are now ready to prove Theorem \[t:es\].
From Lemma \[l:es4\], Lemma \[l:es5\] and Lemma \[l:es6\] we have that for $p=p(\beta)$ sufficiently small and $G(0)\in \mathcal{G}^{*}(p)$, we have $R_{SS}+R_{ST}+R_{TT}< 10n^2$ with high probability. This implies after $10n^2$ steps there are no disagreeing edge in the graph, i.e., for $p$ sufficiently small $\tau\leq 10n^2$ w.h.p.. The theorem now follows from Lemma \[l:es1\].
Now we prove Theorem \[t:rtoresplit\] using Theorem \[t:es\].
Since in the set up of this theorem $G(0)$ is distributed as $G(n,\frac{1}{2})$, it follows that with high probability the number of edges in $G(0)$ is in $[\frac{12n^2}{50},\frac{13n^2}{50}]$. Let $p=p(\beta)$ be sufficiently small so that the conclusion of Theorem \[t:es\] holds. It follows that on $\{\tau_{*}(p)<\infty\}$, with high probability $G(\tau_{*}(p))\in \mathcal{G}^{*}(p)$. Since the *rewire-to-random* dynamics is symmetric in the opinions $0$ and $1$, it follows from Theorem \[t:es\] that $\tau<\tau_{*}(\frac{p}{2})$ with high probability. This completes the proof of the theorem.
Modifications for the Rewire-to-same Model {#s:samemod}
==========================================
We can prove Theorem \[t:rewiretorandom\] for the rewire-to-same model in a similar manner. For the sake of not being repetitive, we only point out the main differences here. Notice that, with respect to our proof in previous sections, the major difference between the two dynamics that causes some inconvenience is that in the *rewire-to-same* dynamics a vertex with minority opinion is likely to receive edges at a higher rate than a vertex with majority opinion. But as long as the minority opinion density does not become too small, the difference is of a bounded factor, and it turns out that the arguments can be modified to accommodate this. We now point out the main lemmas from the previous sections that need to be modified for the *rewire-to-same* dynamics.
Small $\beta$ Case:
-------------------
Notice that the only place that needs a modification is Lemma \[l:betasmallincoming\]. One needs to define for this case $L_{i+1}$ as the first entry after $L_{i}$ to which a rewiring move is legal. Here it is not true that $L_{6n^2}\leq 13n^2/2$ with exponentially high probability. But notice that, on $t<\tau_*(1/3)$, $L_{i+1}-L_{i}$ is stochastically dominated by a $\mbox{Geom}(\frac{1}{3})$ variable and hence, one can say $L_{6n^2}<20n^2$ with exponentially large probability. The rest of the proof is in essence same up to some minor changes in constants.
Large $\beta$ Case:
-------------------
The proof in the large $\beta$ case also follows along the similar lines. Instead of the *rewire-to-same* dynamics, we consider the *rewire-to-same-\** dynamics, where instead of a disagreeing edge, at each step we pick an edge at random, and do not do anything if the edge happens to be agreeing. Most of necessary changes occur while bounding the number of incoming edges to a vertex in time $[t,t+\delta n^2]$. But on $\tau<\tau_*(\varepsilon)$, one can bound the number of incoming edges to a vertex $v$ in that time by a $\mbox{Bin}(\delta n^2, \frac{1}{\varepsilon n})$ variable.
The only other place where a somewhat significant modification is necessary is in the bound for large cuts. Instead of Proposition \[p:proportionofbadedges\] and Proposition \[p:proportionofbadedgesss\], one needs to show that for any given cut $S$ and $T$, with $|S|\wedge |T|\geq \varepsilon_2 n$, the number of edges with one endpoint in $S$ and another endpoint in $T$, such that the $S$ end point has opinion $0$ and the $T$ endpoint has opinion $1$, is roughly about $p(1-p)$ fraction of the total number of edges with exponentially high probability, and similar bounds on other similar quantities. It can be checked that all these can be obtained following a similar line of arguments as in the proof of Proposition \[p:proportionofbadedges\]. We omit the details. The rest of the bounds can then be obtained by suitably modifying the martingale calculations in Proposition \[p:bigcut\]. The whole proof can then be carried out with some minor changes of constants.
Open Questions {#s:conclusion}
==============
While we establish some rigorous results for the evolving voter model on dense random graphs, the picture is far from clear. We conclude with the following natural questions, that are still open.
- [**What happens eventually in the *rewire-to-same* model?**]{} Notice that we do not have any result analogous to Theorem \[t:rtoresplit\] in the *rewire-to-same* model. It is a natural question to ask whether in the rewire-to-same model, is there a positive fraction of both opinions present when the process reaches an absorbing state? As we have mentioned before in [@Durrett12] it was conjectured that, for the sparse graphs (with constant average degree), in the rewire-to-same model, one of the opinions take over almost the whole graph, but it is not known rigorously.
- [**Is there a sharp transition in $\beta$?**]{} Another natural question to ask is if there is any value $\beta_0$ such that for $\beta<\beta_0$ we have behaviour as in Theorem \[t:rewiretorandom\]$(i)$ and for $\beta>\beta_0$, we have behaviour as in Theorem \[t:rewiretorandom\]$(ii)$?
- [**What can we say about sparser graphs?**]{} We can prove by an argument similar to proof of Theorem \[t:rtoresplit\]. for sparser graph (i.e., $G(0)\sim G(n,\frac{\lambda}{n})$) that there exists $\beta_0>0$, such that for all $\beta<\beta_0$, with high probability the process stops before the density of the opinions change. But the other side of the phase transition seems harder to prove. The main difficulty seems to be the presence of a few vertices of very high degree. It is another interesting question to investigate whether one could prove results about evolving voter models starting with not too sparse graphs, e.g. $G(0)\sim G(n, n^{1-\alpha})$ for some $0<\alpha <1$?
**Acknowledgements.** The authors would like to thank Rick Durrett for introducing us to the model.
[^1]: Dept. of Statistics, University of California, Berkeley. Supported by UC Berkeley Graduate Fellowship. Email:[email protected]
[^2]: University of California, Berkeley and Australian National University. Supported by an Alfred Sloan Fellowship and NSF grants DMS-1208338, DMS-1352013. Email:[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a new impurity solver for dynamical mean-field theory based on imaginary-time evolution of matrix product states. This converges the self-consistency loop on the imaginary-frequency axis and obtains real-frequency information in a final real-time evolution. Relative to computations on the real-frequency axis, required bath sizes are much smaller and less entanglement is generated, so much larger systems can be studied. The power of the method is demonstrated by solutions of a three band model in the single and two-site dynamical mean-field approximation. Technical issues are discussed, including details of the method, efficiency as compared to other matrix product state based impurity solvers, bath construction and its relation to real-frequency computations and the analytic continuation problem of quantum Monte Carlo, the choice of basis in dynamical cluster approximation, and perspectives for off-diagonal hybridization functions.'
author:
- 'F. Alexander Wolf'
- Ara Go
- 'Ian P. McCulloch'
- 'Andrew J. Millis'
- Ulrich Schollwöck
title: 'Imaginary-time matrix product state impurity solver for dynamical mean-field theory'
---
Introduction
============
Dynamical mean-field theory (DMFT) in its single-site [@metzner89; @georges92; @georges96] and cluster [@maier05; @kotliar06] variants is among the most widely employed computational techniques for solving quantum many-body problems. At the core of a numerical solution of DMFT is an : an algorithm for solving a quantum impurity problem. The most prominent examples of impurity solvers are continuous time quantum Monte Carlo (CTQMC) , exact diagonalization (ED) , the numerical renormalization group (NRG) , and the density matrix renormalization group (DMRG) . While all methods have their strengths, key limitations mean that fundamentally important classes of problems have not yet been adequately addressed. Other recent suggestions for impurity solvers including in particular the computationally inexpensive density matrix embedding theory (DMET) , are promising but have not been tested in detail.
CTQMC is widely employed but its application to situations involving low point symmetry, non-Hubbard interactions or multiple relevant orbitals is limited by the fermionic sign problem. Reaching low temperatures becomes highly computationally expensive while calculating real-frequency information requires analytical continuation, a numerically ill-posed procedure fraught with practical difficulties.
ED makes no assumption on the interaction and does not have a sign problem. It is limited by the size of the Hilbert space that can be studied, meaning in practice that it is restricted to a small number of correlated sites to which only a small number of bath sites can be attached. Recently, improvements have been achieved by considering only restricted subspaces of the Hilbert space , but the size of problem remains a significant limitation.
NRG converges the DMFT loop on the real-frequency axis and very effectively obtains real-frequency information in the low frequency limit. Current applications have been to relatively small problems (the most recent achievement is a solution of the single-site DMFT approximation to a three-band model ) and it remains to be seen how far the method can be extended.
DMRG is a set of algorithms operating on the space of matrix product states (MPS) . It has been found to be extremely powerful for the calculation of ground states of one-dimensional quantum systems ; it was very successfully extended to the calculation of spectral functions which, in contrast to NRG, it obtains with equal resolution across the spectrum, see . In pioneering work the method was applied as a DMFT solver by and with important further work done by these and other authors . However, the method has not been widely accepted, perhaps because high-quality data were presented only for the single-site approximation to the single-band Hubbard model. Recently the method was used to reliably solve a two-site DCA , and insights into the entanglement of the impurity problem made it even more powerful . In view of these advances, DMRG now is a candidate for a highly flexible low-cost impurity solver, which can in addition be efficiently employed in the non-equilibrium formulation of DMFT .
This paper reformulates the DMRG method as an impurity solver for DMFT on the imaginary-time axis (). As we will show, this strongly reduces entanglement and necessary bath sizes. The price to be paid is a reduced resolution on the real-frequency axis, which we study in detail by comparing with calculations that converge the DMFT loop on the real-frequency axis (). The reformulation allows a much larger class of problems to be addressed, including some that are unreachable by other methods, due e.g. to the sign problem, the size of the correlated cluster or the number of bands. We illustrate this with calculations for three band models in the single-site and two-site DMFT approximation (). We append discussions of the optimization of typical DMFT Hamiltonians () and of the entanglement in different representations of the DCA including a discussion of off-diagonal hybridization functions ().
Method {#secMethod}
======
Overview: Green’s functions in DMRG
-----------------------------------
The computational key challenge in DMFT is the computation of the full frequency dependence of the Green’s function of a quantum impurity model involving an essentially arbitrary bath. The “size” (number of correlated sites $L_c$) of the impurity model should be as large as possible and the kinds of interaction that can be treated should be as general as possible. The Green’s function is used in a self-consistency loop, which may require many iterations for convergence. The solution should be as inexpensive as feasible, and must run automatically, without need for manual optimization of parameters or procedures. In this subsection we present a qualitative discussion of the issues involved in computing the Green’s function using DMRG methods, to motivate the work described in detail below.
Within DMRG one computes Green’s functions by first representing the system ground state $\ket{E_0}$ as an MPS. One then generates a one electron (one hole) excitation $\ket{\psi^>_0}=d\dag\ket{E_0}$ $\left(\ket{\psi^<_0}=d\ket{E_0}\right)$ by applying a creation (annihilation) operator $d\dag$ ($d$) to $\ket{E_0}$. While the state $\ket{\psi_0^\gtrless}$ is at most as entangled as the ground state $\ket{E_0}$ , in order to compute a Green’s function, one has to perform further operations on $\ket{\psi_0^\gtrless}$. These operations typically increase entanglement and by that the of an MPS, which ultimately limits all computations.
Let us be more concrete and consider a general MPS of bond dimension $m$ for a system with $L$ sites and open boundary conditions. Defining $A^{\sigma_i}, B^{\sigma_i} \in \mathbb{C}^{m\times m}$ for $i\neq 1,L$ and $A^{\sigma_1} \in \mathbb{C}^{1\times m}$, $B^{\sigma_L} \in \mathbb{C}^{m\times 1}$, where $\sigma_i \in \{0,\uparrow,\downarrow,\uparrow\downarrow\}$ labels a local basis state of the Hilbert space, any MPS can be represented as where $S=\tx{diag}(s_1,...,s_m)$ is a diagonal matrix, and $A^{\sigma_i}$ are and $B^{\sigma_i}$ are , respectively: Here, $I$ are identity matrices. Left- and right-normalization make the Schmidt decomposition of $\ket{\psi\tl{MPS}}$ that is associated with partitioning the system at bond $(l,l+1)$. The bond entanglement entropy for the associated reduced density matrix can therefore simply be read of from When subsequently we refer to an associated with repeated operations on $\ket{\psi\tl{MPS}}$ this implies the need to adjust the bond dimension $m$ such that $\ket{\psi\tl{MPS}}$ still faithfully represents a physical state. If entanglement in the physical state becomes too large, we have to choose $m$ so large that computations with MPS become impractical.
Since the first suggestion for computing spectral functions within DMRG , the field has evolved by the important development of the correction vector method . The subsequent understanding of the connection between DMRG and MPS opened the door to many further approaches to computing spectral and Green’s functions, in particular, time evolution and subsequent Fourier transform , an improved Lanczos algorithm , and the Chebyshev recursion . All of these are formulated for the calculation of spectral functions at $T=0$, as considered in the present paper, and came at much cheaper computational cost than the correction vector method . We note that for $T>0$, there are perspectives for even more powerful algorithms: it was recently demonstrated that the numerically exact spectral function of a molecule consisting of several hundreds of interacting spins could be computed .
These developments (see for more details) make MPS-based solvers an attractive possibility for dynamical mean-field theory. However, the growth of entanglement arising in all calculations of Green’s function has limited the systems sizes that have been addressed to date. Also, in MPS computations manual adjustments, for example choosing optimal broadening [@holzner11] or combining results of different systems sizes , are still common practice. In the rest of this section, we show that these problems can to a large degree be circumvented by computing Matsubara Green’s functions using imaginary-time evolution. The imaginary-time framework naturally extends existing approaches based on real-time evolution , which have recently been shown to provide currently the most efficient algorithmic approach to compute real-frequency spectral functions .
Imaginary-time computation
--------------------------
The central objects of technical interest in this paper are the and the correlation functions $\wt {G}^\gtrless$, which we define for imaginary time $\tau$
\[Ggtr\]
In the second line, we evaluate $\wt G^\gtrless(\tau)\vert_{\tau=it}$ and by that obtain a correlation function for real time $t$, which will be useful later on. The functions $\wt G^\gtrless$ carry spin and orbital indices associated with the spin and orbital indices of the single-particle (hole) excitation $\ket{\psi_0^\gtrless}$, but these indices are not explicitly written here. We will discuss the relationship of $\wt G^\gtrless$ to the physical Green’s functions (which we denote by $G$) below.
While it is not essential in principle, we evaluate using a Krylov algorithm , which represents the time evolution operator in a local Krylov space and is able to treat Hamiltonians with long-ranged interactions. Before performing a time evolution computation, one has to compute the initial state $\ket{\psi_0^\gtrless}$ using an MPS optimization of the ground state. As impurity models come with open boundary conditions, this is well suited for DMRG. We discuss this optimization for typical DMFT Hamiltonians in .
Figure \[figImagTimeEvol\] presents representative results based on parameters obtained from a two-site DMFT solution of the Hubbard model. Figure \[figImagTimeEvol\](a) shows the time evolution of $\wt G^\gtrless(\tau)$ out to times as long as 350 times the basic timescale (inverse half-bandwidth $D$) of the model, which suffices to converge $\wt G^\gtrless(\tau)$ to a precision of $5\cdot10^{-4}$. Figure \[figImagTimeEvol\](b) demonstrates the key advantage that makes this computation possible: the lack of growth of maximal bond dimensions $m$ with time of the associated imaginary-time evolved states $\ket{\psi^\gtrless(\tau)} = e^{-(H-E_0)\tau} \ket{\psi_0^\gtrless}$. The imaginary time-evolution operator does not create entanglement as it projects on the lowly entangled ground state.
Figure \[figImagTimeEvol\](a) reveals additional information about the nature and rate of convergence of $\wt G^\gtrless(\tau)$. In the insulating phase, $H$ has a gap and $\wt G^\gtrless(\tau)$ decays exponentially irrespective of whether one considers a finite system or the thermodynamic limit. In the metallic phase, $\wt G^\gtrless(\tau)$ decays algebraically in the thermodynamic limit. For a finite system though, there always remains a small gap, and even though the decay resembles an algebraic decay for short times, it always becomes exponential at long times. The exponential decay can be exploited to speed up computations considerably by a simple technique known as . This technique is an efficient formulation of the fitting problem for the ansatz function $f(\tau)=\sum_n \alpha_n e^{\beta_n \tau}$, $\alpha_n,\beta_n \in \mathbb{C}, \tau\in \mathbb{R}$, which can then be used to reliably extrapolate functions with an exponentially decaying envelope. This is illustrated by the dashed black line in (a), which has been fitted to match $\wt G^\gtrless(\tau)$ for $\tau D\in[150,200]$ and was then extrapolated for higher times. The solid green line, by contrast, is the result of the MPS computation. Agreement can be seen to be perfect.
Physical Green’s functions
--------------------------
Of particular interest in the rest of this paper are the imaginary time Green’s functions $G\th{mat}(\tau)$ defined via $$G\th{mat}(\tau) = -\theta(\tau) \wt G^>(\tau) + \theta(-\tau) \wt G^<(\tau),
\label{Gtau}$$ whose Fourier transform gives the Matsubara Green’s function $$G\th{mat}(i\omega_n)=\int_{-\infty}^{\infty} d\tau\, e^{i\omega_n \tau}G\th{mat}(\tau),
\label{Gmat}$$ at zero temperature, where $\omega_n=(2n+1)\pi/\beta$ and $\beta\ra\infty$. We shall also be interested in the retarded real-time Green’s function $$G\th{ret}(t) = -i\theta(t) (\wt G^>(it) + \wt G^<(it)),
\label{Gret}$$ from which the retarded frequency-dependent Green’s function is obtained as $$G\th{ret}(\w) = \int_{-\infty}^{\infty} dt\, e^{i(\omega +i0^+)t}G\th{ret}(t).
\label{Gret}$$
In numerical practice, we evaluate the Fourier transforms leading to and approximately as with cutoff times $\tau\tl{max}$ and $t\tl{max}$. This approximation is only controlled if we are able to reach long enough times $\tau\tl{max}$ and $t\tl{max}$, such that $\wt G^\gtrless(\tau)$ and $\wt G^\gtrless(it)$ have converged to zero to any desired accuracy.
In contrast to a computation on the imaginary axis, reaching arbitrarily long times $t\tl{max}$ on the real axis is prohibited by a logarithmic growth of entanglement, which comes with a power-law growth of bond dimensions. In addition, finite-size effects are a severe source of errors because the long-time behavior is determined by the bath size. For a numerically exact computation, one has to choose the system large enough to observe exponential “pseudo-convergence" of $\wt G^\gtrless(it)$ to zero before finite-size effects are resolved . In the context of the present paper, we deal with small system sizes and will never observe pseudo-convergence. In particular there is no exponential pseudo-convergence, so that cannot be employed . Therefore, when computing the real-frequency spectral function after converging the DMFT loop, one has to use the further approximation of the finite-size effects that emerge at long times by computing, instead of $G\th{ret}(\w)$ in , which yields the spectral function $A_\eta(\w) = -\frac{1}{\pi} \tx{Im} G\th{ret}_\eta(\w) = \frac{1}{\sqrt{2\pi}\eta} \int d\w' A(\w') e^{-\frac{(\w-\w')^2}{2\eta^2}}$. Instead of a Gaussian damping and broadening, one could also use an exponential damping leading to Lorentzian broadening, which damps out the original time-evolution information more strongly, though.
Before presenting detailed benchmark results for the solution of DMFT using imaginary-time evolution of MPS, let us clarify which price we have to pay for profiting from the great advantage of not facing entanglement growth. We do this by comparing the imaginary-time approach (itMPS) to approaches that solve the DMFT loop on the real axis.
Comparison of imaginary-axis with real-axis computations {#secBath}
========================================================
The equation in DMFT relates an impurity model specified by a hybridization function and a self energy to a lattice model specified by a lattice Hamiltonian and the same self energy. We discuss the issues using the example of the dynamical cluster approximation to the single-band Hubbard model Here $\varepsilon_k$ denotes the single-particle dispersion of the lattice and $\mu$ is the chemical potential. In the dynamical cluster approximation, the Brillouin zone, consisting in $N$ momentum vectors $k$, is covered by $N_c$ (for single-band $L_c=N_c$) equal-area tiles (patches), labelled here by $P_K$ and the self energy is taken to be piecewise constant, with being a potentially different function of frequency in each tile. The impurity model is specified by the on-site energy $\varepsilon_K$ and the hybridization function $\Lambda_K(z)$, which is to be determined using the fixed point iteration referred to as the . The loop works as follows: make an initial guess for $\Lambda_K(z)$; compute $\Sigma_K(z)=\omega+\mu-\varepsilon_K-\Lambda_K(z) - [G_K\th{imp}(z)]^{-1}$, then update $\Lambda_K$ via $\Lambda_K(z)=z+\mu-\varepsilon_K-\Sigma_K(z)-[G_K\th{latt}(z)]^{-1}$ and repeat this procedure until convergence.
We discuss two aspects of the comparison of real- and imaginary-frequency solutions of the DMFT self-consistency equation . The first has to do with the number of bath sites needed to obtain a solution of the self-consistency equation. The second is the accuracy to which the spectral functions of physical interest can be reproduced.
The DMFT self-consistency equation , defines the hybridization function $\Lambda_K$ as a continuous function in terms of the difference between the computed self energy and the inverse of the lattice Green’s function. In DMRG-type methods, the hybridization function $\Lambda_K$ is approximated as the hybridization function $\Lambda\th{discr}_K$ of a discrete impurity model, which is the sum of a number of poles. If the number of poles is too small one cannot construct a meaningful approximation on the real axis and a DMFT loop cannot be converged. For this reason DMRG-based solutions of DMFT up to now , all of which were real axis computations, have been performed using numbers of bath sites of at least $L_b/L_c \gtrsim 30$, and in the case of the single-band Hubbard model, even much more $L_b/L_c \gtrsim 120$. Use of such a large number of bath sites means that with modest broadening the hybridization function can be reasonably approximated as a continuum, enabling a stable solution of .
By contrast, formulating the problem on the imaginary axis (as is typically done in standard ED solvers where the number of bath sites is strictly limited) automatically smoothens the hybridization function $\Lambda\th{discr}_K$ and permits a stable solution. From the imaginary axis solution one must then determine the discrete set of bath parameters to represent $\Lambda\th{discr}_K$. This is typically done by numerical minimization of a cost function defined as Here, $\alpha$ defines a weighting function $\w_n^{-\alpha}$. Choosing $\alpha>0$, $\alpha=1$, attributes more weight to smaller frequencies , which we find helpful when using small bath sizes $L_b/L_c<5$. To define the frequency grid for the fit $\w_n=(2n+1)\pi/\beta\tl{eff}$, one defines a fictitious inverse temperature $\beta\tl{eff}$, which has no physical significance. We further employ a cutoff frequency $\omega_c$, which implies a finite number $N\tl{fit}$ of fitted Matsubara frequencies.
If one tries to define an analogous cost function for the real axis, the result is useless as then $\Lambda_K\th{discr}(\w+i0^+)$ is a sum of poles, whereas the hybridization function $\Lambda_K(\w+i0^+)$, as encountered in , is continuous . One can overcome this problem only when using a Lindbladt formalism , which increases the complexity of the problem substantially.
The minimization of is done using using standard numerical optimization. The optimization in the initial DMFT iteration should be done using a global optimization scheme , and in subsequent iterations using a local optimization scheme (conjugate gradient), which takes as an initial guess for the new bath parameters the values of the previous iteration. Figure \[figBathFitSites\] shows the convergence of the fit of the hybridization function with the number of bath sites $L_b/L_c$. For $L_b/L_c = 7$ one already obtains errors as little as $\simeq 10^{-3}$ and for values $L_b/L_c \gtrsim 9$ the quality of the fit already stops improving. It is at this point, where we (and all ED-like techniques) face the problem of “analytic continuation" encountered in CTQMC, namely that Green’s functions on the imaginary axis encode information in a much less usable form than on the real axis.
Consider again the example of the two-site DCA for the single-band Hubbard model on the square lattice. In , this problem has been solved entirely on the real axis using $L_b/L_c=39$ bath orbitals. Here, we converge the DMFT loop on the imaginary axis and compute the spectral function in a final real-time evolution using $L_b/L_c=3,5,7$ bath orbitals. We compare both solutions in . Whereas for the (central) momentum patch “$+$" shown in (a), we find satisfactory agreement of the imaginary-axis with the real-axis calculation, this is not the case for the (outer) momentum patch “$-$" shown in (b), even though the corresponding imaginary-axis Green’s function is well reproduced, see (d). Evidently, in (b), the central peak and the pseudo-gap at the Fermi edge are smeared out by a broadening $\eta=0.2D$ that hides finite-size effects to a large degree. Reducing the broadening to $\eta=0.05D$ as shown in (c), again reveals the pseudo-gap and the central peak; but together with unphysical finite-size effects. We observe that the nature of these finite-size effects is qualitatively comparable when using different numbers of bath sites $L_b/L_c=3,5,7$. Already for $L_b/L_c=3$, we obtain reasonable results. On the imaginary axis, by contrast, $L_b/L_c=5,7$ still improve over $L_b/L_c=3$ and almost agree with the numerically exact QMC data for $\beta=200/D$ of , see (d). However we emphasize that even with the modest number of bath sites used here the basic features of the spectral function are reproduced (for example the areas in given frequency ranges).
Three-band calculations {#secChecks}
=======================
Three-band model in single-site DMFT {#sec3bd1s}
------------------------------------
We now demonstrate the power of the method by applying it to three-band problems in the single-site approximation (where comparison to existing calculations can be made) and the two-site approximation. Both was hitherto not accessible to DMFT+DMRG computations.
We study the three-band Hubbard-Kanamori model with Hamiltonian (omitting the site index $i$ in the following definition of $H_{\tx{loc},i}$) where $i$ labels sites in a lattice and $k$ wave vectors in the first Brillouin zone, $n_{i,a,\sigma}=d^\dagger_{i,a,\sigma}d_{i,a,\sigma}$ is the density of electrons of spin $\sigma$ in orbital $a$ on site $i$, $\mu$ is the chemical potential, $\Delta_a$ is a level shift for orbital $a$, $\varepsilon^{ab}_k$ is the band dispersion, $U$ is the intra-orbital and $U'$ the inter-orbital Coulomb interaction, and $J$ is the coefficient of the Hund coupling and pair-hopping terms. We adopt the conventional choice of parameters, $U'=U-2J$, which follows from symmetry considerations for $d$-orbitals in free space and holds (at least for reasonably symmetric situations) for the $t_{2g}$ manifold in solids .
We study the orbital-diagonal and orbital-degenerate case ($\Delta_a=0$) on the Bethe lattice, the non-interacting density of states is semi-elliptic, In the single-site approximation, the impurity Hamiltonian used within DMFT is given by where $c^\dagger_{l,a,\sigma}$ creates a fermion in the bath orbital $l$, $V_{l,a,\sigma}$ describes the coupling of the impurity to the orbital $l$, and $\varepsilon_{l,a.\sigma}$ denotes the potential energy of orbital $l$. The hybridization function is then given by
Figure \[fig3bdMIT\] compares the dependence of the particle density $n$ on the chemical potential $\mu$ obtained by the MPS methods used here to those obtained by numerically exact CT-QMC methods . The plateaus in $n(\mu)$ are the Mott insulating regimes of the phase diagram. The agreement is very good in general, confirming the reliability of our new procedure even with only three bath sites per correlated site. This leads to an extremely cheap computation, for which a single iteration of the DMFT loop took about on two cores (see for more details).
In panel (a) of we show a more stringent test, namely the dependence of the self energy on Matsubara frequency, in a parameter regime where the self energy was previously found [@werner08] to exhibit an anomalous $\omega^\frac{1}{2}$ frequency dependence and (in some regimes) a non-zero intercept as $\omega \rightarrow 0$. These phenomena are associated with a spin-freezing transition [@werner08; @werner09].
Panel (a) of Figure \[fig3bdSpinFreeze\] shows that the known low frequency $\omega \lesssim t$ self energy is already accurately reproduced even for the computationally inexpensive choice of $L_b/L_c=3$, although one observes deviations for the high-frequency behavior. The deviations at high frequency decrease as the number of bath sites is increased, although full convergence at all frequencies has not been demonstrated. Panel (b) of shows that the deviations are linked to the impossibility of fitting the hybridization function equally well for all frequencies using only a small number of bath sites. The large deviations at high frequencies are due to the choice $\alpha=1$ in , which enforces good agreement for low frequencies and allowed to successfully reproduce the MIT transition in . Increasing the number of bath sites to $L_b/L_c=5$ leads to a much better approximation of the hybridization function also for high frequencies, with concomitant improvement in the self-energy ((a)).
Three-band model in two-site DCA
--------------------------------
We now present results obtained using a two-site DCA approximation to the three-band model of the previous section. For this problem there are no low-temperature results available in the literature. The size of the problem is beyond the scope of standard ED. The truncated configuration interaction (CI) impurity solver allows to access a relatively high number of bath sites but is limited in the number of correlated sites: in , a problem with $L_c=3$ and $L_b=30$ was computed, and in , one with $L_c=4$ and $L_b=20$. The three-band two-site DCA though has $L_c=6$ correlated sites and it remains to be seen whether this is in reach for the CI solver. The problem is also challenging for standard CTQMC. Recent technical improvements on mitigating the sign problem enabled to treat this model at the temperature of $T=0.025D$ with $D$ the half bandwidth, which required large computational resources. This temperature is relatively high as the authors stated that in the study of a simpler two-band two-site model, where they reached $T=0.0125D$, it was “intractable” to reach low enough temperatures to clarify whether bad metal/spin freezing behavior was intrinsic or not .
We study the model on the two-dimensional square lattice, i.e. using $\varepsilon_k^{ab}=-2t\left(\cos k_x + \cos k_y\right)\delta^{ab}$. We use the momentum patching of ; this definition is also used in the single-band computations of and \[figDCA\]. We note that this model is not directly relevant to layered materials where the $t_{2g}$ orbitals are relevant, because in the physical situation the two dimensionality will break the three-fold orbital degeneracy. However the system is well defined as a theoretical model and useful to demonstrate the power of our methods.
As is the case for the CI method, the DMRG method used here is easily able to treat a large number of bath sites if the number of correlated sites is small: for $L_c=1$, DMRG has already often be proven to treat $L_b>120$ bath sites, and for $L_c=2$, $L_b>80$ is easily accessible . However for more correlated sites, the number of bath sites that can be added at given computational cost decreases. For $L_c=6$, we use $L_b=18$, $L_b/L_c=3$, which we have shown to be sufficient to produce reliable results in the single-site case without requiring overly large computational resources (computation time of several hours per DMFT iteration on two cores).
We have tested the two-site calculation by converging the DMFT loop for the three-band Hubbard model with $U'$ and $J=0$ and comparing the results with a corresponding two-site single-band DCA. Perfect agreement is obtained (not shown). Non-zero values of $U'$ and $J$ create additional entanglement and make computations more costly. It is then a decisive question whether a real-space or a momentum-space representation of the impurity-cluster is less entangled. We discuss this in finding that for the single-band Hubbard interaction, both representations yield similar entanglement, whereas for the Hubbard-Kanamori interaction, the real-space representation is much less entangled. Computational cost is therefore tremendously reduced by using the real-space representation, which comes with an off-diagonal hybridization function. This is the opposite behavior as observed for QMC, where the off-diagonal hybridization function creates a severe sign problem. We further note that in the real-space representation, strong interactions yield a less and less entangled impurity problem, as electrons become more and more localized.
We now present results for the more physically relevant case $U'=U-2J$ with $J=U/4$. For these parameters, at half filling the critical interaction for the MIT in the single-site DMFT approximation is $U_c\simeq1.3D$ . (a) shows that our results are consistent with this estimate: the dashed lines depict the single-site (1s) results, showing a metallic solution (spectral function non-zero at $\omega=0$) for $U=D$, and an insulating solution (spectral function zero at $\omega=0$) for $U=2D$. In the two-site (2s) DCA (solid lines), by contrast, the critical value $U_c$ for the MIT is lowered. Even at $U=D$ the $\omega=0$ spectral function is zero (the small non-zero value in panel (a) is an effect of broadening, as seen in panel (b)). The different nature of the metallic and insulating solutions is also visible on the imaginary axis in the different nature of the decay of the imaginary-time Green’s function. This is plotted in (c) for $U=D$; clearly, a power-law decay is observed for the metallic solution obtained in the single-site DMFT, whereas an exponential decay is obtained for the insulating solution obtained within the two-site DCA.
Conclusion {#secCon}
==========
This paper introduces an imaginary-time MPS (itMPS) solver for DMFT and shows that it can treat complex models, not easily accessible with other methods, at modest computational cost. This development establishes DMRG as a flexible low-coast impurity solver for realistic problems, such as encountered in the study of strongly-correlated materials. The crucial advance stems from the fact that imaginary-time evolution does not create entanglement, and hence allows to compute Green’s functions numerically exactly, provided a ground state calculation is feasible.
The method can be improved in many ways. In particular, different representations of the impurity problem exhibit different degrees of entanglement, so optimizing the representation of the impurity problem is a promising route. Ideas from ED approaches for constructing relevant subspaces of the Hilbert space may lead to further improvements. Such techniques have been successfully combined with MPS . Another route to reduce computational effort and by that reach even more complex models could consist in performing computations for the reduced dynamics of the impurity , or making use of extremely cheap algorithms for computing the Green’s function at elevated temperatures . Finally, we note that using MPS as an impurity solver makes accessing entanglement as a quantity for understanding the properties of the embedded impurity cluster very easily accessible. Proposals in this direction have been made for cellular DMFT and for impurity models generally .
Acknowledgements
================
FAW thanks G. K.-L. Chan for stressing the relevance of converging the DMFT loop on the imaginary-frequency axis, and N.-O. Linden for helpful discussions. AJM and US acknowledge the hospitality of the Aspen Center for Physics NSF Grant 1066293 during the inception of this work. FAW and US acknowledge funding by [FOR1807](https://for1807.physik.uni-wuerzburg.de/) of the DFG. AJM and AG were supported by the US Department of Energy under grant ER-046169.
Further technical details {#secTech}
=========================
Ground state optimization {#secGS}
-------------------------
The main challenge in solving the ground state problem of a typical cluster-bath Hamiltonian as encountered in DMFT, stems from the fact that DMRG is a variational procedure that is initialized with a random state, which is then optimized locally. A local optimization procedure is slow when optimizing a global energy landscape. In addition, the local optimization is prone to getting stuck in local minima, if no “perturbation steps that mix symmetry sectors" are applied. The standard perturbation techniques for single-site DMRG rely on “perturbation terms" that are produced by contracting the Hamiltonian with the MPS. If the Hamiltonian itself does not contain terms that mix the symmetry sector, these methods do not work.
A typical cluster-bath Hamiltonian has both features, a global variation of the potential energy and parts that are not connected with symmetry mixing terms, such as in the three-band Hubbard Kanamori model at $J=0$. This situation is sketched in .
In , the models under study allowed to solve this problem using the non-interacting solution. For the general models studied in the present paper, an unbiased numerical technique has to be employed. What we do in practice, is to first find the ground state of a system with additional symmetry-mixing couplings (denoted as red solid lines in ) that are then adiabatically switched off. In practice, we sweep 5 to 10 times with additional hoppings of 10% magnitude of the physical hoppings, and another 5 to 10 times with additional hoppings of 1% magnitude. After these preliminary sweeps, the quantum number (particle number) distribution has globally converged, and we can continue with converging the ground state of the exact Hamiltonian.
Convergence of DMFT iteration {#secConv}
-----------------------------
The calculations for the three-band single-site DMFT in are only trivially parallelized using one core to compute the imaginary time evolution of each the particle ($>$) and the hole ($<$) Green’s functions $\wt G_{a}^{\gtrless}(\tau)$.
In , we show the convergence DMFT loop for the single-site DMFT for the three-band Hubbard-Kanamori model as studied in . (a) shows the convergence of the Matsubara Green’s function down to a precision of $10^{-3}$. (b) and (c) show the convergence of the density and of the ground state energy per particle, respectively. (d) shows the computation time. An iteration on the Matsubara axis takes about . The final real-axis computation (iteration 31) is considerably more expensive, but can still be optimized.
Least-entangled representation and off-diagonal hybridization functions {#secBasis}
=======================================================================
Geometry and general considerations
-----------------------------------
In , some of us showed that the star geometry of the impurity problem can have substantially lower entanglement than its chain geometry. In the star geometry, DMRG profits from the small entanglement of the almost occupied states with low potential energy with the almost unoccupied states with high potential energy. A high weight for the superposition of a low- with a high-energy state is physically irrelevant. In the star geometry, DMRG is able to eliminate these superpositions as potential energy is separated locally, i.e. in the same basis in which DMRG optimizes the reduced density matrix in order to discard irrelevant contributions. In principle, as mentioned in , ideas from basis selective approaches in exact diagonalization are a different method to account for the fact that many states in the Hilbert space have a negligible weight for the computation of the Green’s function and only few physically relevant states occupy a small fraction of the Hilbert space. Among these are the truncated configuration interaction (CI) , the basis-selective ED or the coupled cluster methods in quantum chemistry. As these methods can be combined with DMRG , they might be a further route to construct efficient representations of the impurity-cluster problem
In the present paper, the question of the least entangled representation of the impurity problem shall be restricted to the question of which basis to choose in a DCA calculation. This is of high relevance also in another context: In the real-space representation, the hybridization function becomes off-diagonal. For CTQMC, this generates a sign problem. In our approach, this doesn’t affect computational cost much in the single-band Hubbard model. It even leads to a tremendous reduction of computational cost for the three-band Hubbard Kanamori interaction.
DCA in momentum or real space {#secDCArs}
-----------------------------
The complexity of the interaction determines whether the real- or the momentum-space representation of the cluster-bath Hamiltonian is less entangled. In real space, the interaction has a simple form, but the hybridization function has off-diagonal contributions, which result in additional couplings of cluster and bath sites. In momentum space, the hybridization function is diagonal but the interaction becomes off-diagonal. The additional couplings induced by that depend on the complexity of the interaction.
Let us be more concrete. For the two-site case, the discrete Fourier transform yields the even and odd superposition of the real-space cluster. where the index of $\wt d_{K}\dag$ labels momentum patches $K$ and the index of $d_{i}\dag$ labels real space cluster sites $i$. There might be further indices labeling spin or orbital.
In real-space, the hybridization function has the form where the symmetry of the real-space cluster imposes $\Lambda_{ij}(z) = \Lambda_{ji}(z)$. In momentum space, the hybridization function is diagonal and symmetry is reflected in the reduced number of bath sites per patch $L_b'=L_b/L_c$, where $L_c=2$ is the number of momentum patches.
We choose to use the momentum representation for the bath discretization, as was done for the real-axis in . While on the real-frequency axis this is the only viable option, the bath fitting on the imaginary-frequency axis via is possible also for the off-diagonal real-space case. In real space, , particle hole symmetry can be easily imposed in the fitting procedure, while this is not possible in momentum space.
Given the parameters of the momentum space representation obtained by performing a bath fit via , we define the parameters of the equivalent real-space representation as follows: In momentum space, bath parameters are indexed by $l_K=1,...,L_b'$, $L_b'=L_b/L_c$ and in real space, bath parameters are indexed by $l=1,...,L_b$, then
Whereas the momentum-space Hamiltonian has $L_b$ non-zero couplings $V_{Kl_K}$, the real space Hamiltonian has $L_c\times L_b$ couplings $V_{il}$. On the other hand, the interaction part generates $L_c\times(L_c-1)$ additional non-local couplings in the momentum-space representation as compared to the real-space Hamiltonian.
From this one could naively expect that the real-space representation is less entangled if $L_c\times(L_c-1)>L_c\times L_b$. Numerical experiments show that the real-space representation is much more favorable than this estimate. For a single-band Hubbard model, we find about the same entanglement in the real space and the momentum space representation, with slight advantages for momentum space. In the three-band Hubbard Kanamori model, the real-space representation is considerably less entangled and leads to a tremendous reduction of computational cost. In particular, we were able to obtain the results of in the momentum space representation when using $L_b/L_c=3$, only for $L_b/L_c=2$ but then at much higher computational cost.
Green’s functions from matrix product states {#secSpec}
============================================
Even though the following discussion is not to set up the imaginary-time real-time impurity solver, it is of general interest in this context and might stimulate further advancements.
A computation of $A(\w) = \bra{\psi_0}\delta(\w - (H-E_0)) \ket{\psi_0}$ via a computation of eigen states of $H$ is extremely redundant as only a tiny neighborhood $\mathcal{N} = \{\ket{\psi} \big| \bra{\psi}H\ket{\psi_0}\neq 0\}$ of a the single-particle excitation $\ket{\psi_0}$ contributes in the sum (inserting identities $\sum_n \ket{E_n}\bra{E_n}$) in $A(\w)$. In ED, this is exploited by systematically constructing the subspace $\mathcal{N}$ by spanning it using particle-hole excitations , which might also be a viable route for further developments within DMRG . In DMRG, one needs to make a statement about the entanglement of the states in the subspace $\mathcal{N}$: one might note that these are in general more strongly entangled than the single-particle excitation $\ket{\psi_0}$, but should still be much less entangled than the rest of the Hilbertspace. This is illustrated in the sketch .
In , some of us argued that expanding the spectral function in a family of orthogonal functions is a natural way to construct a basis for $\mathcal{N}$, starting from the lowly entangled $\ket{\psi_0}$ and successively increasing entanglement of states and thereby computational complexity in a sequence of basis states $\ket{\psi_n}$. discussed the expansion of $A(\w)$ in Chebyshev polynomials $T_n(\tfrac{\w}{a})=\arccos(n\cos(\tfrac{\w}{a}))$, which are orthogonal with respect to an inner product weighted by $w(x)=\sqrt{1-x^2}$ , and in the plane waves $\exp(i\w \tfrac{n}{a})$ (orthogonal with weight function $w(x)=1$), where the energy $a$ is chosen larger than the support of $A(\w)$. The associated generated sequences of basis states then are and have different entanglement properties. The states $\ket{\psi_n\th{time}}$ associated with time evolution are in general less entangled than the states $\ket{\psi_n\th{che}}$ associated with the Chebyshev recursion . This is due to the observation that error accumulation in the Chebyshev recursion is worse conditioned than in time propagation , which necessitates to keep the error in a single step of the Chebyshev recursion much smaller than in the equivalent time evolution step, which in turn requires to use higher bond dimensions in the Chebyshev recursion making it less efficient. In addition to the statements of , we note here that the sequence produced by the Lanczos algorithm, can be associated with an expansion of the spectral function in polynomials that are orthogonal with respect to an inner product weighted by $w(x) = A(x)$ . This is very efficient but numerically unstable.
In contrast to the previous methods, which generate an increasingly complex basis when determining the spectral function to a higher and higher precision, correction-vector DMRG aims to optimize a state in frequency space, which contains contributions that have undergone an infinitely long time evolution. As time evolution creates entanglement, these states are much too strongly entangled for an efficient treatment. They are “far away” from the controlled, lowly entangled single-particle excitation $\ket{\psi_0}$. In order to still perform a meaningful computation in frequency space, one introduces a so-called (Lorentzian) broadening parameter $\eta$ that damps out contributions from an infinite time evolution. One does then not obtain the exact spectral function but a broadened version as in . The broadening parameter has to be guessed : If it is chosen too small, high entanglement prevents convergence of the calculation. If it is chosen too large, one will be far from the exact version of the spectral function. In the expansion methods discussed above, by contrast, one can stop the computation simply when it becomes too costly. If one has not recovered the exact $A(\w)$ at this point, a broadened version can be systematically constructed with an determined $\eta$ as in .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We exploit the critical structure on the Quot scheme $\Quot_{\BA^3}(\OO^{\oplus r},n)$, in particular the associated symmetric obstruction theory, in order to define rank $r$ *K-theoretic* Donaldson–Thomas invariants of the Calabi–Yau $3$-fold $\BA^3$. We compute the associated partition function as a plethystic exponential, proving a conjecture proposed in string theory by Awata–Kanno and Benini–Bonelli–Poggi–Tanzini. A crucial step in the proof is the fact that the invariants do not depend on the equivariant parameters of the framing torus $(\BC^\ast)^r$.
Reducing from K-theoretic to *cohomological* invariants, we compute the corresponding DT invariants, proving a conjecture of Szabo. Reducing further to *enumerative* DT invariants, we solve the higher rank DT theory of a pair $(X,F)$, where $F$ is an equivariant exceptional vector bundle on a projective toric $3$-fold $X$.
We give a mathematical definition of the chiral elliptic genus studied in physics by Benini–Bonelli–Poggi–Tanzini. This allows us to define *elliptic DT invariants* of $\BA^3$ in arbitrary rank, and to study their first properties.
address:
- 'SISSA Trieste, Via Bonomea 265, 34136 Trieste, ItalyInstitute for Geometry and Physics, Via Beirut 4, 34100 Trieste, ItalyIstituto Nazionale di Fisica Nucleare, Sezione di Trieste, Via Valerio 2, 34127 Trieste, Italy'
- 'Mathematical Institute, Utrecht University, P.O. Box 80010 3508 TA Utrecht, The Netherlands'
- 'SISSA Trieste, Via Bonomea 265, 34136 Trieste, ItalyInstitute for Geometry and Physics, via Beirut 4, 34100 Trieste, Italy'
author:
- Nadir Fasola
- Sergej Monavari
- 'Andrea T. Ricolfi'
bibliography:
- 'bib.bib'
title: 'Higher rank K-theoretic Donaldson–Thomas Theory of points'
---
[ ]{}
Introduction
============
Overview
--------
Classical Donaldson–Thomas (DT in short) invariants of a smooth complex projective Calabi–Yau $3$-fold $Y$, introduced in [@ThomasThesis], are integers that virtually count stable torsion free sheaves on $Y$, with fixed Chern character $\gamma$. However, the theory is much richer than what the bare DT numbers $$\DT(Y,\gamma)\,\in\,\BZ$$ can capture: there are extra symmetries subtly hidden in the local structure of the moduli spaces of sheaves giving rise to the classical DT invariants. This idea has been present in the physics literature for some time [@RefTopVertex; @RMQ].
The existence of these hidden symmetries suggests that there should exist more *refined* invariants, of which the DT numbers are just a shadow. These branch out in two main directions:
- *motivic* Donaldson–Thomas invariants, and
- *K-theoretic* Donaldson–Thomas invariants.
For the former, the papers by Kontsevich and Soibelman [@KS1; @KS2] are a good starting point. For the latter, see some recent developments after Nekrasov–Okounkov, e.g. [@NO16; @Okounkov_Lectures; @arbesfeld2019ktheoretic; @thomas2018equivariant] and [@nekrasov2017magnificent; @Magnificent_colors; @cao2019ktheoretic] for a generalisation to Calabi–Yau 4-folds. In this paper we deal with K-theoretic DT theory. The relationship between motivic and K-theoretic, which we briefly sketch in §\[motivic\_comparison\], will be investigated in future work.
The subtle structure of the DT moduli spaces is most evident in the *local case*, i.e. when the theory is applied to the simplest Calabi–Yau $3$-fold of all, namely the affine space $\BA^3$. See [@BBS; @DavisonR] for the rank $1$ motivic DT theory of $\BA^3$, and [@Quot19] for a higher rank version. This paper solves the K-theoretic Donaldson–Thomas theory of points of $\BA^3$. Part of the mathematical background needed to *define* the theory in rank $r>1$ is contained in [@BR18], where it is shown that the main player, the Quot scheme $$\label{quot}
\Quot_{\BA^3}(\OO^{\oplus r},n)$$ of length $n$ quotients of the free sheaf $\OO^{\oplus r}$, is a global *critical locus*, i.e. it can be realised as $\set{\dd f = 0}$, for $f$ a regular function on a smooth scheme. This structural result, revealing in bright light the symmetries we were talking about, is used to define the *higher rank K-theoretic DT theory of points* that is the central character in this paper. The case $r=1$, corresponding to $\Hilb^n(\BA^3)$, was already defined, and it was solved by Okounkov [@Okounkov_Lectures § 3], proving a conjecture by Nekrasov [@NEKRASOV2005261]. Our main result (Theorem \[mainthm:K-theoretic\]) can be seen as an upgrade of his calculation.
In physics, remarkably, the definition of the K-theoretic DT invariants studied here already existed, and gave rise to a conjecture that our paper — again, Theorem \[mainthm:K-theoretic\] — proves mathematically. More precisely, our formula for the K-theoretic DT partition function $\DT_r^{\KK}$ of $\BA^3$ was first conjectured by Nekrasov [@NEKRASOV2005261] for $r=1$ and by Awata and Kanno [@MR2545054] for arbitrary $r$. A formula expressing $\DT_r^{\KK}$ as a product of $r$ (suitably shifted) copies of $\DT_1^{\KK}$ was conjectured by Nekrasov and Piazzalunga [@Magnificent_colors § 3] as a limit of the $4$-fold theory; we believe it is closely related to the product formula, displayed in , that we obtain in Theorem \[thm: r copies of rank 1 K theory\]. The reader is referred to §\[sec:string\_stuff\] of this introduction for more background on the physics picture.
We also study higher rank *cohomological DT invariants* of $\BA^3$. As we show in Corollary \[limit K theory cor\], these can be obtained as a suitable limit of the K-theoretic invariants. Motivated by [@MR2545054; @Nekrasov_M-theory], a closed formula for their generating function $\DT_r^{\coh}$ was conjectured by Szabo [@Szabo Conj. 4.10] as a generalisation of the $r=1$ case covered by Maulik–Nekrasov–Okounkov–Pandharipande [@MNOP2 Thm 1]. We prove this conjecture as our Theorem \[mainthm:cohomological\]. To get there, in §\[sec:higher\_rank\_vertex\] we develop a *higher rank topological vertex* formalism based on the combinatorics of *colored plane partitions*, generalising the classical vertex formalism of [@MNOP1; @MNOP2].
We pause for a second to explain a key step in this paper. The Quot scheme , which gives rise to most of the invariants we study here, is acted on by an algebraic torus $$\TT = (\BC^*)^3 \times (\BC^*)^r,$$ and by their very definition, both the K-theoretic and the cohomological DT invariants depend, a priori, on the sets $t = (t_1,t_2,t_3)$ and $w = (w_1,\ldots,w_r)$ of equivariant parameters of $\TT$. The main technical result of this paper, which is proved as Theorem \[independence of Z\^K on w\] in the Appendix, states that $$\label{independence_slogan}
\textrm{The K-theoretic DT invariants do not depend on the framing parameters }w.$$ This is what allows us, for instance, to take arbitrary limits to evaluate our formulae. A different proof of Theorem \[independence of Z\^K on w\] is obtained by Arbesfeld–Kononov in [@Noah_Yasha].
Main results
------------
We briefly outline here the main results obtained in this paper.
### K-theoretic DT invariants
As we mentioned above, the Quot scheme is a critical locus, thus it carries a natural symmetric perfect obstruction theory in the sense of Behrend–Fantechi [@BFinc; @BFHilb]. As we recall in §\[sec: twisted structure sheaf\], there is also an induced *twisted virtual structure sheaf* $\widehat{\OO}^{\vir} \in K_0(\Quot_{\BA^3}(\OO^{\oplus r},n))$, which is a twist by an orientation (i.e. a square root of the virtual canonical bundle) of the ordinary virtual structure sheaf $\OO^{\vir}$. The rank $r$ *K-theoretic DT partition function* of the Quot scheme of points of $\BA^3$, encodes the rank $r$ K-theoretic DT invariants of $\BA^3$, and is defined as $$\DT_r^{\KK}(\BA^3,q,t,w) = \sum_{n\geq 0} q^n \chi(\Quot_{\BA^3}(\OO^{\oplus r},n),\widehat{\OO}^{\vir}) \in \BZ(\!(t,(t_1t_2t_3)^{\frac{1}{2}},w)\!)\llbracket q \rrbracket,$$ where the half power is caused by the twist by the chosen orientation (this choice does not affect the invariants, cf. Remark \[rmk:independence\_of\_orientation\]).
Granting Theorem \[independence of Z\^K on w\], that we stated informally in , we shall write $\DT_r^{\KK}(\BA^3,q,t) = \DT_r^{\KK}(\BA^3,q,t,w)$, ignoring the framing parameters $w$. In §\[sec:proof\_of\_K-theoretic\_thm\] we determine a closed formula for this series, proving the conjecture by Awata–Kanno [@MR2545054]. This conjecture was checked for low number of instantons in [@MR2545054 § 4].
To state our first main result, we need to recall the definition of the *plethystic exponential*. Given an arbitrary power series $$f = f(u_1,\ldots,u_\ell) = \sum_{n_1,\ldots,n_\ell>0}a_{n_1,\ldots,n_\ell}u_1^{n_1}\cdots u_\ell^{n_\ell}
\,\in\,\BZ\llbracket u_1,\ldots,u_\ell \rrbracket,$$ one sets $$\label{def_Exp}
\Exp\left(f\right) = \exp\left(\sum_{n>0}\frac{f(u_1^n,\ldots,u_\ell^n)}{n}\right) = \prod_{n_1,\ldots,n_\ell>0}\frac{1}{(1-u_1^{n_1}\cdots u_\ell^{n_\ell})^{a_{n_1,\ldots,n_\ell}}}.$$ Consider, for a formal variable $x$, the operator $[x] = x^{1/2} - x^{-1/2}$. In §\[sec: preliminaries K theory\] we consider this operator on $K_0^{\TT}(\pt)$. See §\[subsec:Witten\] for its physical interpretation. We are now able to state our first main result.
\[mainthm:K-theoretic\] The rank $r$ K-theoretic DT partition function of $\BA^3$ is given by $$\label{eqn:DT^K}
\DT_r^{\KK}(\BA^3,(-1)^rq,t)=\Exp\left(\mathsf F_r( q,t_1,t_2,t_3)\right),$$ where, setting $\mathfrak{t} = t_1t_2t_3$, one defines $$\mathsf F_r( q,t_1,t_2,t_3)= \frac{[\mathfrak{t}^r]}{[\mathfrak{t}][\mathfrak{t}^{\frac{r}{2}}q ][\mathfrak{t}^{\frac{r}{2}} q^{-1} ]}\frac{[t_1t_2][t_1t_3][t_2t_3]}{[t_1][t_2][t_3]}.$$
The case $r=1$ of Theorem \[mainthm:K-theoretic\] was proved by Okounkov in [@Okounkov_Lectures]. The general case was proposed conjecturally in [@MR2545054; @BBPT]. N. Arbesfeld and Ya. Kononov informed us during the writing of this paper that they also obtained a proof of Theorem \[mainthm:K-theoretic\] [@Noah_Yasha].
It is interesting to notice that Formula is equivalent to the product decomposition $$\label{eqn:product}
\DT_r^{\KK}(\BA^3,(-1)^rq,t)=\prod_{i=1}^r \DT_1^{\KK}(\BA^3,-q\mathfrak{t}^{\frac{-r-1}{2} +i},t),$$ that we obtain in Theorem \[thm: r copies of rank 1 K theory\]. This formula is likely related to the product formula [@Magnificent_colors Formula (3.14)] appearing as a limit of the $4$-fold theory developed by Nekrasov and Piazzalunga.
Formula is also related to its motivic cousin: as we observe in §\[motivic\_comparison\], the motivic partition function $\DT_r^{\mot}$ of the Quot scheme of points of $\BA^3$ (see [@BR18 Prop. 2.7] and the references therein) satisfies the same product formula, after the transformation $\mathfrak t^{1/2} \to - \BL^{1/2}$.
### Cohomological DT invariants
The generating function of *cohomological DT invariants* is defined as $$\DT_r^{\coh}(\mathbb{A}^3,q,s,v) = \sum_{n\geq 0}q^n \int_{[\Quot_{\BA^3}(\OO^{\oplus r},n)]^{\vir}} 1$$ where $s=(s_1,s_2,s_3)$ and $v = (v_1,\ldots,v_r)$, with $s_i = c_1^{\TT}(t_i)$ and $v_j = c_1^{\TT}(w_j)$ respectively, and the integral is defined via $\TT$-equivariant residues. It is a consequence of that $\DT_r^{\coh}(\mathbb{A}^3,q,s,v)$ does not depend on $v$, so we will shorten it as $\DT_r^{\coh}(\mathbb{A}^3,q,s)$. In §\[sec: cohom reduction\] we explain how to recover the cohomological invariants out of the K-theoretic ones. This is the limit formula (Corollary \[limit K theory cor\]) $$\DT_r^{\coh}(\BA^3,q,s) = \lim_{b \to 0} \DT_r^{\KK}(\BA^3,q,e^{bs}),$$ essentially a formal consequence of our explicit expression for the K-theoretic *higher rank vertex* (cf. §\[sec:higher\_rank\_vertex\]) attached to the Quot scheme $\Quot_{\BA^3}(\OO^{\oplus r},n)$.
\[mainthm:cohomological\] The generating series of rank $r$ cohomological Donaldson–Thomas invariants of $\BA^3$ is given by $$\DT_r^{\coh}(\mathbb{A}^3,q,s)=\mathsf M((-1)^rq)^{-r\frac{(s_1+s_2)(s_1+s_3)(s_2+s_3)}{s_1s_2s_3}},$$ where $\mathsf M(t) = \prod_{m\geq 1}(1-t^m)^{-m}$ is the MacMahon function.
The case $r=1$ of Theorem \[mainthm:cohomological\] was proved by Maulik–Nekrasov–Okounkov–Pandharipande [@MNOP2 Thm. 1]. Theorem \[mainthm:cohomological\] was conjectured in [@Nekrasov_M-theory; @Szabo] and confirmed for $r\leq 8$ and $n\leq 8$ in [@BBPT]. The specialisation $\DT_r^{\coh}(\mathbb{A}^3,q,s)\big|_{s_1+s_2+s_3=0}=\mathsf M((-1)^rq)^r$ was already computed in physics [@CSS].
### Elliptic DT invariants
In §\[sec:elliptic\_invariants\] we define the *virtual chiral elliptic genus* and we introduce a refinement $\DT_r^{\rm ell}$ of the generating series $\DT_r^{\KK}$, which can be seen as a mathematical definition of the *elliptic DT invariants* studied in [@BBPT]. We propose a conjecture (Conjecture \[conj: non dependence under some limits of elliptic\]) about the behaviour of $\DT_r^{\rm ell}$ and, granting this conjecture, we prove a conjecture formulated by Benini–Bonelli–Poggi–Tanzini (Theorem \[thm:elliptic\]).
### Global geometry
So far we have only discussed results concerning *local* geometry. When $X$ is a *projective* toric $3$-fold and $F$ is an equivariant exceptional locally free sheaf, by [@Quot19 Thm. A] there is a $0$-dimensional torus equivariant (cf. Proposition \[pot\_global\_equivariant\]) perfect obstruction theory on $\Quot_X(F,n)$. Therefore the higher rank Donaldson–Thomas invariants $$\DT_{F,n} = \int_{[\Quot_X(F,n)]^{\vir}}1 \,\in \,\BZ$$ can be computed via the Graber–Pandharipande virtual localisation formula [@GPvirtual]. The next result confirms (in the toric case) a prediction [@Quot19 Conj. 3.5] for their generating function.
\[mainthm:projective\_toric\] Let $(X,F)$ be a pair consisting of a smooth projective toric $3$-fold $X$ along with an exceptional equivariant locally free sheaf $F$ of rank $r$. Then $$\sum_{n\geq 0} \DT_{F,n}q^n= \mathsf M((-1)^rq)^{r\int_{X}c_3(T_X\otimes K_X)}.$$
The corresponding formula in the Calabi–Yau case was proved in [@Quot19 §3.2], whereas the general rank $1$ case was proved in [@MNOP2 Thm. 2] and [@JLI Thm. 0.2].
Relation to string theory {#sec:string_stuff}
-------------------------
An interpretation for Donaldson–Thomas invariants is available also in the context of supersymmetric string theories. In this framework, one is interested in countings of BPS-bound states on (Calabi–Yau) $3$-folds. The interest in studying the BPS sector of string theories lies in the fact the it consists of quantities which are usually protected from quantum corrections, and which can sometimes be studied non-perturbatively. On the other hand, BPS countings have been shown in many occasions to have a precise mathematical interpretation rooted in counting problems in enumerative geometry. In the case of Donaldson–Thomas theory the interpretation is that of a type IIA theory compactified on $X$, so that the four dimensional effective theory has an $\mathcal N=2$ supersymmetry content. BPS states then preserve half of this supersymmetry and are indexed by a charge vector living on a lattice determined by the cohomology (with compact support) of the $3$-fold $\Gamma\cong H^0(X,\BZ)\oplus H^2(X,\BZ)\oplus H^4(X,\BZ)\oplus H^6(X,\BZ)$. These lattices can also be interpreted as the charge lattices of D-branes wrapping $p$-cycles on $X$, where $${\rm D}p \longleftrightarrow H^{6-p}(X,\BZ)\,\cong\, H_p(X,\BZ),$$ for $p=0,2,4,6$. The Witten index $$\Ind_X(\gamma)=\Tr_{\mathcal H^{(\gamma)}_{{\rm BPS}}}(-1)^F$$ provides a measure of the degeneracy of the BPS states, where the trace is over the fixed charge sector of the single-particle Hilbert space $H_{{\rm BPS}}=\bigoplus_{\gamma\in\Gamma}\mathcal H^{(\gamma)}_{\rm BPS}$ and $F$ is a given one–particle operator acting on the $\gamma$-component of the Hilbert space. The choice of a charge vector of the form $\gamma=(r,0,-\beta,n)$ is then equivalent to a system of (D0-D4-D6)-branes, where $r$ D6-branes wrap the whole $3$-fold $X$. The index of the theory is then defined via integration over the virtual class of the moduli space of BPS states $\mathcal M_{\rm BPS}^{(\gamma)}(X)$, $$\Ind_X(\gamma)=\int_{[\mathcal M_{\rm BPS}^{(\gamma)}(X)]^{\vir}}1.$$ The moduli space of BPS states on $X=\BA^3$ with charge vector $\gamma=(r,0,0,n)$ is then identified with the Quot-scheme $\Quot_{\BA^3}(\OO^{\oplus r},n)$, and the partition function of the theory reproduces the generating function of degree 0 DT invariants on $\BA^3$ $$\sum_{n\ge 0}\Ind_{\BA^3}(r,0,0,n)q^n=\mathsf{DT}_r^{\coh}(\BA^3,q,s).$$ An equivalent interpretation can be also given in type IIB theories, where the relevant systems will be those of (D(-1)-D3-D5)-branes. The effective theory on the D(-1)-branes, which in this case are $0$-dimensional objects, is a quiver matrix model, which encodes the critical structure of $\Quot_{\BA^3}(\OO^{\oplus r},n)$. K-theoretic and elliptic versions of DT invariants can also be studied by suitably generalising the D-branes construction. One can study for instance a D0-D6 brane system with $r$ branes on $X\times S^1$, $S^1$ being the worldvolume of the D0-branes. In this case the D0-branes quantum mechanics describe the K-theoretic generalisation of DT theory, so that the supersymmetric partition function computes the equivariant Euler characteristic of the twisted virtual structure sheaf $\widehat{\OO}^{\vir}$ of Quot scheme identified with the moduli space of BPS vacua of the theory. Analogously, the D1 string theory of a D1-D7 system on the product of $X$ by an elliptic curve, say $X\times T^2$, provides the elliptic generalisation of higher rank DT invariants, [@BBPT], while the superconformal index realises the virtual elliptic genus of the moduli space of BPS states. In this context, a plethystic formula for the generating function of the K-theoretic DT invariants was conjectured in [@Nekrasov_M-theory; @MR2545054] for rank $1$ and $r$, respectively. The elliptic generalisation of DT theory was studied in [@BBPT], where a plethystic formula for the virtual elliptic genus was conjectured in the Calabi–Yau specialisation $t_1t_2t_3=1$ and its generalisation $(t_1t_2t_3)^r=1$ (cf. Theorem \[thm:elliptic\]).
Related works on virtual invariants of Quot schemes
---------------------------------------------------
The study of virtual invariants of Quot schemes is an active research area in enumerative geometry. See for instance the seminal work of Marian–Oprea [@MR2271296], where a virtual fundamental class on the Quot scheme of the trivial bundle on a smooth projective curve was constructed. More recently, there has been a lot of activity on surfaces, partially motivated by Vafa–Witten: tautological integrals over Quot schemes were computed by Oprea–Pandharipande in [@Oprea:2019ab], where the case of $1$-dimensional quotients was also studied, in connection with Seiberg–Witten invariants. Rationality questions of various descendent series with insertions have been investigated by Johnson–Oprea–Pandharipande in [@Johnson:2020aa]. Very recently, Lim [@Lim:2020aa] studied virtual $\chi_{-y}$-genera of Quot schemes on surfaces, presenting a universal formula in terms of Seiberg–Witten invariants. In connection to physics, flags of framed torsion-free sheaves on $\BP^2$, as a generalisation of Quot schemes of points on $\BA^2$, were studied in [@bonelli2019defects; @bonelli2019flags]. Many of the results mentioned above use virtual localisation to study virtual invariants of Quot schemes: our paper exploits virtual localisation as well, though in the K-theoretic setup. Another remarkable application of localisation of virtual classes of Quot schemes (on K3 surfaces) is the work of Marian–Oprea–Pandharipande [@MR3621431], where a conjecture formulated by M. Lehn in 1999 is proved exploiting such techniques.
Generating functions of Euler characteristics of Quot schemes on $3$-folds were computed by Gholampour–Kool in [@Gholampour2017] for quotients of sheaves of homological dimension at most $1$. The wall-crossing type formula obtained in loc. cit. was interpreted, in the Calabi–Yau case, as an Euler characteristic shadow of a DT/PT Hall algebra identity by Beentjes and the third author [@BR18]. Quot schemes of *reflexive* sheaves (of rank $2$) also appeared in the work of Gholampour–Kool–Young [@Gholampour2017a], as fibres of “double dual maps”, exploited to compute generating functions of Euler characteristics of more complicated moduli spaces of (stable, torsion free) sheaves. Virtual invariants are also defined (again, in the rank $2$ case) via localisation in [@Gholampour2017a]. In the rank $1$ case, quotients of the ideal sheaf of a smooth curve $C$ in a $3$-fold $Y$ form a closed subscheme of the Hilbert scheme of curves in $Y$, the key player in rank $1$ DT theory. The series of virtual Euler characteristics of the associated Quot scheme was computed by the third author in [@LocalDT], and later upgraded to a $C$-local DT/PT wall-crossing formula [@Ricolfi2018]. The motivic refinement of this DT/PT correspondence was established by Davison and the third author in [@DavisonR].
Virtual classes of dimension $0$ of Quot schemes (for locally free sheaves) on projective $3$-folds are constructed under certain assumptions in [@Quot19]. In this paper we work also with the virtual class on $\Quot_{\BA^3}(\OO^{\oplus r},n)$ coming from the critical obstruction theory found in [@BR18].
Plan of the paper
-----------------
In §\[sec:background\] we recall the basics on perfect obstruction theories, virtual classes, virtual structure sheaves and how to produce virtual invariants out of these data; we review the K-theoretic virtual localisation theorem in §\[sec:vir\_loc\]. Sections \[sec: quot scheme local model\]–\[sec:elliptic\_invariants\] are devoted to the “local Quot scheme” $\Quot_{\BA^3}(\OO^{\oplus r},n)$. In §\[sec: quot scheme local model\] we recall its critical structure and we define a $\TT$-action on it, whose fixed locus is parametrised by the finitely many $r$-colored plane partitions (Proposition \[prop:fixedlocus\_indexed\_by\_colored\_partitions\]); we study the equivariant critical obstruction theory on the Quot scheme and prove that the induced virtual class on the $\TT$-fixed locus is trivial (Proposition \[trivial pot on fixed locus\]). In §\[subsec:virtual\_invariants\_quot\] we introduce cohomological and K-theoretic DT invariants of $\Quot_{\BA^3}(\OO^{\oplus r},n)$. In §\[sec:higher\_rank\_vertex\] we develop a higher rank vertex formalism which we exploit to write down a formula (Proposition \[prop:Tangent\^vir\]) for the virtual tangent space of a $\TT$-fixed point in the Quot scheme. In §\[sec:K-theory\] we prove Theorem \[mainthm:K-theoretic\] as well as Formula . In §\[sec:cohomological\_invariants\] we prove Theorem \[mainthm:cohomological\] and we show that $\DT_r^{\coh}$ does not depend on any choice of possibly nontrivial $(\BC^\ast)^3$-weights on $\OO^{\oplus r}$. In §\[sec:elliptic\_invariants\], we give a mathematically rigorous definition of a “chiral” version of the virtual elliptic genus of [@Fantechi_Gottsche] and use it in § \[sec:elliptic DT\] to define elliptic DT invariants. In § \[sec:limits of elliptic DT\] we also give closed formulas for elliptic DT invariants in some limiting cases, based on the conjectural independence on the elliptic parameter — see Conjecture \[conj: non dependence under some limits of elliptic\] and Remark \[rem: elliptic indepedence\]. In particular Theorem \[thm:elliptic\] proves a conjecture recently appeared in the physical literature [@BBPT Formula (3.20)]. In §\[sec:compact\_section\] we prove Theorem \[mainthm:projective\_toric\] by gluing vertex contributions from the toric charts of a projective toric $3$-fold. This is made possible by a comparison of global and local equivariant obstruction theories and virtual classes — see Proposition \[restriction of class in K theory\] and Corollary \[cor: pot as box product\]. Finally, Appendix \[sec:framing\_independence\] contains the proof of the slogan , exploiting a functional representation of $\DT_r^{\KK}$ in terms of Jeffrey–Kirwan residues — see Lemma \[contour integration is equivalent to K theory invariants\] — and an explicit characterisation of the poles of the partition function.
We work over $\BC$. A *scheme* is a separated scheme of finite type over $\BC$. If $Y$ is a scheme, we let $\derived^{[a,b]}(Y)$ denote the derived category of coherent sheaves on $Y$, whose objects are complexes with vanishing cohomology sheaves outside the interval $[a,b]$. We let $K^0(Y)$ be the K-group of vector bundles on $Y$. When $Y$ carries an action by an algebraic torus $\TT$, we let $K^0_\TT(Y)$ be the K-group of $\TT$-*equivariant* vector bundles on $Y$. Similarly, we let $K_0(Y)$ denote the K-group of *coherent sheaves* on $Y$, and we let $K_0^{\TT}(Y)$ be the K-group of (the abelian category of) $\TT$-*equivariant* coherent sheaves on $Y$. When $Y$ is smooth, the natural $\BZ$-linear map $K^0(Y) \to K_0(Y)$, resp. $K^0_{\TT}(\pt)$-linear map $K^0_{\TT}(Y) \to K_0^{\TT}(Y)$, is an isomorphism. Chow groups $A^\ast (Y)$ and cohomology groups $H^\ast(Y)$ are taken with rational coefficients.
Acknowledgements {#acknowledgements .unnumbered}
----------------
We thank Noah Arbesfeld and Yakov Kononov for generously sharing with us their progress on their paper [@Noah_Yasha], for reading a first draft of this paper and for sending us very interesting comments. Thanks to Alberto Cazzaniga for suggesting the comparison between motivic and K-theoretic factorisations. We thank Barbara Fantechi and Richard Thomas for enlightening discussions about equivariant obstruction theories. We are grateful to Martijn Kool, Giulio Bonelli and Alessandro Tanzini for the many discussions, insightful suggestions and for commenting on a first draft of this paper.
Background material {#sec:background}
===================
Obstruction theories and virtual classes
----------------------------------------
A *perfect obstruction theory* on a scheme $X$, as defined in [@LiTian; @BFinc], is the datum of a morphism $$\phi\colon \BE\to\BL_X$$ in $\derived^{[-1,0]}(X)$, where $\BE$ is a perfect complex of perfect amplitude contained in $[-1,0]$, such that $h^0(\phi)$ is an isomorphism and $h^{-1}(\phi)$ is surjective. Here, $\BL_X = \tau_{\geq -1}L_X^\bullet$ is the cut-off at $-1$ of the full cotangent complex $L_X^\bullet \in \derived^{[-\infty,0]}(X)$ defined by Illusie [@Ill]. A perfect obstruction theory is called *symmetric* (see [@BFHilb]) if there exists an isomorphism $\theta\colon \BE \simto \BE^\vee[1]$ such that $\theta = \theta^\vee[1]$.
The *virtual dimension* of $X$ with respect to $(\BE,\phi)$ is the integer $\vd = \rk \BE$. This is just $\rk E^0 - \rk E^{-1}$ if one can write $\BE = [E^{-1}\to E^0]$.
\[rmk:K-class\_of\_symmetric\_pot\] A symmetric obstruction theory $\BE \to \BL_X$ has virtual dimension $0$, and moreover the *obstruction sheaf*, defined as $\Ob = h^1(\BE^\vee)$, is canonically isomorphic to the cotangent sheaf $\Omega_X$. In particular, one has the K-theoretic identity $$\BE = h^0(\BE) - h^{-1}(\BE) = \Omega_X - T_X\,\in\,K^0(X).$$
A perfect obstruction theory determines a cone $$\mathfrak C \into E_1 = (E^{-1})^\vee.$$ Letting $\iota\colon X \into E_1$ be the zero section of the vector bundle $E_1$, the induced *virtual fundamental class* on $X$ is the refined intersection $$[X]^{\vir} = \iota^![\mathfrak C]\, \in\, A_{\vd}(X).$$ By a result of Siebert [@Siebert Thm. 4.6], the virtual fundamental class depends only on the K-theory class of $\BE$.
\[def:equivariant\_pot\] Let $G$ be an algebraic group acting on a scheme $X$. A perfect obstruction theory $\phi\colon \BE \to \BL_X$ is $G$-*equivariant* if $\phi$ admits a lift to a morphism in the derived category $\derived^{[-1,0]}(\Coh_X^G)$ of $G$-equivariant coherent sheaves.
From obstruction theories to virtual invariants
-----------------------------------------------
On a proper scheme $X$ with a perfect obstruction theory, one can define virtual enumerative invariants by $$\int_{[X]^{\vir}}\alpha \,\in\, \BQ,$$ where $\alpha\in A^i (X)$. These intersection numbers are going to vanish if $i \neq \vd$.
On the K-theoretic side, it was observed in [@BFinc §5.4] that a perfect obstruction theory $\BE \to \BL_X$ not only induces a virtual fundamental class, but also a *virtual structure sheaf* $$\OO_X^{\vir} = [\mathbf L \iota^\ast \OO_{\mathfrak C}]\,\in\, K_0(X).$$ Its construction first appeared in [@kontsevich_94; @kapranov_2009] in the context of dg-manifolds and in [@BFinc; @Fantechi_Gottsche] in the language of perfect obstruction theories. More recently, Thomas gave a description of $\OO_X^{\vir}$ in terms of the *$K$-theoretic Fulton class*, showing that it only depends on the K-theory class of $\BE$ [@thomas2018ktheoretic Cor. 4.5]. If $\pi\colon X \to \pt$ is proper, one can use $\OO_X^{\vir}$ and K-theoretic pushfoward $\pi_\ast = \chi(X,-)$ to define virtual invariants by $$\chi^{\vir}(X,V)=\chi(X, V\otimes \OO_X^{\vir})\, \in\, K_0(\pt) = \mathbb{Z},\quad V\in K^0(X).$$ A virtual version [@Fantechi_Gottsche Cor. 3.6] of the Hirzebruch–Riemann–Roch theorem holds: one has $$\chi^{\vir}(X,V)={\int_{[X]^{\vir}}} \ch(V)\cdot \td(T_X^{\vir}),$$ where, setting $E_i = (E^{-i})^\vee$, one defines the *virtual tangent bundle* of $X$ by the formula $$T^{\vir}_X = \BE^{\vee} = E_0 - E_1\,\in\,K^0(X).$$
Torus representations and their weights
---------------------------------------
Let $\TT = (\BC^*)^g$ be an algebraic torus, with character lattice $\widehat{\TT} = \Hom(\TT,\BC^\ast) \cong \BZ^g$. Let $K_0^{\TT}(\pt)$ be the K-group of the category of $\TT$-representations. Any finite dimensional $\TT$-representation $V$ splits as a sum of $1$-dimensional representations called the *weights* of $V$. Each weight corresponds to a character $\mu \in \widehat{\TT}$, and in turn each character corresponds to a monomial $t^\mu = t_1^{\mu_1}\cdots t_g^{\mu_g}$ in the coordinates of $\TT$. The map $$\label{eqn:trace}
\tr\colon K_0^{\TT}(\pt) \to \BZ \left[t^\mu \mid \mu \in \widehat{\TT}\right],\quad V\mapsto \tr_V,$$ sending the class of a $\TT$-module to its decomposition into weight spaces is a ring isomorphism, where tensor product on the left corresponds to the natural multiplication on the right. We will therefore sometimes identify a (virtual) $\TT$-module with its character. Sometimes, to ease notation, we shall write $\tr(V)$ instead of $\tr_V$.
Let $V=\sum_\mu t^\mu$ be a $\TT$-module. Define $\Lambda^\bullet_p V = \sum_{i=0}^{\rk V}p^i\Lambda^i V$ to be the *total wedge of* $V$. We shall write $\Lambda^\bullet V = \Lambda^\bullet_{-1} V$. As shown for instance in [@Okounkov_Lectures Ex. 2.1.5], its trace satisfies $$\tr_{\Lambda^\bullet_{-p} V}=\prod_\mu\,(1-pt^\mu).$$
Virtual normal bundle and virtual tangent space {#subsec:Nvir}
-----------------------------------------------
Let $\TT = (\BC^\ast)^g$ be an algebraic torus. If $Y$ is a scheme carrying the trivial $\TT$-action, any coherent sheaf $B \in \Coh Y$ admits a direct sum decomposition $B = \bigoplus_{\mu} B^{\mu}$ into eigensheaves, where $\mu \in \widehat \TT \cong \BZ^g$ ranges over the characters of $\TT$. The $\TT$-*fixed part* and the $\TT$-*moving part* of $B$ are defined as $$B^{\fix} = B^0,\quad B^{\mov} = \bigoplus_{\mu\neq 0}B^{\mu}.$$ This definition extends to complexes of coherent sheaves.
If $X$ is a scheme carrying a $\TT$-action and a $\TT$-equivariant perfect obstruction theory $\BE \to \BL_{X}$, and $Y \subset X^{\TT}$ is a component of the fixed locus, then the *virtual normal bundle* of $Y\subset X$ is the complex $$N^{\vir}_{Y/X}=T_X^{\vir}\big|_{Y}^{\mov} = \BE^\vee \big|_{Y}^{\mov}.$$ If $p \in X^{\TT}$ is a fixed point, the *virtual tangent space* of $X$ at $p$ is $T_p^{\vir} = \BE^\vee \big|_{p} \in K_0^{\TT}(\pt)$.
Virtual localisation {#sec:vir_loc}
--------------------
A very useful tool to compute virtual K-theoretic invariants is the *virtual localisation theorem*. Its first version was proven in equivariant Chow theory in [@GPvirtual], and a K-theoretic version appeared in [@Fantechi_Gottsche § 7] and [@Qu_virtual § 3]. Suppose that $X$ carries an action of an algebraic torus $\TT$ and a $\TT$-equivariant perfect obstruction theory. Then $\OO_X^{\vir}$ is naturally an element of $K_0^{\TT}(X)$ and if $X$ is proper one can define $\chi^{\vir}(X,-) = \chi(X,-\otimes \OO_X^{\vir})$ by means of the equivariant pushforward $\chi(X,-)\colon K_0^{\TT}(X)\to K_0^{\TT}(\pt)$. The K-theoretic virtual localisation formula states that, for an arbitrary element $V\in K_\TT^0(X)$ of the $\TT$-equivariant K-theory, one has the identity $$\label{K-theoretic_localisation}
\chi^{\vir}(X,V)=\chi^{\vir}\left(X^\TT, \frac{V|_{X^\TT}}{\Lambda^\bullet N^{\vir,\vee}} \right)\,\,\in\,\, K^\TT_0(\pt)\left[\frac{1}{1-t^{\mu}}\, \Bigg{|}\, \mu \in \widehat{\TT}\right],$$ where $X^\TT\subset X$ is the $\TT$-fixed locus, $N^{\vir} \in K^0_{\TT}(X^\TT)$ is K-theory class of the virtual normal bundle, i.e. of the $\TT$-moving part of $T_X^{\vir}=\BE^\vee$ restricted to $X^{\TT}$.
More than just being a powerful theorem, the localisation formula allows one to define invariants for *quasi-projective* $\TT$-varieties, provided that they have proper $\TT$-fixed locus: if this is the case, one defines $\chi^{\vir}(X,V)$ to be the right hand side of . If $X$ is proper, this definition coincides with the usual one thanks to the localisation theorem.
We thank Noah Arbesfeld for pointing out to us that for quasi-projective schemes, in contrast to the case of equivariant *cohomology*, one can directly define the equivariant Euler characteristic of a coherent sheaf via its character as a torus representation, without invoking virtual localisation.
The local Quot scheme: critical and equivariant structure {#sec: quot scheme local model}
=========================================================
Overview
--------
In this section we start working on the local Calabi–Yau $3$-fold $\BA^3$. Fix integers $r\geq 1$ and $n\geq 0$. Our focus will be on the *local Quot scheme* $$\Quot_{\BA^3}(\mathscr O^{\oplus r},n),$$ whose points correspond to short exact sequences $$0\to S\to \mathscr O^{\oplus r}\to T \to 0$$ where $T$ is a $0$-dimensional $\OO_{\BA^3}$-module with $\chi(T)=n$.
We shall use the following notation throughout.
If $F$ is a locally free sheaf on a variety $X$, and $F\onto T$ is a surjection onto a $0$-dimensional sheaf of length $n$, with kernel $S\subset F$, we denote by $$[S] \in \Quot_{X}(F,n)$$ the corresponding point in the Quot scheme.
In this section, we will:
- recall from [@BR18] the description of the Quot scheme as a critical locus (§\[sec:critical\_structure\_Quot\]),
- describe a $\TT$-action (for $\TT = (\BC^*)^3\times (\BC^*)^r$ a torus of dimension $3+r$) on $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)$, with isolated fixed locus consisting of direct sums of monomial ideals (§\[sec:torus\_actions\]),
- reinterpret the fixed locus $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\TT}$ in terms of colored partitions (§\[sec: combinatorial description of fixed locus\]),
- prove that the critical perfect obstruction theory on $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)$ is $\TT$-equivariant (Lemma \[lemma:T\_equivariance\_of\_POT\]), and that the induced $\TT$-fixed obstruction theory on the fixed locus is trivial (Proposition \[trivial pot on fixed locus\]).
The content of this section is the starting point for the definition (see §\[subsec:virtual\_invariants\_quot\]) of virtual invariants on $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)$, as well as our construction (see §\[sec:higher\_rank\_vertex\]) of the *higher rank vertex formalism*.
The critical structure on the Quot scheme {#sec:critical_structure_Quot}
-----------------------------------------
Let $V$ be an $n$-dimensional complex vector space. Consider the space $R_{r,n} = \Rep_{(n,1)}(\widetilde{\mathsf L}_3)$ of $r$-framed $(n,1)$-dimensional representations of the $3$-loop quiver $\mathsf L_3$, depicted in Figure \[fig:3loopquiver\_framed\]. The notation “$(n,1)$” means that the main vertex (the one belonging to the $3$-loop quiver, labelled “$0$” in the figure) carries a copy of $V$, whereas the framing vertex (labelled “$\infty$”) carries a copy of $\BC$.
; =\[font=*\] (B) edge \[loop above\] node [$A_1$]{} () edge \[loop right\] node [$A_2$]{} () edge \[loop below\] node [$A_3$]{} ();*
at (-0.2,0.1) [$\vdots$]{}; at (-0.3,0.95) [$u_1$]{}; at (-0.3,-0.85) [$u_r$]{}; (A) to \[bend left=25,looseness=1\] (B) node \[midway,above\] ; (A) to \[bend left=40,looseness=1\] (B) node \[midway\] ; (A) to \[bend right=35,looseness=1\] (B) node \[midway,below\] ;
We have that $R_{r,n}$ is an affine space of dimension $3n^2+rn$, with an explicit description as $$\begin{aligned}
R_{r,n} &= \Set{(A_1,A_2,A_3,u_1,\ldots,u_r)|A_j \in \End(V),\,u_i \in V} \\
&=\End(V)^{\oplus 3}\oplus V^{\oplus r}.\end{aligned}$$ By [@BR18 Prop. 2.4], there exists a stability parameter $\theta$ on the $3$-loop quiver such that $\theta$-stable framed representations $(A_1,A_2,A_3,u_1,\ldots,u_r) \in R_{r,n}$ are precisely those satisfying the condition: $$\textrm{the vectors }u_1,\ldots,u_r \in V \textrm{ jointly generate }(A_1,A_2,A_3) \in \Rep_n (\mathsf L_3).$$ Imposing this stability condition on $R_{r,n}$ we obtain an open subscheme $$U_{r,n} \subset R_{r,n}$$ on which $\GL(V)$ acts freely by the rule $$g \cdot (A_1,A_2,A_3,u_1,\ldots,u_r) = (A_1^g,A_2^g,A_3^g,gu_1,\ldots,gu_r),$$ where $A_i^g = gA_ig^{-1}$ denotes conjugation by $g\in \GL(V)$. The quotient $$\label{def:NCquot}
\NCQuot_{r}^n = U_{r,n}/\GL(V)$$ is a smooth quasi-projective variety of dimension $2n^2+rn$. In [@BR18] the scheme $\NCQuot_{r}^n$ is referred to as the *non-commutative Quot scheme*, by analogy with the *non-commutative Hilbert scheme* [@Nori1], i.e. the moduli space of left ideals of codimension $n$ in $\BC\langle x_1,x_2,x_3\rangle$ (which of course exists for an arbitrary number of free variables).
On $R_{r,n}$ one can define the function $$h_n\colon R_{r,n} \to \BA^1, \quad (A_1,A_2,A_3,u_1,\ldots,u_r) \to \Tr A_1[A_2,A_3],$$ induced by the superpotential $\mathsf W=A_1[A_2,A_3]$ on the $3$-loop quiver. Note that this function
- is symmetric under cyclic permutations of $A_1$, $A_2$ and $A_3$, and
- does not touch the vectors $u_i$, which are only used to define its domain.
Moreover, $h_n|_{U_{r,n}}$ is $\GL(V)$-invariant, and thus descends to a regular function $$\label{eqn:superpotential_ncquot}
f_n \colon \NCQuot_{r}^n \to \BA^1.$$
\[prop:SPOT\_A\^3\] There is an identity of closed subschemes $$\Quot_{\BA^3}(\mathscr O^{\oplus r},n) = \crit(f_n) \subset \NCQuot_{r}^n.$$ In particular, $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)$ carries a symmetric perfect obstruction theory.
We use the notation $\crit(f)$ for the zero scheme $\set{\dd f = 0}$, for $f$ a function on a smooth scheme. The embedding of the Quot scheme inside a non-commutative quiver model had appeared (conjecturally, and in a slightly different language) in the physics literature [@CSS].
Every critical locus $\crit(f)$ has a canonical symmetric obstruction theory, determined by the Hessian complex attached to the function $f$. It will be referred to as the *critical obstruction theory*. In the case of $\mathrm{Q} = \Quot_{\BA^3}(\mathscr O^{\oplus r},n)$, this symmetric obstruction theory is the morphism $$\label{symmetric_POT_quot}
\begin{tikzpicture}[baseline=(current bounding box.center)]
\node (N1) at (-1.98,0.95) {$\mathbb E_{\crit}$};
\node (N2) at (-1.37,0.94) {$=$};
\node (N3) at (-1.97,-0.88) {$\BL_{\mathrm{Q}}$};
\node (N4) at (-1.38,-0.88) {$=$};
\node (O1) at (-0.1,0.93) {$\big[T_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}$};
\node (O2) at (2.99,0.93) {$\Omega_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}\big]$};
\node (O3) at (-0.1,-0.88) {$\big[\mathscr I/\mathscr I^2$};
\node (O4) at (2.99,-0.88) {$\Omega_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}\big]$};
\path[commutative diagrams/.cd, every arrow, every label]
(N1) edge node[swap] {$\phi$} (N3)
(O1) edge node {$\mathsf{Hess}(f_n)$} (O2)
(O1) edge node[swap] {$(\dd f_n)^\vee|_{\mathrm{Q}}$} (O3)
(O3) edge node {$\dd$} (O4)
(O2) edge node {$\mathrm{id}$} (O4);
\end{tikzpicture}$$ in $\derived^{[-1,0]}(\mathrm{Q})$, where we represented the truncated cotangent complex by means of the exterior derivative $\dd$ constructed out of the ideal sheaf $\mathscr I \subset \OO_{\NCQuot^n_r}$ of the inclusion $\mathrm{Q} \into \NCQuot^n_r$.
Torus actions on the local Quot scheme {#sec:torus_actions}
--------------------------------------
In this section we define a torus action on the Quot scheme. Set $$\label{tori_T1_and_T2}
\BT_1 = (\BC^\ast)^3, \quad \BT_2 = (\BC^\ast)^r,\quad \TT = \BT_1 \times \BT_2.$$ The torus $\BT_1$ acts on $\BA^3$ by the standard action $$\label{standard_action}
t\cdot x_i = t_ix_i,$$ and this action lifts to an action on $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)$. At the same time, the torus $\BT_2 = (\BC^\ast)^r$ acts on the Quot scheme by scaling the fibres of $\OO^{\oplus r}$. Thus we obtain a $\TT$-action on $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)$.
\[lemma:compact\_fixed\_locus\_Quot\] The fixed locus $\Quot_{\BA^3}(\OO^{\oplus r},n)^{\BT_1}$ is proper. Indeed, a $\BT_1$-invariant surjection $\OO^{\oplus r}\onto T$ necessarily has the quotient $T$ entirely supported at the origin $0\in \BA^3$. Hence $$\Quot_{\BA^3}(\OO^{\oplus r},n)^{\BT_1} \into \Quot_{\BA^3}(\OO^{\oplus r},n)_0$$ sits inside the *punctual Quot scheme* as a closed subscheme. But $\Quot_{\BA^3}(\OO^{\oplus r},n)_0$ is proper, since it is a fibre of the Quot-to-Chow morphism $\Quot_{\BA^3}(\OO^{\oplus r},n)\to \Sym^n\BA^3$, which is a proper morphism.
We recall, verbatim from [@BR18], the description of the full $\TT$-fixed locus induced by the product action on the local Quot scheme.
\[lemma:T\_fixed\_points\] There is an isomorphism of schemes $$\Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\TT} = \coprod_{n_1+\cdots+n_r=n}\prod_{i=1}^r \Hilb^{n_i}(\BA^3)^{\BT_1}.$$ In particular, the $\TT$-fixed locus is isolated and compact.
The main result proved by Bifet in [@Bifet] (in greater generality) implies that $$\label{Bifet_theorem}
\Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\BT_2} = \coprod_{n_1+\cdots+n_r=n}\prod_{i=1}^r \Hilb^{n_i}(\BA^3).$$ The claimed isomorphism follows by taking $\BT_1$-invariants. Since $\Hilb^{k}(\BA^3)^{\BT_1}$ is isolated (a disjoint union of reduced points, each corresponding to a monomial ideal of colength $k$), the result follows.
The $\TT$-action on $\Quot_{\BA^3}(\OO^{\oplus r},n)$ just described can be seen as the restriction of a $\TT$-action on the larger space $\NCQuot^n_r$. Indeed, a torus element $\mathbf t = (t_1,t_2,t_3,w_1,\ldots,w_r) \in \TT$ acts on $ (A_1,A_2,A_3,u_1,\ldots,u_r) \in \NCQuot^n_r$ by $$\label{T-action_on_NCQUOT}
\mathbf t\cdot P = (t_1A_1,t_2A_2,t_3A_3,w_1u_1,\ldots,w_ru_r).$$ The critical locus $\crit(f_n)$ is $\TT$-invariant, and the induced action is precisely the one we described earlier in this section.
Combinatorial description of the $\TT$-fixed locus {#sec: combinatorial description of fixed locus}
--------------------------------------------------
The $\TT$-fixed locus $\Quot_{\BA^3}(\OO^{\oplus r},n)^\TT$ is described purely in terms of $r$-*colored plane partitions* of size $n$, as we now explain.
Here is the general definition of a partition of arbitrary dimension.
\[def:partitions\_arbitrary\_dim\] Let $d\geq 1$ and $n \geq 0$ be integers. A $(d-1)$-dimensional partition of $n$ is a collection of $n$ points $\CA=\set{\mathbf a_1,\ldots,\mathbf a_n}\subset \BZ_{\geq 0}^{d}$ with the following property: if $\mathbf a_i = (a_{i1},\ldots,a_{id}) \in \CA$, then whenever a point $\mathbf y = (y_1,\ldots,y_d)\in\BZ_{\geq 0}^{d}$ satisfies $0\leq y_j\leq a_{ij}$ for all $j=1,\ldots,d$, one has $\mathbf y \in \CA$. The integer $n = \lvert \mathcal A \rvert$ is called the *size* of the partition.
There is a bijective correspondence between the sets of
- $(d-1)$-dimensional partitions of size $n$,
- $(\BC^\ast)^d$-fixed points of $\Hilb^n(\BA^d)$, and
- monomial ideals $I \subset \BC[x_1,\ldots,x_d]$ of colength $n$.
We will be interested in the case $d=3$, corresponding by definition to *plane partitions*. These can be visualised (cf. Figure \[fig:2D\_partition\]) as configurations of $n$ boxes stacked in the corner of a room (with gravity pointing in the $(-1,-1,-1)$ direction).
If $\mathcal A$ is a $(d-1)$-dimensional partition of size $n$ as in Definition \[def:partitions\_arbitrary\_dim\], the associated monomial ideal is $$I_{\mathcal A} = \Braket{x_1^{i_1}\cdots x_d^{i_d}|(i_1,\ldots,i_d) \in \BZ^d_{\geq 0}\setminus \mathcal A}\,\subset\,\BC[x_1,\ldots,x_d].$$ For instance, if $d=3$, in the case of the plane partition pictured in Figure \[fig:2D\_partition\], the associated monomial ideal of colength $16$ is generated by the monomials shaping the staircase of the partition, and is thus equal to $$\Braket{x_3^4,x_1x_3^2,x_1^2x_3,x_1^5,x_1^2x_2,x_1x_2x_3,x_2x_3^2,x_1x_2^3,x_2^3x_3,x_2^4}\,\subset\, \BC[x_1,x_2,x_3].$$
Here is an alternative definition of plane partitions.
\[def:plane\_partition\] A (finite) *plane partition* is a sequence $\pi=\set{\pi_{ij}|i,j\geq 0} \subset \BZ_{\geq 0}$ such that $\pi_{ij}=0$ for $i,j\gg 0$ and $$\pi_{ij}\geq \pi_{i+1,j}, \quad \pi_{ij}\geq \pi_{i,j+1}\quad \textrm{for all } i,j\geq 0.$$
An *$r$-colored plane partition*[^1] is a tuple $\overline{\pi}=(\pi_1,\ldots,\pi_r)$, where each $\pi_i$ is a plane partition.
Denote by $$|\pi|=\sum_{i,j\geq 0}\pi_{ij}$$ the *size* of a plane partition (i.e. the number $n$ in Definition \[def:partitions\_arbitrary\_dim\]) and by $|\overline{\pi}|=\sum_{i=1}^r|\pi_i|$ the size of an $r$-colored plane partition.
In the light of Definition \[def:plane\_partition\], the monomial ideal associated to a plane partition $\pi$ is $$I_\pi = \Braket{x_1^ix^j_2x^{\pi_{ij}}_3 | i,j\geq 0}\,\subset\, \BC[x_1,x_2,x_3].$$ It is clear that the colength of the ideal $I_{\pi}$ is $|\pi|$.
A general plane partition may have infinite legs, each shaped by (i.e. asymptotic to) a standard ($1$-dimensional) partition, or Young diagram. We are not concerned with infinite plane partitions here, since we only deal with quotients $\OO^{\oplus r} \onto T$ with finite support.
\[prop:fixedlocus\_indexed\_by\_colored\_partitions\] There is a bijection between $\TT$-fixed points $[S]\in\Quot_{\BA^3}(\OO^{\oplus r},n)^\TT$ and $r$-colored plane partitions $\overline{\pi}$ of size $n$.
For $r = 1$ this is well known: as we recalled above, monomial ideals $I\subset \BC[x_1,x_2,x_3]$ are in bijective correspondence with plane partitions. Similarly, to each $r$-colored plane partition $\overline{\pi}=(\pi_1,\ldots,\pi_r)$ there corresponds a subsheaf $S_{\overline{\pi}}=\bigoplus_{i=1}^r I_{\pi_i} \subset \OO^{\oplus r}$. But these are the $\TT$-fixed points of the Quot scheme by Lemma \[lemma:T\_fixed\_points\].
### Computing the trace of a monomial ideal
Recall the map sending a torus representation to its weight space decomposition. Consider the $3$-dimensional torus $\BT_1$ acting on the coordinate ring $R = \BC[x_1,x_2,x_3]$ of $\BA^3$. Then we have $$\tr_{R} = \sum_{\textrm{\ding{114}}\,\,\in\, \BZ_{\geq 0}^3}t^{\textrm{\ding{114}}} =\sum_{(i,j,k) \,\in\, \BZ_{\geq 0}^3}t_1^it_2^jt_3^k = \frac{1}{(1-t_1)(1-t_2)(1-t_3)}.$$ For a cyclic monomial ideal $\mathfrak m_{abc} = x_1^ax_2^bx_3^c\cdot R \subset R$, one has $$\tr_{\mathfrak m_{abc}} = \sum_{i\geq a}\sum_{j\geq b}\sum_{k\geq c} t_1^it_2^jt_3^k = \frac{t_1^at_2^bt_3^c}{(1-t_1)(1-t_2)(1-t_3)}.$$ More generally, for a monomial ideal $I_{\pi}\subset \BC[x_1,x_2,x_3]$, one has $$\label{eqn:trace_ideal}
\tr_{I_\pi} = \sum_{(i,j,k)\,\notin\,\pi}t_1^it_2^jt_3^k.$$ These are the building blocks needed to compute $\tr_S$ for an arbitrary sheaf $S = \bigoplus_{i=1}^r I_{\pi_i}$ corresponding to a $\TT$-fixed point $[S] \in \Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}$.
The -fixed obstruction theory
-----------------------------
Recall from [@GPvirtual Prop. 1] that a torus equivariant obstruction theory on a scheme $Y$ induces a canonical perfect obstruction theory, and hence a virtual fundamental class, on each component of the torus fixed locus. In this subsection we show that the reduced isolated locus $$\Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT} \into \Quot_{\BA^3}(\OO^{\oplus r},n)$$ carries the trivial $\TT$-fixed perfect obstruction theory, so the induced virtual fundamental class agrees with the actual ($0$-dimensional) fundamental class.
We first need to check the equivariance (cf. Definition \[def:equivariant\_pot\]) of the critical obstruction theory $\BE_{\crit}$ obtained in Proposition \[prop:SPOT\_A\^3\].
\[lemma:T\_equivariance\_of\_POT\] The critical obstruction theory on $\mathrm{Q} = \Quot_{\BA^3}(\OO^{\oplus r},n)$ is $\TT$-equivariant.
We start with two observations:
1. The potential $f_n = \Tr A_1[A_2,A_3]$ recalled in is homogeneous (of degree $3$) in the matrix coordinates of the non-commutative Quot scheme.\[obs1\]
2. The potential $f_n$ satisfies the relation $$\label{equivariance_of_f}
f_n(t\cdot P) = t_1t_2t_3\cdot f_n(P)$$ for every $t = (t_1,t_2,t_3)\in \BT_1$ and $P\in \NCQuot^n_r$.\[obs2\]
Fix a point $p \in \mathrm{Q} = \crit (f_n) \subset \NCQuot^n_r$. Then, setting $N = 2n^2+rn = \dim \NCQuot^n_r$, let $x_1,\ldots,x_N$ be local holomorphic coordinates of $\NCQuot^n_r$ around $p$. Let the torus $\TT$ act on these coordinates as prescribed by Equation , i.e. $t_1$ (resp. $t_2$ and $t_3$) rescales each $x_k$ corresponding to the entries of the first (resp. second and third) matrix, and $w_l$ rescales the coordinates of the vector $u_l$, for $l=1,\ldots,r$. Formally, for a matrix coordinate $x_k$, we set $$(t_1,t_2,t_3,w_1,\ldots,w_r) \cdot x_k = t_{\ell(k)} x_k$$ where $\ell(k)\in \set{1,2,3}$ depends on whether $x_k$ comes from an entry of $A_1$, $A_2$ or $A_3$. We also have to prescribe an action on tangent vectors and $1$-forms. For a matrix coordinate $x_k$, we set $$\label{eqn:T-action_forms_vectors}
\begin{split}
(t_1,t_2,t_3,w_1,\ldots,w_r) \cdot \frac{\partial}{\partial x_k} &= \frac{t_1t_2t_3}{t_{\ell(k)}} \frac{\partial}{\partial x_k}\\
(t_1,t_2,t_3,w_1,\ldots,w_r) \cdot \dd x_k &= t_{\ell(k)} \dd x_k.
\end{split}$$ If $x_k$ comes from a vector component of the $l$-th vector, we set $$\label{T2actiondimerda}
\begin{split}
(t_1,t_2,t_3,w_1,\ldots,w_r) \cdot \frac{\partial}{\partial x_k} &= w_l^{-1}\frac{\partial}{\partial x_k}\\
(t_1,t_2,t_3,w_1,\ldots,w_r) \cdot \dd x_k &=w_l \dd x_k.
\end{split}$$ However, the $\BT_2$-action will be invisible in the Hessian since the function $f_n$ does not touch the vectors.
The Hessian can be seen as a section $$\mathsf{Hess}(f_n) \in \Gamma\left(\mathrm{Q},T^\ast_{\NCQuot^n_r}\big|_\mathrm{Q}\otimes T^\ast_{\NCQuot^n_r}\big|_\mathrm{Q}\right).$$ In checking the equivariance relation $$\mathbf t\cdot \mathsf{Hess}(f_n)(\xi) = \mathsf{Hess}(f_n)(\mathbf t\cdot\xi), \quad \mathbf t \in \TT,$$ we may ignore local coordinates $x_k$ corresponding to vector entries, because the Hessian is automatically equivariant in these coordinates (equivariance translates into the identity $0=0$).
So, let us fix an $x_k$ coming from one of the matrices. The $(i,j)$-component of the Hessian applied to $\partial / \partial x_k$ is given by $$\mathsf{Hess}_{ij}(f_n)\left(\frac{\partial}{\partial x_k} \right) =
\frac{\partial^2 f_n}{\partial x_i\partial x_j}(x_1,\ldots,x_N)\,\dd x_j.$$ This will vanish unless $k \in \set{i,j}$. Without loss of generality we may assume $k=i$. In this case we obtain, up to a sign convention, $$\label{equivariance_Hess_LHS}
\mathsf{Hess}_{ij}(f_n)\left(\frac{t_1t_2t_3}{t_{\ell(k)}} \frac{\partial}{\partial x_k} \right) = \frac{t_1t_2t_3}{t_{\ell(k)}}
\frac{\partial^2 f_n}{\partial x_k\partial x_j}(x_1,\ldots,x_N)\,\dd x_j.$$ On the other hand, we have $$\begin{aligned}
\mathbf t\cdot \mathsf{Hess}_{ij}(f_n)\left(\frac{\partial}{\partial x_k} \right)&= \mathbf t\cdot \left(\frac{\partial^2 f_n}{\partial x_k\partial x_j}(x_1,\ldots,x_N)\,\dd x_j\right) \\
&=\frac{\partial^2 f_n}{(\partial t_{\ell(k)}x_k)(\partial t_{\ell(j)}x_j)}(t_{\ell(1)}x_1,\ldots,t_{\ell(N)}x_N)t_{\ell(j)}\dd x_j\\
&=\frac{t_1t_2t_3}{t_{\ell(k)}t_{\ell(j)}}t_{\ell(j)}\frac{\partial^2 f_n}{\partial x_k\partial x_j}(x_1,\ldots,x_N)\,\dd x_j\\
&=\frac{t_1t_2t_3}{t_{\ell(k)}}\frac{\partial^2 f_n}{\partial x_k\partial x_j}(x_1,\ldots,x_N)\,\dd x_j,\end{aligned}$$ where we have combined the observations and with the second identity in to obtain the third equality. The latter expression agrees with the right hand side of Equation . Thus we conclude that the Hessian complex is $\TT$-equivariant, as well as the morphism to the cotangent complex. This finishes the proof.
The property of $f_n$ exhibits the differential $\dd f_n$ as a $\GL_3$-equivariant section $$\dd f_n \otimes \mathfrak{t}^{-1} \colon \OO_{\NCQuot^n_r} \to \Omega_{\NCQuot^n_r}\otimes \mathfrak{t}^{-1},$$ where $\mathfrak{t}^{-1} = (t_1t_2t_3)^{-1}$ is the determinant representation of $\BC^3 = \bigoplus_{1\leq i\leq 3}t_i^{-1}\cdot \BC$. Therefore, explicitly, the morphism in $\derived^{[-1,0]}(\Coh^{\TT}_{\mathrm{Q}})$ lifting the critical obstruction theory is $$\label{equivariant_symmetric_POT_quot}
\begin{tikzpicture}[baseline=(current bounding box.center)]
\node (O1) at (-0.1,0.93) {$ \big[\mathfrak t\otimes T_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}$};
\node (O2) at (2.99,0.93) {$\Omega_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}\big]$};
\node (O3) at (-0.1,-0.88) {$\big[\mathscr I/\mathscr I^2$};
\node (O4) at (2.99,-0.88) {$\Omega_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}\big]$};
\path[commutative diagrams/.cd, every arrow, every label]
(O1) edge node {$\mathsf{Hess}(f_n)$} (O2)
(O1) edge node[swap] {$(\dd f_n)^\vee|_{\mathrm{Q}}$} (O3)
(O3) edge node {$\dd$} (O4)
(O2) edge node {$\mathrm{id}$} (O4);
\end{tikzpicture}$$ so that, in particular, the equivariant K-theory class of the virtual tangent bundle attached to the (equivariant) perfect obstruction theory is $$\label{eqn:virtual_tg_quot}
T_{\mathrm{Q}}^{\vir} = T_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}-\Omega_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}\otimes\mathfrak{t}^{-1} \,\in\,K^0_{\TT}(\mathrm{Q}).$$ This fact will be recalled and used in Propositions \[prop:Tangent\^vir\] and \[restriction of class in K theory\].
Lemma \[lemma:T\_equivariance\_of\_POT\] implies the existence of a “$\TT$-fixed” obstruction theory $$\label{T-fixed_pot}
\BE_{\crit}\big|_{\Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}}^{\fix} \to \BL_{\Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}}$$ on the fixed locus of the Quot scheme. We proved in Lemma \[lemma:T\_fixed\_points\] that this fixed locus is $0$-dimensional, isolated and reduced. The next result shows that the virtual fundamental class induced by on the fixed locus agrees with the actual fundamental class.
\[trivial pot on fixed locus\] There is an identity $$\left[\Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}\right]^{\vir} = \left[\Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}\right]$$ in the Chow group $A_0(\Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT})$.
The perfect obstruction theory $\BE_{\crit}$ on $\mathrm{Q}=\Quot_{\BA^3}(\OO^{\oplus r},n)$ made explicit in is symmetric, induced by the critical structure. Its K-theory class is therefore (cf. Remark \[rmk:K-class\_of\_symmetric\_pot\]) $$\BE_{\crit} = \Omega_{\mathrm{Q}}-T_{\mathrm{Q}} \in K^0(\mathrm{Q}).$$ We know by Lemma \[lemma:T\_fixed\_points\] that the fixed locus is reduced and isolated, which means that $$\label{no_fixed_tangents}
T_{\mathrm{Q}}\big|_p^{\fix} = 0$$ for all $p\in \mathrm{Q}^{\TT}$. Every such fixed point inherits a perfect obstruction theory whose associated virtual fundamental class is of dimension at most $\dim \set{p}=0$. Explicitly, the obstruction theory is given by the complex $$\BE_{\crit}\big|_{p}^{\fix},$$ whose K-theory class is $$\Omega_{\mathrm{Q}}\big|_p^{\fix} - T_{\mathrm{Q}}\big|_p^{\fix} = \Omega_{\mathrm{Q}}\big|_p^{\fix}$$ thanks to the vanishing . For this to have rank at most $0$, we must have $$\Omega_{\mathrm{Q}}\big|_p^{\fix} = 0.$$ So there are neither fixed tangents nor fixed obstructions at $p$, which implies the obstruction theory is trivial. The claim follows.
\[cor:tangents\_are\_T\_movable\] Let $[S] \in \Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}$ be a torus fixed point. The tangent and cotangent spaces of $\Quot_{\BA^3}(\OO^{\oplus r},n)$ at $[S]$ are entirely $\TT$-movable. In particular, the virtual tangent space at $[S]$ can be written as $$\label{tg_vir}
T^{\vir}_S = \BE_{\crit}^\vee\big|^{\mov}_{[S]} = T_{\mathrm{Q}}\big|_{[S]} - \Omega_{\mathrm{Q}}\big|_{[S]} \otimes\mathfrak{t}^{-1} \in K^0_{\TT}(\pt).$$
Invariants attached to the local Quot scheme {#sec:invariants}
============================================
In this section we introduce cohomological and K-theoretic DT invariants of $\BA^3$, the main object of study of this paper, starting from the Quot scheme $\Quot_{\BA^3}(\OO^{\oplus r},n)$ studied in the previous section. We first need to introduce some notation and terminology.
Some notation
-------------
Recall the tori $\BT_1 = (\BC^\ast)^3$ and $\BT_2 = (\BC^\ast)^r$ from . We let $t_1,t_2,t_3$ and $w_1,\ldots,w_r$ be the generators of the representation rings $K^0_{\BT_1}(\pt)$ and $K^0_{\BT_2}(\pt)$, respectively. Then one can write the equivariant cohomology rings of $\BT_1$ and $\BT_2$ as $$H^\ast_{\BT_1} = \BQ[s_1,s_2,s_3],\quad H^\ast_{\BT_2} = \BQ[v_1,\ldots,v_r],$$ where $s_i = c_1^{\BT_1}(t_i)$ and $v_j=c_1^{\BT_2}(w_j)$. For a virtual $\TT$-module $V \in K^0_{\TT}(\pt)$, we let $$\tr_{V} \in \BZ(\!(t_1,t_2,t_3,w_1,\ldots,w_r)\!)$$ denote its character, i.e. its decomposition into weight spaces. We denote by $\overline{(\,\cdot\,)}$ the involution of $\BZ(\!(t_1,t_2,t_3,w_1,\ldots,w_r)\!)$ defined by $$\overline P(t_1,t_2,t_3,w_1,\ldots,w_r) = P(t_1^{-1},t_2^{-1},t_3^{-1},w_1^{-1},\ldots,w_r^{-1}).$$
### Twisted virtual structure sheaf {#sec: twisted structure sheaf}
For any scheme $X$ endowed with a perfect obstruction theory $\BE \to \BL_X$, define as in [@Fantechi_Gottsche Def. 3.12], the *virtual canonical bundle* $$\mathcal{K}_{X,\vir}=\det \BE = \det(T_X^{\vir})^\vee.$$ This is just $\det E^0 \otimes (\det E^{-1})^\vee$ if $\BE = E^0 - E^{-1} \in K^0(X)$. We will simply write $\mathcal{K}_{\vir}$ when $X$ is clear from the context.
\[prop:there\_is\_a\_square\_root\] Let $A$ be a smooth variety equipped with a regular function $f\colon A\to \BA^1$, and let $X= \crit(f) \subset A$ be the critical locus of $f$, with its critical (symmetric) perfect obstruction theory $\BE_{\crit} \to \BL_{X}$. Then $\mathcal{K}_{X,\vir} \in \Pic(X)$ admits a square root, i.e. there exists a line bundle $$\mathcal{K}_{X,\vir}^{\frac{1}{2}}\in \Pic(X)$$ whose second tensor power equals $\det \BE_{\crit}$.
The K-theory class of the critical perfect obstruction theory is $$\BE_{\crit} = \Omega_A\big|_X - T_A\big|_{X},$$ and by definition one has $$\mathcal{K}_{X,\vir}=\frac{\det \Omega_A|_X}{\det {T_A}|_{X}}=\frac{\det \Omega_A|_X}{(\det \Omega_A|_X)^{-1}}=K_A\big|_{X}\otimes K_A\big|_{X}.\qedhere$$
Let $X$ be a scheme endowed with a perfect obstruction theory, and let $\OO_X^{\vir} \in K_0(X)$ be the induced virtual structure sheaf. Assume the virtual canonical bundle admits a square root. Following [@NO16], we define the *twisted* (or *symmetrised*) *virtual structure sheaf* as $$\widehat{\OO}_X^{\vir}=\OO_X^{\vir}\otimes \mathcal{K}_{X,\vir}^{\frac{1}{2}}.$$ In case $X$ carries a torus action, we will see in Remark \[rmk: weight of twisted sheaf\] that $\widehat{\OO}_X^{\vir}$ acquires a canonical weight.
Classical enumerative invariants
--------------------------------
The naive (topological) Euler characteristic of the Quot scheme $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)$ is computed via the Gholampour–Kool formula [@Gholampour2017 Prop. 2.3] $$\sum_{n\geq 0}e(\Quot_{\BA^3}(\mathscr O^{\oplus r},n))q^n = \mathsf M(q)^r,$$ where $\mathsf M(q) = \prod_{m\geq 1}(1-q^m)^{-m}$ is the MacMahon function, the generating function for the number of plane partitions of non-negative integers. On the other hand, the Behrend weighted Euler characteristic of the Quot scheme can be computed via the formula $$\label{eqn:chi_vir_quot_affinespace}
\sum_{n\geq 0}e_{\vir}(\Quot_{\BA^3}(\mathscr O^{\oplus r},n))q^n = \mathsf M((-1)^rq)^r,$$ proved in [@BR18 Cor. 2.8]. For a complex scheme $Z$ of finite type over $\BC$, we have set $e_{\vir}(Z) = e(Z,\nu_Z)$, where $\nu_Z$ is Behrend’s constructible function [@Beh]. See [@BR18 Thm. A] for a proof of the analogue of for an arbitrary pair $(Y,F)$ consisting of a smooth $3$-fold $Y$ along with a locally free sheaf $F$ on it. It was shown by Toda [@Toda2] that, on a projective Calabi–Yau $3$-fold $Y$, the wall-crossing factor in the higher rank DT/PT correspondence is precisely $\mathsf M((-1)^rq)^{re(Y)}$. The relationship between Toda’s wall-crossing formula [@Toda2] and the Gholampour–Kool’s formula for Euler characteristics of Quot schemes on $3$-folds [@Gholampour2017] was clarified in [@BR18] via a Hall algebra argument.
Virtual invariants of the Quot scheme {#subsec:virtual_invariants_quot}
-------------------------------------
The scheme $\Quot_{\BA^3}(\mathscr O^{\oplus r},n)$ is not proper, but carries a torus action with proper fixed locus. Thus we may define virtual invariants via equivariant residues, by setting $$\int_{[\Quot_{\BA^3}(\mathscr O^{\oplus r},n)]^{\vir}}1\defeq \sum_{[S]\in \Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\TT}}\frac{1}{e^{\TT}(T_S^{\vir})}\in \mathbb{Q}(\!(s,v)\!),$$ where $s=(s_1,s_2,s_3) $ and $v=(v_1,\ldots,v_r)$ are the equivariant parameters of the torus $\TT$ and $T_S^{\vir}$ is the virtual tangent space . The sum runs over all $\TT$-fixed points $[S]$, which are isolated, reduced and with the trivial perfect obstruction theory induced from the critical obstruction theory on the Quot scheme (cf. Proposition \[trivial pot on fixed locus\]). We refer to these invariants as *cohomological rank $r$ DT invariants*, as they take value in (an extension of) the fraction field $\mathbb{Q}(s,v)$ of the $\TT$-equivariant cohomology ring $H^\ast_{\TT}$. We will study their generating function $$\label{eqn:cohomological_DT_series}
\DT_r^{\coh}(\BA^3,q,s,v)=\sum_{n\geq 0} q^n\int_{[\Quot_{\BA^3}(\mathscr O^{\oplus r},n)]^{\vir}}1 \in \mathbb{Q}(\!(s,v)\!)\llbracket q \rrbracket$$ in §\[sec:cohomological\_invariants\]. On the other hand, K-theoretic invariants arise as natural refinements of their cohomological counterpart. Naively, one would like to study the virtual holomorphic Euler characteristic $$\begin{aligned}
\chi(\Quot_{\mathbb{A}^3}(\OO^{\oplus r},n), \OO^{\vir})&=\sum_{[S]\in \Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\TT}}\tr\left(\frac{1}{\Lambda^\bullet T_S^{\vir,\vee}} \right) \in \BZ(\!(t,w)\!),\end{aligned}$$ where $t = (t_1,t_2,t_3)$, $w=(w_1,\ldots, w_r)$, and via the trace map $\tr$ we identify a (possibly infinite-dimensional) virtual $\TT$-module with its decomposition into weight spaces. It turns out that guessing a closed formula for these invariants is incredibly difficult and, after all, not what one should look at. Instead, Nekrasov–Okounkov [@NO16] teach us that we should focus our attention on $$\label{eqn: K theoretic invariants}
\chi(\Quot_{\mathbb{A}^3}(\OO^{\oplus r},n), \widehat{\OO}^{\vir})
= \sum_{[S]\in \Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\TT}}\tr\left(\frac{\mathcal{K}_{\vir}^{\frac{1}{2}}}{\Lambda^\bullet T_S^{\vir,\vee}}\right) \in \BZ(\!(t, (t_1t_2t_3)^{\frac{1}{2}},w)\!),\footnote{In theory all equivariant weights could appear with half powers. However, by Remark \ref{rmk: weight of twisted sheaf}, $(t_1t_2t_3)^{\frac{1}{2}}$ is the only one that truly appears.}$$ where the *twisted* *virtual structure sheaf* $\widehat{\OO}^{\vir}$ is defined in §\[sec: twisted structure sheaf\] — a square root of the virtual canonical bundle exists by Lemma \[prop:there\_is\_a\_square\_root\] and Proposition \[prop:SPOT\_A\^3\]. The generating function of rank $r$ K-theoretic DT invariants $$\label{eqn:K_partition_function}
\DT^{\KK}_r(\BA^3,q,t,w)=\sum_{n\geq 0} q^n\chi(\Quot_{\mathbb{A}^3}(\OO^{\oplus r},n), \widehat{\OO}^{\vir})\in \BZ(\!(t, (t_1t_2t_3)^{\frac{1}{2}},w)\!)\llbracket q \rrbracket$$ will be studied in §\[sec:K-theory\].
\[rmk: weight of twisted sheaf\] To be precise, we should replace the torus $\TT$ with its double cover $\TT_{\mathfrak t}$, the minimal cover of $\TT$ where the character $\mathfrak{t}^{-1/2}$ is defined, as in [@NO16 §7.1.2]. Then $\mathcal K_{\vir}^{1/2}$ is a $\TT_{\mathfrak t}$-equivariant sheaf with character $\mathfrak{t}^{-(\dim \NCQuot_{r}^n)/2}$. To ease the notation, we keep denoting the torus acting as $\TT$.
\[rmk:independence\_of\_orientation\] As remarked in [@NO16; @arbesfeld2019ktheoretic], choices of square roots of $\mathcal{K}_{\vir}$ differ by a 2-torsion element in the Picard group, which implies that $\chi(\Quot_{\mathbb{A}^3}(\OO^{\oplus r},n), \widehat{\OO}^{\vir})$ does not depend on such choices of square roots. Thus there is no ambiguity in .
Higher rank vertex on the local Quot scheme {#sec:higher_rank_vertex}
===========================================
The virtual tangent space of the local Quot scheme {#sec: virtual tangent space}
--------------------------------------------------
By Lemma \[lemma:T\_fixed\_points\], we can represent the sheaf corresponding to a $\TT$-fixed point $$[S] \in \Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}$$ as a direct sum of ideal sheaves $$S = \bigoplus_{\alpha=1}^r\mathscr I_{Z_\alpha} \,\subset\,\OO^{\oplus r},$$ with $Z_\alpha \subset \BA^3$ a finite subscheme of length $n_\alpha$ supported at the origin, and satisfying $n = \sum_{1\leq \alpha\leq r} n_\alpha$. In this subsection we derive a formula for the character of $$\chi(\OO^{\oplus r},\OO^{\oplus r})-\chi(S,S) \in K^0_{\TT}(\pt),$$ where for $F$ and $G$ in $K^0_{\TT}(\pt)$, we set $$\chi(F,G) =\mathbf{R}\Hom(F,G)=\sum_{i\geq 0}(-1)^{i}\Ext^{i}(F,G).$$ Our method follows the approach of [@MNOP1 § 4.7]. Moreover, we show in Proposition \[prop:Tangent\^vir\] that such character agrees with the virtual tangent space $T_S^{\vir}$ induced by the critical obstruction theory.
Let $\mathsf Q_\alpha$ be the $\BT_1$-character of the $\alpha$-summand of $\OO^{\oplus r}/S$, i.e. (cf. Equation ) $$\mathsf Q_\alpha = \tr_{\OO_{Z_\alpha}}= \sum_{(i,j,k) \,\in\, \pi_\alpha}t_1^{i}t_2^{j}t_3^{k},$$ where $\pi_\alpha \subset \BZ_{\geq 0}^3$ is the plane partition corresponding to the monomial ideal $\mathscr I_{Z_\alpha} \subset R = \BC[x,y,z]$. Let $\mathsf P_\alpha(t_1,t_2,t_3)$ be the Poincaré polynomial of $\mathscr I_{Z_\alpha}$. This can be computed via a $\BT_1$-equivariant free resolution $$0 \to E_{\alpha,s} \to \cdots \to E_{\alpha,1} \to E_{\alpha,0} \to \mathscr I_{Z_\alpha} \to 0.$$ Writing $$E_{\alpha,i} = \bigoplus_{j}R(d_{\alpha,ij}),\quad d_{\alpha,ij} \in \BZ^3,$$ one has, independently of the chosen resolution, the formula $$\mathsf P_\alpha(t_1,t_2,t_3) = \sum_{i,j}(-1)^it^{d_{\alpha,ij}}.$$ By [@MNOP1 §4.7] we know that there is an identity $$\label{eqn_Q_and_P}
\mathsf Q_\alpha = \frac{1+\mathsf P_\alpha}{(1-t_1)(1-t_2)(1-t_3)}.$$ For each $1\leq \alpha,\beta\leq r$, we can compute $$\begin{aligned}
\chi(\mathscr I_{Z_\alpha},\mathscr I_{Z_\beta}) &= \sum_{i,j,k,l}(-1)^{i+k} \Hom_R(R(d_{\alpha,ij}),R(d_{\beta,kl})) \\
&=\sum_{i,j,k,l}(-1)^{i+k} R(d_{\beta,kl}-d_{\alpha,ij}),\end{aligned}$$ which immediately yields the identity $$\tr_{\chi(\mathscr I_{Z_\alpha},\mathscr I_{Z_\beta})} = \frac{\mathsf P_\beta \overline{\mathsf P}_\alpha}{(1-t_1)(1-t_2)(1-t_3)} \in \BZ(\!(t_1,t_2,t_3)\!).$$ It follows that, as a $\TT$-representation, one has $$\begin{aligned}
\chi(S,S)
&= \chi\left(\sum_\alpha w_\alpha\otimes \mathscr I_{Z_\alpha},\sum_\beta w_\beta\otimes \mathscr I_{Z_\beta}\right) \\
&= \sum_{1\leq \alpha,\beta\leq r}\chi(w_\alpha\otimes \mathscr I_{Z_\alpha},w_\beta\otimes \mathscr I_{Z_\beta}),\end{aligned}$$ which yields $$\label{eqn:trace_chi_FF}
\tr_{\chi(S,S)} = \sum_{1\leq \alpha,\beta\leq r} \frac{w_\alpha^{-1}w_\beta\cdot \mathsf P_\beta \overline{\mathsf P}_\alpha}{(1-t_1)(1-t_2)(1-t_3)}.$$ On the other hand, $$\label{eqn:trace_chi_OO}
\tr_{\chi(\OO^{\oplus r},\OO^{\oplus r})} = \sum_{1\leq \alpha,\beta\leq r} \frac{w_\alpha^{-1}w_\beta}{(1-t_1)(1-t_2)(1-t_3)}.$$
\[prop:character\_virtual\_tangent\] Let $[S] \in \Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}$ be a torus fixed point. There is an identity $$\label{eqn:trace_T_vir}
\tr_{\chi(\OO^{\oplus r},\OO^{\oplus r})-\chi(S,S)} = \sum_{1\leq \alpha,\beta\leq r}w_\alpha^{-1}w_\beta\left(\mathsf Q_\beta-\frac{\overline{\mathsf Q}_\alpha}{t_1t_2t_3}+\frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3}\mathsf Q_\beta \overline{\mathsf Q}_\alpha \right)$$ in $\BZ(\!(t_1,t_2,t_3,w_1,\ldots,w_r)\!)$.
We compute $\tr_{\chi(\OO^{\oplus r},\OO^{\oplus r})}-\tr_{\chi(S,S)}$ via Formulas and . It is enough to observe that $$\begin{aligned}
\frac{1-\mathsf P_\beta \overline{\mathsf P}_\alpha}{(1-t_1)(1-t_2)(1-t_3)}
&= \frac{1-\mathsf P_\beta \overline{\mathsf P}_\alpha+\mathsf P_\beta-\mathsf P_\beta}{(1-t_1)(1-t_2)(1-t_3)} \\
&= \mathsf Q_\beta + \mathsf P_\beta\frac{(-1-\overline{\mathsf P}_\alpha)}{(1-t_1)(1-t_2)(1-t_3)} \\
&= \mathsf Q_\beta + \mathsf P_\beta\frac{\overline{\mathsf Q}_\alpha}{t_1t_2t_3} \\
&= \mathsf Q_\beta + \left(\mathsf Q_\beta (1-t_1)(1-t_2)(1-t_3)-1\right)\frac{\overline{\mathsf Q}_\alpha}{t_1t_2t_3}.\end{aligned}$$ We have used Formula in the second and fourth equality.
For every $\TT$-fixed point $[S]$ we define associated *vertex* terms $$\label{eqn:vertex_terms}
\mathsf V_{ij}=w_i^{-1}w_j\left(\mathsf Q_j-\frac{\overline{\mathsf Q}_i}{t_1t_2t_3}+\frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3}\mathsf Q_j \overline{\mathsf Q}_i \right)$$ for every $i,j=1,\ldots, r$. It is immediate to see that for $r=1$ (forcing $i=j$) we recover the vertex formalism developed in [@MNOP1].
Proposition \[prop:character\_virtual\_tangent\] can then be restated as $$\tr_{\chi(\OO^{\oplus r},\OO^{\oplus r})}-\tr_{\chi(S,S)} = \sum_{1\leq i,j\leq r} \mathsf V_{ij}.$$ We now relate this to the virtual tangent space (cf. §\[subsec:Nvir\]) $T_{S}^{\vir}$ of a point $[S] \in \Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}$.
\[prop:Tangent\^vir\] Let $[S] \in \Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}$ be a $\TT$-fixed point. Let $T^{\vir}_S = \BE_{\crit}^\vee\big|_{[S]}$ be the virtual tangent space induced by the $\TT$-equivariant critical obstruction theory. Then there are identities $$\begin{aligned}
T^{\vir}_S &= \chi(\OO^{\oplus r},\OO^{\oplus r}) - \chi(S,S)=\sum_{1\leq i,j\leq r} \mathsf V_{ij}\in K_0^{\TT}(\pt).\end{aligned}$$
Let $\NCQuot_{r}^n$ be the non-commutative Quot scheme defined in . The superpotential $f_n\colon \NCQuot_{r}^n \to \BA^1$ defined in is equivariant with respect to the character $(t_1,t_2,t_3)\mapsto t_1t_2t_3$, so it gives rise to a $\GL_3$-equivariant section $$\label{GL3_equiv_section}
\OO_{\NCQuot_{r}^n} \xrightarrow{\dd f_n \otimes \mathfrak{t}^{-1}} \Omega_{\NCQuot_{r}^n}\otimes \mathfrak{t}^{-1}$$ where, starting from the representation $\BC^3 = \bigoplus_{1\leq i\leq 3}t_i^{-1}\cdot \BC \in K^0_{\BT_1}(\pt)$, we have set $$\mathfrak{t}^{-1} = \det \BC^3 = (t_1t_2t_3)^{-1}.$$ Here, and throughout this proof, we are identifying a representation with its own character via the isomorphism . The zero locus of the section is our Quot scheme $$\mathrm Q=\Quot_{\BA^3}(\OO^{\oplus r},n),$$ endowed with the $\TT$-action described in §\[sec:torus\_actions\]. According to Equation , the virtual tangent space computed with respect to the critical $\TT$-equivariant obstruction theory on $\mathrm Q$ is $$\label{tg_vir_definition}
T^{\vir}_S = \left(T_{\NCQuot_{r}^n} - \Omega_{\NCQuot_{r}^n}\otimes \mathfrak{t}^{-1}\right)\big|_{[S]}.$$ The tangent space to the smooth scheme $\NCQuot_{r}^n$ can be written, around a point $S \into \OO^{\oplus r} \onto V$, as $$\label{eqn:tangents_NCQUOT}
T_{\NCQuot_{r}^n}\big|_{[S]} = (\BC^3-1)\otimes (\overline{V}\otimes V) + \bigoplus_{\alpha=1}^r \Hom(w_\alpha \BC,V),$$ where $$\bigoplus_{\alpha=1}^r \Hom(w_\alpha \BC,V) = \bigoplus_{\alpha=1}^r w_\alpha^{-1}\otimes V$$ represents the $r$ framings on the $3$-loop quiver. Let $V$ be written as a direct sum of structure sheaves $$V = \bigoplus_{\alpha=1}^r \OO_{Z_\alpha},$$ where the $\alpha$-summand has $\TT$-character $w_\alpha \mathsf Q_\alpha$. Substituting $$\begin{aligned}
\BC^3 -1 &= t_1^{-1}+t_2^{-1}+t_3^{-1} - 1 = \frac{t_1t_2+t_1t_3+t_2t_3-t_1t_2t_3}{t_1t_2t_3}\\
V &= \sum_{\alpha=1}^r w_\alpha \mathsf Q_\alpha \end{aligned}$$ into Formula yields $$T_{\NCQuot_{r}^n}\big|_{[S]} = \frac{t_1t_2+t_1t_3+t_2t_3-t_1t_2t_3}{t_1t_2t_3} \sum_{1\leq \alpha,\beta\leq r} w_\alpha^{-1}w_\beta \overline{\mathsf Q}_\alpha \mathsf Q_\beta+\sum_{1\leq \alpha,\beta\leq r}w_\alpha^{-1}w_\beta \mathsf Q_\beta,$$ and hence $$\left(\Omega_{\NCQuot_{r}^n}\otimes \mathfrak{t}^{-1}\right) \big|_{[S]} =
\frac{t_1+t_2+t_3-1}{t_1t_2t_3} \sum_{1\leq \alpha,\beta\leq r} w_\alpha^{-1}w_\beta \overline{\mathsf Q}_\alpha \mathsf Q_\beta + \sum_{1\leq \alpha,\beta\leq r}w_\alpha^{-1}w_\beta \frac{\overline{\mathsf Q}_\alpha}{t_1t_2t_3},$$ which by Formula yields $$T^{\vir}_S = \sum_{1\leq \alpha,\beta\leq r} w_\alpha^{-1}w_\beta\left(\mathsf Q_\beta - \frac{\overline{\mathsf Q}_\alpha}{t_1t_2t_3} + \frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3}\overline{\mathsf Q}_\alpha \mathsf Q_\beta \right).$$ The right hand side is shown to be equal to $\chi(\OO^{\oplus r},\OO^{\oplus r}) - \chi(S,S)$ in Proposition \[prop:character\_virtual\_tangent\].
A small variation of the vertex formalism {#sec: variation vertex}
-----------------------------------------
All locally free sheaves on $\mathbb{A}^3$ are trivial, but this is not true equivariantly. For example, we have $K_{\mathbb{A}^3}=\OO_{\mathbb{A}^3}\otimes t_1t_2t_3\in K^0_{\BT_1}({\mathbb{A}^3})$, even though the canonical bundle is trivial. Consider $$\label{eqn:sheaf_with_equiv_weights}
F=\bigoplus_{i=1}^r \OO_{\mathbb{A}^3}\otimes \lambda_i\in K^0_{\BT_1}({\mathbb{A}^3})$$ where $\lambda = (\lambda_i)_i$ are weights of the $\BT_1$-action, i.e. monomials in the representation ring of $\BT_1$. Let $[S] \in \Quot_{\BA^3}(F,n)^{\TT}$ be a $\TT$-fixed point. It decomposes as $S = \bigoplus_{i=1}^r\mathscr I_{Z_i}\otimes \lambda_i\in K_{\BT_1}^0 (\BA^3 )$, where the weights $\lambda_i$ are naturally inherited from $F$. This generalises the material in §\[sec: virtual tangent space\], which can be recovered by setting all weights $\lambda_i$ to be trivial. Just as in Proposition \[prop:Tangent\^vir\], we can compute $$T^{\vir}_{S,\lambda}=\chi(F,F)-\chi(S,S)\in K_{\TT}^0 (\pt).$$ We find $$\begin{aligned}
T^{\vir}_{S,\lambda}&=\chi\left(\bigoplus_{i=1}^r \OO_{\mathbb{A}^3}\otimes \lambda_i w_i,\bigoplus_{j=1}^r \OO_{\mathbb{A}^3}\otimes \lambda_j w_j \right)-\chi\left(\bigoplus_{j=1}^r \mathscr I_{Z_i}\otimes \lambda_i w_i, \bigoplus_{i=1}^r\mathscr I_{Z_j}\otimes \lambda_j w_j \right).
\end{aligned}$$
Therefore we derive the same expression for the vertex formalism as before, just substituting $w_i$ with $\lambda_i w_i$. This variation will be crucial in §\[section on compact invariants\] where we indentify the Quot scheme of the local model with the restriction of $\Quot_X(F,n)$ to an open toric chart, with $X$ a toric projective 3-fold and $F$ a $\BT_1$-equivariant exceptional locally free sheaf.
Define, for $\lambda = (\lambda_1,\ldots,\lambda_r)$ as above and $F$ as in , the equivariant integral $$\label{eqn:lambda_invariants}
\int_{[\Quot_{\BA^3}( F,n)]^{\vir}}1\defeq \sum_{[S]\in \Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\TT}}\frac{1}{e^{\TT}(T_{S,\lambda}^{\vir})}\in \mathbb{Q}(\!(s,v)\!),$$ and let $$\label{eqn:cohomological_DT_series_lambda}
\DT_r^{\coh}(\BA^3,q,s,v)_\lambda = \sum_{n\geq 0}q^n \int_{[\Quot_{\BA^3}( F,n)]^{\vir}}1$$ be the generating function of the invariants . We shall see (cf. Corollary \[cor:independence\_on\_lambda\_w\]) that this expression does not depend on the equivariant weights $\lambda_i$.
The higher rank K-theoretic DT partition function {#sec:K-theory}
=================================================
Symmetrised exterior algebras and brackets {#sec: preliminaries K theory}
------------------------------------------
We recall some constructions is equivariant K-theory which will be used to prove Theorem \[mainthm:K-theoretic\]. For a recent and more complete reference, the reader may consult [@Okounkov_Lectures § 2].
Let $\TT$ be a torus, $V=\sum_\mu t^\mu$ a $\TT$-module. Assume that $\det(V)$ is a square in $K_{\TT}^0(\pt)$. Define the *symmetrised exterior algebra* of $V$ as $$\widehat{\Lambda}^\bullet V \defeq \frac{\Lambda^\bullet V}{\det(V)^{\frac{1}{2}}}.$$ It satisfies the relation $$\widehat{\Lambda}^\bullet V^{\vee}=(-1)^{\rk V}\widehat{\Lambda}^\bullet V.$$ Define the operator $[\,\cdot\,]$ by $$[t^{\mu}]=t^{\frac{\mu}{2}}-t^{-\frac{\mu}{2}}.$$ One can compute $$\tr( \widehat{\Lambda}^\bullet V^\vee)=\prod_{\mu}\frac{1-t^{-\mu}}{t^{-\frac{\mu}{2}}}=\prod_{\mu}\,(t^{\frac{\mu}{2}}-t^{-\frac{\mu}{2}})=\prod_{\mu}\,[t^{\mu}],$$ Extending $ \widehat{\Lambda}^\bullet$ and $[\,\cdot\,]$ by linearity to the whole of $K^0_{\TT}(\pt)$, for any virtual $\TT$-representation $V=\sum_{\mu}t^{\mu} -\sum_{\nu}t^{\nu}$ we find $$\tr( \widehat{\Lambda}^\bullet V^\vee)=\frac{\prod_{ \mu}[t^{\mu}]}{\prod_{ \nu}[t^{\nu}]}.$$
Proof of Theorem \[mainthm:K-theoretic\] {#sec:proof_of_K-theoretic_thm}
----------------------------------------
By the description of the $\TT$-fixed locus $\Quot_{\mathbb{A}^3}(\OO^{\oplus r},n)^{\TT}$ given in §\[sec: combinatorial description of fixed locus\], every colored plane partition $ \overline{\pi}=(\pi_1,\dots, \pi_r)$ corresponds to a unique $\TT$-fixed point $S = \bigoplus_{i=1}^r\mathscr I_{Z_i}$, for which we defined in Equation the vertex terms $\mathsf V_{ij}$ by $$\mathsf V_{ij}=w_i^{-1}w_j\left(\mathsf Q_j-\frac{\overline{\mathsf Q}_i}{t_1t_2t_3}+\frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3}\mathsf Q_j \overline{\mathsf Q}_i \right)$$ with notation as in §\[sec: virtual tangent space\]. The generating function of higher rank cohomological DT invariants can be rewritten in a purely combinatorial fashion as $$\begin{aligned}
\DT_r^{\coh}(\BA^3, q,s,v)=\sum_{\overline{\pi}} q^{\lvert \overline \pi\rvert} \prod_{i,j=1}^r e^{\TT}(-\mathsf V_{ij}).\end{aligned}$$ Similarly, the generating function of the K-theoretic invariants can be rewritten as $$\begin{aligned}
\DT_r^{\KK}(\BA^3, q,t,w)=\sum_{\overline{\pi}} q^{\lvert \overline \pi\rvert} \prod_{i,j=1}^r [-\mathsf V_{ij}].\end{aligned}$$
A closed formula for $\DT_1^{\KK}(\BA^3, q,t,w)$ was conjectured in [@NEKRASOV2005261] and has recently been proven by Okounkov.
\[thm: Okounkov rank 1\] The generating series of rank $1$ K-theoretic Donaldson–Thomas invariants of $\BA^3$ is given by $$\begin{aligned}
\DT_1^{\KK}(\BA^3,-q,t,w)=\Exp\left(\mathsf F_1( q,t_1,t_2,t_3)\right)\end{aligned}$$ where, setting $\mathfrak{t}=t_1t_2t_3$, one defines $$\mathsf F_1( q,t_1,t_2,t_3)= \frac{1}{[\mathfrak{t}^{\frac{1}{2}}q ][\mathfrak{t}^{\frac{1}{2}} q^{-1}]}\frac{[t_1t_2][t_1t_3][t_2t_3]}{[t_1][t_2][t_3]}.$$
It is clear from the expression of the vertex in rank 1 that there is no dependence on $w_1$. As pointed out to us by N. Arbesfeld, this can in fact be seen as a shadow of the fact that $\BT_2 = \BC^\ast$ acts trivially on $\NCQuot^n_1$ and on $\dd f_n$.
We devote the rest of this section is to prove a generalisation of Theorem \[thm: Okounkov rank 1\] to higher rank.
\[thm: K theoretic of points higher rank\] The generating series of rank $r$ K-theoretic Donaldson–Thomas invariants of $\BA^3$ is given by $$\begin{aligned}
\DT_r^{\KK}(\BA^3,(-1)^r q,t,w)=\Exp\left(\mathsf F_r( q,t_1,t_2,t_3)\right),\end{aligned}$$ where, setting $\mathfrak t = t_1t_2t_3$, one defines $$\mathsf F_r( q,t_1,t_2,t_3)= \frac{[\mathfrak{t}^r]}{[\mathfrak{t}][\mathfrak{t}^{\frac{r}{2}}q ][\mathfrak{t}^{\frac{r}{2}} q^{-1} ]}\frac{[t_1t_2][t_1t_3][t_2t_3]}{[t_1][t_2][t_3]}.$$
This result was conjectured in [@MR2545054] by Awata and Kanno, who also proved it $\bmod q^4$, i.e. up to $3$ instantons.
As we mentioned in the Introduction, N. Arbesfeld and Ya. Kononov also obtained, simultaneously and independently, a different proof of Theorem \[thm: K theoretic of points higher rank\] [@Noah_Yasha].
The proof of Theorem \[thm: K theoretic of points higher rank\] will follow essentially by taking suitable limits of the weights $w_i$. To perform such limits, we grant for now the following result, anticipated in in the Introduction.
\[thm:independence\] The generating function $\DT_r^{\KK}(\BA^3,q,t,w)$ does not depend on the weights $w_1,\dots,w_r$.
We devote Appendix \[sec:framing\_independence\] to the proof of this crucial fact. Granting for now the independence of the generating series of the weights $w_i$, we specialise them to arbitrary values and take arbitrary limits. We set $w_i=L^{i}$ for $i=1,\ldots,r$ and compute the limit for $L\to \infty$.
\[lemma: lim in K theory Vij\] Let $i< j$. Then we have $$\lim_{L\to \infty}[-\mathsf V_{ij}][-\mathsf V_{ji}]\big|_{w_i=L^{i}}= (-\mathfrak{t}^{\frac{1}{2}})^{|\pi_j|-|\pi_i|}.$$
Notice that all monomials in $\mathsf V_{ij}$ are of the form $w_i^{-1}w_j \lambda$ for $\lambda$ a monomial in $t_1,t_2,t_3$. Then $$[w_i^{-1}w_j\lambda]\big|_{w_i=L^{i}}=(L^{j-i}\lambda)^{\frac{1}{2}}(1- L^{i-j}\lambda^{-1}).$$ Write $\mathsf Q_i=\sum_{\mu}t^{\mu}$ and $\mathsf Q_j=\sum_{\nu}t^{\nu}$. Taking limits, we obtain $$\begin{aligned}
\lim_{L\to \infty}[-\mathsf V_{ij}]\big|_{w_i=L^{i}}&=\lim_{L\to \infty}[-w_i^{-1}w_j(\mathsf Q_j- \overline{\mathsf Q}_i\mathfrak{t}^{-1}+\overline{\mathsf Q}_i \mathsf Q_j\mathfrak{t}^{-1}(1-t_1)(1-t_2)(1-t_3) ) ]\big|_{w_i=L^{i}} \\
&=\lim_{L\to \infty} L^{\frac{j-i}{2}(|\pi_i|-|\pi_j|)}\frac{\prod_\mu (t^{-\frac{\mu}{2}}\mathfrak{t}^{-\frac{1}{2}})}{\prod_{\nu} t^{\frac{\nu}{2}}}.\end{aligned}$$ Similarly, we obtain $$\begin{aligned}
\lim_{L\to \infty}[-\mathsf V_{ji}]\big|_{w_i=L^{i}}&= (-1)^{\rk (-\mathsf V_{ji})}\lim_{L\to \infty}[-\overline{\mathsf V}_{ji}]\big|_{w_i=L^{i}} \\
&=(-1)^{|\pi_i|-|\pi_j|}\lim_{L\to \infty}[-w_i^{-1}w_j(\overline{\mathsf Q}_i- {\mathsf Q}_j\mathfrak{t}-\overline{\mathsf Q}_i \mathsf Q_j(1-t_1)(1-t_2)(1-t_3) ) ]\big|_{w_i=L^{i}}\\
&=(-1)^{|\pi_i|-|\pi_j|}\lim_{L \to \infty} L^{\frac{j-i}{2}(|\pi_j|-|\pi_i|)}\frac{\prod_\nu (t^{\frac{\nu}{2}}\mathfrak{t}^{\frac{1}{2}})}{\prod_{\mu} t^{-\frac{\mu}{2}}}.\end{aligned}$$ We conclude, as required, that $$\lim_{L\to \infty}[-\mathsf V_{ij}][-\mathsf V_{ji}]\big|_{w_i=L^{i}}= (-\mathfrak{t}^{\frac{1}{2}})^{|\pi_j|-|\pi_i|}.\qedhere$$
\[lemma: combinatorical trick\] Let $x$ be a variable and $c_i\in \mathbb{Z}$, for $i=1,\ldots,r$. Then we have $$\prod_{1\leq i<j\leq r}x^{c_j-c_i}=\prod_{i=1}^r x^{(-r-1+2i)c_i}.$$
The assertion holds for $r=1$ as the productory on the left hand side is empty. Assume it holds for $r-1$. Then we have: $$\begin{aligned}
\prod_{1\leq i<j\leq r}x^{c_j-c_i}&= \prod_{1\leq i<j\leq r-1}x^{c_j-c_i}\prod_{i=1}^{r-1}x^{c_r-c_i}\\
&= x^{(r-1)c_r}\prod_{i=1}^{r-1} x^{(-r-1+2i)c_i}\\
&= \prod_{i=1}^r x^{(-r-1+2i)c_i}. \qedhere\end{aligned}$$
Combining Lemma \[lemma: lim in K theory Vij\] with Lemma \[lemma: combinatorical trick\] we can express the rank $r$ K-theoretic DT theory of $\BA^3$ as a product of $r$ copies of the rank $1$ K-theoretic DT theory.
\[thm: r copies of rank 1 K theory\] There is an identity $$\DT_r^{\KK}(\BA^3,(-1)^rq,t,w)=\prod_{i=1}^r \DT_1^{\KK}(\BA^3,-q\mathfrak{t}^{\frac{-r-1}{2} +i},t).$$
Set $w_i=L^i$. The generating series $\DT_r^{\KK}(\BA^3,q,t,w) $ can be computed in the limit $L\to \infty$: $$\begin{aligned}
\lim_{L\to \infty} \DT_r^{\KK}(\BA^3,q,t,w) &= \lim_{L\to \infty}\sum_{\overline{\pi}}q^{|\overline{\pi}|}\prod_{i,j=1}^r [-\mathsf V_{ij}]\\
&= \lim_{L\to \infty}\sum_{\overline{\pi}}\prod_{i=1}^r q^{|\pi_i|}[-\mathsf V_{ii}]\prod_{1\leq i<j\leq r}[-\mathsf V_{ij}][-\mathsf V_{ji}]\\
&= \sum_{\overline{\pi}}\prod_{i=1}^r q^{|\pi_i|}[-\mathsf V_{ij}]\prod_{1\leq i<j\leq r}(-\mathfrak{t}^{\frac{1}{2}})^{|\pi_j|-|\pi_i|}\\
&= \sum_{\overline{\pi}}\prod_{i=1}^r q^{|\pi_i|}[-\mathsf V_{ii}]\prod_{i=1}^r (-\mathfrak{t}^{\frac{1}{2}})^{(-r-1+2i)|\pi_i|}\\
&= \sum_{\overline{\pi}}\prod_{i=1}^r[-\mathsf V_{ii}] q^{|\pi_i|}(-1)^{(r+1)|\pi_i|}\mathfrak{t}^{(\frac{-r-1}{2} +i)|\pi_i|}\\
&=\sum_{\overline{\pi}}\prod_{i=1}^r[-\mathsf V_{ii}] \left( (-1)^{(r+1)}q\mathfrak{t}^{\frac{-r-1}{2} +i} \right)^{|\pi_i|}\\
&=\prod_{i=1}^r \DT_1^{\KK}(\BA^3,(-1)^{(r+1)}q\mathfrak{t}^{\frac{-r-1}{2} +i}, t). \qedhere\end{aligned}$$
We can now prove Theorem \[thm: K theoretic of points higher rank\] (i.e. Theorem \[mainthm:K-theoretic\] from the Introduction).
[Theorem \[thm: K theoretic of points higher rank\]]{}
Define $$G_{r,i}(q,t_1,t_2,t_3)=\mathsf F_1(q\mathfrak{t}^{\frac{-r-1}{2} +i}, t_1,t_2,t_3).$$ We have $$\begin{aligned}
\DT_1^{\KK}(\BA^3,-q\mathfrak{t}^{\frac{-r-1}{2} +i},t)
&=\exp{\left(\sum_{n\geq 1}\frac{1}{n}\frac{1}{[\mathfrak{t}^{\frac{n}{2}}q^n\mathfrak{t}^{n(\frac{-r-1}{2} +i)}][\mathfrak{t}^{\frac{n}{2}}q^{-n}\mathfrak{t}^{n(\frac{r+1}{2} -i)}]}\frac{[t^n_1t^n_2][t^n_1t^n_3][t^n_2t^n_3]}{[t^n_1][t^n_2][t^n_3]} \right)} \\
&=\Exp\left(G_{r,i}(q,t_1,t_2,t_3)\right).\end{aligned}$$ By Theorem \[thm: r copies of rank 1 K theory\] and Theorem \[thm: Okounkov rank 1\] it is enough to show that $\mathsf F_r=\sum_{i=1}^r G_{r,i} $, or equivalently $$\sum_{i=1}^r \frac{1}{[\mathfrak{t}^{\frac{1}{2}}q\mathfrak{t}^{\frac{-r-1}{2} +i}][\mathfrak{t}^{\frac{1}{2}}q^{-1}\mathfrak{t}^{\frac{r+1}{2} -i}]} = \frac{[\mathfrak{t}^r]}{[\mathfrak{t}][\mathfrak{t}^{\frac{r}{2} }q][\mathfrak{t}^{\frac{r}{2}}q^{-1}]}.$$ It is easy to check this is true for $r=1,2$. Let now $r\geq 3 $: we perform induction separately on even and odd cases. Assume the claimed identity holds for $r-2$. In both cases we have $$\begin{aligned}
\sum_{i=1}^r&\frac{1}{[\mathfrak{t}^{\frac{1}{2}}q\mathfrak{t}^{\frac{-r-1}{2} +i}][\mathfrak{t}^{\frac{1}{2}}q^{-1}\mathfrak{t}^{\frac{r+1}{2}-i}]}\\
&=\sum_{i=1}^{r-2} \frac{1}{[\mathfrak{t}^{\frac{1}{2}}q\mathfrak{t}^{\frac{-(r-2)-1}{2} +i}][\mathfrak{t}^{\frac{1}{2}}q^{-1}\mathfrak{t}^{\frac{(r-2)+1}{2} -i}]}+ \frac{1}{[q\mathfrak{t}^{-\frac{r}{2} +1}][q^{-1}\mathfrak{t}^{\frac{r}{2}}]}+\frac{1}{[q\mathfrak{t}^{\frac{r}{2} }][q^{-1}\mathfrak{t}^{-\frac{r}{2}+1}]}\\
&= \frac{[\mathfrak{t}^{r-2}]}{[\mathfrak{t}][\mathfrak{t}^{\frac{r-2}{2}}q][\mathfrak{t}^{\frac{r-2}{2}}q^{-1}]}- \frac{1}{[\mathfrak{t}^{\frac{r-2}{2}}q^{-1}][\mathfrak{t}^{\frac{r}{2}}q^{-1}]}-\frac{1}{[\mathfrak{t}^{\frac{r}{2} }q][\mathfrak{t}^{\frac{r-2}{2}}q]}\\
&= \frac{1}{[\mathfrak{t}][\mathfrak{t}^{\frac{r}{2} }q][\mathfrak{t}^{\frac{r}{2}}q^{-1}]}\cdot \frac{[\mathfrak{t}^{r-2}][\mathfrak{t}^{\frac{r}{2} }q][\mathfrak{t}^{\frac{r}{2}}q^{-1}]-[\mathfrak{t}][\mathfrak{t}^{\frac{r}{2} }q][\mathfrak{t}^{\frac{r-2}{2}}q]-[\mathfrak{t}][\mathfrak{t}^{\frac{r}{2} }q^{-1}][\mathfrak{t}^{\frac{r-2}{2}}q^{-1}] }{[\mathfrak{t}^{\frac{r-2}{2}}q][\mathfrak{t}^{\frac{r-2}{2}}q^{-1}]}\\
&= \frac{[\mathfrak{t}^r]}{[\mathfrak{t}][\mathfrak{t}^{\frac{r}{2} }q][\mathfrak{t}^{\frac{r}{2}}q^{-1}]}\end{aligned}$$ by which we conclude the proof.
Even though the work of Arbesfeld–Kononov [@Noah_Yasha] has different motivations than ours, their proof of Theorem \[thm: K theoretic of points higher rank\] is also obtained as a consequence of the statement of Theorem \[thm:independence\], whose proof in loc. cit. differs from the one we present in §\[sec: functional repr. K teoria\].
Comparison with motivic DT invariants {#motivic_comparison}
-------------------------------------
Let $f\colon U\to \BA^1$ be a regular function on a smooth scheme $U$, and let $\hat\mu$ be the group of all roots of unity. The critical locus $Z = \crit (f) \subset U$ inherits a canonical *virtual motive* [@BBS], i.e. a $\hat\mu$-equivariant motivic class $$[Z]_{\vir} = -\BL^{-\frac{\dim U}{2}}\left[\phi_f\right]\,\in\,\CM^{\hat\mu}_{\BC} = K_0^{\hat\mu}(\Var_{\BC})\bigl[\BL^{-\frac{1}{2}}\bigr]$$ such that $e [Z]_{\vir} = e_{\vir}(Z)$, where $e_{\vir}(-)$ is Behrend weighted Euler characteristic and the Euler number specialisation prescribes $e(\BL^{-1/2}) = -1$. The motivic class $[\phi_f]$ is the (absolute) motivic vanishing cycle class introduced by Denef and Loeser [@DenefLoeser1].
The virtual motive of $\Quot_{\BA^3}(\OO^{\oplus r},n) = \crit(f_n)$, with respect to the critical structure of Proposition \[prop:SPOT\_A\^3\], was computed in [@ThesisR Prop. 2.3.6]. The result is as follows. Let $\DT_r^{\mot}(\BA^3,q)\in \CM_{\BC} \llbracket q \rrbracket$ be the generating function of the virtual motives $[\Quot_{\BA^3}(\OO^{\oplus r},n)]_{\vir}$. Then one has $$\DT_r^{\mot}(\BA^3,q)
= \prod_{m\geq 1}\prod_{k=0}^{rm-1}\left(1-\BL^{2+k-\frac{rm}{2}}q^m\right)^{-1}.$$ The case $r=1$ was computed in [@BBS]. The general case is obtained in a similar fashion [@ThesisR; @Cazzaniga_Thesis]. Moreover, it is immediate to verify that $\DT_r^{\mot}$ satisfies a product formula analogous to the one proved in Theorem \[thm: r copies of rank 1 K theory\] for the K-theoretic invariants: we have $$\label{eqn:motivic_factorisation}
\DT_r^{\mot}(\BA^3,q)
=\prod_{i=1}^r \DT_1^{\mot}\left(\BA^3,q \BL^{\frac{-r-1}{2}+i}\right).$$ In particular, up to the substitution $\mathfrak t^{\frac{1}{2}} \to -\BL^{\frac{1}{2}}$, the factorisation is equivalent to the K-theoretic one (Theorem \[thm: r copies of rank 1 K theory\]). As observed in [@Quot19 § 4], the (signed) motivic partition function admits an expression in terms of the motivic exponential, namely $$\label{eqn:motivic_exp}
\DT_r^{\mot}(\BA^3,(-1)^rq)
= \Exp\left(\frac{(-1)^rq \BL^{\frac{3}{2}}}{\bigl(1-(-1)^rq\BL^{\frac{r}{2}}\bigr)\bigl(1-(-1)^rq\BL^{-\frac{r}{2}}\bigr)}\frac{\BL^{\frac{r}{2}}-\BL^{-\frac{r}{2}}}{\BL^{\frac{1}{2}}-\BL^{-\frac{1}{2}}} \right).$$
Given their structural similarities, we believe it is an interesting problem to compare the K-theoretic partition function with the motivic one.
It is worth noticing that Formula can be recovered from the factorisation , just as we discovered in the K-theoretic case during the proof of Theorem \[thm: K theoretic of points higher rank\]. This fact follows immediately by the properties of the plethystic exponential.
A virtual motive for $\Quot_Y(F,n)$ was defined in [@Quot19 §4] for every locally free sheaf $F$ on a $3$-fold $Y$. Just as in the case of the naive motives of the Quot scheme [@ricolfi2019motive], the partition function only depends on the motivic class $[Y] \in K_0(\Var_{\BC})$ and on $r = \rk F$.
The higher rank cohomological DT partition function {#sec:cohomological_invariants}
===================================================
Cohomological reduction {#sec: cohom reduction}
-----------------------
One should think of K-theoretic invariants as refinements of the cohomological ones, as by taking suitable limits one fully recovers $\DT_r^{\coh}(\mathbb{A}^3,q,s)$ from $\DT^{\KK}_r(\mathbb{A}^3,q,t)$. We make this precise in the remainder of this section.
Let $\TT\cong \BC^g$ be an algebraic torus and let $t_1,\dots, t_g$ be its coordinates. Recall that the Chern character gives a natural transformation from (equivariant) K-theory to the (equivariant) Chow group with rational coefficients by sending $t_i\mapsto e^{s_i}$, where $s_i=c_1^\TT(t_i)$. We can formally extend it to $$\begin{tikzcd}
\mathbb{Z}[t_1^{\pm 1},\dots,t_g^{\pm 1}]\arrow{r}{\ch}\arrow{d}& \BQ \llbracket s_1,\dots,s_g\rrbracket\arrow{d}\\
\mathbb{Z}[t_1^{\pm b},\dots,t_g^{\pm b} | b\in \BC]\arrow{r}{ \ch} & \BC\llbracket s_1,\dots,s_g\rrbracket
\end{tikzcd}$$ by sending $t_i^b\mapsto e^{b s_i}$, where $b\in \BC$.
In §\[sec: preliminaries K theory\] we defined the symmetrised transformation $[t^\mu]=t^{\frac{\mu}{2}}-t^{-\frac{\mu}{2}}$, which enjoys the following *linearisation* property: $$\begin{aligned}
[\ch(t^{b \mu})]&= e^{\frac{b \mu\cdot s}{2}}(1-e^{-b \mu \cdot s})= b e^{\TT}(t^{\mu} )+ o(b^2).\end{aligned}$$ In other words, $ e^{\TT}(\,\cdot\,)$ is the first-order approximation of $[\,\cdot\,]$ in $\TT$-equivariant Chow groups. For a virtual representation $V=\sum_\mu t^{\mu}- \sum_\nu t^\nu\in K^\TT_0(\pt)$, denote by $V^b=\sum_\mu t^{b\mu}- \sum_\nu t^{b\nu} $ the virtual representation where we formally substitute each weight $t^\mu$ with $t^{b \mu}$. We have the identity $$\begin{aligned}
[\ch(V^b)]=\frac{\prod_\mu[\ch(t^{b\mu})]}{\prod_\nu[\ch(t^{b\nu})]} =b^{\rk V} \frac{\prod_{\mu} ( e^{\TT}(t^{\mu} ) + o(b))}{\prod_{\nu} ( e^{\TT}(t^{\nu} ) + o(b))}.\end{aligned}$$ If $\rk V=0$, by taking the limit for $b\to 0$ we conclude $$\begin{aligned}
\lim_{b\to 0}\, [\ch(V^b)]= e^{\TT}(V).\end{aligned}$$
It is clear from the definition of $\ch(\,\cdot\,)$ and $[\,\cdot\,]$ that these two transformations commute with each other. This proves the following relation between K-theoretic invariants and cohomological invariants of the local model.
\[limit K theory cor\] There is an identity $$\DT_r^{\coh}(\mathbb{A}^3,q,s,v)=\lim_{b\to 0}\DT_r^{\KK}(\mathbb{A}^3,q,e^{b s},e^{bv}).$$
Follows from the description of the generating series of K-theoretic invariants as $$\DT_r^{\KK}(\mathbb{A}^3,q,t,w) =\sum_{n\geq 0}q^n\sum_{[S]\in \Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\TT}} [-T^{\vir}_S]$$ and by noticing that $\rk T_S^{\vir}=0$.
Thanks to the $v$-independence, we can now rename $$\DT_r^{\coh}(\BA^3,q,s) = \DT_r^{\coh}(\BA^3,q,s,v).$$ We are ready to prove Theorem \[mainthm:cohomological\] from the Introduction.
\[thm:cohomological\] The generating function of rank $r$ cohomological Donaldson–Thomas invariants of $\BA^3$ is given by $$\DT_r^{\coh}(\mathbb{A}^3,q,s)=\mathsf M((-1)^rq)^{-r\frac{(s_1+s_2)(s_1+s_3)(s_2+s_3)}{s_1s_2s_3}}.$$
By Corollary \[limit K theory cor\] and Theorem \[thm: K theoretic of points higher rank\], we just need to compute the limit $$\lim_{b\to 0}\DT_r^{\KK}\left(\mathbb{A}^3,(-1)^rq,e^{b s}\right)= \lim_{b\to 0}\Exp\left(\mathsf F_r( q,t_1^b,t_2^b,t_3^b)\right).$$ Denote for ease of notation $\mathfrak{s}=c_1^{\TT}(\mathfrak{t})=s_1+s_2+s_3$. By the definition of plethystic exponential, recalled in , we have $$\begin{gathered}
\lim_{b\to 0}\Exp\left(\mathsf F_r(q,t_1^b,t_2^b,t_3^b)\right) \\
=\exp \sum_{k\geq 1}\frac{1}{k}\left( \lim_{b\to 0} \frac{[e^{bkr\mathfrak{s} }]}{[ e^{bk\mathfrak{s} }][e^{\frac{bkr}{2}\mathfrak{s} } q^k ][ e^{\frac{bkr}{2}\mathfrak{s} } q^{-k} ]}\frac{[e^{bk (s_1+s_2)}][e^{bk (s_1+s_3)}][e^{bk (s_2+s_3)}]}{[e^{bk s_1}][e^{bk s_2}][e^{bk s_3}]} \right).\end{gathered}$$ We have $$\begin{aligned}
\lim_{b\to 0}\frac{[e^{bk (s_1+s_2)}][e^{bk (s_1+s_3)}][e^{bk (s_2+s_3)}]}{[e^{bk s_1}][e^{bk s_2}][e^{bk s_3}]}=
\frac{(s_1+s_2)(s_1+s_3)(s_2+s_3)}{s_1s_2s_3},\end{aligned}$$ and $$\begin{aligned}
\lim_{b\to 0} \frac{[e^{bkr\mathfrak{s} }]}{[ e^{bk\mathfrak{s} }][e^{\frac{bkr}{2}\mathfrak{s} } q^k ][ e^{\frac{bkr}{2}\mathfrak{s} } q^{-k} ]}&= \frac{r}{[q^{k} ][ q^{-k} ]}
= -r\cdot \frac{q^{k}}{(1-q^{k})^2}.\end{aligned}$$ Recall the plethystic exponential form of the MacMahon function $$\begin{aligned}
\mathsf M(q)= \prod_{n\geq 1} (1-q^n)^{-n}=\mathrm{Exp}\bigg(\frac{q}{(1-q)^2} \bigg).\end{aligned}$$ We conclude $$\begin{aligned}
\lim_{b\to 0} \DT_r^{\KK}(\mathbb{A}^3,(-1)^r q, e^{b s})&= \exp\left(-r\cdot \frac{(s_1+s_2)(s_1+s_3)(s_2+s_3)}{s_1s_2s_3} \sum_{k\geq 1} \frac{1}{k} \frac{q^{k}}{(1-q^{k})^2}\right)\\
&= \mathsf M(q)^{-r\frac{(s_1+s_2)(s_1+s_3)(s_2+s_3)}{s_1s_2s_3}}.\qedhere\end{aligned}$$
Thus we proved Szabo’s conjecture [@Szabo Conj. 4.10].
The specialisation $$\DT_r^{\coh}(\A^3,q,s)\big|_{s_1+s_2+s_3 = 0} = \mathsf M((-1)^rq)^r,$$ recovering Formula , was already known in physics, see e.g. [@CSS].
We end this section with a small variation of Theorem \[mainthm:cohomological\].
\[cor:independence\_on\_lambda\_w\] Fix an $r$-tuple $\lambda = (\lambda_1,\ldots,\lambda_r)$ of $\BT_1$-equivariant line bundles on $\BA^3$. Then there is an identity $$\DT_r^{\coh}(\BA^3,q,s) = \DT_r^{\coh}(\BA^3,q,s,v)_\lambda,$$ where the right hand side was defined in .
We have $$T_{S,\lambda}^{\vir}=\sum_{i,j}\lambda_i^{-1}\lambda_j \mathsf V_{ij}.$$ Let $\mathsf V_{ij}=\sum_\mu w_i^{-1}w_j t^\mu $ be the decomposition into weight spaces. A monomial in $T_{S,\lambda}^{\vir}$ is of the form $ \lambda_i^{-1}\lambda_jw_i^{-1}w_jt^{\mu}$ and its Euler class is $$\begin{aligned}
e^{\TT}( \lambda_i^{-1}\lambda_jw_i^{-1}w_jt^{\mu})&=\mu\cdot s + v_j+c^{\TT}_1(\lambda_j)-v_i-c^{\TT}_1(\lambda_i)\\
&= \mu\cdot s + \overline{v}_j-\overline{v}_i\end{aligned}$$ where we define $ \overline{v}_i= v_i+c^{\TT}_1(\lambda_i)$. We conclude that $$\DT_r^{\coh}(\BA^3,q,s,v)_\lambda=\DT_r^{\coh}(\BA^3,q,s,\overline{v}),$$ which does not depend on $\overline{v}$ by Theorem \[independence of Z\^K on w\].
\[example r=2, n=1\] Set $r=2$, $n=1$, so that the only $\TT$-fixed points in $\Quot_{\BA^3}(\OO^{\oplus 2},1)$ are the direct sums of ideal sheaves $$S_1 = \mathscr I_{\pt}\oplus \OO \subset \OO^{\oplus 2},\quad S_2 = \OO \oplus \mathscr I_{\pt} \subset \OO^{\oplus 2},$$ where $\pt = (0,0,0) \in \BA^3$ is the origin. One computes $$\begin{aligned}
T^{\vir}_{S_1} &= 1-\frac{1}{t_1t_2t_3}+\frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3} - w_1^{-1}w_2\frac{1}{t_1t_2t_3} + w_2^{-1}w_1 \\
T^{\vir}_{S_2} &= 1-\frac{1}{t_1t_2t_3}+\frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3}- w_2^{-1}w_1\frac{1}{t_1t_2t_3} + w_1^{-1}w_2.\end{aligned}$$ Therefore, the cohomological DT invariant is $$\int_{[\Quot_{\BA^3}(\OO^{\oplus 2},1)]^{\vir}} 1 = e^{\TT}(-T^{\vir}_{S_1}) + e^{\TT}(-T^{\vir}_{S_2}).$$ The part that could possibly depend on the framing parameters $v_1$ and $v_2$ is, in fact, constant: $$\begin{gathered}
e^{\TT}\left(w_1^{-1}w_2\frac{1}{t_1t_2t_3} - w_2^{-1}w_1\right) + e^{\TT}\left(w_2^{-1}w_1\frac{1}{t_1t_2t_3} - w_1^{-1}w_2\right) \\
= \frac{-v_1+v_2-\mathfrak s}{v_1-v_2} + \frac{v_1-v_2-\mathfrak s}{v_2-v_1} = -2.\end{gathered}$$ Let $\lambda_1$ and $\lambda_2$ be two $\BT_1$-equivariant line bundles. After the substitutions $w_i \to w_i\lambda_i$, and setting $\overline v_i = v_i + c_1(\lambda_i)$, the final sum of Euler classes depending on $\overline v$ becomes $$2\frac{\overline v_2-\overline v_1}{\overline v_1-\overline v_2} = -2.$$
Elliptic Donaldson–Thomas invariants {#sec:elliptic_invariants}
====================================
Chiral elliptic genus
---------------------
In [@BBPT] an elliptic generalisation of the residue forms studied in Appendix \[sec:framing\_independence\] is given. In physics the invariants computed in loc. cit. are obtained as the superconformal index of a D1-D7 brane system on a type IIB $\mathcal N=1$ supersymmetric background, where $r$ D7-branes wrap the product of a $3$-fold by a torus, i.e. $X_3\times T^2$, while $n$ D1-branes wrap $T^2$. The connection with enumerative geometry is then given via BPS-bound states counting, as brane systems often provide interesting constructions of relevant moduli spaces. In the case at hand, when $X_3=\BA^3$, the D1-D7 brane system considered has a BPS moduli space which can be naturally identified with the moduli space parametrising quotients of length $n$ of $\OO_{\BA^3}^{\oplus r}$. The superconformal index is usually identified in the physics literature with the elliptic genus of such a moduli space. This however does not coincide with the usual notion of (virtual) elliptic genus, as the coupling to the D7-branes breaks half of the chiral supersymmetry, thus leading to an effective $2d$ $\mathcal N=(0,2)$ GLSM on $T^2$ for the D1-brane dynamics, whose Witten index generalises the K-theoretic DT invariants of $\BA^3$ and provide a sort of chiral (or $1/2$ BPS) version of the elliptic genus of the Quot scheme. In this section we give a mathematical definition of the elliptic invariants computed in [@BBPT], and show that our definition leads precisely to the computation of the same quantities found in [@BBPT §3].
Let $X$ be a scheme carrying a perfect obstruction theory $\BE\to \BL_X$ of virtual dimension $\vd=\rk \BE$.
If $F$ is a rank $r$ vector bundle on $X$ we define $$\label{E_1/2}
\mathcal E_{1/2}(F)=\bigotimes_{n\ge 1}\Sym_{p^n}^\bullet\left(F\oplus F^\vee\right)\in 1+p\cdot K^0(X)\llbracket p\rrbracket$$ where the total symmetric algebra $\Sym^\bullet_p(F)=\sum_{i\ge 0}p^i[S^iF]\in K^0(X)\llbracket p \rrbracket$ satisfies $\Sym^\bullet_p(F)=1/\Lambda^\bullet_{-p}(F)$. Note that $\mathcal E_{1/2}$ defines a homomorphism from the additive group $K^0(X)$ to the multiplicative group $1+p\cdot K^0(X)\llbracket p\rrbracket$. Set $$\ELL_{1/2}(F;p)=(-p^{-\frac{1}{12}})^{\rk F}\ch\left(\mathcal E_{1/2}(F)\right)\cdot\td(F)\in A^*(X)\llbracket p\rrbracket[p^{\pm \frac{ 1}{12}}],$$ so that $\ELL_{1/2}(-;p)$ extends to a group homomorphism from $K^0(X)$ to $A^*(X)\llbracket p\rrbracket[p^{\pm\frac{1}{12}}]$.
We can then define the virtual chiral elliptic genus as follows.
\[half bps elliptic genus\] Let $X$ be a proper scheme with a perfect obstruction theory and $V\in K^0(X)$. The *virtual chiral elliptic genus* is defined as $$\Ell_{1/2}^{\vir}(X,V;p)=(-p^{-\frac{1}{12}})^{\vd}\chi^{\vir}\left(X,\mathcal E_{1/2}(T_X^{\vir})\otimes V\right)\in\BZ \llbracket p\rrbracket[p^{\pm\frac{1}{12}}].$$ By the virtual Riemann–Roch theorem of [@Fantechi_Gottsche] we can also compute the virtual chiral elliptic genus as $$\Ell_{1/2}^{\vir}(X,V;p)=\int_{[X]^{\vir}}\ELL_{1/2}(T_X^{\vir};p)\cdot\ch(V).$$
One may give a more general definition by adding a “mass deformation” and defining $\mathcal E_{1/2}^{(y)}(F)$ for $F\in K^0(X)$ as $$\mathcal E_{1/2}^{(y)}(F;p)=\bigotimes_{n\ge 1}\Sym_{y^{-1}p^n}^\bullet\left(F\right)\otimes \Sym_{yp^n}^\bullet\left(F^\vee\right)\in 1+p\cdot K^0(X)[y,y^{-1}]\llbracket p \rrbracket,$$ so we recover the standard definition of virtual elliptic genus by taking $\mathcal E(F)=\mathcal E_{1/2}^{(1)}(F;p)\otimes\mathcal E_{1/2}^{(y)}(-F;p)$, cf. [@Fantechi_Gottsche § 6].
We have: $$\bigwedge_{y}(F)=S_{-y}(-F)$$ and so $$E_{1/2}^{(1)}(F;p)E_{1/2}^{(y)}(-F;p)= \bigotimes_{n\ge 1}\bigwedge_{-y^{-1}p^n}^\bullet\left(F\right)\otimes \bigwedge_{-yp^n}^\bullet\left(F^\vee\right) \otimes \Sym_{p^n}^\bullet\left(F\oplus F^\vee\right)$$
Let $X$ be a proper scheme with a perfect obstruction and let $V\in K^0(X)$. Then the virtual chiral elliptic genus $\Ell_{1/2}^{\vir}(X,V;p)$ is deformation invariant.
The statement follows directly from Definition \[half bps elliptic genus\] and [@Fantechi_Gottsche Theorem 3.15].
Let $X$ be equipped with a symmetric perfect obstruction theory $[E^{-1}\to E^0]$. Then $\Ell_{1/2}^{\vir}(X,(\mathcal K_X^{\vir})^{1/2};p)$ is $SL(2,\BZ)$-invariant.
Let $p=e^{2\pi i\tau}$, $\Im\tau>0$. The modular group $SL(2,\BZ)$ acts as $$\tau\mapsto\frac{a\tau+b}{c\tau+d},\qquad \begin{pmatrix}
a & b\\
c & d
\end{pmatrix}\in SL(2,\BZ).$$ In order to prove the statement, let $F$ be any vector bundle on $X$, with Chern roots $f_i$, $i=1,\dots,r$. Assume $\det F^\vee$ admits a square root in $K^0(X)$, so we have $$\ELL_{1/2}(F;p)\cdot\ch(-p^{2/3}\det F^\vee)^{1/2}=\prod_{\ell=1}^rf_\ell\frac{\eta(p)}{\theta\left(\frac{f_\ell}{2\pi i},p\right)},$$ where $\theta(y,z)$ denotes the Jacobi theta function $$\theta(y,z)=-iz^{1/8}(y^{1/2}-y^{-1/2})\prod_{n=1}^\infty(1-z^n)(1-yz^n)(1-y^{-1}z^n).$$ It is then known that $\theta(y,z)$ is a Jacobi form of weight $1/2$ and index $1/2$ and that $\eta(z)$ is a modular form of weight $1/2$ with character. Indeed we have $$\begin{aligned}
&\eta(\tau+1)=e^{i\pi/12}\eta(\tau)\\
&\eta(-1/\tau)=\sqrt{-i\tau}\eta(\tau)\end{aligned}$$ and $$\begin{aligned}
&\theta(z|\tau+1)=e^{i\pi/4}\theta(z|\tau)\\
&\theta(z/\tau|-1/\tau)=-i\sqrt{-i\tau}e^{\pi i z^2/\tau}\theta(z|\tau)\end{aligned}$$ Moreover, as $T_X^{\vir}=[E_0-E_1]$ we have $\ELL_{1/2}(T_X^{\vir};p)=\ELL_{1/2}(E_0;p)/\ELL_{1/2}(E_1;p)$. The fact that the perfect obstruction theory on $X$ is symmetric is enough to conclude.
Let now $V=\sum_\mu t^\mu$ be a $\TT$-module as in §\[sec: preliminaries K theory\]. The trace of its symmetric algebra is given by $$\tr\left(\Sym^\bullet_p(V)\right)=\tr\left(\frac{1}{\Lambda^\bullet_{-p}(V)}\right)=\prod_\mu\frac{1}{1-pt^\mu}.$$ Let us now assume as in §\[sec: preliminaries K theory\] that $\det V$ is a square in $K_{\TT}^0(\pt)$. We can then compute the trace of the symmetric product in as $$\begin{aligned}
\tr\left(\bigotimes_{n\ge 1}\Sym_{p^n}^\bullet\left(V\oplus V^\vee\right)\right)&=\prod_{\mu}\prod_{n\ge 1}\frac{1}{(1-p^nt^\mu)(1-p^nt^{-\mu})},\end{aligned}$$ $$\begin{aligned}
\nonumber\tr\left(\frac{\bigotimes_{n\ge 1}\Sym_{p^n}^\bullet\left(V\oplus V^\vee\right)}{\Lambda^\bullet V^\vee}\right)&=\prod_{\mu}\frac{1}{1-t^{-\mu}}\prod_{n\ge 1}\frac{1}{(1-p^nt^\mu)(1-p^nt^{-\mu})}\\
&=\left(-ip^{1/8}\phi(p)\right)^{\rk V}\prod_\mu\frac{t^{\mu/2}}{\theta(p;t^\mu)},\label{trace_thetas}\end{aligned}$$ where $\phi(p)$ is the Euler function, i.e. $\phi(p)=\prod_n(1-p^n)$, and $\theta(p;y)$ denotes the Jacobi theta function $$\theta(p;y)=-ip^{1/8}(y^{1/2}-y^{-1/2})\prod_{n=1}^\infty(1-p^n)(1-yp^n)(1-y^{-1}p^n).$$ Combining everything together we get the identity $$\begin{aligned}
(-p^{-\frac{1}{12}})^{\rk V}\tr\left(\frac{\bigotimes_{n\ge 1}\Sym_{p^n}^\bullet\left(V\oplus V^\vee\right)\otimes\det (V^\vee)^{1/2}}{\Lambda^\bullet V^\vee}\right)&=(-p^{-\frac{1}{12}})^{\rk V}\left(-ip^{1/8}\phi(p)\right)^{\rk V}\prod_\mu\frac{1}{\theta(p;t^\mu)}\\
&=\prod_\mu i\frac{\eta(p)}{\theta(p;t^\mu)},\end{aligned}$$ where $\eta(p)$ is the Dedekind eta function $$\eta(p)=p^{\frac{1}{24}}\prod_{n\ge 1}(1-p^n).$$ We extend it by linearity to the case of a virtual $\TT$-representation $ V=\sum_\mu t^\mu-\sum_\nu t^\nu\in K_{\TT}^0(\pt)$, with $$\begin{aligned}
(-p^{-\frac{1}{12}})^{\rk V}\tr\left(\frac{\bigotimes_{n\ge 1}\Sym_{p^n}^\bullet\left( V\oplus V^\vee\right)\otimes\det ( V^\vee)^{1/2}}{\Lambda^\bullet V^\vee}\right)= (i\cdot \eta(p))^{\rk V} \frac{\prod_{\nu}\theta(p;t^\nu)}{\prod_{\mu}\theta(p;t^\mu)}.\end{aligned}$$ For the remainder of the section we set $p=e^{2\pi i \tau}$, with $\tau\in \mathbb{H}=\set{\tau\in \BC \colon \Im(\tau)>0}$. Denoting $\theta(\tau|z)=\theta(e^{2\pi i\tau};e^{2\pi iz})$, $\theta$ enjoys the modular behaviour $$\theta(\tau|z+a+b\tau)=(-1)^{a+b}e^{-2\pi ibz}e^{-i\pi b^2\tau}\theta(\tau|z),\qquad a,b\in\BZ.$$
Analogously to the measure $[\,\cdot\,]$ for K-theoretic invariants, we define the *elliptic measure* $$\theta[V]= (i\cdot \eta(p))^{-\rk V} \frac{\prod_{\mu}\theta(p;t^\mu)}{\prod_{\nu}\theta(p;t^\nu)},$$ which satisfies $\theta[\overline V]=(-1)^{\rk V}\theta[V]$. Notice that, if $\rk V=0$, the elliptic measure refines both $[\,\cdot\,]$ and $e^\TT(\,\cdot\,)$ $$\theta[V]\xrightarrow{p\to 0} [V]\xrightarrow{b\to 0} e^{\TT}(V)$$ where the second limit was dicussed in §\[sec: cohom reduction\].
The definition we gave for virtual chiral elliptic genus is reminiscent of what is commonly known in physics as the elliptic genus (or superconformal index) of a $2d$ $\mathcal N=(0,2)$ sigma model. Our definition actually matches the one in [@KawaiMohri] for $\mathcal N=(0,2)$ Landau–Ginzburg models. Indeed, in this case we are given an $n$-dimensional (compact) Kähler manifold $X$ together with a holomorphic vector bundle $E\to X$ such that $c_1(E)-c_1(T_X)=0\mod 2$. If we then consider the K-theory class $[V]=[T_X]-[E]$, the superconformal index of [@KawaiMohri] can be written as $$\boldsymbol{\mathcal I}(X,E;p)=(-ip^{-\frac{1}{12}})^{\rk V}\chi\left(X,\mathcal E_{1/2}(V)\otimes\det(V^\vee)^{\frac{1}{2}}\right)$$ and in terms of the Chern roots $v_i$, $w_j$ of $T_X$ and $E$, respectively, we also have (cf. [@FrancoGhimLeeSeong]) $$\boldsymbol{\mathcal I}(X,E;p)=\int_X\prod_{i=1}^r\frac{\theta(\tau|\frac{w_i}{2\pi i})}{\eta(p)}\prod_{j=1}^n\frac{v_j\eta(p)}{\theta(\tau|\frac{-v_j}{2\pi i})}.$$
Elliptic DT invariants {#sec:elliptic DT}
----------------------
\[def:elliptic DT of A\^3\] The generating series of elliptic DT invariants $\DT_r^{\rm ell}(\BA^3,q,t,w;p)$ is defined as $$\DT_r^{\rm ell}(\BA^3,q,t,w;p)=\sum_{n\ge 0}q^n\Ell^{\vir}_{1/2}(\Quot_{\BA^3}(\OO^{\oplus r},n),\mathcal K_{\vir}^{\frac{1}{2}};p)\in \BZ(\!(t,\mathfrak t^{\frac{1}{2}},w)\!)\llbracket p,q \rrbracket.$$ Being that $\Quot_{\BA^3}(\OO^{\oplus r},n)$ is not projective, but nevertheless carries the action of an algebraic torus $\TT$ with proper $\TT$-fixed locus, we define the invariants by means of virtual localisation, as we explained in §\[sec:vir\_loc\].
At each $\TT$-fixed point $[S]\in\Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}$, the localised contribution is $$\tr\left(\frac{\bigotimes_{n\ge 1}\Sym_{p^n}^\bullet\left(T_S^{\vir}\oplus T_S^{\vir,\vee}\right)}{\widehat{\Lambda}^\bullet T_S^{\vir,\vee}}\right)$$ from which we deduce that we can recover the K-theoretic invariants $\chi(\Quot_{\BA^3}(\OO^{\oplus r},n),\widehat{\OO}^{\vir})$ in the limit $p\to 0$. As for K-theoretic invariants, we have $$\begin{aligned}
\DT_r^{\rm ell}(\BA^3,q,t,w;p)\,&=\,\sum_{n\geq 0}q^n\sum_{[S]\in \Quot_{\BA^3}(\mathscr O^{\oplus r},n)^{\TT}}\theta[-T^{\vir}_S]\\
\,&=\,\sum_{\overline\pi}q^{\lvert \overline\pi\rvert}\prod_{i,j=1}^r\theta[-\mathsf V_{ij}],\end{aligned}$$ where $\overline{\pi}$ runs over all $r$-colored plane partitions.
Contrary to the case of K-theoretic and cohomological invariants, there exists no conjectural closed formula for elliptic DT invariants yet, even for the rank 1 case. Moreover, the generating series depends on the equivariant parameters of the framing torus, as shown in the following example.
\[example dependence elliptic\] Consider $Q_1^3=\Quot_{\BA^3}(\OO^{\oplus 3},1)$, whose only $\TT$-fixed points are $$S_1=\mathscr I_\pt\oplus\OO\oplus\OO\subset\OO^{\oplus 3},\quad S_2= \OO\oplus\mathscr I_\pt\oplus\OO\subset\OO^{\oplus 3},\quad S_3=\OO\oplus\OO\oplus\mathscr I_\pt\subset\OO^{\oplus 3},$$ with $\pt=(0,0,0)\in\BA^3$ as in Example \[example r=2, n=1\]. We have $$\begin{aligned}
T^{\vir}_{S_1} &= 1-\frac{1}{t_1t_2t_3}+\frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3} - w_1^{-1}w_2\frac{1}{t_1t_2t_3} + w_2^{-1}w_1 - w_1^{-1}w_3\frac{1}{t_1t_2t_3} + w_3^{-1}w_1 \\
T^{\vir}_{S_2} &= 1-\frac{1}{t_1t_2t_3}+\frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3}- w_2^{-1}w_1\frac{1}{t_1t_2t_3} + w_1^{-1}w_2 - w_2^{-1}w_3\frac{1}{t_1t_2t_3} + w_3^{-1}w_2 \\
T^{\vir}_{S_3} &= 1-\frac{1}{t_1t_2t_3}+\frac{(1-t_1)(1-t_2)(1-t_3)}{t_1t_2t_3}- w_3^{-1}w_1\frac{1}{t_1t_2t_3} + w_1^{-1}w_3 - w_3^{-1}w_2\frac{1}{t_1t_2t_3} + w_2^{-1}w_3\end{aligned}$$ by which we may compute the corresponding elliptic invariant. Set $w_j=e^{2\pi iv_j}$ and $t_\ell=e^{2\pi is_\ell}$, so that $$\begin{gathered}
\Ell^{\vir}_{1/2}(Q_1^3,\mathcal K_{\vir}^{\frac{1}{2}},t,w;p)=\diamondsuit\cdot\left( \frac{\theta(\tau|v_2-v_1-s)\theta(\tau|v_3-v_1-s)}{\theta(\tau|v_1-v_2)\theta(\tau|v_1-v_3)}\right.\\
\left.+\frac{\theta(\tau|v_1-v_2-s)\theta(\tau|v_3-v_2-s)}{\theta(\tau|v_2-v_1)\theta(\tau|v_2-v_3)}+
\frac{\theta(\tau|v_1-v_3-s)\theta(\tau|v_2-v_3-s)}{\theta(\tau|v_3-v_1)\theta(\tau|v_3-v_2)}\right),\end{gathered}$$ where $s=s_1+s_2+s_3$, with the overall factor $$\diamondsuit=\frac{\theta(\tau|s_1+s_2)\theta(\tau|s_1+s_3)\theta(\tau|s_2+s_3)}{\theta(\tau|s_1)\theta(\tau|s_2)\theta(\tau|s_3)}.$$ Moreover, by evaluating residues in $v_i-v_j=0$ one can realise that $\Ell^{\vir}_{1/2}(Q_1^3,\mathcal K_{\vir}^{1/2},t,w;p)$ has no poles in $v_i$. Indeed $$\Res_{v_1-v_2=0}\Ell^{\vir}_{1/2}(Q_1^3,\mathcal K_{\vir}^{\frac{1}{2}};p)=\diamondsuit\cdot\left(\frac{\theta(\tau|-s)\theta(\tau|v_3-v_2-s)}{\theta(\tau|v_2-v_3)}-\frac{\theta(\tau|-s)\theta(\tau|v_3-v_2-s)}{\theta(\tau|v_2-v_3)}\right)=0,$$ and the same occurs for any other pole involving the $v_i$’s. However, this does not imply the independence of the elliptic invariants from $v$, as we now suggest. Set $\overline{v}_i=v_i + a_i+b_i\tau$, with $a_i, b_i\in \BZ$, for $i=1,2,3$. Applying the quasi-periodicity of theta functions, we get $$\begin{gathered}
\Ell^{\vir}_{1/2}(Q_1^3,\mathcal K_{\vir}^{\frac{1}{2}},t,\overline{w};p)=\\
\frac{\diamondsuit}{\theta(\tau|\overline{v}_1-\overline{v}_2)\theta(\tau|\overline{v}_1-\overline{v}_3)\theta(\tau|\overline{v}_2-\overline{v}_3)}\cdot\left(\theta(\tau|\overline{v}_2-\overline{v}_1-s)\theta(\tau|\overline{v}_3-\overline{v}_1-s)\theta(\tau|\overline{v}_2-\overline{v}_3) \right.\\
\left. -\theta(\tau|\overline{v}_1-\overline{v}_2-s)\theta(\tau|\overline{v}_3-\overline{v}_2-s)\theta(\tau|\overline{v}_1-\overline{v}_3)+\theta(\tau|\overline{v}_1-\overline{v}_3-s)\theta(\tau|\overline{v}_2-\overline{v}_3-s)\theta(\tau|\overline{v}_1-\overline{v}_2) \right)\\
=\frac{\diamondsuit}{\theta(\tau|v_1-v_2)\theta(\tau|v_1-v_3)\theta(\tau|v_2-v_3)}\cdot\left(e^{2\pi is (b_2+b_3-2b_1)}\theta(\tau|v_2-v_1-s)\theta(\tau|v_3-v_1-s)\theta(\tau|v_2-v_3) \right.\\
\left. -e^{2\pi is (b_1+b_3-2b_2)}\theta(\tau|v_1-v_2-s)\theta(\tau|v_3-v_2-s)\theta(\tau|v_1-v_3) \right.\\
\left.
+e^{2\pi is (b_1+b_2-2b_3)}\theta(\tau|v_1-v_3-s)\theta(\tau|v_2-v_3-s)\theta(\tau|v_1-v_2) \right).\end{gathered}$$ Hence, notice that for general values of $s$ the above expression is different from the series $ \Ell^{\vir}_{1/2}(Q_1^3,\mathcal K_{\vir}^{1/2},t,w;p)$. However, if we specialise $v_1+v_2+v_3=0$ and $s\in \frac{1}{3}\BZ$, we see that in the previous example $ \Ell^{\vir}_{1/2}(Q_1^3,\mathcal K_{\vir}^{1/2},t,w;p)$ becomes constant with respect to $v$ on the lattice $\BZ+ \tau \BZ$ and is holomorphic in $v$, from which we conclude that it is constant on $v$ under this specialisation. Therefore, by choosing $w_j=e^{2\pi i \frac{j}{3}}$ to be third roots of unity, one can show $$\left.\Ell^{\vir}_{1/2}(Q_1^3,\mathcal K_{\vir}^{\frac{1}{2}},t,w;p)\right|_{\mathfrak{t}=e^{2\pi i \frac{k}{3}}}= \begin{cases} (-1)^{m+1}3, & \mbox{if } k=3m, \quad m\in \BZ \\ 0, & \mbox{if } k\notin 3\BZ. \end{cases}$$
Limits of elliptic DT invariants {#sec:limits of elliptic DT}
--------------------------------
Even if a closed formula for the higher rank generating series of elliptic DT invariants is not available, we can still study its behaviour by looking at some particular limits of the variables $p,t_i,w_j$.
It is easy to see that, under the Calabi–Yau restriction $\mathfrak{t}=1$, the generating series of elliptic DT invariants does not carry any more refined information than the cohomological one; in particular, we have no more dependence on the framing parameters $w_j$ and the elliptic parameter $p$. We generalise this phenomenon in the following setting. Denote by $\TT_k\subset\BT_1$ the subtorus where $\mathfrak t^{\frac{1}{2}}=e^{\pi ik/r}$, $k\in\BZ$. Define by $$\DT_{r,k}^{\rm ell}(\BA^3,q,t,w;p)=\left.\DT_{r}^{\rm ell}(\BA^3,q,t,w;p)\right|_{\TT_k}$$ the restriction of the generating series to the subtorus $\TT_k\subset\BT_1$, which is well-defined as no powers of the Calabi–Yau weight appear in the vertex terms (cf. also [@MNOP1 pag. 1279]).
\[corollary to elliptic macmahon prop\] If $k=rm\in r\BZ$, then $$\DT_{r, k}^{\rm ell}(\BA^3,q,t,w;p)=\mathsf M((-1)^{r(m+1)}q)^{r}.$$ In particular, the dependence on $t_i, w_j$ and $p$ drops.
Let $S\in \Quot_{\BA^3}(\OO^{\oplus r},n)^{\TT}$. Denote $T_S^{\vir}=T^{nc}_{S}-\overline{T^{nc}_{S}}\mathfrak{t}^{-1}$ as in Equation , where $T^{nc}_{S}$ is the tangent space of $\NCQuot_{r}^n$ at $S$. Denote by $T^{nc}_{S,l}$ the sub-representation of $ T^{nc}_{S} $ corresponding to $\mathfrak{t}^l$, with $l\in \mathbb{Z}$. As there are no powers of the Calabi–Yau weight in $T_S^{\vir}$, we have an identity $T^{nc}_{S,l}=\overline{T^{nc}_{S,-l-1}}\mathfrak{t}^{-1}$. Set $$W=T^{nc}_{S}-T^{nc}_{S,0}-T^{nc}_{S,-1}.$$ We have that $$T^{nc}_{S}-\overline{T^{nc}_{S}}\mathfrak{t}^{-1}=W-\overline{W}\mathfrak{t}^{-1}$$ and, in particular, neither $W$ nor $\overline{W}\mathfrak{t}^{-1}$ contain constant terms. Using the quasi-periodicity of the theta function $\theta(\tau|z)$, we have $$\begin{aligned}
\theta[-T_S^{\vir}]=\frac{\theta[\overline{W}\mathfrak{t}^{-1}]}{\theta[W]}=(-1)^{m\rk W}\frac{\theta[\overline{W}]}{\theta[W]}=(-1)^{\rk W(m+1)}.\end{aligned}$$ We conclude by noticing that $$\rk W=\rk T_S^{nc}=rn \, \mod{2}.\qedhere$$
Motivated by Example \[example dependence elliptic\] and Proposition \[corollary to elliptic macmahon prop\], we propose the following conjecture.
\[conj: non dependence under some limits of elliptic\] Under the restriction to the subtorus $\BT_{2,0}=\set{ \prod_{j=1}^rw_j=1}\subset \BT_2$, the series $\DT_{r,k}^{\rm ell}(\BA^3,q,t,w;p)$ does not depend on the elliptic parameter $p$.
\[rem: independence of p then equal to K theoretic\] Assuming $\prod_{j=1}^rw_j=1$, notice that the independence from the elliptic parameter $p$ implies that we can reduce our invariants to the K-theoretic ones by setting $p=0$, i.e. $$\DT_{r,k}^{\rm ell}(\BA^3,q,t,w;p)=\DT_{r}^{ \KK}\left.(\BA^3,q,t)\right|_{\TT_k}.$$
Assuming Conjecture \[conj: non dependence under some limits of elliptic\], we derive a closed expression for $\DT_{r,k}^{\rm ell}(\BA^3,q,t,w;p)$, which was conjectured in [@BBPT Equation (3.20)], motivated by string-theoretic phenomena.
\[thm:elliptic\] Assume Conjecture \[conj: non dependence under some limits of elliptic\] holds and set $\prod_{j=1}^rw_j=1$. Let $k\in \BZ$. Then there is an identity $$\label{eqn:conjecture elliptic}
\DT_{r,k}^{\rm ell}(\BA^3,q,t,w;p)=\mathsf M\left((-1)^{kr}((-1)^rq)^{\frac{r}{\gcd(k,r)}}\right)^{\gcd(k,r)}.$$
Assuming Conjecture \[conj: non dependence under some limits of elliptic\], by Remark \[rem: independence of p then equal to K theoretic\] we just have to prove the result for K-theoretic invariants. By Theorem \[thm: K theoretic of points higher rank\], $$\begin{gathered}
\DT_r^{\KK}(\BA^3,(-1)^r q,t)\\
=\exp{\sum_{n\geq 1}\frac{1}{n}\frac{(1-t_1^{-n}t_2^{-n})(1-t_1^{-n}t_3^{-n})(1-t_2^{-n}t_3^{-n})}{(1-t_1^{-n})(1-t_2^{-n})(1-t_3^{-n})} \frac{1-\mathfrak{t}^{-rn}}{1-\mathfrak{t}^{-n}}\frac{1}{(1-\mathfrak{t}^{-\frac{rn}{2}}q^{-n})(1-\mathfrak{t}^{-\frac{rn}{2}}q^{n})}}.\end{gathered}$$ Assume now that $\mathfrak{t}^{\frac{1}{2}}=e^{\pi i \frac{k}{r}}$, with $k\in \mathbb{Z}$; we have clearly that $\mathfrak{t}^{-\frac{rn}{2}}=(-1)^{kn} $. Moreover, we have $$\frac{1-\mathfrak{t}^{-rn}}{1-\mathfrak{t}^{-n}}= \begin{cases} r, & \mbox{if } n\in \frac{r}{\gcd(r,k)}\BZ \\ 0, & \mbox{if } n\notin \frac{r}{\gcd(r,k)}\BZ \end{cases}$$ In particular, if $n\in \frac{r}{\gcd(r,k)}\BZ$, we have $$\frac{(1-t_1^{-n}t_2^{-n})(1-t_1^{-n}t_3^{-n})(1-t_2^{-n}t_3^{-n})}{(1-t_1^{-n})(1-t_2^{-n})(1-t_3^{-n})} =-1$$ Setting $n=\frac{r}{\gcd(r,k)} m$, with $m\in \BZ$, we have $$\begin{aligned}
\DT_r^{\KK}(\BA^3,(-1)^r q,t)&=\exp{\sum_{m\geq 1}\frac{1}{m}\gcd(r,k)\cdot \frac{-1}{(1-\overline{q}^{-m})(1-\overline{q}^m)}}\end{aligned}$$ where to ease notation we have set $\overline{q}=((-1)^kq)^{\frac{r}{\gcd(r,k)}}$. We conclude by using the description of the MacMahon function as a plethystic exponential $$\DT_r^{\KK}(\BA^3,(-1)^r q,t)=\mathsf M\left((-1)^{kr}q^{\frac{r}{\gcd(r,k)}}\right)^{\gcd(r,k)}.\qedhere$$
\[rem: elliptic indepedence\] A key technical point in the proof of the conjecture proposed in [@BBPT Equation (3.20)] was the assumption of the independence of $\DT_{r,k}^{\rm ell}(\BA^3)$ on $p$, as in Conjecture \[conj: non dependence under some limits of elliptic\]. We strongly believe it should be possible to prove this assumption by exploiting modular properties of the generating series of elliptic DT invariants. One should proceed, in the same spirit of Appendix \[sec: functional repr. K teoria\], by considering the integral representation for the generating series, given in [@BBPT Equation (3.1)]. The analysis of the K–thoretic case, which we carried out in the proof of Theorem \[thm:elliptic\], shows that no dependence whatsoever is present in the limit $\mathfrak t^{1/2}=e^{\pi ik/r}$. As elliptic DT invariants take the form of meromorphic Jacobi forms, given by quotients of theta functions, poles in the equivariant parameters are only given by shifts along the lattice $\BZ+\tau\BZ$ of the poles found in K–theoretic DT invariants. Then $\DT_{r,k}^{\rm ell}(\BA^3)$, as a function of each of the equivariant parameters $v_i$, $i=1,\dots,r$, and $s_j$, $j=1,2,3$, is holomorphic on the torus $\BC/\BZ+\tau\BZ$, so it also carries no dependence on them. This observation may be not very surprising, if one considers the striking resemblance of the chiral virtual elliptic genus to the usual level$-N$ elliptic genus of almost complex manifolds, which is known to be often rigid. In our case, each $q$-term in the elliptic generating series restricted to $\TT_k$ would be now invariant under modular transformations on $\tau$, hence a constant in $p=e^{2\pi i\tau}$.
Relation to string theory {#subsec:Witten}
-------------------------
The definition for the elliptic version of Donaldson–Thomas invariants is motivated by an argument due to Witten [@Dirac_index], which goes as follows: let $M$ be a $2k$-dimensional spin manifold, and take $\mathscr LM=C^0(S^1,M)$ to be the free loop space on $M$. Then $\mathscr LM$ always carries a natural $S^1$-action, given by the rotation of loops, so that fixed points under this action of $S^1$ will only be constant maps $S^1\to M$, and $(\mathscr LM)^{S^1}\cong M\hookrightarrow\mathscr LM$. One can then study the Dirac operator on $\mathscr LM$ by formally computing its index using fixed point formulæ. In particular, if $D\colon \Gamma(S_+)\to\Gamma(S_-)$ is the Dirac operator on $M$ then $$\Ind(D)=\dim\ker D-\dim\coker D,$$ and whenever $M$ admits the action of a compact Lie group $G$ one can define the $G$-equivariant index of $D$ as the virtual character $$\Ind_G(D)(g)=\Tr_{\ker D}g-\Tr_{\coker D}g,\qquad g\in G,$$ which only depends on the conjugacy class of $g$ in $G$. In the case of Dirac operators on loop spaces over spin manifolds, a formal computation yields $$\begin{aligned}
\Ind_{S^1}(D)(q)&=q^{-\frac{d}{24}}\hat A(M)\cdot\ch\left(\bigotimes_{n\ge 1}\Sym^\bullet_{q^n}T_M\right)\cap[M],\end{aligned}$$ where $q$ denotes a topological generator of $S^1$. The $\hat A$-genus is the characteristic class which computes the index of the Dirac complex on a spin manifold. In general, if $E$ is any rank $r$ complex vector bundle on $M$, one can define $\hat A(E)$ in terms of the Chern roots $x_i$ of $E$ as $$\hat A(E)=\prod_{i=1}^r\frac{x_i/2}{\sinh x_i/2},$$ and $\hat A(M)=\hat A(TM)$, which is completely analogous to the K–theoretic invariants we have been studying so far. The previous formula can also be classically interpreted as the index of a twisted Dirac operator over the spin structure of $M$. It is worth noticing that the DT partition functions coming from physics are indeed interpreted as being computing indices of twisted Dirac operators, where the twist by a vector bundle $V\to M$ makes sense only if $w_2(T_M)=w_2(V)$ so as to extend $D$ to an operator $D\colon \Gamma(S_+\otimes V)\to\Gamma(S_-\otimes V)$. In this same spirit one might also justify the definition of the half-BPS elliptic genus in terms of computations of Euler characteristics of loop spaces over (compact) almost complex manifolds. Let then $X$ be a $d$-dimensional almost complex manifold, with holomorphic tangent bundle $T_X$, and whose corresponding free loop space will be denoted by $\mathscr LX$. As it was the case also in the previous situation, $\mathscr LX$ is naturally equipped with an $S^1$ action, whose fixed point will be $(\mathscr LX)^{S^1}\cong X\hookrightarrow\mathscr LX$. By formally applying the virtual localisation formula to the computation of the Euler characteristic of $\mathscr LX$ one gets $$\chi_{S^1}(\mathscr LX)=q^{-\frac{d}{12}}\td(X)\cdot\ch\left(\bigotimes_{n\ge 1}\Sym_{q^n}^\bullet (T_X\oplus\Omega_X)\right)\cap[X],$$ which can also be seen as the index of a twisted ${\rm Spin}^c$-Dirac operator $\overline\partial+\overline\partial^\ast$. Moreover, if $c_1(T_X)=0\mod 2$, $X$ is also spin, and it is possible to compute the index of the Dirac operator on $\mathscr LX$ as before.
Higher rank DT invariants of compact toric 3-folds {#sec:compact_section}
==================================================
Let $X$ be a smooth *projective* toric $3$-fold, along with an exceptional locally free sheaf $F$ of rank $r$. By [@Quot19 Thm. A], the Quot scheme $\Quot_X(F,n)$ has a $0$-dimensional perfect obstruction theory, so that the rank $r$ Donaldson–Thomas invariant $$\DT_{F,n} = \int_{[\Quot_X(F,n)]^{\vir}}1\,\in\,\BZ$$ is well defined. In this section we confirm the formula $$\sum_{n\geq 0}\DT_{F,n} q^n = \mathsf M((-1)^rq)^{r\int_X c_3(T_X \otimes K_X)},$$ suggested in [@Quot19 Conj. 3.5], in the case where $F$ is *equivariant*. This will prove Theorem \[mainthm:projective\_toric\] from the Introduction. The next subsection is an interlude on how to induce a torus action on the Quot scheme and on the associated universal short exact sequence starting from an equivariant structure on $F$. More details are given in [@Equivariant_Atiyah_Class], including a proof that the obstruction theory obtained in [@Quot19 Thm. A] is equivariant (but see Proposition \[pot\_global\_equivariant\] for a sketch).
Inducing a torus action on the Quot scheme {#sec:induced_T-action_quot}
------------------------------------------
Let $X$ be a quasi-projective toric variety with torus $\BT\subset X$. Let $\sigma_X\colon \BT \times X \to X$ denote the action. If $F$ is a $\BT$-equivariant coherent sheaf on $X$, and $\mathrm{Q} = \Quot_X(F,n)$, then $\sigma_X$ has a canonical lift $$\sigma_{\mathrm{Q}}\colon \BT \times \mathrm{Q}\to \mathrm{Q}.$$ This is proved in [@Kool_Fixed_Point_Loci Prop. 4.1], but we sketch here the argument for the sake of completeness. We thank Martijn Kool for guiding us through the details of this construction. Let $p_2\colon \BT \times X \to X$ be the second projection and let $\vartheta\colon p_2^\ast F \simto \sigma_X^\ast F$ denote the chosen $\BT$-equivariant structure on $F$. Let $\pi_X\colon X\times \mathrm{Q} \to \mathrm{Q}$ and $\pi_{\mathrm{Q}}\colon X\times \mathrm{Q}\to \mathrm{Q}$ be the projections, and set $F_Q = \pi_X^\ast F$. Consider the universal exact sequence $$\label{diag_uni}
\begin{tikzcd}
& \mathcal S \into F_{\mathrm{Q}} \overset{u}{\onto} \mathcal T\arrow[dash]{d} & \\
& X\times \mathrm{Q}\arrow[swap]{dl}{\pi_X}\arrow{dr}{\pi_{\mathrm{Q}}} & \\
X & & \mathrm{Q}
\end{tikzcd}$$ and note that there is a commutative diagram $$\begin{tikzcd}[column sep = large,row sep = large]
\BT\times X\times \mathrm{Q}\arrow[swap]{d}{p_{12}}\arrow{r}{\sigma_X\times\id_{\mathrm{Q}}} & X\times \mathrm{Q}\arrow{d}{\pi_X} \\
\BT\times X\arrow{r}{\sigma_X} & X
\end{tikzcd}$$ yielding an identity $(\sigma_X\times \id_{\mathrm{Q}})^\ast F_{\mathrm{Q}} = p_{12}^{\ast}\sigma_X^\ast F$. The induced surjection $$(\sigma_X\times \id_{\mathrm{Q}})^\ast u \circ p_{12}^\ast\vartheta\colon F_{\BT\times \mathrm{Q}} \simto p_{12}^\ast\sigma_X^\ast F \onto (\sigma_X\times \id_{\mathrm{Q}})^\ast \mathcal T$$ defines a $\BT\times \mathrm{Q}$-valued point of $\mathrm{Q}$, i.e. a morphism $$\sigma_{\mathrm{Q}}\colon \BT \times \mathrm{Q} \to \mathrm{Q}.$$ It is straightforward to verify that $\sigma_{\mathrm{Q}}$ satisfies the axioms for a $\BT$-action.
Next, we explain how to make the universal exact sequence $\BT$-equivariant. The universal property of the Quot scheme applied to $\sigma_{\mathrm{Q}}$ implies that there is an isomorphism of surjections $$(\id_X \times \sigma_{\mathrm{Q}})^\ast u \,\simto\,(\sigma_X\times \id_{\mathrm{Q}})^\ast u \circ p_{12}^\ast\vartheta.$$ This means that there is a commutative diagram $$\label{eqn:univ_property_quot}
\begin{tikzcd}
(\id_X \times \sigma_{\mathrm{Q}})^\ast F_{\mathrm{Q}}\arrow[equal]{d}\arrow[two heads]{rrr}{(\id_X \times \sigma_{\mathrm{Q}})^\ast u} & & & (\id_X \times \sigma_{\mathrm{Q}})^\ast\mathcal T \isoarrow{d} \\
p_{12}^\ast p_2^\ast F\arrow{r}{p_{12}^\ast\vartheta} & p_{12}^\ast \sigma_X^\ast F\arrow[two heads]{rr}{(\sigma_X\times \id_{\mathrm{Q}})^\ast u} & & (\sigma_X\times \id_{\mathrm{Q}})^\ast \mathcal T
\end{tikzcd}$$ where we used the identity $(\sigma_X\times \id_{\mathrm{Q}})^\ast F_{\mathrm{Q}} = p_{12}^{\ast}\sigma_X^\ast F$ in the bottom row.
Consider the morphism $$\varphi \colon X\times \mathrm{Q} \times \BT \to X\times \mathrm{Q}, \quad (x,f,t) \mapsto (\sigma_X(t, x),\sigma_{\mathrm{Q}}(t^{-1},f)).$$ We view this as a $\BT$-action on $X\times \mathrm{Q}$. Note that $\pi_X\circ \varphi = \sigma_X\circ p_{12}$. The moduli map $\mathrm{Q} \times \BT \to \mathrm{Q}$ corresponding to the family of quotients $$\begin{tikzcd}
\varphi^*u\circ p_{12}^\ast\vartheta\colon F_{\mathrm{Q} \times \BT} = p_{12}^\ast p_2^\ast F \simto p_{12}^\ast \sigma_X^\ast F = \varphi^\ast F_{\mathrm{Q}} \arrow[two heads]{r}{\varphi^*u} & \varphi^\ast \mathcal T
\end{tikzcd}$$ is easily seen to agree with the first projection $p_1\colon \mathrm{Q} \times \BT \to \mathrm{Q}$. Indeed, if $\mathcal T_f$ denotes the quotient of $F$ corresponding to a point $f = [F \onto \mathcal T_f] \in \mathrm{Q}$, it is immediate to see that $$\varphi^\ast \mathcal T\big|_{X\times f\times t} = \mathcal T_{\sigma_{\mathrm{Q}}(\sigma_{\mathrm{Q}}(t,f),t^{-1})} = \mathcal T_{\sigma_{\mathrm{Q}}(tt^{-1},f)} = \mathcal T_f$$ for all $t \in \BT$.
Let then $q = \id_X \times p_1\colon X\times \mathrm{Q} \times \BT \to X\times \mathrm{Q}$ be the projection. Since $\varphi^*u\circ p_{12}^\ast\vartheta$ corresponds to the projection $p_1\colon \mathrm{Q} \times \BT \to \mathrm{Q}$, by the universal property of $(\mathrm{Q},u)$ we obtain an isomorphism of surjections $q^\ast u \simto \varphi^\ast u$ that, after setting $\mathcal S = \ker (u\colon F_{\mathrm{Q}}\onto \mathcal T)$, we can extend to an isomorphism of short exact sequences $$\begin{tikzcd}
q^\ast \mathcal S \arrow[hook]{rr}\isoarrow{d} & &
F_{\mathrm{Q} \times \BT} \arrow[two heads]{rr}{q^\ast u}\isoarrow{d} & & q^\ast \mathcal T\isoarrow{d} \\
\varphi^\ast \mathcal S \arrow[hook]{rr} & & \varphi^\ast F_{\mathrm Q}\arrow[two heads]{rr}{\varphi^*u} & & \varphi^\ast \mathcal T
\end{tikzcd}$$ on $X\times \mathrm{Q}\times \mathbb T$, where the middle vertical isomorphism is $p_{12}^\ast\vartheta$ and is a $\BT$-equivariant structure on $F_{\mathrm{Q}}$ because $\vartheta$ is. The diagram allows us to conclude that a $\BT$-equivariant structure on $F$ induces a canonical $\BT$-equivariant structure on the universal short exact sequence $$0 \to \mathcal S \to F_{\mathrm{Q}} \to \mathcal T \to 0.$$
The (equivariant) obstruction theory
------------------------------------
Throughout this subsection, $F$ denotes an exceptional locally free sheaf of rank $r$ on a smooth projective toric $3$-fold $X$. In other words, $F$ is simple, i.e. $\End_{\OO_X}(F) = \BC$, and $\Ext^i(F,F)=0$ for $i>0$.
By [@Quot19 Thm. A], there is a $0$-dimensional perfect obstruction theory $$\label{global_pot}
\BE \to \BL_{\Quot_X(F,n)},$$ governed by $$\Def\big|_{[S]} = \Ext^1(S,S),\quad
\Obs\big|_{[S]} =\Ext^2(S,S)$$ around a point $[S] \in \Quot_X(F,n)$. We set $\mathrm{Q} = \Quot_X(F,n)$ for brevity, and we denote by $\pi_{\mathrm{Q}}$ and $\pi_X$ the projections from $X\times \mathrm Q$, as in . Note that $\omega_{\pi_{\mathrm{Q}}} = \pi_X^\ast \omega_X$.
As we recall during the (sketch of) proof of Proposition \[pot\_global\_equivariant\] below, we have $$\BE = \RR\pi_{\mathrm Q\,\ast}(\RRlHom(\mathcal S,\mathcal S)_0\otimes \omega_{\pi_{\mathrm Q}})[2]$$ where $\RRlHom(\mathcal S,\mathcal S)_0$ is the shifted cone of the trace map $\tr\colon\RRlHom(\mathcal S,\mathcal S) \to \OO_{X\times \mathrm Q}$.
\[prop:global pot K theory\] There is an identity $$\BE^\vee=\mathbf{R}\pi_{\mathrm Q\,\ast}\RRlHom(F_{\mathrm Q},F_{\mathrm Q})-\mathbf{R}\pi_{\mathrm Q\,\ast}\RRlHom\mathcal (\mathcal S,\mathcal S)\in K_0(\Quot_X(F,n)).$$
As in the proof of [@Quot19 Theorem 2.5] we have $$\begin{aligned}
\BE\,&=\, (\RR\pi_{\mathrm Q\,\ast}\RRlHom(\mathcal S,\mathcal S)_0)^\vee[-1]\\
\,&=\, (\RR\pi_{\mathrm Q\,\ast}\RRlHom(\mathcal S,\mathcal S))^\vee[-1]- (\RR\pi_{\mathrm Q\,\ast}\RRlHom(\OO,\OO))^\vee[-1]\\
\,&=\,( \RR\pi_{\mathrm Q\,\ast}\RRlHom(\mathcal S,\mathcal S))^\vee[-1]- (\RR\pi_{\mathrm Q\,\ast}\RRlHom(F_{\mathrm Q},F_{\mathrm Q}))^\vee[-1],\end{aligned}$$ where the last identity uses that $F$ is an exceptional sheaf.
The following result is proved in [@Equivariant_Atiyah_Class Thm. B] in greater generality, but we sketch a proof here for the reader’s convenience. Denote by $\BT = (\BC^\ast)^3$ the torus of $X$.
\[pot\_global\_equivariant\] Let $(X,F)$ be a pair consisting of a smooth projective toric $3$-fold $X$ along with an exceptional locally free $\BT$-equivariant sheaf $F$. Then the perfect obstruction theory on $\Quot_X(F,n)$ is $\BT$-equivariant.
The definition of equivariant obstruction theory was recalled in Definition \[def:equivariant\_pot\].
As we explained in §\[sec:induced\_T-action\_quot\], by equivariance of $F$ the Quot scheme $\mathrm Q = \Quot_X(F,n)$ inherits a canonical $\BT$-action with respect to which the universal short exact sequence $\mathcal S \into F_{\mathrm{Q}} \onto \mathcal T$ can be made $\BT$-equivariant. On the other hand, the perfect obstruction theory is obtained by projecting the truncated Atiyah class $$\At_{\mathcal S} \,\in\, \Ext^1_{X \times \mathrm Q}(\mathcal S,\mathcal S\otimes \BL_{X \times \mathrm Q})$$ onto the Ext group $$\label{Atiyah_Journey}
\begin{split}
\Ext^1_{X \times \mathrm Q}(\RRlHom(\mathcal S,\mathcal S)_0,\pi_{\mathrm Q}^\ast \BL_{\mathrm Q}) \,\,&=\,\, \Ext^{-2}_{X\times \mathrm Q}(\RRlHom(\mathcal S,\mathcal S)_0 \otimes \omega_{\pi_{\mathrm Q}},\pi_{\mathrm Q}^\ast \BL_{\mathrm Q} \otimes \omega_{\pi_{\mathrm Q}}[3]) \\
\,\,&=\,\, \Ext^{-2}_{X\times \mathrm Q}(\RRlHom(\mathcal S,\mathcal S)_0 \otimes \omega_{\pi_{\mathrm Q}},\pi_{\mathrm Q}^! \BL_{\mathrm Q}) \\
\,\,&\cong \,\,\Ext^{-2}_{\mathrm Q}(\RR\pi_{\mathrm Q\,\ast}(\RRlHom(\mathcal S,\mathcal S)_0 \otimes \omega_{\pi_{\mathrm Q}}),\BL_{\mathrm Q}).
\end{split}$$ The last isomorphism is Grothendieck duality along the smooth projective morphism $\pi_{\mathrm Q}$. Now we need three ingredients to finish the proof:
- The Atiyah class $\At_{\mathcal S}$ is a $\BT$-invariant extension,
- Grothendieck duality preserves $\BT$-invariant extensions, and
- $\BT$-invariant extensions correspond to morphisms in the equivariant derived category.
These assertions are proved in [@Equivariant_Atiyah_Class].
We let $\Delta(X)$ denote the set of vertices in the Newton polytope of the toric $3$-fold $X$. Then $$X^{\BT} = \Set{p_\alpha | \alpha \in \Delta(X)} \subset X$$ will denote the fixed locus of $X$. For a given vertex $\alpha$, let $U_\alpha \cong \BA^3$ be the canonical chart containing the fixed point $p_\alpha$. The $\BT$-action on this chart can be taken to be the standard action . For every $\alpha$, there is a $\BT$-equivariant open immersion $$\iota_{n,\alpha}\colon \Quot_{U_\alpha}(F|_{U_\alpha},n)\into \mathrm Q = \Quot_X(F,n)$$ parametrising quotients whose support is contained in $U_\alpha$. We think of $F|_{U_\alpha}$ as an equivariant sheaf on $\BA^3$, hence of the form described in . We denote by $\BE_{n,\alpha}^{\crit}$ the critical obstruction theory on $\Quot_{U_\alpha}(F|_{U_\alpha},n)$ from Proposition \[prop:SPOT\_A\^3\].
It is natural to ask whether the restriction of the global perfect obstruction theory along $\iota_{n,\alpha}$ agrees with the critical symmetric perfect obstruction theory described in §\[sec:critical\_structure\_Quot\] (see Conjecture \[conj:pot\_restricted\]). However, what we really need is the following weaker result.
\[restriction of class in K theory\] Let $\BE \in K^0(\mathrm{Q})$ be the class of the global perfect obstruction theory . Then $$\BE_{n,\alpha}^{\crit}\,=\,\iota_{n,\alpha}^\ast \BE\,\in\,K^0(\Quot_{U_\alpha}(F|_{U_\alpha},n)).$$ Considering the two obstruction theories as $\BT$-equivariant, the same identity holds in equivariant K-theory: $$\BE_{n,\alpha}^{\crit}\,=\,\iota_{n,\alpha}^\ast \BE\,\in\,K^0_{\BT}(\Quot_{U_\alpha}(F|_{U_\alpha},n)).$$
The chart $U_\alpha$ is Calabi–Yau, so by [@Quot19 Prop. 2.9] the induced perfect obstruction theory $\iota_{n,\alpha}^\ast \BE$ is symmetric. Since by Remark \[rmk:K-class\_of\_symmetric\_pot\] all symmetric perfect obstruction theories share the same class in K-theory, the first statement follows.
To prove the K-theoretic equality, we need a slightly more refined analysis. Just for this proof, let us shorten $$\BE_{\crr} = \BE_{n,\alpha}^{\crit}\quad\textrm{and}\quad \BE = \iota_{n,\alpha}^\ast \BE,$$ to ease notation. We know by Diagram that we can write $$\label{fucking_E_cr}
\begin{split}
\BE_{\crr} &= \big[\mathfrak t\otimes T_{\NCQuot_{r}^n}\big|_{\mathrm{Q}} \to \Omega_{\NCQuot_{r}^n}\big|_{\mathrm{Q}}\big]\\
&=\Omega - \mathfrak t\otimes T\,\in\,K^0_{\BT}(\Quot_{U_\alpha}(F|_{U_\alpha},n)),
\end{split}$$ where $\Omega$ (resp. $T$) denotes the cotangent sheaf (resp. the tangent sheaf) of $\Quot_{U_\alpha}(F|_{U_\alpha},n)$, equipped with its natural equivariant structure.
Let $\pi\colon U_\alpha \times \Quot_{U_\alpha}(F|_{U_\alpha},n) \to \Quot_{U_\alpha}(F|_{U_\alpha},n)$ be the projection, let $S$ be the universal kernel living on the product and set $\mathfrak{t}_{\pi} = \pi^\ast \mathfrak{t}^{-1}$. By definition, $$\BE = \RR \pi_\ast (\RRlHom(S,S)_0\otimes \omega_\pi)[2].$$ The equivariant isomorphism $\omega_\pi \simto \OO \otimes \mathfrak{t}_{\pi}^{-1}$ along with the projection formula yield $$\label{eqn:equivariant_POT1837}
\mathfrak{t}^{-1} \otimes \BE \simto \RR \pi_\ast \RRlHom(S,S)_0[2].$$ We next show the right hand side is canonically isomorphic to $\BE^\vee[1]$. We have $$\begin{aligned}
\BE^\vee[1] &= \RRlHom(\RR \pi_\ast (\RRlHom(S,S)_0 \otimes \omega_\pi),\OO)[-1] & \small{\textrm{definition of }(-)^\vee}\\
&= \RR\pi_\ast \RRlHom(\RRlHom(S,S)_0 \otimes \omega_\pi,\omega_{\pi}[3])[-1] & \small{\textrm{Grothendieck duality}} \\
&= \RR\pi_\ast \RRlHom(\RRlHom(S,S)_0,\OO)[2] & \small{\textrm{shift}}\\
&= \RR\pi_\ast\RRlHom(S,S)_0^\vee [2] & \small{\textrm{definition of }(-)^\vee} \\
&=\RR\pi_\ast\RRlHom(S,S)_0 [2] & \small{\RRlHom(S,S)_0\textrm{ is self-dual}}\end{aligned}$$ in the derived category of $\BT$-equivariant coherent sheaves on $\Quot_{U_\alpha}(F|_{U_\alpha},n)$, which by proves that $$\mathfrak{t}^{-1} \otimes \BE \cong \BE^\vee[1].$$ We thus have $$\Omega \cong h^0(\BE) \cong \mathfrak t \otimes h^0(\BE^\vee[1]) \cong \mathfrak t \otimes \lExt^2_\pi(S,S),$$ where we use the standard notation $\lExt^i_\pi(-,-)$ for the $i$th derived functor of $\pi_\ast\circ \lHom(-,-)$. We conclude $$\begin{aligned}
\BE
&= h^0(\BE) - h^{-1}(\BE) \\
&= \Omega - h^1(\BE^\vee)^\vee \\
&= \Omega - h^0(\BE^\vee[1])^\vee \\
&= \Omega - \lExt^2_\pi(S,S)^\vee \\
&= \Omega - (\mathfrak{t}^{-1}\otimes \Omega)^\vee \\
&= \Omega - \mathfrak t \otimes T \\
&= \BE_{\crr}.\qedhere\end{aligned}$$
The fixed locus of the Quot scheme and its virtual class
--------------------------------------------------------
In this subsection we describe $\Quot_X(F,n)^{\BT}$ and we compute its virtual fundamental class, obtained via Proposition \[pot\_global\_equivariant\].
If $\mathbf n$ denotes a generic tuple $\set{n_\alpha | \alpha \in \Delta(X)}$ of non-negative integers, we set $|\mathbf n| = \sum_{\alpha \in \Delta(X)}n_\alpha$.
\[T\_1 fixed locus projective\] There is a scheme-theoretic identity $$\Quot_X(F,n)^{\BT} = \coprod_{|\mathbf n| = n}\prod_{\alpha \in \Delta(X)} \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}.$$
Let $B$ be a (connected) scheme over $\BC$. Let $F_B$ be the pullback of $F$ along the first projection $X\times B \to X$, and fix a $B$-flat family of quotients $$\rho\colon F_B \onto \mathcal T$$ defining a $B$-valued point $B \to \Quot_X(F,n)^{\BT}$. Then, by restriction, we obtain, for each $\alpha \in\Delta(X)$, a $B$-flat family of quotients $$\label{F_restricted_to_Ualpha}
\rho_\alpha \colon F_B\big|_{U_\alpha\times B} \onto \mathcal T_\alpha = \mathcal T\big|_{U_\alpha\times B},$$ and we let $n_\alpha$ be the length of the fibres of $\mathcal T_\alpha$. Each $\rho_\alpha$ corresponds to a $B$-valued point $g_\alpha\colon B \to \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}$, thus we obtain a $B$-valued point $$(g_\alpha)_\alpha \colon B \to \prod_{\alpha \in \Delta(X)} \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}.$$ Note that the original family $\mathcal T$ is recovered as the direct sum $\oplus_\alpha \mathcal T_\alpha$, in particular $n = \sum_\alpha n_\alpha$. Conversely, suppose given a tuple of $B$-families of $\BT$-fixed quotients $$\left(\left(F|_{U_{\alpha}}\right)_B
\onto \mathcal T_\alpha\right)_\alpha.$$ We obtain $B$-valued points $$B \to \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT} \subset \Quot_X(F,n_\alpha)^{\BT}.$$ Since the support of these families is disjoint, we can form the direct sum $$\mathcal T = \bigoplus_\alpha \mathcal T_\alpha$$ to obtain a new $B$-flat family, representing a $B$-valued point of $\Quot_X(F,n)^{\BT}$, as required.
Our next goal is to show that, under the identification of Lemma \[T\_1 fixed locus projective\], the induced virtual fundamental class of the $\mathbf n$-th connected component of $\Quot_X(F,n)^{\BT}$ is the box product of the virtual fundamental classes of $\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}$, whose perfect obstruction theory is the $\BT$-fixed part of the critical one, studied in §\[sec:critical\_structure\_Quot\]. For the rest of the section we restrict our attention to each connected component $$\label{fixed_n}
\Quot_X(F,n)^{\BT}_{\mathbf n} = \prod_{\alpha \in \Delta(X)} \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT} \subset \Quot_X(F,n)^{\BT},$$ and we denote by $$\label{universal_structures}
\begin{tikzcd}
\mathcal S \into \mathcal F \onto \mathcal{T}\arrow[dash]{d} & \mathcal S_\alpha \into \mathcal F_\alpha \onto \mathcal{T}_\alpha \arrow[dash]{d}\\
X\times \Quot_X(F,n)^{\BT}_{\mathbf n}\arrow{d}{\pi} & U_\alpha \times \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}\arrow{d}{\pi_\alpha} \\
\Quot_X(F,n)^{\BT}_{\mathbf n} \arrow{r}{p_\alpha} & \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}
\end{tikzcd}$$ the various universal structures and projection maps between these moduli spaces. For instance, $\mathcal F_\alpha$ is the pullback of $F|_{U_\alpha}$ along the projection $U_\alpha \times \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}\to U_\alpha$. Let $\BE_{\mathbf n}$ be the restriction of $\BE \in \derived(\Quot_X(F,n))$ to the closed subscheme $\Quot_X(F,n)^{\BT}_{\mathbf n} \subset \Quot_X(F,n)$.
\[prop: relative cech\] There is an identity in $K^0_{\BT}(\Quot_X(F,n)_{\mathbf n}^{\BT})$ $$\begin{gathered}
\BE_{\mathbf n}^\vee = \mathbf{R}\pi_*\RRlHom(\mathcal F,\mathcal F)-\mathbf{R}\pi_*\RRlHom(\mathcal{S},\mathcal{S}) \\=\sum_{\alpha\in \Delta(X)}p_\alpha^*\left(\mathbf{R}{\pi_{\alpha}}_*\RRlHom(\mathcal F_\alpha,\mathcal F_\alpha)-\mathbf{R}{\pi_\alpha}_*\RRlHom(\mathcal{S}_\alpha,\mathcal{S}_\alpha)\right).\end{gathered}$$
Exploiting the universal short exact sequence $$0\to \mathcal{S}\to \mathcal F\to \mathcal{T}\to 0$$ on $X\times \Quot_X(F,n)^{\BT}_{\mathbf n} \subset X\times \Quot_X(F,n)$, and Proposition \[prop:global pot K theory\], we may write $$\begin{gathered}
\BE_{\mathbf n}^\vee = \BE^\vee\big|_{\Quot_X(F,n)^{\BT}_{\mathbf n}} = \mathbf{R}\pi_*\RRlHom(\mathcal F,\mathcal F)-\mathbf{R}\pi_*\RRlHom(\mathcal{S},\mathcal{S}) \\
= \mathbf{R}\pi_*\RRlHom(\mathcal{S},\mathcal{T})+\mathbf{R}\pi_*\RRlHom(\mathcal{T},\mathcal{S})+\mathbf{R}\pi_*\RRlHom(\mathcal{T},\mathcal{T}).
\end{gathered}$$ Similarly, we have $$\begin{gathered}
\label{eqn:rhom_alpha}
\mathbf{R}\pi_{\alpha*}\RRlHom(\mathcal F_{\alpha},\mathcal F_{\alpha})-\mathbf{R}\pi_{\alpha*}\RRlHom(\mathcal{S}_{\alpha},\mathcal{S}_{\alpha}) \\
= \mathbf{R}\pi_{\alpha*}\RRlHom(\mathcal{S}_{\alpha},\mathcal{T}_{\alpha})+\mathbf{R}\pi_{\alpha*}\RRlHom(\mathcal{T}_{\alpha},\mathcal{S}_{\alpha})+\mathbf{R}\pi_{\alpha*}\RRlHom(\mathcal{T}_{\alpha},\mathcal{T}_{\alpha}).
\end{gathered}$$ In the following, we write $(G_1,G_2)$ for any of the three pairs $(\mathcal{S},\mathcal{T})$, $(\mathcal{T},\mathcal{S})$ or $(\mathcal{T},\mathcal{T})$. Applying the Grothendieck spectral sequence yields $$\begin{aligned}
\mathbf{R}\pi_*\RRlHom(G_1,G_2)&= \sum_{i,j}(-1)^{i+j}\mathbf{R}^i\pi_* \lExt^j(G_1,G_2)\\
&=\sum_{j}(-1)^{j}\pi_* \lExt^j(G_1,G_2),\end{aligned}$$ where we used cohomology and base change along with the fact that $\RR^{i} \pi_\ast$ of a $0$-dimensional sheaf vanishes for $i>0$. The standard Čech cover $\{U_\alpha\}_{\alpha\in \Delta(X)}$ of $X$ pulls back to a Čech cover $\{V_\alpha\}_{\alpha\in \Delta(X)}$ of $X\times \Quot_X(F,n)^{\BT}_{\mathbf n}$, where $V_\alpha=U_\alpha\times \Quot_X(F,n)^{\BT}_{\mathbf n} $. For a finite family of indices $I\subset \BN$, set $V_I=\bigcap_{\alpha\in I} V_\alpha$ and let $j_I\colon V_I\to X\times \Quot_X(F,n)^{\BT}_{\mathbf n}$ be the natural open immersion. We have a Čech resolution $\lExt^j(G_1,G_2)\to \mathfrak{C}^\bullet $, where $\mathfrak{C}^\bullet$ is defined degree-wise (see e.g. [@Hartshorne_AG Lemma III.4.2]) by $$\mathfrak{C}^k=\bigoplus_{|I|=k+1}{j_I}_\ast j_I^\ast \lExt^j(G_1,G_2).$$ Notice that $\mathcal{T}$ vanishes on the restriction to any double intersection $U_{\alpha\beta}\times \Quot_X(F,n)^{\BT}_{\mathbf n}$, where $U_{\alpha\beta}=U_\alpha\cap U_\beta$. This implies that the only contribution of the Čech cover is given by $\mathfrak{C}^0$, thus $$\begin{aligned}
\mathbf{R}\pi_*\RRlHom(G_1,G_2)
&= \sum_{j} (-1)^{j}\pi_* \sum_{\alpha\in \Delta(X)}{j_\alpha}_\ast j_\alpha^\ast\lExt^j(G_1,G_2)\\
&= \sum_{\alpha\in \Delta(X)}
\sum_{j}(-1)^{j}(\pi\circ {j_\alpha})_\ast j_\alpha^\ast\lExt^j(G_1,G_2).\end{aligned}$$ Consider the following cartesian diagram $$\begin{tikzcd}
U_\alpha\times \Quot_X(F,n)^{\BT}_{\mathbf n}\arrow[hook]{d}{j_\alpha} \arrow{r}{\widetilde p_\alpha} & U_\alpha \times \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}\arrow[hook]{d}{} \arrow[dd,bend left=75, "\pi_\alpha "]\\
X\times \Quot_X(F,n)^{\BT}_{\mathbf n}\arrow{d}{\pi} \arrow{r}{} & X\times \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT} \arrow{d}\\
\Quot_X(F,n)^{\BT}_{\mathbf n} \arrow{r}{p_\alpha}& \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}
\end{tikzcd}$$ As it was already clear from the proof of Lemma \[T\_1 fixed locus projective\], the universal short exact sequences in Diagram satisfy $j_\alpha^\ast(\mathcal S \into \mathcal F \onto \mathcal T) = \widetilde p_\alpha^\ast(\mathcal S_\alpha \into \mathcal F_\alpha \onto \mathcal T_\alpha)$. If $(G_{1 \alpha},G_{2 \alpha})$ denotes any of the pairs belonging to the set $\set{(\mathcal{S}_{\alpha},\mathcal{T}_{\alpha}), (\mathcal{T}_{\alpha},\mathcal{S}_{\alpha}),(\mathcal{T}_{\alpha},\mathcal{T}_{\alpha})}$, we can write $$\begin{aligned}
j_\alpha^\ast\lExt^j(G_1,G_2)
&= \mathbf{L} j_\alpha^\ast\lExt^j(G_1,G_2) \\
&= \lExt^j(\mathbf{L} j_\alpha^\ast G_1,\mathbf{L} j_\alpha^\ast G_2) \\
&=\lExt^j(\widetilde{p}_\alpha^\ast G_{1,\alpha},\widetilde{p}_\alpha^\ast G_{2,\alpha}) \\
&=\widetilde p_\alpha^*\lExt^j({G_1}_\alpha,{G_2}_\alpha).\end{aligned}$$ We deduce, by flat base change, $$(\pi\circ {j_\alpha})_\ast j_\alpha^\ast\lExt^j(G_1,G_2)=(\pi\circ {j_\alpha})_\ast\widetilde p_\alpha^*\lExt^j({G_1}_\alpha,{G_2}_\alpha) = p_\alpha^{\ast}\pi_{\alpha\,\ast} \lExt^j({G_1}_\alpha,{G_2}_\alpha).$$ Combining again the Grothendieck spectral sequence, cohomology and base change and the vanishing of higher derived pushforwards on $0$-dimensional sheaves, we conclude that $$\begin{aligned}
\mathbf{R}\pi_*\RRlHom(G_1,G_2)&\,=\,
\sum_{\alpha\in \Delta(X)}
p_\alpha^{\ast}\sum_{j}(-1)^{j}\pi_{\alpha\,\ast} \lExt^j({G_1}_\alpha,{G_2}_\alpha)\\
&\,=\,\sum_{\alpha\in \Delta(X)}p_\alpha^* \mathbf{R}{\pi_\alpha}_* \mathbf{R}\lHom({G_1}_\alpha,{G_2}_\alpha).\end{aligned}$$ Now the result follows from Equation .
\[cor: pot as box product\] The virtual fundamental class of $\Quot_X(F,n)^{\BT}_{\mathbf n}$ is expressed as the product of the virtual fundamental classes $$\left[\Quot_X(F,n)^{\BT}_{\mathbf n}\right]^{\vir} = \prod_{\alpha \in \Delta(X)} p_\alpha^\ast \left[ \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}\right]^{\vir}.$$
Before we prove the corollary, let us explain what virtual classes are involved. The left hand side is the virtual class induced by the $\BT$-fixed obstruction theory $$\BE_{\mathbf n}^{\BT\textrm{-}\fix} \to \BL_{\Quot_X(F,n)^{\BT}_{\mathbf n}},$$ whereas $\left[ \Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}\right]^{\vir}$ is the virtual class induced by the obstruction theory $$\iota_{n_\alpha,\alpha}^\ast \BE \to \BL_{\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)}$$ by restricting to the $\BT$-fixed locus and taking the $\BT$-fixed part. Note that by Proposition \[restriction of class in K theory\], the perfect obstruction theory $$\BE_{n_\alpha,\alpha}^{\crit}\big|_{\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}}^{\BT\textrm{-}\fix} \to \BL_{\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}}$$ induces the same virtual class. This follows from the general fact that the (equivariant) virtual fundamental class depends only on the class in (equivariant) K-theory of the perfect obstruction theory — cf. [@Siebert Theorem 4.6], where all the ingredients are naturally equivariant.
The statement follows by taking $\BT$-fixed parts in Proposition \[prop: relative cech\] and by Siebert’s result [@Siebert Theorem 4.6] mentioned above.
Higher rank Donaldson–Thomas invariants of compact 3-folds {#section on compact invariants}
----------------------------------------------------------
For a pair $(X,F)$ consisting of a smooth projective toric $3$-fold $X$ and an exceptional locally free sheaf $F$, the perfect obstruction theory gives rise to a $0$-dimensional virtual fundamental class $$\left[\Quot_X(F,n)\right]^{\vir} \in A_0(\Quot_X(F,n)),$$ allowing one to define higher rank Donaldson–Thomas invariants $$\DT_{F,n} = \int_{[\Quot_X(F,n)]^{\vir}}1 \in \BZ.$$ Define the generating function $$\DT_F(q) = \sum_{n\geq 0} \DT_{F,n} q^n.$$
We next compute this series in the case of a $\BT$-equivariant exceptional locally free sheaf, thus proving Theorem \[mainthm:projective\_toric\] from the Introduction.
\[thm for toric proj\] Let $(X,F)$ be a pair consisting of a smooth projective toric $3$-fold $X$ along with an exceptional locally free $\BT$-equivariant sheaf $F$. Then $$\DT_F(q) = \mathsf M((-1)^rq)^{r\int_{X}c_3(T_X\otimes K_X)}.$$
Set $Q=\Quot_X(F,n) $ and $Q_\alpha=\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)$. Since by Proposition \[pot\_global\_equivariant\] the perfect obstruction theory on $Q$ is $\BT$-equivariant, we can apply the virtual localisation formula $$\DT_{F,n} =
\int_{[Q^{\BT}]^{\vir}}e^{\BT}(-N_{Q^{\BT}/Q}^{\vir}),$$ where $N_{Q^{\BT}/Q}^{\vir} $ is the virtual normal bundle on the $\BT$-fixed locus computed in Lemma \[T\_1 fixed locus projective\]. By taking $\BT$-moving parts in Proposition \[prop: relative cech\], we obtain the K-theoretic identity $$N_{Q^{\BT}/Q}^{\vir} = \sum_{\alpha \in \Delta(X)} p_\alpha^\ast N_{Q_\alpha^{\BT}/Q_\alpha}^{\vir}$$ of virtual normal bundles. Thus by Corollary \[cor: pot as box product\] we have $$\begin{aligned}
\int_{[Q^{\BT}]^{\vir}}e^{\BT}(-N_{Q^{\BT}/Q}^{\vir})&= \sum_{|{\bf n}|=n}\prod_{\alpha\in \Delta(X)} \int_{[\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT_1}]^{\vir}} e^{\BT}(-N_{Q_\alpha^{\BT}/Q_\alpha}^{\vir}).\end{aligned}$$ In particular, the virtual fundamental class $[\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}]^{\vir}$ agrees with the one coming from the critical structure. Moreover, by the virtual localisation formula applied with respect to $(\BC^\ast)^r$, we have $$\begin{aligned}
\int_{[\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)^{\BT}]^{\vir}} e^{\BT}(-N_{Q_\alpha^{\BT}/Q_\alpha}^{\vir})=\int_{[\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)]^{\vir}} 1,\end{aligned}$$ where the right hand side is defined equivariantly in §\[sec: variation vertex\]. Finally, by Corollary \[cor:independence\_on\_lambda\_w\], we have an identity $$\int_{[\Quot_{U_\alpha}(F|_{U_\alpha},n_\alpha)]^{\vir}} 1=\int_{[\Quot_{U_\alpha}(\OO_{U_\alpha}^{\oplus r},n_\alpha)]^{\vir}} 1$$ of equivariant integrals, where in the right hand side we take $\OO^{\oplus r}_{U_\alpha}$ with the trivial $\BT$-equivariant weights. Therefore we conclude $$\begin{aligned}
\DT_F(q)
&= \sum_{n\geq 0} q^n \sum_{|{\bf n}|=n}\prod_{\alpha\in \Delta(X)} \int_{[\Quot_{U_\alpha}(\OO_{U_\alpha}^{\oplus r},n_\alpha)]^{\vir}} 1\\
&= \prod_{\alpha\in \Delta(X)} \sum_{n_\alpha\geq 0} q^{n_\alpha} \int_{[\Quot_{U_\alpha}(\OO_{U_\alpha}^{\oplus r},n_\alpha)]^{\vir}} 1\\
&=\prod_{\alpha\in \Delta(X)} \mathsf M((-1)^rq)^{-r\frac{(s^{\alpha}_1+s^{\alpha}_2)(s^{\alpha}_1+s^{\alpha}_3)(s^{\alpha}_2+s^{\alpha}_3)}{s^{\alpha}_1s^{\alpha}_2s^{\alpha}_3}}.\end{aligned}$$ We have used Theorem \[thm:cohomological\] to obtain the last identity, in which we have denoted $s_1^\alpha, s_2^\alpha, s_3^\alpha$ the tangent weights at $p_\alpha$. We conclude taking logarithms: $$\begin{aligned}
\log \DT_F(q)&=\sum_{\alpha\in \Delta(X)}-r\frac{(s^{\alpha}_1+s^{\alpha}_2)(s^{\alpha}_1+s^{\alpha}_3)(s^{\alpha}_2+s^{\alpha}_3)}{s^{\alpha}_1s^{\alpha}_2s^{\alpha}_3}\log \mathsf M((-1)^rq)\\
&= r\int_{X}c_3(T_X\otimes K_X)\cdot \log \mathsf M((-1)^rq)\end{aligned}$$ where the prefactor is computed through ordinary Atiyah–Bott localisation.
We have thus proved Conjecture 3.5 in [@Quot19] in the toric case. The general case is still open and will be investigated in future work.
Conjecture: two obstruction theories are the same
-------------------------------------------------
We close this subsection with a couple of conjectures relating the different obstruction theories appeared in the previous section.
\[conj:pot\_restricted\] Let $\BE$ be the perfect obstruction theory . Then its restriction along the open subscheme $\iota_{n,\alpha}\colon \Quot_{U_\alpha}(F|_{U_\alpha},n)\into \Quot_X(F,n)$ agrees, as a symmetric perfect obstruction theory, with the critical obstruction theory $\BE_{\crit}$ of Proposition \[prop:SPOT\_A\^3\].
One can also ask whether $\iota_{n,\alpha}^\ast \BE$ and $\BE_{\crit}$ are $\BT$-*equivariantly* isomorphic over the cotangent complex of $\Quot_{U_\alpha}(F|_{U_\alpha},n)$. This is of course stronger than the statement of Proposition \[restriction of class in K theory\].
A similar conjecture (essentially the rank $1$ specialisation of Conjecture \[conj:pot\_restricted\]) can be stated for the moduli space $\Hilb^n(\BA^3) = \Quot_{\BA^3}(\OO,n)$, without reference to a compactification $\BA^3 \subset X$. The Hilbert scheme of points has two symmetric perfect obstruction theories: the critical obstruction theory $\BE_{\crit}$ (Proposition \[prop:SPOT\_A\^3\]) and the one coming from moduli of ideal sheaves: if $\mathsf p\colon \BA^3 \times \Hilb^n(\BA^3) \to \Hilb^n(\BA^3)$ is the projection and $\mathfrak I$ is the universal ideal sheaf, one has the obstruction theory $$\RR \mathsf p_\ast \RRlHom(\mathfrak I,\mathfrak I)_0 [2] \to \BL_{\Hilb^n(\BA^3)}$$ obtained from the Atiyah class $\At_{\mathfrak I}$ the way we sketched in .
There is an isomorphism of perfect obstruction theories $$\begin{tikzcd}
\BE_{\crit} \arrow{rr}{\sim}\arrow{dr} & & \RR \mathsf p_\ast \RRlHom(\mathfrak I,\mathfrak I)_0 [2]\arrow{dl} \\
& \BL_{\Hilb^n(\BA^3)} &
\end{tikzcd}$$ on the Hilbert scheme of points $\Hilb^n(\BA^3)$.
Independence of the framing parameters {#sec:framing_independence}
======================================
Functional representation of the generating series {#sec: functional repr. K teoria}
--------------------------------------------------
In this appendix, we prove that the generating series of K-theoretic DT invariant does not depend on the weights of the framing torus $\BT_2 = (\BC^\ast)^r$. The strategy is as follow. Given a $\TT$-fixed point $[S]\in \Quot_{\BA^3}(\OO^{\oplus r},n)$ corresponding to an $r$-colored plane partition $\overline{\pi}$, we identify the localised contribution $[-T^{\vir}_S]$ with the Jeffrey–Kirwan (JK, in short) residue studied and computed in [@BBPT Equation 3.10]. Then we realise $\DT_r^{\KK}(\BA^3,q,t,w)$ as a contour integral (over a suitable cycle) of a meromorphic function, whose only contributing residues are precisely the JK residues mentioned above. The independence on the framing parameters follows from an analysis of this contour integral.
Let $n\geq 0$ and denote by $\hat P,\hat Q, \hat N, \hat{\mathcal{I}}$ the following meromorphic functions $$\begin{aligned}
& \hat P(\underline z;\underline v)=\prod_{i=1}^n\prod_{\alpha=1}^r\sinh\left(\frac{(z_i-v_\alpha)}{2}\right)\\
& \hat Q(\underline z;\underline s)=\prod_{1\le i\neq j\le n}\left[\frac{\sinh{\frac{(z_i-z_j)}{2}}\sinh{\frac{(z_i-z_j+s_{12})}{2}}\sinh{\frac{(z_i-z_j+s_{13})}{2}}\sinh{\frac{(z_i-z_j+s_{23})}{2}}}{\sinh{\frac{(z_i-z_j-s_1)}{2}}\sinh{\frac{(z_i-z_j-s_2)}{2}}\sinh{\frac{(z_i-z_j-s_3)}{2}}\sinh{\frac{(z_i-z_j+s)}{2}}}\right]\\
& \hat N(\underline s)=\frac{1}{n!}\left(-\frac{\sinh{\frac{s_{12}}{2}}\sinh{\frac{s_{13}}{2}}\sinh{\frac{s_{23}}{2}}}{4\pi i\sinh{\frac{s_1}{2}}\sinh{\frac{s_2}{2}}\sinh{\frac{s_3}{2}}\sinh{\frac{s}{2}}}\right)^n\\
& \hat{\mathcal I}(\underline z;\underline s,\underline v)=(-1)^{rn}\frac{\hat P(z_1,\dots,z_n;v_1-s,\dots,v_r-s)}{\hat P(z_1,\dots,z_n;v_1,\dots,v_r)}\hat Q(z_1,\dots,z_n;s_1,s_2,s_3)\end{aligned}$$ with $s_{ij}=s_i+s_j$ and $s=s_1+s_2+s_3$. These functions can be easily expressed in terms of $t_i=e^{s_i},w_i=e^{v_i}$ and $u_i=e^{z_i} $ via the operator $[x]=x^{1/2}-x^{-1/2}$ introduced in §\[sec: preliminaries K theory\], by means of the identity $$[e^{z}]=2\sinh{\frac{z}{2}}.$$
Consider the following *Jeffrey–Kirwan integral* $$\label{JK representation of the K-th invariant}
Z_{JK,r}^{\KK,(n)}(\underline s,\underline v)=\hat N(\underline s)\int_{JK}\left(\prod_{i=1}^n dz_i \right)\hat{\mathcal I}(\underline z;\underline s,\underline v).$$ This integral is, by definition, the sum over the iterated residues of the meromorphic differential form $\hat{\mathcal I}(\underline z;\underline s,\underline v)dz_1\wedge\cdots\wedge dz_n$ computed in the poles satisfying the Jeffrey–Kirwan prescription, as explained in [@JK; @JK_localization_quantization].
\[rmk:integration in BBPT\] Integral presentations of the kind are very well known in physics, as they compute partition functions for gauged linear sigma models in supersymmetric gauge theories by means of supersymmetric localisation. The same expression we are studying, though with different conventions for the equivariant parameters, can be found in [@BBPT] (see e.g. equation (2.24) in loc. cit. for the case $r=1$ and equation (3.1) for the higher rank generalisation to the elliptic case).
The classification of the poles contributing to the computation of the Jeffrey–Kirwan residues was carried out in [@BBPT].
The poles contributing to the computation of are classified in terms of $r$-colored plane partitions or, equivalently, in terms of sums of monomial ideals.
Consider the $r=1$ case. If $n=1$ the only pole to be taken into account in the computation of $Z_{JK,1}^{\KK,(1)}$ is $z_1=v_1$, whose residue is just $1$. We represent the pole corresponding to $v_1$ by a single box.
$z_1=v_1\leftrightarrow
\vcenter{\hbox{\begin{tikzpicture}[scale=.45]
\planepartition{1}
\end{tikzpicture}}}$
If $n=2$, we have three different configurations to take into account, namely $\underline z^{(i)}=(v_1,v_1+s_i)$, $i=1,2,3$. As before we represent the pole for $z_1=v_1$ by a single box, while the poles for $z_2=v_1+s_i$ are denoted by attaching a box to the previous one, along the corresponding $s_i$-direction.
$\underline z^{(1)}=(v_1,v_1+s_1)\leftrightarrow\vcenter{\hbox{\begin{tikzpicture}[scale=.45]
\planepartition{{1},{1}}
\end{tikzpicture}}}$ $,\quad\underline z^{(2)}=(v_1,v_1+s_2)\leftrightarrow
\vcenter{\hbox{\begin{tikzpicture}[scale=.45]
\planepartition{{1,1}}
\end{tikzpicture}}}$ $,\quad\underline z^{(3)}=(v_1,v_1+s_3)\leftrightarrow
\vcenter{\hbox{\begin{tikzpicture}[scale=.45]
\planepartition{{2}}
\end{tikzpicture}}}$
In the same way one can see how for a generic $n$, a certain configuration of poles contributing to the residue formula can be represented by a plane partition of size $n$. For instance
$
\underline z=\begin{pmatrix}
v_1\\
v_1+s_1\\
v_1+s_2\\
v_1+s_1+s_2\\
v_1+2s_2\\
v_1+s_3\\
v_1+s_2+s_3\\
v_1+2s_3\\
\end{pmatrix}\leftrightarrow
\vcenter{\hbox{
\begin{tikzpicture}[scale=.45]
\planepartition{{3,2,1},{1,1}}
\end{tikzpicture}}}
$
When $r>1$ poles originating from different $v_i$’s build up $r$-colored partitions. For example, if $n=1$, the pole $z_1=v_\alpha$ corresponds to a single box in the $\alpha$-th color
$$\begin{split}
z_1=v_\alpha\leftrightarrow
\emptyset\oplus\dots\oplus\emptyset\,\oplus\,&\vcenter{\hbox{\begin{tikzpicture}[scale=.25]
\planepartition{1}
\end{tikzpicture}}}\oplus\emptyset\oplus\dots\oplus\emptyset.
\end{split}$$
Similarly, if $n=2$ we have two classes $\underline z_\alpha$ and $\underline z_{\alpha,\beta}$ of contributing pole configurations. For instance we can grow a single plane partition along $v_\alpha$ or place two single boxes, one along $v_\alpha$, the other one along $v_\beta$. As in the previous example we have three configurations within the first class of poles, namely
$\underline z^{(1)}_\alpha=(v_\alpha,v_\alpha+s_1)\leftrightarrow\emptyset\oplus\dots\emptyset\oplus\vcenter{\hbox{\begin{tikzpicture}[scale=.25]
\planepartition{{1},{1}}
\end{tikzpicture}}}\oplus\emptyset\oplus\dots\oplus\emptyset$
$\underline z^{(2)}_\alpha=(v_\alpha,v_\alpha+s_2)\leftrightarrow\emptyset\oplus\dots\emptyset\oplus\vcenter{\hbox{\begin{tikzpicture}[scale=.25]
\planepartition{{1,1}}
\end{tikzpicture}}}\oplus\emptyset\oplus\dots\oplus\emptyset$
$\underline z^{(3)}_\alpha=(v_\alpha,v_\alpha+s_3)\leftrightarrow\emptyset\oplus\dots\emptyset\oplus\vcenter{\hbox{\begin{tikzpicture}[scale=.25]
\planepartition{{2}}
\end{tikzpicture}}}\oplus\emptyset\oplus\dots\oplus\emptyset$
while for the second one we just have a single choice, namely
$$\begin{split}
\underline z_{\alpha,\beta}=(v_\alpha,v_\beta)\leftrightarrow
\emptyset\oplus\dots\oplus\emptyset\oplus\vcenter{\hbox{\begin{tikzpicture}[scale=.25]
\planepartition{1}
\end{tikzpicture}}}\oplus\emptyset\oplus\dots\oplus\emptyset\oplus\vcenter{\hbox{\begin{tikzpicture}[scale=.25]
\planepartition{1}
\end{tikzpicture}}}\oplus\emptyset\oplus\dots\oplus\emptyset.
\end{split}$$
For $n>1$ we would get more classes of pole configurations, corresponding to the different ways of building $r$-colored plane partition of size $n$, each plane partition labelled by one of the $r$ different colors.
\[contour integration is equivalent to K theory invariants\] There is an identity $$\label{K-th invariant equals integration}
\sum_{n\ge 0}Z_{JK,r}^{\KK,(n)}(s,v)q^n=\DT_r^{\KK}(\BA^3,q,t,w),$$ by setting $t_i=e^{ s_i}$ and $w_\alpha=e^{ v_\alpha}$.
The left-hand-side was computed in [@BBPT Equation 3.10] as a sum of JK residues, classified by $r$-colored plane partitions. Such residues are equal to the localised contributions obtained by applying localisation formula on $\DT_r^{\KK}(\BA^3,q,t,w)$, through the vertex formalism developed in Section \[sec: virtual tangent space\], which are again classified by $r$-colored partitions. The statement follows by comparing the coefficients of $q^n$.
In the next lemma, we identify the JK integral as a contour integral, over a cycle encircling all the poles contributing to the JK residues.
\[JK is equivalent to contour integration for K theory\] The JK-integration in is equivalent to an iterated contour integration on a product of infinite cylinders $\Cyl^n$, with $\Cyl\cong\BR\times[-\pi i,\pi i]\subset\BC$, along an $n$-dimensional cycle $\mathcal C$ encircling all the poles of the integrand lying in $(\mathbb H^-)^n\cap\Cyl^n$, where $\mathbb H^-$ denotes the lower half–plane in $\mathbb C$: $$\label{contour integration in K theory}
\int_{JK}\left(\prod_{i=1}^n dz_i \right)\hat{\mathcal I}(\underline z;\underline s,\underline v)=\lim_{\delta,\varepsilon\searrow 0}\oint_{\mathcal C}\left(\prod_{i=1}^n dz_i \right)\hat{\mathcal I}(\underline z;\underline s-i\delta,\underline v-i\varepsilon),$$ with $0<\varepsilon<\delta<\pi$ and the integration is to be intended as $$\oint_{\mathcal C}\left(\prod_{i=1}^ndz_i\right)=\idotsint_\mathcal Cdz_n\cdots dz_1.$$
Proving the statement is equivalent to proving that the poles contributing to the right hand side of are the same as the ones selected by the JK-prescription. This fact can easily be seen by listing zeros and poles of $\hat{\mathcal I}(\underline z;\underline{\hat s},\underline{\hat v})$ lying in the lower-half plane $\BH^-\cap\Cyl=\Cyl^-\subset\BC$, where $\hat s_i=s_i-i \delta$ and $\hat v_i=v_i-i \varepsilon$. By the continuity of the $\sinh$ function and the fact that $\sinh(x)=0$ if and only if $x\in 2\pi i \BZ $, such poles and zeros are then displayed in Table \[table poles and zeros\] where $\alpha=1,\dots,r$ and $i,j=1,\dots,n$.
[**poles**]{} [**zeros**]{}
--------------------- ----------------------------
$z_i=\hat v_\alpha$ $z_i=\hat v_\alpha-\hat s$
$z_i-z_j=\hat s_1$ $z_i-z_j=-\hat s_{12}$
$z_i-z_j=\hat s_2$ $z_i-z_j=-\hat s_{23}$
$z_i-z_j=\hat s_3$ $z_i-z_j=-\hat s_{13}$
$z_i-z_j=-\hat s$ $z_i-z_j=0$
: Poles and zeros of $\hat{\mathcal I}(\underline z;\underline{\hat s},\underline{\hat v})$.[]{data-label="table poles and zeros"}
It is then immediate to see that the only poles with non-zero residues in $(\mathbb H^-)^n$ correspond to plane partitions. Indeed, choose the pole for the first integration to be fixed at $z_1=v_\alpha$, then the pole for $z_2$ can’t be taken again at $v_\alpha$, otherwise we would get zero residue. In the same way, we can’t pick $z_2=z_1+s$, so the only possibilities are to choose $z_2=v_\beta$, $\beta\neq\alpha$, or $z_2=z_1+s_i$, $i=1,2,3$. The construction then proceeds in the same way, so poles with non-zero residue can only be organised in $r$-colored plane partitions. A more detailed analysis of the contributing pole configurations is contained in [@BBPT Appendix D.3].
\[rep of C\] A good choice for $\mathcal C$ is $\mathcal C=\mathcal C_R\times\cdots\times\mathcal C_R$, with $\mathcal C_R$ the contour depicted in Figure \[fig:contour k theory\], taking $R\in\BR_{>0}$ sufficiently large.
(0,-2.5) ellipse (2 and .7); (2\*cos[80]{},-2.5+.7\*sin[80]{}) arc (80:180:2 and .7) – (-2,2.5) –(-2,2.5) arc (180:80:2 and .7) – cycle; (0,2.5) ellipse (2 and .7); (-2,-2.5) arc (180:260:2 and .7) – (2\*cos[260]{},2.5+.7\*sin[260]{}) –(2\*cos[260]{},2.5+.7\*sin[260]{}) arc (260:180:2 and .7) – cycle; (0,0) +(180:2) arc (180:360:2 and .7) ; (0,0) +(0:2) arc (0:180:2 and .7) ; (2\*cos[260]{},-3+.7\*sin[260]{})–(2\*cos[260]{},3+.7\*sin[260]{}); (2\*cos[80]{},-3+.7\*sin[80]{})–(2\*cos[80]{},3+.7\*sin[80]{}); (2\*cos[260]{},-3.5+.7\*sin[260]{})–(2\*cos[260]{},-3+.7\*sin[260]{}); (2\*cos[260]{},3+.7\*sin[260]{})–(2\*cos[260]{},3.5+.7\*sin[260]{}); (2\*cos[80]{},-3.5+.7\*sin[80]{})–(2\*cos[80]{},-3+.7\*sin[80]{}); (2\*cos[80]{},3+.7\*sin[80]{})–(2\*cos[80]{},3.5+.7\*sin[80]{}); (2\*cos[260]{},-2+.7\*sin[260]{})–(2\*cos[260]{},-1.5+.7\*sin[260]{}); (2\*cos[260]{},1+.7\*sin[260]{})–(2\*cos[260]{},1.7+.7\*sin[260]{}); (2\*cos[80]{},-2+.7\*sin[80]{})–(2\*cos[80]{},-1.5+.7\*sin[80]{}); (2\*cos[80]{},.8+.7\*sin[80]{})–(2\*cos[80]{},1.7+.7\*sin[80]{}); (2,-2.5)–(2,2.5); at (4.3,2.7) [$\Cyl\cong\mathbb R\times\left[-\pi i,\pi i\right]$]{}; at (2\*cos[260]{}+.7,-3.2+.7\*sin[260]{}) [$\Im z=0$]{}; at (2\*cos[80]{}+1.1,3.2+.7\*sin[80]{}) [$\Im z=-\pi$]{}; at (2.8,0) ; at (2\*cos[260]{}-.35,-2.2+.7\*sin[260]{}) [$+R$]{}; at (2\*cos[260]{}-.35,2.2+.7\*sin[260]{}) [$-R$]{};
We then define $$Z_{\mathcal C,r}^{\KK,(n)}(\underline s,\underline v)=\hat N(\underline{s})\lim_{\delta,\varepsilon\searrow 0}\oint_{\mathcal C}\left(\prod_{i=1}^n dz_i \right)\hat{\mathcal I}(\underline z;\underline s-i\delta,\underline v-i\varepsilon),$$ Thanks to Lemma \[contour integration is equivalent to K theory invariants\], this integral computes the K-theoretic invariants we are interested in.
The fact that the integration of suitable meromorphic forms over a given cycle reproduces the results of equivariant localisation formulæ is not striking if one takes into account how they come to be. As we already pointed out in Remark \[rmk:integration in BBPT\], formulæ of the form of naturally emerge in the context of supersymmetric gauge theories, whose partition functions are known to compute indices of (usually twisted) Dirac operators. The similarity is made even more manifest by the fact that integral formulæ of the kind , or the ones found in [@BBPT] , are in fact obtained by exploiting the technique of supersymmetric localisation, which can be roughly understood as an infinite dimensional analogue of standard equivariant localisation. The usefulness of such a technique resides in its ability of often reducing the problem of computing path integrals to the computation of ordinary integrals. Indeed it can be shown that the path integration in a supersymmetric theory is reduced from the (infinite-dimensional) space of field configurations to the BPS locus, which is to say the space of field configurations supersymmetric under the SUSY conserved charges. In the case at hand, one starts from a D6-D0 brane system on a threefold $X$ in a type IIb supersymmetric background, such that the D6-branes wrap the threefold while the worldvolume of the D0-branes is a circle $S^1$. If $X=\BA^3$ the D0 effective theory on $S^1$ is a supersymmetric quantum mechanical model whose BPS locus is exactly the DT moduli space, equipped with its critical structure.
\[independence of Z\^K on w\] The generating function $\DT_r^{\KK}(\BA^3,q,t,w)$ does not depend on the weights $w_1,\dots,w_r$.
By Lemmas \[contour integration is equivalent to K theory invariants\] and \[JK is equivalent to contour integration for K theory\], it is enough to compute $$\frac{\partial}{\partial v_\alpha}\DT_r^{\KK}(\BA^3,q,t,w)=\sum_{n\ge 0}\frac{\partial}{\partial v_\alpha}Z^{\KK,(n)}_{\mathcal C,r}(s,v)q^n$$ and show it vanishes. Assume for now that we can apply the dominated convergence theorem and exchange the order of derivation and integration. The derivative with respect to $v_\alpha$ of $\hat{\mathcal I}(\underline z;\underline s,\underline v)$ takes a very simple form: $$\label{derivative of hat I}
\frac{\partial}{\partial v_\alpha}\hat{\mathcal I}(\underline z;\underline s,\underline v)=-\left(\sum_{i=1}^n\frac{\sinh{\frac{s}{2} }}{\sinh{\frac{(z_i-v_\alpha)}{2}}\sinh{\frac{(z_i-v_\alpha+s)}{2}}}\right)\hat{\mathcal I}(\underline z;\underline s,\underline v).$$ The integration is performed as a sum over the (ordered) iterated residues of the integrand over the poles of $z_i$, $i=1,\dots,n$ in $(\BH^-)^n\cap\Cyl^n$. As an application of the global residue theorem, we reverse each contour integration and compute the iterated residues over the poles in the region outside of $\mathcal C\subset\Cyl^n$. Moreover, for each contributing configuration, there is a $j$ in $1,\ldots,n$ such that the only poles of $z_j$ lie in $\mathbb H^-\cap\Cyl^n$, and all the corresponding contributions are then manifestly zero due to the particular form of .
To apply dominated convergence theorem, it is enough to choose a concrete representative of $\mathcal C$, say for instance $\mathcal C_R$ as in Remark \[rep of C\]. Notice that $\mathcal C_R$ is compact, $\hat{\mathcal I}(\underline z;\underline s-i\delta,\underline v-i\varepsilon)$ is continuous on $\mathcal C_R$ and $\hat{\mathcal I}(\underline z;\underline s-i\delta,\underline v-i\varepsilon)$ is bounded in $\underline z$. In fact, one can find an explicit majorisation of $|\hat{\mathcal I}(\underline z;\underline s-i\delta,\underline v-i\varepsilon)|$ on $\mathcal C_R$ by an integrable function.
A different proof of Theorem \[independence of Z\^K on w\] is given in the work of Arbesfeld–Kononov [@Noah_Yasha].
The function $Z_{JK,r}^{\KK,(n)}$ and its cohomological reduction are known in physics as partition functions of $U(r)$ instantonic objects constructed starting from low energy effective theories of D0-D6 and D(-1)-D5 brane systems, respectively. In this context the action of $\BT_2=(\BC^*)^r$ is translated in a global $U(r)$ symmetry of the theory, which on the SUSY fixed locus is completely higgsed to a $U(1)^r$ symmetry. Similarly the action of $\BT_1=(\BC^*)^3$ is translated in a global spacetime $U(3)$ symmetry, which is commonly known as $\Omega$-background.
[^1]: We use the terminology *colored partitions* as used in the physics literature, differently from the pre–existing one in combinatorics.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'App stores include an increasing amount of user feedback in form of app ratings and reviews. Research and recently also tool vendors have proposed analytics and data mining solutions to leverage this feedback to developers and analysts, e.g., for supporting release decisions. Research also showed that positive feedback improves apps’ downloads and sales figures and thus their success. As a side effect, a market for fake, incentivized app reviews emerged with yet unclear consequences for developers, app users, and app store operators. This paper studies fake reviews, their providers, characteristics, and how well they can be automatically detected. We conducted disguised questionnaires with 43 fake review providers and studied their review policies to understand their strategies and offers. By comparing 60,000 fake reviews with 62 million reviews from the Apple App Store we found significant differences, e.g., between the corresponding apps, reviewers, rating distribution, and frequency. This inspired the development of a simple classifier to automatically detect fake reviews in app stores. On a labelled and imbalanced dataset including one-tenth of fake reviews, as reported in other domains, our classifier achieved a recall of 91% and an AUC/ROC value of 98%. We discuss our findings and their impact on software engineering, app users, and app store operators.'
author:
- Daniel Martens
- Walid Maalej
bibliography:
- 'refs.bib'
date: 'Received: date / Accepted: date'
title: Towards Understanding and Detecting Fake Reviews in App Stores
---
Introduction
============
In app stores, users can rate downloaded apps on a scale from 1 to 5 stars and write a review message. Thereby, they can express satisfaction or dissatisfaction, report bugs, or suggest new features [@6606604; @Pagano:2013jn; @Maalej:2016:Automatic]. Similar to other online stores, before downloading an app, users often read through the reviews. Research found that ratings and reviews correlate with sales and download ranks [@Harman:2012gw; @Pagano:2013jn; @Svedic:aG1qEZuT; @Martin:2016fe; @FINKELSTEIN2017119]. Stable numerous ratings lead to higher downloads and sales numbers.
As a side effect, an illegal market for fake app reviews has emerged, with the goal to offer services that help app vendors improve their ratings and ranking in app stores. According to app store operators, in regular app reviews, real users are supposed to be triggered by their satisfaction or dissatisfaction of using the app to provide feedback. Fake reviewers, however, get paid or similarly rewarded to submit reviews. They might or might not be real users of the app. Their review might or might not be correct and reflecting their opinion.
We refer to this type of non-spontaneous, requested, and rewarded reviews as **fake reviews**. Fake reviews are prohibited in popular app stores such as in Google Play [@Weblink:10006] or Apple App Store [@Weblink:10005]. For instance, Apple states: *“If we find that you have attempted to manipulate reviews, inflate your chart rankings with paid, incentivized, filtered, or fake feedback, or engage with third party services to do so on your behalf, we will take steps to preserve the integrity of the App Store, which may include expelling you from the Developer Program.”*.
Recently, Google highlighted the negative effects of fake reviews in an official statement and explicitly asks developers to not buy and users to not accept payments to provide fake reviews [@Weblink:10021]. Even governmental competition authorities started taking actions against companies using fake reviews to embellish their apps. For instance, the Canadian telecommunication provider Bell was fined \$1.25 million [@Weblink:10007] for faking positive reviews to their apps. Vice versa, the CNN app was affected by thousands of negative fake reviews to decrease its rating and ranking within the Apple App Store [@Weblink:20001].
While the phenomena of fake participation (e.g., in form of commenting, reporting or reviewing) is well-known in domains such as online journalism [@Lee:2010:USS:1835449.1835522; @DBLP:journals/corr/FerraraVDMF14; @6921650; @DBLP:journals/corr/SubrahmanianADK16] or on business and travel portals [@Jindal:2008:OSA:1341531.1341560; @Ott:2011:FDO:2002472.2002512; @feng2012distributional; @Mukherjee2013WhatYF], it remains understudied in software engineering – in spite of recent significant research on app store analysis and feedback analytics [@Martin:2016fe].
Fake reviews threaten the integrity of app stores. If real users don’t trust the reviews, they probably will refrain from reading and writing reviews themselves. This can result into a problem for app store operators, as app reviews is a central concept of the app store ecosystem. Fake reviews can have negative implications for app developers and analysts as well. Numerous software and requirements engineering researchers studied app reviews, e.g., to derive useful development information such as bug reports [@Khalid:2013:IUC:2486788.2487044; @Maalej:2016:Automatic] or to understand and steer the dialogue between users and developers [@6624001; @Oh:2013:FDI:2468356.2468681; @Johann:RE:2017; @7886888]. Further, researchers [@6606604; @Chen:2014:AMI:2568225.2568263; @Maalej:2015:Software] and more recently tool vendors [@Weblink:10008] suggested tools that derive actionable information for software teams from reviews such as release priorities and app feature co-existence. None of these works considers fake reviews and their implications. Negative fake reviews, e.g., by competitors reporting false issues, can lead to confusion and waste of developers’ time. Positive fake reviews might also lead to wrong insights about real users needs and requirements.
In this paper, we study fake app reviews, focusing on three research questions:
1. **How and by whom are app ratings and reviews manipulated?**\
Through online research and an investigative disguised questionnaire, we identified 43 fake review providers and gathered information about their fake reviewing strategies and offers.
2. **How do fake reviews differ from regular app reviews?**\
We crawled $\sim$60,000 fake reviews, empirically analyzed and compared them with $\sim$62 million official app reviews from the Apple App Store. We report on quantitative differences of fake reviewers and concerned apps.
3. **How accurate can fake reviews be automatically detected?**\
We developed a supervised classifier to detect fake reviews. Within an in-the-wild experiment, we evaluated the performance of multiple classification algorithms, configurations, and classification features.
In Section \[sec:design\] we introduce the research questions, method, and data. We then report on the results along the research questions: fake review market in Section \[sec:strategies\], characteristics in Section \[sec:char\], and automated detection in Section \[sec:detection\]. Afterwards, we discuss the implications and limitations of our findings in Section \[sec:discussion\]. Finally, we survey related work in Section \[sec:relwork\] and conclude the paper in Section \[sec:conclusion\].
Study Design {#sec:design}
============
We first introduce our research questions. Then, we describe our research method and data along the data collection, preparation, and analysis phase.
Research Questions
------------------
We aim to qualitatively and quantitatively understand fake app reviews including their market, characteristics, and potential automated detection. In the following we detail our research questions by listing the sub-questions we aim to answer.
1. **Fake review market** reveals how app sales and downloads are manipulated and to which conditions. We investigate the following questions:
1. By whom are fake reviews offered? What strategies do fake review providers follow?
2. What exact services do fake review providers offer and under which conditions?
3. What are providers policies for submitting fake reviews? Do these reveal indicators to detect fake reviews?
2. **Fake review characteristics** reveal empirical differences between official and fake reviews, including reviewed apps and reviewers.
1. Which apps are typically affected by fake reviews? What are their categories, prices, and deletion ratio?
2. What is a typical fake reviewer, e.g., in terms of number of reviews provided and review frequency?
3. How do official and fake reviews differ, e.g., with regard to rating, length, votes, submission date, and content?
3. **Fake review detection** examines how well supervised machine learning algorithms can detect fake reviews. We focus on the following questions:
1. Which machine learning features can be used to automatically detect fake reviews?
2. Which machine learning algorithms perform best to classify reviews as fake/non-fake?
3. How can classifiers further be optimized? What is the relative importance of the classification features?
4. How do the classifiers perform in practice on imbalanced datasets with different proportional distributions of fake and regular reviews?
Research Method and Data
------------------------
Our research method consists of a data collection, preparation, and analysis phase, as depicted in Figure \[fig:researchmethod\]. We detail on each of the three phases in the following.
![Research method including data collection, preparation, and analysis phases.[]{data-label="fig:researchmethod"}](02-researchmethod.pdf){width="\columnwidth"}
### Data Collection Phase
For this study we collected two datasets: an official reviews dataset including app metadata and reviews from the Apple App Store; as well as a fake reviews dataset including metadata of apps affected by fake reviews, and fake reviews itself.
The **official reviews dataset** created in March 2017 consists of 1,430,091 apps, their metadata, and reviews. To collect the data, we implemented a distributed crawling tool using GoLang based on iTunes APIs, which we deployed on hundreds of cloud servers. The data collection included three steps. First, we crawled a list of all app identifiers available on the Apple App Store of the United States, as we focus on English reviews. Second, we obtained the metadata for each app, including the category, price, and number of reviews, using the iTunes Search API[^1]. Last, we retrieved the app reviews using an internal iTunes API.
Overall, the Apple App Store included 207,782,199 ratings of which 67,727,914 (24.6%) have a review. We were able to crawl 62,617,037 (92.4%) of these reviews, as iTunes does not allow to receive more than 30,000 reviews per app. The size of the dataset is 36.58 GB. The crawled reviews were written by 25,333,786 distinct reviewers, i.e., users with different Apple IDs[^2]. On average every reviewer submits around 2.47 reviews. The oldest app review was entered on 10/07/2008, therefore our dataset spans for nearly 9 years.
![Screenshot of a fake review provider, offering a) app developers to buy reviews and b) non-/developers to sign-up and write rewarded reviews. Section c) shows a list of apps requesting fake reviews against monetary reward (obfuscated by the authors).[]{data-label="fig:providermulti"}](02-providermulti){width=".8\columnwidth"}
The **fake reviews dataset** was collected in April 2017 following three steps. First, we identified 43 fake review providers by performing a structured manual Google web search. To identify relevant search terms, we initially searched for the phrase “buy app reviews”. We extracted related search terms suggested by the search engine. For those we repeated the previous step, resulting in 39 unique search terms, which are included in the replication package. Afterwards, we crawled the results of the ten first pages for each search term. We removed duplicate results and marked each result as fake review provider, relevant discussion about fake reviews (e.g., in forums), or irrelevant result. From relevant discussions we extracted additional fake review providers by reading through all sub-pages of the discussions. Then, we manually extracted the provider’s offers from their websites.
In the second step, we conducted a disguised questionnaire to collect initial indicators for fake reviews such as the minimum star-rating and length. The questionnaire was presented as a request for buying fake reviews and sent to all providers per email on 26/04/2017[^3]. For providers offering users to sign-up as fake reviewers to exchange or get paid for providing fake reviews, we created accounts and extracted the policies submitted fake reviews must comply with (cf. Figure \[fig:providermulti\]).
[XlrrX]{} Provider Id & Provider Type & \# Apps & \# Reviews & Approach\
PRP10 & Paid Review Provider & 77 & - & Crawl\
PRP16 & Paid Review Provider & 19 & 4 & Crawl, Social\
PRP21 & Paid Review Provider & 3 & - & Social\
PRP25 & Paid Review Provider & - & 3 & Social\
PRP26 & Paid Review Provider & - & 10 & Social\
PRP28 & Paid Review Provider & - & 3 & Social\
REP1 & Review Exchange Portal & 268 & - & Crawl\
REP2 & Review Exchange Portal & 277 & - & Crawl\
REP3 & Review Exchange Portal & 2,007 & 60,411 & API, Crawl\
REP5 & Review Exchange Portal & 7 & - & Crawl\
REP6 & Review Exchange Portal & 9 & - & Crawl\
REP8 & Review Exchange Portal & 182 & - & Crawl\
REP9 & Review Exchange Portal & 4 & - & Crawl\
& & $\sum$ = 2,853 & $\sum$ = 60,431 &\
In the third and last step, we collected the fake review data, i.e., lists of apps requesting fake reviews and fake reviews itself. For this, we used three approaches:
1. *Social investigation*, i.e., asking the providers for fake review examples, while pretending to be interested in their services. We contacted the providers via email or live-chats on their websites. Using this strategy we received 3 apps and 20 fake reviews from 5 providers.
2. *Crawling*, for providers offering to sign-up as fake reviewers, we checked if the lists of apps requesting fake reviews are available (see Figure \[fig:providermulti\], part c). To extract the apps we implemented crawlers. A sample crawler is included in the replication package. Overall, we collected 2,850 apps from 9 providers.
3. *APIs*, we found that providers require reviewers to upload screenshots of their reviews as a proof. We searched for publicly accessible screenshots and downloaded them. Based on this, we gathered 60,411 reviews from a single provider.
**Overall we identified 2,853 apps and 60,431 reviews** from 13 providers, see Table \[tab:fakereviewdataset\]. Per provider the number of extracted apps and reviews is given, in case we could successfully apply at least one of the introduced approaches. The size of the collected data is 11.29 GB. We refer to this as **unfiltered fake data** (cf. Figure \[fig:researchmethod\]), as it needs to be prepared for further analysis. For example, reviews within the dataset could have already been removed from the app store. For data preparation and analysis, all data, except the screenshots, is persisted as Parquet files and analyzed with Apache Zeppelin and Spark[^4].
### Data Preparation Phase
Most fake reviews were collected in form of screenshots as shown in Figure \[fig:reviewscreen\]. We converted the screenshots into text using the Tesseract OCR engine[^5]. We removed incomplete reviews that do not include a full readable title and body, e.g., if the title was outside the screen’s visible area. Then, we used the Language Identification (LangID) library[^6] to retrieve fake reviews in English language only. We removed 7,445 reviews (12.32%) resulting in 52,986 fake reviews in English language.
![Screenshots of fake reviews before submission to the app store used as proof for fake review providers, depicting a) fake review included in our study, b) cut-off fake review excluded from the study, and c) non-English fake review also excluded from the study.[]{data-label="fig:reviewscreen"}](02-reviewscreenshots){width=".8\columnwidth"}
Since the screenshots show the review edit screens before submitting the reviews to the app store, we further filtered the collected fake reviews for three possible reasons. First, we cannot assure that the reviews were actually submitted. Second, the reviews could have not been unlocked by the app store operator. Third, the reviews could have been deleted. Therefore, we only considered fake reviews that have been published to and still exist within the Apple App Store, i.e., which we could identify in the official reviews dataset as well.
For uniquely identifying (i.e., matching) reviews from the fake reviews dataset within the official reviews dataset, we removed duplicate reviews which consist of the same title and body within both datasets. Thereby, we removed 4,298 (8.11%) fake reviews leaving 48,688 items. The percentage of duplicate reviews within the official reviews dataset is with 16.08% nearly twice as high, which may be an indicator for the high diversity of fake reviews. We performed the matching using exact text comparison and by comparing the reviews’ Levensthein distances. We used the Levensthein distance as single characters on the screenshots were sometimes not parsed correctly by the OCR engine. We searched for all fake reviews within the official reviews dataset by using an edit distance of up to 10 characters. For possible matches identified using the Levensthein distance, we manually verified if one of the suggested pairs is a match using two human annotators comparing the screenshot of the fake review and the possible matches. In case of disagreements (3% of all cases), a third annotator resolved the conflict. We matched 6,020 reviews by exact text comparison and 2,584 reviews by comparing the Levensthein distance.
Overall, we were able to identify 8,607 of the 60,431 (14.2%) collected fake reviews within the Apple App Store. These reviews were extracted from 5 providers. We also matched *apps* affected by fake reviews against the official reviews dataset, as the apps might not be available in the US App Store or might have been deleted. Of the 2,853 collected apps we found 2,174 apps (76.2%) in the official reviews dataset. Further, we identified 898 additional apps by extracting the app identifiers from previously matched fake reviews, resulting in 3,072 apps. We removed all apps that did not receive reviews within the app store, resulting in 1,929 of 3,072 (62.8%) apps provided by 10 different providers. Finally, we identified 721 fake *reviewers*, i.e., accounts of persons submitting fake reviews to the app store, by extracting their user identifiers from fake reviews.
In summary, after data cleaning the **fake reviews dataset** consists of 43 providers and structural information about their offers and policies, as well as **8,607 fake reviews, 1,929 apps affected by fake reviews, and 721 fake reviewers**. The dataset spans for nearly 7 years, as the oldest fake review was entered on 16/10/2010. Table \[tab:datasets\] summarizes the official and fake reviews datasets.
[Xrr]{} & Official Reviews Dataset & Fake Reviews Dataset\
\# of reviews & 62,617,037 & 8,607\
\# of apps & 1,430,091 & 1,929\
\# of reviewers & 25,333,786 & 721\
### Data Analysis Phase
The data analysis phase consists of three steps which respectively answer the research questions. To answer our first research question regarding the fake review **market**, we analyzed the qualitative data aggregated from the providers’ websites, collected during the questionnaire, and extracted from the review policies.
To explore the fake review **characteristics** we applied a statistical analysis of the reviews, apps, and reviewers. We compare the figures from the fake reviews dataset to those from the official reviews dataset and run statistical tests whenever applicable. For example, we found that most fake reviews are provided for games. While regular apps receive most reviews on the day of an app release, apps affected by fake reviews receive most reviews eleven days after the update. This could indicate that apps affected by fake reviews do not have a real users basis that intrinsically provides reviews in reaction to changes introduced by app updates. Further, we found that fake reviewers provide twelve times as much reviews, with a four times higher frequency. Also, fake reviews have more positive ratings compared to official reviews, however the biggest difference exists between the amount of 1-star ratings.
To **detect** fake reviews, we created a labeled and balanced dataset of fake reviews and official reviews and used it to train and evaluate multiple classifiers, based on machine learning features that we derived from the analysis of characteristics. We conducted a hyperparameter tuning of the classifiers and evaluated the importance of the classification features. Finally, we performed an in-the-wild experiment to get more realistic results of how the classifiers perform in practice. Therefore, we used imbalanced datasets of fake and regular reviews. We varied the skewness of the datasets between 90% to 0.1% fake reviews and compared the classification results.
We detail each of these analysis steps in the following chapters. To support replication, our dataset and the analyses source code as Zeppelin notebooks are publicly available on our website[^7].
Fake Review Market (RQ1) {#sec:strategies}
========================
This section describes fake review providers and their market strategies, as well as offers and pricing models. Afterwards, pretended characteristics of fake reviews are summarized based on the results of the disguised questionnaire and analysis of reviewing policies.
Review Providers and Market Strategies
--------------------------------------
We identified 43 providers offering fake reviews. These can be separated into two groups by their strategies used to supply reviews.
![A fake review request with predefined rating and review on a review exchange portal.[]{data-label="fig:reviewrequest"}](02-reviewwrite){width="\columnwidth"}
**Paid review providers (PRP) accept payments to provide fake reviews.** This applies for 34 out of 43 (79%) providers. User can select a package of, e.g., 50 reviews, specify their app name and identifier, and purchase it via Paypal or similar services. Afterwards, the fake reviews are submitted to the app store.
**Review exchange portals (REP) allow app developers to sign-up and exchange reviews.** The remaining 9 providers (21%) belong to this group. After sign-up developers browse through a list of apps requesting fake reviews. Figure \[fig:reviewrequest\] shows a sample request for fake reviews. Depending on their policies, review exchange portals ask users to submit fake reviews, e.g., with predefined ratings and review messages. For each fake review the developer submits, one credit is given as a reward. Developers with at least one credit can add their app to the list. Then, the credits are redeemed into reviews written by other developers.
In some cases, review exchange portals allow developers to buy credits and non-developers to sign-up and submit fake reviews. Non-developers are rewarded using micro-payments, typically between \$0.20 to \$1.50 per fake review.
Figure \[fig:reviewingstrategy\] shows the strategies of the fake review providers. After deciding to buy fake reviews at a paid review provider or to exchange (or buy) reviews at a review exchange portal, the developer provides basic information, such as the application identifier and whether the reviews should be positive or negative. Optionally, further information, such as keywords to be included within the reviews or predefined review messages, can be submitted. Using this information, the provider creates a review request (see, e.g., Figure \[fig:reviewrequest\]).
Review exchange portals publish these requests on their internal platform to recruit fake reviewers. For paid review providers the publishing process is not transparent. Using social investigation and by offering our service as fake reviewer, we identified that at least five providers publish their review requests on invite-only Slack or Telegram channels. By observing the communication within these channels, we found that paid review providers occasionally cross-post review requests on review exchange portals while offering micro payments.
Fake reviewers can browse through and assign review requests to themselves. Afterwards, they are presented a review policy that regulates what information or rating the review should include. The fake reviewer submits an appropriate fake review to the app store. As a proof, the fake reviewer uploads a screenshot of the review edit screen showing their rating and review to the provider.
Last, the provider compares if the provided review meets the reviewing policy. If this applies, and the review has been published within the app store, the fake reviewer is rewarded. Reviewers providing reviews that do not meet the policies are excluded from the channels or portals and are not rewarded.
![Fake reviewing strategies.[]{data-label="fig:reviewingstrategy"}](03-approach){width="\columnwidth"}
[llXXXXXXXXXXXX]{} PRP & Co. & & &\
& & & & & & &\
& & Min & Max & Min & Max & Min & Max & Min & Max & Min & Max & Min & Max\
1 &IN & & & 1.35 & 1.50 & & & & & & & &\
2 &DK & 4.63 & 4.90 & 0.98 & 1.00 & & & & & 0.25 & 0.25 & 0.05 & 0.06\
3 &IN & & & 0.25 & 0.25 & & & 0.20 & 0.20 & & & 0.09 & 0.09\
4 &IN & & & 1.50 & 1.98 & & & & & & & &\
5 &GB & & & 1.11 & 1.50 & & & & & & & 0.10 & 0.10\
6 &US & & & 2.90 & 2.95 & & & 1.28 & 1.58 & & & 1.30 & 1.36\
7 &RU & & & 0.25 & 0.25 & & & 0.20 & 0.20 & & & 0.10 & 0.10\
8 &US & 6.00 & 9.00 & **6.00** & **9.00** & & & & & & & &\
9 &US & 3.33 & 4.17 & 1.00 & 1.50 & & & & & & & 0.09 & 0.15\
10 &NL & & & 1.55 & 1.55 & & & 1.00 & 1.00 & 0.65 & 0.65 & 0.20 & 0.20\
11 &US & 2.50 & 4.00 & 3.50 & 5.00 & & & & & 0.49 & 0.90 & 0.08 & 0.12\
12 &CA & & & 1.00 & 1.00 & & & & & & & &\
13 &US & 2.15 & 2.50 & 1.59 & 2.50 & & & & & 0.34 & 0.46 & 0.13 & 0.20\
14 &US & 4.30 & 5.00 & 0.85 & 1.20 & & & & & 0.35 & 0.38 & 0.35 & 0.38\
15 &IN & & & 0.15 & 0.15 & & & 0.08 & 0.08 & & & 0.10 & 0.10\
16 &RU & 2.09 & 2.99 & 2.99 & 2.99 & & & & & & & &\
17 &US & 5.02 & 5.20 & 2.00 & 2.60 & & & & & 0.40 & 0.45 & 0.40 & 0.46\
18 &DE & & & 2.50 & 2.50 & & & & & & & 0.17 & 0.17\
19 &US & **8.69** & **10.00** & 3.60 & 4.00 & & & 1.28 & 1.60 & & & 1.36 & 1.58\
20 &VN & & & 0.05 & 0.05 & & & 0.05 & 0.05 & 0.10 & 0.10 & 0.05 & 0.05\
21 &US & 2.00 & 2.00 & 1.40 & 2.00 & & & & & & & &\
22 &US & & & 1.45 & 2.00 & & & & & 0.29 & 0.32 & 0.08 & 0.15\
23 &RU & 3.40 & 4.00 & 2.75 & 2.75 & & & & & & & &\
24 &US & & & 1.00 & 1.00 & & & 0.80 & 0.80 & & & 0.15 & 0.15\
25 &NL & 1.78 & 3.30 & 1.78 & 3.30 & & & & & 0.50 & 0.50 & 0.08 & 0.10\
26 &RU & 3.00 & 3.00 & & & & & & & & & &\
27 &IN & & & 1.99 & 2.40 & & & & & 0.39 & 0.46 & 0.39 & 0.46\
28 &CN & 2.09 & 2.99 & 2.39 & 2.99 & 1.00 & 1.99 & & & & & &\
29 &SG & 3.00 & 3.00 & 3.00 & 3.00 & & & & & & & &\
30 &DE & 1.93 & 4.00 & & & & & & & 0.45 & 0.50 & 0.06 & 0.14\
31 &US & & & 1.00 & 2.00 & & & & & & & 0.50 & 1.00\
32 &IN & & & 0.50 & 0.50 & & & & & & & &\
33 &AE & & & 0.90 & 1.00 & & & 0.75 & 0.80 & & & 0.15 & 0.40\
34 &IN & 2.00 & 2.00 & 2.00 & 2.00 & **2.00** & **2.00** & **2.00** & **2.00** & **1.60** & **1.67** & **1.60** & **1.67**\
NUM & & 17 & 17 & 32 & 32 & 2 & 2 & 10 & 10 & 12 & 12 & 23 & 23\
AVG & & 3.41 & 4.24 & 1.73 & 2.14 & 1.50 & 2.00 & 0.76 & 0.83 & 0.48 & 0.55 & 0.33 & 0.40\
SD & & 1.85 & 2.21 & 1.24 & 1.70 & 0.71 & 0.01 & 0.64 & 0.71 & 0.38 & 0.40 & 0.45 & 0.50\
Offers and Pricing Models
-------------------------
To increase app downloads and sales, paid review providers offer fake reviews, ratings, and installs. Table \[tab:paidprovider\] shows the prices of these offers for both the Android and iOS platform. The table lists the offers’ minimum and maximum price, which varies, e.g., depending on the amount of reviews bought.
**Paid fake reviews** are offered by 17 of 34 (50%) providers for iOS and by 32 of 34 (94.1%) for Android. Reviews always include a rating. Among all offers, reviews are the most expensive. The price of a review for iOS is, on average, between \$3.41 and \$4.24 with a standard deviation from 1.85 to 2.21. The price of a review for Android is less expensive, on average, between \$1.73 and \$2.14 with a standard deviation from 1.73 to 2.14. The price for iOS is $\sim$97 to 98% higher.
**Paid fake ratings** are offered by two of 34 (5.9%) providers for iOS and by 10 of 34 (29.4%) for Android. With this offer fake reviewers rate the app without submitting a written review. The price of a rating for iOS is on average between \$1.50 and \$2.00 with a standard deviation from 0.01 to 0.71. The price of a rating for Android is on average between \$0.76 to \$0.83 with a standard deviation of 0.76 to 0.83. Compared to Android the price for iOS ratings is $\sim$97 to 140% higher.
**Paid installs** are generated, e.g., by advertising the app on blogs. Also, users can be paid to install the app. The acquired new app users decide by themselves to rate and review the app. According to our definition these reviews are not considered as fake, as these are not directly requested or paid for. Installs are offered by 12 of 34 (35.4%) providers for iOS and by 23 of 34 (67.6%) for Android. Among all offers installs are the least expensive. The price of an iOS app install is between \$0.48 to \$0.55 with a standard deviation of 0.38 to 0.40. For Android the price is between \$0.33 to \$0.40 with a standard deviation of 0.45 to 0.50. Comparing both platforms the price difference is $\sim$37 to 45%.
Overall, the offers are rather expensive, e.g., compared to an average crowdsourcing task or buying followers on Twitter. 10,000 Twitter followers can be bought for a price of \$4 [@Stringhini:2013:FGG:2504730.2504731]. This might depend on the fact that fake ratings and reviews are generated manually, e.g., due to a strict moderation by app store operators, while Twitter followers can be generated automatically.
We further tried to identify the popularity of the three different types of offers. Unfortunately, we were only able to extract usage numbers from paid review provider 10 (PRP10) and thus cannot provide generalizable information. Overall, this provider sold 354,000 offers, of which 20,750 (5.9%) were fake reviews, 29,150 (8.2%) fake ratings, and 304,100 (85.9%) paid installs. We cannot give any numbers on how many paid installs result into a rating or review, and if these are comparable to fake ratings and reviews.
Pretended Fake Review Characteristics
-------------------------------------
To understand the rules and conditions of providing fake reviews, we conducted a disguised questionnaire with the paid review providers. We also extracted the policies, with which submitted reviews must comply in review exchange portals.
### Disguised Questionnaire
The disguised questionnaire consists of eleven questions and was presented to providers in a request for buying fake reviews. A sample question is shown below:
*We have several competitors which gain more and more market share. For this reason we are looking for both positive and negative reviews, positive for our apps and negative for our competitors’ apps. \[...\]*
We decided against open questions as we noticed during a pre-run of the questionnaire, conducted using different identities, that providers returned incomplete answers. The questionnaire is included in our replication package.
Eleven out of 34 paid review providers (32.3%) answered our questionnaire. Table \[tab:survey\] summarizes their answers. Even upon request, not all providers answered all of our questions. Therefore, the total answers refer to the number of providers that explicitly answered the specific question.
[lXXXXXX]{} PRP & Positive Ratings & Negative Ratings & Custom Keywords & Predefined Reviews & Real Users & Guarantee\
9 & Yes & No & Yes & Yes & Yes & Yes\
10 & Yes & Yes & & & Yes & No\
12 & Yes & Yes & Yes & No & Yes & Yes\
15 & Yes & Yes & & & & No\
16 & Yes & No & Yes & Yes & Yes & No\
22 & Yes & Yes & & & Yes &\
23 & Yes & No & No & Yes & Yes & No\
25 & Yes & Yes & Yes & Yes & Yes & Yes\
26 & Yes & Yes & Yes & Yes & Yes & Yes\
28 & Yes & No & No & Yes & Yes & No\
29 & Yes & & Yes & Yes & Yes & Yes\
While all 11 providers offer positive ratings and reviews, 6 also offer negative, e.g., to lower the reputation of competing apps. Regarding the content, 6 of 8 providers reported to accept keywords, which will be included in their reviews. Seven of 8 providers accept predefined reviews to be submitted by their reviewers. All providers state their reviews are written by humans and not generated using algorithms. Five of 10 providers gave a guarantee to replace deleted fake reviews.
Regarding the geographical origin of fake reviews, PRP10 and PRP15 provide reviews from the US. PRP23 and PRP26 additionally provide reviews from Russia. P25 also provides reviews from India. PRP28 specified 13 countries from which the reviews are submitted, these are Austria, Canada, China, France, Germany, India, Italy, Japan, Russia, Taiwan, United Kingdom, United States, and Vietnam. Four providers (PRP9, PRP12, PRP16, and PRP29) reported to submit reviews from all over the world. According to the results, the top three countries reviews are provided from are: United States (54.5%), Russia (36.4%), and India (18.2%). Regarding the language, PRP9 and PRP29 reported to provide reviews in all languages. PRP10, PRP12, PRP15, PRP25, and PRP28 only provide reviews in English. PRP23 and PRP26 also provide reviews in Russian language. By analyzing all 60,431 fake reviews, initially collected, using LangID, we found that these are written in 70 languages. The five most common are English (87.7%), French (2.3%), German (2.2%), Italian (1.3%), and Spanish (1%).
### Review Policies
For all review exchange portals, we were able to extract policies that state the requirements submitted fake reviews have to confirm to, see Table \[tab:guidelines\]. The policies have different levels of details. Thus, not every requirement is stated by each policy. With the total number we refer to the policies that explicitly state a requirement.
[llXXllXXlX]{} REP & Co. & Real Dev. & Install App & Use App & Keep App & Honest & Rating & Length & Copy\
1 & IN & & Yes & No & & & 1-5 & &\
2 & ES & Yes & Yes & & 5 days & Yes & 3-5 & $>$10 words & No\
3 & US & Yes & Yes & & 1-2 days & Yes & & 2-3 sentences &\
4 & US & & & & & & & 1-2 sentences & No\
5 & GB & Yes & Yes & & 1 day & Yes & 4-5 & 1-2 sentences &\
6 & CN & Yes & Yes & 4 min & 5 days & & & &\
7 & GB & & & & & Yes & & 1-2 sentences &\
8 & SE & Yes & Yes & & 2 days & Yes & 3-5 & $>$10 words & No\
9 & RU & & Yes & 10 min & 7 days & & & &\
Five portals require to use a real device to submit a review. The installation of the app is explicitly required by seven portals. Only two portals request the reviewers to use the app before submitting a review. REP6 requires the reviewer to use the app for at least 4 minutes and REP9 for at least 10 minutes. REP1 explicitly states that the usage of the app is not required. Six portals state that the app should be kept on the phone for a specific amount of time after leaving the review. The minimum amount of time is one (REP3, REP5) up to 7 days (REP9).
Regarding the rating, four providers specify a range the rating should follow. REP2 and REP8 request 3-5 stars, and REP5 4-5 stars. REP1 is the only provider explicitly allowing positive and negative (1-5 stars) ratings. However, reviews from REP1 are not included in our fake reviews dataset. Five providers state that the review should be honest, although three of those allow only positive ratings with at least 3-4 stars. These providers state that reviewers should skip apps if they are unable to submit a positive review.
Regarding the review length, six portals make a statement: two require at least 10 words, three portals 1-2 sentences, and one portal 2-3 sentences. Three providers explicitly state that the review should not contain content copied from the app description. REP5 and REP9 additionally require that the reviews are sufficiently detailed, e.g., “should describe app features instead of providing only praise”.
REP2 allows reviewers to only rate up to 10 apps per day. Further, the app should not be immediately reviewed after its installation. Reviewers should randomly download apps from the app store without leaving a review. Before leaving another review the reviewer should wait a few minutes. Last, reviewers should not only provide 5-star ratings, but vary between 3-5 star ratings. REP7 requires the ratings to match review content. Finally, REP9 requires to launch reviewed apps periodically within the next 7 days. A possible reason for this, might be to hide the suspicious behavior of quitting to use an app after providing a positive review from app store operators.
### Initial Fake Review Indicators
Considering the questionnaire results, the review policies, and taking into account the efforts of providers to disguise fake reviews, we can hypothesize that fake reviews are highly diverse. For example, the *rating* of fake reviews can be positive or negative. The *length* of a review can also vary. In addition, the quality of the *content* may strongly differ. Overall, fake reviews do not mean short and low quality reviews, as our initial results reveal. These reviews could be either written by paid reviewers (whether or not they have to use specific keywords), or they can be predefined reviews that are written by the app developers and that have to be published by the reviewers.
Fake Review Characteristics (RQ2) {#sec:char}
=================================
We investigate apps affected by fake reviews, reviewers providing fake reviews, and fake reviews themselves. In particular, we study the differences of fake reviews to the reviews from the official reviews dataset.
Apps
----
We identified 3,072 apps requesting fake reviews. As these apps could have, e.g., been entered on review exchange portals for testing purposes either or not by their developers, we only consider apps that received fake reviews to strengthen our results. Overall, we analyzed 1,929 (62.8%) of the identified apps.
**Most apps with fake reviews fall into in the category games, nearly twice as much as regular apps.** Table \[fig:appscategory\] lists the 25 app categories of the Apple App Store. It compares apps from the fake reviews and the official reviews dataset per category. The table depicts each categories’ rank, number and percentage of apps, and percentage of reviews within the datasets. The highest rank is assigned to the category with the most apps included.
We found that more than half of the apps with fake reviews (53%) belong to the category “Games”, followed by the categories “Photo & Video” (5.8%), “Education” (4.8%), “Entertainment” (4.5%), and “Health & Fitness” (4.4%). The categories with least apps with fake reviews are “Stickers” (0.05%), “News” (0.1%), “Catalogs” (0.16%), “Newsstand” (0.16%), and “Books” (0.21%).
Between both datasets, we found a strong, positive correlation for the distribution of apps over the categories. We compared the category ranks using the Spearman rank correlation coefficient (*r$_s$* = 0.74, two-tailed p-value = 0.00002). The coefficient is used to measure the rank correlation, i.e., the statistical dependence between the rankings of two variables.
To identify categories with the highest difference between both datasets, we calculated the rank difference using the following formula.
$$r\textsubscript{diff}=|rank\textsubscript{official}-rank\textsubscript{fake}|$$
The highest differences exis for the categories “Social Networking” (*r~diff~=11*), “Photo & Video” (*r~diff~=10*), “Business” (*r~diff~=9*), and “Shopping” (*r~diff~=9*). The 3rd category “Business” within the app store is, for example, only ranked 12th in the fake reviews dataset. Vice versa, the category “Photo & Video” contains more apps in the fake reviews dataset.\
[lrrrrrr]{} Category & &\
& Rank & Apps & Reviews & Rank & Apps & Reviews\
Books & 21 & 4 (0.21%) & 0.06% & 19 & 25069 (1.75%) & 0.89%\
Business & 12 & 33 (1.71%) & 1.72% & 3 & 130825 (9.15%) & 1.35%\
Catalogs & 23 & 3 (0.16%) & 0.09% & 23 & 10951 (0.77%) & 0.29%\
Education & 3 & 92 (4.77%) & 3.80% & 2 & 131302 (9.18%) & 1.74%\
Entertainment & 4 & 87 (4.51%) & 4.03% & 5 & 79504 (5.56%) & 5.68%\
Finance & 17 & 17 (0.88%) & 1.11% & 14 & 34684 (2.43%) & 1.66%\
Food & Drink & 15 & 19 (0.98%) & 0.65% & 9 & 50944 (3.56%) & 1.34%\
Games & 1 & 1023 (53.03%) & 47.57% & 1 & 326864 (22.86%) & 49.95%\
Health & Fitn. & 5 & 85 (4.41%) & 5.81% & 8 & 54410 (3.80%) & 3.23%\
Lifestyle & 8 & 69 (3.58%) & 4.82% & 4 & 102183 (7.15%) & 2.46%\
Medical & 20 & 13 (0.67%) & 0.65% & 16 & 30101 (2.10%) & 0.42%\
Music & 11 & 36 (1.87%) & 1.99% & 11 & 43874 (3.07%) & 2.72%\
Navigation & 20 & 13 (0.67%) & 1.06% & 20 & 21559 (1.51%) & 0.80%\
News & 24 & 2 (0.10%) & 0.06% & 18 & 26358 (1.84%) & 1.64%\
Newsstand & 23 & 3 (0.16%) & 0.18% & 25 & 1021 (0.07%) & 0.00%\
Photo & Video & 2 & 112 (5.81%) & 7.66% & 12 & 40034 (2.80%) & 6.54%\
Productivity & 10 & 42 (2.18%) & 1.92% & 10 & 44191 (3.09%) & 3.35%\
Reference & 18 & 14 (0.73%) & 0.63% & 15 & 34465 (2.41%) & 1.50%\
Shopping & 13 & 25 (1.30%) & 2.50% & 22 & 15253 (1.07%) & 2.15%\
Social Netw. & 6 & 82 (4.25%) & 3.69% & 17 & 27488 (1.92%) & 5.41%\
Sports & 9 & 45 (2.33%) & 1.79% & 13 & 37060 (2.59%) & 1.14%\
Stickers & 25 & 1 (0.05%) & 0.01% & 21 & 20979 (1.47%) & 0.01%\
Travel & 14 & 22 (1.14%) & 2.08% & 7 & 64846 (4.53%) & 1.30%\
Utilities & 8 & 69 (3.58%) & 4.69% & 6 & 71680 (5.01%) & 3.61%\
Weather & 16 & 18 (0.93%) & 1.43% & 24 & 4446 (0.31%) & 0.82%\
& & $\sum$ = 1,929 & & & $\sum$ = 1,430,091 &\
**Apps with fake reviews are on average three times less offered as paid, compared to regular apps.** Developers might invest a lot of money buying fake reviews. Therefore, we analyzed the monetization of apps in the fake reviews dataset. We focused on the app price and in-app purchases. Regarding the app price, we found that 1,799 apps (93.3%) are offered for free, while 130 apps (6.7%) are paid. In comparison, the app store includes 1,167,377 (81.6%) free and 262,714 (18.4%) paid apps. The mean price of an app is \$2.16 with a standard deviation of 2.5. For the app store the mean price is \$4.07 with a standard deviation of 16.7. This difference between the mean prices is statistically significant (two-sample t-test, p$<$0.001, CI=0.99). However, the magnitude between the differences is slightly below small, found by calculating the effect size (d=-0.160) [@cohen1988spa]. Of the paid apps, 62 apps (47.7%, cf. 39.6% in app store) are offered for \$0.99, 39 apps (30%, cf. 18.9% in app store) for \$1.99, and 16 apps (12.3%, cf. 16.2% in app store) for \$2.99. The remaining apps (10%) cost between \$3.99-\$24.99. In the app store the price of the remaining apps (25.4%) is between \$3.99-\$999.99.
In-app purchases are offered by 759 fake-reviewed apps (39.4%). These apps contain 3,845 in-app offers, of which 3,186 are in-app purchases. On average each app includes around 4.2 in-app purchases with an average price of \$10.46. 26.33% of the in-app purchases are offered for \$0.99, 26.9% are in the range of \$1.99-\$2.99, 26.6% in the range of \$3.99-\$9.99, and 20.2% in the range of \$10.99-\$399.99. We were unable to automatically crawl in-app purchases, therefore we cannot compare the figures from the fake reviews to the official reviews dataset.
Also, apps could be further monetized through advertisements. We were unable to study this aspect, since no publicly available data on the number of advertisement impressions and revenue generated per impression exists.
**Most apps targeted by fake reviews have 2-9 reviews, which is the case for 42.2% of all apps.** We analyzed the total number of reviews for apps affected by fake reviews and apps from the official reviews dataset. Figure \[fig:appsratio\] groups both types of apps into given ranges of fake and official reviews. The result indicates that fake reviews are not necessarily limited to a small and specific group of apps, but could be distributed across the majority of apps.
![Distribution of fake and official reviews over ranges per app.[]{data-label="fig:appsratio"}](04-apps-reviewratio){width=".85\columnwidth"}
**Only less than 7% of apps affected by fake reviews were removed from the Apple App Store.** We studied whether apps with a high percentage of fake reviews rather get removed from the app store, compared to apps with less fake reviews. Therefore, we crawled the apps affected by fake reviews again after three months in June 2017. Of the 1,929 apps, 131 (6.8%) were no longer available on the app store. Most of the deleted apps (68%) belong to the category “Games”, 5% to “Entertainment”, and 5% to “Utilities”. Since no justification is provided by the app store operators, there are two possible reasons for this: Either, the apps have been removed by their own developers, or the app store operators have removed the app due to fake reviews or other compliance reasons, e.g., spam apps [@Seneviratne:2017:SMA:3062397.3007901]. Figure \[fig:delapps\] shows two plots of deleted and non-deleted apps and their percentage of fake reviews.
The upper plot considers all apps affected by fake reviews. We found that deleted apps received 27.5% fake (median: 20%) and 72.5% official reviews. Non-deleted apps received 38.1% fake (median: 30.8%) and 61.9% official reviews. By analyzing the median, we found that non-deleted apps receive 12 reviews (cf. 15 reviews for deleted apps) of which 2 are fake (cf. 2 reviews for deleted apps). A $\chi^2$-test showed that being no longer available on the app store and the percentage of fake reviews are independent ($\chi^2$=2.0906, p=0.1482).
As this gives the impression that the amount of fake reviews does not impact being removed from the app store, we further analyzed apps with at least ten fake reviews only, see lower plot. This applies for 181 apps, of which 11 were deleted. For these, the median of fake reviews for deleted apps is 63.5%. For non-deleted apps the median is 37.1%. Based on medians, deleted apps receive 51 reviews of which 22 are fake. Non-deleted apps receive 49.5 reviews of which 15 are fake. For these apps, a $\chi^2$-test showed both values are no longer independent ($\chi^2$=6.8708, p=0.008762), compared to considering all apps with at least one fake review.
![Ratio of fake reviews for non-/deleted apps.[]{data-label="fig:delapps"}](04-apps-deletedboxplot){width=".85\columnwidth"}
Reviewers
---------
**Fake reviewers submit about 30 reviews on average — 12 times more than regular reviewers.** We identified 721 users providing fake reviews. These fake reviewers provide 29.9 reviews per user, on average, compared to 2.5 reviews per reviewer in the official reviews dataset. This difference is statistically significant (two-sample t-test, p$<$0.001, CI=0.99) and the effect size is large (d=0.802). Overall, these users provided 21,581 reviews in total for 8,429 different apps. Surprisingly, fake reviewers do not seem to use several accounts to hide their activities.
**More than 50% of the reviewers in the official dataset provide only a single review.** The total number of reviews given per fake reviewer varies between 1 and 573. For reviewers within the official reviews dataset this is between 1 and 913. Figure \[fig:reviewersgroups\] groups both fake and regular reviewers according to their number of submitted reviews.
Exactly one review was given by 5.4% of the fake reviewers, compared to 53.1% for regular reviewers. 2-5 reviews were provided by 15.8% fake reviewers (cf. 35.6%), 6-10 reviews by 20.8% (cf. 9.3%), 11-50 reviews by 40.8% (cf. 1.9%), 51-100 reviews by 11.8% (cf. 0.02%), and more than 100 reviews by 5.4% (cf. 0.006%). The highest percentage of fake reviewers (40.8%) is in the range of 11-50 reviews, while most regular reviewers (53.1%) provide a single review.
![Number of reviews provided per fake or official reviewer.[]{data-label="fig:reviewersgroups"}](04-reviewers-reviewratio){width=".85\columnwidth"}
**Fake reviewers review about 4 times more frequently than regular reviewers.** Fake reviewers are more active compared to others. They have a frequency of one review per 78.8 days, compared to 328.9 days for regular reviewers. The difference is statistically significant with 250.1 days, i.e., 417.2% (two-sample t-test, p$<$0.001, CI=0.99). The effect size is large (d=-0.955).
**The lifetime of fake reviewer accounts is nearly twice as long as regular users.** The account lifetime, i.e., the time difference between the first and last review provided, is 622.3 days for fake reviewers, compared to 331.3 days for other app store users. The difference between fake and regular reviewers is 291 days (187.9%) and statistically significant using the previous test (p$<$0.001, CI=0.99). The effect size is near medium (d=0.464). This shows that the accounts of fake reviewers remain undetected in app stores for several years.
Reviews
-------
Although we found 60,431 fake reviews, in the following we only consider the 8,607 fake reviews that we identified and still exist in the Apple App Store. These reviews have not been filtered by mechanism of the app store operators and could impact app developers and users.
**The distribution between ratings of fake and official reviews varies most for 1-star reviews.** Figure \[fig:ratingreviews\] compares the distribution of ratings for reviews from the fake and official reviews dataset. 70% of the fake reviews are rated with 5 stars compared to 65% for official reviews. 23% of fake reviews are rated with 4 stars (cf. 16%), 5% with 3 stars (cf. 6%), 1% with 2 stars (cf. 4%), and 0.6% with 1 star (cf. 10%). Overall, ratings are very positive in both datasets. The greatest difference between fake and official reviews can be observed by the percentage of 1-star ratings. We have evidence that fake review providers explicitly ask their reviewers within reviewing policies to not only provide 5-stars reviews but also 4-stars and even 3-stars reviews. This might result in rather small differences between fake and official reviews regarding extremely positive ratings. Thereby, the suspicious behavior of writing, and also receiving, only 5-stars reviews should be hidden as this could possibly result in the deletion fake reviews by app store operators and in worst case the removal of the affected app from the app store (cf. Section 3.3.2).
![Distribution of star ratings between official and fake reviews.[]{data-label="fig:ratingreviews"}](04-reviews-distribution){width=".85\columnwidth"}
**Compared to official reviews, short reviews are rather uncommon in fake reviews.** The length of a fake review (consisting of title and body) is, on average, 121.3 characters. Official reviews have a length of 110.8 characters, on average – resulting in a difference of 10.5 characters. Considering the median, fake reviews consist of 111 characters while official reviews consist of 63 characters, see Figure \[fig:reviewslength\]. The difference regarding the median is 48 characters.
We further analyzed the number of words per review. Fake reviews have, on average, 22.9 words, with a median of 21 words. Official reviews have 21.3 words, with a median of 12 words. Regarding the amount of average words, the difference is relatively small with 1.7 words. Considering the median the difference is 9 words. A typical fake review is given below.
Great for expense tracking $\star\star\star\star\star$\
*Does a great job for expense tracking. Nice interface and color scheme. Definitely recommend!*
We found that rather short reviews, which constitute a major part of the official reviews, are uncommon for fake reviews (see example below).
Fantastic $\star\star\star\star\star$\
*Great game, my son loves it. Lots of fun.*
We initially assumed fake reviews to be short. However, according to the dataset fake reviews are significantly longer regarding the number of characters and words (Wilcoxon rank sum test, p$<$0.001, CI=0.99). The effect size [@fritz2012effect] however is near zero (r=0.001).
![Distribution of review length (in characters) between regular and fake reviews.[]{data-label="fig:reviewslength"}](04-reviews-length){width=".9\columnwidth"}
**Fake reviews are rated more often helpful compared to official reviews.** In app stores users can rate the helpfulness of reviews through votes. 132 of the 8,607 fake reviews (1.5%) received at least one vote, compared to 2.7% for reviews in the app store. Overall, the reviews received 270 votes. 245 are votes (90.7%) rating the reviews as helpful and 25 votes (9.3%) rate the reviews not helpful. In the app store less helpful votes (67.8%) exist. Both, the number of reviews with votes and the overall number of helpful votes are significantly different (two-sample t-test, p$<$0.001, CI=0.99). Also in this case, the effect size is near zero (d=-0.018).
![Percentage of reviews received per day after app release (day 0 is release of app update).[]{data-label="fig:reviewsdate"}](04-reviews-update){width=".9\columnwidth"}
**After releasing updates, apps affected by fake reviews do not immediately receive more reviews.** We analyzed the relative reviewing frequency by summing up all reviews per day after the apps’ last releases. Figure \[fig:reviewsdate\] shows the percentage of received reviews per day for apps affected by fake reviews and regular apps over a time span of three weeks. After three weeks the amount of received reviews stabilized. We choose only the apps’ last release as we were unable to automatically crawl release dates.
For regular apps most reviews are given on the day of the app release [@Pagano:2013jn]. For apps affected by fake reviews there is only a small peak on the app release day. For these apps the percentage of reviews provided increases on a daily basis, until it decreases on day 11 after the app update. One reason might be that for apps affected by fake reviews no large user basis exist that could, intrinsically motivated, provide reviews. Developers of these apps have to buy fake reviews to promote their updates. The distribution of request for providing fake reviews to the actual reviewers might take time. Compared to that, regular apps with a user basis that matches the amount of reviews received, have enough users that spontaneously provide their feedback after installing the app update.
**Fake reviews include more positives adjectives and less negative words related to software engineering such as “fix” or “crash”.** We analyzed the review content by comparing the 100 most-common words of fake and official reviews. We extracted the lists of most-common words in five steps. First, we removed the punctuation. We transformed all words into lowercase writing. Then, we tokenized the words of each review. We removed stopwords. Last, we counted the occurrences of each word.
Both lists have 63 words in common and 37 unique words. We sorted the lists descending by the occurrences of words. Afterwards, for each word the lists have in common, we calculated the difference between their positions in the lists. The word “simple” is, e.g., on position 14 for fake reviews and on position 97 for official reviews, resulting in a rank of -83. Therefore, negative ranks denote words that are more common for fake reviews. We plotted word ranks in Figure \[fig:reviewswords\].
The top five words that are more common for fake reviews are “simple", “super", “little", “recommend", and “well". The top five words that are less common for fake reviews are “even", “can’t", “don’t", “want", and “free".
![Delta between occurrences of most-common words in official/fake reviews, a negative ranking indicates that the specific word is more common for fake reviews, a positive ranking denotes that the word is less common for fake reviews.[]{data-label="fig:reviewswords"}](04-reviews-words){width="\columnwidth"}
Afterwards, we analyzed the distinct words in both lists. For official reviews the distinct five most common words are (in order): “update”, “ever”, “please”, “fix”, “every”. Also, words possibly related to the functionality of the apps, such as “doesn’t”, “crashes”, “wish”, and “bad” are included. The five most common distinct words for fake reviews are: “graphics”, “useful”, “idea”, “ads”, and “kids”. Also positive words, such as “interesting”, “perfect”, “helpful”, “recommended”, “funny”, and “learn” are popular.
Last, we compared the most common bi-grams for fake and official reviews. As for most common words, again 63 matches exist. We observed that bi-grams possibly pointing to bug reports, such as “please fix”, only exist in the official reviews dataset. Bi-grams indicating feature requests, such as “would like” or “wish could”, exist in both datasets. Negative bi-grams, such as “waste time” or “keeps crashing”, again only exist within the official reviews dataset.
Fake Review Detection (RQ3) {#sec:detection}
===========================
We build a supervised binary classifier to classify reviews as fake or not. Figure \[fig:classificationmethod\] shows the three phases conducted after feature extraction. We begin by preprocessing the data. Then, we compare the results of different classification algorithms. We optimize the algorithms by feature selection and hyperparameter tuning. Last, we evaluate the importance of the classification features. Afterwards, we conduct an in-the-wild experiment to evaluate how our classifier performs in practice, i.e., on imbalanced data.
![Overview of fake review classification.[]{data-label="fig:classificationmethod"}](05-classification){width="\columnwidth"}
Feature Extraction
------------------
We extracted a balanced truthset of 16,000 reviews. Of these reviews, 8,000 are randomly selected fake and 8,000 are randomly selected official reviews. Per review, the truthset includes a vector containing 15 numerical features and a label, which is either “real” or “fake”. The features were selected based on the differences we identified between fake and official reviews, and by experimentation. We decided **not** to use the textual review itself, e.g., in form of TF/IDF representation, for several reasons. Mukherjee et al. [@Mukherjee2013WhatYF] analyzed fake reviews published on Yelp and found that the word distribution of fake reviews does not significantly differ from official reviews. As a result, their text-based classifier only achieved an accuracy of 67.8%. Vice versa, the word distribution within fake reviews could also highly differ. According to the questionnaire with paid review providers, custom keywords can be included in the reviews or predefined reviews can be submitted by the fake reviewers. Finally, when using text, training data would be required for every language, which is difficult to collect.
In contrast, Ferrara et al. [@DBLP:journals/corr/FerraraVDMF14] use non-textual features related to the user, such as the account creation time or total number of followers. By using such features their classifier to detect bots in social networks that, e.g., influence political discussions, achieved better results. Other researchers [@6921650; @Lee:2010:USS:1835449.1835522; @feng2012distributional; @Park:2016:SCM:2858036.2858389] followed this approach and were also able to improve their classification results.
We therefore focused on features that relate to the context of the fake review, i.e., the reviewer and app. Table \[tab:featuresselected\] lists all selected features. For the reviewer we selected four features: the total number of reviews provided, the percentage of reviews per star rating (e.g., the reviewer could have provided 70% of all reviews with a 5-star rating and 30% with a 1-star rating), the review frequency (i.e., the average time in seconds between all reviews provided), and the account usage (which is the lifetime of the reviewers account, i.e., the timespan between the first and the last review provided in seconds). For the app we selected two features: the total number of reviews received for all app versions and the percentage of reviews received per star rating. Finally feature for the review, we selected the length, i.e. the characters count.
[llXXl]{} Category & Name & Type & Null-Values & Example\
Reviewer & \# Reviews (Total) & Int & 0 & 100\
& % Reviews (per Star-Rating) & [\[]{}Float[\]]{} & 0 & [\[]{}0.7, 0.0, 0.0, 0.0, 0.3[\]]{}\
& Review Frequency (in Seconds) & Int & 1,734 & 100\
& Account Usage (in Seconds) & Int & 0 & 600\
App & \# Reviews (Total) & Int & 0 & 100\
& \# Reviews (per Star-Rating) & [\[]{}Float[\]]{} & 0 & [\[]{}0.2, 0.2, 0.2, 0.2, 0.2[\]]{}\
Review & Length (in Characters) & Int & 0 & 100\
Data Preprocessing
------------------
We preprocessed the data in three steps. We began by performing data cleaning, i.e., filling null values instead of removing affected columns. Of the selected features only a single column includes null values, see Table \[tab:featuresselected\]. The review frequency is in 1,734 cases undefined because only a single review was provided by the reviewer. In this case, we set the frequency to lifetime of the app store, which is 9 years.
Then, we normalized the dataset so that individual samples to have unit norm. We used the `normalize()` method with standard parameters of the `preprocessing` module provided by scikit-learn [@scikit-learn].
Last, we standardized the dataset so that the individual features are standard normally distributed, i.e., gaussian with zero mean and unit variance. This is a common requirement for many classification algorithms, such as the radial basis function (RBF) kernel of support vector machines. If not standardizing the data, features with a much higher variance compared to others might dominate the objective function. As a result, the classification algorithm is unable to learn from other features [@Weblink:SciKit:10001]. We used the `scale()` method with standard parameters of the `preprocessing` module.
Classification with Balanced Data
---------------------------------
We compare seven supervised machine learning approaches to classify reviews as fake or not. We use the implementations provided by the scikit-learn [@scikit-learn] library. Supervised approaches need to be trained using a labeled truthset before they can be applied. This truthset is split into a training and testing set. The training set is used by the algorithms to build a model on which unseen instances are classified. In the test phase, the classifier performs a binary classification and decides whether reviews within the test set are fake or not.
To get more reliable measures of the model quality, we apply cross validation on our truthset. This is performed in several folds, i.e., splits of the data, called k-fold cross validation. In this paper we perform 10 folds. Per fold, a randomly selected amount of 1/k of the overall data is held out of the training as a test set for the evaluation. The final performance is the average of the scores computed in all folds. Using cross validation, we also avoid bias which would otherwise be introduced by using only a random train/test split. In addition, although our truthset is balanced, we apply stratification. Stratification ensures that each split contains a balanced amount of fake and official reviews. We repeat the cross validation 30 times per classification algorithm with different seeds. We use the `RepeatedStratifiedKFold` method of the `model_selection` module.
[lXXXXX]{} Classifier & Accuracy & Precision & Recall & F1 & AUC/ROC\
RandomForestClassifier & 0.970 & 0.973 & 0.967 & 0.970 & 0.989\
DecisionTreeClassifier & 0.953 & 0.949 & 0.957 & 0.953 & 0.953\
MLPClassifier & 0.919 & 0.921 & 0.916 & 0.918 & 0.969\
SVC(kernel=’rbf’) & 0.901 & 0.879 & 0.930 & 0.904 & 0.959\
SVC(kernel=’linear’) & 0.899 & 0.878 & 0.926 & 0.902 & 0.960\
LinearSVC & 0.895 & 0.861 & 0.941 & 0.900 & 0.964\
GaussianNB & 0.765 & 0.731 & 0.889 & 0.755 & 0.955\
The seven classification algorithms we compare are the following: Naive Bayes (GaussianNB) is a popular algorithm for binary classification [@Bird:2009:NLP:1717171], which is based on the Bayes theorem with strong independence assumptions between features. Compared to other classifiers it does not require a large training set. Random Forest (RF) [@Ho:1995:RDF:844379.844681] is an ensemble learning method for classification and other tasks. It can build multiple trees in randomly selected subspaces of the feature space. Decision Tree (DT) [@Torgo:2010:DMR:1951702] assumes that all features have finite discrete domains and that there is a single target feature representing the classification (i.e., the tree leaves). Support Vector Machine (SVM) [@Cortes1995] represents the training data as points in space. It creates support vectors for gaps between classes in the space. The test data is classified based on which side of the gap its instances fall. The Gaussian radial basis function (rbf) is used for non-linear classification by applying the kernel trick [@10016858207]. Linear support vector classification (LinearSVC) penalizes the intercept, in comparison to SVM. Multilayer perceptron (MLP) is an artificial neural network which consists of at least three layers of nodes. MLP utilizes the supervised learning technique backpropagation for training.
Table \[tab:classifiercompare\] shows the results of the seven classification algorithms, each with default configuration. The results include accuracy, precision, recall, F1-score, and area under the ROC curve (AUC) value. Among all, the random forest algorithm achieved the best scores.
Optimization
------------
We optimize the classifiers by performing feature selection and hyperparameter tuning. We optimize for **precision only**. In comparison to fake reviews, for regular reviews we were unable to create a gold-standard dataset. To create a gold-standard dataset for regular non-fake reviews, all fake reviews must be identified and removed from the official reviews dataset. This is practically in-feasible as there is currently no measure to ensure that a review is not fake.
Hence we do not know all fake reviews, we can only report on how many of the *known* fake reviews are classified as fake (precision) and not on how many of *all existing* fake reviews were classified as fake (recall). Resultant measures, such as the F1-score, are reported for completeness.
To select features, we apply recursive feature elimination with cross validation. After the classification algorithm assigned a weight to each feature, these are eliminated recursively by considering smaller sets. Per iteration, the least important feature is removed to determine their optimal number. We use the `RFECV` method from the `feature_selection` module. The cross validation is performed as described in the previous phase.
We received the best result with the random forest algorithm using all features. Nearly similar accuracies are already possible with less features, e.g., the precision with three features is 0.969, compared to 0.973 using all features, see Figure \[fig:numfeaturesselected\]. The three selected features are the *1) total number of reviews the app received and 2) the user provided*, as well as the *3) frequency in which the user provides reviews*.
![Precision of random forest classifier by number of features selected.[]{data-label="fig:numfeaturesselected"}](05-classifier-numfeatures){width=".8\columnwidth"}
To tune the hyperparameters, we apply the grid search method `GridSearchCV` from the `model_selection` module. This method performs a cross validated, exhaustive search over a predefined grid of parameters for a classification algorithm. After finding the optimal combination of parameters within the grid, this is further manually tuned by adding more values around the currently best.
We achieved the best result using the random forest algorithm with the parameters *{’criterion’: ’gini’, ’max\_depth’: 30, ’max\_features’: ’sqrt’, ’n\_estimators’: 300}*. The parameter *criterion* measures the quality of a split. We used Gini impurity, which is intended for continuous attributes and faster to compute compared to Entropy. It is recommended to minimize the number of misclassifications. *max\_depth* defines the depth of the tree. *max\_features* sets the number of features to consider when looking for the best split. *n\_estimators* defines the number of trees in the forest.
Although performing hyperparameter tuning, the classifier’s precision equal the results using the default configuration. Only measures we do not consider, the recall and F1-score, were slightly improved resulting in 98% each.
Feature Importance
------------------
Last, we analyzed the relative importance of the extracted features with respect to the predictability of whether a given review is fake or not, called feature importance. The feature importance is calculated on how often a feature is used in the split points of a tree. More frequently used features are more important.
![Relative importance of extracted features to detect fake reviews.[]{data-label="fig:featureimportance"}](05-classifier-featureimportance){width=".8\columnwidth"}
Figure \[fig:featureimportance\] shows that the three most important feature are the total number of reviews an app received (30%), the total number of reviews a user provided (24%), and the frequency with that a user provides reviews (13%).
We assume the total number of reviews received by an app is the most important feature as apps with a specific amount of reviews, e.g., 2-9 reviews (cf. Section 4.1), are most often targeted by fake reviews. The total number of reviews a user provided as well as review frequency have a high importance as fake reviewers provide much more reviews than regular reviewers, with a higher frequency. The percentage of 1- and 2-star ratings an app received are important, as the difference between those star ratings provided are the highest when comparing apps with fake reviews and regular apps (cf. Figure \[fig:ratingreviews\]).
Classification with Imbalanced Data
-----------------------------------
In practice, fake and regular reviews are imbalanced. For app stores no reliable estimate on the distribution exists. Other domains, such as social media, mark 10% to 15% of their reviews as fake [@Weblink:10019]. The travel portal Yelp filters about 15% of their reviews as suspicious [@doi:10.1287/mnsc.2015.2304; @mukherjee2013fake]. This class imbalance can additionally be affected by numerous factors, such as the selected apps or time period. Free apps, for example, receive more fake reviews than paid apps (cf. Section 4.1). This reveals a skewed distribution of fake and regular reviews in app stores.
Research found that highly imbalanced data often results into poor performing classification models [@drummond2003c4; @chawla2004special; @saito2007large]. To have a more realistic setting of how our classifier can perform in practice, we conduct an in-the-wild experiment by varying the skewness of our dataset. We decided to vary the skewness on a logarithmic scale to depict the classification scores on finer granularities towards extremely imbalanced datasets with fake reviews as the minority class. We keep a fixed amount of 8,000 fake reviews and create 27 datasets including $10^2-10$ (90%) to $10^{-1}$ (0.1%) fake reviews. For a skew of 90% we used 889 regular reviews. With every change of the skewness we added additional regular reviews. All of the about 8 million used regular reviews were randomly selected at once from the official reviews dataset, so that the classification results are comparable.
![Classification scores of machine learning algorithms on imbalanced datasets, including 90% to 0.1% fake reviews, plotted on a logarithmic scale.[]{data-label="fig:classifierskewness"}](05-imbalance-all-v2){width=".95\columnwidth"}
Figure \[fig:classifierskewness\] shows an overview of the results of the seven supervised machine learning approaches studied in this work. Per classification algorithm and skewed dataset, we report on the precision, recall, F1-score, and AUC value.
[lXX]{} & Predicted as fake review & Predicted as regular review\
Actual fake review & True positive (TP) & False negative (FN)\
Actual regular review & False positive (FP) & True negative (TN)\
Since identifying fake reviews is a two-class problem, the performance metrics can be derived from the confusion matrix that is generated with every classification, see Table \[tab:confusionmatrix\]. We explain the reported performance metrics in the following: Precision ($\frac{TP}{TP + FP}$) measures the exactness, i.e., the number of correctly classified fake reviews to the overall number of reviews classified as fake. Recall ($\frac{TP}{TP + FN}$) measures the completeness, i.e., the number of correctly classified fake reviews to the overall number of fake reviews. The F1-score ($\frac{recall * precision}{precision + recall}$) is the harmonic mean between precision and recall. As improving precision and recall can be conflicting, it shows the trade-off between both. The AUC value measures the area under the ROC curve. It varies within the interval \[0, 1\]. The ROC curve itself depicts all possible trade-offs between TP rate, i.e., recall, and FP rate ($\frac{FP}{TN + FP}$). A better classifier produces an ROC curve closer to the top-left corner and therefore a higher AUC value [@8246564]. As we are focusing on skewed datasets within our in-the-wild experiment, we need to select performance metrics that are insensitive to class imbalance. This applies for all measures that use values from only one row of the confusion matrix [@8246564]. Precision and F1-score are sensitive to class imbalance and biased towards the majority class. Therefore, these metrics are inappropriate for our evaluation. Recall and AUC/ROC value are insensitive to class distribution, we use both measures to compare the performance of the classification algorithms within our in-the-wild experiment.
In the further evaluation, we include all classifiers that achieve a recall and AUC value higher than 0.5 for datasets including 90% to 1% fake reviews. This applies for the random forest (RF), decision tree (DT), and MLP algorithm. The remaining algorithms are excluded from the evaluation. The recall of the gaussian naive bayes (GaussianNB) algorithm decreases to nearly 0 for datasets with fake reviews as the majority class. Also, its precision is extremely low for datasets including less than 10% fake reviews. Similarly, the recall of the SVC algorithms decrease to nearly 0 for datasets with less than 10% fake reviews. In addition, the SVC(kernel=’linear’) implementation of sklearn return errors for skews below 5%, so do the remaining two SVC algorithms for skews of 0.1%.
### Classification Results with Imbalanced Data
Figure \[fig:classifierskewnessrel\] shows the performance measures of all three classification algorithms that remain within our evaluation. We chose to depict each measure as a single plot to more easily compare the algorithms. We report the precision and F1-score for reasons of completeness, although these are inappropriate measures for imbalanced data. Within the graphs, we highlight in gray the interval in which fake reviews typically occur in other domains. Further, we mark where the classes are equally distributed, i.e., there exist 50% fake reviews. At this point the algorithms perform well with all measures above 0.9 (cf. Table \[tab:scoresrel\]).
![Classification scores of appropriate machine learning algorithms for datasets with class imbalance, i.e., including 90% to 0.1% fake reviews, plotted on a logarithmic scale.[]{data-label="fig:classifierskewnessrel"}](05-imbalance-relevant){width="\columnwidth"}
[rXXXXXXXXXXXX]{}
Skew & & & &\
& DT & MLP & RF & DT & MLP & RF & DT & MLP & RF & DT & MLP & RF\
90.0 & 0.982 & 0.978 & 0.986 & 0.897 & 0.968 & 0.984 & 0.979 & 0.970 & 0.987 & 0.981 & 0.974 & 0.986\
80.0 & 0.972 & 0.963 & 0.980 & 0.923 & 0.970 & 0.987 & 0.968 & 0.958 & 0.984 & 0.970 & 0.960 & 0.982\
70.0 & 0.966 & 0.943 & 0.976 & 0.941 & 0.973 & 0.987 & 0.964 & 0.954 & 0.980 & 0.965 & 0.949 & 0.978\
60.0 & 0.964 & 0.940 & 0.970 & 0.946 & 0.974 & 0.988 & 0.952 & 0.940 & 0.976 & 0.958 & 0.940 & 0.973\
50.0 & 0.953 & 0.920 & 0.962 & 0.950 & 0.972 & 0.989 & 0.947 & 0.925 & 0.974 & 0.950 & 0.922 & 0.968\
40.0 & 0.945 & 0.902 & 0.956 & 0.953 & 0.972 & 0.989 & 0.941 & 0.914 & 0.972 & 0.943 & 0.908 & 0.964\
30.0 & 0.942 & 0.872 & 0.947 & 0.956 & 0.973 & 0.988 & 0.932 & 0.903 & 0.963 & 0.937 & 0.887 & 0.955\
20.0 & 0.924 & 0.842 & 0.937 & 0.950 & 0.974 & 0.988 & 0.903 & 0.885 & 0.950 & 0.914 & 0.863 & 0.944\
10.0 & 0.896 & 0.758 & 0.912 & 0.941 & 0.977 & 0.986 & 0.872 & 0.859 & 0.933 & 0.884 & 0.802 & 0.922\
9.0 & 0.894 & 0.824 & 0.907 & 0.940 & 0.978 & 0.986 & 0.865 & 0.826 & 0.925 & 0.879 & 0.824 & 0.916\
8.0 & 0.885 & 0.809 & 0.896 & 0.936 & 0.979 & 0.986 & 0.859 & 0.832 & 0.925 & 0.872 & 0.819 & 0.910\
7.0 & 0.884 & 0.785 & 0.892 & 0.936 & 0.979 & 0.984 & 0.855 & 0.830 & 0.920 & 0.869 & 0.807 & 0.906\
6.0 & 0.857 & 0.804 & 0.882 & 0.924 & 0.978 & 0.983 & 0.848 & 0.790 & 0.915 & 0.853 & 0.797 & 0.898\
5.0 & 0.860 & 0.694 & 0.876 & 0.925 & 0.979 & 0.984 & 0.826 & 0.820 & 0.904 & 0.843 & 0.752 & 0.890\
4.0 & 0.857 & 0.714 & 0.869 & 0.925 & 0.979 & 0.982 & 0.816 & 0.776 & 0.902 & 0.836 & 0.744 & 0.885\
3.0 & 0.830 & 0.681 & 0.846 & 0.912 & 0.980 & 0.980 & 0.789 & 0.741 & 0.898 & 0.809 & 0.710 & 0.871\
2.0 & 0.806 & 0.567 & 0.814 & 0.901 & 0.978 & 0.978 & 0.764 & 0.690 & 0.900 & 0.784 & 0.622 & 0.855\
1.0 & 0.775 & 0.395 & 0.756 & 0.886 & 0.978 & 0.971 & 0.723 & 0.589 & 0.900 & 0.748 & 0.473 & 0.822\
0.9 & 0.776 & 0.255 & 0.755 & 0.886 & 0.976 & 0.973 & 0.730 & 0.606 & 0.895 & 0.752 & 0.359 & 0.819\
0.8 & 0.764 & 0.300 & 0.735 & 0.881 & 0.978 & 0.969 & 0.714 & 0.555 & 0.901 & 0.738 & 0.380 & 0.810\
0.7 & 0.765 & 0.252 & 0.731 & 0.881 & 0.977 & 0.968 & 0.714 & 0.559 & 0.906 & 0.738 & 0.334 & 0.809\
0.6 & 0.756 & 0.166 & 0.710 & 0.877 & 0.976 & 0.967 & 0.697 & 0.536 & 0.897 & 0.725 & 0.253 & 0.793\
0.5 & 0.747 & 0.065 & 0.703 & 0.873 & 0.977 & 0.966 & 0.690 & 0.600 & 0.916 & 0.717 & 0.117 & 0.796\
0.4 & 0.735 & 0.021 & 0.661 & 0.867 & 0.975 & 0.963 & 0.675 & 0.496 & 0.912 & 0.704 & 0.038 & 0.766\
0.3 & 0.712 & 0.001 & 0.634 & 0.855 & 0.976 & 0.962 & 0.656 & 0.396 & 0.930 & 0.683 & 0.002 & 0.754\
0.2 & 0.692 & 0.001 & 0.608 & 0.846 & 0.975 & 0.952 & 0.638 & 0.292 & 0.931 & 0.664 & 0.001 & 0.736\
0.1 & 0.680 & 0.000 & 0.560 & 0.840 & 0.973 & 0.945 & 0.712 & 0.100 & 0.962 & 0.696 & 0.000 & 0.707\
Unfortunately, in practice there can exist imbalances towards fake or regular reviews being the majority class. From our results and research in other domains it is more likely that the bias is towards regular reviews. For this reason we choose more detailed results towards fake reviews being the minority class.
However, when fake reviews become the majority class (towards the right of the 50% mark) the recall improves for all algorithms, by up to 6.4% for the MLP algorithm. The best result is achieved by the RF algorithm with 0.986 recall. The AUC value decreases in all cases, for the RF and MLP algorithms the value slightly decreases by up to 0.5%. The value of the DT algorithm decreases more stronly by 5.5%. The best AUC value is achieved by the random forest algorithm with 0.984.
When fake reviews become the minority class most performance measures decrease. With an amount of **10% fake reviews** ($10^1$), as reported to be typical for other domains [@Weblink:10019; @doi:10.1287/mnsc.2015.2304; @mukherjee2013fake], the recall of the RF and DT algorithm are nearly identical (RF: 0.912, DT: 0.896). Compared to the result of the balanced dataset, the recall decreased by 5.3% for the RF algorithm and by 5.9% for the DT algorithm. The recall of the MLP algorithm is significantly less with 0.758 (-17.6%). The AUC value is the highest for the RF algorithm (0.986, +0.3%), followed by the MLP (0.977, +0.6%) and DLT (0.941, -0.9%) algorithms.
For **1% fake reviews** ($10^0$), the recall of the RF and DT algorithms is still nearly identical (RF: 0.756, -21.4%; DT: 0.775, -18.7%), followed by the MLP algorithm (0.395, -57.1%). The AUC value is the highest for the MLP algorithm (0.978, +0.7%), followed by the RF (0.971, -1.8%) and DT (0.886, -6.7%) algorithms.
For an amount of **0.1% fake reviews** ($10^{-1}$) the recall is the highest for the DT algorithm (0.680, -28.6%), followed by the RF algorithm (0.560, -41.8%). The recall of the MLP algorithm dropped to 0 at about 0.3% fake reviews and below within the dataset. The AUC value is the highest for the MLP algorithm (0.973, +0.2%), followed by the RF algorithm (0.945, -4.4%). Last, the AUC value of the DT algorithm significantly decreased to 0.840 (-11.4%).
Comparing all three algorithms using their recall and AUC/ROC value, **the random forest algorithm performs best for imbalanced datasets**. Although, the decision tree algorithm achieves a better recall when the dataset is extremely skewed (less than 1% fake reviews), its AUC/ROC value is significantly lower for all datasets. Similar, the MLP algorithm achieves better AUC/ROC values for datasets with less than 1% fake reviews. However, the recall of the MLP algorithm drops to 0 for extremely skewed datasets. With skews common for other domains, the random forest algorithm performs best with a recall of 0.912 and AUC/ROC value of 0.986.
Discussion {#sec:discussion}
==========
We discuss implications of fake reviews on software engineering, and from the perspective of app users and store operators. Then, we discuss the results’ validity.
Implications
------------
In modern app stores, developers are for the first time able to publicly retrieve customers’ and users’ opinions about their software and to compare its popularity in form of rank or number of downloads. Although app reviews provide a rich source of information, they may not be fully reliable, as customers may leave reviews that do not reflect their true impressions [@FINKELSTEIN2017119]. Our work shed light on one of these cases: fake reviews.
Generally, fake reviews, i.e., paid, incentivized reviews (which can provided either directly or via fake review providers) are prohibited by official app store reviewing policies. The main reason is to **preserve the integrity of app stores** [@Weblink:10005; @Weblink:10020]. Users that do not trust app stores and their reviews will most likely refrain from providing app reviews themselves. This would harm one of the most important advantages of app stores: collecting *real, spontaneous feedback* on software in a channel used by both developers and users.
We applied our fake review classifier to the full official Apple App Store dataset. As a results, 22,207,782 (35,5%) of all 62,617,037 reviews were classified as fake. This number seems very high at first and can only be used as a first indication. Further studies need to be carried out to give a precise approximation of the amount of fake reviews in official app stores. Still, multiple indices indicates a non-trivial amount of fake reviews in app stores. We identified about 60,000 reviews from only a single provider. Overall, we identified 43 providers while much more might exists or have existed before. If every provider would provide the same number of reviews, the amount would sum up to 2.58 million fake reviews. We hypothesize that the majority of fake reviews is written by persons, who get directly asked by developers. Although not generalizable, we repeatedly observed this phenomena in our professional app development settings. When apps are developed privately, friends were asked to provide fake reviews. When programmed in a commercial environment, either employees of the developing company or of the ordering company (cf. Bell [@Weblink:10007]) are asked to provide fake reviews. Given that 1.4 million apps exists within the dataset the number of fake reviews does no longer seem unattainably high. Such amount of fake reviews are also presumed in other domains. Streitfeld [@Weblink:10015] report that every fifth review submitted to Yelp is detected as dubious by internal filters.
**App users** might, by using positive or negative fake reviews, get mislead to either downloading an app or not. As shown by Ott et al. [@Ott:2011:FDO:2002472.2002512] fake reviews sound authentic and are hard to detect by humans. In an experiment, humans at most scored an accuracy of 61% identifying fake reviews, even as the word distribution of the used fake reviews differed from regular reviews. We think that this also applies for apps. Users and developers might not be able to identify fake reviews only based on their text.
Measures and tools should enable users (and developers) to identify fake reviews and affected apps. Such tools already exist for products sold on Amazon, e.g., Fakespot [@Weblink:10014]. Users enter the name of a product to determine if its reviews are trustworthy. Fakespot also takes a step towards analyzing fake reviews in the app store. The features used to classify reviews as fake also related to the review context, such as if a large number of positive reviews is provided within a short period of time. However, the selected criteria to classify reviews as fake are not completely transparent nor empirically validated. Also, we assume that no gold-standard dataset has been used for fake reviews. For the Instagram app the site, e.g., classifies 50% (about 600,000) of the reviews as fake, which raises accuracy concerns for this approach.
The **research area App Store Analysis** covers work to mine apps and their reviews and extract relevant information for software and requirements practitioners, e.g., to get inspirations about what should be developed and to guide the development process [@Harman:2012gw; @Pagano:2013:UIS:2486788.2486920; @Pagano:2013jn; @7886888; @Chen:2014:AMI:2568225.2568263; @6912257; @6624001; @Khalid:2013:IUC:2486788.2487044; @6606604]. Martin et al. provide a comprehensive literature review of this area [@Martin:2016fe]. The majority of papers identified in their study (127 of 187 papers, 68%) analyze non-technical information, such as app reviews.
Review Analysis itself is one of the largest sub-fields of app store analysis, which receives a significant and increasing number of publications each year. Work in this sub-field started by analyzing the content of app reviews (2012 – 2013), afterwards focusing on adding additional features such as sentiments (2013 – 2014). Then, app reviews were summarized to extract app requirements. Although, information extracted from app reviews is getting increasingly integrated into the requirements engineering processes [@7332475; @Maalej:2015:Software], **none of the papers discusses the impact of fake reviews**. In the following, we discuss the potential impact of fake reviews on software engineering along with the main review analysis topics according to Martin et al. [@Martin:2016fe].
Fake reviews, similar to official reviews, include **requirements-related information, such as feature requests.** Oh et al. automatically categorize app reviews into bug reports and non-/functional requests to produce a digest for developers including the most informative reviews [@Oh:2013:FDI:2468356.2468681]. Additional work focuses on extracting requirements-related information from app reviews [@10.1007/978-3-319-05452-0_4; @Iacob:2013:YCS:2578048.2578086; @7332474]. Iacob and Harrison found that 23.3% of the app reviews include feature requests [@6624001]. We applied the classifier of Maalej et al. [@Maalej:2016:Automatic] to extract bug reports and feature request from the 8,000 official and 8,000 fake reviews included in our truthset (see Figure \[fig:truthsetclassification\]). While fake and regular reviews are imbalanced within the overall app store dataset, when applying review analysis approaches on a subset of reviews the distribution is unknown. For this reason, we did not set a specific distribution of fake and regular reviews. Within the official reviews, we identified 1,297 bug reports and 921 feature requests, while we found that the fake reviews contain 362 bug reports and 521 feature requests. We include an example feature request within the fake review dataset below.
![Reviews within truthset classified as bug report and/or feature request.[]{data-label="fig:truthsetclassification"}](06-reclassifier){width=".85\columnwidth"}
Nice UI $\star\star\star\star$
*Very clean and beautiful UI. I like the goal setting and the reminders. I would like to see some animation when scrolling the weekly progress bars.*
We assume that most fake reviewers did not use the reviewed app before and are unfamiliar with it. In addition, review policies ask fake reviewers to explicitly talk about features rather than providing praise only. For this reasons, it is unclear if those feature requests are really relevant or only thought up to make the review sound more authentic. Fake reviews might thus impact the results of existing classifiers. When not removing or at least flagging fake reviews with information relevant for developers, wrong assumptions for the future decisions might be drawn.
Researchers showed that nearly **half of the feature requests included in app reviews are implemented.** Hoon et al. highlight that user expectations are changing rapidly, as observable through app reviews. Developers must keep up with the demand to stay competitive [@hoon2013analysis]. Palomba et al. studied the reviews of 100 open-source apps. By linking reviews to code changes the authors showed that 49% of the changes requested were implemented in app updates [@7332475]. These results show that a significant amount of changes proposed by users are integrated into software. Some of these changes might be inspired and prioritized based on fake reviews.
Recent approaches summarize and extract requirements-related information from **app reviews of related or competing apps.** Fu et al. present WisCom to analyze app reviews per app/market level, e.g., to get an overview of competing apps [@fu2013people]. Gao et al. present AR-Tracker to summarize app reviews to real issues and prioritize them by their frequency and importance [@gao2015ar]. Nayebi et al. mine app reviews and tweets of similar apps within a specific domain [@8049145; @7961668]. These approaches monitor and extract information from app reviews of related or competing apps. While developers will probably know when their apps receive fake reviews, i.e., when they bought those instead of being affected by negative fake reviews bought by competitors, developers cannot be fully sure if competing apps receive fake reviews. Wrong conclusions can also be drawn from fake reviews including the honest opinion of reviewers. Fake reviewers can just copy and modify a regular review they honestly agree with and resubmit those on review exchange portals. This way the frequency (and hence the priority) of, e.g., a feature request, might be fake (i.e. incentivized) and thus biased. Table \[tab:guidelines\] highlights that only three review exchange portals forbid fake reviewers to copy and modify existing reviews. Using those reviews as input for summarization approaches, “wrong” features could emerge as a result.
**App store operators** try to prevent fake reviews by providing review policies. The Apple App Store policy [@Weblink:10005] states that apps will be removed and that the developers may be expelled from the app store’s developer program *“if we find that you have attempted to manipulate reviews, inflate your chart rankings with paid, incentivized, filtered, or fake feedback”*. This is the case when app developers buy fake reviews. However, the concrete actions taken to identify fake reviews are non-transparent. We also noticed that larger, popular apps try to prevent negative feedback from being submitted to the app store. These apps ask users for submitting feedback within the app in form of star ratings. The rating is not directly forwarded to the app store. In case of a one star rating, a mail form appears asking the user to submit the review directly to the app developer instead of forwarding it to the app store, where the review is publicly visible. Such actions might also manipulate the app ratings and reviews as well.
We think that researchers should carefully sample apps and perform data cleanings before studying and mining app reviews. For example, apps with an unusual distribution of ratings might be affected by fake reviews. Similar, reviews of users with an amount or frequency above average might have to be removed or considered separately during data cleaning. Otherwise, wrong assumptions for the future development of an app could be drawn.
Collecting and analyzing **context and usage information** can help substantiate decisions and check the quality of reviews [@Maalej:2015:Software]. App developers can, e.g., utilize the number of users or the average time users spend with a specific feature to decide which parts of their apps to improve and which suggestions should be taken into consideration in the next release. App store operators can take measures, such as the number of times a user opened an app or the daily app usage time to decide the trustworthiness and weight of reviews within an app’s overall rating. Instead of limiting the amount of users who can participate in the reviewing process, one might think about to weight or consider incentivized reviews differently and in a transparent manner. Since the overall app store ecosystem is designed that more positive reviews lead to more downloads and thus increase the app’s success [@Harman:2012gw], developers will likely continue to ask “friends” to rate their apps. Even if incentivized and not independent, such reviews can also include useful information. Instead of excluding the reviews and their reviewers, a possible alternative might be to highlight these review with badges (e.g., friend, expert, or crowd testers).
Results Validity
----------------
The results of our study might have limitations and should be considered within the study context.
The process of extracting actual fake reviews was challenging. However, we did not try to generate fake reviews using a crowdsourced approach, as we wanted to only rely on fake reviews that have been published to and still remain unidentified within the app store by its users and operators. For this reason, we decided only using a subset of our about 60,000 collected fake reviews.
Nearly all reviews were extracted from a single provider. This provider is a review exchange portal (see Table \[tab:fakereviewdataset\], REP3). On these portals reviewers could submit their honest opinion. However, **we consider the collected reviews as fake for the following reasons:**
First, app store operators strongly require that app reviews must be 1) written by *real users* of the app and 2) cannot be *incentivized*. Both conditions are not given in review exchange portals. For this reason, these reviews are fake according to the definition and agreement or app store providers. Even if reviewers are allowed to submit their honest opinion according to the review policy of this exchange portal, *rewarded, incentivized, or non-spontaneous reviews* are prohibited by the official Google and Apple App Store Review Guidelines [@Weblink:10005; @Weblink:10020].
Second, review exchange portals provide predefined ratings and review messages. These ratings do not necessarily correspond to the opinion of the reviewers. The providers’ review policies ensure that reviewers that post their honest opinion are not being rewarded and are excluded from reviewing portals. The review policy of REP3 does not include a general rating, e.g., 3-stars or above (cf. Table \[tab:guidelines\]). The provider uses individual policies per app (cf. Figure \[fig:reviewrequest\]). We could not extract historical data to say if all individual policies included a predefined rating. However, for active review requests at the time of data collection, predefined ratings were included in all cases.
Third, paid review providers and review exchange portals share reviewers. As identified, paid review providers cross-post their review requests on review exchange portals (cf. Figure \[fig:reviewingstrategy\]). This also applies for REP3. Paid review providers would not cross-post their review requests on these portals, if the app ratings would not change as desired by app developers.
Fourth, per app we compared the collected fake reviews to each other. We searched for apps that received fake reviews with a rating of 1-2 stars as well as fake reviews with a rating of 4-5 stars. These reviews are most likely written by reviewers that posted their honest opinion about an app, either positive or negative. This applies for only 32 of the 1,890 (1.69%) collected apps. For these apps, 41 1-star and 2-star reviews out of 8,607 reviews (0.48%) were provided.
Another limitation is that although we filtered fake reviews for reviews in English language only and targeted the US storefront of the Apple App Store, reviews in English language could have been submitted to other storefronts. This could be a possible reason why we were only able to identify 8,607 of the initially collected 60,431 fake reviews within the app store, i.e., the official reviews dataset. In this case, the moderation of reviews by app store operators is less strict than observed.
Further, review exchange portals are also used by app developers, since the reward of providing a fake review is a credit which can be redeemed into another fake review for an app specified. As a result, the amount of requirements-related information included in fake reviews could be influenced since some users of the portals might be app developers.
For in-app purchases, we were unable to receive the offers programatically. Therefore, we could not compare the manually collected in-app purchases for apps affected by fake reviews against other app store in-app purchases. Also, we could not identify statistics which we could have used alternatively, or statistics on the monetization through ads.
However, we decided to focus on the Apple App Store because of our prior experience with the technology and because this app store does not impose major API limitations to retrieve its data, e.g., compared to Google Play with limits the number of accessible reviews to 2,000 per app and uses captchas. To have reliable results for the Apple App Store itself we crawled the largest dataset of about 62 million app reviews which has been analyzed so far to our knowledge. Thereby, we also avoid the App Sampling Problem for app store mining, described by Martin et al. [@Martin:2015:ASP:2820518.2820535].
The questionnaire we conducted with the paid review providers was hidden as a request for buying app reviews. We cannot assure that the responses only contain true statements. Therefore, we contacted several providers again after a few weeks using a different identity and communication channel, such as Skype. For providers we contacted again, their responses did not change.
When manually labelling data, such as when finding agreements for potential matches between reviews within the fake reviews dataset and official reviews dataset, we used two human annotators which independently solved the task. In case of mismatches (3%), we resolved the conflicts using a third human annotator. However, single reviews could have been mismatched.
For statistical tests, in addition to reporting the p value we also calculated the effect size. For t-tests we calculated the effect size using Cohen’s d, that is the difference between means divided by the pooled standard deviation [@cohen1988spa]. For Wilcoxon tests we calculated the r value, dividing the z distribution by the square root of the number of samples [@fritz2012effect]. We report the effect size considering the following values, for Cohen’s d (0.2 = small, 0.5 = medium, 0.8 = large) and for the correlation coefficient r (0.10 = small, 0.30 = medium, 0.50 = large). In two cases, although a statistical difference was observed the effect size revealed that the magnitude between differences is near zero. For these, the tests need to be repeated with additional samples (i.e., fake reviews) to show a statistical difference.
We want to stress that our classifier is only a first attempt to automatically identify fake reviews and not the main contribution of our paper. We wanted to verify if the features identified in our study are relevant for identifying fake reviews.
The machine learning model could be overfitted. This may be due to the small amount of fake reviews. More fake reviews need to be collected to improve the results. We tried to minimize overfitting by using k-fold cross validation. To minimize the impact of randomly chosen data, we used another 8,000 randomly selected official reviews and were able to reproduce our results reported in the paper.
Moreover, we cannot ensure that all official reviews are non-fake reviews. As stated before, to provide a gold-standard dataset for regular reviews as well, all fake reviews must be known and removed which is the problem we are trying to solve in this paper. For this reason, we optimized the classifier for precision only.
We leave the development of an advanced classifier for future research. Additional features, e.g., the emotion of users [@calefato2017emotxt; @Martens:2017:EUA:3105556.3105559], have to be analyzed to strengthen the results. For that we publicly share our gold-standard fake reviews dataset within our replication package.
Related Work {#sec:relwork}
============
User reviews are a valuable resource for decision making – both to other users and developers. To our best knowledge no published studies on fake reviews for software products exist.
A similar phenomena to fake reviews for software products, called web spamming, has been studied earlier when web pages began to compete for the rank within search engines’ results. This is comparable to apps competing within app stores today. Web spamming is defined as the act of misleading search engines to rank pages higher than they deserve. Website operators edit their pages, e.g., by repeatedly adding specific terms that improve their ranking in search results [@gyongyi2005web]. Based on this a definition for user-generated content emerged, called opinion spam. The authors divide opinion spam into three categories, of which the first category are untruthful opinions. These mislead readers and opinion mining systems by giving undeserved/unjust either positive reviews to promote or negative reviews to damage the reputation of a target object. Untruthful opinions are also commonly known as fake reviews [@Jindal:2008:OSA:1341531.1341560]. This definition has been refined several times, e.g., by adding that fake reviews are written by persons as if they were real customers [@Ott:2011:FDO:2002472.2002512].
Recent research in other areas studied fake reviews, e.g., for products sold on Amazon, hotels rated on TripAdvisor, and businesses rated on Yelp. Jindal and Liu [@Jindal:2008:OSA:1341531.1341560] first analyzed opinion spam. The authors analyze 5.8 million reviews and 2.14 million reviewers from Amazon to detect spam activities and present techniques to detect those. Due to the difficulty to create a fake reviews dataset, the authors used duplicate and near-duplicate reviews written by the same reviewers on different products. In our work we were able to extract data from fake review providers to achieve more reliable results.
Ott et al. [@Ott:2011:FDO:2002472.2002512] state that an increasing amount of user reviews is provided. Due to their value, platforms containing user reviews are becoming targets of opinion spam for potential monetary gain – our work confirms this for app stores and provides further insight on the fake review offers and policies. The authors focus on analyzing deceptive fake reviews, which are reviews that have been written to sound authentic, instead of disruptive fake reviews. The authors highlighted that there are few sources for deceptive fake reviews. To overcome the issue they hired 400 humans using a crowd-sourcing platform to write fake hotel reviews. Their classifier integrates work from psychology and computational linguistics. It has an accuracy of nearly 90% on the crowdsourced dataset. The authors showed that classifiers are better in recognizing deceptive fake reviews compared to humans which scored an accuracy of 61% at most.
Feng et al. [@feng2012distributional] analyzed fake reviews on TripAdvisor and Amazon. They identified fake reviews based on the hypothesis that for a given domain a representative distribution of review rating scores exist which is distorted by fake reviews. The authors used an unsupervised learning approach to create a review dataset that is labeled automatically based on rating distributions. Using a statistical classifier trained on that dataset the authors were able to detect fake reviews with an accuracy of 72%.
Mukherjee et al. [@Mukherjee2013WhatYF], compared to existing studies, used real fake reviews instead of pseudo-fake reviews, e.g., generated using crowdsourcing platforms. Their dataset consists of fake reviews published on Yelp, filtered and marked by the platform itself. The authors used the supervised approach of Ott et al. [@Ott:2011:FDO:2002472.2002512] on their dataset an achieved a significantly lower accuracy of 67.8%, compared to 89.6%. The authors found that the word distribution of pseudo-fake reviews is different to the word distribution of real reviews. However, this does not apply for fake reviews within their dataset. Instead of using linguistic features, the authors suggest to use behavioral features. These include numeric values, such as the maximum number of reviews of a reviewer per day, or the review length. We followed this approach when designing our classifier and achieved encouraging results of up to 97% precision.
Fake reviews have also been frequently discussed within the media. Streitfeld [@Weblink:10015] reported that every fifth review submitted to Yelp is detected as dubious by its internal filters. Instead of removing dubious reviews, these are moved to the second page where they are read by less users. As fake reviews further increase, Yelp began a ’sting’ campaign to publicly expose businesses buying fake reviews.
Conclusion {#sec:conclusion}
==========
App reviews can be a valuable, unique source of information for software engineering teams reflecting the opinions and needs of actual users. Also potential users read through the reviews before deciding to download an app, similar to buying other products on the Internet. Our work shows that part of app reviews in app stores are fake – that is, they are incentivized and might not reflect spontaneous, unbiased opinions.
We analyzed the market of fake review providers and their fake reviewing strategies and found that developers buy reviews to relatively expensive prices of a few dollars or deal with reviews in exchange portals. Fake reviews are written to look authentic and are hard to recognize by humans. We identified differences between fake and official reviews. We found that properties of the corresponding app and reviewer are most useful to determine if a review is fake. Based on the identified differences, we developed, trained, fine-tuned, and compared multiple supervised machine learning approaches. We found that the Random Forest classifier identifies fake reviews, given a proportional distribution of fake and regular reviews as reported in other domains, with a recall of 91% and AUC/ROC value of 98%. We publicly share our gold-standard fake reviews dataset to enable the development of more accurate classifiers to identify fake reviews. Our work helps app store mining researchers to sample apps and perform data cleaning to achieve more reliable results. Further, tools for app users and store operators can be built based on our findings to detect if app reviews are trustworthy and to take further actions against fake reviewers.
Acknowledgment {#acknowledgment .unnumbered}
==============
This research was partially supported by the European Union Horizon 2020 project OpenReq under grant agreement no. 732463.
[^1]: https://affiliate.itunes.apple.com/resources/documentation/itunes-store-web-service-search-api/
[^2]: https://appleid.apple.com/faq/\#!&page=faq
[^3]: Conducted with permission of the Ethics Committee of the University of Hamburg.
[^4]: https://zeppelin.apache.org/
[^5]: https://github.com/tesseract-ocr/tesseract
[^6]: https://github.com/saffsd/langid.py
[^7]: https://mast.informatik.uni-hamburg.de/app-review-analysis/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We collect several open questions in Banach spaces, mostly related to measure theoretic aspects of the theory. The problems are divided into five categories: miscellaneous problems in Banach spaces (non-separable $L^p$ spaces, compactness in Banach spaces, $w^*$-null sequences in dual spaces), measurability in Banach spaces (Baire and Borel $\sigma$-algebras, measurable selectors), vector integration (Riemann, Pettis and McShane integrals), vector measures (range and associated $L^1$ spaces) and Lebesgue-Bochner spaces (topological and structural properties, scalar convergence).'
address: 'Dpto. de Ingeniería y Tecnología de Computadores, Facultad de Informática, Universidad de Murcia, 30100 Espinardo (Murcia), Spain'
author:
- José Rodríguez
title: Open problems in Banach spaces and measure theory
---
Introduction {#introduction .unnumbered}
============
The interaction between Banach space theory and measure theory has provided truly important results in functional analysis. The topic reached its maturity in the seventies and eighties thanks to the contributions of outstanding mathematicians like J. Bourgain, G.A. Edgar, D.H. Fremlin, I. Namioka, R.R. Phelps, C. Stegall, M. Talagrand, etc. For detailed information we refer to the monographs [@bou-J; @cem-men; @die-uhl-J; @tal; @van]. Among the most celebrated results in this context, one finds the characterization of Asplund spaces as those whose dual has the Radon-Nikodým property, and the characterization of Banach spaces not containing $\ell^1$ as those whose dual has the weak Radon-Nikodým property. Regarding applications to other areas, we can say, for instance, that the techniques of set-valued integration in Banach spaces (initiated by R.J. Aumann and G. Debreu) have been used successfully in mathematical economics, control theory, game theory, etc. (see e.g. [@aub-fra; @kle-tho]).
In this expository paper we collect some open questions (mostly) related to measure theoretic aspects of Banach space theory. The choice of problems is conditioned by our own research interests (and shortcomings) and we do not pretend to be exhaustive. The questions are organized in several sections:
- Miscellaneous problems in [*Banach spaces*]{}: non-separable $L^p$ spaces, compactness in Banach spaces, $w^*$-null sequences in dual spaces.
- Problems in [*measurability in Banach spaces*]{}: Baire and Borel $\sigma$-algebras, measurable selectors.
- Problems in [*vector integration*]{}: Riemann, Pettis and McShane integrals.
- Problems in [*vector measures*]{}: range and associated $L^1$ spaces.
- Problems in [*Lebesgue-Bochner spaces*]{}: topological and structural properties, scalar convergence.
We hope that this paper is a useful source of information as well as inspiration for researchers in Banach space theory and measure theory.
Notation and terminology {#notation-and-terminology .unnumbered}
------------------------
Our standard references are [@alb-kal; @fab-ultimo] (Banach spaces) and [@die-uhl-J] (vector measures). All our topological spaces are Hausdorff and we only consider real Banach spaces. An [*operator*]{} between Banach spaces is a linear continuous map. Given a Banach space $X$, its norm is denoted by $\|\cdot\|_X$ or simply $\|\cdot\|$. We write $B_X=\{x\in X:\|x\|\leq 1\}$ (the closed unit ball of $X$) and $X^*$ denotes the topological dual of $X$. The weak topology on $X$ and the weak$^*$ topology on $X^*$ are denoted by $w$ and $w^*$, respectively. A [*subspace*]{} of $X$ is a closed linear subspace. The subspace generated by a set $A{\subseteq}X$ is denoted by $\overline{{\rm span}}(A)$. Given another Banach space $Y$, we say that $X$ contains $Y$ or that $X$ contains a copy of $Y$ (and we write $X \supseteq Y$) if there is a subspace of $X$ which is isomorphic to $Y$. Recall that $X$ is said to be [*Asplund*]{} if every separable subspace of $X$ has separable dual (equivalently, $X^*$ has the Radon-Nikodým property). A set $\Gamma {\subseteq}X^*$ is said to be [*total*]{} if it separates the points of $X$, i.e. for every $x\in X\setminus \{0\}$ there is $x^*\in \Gamma$ such that $x^*(x)\neq 0$. The topology on $X$ of pointwise convergence on a total set $\Gamma {\subseteq}X^*$ is denoted by $\sigma(X,\Gamma)$.
The unit interval $[0,1]$ is equipped with the Lebesgue measure $\lambda$ on the $\sigma$-algebra of all Lebesgue measurable subsets. The symbols $\mathfrak{c}$ and $\omega_1$ denote the cardinality of the continuum and the first uncountable ordinal, respectively. The [*density character*]{} of a topological space $T$ is the cardinal ${\rm dens}(T)=\min\{|A|: A {\subseteq}T, \, \overline{A}=T\}$, where $|A|$ denotes the cardinality of the set $A$. We use the term “compact” as an abbreviation of “compact topological space”.
Miscellaneous problems in Banach spaces {#section:Banach}
=======================================
Non-separable $L^p$ spaces {#subsection:Lp-nosetoble}
--------------------------
Several aspects of the structure of non-separable $L^p$ spaces are still unclear. In the separable case, it is known that $L^1[0,1]$ has no unconditional basis and, moreover, it is not a subspace of a Banach space with unconditional basis. On the other hand, $L^p[0,1]$ has unconditional basis for every $1<p<\infty$. The concept of unconditional basis admits a natural extension to the non-separable case (see e.g. [@fab-alt-JJ Section 7.3]). Enflo and Rosenthal [@enf-ros] proved that $L^p(\mu)$ (for $1<p<\infty$, $p\neq 2$, and a finite measure $\mu$) is not a subspace of a Banach space with unconditional basis if ${\rm dens}(L^p(\mu))\geq \aleph_\omega$. They also conjectured that the same conclusion holds whenever $L^p(\mu)$ is not separable. The assumption on the density character is needed to use a combinatorial lemma that works only for “big” cardinals. It is worth mentioning that a weakening of that lemma, for arbitrary uncountable sets, has been recently used in [@avi-mar] to study certain extension operators between spaces of continuous functions.
\[problem:EnfloRosenthal\] Let $1<p<\infty$, $p\neq 2$, and let $\mu$ be a finite measure such that $L^p(\mu)$ is not separable. Can $L^p(\mu)$ be a subspace of a Banach space having unconditional basis?
In an attempt to attack the previous problem, Johnson and Schechtman [@joh-sch] got interesting results on non-separable $L^p$ spaces. A simple application of Pitt’s theorem shows that, for $2<p<\infty$, the space $\ell^p(\omega_1)$ is not a subspace of $L^p(\mu)$ for any finite measure $\mu$. It is more complicated to prove the same statement for the range $1<p<2$, [@enf-ros]. On the other hand, if $1<p<2$ and $X$ is a subspace of an $L^p$ space (over an arbitrary non-negative measure) such that $X\not\supseteq \ell^p(\omega_1)$, then $X {\subseteq}L^p(\mu)$ for some finite measure $\mu$, [@joh-sch]. It is conjectured that this result also holds true for $2<p<\infty$.
\[problem:JS\] Let $2<p<\infty$ and let $X$ be a subspace of an $L^p$ space such that $X\not\supseteq \ell^p(\omega_1)$. Does $X {\subseteq}L^p(\mu)$ for some finite measure $\mu$?
A Banach space $X$ is said to be [*Hilbert generated*]{} if there exist a Hilbert space $H$ and an operator $T:H \to X$ having dense range (see e.g. [@fab-alt-JJ Section 6.3]). The basic examples of Hilbert generated spaces are: separable spaces, $c_0(\Gamma)$ (for any set $\Gamma$) and $L^p(\mu)$ for any $1\leq p\leq 2$ and any finite measure $\mu$ (in this case we can take $H=L^2(\mu)$ and $T$ the inclusion operator). Every super-reflexive space is a [*subspace*]{} of a Hilbert generated space and there are examples of super-reflexive spaces which are not Hilbert generated, [@fab-alt-J-4]. However, it seems to be unknown if every space $L^p(\mu)$ with $2<p<\infty$ (and a finite measure $\mu$) is Hilbert generated. Of course, this is true if $L^p(\mu)$ is separable, but also if ${\rm dens}(L^p(\mu))=\omega_1$, see [@fab-alt-J-4].
\[problem:LpHilbertGenerated\] Let $2<p<\infty$ and let $\mu$ be a finite measure. Is $L^p(\mu)$ Hilbert generated?
Compactness in Banach spaces {#subsection:compacidad}
----------------------------
A Banach space $X$ is said to be [*weakly compactly generated*]{} (WCG) if there is a weakly compact set $K {\subseteq}X$ such that $X=\overline{{\rm span}}(K)$. This class of Banach spaces includes all separable spaces, reflexive spaces, Hilbert generated spaces, etc. and plays a very important role in non-separable Banach space theory (see e.g. [@fab-ultimo Chapter 13] and [@fab-alt-JJ Chapter 6]). In the particular case of a Banach [*lattice*]{} $X$, Diestel [@die6] asked whether the property of being WCG is equivalent to the existence of a weakly compact set $L {\subseteq}X$ such that the [*sublattice*]{} generated by $L$ is dense in $X$. In [@avi-alt-3] we answered in the affirmative this question for order continuous Banach lattices, but the general case still remains open.
\[problem:Diestel\] Let $X$ be a Banach lattice for which there is a weakly compact set $L {\subseteq}X$ such that the sublattice generated by $L$ is dense in $X$. Is $X$ WCG?
The concept of [*weakly precompactly generated*]{} (WPG) Banach space was introduced by Haydon in [@hay10]. A subset $C$ of a Banach space is called [*weakly precompact*]{} if every sequence in $C$ admits a weakly Cauchy subsequence or, equivalently, if $C$ is bounded and contains no sequence which is equivalent to the usual basis of $\ell^1$ (thanks to Rosenthal’s $\ell^1$-theorem, [@ros-J-3]). A Banach space $X$ is said to be WPG if there is a weakly precompact set $C {\subseteq}X$ such that $X=\overline{{\rm span}}(C)$. This class includes all WCG spaces and all Banach spaces not containing $\ell^1$. Some results on these spaces can also be found in [@sch-PhD Section 2.3]. While preparing his PhD Thesis, G. Martínez-Cervantes [@gon2] is studying WPG spaces and the class of [*weak Radon-Nikodým*]{} (WRN) compacta introduced in [@gla-meg-1]. A compact $K$ is said to be WRN if it is homeomorphic to a $w^*$-compact subset of the dual of a Banach space not containing $\ell^1$; this condition is equivalent to $C(K)$ being WPG. The class of WRN compacta is larger than the well-known class of Radon-Nikodým compacta. Recall that a compact $K$ is called [*Radon-Nikodým*]{} if it is homeomorphic to a $w^*$-compact subset of the dual of an Asplund space.
In [@hay10] Haydon asked, using the language of WPG Banach spaces, whether every WRN (infinite) compact admits a convergent (non-stationary) sequence. This question is related to “Efimov’s problem” on the existence of (infinite) compacta without subsets homeomorphic to $\beta\N$ and without convergent (non-stationary) sequences. Until now such examples have been constructed only under additional axioms of set theory (see e.g. [@dow-she; @har]).
\[problem:Haydon\] Does every WRN (infinite) compact admit a convergent (non-stationary) sequence?
In [@avi-ple-rod-5] we have studied the partially ordered set $\cK(B_X)$ of all weakly compact subsets of the closed unit ball of a separable Banach space $X$ (the order being given by inclusion). To measure the complexity of $\cK(B_X)$ we use Tukey ordering, which is a useful tool to isolate essential properties of ordered structures in measure theory and topology, see [@fre12; @fre13; @sol-tod]. For instance, Fremlin [@fre12] proved that, given a coanalytic separable metric space $E$, the partially ordered set $\cK(E)$ of all compact subsets of $E$ is Tukey equivalent to one of the following: $\{0\}$, $\N$, $\N^\N$, $\cK(\mathbb{Q})$. In [@avi-ple-rod-5] we use advanced techniques of descriptive set theory to show, under the axiom of analytic determinacy, that $\cK(B_X)$ is Tukey equivalent to $\{0\}$, $\N^\N$, $\cK(\mathbb{Q})$ or $[\mathfrak{c}]^{<\mathbb{N}}$ (the family of all finite subsets of $\mathfrak{c}$). This classification result is valid in ZFC if $X\not\supseteq \ell^1$, but we do not know what happens in general.
\[problem:Tukey\] Is it relatively consistent that there is a separable Banach space $X$ such that $\cK(B_X)$ is not Tukey equivalent to $\{0\}$, $\N^\N$, $\cK(\mathbb{Q})$ or $[\mathfrak{c}]^{<\mathbb{N}}$?
$w^*$-null sequences in dual spaces {#subsection:SequencesDual}
-----------------------------------
A subset $C$ of a Banach space $X$ is said to be [*limited*]{} if $\lim_{n\to \infty} \, \sup_{x\in C}|x_n^*(x)|=0$ for every $w^*$-null sequence $(x_n^*)$ in $X^*$. This concept is intimately related to different notions of compactness. It is easy to check that every relatively norm compact set is limited. Bourgain and Diestel [@bou-die] proved that, in general, every limited set is weakly precompact; they also showed that every limited subset of a Banach space not containing $\ell^1$ is relatively weakly compact. This last result extends to any WPG space, as pointed out in [@sch-PhD Cor. 2.3.3]. In another direction, a Banach space is said to have the [*Gelfand-Phillips property*]{} if every limited subset is relatively norm compact (see e.g. [@bou-die; @dre2; @sch-2]). For instance, every Banach space having $w^*$-sequentially compact dual ball (e.g. WCG or Asplund spaces, see [@die-J Chapter XIII]) or having the separable complementation property enjoys the Gelfand-Phillips property. Recall that a Banach space $X$ has the [*separable complementation property*]{} (SCP) if every separable subspace of $X$ is contained in a separable complemented subspace of $X$. Typical examples of Banach spaces with the SCP are WCG spaces and Banach lattices not containing $c_0$. The following open problem is directly connected with a question on Pettis integration raised by Talagrand [@tal 4-2-6] (see Subsection \[subsection:Pettis\]):
\[problem:GelfandPhillips\] Let $X$ be a Banach space not containing $c_0$. Does $X$ have the Gelfand-Phillips property?
Mazur’s theorem ensures that every weakly convergent sequence in a Banach space admits a convex block subsequence which converges in norm. Recall that a [*convex block subsequence*]{} of a sequence $(x_n)$ in a Banach space is a sequence $(y_k)$ of vectors of the form $y_k=\sum_{n\in I_k}a_n x_n$, where $I_1,I_2,\dots$ are finite subsets of $\N$ with $\max(I_k)<\min(I_{k+1})$ and the scalars $a_n\geq 0$ satisfy $\sum_{n\in I_k}a_n=1$ for every $k\in \N$. A Banach space $X$ has [*property (K)*]{} if every $w^*$-null sequence $(x_n^*)$ in $X^*$ admits a convex block subsequence $(y^*_k)$ which converges in the Mackey topology $\mu(X^*,X)$, that is, $\lim_{k\to \infty} \, \sup_{x\in K}|y_k^*(x)|=0$ for every weakly compact set $K {\subseteq}X$. This concept, attributed to Kwapień, was used in [@kal-pel] to study certain questions about subspaces of $L^1[0,1]$. A variant of property (K) was employed in [@fig-alt2] to prove that, in general, the SCP is not inherited by subspaces. The basic examples of Banach spaces with property (K) are the reflexive ones and, more generally, Grothendieck spaces (e.g. $\ell^\infty$) and the strongly WCG spaces of [@sch-whe] (e.g. $L^1(\mu)$ for any finite measure $\mu$). A Banach space $X$ is said to be [*strongly WCG*]{} (SWCG) if there is a weakly compact set $K_0 \subseteq X$ such that, for every weakly compact $K \subseteq X$ and every $\varepsilon>0$, we have $K \subseteq nK_0 + \varepsilon B_X$ for some $n\in \N$.
Pełczyński (see e.g. [@fig-alt2; @fra-ple]) proved that the $\ell^1$-sum of $\mathfrak{b}$ copies of $L^1[0,1]$ fails property (K). Extending a result of [@fra-ple], we have recently shown that property (K) is preserved by $\ell^1$-sums of less than $\mathfrak{p}$ summands, [@avi-ple-rod-6]. For more information about cardinals $\mathfrak{p}$ and $\mathfrak{b}$ (that satisfy $\omega_1\leq \mathfrak{p}\leq \mathfrak{b}\leq \mathfrak{c}$), see e.g. [@dou2].
\[problem:l1sumasK\] What is the least cardinality of a family of Banach spaces with property (K) whose $\ell^1$-sum fails such property?
\[problem:BaseIncondicionalK\] Let $X$ be a separable Banach space which is weakly sequentially complete. Does $X$ have property (K)?
Problems in measurability in Banach spaces {#section:measurability}
==========================================
Baire $\sigma$-algebras {#subsection:Baire}
-----------------------
The *Baire $\sigma$-algebra* of a topological space $T$, denoted by ${{\rm Ba}}(T)$, is the one generated by all continuous functions from $T$ to ${\mathbb{R}}$. It is contained in the *Borel $\sigma$-algebra* of $T$ (denoted by ${{\rm Bo}}(T)$) and, in general, such inclusion is strict. If $E$ is a locally convex space, a result of Edgar [@edgar1] says that the Baire $\sigma$-algebra of $(E,weak)$ is the $\sigma$-algebra generated by all elements of the topological dual of $E$. In particular, if $X$ is a Banach space and $\Gamma {\subseteq}X^*$ is a total set, then ${{\rm Ba}}(X,\sigma(X,\Gamma))$ is the $\sigma$-algebra generated by $\Gamma$; of course, one has ${{\rm Ba}}(X,\sigma(X,\Gamma)) {\subseteq}{{\rm Ba}}(X,w)$. A Banach space $X$ is said to have [*property $\cD$*]{} if ${{\rm Ba}}(X,\sigma(X,\Gamma)) = {{\rm Ba}}(X,w)$ for every total set $\Gamma {\subseteq}X^*$. This property is fulfilled by every Banach space $X$ such that $(X^*,w^*)$ is [*angelic*]{} (e.g. if $X$ is WCG), [@gul-J]. The angelicity of $(X^*,w^*)$ means that every element of the $w^*$-closure of a bounded set $A{\subseteq}X^*$ is the $w^*$-limit of a sequence contained in $A$. The space of Johnson-Lindenstrauss $JL_2$ has property $\cD$ but fails to have $w^*$-angelic dual, [@pli3]. A technical condition between the $w^*$-angelicity of the dual and property $\cD$ is the following:
- every $w^*$-sequentially closed linear subspace of $X^*$ is $w^*$-closed.
\[problem:Plichko\] Is property $\cD'$ equivalent to property $\cD$?
Let $K$ be compact. The dual space $C(K)^*$ is identified with the space of all regular Borel signed measures on $K$. The subset of $C(K)^*$ consisting of all probabilities is denoted by $P(K)$. Let $\mathfrak{T}_p$ be the topology on $C(K)$ of pointwise convergence on $K$. In view of the above, the $\sigma$-algebra ${{\rm Ba}}(C(K),\mathfrak{T}_p)$ is generated by the set of evaluation functionals $\Delta_K:=\{\delta_t:t\in K\} {\subseteq}P(K)$, where $\delta_t(f):=f(t)$ for all $t\in K$ and $f\in C(K)$. The equality ${{\rm Ba}}(C(K),\mathfrak{T}_p)={{\rm Ba}}(C(K),w)$ holds true if every $\mu \in P(K)$ admits a [*uniformly distributed sequence*]{}, i.e. a sequence $(t_n)$ in $K$ such that, for every $f\in C(K)$, $$\int_K f \, d\mu=\lim_{m\to \infty}\frac{1}{m}\sum_{n=1}^m f(t_n).$$ This happens in several cases, for instance, when $K=\{0,1\}^{\mathfrak{c}}$ or when $K$ is Eberlein, Radon-Nikodým, etc. (see [@freMT-4 §491] and [@mer-J]). Recall that $K$ is said to be [*Eberlein*]{} if it is homeomorphic to a weakly compact subset of a Banach space.
However, the coincidence of ${{\rm Ba}}(C(K),\mathfrak{T}_p)$ and ${{\rm Ba}}(C(K),w)$ does not guarantee the existence of uniformly distributed sequences for every $\mu\in P(K)$, at least under the Continuum Hypothesis [@avi-ple-rod-4]. Now, write ${\rm co}(\Delta_K)$ for the convex hull of $\Delta_K$ and let ${\rm Seq}({\rm co}(\Delta_K))$ be its $w^*$-sequential closure in $C(K)^*$ (i.e. the smallest $w^*$-sequentially closed subset of $C(K)^*$ containing ${\rm co}(\Delta_K)$).
\[problem:BaireSequential\] Let $K$ be compact such that ${{\rm Ba}}(C(K),\mathfrak{T}_p)={{\rm Ba}}(C(K),w)$. Does the equality ${\rm Seq}({\rm co}(\Delta_K)) = P(K)$ hold?
An equivalent norm $\|\cdot\|$ in a Banach space $X$ is ${{\rm Ba}}(X,w)$-measurable (as a function from $X$ to ${\mathbb{R}}$) if and only if its balls belong to ${{\rm Ba}}(X,w)$. The following implications are easy to check:
$B_{X^*}$ is $w^*$-separable $\Longrightarrow$ $\|\cdot\|$ is ${{\rm Ba}}(X,w)$-measurable $\Longrightarrow$ $X^*$ is $w^*$-separable.
The converse implications do not hold in general, as we shown in [@rod9] by using certain equivalent norms in $\ell^\infty$. However, we do not know what happens with the converse implications when $X$ is a $C(K)$ space equipped with the [*supremum norm*]{} $\|\cdot\|_\infty$. Under the Continuum Hypothesis, Talagrand [@tal12] constructed a compact $K$ such that $C(K)^\ast$ is $w^\ast$-separable while $B_{C(K)^\ast}$ is not. In [@avi-ple-rod-3] we gave a ZFC example of a compact with these properties and we also proved some partial results about the measurability of $\|\cdot\|_\infty$ in the corresponding $C(K)$ space.
\[problem:NormaSupremo\] Let $K$ be compact. Is the measurability of $\|\cdot\|_\infty$ with respect to ${{\rm Ba}}(C(K),w)$ equivalent to the $w^*$-separability of $B_{C(K)^*}$ or $C(K)^*$?
Borel $\sigma$-algebras {#subsection:Borel}
-----------------------
Given a compact $K$, the space $C(K)$ admits several topologies (of uniform, weak or pointwise convergence) leading to different $\sigma$-algebras: $$\label{eqn:sigma-algebras}
\begin{array}{c c c c c}
{{\rm Ba}}(C(K),\mathfrak{T}_p) & \subset & {{\rm Bo}}(C(K),\mathfrak{T}_p) & \mbox{ } & \mbox{ } \\
\cap & \mbox{ } & \cap & \mbox{ } & \mbox{ } \\
{{\rm Ba}}(C(K),w) & \subset & {{\rm Bo}}(C(K),w) & \subset & {{\rm Bo}}(C(K))
\end{array}$$ All these $\sigma$-algebras coincide if $K$ is metrizable, but also in some cases beyond metrizability, like $K=\{0,1\}^{\omega_1}$, [@avi-ple-rod-2]. On the other hand, $K=\beta\N$ is an example for which all the $\sigma$-algebras of are different, see [@mar-ple; @tal9]. In general, the equalities between such $\sigma$-algebras are closely related with topological properties of $K$ and Banach space properties of $C(K)$. For instance, it is known that the equality ${{\rm Bo}}(C(K),\mathfrak{T}_p) = {{\rm Bo}}(C(K))$ holds true if $K$ is an Eberlein compact or, more generally, a Valdivia compact (see e.g. [@edgar1; @edgar2] and [@dev-alt-J Chapter VII]). The following questions remain open:
\[problem:Burke-Pol\] Let $K$ be compact such that $${{\rm Bo}}(C(K),w) \neq {{\rm Bo}}(C(K)).$$ Is there a norm discrete subset of $C(K)$ which does not belong to ${{\rm Bo}}(C(K),w)$?
\[problem:Marciszewski-Pol\] Let $K$ be compact such that $${{\rm Bo}}(C(K),w) = {{\rm Bo}}(C(K)).$$ Does the equality ${{\rm Bo}}(C(K),\mathfrak{T}_p)={{\rm Bo}}(C(K))$ hold?
A topological space $T$ is said to be [*Radon*]{} if every probability $\mu$ on ${{\rm Bo}}(T)$ is Radon (i.e. $\mu(E)=\sup\{\mu(L): \, L {\subseteq}E, \, L \mbox{ compact}\}$ for every $E\in {{\rm Bo}}(T)$). A classical result states that a complete metric space $T$ is Radon if ${\rm dens}(T)$ is a cardinal of measure zero (e.g. if ${\rm dens}(T)\leq \omega_1$). L. Schwartz asked if $(X,w)$ is Radon for any Banach space $X$. If $X$ is separable, then ${{\rm Bo}}(X,w)={{\rm Bo}}(X)$ and therefore $(X,w)$ is Radon; the same happens if $X$ is WCG and ${\rm dens}(X)=\omega_1$. However, there exist Banach spaces with density character $\mathfrak{c}$ which are not Radon with its weak topology, like $\ell^\infty/c_0$. The references [@freMT-4 §466], [@jay-alt10] and [@tal] contain more information about this topic. For any Banach space $X$, every Radon probability on ${{\rm Bo}}(X,w)$ extends to a Radon probability on ${{\rm Bo}}(X)$ (Phillips-Grothendieck). Thanks to this result, an affirmative answer to the following question would imply that $(\ell^\infty,w)$ is not Radon, thus solving a long-standing conjecture:
\[problem:FremlinRadon\] Is there a probability on ${{\rm Bo}}(\ell^\infty,w)$ which cannot be extended to a probability on ${{\rm Bo}}(\ell^\infty)$?
Measurable selectors {#subsection:selectors}
--------------------
Let $(\Omega,\Sigma,\mu)$ be a complete probability space and $X$ a Banach space. Denote by $\mathcal{P}_0(X)$ the family of all non-empty subsets of $X$. A *multi-function* is a map $F: \Omega \to \mathcal{P}_0(X)$ and a *selector* of $F$ is a function $f: \Omega \to X$ such that $f(t) \in F(t)$ for every $t \in \Omega$. Most of the results about the existence of “measurable” selectors are restricted to the context of separable spaces, that allows to use classical descriptive set theory. For instance, the Kuratowski and Ryll-Nardzewski selection theorem [@kur-ryl] ensures that, for separable $X$, a multi-function $F:\Omega \to \mathcal{P}_0(X)$ has ${{\rm Bo}}(X)$-measurable selectors if it takes closed values and satisfies: $$F^{-}(G):=\{t\in \Omega: \, F(t)\cap G \neq \emptyset\}\in \Sigma\quad
\mbox{for every open }G {\subseteq}X.$$ A multi-function $F:\Omega \to \mathcal{P}_0(X)$ is said to be [*scalarly measurable*]{} if $F^{-}(H)\in \Sigma$ for every open half-space $H{\subseteq}X$. For arbitrary Banach spaces, we showed in [@cas-kad-rod-3] that every scalarly measurable multi-function with [*weakly compact*]{} values admits scalarly measurable selectors (cf. [@fre-CKR]). Recall that a function $f:\Omega\to X$ is said to be [*scalarly measurable*]{} if the composition $x^*f:\Omega \to {\mathbb{R}}$ is measurable for every $x^*\in X^*$ (equivalently, $f$ is ${{\rm Ba}}(X,w)$-measurable). The result of [@cas-kad-rod-3] allowed to develop the set-valued Pettis integral theory in non-separable Banach spaces, [@cas-kad-rod-2; @mus9; @mus11].
\[problem:multi\] Let $F:\Omega \to \mathcal{P}_0(X)$ be a scalarly measurable multi-function with convex closed values. Does $F$ admit scalarly measurable selectors?
A [*set selector*]{} of a family of sets $\cC {\subseteq}\mathcal{P}_0(X)$ is a map $\psi:\cC \to X$ such that $\psi(C)\in C$ for every $C\in \cC$. If $F:\Omega \to \cC$ is any multi-function, then the composition $\psi\circ F:\Omega \to X$ is a selector of $F$, and it is natural to ask whether such selector enjoys some nice measurability or continuity property provided $F$ does. Some important theorems about measurable selectors (Kuratowski and Ryll-Nardzewski) or first Baire class selectors (Jayne and Rogers) can be deduced from the existence of suitable set selectors, [@gho-alt2]. We say that the family $\cC$ has [*property (SC)*]{} if there is a set selector $\psi:\cC\to X$ such that $\psi\circ F$ is scalarly measurable for every scalarly measurable multi-function $F:\Omega \to \cC$. Denote by $wk(X)$ the family of all weakly compact non-empty subsets of $X$. Then $wk(X)$ has property (SC) if $X^*$ is $w^*$-separable or if $(X^*,w^*)$ is angelic and has density character $\omega_1$, see [@vld] and [@cas-kad-rod-2] (cf. [@fre-CKR]). Fremlin [@fre-CKR] proved that, in general, $wk(X)$ may fail property (SC), even for Hilbert spaces. The following questions remain open:
\[problem:selectorsets1\] Suppose ${\rm dens}(X)=\omega_1$. Does $wk(X)$ have property (SC)?
\[problem:selectorsets2\] Does the family $cwk(X)$ of all convex weakly compact non-empty subsets of $X$ have property (SC)?
Problems in vector integration {#section:integrals}
==============================
Riemann integral {#section:Riemann}
----------------
Lebesgue’s criterion of Riemann integrability does not work in general for vector-valued functions: there exist Riemann integrable functions defined on $[0,1]$ with values in a Banach space which are not continuous a.e. (see e.g. [@gor-2]). This phenomenon occurs in almost all classical Banach spaces. A Banach space $X$ is said to have the [*Lebesgue property*]{} (LP) if every Riemann integrable function $f:[0,1] \to X$ is continuous a.e. For instance, $\ell^{1}(\Gamma)$ has the LP for any set $\Gamma$. A complete characterization of Banach spaces having the LP is missing, although there are partial results related to fine properties of the structure of Banach spaces. A result attributed to da Rocha and Pełczyński asserts that every asymptotic $\ell^1$ Banach space (e.g. Tsirelson’s space) has the LP. On the other hand, Haydon [@hay9] proved that Banach spaces having the LP share a feature with Banach spaces having the Schur property, namely: every spreading model is equivalent to the usual basis of $\ell^1$. However, there exist separable Banach spaces having the Schur property and failing the LP, [@hay9]. The paper [@nar] contains alternative proofs of these results.
\[problem:LP\] Characterize Banach spaces having the LP.
Alexiewicz and Orlicz [@ale-orl-J] showed an example of a Riemann integrable function with values in $C[0,1]$ without points of continuity with respect to the weak topology. A Banach space $X$ is said to have the [*weak Lebesgue property*]{} (WLP) if every Riemann integrable function $f:[0,1] \to X$ is [*weakly*]{} continuous a.e. It is easy to see that all spaces with separable dual have the WLP. Beyond this case, the space $L^{1}[0,1]$ has the WLP, [@wan-2]. This result was extended to certain Lebesgue-Bochner spaces in [@cal-alt-3] and, recently, it has been improved by G. Martínez-Cervantes [@gon], showing that $C(K)^*$ has the WLP for a wide class of compacta $K$ including those which are Eberlein or Radon-Nikodým. Some partial answers to the following question have also been given in [@gon]:
\[problem:WLP\] For which cardinals $\kappa$ does the space $\ell^2(\kappa)$ have the WLP?
Given a Banach space $X$ and a bounded function $f: [0,1] \to X$, non-necessarily Riemann integrable, one can consider the subset $I(f)$ of $X$ consisting of all limits of sequences of Riemann sums of $f$ based on partitions with diameter tending to $0$. The properties of the set $I(f)$ have been widely studied by several authors, see [@kad-kad-J Appendix]. It is known that $I(f)\neq \emptyset$ if $X$ is separable or if $X$ is super-reflexive; in the second case, the set $I(f)$ is convex. Recently a similar concept of “limit set” based on the Henstock-Kurzweil integral has been discussed in [@cap-alt]. One of the open questions along this line is:
\[problem:KadetsKadets\] Is the set $I(f)$ convex for any bounded function $f:[0,1]\to c_0$?
Pettis integral {#subsection:Pettis}
---------------
Let $(\Omega,\Sigma,\mu)$ be a complete probability space and $X$ a Banach space. A function $f:\Omega \to X$ is said to be [*scalarly bounded*]{} if there is $M>0$ such that for each $x^*\in B_{X^*}$ we have $|x^*f|\leq M$ $\mu$-a.e. A function $f:\Omega \to X$ is called [*Pettis integrable*]{} if: (i) $x^{*}f$ is integrable for every $x^{*}\in X^{*}$; (ii) for each $A\in \Sigma$ there is a vector $\int_A f \, d\mu \in X$ such that $x^{*}(\int_A f \, d\mu)=\int_{A}x^{*} f \, d\mu$ for all $x^{*}\in X^{*}$. The Pettis integral theory is closely related to deep questions about pointwise compact sets of measurable real-valued functions (see e.g. [@tal]). The space $X$ has the [*Pettis integrability property*]{} (PIP) if every scalarly bounded and scalarly measurable function taking values in $X$ is Pettis integrable. For instance, if $(X^*,w^*)$ is angelic, then $X$ has the PIP. On the other hand, Fremlin and Talagrand [@fre-tal] proved that $\ell^\infty$ fails the PIP.
For any Pettis integrable function $f:\Omega \to X$, the map $\nu_f:\Sigma \to X$ defined by $\nu_f(A):=\int_A f \, d\mu$ is countably additive and, in particular, its range $$\mathcal{R}(\nu_f)=\Big\{\int_A f \, d\mu: \ A\in \Sigma\Big\} {\subseteq}X$$ is relatively weakly compact (see Subsection \[subsection:range\]). One can approximate $f$ nicely by simple functions whenever the set $\mathcal{R}(\nu_f)$ is separable or, even better, relatively norm compact. The paper [@fre-tal] exhibited the first example of a Pettis integrable function $f$ such that $\mathcal{R}(\nu_f)$ is not relatively norm compact. Such examples are very pathological, since the relative norm compactness of $\mathcal{R}(\nu_f)$ is guaranteed in many cases, for instance: if $\mu$ is a perfect measure (e.g. a Radon measure), if $X\not\supseteq \ell^{1}(\omega_{1})$ (e.g. if $(X^*,w^*)$ is angelic), if $X$ is the dual of a Banach space not containing $\ell^1$, etc. (see e.g. [@Musial; @mus3; @tal]). There are no characterizations of those Banach spaces $X$ and complete probability spaces $(\Omega,\Sigma,\mu)$ for which $\mathcal{R}(\nu_f)$ is relatively norm compact (resp. separable) for every Pettis integrable function $f:\Omega \to X$.
\[problem:Plebanek\] Suppose $X$ has the PIP. Is $\mathcal{R}(\nu_f)$ separable for every Pettis integrable function $f:\Omega \to X$?
A *quasi-Radon* probability space (see e.g. [@freMT-4 Chapter 41]) is a quadruple $(\Omega,\mathfrak{T},\Sigma,\mu)$, where $(\Omega,\Sigma,\mu)$ is a complete probability space and $\mathfrak{T}\subset \Sigma$ is a topology on $\Omega$ such that $\mu$ is inner regular with respect to the collection of all closed sets, and $\mu(\bigcup \mathcal{G})=\sup\{\mu(G):
G\in \mathcal{G}\}$ for every upwards directed family $\mathcal{G}\subset
\mathfrak{T}$. Of course, every Radon probability space is quasi-Radon.
\[problem:FremlinQuasiRadon\] Suppose $\mu$ is quasi-Radon. Is $\mathcal{R}(\nu_f)$ relatively norm compact for every Pettis integrable function $f:\Omega \to X$?
It is easy to check that all sets of the form $\mathcal{R}(\nu_f)$ are limited. Therefore, an affirmative answer to Problem \[problem:GelfandPhillips\] (see Subsection \[subsection:SequencesDual\]) would solve the following:
\[problem:Talagrand\] Suppose $X\not\supseteq c_0$. Is $\mathcal{R}(\nu_f)$ relatively norm compact for every Pettis integrable function $f:\Omega \to X$?
McShane integral {#subsection:McShane}
----------------
Let $X$ be a Banach space. A function $f:[0,1] \to X$ is said to be [*McShane integrable*]{}, with integral $x\in X$, if it satisfies the following condition: for every $\epsilon>0$ there is $\delta:[0,1] \to (0,\infty)$ such that $\|\sum_{j=1}^{p}\lambda(I_{j})f(t_{j})-x\|< \epsilon$ for every partition $\{I_1,\dots,I_p\}$ of $[0,1]$ into finitely many non-overlapping closed intervals and every choice of points $t_j\in [0,1]$ such that $I_j{\subseteq}(t_j-\delta(t_j),t_j+\delta(t_j))$. The notion of McShane integrability lies between Bochner and Pettis integrability; moreover, McShane and Pettis integrability coincide for functions with values in separable Banach spaces, [@fre-men]. Let $\mathcal{U}$ be the class of those Banach spaces $X$ such that every Pettis integrable function $f:[0,1]\to X$ is McShane integrable. Within the context of non-separable Banach spaces, Di Piazza and Preiss [@dip-pre] proved that super-reflexive spaces and $c_0(\Gamma)$ (for any set $\Gamma$) belong to $\mathcal{U}$, raising the problem of whether the same holds for any WCG space. In [@rod10] we proved that the space $L^1(\mu)$ (for any finite measure $\mu$) belongs to $\mathcal{U}$. Later we showed that all (subspaces of) Hilbert generated spaces belong to the class $\mathcal{U}$ (see [@dev-rod]), thus generalizing all previous results in this direction. In general, the answer to the question of Di Piazza and Preiss is negative: in [@avi-alt] we constructed Pettis integrable functions, with values in reflexive Banach spaces, that are not McShane integrable. The solution uses techniques of infinite combinatorics related to Fremlin’s “problem DU” [@freDU].
Let $\cK$ denote the class of all compacta $K$ such that $C(K)\in \mathcal{U}$. The results of [@avi-alt; @dev-rod] show that $\cK$ contains every uniform Eberlein compact, but not every Eberlein compact. Recall that a compact $K$ is said to be [*uniform Eberlein*]{} if it is homeomorphic to a weakly compact subset of a Hilbert space or, equivalently, if $C(K)$ is Hilbert generated. The paper [@fab-mcs] provides another approach to the results of [@avi-alt; @dev-rod] and discusses some classical examples of Eberlein compacta which are not uniform Eberlein, in connection with the following open problem:
\[problem:McShane\] Are there non-uniform Eberlein compacta in the class $\cK$?
The McShane integral can be generalized to the case of functions defined on a quasi-Radon probability space [@fre5]. In this context the following question arises (it has affirmative answer if $\mathfrak{T}_1$ has a countable basis [@avi-alt]):
\[problem:FremlinMcShane\] Let $f:\Omega \to X$ be a function defined in a complete probability space $(\Omega,\Sigma,\mu)$ and taking values in a Banach space $X$. Let $\mathfrak{T}_1$ and $\mathfrak{T}_2$ be two topologies on $\Omega$ for which $\mu$ is quasi-Radon. Suppose $f$ is McShane integrable with respect to $\mathfrak{T}_1$. Is $f$ McShane integrable with respect to $\mathfrak{T}_2$?
Problems in vector measures {#section:MedidasVectoriales}
===========================
Range of a vector measure {#subsection:range}
-------------------------
By a [*vector measure*]{} we mean a countably additive map $\nu:\Sigma\to X$ defined on a $\sigma$-algebra $\Sigma$ and taking values in a Banach space $X$. We denote by $|\nu|$ (resp. $\|\nu\|$) the variation (resp. semivariation) of $\nu$. A theorem of Bartle, Dunford and Schwartz [@bar-alt] ensures that the range of $\nu$, i.e. the set $\mathcal{R}(\nu)=\{\nu(E):E\in \Sigma\}$, is relatively weakly compact. In fact, the range has the Banach-Saks property (every sequence in it admits a subsequence whose arithmetic means are norm convergent), [@die-sei]. It is natural to ask which subsets of a Banach space (e.g. sequences with some convergence or summability property) are contained in the range of a vector measure (see e.g. [@sof]). It is known that the closed unit ball $B_X$ of a Banach space $X$ [*is*]{} the range of a vector measure if and only if $X^*$ is isometric to a reflexive subspace of $L^1(\mu)$ for some finite measure $\mu$ (e.g. $X=L^p[0,1]$ or $X=\ell^p$ with $2\leq p < \infty$), [@ana-die]. However, it seems that there is no characterization of those Banach spaces $X$ for which $B_X$ is [*contained*]{} in the range of an $X$-valued vector measure. This class of spaces is strictly contained in that of super-reflexive spaces, [@cas-san].
\[problem:ball\] Characterize the Banach spaces $X$ for which $B_X$ is contained in the range of an $X$-valued vector measure.
\[problem:Sofi\] Let $K$ be a weakly compact subset of $c_0$. Is $K$ contained in the range of a $c_0$-valued vector measure?
Certain properties of a vector measure are determined by its range. For instance, if $\nu_1$ and $\nu_2$ are vector measures such that ${\mathcal{R}}(\nu_1)={\mathcal{R}}(\nu_2)$, then $|\nu_1|$ is finite if and only if $|\nu_2|$ is finite, [@rodpiazza-2]. In such case, $\nu_1$ is Bochner differentiable with respect to $|\nu_1|$ if and only if $\nu_2$ is Bochner differentiable with respect to $|\nu_2|$, [@rodpiazza-1]. An example of Fremlin and Talagrand [@fre-tal] was used in [@rod-rom] to show that the previous result is not true in general for the Pettis integral.
\[problem:RPRM\] Let $\nu_1$ and $\nu_2$ be vector measures of finite variation such that ${\mathcal{R}}(\nu_1)={\mathcal{R}}(\nu_2)$. Suppose further that $|\nu_1|$ and $|\nu_2|$ are perfect measures. Do $\nu_1$ and $\nu_2$ have the same Pettis differentiability character with respect to their variations?
$L^1$ spaces of a vector measure {#subsection:L1m}
--------------------------------
Let $(\Omega,\Sigma)$ be a measurable space, $X$ a Banach space and $\nu: \Sigma \to X$ a vector measure. A measurable function $f:\Omega \to {\mathbb{R}}$ is said to be [*$\nu$-integrable*]{} if: (i) it is integrable with respect to the composition $x^{*}\nu:\Sigma \to {\mathbb{R}}$ for every $x^{*}\in X^{*}$; (ii) for each $A \in \Sigma$ there is $\int_{A}f \, d\nu \in X$ such that $x^{*}(\int_{A}f \, d\nu)=\int_{A}f \, d(x^{*}\nu)$ for all $x^{*}\in X^{*}$. By identifying functions which coincide $\|\nu\|$-a.e. we obtain the Banach lattice $L^1(\nu)$ of all (equivalence classes of) $\nu$-integrable functions, with the norm $\|f\|_{L^{1}(\nu)}=\sup\{\int_{\Omega}|f| \, d|x^{*}\nu|: x^{*} \in B_{X^{*}}\}$. Every $L^1$ space of a vector measure is both WCG and a subspace of a Hilbert generated space, [@cur1; @rod15]. Several classical spaces can be represented in this way, since any order continuous Banach lattice with a weak order unit is lattice isometric to the $L^1$ space of a vector measure, [@cur1; @dep-alt].
In this theory the integration operator $I_{\nu}:L^1(\nu) \to X$, $I_{\nu}(f):=\int_\Omega f \, d\nu$, plays an important role (see e.g. [@oka-alt]). In [@rod15] we improved results of [@oka-alt2; @oka-alt3] by showing that if $I_\nu$ is completely continuous and Asplund (e.g. if $I_\nu$ is compact or absolutely $p$-summing, $1\leq p <\infty$), then $|\nu|$ is finite and $L^1(\nu)=L^1(|\nu|)$; therefore, in such case the space $L^1(\nu)$ is an “ordinary” $L^1$ space. Recall that an operator between Banach spaces is said to be [*Asplund*]{} if it factors through an Asplund space (for instance, every weakly compact operator is Asplund). The particular case in which $X$ is an Asplund space was considered in [@cal-alt-5] to give a partial answer to the following question:
Let $X$ be a Banach space not containing $\ell^1$ and let $\nu$ be an $X$-valued vector measure such that $I_\nu$ is completely continuous. Does $\nu$ have finite variation?
Curbera [@cur2] (cf. [@lip]) proved that if $\nu$ is non-atomic and $L^1(\nu)$ is separable, then there is a vector measure $\tilde{\nu}:\Sigma\to c_0$ such that $L^1(\nu)=L^1(\tilde{\nu})$ with equal norms. Lipecki [@lip2] asked if there is a “non-separable version” of that result by using $c_0(\Gamma)$ as range space for a suitable set $\Gamma$. In [@rod16] we have shown that if $\nu$ is non-atomic and ${\rm dens}(L^1(\nu))=\omega_1$, then there is a vector measure $\tilde{\nu}:\Sigma\to \ell^\infty_c(\omega_1)$ such that $L^1(\nu)=L^1(\tilde{\nu})$ with equivalent norms. Here $\ell^\infty_c(\omega_1)$ denotes the subspace of $\ell^\infty(\omega_1)$ made up of all countably supported vectors.
\[problem:Lipecki\] Suppose $\nu$ is non-atomic. Are there a set $\Gamma$ and a vector measure $\tilde{\nu}:\Sigma \to c_0(\Gamma)$ such that $L^1(\nu)=L^1(\tilde{\nu})$?
For any $g \in B_{L^{\infty}(\nu)}$ and $x^{*}\in B_{X^{*}}$ we have a functional $\gamma_{g,x^{*}} \in B_{L^{1}(\nu)^{*}}$ defined by the formula $\gamma_{g,x^{*}}(f):=\int_{\Omega} f g \, d(x^{*}\nu)$. The set $\Gamma$ consisting of all such functionals is norming and the topology $\sigma(L^1(\nu),\Gamma)$ is coarser than the weak topology of $L^1(\nu)$. However, there are non-trivial connections between both topologies, specially when $\mathcal{R}(\nu)$ is relatively norm compact. In such case, every $\sigma(L^1(\nu),\Gamma)$-convergent sequence is weakly convergent (see [@lew4; @oka]) and, if in addition $X$ is Asplund, then both topologies coincide on bounded sets [@gra-rue]. These statements can fail in the absence of some of the hypotheses, [@cur2; @cur3]. Along this line, it is relevant to know conditions ensuring that $\Gamma$ is a [*James boundary*]{} of $L^1(\nu)$, meaning that for every $f\in L^1(\nu)$ there is $\gamma_{g,x^{*}}\in \Gamma$ such that $\|f\|_{L^{1}(\nu)}=\gamma_{g,x^{*}}(f)$. This happens when $\mathcal{R}(\nu)$ is relatively norm compact or $\nu$ is a positive vector measure with values in a Banach lattice, [@cal-alt-6; @man-J]. We stress that, in an arbitrary Banach space, every bounded subset which is compact with respect to the topology of pointwise convergence on a James boundary is weakly compact: this striking result of Pfitzner [@pfi-J] answered a long-standing question known as “the boundary problem”. It is also worth mentioning that James boundaries are useful to study summability in Banach spaces, see [@cal-alt-7; @fon2-J; @gas].
\[problem:frontera\] Characterize when $\Gamma$ is a James boundary of $L^1(\nu)$.
Problems in Lebesgue-Bochner spaces {#section:Bochner}
===================================
Let $(\Omega,\Sigma,\mu)$ be a probability space, $X$ a Banach space and $1\leq p<\infty$. Denote by $L^p(\mu,X)$ the Banach space of all (equivalence classes of) strongly measurable functions $f:\Omega\to X$ such that $$\|f\|_{L^p(\mu,X)}=\Big(\int_\Omega \|f(\cdot)\|^p \, d\mu\Big)^{1/p}<\infty.$$ These spaces are usually called Lebesgue-Bochner spaces. We refer the reader to [@cem-men; @die-uhl-J; @lin-J] for complete information on this topic.
Topological properties {#subsection:TopologiaBochner}
----------------------
Two important contributions on the weak topology of Lebesgue-Bochner spaces are the characterization of weakly compact sets of Diestel, Ruess and Schachermayer [@die-alt2], and the parametric version of Rosenthal’s $\ell^1$-theorem due to Talagrand [@tal11]. Nowadays the topic still offers challenging problems. A generic question is as follows: [*if $X$ satisfies a certain property (P), does $L^p(\mu,X)$ enjoy (P) as well?*]{} The factorization theorem of Davis, Figiel, Johnson and Pełczyński paves the way to show that $L^p(\mu,X)$ is WCG if $X$ is, [@die7]. Talagrand [@tal1] proved that the properties of being weakly K-analytic (WKA) and weakly countably determined (WCD) also pass from $X$ to $L^p(\mu,X)$. It seems to be unknown if the same happens with weaker properties like having $w^*$-sequentially compact dual ball.
\[problem:ballSC\] Is $B_{L^p(\mu,X)^*}$ $w^*$-sequentially compact if $B_{X^*}$ is?
Schlüchtermann and Wheeler [@sch-whe] asked if the space $L^1(\mu,X)$ is SWCG when $X$ is SWCG. The answer is affirmative for reflexive spaces and $L^1(\mu)$ (for any finite measure $\mu$), but the general case remains open. In [@laj-rod] we gave a positive answer to the problem when restricted to weakly compact decomposable sets (e.g. sets of selectors of multi-functions). Recently we provided related results by showing, for instance, that for $1<p<\infty$ (and infinite-dimensional $L^p(\mu)$), the space $L^p(\mu,X)$ is a subspace of an SWCG space if and only if $X$ is reflexive, [@rod13]. On the other hand, the “strong” versions of properties WKA and WCD were discussed in [@kam-mer; @kam-mer2; @mer-sta-2] and it is natural to ask if they pass to the corresponding Lebesgue-Bochner space.
\[problem:SW\] Is the space $L^1(\mu,X)$ SWCG whenever $X$ is?
\[problem:SWKA\] Is the space $L^p(\mu,X)$ strongy WKA (resp. strongly WCD) whenever $X$ is?
Structural properties {#subsection:EstructuraBochner}
---------------------
Kwapień proved that $L^p(\mu,X) \not\supseteq c_0$ if (and only if) $X\not \supseteq c_0$. For $1<p<\infty$ we have an analogous result about copies of $\ell^1$, due to Bourgain, Maurey and Pisier. Nice proofs of these results can be found e.g. in [@cem-men Chapter 2]. In another direction, Diestel and Uhl [@die-uhl-J] asked if the space $L^p([0,1],X)$ has unconditional basis when $1<p<\infty$ and $X$ has unconditional basis. Aldous [@ald] answered in the negative this question by showing that the existence of an unconditional basis in $L^p([0,1],X)$ implies that $X$ is super-reflexive.
\[problem:Aldous\] Characterize those Banach spaces $X$ such that $L^p([0,1],X)$ has unconditional basis ($1<p<\infty$).
While $\ell^2$ and $L^2[0,1]$ are isomorphic, the previous result of Aldous implies that the spaces $\ell^2(X)$ and $L^2([0,1],X)$ are not isomorphic in general (take $X=c_0$, for instance). On the other hand, $\ell^2$ is isomorphic to a subspace of $L^1[0,1]$, but in general $\ell^2(X)$ is not isomorphic to a subspace of $L^1([0,1],X)$. An example of such situation is obtained by taking $X=\ell^1$, since $L^1([0,1],\ell^1)$ is SWCG but $\ell^2(\ell^1)$ is not a subspace of an SWCG space (see the previous subsection).
\[problem:DiestelQ7\] For which separable Banach spaces $X$ the spaces $\ell^2(X)$ and $L^2([0,1],X)$ are isomorphic?
\[problem:DiestelQ8\] For which separable Banach spaces $X$ the space $L^1([0,1],X)$ contains a copy of $\ell^2(X)$?
The Banach space $L^\infty(\mu,X)$, consisting of all (equivalence classes of) essentially bounded and strongly measurable functions, also plays a relevant role in the theory (see e.g. [@cem-men Chapter 5]). While $\ell^\infty$ and $L^\infty[0,1]$ are isomorphic, in general the spaces $\ell^\infty(X)$ and $L^\infty([0,1],X)$ are not. For instance, $L^\infty([0,1],\ell^\infty)$ contains a [*complemented*]{} copy of $c_0$, so it cannot be isomorphic to $\ell^\infty(\ell^\infty) \simeq \ell^\infty$.
\[problem:CM\] When are $\ell^\infty(X)$ and $L^\infty([0,1],X)$ isomorphic?
Scalar convergence {#subsection:DilworthGirardi}
------------------
Edgar [@edg6] proved that $X$ is Asplund if and only if the following condition holds: every sequence of strongly measurable functions $f_n:[0,1] \to X$ which is bounded in $L^\infty([0,1],X)$ and [*converges scalarly*]{} to $0$ a.e. (i.e. for every $x^* \in X^*$ the sequence $(x^*f_n)$ converges to $0$ a.e.) is weakly convergent to $0$ a.e., that is, there is a $\lambda$-null set $E{\subseteq}[0,1]$ such that the sequence $(f_n(t))$ is weakly null in $X$ for every $t\in [0,1]\setminus E$. On the other hand, Dilworth and Girardi [@dil-gir] studied the scalar version of the classical result relating convergence in measure with a.e. convergence. Namely, they discussed the following property (for $1\leq p \leq \infty$):
- Every sequence of strongly measurable functions $f_n:[0,1] \to X$ which is bounded in $L^p([0,1],X)$ and converges [*scalarly in measure*]{} to $0$ (i.e. for every $x^* \in X^*$ the sequence $(x^*f_n)$ converges to $0$ in measure) admits a subsequence that converges scalarly to $0$ a.e.
A couple of open questions along this line are:
\[problem:DG1\] Does $L^1[0,1]$ have property ($D_\infty$)?
\[problem:DG2\] Are there reflexive (infinite-dimensional) Banach spaces with property ($D_1$)?
Acknowledgements {#acknowledgements .unnumbered}
----------------
Research partially supported by [*Ministerio de Economía y Competitividad - FEDER*]{} (project MTM2014-54182-P). This work was also partially supported by the research project 19275/PI/14 funded by [*Fundación Séneca - Agencia de Ciencia y Tecnología de la Región de Murcia*]{} within the framework of [*PCTIRM 2011-2014*]{}.
\#1[0=]{} \[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{}
[100]{}
F. Albiac and N. J. Kalton, *Topics in [B]{}anach space theory*, Graduate Texts in Mathematics, vol. 233, Springer, New York, 2006.
D. J. Aldous, *Unconditional bases and martingales in [$L\sb{p}(F)$]{}*, Math. Proc. Cambridge Philos. Soc. **85** (1979), no. 1, 117–123.
A. Alexiewicz and W. Orlicz, *Remarks on [R]{}iemann-integration of vector-valued functions*, Studia Math. **12** (1951), 125–132.
R. Anantharaman and J. Diestel, *Sequences in the range of a vector measure*, Comment. Math. Prace Mat. **30** (1991), no. 2, 221–235.
J. P. Aubin and H. Frankowska, *Set-valued analysis*, Systems & Control: Foundations & Applications, vol. 2, Birkhäuser Boston Inc., Boston, MA, 1990.
A. Avil[é]{}s, A. J. Guirao, S. Lajara, J. Rodr[í]{}guez, and P. Tradacete, *Weakly compactly generated [B]{}anach lattices*, Studia Math. (to appear), arXiv:1512.08628.
A. Avil[é]{}s and W. Marciszewski, *Extension operators on balls and on spaces of finite sets*, Studia Math. **227** (2015), no. 2, 165–182.
A. Avil[é]{}s, G. Plebanek, and J. Rodr[í]{}guez, *The [M]{}c[S]{}hane integral in weakly compactly generated spaces*, J. Funct. Anal. **259** (2010), no. 11, 2776–2792.
A. Avil[é]{}s, G. Plebanek, and J. Rodr[í]{}guez, *Measurability in [$C(2\sp \kappa)$]{} and [K]{}unen cardinals*, Israel J. Math. **195** (2013), no. 1, 1–30.
A. Avil[é]{}s, G. Plebanek, and J. Rodr[í]{}guez, *On [B]{}aire measurability in spaces of continuous functions*, J. Math. Anal. Appl. **398** (2013), no. 1, 230–238.
A. Avil[é]{}s, G. Plebanek, and J. Rodr[í]{}guez, *A weak[$\sp *$]{} separable [$C(K)\sp *$]{} space whose unit ball is not weak[$\sp *$]{} separable*, Trans. Amer. Math. Soc. **366** (2014), no. 9, 4733–4753.
A. Avil[é]{}s, G. Plebanek, and J. Rodr[í]{}guez, *Tukey classification of some ideals on $\omega$ and the lattices of weakly compact sets in [B]{}anach spaces*, preprint, arXiv:1406.5526v3.
A. Avil[é]{}s and J. Rodr[í]{}guez, *Convex combinations of weak[$\sp
*$]{}-convergent sequences and the [M]{}ackey topology*, preprint, arXiv:1601.05825.
R. G. Bartle, N. Dunford, and J. Schwartz, *Weak compactness and vector measures*, Canad. J. Math. **7** (1955), 289–305.
J. Bourgain and J. Diestel, *Limited operators and strict cosingularity*, Math. Nachr. **119** (1984), 55–58.
R. D. Bourgin, *Geometric aspects of convex sets with the [R]{}adon-[N]{}ikodým property*, Lecture Notes in Mathematics, vol. 993, Springer-Verlag, Berlin, 1983.
D. K. Burke and R. Pol, *On [B]{}orel sets in function spaces with the weak topology*, J. London Math. Soc. (2) **68** (2003), no. 3, 725–738.
J. M. Calabuig, S. Lajara, J. Rodr[í]{}guez, and E. A. S[á]{}nchez-P[é]{}rez, *Compactness in [$L^1$]{} of a vector measure*, Studia Math. **225** (2014), no. 3, 259–282.
J. M. Calabuig, J. Rodr[í]{}guez, and E. A. S[á]{}nchez-P[é]{}rez, *Weak continuity of [R]{}iemann integrable functions in [L]{}ebesgue-[B]{}ochner spaces*, Acta Math. Sin. (Engl. Ser.) **26** (2010), no. 2, 241–248.
J. M. Calabuig, J. Rodr[í]{}guez, and E. A. S[á]{}nchez-P[é]{}rez, *On completely continuous integration operators of a vector measure*, J. Convex Anal. **21** (2014), no. 3, 811–818.
J. M. Calabuig, J. Rodr[í]{}guez, and E. A. S[á]{}nchez-P[é]{}rez, *Summability in [$L^1$]{} of a vector measure*, Math. Nachr. (to appear), DOI: 10.1002/mana201600020
D. Caponetti, L. Di Piazza, and V. Kadets, *Description of the limit set of [H]{}enstock-[K]{}urzweil integral sums of vector-valued functions*, J. Math. Anal. Appl. **421** (2015), no. 2, 1151–1162.
B. Cascales, V. Kadets, and J. Rodr[í]{}guez, *Measurable selectors and set-valued [P]{}ettis integral in non-separable [B]{}anach spaces*, J. Funct. Anal. **256** (2009), no. 3, 673–699.
B. Cascales, V. Kadets, and J. Rodr[í]{}guez, *Measurability and selections of multi-functions in [B]{}anach spaces*, J. Convex Anal. **17** (2010), no. 1, 229–240.
J. M. F. Castillo and F. S[á]{}nchez, *Remarks on the range of a vector measure*, Glasgow Math. J. **36** (1994), no. 2, 157–161.
P. Cembranos and J. Mendoza, *Banach spaces of vector-valued functions*, Lecture Notes in Mathematics, vol. 1676, Springer-Verlag, Berlin, 1997.
G. P. Curbera, *Operators into [$L\sp 1$]{} of a vector measure and applications to [B]{}anach lattices*, Math. Ann. **293** (1992), no. 2, 317–330.
G. P. Curbera, *When [$L\sp 1$]{} of a vector measure is an [AL]{}-space*, Pacific J. Math. **162** (1994), no. 2, 287–303.
G. P. Curbera, *Banach space properties of [$L\sp 1$]{} of a vector measure*, Proc. Amer. Math. Soc. **123** (1995), no. 12, 3797–3806.
R. Deville, G. Godefroy, and V. Zizler, *Smoothness and renormings in [B]{}anach spaces*, Pitman Monographs and Surveys in Pure and Applied Mathematics, vol. 64, Longman Scientific & Technical, Harlow, 1993.
R. Deville and J. Rodr[í]{}guez, *Integration in [H]{}ilbert generated [B]{}anach spaces*, Israel J. Math. **177** (2010), 285–306.
L. Di Piazza and D. Preiss, *When do [M]{}c[S]{}hane and [P]{}ettis integrals coincide?*, Illinois J. Math. **47** (2003), no. 4, 1177–1187.
J. Diestel, *[$L\sp{1}\sb{X}$]{} is weakly compactly generated if [$X$]{} is*, Proc. Amer. Math. Soc. **48** (1975), 508–510.
J. Diestel, *Some problems arising in connection with the theory of vector measures*, Séminaire [C]{}hoquet, 17e année (1977/78), [I]{}nitiation à l’analyse, [F]{}asc. 2, Secrétariat Math., Paris, 1978, pp. Exp. No. 23, 11.
J. Diestel, *Sequences and series in [B]{}anach spaces*, Graduate Texts in Mathematics, vol. 92, Springer-Verlag, New York, 1984.
J. Diestel, *Weakly compactly generated [B]{}anach lattices*, Problems presented at the conference “Integration, Vector Measures and Related Topics IV”, La Manga del Mar Menor, March 2011, http://www.um.es/beca/Murcia2011/.
J. Diestel, W. M. Ruess, and W. Schachermayer, *On weak compactness in [$L\sp 1(\mu, X)$]{}*, Proc. Amer. Math. Soc. **118** (1993), no. 2, 447–453.
J. Diestel and C. J. Seifert, *An averaging property of the range of a vector measure*, Bull. Amer. Math. Soc. **82** (1976), no. 6, 907–909.
J. Diestel and J. J. Uhl, Jr., *Vector measures*, American Mathematical Society, Providence, R.I., 1977, With a foreword by B. J. Pettis, Mathematical Surveys, No. 15.
S. J. Dilworth and M. Girardi, *On various modes of scalar convergence in [$L^0(X)$]{}*, J. Math. Anal. Appl. **259** (2001), no. 2, 660–684.
P. G. Dodds, B. de Pagter, and W. Ricker, *Reflexivity and order properties of scalar-type spectral operators in locally convex spaces*, Trans. Amer. Math. Soc. **293** (1986), no. 1, 355–380.
A. Dow and S. Shelah, *An [E]{}fimov space from [M]{}artin’s axiom*, Houston J. Math. **39** (2013), no. 4, 1423–1435.
L. Drewnowski, *On [B]{}anach spaces with the [G]{}elfand-[P]{}hillips property*, Math. Z. **193** (1986), no. 3, 405–411.
G. A. Edgar, *Measurability in a [B]{}anach space*, Indiana Univ. Math. J. **26** (1977), no. 4, 663–677.
G. A. Edgar, *Measurability in a [B]{}anach space. [II]{}*, Indiana Univ. Math. J. **28** (1979), no. 4, 559–579.
G. A. Edgar, *Asplund operators and a.e. convergence*, J. Multivariate Anal. **10** (1980), no. 3, 460–466.
P. Enflo and H. P. Rosenthal, *Some results concerning [$L\sp{p}(\mu
)$]{}-spaces*, J. Funct. Anal. **14** (1973), 325–348.
M. Fabian, *On coincidence of [P]{}ettis and [M]{}c[S]{}hane integrability*, Czechoslovak Math. J. **65(140)** (2015), no. 1, 83–106.
M. Fabian, G. Godefroy, P. H[á]{}jek, and V. Zizler, *Hilbert-generated spaces*, J. Funct. Anal. **200** (2003), no. 2, 301–323.
M. Fabian, P. Habala, P. H[á]{}jek, V. Montesinos, and V. Zizler, *Banach space theory*, CMS Books in Mathematics/Ouvrages de Mathématiques de la SMC, Springer, New York, 2011, The basis for linear and nonlinear analysis.
T. Figiel, W. B. Johnson, and A. Pe[ł]{}czy[ń]{}ski, *Some approximation properties of [B]{}anach spaces and [B]{}anach lattices*, Israel J. Math. **183** (2011), 199–231.
V. P. Fonf, *Weakly extremal properties of [B]{}anach spaces*, Mat. Zametki **45** (1989), no. 6, 83–92, 112, English translation in Math. Notes 45 (1989), no. 5-6, 488–494.
R. Frankiewicz and G. Plebanek, *Convex combinations and weak[$\sp \ast$]{} null sequences*, Bull. Polish Acad. Sci. Math. **45** (1997), no. 3, 221–225.
D. H. Fremlin, *Families of compact sets and [T]{}ukey’s ordering*, Atti Sem. Mat. Fis. Univ. Modena **39** (1991), no. 1, 29–50.
D. H. Fremlin, *The partially ordered sets of measure theory and [T]{}ukey’s ordering*, Note Mat. **11** (1991), 177–214, Dedicated to the memory of Professor Gottfried K[ö]{}the.
D. H. Fremlin, *The generalized [M]{}c[S]{}hane integral*, Illinois J. Math. **39** (1995), no. 1, 39–67.
D. H. Fremlin, *Measure theory. [V]{}ol. 4*, Torres Fremlin, Colchester, 2006, Topological measure spaces. Part I, II, Corrected second printing of the 2003 original.
D. H. Fremlin, *Problem [D]{}[U]{}*, note of 30.9.2009,\
http://www.essex.ac.uk/maths/people/fremlin/problems.htm.
D. H. Fremlin, *The [C]{}ascales-[K]{}adets-[R]{}odriguez selection theorem*, note of 27.9.2012,\
http://www.essex.ac.uk/maths/people/fremlin/preprints.htm.
D. H. Fremlin and J. Mendoza, *On the integration of vector-valued functions*, Illinois J. Math. **38** (1994), no. 1, 127–147.
D. H. Fremlin and M. Talagrand, *A decomposition theorem for additive set-functions, with applications to [P]{}ettis integrals and ergodic means*, Math. Z. **168** (1979), no. 2, 117–142.
I. Gasparis, *On a problem of [H]{}. [P]{}. [R]{}osenthal concerning operators on [$C[0,1]$]{}*, Adv. Math. **218** (2008), no. 5, 1512–1525.
N. Ghoussoub, B. Maurey, and W. Schachermayer, *Slicings, selections and their applications*, Canad. J. Math. **44** (1992), no. 3, 483–504.
E. Glasner and M. Megrelishvili, *Representations of dynamical systems on [B]{}anach spaces not containing [$l\sb 1$]{}*, Trans. Amer. Math. Soc. **364** (2012), no. 12, 6395–6424.
R. A. Gordon, *Riemann integration in [B]{}anach spaces*, Rocky Mountain J. Math. **21** (1991), no. 3, 923–949.
W. H. Graves and W. Ruess, *Compactness and weak compactness in spaces of compact-range vector measures*, Canad. J. Math. **36** (1984), no. 6, 1000–1020.
A. B. Gulisashvili, *Estimates for the [P]{}ettis integral in interpolation spaces, and a generalization of some imbedding theorems*, Soviet Math., Dokl. **25** (1982), 428–432.
P. H[á]{}jek, V. Montesinos Santaluc[í]{}a, J. Vanderwerff, and V. Zizler, *Biorthogonal systems in [B]{}anach spaces*, CMS Books in Mathematics/Ouvrages de Mathématiques de la SMC, 26, Springer, New York, 2008.
K. P. Hart, *Efimov’s problem*, Open Problems in Topology II, (E. Pearl, ed.), Elsevier Science, 2007, pp. 171–177.
R. Haydon, *Nonseparable [B]{}anach spaces*, Functional analysis: surveys and recent results, [II]{} ([P]{}roc. [S]{}econd [C]{}onf. [F]{}unctional [A]{}nal., [U]{}niv. [P]{}aderborn, [P]{}aderborn, 1979), Notas Mat., vol. 68, North-Holland, Amsterdam-New York, 1980, pp. 19–30.
R. Haydon, *Darboux integrability and separability of types in stable [B]{}anach spaces*, Seminar on functional analysis, 1983/1984, Publ. Math. Univ. Paris VII, vol. 20, Univ. Paris VII, Paris, 1984, pp. 95–115.
J. E. Jayne and C. A. Rogers, *Radon measures on [B]{}anach spaces with their weak topologies*, Serdica Math. J. **21** (1995), no. 4, 283–334.
W. B. Johnson and G. Schechtman, *Subspaces of [$L\sb p$]{} that embed into [$L\sb p(\mu)$]{} with [$\mu$]{} finite*, Israel J. Math. **203** (2014), no. 1, 211–222.
M. I. Kadets and V. M. Kadets, *Series in [B]{}anach spaces*, Operator Theory: Advances and Applications, vol. 94, Birkhäuser Verlag, Basel, 1997, Conditional and unconditional convergence, Translated from the Russian by Andrei Iacob.
N. J. Kalton and A. Pe[ł]{}czy[ń]{}ski, *Kernels of surjections from [$\mathcal{L}_1$]{}-spaces with an application to [S]{}idon sets*, Math. Ann. **309** (1997), no. 1, 135–158.
K. K. Kampoukos and S. K. Mercourakis, *A new class of weakly countably determined [B]{}anach spaces*, Fund. Math. **208** (2010), no. 2, 155–171.
K. K. Kampoukos and S. K. Mercourakis, *On a certain class of [$\mathcal{K}_{\sigma\delta}$]{} [B]{}anach spaces*, Topology Appl. **160** (2013), no. 9, 1045–1060.
E. Klein and A. C. Thompson, *Theory of correspondences*, Canadian Mathematical Society Series of Monographs and Advanced Texts, John Wiley & Sons Inc., New York, 1984, Including applications to mathematical economics, A Wiley-Interscience Publication.
K. Kuratowski and C. Ryll-Nardzewski, *A general theorem on selectors*, Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys. **13** (1965), 397–403.
S. Lajara and J. Rodr[í]{}guez, *Lebesgue-[B]{}ochner spaces, decomposable sets and strong weakly compact generation*, J. Math. Anal. Appl. **389** (2012), no. 1, 665–669.
D. R. Lewis, *Conditional weak compactness in certain inductive tensor products*, Math. Ann. **201** (1973), 201–209.
P.-K. Lin, *Köthe-[B]{}ochner function spaces*, Birkhäuser Boston Inc., Boston, MA, 2004.
Z. Lipecki, *Semivariations of a vector measure*, Acta Sci. Math. (Szeged) **76** (2010), no. 3-4, 411–425.
Z. Lipecki, *The variations and semivariations of a vector measure*, Invited lecture at the conference “Integration, Vector Measures and Related Topics VI”, Bedlewo, June 2014.
G. Manjabacas, *Topologies associated to norming sets in [B]{}anach spaces*, Ph.D. Thesis (Spanish), Universidad de Murcia, 1998. Available at http://webs.um.es/beca/dissertationstudents.html.
W. Marciszewski and G. Plebanek, *On [B]{}orel structures in the [B]{}anach space [$C(\beta\omega)$]{}*, J. Funct. Anal. **266** (2014), no. 6, 4026–4036.
W. Marciszewski and R. Pol, *On some problems concerning [B]{}orel structures in function spaces*, Rev. R. Acad. Cienc. Exactas Fís. Nat. Ser. A Math. RACSAM **104** (2010), no. 2, 327–335.
G. Mart[í]{}nez-Cervantes, *Riemann integrability versus weak continuity*, J. Math. Anal. Appl. **438** (2016), no. 2, 840–855.
G. Mart[í]{}nez-Cervantes, *On weakly [R]{}adon-[N]{}ikodým compact spaces*, Israel J. Math. (to appear), arXiv:1509.05324.
S. Mercourakis, *Some remarks on countably determined measures and uniform distribution of sequences*, Monatsh. Math. **121** (1996), no. 1-2, 79–111.
S. Mercourakis and E. Stamati, *A new class of weakly [K]{}-analytic [B]{}anach spaces*, Comment. Math. Univ. Carolin. **47** (2006), no. 2, 291–312.
K. Musia[ł]{}, *Topics in the theory of [P]{}ettis integration*, Rend. Istit. Mat. Univ. Trieste **23** (1991), no. 1, 177–262 (1993), School on Measure Theory and Real Analysis (Grado, 1991).
K. Musia[ł]{}, *Pettis integral*, Handbook of measure theory, Vol. I, II, North-Holland, Amsterdam, 2002, pp. 531–586.
K. Musia[ł]{}, *Pettis integrability of multifunctions with values in arbitrary [B]{}anach spaces*, J. Convex Anal. **18** (2011), no. 3, 769–810.
K. Musia[ł]{}, *Approximation of [P]{}ettis integrable multifunctions with values in arbitrary [B]{}anach spaces*, J. Convex Anal. **20** (2013), no. 3, 833–870.
K. M. Naralenkov, *Asymptotic structure of [B]{}anach spaces and [R]{}iemann integration*, Real Anal. Exchange **33** (2008), no. 1, 111–124.
S. Okada, *The dual space of [$L\sp 1(\mu)$]{} for a vector measure [$\mu$]{}*, J. Math. Anal. Appl. **177** (1993), no. 2, 583–599.
S. Okada, W. J. Ricker, and L. Rodr[í]{}guez-Piazza, *Compactness of the integration operator associated with a vector measure*, Studia Math. **150** (2002), no. 2, 133–149.
S. Okada, W. J. Ricker, and L. Rodr[í]{}guez-Piazza, *Operator ideal properties of vector measures with finite variation*, Studia Math. **205** (2011), no. 3, 215–249.
S. Okada, W. J. Ricker, and E. A. S[á]{}nchez P[é]{}rez, *Optimal domain and integral extension of operators*, Operator Theory: Advances and Applications, vol. 180, Birkhäuser Verlag, Basel, 2008, Acting in function spaces.
H. Pfitzner, *Boundaries for [B]{}anach spaces determine weak compactness*, Invent. Math. **182** (2010), no. 3, 585–604.
G. Plebanek, *On [P]{}ettis integrals with separable range*, Colloq. Math. **64** (1993), no. 1, 71–78.
A. Plichko, *Three sequential properties of dual [B]{}anach spaces in the weak[$\sp *$]{} topology*, Topology Appl. **190** (2015), 93–98.
J. Rodr[í]{}guez, *On the equivalence of [M]{}c[S]{}hane and [P]{}ettis integrability in non-separable [B]{}anach spaces*, J. Math. Anal. Appl. **341** (2008), no. 1, 80–90.
J. Rodr[í]{}guez, *Weak [B]{}aire measurability of the balls in a [B]{}anach space*, Studia Math. **185** (2008), no. 2, 169–176.
J. Rodr[í]{}guez, *On the [SWCG]{} property in [L]{}ebesgue-[B]{}ochner spaces*, Topology Appl. **196** (2015), no. part A, 208–216.
J. Rodr[í]{}guez, *Factorization of vector measures and their integration operators*, Colloq. Math. **144** (2016), no. 1, 115–125.
J. Rodr[í]{}guez, *On non-separable [$L^1$]{}-spaces of vector measures*, preprint.
L. Rodr[í]{}guez-Piazza, *The range of a vector measure determines its total variation*, Proc. Amer. Math. Soc. **111** (1991), no. 1, 205–214.
L. Rodr[í]{}guez-Piazza, *Derivability, variation and range of a vector measure*, Studia Math. **112** (1995), no. 2, 165–187.
L. Rodr[í]{}guez-Piazza and M. C. Romero-Moreno, *Conical measures and properties of a vector measure determined by its range*, Studia Math. **125** (1997), no. 3, 255–270.
H. P. Rosenthal, *A characterization of [B]{}anach spaces containing [$l\sp{1}$]{}*, Proc. Nat. Acad. Sci. U.S.A. **71** (1974), 2411–2413.
G. Schl[ü]{}chtermann and R. F. Wheeler, *On strongly [WCG]{} [B]{}anach spaces*, Math. Z. **199** (1988), no. 3, 387–398.
Th. Schlumprecht, *Limited sets in [B]{}anach spaces*, Ph.D. Thesis, Ludwig-Maximilians-Universität München, 1987.
Th. Schlumprecht, *Limited sets in [$C(K)$]{}-spaces and examples concerning the [G]{}elfand-[P]{}hillips-property*, Math. Nachr. **157** (1992), 51–64.
M. A. Sofi, *Structural properties of [B]{}anach and [F]{}réchet spaces determined by the range of vector measures*, Extracta Math. **22** (2007), no. 3, 257–296.
S. Solecki and S. Todorcevic, *Cofinal types of topological directed orders*, Ann. Inst. Fourier (Grenoble) **54** (2004), no. 6, 1877–1911 (2005).
M. Talagrand, *Comparaison des boreliens d’un espace de [B]{}anach pour les topologies fortes et faibles*, Indiana Univ. Math. J. **27** (1978), no. 6, 1001–1004.
M. Talagrand, *Espaces de [B]{}anach faiblement [${\mathcal K}$]{}-analytiques*, Ann. of Math. (2) **110** (1979), no. 3, 407–438.
M. Talagrand, *Séparabilité vague dans l’espace des mesures sur un compact*, Israel J. Math. **37** (1980), no. 1-2, 171–180.
M. Talagrand, *Pettis integral and measure theory*, Mem. Amer. Math. Soc. **51** (1984), no. 307, ix+224.
M. Talagrand, *Weak [C]{}auchy sequences in [$L\sp{1}(E)$]{}*, Amer. J. Math. **106** (1984), no. 3, 703–724.
M. Valadier, *Multi-applications mesurables à valeurs convexes compactes*, J. Math. Pures Appl. (9) **50** (1971), 265–297.
E. K. van Douwen, *The integers and topology*, Handbook of set-theoretic topology, North-Holland, Amsterdam, 1984, pp. 111–167.
D. van Dulst, *Characterizations of [B]{}anach spaces not containing [$l{\sp
1}$]{}*, CWI Tract, vol. 59, Centrum voor Wiskunde en Informatica, Amsterdam, 1989.
C. Wang and K. Wan, *On the weak property of [L]{}ebesgue of [$L\sp
1(\Omega,\Sigma,\mu)$]{}*, Rocky Mountain J. Math. **31** (2001), no. 2, 697–703.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Properties of the Roper resonance, the first scalar excitation of the nucleon, are determined. Pole positions and residues of the $P_{11}$ partial wave are studied in a combined analysis of pion- and photo-induced reactions. We find the Roper pole at $\{(1371\pm7)-i(92\pm10)\}$MeV and an elasticity of $0.61\pm 0.03$. The largest decay coupling is found for the $N\sigma$ ($\sigma=(\pi\pi)$-$S$-wave). The analysis is based on new data on $\gamma p\to p\pi^0\pi^0$ for photons in the energy range from the two-pion threshold to 820MeV from TAPS at Mainz and from 0.4 to 1.3GeV from Crystal Barrel at Bonn and includes further data from other experiments. The partial wave analysis excludes the possibility that the Roper resonance is split into two states with different partial decay widths.\
[*PACS: 11.80.Et, 13.30.-a, 13.40.-f, 13.60.Le*]{}
address:
- 'Helmholtz-Institut für Strahlen- und Kernphysik der Universität Bonn, Germany'
- 'Petersburg Nuclear Physics Institute, Gatchina, Russia'
- 'Physikalisches Institut, Universität Basel, Switzerland'
- 'II. Physikalisches Institut, Universität Giessen'
- 'Institut für Kernphysik, Universität Mainz, Germany'
- 'Department of Physics and Astronomy, University of Glasgow, UK'
- 'Physikalisches Institut, Universität Erlangen, Germany'
- 'Physikalisches Institut, Universität Bonn, Germany'
- 'KVI, Groningen, Netherlands'
- 'Department of Physics, Florida State University, USA'
- 'Physikalisches Institut, Universität Bochum, Germany'
author:
- 'A.V. Sarantsev'
- 'M. Fuchs'
- 'M. Kotulla'
- 'U. Thoma'
- 'J. Ahrens'
- 'J.R.M. Annand'
- 'A.V. Anisovich'
- 'G. Anton'
- 'R. Bantes'
- 'O. Bartholomy'
- 'R. Beck'
- 'Yu. Beloglazov'
- 'R. Castelijns'
- 'V. Crede'
- 'A. Ehmanns'
- 'J. Ernst'
- 'I. Fabry'
- 'H. Flemming'
- 'A. Fösel'
- 'Chr. Funke'
- 'R. Gothe'
- 'A. Gridnev'
- 'E. Gutz'
- 'St. Höffgen'
- 'I. Horn'
- 'J. Hößl'
- 'D.Hornidge'
- 'S.Janssen'
- 'J. Junkersfeld'
- 'H. Kalinowsky'
- 'F. Klein'
- 'E. Klempt'
- 'H. Koch'
- 'M. Konrad'
- 'B. Kopf'
- 'B. Krusche'
- 'J. Langheinrich'
- 'H. Löhner'
- 'I. Lopatin'
- 'J. Lotz'
- 'J.C. McGeorge'
- 'I.J.D. MacGregor'
- 'H. Matthäy'
- 'D. Menze'
- 'J.G.Messchendorp'
- 'V. Metag'
- 'V.A. Nikonov'
- 'D. Novinski'
- 'R. Novotny'
- 'M. Ostrick'
- 'H. van Pee'
- 'M. Pfeiffer'
- 'A. Radkov'
- 'G.Rosner'
- 'M.Rost'
- 'C. Schmidt'
- 'B. Schoch'
- 'G. Suft'
- 'V. Sumachev'
- 'T. Szczepanek'
- 'D. Walther'
- 'D.P.Watts'
- 'Chr. Weinheimer'
title: 'New results on the Roper resonance and the $P_{11}$ partial wave'
---
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
, and
\
The lowest-mass excitation of the nucleon, the Roper $N(1440)P_{11}$ resonance with spin and parity $J^P=1/2^+$, and the second scalar nucleon excitation $N(1710) P_{11}$ [@Yao:2006px], remain to be the most enigmatic states in baryon spectroscopy. In the bag model [@Meissner:1984un] and in the Skyrme model [@Hajduk:1984ry], the Roper resonance was interpreted as surface oscillation, also called breathing mode. In quark models, two low-mass scalar excitations of the nucleon are predicted. Using a linear confining potential and one-gluon-exchange [@Capstick:bm] or instanton-induced interactions [@Loring:2001kx], a level ordering is calculated in which the mass of the $N(1440)P_{11}$ exceeds the mass of the negative-parity state $N(1535)S_{11}$ by 80MeV; experiments find it $\sim 100$MeV below. The spacing between the two scalar excitations is predicted to be $\sim220$MeV [@Capstick:bm; @Loring:2001kx; @Loring:2001ky] while experiments find 270MeV [@Yao:2006px]. When one-gluon exchange interactions are replaced by exchanges of Goldstone bosons, the $N(1440)P_{11}$ mass can well be reproduced [@Glozman:1997ag], the $N(1710) P_{11}$ mass was not calculated. Lattice gauge calculations indicate that the first scalar excitation of the nucleon should be expected above $N(1535)S_{11}$ [@Burch:2006cc]. Compared to model and lattice predictions, the mass of the Roper resonance is too small; compared to other low-mass resonances, its width too large.
These problems would not occur if $N(1710)P_{11}$ were the first radial scalar excitation of the proton. The Roper resonance can then be interpreted within a coupled-channel meson exchange model based on an effective chiral-symmetric Lagrangian [@Krehl:1999km]; no genuine $N(1440)$ (3 quark) resonance was needed to fit $\pi N$ phase shifts and inelasticities, in agreement with [@Schneider:2006bd]. Motivated by the $Q^2$ dependence of the Roper helicity amplitude which would seem to suggest a hybrid nature [@Li:1991yb], Capstick and Page [@Capstick:1999qq] calculated masses of baryonic hybrids. Their masses were, however, too large to interpret the Roper resonance as a hybrid. The $\Theta^+(1530)$, a baryon with positive strangeness, which may have been observed in low-statistics photo-production experiments [@Nakano:2003qx; @Barmin:2003vv; @Stepanyan:2003qr; @Barth:2003es], made the Roper resonance [@Jaffe:2003sg] and/or the $N(1710)P_{11}$ [@Diakonov:1997mm] to viable pentaquark candidates. The existence of a very narrow $P_{11}$ state in the mass region 1650-1750 MeV was investigated in [@Arndt:2003ga]. The fading evidence for $\Theta^+(1530)$ [@Battaglieri:2005er; @Niccolai:2006td] makes this interpretation less attractive. Morsch and Zupranski [@Morsch:2000xi] suggested the Roper mass region might house two resonances, one at 1390MeV with a small elastic width and large coupling to $N\pi\pi$, and a second one at higher mass – around 1460MeV – with a large elastic width and small $N\pi\pi$ coupling. The former resonance was found to be produced in $\pi N$ scattering, and in $\alpha$-proton scattering using an $\alpha$ beam of $E_{\alpha}=4.2$GeV kinetic energy; the latter resonance was suggested to be excited by $\gamma
N$. The two resonances may have rather different wave functions [@Morsch:2004kn]. Studies of the reaction $pp\to pp\pi^+\pi^-$ suggested that the low-energy tail of the Roper resonance might decay to both $N\sigma$ and $\Delta\pi$ [@Patzold:2003tr]. Obviously, the $P_{11}$ partial wave is not sufficiently constrained by precision data and supports very different interpretations.
In this letter, we present data on $\gamma p \to p\pi^0\pi^0$ of the A2-TAPS collaboration at the Mainz Microtron (MAMI) electron accelerator [@ahrens:mami] and of the CB-ELSA collaboration at the Bonn ELectron Stretcher Accelerator (ELSA) [@Hillert:2006yb]. The Bonn set up and the analysis method was described briefly in the preceding paper [@Thoma:2006]. Here we give only a short summary of the TAPS setup, for more details see [@kottu:2pi0thres]. Earlier data taken at MAMI [@Harter:1997jq; @Wolf:2000qt] have smaller statistics and are not discussed here.
The photon energy at MAMI covered the range 285–820MeV. The photon energies were measured in the Glasgow tagged photon facility [@hall:tagger] with an average energy resolution of 2MeV. The TAPS detector [@novotny:taps; @gabler:response] consisted of six blocks each with 62 hexagonally shaped BaF$_2$ crystals arranged in an 8$\times$8 matrix and a forward wall with 138 BaF$_2$ crystals arranged in a 11$\times$14 rectangle. This setup covered $\approx$40% of $4\pi$. The $\gamma p \rightarrow \pi^0 \pi^0 p$ reaction channel was identified by constructing the 4-momenta of the two neutral pions from their $\gamma\gamma$ decays; proton detection was not required in the analysis. The $\pi^0$ mesons were detected via their $2\gamma$ decay and identified by their invariant mass. The mass of the missing particle was calculated from the four-momenta of the pions, and the beam energy $E_{\gamma}$ using the mass of the target proton. The resulting distribution is shown in Fig. \[gg\_vs\_gg\]a and demonstrates the unambiguous identification of the reaction $\gamma p \rightarrow \pi^0
\pi^0 p$. At incident beam energies above the $\eta$ production threshold of $E_\gamma = 707$ MeV, a possible background from the $\eta
\rightarrow 3\pi^0$ decay with one undetected $\pi^0$ can be cleanly separated from the reaction of interest (see Fig. \[gg\_vs\_gg\]a). Further details are given in [@kottu:mdm_prl].
In Fig. \[gg\_vs\_gg\]b the total cross section is displayed. Two peaks due to the second and third resonance region are observed, with peaks at $\sim 1500$ and $\sim 1700$MeV. There is good general agreement between the three data sets. The GRAAL data [@Assafiri:mv] fall off at high masses more rapidly than the CB-ELSA data. At low energies, the TAPS data fall below the CB-ELSA data while the peak cross sections of all 3 experiments agree reasonably well. The discrepancies show the difficulties of extracting total cross sections when the full phase space is not covered by the detector. Note that the extrapolation was done differently: the CB-ELSA and the A2-TAPS collaborations used the result of this partial wave analysis; the GRAAL collaboration used a simulation based on $\gamma p\to \Delta^+\pi^0$ and $\gamma p\to
p\pi^0\pi^0$ phase space. The inclusion of both, CB-ELSA and TAPS data, provides an additional tool to estimate the systematic error of the experimental data. The fit curves in Fig. \[gg\_vs\_gg\]b are discussed below.
-- -- --
-- -- --
The total cross section gives only a very superficial view of the reaction. Fig. \[masang\]a shows the experimental $p\pi^0\pi^0$ Dalitz plots, Fig. \[masang\]b,c the $p\pi^0$ and $\pi^0\pi^0$ mass distributions. The solid line represents the result of a fit, the dashed line represents the distribution of reconstructed phase space events. The projections are not corrected for detection efficiency to allow the reader to compare data and fit directly. From the $\pi^0 p$ mass distributions we conclude that the $\Delta$ isobar plays an important role in the two-pion photoproduction dynamics. The $\pi^0\pi^0$ mass distributions are featureless but show strong deviations from phase space.
Even the Dalitz plot and the projections do not carry the full information on the reaction dynamics. The full sensitivity of the data can only be exploited using an event-based likelihood fit. The data presented here were subjected to a partial wave analysis based on the Bonn-Gatchina approach [@Anisovich:2004zz; @Anisovich:2006bc]. Compared to a previous analysis [@Anisovich:2005tf; @Sarantsev:2005tg], several new data sets are included in this analysis. A list of additional data and a description of the partial wave analysis method can be found in [@cxcz]. In the case of two-particle final states (including $\gamma p\to\Lambda K^+$ and $\gamma p\to\Sigma K$), angular distributions are fitted; three-body final states like $N\pi^0\pi^0$ and $p\pi^0\eta$ [@Horn:2007] undergo an event-based likelihood fit.
--- -------------------------- --------------------------------- ----------------------------------------------------------------------------
1 $\gamma Figs. \[gg\_vs\_gg\],\[masang\] this work
p\to p\pi^0\pi^0$
2 $\gamma p\to p\pi^0$ Figs. \[cb-pi0\]-\[r-pol\] [@Bartholomy:2004uz; @van; @Pee:2007tw; @Bartalini:2005wx; @Dugger:2007bt]
3 $\pi N Fig. \[p11-amp\] [@Arndt:2006bf]
\to N\pi$
4 $\pi^- p\to n\pi^0\pi^0$ Fig. \[ball\] [@Prakhov:2004zv]
--- -------------------------- --------------------------------- ----------------------------------------------------------------------------
: \[reactions\]The reactions most important for the study of properties of the Roper resonance.
The reactions most relevant for the present analysis are collected in Table \[reactions\]. The $\pi N$ elastic scattering [@Arndt:2006bf] amplitude provides a strong constraint for $N\pi$ partial decay widths of resonances in this partial wave. The inclusion of data on $\gamma p \to p\pi^0$ [@Bartholomy:2004uz; @van; @Pee:2007tw; @Bartalini:2005wx; @Dugger:2007bt] and on $\pi^-p\to
n\pi^0\pi^0$ [@Prakhov:2004zv] over-constrains resonance properties: the three partial decay widths, $\Gamma_{N\gamma}$, $\Gamma_{N\pi}$, and $\Gamma_{N\pi\pi}$, of the Roper $N(1440)P_{11}$ resonance have to describe its properties in four reactions. In addition, the partial widths define the number of events ascribed to the Roper resonance in the reactions above. Their sum must equal the total width since the only missing channel, $N(1440)P_{11}\to N\rho$, is expected to provide a very small contribution due to the small available phase space. We believe that the tight constraints due to the use of four different reactions defines the Roper mass, width, and coupling constants with much higher reliability than analyses of individual reactions can do. The background amplitudes in the four reactions are treated independently. The data and the quality of the description are shown in Figs. \[gg\_vs\_gg\]-\[ball\].
We started the partial wave analysis from the solution given in [@Anisovich:2005tf; @Sarantsev:2005tg]. Including the new data, we found good compatibility for masses and widths of the contributing resonances. The new description of single $\pi^0$ photoproduction is shown in Fig. \[cb-pi0\]. In Figs. \[t-pol\] and \[r-pol\] we present a comparison of fit and data on target asymmetry and on the proton recoil polarization from different experiments [@SAID]. Inclusion of the latter data had an impact on the size of couplings but did not change the pole positions; the properties of the Roper resonance were nearly unaffected. The figures (not the fits) are restricted to the mass range below 1800MeV.
The $\gamma p\to p\pi^0\pi^0$ data provide new information on the N$\pi\pi$ decay modes. The $D_{33}$ amplitude gives the largest contribution to this data (see Fig. \[gg\_vs\_gg\]b). Its interference with $N(1520)D_{13}$, constructive at $\sim$1500MeV, destructive at $\sim$1600MeV, generates the dip between the two peaks in the total cross section.\
The $\Delta(1700)D_{33}$ has a large coupling to $\Delta\pi$. In the main solution, $\Delta(1700)D_{33}$ decays into $\Delta\pi$ in a relative $S$ wave. There is however a second solution with very similar likelihood in which the $\Delta(1700)D_{33}\to\Delta\pi$ decays proceed via
-- -- --
-- -- --
$D$-wave. This ambiguity results in different contributions of all $p\pi^0\pi^0$ partial waves as shown in Fig. \[gg\_vs\_gg\]b. The interference between $\Delta(1700)D_{33}$ and background contributions is responsible for the shallow dent in solution 2 of the $D_{33}$ contribution visible in Fig. \[gg\_vs\_gg\]b.
The importance of the $(\pi\pi)$-$S$-wave was already hinted at by Murphy and Laget (quoted by the GRAAL collaboration in [@Assafiri:mv]), although in a very different framework based on an effective isobaric Lagrangian. In the Laget model, the Roper resonance provided the largest contribution to $N\pi\pi$, followed by the $D_{13}$ partial wave while $D_{33}(1700)$ was very weak. The Valencia model [@GomezTejedor:1995pe; @Nacher:2000eq] is limited to $E_{\gamma}<0.84$GeV; it predicted strong contributions of $D_{13}(1520)$ and small $N(1440)P_{11}$ and $D_{33}(1700)$ contributions. Our analysis finds a very strong $D_{33}(1700)$ contribution. However, the dominant orbital angular momentum in the $D_{33}(1700)\to \Delta\pi$ decay is ambiguous giving rise to two acceptable solutions. Both solutions are fully compatible with the $D_{33}$ $\pi N$ elastic scattering amplitude (see Fig. 5 in [@Horn:2007]). The analysis presented here is constrained by a large number of data sets and exploits all two-particle correlations within the $N\pi\pi$ final state. These technical differences may very well be a reason for the discrepant results.
The $P_{11}$ amplitude for $\pi N$ elastic scattering is written in the form of a K-matrix containing three constants, describing non-resonant contributions to elastic and inelastic reactions, and a series of poles representing resonant contributions. The $P_{11}$ photoproduction amplitude is written as a K-matrix in P-vector approach (which neglects $p\gamma$ loops in the rescattering series). The photoproduction amplitude has the same poles as the scattering matrix. One constant each is introduced for reaction (1) and (2) in Table \[reactions\] describing direct $p\pi^0$ and $p\pi^0\pi^0$ production. The constants, pole positions and couplings $g_{Nx}$ to a final state $N\,x$ are free parameters of the fit. The Born term is described by a pole at the proton mass. At least two poles were required, with pole positions at 1370 and 1850MeV, respectively. In Fig. \[p11-amp\]a,b we show the $P_{11}$ amplitude for the two-pole solution. The data are well described.
As a next step, we introduced a second pole in the Roper region, a pion-induced resonance R and a second photo-induced R’. This attempt failed. The fit reduced the elastic width to the minimal allowed value of 50MeV; the overall probability of the fit became unacceptable. The resulting elastic amplitude is shown in Fig. \[p11-amp\]a,b as dashed line. We did not find any meaningful solution where the Roper region could comprise two resonances.
In [@Anisovich:2005tf; @Sarantsev:2005tg], no evidence for $N(1710)P_{11}$ was found. The increased sensitivity due to new data encouraged us to introduce a third pole in the $P_{11}$ amplitude. Fig. \[p11-amp\]c,d show the result of this fit. A small improvement due to $N(1710)P_{11}$ is observed, and also other data sets are slightly better described. The parameters of the resonance are not well defined, the pole position is found in the 1580 to 1700MeV mass range.
Introduction of the $N(1710)P_{11}$ as third pole changes the $N(1840)P_{11}$ properties. In the two-pole solution, the $N(1840)P_{11}$ resonance is narrow ($\sim 150$MeV), in the three-pole solution, the $N(1710)P_{11}$ and a $\sim 250$MeV wide $N(1840)P_{11}$ resonance interfere to reproduce the structure. Data with polarized photons and protons will hopefully clarify existence and properties of these additional resonances. Further $P_{11}$ poles are expected at larger masses. Introducing such a pole does not lead to a significant improvement of the fit.
The properties of the $N(1440)P_{11}$ resonance determined here are listed in Table \[1440\]. From the K-matrix poles and their couplings, the poles of the scattering matrix T were deduced. The speed plot $|dT|/dm$ gives $M_{\rm speed}\sim 1340$MeV. The Breit-Wigner parameters are deduced by the following method. The helicity coupling and the coupling constant for a given decay mode are calculated as residues of the T-matrix pole in the corresponding complex $s=M^2$ plane. These are complex numbers. The partial decay widths are calculated from the coupling constants and the available phase space including centrifugal barrier factor and Blatt-Weisskopf corrections [@Anisovich:2004zz]. These partial widths (including the missing width) are scaled by a common factor to reproduce the T-matrix pole position. The errors cover the range of a large variety of different PWA solutions.
[lcccccc]{} M&=&$1436\pm15$MeV&&M$_{\rm pole}$&=&$1371\pm7$MeV\
$\Gamma$&=&$335\pm40$MeV&& $\Gamma_{\rm pole}$&=&$192\pm20$MeV\
$\Gamma_{\pi N}$&=&$205\pm25$MeV&& $g_{\pi
N}$&=&$(0.51\pm0.05)\cdot e^{-i\pi\frac{(35\pm5)}{180}}$\
$\Gamma_{\sigma N}$&=&$71\pm17$MeV&&$g_{\sigma
N}$&=&$(0.82\pm0.16)\cdot e^{-i\pi\frac{(20\pm13)}{180}}$\
$\Gamma_{\pi\Delta}$&=&$59\pm15$MeV&&$g_{\pi\Delta}$&= &$(-0.57\pm0.08)\cdot e^{i\pi\frac{(25\pm20)}{180}}$\
\
The fractional yields of resonant and non-resonant parts are of course ill-defined quantities. To allow the reader to appreciate better the meaning of the results, we have set to zero the resonant or non-resonant part of the amplitude and calculated the corresponding cross sections, integrated over the Roper region (1300-1500MeV). Interferences are neglected. The results on the different photoproduction reactions are presented in Table \[1440\_contr\]. Resonant and non-resonant contributions are comparatively large and interfere destructively to yield the observed $P_{11}$ wave.
Reaction $P_{11}, obs$ (%) $P_{11}, res$ (%) $P_{11}, nonres$ (%)
--------------------------------- ------------------- ------------------- ----------------------
$\gamma p\to\pi^0 p$ $2.4\pm 0.8$ $4 \pm 1$ $ 7\pm 2$
$\gamma p\to\pi^0\pi^0 p$ $7\pm 2$ $6 \pm 2$ $11\pm 3$
$\quad\gamma p\to\Delta^+\pi^0$ $5\pm 1$ $4 \pm 1$ $7 \pm 2$
$\quad\gamma p\to p\sigma$ $4\pm 1$ $3\pm 1$ $7 \pm 2$
: \[1440\_contr\] Contributions of the $P_{11}$-wave to different photoproduction reactions, integrated over the 1300-1500MeV mass range.
This is different in pion scattering. The largest contribution to the $n\pi^0\pi^0$ final state goes via the $N(1440)P_{11}$ resonance. The complicated interference between resonant and non-resonant amplitudes may be the reason why the Roper resonance is so difficult to identify in photoproduction reactions.
$p_{\pi}$ (MeV/c) 472 551 655
-------------------------------------- ----------- ----------- -----------
$\pi^- p\to P_{11}\to n\pi^0\pi^0$ $95\pm 3$ $88\pm 3$ $60\pm 5$
$\pi^- p\to P_{11}\to \Delta^0\pi^0$ $22\pm 3$ $29\pm 3$ $25\pm 3$
$\pi^- p\to P_{11}\to p\sigma$ $70\pm 5$ $53\pm 2$ $32\pm 6$
: \[1440\_pion\] Fractional contributions (in %) of the $N(1440)P_{11}$ and its isobars to $\pi^- p\to p\pi^0\pi^0$ for 3 different $\pi^-$ energies.
The properties of the Roper derived here are mostly consistent with previous determinations. Pole position and Breit-Wigner mass and width fall into the range of values given by the Particle Data Group (PDG [@Yao:2006px]),
---------------------- ------ -- ----------------------------- -----
$M_{BW} = 1430-1470$ MeV; $\Gamma_{BW} MeV
=250-450$
$M_{\rm pole}= MeV; $\Gamma_{\rm pole}=160-260$ MeV
1345-1385$
---------------------- ------ -- ----------------------------- -----
but are defined more precisely here. The helicity coupling agrees with the PDG mean value but from the variety of different solutions we estimate a larger error. Note that our helicity amplitude is defined in the complex $s=M^2$ plane at the pole position of the Roper. The elastic width ($\Gamma_{N\pi}/\Gamma_{tot}= 0.612\pm 0.020$) is compatible with previous findings (60-70%). Its decay fraction into $\Delta\pi$ ($\Gamma_{\Delta\pi}/\Gamma_{tot}= 0.176\pm 0.020$) is not in conflict with the PDG mean value (20-30%); only the $N\sigma$ partial decay width deviates significantly from PDG. We find $\Gamma_{N\sigma}/\Gamma_{tot}= 0.212\pm 0.030$ while PDG gives 5-10%.
Due to its larger phase space, decays into $N\pi$ are more frequent than those into $N\sigma$ even though the latter decay mode provides the largest coupling. For a radial excitation, this is not unexpected: about 50% of all $\psi(2S)$ resonances decay into J/$\psi\,\sigma$, more than 25% of $\Upsilon(2S)$ resonances decay via $\Upsilon(1S)\,\sigma$ [@Yao:2006px]. The large value of $g_{\sigma N}$ might therefore support the interpretation of the Roper resonance as radial excitation.
An alternative interpretation of the $N\pi\pi$ decay is offered by Hernandez, Oset and Vicente-Vacas [@Hernandez:2002xk] who take into account the re-scattering of the two final state pions. The authors of [@Hernandez:2002xk] do not fit data; instead they show that they can reproduce qualitatively the phenomenology of $N(1440)P_{11}\to N\pi\pi$ decays by rescattering thus avoiding the need to introduce a genuine $N(1440)P_{11}\to N\sigma$ amplitude.
In this letter, we have presented new data on photoproduction of two neutral pions in the energy range from the 2$\pi^0$ production threshold up to a photon energy of 820MeV (Mainz) and up to 1300MeV (Bonn) and reported results from a partial wave analysis of this and of related reactions.
The focus of this letter is the Roper resonance. We show that the data are incompatible with the conjecture that conflicting results on its properties could originate from the presence of two similar resonances and their interference, where both are in the $P_{11}$ wave and both fall into the 1300 to 1500MeV mass region. Due to the fact that the Roper properties are over-constrained by the data, we can rule out this possibility. The decay pattern is consistent with an interpretation of the Roper resonance as first radial excitation of the nucleon.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank the technical staff of the ELSA and MAMI machine groups and of all the participating institutions of their invaluable contributions to the success of the experiment. We acknowledge financial support from the Deutsche Forschungsgemeinschaft (DFG) within the SFB/TR16 and from the Schweizerische Nationalfond. The collaboration with St. Petersburg received funds from DFG and RFBR. U.Thoma thanks for an Emmy Noether grant from the DFG. A.V. Sarantsev acknowledges support from RSSF.
[20]{}
W.M. Yao [*et al.*]{}, J. Phys. G [**33**]{} (2006) 1. U.G. Meißner, J.W. Durso, Nucl. Phys. A [**430**]{} (1984) 670. C. Hajduk, B. Schwesinger, Phys. Lett. B [**140**]{} (1984) 172. S. Capstick, N. Isgur, Phys. Rev. D [**34**]{} (1986) 2809. U. Löring [*et al.*]{}, Eur. Phys. J. A [**10**]{} (2001) 395. U. Löring [*et al.*]{}, Eur. Phys. J. A [**10**]{} (2001) 447. L.Y. Glozman [*et al.*]{}, Phys. Rev. D [**58**]{} (1998) 094030. T. Burch [*et al.*]{}, Phys. Rev. D [**74**]{} (2006) 014504. O. Krehl [*et al.*]{}, Phys. Rev. C [**62**]{} (2000) 025207. S. Schneider [*et al.*]{}, Eur. Phys. J. A [**28**]{} (2006) 107. Z.P. Li [*et al.*]{}, Phys. Rev. D [**46**]{} (1992) 70. S. Capstick, P.R. Page, Phys. Rev. D [**60**]{} (1999) 111501. T. Nakano [*et al.*]{}, Phys. Rev. Lett. [**91**]{} (2003) 012002. V.V. Barmin [*et al.*]{}, Phys. Atom. Nucl. [**66**]{} (2003) 1715 \[Yad. Fiz. [**66**]{} (2003) 1763\]. S. Stepanyan [*et al.*]{}, Phys. Rev. Lett. [**91**]{} (2003) 252001. J. Barth [*et al.*]{}, Phys. Lett. B [**572**]{} (2003) 127. R.L. Jaffe and F. Wilczek, Phys. Rev. Lett. [**91**]{} (2003) 232003. D. Diakonov [*et al.*]{}, Z. Phys. A [**359**]{} (1997) 305. R. A. Arndt [*et al.*]{}, Phys. Rev. C [**69**]{} (2004) 035208. M. Battaglieri [*et al.*]{}, Phys. Rev. Lett. [**96**]{} (2006) 042001. S. Niccolai [*et al.*]{}, Phys. Rev. Lett. [**97**]{} (2006) 032001. H.P. Morsch, P. Zupranski, Phys. Rev. C [**61**]{} (2000) 024002. H. P. Morsch and P. Zupranski, Phys. Rev. C [**71**]{} (2005) 065203. J. Patzold [*et al.*]{}, Phys. Rev. C [**67**]{} (2003) 052202. J. Ahrens [*et al.*]{}, Nucl. Phys. News [**4**]{} (1994) 5. W. Hillert, Eur. Phys. J. A [**28S1**]{} (2006) 139. U. Thoma [*et al.*]{}, “N\* and $\rm\Delta$\* decays into $\rm N\pi^0\pi^0$”, arXiv:0707.3592. M. Kotulla [*et al.*]{}, Phys. Lett. B [**578**]{} (2004) 63. F. Harter [*et al.*]{}, Phys. Lett. B [**401**]{} (1997) 229. M. Wolf [*et al.*]{}, Eur. Phys. J. A [**9**]{} (2000) 5. S.J. Hall [*et al.*]{}, Nucl. Instr. Meth. A [**368**]{} (1996) 698. R. Novotny, IEEE Trans. Nucl. Sci. [**38**]{} (1991) 379. A.R. Gabler [*et al.*]{}, Nucl. Instr. Meth. A [**346**]{} (1994) 168. M. Kotulla [*et al.*]{}, Phys. Rev. Lett. [**89**]{} (2002) 272001. Y. Assafiri [*et al.*]{}, Phys. Rev. Lett. [**90**]{} (2003) 222001. A.V. Anisovich [*et al.*]{}, Eur. Phys. J. A [**24**]{} (2005) 111. A.V. Anisovich and A.V. Sarantsev, Eur. Phys. J. A [**30**]{} (2006) 427. A.V. Anisovich [*et al.*]{}, Eur. Phys. J. A [**25**]{} (2005) 427. A.V. Sarantsev [*et al.*]{}, Eur. Phys. J. A [**25**]{} (2005) 441. A.V. Anisovich [*et al.*]{}, “Baryon resonances and polarization transfer in hyperon photoproduction", arXiv:0707.3596. I. Horn [*et al.*]{}, “Evidence for two negative-parity $\Delta D_{33}$ excitations decaying into $\Delta\eta$”, in preparation. O. Bartholomy [*et al.*]{}, Phys. Rev. Lett. [**94**]{} (2005) 012003. H. van Pee [*et al.*]{}, Eur. Phys. J. A [**31**]{} (2007) 61. O. Bartalini [*et al.*]{}, Eur. Phys. J. A [**26**]{} (2005) 399. M. Dugger [*et al.*]{}, Phys. Rev. C [**76**]{} (2007) 025211. R. A. Arndt [*et al.*]{}, Phys. Rev. C [**74**]{} (2006) 045205. S. Prakhov [*et al.*]{}, Phys. Rev. C [**69**]{} (2004) 045202. R.A. Arndt *et al.*, http://gwdac.phys.gwu.edu. J.A. Gomez Tejedor and E. Oset, Nucl. Phys. A [**600**]{} (1996) 413. J.C. Nacher [*et al.*]{}, Nucl. Phys. A [**695**]{} (2001) 295. E. Hernandez, E. Oset and M. J. Vicente Vacas, Phys. Rev. C [**66**]{}, (2002) 065201.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For an arbitrary $q$-polynomial $f$ over ${{\mathbb F}}_{q^n}$ we study the problem of finding those $q$-polynomials $g$ over ${{\mathbb F}}_{q^n}$ for which the image sets of $f(x)/x$ and $g(x)/x$ coincide. For $n\leq 5$ we provide sufficient and necessary conditions and then apply our result to study maximum scattered linear sets of $\operatorname{{PG}}(1,q^5)$.'
author:
- 'Bence Csajbók, Giuseppe Marino, Olga Polverino[^1]'
title: A Carlitz type result for linearized polynomials
---
Introduction
============
Let ${{\mathbb F}}_{q^n}$ denote the finite field of $q^n$ elements where $q=p^h$ for some prime $p$. For $n>1$ and $s \mid n$ the trace and norm over ${{\mathbb F}}_{q^s}$ of elements of ${{\mathbb F}}_{q^n}$ are defined as $\operatorname{Tr}_{q^n/q^s}(x)=x+x^{q^s}+\ldots+x^{q^{n-s}}$ and $\operatorname{N}_{q^n/q^s}(x)=x^{1+q^s+\ldots+q^{n-s}}$, respectively. When $s=1$ then we will simply write $\operatorname{Tr}(x)$ and $\operatorname{N}(x)$. Every function $f \colon {{\mathbb F}}_{q^n} \rightarrow {{\mathbb F}}_{q^n}$ can be given uniquely as a polynomial with coefficients in ${{\mathbb F}}_{q^n}$ and of degree at most $q^n-1$. The function $f$ is ${{\mathbb F}}_q$-linear if and only if it is represented by a *$q$-polynomial*, that is, $$\label{fff}
f(x)=\sum_{i=0}^{n-1}a_i x^{q^i}$$ with coefficients in ${{\mathbb F}}_{q^n}$. Such polynomials are also called *linearized*. If $f$ is given as in , then its adjoint (w.r.t. the symmetric non-degenerate bilinear form defined by ${\langle}x, y{\rangle}=\operatorname{Tr}(xy)$) is $$\hat f(x):=\sum_{i=0}^{n-1}a_i^{q^{n-i}} x^{q^{n-i}},$$ i.e. $\operatorname{Tr}(xf(y))=\operatorname{Tr}(y\hat f(x))$ for any $x,y\in{{\mathbb F}}_{q^n}$.
The aim of this paper is to study what can be said about two $q$-polynomials $f$ and $g$ over ${{\mathbb F}}_{q^n}$ if they satisfy $$\label{start}
Im\,\left(\frac{f(x)}{x}\right)=Im\,\left(\frac{g(x)}{x}\right)\ \footnote{By $Im\,(f(x)/x)$ we mean the image of the rational function $f(x)/x$, i.e. $\{f(x)/x \colon x\in {{\mathbb F}}_{q^n}^*\}$.}.$$
For a given $q$-polynomial $f$, the equality clearly holds with $g(x)=f(\lambda x)/\lambda$ for each $\lambda\in {{\mathbb F}}_{q^n}^*$. A less obvious choice when holds is when $g(x)=\hat f(\lambda x)/\lambda$, see [@BGMP2015 Lemma 2.6] and the first part of [@CSMP2016 Section 3].
When one of the functions in is a monomial then the answer to the question posed above follows from McConnel’s generalization [@McConnel Theorem 1] of a result due to Carlitz [@Carlitz] (see also Bruen and Levinger [@BL]).
[[@McConnel Theorem 1]]{} \[Carlitz\] Let $p$ denote a prime, $q=p^h$, and $1<d$ a divisor of $q-1$. Also, let $F \colon {{\mathbb F}}_{q} \rightarrow {{\mathbb F}}_{q}$ be a function such that $F(0)=0$ and $F(1)=1$. Then $$\left(F(x)-F(y)\right)^{\frac{q-1}{d}}=\left(x-y\right)^{\frac{q-1}{d}}$$ for all $x,y\in {{\mathbb F}}_{q}$ if and only if $F(x)=x^{p^j}$ for some $0 \leq j < h$ and $d\mid p^j-1$.
Indeed, when the function $F$ of Theorem \[Carlitz\] is ${{\mathbb F}}_q$-linear, we easily get the following corollary (see Section \[BPR\] for the proof, or [@Praha Corollary 1.4] for the case when $q$ is an odd prime).
\[norm\] Let $g(x)$ and $f(x)=\alpha x^{q^k}$, $q=p^h$, be $q$-polynomials over ${{\mathbb F}}_{q^n}$ such that $$\label{form:im}
Im\,\left(\frac{f(x)}{x}\right)=Im\,\left(\frac{g(x)}{x}\right).$$ Denote $\gcd(k,n)$ by $t$. Then $g(x)=\beta x^{q^s}$ with $\gcd(s,n)=t$ for some $\beta$ with $\operatorname{N}_{q^n/q^t}(\alpha)=\operatorname{N}_{q^n/q^t}(\beta)$.
Another case for which we know a complete answer to our problem is when $f(x)=\operatorname{Tr}(x)$.
\[trace\] Let $f(x)=\operatorname{Tr}(x)$ and let $g(x)$ be a $q$-polynomial over ${{\mathbb F}}_{q^n}$ such that $$Im\,(f(x)/x)=Im \,(g(x)/x).$$ Then $g(x)=\operatorname{Tr}(\lambda x)/\lambda$ for some $\lambda \in {{\mathbb F}}_{q^n}^*$.
Note that in Theorem \[trace\] we have $\hat f(x) = f(x)$ and the only solutions for $g$ are $g(x)=f(\lambda x)/\lambda$, while in Corollary \[norm\] we have (up to scalars) $\varphi(n)$ different solutions for $g$, where $\varphi$ is the Euler’s totient function.
The problem posed in is also related to the study of the directions determined by an additive function. Indeed, when $f$ is additive, then $$Im\,(f(x)/x)=\left\{\frac{f(x)-f(y)}{x-y} \colon x\neq y,\, x,y\in {{\mathbb F}}_{q^n}\right\},$$ is the *set of directions* determined by the graph of $f$, i.e. by the point set ${\mathcal G}_f:=\{(x,f(x)) \colon x\in {{\mathbb F}}_{q^n}\}\subset \operatorname{{AG}}(2,q^n)$. Hence, in this setting, the problem posed in corresponds to finding the ${{\mathbb F}}_q$-linear functions whose graph determines the same set of directions. The maximum cardinality of the image set of $f(x)/x$, where $f$ is a $q$-polynomial over ${{\mathbb F}}_{q^n}$, is $(q^n-1)/(q-1)$, whereas, if ${{\mathbb F}}_q$ is the maximum field of linearity of $f$, then by [@BBBSSz; @B2003; @Cs2017] $Im\,(f(x)/x)$ contains at least $q^{n-1}+1$ elements, hence $$\label{directions}
q^{n-1}+1\leq |Im\,(f(x)/x)| \leq \frac{q^{n}-1}{q-1}.$$
The classical examples which show the sharpness of these bounds are the monomial functions $x^{q^s}$, with $\gcd (s,n)=1$, and the $\operatorname{Tr}(x)$ function. However, these bounds are also achieved by other polynomials which are not “equivalent” to these examples (see Section 2 for more details).
Two ${{\mathbb F}}_q$-linear polynomials $f(x)$ and $h(x)$ of ${{\mathbb F}}_{q^n}[x]$ are *equivalent* if the two graphs ${\mathcal G}_f$ and ${\mathcal G}_h$ are equivalent under tha action of the group $\Gamma {\mathrm L}(2,q^n)$, i.e. if there exists an element $\varphi\in\Gamma {\mathrm L}(2,q^n)$ such that ${\mathcal G}_f^\varphi={\mathcal G}_h$. In such a case, we say that $f$ and $h$ are *equivalent* (via $\varphi$) and we write $h=f_\varphi$. It is easy to see that in this way we defined an equivalence relation on the set of $q$-polynomials over ${{\mathbb F}}_{q^n}$. If $f$ and $g$ are two $q$-polynomials such that $Im\,(f(x)/x)=Im\,(g(x)/x)$, then $Im\,(f_\varphi(x)/x)=Im\,(g_\varphi(x)/x)$ for any admissible $\varphi\in{\mathrm{\Gamma L}}(2,q^n)$ (see Proposition ). This means that the problem posed in can be investigated up to equivalence.
For $n\leq 4$, the only solutions for $g$ in Problem are the trivial ones, i.e. either $g(x)=f(\lambda x)/x$ or $g(x)=\hat f(\lambda x)/x$ (cf. Theorem \[n=4\]).
For the case $n=5$, in Section \[section-proof\], we prove the following main result.
\[main\] Let $f(x)$ and $g(x)$ be two $q$-polynomials over ${{\mathbb F}}_{q^5}$, with maximum field of linearity ${{\mathbb F}}_q$, such that $Im\,(f(x)/x)=Im\,(g(x)/x)$. Then either there exists $\varphi\in {\mathrm{\Gamma L}}(2,q^5)$ such that $f_\varphi(x)=\alpha x^{q^i}$ and $g_\varphi(x)=\beta x^{q^j}$ with $\operatorname{N}(\alpha)=\operatorname{N}(\beta)$ for some $i,j\in \{1,2,3,4\}$, or there exists $\lambda\in {{\mathbb F}}_{q^5}^*$ such that $g(x)=f(\lambda x)/\lambda$ or $g(x)=\hat{f}(\lambda x)/\lambda$.
Finally, the relation between $Im\,(f(x)/x)$ and the linear sets of rank $n$ of the projective line $\operatorname{{PG}}(1,q^n)$ will be pointed out in Section \[geom\]. As an application of Theorem \[main\] we get a criteria of $\mathrm{P}\Gamma\mathrm{L}(2,q^5)$-equivalence for linear sets in $\operatorname{{PG}}(1,q^5)$ and this allows us to prove that the family of (maximum scattered) linear sets of rank $n$ of size $(q^n-1)/(q-1)$ in $\operatorname{{PG}}(1,q^n)$ found by Sheekey in [@Sh] contains members which are not-equivalent to the previously known linear sets of this size.
Background and preliminary results {#BPR}
==================================
Let us start this section by the following immediate corollary of .
\[fieldoflinearity\] If $Im\,(f(x)/x)=Im\,(g(x)/x)$ for two $q$-polynomials $f$ and $g$ over ${{\mathbb F}}_{q^n}$, then their maximum field of linearity coincide.
Let ${{\mathbb F}}_{q^m}$ and ${{\mathbb F}}_{q^k}$ be the maximum fields of linearity of $f$ and $g$, respectively. Suppose to the contrary $m<k$. Then $|Im\,(g(x)/x)|\leq (q^n-1)/(q^k-1)<q^{n-k+1}+1\leq q^{n-m}+1 \leq |Im\,(f(x)/x)|$, a contradiction.
Now we are able to prove Corollary \[norm\].
The maximum field of linearity of $f(x)$ is ${{\mathbb F}}_{q^t}$, thus, by Proposition \[fieldoflinearity\], $g(x)$ has to be a $q^t$-polynomial as well. Then for $t>1$ the result follows from the $t=1$ case (after substituting $q$ for $q^t$ and $n/t$ for $n$) and hence we can assume that $f(x)$ and $g(x)$ are strictly ${{\mathbb F}}_q$-linear. By , we note that $g(1)=\alpha z_0^{q^k-1}$, for some $z_0\in{{\mathbb F}}_{q^n}^*$. Let $F(x):=g(x)/g(1)$, then $F$ is a $q$-polynomial over ${{\mathbb F}}_{q^n}$, with $F(0)=0$ and $F(1)=1$. Also, from , for each $x\in{{\mathbb F}}_{q^n}^*$ there exists $z\in{{\mathbb F}}_{q^n}^*$ such that $$\frac{F(x)}{x}=\left(\frac{z}{z_0}\right)^{q^k-1}.$$ This means that for each $x \in{{\mathbb F}}_{q^n}^*$ we get $\operatorname{N}\left(\frac{F(x)}{x}\right)=1$. From Theorem \[Carlitz\] (with $d=q-1$) it follows that $F(x)=x^{p^j}$ for some $0\leq j<nh$. Then Proposition \[fieldoflinearity\] yields $p^j=q^s$ with $\gcd(s,n)=1$. We get the first part of the statement by putting $\beta:=f(1)$. Again from it follows that $\operatorname{N}(\alpha)=\operatorname{N}(\beta)$.
Let $f$ and $g$ be two equivalent $q$-polynomials over ${{\mathbb F}}_{q^n}$ via the element $\varphi\in\Gamma{\mathrm L}(2,q^n)$ represented by the invertible matrix $\begin{pmatrix}
a&b\\
c&d\\
\end{pmatrix}$ and with companion automorphism $\sigma$ of ${{\mathbb F}}_{q^n}$. Then $$\label{form:equiv}
\left\{
\begin{pmatrix}
x\\
g(x)
\end{pmatrix}
\colon x\in {{\mathbb F}}_{q^n}
\right\}=
\left\{
\begin{pmatrix}
a&b\\
c&d\\
\end{pmatrix}
\begin{pmatrix}
x^\sigma\\
f(x)^\sigma
\end{pmatrix}
\colon x \in {{\mathbb F}}_{q^n}\right\}.$$ By we get that the function $k_f(x):=ax^\sigma+bf(x)^\sigma$ is invertible and, with $h_f(x):=cx^\sigma+df(x)^\sigma$, it follows that $g(x)=h_f(k_f^{-1}(x))$. Hence, we have proved the following proposition.
\[fvarphi\] Let $f$ and $g$ be $q$-polynomials over ${{\mathbb F}}_{q^n}$ such that $g=f_\varphi$ for some $\varphi\in{\mathrm{\Gamma L}}(2,q^n)$ with linear part represented by $\left(
\begin{array}{cc}
a & b \\
c & d \\
\end{array}
\right)$ and with companion automorphism $\sigma$. Let $k_f(x)=ax^\sigma+bf(x)^\sigma$ and $h_f(x)=cx^\sigma+df(x)^\sigma$, then $k_f$ is invertible and $g(x)=f_\varphi(x)=h_f(k_f^{-1}(x))$.
From it is also clear that $$\label{map}
Im\,\left(\frac{f_\varphi(x)}x\right)=\left\{\frac{c+dz^\sigma}{a+bz^\sigma}
\colon z \in Im\,\left(\frac{f(x)}x\right)\right\}$$ and hence $$\label{size}
|Im\,(f_\varphi(x)/x)|=|Im\,(f(x)/x)|.$$
From Equations and the next result easily follows.
\[prop:lin\] If two $q$-polynomials over ${{\mathbb F}}_{q^n}$ are equivalent, then their maximum field of linearity coincide.
Note that $|Im\,(g(x)/x)|=|Im\,(f(x)/x)|$ does not imply the equivalence of $f$ and $g$. In fact, in the last section we will list the known examples of $q$-polynomials $f$ which are not equivalent to monomials but the size of $Im\,(f(x)/x)$ is maximal. To find such functions was also proposed in [@Praha] and, as it was observed by Sheekey, they determine certain MRD-codes [@Sh].
The following results will be useful later in the paper.
\[imtrans\] If $Im\, (f(x)/x)=Im\, (g(x)/x)$ for some $q$-polynomials over ${{\mathbb F}}_{q^n}$, then $Im\, (f_\varphi(x)/x)=Im\, (g_\varphi(x)/x)$ holds for each admissible [^2] $\varphi\in {\mathrm{\Gamma L}}(2,q^n)$.
From $Im\,(f(x)/x)=Im\, (g(x)/x)$ it follows that any $\varphi\in {\mathrm{\Gamma L}}(2,q^n)$ admissible w.r.t. $f$ is admissible w.r.t. $g$ as well. Hence $k_f$ and $k_g$ are both invertible and we may construct $f_\varphi$ and $g_\varphi$ as indicated in Proposition \[fvarphi\]. The statement now follows from Equation .
\[lambda0\] Let $f$ and $g$ be $q$-polynomials over ${{\mathbb F}}_{q^n}$ and take some $\varphi\in {\mathrm{\Gamma L}}(2,q^n)$ with companion automorphism $\sigma$. Then $g_\varphi(x)=f_\varphi(\lambda^\sigma x)/\lambda^\sigma$ for some $\lambda\in {{\mathbb F}}_{q^n}^*$ if and only if $g(x)=f(\lambda x)/\lambda$.
First we prove the “if” part. Since $g(x)=f(\lambda x)/\lambda=(\omega_{1/\lambda}\circ f \circ \omega_\lambda)(x)$, where $\omega_\alpha$ denotes the scalar map $x\in{{\mathbb F}}_{q^n}\mapsto\alpha x\in{{\mathbb F}}_{q^n}$, direct computations show that $h_g=\omega_{1/\lambda^\sigma}\circ h_f\circ\omega_\lambda$ and $k_g=\omega_{1/\lambda^\sigma}\circ k_f\circ\omega_\lambda$. Then $g_\varphi=\omega_{1/\lambda^\sigma}\circ f_\varphi\circ\omega_{\lambda^\sigma}$ and the first part of the statement follows. The “only if” part follows from the “if” part applied to $g_\varphi(x)=f_\varphi(\lambda^\sigma x)/\lambda^\sigma$ and $\varphi^{-1}$; and from $(f_\varphi)_{\varphi^{-1}}=f$ and $(g_\varphi)_{\varphi^{-1}}=g$.
Next we summarize what is known about Problem for $n\leq 4$.
\[n=4\] Suppose $Im\, (f(x)/x)=Im\, (g(x)/x)$ for some $q$-polynomials over ${{\mathbb F}}_{q^n}$, $n\leq 4$, with maximum field of linearity ${{\mathbb F}}_q$. Then there exist $\varphi\in \mathrm{GL}(2,q^n)$ and $\lambda \in {{\mathbb F}}_{q^n}^*$ such that the following holds.
- If $n=2$ then $f_\varphi(x) = x^q$ and $g(x) = f(\lambda x)/\lambda$.
- If $n=3$ then either $$f_\varphi(x)=\operatorname{Tr}(x) \mbox{ and } g(x)=f(\lambda x)/\lambda$$ or $$f_\varphi(x) = x^q \mbox{ and } g(x)=f(\lambda x)/\lambda \mbox{ or } g(x)=\hat{f}(\lambda x)/\lambda.$$
- If $n=4$ then $g(x) = f(\lambda x)/\lambda$ or $g(x) = \hat{f}(\lambda x)/\lambda$.
In the $n=2$ case $f(x)=ax+bx^q$, $b\neq 0$. Then $\varphi:=\begin{pmatrix}
1 & 0 \\
-a/b & 1/b \\
\end{pmatrix}$ maps $f(x)$ to $x^q$. Then Proposition \[imtrans\] and Corollary \[norm\] give $g_\varphi(x)=f_\varphi(\mu x)/\mu$ and hence Proposition \[lambda0\] gives $g(x)=f(\lambda x)/\lambda$ for some $\lambda\in{{\mathbb F}}_{q^n}$. If $n=3$ then according to [@LaVa2010 Theorem 5] and [@CSMP2016 Theorem 1.3] there exists $\varphi\in \mathrm{GL}(2,q^3)$ such that either $f_\varphi(x)=\operatorname{Tr}(x)$ or $f_\varphi(x)=x^q$. In the former case Proposition \[imtrans\] and Theorem \[trace\] give $g_\varphi(x)=f_\varphi(\mu x)/\mu$ and the assertion follows from Proposition \[lambda0\]. In the latter case Proposition \[imtrans\] and Corollary \[norm\] give $g_\varphi(x)=\alpha x^{q^i}$ where $i\in \{1,2\}$ and $\operatorname{N}(\alpha)=1$. If $i=1$, then $g_\varphi(x)=f_\varphi(\mu x)/\mu$ where $\mu^{q-1}=\alpha$ and the assertion follows from Proposition \[lambda0\]. Let now $i=2$ and denote by $\begin{pmatrix}
A&B\\
C&D\\
\end{pmatrix}$ the matrix of $\varphi^{-1}$. Also, let $\Delta$ denote the determinant of this matrix. Since $\operatorname{N}(\alpha)=1$, direct computations show $$f(x)=(f_{\varphi})_{\varphi^{-1}}(x)=\frac{(A^{q+q^2}C+B^{q+q^2}D)x+A^{q^2}\Delta x^q-B^q\Delta x^{q^2}}{\operatorname{N}(A)+\operatorname{N}(B)},$$ $$g(x)=(g_{\varphi})_{\varphi^{-1}}(x)=\frac{(A^{q+q^2}C+B^{q+q^2}D)x-B^{q^2}\Delta \alpha^{q^2+1}x^q+A^q\Delta\alpha x^{q^2}}{\operatorname{N}(A)+\operatorname{N}(B)},$$ and hence $g(x)=\hat{f}(\lambda x)/\lambda$ for each $\lambda\in {{\mathbb F}}_{q^3}^*$ with $\lambda^{q-1}=\Delta^{1-q}/\alpha^q$.
The case $n=4$ is [@CSMP2016 Proposition 4.2 ].
[Theorem \[n=4\] yields that there is a unique equivalence class of $q$-polynomials, with maximum field of linearity ${{\mathbb F}}_q$, when $n=2$. For $n=3$ there are two non-equivalent classes and they correspond to the classical examples: $\operatorname{Tr}(x)$ and $x^q$. Whereas, for $n=4$, from [@CSMP2016 Sec. 5.3] and [@BoPo2005 Table p. 54], there exist at least eight non-equivalent classes. The possible sizes for the sets of directions determined by these strictly ${{\mathbb F}}_q$-linear functions are $q^3+1$, $q^3+q^2-q+1$, $q^3+q^2+1$ and $q^3+q^2+q+1$ and each of them is determined by at least two non-equivalent $q$-polynomials. Also, by [@CSZ2017 Theorem 3.4], if $f$ is a $q$-polynomial over ${{\mathbb F}}_{q^4}$ for which the set of directions is of maximum size then $f$ is equivalent either to $x^q$ or to $\delta x^q+x^{q^3}$, for some $\delta\in{{\mathbb F}}_{q^4}^*$ with $N(\delta)\ne 1$ (see [@LP2001]).]{}
Preliminary results about Tr(x) and the monomial q-polynomials over Fq5
=======================================================================
Let $q$ be a power of a prime $p$. We will need the following results.
\[trace5\] Let $f(x)=\sum_{i=0}^4a_ix^{q^i}$ and $g(x)=\operatorname{Tr}(x)$ be $q$-polynomials over ${{\mathbb F}}_{q^5}$. Then there is an element $\varphi\in\Gamma\mathrm{L}(2,q^5)$ such that $Im\,(f_\varphi(x)/x)=Im\,(g(x)/x)$ if and only if $a_1a_2a_3a_4\neq0$, $(a_1/a_2)^q=a_2/a_3$, $(a_2/a_3)^q=a_3/a_4$ and $\operatorname{N}(a_1)=\operatorname{N}(a_2)$.
Let $\varphi\in\Gamma\mathrm{L}(2,q^5)$ such that $Im\,(f_\varphi(x)/x)=Im\,(g(x)/x)$. By Proposition \[prop:lin\], the maximum field of linearity of $f$ is ${{\mathbb F}}_q$ and by Theorem \[trace\] there exists $\lambda\in{{\mathbb F}}_{q^5}^*$ such that $f_\varphi(x)=\operatorname{Tr}(\lambda x)/\lambda$.This is equivalent to say that there exist $a,b,c,d$, $ad-bc\neq 0$ and $\sigma=p^h$ such that $$\left\{
\begin{pmatrix}
y\\
\operatorname{Tr}(y)
\end{pmatrix}
\colon y\in {{\mathbb F}}_{q^5}
\right\}=
\left\{
\begin{pmatrix}
a&b\\
c&d\\
\end{pmatrix}
\begin{pmatrix}
x^\sigma\\
f(x)^\sigma
\end{pmatrix}
\colon x \in {{\mathbb F}}_{q^5}\right\}.$$ Then $c x^\sigma+d f(x)^\sigma \in {{\mathbb F}}_q$ for each $x\in {{\mathbb F}}_{q^5}$. Let $z=x^{\sigma}$. Then $$cz+d\sum_{i=0}^4a_i^{\sigma}z^{q^i}=c^q z^{q}+d^q\sum_{i=0}^4 a_i^{\sigma q}z^{q^{i+1}},$$ for each $z$. As polynomials of $z$ the left and right-hand sides of the above equation coincide modulo $z^{q^5}-z$ and hence comparing coefficients yield $$c+da_0^\sigma = d^q a_4^{\sigma q},$$ $$d a_1^\sigma = c^q + d^q a_0^{\sigma q},$$ $$d a_{k+1}^\sigma=d^q a_k^{\sigma q},$$ for $k=1,2,3$. If $d=0$, then $c=0$, a contradiction. Since $d\neq 0$, if one of $a_1,a_2,a_3,a_4$ is zero, then all of them are zero and hence $f$ is ${{\mathbb F}}_{q^5}$-linear. This is not the case, so we have $a_1a_2a_3a_4\neq 0$. Then the last three equations yield $$\left(\frac{a_1}{a_2}\right)^q=\frac{a_2}{a_3},$$ $$\left(\frac{a_2}{a_3} \right)^q=\frac{a_3}{a_4},$$ and also $\operatorname{N}(a_1)=\operatorname{N}(a_2)$.
Now assume that the conditions of the assertion hold. It follows that $a_3=a_2^{q+1}/a_1^q$ and $a_4=a_3^{q+1}/a_2^q=a_2^{q^2+q+1}/a_1^{q^2+q}$. Let $\alpha_i=a_i/a_1$ for $i=0,1,2,3,4$. Then $\alpha_1=1$, $\operatorname{N}(\alpha_2)=1$, $\alpha_3=\alpha_2^{q+1}$ and $\alpha_4=\alpha_2^{1+q+q^2}$. We have $\alpha_2=\lambda^{q-1}$ for some $\lambda\in {{\mathbb F}}_{q^5}^*$. If $$\begin{pmatrix}
a&b\\
c&d
\end{pmatrix}
=
\begin{pmatrix}
1&0\\
1-\lambda^{1-q^4}a_0/a_1& \lambda^{1-q^4}/a_1
\end{pmatrix},$$ then $$\begin{pmatrix}
a&b\\
c&d
\end{pmatrix}
\begin{pmatrix}
x\\
f(x)
\end{pmatrix}
=$$ $$\begin{pmatrix}
x\\
x+\lambda^{1-q^4}x^q+\lambda^{q-q^4}x^{q^2}+\lambda^{q^2-q^4}x^{q^3}+\lambda^{q^3-q^4}x^{q^4}
\end{pmatrix}
=
\begin{pmatrix}
x\\
\operatorname{Tr}(x \lambda^{q^4})/\lambda^{q^4}
\end{pmatrix},$$ i.e. $f_\varphi(x)=\operatorname{Tr}(\lambda^{q^4}x)/\lambda^{q^4}$, where $\varphi$ is defined by the matrix $\begin{pmatrix}
a&b\\
c&d
\end{pmatrix}$.
\[pseudoalg\] Let $f(x)=\sum_{i=0}^4a_ix^{q^i}$, with $a_1a_2a_3a_4\neq 0$. Then there is an element $\varphi\in\Gamma\mathrm{L}(2,q^5)$ such that $Im\,(f_\varphi(x)/x)=Im\,(x^{q-1})$ if and only if one of the following holds:
1. $(a_1/a_2)^q=a_2/a_3$, $(a_2/a_3)^q=a_3/a_4$ and $\operatorname{N}(a_1)\neq \operatorname{N}(a_2)$, or
2. $(a_4/a_1)^{q^2}=a_1/a_3$, $(a_1/a_2)^{q^2}=a_3/a_4$ and $\operatorname{N}(a_1)\neq \operatorname{N}(a_3)$.
In both cases, if the condition on the norms does not hold, then $Im\,(f_\varphi(x)/x)=Im\,(\operatorname{Tr}(x)/x)$.
We first note that the monomials $x^{q^i}$ and $x^{q^{5-i}}$ are equivalent via the map $\psi:=\begin{pmatrix}
0&1\\
1&0
\end{pmatrix}$. Hence, by Corollary \[norm\], the statement holds if and only if there exist $a,b,c,d$, $ad-bc\neq 0$, $\sigma=p^h$ and $i\in \{1,2\}$ such that $$\label{pseudoxx}
\left\{
\begin{pmatrix}
y\\
y^{q^i}
\end{pmatrix}
\colon y\in {{\mathbb F}}_{q^5}
\right\}
=
\left\{
\begin{pmatrix}
a&b\\
c&d\\
\end{pmatrix}
\begin{pmatrix}
x^\sigma\\
f(x)^\sigma
\end{pmatrix}
\colon x \in {{\mathbb F}}_{q^5}\right\}
.$$
If Condition 1 holds then let $\alpha_j=a_j/a_1$ for $j=0,1,2,3,4$. So $\alpha_1=1$, $\operatorname{N}(\alpha_2)\neq 1$, $\alpha_3=\alpha_2^{q+1}$, $\alpha_4=\alpha_2^{1+q+q^2}$ and is satisfied with $$\begin{pmatrix}
a&b\\
c&d
\end{pmatrix}
=
\begin{pmatrix}
1&\alpha_2^{q^4}\\
\alpha_2^{1+q+q^2+q^3}&1
\end{pmatrix}
\begin{pmatrix}
1&0\\
-\alpha_0 & 1/a_1
\end{pmatrix}.$$ If Condition 2 holds then let $\alpha_j=a_j/a_3$ for $j=0,1,2,3,4$. So $\alpha_3=1$, $\operatorname{N}(\alpha_1)\neq 1$, $\alpha_2=\alpha_1^{1+q+q^3}$, $\alpha_4=\alpha_1^{1+q^3}$ and is satisfied with $$\begin{pmatrix}
a&b\\
c&d
\end{pmatrix}
=
\begin{pmatrix}
\alpha_1^{1+q+q^3+q^4}&1\\
1&\alpha_1^{q^2}
\end{pmatrix}
\begin{pmatrix}
1&0\\
-\alpha_0 & 1/a_3
\end{pmatrix}.$$ Suppose now that holds and put $z=x^\sigma$. Then $$(za+b\sum_{j=0}^4 a_j^\sigma z^{q^j})^{q^i}=cz+d\sum_{j=0}^4a_j^\sigma z^{q^j}$$ for each $z\in {{\mathbb F}}_{q^5}$ and hence, as polynomials in $z$, the left-hand side and right-hand side of the above equation coincide modulo $z^{q^5}-z$. The coefficients of $z$, $z^{q^i}$ and $z^{q^k}$ with $i\in \{1,2\}$ and $k\in \{1,2,3,4\}\setminus \{i\}$ give $$b^{q^i} a_{-i}^{\sigma q^i} =c+da_0^\sigma,$$ $$a^{q^i} +b^{q^i} a_0^{\sigma q^i}=d a_i^\sigma,$$ $$b^{q^i} a_{k-i}^{\sigma q^i}=d a_{k}^\sigma,$$ respectively, where the indices are considered modulo 5. Note that $db\neq 0$ since otherwise also $a=c=0$ and hence $ad-bc=0$. With $\{r,s,t\}=\{1,2,3,4\} \setminus \{i\}$, the last three equations yield: $$\left(\frac{a_{r-i}}{a_{s-i}}\right)^{q^i}=\frac{a_r}{a_s},$$ $$\left(\frac{a_{s-i}}{a_{t-i}} \right)^{q^i}=\frac{a_s}{a_t}.$$ First assume $i=1$. Then we have $$\left(\frac{a_1}{a_2}\right)^q=\frac{a_2}{a_3} \quad \text{and} \quad \left(\frac{a_2}{a_3} \right)^q=\frac{a_3}{a_4}.$$ If $\operatorname{N}(a_1)=\operatorname{N}(a_2)$, from Proposition \[trace5\] and Equation it follows that $|Im\,(x^{q-1})|=|Im\,(\operatorname{Tr}(x)/x)|$, a contradiction.
Now assume $i=2$. Then we have $(a_4/a_1)^{q^2}=a_1/a_3$ and $$\label{form1}
\left(\frac{a_1}{a_2} \right)^{q^2}=\frac{a_3}{a_4}.$$ Multiplying these two equations yields $a_4^{q^2+1}=a_1a_2^{q^2}$ and hence $$\label{form2}
a_2=a_1^{1+q+q^3}/a_3^{q^3+q}.$$ By this implies $$\label{form3}
a_4=a_1^{q^3+1}/a_3^{q^3}.$$ If $\operatorname{N}(a_1)=\operatorname{N}(a_3)$, then also $\operatorname{N}(a_1)=\operatorname{N}(a_2)=\operatorname{N}(a_3)=\operatorname{N}(a_4)$. We show that in this case $Im\,(f_\varphi(x)/x)=Im\,(\operatorname{Tr}(x)/x)$, so we must have $\operatorname{N}(a_1)\neq \operatorname{N}(a_3)$. According to Proposition \[trace5\] it is enough to show $(a_1/a_2)^q=a_2/a_3$ and $(a_2/a_3)^q=a_3/a_4$. By we have $(a_1/a_2)^q=(a_3/a_4)^{q^4}$, which equals $a_2/a_3$ if and only if $(a_2/a_3)^q=a_3/a_4$, i.e. $a_3^{1+q}=a_4a_2^q$. Taking into account and , this equality follows from $\operatorname{N}(a_1)=\operatorname{N}(a_3)$.
Proof of the main theorem {#section-proof}
=========================
In this section we prove Theorem \[main\]. In order to do this, we use the following two results and the technique developed in [@CSMP2016].
\[lmain\] Let $f$ and $g$ be two linearized polynomials over ${{\mathbb F}}_{q^n}$. If $Im\,(f(x)/x)=Im \,(g(x)/x)$, then for each positive integer $d$ the following holds $$\sum_{x\in {{\mathbb F}}_{q^n}^*} \left(\frac{f(x)}{x}\right)^d = \sum_{x\in {{\mathbb F}}_{q^n}^*} \left(\frac{g(x)}{x}\right)^d.$$
\[lfolk\] For any prime power $q$ and integer $d$ we have $\sum_{x\in {{\mathbb F}}_{q}^*} x^d=-1$ if $q-1 \mid d$ and $\sum_{x\in {{\mathbb F}}_{q}^*} x^d=0$ otherwise.
\[prop\] Let $f(x)=\sum_{i=0}^4 a_i x^{q^i}$ and $g(x)=\sum_{i=0}^4 b_i x^{q^i}$ be two $q$-polynomials over ${{\mathbb F}}_{q^5}$ such that $Im\,(f(x)/x)=Im \,(g(x)/x)$. Then the following relations hold between the coefficients of $f$ and $g$: $$\label{e0}
a_0=b_0,$$ $$\label{e1}
a_1a_4^q=b_1 b_4^q,$$ $$\label{e2}
a_2a_3^{q^2}=b_2 b_3^{q^2},$$ $$\label{e3}
a_1^{q+1}a_3^{q^2}+a_2a_4^{q+q^2}=b_1^{q+1}b_3^{q^2}+b_2b_4^{q+q^2},$$ $$\label{e4}
a_1a_2^{q+q^3}+a_3^{1+q^3}a_4^q=b_1b_2^{q+q^3}+b_3^{1+q^3}b_4^q,$$ $$\label{e5}
a_1^{1+q+q^2}a_2^{q^3}+a_2^{1+q}a_3^{q^2+q^3}+a_1^qa_3^{1+q^2+q^3}+a_1^{q^2}a_2a_3^{q^3}a_4^q+a_2^{1+q+q^3}a_4^{q^2}+$$ $$a_1^qa_2^{q^3}a_3a_4^{q^2}+a_1a_2^qa_3^{q^2}a_4^{q^3}+a_1^{1+q^2}a_4^{q+q^3}+a_3a_4^{q+q^2+q^3}=$$ $$b_1^{1+q+q^2}b_2^{q^3}+b_2^{1+q}b_3^{q^2+q^3}+b_1^qb_3^{1+q^2+q^3}+b_1^{q^2}b_2b_3^{q^3}b_4^q+b_2^{1+q+q^3}b_4^{q^2}+$$ $$b_1^qb_2^{q^3}b_3b_4^{q^2}+b_1b_2^qb_3^{q^2}b_4^{q^3}+b_1^{1+q^2}b_4^{q+q^3}+b_3b_4^{q+q^2+q^3},$$ $$\label{e6}
\operatorname{N}(a_1)+\operatorname{N}(a_2)+\operatorname{N}(a_3)+\operatorname{N}(a_4)+\operatorname{Tr}(a_1^qa_2^{q^2+q^3+q^4}a_3+a_1^{q+q^3}a_2^{q^4}a_3^{1+q^2}+$$ $$a_1^{q+q^2}a_2^{q^3+q^4}a_4+a_1^{q+q^2+q^4}a_3^{q^3}a_4+a_2^qa_3^{q^2+q^3+q^4}a_4+a_1^{q^2}a_3^{q^3+q^4}a_4^{1+q}+$$ $$a_2^{q+q^3}a_3^{q^4}a_4^{1+q^2}+a_1^{q^2}a_2^{q^4}a_4^{1+q+q^3})=$$ $$\operatorname{N}(b_1)+\operatorname{N}(b_2)+\operatorname{N}(b_3)+\operatorname{N}(b_4)+\operatorname{Tr}(b_1^qb_2^{q^2+q^3+q^4}b_3+b_1^{q+q^3}b_2^{q^4}b_3^{1+q^2}+$$ $$b_1^{q+q^2}b_2^{q^3+q^4}b_4+b_1^{q+q^2+q^4}b_3^{q^3}b_4+b_2^qb_3^{q^2+q^3+q^4}b_4+b_1^{q^2}b_3^{q^3+q^4}b_4^{1+q}+$$ $$b_2^{q+q^3}b_3^{q^4}b_4^{1+q^2}+b_1^{q^2}b_2^{q^4}b_4^{1+q+q^3}).$$
Equations – follow from [@CSMP2016 Lemma 3.6]. To prove we will use Lemma \[lfolk\] together with Lemma \[lmain\] with $d=q^3+q^2+q+1$. This gives us $$\sum_{1\leq i,j,m,n \leq 4} a_ia_j^qa_m^{q^2}a_n^{q^3} \sum_{x\in {{\mathbb F}}_{q^5}^*} x^{q^i-1+q^{j+1}-q+q^{m+2}-q^2+q^{n+3}-q^3}=$$ $$\sum_{1\leq i,j,m \leq 4} b_ib_j^qb_m^{q^2}b_n^{q^3} \sum_{x\in {{\mathbb F}}_{q^5}^*} x^{q^i-1+q^{j+1}-q+q^{m+2}-q^2+q^{n+3}-q^3}.$$ We have $\sum_{x\in {{\mathbb F}}_{q^n}^*} x^{q^i-1+q^{j+1}-q+q^{m+2}-q^2+q^{n+3}-q^3}=-1$ if and only if $$\label{eki}
q^i+q^{j+1}+q^{m+2}+q^{n+3} \equiv 1+q+q^2+q^3 \pmod {q^5-1},$$ and zero otherwise. Suppose that the former case holds. The right-hand side of is smaller than the left-hand side, thus $$q^i+q^{j+1}+q^{m+2}+q^{n+3} = 1+q+q^2+q^3 + k(q^5-1),$$ for some positive integer $k$. We have $q^i+q^{j+1}+q^{m+2}+q^{n+3}\leq q^4+q^5+q^6+q^7< 1+q+q^2+q^3+(q^2+q+2)(q^5-1)$ and hence $k\leq q^2+q+1$. If $i=1$, then $q^2 \mid 1-k$ and hence $k=1$, $j=m=1$ and $n=2$, or $k=q^2+1$, $n=4$ and either $j=2$ and $m=3$, or $j=4$ and $m=1$. If $i>1$, then $q^2$ divides $q+1-k$ and hence $k=q+1$, or $k=q^2+q+1$. In the former case $i=j=n=2$ and $m=4$, or $i=j=2$ and $n=m=3$, or $i=3$, $j=1$, $m=4$ and $n=2$, or $i=3$, $j=1$ and $m=n=3$, or $m=1$, $i=2$, $j=4$ and $n=3$. In the latter case $i=3$ and $n=m=j=4$. Then follows.
To prove we follow the previous approach with $d=q^4+q^3+q^2+q+1$. We obtain $$\sum a_ia_j^qa_m^{q^2}a_n^{q^3}a_r^{q^4}=\sum b_ib_j^qb_m^{q^2}b_n^{q^3}b_r^{q^4},$$ where the summation is on the quintuples $(i,j,m,n,r)$ with elements taken from $\{1,2,3,4\}$ such that $L_{i,j,m,n,r}:=(q^i-1)+(q^{j+1}-q)+(q^{m+2}-q^2)+(q^{n+3}-q^3)+(q^{r+4}-q^4)$ is divisible by $q^5-1$. Then $$L_{i,j,m,n,r} \equiv K_{i,j',m',n',r'} \pmod {q^5-1},$$ where $$K_{i,j',m',n',r'}=(q^i-1)+(q^{j'}-q)+(q^{m'}-q^2)+(q^{n'}-q^3)+(q^{r'}-q^4),$$ such that $$\label{trex}
j'\in \{0,2,3,4\}, \quad m'\in \{0,1,3,4\}, \quad n'\in \{0,1,2,4\}, \quad r'\in \{0,1,2,3\}$$ with $j'\equiv j+1,\, m'\equiv m+2,\, n'\equiv n+3,\, r'\equiv r+4 \pmod 5$.
For $q=2$ and $q=3$ we can determine by computer those quintuples $(i,j',m',n',r')$ for which $K_{i,j',m',n',r'}$ is divisible by $q^5-1$ and hence follows. So we may assume $q>3$. Then $$3-q^2-q^3-q^4=(q-1)+(1-q)+(1-q^2)+(1-q^3)+(1-q^4) \leq$$ $$K_{i,j',m',n',r'}\leq$$ $$(q^4-1)+(q^4-q)+(q^4-q^2)+(q^4-q^3)+(q^3-q^4)=3q^4-1-q-q^2,$$ and hence $L_{i,j,m,n,r}$ is divisible by $q^5-1$ if and only if $K_{i,j',m',n',r'}=0$. It follows that $$q^i+q^{j'}+q^{m'}+q^{n'}+q^{r'}=1+q+q^2+q^3+q^4.$$ So $\sum_{h=0}^4 c_h q^h=1+q+q^2+q^3+q^4$ for some $0\leq c_h \leq 4$ with $\sum_{h=0}^4 c_h=5$. For $q>3$ this happens only if $c_h=1$ for $h=0,1,2,3,4$ thus we have to find those quintuples $(i,j',m',n',r')$ for which $i\in \{1,2,3,4\}$, $\{i,j',m',n',r'\}=\{0,1,2,3,4\}$ and are satisfied. This can be done by computer and the 44 solutions yield .
Proof of Theorem \[main\] {#proof-of-theorem-main .unnumbered}
-------------------------
Since $f$ has maximum field of linearity ${{\mathbb F}}_q$, we cannot have $a_1=a_2=a_3=a_4=0$. If three of $\{a_1,a_2,a_3,a_4\}$ are zeros, then $f(x)=a_0x+a_ix^{q^i}$, for some $i\in\{1,2,3,4\}$. Hence with $\varphi$ represented by $\begin{pmatrix}
1 & 0 \\
-a_0/a_i & 1/a_i \\
\end{pmatrix}$ we have $f_{\varphi}(x)=x^{q^i}$. Then Proposition \[imtrans\] and Corollary \[norm\] give $g_\varphi(x)=\beta x^{q^j}$ where $\operatorname{N}(\beta)=1$ and $j\in\{1,2,3,4\}$. Now, we distinguish three main cases according to the number of zeros among $\{a_1,a_2,a_3,a_4\}$.
Two zeros among $\{a_1, a_2, a_3, a_4\}$ {#two-zeros-among-a_1-a_2-a_3-a_4 .unnumbered}
----------------------------------------
Applying Proposition \[prop\] we obtain $a_0=b_0$. The two non-zero coefficients can be chosen in six different ways, however the cases $a_1a_2\neq 0$ and $a_1a_3\neq 0$ correspond to $a_3a_4\neq 0$ and $a_2a_4\neq 0$, respectively, since $Im\,(f(x)/x)=Im \,(\hat f(x)/x)$. Thus, after possibly interchanging $f$ with $\hat f$, we may consider only four cases.
First let $f(x)=a_0x+a_1x^q+a_4x^{q^4}$, $a_1a_4\neq 0$. Applying Proposition \[prop\] we obtain $$\label{f1}
a_1a_4^q=b_1 b_4^q,$$ $$0=b_2 b_3^{q^2}.$$ Since $b_1b_4\neq 0$, from we get $b_2=b_3=0$ and hence gives $$\operatorname{N}(a_1)+\operatorname{N}(a_4)=\operatorname{N}(b_1)+\operatorname{N}(b_4).$$ Also, from we have $\operatorname{N}(a_1a_4)=\operatorname{N}(b_1b_4)$. It follows that either $\operatorname{N}(a_1)=\operatorname{N}(b_1)$ and $\operatorname{N}(a_4)=\operatorname{N}(b_4)$, or $\operatorname{N}(a_1)=\operatorname{N}(b_4)$ and $\operatorname{N}(a_4)=\operatorname{N}(b_1)$. In the first case $b_1=a_1\lambda^{q-1}$ for some $\lambda\in {{\mathbb F}}_{q^5}^*$ and by we get $g(x)=f(\lambda x)/\lambda$. In the latter case $b_1=a_4^q\lambda^{q-1}$ for some $\lambda\in {{\mathbb F}}_{q^5}^*$ and by we get $g(x)=\hat{f}(\lambda x)/\lambda$.
Now consider $f(x)=a_1x^q+a_3x^{q^3}$, $a_1a_3\neq 0$. Applying Proposition \[prop\] and arguing as above we have either $b_2=b_4=0$ or $b_1=b_3=0$. In the first case from we obtain $$a_1^qa_3^{1+q^2+q^3}=b_1^qb_3^{1+q^2+q^3}$$ and together with this yields $\operatorname{N}(a_1)=\operatorname{N}(b_1)$ and $\operatorname{N}(a_3)=\operatorname{N}(b_3)$. In this case $g(x)=f(\lambda x)/\lambda$ for some $\lambda\in {{\mathbb F}}_{q^5}^*$. If $b_1=b_3=0$, then in $\hat{g}(x)$ the coefficients of $x^{q^2}$ and $x^{q^4}$ are zeros thus applying the result obtained in the former case we get $\lambda\hat{g}(x)=f(\lambda x)$ and hence after substituting $y=\lambda x$ and taking the adjoints of both sides we obtain $g(y)=\hat{f}(\mu y)/\mu$, where $\mu=\lambda^{-1}$.
The cases $f(x)=a_1x^q+a_2x^{q^2}$ and $f(x)=a_2x^{q^2}+a_3x^{q^3}$ can be handled in a similar way, applying Equations – of Proposition \[prop\].
One zero among $\{a_1, a_2, a_3, a_4\}$ {#one-zero-among-a_1-a_2-a_3-a_4 .unnumbered}
---------------------------------------
Since $Im\,(f(x)/x)=Im \,(\hat f(x)/x)$, we may assume $a_1=0$ or $a_2=0$. First suppose $a_1=0$. Then by either $b_1=0$ or $b_4=0$. In the former case putting together Equations , , we get $\operatorname{N}(a_i)=\operatorname{N}(b_i)$ for $i\in \{2,3,4\}$ and hence there exists $\lambda\in {{\mathbb F}}_{q^5}^*$ such that $g(x)=f(\lambda x)/\lambda$. If $a_1=b_4=0$, then in $\hat{g}(x)$ the coefficients of $x^q$ is zero thus applying the previous result we get $g(x)=\hat{f}(\mu x)/\mu$, where $\mu=\lambda^{-1}$.
Now suppose $a_2=0$. Then by either $b_2=0$ or $b_3=0$. Using the same approach but applying , and we obtain $g(x)=f(\lambda x)/\lambda$ or $g(x)=\hat{f}(\lambda x)/\lambda$.
Case $a_1a_2a_3a_4\neq 0$ {#case-a_1a_2a_3a_4neq-0 .unnumbered}
-------------------------
We will apply - of Proposition \[prop\]. Note that Equations and yield $a_1a_2a_3a_4\neq 0 \Leftrightarrow b_1b_2b_3b_4\neq 0$. Multiplying by $a_2$ and applying yield $$a_2^2a_4^{q+q^2}-a_2(b_1^{q+1}b_3^{q^2}+b_2b_4^{q+q^2})+a_1^{q+1}b_3^{q^2}b_2=0.$$ Taking into account, this is equivalent to $$(a_2a_4^{q+q^2}-b_1^{q+1}b_3^{q^2})(a_2a_4^{q+q^2}-b_2b_4^{q+q^2})=0.$$ Multiplying by $a_1$ and applying yield $$a_1^2a_2^{q+q^3}-a_1(b_1b_2^{q+q^3}+b_3^{1+q^3}b_4^q)+a_3^{1+q^3}b_4^qb_1=0.$$ Taking into account, this is equivalent to $$(a_1a_2^{q+q^3}-b_1b_2^{q+q^3})(a_1a_2^{q+q^3}-b_3^{1+q^3}b_4^q)=0.$$ We distinguish four cases:
1. $a_2a_4^{q+q^2}=b_1^{q+1}b_3^{q^2}$ and $a_1a_2^{q+q^3}=b_1b_2^{q+q^3}$,
2. $a_2a_4^{q+q^2}=b_1^{q+1}b_3^{q^2}$ and $a_1a_2^{q+q^3}=b_3^{1+q^3}b_4^q$,
3. $a_2a_4^{q+q^2}=b_2b_4^{q+q^2}$ and $a_1a_2^{q+q^3}=b_1b_2^{q+q^3}$,
4. $a_2a_4^{q+q^2}=b_2b_4^{q+q^2}$ and $a_1a_2^{q+q^3}=b_3^{1+q^3}b_4^q$.
We show that these four cases produce the relations: $$\label{1per4}
\operatorname{N}\left(\frac{b_1}{a_4}\right)=\frac{a_1a_2^{q+q^3}}{a_4^qa_3^{q^3+1}}=\frac{b_1b_2^{q+q^3}}{b_4^qb_3^{q^3+1}},$$ $$\label{2per4}
\operatorname{N}\left(\frac{b_1}{a_4}\right)=1,$$ $$\label{3per4}
\operatorname{N}\left(\frac{b_1}{a_1}\right)=1,$$ $$\label{4per4}
\operatorname{N}\left(\frac{b_1}{a_1}\right)=\frac{a_3^{q^3+1}a_4^q}{a_1 a_2^{q+q^3}}=\frac{b_1b_2^{q+q^3}}{b_3^{q^3+1}b_4^q},$$ respectively. To see observe that from $a_2a_4^{q+q^2}=b_1^{q+1}b_3^{q^2}$ and we get $$\label{new}
\operatorname{N}\left(\frac{b_1}{a_4}\right)=\left(\frac{b_1^{q+1}}{a_4^{q+q^2}}\right)^{q^2+1}\frac{b_1^{q^4}}{a_4}=
\left(\frac{a_2^{q^2+1}}{b_3^{q^2+q^4}}\right)\frac{a_1^{q^4}}{b_4}=\frac{a_1a_2^{q+q^3}}{b_4^qb_3^{q^3+1}},$$ and hence by $a_1a_2^{q+q^3}=b_1b_2^{q+q^3}$ and we get . Equation immediately follows from taking $a_1a_2^{q+q^3}=b_3^{1+q^3}b_4^q$ into account. Similarly, using and we get and .
In Case 3 by we get $b_1=a_1\lambda^{q-1}$ for some $\lambda\in {{\mathbb F}}_{q^5}^*$ and by and we have $g(x)=f(\lambda x)/\lambda$. Analogously, in Case 2 $g(x)=\hat{f}(\lambda x)/\lambda$. Note that Case 4 is just Case 3 after replacing $g$ by $\hat{g}$ since $Im\,(g(x)/x)=Im \,(\hat g(x)/x)$. This allows us to restrict ourself to Case 1. It will be useful to express $a_1$, $a_2$, $a_3$ as follows: $$\label{ais}
a_1=\frac{b_1 b_4^q}{a_4^q}, \quad a_2=\frac{b_1^{q+1}b_3^{q^2}}{a_4^{q+q^2}}, \quad a_3=\frac{b_2^{q^3}b_4^{1+q^4}}{a_1^{q^3+q^4}}.$$ We are going to simplify . Using Equations and it is easy to see that $a_2^{1+q}a_3^{q^2+q^3}=b_2^{1+q}b_3^{q^2+q^3}$, $a_1^{1+q^2}a_4^{q+q^3}=b_1^{1+q^2}b_4^{q+q^3}$, $a_1^{q^2}a_2a_3^{q^3}a_4^q=b_1b_2^qb_3^{q^2}b_4^{q^3}$, $a_1^qa_2^{q^3}a_3a_4^{q^2}=b_1^qb_2^{q^3}b_3b_4^{q^2}$, $a_1a_2^qa_3^{q^2}a_4^{q^3}=b_1^{q^2}b_2b_3^{q^3}b_4^q$ and hence $$\label{eq5b}
a_1^{1+q+q^2}a_2^{q^3}+a_1^qa_3^{1+q^2+q^3}+a_2^{1+q+q^3}a_4^{q^2}+a_3a_4^{q+q^2+q^3}=$$ $$b_1^{1+q+q^2}b_2^{q^3}+b_1^qb_3^{1+q^2+q^3}+b_2^{1+q+q^3}b_4^{q^2}+b_3b_4^{q+q^2+q^3}.$$ The following equations can be proved applying , and : $$\label{ea}
\operatorname{N}\left(\frac{b_1}{a_4}\right)b_3b_4^{q+q^2+q^3}=a_2^{q^3}a_1^{1+q+q^2},$$ $$\label{eb}
\operatorname{N}\left(\frac{a_4}{b_1}\right)b_4^{q^2}b_2^{1+q+q^3}=a_1^qa_3^{1+q^2+q^3},$$ $$\label{ec}
\operatorname{N}\left(\frac{b_1}{a_4}\right)b_1^qb_3^{1+q^2+q^3}=a_2^{1+q+q^3}a_4^{q^2},$$ $$\label{ed}
\operatorname{N}\left(\frac{a_4}{b_1}\right)b_2^{q^3}b_1^{1+q+q^2}=a_3a_4^{q+q^2+q^3}.$$
Then can be written as $$(\operatorname{N}(b_1/a_4)-1)(b_3b_4^{q+q^2+q^3}+b_1^qb_3^{1+q^2+q^3})=\frac{\operatorname{N}(b_1/a_4)-1}{\operatorname{N}(b_1/a_4)}(b_4^{q^2}b_2^{1+q+q^3}+b_2^{q^3}b_1^{1+q+q^2}).$$ If $\operatorname{N}(b_1/a_4)=1$, then we are in Case 2 and hence the assertion follows. Otherwise dividing by $\operatorname{N}(b_1/a_4)-1$ and substituting $\operatorname{N}(b_1/a_4)=b_1b_2^{q+q^3}/b_4^qb_3^{q^3+1}$ we obtain $$b_1b_2^{q+q^3}(b_3b_4^{q+q^2+q^3}+b_1^qb_3^{1+q^2+q^3})=b_4^qb_3^{q^3+1}(b_4^{q^2}b_2^{1+q+q^3}+b_2^{q^3}b_1^{1+q+q^2}).$$ Substituting $\operatorname{N}(b_1/a_4) b_4^qb_3^{q^3+1}/b_2^{q+q^3}$ for $b_1$ and using the fact that $\operatorname{N}(b_1/a_4)\in {{\mathbb F}}_q$ we obtain $$\left(1-\operatorname{N}\left(\frac{b_1}{a_4}\right)^2\operatorname{N}\left(\frac{b_3}{b_2}\right)\right)\left(\operatorname{N}\left(\frac{b_1}{a_4}\right)b_4^{q+q^3}b_3-b_2^{1+q+q^3}\right)=0.$$ This gives us two possibilities: $$\label{sec}
\operatorname{N}\left(\frac{b_1}{a_4}\right)b_4^{q+q^3}b_3=b_2^{1+q+q^3},$$ or $$\label{sec2}
\operatorname{N}\left(\frac{b_2}{b_3}\right)=\operatorname{N}\left(\frac{b_1}{a_4}\right)^2.$$ First consider the case when holds. We show $\operatorname{N}(a_1)=\operatorname{N}(b_1)$, that is, . We have $a_2a_4^{q+q^2}=b_1^{q+1}b_3^{q^2}$ from and hence $\operatorname{N}(a_2)\operatorname{N}(a_4)^2=\operatorname{N}(b_1)^2\operatorname{N}(b_3)$. It follows that $$\operatorname{N}\left(\frac{b_1}{a_4}\right)^2=\operatorname{N}\left(\frac{a_2}{b_3}\right).$$ Combining this with we obtain $\operatorname{N}(b_2)=\operatorname{N}(a_2)$. Then $\operatorname{N}(b_1)=\operatorname{N}(a_1)$ follows from $a_1a_2^{q+q^3}=b_1b_2^{q+q^3}$ since we are in Case 1.
From now on we can suppose that holds. Then yields $$\label{elso}
\left(\frac{b_1}{b_2}\right)^{q^2}=\frac{b_3}{b_4}.$$ Multiplying both sides of by $b_4^{q^2}$ and applying gives $$\label{nuovo}
a_2^{q^3}a_1^{1+q+q^2}=b_2^{1+q+q^3}b_4^{q^2}.$$ Then multiplying by and taking into account we obtain $$a_1^qa_3^{1+q^2+q^3}=b_3b_4^{q+q^2+q^3}.$$ Multiplying and yield $$(b_1^qb_3^{1+q^2+q^3})(b_2^{q^3}b_1^{1+q+q^2})=(a_2^{1+q+q^3}a_4^{q^2})(a_3a_4^{q+q^2+q^3}).$$ On the other hand, from it follows that $$b_1^qb_3^{1+q^2+q^3}+b_2^{q^3}b_1^{1+q+q^2}=a_2^{1+q+q^3}a_4^{q^2}+a_3a_4^{q+q^2+q^3}.$$ Hence $b_1^qb_3^{1+q^2+q^3}=a_2^{1+q+q^3}a_4^{q^2}$ and $b_2^{q^3}b_1^{1+q+q^2}=a_3a_4^{q+q^2+q^3}$, or $b_1^qb_3^{1+q^2+q^3}=a_3a_4^{q+q^2+q^3}$ and $b_2^{q^3}b_1^{1+q+q^2}=a_2^{1+q+q^3}a_4^{q^2}$. In the former case yields $\operatorname{N}(b_1/a_4)=1$, which is . In the latter case and gives $$\frac{b_4^qb_3^{q^3+1}}{b_1b_2^{q+q^3}}b_2^{q^3}b_1^{1+q+q^2}=\operatorname{N}(a_4/b_1)b_2^{q^3}b_1^{1+q+q^2}=b_1^qb_3^{1+q^2+q^3},$$ and hence $$\label{masodik}
\frac{b_4}{b_2}=\left(\frac{b_3}{b_1}\right)^q.$$ Equation is equivalent to $$b_4b_1^{q^2}=b_3b_2^{q^2},$$ while is equivalent to $$b_4b_1^q=b_3^qb_2.$$ Dividing these two equations by each other yield $$b_2^{q^2-1}=b_3^{q-1}b_1^{q^2-q}.$$ It follows that there exists $\lambda\in {{\mathbb F}}_q^*$ such that $$\label{lambda}
b_2^{q+1}=\lambda b_3 b_1^q,$$ thus $$b_3=b_2^{q+1}/(b_1^q\lambda)$$ and $$b_4=b_2^{1+q+q^2}/(b_1^{q+q^2}\lambda).$$ Then can be written as $$\operatorname{N}\left(\frac{b_1}{a_4}\right)=\frac{b_1b_2^{q+q^3}}{b_4^qb_3^{q^3+1}}=\operatorname{N}\left(\frac{b_1}{b_2}\right)\lambda^3,$$ and hence $$\operatorname{N}\left(\frac{b_2}{a_4}\right)=\lambda^3.$$ Then, using the previous expressions for $b_3$ and $b_4$ and taking and into account, we can express $\operatorname{N}(a_i)$ for $i=1,2,3,4$ as follows $$\label{bp1}
\operatorname{N}(a_1)=\operatorname{N}(b_2)^2/(\operatorname{N}(b_1)\lambda^2),$$ $$\label{bp2}
\operatorname{N}(a_2)=\operatorname{N}(b_1)\lambda,$$ $$\operatorname{N}(a_3)=\operatorname{N}(b_2)^3/(\operatorname{N}(b_1)^2\lambda^6),$$ $$\operatorname{N}(a_4)=\operatorname{N}(b_2)/\lambda^3.$$
Before we go further, we simplify and prove $$\label{sumofnorms}
\operatorname{N}(a_1)+\operatorname{N}(a_2)+\operatorname{N}(a_3)+\operatorname{N}(a_4)=\operatorname{N}(b_1)+\operatorname{N}(b_2)+\operatorname{N}(b_3)+\operatorname{N}(b_4).$$ It is enough to show $$\operatorname{Tr}(\overbrace{\hbox{$a_1^qa_2^{q^2+q^3+q^4}a_3$}}^{A_1}+\overbrace{\hbox{$a_1^{q+q^3}a_2^{q^4}a_3^{1+q^2}$}}^{A_2}+
\overbrace{\hbox{$a_1^{q+q^2}a_2^{q^3+q^4}a_4$}}^{A_3}+\overbrace{\hbox{$a_1^{q+q^2+q^4}a_3^{q^3}a_4$}}^{A_4}+$$ $$\overbrace{\hbox{$a_2^qa_3^{q^2+q^3+q^4}a_4$}}^{A_5}+\overbrace{\hbox{$a_1^{q^2}a_3^{q^3+q^4}a_4^{1+q}$}}^{A_6}+
\overbrace{\hbox{$a_2^{q+q^3}a_3^{q^4}a_4^{1+q^2}$}}^{A_7}+\overbrace{\hbox{$a_1^{q^2}a_2^{q^4}a_4^{1+q+q^3}$}}^{A_8})=$$ $$\operatorname{Tr}(\overbrace{\hbox{$b_1^qb_2^{q^2+q^3+q^4}b_3$}}^{B_1}+\overbrace{\hbox{$b_1^{q+q^3}b_2^{q^4}b_3^{1+q^2}$}}^{B_7}+
\overbrace{\hbox{$b_1^{q+q^2}b_2^{q^3+q^4}b_4$}}^{B_3}+\overbrace{\hbox{$b_1^{q+q^2+q^4}b_3^{q^3}b_4$}}^{B_8}+$$ $$\overbrace{\hbox{$b_2^qb_3^{q^2+q^3+q^4}b_4$}}^{B_5}+\overbrace{\hbox{$b_1^{q^2}b_3^{q^3+q^4}b_4^{1+q}$}}^{B_6}+
\overbrace{\hbox{$b_2^{q+q^3}b_3^{q^4}b_4^{1+q^2}$}}^{B_2}+\overbrace{\hbox{$b_1^{q^2}b_2^{q^4}b_4^{1+q+q^3}$}}^{B_4}),$$ which can be done by proving $\operatorname{Tr}(A_i)=\operatorname{Tr}(B_i)$ for $i=1,2,\ldots,8$. Expressing $a_3$ with $a_4$ in gives $a_3=b_2^{q^3}a_4^{q^4+1}/b_1^{q^3+q^4}$. Then $a_1,a_2,a_3$ can be eliminated in all of the $A_i$, $i\in \{1,2\ldots,8\}$. It turns out that this procedure eliminates also $a_4$ when $i\in\{2,4,7,8\}$ and we obtain $A_2=B_2^{q^2}$, $A_4=B_4^{q^2}$, $A_7=B_7^{q^3}$ and $A_8=B_8^{q^2}$. In each of the other cases what remains is $\operatorname{N}(a_4)$ times an expression in $b_1,b_2,b_3,b_4$. Then by using we can also eliminate $\operatorname{N}(a_4)$ and hence $A_i$ can be expressed in terms of $b_1,b_2,b_3,b_4$. This gives $A_1=B_1$ and $A_5=B_5$. Applying also $\eqref{elso}$ and $\eqref{masodik}$ we obtain $A_3=B_3^{q^2}$ and $A_6=B_6$.
Let $x=\operatorname{N}(b_2/b_1)$, then gives us the following equation $$x^2\lambda^4+\lambda^7+x^3+x\lambda^3=\lambda^6+x\lambda^6+x^2\lambda+\lambda x^3.$$ After rearranging we get: $$(1-\lambda)(x-\lambda)(x-\lambda^2)(x-\lambda^3)=0.$$ First suppose $\lambda\neq 1$, then we have three possibilities: $$x=\lambda,$$ in which case $N(b_1)=N(a_1)$ follows from , which is Case 3; $$x=\lambda^3,$$ in which case $N(a_4)=N(b_1)$ follows from , which is Case 2; $$x=\lambda^2,$$ in which case we show that there exists $\varphi\in\Gamma\mathrm{L}(2,q^5)$ such that either $Im\,(g_\varphi(x)/x)=Im\,(x^{q-1})$ or $Im\,(g_\varphi(x)/x)=Im\,(\operatorname{Tr}(x)/x)$. In the former case by Proposition \[imtrans\] and Corollary \[norm\] we get $f_\varphi(x)=\alpha x^{q^i}$ and $g_\varphi(x)=\beta x^{q^j}$ for some $i,j\in\{1,2,3,4\}$, with $\operatorname{N}(\alpha)=\operatorname{N}(\beta)=1$. In the latter case, by Theorem \[trace\] and by Propositions \[imtrans\] and \[lambda0\], there exists $\lambda\in{{\mathbb F}}_{q^5}^*$ such that $g(x)=f(\lambda x)/\lambda$.
According to Proposition \[pseudoalg\] part 2, it is enough to show $$(b_4/b_1)^{q^2}=b_1/b_3, \quad (b_1/b_2)^{q^2}=b_3/b_4.$$ The second equation is just , thus it is enough to prove the first one. First we show $$\label{meta}
b_2b_3^{q+q^3}=b_1^{1+q+q^3}.$$ From we have $$\operatorname{N}\left(\frac{b_2}{b_1}\right)=\lambda^2=\left(\frac{b_2^{q+1}}{b_3b_1^q}\right)^2,$$ and hence after rearranging $$\frac{b_2^{q^2+q^3+q^4}b_3}{b_1^{1+q^2+q^3+q^4}}=\frac{b_2^{q+1}}{b_3b_1^q}.$$ On the right-hand side we have $\lambda$, which is in ${{\mathbb F}}_q$, thus, after taking $q$-th powers on the left and $q^3$-th powers on the right, the following also holds $$\frac{b_2^{q^3+q^4+1}b_3^q}{b_1^{q+q^3+q^4+1}}=\frac{b_2^{q^3+q^4}}{b_3^{q^3}b_1^{q^4}}.$$ After rearranging we obtain . Now we show that $(b_4/b_1)^{q^2}=b_1/b_3$ is equivalent to . Expressing $b_4$ from we get $$(b_4/b_1)^{q^2}=b_1/b_3 \Leftrightarrow b_3^{1+q^2}b_2^{q^4}=b_1^{1+q^2+q^4},$$ where the equation on the right-hand side is just the $q^4$-th power of .
Now consider the case when $\lambda=1$. Then $b_3=b_2^{q+1}/b_1^q$, $b_4=b_2^{1+q+q^2}/b_1^{q+q^2}$ and it follows from Proposition \[pseudoalg\] that there exists $\varphi\in\Gamma\mathrm{L}(2,q^5)$ such that either $Im\,(g_\varphi(x)/x)=Im\,(x^{q-1})$ or $Im\,(g_\varphi(x)/x)=Im\,(\operatorname{Tr}(x)/x)$. As above, the assertion follows either from Proposition \[imtrans\] and Corollary \[norm\] or from Theorem \[trace\] and by Propositions \[imtrans\] and \[lambda0\].
This finishes the proof when $\prod_{i=1}^4 a_ib_i \neq 0$.
New maximum scattered linear sets of PG(1,q5) {#geom}
=============================================
A point set $L$ of a line $\Lambda=\operatorname{{PG}}(W,{{\mathbb F}}_{q^n})\allowbreak=\operatorname{{PG}}(1,q^n)$ is said to be an *${{\mathbb F}}_q$-linear set* of $\Lambda$ of rank $n$ if it is defined by the non-zero vectors of an $n$-dimensional ${{\mathbb F}}_q$-vector subspace $U$ of the two-dimensional ${{\mathbb F}}_{q^n}$-vector space $W$, i.e. $$L=L_U:=\{{\langle}{\bf u} {\rangle}_{\mathbb{F}_{q^n}} \colon {\bf u}\in U\setminus \{{\bf 0} \}\}.$$ One of the most natural questions about linear sets is their equivalence. Two linear sets $L_U$ and $L_V$ of $\operatorname{{PG}}(1,q^n)$ are said to be *$\mathrm{P\Gamma L}$-equivalent* (or simply *equivalent*) if there is an element in $\mathrm{P\Gamma L}(2,q^n)$ mapping $L_U$ to $L_V$. In the applications it is crucial to have methods to decide whether two linear sets are equivalent or not. This can be a difficult problem and some results in this direction can be found in [@CSZ2015; @CSMP2016]. If $L_U$ and $L_V$ are two equivalent ${{\mathbb F}}_q$-linear sets of rank $n$ in $\operatorname{{PG}}(1,q^n)$ and $\varphi$ is an element of ${\mathrm{\Gamma L}}(2,q^n)$ which induces a collineation mapping $L_U$ to $L_V$, then $L_{U^\varphi}=L_V$. Hence the first step to face with the equivalence problem for linear sets is to determine which ${{\mathbb F}}_q$-subspaces can define the same linear set.
For any $q$-polynomial $f(x)=\sum_{i=0}^{n-1} a_i x^{q^i}$ over ${{\mathbb F}}_{q^n}$, the graph ${\mathcal G}_f=\{(x,f(x)) \colon x\in {{\mathbb F}}_{q^n}\}$ is an ${{\mathbb F}}_q$-vector subspace of the 2-dimensional vector space $V={{\mathbb F}}_{q^n}\times{{\mathbb F}}_{q^n}$ and the point set $$L_f:=L_{{\mathcal G}_f}=\{{\langle}(x,f(x)){\rangle}_{\mathbb{F}_{q^n}} \colon x\in {{\mathbb F}}_{q^n}^*\}$$ is an ${{\mathbb F}}_q$-linear set of rank $n$ of $\operatorname{{PG}}(1,q^n)$. In this context, the problem posed in corresponds to find all ${{\mathbb F}}_q$-subspaces of $V$ of rank $n$ (cf. [@CSMP2016 Proposition 2.3]) defining the linear set $L_f$. The maximum field of linearity of $f$ is the maximum field of linearity of $L_f$, and it is well-defined (cf. Proposition \[fieldoflinearity\] and [@CSMP2016 Proposition 2.3]). Also, by the Introduction from any $q$-polynomial $f$ over ${{\mathbb F}}_{q^n}$, the linear sets $L_f$, $L_{f_\lambda}$ (with $f_\lambda(x):=f(\lambda x)/\lambda$ for each $\lambda\in{{\mathbb F}}_{q^n}^*$) and $L_{\hat f}$ coincide (cf. [@BGMP2015 Lemma 2.6] and the first part of [@CSMP2016 Section 3]). If $f$ and $g$ are two equivalent $q$-polynomials over ${{\mathbb F}}_{q^n}$, i.e. ${\mathcal G}_f$ an ${\mathcal G}_g$ are equivalent w.r.t. the action of the group $\Gamma{\mathrm L}(2,q^n)$, then the corresponding ${{\mathbb F}}_q$-linear sets $L_f$ and $L_g$ of $\operatorname{{PG}}(1,q^n)$ are ${\mathrm P}\Gamma{\mathrm L}(2,q^n)$-equivalent. The converse does not hold (see [@CSZ2015] and [@CSMP2016] for further details). More precisely,
\[pgamma-equiv\] Let $L_f$ and $L_g$ be two ${{\mathbb F}}_q$-linear sets of rank $n$ of $\operatorname{{PG}}(1,q^n)$. Then $L_f$ and $L_g$ are $\mathrm{P\Gamma L}(2,q^n)$-equivalent if and only if there exists and element $\varphi\in\Gamma\mathrm{L}(2,q^n)$ such that $Im\,(f_\varphi(x)/x)=Im\,(g(x)/x)$.
Linear sets of rank $n$ of $\operatorname{{PG}}(1,q^n)$ have size at most $(q^n-1)/(q-1)$. A linear set $L_U$ of rank $n$ whose size achieves this bound is called *maximum scattered*. For applications of these objects we refer to [@OP2010] and [@Lavrauw].
Following [@LuMaPoTr2014] and [@DoDu9] a maximum scattered ${{\mathbb F}}_q$-linear set $L_U$ of rank $n$ in $\operatorname{{PG}}(1,q^n)$ is of *pseudoregulus type* if it is $\mathrm{P}\Gamma\mathrm{L}(2,q^n)$-equivalent to $L_f$ with $f(x)=x^q$ or, equivalently, if there exists $\varphi\in\mathrm{GL}(2,q^n)$ such that $$L_{U^\varphi}=\{{\langle}(x,x^q){\rangle}_{{{\mathbb F}}_{q^n}}\colon x\in{{\mathbb F}}_{q^n}^*\}.$$ By Proposition \[pgamma-equiv\] and Corollary \[norm\], it follows
\[peseud-equiv\] An ${{\mathbb F}}_q$-linear set $L_f$ of rank $n$ of $\operatorname{{PG}}(1,q^n)$ is of pseudoregulus type if and only if $f(x)$ is equivalent to $x^{q^i}$ for some $i$ with $\gcd(i,n)=1$.
For the proof of the previous result see also [@LaShZa2013].
The known pairwise non-equivalent families of $q$-polynomials over ${{\mathbb F}}_{q^n}$ which define maximum scattered linear sets of rank $n$ in $\operatorname{{PG}}(1,q^n)$ are
1. $f_{s}(x)= x^{q^s}$, $1\leq s\leq n-1$, $\gcd(s,n)=1$ ([@BL2000; @CSZ2016]),
2. $g_{s,\delta}(x)= \delta x^{q^s} + x^{q^{n-s}}$, $n\geq 4$, $\operatorname{N}_{q^n/q}(\delta)\notin \{0,1\}$ [^3], $\gcd(s,n)=1$ ([@LP2001] for $s=1$, [@Sh; @LTZ] for $s\neq 1$),
3. $h_{s,\delta}(x):= \delta x^{q^s}+x^{q^{s+n/2}}$, $n\in \{6,8\}$, $\gcd(s,n/2)=1$, $\operatorname{N}_{q^n/q^{n/2}}(\delta) \notin \{0,1\}$, for the precise conditions on $\delta$ and $q$ see [@CMPZ Theorems 7.1 and 7.2] [^4],
4. $k_b(x):=x^q+x^{q^3}+bx^{q^5}$, $n=6$, with $b^2+b=1$, $q\equiv 0,\pm 1 \pmod 5$ ([@CSMZ2017]).
All the previous polynomials in cases 2.,3.,4. above are examples of functions which are not equivalent to monomials but the set of directions determined by their graph has size $(q^n-1)/(q-1)$, i.e. the corresponding linear sets are maximum scattered. The existence of such linearized polynomials is briefly discussed also in [@Praha p. 132].
For $n=2$ the maximum scattered ${{\mathbb F}}_q$-linear sets coincide with the Baer sublines. For $n=3$ the maximum scattered linear sets are all of pseudoregulus type and the corresponding $q$-polynomials are all $\mathrm{GL}(2,q^3)$-equivalent to $x^q$ (cf. [@LaVa2010]). For $n=4$ there are two families of maximum scattered linear sets. More precisely, if $L_f$ is a maximum scattered linear set of rank $4$ of $\operatorname{{PG}}(1,q^4)$, with maximum field of linearity ${{\mathbb F}}_{q}$, then there exists $\varphi\in\mathrm{GL}(2,q^4)$ such that either $f_\varphi(x)=x^q$ or $f_\varphi(x)=\delta x^q+x^{q^3}$, for some $\delta\in{{\mathbb F}}_{q^4}^*$ with $\operatorname{N}_{q^4/q}(\delta)\notin\{0,1\}$ (cf. [@CSZ2017]). It is easy to see that $L_{f_{1}}=L_{f_s}$ for any $s$ with $\gcd(s,n)=1$, and $f_i$ is equivalent to $f_j$ if and only if $j\in\{i,n-i\}$. Also, the graph of $g_{s,\delta}$ is $\mathrm{GL}(2,q^n)$-equivalent to the graph of $g_{n-s,\delta^{-1}}$.
In [@LP2001 Theorem 3] Lunardon and Polverino proved that $L_{g_{1,\delta}}$ and $L_{f_1}$ are not $\mathrm{P}\Gamma \mathrm{L}(2,q^n)$-equivalent when $q>3$, $n\geq 4$. This was extended also for $q=3$ [@CSMZ2017 Theorem 3.4]. Also in [@CSMZ2017], it has been proven that for $n=6,8$ the linear sets $L_{f_1}$, $L_{g_{s,\delta}}$, $L_{h_{s',\delta'}}$ and $L_{k_b}$ are pairwise non-equivalent for any choice of $s,s',\delta, \delta',b$.
In this section we prove that one can find for each $q>2$ a suitable $\delta$ such that $L_{g_{2,\delta}}$ of $\operatorname{{PG}}(1,q^5)$ is not equivalent to the linear sets $L_{g_{1,\mu}}$ of $\operatorname{{PG}}(1,q^5)$ for each $\mu\in{{\mathbb F}}_{q^5}^*$, with $\operatorname{N}_{q^5/q}(\mu)\notin\{0,1\}$. In order to do this, we first reformulate Theorem \[main\] as follows.
\[main1\] Let $f(x)$ and $g(x)$ be two $q$-polynomials over ${{\mathbb F}}_{q^5}$ such that $L_f=L_g$. Then either $L_f=L_g$ is of pseudoregulus type or there exists $\lambda\in {{\mathbb F}}_{q^5}^*$ such that $g(x)=f(\lambda x)/\lambda$ or $g(x)=\hat{f}(\lambda x)/\lambda$ holds.
Let $g_{2,\delta}(x)=\delta x^{q^2}+x^{q^3}$ for some $\delta\in {{\mathbb F}}_{q^5}^*$ with $\operatorname{N}(\delta)^5 \neq 1$. Then $L_{g_{2,\delta}}$ is not ${\mathrm P}\Gamma{\mathrm L}(2,q^5)$-equivalent to any linear set $L_{g_{1,\mu}}$ and hence it is a new a maximum scattered linear set.
Suppose, contrary to our claim, that $L_{g_{2,\delta}}$ is ${\mathrm P}\Gamma{\mathrm L}(2,q^5)$-equivalent to a linear set $L_{g_{1,\mu}}$. From Proposition \[pgamma-equiv\] and Theorem \[main1\], taking into account that $L_{g_{1,\mu}}$ is not of pseudoregulus type, it follows that there exist $\varphi\in\Gamma{\mathrm L}(2,q^5)$ and $\lambda\in{{\mathbb F}}_{q^5}^*$ such that either $(g_{2,\delta})_{\varphi}(x)=g_{1,\mu}(\lambda x)/\lambda$ or $(g_{2,\delta})_{\varphi}(x)=\hat g_{1,\mu}(\lambda x)/\lambda$. This is equivalent to say that there exist $A,B,C,D \in {{\mathbb F}}_{q^5}$ with $AD-BC \neq 0$ and a field automorphism $\tau$ of ${{\mathbb F}}_{q^5}$ such that $$\left\{
\begin{pmatrix}
A & B \\
C & D \\
\end{pmatrix}
\begin{pmatrix}
x^\tau \\
h(x)^\tau \\
\end{pmatrix}
\colon x\in {{\mathbb F}}_{q^5}
\right\}
=
\left\{
\begin{pmatrix}
z\\
\alpha z^q + \beta z^{q^4}\\
\end{pmatrix}
\colon z\in {{\mathbb F}}_{q^5}
\right\}
,$$ where $N(\alpha) \neq N(\beta)$ and $\alpha \beta \neq 0$. We may substitute $x^{\tau}$ by $y$, then $$\alpha (Ay+B\delta^\tau y^{q^2}+By^{q^3})^q+\beta (Ay+B\delta^\tau y^{q^2}+By^{q^3})^{q^4} =$$ $$Cy+D\delta^\tau y^{q^2}+Dy^{q^3}$$ for each $y\in {{\mathbb F}}_{q^5}$. Comparing coefficients yields $C=0$ and $$\label{1}
\alpha A^q+\beta B^{q^4}\delta^{q^4 \tau}=0,$$ $$\label{2}
\beta B^{q^4}=D\delta^\tau,$$ $$\label{3}
\alpha B^q \delta^{q\tau} =D,$$ $$\label{4}
\alpha B^q +\beta A^{q^4}=0.$$ Conditions and give $$\label{primo}
B^{q^4-q}= \delta^{(q+1)\tau} \alpha/\beta.$$ On the other hand from we get $A^q=- B^{q^3}\alpha^{q^2}/\beta^{q^2}$ and substituting this into we have $$\label{secondo}
B^{q^3-q^4}=\delta^{q^4 \tau} \beta^{q^2+1}/\alpha^{q^2+1}.$$ Equations and give $N(\beta/\alpha)=N(\delta)^{2\tau}$ and $N(\alpha/\beta)^2=N(\delta)^{\tau}$, respectively. It follows that $N(\delta)^{5\tau}=1$ and hence $N(\delta)^5=1$, a contradiction.
Open problems {#open-problems .unnumbered}
=============
We conclude the paper by the following open problems.
1. Is it true also for $n>5$ that for any pair of $q$-polynomials $f(x)$ and $g(x)$ of ${{\mathbb F}}_{q^n}[x]$, with maximum field of linearity ${{\mathbb F}}_q$, if $Im\,(f(x)/x)=Im\,(g(x)/x)$ then either there exists $\varphi\in \Gamma\mathrm{L}(2,q^n)$ such that $f_{\varphi}(x)=\alpha x^{q^i}$ and $g_{\varphi}(x)=\beta x^{q^j}$ with $\operatorname{N}(\alpha)=\operatorname{N}(\beta)$ and $\gcd(i,n)=\gcd(j,n)=1$, or there exists $\lambda\in {{\mathbb F}}_{q^n}^*$ such that $g(x)=f(\lambda x)/\lambda$ or $g(x)=\hat f(\lambda x)/\lambda$?
2. Is it possible, at least for small values of $n>4$, to classify, up to equivalence, the $q$-polynomials $f(x)\in {{\mathbb F}}_{q^n}[x]$ such that $|Im\,(f(x)/x)|=(q^n-1)/(q-1)$? Find new examples!
3. Is it possible, at least for small values of $n$, to classify, up to equivalence, the $q$-polynomials $f(x)\in {{\mathbb F}}_{q^n}[x]$ such that $|Im\,(f(x)/x)|=q^{n-1}+1$? Find new examples!
4. Is it possible, at least for small values of $n$, to classify, up to equivalence, the $q$-polynomials $f(x)\in {{\mathbb F}}_{q^n}[x]$ such that in the multiset $\{f(x)/x \colon x\in {{\mathbb F}}_{q^n}^*\}$ there is a unique element which is represented more than $q-1$ times? In this case the linear set $L_f$ is an *$i$-club* of rank $n$ and when $q=2$, then such linear sets correspond to translation KM-arcs cf. [@KMarcs2] (a KM-arc, or $(q+t)$-arc of type $(0,2,t)$, is a set of $q+t$ points of $\operatorname{{PG}}(2,2^n)$, such that each line meet the point set in 0,2 or in $t$ points, cf. [@KMarcs]). Find new examples!
5. Determine the equivalence classes of the set of $q$-polynomials in ${{\mathbb F}}_{q^4}[x]$.
6. Determine, at least for small values of $n$, all the possible sizes of $Im\,(f(x)/x)$ where $f(x)\in {{\mathbb F}}_{q^n}[x]$ is a $q$-polynomial.
[pippo]{}
The number of directions determined by a function over a finite field, J. Combin. Theory Ser. A [**104**]{} (2003), 341–350.
On the number of slopes of the graph of a function definied over a finite field, J. Combin. Theory Ser. A [**86**]{} (1999), 187–196.
Maximum scattered linear sets and complete caps in Galois spaces, Combinatorica [**38**]{} (2018), to appear. DOI: 10.1007/s00493-016-3531-6.
$\mathbb{F}_q$-linear blocking sets in $\mathrm{PG}(2,q^4)$, Innov. Incidence Geom. [**2**]{} (2005), 35–56.
Scattered spaces with respect to a spread in $\mathrm{PG}(n,q)$, Geom. Dedicata [**81**]{} (2000), 231–243.
A theorem on permutations of a finite field, Canad. J. Math. [**25**]{} (1973), 1060–-1065.
A theorem on permutations in a finite field, Proc. Amer. Math. Soc. [**11**]{} (1960), 456-–459.
On bisecants of Rédei type blocking sets and applications, Combinatorica (2018), to appear. DOI: 10.1007/s00493-016-3442-6
, J. Combin. Theory Ser. A [**157**]{} (2018), 402–426.
A new family of MRD-codes, Linear Algebra Appl. [**548**]{} (2018), 203–220.
New maximum scattered linear sets of the projective line, <https://arxiv.org/abs/1709.00926>
On the equivalence of linear sets, Des. Codes Cryptogr. [**81**]{} (2016), 269–281.
On linear sets of pseudoregulus type in $\mathrm{PG}(1,q^t)$, Finite Fields Appl. 41 (2016), 34–54.
Maximum scattered $\mathbb{F}_q$-linear sets of $\mathrm{PG}(1,q^4)$, Discrete Math. [**341**]{} (2018), 74–80.
A linear set view on KM-arcs, J. Algebraic. Combin. [**44**]{}, n.1 (2016), 131–164.
Scattered linear sets generated by collineations between pencils of lines, J. Algebraic. Combin. [**40**]{}, n. 4 (2014), 1121–1131.
On theorems of Carlitz and Payne on permutation polynomials over finite fields with an application to $x^{-1}+L(x)$, Finite Fields Appl. [**27**]{} (2014), 130-–142.
G. Korchmáros and F. Mazzocca, [*On $(q+t)$-arcs of type $(0,2,t)$ in a desarguesian plane of order $q$*]{}, Math. Proc. Cambridge Philos. Soc. 108 (1990), 445–459.
Scattered spaces in Galois Geometry, [*Contemporary Developments in Finite Fields and Applications*]{}, 2016, 195–216.
On embeddings of minimum dimension of $\mathrm{PG}(n,q)\times \mathrm{PG}(n,q)$, Des. Codes Cryptogr. [**74**]{} (2015), 427–440.
On linear sets on a projective line, Des. Codes Cryptogr. [**56**]{} (2010), 89–104.
Maximum scattered linear sets of pseudoregulus type and the Segre Variety ${\cal S}_{n,n}$, J. Algebr. Comb. **39** (2014), 807–831.
Blocking Sets and Derivable Partial Spreads, J. Algebraic Combin. [**14**]{} (2001), 49–56.
Generalized Twisted Gabidulin Codes, <http://arxiv.org/abs/1507.07855>.
Pseudo-ordered polynomials over a finite field, Acta Arith. [**8**]{} (1962–1963), 127–-151.
A complete determination of translation ovoids in finite Desarguian planes, Atti Accad. Naz. Lincei, Rend. Cl. Sci. Fis. Mat. Nat. [**51**]{} (8) (1971), 328-–331.
Linear sets in finite projective spaces, Discrete Math. [**310**]{} (2010), 3096–3107.
A new family of linear maximum rank distance codes, [*Adv. Math. Commun.*]{} [**10**]{}(3) (2016), 475–488.
Bence Csajbók\
MTA–ELTE Geometric and Algebraic Combinatorics Research Group\
ELTE Eötvös Loránd University, Budapest, Hungary\
Department of Geometry\
1117 Budapest, Pázmány P. stny. 1/C, Hungary\
[[*[email protected]*]{}]{}
Giuseppe Marino, Olga Polverino\
Dipartimento di Matematica e Fisica,\
Università degli Studi della Campania “Luigi Vanvitelli”,\
Viale Lincoln 5, I-81100 Caserta, Italy\
[[*[email protected]*]{}, [*[email protected]*]{}]{}
[^1]:
[^2]: An element $\varphi\in {\mathrm{\Gamma L}}(2,q^n)$ represented by the invertible matrix $\begin{pmatrix}
a&b\\
c&d\\
\end{pmatrix}$ and with companion automorphism $\sigma$ of ${{\mathbb F}}_{q^n}$ is said to be *admissible* w.r.t. a given $q$-polynomial $f$ if either $b=0$ or $-(a/b)^{\sigma^{-1}}\notin Im\, (f(x)/x)$.
[^3]: This condition implies $q\neq 2$.
[^4]: Also here $q>2$, otherwise the linear set defined by $h_{s,\delta}$ is never scattered.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
Jan Luts[^1] ,\
School of Mathematical Sciences, University of Technology Sydney\
\
bibliography:
- 'distributedRTVB.bib'
title: '**Real-time semiparametric regression for distributed data sets**'
---
**Abstract**
This paper proposes a method for semiparametric regression analysis of large-scale data which are distributed over multiple hosts. This enables modeling of nonlinear relationships and both the batch approach, where analysis starts after all data have been collected, and the real-time setting are addressed. The methodology is extended to operate in evolving environments, where it can no longer be assumed that model parameters remain constant over time. Two areas of application for the methodology are presented: regression modeling when there are multiple data owners and regression modeling within the MapReduce framework. A website, `realtime-semiparametric-regression.net`, illustrates the use of the proposed method on United States domestic airline data in real-time.
<span style="font-variant:small-caps;">Keywords</span>: [distributed learning; semiparametric regression; variational Bayes; data streams; evolving environments; real-time; MapReduce; big data]{}
Introduction
============
In recent years, advances in the field of electronics, telecommunication, computer and engineering sciences have led to a very strong increase, both in terms of speed and volume, in data being generated. Popular exponents of the present large-scale data era are companies as Facebook Inc. and Google Inc., of which the latter has already been processing more than 20 petabytes of data per day since 2008, but government organizations are also important players [@DEAN:2008; @KALIL:2012].
Due to the declining costs of bandwidth, computing power and storage of data, it is expected that this trend will persist in the future. Inevitably, this necessitates the design of tools to gain insights into these large-scale data sets. Therefore, the design of data mining, statistical and machine learning algorithms to examine large amounts of data and support decision making is of key interest. Commonly used approaches in this research field are clustering, dimensionality reduction, filtering, classification and regression modeling. The focus of this paper is regression modeling, more specifically, semiparametric regression which represents a large class of regression models that allow for nonlinear effects in predictive models [@RUPPERT:2003]. The typical approach to address semiparametric regression modeling is by analyzing the data in one batch. This requires collecting all data before analysis and storing it on one machine. Having all data available at a central location is, however, unrealistic or not feasible for the large-scale setting. For example, Google Inc. designed and implemented a scalable distributed file system to meet its storage needs and, more recently, reported about Spanner, its scalable, multi-version, globally distributed, and synchronously-replicated database, having data centers spread all over the world [@GHEMAWAT:2003; @CORBETT:2012]. Other examples of organizations that have their data distributed over multiple locations are federal departments and agencies, retail companies, but also peer-to-peer networks are part of this scenario. Moreover, in many contexts it can be in the interest of the different organizations (e.g. networks of retailers, hospitals) to combine their individual, potentially non-distributed, data sets to discover new knowledge for improved decision making.
When large data sets are distributed over multiple machines or locations, moving the actual data is usually not a solution due to the associated communication complexity. Apart from the inefficiency, data confidentiality is another important reason to develop feasible alternatives, since different cooperating organizations may not be allowed or willing to share raw data. In this paper, the focus is on so-called horizontally distributed data, meaning that each data host stores different data subjects, but holds the same attributes [@DU:2004]. In the regression context this means that the different hosts store different samples and have all corresponding predictor variables available. The literature on regression modeling for horizontally distributed data sets is largely concerned with multivariate linear regression [@KARR:2005; @KARR:2007; @GHOSH:2012]. An exception to this is the study by @GHOSH:2007 that presented an approach based on multivariate adaptive regression splines to incorporate more flexibility in the model. These approaches typically combine the output from local regression models (cf. ensemble learning) or combine local summary statistics to find the global regression model. Other notable studies are those by @BOYD:2011 and @RAM:2012 in the context of distributed convex optimization and @PREDD:2006.
The focus of this paper, by contrast, is on semiparametric regression modeling for data sets that are horizontally partitioned over multiple hosts and the use of mean field variational Bayes (MFVB) for approximate inference [@WAINWRIGHT:2008; @ORMEROD:2010]. Interpreting semiparametric regression in terms of graphical models offers an elegant and unified way to handle, for example, generalized additive models, geostatistical models, wavelet nonparametric regression models and their various combinations [@WAND:2011]. Moreover, MFVB provides a fast alternative to Markov chain Monte Carlo (MCMC) for fitting these models while it exhibits excellent accuracy for the models that this paper deals with. The methodology also enables handling of grouped data, within-subject correlation, automated regularization parameter inference and various (hierarchical) priors. Importantly, apart from point estimates, measures of uncertainty can be obtained in a straightforward way.
While data sets have typically been processed in batch, nowadays, there is increasing interest in real-time systems that require so-called online stream-processing [@MICHALAK:2012]. Other studies outside of semiparametric regression that developed online methods for horizontally partitioned data are @GUESTRIN:2004, @BHADURI:2008, @POZDNOUKHOV:2011 and @YAN:2013. Interestingly, the former three studies included mechanisms for concept drift in the algorithms for distributed regression. Online MFVB algorithms that make a single pass through the data have recently been developed. @HOFFMAN:2010 and @WANG:2011 proposed MFVB methods for latent Dirichlet allocation and the hierarchical Dirichlet process for topic modeling, respectively. @TCHUMTCHOUA:2011 used online MFVB inference for high-dimensional correlated data and, very recently, @LUTS:2013 proposed real-time semiparametric regression through MFVB approximate inference. However, it appears that these MFVB-based studies only dealt with non-distributed data sets for inference. Therefore, this paper demonstrates batch semiparametric regression for large-scale horizontally distributed data sets and real-time semiparametric regression for processing of horizontally distributed infinite data streams. The proposed algorithms provide exact solutions in the sense that an identical solution is obtained as when all data would have been available at a central location. In addition, this study proposes approaches for temporal adaptation for real-time semiparametric regression of distributed data streams, offering fully-automated regularization for evolving environments. A website is created for real-time demonstration of these methods on live airline data.
Section \[section2\] provides background material on semiparametric regression and MFVB approximate inference. Semiparametric regression for distributed data sets is introduced in Section \[section3\]. Both batch and real-time processing are treated. In Section \[section4\] two approaches are presented to handle the issue of evolving environments for real-time semiparametric regression. Section \[section5\] deals with two application areas of the proposed methodology: semiparametric regression in case of multiple data owners and within the MapReduce framework [@DEAN:2008; @WHITE:2009]. A dynamic website that illustrates the methodology on live airline data is the focus of Section \[section6\]. Closing remarks are made in Section \[section7\].
Variational Bayesian inference for semiparametric regression {#section2}
============================================================
Penalized splines are often used in the semiparametric regression literature to model nonlinear functional relationships [@RUPPERT:2009]. Consider the simple model
$$f(x_i)=\beta_0+\beta_1 x_i+\sum_{k=1}^K u_k z_k(x_i),\quad u_k\ \simind\ N(0,\sigma_u^2), \quad 1 \leq i \leq n,\\
\label{simpleSemipar1}$$
with model parameters $\beta_0$, $\beta_1$, $u_1, \ldots, u_K$ and smoothing parameter $\sigma_u^2$, while $\simind$ denotes distributed independently. The $z_1(\cdot), \ldots, z_K(\cdot)$ represent spline basis functions and in this paper O’Sullivan splines, providing a close approximation to smoothing splines, are used for this purpose [@WAND:2008]. Note that (\[simpleSemipar1\]) can be interpreted as a linear mixed model and leads to the following Bayesian Gaussian response model
$$\begin{array}{c}
y_i|\beta_0,\beta_1,u_1, \ldots , u_K \simind N\left(\beta_0+\beta_1 x_i+\sum_{k=1}^Ku_k z_k(x_i),\sigeps^2\right),\\
\null\\
1 \leq i \leq n,\quad u_k \simind N(0,\sigma_u^2),\quad \beta_0,\beta_1\simind N(0,\sigma_{\beta}^2),\\
\end{array}
\label{simpleSemipar2}$$
for given smoothing parameter $\sigma_u^2$, error variance $\sigeps^2$ and positive hyperparameter $\sigma_{\beta}^2$. By assuming an arbitrary number of predictor variables and introducing uninformative priors for $\sigma_{u1}^2, \ldots, \sigma_{ur}^2$ and $\sigeps^2$ the more general representation of (\[simpleSemipar2\]) becomes
$$\begin{array}{c}
\by|\,\bbeta,\bu,\sigeps^2 \sim
N(\bX\bbeta+\bZ\bu,\sigeps^2\,\bI_n), \quad \bbeta\sim N(\bzero,\sigma_{\beta}^2\bI_p),\\
\null\\
\bu|\,\sigma_{u1}^2, \ldots,\sigma_{ur}^2\sim N(\bzero,\mbox{blockdiag}(\sigma_{u1}^2\,\bI_{K_1},\ldots,\sigma_{ur}^2\,\bI_{K_r})),\\
\null\\
\sigma_{u\ell}\simind\mbox{Half-Cauchy}(A_{u\ell}),\ 1\le\ell\le r,\quad
\sigeps\sim\mbox{Half-Cauchy}(A_{\varepsilon}),\\
\end{array}
\label{lmm}$$
where $\by$ is an $n\times1$ vector of response variables, $A_{\varepsilon}$ and $A_{u\ell}$ are positive hyperparameters, $\bbeta$ is a $p\times1$ vector of fixed effects, $\bu$ is a $(\sum^{r}_{l=1}K_l)\times1$ vector of random effects and $\bX$ and $\bZ$ corresponding design matrices. In this paper, all examples are based on the following values for the hyperparameters: $\sigma_{\beta}^2=10^8$ and $A_{\varepsilon}=A_{u\ell}=10^5$. Note that the variance parameters $\sigma_{u1}^2, \ldots, \sigma_{ur}^2$ correspond to the sub-blocks of $\bu$ having size $K_1,\ldots,K_r$, respectively. In this model the $\text{Half-Cauchy}(A)$ prior is such that the prior density of $\sigma$ is $p(\sigma)\propto\{1+(\sigma/A)^2\}^{-1},\ \sigma>0$. To obtain an equivalent, but more tractable model, the form $\sigma\sim\text{Half-Cauchy}(A)$ is replaced in (\[lmm\]) by the auxiliary variable representation introduced in @WAND:2011b
$$\sigma^2|\,a\sim\text{Inverse-Gamma}\left(1/2,1/a\right),\quad a\sim\text{Inverse-Gamma}\left(1/2,1/A^2\right),\\ \nonumber$$
where $v\sim\mbox{Inverse-Gamma}(A,B)$ if and only if its density function is
$$p(v)=B^A\Gamma(A)^{-1}\,v^{-A-1}\,\exp(-B/v),\quad v>0.\\ \nonumber$$
As @ZHAO:2006 reported, model (\[lmm\]) is quite general and encompasses a large class of models, including simple random effects models, cross random effects models, nested random effects models, generalized additive models, semiparametric mixed models, bivariate smoothing and geoadditive models. For example, in the case of a simple semiparametric model with one predictor and a random intercept, (\[lmm\]) reduces to
$$\begin{array}{c}
y_{ij}|\beta_0,\beta_1,u_1, \ldots , u_K,U_i,\sigeps^2\simind N(\beta_0+\beta_1\,x_{ij}+\sum_{k=1}^Ku_k z_k(x_{ij})+U_i,\sigeps^2),\\
\null\\
1\le i\le m, \quad 1\le j\le n_i, \quad
\beta_0,\beta_1\simind N(0,\sigma_{\bbeta}^2),\quad u_k \simind N(0,\sigma_u^2),\\
\null\\
U_i|\,\sigma_U^2\simind N(0,\sigma_U^2), \quad \sigma_u\sim\mbox{Half-Cauchy}(A_u),\\
\null\\
\sigma_U\sim\mbox{Half-Cauchy}(A_U), \quad \sigeps\sim\mbox{Half-Cauchy}(A_{\varepsilon}),\\
\vspace{-1.6cm}
\end{array}
\vspace{1.6cm}
\label{eq:randInt}$$
where $(x_{ij},y_{ij})$ represents the $j$th predictor/response pair for the $i$th group, with $n_{i}$ denoting the number of subjects in group $i$ and $m$ the total number of groups. An extension of random intercept model (\[eq:randInt\]) is used for the real-life example in Section \[section6\].
MFVB is a class of methods relying on approximate inference of posterior density functions [@WAINWRIGHT:2008; @ORMEROD:2010]. A mean field approximation is founded upon approximating the posterior density function $p(\btheta|\by)$, e.g. parameter vector $\btheta = [\bbeta,\bu,a_{u1},\ldots,$ $a_{ur},a_{\varepsilon},
\sigma_{u1}^2,\ldots,\sigma_{ur}^2,$ $\sigeps^2]^{T}$ for model (\[lmm\]), by a product form $q(\btheta) = \prod_{i=1}^{d} q_i(\btheta_i)$. The choice of the $q_i(\btheta_i)$ density functions is guided by the notion of Kullback-Leibler divergence
$$\int q(\btheta) \log\left\{\frac{q(\btheta)} {p(\btheta|\by)}\right\}\,d\btheta,\\
\label{KL}$$
such that the distance between $\prod_{i=1}^{d} q_i(\btheta_i)$ and $p(\btheta|\by)$ is minimized. It can be shown that an equivalent optimization problem corresponds to maximizing the so-called lower bound on the marginal likelihood $p(\by)$,
$$\punder(\by;q)\equiv\exp \left[ \int q(\btheta) \log\left\{\frac{p(\btheta,\by)} {q(\btheta)}\right\}\,d\btheta\right].\\ \nonumber$$
The optimal $q^*_i(\btheta_i)$ density functions, in terms of minimizing the Kullback-Leibler divergence in (\[KL\]), are known to satisfy
$$q^*_i(\btheta_i) \propto \exp \left[ \mathop{\mathlarger{\mathlarger{\mathlarger{\int}}}} {\left\{ \prod_{j\neq i} q_{j}(\btheta_j) \right\} \log p(\btheta,\by) \, d\btheta_{-i}} \right]\nonumber,\\$$
where $\btheta_{-i} = [\btheta_{1},\ldots,\btheta_{i-1},\btheta_{i+1},\ldots,\btheta_{d}]^T$.
Although MFVB is limited in its approximation accuracy when compared to MCMC, which can be made arbitrarily accurate by increasing the Monte Carlo sample sizes, the latter is much slower than MFVB. Moreover, the accuracy of MFVB for the models that are considered in this paper is typically excellent.
For the mixed model in (\[lmm\]) the mean field approximation and chosen product form
$$\begin{array}{l}
p(\bbeta,\bu,a_{u1},\ldots,a_{ur},a_{\varepsilon},\sigma_{u1}^2,\ldots,\sigma_{ur}^2,\sigeps^2|\by) \\
\null\\
\hspace{3cm} \approx q(\bbeta,\bu,a_{u1},\ldots,a_{ur},a_{\varepsilon},\sigma_{u1}^2,\ldots,\sigma_{ur}^2,\sigeps^2)\\
\null\\
\hspace{3cm} \approx q(\bbeta,\bu,a_{u1},\ldots,a_{ur},a_{\varepsilon}) \,q(\sigma_{u1}^2,\ldots,\sigma_{ur}^2,\sigeps^2),\nonumber \\
\end{array}$$
lead to the following optimal product density functions: $q^*(\bbeta,\bu,a_{u1},\ldots,a_{ur},a_{\varepsilon})$ is the product of the $
N(\bmu_{q(\bbeta,\bu)},\bSigma_{q(\bbeta,\bu)})$ density function, Inverse-Gamma$(1,$ $B_{q(a_{u\ell})})$ density functions, $1\le\ell\le r$, and the Inverse-Gamma$(1,B_{q(a_{\varepsilon})})$ density function, while $q^*(\sigma_{u1}^2,\ldots,\sigma_{ur}^2,\sigeps^2)$ is the product of Inverse-Gamma$(\smhalf(K_{\ell}+1),B_{q(\sigma^2_{u\ell})})$ density functions for $1\le\ell\le r$ and the Inverse-Gamma$(\smhalf(n+1),B_{q(\sigeps^2)})$ density function. Notice that this solution results in so-called induced factorizations. For example, the factorization $q(\bbeta,\bu,a_{u1},\ldots,a_{ur},a_{\varepsilon})=q(\bbeta,\bu)\,q(a_{u1}),\ldots,q(a_{ur})\,q(a_{\varepsilon})$ is not assumed a priori.
Since the optimal parameters in the $q^*$ density functions are interrelated, for example,
$$\bSigma_{q(\bbeta,\bu)}= \left[\mu_{q(1/\sigeps^2)}\,\bC^T\bC+\mbox{blockdiag}\{\sigma_{\beta}^{-2}\,\bI_p,\mu_{q(1/\sigma_{u1}^2)}\bI_{K_1},\ldots, \mu_{q(1/\sigma_{ur}^2)}\bI_{K_r}\} \right]^{-1}, \nonumber\\$$
with $\bC = [\bX \, \bZ]$, the iterative coordinate ascent Algorithm \[algNonDistBatch\] is used to compute the optimal densities where the logarithm of the lower bound equals
$$\begin{aligned}
\log\,\punder(\by;q)&=& \frac{p+\sum_{\ell=1}^{r} K_\ell}{2} - \frac{n}{2} \log(2\pi) - (r+1) \log(\pi) - \frac{p}{2} \log(\sigma_{\beta}^{2}) + \frac{1}{2} \log(|\bSigma_{q(\bbeta,\bu)}|) \\[0.75ex]
&& + \log\left(\Gamma\left(\frac{n+1}{2}\right)\right) - \frac{1}{2\sigma_{\beta}^2} \{ ||\bmu_{q(\bbeta)}||^2 +\mbox{tr}(\bSigma_{q(\bbeta)}) \} - \left(\frac{n+1}{2} \right) \log(B_{q(\sigeps^2)}) \\[0.75ex]
&& + \mu_{q(1/\aeps)} \mu_{q(1/\sigeps^2)} - \log(\Aeps) - \log(B_{q(a_{\varepsilon})}) + \sum_{\ell=1}^{r} \Bigg\{ \log\left(\Gamma\left(\frac{K_\ell+1}{2}\right)\right) - \log(A_{u\ell}) \\[0.75ex]
&& - \log(B_{q(a_{u\ell})}) - \left(\frac{K_\ell+1}{2}\right) \log(B_{q(\sigma^2_{u\ell})}) + \mu_{q(1/a_{u\ell})} \mu_{q(1/\sigma^2_{u\ell})} \Bigg\}.\\ \nonumber\end{aligned}$$
$\bC, \by, n, p, K_{\ell}, \mu_{q(1/\sigeps^2)}, \Aeps, \mu_{q(1/\sigma_{u\ell}^2)}, A_{u\ell}, \sigma_{\beta}^{2} \, \text{with} \, 1\le\ell\le r$ $\bSigma_{q(\bbeta,\bu)}\leftarrow \left[\mu_{q(1/\sigeps^2)}\,\bC^T\bC+ \mbox{blockdiag}\{\sigma_{\beta}^{-2}\,\bI_p, \mu_{q(1/\sigma_{u1}^2)}\bI_{K_1},\ldots, \mu_{q(1/\sigma_{ur}^2)}\bI_{K_r}\} \right]^{-1}$ $\bmu_{q(\bbeta,\bu)}\leftarrow
\mu_{q(1/\sigeps^2)}\,\bSigma_{q(\bbeta,\bu)}\,\bC^T\by;\quad \mu_{q(1/\aeps)}\leftarrow 1/\{\mu_{q(1/\sigeps^2)}+\Aeps^{-2}\}$ $\mu_{q(1/\sigeps^2)}\leftarrow
\displaystyle{\frac{n+1}{2\,\mu_{q(1/\aeps)}
+\by^T\by-2\bmu_{q(\bbeta,\bu)}^T\bC^T\by
+\mbox{tr}[(\bC^T\bC)\{\bSigma_{q(\bbeta,\bu)}+\bmu_{q(\bbeta,\bu)}
\bmu_{q(\bbeta,\bu)}^T\}]}}$ $\mu_{q(1/a_{u\ell})}\leftarrow 1/\{\mu_{q(1/\sigma_{u\ell}^2)}+A_{u\ell}^{-2}\};\quad \mu_{q(1/\sigma_{u\ell}^2)}\leftarrow
\displaystyle{\frac{K_{\ell}+1}{2\,\mu_{q(1/\auell)}
+\Vert\bmu_{q(\bu_{\ell})}\Vert^2
+\mbox{tr}(\bSigma_{q(\bu_{\ell})})}}$
Semiparametric regression for distributed data sets {#section3}
===================================================
Specifying an appropriate regression model might be difficult when data are spread over multiple hosts and there is no opportunity to inspect the combined data set. In these circumstances, semiparametric regression represents a viable alternative to multivariate linear regression, as the latter requires having to choose in advance which polynomial terms to include or transformations to apply. On the other hand, the MFVB approach that was presented in Section \[section2\] includes inference for the smoothing parameters $\sigma^2_{u\ell}, \,1\le\ell\le r$ and, as a consequence, offers fully-automated fitting of flexible relationships between predictors and the dependent variable. The following sections explain how to perform semiparametric regression for distributed data in the batch and the real-time setting.
Batch processing
----------------
Algorithm \[algNonDistBatch\] relies on having the data for all $n$ samples in one location and receives these as input via $\bC$ and $\by$. The crucial piece that allows extending Algorithm \[algNonDistBatch\] towards the distributed setting is how it uses the data: it only depends on the data through the quantities $\bC^T\bC$, $\bC^T\by$, $\by^T\by$ and $n$. Assuming that there are $h$ different locations that host data, i.e. $\bC_g$, $\by_g$ and $n_g$, $1\leq g \leq h$, the following straightforward relationships can be used: $\bC^T\bC = \sum_{g=1}^{h}\bC_g^T\bC_g$, $\bC^T\by = \sum_{g=1}^{h}\bC_g^T\by_g$, $\by^T\by = \sum_{g=1}^{h}\by_g^T\by_g$ and $n = \sum_{g=1}^{h}n_g$.
Algorithm \[algDistBatch\] summarizes the procedure for batch semiparametric regression for distributed data sets. Note that $P=p+\sum_{l=1}^{r}K_l$ denotes the number of columns of $\bC$. Each host performs the computation of the summary statistics locally and Algorithm \[algDistBatch\] only relies on those summaries. Therefore, there is no need to send the actual raw data over the network, thereby saving bandwidth and speeding up the algorithm. This approach is particularly useful for large-scale data sets having large sample sizes and it avoids security risks by data being flooded through the network. In addition, all hosts can generate the local summary statistics simultaneously, but Algorithm \[algDistBatch\] can only start from the moment that all hosts have finished their local computations. The total number of parameters that each host has to send equals $P(P+1)/2+P+2$ since $\bC^T\bC$ is symmetric. Depending on the structure of $\bC$, this number can further be reduced. Section \[section3:b\] illustrates this for the random intercept model (\[eq:randInt\]). Note that Algorithm \[algDistBatch\] assumes the existence of another party, called *combiner* in this paper, which receives the local summary statistics from the data hosts and manages the global semiparametric regression. However, as Section \[section5:a\] points out, the existence of a separate party is in fact not a requirement.
A potential issue with the proposed method is that the spline basis functions have to be set without having the combined data set available. For example, a set of knot positions may need to be specified. For many applications it is simple to specify the range of possible values beforehand. For example, for a predictor variable corresponding to outside temperature in degrees Celsius or wind speed in knots, equidistantly positioning knots within a reasonable range represents an effective approach. Dealing with grouped data, e.g. within the context of random intercept model (\[eq:randInt\]), involves similar issues. Algorithm \[algDistBatch\] requires specifying the number and kind of groups a priori. Again, for many applications this is not a problem. For example, when flights are grouped per airline, the total number of possible airlines can be determined beforehand. If these assumptions are not reasonable, some adjustments have to be made.
$p, P, K_{\ell}, \mu_{q(1/\sigeps^2)}, \Aeps, \mu_{q(1/\sigma_{u\ell}^2)}, A_{u\ell}, \sigma_{\beta}^{2} \, \text{with} \, 1\le\ell\le r$ $\bC^T\bC \leftarrow \bzero_{P \times P}; \quad \bC^T\by \leftarrow \bzero_{P \times 1}; \quad \by^T\by \leftarrow 0; \quad n \leftarrow 0$ $\text{retrieve} \, \, \bC_g^T\bC_g, \, \, \bC_g^T\by_g, \, \, \by_g^T\by_g \, \, \text{and} \, \, n_g \, \, \text{from host} \, g$ $\bC^T\bC \leftarrow \bC^T\bC + \bC_g^T\bC_g; \quad \bC^T\by \leftarrow \bC^T\by + \bC_g^T\by_g ; \quad \by^T\by \leftarrow \by^T\by + \by_g^T\by_g; \quad n \leftarrow n + n_g$ $\bSigma_{q(\bbeta,\bu)}\leftarrow \left[\mu_{q(1/\sigeps^2)}\,\bC^T\bC+ \mbox{blockdiag}\{\sigma_{\beta}^{-2}\,\bI_p, \mu_{q(1/\sigma_{u1}^2)}\bI_{K_1},\ldots, \mu_{q(1/\sigma_{ur}^2)}\bI_{K_r}\} \right]^{-1}$ $\bmu_{q(\bbeta,\bu)}\leftarrow
\mu_{q(1/\sigeps^2)}\,\bSigma_{q(\bbeta,\bu)}\,\bC^T\by;\quad \mu_{q(1/\aeps)}\leftarrow 1/\{\mu_{q(1/\sigeps^2)}+\Aeps^{-2}\}$ $\mu_{q(1/\sigeps^2)}\leftarrow
\displaystyle{\frac{n+1}{2\,\mu_{q(1/\aeps)}
+\by^T\by-2\bmu_{q(\bbeta,\bu)}^T\bC^T\by
+\mbox{tr}[(\bC^T\bC)\{\bSigma_{q(\bbeta,\bu)}+\bmu_{q(\bbeta,\bu)}
\bmu_{q(\bbeta,\bu)}^T\}]}}$ $\mu_{q(1/a_{u\ell})}\leftarrow 1/\{\mu_{q(1/\sigma_{u\ell}^2)}+A_{u\ell}^{-2}\};\quad \mu_{q(1/\sigma_{u\ell}^2)}\leftarrow
\displaystyle{\frac{K_{\ell}+1}{2\,\mu_{q(1/\auell)}
+\Vert\bmu_{q(\bu_{\ell})}\Vert^2
+\mbox{tr}(\bSigma_{q(\bu_{\ell})})}}$
### Illustration for Sydney property rental data {#section3:a2}
This section illustrates distributed batch semiparametric regression by analyzing data from the residential property rental market in Sydney, Australia. With more than a thousand real estate offices, the Sydney real estate market is a highly competitive one. All together, 1447 real estate offices hosted data belonging to 150471 properties during the period 9th May, 2012 and 25th May, 2013. In this example, these data are processed by Algorithm \[algDistBatch\] using the model $$\begin{array}{l}
\log(\mbox{\texttt{weekly\,\,rent}}_{i})|\,\bbeta,
\bu_2,\bu_3,\bu_4,\bu_5,
\sigeps^2\simind\\[1ex]
\qquad
N(\beta_0 + \beta_1\,\mbox{\texttt{house}}_{i} +
f_2(\mbox{\texttt{number\,\,of\,\,bedrooms}}_{i})\\[1ex]
\qquad + f_3(\mbox{\texttt{number\,\,of\,\,bathrooms}}_{i} )
+f_4(\mbox{\texttt{number\,\,of\,\,car\,\,spaces}}_{i} )\\[1ex]
\qquad+ f_5(\mbox{\texttt{longitude}}_{i},\mbox{\texttt{latitude}}_{i}),\sigeps^2),
\end{array}
\label{eq:SydneyRealEstate}$$ where $\mbox{\texttt{weekly rent}}_{i}$ is the weekly rental amount in Australian dollars of the $i$th property, $\mbox{\texttt{house}}_{i}$ is an indicator of the $i$th property being a house, townhouse or villa versus an apartment, and $\mbox{\texttt{number\,\,of\,\,bedrooms}}_{i}$ is the number of bedrooms in the $i$th property. Variables concerning the number of bathrooms and car spaces are defined in a similar way. The geographical location of the $i$th property is included by the variables $\texttt{longitude}_{i}$ and $\texttt{latitude}_{i}$. To execute Algorithm \[algDistBatch\], the fixed effect regression coefficients $\beta_0$, $\beta_1$ and the linear contributions to $f_2,\ldots,f_5$ are stored in $\bbeta$, while the spline basis coefficients for $f_2,\ldots,f_5$ are stored in $\bu_2,\ldots,\bu_5$. The estimate of $f_5$ is based on bivariate thin plate splines [@RUPPERT:2003].
Figure \[fig:batch1\] shows various regression summaries resulting from fitting of (\[eq:SydneyRealEstate\]) using Algorithm \[algNonDistBatch\] on data from real estate agency `McGrath Leichhardt` only, corresponding to 436 properties, and the combined result from Algorithm \[algDistBatch\] based on data hosted by 1447 real estate offices (i.e. 150471 properties). As expected, the estimates based on combining information from multiple hosts are more reliable and the figure shows that more narrow 95% credible sets are obtained. The approximate posterior density function for $\beta_1$ shows that the average rental amount for houses is 9.5% higher than for apartments after correcting for all other covariates. The remaining panels show the increase in rental amount when the property includes more bedrooms, bathrooms or car spaces. Finally, a color-coded geographical map of Sydney, based on the data from 1447 hosts, displays the weekly rent for a two bedroom apartment with one bathroom and one car space for various geographical locations (Figure \[fig:batch2\]).
\
**** [![*Approximate posterior density functions, regression fits and corresponding 95% credible sets for the Sydney property rental data example in Section \[section3:a2\]. The first column displays the impact of the property being a house or not, while the other three columns visualize the effects of the number of bedrooms, bathrooms and car spaces on the weekly rent for apartments. The top row results are based on data from the real estate agency `McGrath Leichhardt` only, whereas the bottom row displays results based on data hosted by 1447 real estate offices.*[]{data-label="fig:batch1"}](result3All.eps "fig:"){width="\textwidth"}]{}
Real-time processing {#section3:b}
--------------------
The implementation of online regression methods in real-time systems supports incremental calculations of regression results when new data arrive [@LUTS:2013]. This section focuses on the more complicated setting where $h$ different hosts simultaneously receive different streams of data, independent from each other. While the hosts can individually monitor their data streams and perform semiparametric regression in real-time, the real-time combined regression result based on all $h$ distributed data streams is of primary interest in this paper. Algorithm \[algDistOnline\] summarizes how the combiner manages the global semiparametric regression in an incremental manner for such a setting.
First, the individual data hosts process their data streams locally in real-time. This includes repeatedly extracting the dependent variable and the predictor variables from the stream, but may involve additional preprocessing. Online semiparametric regression at each host is performed according to Algorithm 3 of @LUTS:2013. In addition to this, each host stores the summary statistics for its incoming data $\bc_{\text{new}}\,\bc_\text{new}^T$, $\bc_\text{new}\, y_{\text{new}}$ and $y^2_{\text{new}}$ in a local buffer. Once the local buffer of a host exceeds a threshold size, the sum of the buffer’s local summary statistics, i.e. $\bC_b^T\bC_b$, $\bC_b^T\by_b$, $\by_b^T\by_b$ and $n_b$, is sent to the combiner and the buffer is emptied. A buffer at the site of the data host enables it to regulate data traffic and, depending on its size and the rate at which data are coming in, the combiner will receive the local summary statistics with a certain delay. The combiner has its own buffer where the summaries $\bC_b^T\bC_b$, $\bC_b^T\by_b$, $\by_b^T\by_b$ and $n_b$ from the different hosts are stored. The use of a buffer at the combiner site allows a difference between the rate at which summary statistics are received from the data hosts and the rate at which they can be processed by the combiner via Algorithm \[algDistOnline\]. This setup enables the data hosts to operate independently and asynchronously from each other. In addition, they can simultaneously process the different streams, speeding up the computation of the combined semiparametric regression result.
$p, P, K_{\ell}, \mu_{q(1/\sigeps^2)}, \Aeps, \mu_{q(1/\sigma_{u\ell}^2)}, A_{u\ell}, \sigma_{\beta}^{2} \, \text{with} \, 1\le\ell\le r$ $\bC^T\bC \leftarrow \bzero_{P \times P}; \quad \bC^T\by \leftarrow \bzero_{P \times 1}; \quad \by^T\by \leftarrow 0; \quad n \leftarrow 0$ $B \leftarrow \text{number of items in buffer}$ $\text{read items from buffer and compute} \, \, \sum_{b=1}^{B}\bC_b^T\bC_b, \, \, \sum_{b=1}^{B} \bC_b^T\by_b, \, \, \sum_{b=1}^{B} \by_b^T\by_b \, \, \text{and} \, \, \sum_{b=1}^{B} n_b$ $\bC^T\bC \leftarrow \bC^T\bC + \sum_{b=1}^{B}\bC_b^T\bC_b; \quad \bC^T\by \leftarrow \bC^T\by + \sum_{b=1}^{B} \bC_b^T\by_b$ $\by^T\by \leftarrow \by^T\by + \sum_{b=1}^{B} \by_b^T\by_b; \quad n \leftarrow n + \sum_{b=1}^{B} n_b$ $\bSigma_{q(\bbeta,\bu)}\leftarrow \left[\mu_{q(1/\sigeps^2)}\,\bC^T\bC+ \mbox{blockdiag}\{\sigma_{\beta}^{-2}\,\bI_p, \mu_{q(1/\sigma_{u1}^2)}\bI_{K_1},\ldots, \mu_{q(1/\sigma_{ur}^2)}\bI_{K_r}\} \right]^{-1}$ $\bmu_{q(\bbeta,\bu)}\leftarrow
\mu_{q(1/\sigeps^2)}\,\bSigma_{q(\bbeta,\bu)}\,\bC^T\by;\quad \mu_{q(1/\aeps)}\leftarrow 1/\{\mu_{q(1/\sigeps^2)}+\Aeps^{-2}\}$ $\mu_{q(1/\sigeps^2)}\leftarrow
\displaystyle{\frac{n+1}{2\,\mu_{q(1/\aeps)}
+\by^T\by-2\bmu_{q(\bbeta,\bu)}^T\bC^T\by
+\mbox{tr}[(\bC^T\bC)\{\bSigma_{q(\bbeta,\bu)}+\bmu_{q(\bbeta,\bu)}
\bmu_{q(\bbeta,\bu)}^T\}]}}$ $\mu_{q(1/a_{u\ell})}\leftarrow 1/\{\mu_{q(1/\sigma_{u\ell}^2)}+A_{u\ell}^{-2}\};\quad \mu_{q(1/\sigma_{u\ell}^2)}\leftarrow
\displaystyle{\frac{K_{\ell}+1}{2\,\mu_{q(1/\auell)}
+\Vert\bmu_{q(\bu_{\ell})}\Vert^2
+\mbox{tr}(\bSigma_{q(\bu_{\ell})})}}$
Note that Algorithm \[algDistOnline\] initializes the summary statistics in line 1 to zero and requires starting values for $\mu_{q(1/\sigeps^2)}$ and $\mu_{q(1/\sigma_{u\ell}^2)}$. Section 2.1.1 of @LUTS:2013 explains that good initialization by means of a so-called warm-up step can be important for convergence of the real-time semiparametric regression approach. For clarity of presentation, this warm-up step was not included in Algorithm \[algDistOnline\]. Although experiments have shown that warming-up is in the first place important for wavelet regression and logistic regression (cf. @LUTS:2013), it can also easily be incorporated in Algorithm \[algDistOnline\]. All it requires is running batch Algorithm \[algNonDistBatch\] on a subset of data and using the summary statistics and obtained estimates as starting values for Algorithm \[algDistOnline\].
Closer inspection of Algorithm \[algDistOnline\] reveals that line 6 involves inverting a matrix of size $P \times P$, with $P = p+\sum_{\ell=1}^{r}K_\ell$. As also noted by @SMITH:2008 in the context of frequentist inference for additive mixed models, na[ï]{}ve implementation of line 6 can be extremely inefficient for grouped data as in (\[eq:randInt\]). Moreover, since Algorithm \[algDistOnline\] aims to run in an online fashion on large-scale data with potentially many groups, and as a consequence large $P$, it is important to optimize this line of code. @SMITH:2008 outline a procedure for which the variance calculations are linear in the number of groups, but omit the computation of correlations between any two groups. Algorithm \[algDistOnline\], however, does require calculating these inter-group correlations since the full matrix $\bSigma_{q(\bbeta,\bu)}$ is needed to compute $\bmu_{q(\bbeta,\bu)}$, for example. The following paragraphs explain how line 6 can be solved in more efficient way for grouped data as for example the live example in Section \[section6\].
Assume that $\bC = [\bX \, \bZ_1 \, \bZ_2]$, where the original design matrix $\bZ$ is divided into a design matrix that is only related to the $K_r$ random intercepts, i.e. $\bZ_2$, and a design matrix for all the rest, i.e. $\bZ_1$, including spline basis functions. This enables the block decomposition
$$\renewcommand{\arraystretch}{1.7}
\begin{array}{rl}
\bM \ \equiv \ \bSigma_{q(\bbeta,\bu)}^{-1} & = \ \mu_{q(1/\sigeps^2)}
\left[ \begin{array}{cc|c}
\bX^T \bX + \mu_{q(1/\sigeps^2)}^{-1} \bG_1 & \bX^T \bZ_1 & \bX^T \bZ_2 \\
\bZ^T_1 \bX & \bZ_1^T \bZ_1 + \mu_{q(1/\sigeps^2)}^{-1} \bG_2 & \bZ_1^T \bZ_2 \\ \hline
\bZ_2^T \bX & \bZ_2^T \bZ_1 & \bZ_2^T \bZ_2 + \mu_{q(1/\sigeps^2)}^{-1} \bG_3
\end{array} \right]\\
\null\\
& = \ \mu_{q(1/\sigeps^2)}
\renewcommand{\arraystretch}{1}
\left[ \begin{array}{cc}
\bM_{11} & \bM_{12}\\[0.3em]
\bM_{21} & \bM_{22}\\[0.3em]
\end{array} \right], \nonumber
\end{array}$$
where $\bG_1 = \sigma_{\beta}^{-2}\,\bI_p$, $\bG_2 = \mbox{blockdiag}\{\mu_{q(1/\sigma_{u1}^2)}\bI_{K_1},\ldots, \mu_{q(1/\sigma_{ur-1}^2)}\bI_{K_{r-1}}\}$ and $\bG_3 = \mu_{q(1/\sigma_{ur}^2)}$ $\bI_{K_r}$. The rules for computing the inverse of a block-partitioned matrix give
$$\bM^{-1} \ \equiv \ \bSigma_{q(\bbeta,\bu)} = \mu_{q(1/\sigeps^2)}^{-1} \left[ \begin{array}{cc}
\bM^{11} & \bM^{12}\\[0.3em]
\bM^{21} & \bM^{22}\\[0.3em]
\end{array} \right],
\label{inverse}$$
with $\bM^{11}=(\bM_{11}-\bM_{12}\bM_{22}^{-1}\bM_{21})^{-1}$, $\bM^{12}=-\bM^{11}\bM_{12}\bM_{22}^{-1}$, $\bM^{21}=(\bM^{12})^T$ and $\bM^{22}=\bM_{22}^{-1}+\bM_{22}^{-1}\bM_{21}\bM^{11}\bM_{12}\bM_{22}^{-1}$ [@HARVILLE:2000]. Dealing with a large number of groups results in the relationship $K_r \gg p + \sum_{\ell=1}^{r-1} K_l$. In these circumstances, the straightforward matrix multiplications $\bX^T \bX$, $\bX^T \bZ_1$ and $\bZ_1^T \bZ_1 $ are relatively inexpensive. As also explained in @SMITH:2008, $\bZ_2$ has a special structure because of the random intercept design, thereby making the computation of $\bX^T \bZ_2$ and $\bZ_1^T \bZ_2$ efficient. The biggest inverse that is needed in (\[inverse\]) is $\bM_{22}^{-1}$, but since $\bM_{22}$ is diagonal it can be obtained in $K_r$ steps. The final step to obtain $ \bSigma_{q(\bbeta,\bu)}^{-1} $ is computing $\bM^{22}$. Whereas @SMITH:2008 only compute the diagonal elements of this matrix, Algorithm \[algDistOnline\] requires all unique entries of this symmetric matrix. Denoting $\bM_{12}=[\bh_1 , \ldots , \bh_{K_r}]$, the elements of $\bM^{22}$ are
$$\begin{array}{rl}
\bM^{22}_{ii} & = \frac{\mu_{q(1/\sigeps^2)}}{n_i \mu_{q(1/\sigeps^2)}+\mu_{q(1/\sigma_{ur}^2)}} \left(1+ \frac{\mu_{q(1/\sigeps^2)} \, \bh_i^T\bM^{11}\bh_i}{n_i \mu_{q(1/\sigeps^2)}+\mu_{q(1/\sigma_{ur}^2)}} \right),\\
\null\\
\bM^{22}_{ij} & = \frac{\mu^2_{q(1/\sigeps^2)} \, \bh_i^T\bM^{11}\bh_j}{\left(n_i \mu_{q(1/\sigeps^2)}+\mu_{q(1/\sigma_{ur}^2)}\right)\left(n_j \mu_{q(1/\sigeps^2)}+\mu_{q(1/\sigma_{ur}^2)}\right)},
\quad i\neq j,\\ \nonumber
\end{array}$$
with $n_i$ the number of subjects in group $i$. Observe that the $K_r(K_r + 1)/2$ unique entries of $\bM^{22}$ can be computed in parallel. In addition, grouped data sets enable a further, significant reduction in unique entries to be transferred from host to combiner as $\bZ^T_2 \bZ_2$ is diagonal.
### Illustration for simulated data {#section3:b2}
Consider the following model for a synthetic data example to illustrate Algorithm \[algDistOnline\],
$$y_i|\bbeta,\bu_4,\bu_5,\bu_6,\sigeps^2 \simind N\Big(\beta_1\,x_{1i}+\beta_2\,x_{2i}+\beta_3\,x_{3i} +f_4(x_{4i})+f_5(x_{5i})+f_6(x_{6i}),\sigeps^2\Big), \nonumber \\$$
where $\bu_\ell$ is the vector of spline coefficients for $f_\ell(\cdot)$ with $\ell=4,5$ and $6$. The number of hosts is fixed at $h=9$ and each of these hosts processes 1000 samples, generated according to the model above with $x_{1i},x_{2i},x_{3i}\simind\mbox{Bernoulli}\,(1/2)$ and $x_{4i},x_{5i},x_{6i}\simind N(0,1)$. The true values were set at $\beta_1=0.2$, $\beta_2=-0.3$, $\beta_3=0.6$, $f_4(x)=2\Phi(6x-3)$, $f_5(x)=\sin(3\pi x^3)$, $f_6(x)=\cos(4\pi x)$ and $\sigeps^2=1$. Each host individually processes its incoming data in an online manner and, when its local buffer contains summary statistics from 10 samples, it sends the corresponding sums to the combiner. Assuming that all hosts simultaneously process their data at the same rate, the combiner receives summary statistics from 90 samples at each time instance. Figure \[fig:realTimeDistributed\] visualizes the approximate posterior density functions for the regression coefficients and the regression fits at the site of the combiner and host 1. The approximate posterior density functions are first, i.e. time = 1, flat at host 1 and the combiner and regression fits show noisy relationships. As time progresses, i.e. time = 20 and 100, these regression summaries start to approximate the true underlying values and relationships. The link `Real-time Gaussian additive model for distributed data` on the website `realtime-semiparametric-regression.net` points to a movie showing summaries of the regression fits when the streaming data are simultaneously processed at 9 hosts and the combiner. Convergence to the true values and nonlinear relationships is faster at the combiner than at an individual host, illustrating the power of a real-time distributed semiparametric regression analysis. The difference in rate of convergence is dependent on the number of hosts $h$. Note that a warm-up sample of size 100 was used at the combiner and all 9 hosts, providing starting values for $\mu_{q(1/\sigeps^2)}$, $\mu_{q(1/\sigma_{u\ell}^2)}$, $\bC^T\bC$, $\bC^T\by$, $\by^T\by$ and $n$. This explains the sample sizes $110$ and $190$ for host 1 and the combiner at time = 1, respectively.
Evolving environments {#section4}
=====================
The previous section outlined an algorithm for continuous learning for distributed data sets based on the assumption that the underlying true model (e.g. $\bbeta$, $\bu$ and $\sigeps^2$) does not change over time. In this section, two approaches are proposed to relax this assumption since handling evolving environments is almost inherently connected with real-time streaming data analysis: the characteristics of the new incoming data can change over time in a data stream. The first approach relies on the definition of a time window, while the second is based on reweighting older data.
Adaptation through a time window {#section4a}
--------------------------------
When data arrive in a stream, newer samples are often more relevant for the present situation than older samples. For example, housing market data from the last month might be more informative than data from 24 months ago if one aims to create a predictive model for the near future. However, in some situations data from the same month, season (or quarter in economics) from the previous year might be more relevant than the previous month or season of the current year. In both situations it is often possible to define an appropriate time period of interest, such that only samples from within that specific time frame contribute to the regression fit. In this paper the time period of interest is called the time window and, as time evolves, the time window is shifted so that older samples leave the window and new samples enter the window. Although this section assumes a fixed window width, the methodology can be generalized to a time-variable window width.
Extending Algorithm \[algDistOnline\] towards evolving environments using a time window simply requires modifying lines 4–5 to
$$\bC^T\bC \leftarrow \bC^T\bC + \sum_{b=1}^{B}\bC_b^T\bC_b - \bC_{\text{old}}^T\bC_{\text{old}}; \quad \bC^T\by \leftarrow \bC^T\by + \sum_{b=1}^{B} \bC_b^T\by_b - \bC_{\text{old}}^T\by_{\text{old}}
\label{window1}$$
and
$$\by^T\by \leftarrow \by^T\by + \sum_{b=1}^{B} \by_b^T\by_b - \by_{\text{old}}^T\by_{\text{old}}; \quad n \leftarrow n + \sum_{b=1}^{B} n_{\text{old}} - n_{\text{old}}
\label{window2}$$
where $\bC_{\text{old}}$, $\by_{\text{old}}$ and $n_{\text{old}}$ correspond to the data and number of samples that leave the time window at a certain point in time, respectively. Note that this extension, in contrast to the original Algorithm \[algDistOnline\] in Section \[section3:b\], assumes that new data are temporally stored such that their contribution to the summary statistics can later be removed.
For illustrative purposes, the first synthetic data example in this section considers the simple linear regression model,
$$y_i|\beta_0,\beta_1,\sigeps^2 \simind N\Big(\beta_0+\beta_1\,x_{i},\sigeps^2\Big)\quad \text{with} \quad x_i \simind \text{Uniform}\left(0,1\right). \\
\label{generatedFromLinearRegression}$$
The true values for the different parameters are gradually decreased over time: $\beta_0 \in \left\{4,3.665,3.33\right\}$, $\beta_1 \in \left\{3,2.72,2.44\right\}$ and $\sigeps^2 \in \left\{0.350,0.325,0.300\right\}$. For the first, second and third combination of parameters 300, 500 and 400 $(x_i,y_i)$-pairs were generated, respectively. Figure \[fig:forgetWindowLin\] visualizes the result from applying a simplified version (i.e. the linear regression model above being a simplification of model (\[lmm\])) of Algorithm \[algDistOnline\] with the extension in (\[window1\])–(\[window2\]) to this data set. A fixed window size of 100 samples is used and the $i$th time instance on the horizontal axis represents the moment when the $i$th and $(i-100)$th sample enters and leaves the time window, respectively. For example, the estimated value for $\beta_0$ at time = 200 is exclusively based on samples 101 to 200. The online algorithm is compared with batch Algorithm \[algDistBatch\] which is used on all samples in the current time window. Each time an old (new) sample leaves (enters) the time window the full batch analysis needs to be repeated entirely. The results show that the mean estimates and 95% credible sets from the online algorithm and the batch algorithm coincide. In addition, the underlying truth is contained in the 95% credible sets for all parameters. In case the time window starts to contain samples being generated from different $\bbeta$ or $\sigeps^2$ values (i.e. when the red horizontal lines overlap in time), the estimates tend to enter a transition phase between the true, underlying values after which stable estimates are again obtained.
The next example fits a semiparametric regression model on synthetic data that were generated according to
$$y_i|\alpha_0,\alpha_1,\alpha_2,\sigeps^2 \simind N\Big(\alpha_0+\alpha_1\,\sin(6\pi x_{i}+\alpha_2),\sigeps^2\Big) \quad \text{and} \quad x_i \simind \text{Uniform}\left(0,1\right), \\
\label{generatedFromMixedRegression}$$
with $\alpha_0=4$, $\alpha_1 \in \left\{0.5,\ldots,3\right\}$, $\alpha_2 \in \left\{0,\ldots,5 \right\}$ and $\sigeps^2 \in \left\{0.1,\ldots,0.4 \right\}$. The values for $\alpha_1$, $\alpha_2$ and $\sigeps^2$ gradually evolve in 10 equally spaced steps between these boundaries and for each combination 600 $(x_i,y_i)$-pairs were generated. Figure \[fig:forgetWindowMixed\] visualizes 95% credible sets for the mean at 6 different time points for this data set. Samples within the time window of size 400 are denoted by black dots while older (i.e. outside the time window) data are indicated by small grey dots. Comparing the 95% credible sets with the true, underlying model (i.e. red curve) at each time point shows that the estimates capture the evolving true nonlinear relationship.
Forgetting by reweighting data
------------------------------
The methodology from Section \[section4a\] represents an appropriate approach when one is interested in real-time modeling using data from a predefined period of time. For example, one aims to continuously update a regression model such that only (distributed) data from the last 30 days contribute to the fit. However, a potential issue with the method in the previous section is that the summary statistics from data need to be stored until those data fall out of the time window. This section presents a different approach for incorporating a mechanism that enables real-time semiparametric regression for evolving environments without having to store the data or corresponding individual summary statistics.
Closer inspection of Algorithm \[algDistOnline\] for online MFVB semiparametric regression of distributed data reveals that all summary statistics contribute with equal weight to the total sums (cf. line 4 and 5). In order to forget older information and focus on more recent data Algorithm \[algDistOnlineReweighting\] uses a decaying window through the introduction of reweighting for the summary statistics. For example, reweighting for the summary statistic $\by^T\by$ at time $t$ can be imposed via the assignment
$$\by^T\by \leftarrow (1-\rho_t)\by^T\by +\rho_t \sum_{b=1}^{B} \by_b^T\by_b,$$
where $\rho_t$ denotes the learning rate at time $t$. Various ways exist to define the learning rate: it can be kept constant over time or an adaptive approach can be used. An example of a decreasing learning rate is $\rho_t=(\tau+t)^{-\kappa}$, with fixed parameters $\tau>0$ and $\kappa \in \{k \in \mathbb{R}\,|\,0.5<k\leq 1\}$. With this decreasing learning rate, larger values for $\tau$ and $\kappa$ result in less forgetting of older samples. In addition, the level of forgetting is decreased as time evolves (i.e. for increasing $t$). On the other hand, a constant learning rate can be used to impose a constant level of forgetting over time. Similarly to the time window approach from the previous section, a constant learning rate enables us to specify that only a fixed number of most recent samples contribute to the sum of summary statistics. Whereas the samples in the time window have an equal contribution to the sum, the constant learning rate approach implies an additional weighting such that the most recent samples have higher weights. In this way the assumption that samples need to be stored can be omitted. Note that Algorithm \[algDistOnlineReweighting\] incorporates a decreasing learning rate $\rho_t$. Using a constant learning rate simply requires to fix $\rho_t = \rho$ beforehand, where $0<\rho<1$.
$p, P, K_{\ell}, \mu_{q(1/\sigeps^2)}, \Aeps, \mu_{q(1/\sigma_{u\ell}^2)}, A_{u\ell}, \sigma_{\beta}^{2}, \tau, \kappa \, \, \text{with} \, 1\le\ell\le r$ $\bC^T\bC \leftarrow \bzero_{P \times P}; \quad \bC^T\by \leftarrow \bzero_{P \times 1}; \quad \by^T\by \leftarrow 0; \quad n \leftarrow 0; \quad t \leftarrow 0$ $\text{retrieve and remove} \, \, \sum_{b=1}^{B}\bC_b^T\bC_b, \, \, \sum_{b=1}^{B} \bC_b^T\by_b, \, \, \sum_{b=1}^{B} \by_b^T\by_b \, \, \text{and} \, \, \sum_{b=1}^{B} n_b \, \, \text{from buffer}$ $t \leftarrow t + 1; \quad \rho_t \leftarrow (\tau+t)^{-\kappa}$ $\bC^T\bC \leftarrow (1-\rho_t) \bC^T\bC + \rho_t \sum_{b=1}^{B}\bC_b^T\bC_b; \quad \bC^T\by \leftarrow (1-\rho_t) \bC^T\by + \rho_t \sum_{b=1}^{B} \bC_b^T\by_b$ $\by^T\by \leftarrow (1-\rho_t)\by^T\by +\rho_t \sum_{b=1}^{B} \by_b^T\by_b; \quad n \leftarrow n + \sum_{b=1}^{B} n_b; \quad \gamma \leftarrow n/\{\sum_{b=1}^{B} n_b\}$ $\bSigma_{q(\bbeta,\bu)}\leftarrow \left[\mu_{q(1/\sigeps^2)}\,\gamma\,\bC^T\bC+ \mbox{blockdiag}\{\sigma_{\beta}^{-2}\,\bI_p, \mu_{q(1/\sigma_{u1}^2)}\bI_{K_1},\ldots, \mu_{q(1/\sigma_{ur}^2)}\bI_{K_r}\} \right]^{-1}$ $\bmu_{q(\bbeta,\bu)}\leftarrow
\mu_{q(1/\sigeps^2)}\,\bSigma_{q(\bbeta,\bu)}\,\gamma\,\bC^T\by;\quad \mu_{q(1/\aeps)}\leftarrow 1/\{\mu_{q(1/\sigeps^2)}+\Aeps^{-2}\}$ $\mu_{q(1/\sigeps^2)}\leftarrow
\displaystyle{\frac{n+1}{2\,\mu_{q(1/\aeps)}
+\gamma(\by^T\by-2\bmu_{q(\bbeta,\bu)}^T\bC^T\by
+\mbox{tr}[(\bC^T\bC)\{\bSigma_{q(\bbeta,\bu)}+\bmu_{q(\bbeta,\bu)}
\bmu_{q(\bbeta,\bu)}^T\}])}}$ $\mu_{q(1/a_{u\ell})}\leftarrow 1/\{\mu_{q(1/\sigma_{u\ell}^2)}+A_{u\ell}^{-2}\};\quad \mu_{q(1/\sigma_{u\ell}^2)}\leftarrow
\displaystyle{\frac{K_{\ell}+1}{2\,\mu_{q(1/\auell)}
+\Vert\bmu_{q(\bu_{\ell})}\Vert^2
+\mbox{tr}(\bSigma_{q(\bu_{\ell})})}}$
To illustrate Algorithm \[algDistOnlineReweighting\], data were generated according to
$$\begin{array}{c}
y_i|\beta_0,\beta_1,u_1, \ldots , u_{24} \simind N\left(\beta_0+\beta_1 x_i+\sum_{k=1}^{24} u_k z_k(x_i),0.25\right),\quad x_i \simind \text{Uniform}\left(0,1\right), \\
\vspace{-0.4cm}
\end{array}
\label{generatedFromMixedRegressionWeighting}
\vspace{0.4cm}$$
where the true values for $\beta_0,\beta_1,u_1, \ldots , u_{24}$ were gradually modified using linear interpolation. The number of data hosts, i.e. $B$, was fixed at 10 and $n_b=1$ was kept constant. At each time instance 10 samples were processed and the total number of time instances equaled 100000. The true values for the model parameters were modified each 12500 time instances. Figure \[fig:forgetReweightingMixed\] displays the evolution of the true relationship between the independent and response variable as a red curve at six time points. The 100 most recent samples are plotted as black dots while older data are visualized as small grey dots. The thin blue curve shows the estimate of the mean by using Algorithm \[algDistOnlineReweighting\] with a fixed learning rate $\rho_t=0.001$. Figure \[fig:forgetReweightingMixed\] shows that the estimated mean adapts itself to the time-evolving data.
Application areas {#section5}
=================
The algorithms presented in the previous sections have assumed that there exists one combiner that merges all the contributions from the individual data hosts to obtain the global regression result. This setting is potentially useful for a wide range of applications. For example, a large retail company with several local stores wishes to analyze consumer behavior and local store performance in real-time. Each local store collects data about individual consumer purchases every few seconds and corresponding local summary statistics are continuously combined in the analytics department of the company. Algorithm \[algDistOnlineReweighting\] allows us to handle such grouped (e.g. the local store and/or the individual consumer) data: patterns of behavior over time can be interpreted and the performance of each individual store can be monitored in real-time for presentation to management.
This section further explains other possible scenarios for the use of the proposed algorithms. The first example deals with the situation where there exist multiple data owners that want to do cooperative semi-parametric regression, but without disclosure of their data or summary statistics. The second example addresses the use of the algorithms within the context of the MapReduce programming model for distributed computing.
Multiple data owners: cooperative analysis {#section5:a}
------------------------------------------
When mutually untrusted parties or competitors jointly aim to conduct the proposed semi-parametric regression analysis, privacy becomes an important issue. Even though the various parties are likely to benefit from a cooperative analysis, their highest priority might still be protecting the confidentiality of their own data. For example, the parties might not be willing to share individual records, nor to reveal the origin of the data. Assuming the existence of a trusted third party that performs the analysis is not always realistic, secure multiparty computation has a role to play [@DU:2001].
The algorithms that were presented in the previous sections do not require sharing individual data records (i.e. samples), but are based on sharing summary statistics. Secure multiparty computation in such a context requires a method for secure summation of these summary statistics. The outcome of such a secure summation is that the different parties, or data hosts as described in Algorithm \[algDistBatch\], obtain the combined results, i.e. $\bC^T\bC, \bC^T\by, \by^T\by$ and $n$, but gain no information about the individual summary statistics of the other parties. This includes both protecting the summary statistics and their origin. A simple secure summation protocol to compute $n$ for $B>2$ parties is as follows:
- Party 1 generates a large random integer $n_{\text{random}}$ and sends $n_{\text{random}}+n_1$ to Party 2.
- Party 2 adds $n_2$ to the input it received from Party 1 and sends the result to Party 3, etc.
- Party 1 subtracts $n_{\text{random}}$ from the number it received from Party $B$ and shares the result with all other parties.
An identical protocol can be followed to compute $\by^T\by$ and, similarly, it can be used to compute the (unique) entries of $\bC^T\bC$ and $\bC^T\by$. Remark that this secure summation protocol assumes that the different parties correctly follow the protocol specification and that they use their true data. @GHOSH:2007 used such a protocol for secure multiparty computation for multivariate adaptive regression splines to model nonlinear relationships. Compared to @GHOSH:2007 the Bayesian penalized splines approach in this paper offers the advantage of automated regularization parameter inference, providing measures of uncertainty (e.g. credible sets) and extensions to more complicated graphical models (e.g. grouped data, geostatistical data or sparse priors) are straightforward. In addition, online (cf. Algorithm \[algDistOnline\]) instead of batch computation can also be used in a secure multiparty computation context. However, this requires repeatedly applying the protocol above, which might be time-consuming.
Alternatively, a network-based client-server model can be used for secure computation as in @KARR:2007, where it was used for linear regression. In this way, the parties do not directly interact with each other but only through a server, having the advantage of randomizing the order in which messages are sent between the clients, i.e. parties. Encryption technology prevents the server from actually reading the summary statistics it passes between the clients.
MapReduce for processing large data sets {#section5:b}
----------------------------------------
The MapReduce programming model was developed at the Internet technology company Google Inc. for distributed processing of very large data sets [@DEAN:2008]. Being confronted with huge computing tasks Google Inc. decided to take advantage of a distributed computing environment, where large clusters of hundreds or thousands of commodity computers are connected together. Such a setting requires a system for taking care of partitioning the input data, scheduling the execution across the commodity computers, handling failure of computers and managing communication between the machines. The MapReduce framework provides a convenient way to handle these tasks and enables programmers without any experience with parallel and distributed systems to make use of the resources of distributed processing. In essence, MapReduce can be used in conjunction with various architectures. For example, @CHU:2007 presented a MapReduce implementation based on multicore computers, thereby taking advantage of the shared memory. In this section no assumptions are made about the underlying architecture as the main aim is to provide the general flavor of how the proposed algorithms fit into the MapReduce programming paradigm. In addition, the issue of when to opt for MapReduce over another distributed system is out of the scope of this paper.
A MapReduce task consists of a map phase and a reduce phase and the user has to specify the corresponding map and reduce function. The map function processes key-value pairs and outputs intermediate key-value pairs. Typically, the map task can be distributed over multiple machines, each operating in parallel on a small subset of the total data set. The reduce function then processes all the intermediate values that share the same intermediate key and outputs the final result. Essentially, the reduce function combines the intermediate results from the map function. Optionally, there is the possibility to implement a combiner function, which operates before the reduce phase starts. The combiner function is typically identical to the reduce function, but it is executed on each computer that performs a map task. This has the advantage of speeding up the computations when there exists significant repetition in the intermediate keys.
A factor that strongly popularized the use of MapReduce, was the development of an open-source implementation called Hadoop [@WHITE:2009]. While Hadoop was directly derived from Google Inc.’s MapReduce and the Google File System, a number of related projects have emerged in recent years. For example, the Mahout project is concerned with free implementations of distributed or otherwise scalable machine learning algorithms on the Hadoop platform (`http://mahout.apache.org/`). Most of the algorithms that are implemented in Mahout are concerned with clustering or classification. Here, a map and reduce function are presented for batch semiparametric regression when the data are stored in a distributed file system (Algorithm \[algMapReduce\]). The map function basically computes the summary statistics (as needed for executing Algorithm \[algDistBatch\]) based on subset $D_i$ of the total data set. Firstly, $\bC, \by$ and $n$ are extracted from $D_i$ and then the summary statistics based on this subset of samples are emitted together with the corresponding intermediate key. The reduce function simply sums the summary statistics with the same intermediate key together and outputs this result. Finally, lines 6–13 of Algorithm \[algDistBatch\] are used to compute $\bSigma_{q(\bbeta,\bu)}$ and $\bmu_{q(\bbeta,\bu)}$. Note that although lines 6–13 are iterative, the individual steps can again be parallelized (e.g. computing $\bSigma_{q(\bbeta,\bu)}$ or computing $\mu_{q(1/a_{u\ell})}$ and $\mu_{q(1/\sigma_{u\ell}^2)}$ for different values of $\ell$) as explained in Section \[section3:b\].
Although MapReduce was originally developed for computing batch jobs, a lot of research is going on to adapt it to process data streams. @CONDIE:2010 proposed a modified MapReduce architecture called Online MapReduce, that allows mappers to push data to reducers as soon as it is generated. This type of pipelining between mappers and reducers enables online aggregation and continuous queries. Online aggregation means that an intermediate result is generated during the course of execution, instead of having to wait for obtaining the final result till the job is completely finished. In addition, their architecture allows for real-time processing of data streams: the Online MapReduce jobs can run continuously, accept new data as it becomes available and process it immediately. Algorithm \[algDistOnline\] for real-time semiparametric regression of distributed data sets nicely fits into this architecture. By using the map and reduce function from Algorithm \[algMapReduce\], Online MapReduce repeatedly generates updated summary statistics. Each time the updated summary statistics are being outputted, lines 6–10 of Algorithm \[algDistOnline\] are executed, continuously producing estimates for $\bSigma_{q(\bbeta,\bu)}$ and $\bmu_{q(\bbeta,\bu)}$.
**function** map(key $i$, data set $D_i$) extract $\bC, \by$ and $n$ from data set $D_i$ compute $\bC^T\bC, \bC^T\by$ and $\by^T\by$ **emit**(1,$\bC^T\bC$) **emit**(2,$\bC^T\by$) **emit**(3,$\by^T\by$) **emit**(4,$n$) **function** reduce(key $i$, list $L$) $S \leftarrow$ sum($L$) **emit**($i$,$S$)
Example: real-time processing of U.S. domestic flight data {#section6}
==========================================================
This section demonstrates the proposed methodology by processing U.S. domestic flight data with the goal to analyze air traffic delays in real time. Nowadays, the status of a flight is continuously monitored and airports generate data streams which contain information about, among other things, the actual runway and gate arrival times for thousands of flights per day. For this example, the website `www.flightstats.com` is used to get access to these real-time data on flight delays, flight distances, operating airlines and flight paths. In addition, air temperature, wind speed and aviation flight category observations are continuously made at airports and nearby weather observation stations. These weather reports can be produced by automated airport weather stations or by trained observers or forecasters who manually observe and encode their observations. Here, these data are obtained through the `aviationweather.gov` website.
In this example the real-time flight data consist of the flight delay, flight distance, operating airline and the flight path. The real-time weather data consist of air temperature, wind speed and aviation flight category measurements. The flight category is a combined measure for the visibility and ceiling and there exist four categories: visual flight rules (VFR, i.e. ceiling $>$ 3000 feet and visibility $>$ 5 miles), marginal visual flight rules (MVFR, i.e. 1000 feet $\leq$ ceiling $\leq$ 3000 feet and/or 3 miles $\leq$ visibility $\leq$ 5 miles), instrument flight rules (IFR, i.e. 500 feet $\leq$ ceiling $<$ 1000 feet and/or 1 mile $\leq$ visibility $<$ 3 miles) and low instrument flight rules (LIFR, i.e. ceiling $<$ 500 feet and/or visibility $<$ 1 mile).
An extension of semiparametric regression model (\[eq:randInt\]) is used to demonstrate the methodology: $$\begin{array}{l}
\log(\mbox{\texttt{delay}}_{ijk}+120)|\,\bbeta,U_i, V_j,\bu_7,
\bu_8,\bu_9,\bu_{10},\bu_{11},\sigeps^2\simind \\[1ex]
\hspace{2mm} N(\beta_0 + \beta_1 {\mbox{\texttt{MVFRdep}}}_{ijk}
+ \beta_2 {\mbox{\texttt{IFRdep}}}_{ijk} + \beta_3 {\mbox{\texttt{LIFRdep}}}_{ijk}
+ \beta_4 {\mbox{\texttt{MVFRarr}}}_{ijk}\\[1ex]
\hspace{2mm} + \beta_5 {\mbox{\texttt{IFRarr}}}_{ijk}
+ \beta_6 {\mbox{\texttt{LIFRarr}}}_{ijk} + f_7({\mbox{\texttt{flight distance}}}_{j} )
\\[1ex]
\hspace{2mm} + f_8({\mbox{\texttt{departure temperature}}}_{ijk} )
+ f_9({\mbox{\texttt{arrival temperature}}}_{ijk} )
\null\\[1ex]
\hspace{2mm} + f_{10}({\mbox{\texttt{departure wind speed}}}_{ijk} )
+ f_{11}({\mbox{\texttt{arrival wind speed}}}_{ijk} )
\null\\[1ex]
\hspace{2mm} + U_i + V_j,\sigeps^2),
\hspace{3mm} U_1,\hdots,U_{171} |\,\sigma^2_{U} \simind N(0,\sigma^2_{U}),
\hspace{3mm} V_1,\hdots,V_{2000} |\,\sigma^2_{V} \simind N(0,\sigma^2_{V}).
\end{array}
\label{eq:DomesticFlightData}$$ Here, `delay`$_{ijk}$ is the difference between the actual and scheduled runway arrival time in minutes for the $k$th flight of airline $i$ on flight path $j$. `MVFRdep`$_{ijk}$ is an indicator which equals 1 if MVFR are applied at the scheduled runway departure time of the $k$th flight of airline $i$ on flight path $j$ and 0 otherwise. The variable $\texttt{MVFRarr}_{ijk}$ is defined in an analogous way, but for the scheduled runway arrival time. The other aviation flight category variables are defined similarly. The variable `flight` `distance`$_{j}$ represents the distance of flight path $j$ in kilometers. Variables `departure temperature`$_{ijk}$ and `arrival temperature`$_{ijk}$ denote the air temperature in degrees Celsius at the scheduled runway departure and arrival time of the $k$th flight of airline $i$ on flight path $j$, respectively. Variables `departure wind speed`$_{ijk}$ and `arrival wind speed`$_{ijk}$ denote the wind speed in knots at the scheduled runway departure and arrival time of the $k$th flight of airline $i$ on flight path $j$, respectively. Random intercepts for each of the 171 airlines are denoted by $U_i$,$\, 1 \leq i \leq 171$, and random intercepts for each of the 2000 flight paths are defined by $V_j$,$\, 1\leq j \leq$ 2000. $\bbeta$ stores the fixed effect regression coefficients $\beta_0,\hdots,\beta_6$ and the linear contribution to $f_7, \hdots , f_{11}$. The spline basis coefficients for $f_7, \hdots , f_{11}$ are stored in $\bu_7, \hdots ,\bu_{11}$.
Algorithm \[algDistOnline\] is used to fit model (\[eq:DomesticFlightData\]) and the time window extension in assignments (\[window1\])–(\[window2\]) is implemented to focus only on the 30 most recent days, i.e. a time window of 30 days is used. The website `realtime-semiparametric-regression.net/FlightDataForgetting/` demonstrates fitting of (\[eq:DomesticFlightData\]) using this methodology and presents continuously updated results in real time. To highlight the advantage of distributed processing through Algorithm \[algDistOnline\] the combined results for data generated by 415 U.S. airports (i.e. the data hosts) are presented together with the separate results, obtained by independently using Algorithm 3 of @LUTS:2013 extended with time window assignments (\[window1\])–(\[window2\]), for `O’Hare International Airport` and `Dallas/Fort Worth International Airport`. The first table shows the influence of flight distance, airline and the weather at the departure and arrival airport on the flight delay using summary statistics from all airports, from only `O’Hare International Airport` and from only `Dallas/Fort Worth International Airport`. Particularly interesting are the top 10 airlines having lowest and highest delays during the last 30 days after correcting for all other covariates such as weather circumstances and airports through the flight path variable. Similarly, the second table provides the top 10 flight paths having lowest and highest delays during the last 30 days. All these regression summaries are computed in real-time and the figures are updated every few minutes. The figure entitled `airline delay evolution over time` visualizes the on-time performance for the major airlines `Delta Air Lines`, `United Airlines` and `Southwest Airlines` based on the estimates for the random intercepts $U_i$ by combining summary statistics from 415 airports. Each day a new data point is added to this figure for each of these airlines.
Conclusion {#section7}
==========
This paper proposes methodology for semiparametric regression analysis when the samples are spread over multiple data hosts. Often it is not possible to move the raw data itself due to their large-scale nature or due to confidentiality issues. Mean field variational Bayes semiparametric regression algorithms are presented for this setting, thereby allowing data to be processed in batch or in an online manner. The key aspect of these approaches is to combine summary statistics instead of actual data. Compared to earlier work on regression for distributed data sets, this allows modeling of nonlinear relationships, enables fully-automated inference for smoothing parameters and provides measures of uncertainty. Furthermore, the presented model handles complications as grouped data and the Bayesian approach permits extensions to a wider variety of models.
An important aspect of analyzing distributed streaming data is to adapt to changes in the target over time. Two approaches have been proposed to deal with evolving environments. The first approach uses a time window to let the real-time regression estimates only depend on the most recent samples. This requires defining the window width and storing the summary statistics belonging to the time window. The second approach uses a decaying window by reweighting the summary statistics of older data and new data to handle a changing environment. This approach requires choosing a learning rate.
In order to illustrate the practical relevance of the proposed method, two types of application areas are discussed: semiparametric regression when there are multiple data owners requiring secure multiparty computation and the use of semiparametric regression within the MapReduce programming model. Finally, the method has been demonstrated on a real-life data set. An Internet site attached to this paper visualizes semiparametric regression analysis for infinite streams of data that are generated at 415 U.S. airports in real time. Future work includes extensions to other types of regression models as for example logistic regression or models with sparsity-inducing penalties for automated variable selection.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research was supported by Australian Research Council Discovery Project DP110100061. The author is grateful to Alan Huang and Matt Wand for their comments.
[^1]: Jan Luts (E-mail: *[email protected]*; Tel.: +61 2 9514 2267; Fax: +61 2 9514 2260)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In online advertising, the Internet users may be exposed to a sequence of different ad campaigns, i.e., display ads, search, or referrals from multiple channels, before led up to any final sales conversion and transaction. For both campaigners and publishers, it is fundamentally critical to estimate the contribution from ad campaign touch-points during the customer journey (conversion funnel) and assign the right credit to the right ad exposure accordingly. However, the existing research on the multi-touch attribution problem lacks a principled way of utilizing the users’ pre-conversion actions (i.e., clicks), and quite often fails to model the sequential patterns among the touch points from a user’s behavior data. To make it worse, the current industry practice is merely employing a set of arbitrary rules as the attribution model, e.g., the popular *last-touch* model assigns 100% credit to the final touch-point regardless of actual attributions. In this paper, we propose a Dual-attention Recurrent Neural Network (DARNN) for the multi-touch attribution problem. It *learns* the attribution values through an attention mechanism directly from the conversion estimation objective. To achieve this, we utilize sequence-to-sequence prediction for user clicks, and combine both post-view and post-click attribution patterns together for the final conversion estimation. To quantitatively benchmark attribution models, we also propose a novel yet practical attribution evaluation scheme through the proxy of budget allocation (under the estimated attributions) over ad channels. The experimental results on two real datasets demonstrate the significant performance gains of our attribution model against the state of the art.'
author:
- |
Kan Ren, Yuchen Fang, Weinan Zhang, Shuhao Liu,\
Jiajun Li, Ya Zhang, Yong Yu, Jun Wang
bibliography:
- 'multi-attr.bib'
title: |
Learning Multi-touch Conversion Attribution\
with Dual-attention Mechanisms for Online Advertising
---
Introduction {#sec:intro}
============
A benefit for online advertising is that advertisers would be able to get a significant amount of user feedbacks to measure the successfulness of their ad campaigns and optimize them accordingly. Aiming at delivering the optimization above, computational advertising has gained a large attraction and achieved great progress in many technical fields, including user targeting [@mcmahan2013ad; @ren2016user], bidding strategy [@perlich2012bid; @zhang2014optimal; @ren2017bidding] and budget pacing [@amin2012budget; @agarwal2014budget; @lee2013real].
As illustrated in Figure \[fig:hori-vert-ads\], with online advertising, an Internet user may be exposed to a sequence of ad campaigns from multiple channels, such as search engines, social media, mobile platforms, before reaching to any final conversion and transaction. It is crucial for advertisers to attribute the right conversion credit onto each *touch point* (i.e., the interaction between the user and the ad content, along the customer journey). The reasons are threefold. First, advertisers should know the contribution of each single touch point to the final conversion so as to make informed impression-level ad buying decisions [@lee2012estimating]. Second, if the attribution of each conversion over multiple ad exposures can be accurately and reliably estimated, a more quantitative credit-based ad pricing scheme can be established between advertisers and publishers (and ad tech providers). Last but not least, the attribution aggregated over ad channels may provide useful guidance for advertisers to allocate their budgets over these ad channels so as to acquire more positive user actions with lower cost in the next-round campaigns [@geyik2014multi].
![An illustration of different user activity sequences over multiple channels. Here we illustrate two user interactions with the ad contents, i.e., impressions and clicks, over three typical ad channels. Each user would gain an impression and then would probably click on that. After a sequence of user actions, the final conversion may be drawn according to the comprehensive experiments.[]{data-label="fig:hori-vert-ads"}](figures/hori-vert-ads-3.pdf){width="0.8\columnwidth"}
Traditionally, the attribution problem is addressed simply by a rule-based approach among most advertisers, such as first touch, last touch and other simple mechanisms [@wang2017display]. Specifically, the first (last) touch method generally attributes the conversion credits to the first (last) user interaction with the ad content. While these methods are easy to deploy, they obviously lack adequate capability of useful pattern recognition to support the higher-level budget allocation [@chandler2009measuring]. @shao2011data proposed the first data-driven multi-touch attribution model to allocate the credits to all the user touch points. Whereafter, many works have been published including probabilistic models using some distributional assumptions [@xu2016lift; @dalessandro2012causally] and additive exciting process [@zhang2014multi; @ji2017additional; @xu2014path]. Despite of the claimed advantages, there are three important factors missing in the above solutions.
**Sequential Pattern Modeling.** These methods are all based on the assumption that the user conversion would be driven by the individual advertising touch point of positive influence [@zhang2014multi; @shao2011data], which may not be realistic for the user journey (conversion funnel). In fact, sequential patterns within the user browsing behavior are of great value for response prediction or decision making in many fields such as recommender systems [@rendle2010factorizing], information retrieval [@song2017hierarchical] and search advertising [@zhang2014sequential].
**Data-driven Credit Assignment.** The attribution credits obtained in these models are heuristically assigned to each user interaction with the advertiser’s contents, rather than statistically learned from the data. For example, @ji2017additional proposed that the final conversion would be driven by the additive hazard rate of being converted at the time of each previous touch point, which pre-assumes that the more user exposures, the higher probability of the final user action. As is illustrated in Figure \[fig:user-sequence\] of one real-world dataset used in our experiments, the conversion rate does not necessarily increase w.r.t. the user action sequence length. This assumption may cause unconscionable ad exposures and may destruct user experience in online service.
**Different Pre-conversion Behaviors.** Almost all the related works ignore the difference between various types of user behaviors. Specifically, they assume the attribution are solely based on post-view or post-click, or even simply treat these behavior types equally for conversion attribution, where the credits are placed solely on impressions or clicks, or even discard the difference between them. These treatments are not effective since the user shows apparently different preferences behind the different interactive actions, which may (not) lead to the final conversion in different degrees.
To address the above limitations, in this paper, we propose a Dual-attention Recurrent Neural Network (DARNN) to capture the sequential user behavior patterns and learn the optimal attentions as the conversion attributions. Specificall, our model has two learning objectives. On one hand, we utilize sequence-to-sequence architecture to model the relationship between the impressions and the click actions, where the click behavior modeling is handled in this procedure. On the other hand, the final sequence prediction is the probability of the user conversion with the attention learned from the sequential modeling. The advantage of the attention mechanism is that it not only contributes to the sequence prediction accuracy, but also naturally *learns* the attribution of the conversion action over the whole sequence of the touch points. Moreover, DARNN applies the attention mechanism not merely on the features of the original touch point, but additionally over the learned hidden states of click actions, and then dynamically combines both attentions to predict the final conversion, which is the reason of *dual* attention. By this means the conversion estimation has captured both impression-level and click-level patterns. We also note that both dual-attention and the dynamic combination for the final conversion prediction are statistically learned from the data.
In addition, we also propose an offline evaluation framework for conversion attribution mechanisms. Since the obtained attribution credits over different channels could direct the budget allocation for the subsequent ad campaigns. However, none of the related work has empirically shown the effectiveness of the calculated attributions [@shao2011data; @dalessandro2012causally; @zhang2014multi; @ji2016probabilistic; @ji2017additional; @xu2014path] unless spending a huge budget to conduct online A/B test [@xu2016lift; @geyik2014multi]. Since budget allocation over different ad channels is always an important decision for advertisers to make, it is crucial for them to evaluate the performance of their multi-touch attribution methods, before the online A/B testing phase.
![The statistics of action sequence lengths and the conversion rates over Criteo dataset. The left plot shows the number of sequences and the converted sequences w.r.t. sequence length; the right plot presents the user conversion rate w.r.t. the user action sequence length.[]{data-label="fig:user-sequence"}](figures/length_num_cvr.pdf){width="1.0\columnwidth"}
To sum up, the novelty of our work is fourfold. (i) We build sequential pattern learning models for user behavior sequence. (ii) Our model learns the attribution from the final conversion estimation rather than heuristically assigning credits. (iii) We combine different attributions on various user action types. (iv) We also propose an offline evaluation protocol to measure the effectiveness of the attribution model through ad budget allocation and campaign data replay. Our experimental results prove the significant improvement (over 5.5%) of our DARNN model on conversion estimation performance against state-of-the-art baselines. The back replay evaluation also illustrates that the proposed model achieves the best cost effective performance.
Related Works {#sec:related-work}
=============
In online advertising, conversion attribution is commonly calculated by some rule-based methods, such as first-touch and last-touch, whereafter the return-on-investment (ROI) is gained based on the achieved attribution results which may result in some bias [@chandler2009measuring]. In recent years, many works based on multi-touch attribution (MTA) have been proposed for modeling the attribution for the sequential touch points over various channels [@berman2017beyond; @sinha2014estimating]. @shao2011data proposed the first work for data-driven multi-touch attribution model, which estimates the conversion rate based on the viewed ads of the user by the bagged logistic regression model. Some other works are mainly based on the probabilistic models with some distributional assumptions. @dalessandro2012causally proposed a causally motivated methodology that the conversion credits should be assigned by a causal cooperation model such as Shapley value. @xu2016lift argued that the user behavior has different additional effect on the final user decision of conversion and proposed a lift-based prediction model for real-time ad delivery. However, these methods did not take all the touch points of a user into the whole consideration so that the temporal and sequential factors were ignored [@ji2017additional]. Moreover, these models did either not consider much of sequential pattern modeling, which has been shown great effectiveness of user modeling [@zhang2014sequential].
As for the multiple interactions between advertisers and users, many works proposed the exciting point process methods for user behavior modeling. @yan2015machine developed a two-dimensional Hawkes process model to capture the influences from sellers’ activities on their contributions to the winning outcome in sales pipeline analytics. @xu2014path presented an MTA model based on mutually exciting point process which independently considered the impressions and clicks as random events along the continuous time. These exciting point process methods only considered the occurrence of the event which ignored the data of non-conversion cases. For analyzing the cumulative effects of the touch points, many works [@ji2017additional; @zhang2014multi; @xu2016lift; @sinha2014estimating] made an assumption that the final conversion was influenced by the additive contributions from the touch points along the user browsing history. However, it might result in a trend that more ad exposures were better which severely destroyed the user experience [@yuan2015supply]. The reason is that the attribution of each touch points along the user behavior sequence may positively contribute or counteract the final conversion. Thus, it is more reasonable to dynamically calculate and assign the attribution credits over the user behaviors.
Another school of MTA modeling is based on the survival theory [@zhang2014multi; @ji2016probabilistic; @ji2017additional], which models the conversion event as the predictive goal and estimates the probability for the event occurrence at the specific time while considering the censored data, i.e., the true occurrence time is later than the observation time. Nevertheless, these methodologies focus more on single point prediction and fail to consider the sequential patterns embedded in the user browsing history. Moreover, the obtained attribution credits are mainly calculated based on heuristic additive assumptions, which may not be effective in practice. They also made assumptions about the survival function such as exponential hazard function [@zhang2014multi; @ji2017additional] and Weibull distribution for hazard rate estimation [@ji2016probabilistic] to make their model parameterized and thus optimizable. However, such parameterization could severely constrain the capacity of the model to fit various real-world data.
Considering all the limitations above, we propose a dual-attention recurrent neural network for both conversion estimation and attribution. Attention mechanism is originally proposed for machine translation tasks, where a sequence-to-sequence model samples the next output word by attend each word of the input sentence [@bahdanau2014neural]. In our problem, the attention is modeled as the attribution which may dynamically *learn* to assign the credits over all the historical touch points for a specific user. The sequential patterns have been efficiently captured by the recurrent mechanism.
Moreover, few works have discussed the budget allocation from the obtained attribution model. @DiemertMeynet2017 proposed a bidding strategy based on the attribution credits for each real-time auction, which is not appropriate in general applications of online advertising. @geyik2014multi presented a method for online budget allocation based on the obtained ROI from conversion attribution. We borrow the idea of the ROI calculation from [@geyik2014multi] and devise an offline evaluation framework for multi-touch conversion attribution, which is the first offline experimental evaluation methodology for attribution models.
Methodology {#sec:methodology}
===========
In this section, we firstly formulate the problem of the multi-touch conversion attribution, and then propose our sequential behavior modeling with dual-attention mechanism. Finally we present our evaluation protocol for conversion attribution guided budget allocation.
Problem Definition {#sec:problem}
------------------
Without loss of generality, let us focus our study on the advertiser side. When a user $u_i$ is taking Internet activities, e.g., browsing online contents, querying search engines or playing on social media, etc., there would be many sequential interactions between this user and the ad contents of an advertiser, which are called touch points for the ad campaign. The observations are the user browsing sequences $\{ u_i, \{ q_j \}_{j=1}^{m_i}, y_i, T_i \}_{i=1}^{n}$ for each user $u_i$ who generates totally $m_i$ browsing activities with the ad of the advertiser. $y_i$ is the indicator of whether the user converts and $T_i$ is the conversion time if the conversion occurs, otherwise null. Each touch point $q_j = ({{\boldsymbol}{x}}_j, z_j)$ contains the feature vector ${{\boldsymbol}{x}}_j$ of the this touch point and the binary action type $z_j$, i.e., non-click impression or click. Among them, the feature ${{\boldsymbol}{x}}_j$ includes the side information of the user and the ad contents, e.g., user ID, advertising form, website, the operation systems and browser information, also with the channel ID feature $c_j$ over which this touch point is delivered and the time $t_j$ of the interaction occurrence.
The goal is to model the sequential user patterns and derive efficient conversion attribution credits for all the touch points $\{q_j\}_{j=1}^{m_i}$ along the user browsing sequence. In return, the better conversion attribution obtained, the higher accuracy of the user conversion estimation for each browsing sequence. Similar formulation has been adopted in many literatures [@zhang2014multi; @ji2016probabilistic; @ji2017additional]. In Sec. \[sec:seq-model\], we present a recurrent neural network to model the sequential patterns and the final conversion rate. We also apply sequence-to-sequence modeling for user click pattern mining and jointly learn impression and click patterns for conversion estimation. The key component in this sequence modeling methodology is the dual-attention mechanism which takes two types of the user actions (i.e., impressions and clicks) into a unified comprehensive framework and facilitate the conversion modeling, which is described in Sec. \[sec:attention\]. As a result, the obtained attention from the sequence modeling is naturally the conversion attribution over the whole user browsing history.
Interestingly, the derived conversion attribution also contributes to budget allocation for the subsequent ad delivery [@geyik2014multi]. In Sec. \[sec:eval-alg\], we propose an evaluation protocol for budget allocation with offline campaign data.
Sequential Modeling {#sec:seq-model}
-------------------
We utilize recurrent neural network (RNN) for sequential user modeling, as illustrated in Figure \[fig:model\]. Leveraging RNN for sequential modeling and time-series prediction has been widely applied in information retrieval systems [@song2017hierarchical; @qin2017dual; @zhai2016deepintent]. Note that our methodology aims at final conversion estimation rather than sequential prediction for click at each touch point.
The whole structure can be divided as three separate parts that (i) the encoder for the impression-level behavior modeling; (ii) the decoder and sequential prediction for click probability; (iii) taking the above modeling output we implement dual-attention for jointly modeling impression and click behavior and produce the final conversion estimation. We will clarify the first two parts in this section and discuss the attention mechanism later.
[**Impression-level Behavior Modeling.**]{} For the $i^{\text{th}}$ user behavior sequence $\{ u_i, \{ q_j \}_{j=1}^{m_i}, y_i, T_i \}$ where $q_j = ({{\boldsymbol}{x}}_j, z_j)$, the input feature sequence to the RNN model is $\mathbf{x} = ({{\boldsymbol}{x}}_1, \dots, {{\boldsymbol}{x}}_j, \ldots, {{\boldsymbol}{x}}_{m_i})$. Since the side-information feature vector is mostly categorical [@zhang2016deep], we firstly utilize an embedding layer to transform the sparse input feature into dense representation vector for subsequent training, which has been widely used in the related literatures [@qu2016product; @wang2017dynamic].
Then we feed the embedded feature vectors through the encoder RNN function $f_e$ approach as $$\label{eq:encoder}
\begin{aligned}
{{\boldsymbol}{h}}_j &= f_e({{\boldsymbol}{x}}_j, {{\boldsymbol}{h}}_{j-1}) ~, \\
\end{aligned}$$ where ${{\boldsymbol}{h}}$ is the hidden vector of each time step $j$. We implement $f_e$ as a standard long short-term memory (LSTM) model described in [@hochreiter1997long], which has been widely used in natural language processing fields.
![Sequential modeling with dual-attention.[]{data-label="fig:model"}](figures/darnn.pdf){width="1.0\columnwidth"}
[**Click-level Sequential Prediction.**]{} In this part, our goal is to model the click action at each time when each ad is shown to the user.
In the sequence-to-sequence model, the decoder defines a probability over the click outcomes $\mathbf{z}$ by decomposing the *joint probability* into the ordered conditionals as $$p(\mathbf{z}) = \prod_{j=1}^{m_i} p( z_j=1 ~|~ \{ z_1, \ldots, z_{j-1} \}, \mathbf{x}) ~.$$ Note that $\mathbf{z} = (z_1, \ldots, z_{m_i})$ and $\mathbf{x} = ({{\boldsymbol}{x}}_1, \ldots, {{\boldsymbol}{x}}_{m_i})$. With this decoder each conditional probability is modeled as $$\hat{z}_j = p(z_j=1 ~|~ \{z_1, \ldots, z_{j-1} \}, \mathbf{x}) = g(z_{j-1}, {{\boldsymbol}{s}}_j) ~.$$ Here $g$ is the output function which is a multi-layer fully connected perceptron with sigmoid activation function $\text{sigmoid}(x) = \frac{1}{1+e^{-x}}$ that outputs the probability of $z_j=1$. And ${{\boldsymbol}{s}}_j$ is the hidden vector at click-level of the $j^{\text{th}}$ touch point, calculated by $$\label{eq:decoder}
{{\boldsymbol}{s}}_j = f_d({{\boldsymbol}{s}}_{j-1}, z_{j-1}, {{\boldsymbol}{h}}_{m_i})~,$$ where $f_d$ is the nonlinear decoder RNN function, potentially multi-layered, that models the sequential click patterns for user behavior sequence. We utilize the same structure of LSTM model as the encoder $f_e$. Each hidden state ${{\boldsymbol}{s}}_j$ in the decoder uses the last hidden state ${{\boldsymbol}{h}}_{m_i}$ from the encoder. Our first loss is based on the sequential prediction for click probabilities as $$\label{eq:click-obj-func}
L^c = \sum_{i=1}^n \sum_{j=1}^{m_i} -z_j \log \hat{z}_j - (1-z_j) \log (1-\hat{z}_j) ~.$$
There are two rationales for the sequence-to-sequence click prediction in this work. The first is to some extent similar with the idea of multi-task learning to alleviate the data sparsity problem and conduct a shared base representation of user behavior features. As is known that the users follow a pattern of actions that they may click after impression of the ad and after a sequence of ad delivery they may (not) drive the final conversion, which derives the data sparsity problem behind the “impression-click-conversion” action pattern [@ma2018entire]. Specifically, clicks are less frequent events than impressions and conversions are much rarer than clicks. It is necessary to conduct a methodology to tackle with the data sparsity challenge. Our intuition is to utilize the signal of click behavior to boost the estimation capacity for the sparse conversion behaviors. Another reason for the sequential click pattern mining is to obtain the click-level attribution modeling for multiple pre-conversion behavior modeling, which has shown statistically more important attribution credits than impression-level behaviors in our experimental results.
Learning Attribution with Dual-attention {#sec:attention}
----------------------------------------
Our final goal is to model the sequential user patterns and predict the conversion probability. The final output is calculated as $$\label{eq:final-conv-estimation}
\begin{aligned}
{{\boldsymbol}{c}}^{i2v} &= A^{i2v}({{{\boldsymbol}{h}}_1, \ldots, {{\boldsymbol}{h}}_j, \ldots, {{\boldsymbol}{h}}_{m_i}}) ~,\\
{{\boldsymbol}{c}}^{c2v} &= A^{c2v}({{{\boldsymbol}{s}}_1, \ldots, {{\boldsymbol}{s}}_j, \ldots, {{\boldsymbol}{s}}_{m_i}}) ~, \\
\hat{y}_i &= p(y=1 ~|~ \mathbf{x}, \mathbf{z}) = r({{\boldsymbol}{x}}_{m_i}, {{\boldsymbol}{c}}^{i2v}, {{\boldsymbol}{c}}^{c2v}) ~,
\end{aligned}$$ where $r$ contains a weighting function for balancing impression-level and click-level attribution, which will be described in detail later in this section, and a dense multi-layer neural network for the final conversion prediction. ${{\boldsymbol}{x}}_{m_i}$ is the feature vector of the last touch point which would be fed through the same embedding layer as that in Sec. \[sec:seq-model\]. ${{\boldsymbol}{c}}^{i2v}$ is the context vectors of all the input user behavior vectors capturing impression patterns and ${{\boldsymbol}{c}}^{c2v}$ is the context vector from modeling the click patterns for conversion estimation.
[**Learning Attention through Conversion Estimation.**]{} The loss is calculated by the cross entropy for the conversion estimation that $$\label{eq:obj-func}
L^v = \sum_{i=1}^n - y_i \log \hat{y}_i - (1-y_i) \log(1 - \hat{y}_i) ~.$$
![Attention calculation mechanism.[]{data-label="fig:attention"}](figures/attention.pdf){width="0.65\columnwidth"}
The key component is the attention input ${{\boldsymbol}{c}}^{i2v}$ and ${{\boldsymbol}{c}}^{c2v}$ from impression-level and click-level, respectively. The mechanism of attention function $A$ is illustrated in Figure \[fig:attention\].
To calculate the impression-to-conversion attention ${{\boldsymbol}{c}}^{i2v}$ and the click-to-conversion attention ${{\boldsymbol}{c}}^{c2v}$, we propose a unified energy-based function as $$\begin{aligned}
{{\boldsymbol}{c}}&= l({{{\boldsymbol}{h}}_1, \ldots, {{\boldsymbol}{h}}_j, \ldots, {{\boldsymbol}{h}}_{m_i}}) = \sum_{j=1}^{m_i} a_j {{\boldsymbol}{h}}_j ~.
\end{aligned}$$ Note that this formulation is expressed to calculate ${{\boldsymbol}{c}}^{i2v}$ while without losing generality by replacing ${{\boldsymbol}{x}}$ and ${{\boldsymbol}{h}}$ in Eq. (\[eq:encoder\]) with ${{\boldsymbol}{s}}$, $z$ and ${{\boldsymbol}{h}}_{m_i}$ in Eq. (\[eq:decoder\]) for ${{\boldsymbol}{c}}^{c2v}$ calculation.
And the weight $a_j$ is calculated based on the softmax operated energy value $e_j$ as $$\label{eq:attn-weight}
a_j = \frac{\exp(e_j)}{\sum_{k=1}^{m_i} \exp(e_k)} ~,$$ where $$\label{eq:attn-energy}
e_j = E({{\boldsymbol}{h}}_j, {{\boldsymbol}{x}}_{m_i})$$ is an energy model which scores the credit of each touch point to the final conversion. Note that the energy function $E$ is a nonlinear multi-layer deep neural network with tanh activation function $\text{tanh}(x)=\frac{e^x - e^{-x}}{e^x + e^{-x}}$. The way we calculate the attention through the energy function $E$ is similar to that in the natural language processing field [@bahdanau2014neural; @gehring2017convolutional].
As a result, the dual-attention mechanism is expressed as $$\begin{aligned}
{{\boldsymbol}{c}}^{i2v} &= \sum_{j=1}^{m_i} a^{i2v}_j {{\boldsymbol}{h}}_j ~, \\
{{\boldsymbol}{c}}^{c2v} &= \sum_{j=1}^{m_i} a^{c2v}_j {{\boldsymbol}{s}}_j ~, \\
\end{aligned}$$ and the values of $a$ in both attention calculation are obtained through Eqs. (\[eq:attn-weight\]) and (\[eq:attn-energy\]).
[**Attribution Calculation with Dual-attention.**]{} Till now, we have obtained the estimated conversion probability $p(y|\mathbf{x}, \mathbf{z})$ and the attention results $a^{i2v}_j$ and $a^{c2v}_j$ for each touch point, based on which we can naturally assign the credits for each touch point $j$.
Recall that the final conversion estimation is based on the learned dual-attention vector, i.e., ${{\boldsymbol}{c}}^{i2v}_j$ and ${{\boldsymbol}{c}}^{c2v}_j$, and the final touch point feature vector ${{\boldsymbol}{x}}_{m_i}$, here we adopt a dynamic weighting function $f_{\lambda}$ to balance the effcts of the two attentions that $$\label{eq:lambda-calculation}
\lambda = \frac{\exp\left[f_{\lambda}({{\boldsymbol}{x}}_{m_i}, {{\boldsymbol}{c}}^{c2v})\right]}{\exp\left[f_{\lambda}({{\boldsymbol}{x}}_{m_i}, {{\boldsymbol}{c}}^{i2v})\right] + \exp\left[f_{\lambda}({{\boldsymbol}{x}}_{m_i}, {{\boldsymbol}{c}}^{c2v})\right]} ~,$$ where $\lambda$ measures the importance of the click-level attention w.r.t. that from the impression-level and $f_{\lambda}$ is a multi-layer perceptron whose goal is to learn the weight of two attention results for the final conversion estimation.
Thus, the estimation function $r$ mentioned first in Eq. (\[eq:final-conv-estimation\]) is that $$r({{\boldsymbol}{x}}_{m_i}, {{\boldsymbol}{c}}^{i2v}, {{\boldsymbol}{c}}^{c2v}) = r_{\text{conv}}\left( (1 - \lambda) \cdot {{\boldsymbol}{c}}^{i2v} + \lambda \cdot {{\boldsymbol}{c}}^{c2v} \right) ~.$$ Here $r_{\text{conv}}(\cdot)$ is a multi-layer neural network for conversion estimation with sigmoid activation function.
Now that we have weighted the contribution of impression-level and click-level attentions for final conversion estimation, we can naturally obtain the attribution for each touch point through these learned patterns as $$\label{eq:attention-calc}
\text{Attr}_j = (1-\lambda) \cdot a^{i2v}_j + \lambda \cdot a^{c2v}_j ~.$$
The motivation for building such a dual-attention mechanism is that we care both the impression-level and the click-level user behavior patterns to facilitate conversion estimation and the subsequent attribution results.
Evaluation Protocol {#sec:eval-alg}
-------------------
With the attribution credits allocated to the touch points along the user behavior sequences, our focus moves onto the efficiency of budget allocation based on the calculated attribution credits. Note that almost all the related works report only the conversion estimation performance; few of them test the budget allocation under the obtained attribution credits, except online A/B testing which is expensive and risky. Here we propose a framework to offline evaluate the conversion attribution model based on the historic data of a campaign.
In online advertising, the guideline of the advertiser to allocate budgets for the subsequent ad delivery on different channels of the ad campaign is intuitively based on the past performance. The performance here means the effectiveness, i.e., return on investment (ROI), of the ad delivery onto each channel, which is measured as the obtained positive user conversions w.r.t. the delivered ad costs. The most intuitive idea is to allocate more budgets for the channels or sub-campaigns with higher ROI than others, to gain more user conversions. However, different attribution methods substantially influence the ROI calculation results [@geyik2014multi]. Specifically, the idea of our evaluation protocol is to first calculate the ROI performance results for each channel under different attribution models, and then utilize the offline replay of ad delivery history to measure the performance of the obtained fresh conversions and, considering the costs in the offline replay, calculate the effectiveness results of the ad delievery for different evaluation baselines. So that, under this evaluation, the more proper attribution credits one model proposed, the better performance it would obtain in the subsequent budget allocation for different ad channels and naturally obtain better performance through the ad replay evaluation.
Next we will first present the budget allocation method based on attribution-guided ROI results. Then we illustate our back evaluation algorithm w.r.t. the allocated budget scheme for later performance comparison.
[**ROI-based Budget Allocation.**]{} In this stage, the first problem is to allocate the budget $\{ b_1, \ldots, b_K \}$ across $K$ channels according to the obtained attribution credits. Here we follow the idea presented in [@geyik2014multi] that $$\label{eq:roi-calc}
\text{ROI}_{c_k} = \frac{\sum_{\forall y_i=1} \text{Attr}(c_k | y_i) ~ V(y_i)}{\text{Money spent on channel} ~ c_k} ~,$$ where $$\begin{aligned}
\text{Attr}(c_k | y_i) = \sum_{j=1}^{m_i} \text{Attr}_j \cdot 1(c_j = c_k)\end{aligned}$$ is the overall credit attributed on the channel $c_k$ by aggregating the credit $\text{Attr}_j$ of all touch points $j$’s within this channel, $1(\cdot)$ is the indicator function, and $V(y_i)$ is the value of the conversion. After the ROI calculation we allocate budgets for different channels w.r.t. the obtained ROI proportion as $b_k = \frac{\text{ROI}_{c_k}}{\sum_{v=1}^K ROI_{c_v}} \times B$ for channel $c_k$, where $B$ is the total budget.
[**Back Evaluation under Reallocated Budgets.**]{} The historic data is a series of event sequences $\{\text{seq}^s\}$ and each sequence is represented as $\{ (q_i, {{\boldsymbol}{x}}_i, t_i, y_i, c_i, o_i) \}$ where each user interaction identified by $q_i$ is on the specific channel $c_{i}$ at time $t_i$ with cost $o_i$ and the feature vector ${{\boldsymbol}{x}}_i$ includes the click label information. Each event is either an ad serving event without conversion ($y_i = 0$) or a user conversion event ($y_i = 1$).
In addition, we introduce the concept of *conversion blacklist*. If the budget of one channel $c_k$ is exhausted at moment $t$, then the conversion events of all the unfinished sequence with ad serving event after $t$ on channel $c_k$ become invalid. These conversions should be put into the conversion blacklist. This is reasonable because if the user cannot observe the ad touch point, it is no guarantee that she will finally convert at the end of the sequence. Such a back test result serves as a lower bound estimation of the true but unknown performance.
Given the budget allocation $(b_1, b_2, \ldots, b_K)$ across $K$ channels, we can make the following back test as presented in Algorithm \[alg:eval\]. Specifically, the back test goes over the historic events by their recording time $t_i$. If there is no budget left for the channel $c_k$ for the back playing event $(q_i, {{\boldsymbol}{x}}_i, t_i, y_i, c_i=c_k, o_i) $, then put the sequence indicator $\text{seq}^s$ into conversion blacklist. After the back test, we can evaluate the attribution models by the cost $O$ and the obtained valid conversion number $Y$.
The events $\{ (q_i, {{\boldsymbol}{x}}_i, t_i, y_i, c_i, o_i) \}$ ordered by the\
serving time $t_i$ and the budget allocation $\{ b_1, \ldots, b_K \}$. The total conversion number $Y$ and the total cost $O$. Initially set the blacklist of sequence list $\mathcal{B} = \{\}$ and the obtained conversion number $Y=0$, total cost $O=0$. $O = O + o_i$ $Y = Y + y_i$ $b_{c_i} = b_{c_i} - o_i$ Put $\text{seq}^s$ into $\mathcal{B}$
Experiments {#sec:exp}
===========
In this section, we firstly present the experiment setup including the description of two real-world datasets, the evaluation measurements and the compared models used in our experiments. Then we illustrate the corresponding results for the two-staged experiment settings. The first stage is for the conversion estimation accuracy while the second one is for the attribution guided budget allocation performance over history data. In addition, we have published our code[^1] for repeatable experiments.
Datasets
--------
In our experiments, we apply our model and the compared baselines over two real-world datasets.
**Miaozhen** is a leading marketing technique company in China. This dataset [@zhang2014multi] includes almost 1.24 billion advertising logs from May 1 to June 30 and April 4 to June 9 in 2013. Specifically it contains about 59 million users and 1044 conversions. These ad contents have been exposed over 2498 channels with 40 advertising forms, such as button ads and social ads. In the dataset, every time a user is exposed to the ad or click on the ad contents, the exact time of the user action with the side information will be recorded. Moreover, it also contains the purchasing information as the conversion of the user with the corresponding timestamp. The user is tracked according to the user cookie identifier which is anonymized in the dataset. With these logs, we are able to reconstruct the time line of the user action sequence including impression and click information, the exposure ad channels and the conversion labels for each sequence.
**Criteo** is a pioneering company in online advertising research. They have published this dataset[^2] for attribution modeling in real-time auction based advertising [@DiemertMeynet2017]. This dataset is formed of Criteo live traffic data in a period of 30 days. It has more than 16 million impressions and 45 thousand conversions over 700 campaigns. The impressions in this dataset may derive click actions so each touch point along the user action sequence has a label of whether a click has occurred, and the corresponding conversion ID if this sequence of touch points leads to a conversion event. Each impression log also contains the cost information, which will be used in our second state experiment for attribution effectiveness evaluation. Since the channel data are missing so we take campaign as the budget allocation targets.
[**Data Preprocessing and Sampling.**]{} Since the user conversion is a rare event, we perform negative sampling in data preprocessing. Following [@zhang2014multi; @ji2017additional], the sequence preparation and sampling rules are that (i) if the user has multiple conversion events, her action sequence will be split according to the conversion time to guarantee that each sequence has at most one conversion; (ii) we extract the user action sequences with the minimal length of 3 and maximal length of 20 with the sequence duration within 14 days; (iii) all of the user sequences leading to conversion events have been retained and we uniformly sample the sequences without conversions to 20 times of the number of converted sequences.
Evaluation Pipeline and Metrics
-------------------------------
Here we present the evaluation pipeline and the measurements over the compared settings. Overall, we have two stages of the experiments.
The first stage focuses on the conversion estimation performance which has been widely adopted in the conversion attribution task [@zhang2014multi; @ji2016probabilistic; @ji2017additional]. Specifically, given the evaluation samples $\{ u_i, \{ q_j \}_{j=1}^{m_i}, y_i, T_i \}_{i=1}^{n}$ in the test dataset, the model predicts the output of conversion probability $\hat{y}_i$ after the user going through each sequence of touch points. There are two evaluation metrics for measuring the performance of each model. **Log-loss** is the common measurement to estimate the classification performance for the event probability prediction which is the cross entropy as is expressed in Eq. (\[eq:obj-func\]). The other metric is **AUC** (area under ROC curve) which measures the pairwise ranking performance of the classification results between the converted and nonconverted sequence samples.
The second stage aims at the performance of budget allocation, with the calculated conversion attributions, for various channels or sub-campaigns. According to Algorithm \[alg:eval\], we replay all the test campaign data w.r.t. the recorded timestamp, and calculate the performance for the below metrics. Note that we set the evaluation budgets for each model as 1/2, 1/4, 1/8, 1/16, 1/32 of the total budget in the whole test dataset. The similar evaluation setting has been widely adopted in online advertising researches [@zhang2014optimal; @ren2016user; @ren2017bidding]. The **number of conversions** is the total number of the achieved conversions. **Profit** is the total gains, i.e., total value of the obtained conversions. The other two metrics are **CVR** (conversion rate) and **CPA** (cost per conversion action). CVR is the ratio of the converted sequences among all the touched user impression sequences which reflects the ratio of gain for the ad delivery. And CPA is the cost averaged by the obtained conversion numbers which mesures the efficiency for the ad campaign. Note that only Criteo dataset contains the cost information so that our second stage experiments is conducted on Criteo dataset.
Compared Settings
-----------------
In this section, we discuss the compared baselines and our model settings. We compare four baseline models with our dual-attention model. We also discuss the advantages of our dual-attention mechanism against the normal RNN model with single attention mechanism. Note that the click label ground truth $z$ has been included in the input feature ${{\boldsymbol}{x}}$ in the other baseline models, except for our proposed model which utilizes click as the prediction label, for equally comparison.
- **LR** is the Logistic Regression model proposed in [@shao2011data] and the attribution is calculated as the learned coefficient values of the regression model parameter for each channel.
- **SP** is a Simple Probabilistic model whose idea is derived from [@dalessandro2012causally] and the conversion rate of each user action sequence is calculated as in [@zhang2014multi] that $$p( y = 1 | \{ c_j \}_{j=1}^{m_i} ) = 1 - \prod_j^{m_i} ( 1 - \text{Pr}(y = 1 | c_j = k) ) ~,
\vspace{-5pt}$$ where $\text{Pr}(y = 1 | c_j = k)$ is the conversion probability from the observed data w.r.t. the $k\text{th}$ channel.
- **AH** (AdditiveHazard) model is the first work [@zhang2014multi] using survival analysis and additive hazard function of conversion with the consideration of the touch point time to predict the final conversion rate. More details could be found in the paper.
- **AMTA** is the Additional Multi-touch Attribution model proposed in [@ji2017additional] which was state-of-the-art for this conversion attribution problem. It applies survival analysis to model the conversion estimation and utilizes the hazard rate of conversion at the specific time to model the conversion attribution.
- **ARNN** is the normal Recurrent Neural Network (i.e., only encoder part) method with the single Attention merely based on impression-level patterns to model the conversion attribution that $\hat{y}_i = r'({{\boldsymbol}{c}}^{i2v})$, rather than sequence-to-sequence modeling in Eq. (\[eq:decoder\]). This model is to illustrate the advantage of our dual-attention mechanism for data sparsity problem and multi-view learning schema.
- **DARNN** is our proposed model with dual-attention mechanism, which has been described in Section \[sec:methodology\].
All the deep models are trained separately over one NVIDIA GeForce GTX 1080 Ti with Intel Core i7 processor for five hours. The detailed hyperparameter settings have been described in our published code, including learning rate, feature embedding size, hidden state size in RNN cell, etc.
Conversion Estimation Performance
---------------------------------
Our first evaluation is to measure the performance of user conversion estimation. Table \[tab:conv\_est\] presents the detailed evaluation results under different models. From the statistics in the table, our model outperforms other baselines under both evaluation metrics. The results also reflect the other findings as below. (i) Both of the attention-based methods, i.e., DARNN and ARNN, achieve much better performance for sequential prediction than other compared models, which reflects the great pattern mining capability of deep neural networks. (ii) The exciting point process based methods AH and AMTA has poor classification performance for the conversion estimation. The reason is that they are designed to model the additive hazard ratio of conversion for each touch point. Though they learn the conversion prediction for the whole sequence, they do not consider much of the sequential patterns within the user behavior sequence. (iii) For the log-loss metric, the baselines get relatively higher (i.e., poorer) values than the deep models, which reflects that these baselines predict the conversion probability with totally large or small absolute values. Note that, however, AUC has no relationship to the direct output value of the model but considers the pairwise ranking performance. So almost all the baselines get considerably acceptable AUC results.
![Learning curves on two datasets. Here one “epoch” means one whole iteration over the train dataset. The vertical purple line means the conversion estimation optimization starts in the second training stage.[]{data-label="fig:learning_curve"}](figures/learning_curve.pdf){width="1\columnwidth"}
As for the learning procedure, since our proposed DARNN model captures both impression-level and click-level patterns and optimizes under two types of losses as that in Eqs. (\[eq:click-obj-func\]) and (\[eq:obj-func\]), the training procedure of our DARNN model generates two learning curves in Figure \[fig:learning\_curve\]. In model training, we firstly make the model learn the click patterns, i.e., only optimize under the sequential click prediction loss as that in Eq. (\[eq:click-obj-func\]) and then, after the convergence of the first objective, we turn on the conversion estimation training, i.e., optimize under both two losses till convergence of the conversion loss. The convergence of each objective is defined as two successive rising of the optimization loss. The reason of two-stage training is to stabilize the model optimization under these two learning objectives. The similar training procedure has been studied in multi-task learning for recommender systems [@cao2018neural].
We may easily find from Figure \[fig:learning\_curve\] that our model not only optimizes the sequential click prediction, but also learns the conversion estimation. Moreover, the two learning objectives have been alternatively optimized to convergence at the second stage. Both the click prediction and the conversion estimation achieve excellent prediction performance.
Attribution Guided Budget Allocation
------------------------------------
In the second stage of the experiments, we evaluate the effectiveness of different conversion attribution models for budget allocation. After replaying the historic touch points along the ordered timestamps, we calculated the total costs and the obtained conversion numbers of the compared model settings. To calculate the obtained profits for each model, we make the conversion value $V(y_i=1)$ in Eq. (\[eq:roi-calc\]) as eCPA (effective cost-per-action) which is constant for each model and calculated as $V(y_i=1) = \text{eCPA}_{\text{train}} = (\text{total cost} / \text{conversion number})$ in the training data. The detailed results are presented in Table \[tab:budget\_allo\] and Figure \[fig:budget\_allocation\]. As is presented in the table, since LR performs quite poor, we eliminate LR results in the figure for better illustration. Moreover, note that, we also compared simple *last-touch* attribution method in the second-stage experiment. We did not report this heuristic method in our experiments since the result showed that AH baseline model performed almost the same as the last-touch attribution method which is quite interesting and needs further investigations in the future work.
![Performance with budgets on Criteo.[]{data-label="fig:budget_allocation"}](figures/stage_2.pdf){width="1.0\columnwidth"}
From Table \[tab:budget\_allo\] and Figure \[fig:budget\_allocation\], we may find that: (i) As the budget increases, all the models spend more to earn each user action, i.e., the CPA value of each model is increasing, which is reasonable. (ii) Both of the attention-based neural network models, i.e., ARNN and DARNNs, achieve relatively better performance compared with the other models over all the evaluation metrics. The reason is probably the sequential pattern mining of these two models. (iii) The two baselines AMTA and AH achieve very similar performance, which is probably accounted for the similar idea of the additional conversion probability modeling within their models. (iv) DARNN model achieves the best performance under CPA and CVR, which reflects the effectiveness our learned attribution values of dual-attention mechanism. Moreover, this result also shows the advantage of the dual attention mechanism over single attention model ARNN. (v) ARNN spends money more aggressive than other models thus getting poor CPA result. The reason may be that its attribution is based merely on impression-level and the pattern captured tends to long-term investment on the user behavior. However, our DARNN model spends the budget more economically which leads to more efficient budget pacing, i.e., lower CPA. This indicates that combining both impression-level and click-level attention will take advantages of both long-term (impression to conversion) and short-term (click to conversion) behavior patterns.
Comprehensive Analysis
----------------------
In this part, we look deeper into the learned attribution model. We first discuss the calculated attribute credits over both touch point level and channel level, and then analyze the results of the learned weighting parameter $\lambda$ according to Eq. (\[eq:lambda-calculation\]) which controls the influence of the two types of user actions, i.e., impressions and clicks.
First, we illustrate the touch point attribution in Miaozhen dataset which calculates the averaged attribution credits over all the sequence samples with fixed sequence length, on each touch point. Specifically, here the credits on the $j^{\text{th}}$ touch point is averaged over all the converted sequences with the fixed sequence length $m$ as $\overline{\text{Attr}}_j = \frac{1}{N_m} \sum_{i=1}^{N_m} y_i \cdot \text{Attr}_{ij}$, where $N_m$ is the total number of the sequences with length $m$ and $y_i$ is the conversion indicator of the sequence sample.
![Touch point level attribution statistics (Miaozhen).[]{data-label="fig:attribution_over_sequence"}](figures/horizontal_attr.pdf){width="1.0\columnwidth"}
Figure \[fig:attribution\_over\_sequence\] illustrates the touch point conversion attribution results on the sequences with length of 5 and 10 respectively. Since LR and SP calculate the attribution based on different channels rather than each touch point, so we cannot get the specific result of these two models at the touch point level. From the figure we may find that the credits attributed on each touch point varies over different models. When sequence length is relatively short, DARNN learns that the touch point closer to the final touch may more likely derive the final conversion. In longer sequences (with length of 10), our DARNN model place higher credits for the touch points in the middle process while the attribution drops a little later and consequently rises to much higher when final conversion approaches. This phenomenon is reasonable since it is not always correct about endless ad delivery for the user and, moreover, it reflects the tradeoff between the ad effectiveness and the user experience of the Internet service. However, ARNN seems to “average” the credits over all the touch points within the sequence. Note that ARNN only concerns impression-level contributions, which in contrast shows the great effects of click-level patterns in our proposed dual-attention mechanism on the final conversion attribution.
![Attribution of different channels on Miaozhen.[]{data-label="fig:attribution_miaozhen"}](figures/attribution_miaozhen.pdf){width="1.0\columnwidth"}
The next analysis is based on the channel level attribution distribution. We illustrate the calculated attribution credits of converted sequences over multiple channels in Miaozhen dataset as that in Figure \[fig:attribution\_miaozhen\]. The horizontal axis is the channel information varying from social media to music platforms. Since there is no conversions on music channel, no credit has been assigned by the models except LR which takes the learned parameter coefficient of the channel feature for conversion attribution. From the illustration we can find that (i) LR, SP, AH and DARNN models assign the highest attribution credits to search channel, while ARNN and AMTA attribute the most onto video channel. (ii) SP and AH assign relatively much higher credits on search channel, while the other models distribute attribution more smoothly. (iii) Vertical and community channels have low credits while union channel has much higher attribution under attention-based models. From these findings we find the significance of the replay evaluation for attribution guided budget allocation in the second stage of experiments, since the calculated conversion attribution credits over different channels vary from different models.
![Attribution distribution over channels on Criteo.[]{data-label="fig:attribution_criteo"}](figures/attribution_criteo.pdf){width="1.0\columnwidth"}
![$\text{ROI}$ distribution over channels on Criteo.[]{data-label="fig:ROI_criteo"}](figures/ROI_criteo.pdf){width="1.0\columnwidth"}
In addition, we illustrate the attribution credits on Criteo dataset in Figure \[fig:attribution\_criteo\] and the corresponding $\text{ROI}_{c_k}$ of each channel $c_k$ in Figure \[fig:ROI\_criteo\] which is calculated according to Eq. (\[eq:roi-calc\]) under different models. From channel level, our DARNN model assigns the highest credits onto channel 5. However, the ROI calculation derives that all models allocate the most budget credits onto channel 7. This is reasonable since the ROI is based on both channel level and touch point level information as that in Eq. (\[eq:roi-calc\]).
![The distribution of $\lambda$ over Criteo dataset.[]{data-label="fig:lambda-tuning"}](figures/lambda_distribution.pdf){width="1.0\columnwidth"}
Finally, in Figure \[fig:lambda-tuning\], we statistically visualize the value distribution of $\lambda$ which controls the impact from impression-level and click-level patterns, respectively. As is calculated in Eq. (\[eq:lambda-calculation\]), note that, when $\lambda$ gets larger the click-level patterns get higher impact on the conversion attribution. We may find from the figure that the click-level patterns relatively contribute more to the final conversion estimation, which reflects that the effectiveness of our dual-attention mechanism for different action pattern mining, as is described in Sec. \[sec:attention\]. Generally speaking, the results illustrate the importance of combining both impression pattern and click pattern through dual-attention mechanism, especially that the click-level patterns contribute better under tight budget cases.
Conclusion and Future Work {#sec:conclusion}
==========================
In this paper, we proposed a dual-attention recurrent neural network model for learning to assign conversion credits over the ad touch point sequences. Our model not only captures sequential user patterns, but also pays attention to both impression-level and click-level user actions and derives an effective conversion attribution methodology. The experiments show the significant improvement over the other state-of-the-art baselines.
One of the limitations of this work is that we have not taken the cost of ad impressions into account in the attention mechanism. It is of great interest to take the cost factor into modeling and improve the cost-effectiveness performance in the future as that in the works [@ren2016user; @ren2017bidding] of real-time auction advertising.
[^1]: Repeatable experiment code link: <https://github.com/rk2900/deep-conv-attr>.
[^2]: Processed dataset link: <http://apex.sjtu.edu.cn/datasets/13>.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[*A Banach algebra $\A$ for which the natural embedding $x\mapsto \hat{x}$ of $\A$ into $WAP(\A)^*$ is bounded below; that is, for some $m\in \mathbb{R}$ with $m>0$ we have $||\hat{x}||\geq m||x||$, is called a WAP-algebra. Through we mainly concern with weighted measure algebra $M_b(S,\omega),$ where $\omega$ is a weight on a semi-topological semigroup $S$. We study those conditions under which $M_b(S,\omega)$ is a WAP-algebra (respectively dual Banach algebra). In particular, $M_b(S)$ is a WAP-algebra (respectively dual Banach algebra) if and only if $wap(S)$ separates the points of $S$ (respectively $S$ is [*compactly cancellative semigroup*]{}). We apply our results for improving some older results in the case where $S$ is discrete.*]{}'
address:
- '$^1$Department of Pure Mathematics, Center of Excellence in Analysis on Algebraic Structures (CEAAS), Ferdowsi University of Mashhad, IRAN, e-mail: [[email protected]]{}'
- '$^2$ Corresponding author, Department of Mathematics, University of Isfahan, Isfahan, IRAN. e-mail: [${\texttt {b}}_{-}[email protected]]{}'
- '$^3$Department of Mathematics, University of Isfahan, Isfahan, IRAN. e-mail: [[email protected]]{}'
author:
- 'H.R. Ebrahimi Vishki'
- 'B. Khodsiani'
- 'A. Rejali'
title: ' WEIGHTED SEMIGROUP MEASURE ALGEBRA AS A ${\rm WAP}$-ALGEBRA'
---
[ WAP-algebra, dual Banach algebra, Arens regularity, weak almost periodicity.]{}
43A10, 43A20, 46H15, 46H25.
Introduction and Preliminaries
==============================
Throughout this paper, we study those conditions under which $M_b(S,\omega)$ is either a ${\rm WAP}$-algebra or a dual Banach algebra. Our main result in section 2 is that for a locally compact topological semigroup and a continuous weight $\omega$ on $S$, the measure algebra $M_b(S,\omega)$ is a dual Banach algebra with respect to $C_0(S, 1/{\omega})$ if and only if for all compact subsets $F$ and $K$ of $S$ , the maps $\frac{{\displaystyle \chi}_{F^{-1}K}}{\omega}$ and $\frac{{\displaystyle \chi}_{KF^{-1}}}{\omega}$ vanishes at infinity. This improved the result of Abolghasemi, Rejali, Ebrahimi Vishki [@ARV] to include the case where $S$ is not necessarily discrete. As a consequence in non-weighted case, we conclude for a locally compact topological semigroup $S$, the measure algebra $M_b(S)$ is a dual Banach algebra with respect to $C_0(S)$ if and only if $S$ is a compactly cancellative semigroup. The later result improved the well known result of Dales, Lau and Strauss [@DLS Theorem4.6], $\ell_1(S)$ is dual Banach algebra with respect to $c_0(S)$ if and only if $S$ is weakly cancellative semigroup.
Section 3 is devoted to study $WAP$-algebras on a semigroup $S$. For every weighted locally compact semi-topological semigroup $(S,\omega)$, $M_b(S,\omega)$ is a ${\rm WAP}$-algebra if and only if the evaluation map $\epsilon:S\longrightarrow \tilde {X}$ is one to one, where $\tilde{X}=MM(wap(S,1/\omega))$. Our main result of this section is that $M_b(S)$ is ${\rm WAP}$-algebra if and only if $wap(S)$ separate the points of $S$. If $C_0(S,1/\omega)\subseteq wap(S,1/\omega)$ then $wap(S,1/\omega)$ separate the points of $S$. Thus $M_b(S,\omega)$ is a ${\rm WAP}$-algebra. We may ask whether, if $M_b(S,\omega)$ is a ${\rm WAP}$-algebra then $C_0(S,1/\omega)\subseteq wap(S,1/\omega)$. We answer to this question negatively by a counter example. Then we exhibit some necessary and sufficient condition for $c_0(S)\subseteq wap(S)$. we end the paper by some examples which show that our results cannot be improved.
The dual $\A^*$ of a Banach algebra $\A$ can be turned into a Banach $\A-$module in a natural way, by setting
$$\langle f\cdot a, b\rangle=\langle f ,ab\rangle \ \ \ \ \ {\rm and} \ \ \ \ \langle a\cdot f, b\rangle=\langle f, ba\rangle\ \ \ \ (a,b\in \A, f\in \A^*).$$
A [*dual Banach algebra*]{} is a Banach algebra $\A$ such that $\A=(\A_*)^*$, as a Banach space, for some Banach space $\A_*$, and such that $\A_*$ is a closed $\A-$submodule of $\A^*;$ or equivalently, the multiplication on $\A$ is separately weak\*-continuous. We call $\A_*$ the predual of $\A$. It should be remarked that the predual of a dual Banach algebra need not be unique, in general (see [@D; @DPW]); so we usually point to the involved predual of a dual Banach algebra.
A functional $f\in \A^*$ is said to be [*weakly almost periodic*]{} if $\{f\cdot a: \|a\|\leq 1\}$ is relatively weakly compact in $\A^*$. We denote by $WAP(\A)$ the set of all weakly almost periodic elements of $\A^*.$ It is easy to verify that, $WAP(\A)$ is a (norm) closed subspace of $\A^*$.
It is known that the multiplication of a Banach algebra $\A$ has two natural but, in general, different extensions (called Arens products) to the second dual $\A^{**}$ each turning $\A^{**}$ into a Banach algebra. When these extensions are equal, $\A$ is said to be (Arens) regular. It can be verified that $\A$ is Arens regular if and only if $WAP(\A)=\A^*$. Further information for the Arens regularity of Banach algebras can be found in [@D; @DL].
WAP-algebras, as a generalization of the Arens regular algebras, has been introduced and intensively studied in [@Da2]. A Banach algebra $\A$ for which the natural embedding $x\mapsto \hat{x}$ of $\A$ into $WAP(\A)^*$ where $\hat{x}(\gamma)=\gamma(x)$ for $\gamma\in WAP(\A)$, is bounded below; that is, for some $m\in \mathbb{R}$ with $m>0$ we have $||\hat{x}||\geq m||x||$, is called a WAP-algebra. When $\A$ is Arens regular or dual Banach algebra, the natural embedding of $\A$ into $WAP(\A)^*$ is isometric [@Runde Corollary4.6]. Also Theorem \[mwap\] shows that $M_b(S,\omega)$ is a WAP-algebra if and only if this embedding is isometric and of course bounded below, however in general $M_b(S,\omega)$ is neither Arens regular nor dual Banach algebra. It has also known that $\A$ is a WAP-algebra if and only if it admits an isometric representation on a reflexive Banach space.
Moreover, group algebras are also always WAP-algebras, however; they are neither dual Banach algebras, nor Arens regular in the case where the underlying group is not discrete, see [@y1]. Ample information about WAP-algebras with further details can be found in the impressive paper [@Da2].
A character on an ablian algebra $\A$ is a non-zero homomorphism $\tau:\A\rightarrow \mathbb{C}.$ The set of all characters on $\A$ endowed with relative weak$^*$- topology is called character space of $\A$.
Following [@BJM], a semi-topological semigroup is a semigroup $S$ equipped with a Hausdorff topology under which the multiplication of $S$ is separately continuous. If the multiplication of $S$ is jointly continuous, then $S$ is said to be a topological semigroup. We write $\ell^{\infty}(S)$ for the commutative $C^*$-algebra of all bounded complex-valued functions on $S$. In the case where $S$ is locally compact we also write $C(S)$ and $C_0(S)$ for the $C^*-$subalgebras of $\ell_\infty(S)$ consist of continuous elements and continuous elements which vanish at infinity, respectively. We also denote the space of all [*weakly almost periodic*]{} functions on $S$ by $wap(S)$ which is defined by $$wap(S)=\{f\in C(S): \{R_sf: s\in S\}\ {\rm is\ relatively\ weakly\ compact}\},$$ where $R_sf(t)=f(ts), \ (s,t\in S).$ Then $wap(S)$ is a $C^*-$subalgebra of $C(S)$ and its character space $S^{wap}$, endowed with the Gelfand topology, enjoys a (Arens type) multiplication that turns it into a compact semi-topological semigroup. The evaluation mapping $\epsilon: S\rightarrow S^{wap}$ is a homomorphism with dense image and it induces an isometric $*-$isomorphism from $C(S^{wap})$ onto $wap(S).$ Many other properties of $wap(S)$ and its inclusion relations among other function algebras are completely explored in [@BJM].
Let $M_b(S)$ be the Banach space of all complex regular Borel measures on $S,$ which is known as a Banach algebra with the total variation norm and under the convolution product $*$ defined by the equation $$\langle \mu*\nu,g\rangle =\int_S\int_Sg(xy)d\mu(x)d\nu(y)\quad (g\in C_0(S))$$ and as dual of $C_0(S)$.
Throughout, a weight on $S$ is a Borel measurable function $\omega:S\rightarrow (0,\infty)$ such that $$\omega(st)\leq \omega(s)\omega(t),\ \ (s,t\in S).$$ For $\mu\in M_b(S)$ we define $(\mu\omega)(E)=\int_E\omega d\mu, \ (E\subseteq S\ {\rm is\ Borel\ set}).$ If $\omega\geq 1$, then $$M_b(S,\omega)=\{\mu\in M_b(S): \mu\omega\in M_b(S)\}$$ is known as a Banach algebra which is called the [*weighted semigroup measure algebra*]{} (see [@DL; @Re1; @Re; @RV] for further details about such algebras and arbitrary weight functions). Let $S$ be a locally compact semigroup, and let $B(S)$ denote the space of all Borel measurable and bounded functions on $S$. Set $B(S,1/\omega)=\{f:S\rightarrow \mathbb{C}:f/\omega\in B(S)\}$. A standard predual for $M_b(S,\omega)$ is $$C_0(S,1/\omega)=\{f\in B(S,1/\omega): f/\omega\in C_0(S)\}.$$ Let $f \in C(S,1/\omega)$ then $f$ is called $\omega$-weakly almost periodic if the set $\{\frac{R_sf}{\omega(s)\omega}: s \in S\}$ is relatively weakly compact in $C(S)$, where $R_s$ is defined as above. The set of all $\omega$-weakly almost periodic functions on $S$ is denoted by $wap(S, 1/\omega)$.
In the case where $S$ is discrete we write $\ell_1(S,\omega)$ instead of $M_b(S,\omega)$ and $c_0(S,\frac{1}{\omega})$ instead of $C_0(S,\frac{1}{\omega}).$ Then the space $$\ell_1(S,\omega)=\{f:f=\sum_{s\in S}f(s)\delta_s,\quad ||f||_{1,\omega}=\sum_{s\in S}|f(s)|\omega(s)<\infty\}$$
(where, $\delta_s\in\ell_1(S,\omega) $ be the point mass at $s$ which can be thought as the vector basis element of $\ell_1(S,\omega)$ ) equipped with the multiplication $$f*g=\sum_{r\in S}\sum_{st=r}f(s)g(t)\delta_r$$ (and also define $f*g=0$ if for each $r\in S$ the equation $st=r$ has no solution;) is a Banach algebra which will be called weighted semigroup algebra. We also suppress $1$ from the notation whenever $w=1.$
Semigroup Measure Algebras as Dual Banach Algebras
====================================================
It is known that the semigroup algebra $\ell_1(S)$ is a dual Banach algebra with respect to $c_0(S)$ if and only if $S$ is weakly cancellative semigroup, see [@DLS Theorem4.6]. Throughout this section $\omega$ is a continuous weight on $S$. This result has been extended for the weighted semigroup algebras $\ell_1(S,\omega)$; [@ARV; @Da1]. In this section we extend this results to the non-discrete case. We provide some necessary and sufficient conditions that the measure algebra $M_b(S, \omega)$ becomes a dual Banach algebra with respect to the predual $C_0(S,1/\omega)$.
Let $F$ and $K$ be nonempty subsets of a semigroup $S$ and $s\in S$. We put $$s^{-1}F=\{t\in S:st\in F\}\mbox{, and}\quad Fs^{-1}=\{t\in S:ts\in F\}$$ and we also write $s^{-1}t $ for the set $s^{-1}\{t\}$, $FK^{-1}$ for $\cup\{Fs^{-1}:s\in K\}$ and $K^{-1}F$ for $\cup\{s^{-1}F:s\in K\}$.
A semigroup $S$ is called left (respectively, right) zero semigroup if $xy=x$ (respectively, $xy=y$), for all $x,y\in S$. A semigroup $S$ is called zero semigroup if there exist $z\in S$ such that $xy=z$ for all $x,y\in S$. A semigroup $S$ is said to be [*left (respectively, right) weakly cancellative semigroup*]{} if $s^{-1}F$ (respectively, $Fs^{-1}$) is finite for each $s\in S$ and each finite subset $F$ of $S$. A semigroup $S$ is said to be [*weakly cancellative semigroup*]{} if it is both left and right weakly cancellative semigroup.
A semi-topological semigroup $S$ is said to be [*compactly cancellative semigroup*]{} if for every compact subsets $F$ and $K$ of $S$ the sets $F^{-1}K $ and $KF^{-1}$ are compact set.
\[ert\][Let $S$ be a topological semigroup. For every compact subsets $F$ and $K$ of $S$ the sets $F^{-1}K $ and $KF^{-1}$ are closed. ]{}
If ${F^{-1}K }$ is empty, then it is closed. Let $x$ be in the closure of ${F^{-1}K }$. Then there is a net $(x_\alpha)$ in $F^{-1}K $ such that $x_\alpha \rightarrow x$. Since $x_\alpha\in F^{-1}K $ there is a net $(f_\alpha)$ in $F$ such that $f_\alpha x_\alpha\in K$. Using the compactness of $F$ and $K$, by passing to a subnet, if necessary, we may suppose that $f_\alpha x_\alpha\rightarrow k$ and $f_\alpha \rightarrow f$, for some $f\in F$ and $k\in K$. So $fx=k\in K$, that is $x\in {F^{-1}K }$. Therefore $F^{-1}K $ is closed. A similar argument shows that $KF^{-1}$ is also closed.
In the next result we study $M_b(S,\omega)$ from the dual Banach algebra point of view.
\[we\][Let $S$ be a locally compact topological semigroup and $\omega$ be a continuous weight on $S$. Then the measure algebra $M_b(S,\omega)$ is a dual Banach algebra with respect to the predual $C_0(S, 1/{\omega})$ if and only if for all compact subsets $F$ and $K$ of $S$ , the maps $\frac{{\displaystyle \chi}_{F^{-1}K}}{\omega}$ and $\frac{{\displaystyle \chi}_{KF^{-1}}}{\omega}$ vanishes at infinity. ]{}
Suppose that $M_b(S,\omega)$ is a dual Banach algebra with respect to $C_0(S, 1/\omega)$ and let $K, F$ be nonempty compact subsets of $S$ with a net $(x_\alpha)$ in $F^{-1}K $. Let $C^+_{00}(S)$ denote the non-negative continuous functions with compact support on $S$ and set $C_{00}^+(S,\frac{1}{\omega})=\{f\in C_{0}(S,\frac{1}{\omega}):f/\omega\in C_{00}^+(S)\}$. Since $\omega$ is continuous we may choose $f\in C_{00}^+(S,\frac{1}{\omega})$ with $f(K)=1$. There is a net $(t_\alpha)\in F$ such that $t_\alpha x_\alpha\in K$ and the compactness of $F$ guaranties the existence of a subnet $(t_\gamma)$ of $(t_\alpha)$ such that $t_\gamma\rightarrow t_0$ for some $t_0$ in $S$. Indeed, for $s\in S$, $$\lim_{\gamma}(\frac{\delta_{t_\gamma}.f}{\omega})(s)=\lim_{\gamma}\frac{f(t_\gamma s)}{\omega(s)}=
\frac{f(t_0s)}{\omega(s)}=\frac{\delta_{t_0}.f}{\omega}(s)$$ there is a $\gamma_0$ such that $$\bigcup_{\gamma\geq\gamma_0}t_\gamma^{-1}K\subseteq\bigcup_{\gamma\geq\gamma_0}\{s\in
S:(\frac{\delta_{t_\gamma}.f}{\omega})(s)\geq1\}\subseteq\{s\in
S:(\frac{\delta_{t_0}.f}{\omega})(s)\geq\frac{1}{2}\}.$$ Let $H=\{t_\gamma: \gamma\geq\gamma_0\}\cup\{t_0\}$. Then $$H^{-1}K =\bigcup_{\gamma\geq\gamma_0} t_\gamma^{-1}K\cup t_0^{-1}K\subseteq
\{s\in S:(\frac{\delta_{t_0}.f}{\omega})(s)\geq\frac{1}{2}\}$$ and so $H^{-1}K $ is compact. Furthermore, $t_\gamma x_\gamma\in K$, that is $(x_\gamma)$ is a net in compact set $H^{-1}K$. This means that $(x_\alpha)$ has a convergent subnet and this is the proof of necessity.
The sufficiency can be adopted from [@ARV Proposition 3.1] with some modifications. Let $f\in C_0(S, 1/\omega)$ , $\mu\in M_b(S,\omega)
$ and $\varepsilon >0$ be arbitrary. There exist compact subsets $F$ and $K$ of S such that $|\frac{f}{\omega}(s)|<\varepsilon$ for all $s\not\in K$ and $|(\mu\omega)|(S\setminus F)<\varepsilon$.
Let $s\not\in \{t\in F^{-1}K:\omega(t)\leq \frac{1}{\varepsilon}\}$ , which is compact by hypothesis. Then $$\begin{aligned}
|\frac{\mu.f}{\omega}(s)|&=&|\int_S\frac{f(ts)}{\omega(s)}d\mu(t)|\\
&\leq&|\int_F\frac{f(ts)}{\omega(s)}d\mu(t)|+|\int_{S\setminus F}\frac{f(ts)}{\omega(s)}d\mu(t)|\\
&\leq&\int_F|\frac{f(ts)}{\omega(ts)}|\omega(t)d|\mu|(t)+\int_{S\setminus F}|\frac{f(ts)}{\omega(ts)}|\omega(t)d|\mu|(t)\\
&\leq&\varepsilon\int_S\omega(t)d|\mu|(t)+\|f\|_{\omega,\infty}\int_{S\setminus F}\omega(t)d|\mu|(t)\\
&\leq&\varepsilon\|\mu\|_{\omega}+\varepsilon\|f\|_{\omega,\infty}\end{aligned}$$ That is, $\mu.f\in C_0(S, 1/\omega)$. Therefore $M_b(S,\omega)$ is a dual Banach algebra with respect to $C_0(S, 1/\omega)$.
The next Corollaries are immediate consequences of Theorem \[we\].
\[w\][Let $S$ be a locally compact topological semigroup. Then the measure algebra $M_b(S)$ is a dual Banach algebra with respect to $C_0(S)$ if and only if $S$ is a compactly cancellative semigroup. ]{}
\[ws\][[@ARV Theorem 2.2]]{}[For a semigroup $S$ the semigroup algebra $\ell_1(S,\omega)$ is a dual Banach algebra with respect to the predual $c_0(S, 1/\omega)$ if and only if for all $s,t\in S$, the maps $\frac{{\displaystyle \chi}_{t^{-1}s}}{\omega}$ and $\frac{{\displaystyle \chi}_{st^{-1}}}{\omega}$ are in $c_0(S)$. ]{}
\[yuu\][For a locally compact topological semigroup $S$, if $M_b(S)$ is a dual Banach algebra with respect to $C_0(S)$ then $M_b(S,\omega)$ is a dual Banach algebra with respect to $C_0(S, 1/\omega)$. ]{}
[Let $S$ be either a left zero (right zero) or a zero locally compact semigroup. There is a weight $\omega$ such that $M_b(S,\omega)$ is a dual Banach algebra with respect to $C_0(S,\frac{1}{ \omega})$ if and only if $S$ is $\sigma$-compact. ]{}
Let $K$ and $F$ be compact subsets of $S$. It can be readily verified that in either cases (being left zero, right zero or zero) the sets $F^{-1}K $ and $KF^{-1}$ are equal to either empty or $S$. Put $$S_m=\{t\in
F^{-1}K:\omega(t)\leq m\}=\{t\in S: \omega(t)\leq m\}\ \ (m\in\mathbb{N}).$$ Then $S=\cup_{m\in \mathbb{N}}S_m$ and so $S$ is $\sigma$-compact. For the converse let $S=\cup_{n\in \mathbb{N}}S_n$ as a disjoint union of compact sets and let $z$ be a (left or right) zero for $S$. Define $\omega(z)=1$ and $\omega(x)=1+n$ for $x\in S_n$ then $\omega $ is a weight on $S$ and $M_b(S,\omega)$ is a dual Banach algebra.
\[Ex\]
1. The set $S=\mathbb{R}^+\times\mathbb{R}$ equipped with the multiplication $$(x,y).(x',y')=(x+x',y')\ \ \ ((x,y),(x',y')\in S)$$ and the weight $\omega(x,y)=e^{-x}(1+|y|)$ is a weighted semigroup. In this example $[a,b]$ denotes a closed interval. As for $F=[a,b]\times[c,d]$ and $K=[e,f]\times[g,h],$ with $[c,d]\cap[g,h]\not=\emptyset$ $$\begin{aligned}
% \nonumber to remove numbering (before each equation)
F^{-1}K &=& \bigcup_{(x,y)\in F}(x,y)^{-1}K \\
&=& \bigcup_{(x,y)\in F}\{(s,t)\in S:(x,y)(s,t)\in K\} \\
&=& \bigcup_{(x,y)\in F}\{(s,t)\in S:(x+s,t)\in K\} \\
&=& \bigcup_{(x,y)\in F}[e-x,f-x]\times[g,h]=[e-b,f-a]\times[g,h]\end{aligned}$$ and $$\begin{aligned}
% \nonumber to remove numbering (before each equation)
K F^{-1}&=& \bigcup_{(x,y)\in F}K(x,y)^{-1} \\
&=& \bigcup_{(x,y)\in F}\{(s,t)\in S:(s,t)(x,y)\in K\} \\
&=& \bigcup_{(x,y)\in F}\{(s,t)\in S:(x+s,y)\in K\} \\
&=& \bigcup_{(x,y)\in F}[e-x,f-x]\times\mathbb{R} =[e-b,f-a]\times\mathbb{R}\end{aligned}$$ Thus $$F^{-1}K=[e-b,f-a]\times[g,h]\mbox{\quad and \quad}KF^{-1}= \left\{
\begin{array}{lr}
[e-b,f-a]\times\mathbb{R} &\mbox{if}\quad [c,d]\cap[g,h]\not=\emptyset \\
\emptyset &\mbox{if}\quad [c,d]\cap[g,h]=\emptyset
\end{array}\right.
,$$ $M_b(S)$ is not a dual Banach algebra by Corollary \[w\]. However, for all compact subsets $F$ and $K$ of $S$ , the maps $\frac{{\displaystyle \chi}_{F^{-1}K}}{\omega}$ and $\frac{{\displaystyle \chi}_{KF^{-1}}}{\omega}$ vanishes at infinity. So $M_b(S,\omega)$ is a dual Banach algebra with respect to $C_0(S, 1/\omega)$. This shows that the converse of Corollary \[yuu\] may not be valid.
2. For the semigroup $S=[0,\infty)$ endowed with the zero multiplication, neither $M_b(S)$ nor $\ell_1(S)$ is a dual Banach algebra. In fact, $S$ is neither compactly nor weakly cancellative semigroup.
Semigroup Measure Algebras as WAP-Algebras
===========================================
In this section, for a weighted locally compact semi-topological semigroup $(S,\omega)$, we investigate some necessary and sufficient condition for $M_b(S,\omega)$ being WAP-algebra. First, we provide some preliminaries.
Let $\tilde{\mathcal{F}}$ be a linear subspace of $B(S,1/\omega)$, and let $\tilde{\mathcal{F}}_r$ denote the set of all real-valued members of $\tilde{\mathcal{F}}$. A mean on $\tilde{\mathcal{F}}$ is a linear functional $\tilde{\mu}$ on $\tilde{\mathcal{F}}$ with the property that $$\inf_{s\in S} \frac{f}{\omega} ( s ) \leq \tilde{\mu} ( f ) \leq \sup_{s\in S}\frac{f}{\omega}(s)\quad (f\in \tilde{\mathcal{F}}_r ).$$ The set of all means on $\tilde{\mathcal{F}}$ is denoted by $M(\tilde{\mathcal{F}})$. If $\tilde{\mathcal{F}}$ is also an algebra with the multiplication given by $f \odot g:= (f. g)/\omega\quad (f, g \in \tilde{\mathcal{F}} )$ and if $\tilde{\mu}\in M(\tilde{\mathcal{F}})$ satisfies $$\tilde{\mu}(f \odot g) = \tilde{\mu}(f)\tilde{\mu}(g)\quad (f, g \in \tilde{\mathcal{F}} ),$$ then $\tilde{\mu}$ is said to be multiplicative. The set of all multiplicative means on $\tilde{\mathcal{F}}$ will be denoted by $MM(\tilde{\mathcal{F}})$.
Let $\tilde{\mathcal{F}}$ be a conjugate closed, linear subspace of $B(S,1/\omega)$ such that $\omega\in \tilde{\mathcal{F}}$.
1. For each $s\in S$ define $\epsilon(s)\in M(\tilde{\mathcal{F}})$ by $\epsilon(s)(f)=(f/\omega)(s)\quad (f\in \tilde{\mathcal{F}})$. The mapping $\epsilon: S\longrightarrow M(\tilde{\mathcal{F}})$ is called the evaluation mapping. If $\tilde{\mathcal{F}}$ is also an algebra, then $\epsilon(S)\subseteq MM(\tilde{\mathcal{F}})$.
2. Let $\tilde X = M(\tilde{\mathcal{F}})$ (resp. $\tilde X = MM(\tilde{\mathcal{F}})$, if $\tilde{\mathcal{F}}$ is a subalgebra) be endowed with the relative weak\* topology. For each $f \in \tilde{\mathcal{F}}$ the function $\hat{f}\in C(\tilde{X})$ is defined by $$\hat{f}(\tilde{\mu}):=\tilde{\mu}(f)\quad (\tilde{\mu}\in \tilde{X}).$$
Furthermore, we define $\hat{\tilde{\mathcal{F}}}:=\{\hat{f}:f\in \tilde{\mathcal{F}}\}$
1. The mapping $f\longrightarrow \hat{f}: \tilde{\mathcal{F}}\longrightarrow C(\tilde{X})$ is clearly linear and multiplicative if $\tilde{\mathcal{F}}$ is an algebra and $\tilde{X }= MM(\tilde{\mathcal{F}})$. Also it preserves complex conjugation, and is an isometry, since for any $f \in \tilde{\mathcal{F}}$ $$\begin{aligned}
% \nonumber to remove numbering (before each equation)
||\hat{f}||&=&\sup\{|\hat{f}(\tilde{\mu})|:\tilde{\mu}\in\tilde{X}\}=\sup\{|\tilde{\mu}(f)|:\tilde{\mu}\in\tilde{X}\}\\
&=&\sup\{|\mu(\frac{f}{\omega})|:\mu\in X\}\leq\sup\{|\mu(\frac{f}{\omega})|:\mu\in C(X)^*, ||\mu||\leq1\} \\
&=&||\frac{f}{\omega}||=||f||_{\omega}=\sup\{|\frac{f}{\omega}(s)|:s\in S\}=\sup\{|\epsilon(s)(f)|:s\in S\}\\
&=& \sup\{|\hat{f}(\epsilon(s))|:s\in S\}\leq||\hat{f}||,\end{aligned}$$ where $X=M(\mathcal{F})$ and $\mathcal{F}=\{ f/\omega:f\in\tilde{\mathcal{F}}\}$. Note that $\hat{f}(\epsilon(s))=\epsilon(s)(f)=(\frac{f}{\omega})(s) (f\in\tilde{F}, s\in S)$. This identity may be written in terms of dual map $\tilde {\epsilon}^*:C(\tilde{X})\longrightarrow C(S,\omega)$ as $\epsilon^*(\hat{f})=f$ for $f\in \tilde {F}$.
2. Let $\tilde{\mathcal{F}}$ be a conjugate closed linear subspace of $B(S,1/\omega)$, containing $\omega$. Then $M(\tilde{\mathcal{F}})$ is convex and weak\* compact, $co(\epsilon(S))$ is weak\* dense in $M(\tilde{\mathcal{F}})$, $\tilde{\mathcal{F}}^*$ is the weak\* closed linear span of $\epsilon(S)$, $\epsilon : S \longrightarrow M(\tilde{\mathcal{F}})$ is weak\* continuous, and if $\tilde{\mathcal{F}}$ is also an algebra, then $MM(\tilde{\mathcal{F}})$ is weak\* compact and $\epsilon(S)$ is weak\* dense in $MM(\tilde{\mathcal{F}})$.
3. Let $\tilde{\mathcal{F}}$ be a $C^*$-subalgebra of $B(S,1/\omega)$, containing $\omega$. If $\tilde {X}$ denotes the space $MM(\tilde{\mathcal{F}})$ with the relative weak\* topology, and if $\epsilon : S \longrightarrow \tilde{X}$ denotes the evaluation mapping, then the mapping $f\longrightarrow\hat{f} :\tilde{\mathcal{F}}\longrightarrow C(\tilde {X})$ is an isometric isomorphism with the inverse $\epsilon^*: C(\tilde {X})\longrightarrow\tilde{\mathcal{F}}$.
Let $\tilde{\mathcal{F}}=wap(S,1/\omega)$. Then $\tilde{\mathcal{F}}$ is a $C^*$-algebra and a subspace of $WAP(M_b(S,\omega))$, see [@Lashkarizadeh Theorem1.6, Theorem3.3]. Set $\tilde{X}=MM(\tilde{\mathcal{F}})$. By the above remark $wap(S, 1/\omega)\cong C(\tilde{X})$ and so $$M_b(\tilde{X})\cong C(\tilde{X})^*\cong wap(S, 1/\omega)^* \subseteq WAP(M_b(S,\omega))^*.$$
Let $\epsilon :S\longrightarrow \tilde{X}$ be the evaluation mapping. We also define $\bar{\epsilon}: M_b(S,\omega)\longrightarrow M_b(\tilde {X})$, by $\langle\bar{\epsilon}(\mu),f\rangle=\int_Sf\omega d\mu$ for $f\in wap(S, 1/\omega)\cong C(\tilde{X})$. Then for every Borel set $B$ in $\tilde {X}$, $$\bar{\epsilon}(\mu)(B)=(\mu\omega)(\epsilon^{-1}(B)).$$ In particular, $\bar{\epsilon}(\frac{\delta_x}{\omega(x)})=\delta_{\epsilon(x)}$.
The next theorem is the main result of this section.
\[mwap\]
For every weighted locally compact semi-topological semigroup $(S,\omega)$ the following statements are equivalent:
1. The evaluation map $\epsilon:S\longrightarrow \tilde {X}$ is one to one, where $\tilde{X}=MM(wap(S,1/\omega))$;
2. $\bar{\epsilon}: M_b(S,\omega)\longrightarrow M_b(\tilde {X})$ is an isometric isomorphism;
3. $M_b(S,\omega)$ is a [WAP]{}-algebra.
\(1) $\Rightarrow$ (2). Take $\mu\in M_b(S,\omega)$, say $\mu=\mu_1-\mu_2+i(\mu_3-\mu_4)$, where $\mu_j\in M_b(S,\omega)^+$. Set $\nu_j=\bar{\epsilon}(\mu_j)\in M_b(\tilde {X})^+$ for $j=1,2,3,4$, and set $$\nu=\bar{\epsilon}(\mu)=\nu_1-\nu_2+i(\nu_3-\nu_4).$$ Take $\delta > 0$. For each $j$, there exists Borel set $B_j$ in $\tilde {X}$ such that $\nu_j(B)\geq 0$ for each Borel subset $B$ of $B_j$ and $\sum_{j=1}^4\nu_j(B_j)>||\nu||-\delta$. In fact, by Hahn decomposition theorem for signed measures $\lambda_1=\nu_1-\nu_2$ and $\lambda_2=\nu_3-\nu_4$ there exist four Borel sets $P_1$, $P_2$, $N_1$ and $N_2$ in $\tilde {X}$ such that $$P_1\cup N_1=\tilde {X},\quad P_1\cap N_1=\emptyset,\quad P_2\cup N_2=\tilde {X},\quad P_2\cap N_2=\emptyset$$ and for every Borel set $E$ of $\tilde {X}$ we have, $$\nu_1(E)=\lambda_1(P_1\cap E),\ \nu_2(E)=-\lambda_1(N_1\cap E),\ \nu_3(E)=\lambda_2(P_2\cap E),\ \nu_4(E)=-\lambda_2(N_2\cap E).$$ that is $\nu_1,\nu_2,\nu_3,\nu_4$ are concentrated respectively on $P_1,N_1,P_2,N_2$.
Set $D_1:=P_1\cap N_2,D_2:=N_1\cap P_2, D_3:=P_2\cap P_1, D_4:=N_2\cap N_1$. Then the family $\{D_1,D_2,D_3,D_4\}$ is a partition of $\tilde {X}$. Also for $\delta > 0$ there is a compact set $K$ for which $$||\nu||-\delta\leq\sum_{j=1}^4||{\nu_j}_{|_{D_j}}||-\delta\leq\sum_{j=1}^4{\nu_j}_{|_{D_j}}(K)
=\sum_{j=1}^4{\nu_j}(D_j\cap K).$$ Set $B_j=D_j\cap K$. Then the sets $B_1,B_2,B_3,B_4$ are pairwise disjoint.
Set $C_j=(\epsilon)^{-1}(B_j)$, a Borel set in $S$. Then $(\mu_j\omega)(C_j)=\nu_j(B_j).$ Since $\epsilon$ is injection, the sets $C_1,C_2,C_3,C_4$ are pairwise disjoint, and so $$||\mu||_{\omega}\geq \sum_{j=1}^4|\mu\omega(C_j)|\geq \sum_{j=1}^4(\mu_j\omega)(C_j)=\sum_{j=1}^4\nu_j(B_j)>||\nu||-\delta$$ This holds for each $\delta> 0$, so $||\mu||_{\omega}\geq ||\nu||$. A similar argument shows that $||\mu||_{\omega}\leq ||\nu||$. Thus $||\mu||_{\omega}= ||\nu||$.
(2)$\Rightarrow$(1). Let $P(S,\omega)$ denote the subspace of all probability measures of $M_b(S,\omega)$ and $ext(P(S,\omega))$ the extreme points of unit ball of $P(S,\omega)$. Then $ext(P(S,\omega))=\{\frac{\delta_x}{\omega(x)}:x\in S\}\cong S$ and $ext(P(\tilde {X})\cong \tilde {X}$, see [@Conway p.151]. By injectivity of $\bar{\epsilon}$, it maps the extreme points of the unit ball onto the extreme points of the unit ball, thus $\epsilon:S\longrightarrow \tilde {X}$ is a one to one map.
(2)$\Rightarrow$(3). Since $\tilde {X}$ is compact, $M_b(\tilde {X})$ is a dual Banach algebra with respect to $C(\tilde {X})$, so it has an isometric representation $\psi$ on a reflexive Banach space $E$, see [@Da2]. In the following commutative diagram,
M\_b(S,) &\^[|]{} &M\_b()\
&\^ &\_&\
& &B(E)
If $\bar{\epsilon}$ is isometric, then so is $\phi$.
Thus $M_b(S,\omega)$ has an isometric representation on a reflexive Banach space $E$ if $\bar{\epsilon}$ is an isometric isomorphism. So $M_b(S,\omega)$ is a [WAP]{}-algebra if $\bar{\epsilon}$ is an isometric isomorphism.
(3)$\Rightarrow$(1). Let $M_b(S,\omega)$ be a [WAP]{}-algebra. Since $\ell_1(S,\omega)$ is a norm closed subalgebra of $M_b(S,\omega)$, then $\ell_1(S,\omega)$ is a [WAP]{}-algebra. Using the double limit criterion, it is a simple matter to check that $wap(S,1/\omega)=WAP(\ell_1(S,\omega))$ (see also [@Lashkarizadeh Theorem3.7]) where we treat $\ell^\infty(S,1/\omega)$ as an $\ell_1(S,\omega)$-bimodule. Then $\bar{\epsilon}:\ell_1(S,\omega)\longrightarrow wap(S,1/\omega)^*$ is an isometric isomorphism. Since $wap(S,1/\omega)$ is a $C^*$-algebra, as (2)$\Rightarrow$(1), $\epsilon:S\longrightarrow \tilde {X}$ is one to one.
The following statement are equivalent.
1. $\ell_1(S,\omega)$ is a [WAP]{}-algebra;
2. $M_b(S,\omega)$ is a [WAP]{}-algebra.
For $\omega=1$, it is clear that $\tilde{X}=S^{wap}$, and the map $\epsilon:S\longrightarrow S^{wap}$ is one to one if and only if $wap(S)$ separates the points of $S$, see [@BJM].
For a locally compact semi-topological semigroup $S$, the following statements are equivalent:
1. $M_b(S)$ is a [WAP]{}-algebra;
2. $\ell_1(S)$ is a [WAP]{}-algebra;
3. The evaluation map $\epsilon:S\longrightarrow S^{wap}$ is one to one;
4. $wap(S)$ separates the points of $S$.
Let $X$, $Y$ be sets and $f$ be a complex-valued function on $X\times Y$.
1. We say that $f$ is a cluster on $X\times Y$ if for each pair of sequences $(x_n)$, $(y_m)$ of distinct elements of $X,Y$, respectively $$\label{fg}
\lim_n\lim_mf(x_n,y_m)=\lim_m\lim_nf(x_n,y_m)$$ whenever both sides of (\[fg\]) exist.
2. If $f$ is cluster and both sides of \[fg\] are zero (respectively positive) in all cases, we say that $f$ is $0$-cluster(respectively positive cluster).
In general $\{f\omega:f\in wap(S)\}\not= wap(S,1/\omega)$. By using [@BR Lemma1.4] the following is immediate.
\[ghjk\]
Let $\Omega(x,y)=\frac{\omega(xy)}{\omega(x)\omega(y)}$, for $x,y\in S$. Then
1. If $\Omega$ is cluster, then $\{f\omega:f\in wap(S)\}\subseteq wap(S,1/\omega)$;
2. If $\Omega$ is positive cluster, then $wap(S,1/\omega)=\{f\omega:f\in wap(S)\}$.
It should be noted that if $M_b(S)$ is Arens regular (resp. dual Banach algebra) then $M_b(S,\omega)$ is so. We don’t know that if $M_b(S)$ is [WAP]{}-algebra, then $M_b(S,\omega)$ is so. The following Lemma give a partial answer to this question.
\[weighted\]
Let $S$ be a locally compact topological semigroup with a Borel measurable weight function $\omega$ such that $\Omega$ is cluster on $S\times S$.
1. If $M_b(S)$ is a [WAP]{}-algebra, then so is $M_b(S,\omega)$;
2. If $\ell_1(S)$ is a [WAP]{}-algebra, then so is $\ell_1(S,\omega)$.
\(1) Suppose that $M_b(S)$ is a [WAP]{}-algebra so $wap(S)$ separates the points of $S$. By lemma\[ghjk\] for every $f\in wap(S)$, $f\omega\in wap(S,1/\omega)$. Thus the evaluation map $\epsilon:S\longrightarrow \tilde {X}$ is one to one.
\(2) follows from (1).
\[homomo\]
For a locally compact semi-topological semigroup $S$,
1. If $C_0(S)\subseteq wap(S)$, then the measure algebra $M_b(S)$ is a [WAP]{}-algebra.
2. If $S$ is discrete and $c_0(S)\subseteq wap(S)$, then $\ell_1(S)$ is a [WAP]{}-algebra.
\(1) By [@BJM Corollary 4.2.13] the map $\epsilon:S\longrightarrow S^{wap}$ is one to one, thus $M_b(S)$ is a [WAP]{}-algebra.
\(2) follows from (1).
Dales, Lau and Strauss [@DLS Theorem 4.6, Proposition 8.3] showed that for a semigroup $S$, $\ell^1(S)$ is a dual Banach algebra with respect to $c_0(S)$ if and only if $S$ is weakly cancellative. If $S$ is left or right weakly cancellative semigroup, then $\ell^1(S)$ is a [WAP]{}-algebra. The next example shows that the converse is not true, in general.
Let $S=(\mathbb{N},\min)$ then $wap(S)=c_0(S) \oplus \mathbb{C} $. So $\ell^1(S)$ is a [WAP]{}-algebra but $S$ is neither left nor right weakly cancellative. In fact, for $f\in wap(S)$ and all sequences $\{a_n\}$, $\{b_m\}$ with distinct element in $S$, we have $\lim_mf(b_m)=\lim_m\lim_nf(a_nb_m)=\lambda=\lim_n\lim_mf(a_nb_m)=\lim_nf(a_n)$, for some $\lambda\in \mathbb{C}$. This means $f-\lambda\in c_0(S)$ and $wap(S)\subseteq c_0(S) \oplus \mathbb{C} $. The other inclusion is clear.
If $\{x_n\}$ and $\{y_m\}$ are sequences in $S$ we obtain an infinite matrix $\{x_ny_m\}$ which has $x_ny_m$ as its entry in the $m$th row and $n$th column. As in [@BR], a matrix is said to be of row type $C$ ( resp. column type $C$) if the rows ( resp. columns ) of the matrix are all constant and distinct. A matrix is of type $C$ if it is constant or of row or column type $C$.
J.W.Baker and A. Rejali in [@BR Theorem 2.7(v)] showed that $\ell^1(S)$ is Arens regular if and only if for each pair of sequences $\{x_n\}$, $\{y_m\}$ with distinct elements in $S$ there is a submatrix of $\{x_ny_m\}$ of type $C$.
A matrix $\{x_ny_m\}$ is said to be upper triangular constant if $x_ny_m=s$ if and only if $m\geq n$ and it is lower triangular constant if $x_ny_m=s$ if and only if $m\leq n$. A matrix $\{x_ny_m\}$ is said to be $W$-type if every submtrix of $\{x_ny_m\}$ is neither upper triangular constant nor lower triangular constant.
Let $S$ be a semigroup. The following statements are equivalent:
1. $c_0(S)\subseteq wap(S)$.
2. For each $s\in S$ and each pair $\{x_n\}$, $\{y_m\}$ of sequences in $S$, $$\{\chi_s(x_ny_m):n<m\}\cap \{\chi_s(x_ny_m):n>m\}\not=\emptyset;$$
3. For each pair $\{x_n\}$, $\{y_m\}$ of sequences in $S$ with distinct elements, $\{x_ny_m\}$ is a $W$-type matrix;
4. For every $s\in S$, every infinite set $B\subset S$ contains a finite subset $F$ such that $\cap\{sb^{-1}:b\in
F\}\setminus (\cap\{sb^{-1}: b\in B\setminus F\})$ and $\cap\{b^{-1}s:b\in F\}\setminus (\cap\{b^{-1}s: b\in B\setminus F\})$ are finite.
(1)$\Leftrightarrow$ (2). For all $s\in S$, $\chi_s\in wap(S)$ if and only if $$\{\chi_s(x_ny_m):n<m\}\cap \{\chi_s(x_ny_m):n>m\}\not=\emptyset.$$
(3)$\Rightarrow$ (1)Let $c_0(S)\not\subseteq wap(S)$ then there are sequences $\{x_n\}$, $\{y_m\}$ in $S$ with distinct elements such that for some $s\in S$, $$1=\lim_m\lim_n\chi_s(x_ny_m)\not=\lim_n\lim_m\chi_s(x_ny_m)=0.$$
Since $\lim_n\lim_m\chi_s(x_ny_m)=0$, for $1>\varepsilon>0$ there is a $N\in\mathbb{N}$ such that for all $n\geq N$, $\lim_m\chi_s(x_my_n)<\varepsilon$. This implies for all $n\geq N$, $\lim_m\chi_s(x_my_n)=0$. Then for $n\geq N$, $1>\varepsilon>0$ there is a $M_n\in\mathbb{N}$ such that for all $m\geq M_n$ we have $\chi_s(x_my_n)<\varepsilon$. So if we omit finitely many terms, for all $n\in \mathbb{N}$ there is $M_n\in\mathbb{N}$ such that for all $m\geq M_n$ we have $x_my_n\not=s$. As a similar argument, for all $m\in \mathbb{N}$ there is $N_m\in\mathbb{N}$ such that for all $n\geq N_m$, $x_my_n=s$.
Let $a_1=x_1$, $b_1$ be the first $y_n$ such that $a_1y_n=s$. Suppose $a_m, b_n$ have been chosen for $1\leq m,n<r$, so that $a_nb_m=s$ if and only if $n\geq m$. Pick $a_r$ to be the first $x_m$ not belonging to the finite set $\cup_{1\leq n\leq r}\{x_m:x_my_n=s\}$. Then $a_rb_n\not=s$ for $n<r$. Pick $b_r$ to be the first $y_n$ belonging to the cofinite set $\cap_{1\leq n\leq r}\{y_n:x_my_n=s\}$. Then $a_nb_m=s$ if and only if $n\geq m$. The sequences $(a_m)$, $(b_n)$ so constructed satisfy $a_mb_n=s$ if and only if $n\geq m$. That is, $\{a_nb_m\}$ is not of $W$-type and this is a contradiction.
(1)$\Rightarrow$ (3). Let there are sequences $\{x_n\}$, $\{y_m\}$ in $S$ such that $\{x_ny_m\}$ is not a $W$-type matrix, (say) $x_ny_m=s$ if and only if $m\leq n$. Then $$1=\lim_m\lim_n\chi_s(x_ny_m)\not=\lim_n\lim_m\chi_s(x_ny_m)=0.$$ So $\chi_s\not\in wap(S)$. Thus $c_0(S)\not\subseteq wap(S)$.
(4)$\Leftrightarrow$ (1)This is Ruppert criterion for $\chi_s\in wap(S)$, see [@Ruppert Theorem 4].
We conclude with some examples which show that some of the above results cannot be improved.
1. Let $S=\mathbb{N}$. Then for $S$ equipped with $\min$ multiplication, the semigroup algebra $\ell_1(S)$ is a WAP-algebra but is not neither Arens regular nor a dual Banach algebra. While, if we replace the $\min$ multiplication with $\max$ then $\ell_1(S)$ is a dual Banach algebra (so a WAP-algebra) which is not Arens regular. If we change the multiplication of $S$ to the zero multiplication then the resulted semigroup algebra is Arens regular (so a WAP-algebra) which is not a dual Banach algebra. This describes the interrelation between the concepts of being Arens regular algebra, dual Banach algebra and WAP-algebra.
2. Let $S$ be the set of all sequences with $0,1 $ values. We equip $S$ with coordinate wise multiplication. We denote by $e_n$ the sequence with all zero unless a $1$ in the $n$-th place. Let $s=\{x_n\}\in S$, and let $F_w(S)$ be the set of all elements of $S$ such that $x_i=0$ for only finitely index $i$. It is easy to see that $F_w(S)$ is countable. Let $F_w(S)=\{s_1,s_2,\cdots\}$ . Recall that, every element $g\in \ell^\infty(S)$ can be denoted by $g=\sum_{s\in S}g(s)\chi_s$, see [@DL p.65]. Suppose $$g=\sum_{s\in S\setminus F_w(S)}g(s)\chi_s$$ be in $wap(S)$, we show that $g=0$. Let $s=\{x_n\}\in S$, and $\{k\in
\mathbb{N}:x_k=0\}=\{k_1,k_2,\cdots\}$ be an infinite set. Put $a_n=s+\sum_{j=1}^n e_{k_j}$ and $b_m=s+\sum_{i=m}^{\infty}e_{k_i}$. Then $$a_nb_m=\left\{
\begin{array}{lr}
\sum_{j=m}^ne_{k_j}+s &\mbox{if}\quad m\leq n \\
s &\mbox{if}\quad m>n
\end{array}\right.$$ Thus $g(s)=\lim_n\lim_m g(a_nb_m)=\lim_m\lim_ng(a_nb_m)=\lim_mg(s+\sum_{i=m}^{\infty}e_{k_i})=0$.
In fact, $$wap(S)=\{f\in \ell^\infty(S):f=\sum_{i=1}^\infty f(s_i)\chi_{s_i},\quad s_i \in F_w(S)\}\oplus \mathbb{C}$$
It is clear that $F_w(S)$ is the subsemigroup of $S$ and $wap(F_w(S))=\ell^\infty(F_w(S))$. So $\ell^1(F_w(S))$ is Arens regular. Let $T$ consists of those sequences $s=\{x_n\}\in S$ such that $x_i=0$ for infinitely index $i$, then $T$ is a subsemigroup of $S$ and $wap(T)=\mathbb{C}$. Since $\epsilon_{|_T}: T\longrightarrow S^{wap}$ isn’t one to one, $\ell^1(S)$ is not a [WAP]{}-algebra. This shows that in general $\ell^1(S)$ need not be a [WAP]{}-algebra.
3. If we equip $S=\mathbb{ R}^2$ with the multiplication $(x,y).(x',y')=(xx',x'y+y')$, then $M_b(S)$ is not a WAP-algebra. Indeed, every non-constant function $f$ over $x$-axis is not in $wap(S)$. Let $f(0,z_1)\not=f(0,z_2)$ and $\{x_m\}$,$\{y_m\}$, $\{\beta_n\}$ be sequences with distinct elements satisfying the recursive equation $$\beta_nx_m+y_m=\frac{mz_1+nz_2}{m+n}$$ Then $$\begin{aligned}
\lim_n\lim_mf((0,\beta_n).(x_m,y_m))&=&\lim_n\lim_mf(0,\beta_nx_m+y_m)\\
&=&\lim_n\lim_mf(0,\frac{mz_1+nz_2}{m+n})\\
&=&f(0,z_1)\end{aligned}$$ and similarly $$\lim_m\lim_nf((0,\beta_n).(x_m,y_m))=f(0,z_2).$$
Thus the map $\epsilon:S\longrightarrow S^{wap}$ isn’t one to one, so $M_b(S)$ is not a WAP-algebra. This shows that in general $M_b(S)$ need not be a [WAP]{}-algebra.
4. Let $S$ be the interval $[\frac{1}{2},1]$ with multiplication $x.y=\max\{\frac{1}{2},xy\}$, where $xy$ is the ordinary multiplication on $\mathbb{R}$. Then for all $s\in S\setminus\{\frac{1}{2}\}$, $x\in S$, $x^{-1}s$ is finite. But $x^{-1}\frac{1}{2}=[\frac{1}{2},\frac{1}{2x}]$. Let $B=[\frac{1}{2},\frac{3}{4})$. Then for all finite subset $F$ of $B$, $$\bigcap_{x\in F}x^{-1}\frac{1}{2}\setminus \bigcap_{x\in B\setminus F}x^{-1}\frac{1}{2}=[\frac{2}{3},\frac{1}{2x_F}]$$ where $x_F=\max F$. By [@Ruppert Theorem 4] $\chi_{\frac{1}{2}}\not\in wap(S)$. So $c_0(S\setminus\{\frac{1}{2}\})\oplus \mathbb{C}\subsetneqq wap(S)$. It can be readily verified that $\epsilon:S\longrightarrow S^{wap}$ is one to one, so $\ell_1(S)$ is a WAP-algebra but $c_0(S)\not\subseteq wap(S)$. This is a counter example for the converse of Corollary \[homomo\].
5. Take $T=(\mathbb{N}\cup\{0\},.)$ with 0 as zero of $T$ and the multiplication defined by
$$n.m=\left\{
\begin{array}{lr}
n & \mbox{if}\quad n=m \\
0 &
\mbox{otherwise}.
\end{array}\right.$$
Then $S=T\times T$ is a semigroup with coordinate wise multiplication. Now let $X=\{(k,0): k\in T\}$, $Y=\{(0,k):k\in T\}$ and $Z=X\cup Y$. We use the Ruppert criterion [@Ruppert] to show that $\chi_z\not\in wap(S)$, for each $z\in Z$. Let $B=\{(k,n): k,n\in
T\}$, then $(k,n)^{-1}(k,0)=\{(k,m):m\not= n\}=B\setminus
\{(k,n)\}$. Thus for all finite subsets $F$ of $B$, $$\begin{aligned}
%\nonumber to remove numbering (before each equation)
\left(\cap\{(k,n)^{-1}(k,0): (k,n)\in F\}\right)&\setminus&(
\cap\{(k,n)^{-1}(k,0):
(k,n)\in B\setminus F\})\\
&=& \left(\cap\{(k,0)(k,n)^{-1}: (k,n)\in F\}\right ) \\
&\setminus &(\cap\{(k,n)^{-1}(k,0): (k,n)\in B\setminus F\}) \\
&=& (B\setminus F) \setminus F=B\setminus F\end{aligned}$$ and the last set is infinite. This means $\chi_{(k,0)}\not\in wap(S)$. Similarly $\chi_{(0,k)}\not\in wap(S)$. Let $f=\sum_{n=0}f(0,n)\chi_{(0,n)}+\sum_{m=1}^\infty f(m,0)\chi_{(m,0)}$ be in $ wap(S)$. For arbitrary fixed $n$ and sequence $\{(n,k)\}$ in $S$, we have $\lim_k f(n,k)=\lim_k\lim_lf(n,l.k)=\lim_l\lim_kf(n,l.k)=f(n,0)$ implies $f(n,0)=0$. Similarly $f(0,n)=0$ and $f(0,0)=0$. Thus $f=0$. In fact $wap(S)\subseteq \ell^\infty(\mathbb{N}\times\mathbb{N})$. Since $wap(S)$ can not separate the points of $S$ so $\ell_1(S)$ is not a WAP-algebra. Let $\omega(n,m)=2^n3^m$ for $(n,m)\in S$. Then $\omega$ is a weight on $S$ such that $\omega\in wap(S,1/\omega)$, so the evaluation map $\epsilon:S\longrightarrow \tilde {X}$ is one to one. This means $\ell_1(S,\omega)$ is a WAP-algebra but $\ell_1(S)$ is not a WAP-algebra. This is a counter example for the converse of Corollary \[weighted\].
[[**Acknowledgments.**]{}]{} This research was supported by the Center of Excellence for Mathematics at Isfahan university.
[99]{}
, Weighted semigroup algebras as dual Banach algebras, arXiv:0808.1404 \[math.FA\].
, On the Arens regularity of weighted convolution algebras, J. London Math. Soc. (2) **40**, 535-546, (1989).
Analysis on Semigroups, Wiley, New York, (1989).
, A Course in Functional Analysis, Springer-Verlag, Berlin, 1985.
, Banach Algebras and Automatic Continuity, Clarendon Press, Oxford, 2000.
, The Second Dual of Beurling Algebras, Mem. Amer. Math. Soc. **177** (2005), no. 836.
, Banach algebras on semigroups and on their compactification, Mem. Amer. Math. Soc. **205** (2010), 1-165.
, Connes-amenability of bidual and weighted semigroup algebras, Math. Scand. **99** (2006) no. 2, 217-246.
, Dual Banach algebras: Representation and injectivity, Studia Math, **178** (2007), 231-275.
, Conditions implying the uniqeness of the weak$^*-$topology on certain group algebras, Houston J. Math. **35** (2009), no. 1, 253�276.
, Function algebras on weighted topological semigroups, Math. Japonica **47**, no.2 (1998), 217-227.
, The analogue of weighted group algebra for semitopological semigroups, J. Sci, I.R. Iran, **6**, no. 2,(1995) 113-120.
, The Arens regularity of weighted semigroup algebras, Sci. Math. Japon, **60**, no.1 (2004), 129-137.
, Weighted convolution measure algebras characterized by convolution algebras, J. Sci. I. Iran, **18** (4), (2007) 345-349 .
, On weakly almost periodic sets, Semigroup Fourm **32**,no.3,(1985), 267-281.
, Dual Banach algebras: Connes-amenability, normal, virtual diagonals, and injectivity of the predual bimodule, Math. Scand. **95** (2004), 124-144.
, Periodicity of functionals and representations of normed algebras on reflexive spaces, Proc. Edinburgh Math. Soc.(2)[**20**]{} (1976/77), 99-120.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present CCD $uvby\beta$ photometry for stars in the nuclei of the young double cluster $h$ and $\chi$ Persei. We find that the reddening is highly variable through the $h$ Per nucleus, increasing from west to east, with values ranging from $E(b-y) = 0.328\pm0.022$ in the western part to $E(b-y) = 0.465\pm0.024$ in the south-east. Towards $\chi$ Per the reddening is fairly constant, with $E(b-y) =
0.398\pm0.025$. Both clusters share a common distance modulus of 11.7$\pm$0.1 mag., and an age of $\log t = 7.10\pm0.05$ years.
author:
- 'G.Capilla'
- 'J.Fabregat'
date: 'Received date; accepted date'
subtitle: 'I.The double cluster $h$ and $\chi$ Persei [^1]'
title: 'CCD $uvby\beta$ photometry of young open clusters'
---
Introduction
============
The precise determination of galactic open clusters main physical parameters plays a central role in the study of the stellar structure and evolution. With accurate photometric data, and once the external variables such as reddening are corrected for, the cluster distances, ages and chemical abundances can be inferred from the study of the photometric colour-magnitude and colour-colour diagrams.
The usual way to obtain the cluster age is by means of isochrone fitting to the main sequence in a colour-magnitude diagram. In the case of young clusters -younger than 50 Myr- isochrone fitting is made difficult by the usual presence of differential reddening across the cluster face, which widens the observed main sequence. Moreover, the presence of emission line stars, like Be o PMS stars, which occupy anomalous positions in the photometric diagrams, additionally contributes to a further main sequence widening. Hence, the fit of a particular isochrone can be a very uncertain process, and it is not difficult to find recent age determinations with very diverging values for a given young cluster.
In a recent paper, Fabregat & Torrejón ([@fabregat00]) propose the use of isochrone fitting in the $V_0 - c_0$ plane of the $uvby$ photometric system as an adequate tool to obtain accurate cluster ages. The range of variation of the $c_0$ index along the B-type sequence amounts more than 1 mag., being significantly larger than most common used photometric colours. Moreover, the $c_0$ index is less affected by reddening, and allows an efficient segregation of emission-line stars.
In order to produce accurate and homogeneous dating for a sample of young galactic clusters, we have undertaken an observational programme to obtain CCD $uvby\beta$ photometry. In this paper we present the first results, related to the clusters $h$ and $\chi$ Persei.
The double cluster and is one of the richest young open clusters in the Galaxy, and also one of the brightest and closest to us. On a clear night, far from the light pollution, it can be easily seen with a naked eye, distinctly shining in the Milky Way between Perseus and Cassiopeia constellations. $h$ and $\chi$ Persei ( and respectively) form the nucleus of the broader Per OB1 association.
The published work on $h$ and $\chi$ Persei is very extensive. A summary of the key papers and discussion on the past work is given by Waelkens et al. ([@waelkens]). Within the literature there is not convergence on the fundamental parameters of the clusters, such as their distances and ages. The discrepancies in the cluster distance moduli are well in excess of 0.5 mag., while there is not agreement about both clusters being of the same or different ages.
Two modern studies based on CCD photometry, presented by Keller et al. ([@keller]) and Marco & Bernabeu ([@marco]), converge in a common distance modulus of about 11.7 mag. for both clusters, but still present controversial results regarding their ages. The former finds a common age of log $t$ = 7.1 years for both clusters and the surrounding population, while the latter claims the existence of at least three different episodes of star formation. To contribute to ascertain this issue is one of the objectives of the present paper.
Observations and reduction procedure
====================================
Observations and image processing
---------------------------------
CCD photometry of the central regions of $h$ and $\chi$ Persei was obtained on the nights 20 to 22 November 1998 at the 1.52m. telescope of the Observatorio Astronómico Nacional, located at the Calar Alto Observatory (Almería, Spain). The chip employed was the Tektronics TK 1024 AB, with a size of 1024x1024 pixels. The 04 unbinned pixels provide a field size of 69x69, which almost entirely covers the clusters ’nuclei’ area, as defined by Maps 2 and 4 in Oosterhoff ([@oost]).
Observations were done through the four Strömgren $uvby$ and Crawford narrow and wide H$\beta$ filters, being every field sequentially measured through the six filters. Two different exposure times were used with each filter, in order to ensure a wide range of stellar magnitudes. Exposure times in each filter were selected so as a B type star produces approximately equal count rates through all filters. Employed exposure times are presented in Table \[t1\].
[lrr]{} Filter & short & large\
$y$ & 10 & 50\
$b$ & 12 & 60\
$v$ & 35 & 175\
$u$ & 120 & 600\
H$\beta_{w}$ & 12 & 60\
H$\beta_{n}$ & 30 & 150\
In order to ensure the atmospheric extinction and standard transformation determination, three additional fields centered on the open clusters , and were also observed at different airmasses. The list of all observations is presented in Table \[t2\].
[lcccl]{} NGC & JD & date & airmass & int. time\
869 & 51138 & 20-11-98 & 1.06 & short\
869 & 51138 & 20-11-98 & 1.07 & long\
869 & 51138 & 20-11-98 & 1.83 & short\
869 & 51139 & 21-11-98 & 1.07 & short\
869 & 51139 & 21-11-98 & 1.07 & long\
869 & 51140 & 22-11-98 & 1.07 & short\
869 & 51140 & 22-11-98 & 1.63 & short\
884 & 51138 & 20-11-98 & 1.10 & short\
884 & 51138 & 20-11-98 & 1.12 & long\
884 & 51138 & 20-11-98 & 1.98 & short\
884 & 51139 & 21-11-98 & 1.11 & short\
884 & 51139 & 21-11-98 & 1.13 & long\
884 & 51140 & 22-11-98 & 1.06 & short\
884 & 51140 & 22-11-98 & 1.70 & short\
1039 & 51138 & 20-11-98 & 1.57 & short\
1039 & 51140 & 22-11-98 & 1.55 & short\
6910 & 51138 & 20-11-98 & 1.11 & short\
6910 & 51138 & 20-11-98 & 1.77 & short\
6910 & 51138 & 20-11-98 & 1.14 & long\
6910 & 51138 & 20-11-98 & 1.83 & long\
6910 & 51140 & 22-11-98 & 1.14 & short\
6910 & 51140 & 22-11-98 & 1.66 & short\
6913 & 51138 & 20-11-98 & 1.25 & short\
6913 & 51138 & 20-11-98 & 1.30 & long\
6913 & 51140 & 22-11-98 & 1.18 & short\
6913 & 51140 & 22-11-98 & 1.81 & short\
Images were processed using IRAF.[^2] A sizeable sample of bias and sky flat frames were obtained at the beginning and at the end of every night. The images were subjected to the usual overscan, bias and flat field corrections.
Photometry was performed using the DAOPHOT package (Stetson [@stetson]). Aperture photometry was obtained for a number of sufficiently clean stars in each frame, through a constant 14 pixel radius which was chosen to contain virtually all the stellar flux in all images, as indicated by a grow curve analysis. PSF-fitting photometry was subsequently obtained for all identified stars in all frames. A constant PSF provided a good representation of the stellar profiles through each frame. The difference between aperture and PSF-based instrumental magnitudes were determined for each frame, and the latter were corrected accordingly.
Extinction and instrumental system
----------------------------------
The atmospheric extinction was determined by the multi-night, multi-star method described by Manfroid ([@manf93]). Computations were done by using the RANBO2 package, written by J. Manfroid. The implementation of this reduction procedure allows the construction of a consistent natural system, which contains the extra-atmospheric instrumental magnitudes of all constant stars included in the computation procedure. Stars from all observed fields were included in the building of the natural system.
The value of the extinction coefficient was determined for each individual frame. For the stars in common, the mean difference between the observed and natural magnitudes was obtained, and divided by the airmass to obtain the corresponding extinction coefficient.
[crrrrrrrrrrr]{} & & & & & & & & $D$ & & &\
star & $V$ & $(b-y)$ & $m_{1}$ & $c_{1}$ & $\beta$ & $V$ & $(b-y)$ & $m_{1}$ & $c_{1}$ & $\beta$ & N\
0869-0837 & 14.090 & - & - & - & - & -0.010 & - & - & - & - & 4\
0869-0843 & 9.317 & 0.286 & -0.065 & 0.172 & 2.593 & 0.003 & -0.009 & 0.015 & -0.006 & -0.026 & 2\
0869-0922 & - & 0.304 & -0.066 & 0.130 & - & - & 0.006 & -0.016 & 0.001 & - & 3\
0869-0963 & - & 0.286 & -0.044 & 0.191 & - & - & -0.003 & -0.010 & -0.003 & - & 5\
0869-0978 & - & 0.301 & -0.033 & 0.167 & 2.627 & - & 0.004 & -0.006 & 0.010 & -0.016 & 4\
0869-0980 & - & 0.284 & -0.042 & 0.178 & - & - & 0.006 & -0.012 & -0.011 & - & 3\
0869-0991 & - & 0.328 & -0.055 & 0.255 & - & - & 0.001 & -0.007 & 0.020 & - & 5\
0869-1004 & - & 0.305 & -0.051 & 0.204 & - & - & 0.013 & -0.010 & 0.010 & - & 5\
0869-1181 & - & 0.350 & -0.049 & 0.374 & 2.730 & - & 0.022 & 0.015 & 0.005 & 0.012 & 5\
0869-1187 & 10.857 & 0.334 & -0.046 & 0.205 & 2.618 & -0.037 & 0.014 & -0.017 & 0.007 & -0.030 & 5\
0884-2167 & 13.401 & 0.346 & -0.057 & 0.604 & 2.762 & -0.041 & 0.006 & 0.001 & 0.023 & 0.010 & 4\
0884-2196 & 11.538 & - & - & - & 2.666 & 0.032 & - & - & - & 0.004 & 5\
0884-2200 & - & - & - & - & 2.721 & - & - & - & - & 0.016 & 5\
0884-2232 & 11.102 & - & - & - & 2.639 & 0.008 & - & - & - & -0.012 & 4\
0884-2235 & 9.361 & 0.330 & -0.095 & 0.134 & 2.608 & -0.001 & -0.014 & 0.007 & 0.016 & 0.003 & 3\
0884-2246 & 9.930 & 0.291 & -0.058 & 0.124 & 2.625 & -0.030 & -0.004 & -0.004 & 0.005 & 0.002 & 4\
0884-2251 & 11.560 & 0.315 & -0.051 & 0.371 & 2.708 & 0.000 & -0.013 & 0.009 & -0.022 & 0.001 & 5\
0884-2296 & 8.499 & 0.323 & -0.104 & 0.136 & - & 0.031 & -0.032 & 0.032 & -0.003 & - & 1\
0884-2311 & 9.363 & 0.299 & -0.074 & 0.146 & 2.601 & 0.017 & -0.017 & 0.010 & 0.014 & -0.020 & 3\
0884-2330 & 11.446 & 0.273 & -0.070 & 0.268 & 2.630 & -0.026 & -0.006 & 0.020 & 0.009 & 0.010 & 5\
1039-0226 & 10.482 & - & - & - & - & -0.002 & - & - & - & - & 2\
1039-0267 & 11.940 & 0.296 & 0.133 & 0.505 & 2.683 & 0.020 & 0.007 & 0.005 & -0.024 & 0.005 & 2\
1039-0274 & 9.745 & 0.092 & 0.159 & 0.994 & 2.876 & -0.005 & -0.006 & 0.017 & -0.021 & -0.014 & 2\
1039-0278 & 11.789 & 0.146 & 0.202 & 0.911 & 2.857 & 0.031 & -0.002 & -0.004 & -0.011 & 0.002 & 2\
1039-0284 & 10.741 & 0.151 & 0.185 & 0.898 & 2.862 & 0.019 & 0.000 & 0.009 & -0.004 & 0.014 & 2\
1039-0294 & 11.185 & 0.186 & 0.195 & 0.782 & - & 0.035 & -0.010 & 0.009 & 0.014 & - & 2\
1039-0301 & 10.041 & 0.040 & 0.185 & 1.014 & 2.887 & -0.011 & 0.026 & -0.033 & -0.001 & -0.029 & 2\
1039-0303 & 9.951 & 0.055 & 0.169 & 1.023 & - & -0.001 & 0.000 & -0.006 & -0.002 & - & 2\
6910-0010 & 10.782 & - & - & - & 2.941 & -0.032 & - & - & - & 0.032 & 5\
6910-0021 & 11.763 & 0.582 & -0.099 & 0.234 & 2.678 & -0.033 & 0.008 & -0.001 & -0.014 & 0.026 & 5\
6910-0024 & 11.710 & - & - & - & 2.637 & 0.010 & - & - & - & -0.010 & 5\
6910-0028 & 12.243 & - & - & - & - & -0.023 & - & - & - & - & 5\
6910-0041 & 12.803 & 0.759 & -0.166 & 0.319 & 2.674 & 0.007 & -0.009 & 0.026 & -0.029 & 0.006 & 5\
6913-0001 & 8.842 & 0.726 & -0.159 & 0.166 & 2.596 & 0.018 & 0.004 & -0.011 & 0.004 & 0.007 & 3\
6913-0002 & 8.912 & 0.644 & -0.148 & 0.112 & 2.625 & -0.002 & -0.014 & 0.008 & 0.008 & 0.029 & 3\
6913-0003 & 8.942 & 0.699 & -0.186 & 0.165 & 2.639 & 0.038 & -0.019 & 0.026 & -0.035 & 0.045 & 3\
6913-0004 & 10.199 & 0.621 & -0.115 & 0.157 & 2.632 & -0.019 & -0.001 & -0.005 & 0.013 & 0.015 & 4\
6913-0008 & 12.190 & 0.546 & -0.062 & 0.365 & 2.618 & -0.020 & 0.024 & -0.028 & -0.025 & -0.044 & 4\
6913-0009 & 11.747 & 0.587 & -0.069 & 0.297 & - & 0.033 & 0.033 & -0.041 & 0.013 & - & 4\
6913-0010 & - & - & - & - & 2.678 & - & - & -& - & 0.015 & 2\
6913-0026 & 13.230 & 0.694 & -0.102 & 0.778 & - & -0.010 & -0.014 & 0.002 & 0.042 & - & 4\
& & & & &Mean: &0.000 &0.000 &0.000 &0.000 &0.001 &\
& & & & &RMS: &0.023 &0.014 &0.017 &0.017 &0.021 &\
$uvby\beta$ transformation
--------------------------
The choice of an adequate set of $uvby\beta$ standard stars for CCD photometry is a very critical issue. By one hand, the primary standard stars of the $uvby$ and H$\beta$ systems (Crawford & Mander [@craw66]; Crawford & Barnes [@craw70a]; Perry et al. [@perry87]) are all bright enough to saturate the CCD chip, even with short exposures. By the other hand, most of the observed stars are reddened B type stars. Manfroid & Sterken ([@manf87]), Delgado & Alfaro ([@delgado]) and Crawford ([@craw94]) have shown that transformations made only with unreddened stars introduce large systematic errors when applied to reddened stars, even if the colour range of the standards brackets that of the programme stars. No such reddened early type stars are included in the primary $uvby\beta$ standard lists.
Our standard list was composed by stars in young open clusters with $uvby\beta$ photometry published by Crawford et al. ([@craw70b]) for $h$ & $\chi$ Persei, Canterna et al. ([@canterna]) for NGC 1039 and Crawford et al. ([@craw77]) for NGC 6910 and NGC 6913. All these photometric lists were obtained with the same telescopes, instrumentation and reduction procedures used to define the standard Crawford & Barnes ([@craw66]) and Crawford & Mander ([@craw70a]) systems, and so the photometric values are in the standard system. As the $V$ values for $h$ and $\chi$ Persei are not included in the corresponding $uvby$ photometric list, we have used the values given by Johnson & Morgan ([@johnson]). The final standard star list is presented in Table \[t3\]. For cluster stars we have adopted the numbering system in the WEBDA database[^3] (Mermilliod [@merm]). Note that these numbers may not correspond with the numbering in the referred to photometric papers.
Individual stars in each list were selected so as to insure as large as possible a range in all photometric indices. However, as the main purpose of this paper is to study the clusters upper main sequence, and in particular the B type spectral range, no late type stars were included in the standard star list. A few intrinsically red field stars have been observed, and their values have been included in the photometry tables for completeness, but it should be noted that photometry for stars redder than $(b-y) \sim 0.750$ is likely to be affected by systematic errors. This is also the case for the $\beta$ index of emission-line stars, which reach a very low value. As no emission line star was included in the standard list due to their well know variability, there is no way to avoid extrapolation.
Transformations were computed from the natural system to the standard $uvby\beta$ system defined by the standard star list described above. The obtained transformation equations are the following:
> $V = -4.793 - 0.023 (b-y)_n + y_n$\
> $(b-y) = -0.413 + 0.977 (b-y)_n$\
> $m_1 = 0.063 + 1.044 m_{1,n} - 0.013 (b-y)$\
> $c_1 = 0.598 + 0.999 c_{1,n} + 0.135 (b-y)$\
> $\beta = -0.079 + 1.692 \beta_n$\
were subscript ’n’ refers to the natural system. All scale coefficients in the $uvby$ transformation are close to unit, while the colour terms are small, indicating a good conformity between the instrumental and standard photometric systems. This is not the case for the H$\beta$ transformation, where the scale coefficient is larger. This may be due to the wide filter used being significantly narrower than the standard one defined by Crawford & Mander ([@craw66]).
A measure of the photometric accuracy is the standard deviation of the mean catalogue minus transformed values for the standard stars. These values are presented in Table \[t3\], bottom line.
Coordinates
-----------
Although precise astrometry is not among the scopes of this paper, we have transformed the instrumental pixel coordinates into equatorial coordinates, in order to facilitate the identification of all observed stars, and their cross-correlation with other photometric lists.
Astronomical coordinates in the field of $h$ Per where derived from 13 stars with positions in the PPM Catalogue (Roeser & Bastian [@roeser]). In the field of $\chi$ Per there are only 7 stars with positions in the PPM Catalogue, which are grouped in the eastern part. We complemented this list with four more stars with positions given by Abad & García ([@abad]). Transformation equations were computed by means of the [*Starlink*]{} program ASTROM (Wallace [@wallace]). The final astrometric accuracy, measured as the RMS of the mean catalogue minus transformed values for the stars used in the transformation, is within 04.
The data
--------
Equatorial coordinates and mean photometric magnitudes, colours and indices for stars in the $h$ and $\chi$ Persei nuclei are presented in Tables 6 and 7. Only frames obtained at airmasses lower than 1.5 have been considered to compute the final photometric values. For this reason, small differences up to a few millimagnitudes may appear between values in Table \[t3\] and Tables 6 and 7 for the stars in common.
In the photometric tables we have also adopted the cluster star numbering system from the WEBDA database. For cluster numbers lower than 3000, WEBDA numbers are coincident with Oosterhoff ([@oost]) numbers, which we will refer to as ’Oo’ hereinafter. A few observed stars have not entry in WEBDA. We have introduced new numbering for them, starting with 7000 in $h$ Per and with 8000 in $\chi$ Per.
Reddening, intrinsic colours and distance.
==========================================
Colour-magnitude diagrams of all observed stars in both cluster regions are presented in Fig. \[f1\], and photometric $V-c_1$ diagrams in Fig. \[f2\]. To obtain the intrinsic colours we have first classified the stars as belonging to the early (earlier than A0), intermediate (A0-A3) and late (A3 onwards) groups defined by Strömgren ([@strom]). The classification was performed by means of the algorithm described by Figueras et al. ([@figueras]). Cluster membership were assigned from the position of each star in the different $uvby\beta$ photometric diagrams. Stars considered as members were marked in the last column of Tables 6 and 7.
Reddening values and intrinsic colours and indices were obtained for stars in the early group by means of the procedure described by Crawford ([@craw78]). We have used the standard $(b-y)_0 - c_0$ relation given in Table X of Perry et al. ([@perry87]), and the following reddening relations:
> $A_V = 4.3 E(b-y)$\
> $E(c_1) = 0.2 E(b-y)$\
> $E(m_1) = -0.32 E(b-y)$\
Known Be stars have not been included in the computation of the interstellar reddening, as they present an additional reddening contribution of circumstellar origin.
For $h$ Per we obtained a mean reddening value of $E(b-y) = 0.420 \pm
0.045$, from 127 stars. The large value of the standard deviation indicates the presence of variable reddening across the cluster nucleus. In Fig. \[f3\] we have represented the reddening values for individual stars as a function of their position. It is apparent a trend of increasing reddening from west to east, and a heavily obscured region in the south-eastern part of the cluster center. In Fig. \[f3\] we have divided the cluster nucleus area in three regions of different reddening. Mean reddening values in these regions are presented in Table \[t6\].
The large absorption in the south-east part has been previously pointed out by Waelkens et al. ([@waelkens]) from photoelectric photometry in the Geneva system, but surprisingly has not been accounted for in recent CCD studies. On the other hand, the low reddening value in the western part was already noted by Fabregat et al. ([@fabregat96]). They considered the Be stars , and as probable nonmembers, due to their reddening values ($E(b-y) =$ 0.313, 0.346 and 0.308 respectively), which are much lower than the cluster average reddening. From the present result we can conclude that the low reddening values are consistent with the position of these stars in the low absorption region at the western part of the cluster.
The mean value of the reddening for $\chi$ Per is $E(b-y) = 0.398 \pm
0.025$, from 82 stars. In Fig. \[f4\] we have also represented the individual reddening values as a function of the stars position. In this case no trend is present. The standard deviation of the mean reddening value is only slightly larger than the deviation of the reddening computation method (namely 0.018, cf. Perry & Johnston [@perry82]; Franco [@franco]), and we conclude that the interstellar absorption is constant towards the $\chi$ Per nucleus, within the accuracy of our photometry.
The different reddening characteristics towards both cluster nuclei are apparent in the photometric diagrams presented in Figs. \[f1\] and \[f2\]. In the $V-(b-y)$ plane, the $h$ Per sequence is much broader than the $\chi$ Per one, due to the strongly variable reddening in the former. Assuming that both clusters are at the same distance (see below), the higher mean reddening in $h$ Per causes that its observed sequence ends at earlier spectral types. This is apparent in Fig. \[f1\], but much more remarkable in the $V-c_1$ plane in Fig. \[f2\]: in $\chi$ Per the lower branch of the cluster sequence is almost complete until $c_1 =
0.45$ (spectral type around F5), while in $h$ Per it ends at $c_1 = 0.85$ (around A7).
[ccc]{} Region & $E(b-y)$ & stars\
A & $0.328 \pm 0.022$ & 12\
B & $0.414 \pm 0.026$ & 81\
C & $0.465 \pm 0.024$ & 34\
In Fig. \[f5\] we present the intrinsic $V_0 - c_0$ diagram for B stars in both cluster nuclei. $\chi$ Per stars have been dereddened by using the mean cluster value of $E(b-y) = 0.398$. For $h$ Per, each star has been dereddened on the basis of its position within Fig. \[f3\], and using the reddening values in Table \[t6\]. The loci of both sequences are undistingishable, indicating that both clusters are at the same distance, and have also the same age, as we will discuss in the next section. The sequence of $h$ Per is somewhat broader, most likely due to the variable reddening.
To obtain the distance we have fitted to the $V_0 - c_0$ diagram of each cluster the ZAMS as presented in Table X of Perry et al. ([@perry87]). Results are shown in Fig. \[f6\]. We found the best fit at a distance modulus of 11.7 mag. To estimate the error of this determination we have also represented in Fig. \[f6\], as dotted lines, the ZAMS shifted by distance moduli of 11.5 and 11.9 respectively. We find that the 11.7 value produces a distinctly better fit that the two latter ones, and hence we give the value of $11.7 \pm 0.1$ mag. as the distance modulus of $h$ and $\chi$ Persei. This value is in good agreement with the recent determinations based on CCD photometry by Keller et al. ([@keller]) and Marco & Bernabeu ([@marco]).
Finally, an indication of the validity of the dereddening procedure and the conformity between our photometry and the standard system can be obtained by comparing the resulting $m_0$ and $(u-b)_0$ indices with those of nearby, unreddened stars as given by Perry et al. ([@perry87]). The results are shown in Figs. \[f7\] and \[f8\]. As it can be seen, the obtained indices are coincident with the mean intrinsic relations, and hence we can conclude that our photometry is in the standard system, and free of systematic effects. This is in turn a proof of the validity of our reduction methods and the adequacy of the standard star selection.
Cluster ages
============
Determination of the cluster ages will be done by isochrone fitting to the upper main sequence. In the $uvby$ system and for stars in the early group, the $(b-y)_0$ colour and the $c_0$ index are temperature indicators, and hence both $V_0 - (b-y)_0$ and $V_0 - c_0$ planes are observational HR diagrams. Following the discussion in Fabregat & Torrejón ([@fabregat00]) we consider more precise and reliable the isochrone fitting to the $V_0 - c_0$ diagram, for the following reasons: i./ the range of variation of the $c_0$ index along the B-type sequence is more than ten times larger than the range of variation of $(b-y)_0$, providing much better discrimination between isochrones of similar ages; ii./ the $c_0$ index is less affected, by a factor of 5, by interstellar reddening; iii./ the $V_0 - c_0$ plane allows an efficient segregation of emission line stars.
In Fig. \[f9\] we present the $V_0 - c_0$ sequences of $h$ and $\chi$ Per, together with isochrones with ages of $\log t =$ 7.0, 7.1 and 7.2 years. The isochrones have been computed with the evolutionary models of Schaller at al. ([@schaller]), and transformed to the observational plane by means of the relations obtained by Torrejón ([@torrejon]). The best fitting in both clusters is obtained for $\log t =$ 7.1, with most of the stars at the turnoff point lying between the $\log t =$ 7.0 and 7.2 isochrones. Hence we propose $\log t = 7.10\pm0.05$ as the common age of both clusters, in good agreement with the recent result of Keller et al. ([@keller]).
Discussion
==========
In a recent paper Marco & Bernabeu ([@marco], hereinafter referred to as MB01) claimed that $h$ Per is significantly older than $\chi$ Per, and go further to propose three epochs of star formation within the double cluster at ages of log $t$ = 7.0, 7.15 and 7.3 years. This is in clear disagreement with our results in the previous section. Their study is also based on CCD $uvby$ photometry, and they use the same methods and techniques that we do in the present paper. Hence, it is worth to compare both studies in order to ascertain the reasons of the diverging results. When comparing both sets of data, it should be noted that: i./ our standard photometry is at least twice as accurate as the MB01 one (cf. Table 9 in MB01 and Table \[t3\] in this paper), and ii./our CCD frames cover a larger area than MB01 ones, and hence our photometric sequences are more populated.
MB01 find an age of log $t$ = 7.10 – 7.15 years for $\chi$ Per, in agreement with our value of log $t$ = 7.10 obtained in the previous section. For $h$ Per they obtain a significantly older age. In their Fig. 10 they show the log $t$ = 7.3 isochrone fitting a clump of stars at about $c_0 = 0.12$, which they consider as the main sequence turnoff. This clump is also present in our data, as can be seen in Fig. \[f10\] where we present the $V_0 - c_0$ plane and the the log $t$ = 7.1 and 7.3 isochrones. However, the distribution of stars in our Fig. \[f10\] near the turnoff point is different. The clump at $c_0 = 0.12$ is less conspicuous, it is only formed for at most six stars, and significantly leftwards to the fitting log $t$ = 7.3 isochrone there are at least eight main sequence cluster stars, which should have a younger age. We consider the latter as defining the actual cluster main sequence turnoff, which is well fitted by the log $t$ = 7.1 isochrone.
The possibility of two distinct populations with log $t$ = 7.1 and 7.3 years seems very unlikely. The occurrence of stars slightly above the main sequence can be justified by several reasons, including binarity, high rotational velocity, or photometric errors. Other small clump above the main sequence is apparent at $c_0 = 0.3$, and it is also most likely caused by the above referred to reasons.
MB01 also propose the presence of a population younger than log $t$ = 7.0 in the $h$ Per region. This claim is based on the position in the $V_0 -
c_0$ plane of six supergiant B stars observed by Crawford et al. ([@craw70b]). In Fig. \[f10\] we have also represented these stars (filled triangles), which clearly fall at the left of the log $t$ = 7.1 isochrone.
But this argument is wrong. The $c_1$ index of the $uvby$ system is defined to be a measure of the Balmer discontinuity depth, which for OB stars is related to the effective temperature. For this reason, the $V_0 - c_0$ photometric plane can be considered as an observational Hertzsprung-Russell diagram. However, this is only true for stars of luminosity classes III to V, and not for supergiants. Crawford et al. ([@craw70b]) already note that Ia type supergiants do not follow the mean $(b-y)_0 - c_0$ calibration. Schild & Chaffee ([@schild75]) show that the Balmer continuum of early type supergiants is affected by an emission or absorption contribution of circumstellar origin, and hence the Balmer discontinuity depth is not correlated with spectral types or effective temperature. Kilkenny & Whittet ([@kilkenny]) present $(b-y)_0 - c_0$ intrinsic relations for early type supergiants. There is a different relation for each luminosity class (Ia, Iab, Ib and II), and all of them are in turn different to the mean relation for III-V classes. If we assume that $(b-y)_0$ is a good temperature indicator for all luminosity classes (see below), this implies that the $c_0$ index cannot define a unique temperature scale valid for all luminosities.
The case is much the same as for classical Be stars. Be stars also have continuum emission or absorption of circumstellar origin at the Balmer discontinuity, and hence their $c_0$ indices are anomalous (Fabregat et al. [@fabregat96], and references therein). For comparison, in Fig. \[f10\] we have represented the four Be stars in the cluster nucleus (filled circles). They also occupy anomalous positions at the left of the isochrones, and even below the ZAMS. Neither supergiants nor Be stars should be used in a $V_0 - c_0$ diagram when it is intended to be an observational HR diagram for isochrone fitting purposes, due that their $c_0$ index is not related with effective temperature.
In order to study the age of the supergiant star population in the $h$ Per area, in Fig. \[f11\] we have represented the $V_0 - (b-y)_0$ photometric diagram. $(b-y)$ is a measure of the Paschen continuum slope, which is correlated with effective temperature for stars of all luminosity classes. Be stars, however, also deviates due to additional reddening of circumstellar origin (eg. Fabregat et al. [@fabregat96]). Two supergiant stars, and , lie in the cluster nucleus, and have been dereddened as a function of their position in Fig. \[f3\]. The other four, , , and , are placed at the west of the cluster nucleus. Following the discussion in Sect. 3, we have used the value of $E(b-y) = 0.328$ in Table \[t6\] to obtain their intrinsic colours. In order to check the reliability of this value, we have calculated the individual reddening for all supergiant stars using the standard relations given by Kilkenny & Whittet ([@kilkenny]). Results are presented in Table \[t7\]. All reddening values for the western stars are compatible with the mean reddening for the west region given in Table \[t6\].
[rc]{} Oo & $E(b-y)$\
3 & 0.306\
16 & 0.340\
612 & 0.350\
662 & 0.357\
1057 & 0.411\
1162 & 0.473\
In Fig. 11 it is apparent that most of the supergiant stars lie close to the log $t$ = 7.1 isochrone, giving further support to this value as the actual age of $h$ Per. There are no stars significantly leftwards to this isochrone, indicating that no younger population is present in the cluster area.
MB01 also presented the $V_0 - (b-y)_0$ plane for $h$ Per (their Fig. 9). In their figure most supergiant stars are placed leftwards to the log $t$ = 7.0 isochrone, and even leftwards to the ZAMS, indicating that these stars could be no members of the cluster. This is due to a reddening overcorrection. They used their mean cluster value, $E(b-y) = 0.44$, to calculate the intrinsic colours. The deviating supergiants are placed at the west of the cluster nucleus, where we have demonstrated that the reddening is lower by more than a tenth of magnitude. With the proper reddening values supergiants are placed close to the cluster isochrone.
We can conclude that there are no different epochs of star formation within the $h$ and $\chi$ Persei clusters area, and that both clusters and the surrounding field stars share the same age of $\log t =
7.10\pm0.05$ years, as determined in the previous sections. We propose that the wrong results obtained by MB01 are due to a number of factors, including the low accuracy of their photometry, the neglecting of the strongly variable reddening across the $h$ Per cluster, the incorrect use of supergiant stars in the $V_0 - c_0$ plane for isochrone analysis, and to a large extent the overinterpretation of the data at their disposal.
As discussed in Sect. 4, and in more detail in Fabregat & Torrejón ([@fabregat00]), the isochrone fitting in the $V_0 - c_0$ plane is an excellent tool for accurate age determination in young open clusters. With accurate photometry, the main sequence turnoff is very well defined, allowing a precise discrimination between isochrones of slightly different ages. But its applicability is limited to the range of spectral types and luminosity classes for which the $c_0$ index is a good temperature indicator. This is restricted to OB stars of luminosity classes III to V, and beyond this limits its use can conduct to misleading results.
Conclusions
===========
We have presented CCD $uvby\beta$ photometry for stars in the nuclei of the young open clusters $h$ and $\chi$ Per. We have shown that our photometry is free of systematic effects and well tied to the standard $uvby$ system.
We have obtained the cluster astrophysical parameters from the analysis of the B type star range. The reddening is highly variable through the $h$ Per nucleus, increasing from west to east. Its value ranges from $E(b-y) = 0.328\pm0.022$ in the western part to $E(b-y) = 0.465\pm0.024$ in the south-east. Towards $\chi$ Per the reddening is fairly constant, with $E(b-y) =
0.398\pm0.025$. Both clusters share a common distance modulus of 11.7$\pm$0.1 mag., and an age of $\log t = 7.10\pm0.05$ years.
We would like to thank Dr. J. Manfroid for providing us with his RANBO2 code to compute atmospheric extinction and photometric natural system. We are grateful to the Observatorio Astronómico National for the allocation of observing time in the 1.5m. telescope, and for support during observations. This research has made use of the WEBDA database, developed and maintained by J.C. Mermilliod, the SIMBAD database, operated at CDS, Strasbourg, France, and the NASA’s Astrophysics Data System Abstract Service. The authors acknowledge the data analysis facilities provided by the IRAF data reduction and analysis system, and by the Starlink Project which is run by CCLRC on behalf of PPARC. This work has been partially supported by the [*Plan Nacional de Investigación Científica, Desarrollo e Innovación Tecnológica del Ministerio de Ciencia y Tecnología*]{} and FEDER, through contract AYA2000-1581-C02-01. JF acknowledges grants from the [*Conselleria de Cultura i Educació de la Generalitat Valenciana*]{} and the [*Secretaría de Estado de Educación y Universidades*]{} of the Spanish Governement.
Abad, C., García, L. 1995, Rev. Mex. Astron. Astrophys. 31, 15
Canterna, R., Perry, C.L., Crawford, D.L. 1979, PASP 91, 263
Crawford, D.L. 1978, AJ 83,48
Crawford, D.L. 1994, PASP 106, 397
Crawford, D.L., Mander, J. 1966, AJ 71, 144
Crawford, D.L., Barnes, J.V. 1970, AJ 75, 978
Crawford, D.L, Glaspey, J.W., Perry, C.L. 1970, AJ 75, 822
Crawford, D.L., Barnes, J.V., Hill G. 1977, AJ 82, 606
Delgado, A.J., Alfaro, E.J. 1989, A&A 219, 121
Fabregat, J., Torrejón, J.M., 2000, A&A 357, 451
Fabregat, J., Torrejón, J.M., Reig, P. et al. 1996, A&AS 119, 271
Figueras, J., Torra, J., Jordi, C. 1991, A&AS 87, 319
Franco, G.A.P. 1989, A&AS 78, 105
Johnson, H.L., Morgan, W.W. 1955, ApJ 122, 429
Keller, S.C., Grebel, E.K., Miller, G.J., Yoss, K.M. 2001, AJ 122, 248
Kilkenny, D., Whittet, D.C.B. 1985, MNRAS 216, 127
Manfroid, J. 1993, A&A 271, 714
Manfroid, J., Sterken, C. 1987, A&AS 71, 539
Marco, A., Bernabeu, G. 2001, A&A 372, 477
Mermilliod, J.C. 1999, in Very Low-Mass Stars and Brown Dwarfs in Stellar Clusters and Associations, Cambridge Univ. Press, eds. Rebolo, R. and Zapatero-Osorio, R.M.
Oosterhoff, P.T. 1937, Ann. Sterrewatch Leiden 17, 1
Perry, C.L., Johnston, L. 1982, ApJS 50, 451
Perry, C.L., Olsen, E.H., Crawford, D.L. 1987, PASP 99, 1184
Roeser, S., Bastian, U. 1988, A&AS 74, 449
Schaller, G., Schaerer, D., Meynet, G., Maeder, A. 1992, A&A 96, 269
Schild, R.E., Chaffee, F.H. 1975, ApJ 196, 503
Stetson, P.R. 1987, PASP 99, 191
Strömgrem, B., 1966, ARA&A 4, 433
Torrejón, J.M. 1997, PhD Thesis, University of Valencia
Waelkens, C., Lampens, P., Heynderickx, D. et al. 1990, A&AS 83, 11
Wallace, P.T. 1998, Starlink User Note 5.17, Rutherford Appleton Laboratory
[^1]: Tables 6 and 7 will be only available in electronic form at the CDS. Now are available at http://pleione.uv.es/recent.html
[^2]: IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation, U.S.A.
[^3]: http://obswww.unige.ch/webda/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present an $\tO{m^{\frac{10}{7}}U^{\frac{1}{7}}}$-time algorithm for the maximum $s$-$t$ flow problem (and the minimum $s$-$t$ cut problem) in directed graphs with $m$ arcs and largest integer capacity $U$. This matches the running time of the $\tO{(mU)^{\frac{10}{7}}}$-time algorithm of Mdry [@Madry13] in the unit-capacity case, and improves over it, as well as over the $\tO{m
\sqrt{n} \log U}$-time algorithm of Lee and Sidford [@LeeS14], whenever $U$ is moderately large and the graph is sufficiently sparse.
By well-known reductions, this also gives us an $\tO{m^{\frac{10}{7}}B^{\frac{1}{7}}}$-time algorithm for the maximum-cardinality bipartite $\bb$-matching problem in which the largest integer demand is $B$. This, again, matches the $\tO{(mB)^{\frac{10}{7}}}$-time algorithm of Mdry [@Madry13], when $B=1$, which corresponds to the maximum-cardinality bipartite matching problem, and outperforms it, as well as the $\tO{m
\sqrt{n} \log B}$-time algorithm of Lee and Sidford [@LeeS14], for moderate values of $B$ and sufficiently sparse graphs.
One of the advantages of our algorithm is that it is significantly simpler than the ones presented in [@Madry13] and [@LeeS14]. In particular, these algorithms employ a sophisticated interior-point method framework, while our algorithm is cast directly in the classic augmenting path setting that almost all the combinatorial maximum flow algorithms use. At a high level, the presented algorithm takes a primal dual approach in which each iteration uses electrical flows computations both to find an augmenting $s$-$t$ flow in the current residual graph and to update the dual solution. We show that by maintain certain careful coupling of these primal and dual solutions we are always guaranteed to make significant progress.
author:
- |
Aleksander Mdry[^1]\
[MIT]{}\
[ [email protected]]{}
bibliography:
- '../main.bib'
title: Computing Maximum Flow with Augmenting Electrical Flows
---
=10000 = 10000
[^1]: Supported by NSF grant CCF-1553428, and a Sloan Research Fellowship.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the thermal response of the atmosphere of a solar-type star to an electron beam injected from a hot Jupiter by performing a 1-dimensional magnetohydrodynamic numerical experiment with non-linear wave dissipation, radiative cooling, and thermal conduction. In our experiment, the stellar atmosphere is non-rotating and is modelled as a 1-D open flux tube expanding super-radially from the stellar photosphere to the planet. An electron beam is assumed to be generated from the reconnection site of the planet’s magnetosphere. The effects of the electron beam are then implemented in our simulation as dissipation of the beam momentum and energy at the base of the corona where the Coulomb collisions become effective. When the sufficient energy is supplied by the electron beam, a warm region forms in the chromosphere. This warm region greatly enhances the radiative fluxes corresponding to the temperature of the chromosphere and transition region. The warm region can also intermittently contribute to the radiative flux associated with the coronal temperature due to the thermal instability. However, owing to the small area of the heating spot, the total luminosity of the beam-induced chromospheric radiation is several orders of magnitude smaller than the observed Ca II emissions from HD 179949.'
author:
- 'Pin-Gao Gu'
- 'Takeru K. Suzuki'
title: |
Thermal Response of A Solar-like Atmosphere to\
An Electron Beam from A Hot Jupiter: A Numerical Experiment
---
Introduction
============
Hot Jupiters are Jupiter-mass planets located within $\sim$ 0.1 AU or less from their parent stars. Because of the close proximity to the parent stars, hot Jupiters have been expected to be able to influence their stellar companions via magnetic [@Cuntz; @RS00] and/or tidal interactions [@Lin96; @Jackson09; @Pfahl08]. The observations of Ca II H & K lines from a number of stars harboring a hot Jupiter have suggested that the chromospheric activities, characterized either by line intensity or short-time variability, sometimes correlate with the orbits of their planets [@Shk03; @Shk05; @Shk08]. These phenomena can be modelled as a hot spot or a more “variable" region, despite residing in the chromosphere, following the planet’s orbital motion with a phase difference. In particular, the observations carried out in 2001, 2002, and 2005 imply that a hot spot on HD 179949 persistently leads the planet by $\sim 70^\circ$ with the intensity of $\sim 10^{27}$ erg/s in Ca II emissions. The similar phase lead of a variable region in optical has been suggested by the MOST satellite photometry for the hot-Jupiter host star $\tau$ Bootis [@Walker08]. Since the planet-induced stellar activities occur only once during one orbital period, the origins of this “spot" have been attributed to magnetic rather than tidal interactions.
One of the commonly adopted scenarios to describe the star-planet magnetic interactions is the magnetic interactions between Jupiter and its Galilean satellites (see Zarka 2007 for a review). In this scenario, the orbital motion of the Galilean satellites relative to Jupiter’s magnetosphere taps the orbital energy of the satellites at a rate that depends on detailed modelling on the magnetic interactions. The interactions can be classified into two types: the unipolar interaction with an unmagnetized satellite such as Io and the magnetic reconnection with a magnetized satellite such as Ganymede. The energy is then transported by the Alfvén waves and/or by a fast electron beam along the field lines from the satellites to Jupiter’s surface where the energy is dissipated, thereby explaining the satellite-induced emissions from Jupiter. In the case of star-planet interactions, this picture has been modified to take into consideration stellar winds along open field lines or to allow for a large static magnetic loop connecting the star and the planet. As a result, the phase differences between the stellar “spot" and the planet are explained by the time lag due to the Alfvén travel time to the star in the Alfven-wave model [@Pre06], or by the large magnetic loop having a geometry across longitudes of the star in the electron-beam case [@Lanza08]. A three-dimensional resistive magnetohydrodynamic simulation was performed to study how the magnetic field-aligned current can develop from a hot Jupiter [@Pre07].
However, how the stellar atmosphere thermally responds to any energy injection from the planet so as to generate the chromospheric emissions of $\gtrsim 10^{27}$ erg/s remains elusive. The magnetic energy at the magnetopause of a hot Jupiter has been estimated to be insufficient to supply the energy rate of $\gtrsim 10^{27}$ erg/s [@Shk05; @zak07] if the strength of the stellar surface field is $\sim$ a few Gauss [@Catala]. Furthermore, if this energy dissipation rate arises entirely from the orbital energy of a hot Jupiter, most hot Jupiters would plunge into their central stars in a few billion years, a timescale comparable to the age of these planetary systems [@Shk05]. However, the gas in a stellar atmosphere is certainly not quiescent but is fluctuating with the free energy that may be liberated as a hot Jupiter encounters the turbulent stellar fields along its orbit. @Cuntz and @Saar took into account the energy contribution from the stellar macroturbulence velocity and compared the strength of the planet-star interactions relative to each other for a number of hot-Jupiter systems. @Gu05 postulated that most of the planet induced emissions may result from stellar turbulent energy to reconcile the energetic problem. To further examine this possibility, a more realistic stellar atmosphere model implemented with an energy equation involving stellar turbulence is required.
In the case of the solar atmosphere, the coronal fields that can reach the typical orbits of hot Jupiters are the open magnetic fields emanating from the coronal holes. The dynamical features of the open fields are the non-linear fluctuations in the corona and the high-speed winds (800 km/s at $\approx 1$ AU). To explain these features, [@SI05] (hereafter SI05) introduced a 1-dimensional magnetohydrodynamic simulation with radiative cooling and thermal conduction. In this simulation, the authors have self-consistently treated the transfer of mass/momentum/energy by solving the magnetic waves propagating from the photosphere to the interplanetary space. The open field lines are assumed to expand super-radially from the photosphere to the corona [@KH], which is consistent with the spectropolarimetric measurements of magnetic structures in the Sun’s polar region [@Tsuneta]. The heating and acceleration of the gas in the coronal holes are achieved by the dissipation and the pressure of the non-linear magnetic waves. The model can explain the structure of the solar atmosphere (i.e. consisting of the chromosphere, the thin transition region, and the corona) as well as the high-speed winds from the coronal holes.
In this paper, we employ the magnetohydrodynamic simulation described above and model the open fields as a 1-dimensional flux tube from the stellar photosphere to the hot Jupiter. The aim is to conduct a numerical experiment as a starting point to study the thermal properties of the stellar chromosphere, transition region, and corona in response to an energy injection. Our study concerns HD 179949, but it is also guided by information about the solar wind due to the fact that various properties of HD 179949, especially in regard to its chromosphere, corona and wind, are poorly constrained. In §2, we describe the stellar atmosphere model in the presence of an inward-propagating electron beam. The numerical results are presented in §3. The paper concludes with a summary and discussion in §4.
Model Description
=================
We consider a magnetized hot Jupiter revolving in a circular orbit on the equatorial plane of its parent star. As the planet orbits through the open flux tubes emanating from the stellar surface, magnetic reconnections occur and generate plasma jets propagating inwards along the flux tubes to the parent star. In the following subsections, we shall describe the model of electron-beam injection and the stellar atmosphere associated with the flux tube.
Electron-Beam Injection {#sec:ebi}
-----------------------
Since HD 179949 has been the canonical planetary system highlighted by most previous studies for magnetic interactions, we adopt the parameters of HD 179949 as an illustrative example for our main study. In other words, we focus on a hot Jupiter orbiting at the radial distance $r=7.8R_*$ around a central star of mass $M_*=1.21M_{\odot}$, radius $R_*=1.22R_{\odot}$, and effective temperature $T_{\rm eff}=6168$ K [@Butler]. In terms of $M_*$, $R_*$, and $T_{\rm eff}$, these stellar parameters are noticeably different from the solar values, although in the broad sense HD 179949 still constitutes a solar-type star.
The typical field strength at the plane orbit is $B\sim 0.1-0.01$ G, for the [*average*]{} radial field strength $\sim 1-10$ G at the stellar surface (i.e. $B\propto r^{-2}$). At $r\approx
10R_{\odot}$, the planet is located inside the Alfvén radius of the star (see the next subsection) and therefore no fast MHD shocks form as the stellar winds encounter the planet’s magnetosphere (cf. Zarka 2001; Ip et al. 2004; Preussue et al. 2005). Assuming a dipole field for the planet’s magnetic field, we can estimate the radius of the magnetopause $R_{mp}$ at which the stellar and the planet’s fields are balanced: $$R_{mp}=4.64 R_J \left({R_p\over R_J}\right) \left[ \left({B_p
\over 1\,{\rm G}} \right) \left( {B_* \over 1\,{\rm G}}
\right)^{-1} \left( {a \over 10R_{\odot}} \right)^2 \left( {R_*
\over R_{\odot}} \right)^{-2} \right]^{1/3}, \label{eq:R_mp}$$ where $B_p$ and $B_*$ are the average surface fields of the planet and the parent star respectively, $R_J$ is the radius of Jupiter, and $a$ is the semi-major axis. As can be clearly seen from the above equation, $R_{mp}<<a=10R_{\odot}$. Therefore the magnetic field strength $B_{mp}$ at the magnetopause is almost equal to the stellar field strength at $r=a$; i.e., $B_{mp}\sim 0.1-0.01 G$.
We consider a jet originating from the reconnection sites at the magnetopause. The energy flux liberated by the reconnection is on the order of $~ (B_{mp}^2/4\pi)v_{rel}$, where $v_{rel}$ is the speed of the planet relative to the stellar fields. For a slowly rotating solar-type star, $v_{rel}$ is close to the orbital speed of the planet $\sqrt{GM_\odot /a}=1.38\times 10^7$ cm/s (cf. $v_{rel}\gtrsim 300$ km s$^{-1}$ in @zak07). Thus, we arrive at a rough estimate of the liberated energy flux $$\frac{B_{mp}^2}{4\pi}v_{rel} \approx 10^2 {\rm erg\;
cm^{-2}s^{-1}}\left(\frac{B_{mp}} {0.01{\rm
G}}\right)^2\left(\frac{v_{rel}}{138 \,\rm{km\;s^{-1}}}\right).
\label{eq:F_b}$$ A fraction of this energy is converted to the kinetic energy of the reconnection jets. The speed of the jets is on the order of the local Alfvén speed $B_{mp}/\sqrt{4\pi\rho}\simeq 900$km s$^{-1}\left(\frac{B}{0.01{\rm G}}\right)
\left(\frac{\rho}{10^{-21}{\rm g cm^{-3}}}\right)^{-1/2}$, where $\rho=10^{-21}$ g/cm$^3$ is the typical mass density at $r\approx
10R_{\odot}$ (e.g. Suzuki & Inutsuka 2006). We assume that the same energy per unit mass is converted to the kinetic energy of the electrons. Therefore the electrons are moving faster than the ions by a factor of the square root of the ion-electron mass ratio ($\approx 40$). As a result, $v_b \approx 3\times 10^4$ km s$^{-1}$ is a typical speed of the electron beam.
We denote the momentum flux and the energy flux of the electron beam as $P_b$ and $F_b$, and their initial values as $P_{bi}$ and $F_{bi}$, respectively. If we neglect thermal fluctuations of the beam particles, the initial momentum flux and energy flux of the electron beam are $$\begin{aligned}
P_{bi}&=&\rho_{bi} v_b^2, \\
F_{bi} &=&\frac{1}{2}\rho_{bi} v_b^3,\end{aligned}$$ where $\rho_{bi}$ is the initial beam mass density. In the present numerical experiment, we restrict ourselves to the cases in which $F_{bi} = 10^2$ and $10^4$ erg cm$^{-2}$s$^{-1}$. We adopt a constant $v_{b}=3\times 10^4$km s$^{-1}$, and vary $\rho_{bi}$ for different $F_{bi}$. Note that according to eq. (\[eq:F\_b\]) and the parameters of the HD 179949 system, $F_{bi} = 10^2$erg cm$^{-2}$s$^{-1}$ corresponds to $B_{mp} \approx 0.005-0.01$ G and hence $B_* \approx 1$ G, and the larger energy flux $F_{bi} =
10^4$erg cm$^{-2}$s$^{-1}$ corresponds to $B_{mp} \approx
0.05-0.1$ G and thus the stronger stellar field $B_* \approx 5$ G.
We assume that the electron beam initially streams freely along the flux tube at a constant $v_b$. As they approach the parent star, the beam particles start to interact with the dense background ions through Coulomb collisions and become more concentrated due to the flux-tube convergence (see the next subsection). The collisions lead to the heating of surrounding plasma. The heating becomes efficient when the mean free path of the beam is comparable to the local density scale height $H_\rho$. The mean free path of an electron colliding with a pool of thermal ions is given by $l_{\rm mfp} \approx 3\times 10^4 {\rm km}
\left(\frac{v_b}{3\times 10^4 {\rm
km\;s^{-1}}}\right)^4\left(\frac{n}{10^9 {\rm
cm^{-3}}}\right)^{-1}$ [@Braginskii], where $n=10^{9}$cm$^{-3}$ is the typical density at the upper transition region or the lower corona (e.g. see SI05, or refer to the density profile in Figure 1 to be discussed in §3). Roughly speaking, for $v_b=3\times 10^4$ km/s, $l_{\rm mfp} \lesssim H_\rho$ when $n>10^9$cm$^{-3}$. Therefore, we anticipate the inward-propagating electron beam to heat up ambient media from the lower corona to the upper chromosphere. This allows us to assume that the incoming electron beam starts dissipating from $r_{\rm max} =1.1R_{\odot}$ (lower corona) to $r_{\rm min}=1.001R_{\odot}$ (upper chromosphere). We model the energy flux of the beam to decrease inwardly according to $$F_{\rm b} \propto \left(\frac{r-r_{\rm min}}{r_{\rm max}-r_{\rm
min}} \right)^k , \label{eq:F_b_diss}$$ where $k$ is a parameter that describes the spacial distribution of the heating. $k=1$ corresponds to constant volumetric heating; namely, the heating rate per unit mass is higher in the upper region (i.e. corona) than in the lower region (i.e. chromosphere). If $k= 0.1$, the heating rate per unit mass is more uniformly distributed. In this work, we adopt $k=0.1$ to resemble the situation of a constant beam-heating rate per unit mass. We also assume that the momentum flux of the beam $P_b$ dissipates in the same manner as that described by eq. (\[eq:F\_b\_diss\]) for the energy flux. This implies that although the beam velocity $v_b$ does not decay in our dissipation model, the beam density $\rho_b$ declines, meaning that more and more beam electrons have been transformed into thermal electrons as the beam progresses downwards in the dissipation region.
The beam heating at the footpoint of one open field line occurs when the planet’s magnetosphere is crossing the field line. Hence, by means of eq. (\[eq:R\_mp\]), the beam heating proceeds on the timescale $$t_{beam}={2 R_{mp} \over v} \approx 80\, {\rm mins} \left( {R_p
\over R_J} \right) \left({v\over 138\,{\rm km/s}} \right)^{-1}
\left[ \left({B_p \over 1\,{\rm G}} \right) \left( {B_* \over
1\,{\rm G}} \right)^{-1} \left( {a \over 10R_{\odot}} \right)^2
\left( {R_* \over R_{\odot}} \right)^{-2} \right]^{1/3}.
\label{eq:t_heat}$$
Stellar Atmosphere Model
------------------------
The stellar atmosphere model is based on the 1-D magnetohydrodynamic simulation with radiative cooling and thermal conduction in an open flux tube (SI05). In the original simulation, the heating is given by the nonlinear dissipation of the Alfvén waves excited by the granulations at the photosphere. In the case of $B_*=1$ G, we set a rms average amplitude $<dv_{\perp}> \sim 1.8$ km/s at the photosphere, which is estimated from the scaling with the surface convective flux (Suzuki 2007) from the Sun (SI05). In the case of the stronger field $B_*=5$ G, the larger rms velocity fluctuation $<dv_{\perp}>
\sim 3.6$ km/s is used for the experiment, which is twice the value for the $B_*=1$ G case. We neglect the effect of rotation for simplicity in order to focus mainly on the energetics influenced by the electron beam.
When an inward-propagating electron beam described in the above subsection is taken into account, the momentum and energy equations in a 1-D open flux tube are modified to (cf. SI05) $$\label{eq:mom} \rho \frac{d v_r}{dt} = -\frac{\partial p}{\partial
r} - \frac{1}{8\pi r^2 f}\frac{\partial}{\partial r} (r^2 f
B_{\perp}^2) + \frac{\rho v_{\perp}^2}{2r^2 f}\frac{\partial
}{\partial r} (r^2 f) -\rho \frac{G M_*}{r^2} - \frac{\partial
P_{\rm b}}{\partial r} ,$$ and $$\begin{aligned}
\label{eq:eng} \rho \frac{d}{dt}(e + \frac{v^2}{2} +
\frac{B^2}{8\pi\rho} - \frac{GM_*}{r}) + \frac{1}{r^2 f}
\frac{\partial}{\partial r}[r^2 f \{ (p + \frac{B^2}{8\pi}) v_r -
\frac{B_r}{4\pi} (\mbf{B \cdot v})\}] \nonumber \\
+ \frac{1}{r^2
f}\frac{\partial}{\partial r}(r^2 f F_{\rm c})
+ \frac{1}{r^2 f}\frac{\partial}{\partial r}(r^2 f F_{\rm b}) +
q_{\rm R} = 0,\end{aligned}$$ where $\rho$, $\mbf v$, $p$, $e$, and $\mbf B$ are the density, velocity, pressure, specific internal energy, and magnetic field strength, respectively; the subscripts $r$ and $\perp$ denote radial and tangential components, respectively; $d/dt$ and $\partial/\partial t$ denote Lagrangian and Eulerian derivatives, respectively; $G$ and $M_*$ are the gravitational constant and the stellar mass, respectively; $F_c$ is thermal conductive flux; $q_R$ is the radiative cooling and $f$ is a super-radial expansion factor. We assume that the super-radial expansion (in addition to the radial expansion $\propto r^2$) is a factor of 240 and 480 from the photosphere to $\approx 2 R_{\odot}$ in the cases of $B_*=1$ and 5 G, respectively. It then follows that to give $B_*=$ 1G and 5G, the radial magnetic field strength at the footpoint of the flux tube at the photosphere $B_{ph}$ is 240G and 2400G, respectively. Note that on account of the convergence of the flux tube toward the central star, the energy flux of the incoming electron beam increases due to the areal focusing. In the case of $B_*=1$G, the flux tube converges by a factor of $\approx$14,600 (a factor of $\approx 61$ from the radial convergence and another factor of 240 from the super-radial convergence) from $7.8R_*$ (the planet’s orbit) to $1R_*$. On the other hand, the converging factor of the flux tube in the $B_*=5G$ case is twice as large. In the absence of dissipation, the energy flux of the beam increases by the same factor as the converging factor of the flux tube. In contrast, the same enhancement effect of the electron beam pressure on stellar thermal ions is almost negligible in our calculations because of the small mass of an electron.
Results of the Numerical Experiment
===================================
We conduct the numerical experiment to study the effects on the stellar atmosphere due to the electron beam with the initial energy fluxes for the two $B_*$ cases as described in the preceding session. The experiment is first carried out in the absence of the beam until the simulated atmosphere attains a quasi-steady state. The beam heating is then added afterwards.
Figure 1 shows the snapshot structures of the stellar atmosphere for $B_*=$ 1 and 5 G denoted respectively by “small" and “large $B$ and $dv$", and compares the results with and without the electron-beam injection. In the $B_*=1$G case, the results for $F_{bi}=10^2$ erg cm$^{-2}$ s$^{-1}$ (blue curves) are almost identical to those for the no-beam case (green curves). In contrast, the atmospheric structures change noticeably in the larger $B_*$ ($=5$G) and $<dv_\perp>$ ($=3.6$ km/s) cases, as illustrated by the red and black curves. In the absence of the beam heating, the stronger dissipation arising from the larger $<dv_\perp>$ raises the chromospheric temperature and therefore causes the chromosphere to evaporate and expand (see the red curve in the middle panel). The local density scale height increases and the density drops more slowly with $r$. Therefore, the locations of the transition region lie at higher altitudes in the larger $<dv_\perp>$ case. A hot and dense region forms in the chromosphere, which we term as “a warm region" (i.e. $T>10^4$K, see the bottom panel). It then becomes difficult to further heat up the warm region to the coronal temperature (i.e. $\gtrsim 10^6$ K) as a result of the efficient radiative cooling at $T\lesssim 10^5$ K [@LM90]. However, when the beam energy of $F_{bi}=10^4$ erg cm$^{-2}$ s$^{-1}$ is added, the chromosphere is further heated up and thus an even larger warm region can form at the location of $r=1.001 - 1.01 R_*$ (see the black curve in the bottom panel). The black curve in the middle panel indicates that the density is on the average much higher in the beam-heated region. Consequently, the heating rate per unit mass becomes smaller and the temperature of the coronal region at $r \gtrsim 1.01R_*$ then becomes lower in the case of $F_{bi}=10^4$ erg cm$^{-2}$ s$^{-1}$ than in the case of $F_{bi}=0$ (see the bottom panel). The sharp transition region disappears.
The top panel of Figure 1 shows the radial profile of the stellar wind velocity $v_r$. The winds in the outer region ($r\gtrsim
2R_*$) are faster in the large $B_*$ case. $B_*$ equals $B_{ph}/f$, which in fact specifies the flux tube properties. The larger the factor $B_{ph}/f$ is, the more the wave energy dissipates in the outer region of the atmosphere, leading to faster winds [@Kojima; @Suzuki06]. However, the winds on the average are not significantly amplified by the beam heating, as indicated by the large overlap between the red and black curves in the plot. The wind speed at the planet’s orbit is $\approx 300$ km s$^{-1}$, which is lower than the Alfvén speeds $\sim 500$ and 1000 km s$^{-1}$ there in our cases for $B_*=1$ and 5 G. That is, the planet lies inside the Alfvén radius.
Having described the snapshot structures, we should note that the thermal properties of the stellar atmosphere actually fluctuate with time owing to wave propagation and dissipation. As a result, the location and thermal properties of the warm region fluctuate with time as well. Figure 2 shows the evolution of the radiative fluxes arising from different temperature components in the beam and no-beam cases. In general, all of the radiative fluxes fluctuate with time. Nevertheless for the case of $B_*=1$G and $<v_\perp>=1.8$ km/s, the electron beam of $F_{bi}=10^2$ erg cm$^{-2}$ s$^{-1}$ gives rise to only tiny effects on the radiations compared to the no-beam case, as expected from the snapshot structures shown in Figure 1. In contrast, the case of $F_{bi}=10^4$ erg cm$^{-2}$ s$^{-1}$ corresponding to $B_*=5$G exhibits noticeable differences from the no-beam case, which can also be expected from the snapshot structures in Figure 1. The radiative fluxes from the hot chromosphere (7000-20000K) and the transition region (20000-$5\times 10^5$K) are greatly enhanced by the warm region in the vicinity of $r=1.001-1.01R_*$. Since the warm region is on the average denser than the usual chromosphere and transition region in $r\gtrsim 1.003 R_*$ (see Figure 1), the radiative flux corresponding to the chromospheric temperature is on average increased by a factor of $\sim10$ and the radiative flux associated with the transition-region temperature is intensified by a factor of $\sim100$.
The bottom panel of Figure 2 shows that the beam of $F_{bi}=10^4$ erg cm$^{-2}$ s$^{-1}$ associated with $B_*=5$G is able to enhance the radiation from the hot gas of $T>5\times 10^5$K occasionally by a factor of $\gtrsim 100$. This radiation is mainly from the warm region that intermittently develops around $r=1.001-1.003R_*$. With this beam heating, the temperature of the warm region can sometimes go up and down between $10^5$K and $10^6$K. This temperature variability is due to the thermal instability [@LM90; @Suzuki07]. The electron beam is energetic enough to continuously heat up the warm region, while the wave dissipation heats it up in a stochastic manner. A small change of the wave heating rate triggers violent fluctuations of temperature in the thermally unstable regime of $10^5 < T < 10^6$K.
As is shown in Figure \[fig2\], the thermal response of the stellar atmosphere to the onset of the beam-heating to evolve to the hot state is nearly instantaneously. In §\[sec:ebi\], we estimated the duration of the electron beam as $\sim 100$ min, which is rather short in comparison with the simulation time presented in Figure \[fig2\]. To estimate the area of the hot spot, it is equally important to examine how long the hot atmosphere can be sustained after the electron beam passes by. Figures \[fig4\] & \[fig5\] show the results for the $B_*=5$G case but the beam is switched off at $t=109$ min. Figure \[fig4\], which presents the snapshot structure at $219$ min, namely 110 min after the beam stops, illustrates that while the warm region in the chromosphere is still present as shown by the temperature profile, its density has dropped to the original lower-density level. Figure \[fig5\] shows that the radiative fluxes from the hot chromosphere (top) and transition regions (middle) have declined for $\approx$ 50-60 minutes since the beam is switched off, which indicates that the atmosphere takes only a fraction of the heating time to revert to the normal state in the absence of the electron beam. Therefore, the size of the hot spot estimated in terms of the projected area of the planet’s magnetosphere along a flux tube onto the chromosphere is a reasonable approximation.
@Shk05 [@Shk08] found that in the case of HD 179949, the planet induced Ca II flux averaged over the stellar disk is $\sim 1.5\times 10^5$ ergs cm$^{-2}$ s$^{-1}$, which amounts to an increase of a factor of about 1.04 compared to the non-planet induced Ca II emissions. Hence the Ca II emissions with no planet induced component is $\sim 1.5/0.04 \times 10^5 = 3.7 \times 10^6$ erg cm$^{-2}$ s$^{-1}$, which is on the similar order of the value given from the model shown by the red-dotted line in Figure \[fig2\]. In this sense, the model for $B_*=5 $G and $<dv_\perp>=3.6$ km/s may mimic a chromospheric condition similar to that of HD 179949, although our cooling function $q_R$ for the chromospheric radiation is based on the observation of the Sun [@AA89]. Having obtained the radiative flux $\sim 10^{6-7}$ erg cm$^{-2}$ s$^{-1}$ from the hot chromosphere in the case for $B_*=5$G and $<dv_\perp>=3.6$ km/s, we can estimate the luminosity of the hot spot. In our model, the cross-sectional area of the converging flux tube in the chromosphere is about 8000 times smaller than the planet’s magnetosphere. In other words, the area of the hot spot in the chromosphere is given by $$A_{chromo}\approx 10^{17-18} \left( {R_{mp} \over 5 R_J} \right)^2
\,{\rm cm^2},$$ which when multiplied by the chromospheric flux gives the luminosity of the chromospheric hot spot $\approx 10^{23-25}$ erg/s. This total chromospheric emission is still 2-4 orders of magnitude weaker than the observational Ca II emissions for HD 179949.
Summary and Discussions
=======================
By conducting a numerical experiment, we study the thermal response of the atmosphere of a solar-type star to the dissipation of an injected electron beam at the coronal base. The experiment is carried out based on the framework of the 1-D magnetohydrodynamic simulation by SI05 with non-linear wave dissipation, radiative cooling, and thermal conduction. We assume that the magnetic stress due to the orbital motion of the planet relative to the stellar coronal fields generates an electron beam, which in turn funnels along the stellar open field lines to the central star. As the beam travels inwards, the energy flux of the incoming electron beam is intensified by the areal focusing of the super-radially converging open flux tube.
We use the stellar parameters of HD 179949 as an illustrative example but ignore possible magnetic properties arising from its stellar rotation. When the average stellar field at the photosphere $B_*$ is about 1 G and the average amplitude of the wave velocity $<v_\perp>$ is about 1.8 km/s, the stellar atmosphere is not considerably altered after the beam dissipation is turned on. In contrast, when $B_*=$5 G and $<v_\perp>=3.6$ km/s, we find that a warm region forms in the chromosphere. The warm region becomes substantially hotter and denser once the electron-beam heating is switched on. As a result, the beam-intensified warm region enhances the chromospheric radiative flux by a factor of $\sim10$, and the radiative flux corresponding to the temperature of the transition region by a factor of $\sim100$. The warm region can also intermittently contribute to the radiative flux associated with the coronal temperature by a factor of $\sim100$ due to the thermal instability [@LM90]. In other words, the planet-induced radiations are not perturbations in the local region of the hot spot compared to the normal state of the stellar emissions. However, owing to the small area of the heating spot, the total luminosity of the beam-induced chromospheric radiation is 2-4 orders of magnitude smaller than the observed Ca II emissions from HD 179949. The energetics of the planet-induced emissions becomes a more serious problem in our numerical experiments when explaining the statistical results of X-rays from planet-host stars: stars with close-in giant planets are on average more X-ray active by a factor $\sim 4$ than those with planets that are more distant [@Kashyap08]. Since the typical X-ray luminosity from a solar-type stars is $\gtrsim 10^{27-28}$ erg/s, the planet-induced X-ray inferred from the statistical analysis is actually even stronger than the planet-induced Ca II emissions from HD 179949. Our simulation results show that an $\sim 100$ times enhancement in X-ray due occasionally to the thermal instability of the small warm region contributes only an even more negligible perturbation to the total X-ray emission, rather than being comparable to it.
We note that our open-field model for the chromospheric emissions is different from those occurring on the Sun where the emissions come primarily from the solar plage regions in closed magnetic loops. Needless to say, our 1-D numerical experiment restricts ourselves to exclude any mechanical and thermal influence of the heating area on neighboring open fields and closed magnetic loops. As such, our results leave an open question as to whether the thermal instability or any other magnetic instabilities (e.g. Lanza 2008; Ishikawa et al. 2008) can be further triggered around the beam-heated spot to liberate more energy. In our numerical experiment $B_*$ has been taken to be 1 and 5 G, which is consistent with the field measurements via spectrapolarimetry for the similar spectral type dwarf $\tau$ Bootis [@Catala]. However, the field strength inferred from the Stokes V observation may be underestimated due to the cancellation of circular polarization arising from the opposite directions of $B_*$ along the line of sight. It is normally expected that a faster rotator and therefore a stronger X-ray emitter may possess stronger $B_*$ than the Sun [@Ruedi; @Gudel]. For instance, $B_*$ of HD 179949 has been assumed to be $\approx$ 8-9$\times$ solar value by scaling with the X-ray flux [@Saar]. Furthermore, [@MJ] related $B_*$ and the filling factor $f_s$ to the Rossby number $Ro$ ($\equiv
P_{rot}/\tau_c$, where $P_{rot}$ is the rotation period and $\tau_c$ is the convection turnover time) of a dwarf star. In the case of HD 179949, $P_{rot}$ has been suggested to be about 7 days [@WH04; @Shk08] and $\tau_c \approx 4.287$ days may be inferred from its color B$-$V=0.503 (Noyes et al. 1984). Thus following [@MJ], we obtain the average field $Bf_s=37.6$ G and the filling factor $f_s=0.02$. If we assume that these magnetic properties are contributed mainly from the open-field region of our model, then $Bf_s = 37.6$ G is equivalent to $B_*$ and the filling factor $f_s=0.02$ may correspond to the super-radial expansion factor $1/f$ in our experiment. In reality, some of the contribution may come from closed-field regions [@MJ], which introduces additional complexity.
The other free parameter that governs our numerical results is the power spectrum of MHD waves. The stellar macroturbulent velocity $v_{mac}$ for HD 179949 is expected to be 2 times larger than that for the Sun [@SO97; @Saar]. In our experiment, we adopt a larger $<dv_{\perp}>$ for the stronger $B_*$ case. However, the correlation between $v_{mac}$ and $<dv_{\perp}>$ is not tested in this work and how these velocities are associated with $B_*$ is not modelled in our numerical experiment. After all, in view of all of the uncertainties and complexities mentioned above, our results serve only as the fiducial examples for future studies. The numerical experiment covering a broader parameter space coupled with more magnetic-field measurements will be essential to further diagnose the problem.
The effect of the centrifugal force is not implemented in the present numerical experiment. While it is a reasonable approximation for thermally driven winds from a slow rotator like the Sun, the magneto-centrifugal winds play an equally important role in accelerating stellar winds for a star rotating $\sim 10$ times faster than the Sun [@BM76; @WS93; @Pre06]. The centrifugal force of a 7 day-period star is $\sim 1/5$ that of a 3 day-period star, as the force is proportional to the square of the rotation frequency. The effect of the centrifugal force on the stellar atmosphere will be investigated in a future work.
While most of the previous magnetohydrodynamic simulations and studies have focused on the planet’s side for the magnetic interactions (Ip et al. 2004; Preusse et al. 2006, 2007; cf. Laine et al. 2008 for a young hot Jupiter), our numerical experiment makes an attempt to investigate how a stellar atmosphere down to the photosphere in the open field region responds to the dissipation likely from a hot Jupiter. The current model is simplified in such a way that we prescribe the energy dissipation and ignore stellar rotation. Nevertheless, the dissipation contributed by the planet is described only by the energy flux $F_b$ at the coronal base, meaning that $F_b$ is not necessarily specific only to an incoming electron beam but can be in the form of other dissipative sources (e.g. damping of Alfven waves) with proper modification. Furthermore, our simulation lays the framework to extend the calculations for other types of dwarf stars if the corresponding magnetic wave amplitudes and power spectra are specified.
We wish to dedicate this work in memory of Chi Yuan who passed away on July 24, 2008. Without him, the authors would not have met each other and finally made the work possible. We are grateful to N. Phan-Bao and E. Shkolnik for useful discussions. We also thank the anonymous referee for valuable comments to improve the paper. P.-G. Gu is supported by the NSC grants in Taiwan through NSC 95-2112-M-001-073MY2 and 97-2112-M-001-017. T. K. Suzuki is supported in part by Grants-in-Aid for Scientific Research from the MEXT of Japan, 19015004 and 20740100 and Inamori Foundation.
Anderson, L. S., & Athay, R. G. 1989, , 346, 1010 Braginskii, S. I. 1965, Rev. Plasma Phys., 1, 205 Belcher, J. W., & MacGregor, K. B. 1976, , 210, 496 Butler, R. P., Wright, J. T., Marcy, G. W., Fischer, D. A., Vogt, S. S., Tinney, C. G., Jones, H. R. A., Carter, B. D., Johnson, J. A., McCarthy, C., & Penny, A. J. 2006, ApJ, 646, 505 Catala, C., Donati, J.-F., Shkolnik, E., Bohlender, D., & Alecian, E. 2007, MNRAS, 374, L42 Cuntz, M., Saar, S. H., & Musielak, Z. E. 2000, , 533, L151 Gu, P.-G., Shkolnik, E., Li, S.-L., & Liu, X.-W. 2005, Astronomische Nachrichten, 326, 909 Güdel, M. 2007, Living Rev. Solar Phys., 4, 3 Ip, W.-H., Kopp, A., & Hu, J.-H. 2004, , 602, L53 Ishikawa, R., Tsuneta, S., Ichimoto, K., Isobe, H., Katsukawa, Y., Lites, B. W., Nagata, S., Shimizu, T., Shine, R. A., Suematsu, Y., Tarbell, T. D., & Title, A. M. 2008, , 481, L25 Jackson, B., Barnes, R., & Greenberg, R. 2009, , 698, 1357 Jacques, S. A. 1977, , 215, 942 Kashyap, V. L., Drake, J. J., & Saar, S. H. 2008, , 687, 1339 Kojima, M., Fujiki, K., Hirano, M., Tokumaru, M., Ohmi, T., and Hakamada K., 2005, “The Sun and the heliosphere as an Integrated System”, G. Poletto and S. T. Suess, Eds. Kluwer Academic Publishers, 147 Kopp, R. A., & Holzer, T. E. 1976, Sol. Phys., 49, 43 Laine, R. O., Lin, D. N. C., & Dong, S.-F. 2008, , 685, 521 Landini, M., & Monsignori-Fossi, B. C. 1990, A&AS, 82, 229 Lanza, A. F. 2008, , 487, 1163 Lin, D. N. C., Bodenheimer, P., & Richardson, D. C. 1996, Nature, 380, 606 Montesinos, B., Jordan, C. 1993, , 264, 900 Noyes, R. W., Hartmann, L. W., Baliunas, S. L., Duncan, D. K., Vaughan, A. H. 1984, , 279, 763 Pfahl, E., Arras, P., & Paxton, B. 2008, , 679, 783 Preusse, S., Kopp, A., Büchner, J., & Motschmann, U. 2005, A&A, 434, 1191 Preusse, S., Kopp, A., Büchner, J., & Motschmann, U. 2006, A&A, 460, 317 Preusse, S., Kopp, A., Büchner, J., & Motschmann, U. 2007, Planetary & Space Science, 55, 589 Rubenstein, E. P. & Schaefer, B. E. 2000, , 529, 1031 Rüedi, I, Solanki, S. K., Mathys, G., & Saar, S. H. 1997, , 318, 429 Saar, S. H., & Osten, R. A. 1997, MNRAS, 284, 803 Saar, S. H., Cuntz, M. & Shkolnik, E. 2004, IAU Symposium, Stars as Suns: Activity, Evolution, and Planets, A. K. Dupree and A. O. Benz, Eds, 219, 355 Shkolnik, E., Walker, G. A.H., & Bohlender, D. A., 2003, , 597, 1092 Shkolnik, E., Walker, G. A.H., Bohlender, D. A., Gu, P.-G., & Kürster, M. 2005, , 622, 1075 Shkolnik, E., Bohlender, D. A., Walker, G. A. H., & Collier Cameron, A. 2008, , 676, 628 Suzuki, T. K. & Inutsuka, S.-I. 2005, , 632, L49 (SI05) Suzuki, T. K. 2006, ApJ, 640, L75 Suzuki, T. K. & Inutsuka, S.-I. 2006, Journal of Geophysical Research, l11, A06101 Suzuki, T. K. 2007, , 659, 1592 Tsuneta, S., Ichimoto, K., Katsukawa, Y., Lites, B. W., Matsuzaki, K., Nagata, S., Orozco Suarez, D., Shimizu, T., Shimojo, M., Shine, R. A., Suematsu, Y., Suzuki, T. K., Tarbell, T. D., & Title, A. M. 2008, , 688, 1374 Walker, G. A. H. et al. 2008, , 482, 691 Washimi, & Shibata 1993, MNRAS, 262, 936 Wolf, M. & Harmanec, P. 2004, Inf. Bull. Variable Stars, 5575, 1 Zarka, P., Treumann, R. A., Ryabov, B. P., & Ryabov, V. B. 2001, Ap&SS, 277, 293 Zarka, P. 2007, Plan. Spa. Sci., 55, 589
| {
"pile_set_name": "ArXiv"
} |
---
author:
- '****'
title: '**The Existence of Type II Singularities for the Ricci Flow on $S^{n+1}$**'
---
7.7in 5.8in 0.3in 0.4in 0.15in \[section\] \[section\] \[section\]
**Hui-Ling Gu and Xi-Ping Zhu**
[Department of Mathematics]{}
[Sun Yat-Sen University ]{}
[Guangzhou, P. R. China]{}
[**Abstract** ]{} In this paper we prove the existence of Type II singularities for the Ricci flow on $S^{n+1}$ for all $n
\geq 2$.
[****[1. Introduction]{}****]{}
In this paper, we consider the Ricci flow $$\left\{
\begin{array}{lll}
\frac{\partial g_{ij}}{\partial t}=-2R_{ij},
\\[4mm]
g_{ij}(0)=\hat g_{ij},
\end{array}
\right.
\eqno(1.1)$$ starting from a given compact Riemannian manifold (${M},\hat g$). This is a nonlinear (degenerate) parabolic system on metrics. In the seminal paper [@Ha82], Hamilton proved the Ricci flow admits a unique solution on a maximal time interval $[0,T)$ so that either $T=+\infty$ or $T<+\infty$ and $|Rm|$ is unbounded as $t\rightarrow T$. We call such a solution $g(t)$ a maximal solution of the Ricci flow. If $T<+\infty$ and the curvature becomes unbounded as $t$ tends to $T$, we say the maximal solution develops singularities as $t$ tends to $T$ and $T$ is a singular time. It is well-known that the Ricci flow generally develops singularity.
If a solution $(M,g(t))$ to the Ricci flow develops singularities at a maximal time $T < +\infty$, according to Hamilton [@Ha95F], we say it develops a if $$\ \sup_{t\in [0,T)}(T-t)K_{max}(t)<+\infty,$$ and say it develops a if $$\ \sup_{t\in [0,T)}(T-t)K_{max}(t)=+\infty,$$ where $K_{max}(t) = \max \{|Rm(x,t)| \ | \ x \in M \}$.
Clearly, a round sphere, or more generally a finite product of several space-forms with positive curvature, shrinks to form Type I singularities. In [@Ha88; @Chow], Hamilton and Chow proved the Ricci flow on two-sphere $S^2$ (with an arbitrary metric) always develops a Type I singularity and shrinks to a round point. In [@Ha82; @Ha86], Hamilton proved the Ricci flow on a compact three-manifold with positive Ricci curvature, or a compact four-manifold with positive curvature operator, develops a Type I singularity and shrinks to a round point; recently, B$\ddot{o}$hm-Wilking [@BW] had shown that the Ricci flow on a general compact $n$-dimensional Riemannian manifold with positive curvature operator also develops a Type I singularity and shrinks to a round point.
Intuitively, a compact manifold with the shape like a dumbbell will develop a Type I singularity in the neck part. In views of the work [@Ha97] of Hamilton on four-manifolds with positive isotropic curvature (see also [@CZ05]), a Type I singularity with neckpinch is expected. Indeed, such examples of Type I singularities with neckpinch for the mean curvature flow were known more than fifteen years ago (see for example [@Gr] and [@An-m]). It is very surprising that the existence of Type I singularities with neckpinch for the Ricci flow was only known very recently. The first rigorous examples of Type I singularity with neckpinch for the Ricci flow were constructed by Miles Simon [@MS] on *noncompact* warped products $R\times_fS^n$. In [@FIK], Feldman-Ilmanen-Knopf constructed another family of rigorous examples of Type I with neckpinch on the holomorpic line bundle $L^{-k}$ over $CP^{n-1}$. Both of these families of examples live on *noncompact* manifolds. For the Type I singularity with neckpinch on compact manifolds, the first rigorous examples were given by Sigurd Angenent and Dan Knopf in [@AK] by constructing suitable rotationally symmetric metrics on $S^{n+1}$, where the definition of a rotationally symmetric metric is the following: 0.3cm [**Definition 1.1**]{} A metric $g$ on $I\times
S^n$, where $I$ is an interval, is called **rotationally symmetric** if it has the following form:$$g=\varphi (x)^2dx^2+\psi
(x)^2g_{can},\qquad x\in I,$$ where $g_{can}$ is the standard metric of the round sphere $S^n$ with constant (sectional) curvature 1.
For the Type II singularity for the Ricci flow, a rigorous example on $R^2$ was recently given by Daskalopoulos and Hamilton in [@DH]. However, no rigorous examples of Type II singularity for the Ricci flow on compact manifolds have yet appeared. We remark that some beautiful intuitions of the forming of Type II singularity were described and explained by Chow-Knopf in [@CK] and Topping in [@T]. (For the mean curvature flow, the existence of Type II singularities was already justified by Altschuler-Angenent-Giga [@AAG] and Angenent-Vel$\acute{a}$zquez [@An-c].)
The purpose of this paper is to demonstrate the existence of Type II singularity on compact manifolds, in particular for rotationally-symmetric initial metrics on $S^{n+1}$. Our main result is the following:
0.3cm [**Theorem 1.2**]{} *For each $n\geq 2$, there exist rotationally-symmetric metrics on $S^{n+1}$ such that the Ricci flow starting at the metrics develop Type II singularities at some times $T<+\infty$.*0.3cm
This paper contains four sections. In Section 2, we recall some useful estimates of Angenent-Knopf [@AK] on rotationally symmetric solutions to the Ricci flow. In general, to understand the structure of singularities, one usually needs to get a classification for gradient shrinking solitons. The recent work [@P2] of Perelman gives a complete classification to positively curved gradient shrinking Ricci soliton in dimension three. In Section 3 we will extend Perelman’s classification to higher dimensions in the class of rotationally symmetric metrics. Finally in Section 4, based on the generalized classification, we will prove the main result Theorem 1.2.
Our work in this paper benefits from a conversation with Professor R. S. Hamilton, who suggested the second author to consider the class of rotationally symmetric metrics. The second author is partially supported by NSFC 10428102 and NKBRPC 2006CB805905.
[ ****[2. Angenent-Knopf’s Estimates]{}****]{}
Consider a rotationally-symmetric metric $$g=\varphi (x)^2dx^2+\psi (x)^2g_{can} \eqno(2.1)$$ on the set $(-1,1)\times{S^n}$, in which $g_{can}$ is the metric of constant sectional curvature 1 on $S^n$. The coordinate $x$ is ungeometric, a more geometric quantity is the distance $s$ to the equator given by $$s(x)=\int_0^x\varphi(x)dx.$$ Then $$\frac{\partial}{\partial s}=\frac{1}{\varphi(x)}\frac{\partial}{\partial
x}$$ and $$ds=\varphi(x)dx.$$ With this notation the metric is $$g=ds^2+\psi^2g_{can}. \eqno(2.2)$$ In order to extend $g$ to be a smooth Riemannian metric on $S^{n+1}$, it is sufficient and necessary to impose the boundary conditions: $$\psi(\pm1)=0,\qquad
\lim_{x\rightarrow\pm1}\psi_{s}(x)=\mp1$$ and $$\lim_{x\rightarrow\pm1}\frac{d^{2k}\psi(x)}{ds^{2k}}=0$$ for all $k = 1, 2, \cdots$. The Riemannian curvature tensor of (2.2) is determined by the sectional curvatures $$K_0=-\frac{\psi_{ss}}{\psi}$$ of the $n$ 2-planes perpendicular to the spheres $\{x\}\times S^n$, and the sectional curvatures $$K_1=\frac{1-\psi_s^2}{\psi^2}$$ of the $\frac{n(n-1)}{2}$ 2-planes tangential to these spheres. In the ungeometric coordinate $x$ the Ricci tensor of the metric $g$ given by (2.1) is $$Ric=n\{-\frac{\psi_{xx}}{\psi}+\frac{\varphi_x\psi_x}{\varphi\psi}\}(dx)^2+
\{-\frac{\psi\psi_{xx}}{\varphi^2}-\frac{(n-1)\psi_x^2}{\varphi^2}+\frac{\psi\varphi_x\psi_x}{\varphi^3}
+n-1\}g_{can}.$$ In the geometric coordinate this simplifies to $$Ric=(nK_0)ds^2+\psi^2[K_0+(n-1)K_1]g_{can}.$$The scalar curvature is given by $$R=2nK_0+n(n-1)K_1.$$ The above computations can be found in [@AK] or the textbook [@PP].
Suppose we have a time dependent family of metrics $g(\cdot,t)$ having the form (2.1). Then the family $g(\cdot,t)$ satisfies the Ricci flow if and only if $\varphi$ and $\psi$ evolve by: $$\varphi_t=n\frac{\psi_{ss}}{\psi}\varphi, \eqno(2.3)$$ $$\psi_t=\psi_{ss}-(n-1)\frac{1-\psi_s^2}{\psi}. \eqno(2.4)$$
According to Angenent-Knopf [@AK], the (interior) local minimal points of the function $x\mapsto \psi(x,t)$ are called “**necks**" and the (interior) local maximal points are called “**bumps**". As long as the solution exists at a time $t$, the radius of the smallest neck is given by $$r_{min}(t)=\min\{\psi(x,t)|\psi_x(x,t)=0\}.$$ Of course, if the solution has no necks at the time $t$, we let $r_{min}$ not be defined. Denote by $x_+(t),x_-(t)$ the right-most bump (i.e. the largest local maximal point on $(-1,+1)$) and left-most bump (i.e. the least local maximal point on $(-1,+1)$) respectively. The region right of $x_+(t)$ and left of $x_-(t)$ are called the “**right polar cap**" and “**left polar cap**" respectively. In [@AK], Angenent and Knopf obtained several useful estimates for the Ricci flow via the equations (2.3) and (2.4). We recall some of their estimates as follows.
0.2cm[**Proposition 2.1**]{} (Angenent-Knopf [@AK]) *Let $g(t)$ be a solution to the Ricci flow of the form (2.2) such that $|\psi_s|\leq 1$ and the scalar curvature $R>0$ and $\psi_s$ has finitely many zeroes initially. Then*
\(1) (Proposition 5.1 of [@AK]) *As long as the solution exists, $|\psi_s|\leq1$.*
\(2) (Lemma 7.1 of [@AK]) *There exists $C=C(n,g(0))$ such that as long as the solution exists,$$|Rm|\leq\frac{C}{\psi^2}.$$*
\(3) (Lemma 5.6 and Lemma 7.2 of [@AK]) *If the left polar cap is strictly concave (i.e., $\psi_{ss}<0$)at initial, then as long as the solution exists, left polar cap exists and remains strictly concave, and $D=\lim_{t\nearrow T}\psi(x_-(t),t)$ exists. Furthermore, no singularity occurs on the left polar cap if $D >
0$.*
\(4) (Lemma 9.1 of [@AK]) *There exists $C=C(n,g_0)$ such that $$\frac{K}{L}[logL+2-logL_{min}(0)]\leq C,$$ where $K=-K_0=\frac{\psi_{ss}}{\psi}$ and $L=K_1=\frac{1-\psi_s^2}{\psi^2}$.*$$\eqno \#$$
[ ****[ 3. Classification of Shrinking Solitons]{}****]{}
To understand the structure of singularities, one usually needs to get a classification for gradient shrinking solitons. In [@P2], Perelman obtained a complete classification for nonnegatively curved gradient shrinking soliton in dimension three. An open question is how to generalize Perelman’s classification to higher dimensions. In the next proposition, we obtain such a classification for the class of rotationally symmetric solitons.
0.2cm[**Proposition 3.1**]{} *Let $(M,
g_{ij}(t))$, $-\infty <t<0$, be a nonflat gradient shrinking soliton to the Ricci flow on a complete $(n+1)$-dimensional manifold and assume the metric $g_{ij}(t)$ is rotationally symmetric. Suppose $(M,g_{ij}(t))$ has bounded and nonnegative sectional curvature and is $\kappa$-noncollapsed on all scales for some $\kappa>0$. Then $(M,g_{ij}(t))$ is one of the followings:*
\(i) *the round sphere $S^{n+1}$;*
\(ii) *the round infinite cylinder $(-\infty,+\infty) \times
S^n $.*
0.1cm [**Proof.**]{} Note that for a rotationally symmetric metric, the nonnegativity of sectional curvatures is equivalent to the nonnegativity of curvature operator. Indeed, we can choose a coordinate system $(x^0,x^1,\cdot\cdot\cdot,x^n)$ (where $x^0$ is the radial direction and $x^i, i = 1,
\cdot\cdot\cdot, n,$ are the spherical directions) on $M$ such that all components of the Riemannian curvature tensor vanish in the coordinate system except the sectional curvatures $R_{i0i0}=\psi^2K_0$ and $R_{ijij}=\psi^4K_1$ $ (i\neq j)$, and then the equivalence follows directly from Proposition 1.1 and 1.2 of [@PP].
Firstly, we consider the case that the gradient shrinking soliton is compact and has strictly positive sectional curvature everywhere. By the Theorem 1 in [@BW] we see that the compact gradient shrinking soliton is getting round and tends to a space form (with positive constant curvature) as the time tends the maximal time $t=0$. Since the shape of a gradient shrinking Ricci soliton is unchanging up to reparameterizations and homothetical scalings, the gradient shrinking soliton has to be the round $(n+1)$-sphere $S^{n+1}$.
Next, we consider the case that the sectional curvature of the nonflat gradient shrinking soliton vanishes somewhere. Note that a rotationally symmetric metric is defined on $I \times S^n$ for some interval $I$. By Hamilton’s strong maximum principle in [@Ha86], we know that the soliton splits off a line and then the soliton is the round cylinder $R \times S^n $. (We remark that $R^k \times
S^{n+1-k}$ is not rotationally symmetric if $k>1$.)
Finally we want to exclude the case that the gradient shrinking soliton is noncompact and has strictly positive sectional curvature everywhere.
Suppose there is a complete $(n+1)$-dimensional noncompact $\kappa$-noncollapsed gradient shrinking soliton $g_{ij}(t)$, $-\infty<t<0$, satisfies $$\nabla_i\nabla_j
f+R_{ij}+\frac{1}{2t}g_{ij}=0,\ \ \mbox{on}\ -\infty<t<0, \eqno
(3.1)$$ everywhere for some function $f$ and $g(t)=ds^2+\psi^2(s,t)g_{can}$ and with bounded and positive sectional curvature at each time $t\in(-\infty,0)$. Let us consider the shrinking soliton at the time $t=-1$. Arbitrarily fix a point $x_0$ in $M$. By the same arguments as in the proof of Lemma 1.2 of Perelman [@P2] (or see the proof of Lemma 6.4.1 of [@CZ] for the details), one has the followings:
\(1) at large distance from the fixed point $x_0$ the function $f$ has no critical point, and its gradient makes small angle with the gradient of the distance function from $x_0$;
\(2) at large distance from $x_0$, the scalar curvature $R$ is strictly increasing along the gradient curves of $f$, and $$\limsup_{d_{(-1)}(x,x_0)\rightarrow+\infty}R(x,-1) \leq \frac{n}{2};$$
\(3) the volume of the level set of $f$ satisfies $$Vol\{f=a\}<Vol(S^n(\sqrt{2(n-1)}))\eqno(3.2)$$ for all large enough $a$.
In the three-dimension case, Perelman (in Lemma 1.2 of [@P2]) argued by using Gauss-Bonnet formula to the level set $\{f=a\}$ to derive a contradiction. But now we are considering the general dimensional case, in particular, the (generalized) Gauss-Bonnet formulas are not available. So we need a new argument in the following.
By using Gauss equation and (3.1), the intrinsic sectional curvature $\tilde{R}_{ijij}$ of the level set $\{f=a\}$ can be compute as $$\arraycolsep=1.5pt\begin{array}{rcl}
\tilde{R}_{ijij}&=&R_{ijij}+(h_{ii}h_{jj}-h_{ij}^2)\\[4mm]
&=&R_{ijij}+\frac{1}{|\nabla f|^2}(f_{ii}f_{jj}-f_{ij}^2)\\[4mm]
&\leq& R_{ijij}+\frac{1}{4|\nabla f|^2}(f_{ii}+f_{jj})^2\\[4mm]
&=&R_{ijij}+\frac{1}{4|\nabla f|^2}(1-R_{ii}-R_{jj})^2.
\end{array}\eqno(3.3)$$
Denote by $X=\frac{\nabla f}{|\nabla f|}$ the unit normal vector to the level set $\{f=a\}$. Then set $$X=\delta^0\frac{\partial}{\partial x^0}+\delta^\alpha\frac{\partial}{\partial
x^\alpha},$$ and $$e_i=u_i^0\frac{\partial}{\partial
x^0}+u_i^\alpha\frac{\partial}{\partial x^\alpha},\qquad
i=1,2,\cdot\cdot\cdot,n.$$ where the summation convention of summing over repeated indices is used and $\{x^0,x^1,\cdot\cdot\cdot,x^n\}$ is the local coordinate on the $(n+1)$-dimensional rotationally symmetric manifold with $g=ds^2+\psi^2g_{can}$ with $x^0=s\in R$ and $P=(x^1,\cdot\cdot\cdot,x^n)\in S^n$ and $g_{\alpha\beta}=\delta_{\alpha\beta}$ at $(s,P)$. In these coordinates all components of the Riemann tensor and Ricci tensor vanish except $R_{\alpha 0\alpha 0}=K_0$ and $R_{\alpha\beta\alpha\beta}=K_1 (\alpha\neq\beta)$ and $R_{00}=nK_{0}$ and $R_{\alpha\alpha}=K_0+(n-1)K_1,$ ($\alpha
=1,2,\cdot\cdot\cdot,n$) where $K_0=-\frac{\psi_{ss}}{\psi}$ and $K_1=\frac{1-\psi_s^2}{\psi^2}$. And then the scalar curvature $R=2nK_0+n(n-1)K_1$. So we have $$\arraycolsep=1.5pt\begin{array}{rcl}
R_{ijij}&=&R(u_i^\alpha \frac{\partial}{\partial
x^\alpha},u_j^\beta\frac{\partial}{\partial
x^\beta},u_i^\gamma\frac{\partial}{\partial x^\gamma},
u_j^\eta\frac{\partial}{\partial x^\eta})\\[4mm]
&=&\sum_{\alpha\beta\gamma\eta}u_i^\alpha u_j^\beta u_i^\gamma
u_j^\eta R_{\alpha\beta\gamma\eta}\\[4mm]
&=&\sum_{\alpha\beta}(u_i^\alpha u_j^\beta )^2 R_
{\alpha\beta\alpha\beta}-\sum_{\alpha\beta}u_i^\alpha u_j^\beta
u_i^\beta u_j^\alpha R_{\alpha\beta\alpha\beta}\\[4mm]
&=&\sum_{\beta=1}^n[(u_i^0)^2(u_j^\beta)^2+(u_i^\beta)^2(u_j^0)^2]K_0+\sum_{\alpha,\beta=1}
^n(u_i^\alpha u_j^\beta)^2K_1\\[4mm]
&&-2\sum_{\beta=1}^n u_i^0u_j^0u_i^\beta u_j^\beta
K_0-\sum_{\alpha,\beta=1}^n u_i^\alpha u_j^\beta u_i^\beta
u_j^\alpha K_1\\[4mm]
&=&(\sum_{\beta=1}^n[(u_i^0)^2(u_j^\beta)^2+(u_i^\beta)^2(u_j^0)^2]+2(u_i^0)^2(u_j^0)^2)K_0\\[4mm]
&&+\sum_{\alpha,\beta=1}^n [(u_i^\alpha u_j^\beta)^2-u_i^\alpha
u_j^\beta u_i^\beta u_j^\alpha]K_1\\[4mm]
&=&[(u_i^0)^2+(u_j^0)^2]K_0+\sum_{\alpha,\beta=1}^n[(u_i^\alpha
u_j^\beta)^2-u_i^\alpha u_j^\beta u_i^\beta u_j^\alpha]K_1\\[4mm]
&=&[(u_i^0)^2+(u_j^0)^2]K_0+[(1-(u_i^0)^2)(1-(u_j^0)^2)-(u_i^0)^2(u_j^0)^2]K_1\\[4mm]
&=&[(u_i^0)^2+(u_j^0)^2]K_0+[1-(u_i^0)^2-(u_j^0)^2]K_1.
\end{array}\eqno(3.4)$$ where in the fifth and sixth equalities we used $$\sum_{\beta=1}^n u_i^\beta u_j^\beta=-u_i^0 u_j^0$$ and $$\sum_{\beta=0}^n (u_j^\beta)^2=1$$ since $\{e_1,\cdots,e_n\}$ is an orthonormal basis of the level set $\{f=a\}$. Then by (3.4) we have $$\arraycolsep=1.5pt\begin{array}{rcl}
R_{ii}&=&R_{iXiX}+\sum_{j=1,j\neq i}^n R_{ijij}\\[4mm]
&=&[(u_i^0)^2+(\delta^0)^2]K_0+[1-(u_i^0)^2-(\delta^0)^2]K_1\\[4mm]
&&+\sum_{j=1,j\neq i}^n [(u_i^0)^2+(u_j^0)^2]K_0+\sum_{j=1,j\neq
i}^n [1-(u_i^0)^2-(u_j^0)^2]K_1\\[4mm]
&=&[n(u_i^0)^2+1-(u_i^0)^2]K_0+[n-n(u_i^0)^2-1+(u_i^0)^2]K_1\\[4mm]
&=&[1+(n-1)\varepsilon]K_0+[(n-1)(1-\varepsilon)K_1]
\end{array}\eqno(3.5)$$ where $\varepsilon=(u_i^0)^2\ll 1$, if $a$ is large enough.
Obviously by (3.5) we get $$R_{ii}<2K_0+(n-1)K_1=\frac{R}{n}<\frac{1}{2},$$ and then $$1-R_{ii}-R_{jj}>0.$$ Again by (3.5) we know $$R_{ii}>K_0+(n-1)(1-\varepsilon)
K_1\eqno(3.6)$$ and by (3.4) we know $$R_{ijij}<2\varepsilon
K_0+(1-\varepsilon)K_1.\eqno(3.7)$$ Hence by (3.3), (3.6) and (3.7) $$\arraycolsep=1.5pt\begin{array}{rcl}
\tilde{R}_{ijij}&\leq& R_{ijij}+\frac{1}{4|\nabla
f|^2}(1-R_{ii}-R_{jj})^2\\[4mm]
&<&2\varepsilon K_0+(1-\varepsilon)K_1+\frac{1}{4|\nabla
f|^2}[1-2(K_0+(n-1)(1-\varepsilon)K_1)]^2\\[4mm]
&=&2\varepsilon
K_0+\frac{1}{2(n-1)}\{1-2(K_0+(n-1)(1-\varepsilon)K_1+2(n-1)(1-\varepsilon)K_1\\[4mm]
&&-[1-2(K_0+(n-1)(1-\varepsilon)K_1)]\}+\frac{1}{4|\nabla
f|^2}[1-2(K_0+(n-1)(1-\varepsilon)K_1)]^2\\[4mm]
\end{array}$$ $$\arraycolsep=1.5pt\begin{array}{rcl}
&=&\frac{1}{2(n-1)}\{1-2(1-2\varepsilon
(n-1))K_0-[1-2(K_0+(n-1)(1-\varepsilon)K_1)]\\[4mm]
&&+\frac{n-1}{2|\nabla f|^2}[1-2(K_0+(n-1)(1-\varepsilon)K_1)]^2\}\\[4mm]
&<&\frac{1}{2(n-1)}
\end{array}\eqno(3.8)$$ for sufficiently large $a$, since $2(1-2\varepsilon
(n-1))K_0>0$ and $1-2(K_0+(n-1)(1-\varepsilon)K_1)>0$ and $|\nabla
f|$ is large as $a$ large. Then by (3.8) and the volume comparison theorem we know $$Vol\{f=a\}>Vol(S^n(\sqrt{2(n-1)}))$$ for large enough $a$ and then it is a contradiction with (3.2).
Therefore we have proved the proposition.$$\eqno \#$$
[ ****[ 4. Type II Singularity Happens]{}****]{}
Suppose we have a family of rotationally symmetric solutions $$\{(S^{n+1},g_\alpha(t))| \alpha\in[0,1]\}$$ of the Ricci flow with $g_\alpha(0)=ds^2+\psi_\alpha^2g_{can}$, $\alpha \in [0,1]$, where $g_{can}$ is the standard metric of constant sectional curvature 1 on $S^{n}$. We specify the initial metrics as follows.
When $\alpha=1$, let the initial metric $g_1(0)$ be a symmetric dumbbell with two equally-sized hemispherical regions joined by a thin neck. By the work in [@AK], we can assume the two hemispheres are suitably large and the neck is suitably thin so that this initial metric $g_1(0)$ leads to a neckpinch singularity of the Ricci flow at some time $T_1<+\infty$. (see Figure 1.)
(350,90) (50,50)(55,85)(90,90) (50,50)(55,15)(90,10)(90,90)(65,50)(90,10)(90,90)(115,50)(90,10) (90,90)(118,83)(130,65) (90,10)(118,17)(130,35)(180,55)(174,50)(180,45)(180,55)(186,50)(180,45) (130,65)(140,60)(180,55) (130,35)(140,40)(180,45) (180,55)(220,60)(230,65) (180,45)(220,40)(230,35) (240,75)(246,83)(270,90)(230,65)(235,72)(240,75) (230,35)(240,17)(270,10)(270,90)(295,50)(270,10)(270,90)(245,50)(270,10) (270,90)(305,85)(310,50) (270,10)(305,15)(310,50)
When $\alpha=0$, let the initial metric $g_0(0)$ be a lopsided and degenerate dumbbell where $g_0(0)=ds^2+\psi_0^2g_{can}$ with $\psi_0(0)$ has only one bump and it is nonincreasing on the right polar cap and strictly concave on the left polar cap. (see Figure 2.)
(350,105) (50,60)(55,95)(90,105) (50,60)(55,25)(90,15)(90,105)(65,60)(90,15)(90,105)(115,60)(90,15) (90,105)(115,100)(130,85)(90,15)(115,20)(130,35)(130,85)(115,60)(130,35) (130,85)(140,80)(200,70) (130,35)(140,40)(200,50)(130,85)(145,60)(130,35) (200,70)(237,68)(240,60) (200,50)(237,52)(240,60)
Clearly, we may choose the $g_1(0), g_0(0)$ to have positive scalar curvatures. Let $\{ g_\alpha(0) \ | \ \alpha \in [0,1] \}$ (see Figure 3.) be a smooth family of dumbbells (including degenerate dumbbells) connecting the $g_1(0)$ to the $g_0(0)$ and satisfying the followings: 0.1cm(i) for each $\alpha \in
[0,1]$, $\psi_\alpha(0)$ has exactly two bumps or one bump, 0.1cm(ii) for each $\alpha \in [0,1]$, $(\psi_\alpha)_s(0)$ has only finitely many zeros, and satisfies $$|(\psi_\alpha)_s|(0) \leq 1,$$ 0.1cm(iii) for each $\alpha \in [0,1]$, $\psi_\alpha(0)$ is strictly concave on the left polar cap, 0.1cm(iv) each initial metric $g_\alpha(0)$, $\alpha \in [0,1]$, has positive scalar curvature.
(350,110) (50,60)(55,95)(90,100) (50,60)(55,25)(90,20)(90,100)(70,60)(90,20)(90,100)(110,60)(90,20) (90,100)(118,93)(130,75) (90,20)(118,27)(130,45) (130,75)(140,70)(180,65) (130,45)(140,50)(180,55) (180,65)(220,70)(230,75) (180,55)(220,50)(230,45) (240,85)(246,93)(270,100)(230,75)(235,82)(240,85) (230,45)(240,27)(270,20) (270,100)(305,97)(310,60) (270,20)(305,23)(310,60) (40,60)(48,105)(90,110) (40,60)(48,15)(90,10) (90,110)(115,105)(130,85)(90,10)(115,15)(130,35) (130,85)(140,80)(205,70) (130,35)(140,40)(205,50) (205,70)(237,68)(240,60) (205,50)(237,52)(240,60) (45,60)(50,100)(90,105) (45,60)(50,20)(90,15) (90,105)(118,98)(130,80) (90,15)(118,22)(130,40) (130,80)(140,75)(180,70) (130,40)(140,45)(180,50)(180,70)(220,75)(230,80) (180,50)(220,45)(230,40)(230,80)(235,85)(260,90) (230,40)(235,35)(260,30)(260,90)(285,85)(290,60) (260,30)(285,35)(290,60) (300,10)[(0,0)\[bl\][$g_1(0)$]{}]{} (310,20)[(-1,1)[10]{}]{} (260,5)[(0,0)\[bl\][$g_\alpha(0)$]{}]{} (260,10)[(0,1)[20]{}]{} (200,20)[(0,0)\[bl\][$g_0(0)$]{}]{} (210,30)[(1,1)[25]{}]{}
Since the scalar curvature is positive, each solution $g_\alpha(t)$, $\alpha \in [0,1]$, will exist up to a maximal time $T_\alpha<+\infty$ and develops a singularity. The main purpose of this section is to show that there exists $\alpha_0\in[0,1)$ such that the solution $g_{\alpha_0}(t)$, with the metric $g_{\alpha_0}(0)$ as initial datum, develops a Type II singularity. We remark that a Type II singularity might occur in such family of metrics had been conjectured and the intuition had already described in [@CK] and [@T].
Let us first consider the case that the solutions with degenerate dumbbells as initial data.
0.2cm[**Lemma 4.1**]{} *Suppose $g_\alpha (t)$ is a rotationally symmetric solution of the Ricci flow on $S^{n+1}$ with $g_\alpha (0)=\varphi(x,0)^2dx^2+\psi_\alpha ^2(x,0)g_{can}$, $x\in[-1,1]$. If at initial, the scalar curvature $R^{(\alpha)}>0$, $\psi_\alpha(x,0)$ has only one bump, it is nonincreasing on the right polar cap and is strictly concave on the left polar cap, and $|(\psi_\alpha)_s|(x,0) \leq 1$ on $[-1,1]$, then either the solution $g_\alpha (t)$ develops a Type II singularity or it shrinks to a round point.*
0.1cm [**Proof.**]{} By the assumption of $R^{(\alpha)}>0$ and apply the maximum principle to the evolution equation of the scalar curvature $$\frac{\partial R}{\partial
t}=\Delta R+2|Ric|^2$$ we know that the maximal time $T<+\infty$.
Now we consider the geometric quantity $s$ defined by $$s(x,t)=\int_0^x \varphi(x,t)dx.$$ Then the metric can be written as $$g=ds^2+\psi_\alpha^2(s,t)g_{can}.$$ In the following if we write a relation of the type $f=f(s)$ , it is to be understood as shorthand for $f=f(s(x,t))$ for evolving metrics. Since $\psi_\alpha (\pm 1,t)=0$, we know that for any time $0\leq t<T$, the bump exists. By the standard Sturmian comparison [@An], we know that $\psi_\alpha(x,t)$ also has a unique bump for each $t \in [0, T)$. Let $x_\ast (t)$ denote the unique bump. By Proposition 2.1(3), we can define $$D=\lim_{t\nearrow T}\psi_\alpha
(x_\ast (t),t).$$ We divide it into two cases: 0.1cm [**Case 1:**]{} $D>0$.
In this case, by Proposition 2.1(3) and the assumption that $\psi_\alpha$ is strictly concave on the left polar cap, we know that no singularity occurs on the left polar cap. Thus the singularity must occur on the right polar cap. Take the maximal points $(\tilde{P}_m,t_m)$, i.e., choose the points $(\tilde{P}_m,t_m)$ such that $$|Rm(\tilde{P}_m,t_m)|=\sup_{t\leq t_m,Q\in
S^{n+1}}|Rm(Q,t)|\rightarrow+\infty,$$ as $m\rightarrow+\infty$.
Since for any time $t\in[0,T)$ we have $(\psi_\alpha)_s(P,t)=-1$, where $P$ is the pole of the right polar cap (i.e. the point with $x=1$), we can choose the nearest point $P_m'$ to $P$ such that $(\psi_\alpha)_s(P_m',t_m)=-\frac{1}{2}$. If $d_{t_m}(P_m',P)>d_{t_m}(\tilde{P}_m,P)$, then we set $P_m=\tilde{P}_m$, otherwise set $P_m=P_m'$. Clearly in the region between $P_m$ and $P$, we have $|(\psi_\alpha)_s|\geq\frac{1}{2}$.
We first claim that the curvature at $(P_m,t_m)$ is comparable to the curvature at the maximal point $(\tilde{P}_m,t_m)$. Indeed, if $P_m=\tilde{P}_m$, then there is nothing to show. If $P_m\neq
\tilde{P}_m$, then by the estimate in Proposition 2.1(2) and by the condition that $\psi_\alpha$ is nonincreasing on the right polar cap and by the choice of the point $P_m$, we know that $$|Rm(\tilde{P}_m,t_m)|\leq
\frac{C}{\psi_\alpha^2(\tilde{P}_m,t_m)}\leq
\frac{C}{\psi_\alpha^2(P_m,t_m)}.$$ On the other hand $$K_1(P_m,t_m)=\frac{1-(\psi_\alpha)_s^2(P_m,t_m)}
{\psi_\alpha^2(P_m,t_m)}=\frac{3}{4\psi_\alpha^2(P_m,t_m)}.$$ So $$|Rm(P_m,t_m)|\geq K_1(P_m,t_m)=
\frac{3}{4\psi_\alpha^2(P_m,t_m)}\geq
\frac{3}{4C}|Rm(\tilde{P}_m,t_m)|.$$ Obviously since $(\tilde{P}_m,t_m)$ is the maximal point, we have $$|Rm(P_m,t_m)|\leq
|Rm(\tilde{P}_m,t_m)|.$$ So the curvatures at $(P_m,t_m)$ and $(\tilde{P}_m,t_m)$ are comparable, where we used the definition of $|Rm|$ to be the largest absolute value of the eigenvalues of the curvature operator $Rm$.
Applying the maximum principle to the evolution equation of the scalar curvature $R$:$$\frac{\partial R}{\partial t}=\Delta
R+2|Ric|^2$$ and using the pinching estimate in Proposition 2.1(4) we get $$\frac{dR_{max}}{dt}\leq CR_{max}^2.$$ Then $$R_{max}(t)\geq
\frac{C}{T-t}$$ for some constant $C$.
We now argue by contradiction to show that the solution develops a Type II singularity in this case.
Suppose not, then the singularity is of Type I. That is, there exists some constant $C>0$ such that $$\frac{C^{-1}}{T-t_m}\leq
R^{(\alpha)} (P_m,t_m)\leq \frac{C}{T-t_m}.\eqno(4.1)$$ Define $$g_{ij}^{(m)}(\cdot,t)=R^{(\alpha)}(P_m,t_m)(g_\alpha)_{ij}(\cdot,t_m+\frac{t}{R^{(\alpha)}(P_m,t_m)}),$$ for $t\in [-t_mR^{(\alpha)}(P_m,t_m),0]$. Then we claim that the distance from $P_m$ to the pole $P$ measured in the rescaled metric $g_{ij}^{(m)}(\cdot,0)$ is bounded. Indeed, by the estimate in Proposition 2.1(2), we know$$|Rm|\leq\frac{C}{\psi_\alpha^2}$$ for some constant $C$. Then we have $$\psi_\alpha^2(P_m,t_m)\leq\frac{C}{|Rm(P_m,t_m)|}$$ and using $|(\psi_\alpha)_s|
\geq\frac{1}{2}$ in the region between $P_m$ and $P$, we have $$d_{t_m}(P_m,P)\leq \frac{\psi_\alpha(P_m,t_m)}{\frac{1}{2}}
\leq\frac{2C}{\sqrt{|Rm(P_m,t_m)|}} \eqno(4.2)$$ where $d_t$ is the distance measured with the metric $g_\alpha(t)$. Therefore by the pinching estimate in Proposition 2.1(4) we know that the distance from $P_m$ to the pole $P$ measured in the rescaled metric $g_{ij}^{(m)}(\cdot,0)$ is bounded.
The rescaled $g_{ij}^{(m)}(t)$ is a solution of the Ricci flow defined for $t\in[-t_mR^{(\alpha)}(P_m,t_m),0]$ and $0<R^{(m)}(\cdot,t)\leq 1$ and $R^{(m)}(P_m,0)=1$ and has bounded curvature. After taking a subsequence of $g_{ij}^{(m)}$, we can assume that the marked manifold $(S^{n+1},g_{ij}^{(m)}(t),P)$ converges to a marked manifold $(R^{n+1},g_{ij}(t),P), -\infty <t\leq 0$, which is a solution of the Ricci flow on $R^{n+1}$ with nonnegative curvature operator ( by the pinching estimate in Proposition 2.1(4) ), has bounded curvature with $R(P_\ast,0)=1$ at some point $P_\ast$, and is $\kappa$-noncollapsed for all scales. So the limit is a nonflat ancient $\kappa$-solution on $R^{n+1}$.
The reduced distance, due to Perelman [@P1], is defined by $$\arraycolsep=1.5pt\begin{array}{rcl}
&&l^{(\alpha)}(q,\tau)=\frac{1}{2\sqrt{\tau}}\inf\{\int_0^\tau\sqrt{s}
(R^{(\alpha)}(\gamma(s),t_m-s)+|\dot{\gamma}(s)|^2_{(g_\alpha)_{ij}(t_m-s)})ds|\\[4mm]
&&\hskip 3cm\gamma:[0,\tau]\rightarrow S^{n+1}\;\mbox{with}
\;\gamma(0)=P,\gamma(\tau)=q\}.
\end{array}$$ where $\tau=t_m-t, \mbox{for }\;t<t_m $. Then by the Type I assumption, we have $$\arraycolsep=1.5pt\begin{array}{rcl}
l^{(\alpha)}(P,\tau)&\leq&\frac{1}{2\sqrt{\tau}}\int_0^\tau\sqrt{s}\frac{C}{T-t_m+s}ds\\[4mm]
&\leq&\frac{C}{2\sqrt{\tau}}\int_0^\tau \frac{1}{\sqrt{s}}ds\\[4mm] &=&
C.
\end{array}\eqno(4.3)$$
We can now use Perelman’s backward limit argument in Proposition 11.2 of [@P1] to choose a sequence of times $t_k\rightarrow
-\infty$ such that the scaling of $g_{ij}(\cdot,t)$ around $P$ with the factor $|t_k|^{-1}$ and with the times $t_k$ shifting to the new time zero converge to a nonflat gradient shrinking soliton in $C_{loc}^\infty$ topology. Indeed, in the Proposition 11.2 of [@P1], Perelman takes a limit around some points $q(\tau)$ where the reduced distance at $q(\tau)$ are uniformly bounded above by $(n+1)/2$. Instead, in our situation, we want to take a backward limit around the fixed point $P$. By inspecting the proof of Proposition 11.2 of [@P1] (see also the proof of Theorem 6.2.1 of [@CZ] for the details), one only needs to have a uniform upper bound for the reduced distance at the fixed point $P$. This is just our estimate (4.3) by the Type I assumption. Then the same argument as Perelman in section 11.2 in [@P1] applies to the present situation.
By combining with the above Proposition 3.1 and noting that the gradient shrinking soliton $\bar{g}_{ij}$ is noncompact, we conclude that the backward limit is $S^n\times R$. But since the limit is taking around the pole and the metric is rotationally symmetric, it can not be $S^n\times R$, so we get a contradiction! Hence we have proved that the singularity is of Type II. 0.1cm [**Case 2:**]{} $D=0$.
In this case, if the singularity is of Type II, then there is nothing to prove. Thus we may assume that the singularity is of Type I. By the same argument as in the case 1, we can first take a rescaling limit around the pole $P$ at the maximal time $T$ to get an ancient $\kappa$-solution and then take a backward limit around the pole $P$ again to get a nonflat gradient shrinking soliton. If the shrinking soliton is compact, then by Proposition 3.1 we know that it is the round $S^{n+1}$. This implies that the original solution shrinks to a round point as the time tends to the maximal time $T$. While if the shrinking soliton is noncompact, then by Proposition 3.1 we know that it is $S^n\times R$; so the same reason in the proof of the case 1 gives a contradiction! Therefore we have proved Lemma 4.1. $$\eqno \#$$
0.2cm[**Lemma 4.2**]{} *The set $A_1$ of $\alpha\in[0,1]$ such that the initial metric $g_\alpha(0)$ leads to a neckpinch singularity of the Ricci flow at some time $T_\alpha<+\infty$ is open in $[0,1]$.*
0.1cm [**Proof.**]{} Obviously it is not empty for $1\in A_1$.
Suppose $\alpha\in A_1$, then we claim that $\psi_\alpha(0)$ has two bumps. Otherwise it has only one bump. Since it satisfies the assumptions of Lemma 4.1 by the above conditions (i), (ii), (iii) and (iv), then we know that either it shrinks to a round point, or forms a Type II singularity. Consequently, the initial metric does not lead to a neckpinch for such $\alpha$. This contradicts with $\alpha\in A_1$. Similarly, $\psi_\alpha(t)$ has two bumps as long as the solution exists. Take a small perturbation $g_\alpha^{(k)}(0)$ of $g_\alpha(0)$ (in $C^3$ topology). Then $g_\alpha^{(k)}(0)$ still has two bumps. We need to show that $g_\alpha^{(k)}(0)$ leads to a neckpinch singularity at the maximal time $T_\alpha^{(k)}<+\infty$.
Since $g_\alpha^{(k)}(0)$ are very close to $g_\alpha(0)$ in $C^3$ topology, the scalar curvatures of the metrics $g_\alpha^{(k)}(0)$ have a uniform positive lower bound. Thus it follows from the evolution equation of the scalar curvature that the maximal times $T_\alpha^{(k)}$ are uniformly bounded. After passing to a subsequence, we can then assume that $T_\alpha^{(k)}\rightarrow
\tilde{T}$ as $k\rightarrow \infty$.
0.2cm[**Claim:**]{} $\tilde{T}\geq T$.
Indeed, suppose not, then there exists $\varepsilon >0$, such that $$\tilde{T}-\varepsilon<T_\alpha^{(k)}<\tilde{T}+\varepsilon<T-\varepsilon<T$$ for all sufficiently large $k$. Consider the time interval $[0,\tilde{T}-\varepsilon]$. By the assumption that $g_\alpha^{(k)}(0)$ is sufficiently close to $g(0)$ and $g(t)$ is smooth on $[0,\tilde{T}-\varepsilon]$, we first show that the curvature of $g_\alpha^{(k)}(t)$ is uniformly bounded on $[0,\tilde{T}-\varepsilon]$ for all sufficiently large $k$.
For each $0\leq t \leq \tilde{T}-\varepsilon$, set $$M(t)=\sup\{|Rm^{(k)}(x,t)| | k\geq 1, x\in S^{n+1}\}$$ and $$t_0=\sup\{t\geq 0| M(t)<+\infty\},$$ where $Rm^{(k)}$ denotes the curvature of $g_\alpha^{(k)}(t)$. We want to show that $t_0=\tilde{T}-\varepsilon$.
By Shi’s local derivative estimate in [@Sh89], we know that $t_0>0$. Suppose $t_0<\tilde{T}-\varepsilon$, then for any small $\varepsilon'>0$, consider the time interval $[0,t_0-\varepsilon']$. By the above definition of $M(t)$, we know that the curvature of $g_\alpha^{(k)}(t)$ is uniformly bounded by $M(t_0-\varepsilon')$ on $[0,t_0-\varepsilon']$. Take a limit of $g_\alpha^{(k)}(t)$ and by the uniqueness of the solution to the Ricci flow [@Ha82] or [@Ha95F], we get the limit must be the original solution $g(t)$ on $[0,t_0-\varepsilon']$. So we have the curvature of $g_\alpha^{(k)}(t)$ is uniformly bounded by some constant $C$ which does not depend on $\varepsilon'$. Then by Shi’s local derivative estimate in [@Sh89] again, we know that the curvature of $g_\alpha^{(k)}(t)$ is uniformly bounded on $[0,t_0-\varepsilon'+\frac{1}{C}]$. By choosing $\varepsilon'$ small enough, we get $t_0-\varepsilon'+\frac{1}{C}>t_0$ and then it is a contradiction! So we have proved that $t_0=\tilde{T}-\varepsilon$, that is the curvature of $g_\alpha^{(k)}(t)$ is uniformly bounded on $[0,\tilde{T}-\varepsilon]$ for all sufficiently large $k$. Similarly as above, we can take a limit of $g_\alpha^{(k)}(t)$ and by the uniqueness of the Ricci flow [@Ha82] or [@Ha95F], we get the limit must be $g(t)$ on $[0,\tilde{T}-\varepsilon]$. Using $g(t)$ is smooth on $[0,\tilde{T}]$, we get the curvature of $g_\alpha^{(k)}(t)$ is uniformly bounded by some constant $C'$ which does not depend on $\varepsilon$. Again by Shi’s local derivative estimate in [@Sh89], we know that the curvature of $g_\alpha^{(k)}(t)$ is uniformly bounded on $[0,\tilde{T}-\varepsilon+\frac{1}{C'}]$ for all sufficiently large $k$. Choose $\varepsilon>0$ sufficiently small and using that $T_\alpha^{(k)}\rightarrow \tilde{T}$ as $k\rightarrow\infty$, we get $\tilde{T}-\varepsilon+\frac{1}{C'}>T_\alpha^{(k)}$ for all sufficiently large $k$, which contradicts with the definition of the maximal time. So $\tilde{T}\geq T$.
Next we show that each $g_\alpha^{(k)}(0)$ leads to a neckpinch singularity at the maximal time $T_\alpha^{(k)}<+\infty$.
For all sufficiently small $\varepsilon>0$, since $g_\alpha(0)$ leads to a neckpinch, we have $$\frac{(r_\alpha)_{min}(T-\varepsilon)}
{\psi_\alpha(x_\pm(T-\varepsilon),T-\varepsilon)}\ll 1.\eqno(4.4)$$ By the assumption that $g_\alpha^{(k)}(0)$ is sufficiently close to $g_\alpha(0)$ and $T_\alpha^{(k)}\rightarrow \tilde{T}\geq T$, we know that as $k$ large enough, $T_\alpha^{(k)}>T-\varepsilon$ and $g_\alpha^{(k)}(T-\varepsilon)$ is sufficiently close to $g_\alpha(T-\varepsilon)$. So $$\frac{(r_\alpha^{(k)})_{min}(T-\varepsilon)}
{\psi_\alpha^{(k)}(x_\pm(T-\varepsilon),T-\varepsilon)}\ll
1.\eqno(4.5)$$ In views of the work of Angenent and Knopf [@AK], we know that if we have a rotationally symmetric $g(0)=ds^2+\psi^2(0)g_{can}$ on $S^{n+1}$ which has two bumps $x_\pm(0)$ and $\frac{r_{min}(0)}{\psi(x_\pm(0),0)}<C^{-1}$ for some universal constant $C>0$, (for example we can take $C=100$), then it leads to a neckpinch singularity. By (4.5), we know that $g_\alpha^{(k)}(0)$ leads to a neckpinch singularity. Therefore we proved that $A_1$ is open in $[0,1]$. $$\eqno \#$$
In the next Proposition, following Perelman’s Theorem 12.1 in [@P1], we will give the singularity structure for the rotationally symmetric solutions of the Ricci flow. 0.2cm[**Proposition 4.3**]{} *Suppose $g_{ij}(t)$, $t\in[0,T)$, is a rotationally symmetric solution of the Ricci flow on $S^{n+1}$ with $g(0)=ds^2+\psi^2(0)g_{can}$. If at initial the scalar curvature $R>0$, then for any given $\varepsilon>0$, there exists $K=K(\varepsilon, g(0))>\max\{2\varepsilon^{-1},
Q(\frac{3}{4}T)\}>0$, where $Q(\frac{3}{4}T)$ denotes the upper bound of the curvature for the times $t\leq\frac{3}{4}T$, such that for any point $(x_0,t_0)$ with $t_0\geq\frac{3}{4}T$ and $Q=R(x_0,t_0)\geq K$, the solution in $\{(y,t)|d_{t_0}^2(y,x_0)<\varepsilon^{-2}Q^{-1},
t_0-\varepsilon^{-2}Q^{-1}\leq t \leq t_0\}$ is, after scaling by the factor $Q$, $\varepsilon$-close to the corresponding subset of some orientable ancient $\kappa$-solution, where $\kappa$ is a positive constant depending only on $T$ and the initial metric $g(0)$. Consequently, in the region, we have the following gradient estimates $$|\nabla(R^{-\frac{1}{2}})|\leq \eta \ \ \mbox{ and } \ \
|\frac{\partial}{\partial t}(R^{-1})|\leq \eta$$ for some constant $\eta=\eta(\kappa)>0.$* 0.1cm [**Proof.**]{} This is just a higher dimensional version of Perelman’s singularity result (Theorem 12.1 of [@P1]) for the rotationally symmetric class. In Theorem 12.1 of [@P1], Perelman obtained this singularity structure result for any three-dimensional solution. For the details, one can consult [@KL] (from page 83 to 88) or [@CZ] (from page 399 to 405). By inspecting Perelman’s argument, when one tries to generalize Perelman’s singularity structure result to higher dimensions, one only needs to have a higher-dimensional version of the (three-dimensional) Hamilton-Ivey curvature pinching estimate and shows the canonical neighborhoods of an ancient $\kappa$-solution consisting $\varepsilon$-necks or $\varepsilon$-caps. For our case, since the metric is rotationally symmetric, the estimate due to Angenent-Knopf in Proposition 2.1(4) gives the desired curvature pinching estimate. While for a rotationally symmetric ancient $\kappa$-solution, it is clear that any canonical neighborhood is either an $\varepsilon$-neck or an $\varepsilon$-cap. So by repeating Perelman’s argument, we obtain the proof of the proposition. $$\eqno \#$$
We can now prove the main theorem.
0.2cm[**Proof of Theorem 1.2.**]{}
Suppose $g_\alpha(t)$ is the family of the solutions to the Ricci flow satisfies the above conditions (i), (ii), (iii) and (iv). We want to show that there exists $\alpha_ 0\in [0,1)$ such that for the solution $g_{\alpha_0}(t)$ of the Ricci flow on $S^{n+1}$ with initial data $g_{\alpha_ 0}(0)=ds^2+\psi_{\alpha_0}^2g_{can}$, exists up to a maximal time $T_{\alpha_0}<+\infty$ and develops a Type II singularity.
Since $g_1(0)=ds^2+\psi_1^2g_{can}$ and by our assumption that $g_1(0)$ leads to a neckpinch singularity. Then by Lemma 4.2, we know that $A_1$ is not empty and open in $[0,1]$. While by Lemma 4.1, we know that the solution $g_0(t)$ with the initial data $g_0(0)$ either develops a Type II singularity or shrinks to a round point, so $0\overline{\in} A_1$. Let $(\alpha,1]$ be a connected component of $A_1$. We want to show that the $\alpha$ is the number we want.
If $g_\alpha(0)$ develops a Type II singularity, then there is nothing to show. So in the following we assume it does not develop a Type II singularity. 0.1cm [**Claim 1:**]{} *$\psi_\alpha(0)$ exactly has two bumps.*
Indeed, if $\psi_\alpha(0)$ has only one bump, then $\psi_\alpha$ is nonincreasing on the right polar cap and by our construction we know that $\psi_\alpha$ is strictly concave on the left polar cap. So by Lemma 4.1, we know either the singularity is of Type II or it shrinks to a round point at the maximal time $T_\alpha<+\infty$. By our assumption that the singularity is not of Type II. So it is shrinking to a round point, and then there exists a time $\tilde{t}<T_\alpha$ close to $T_\alpha$, such that the curvature is positive for $t\geq\tilde{t}$. Whenever $\beta\in (\alpha,1]\subset
A_1$ is sufficiently close to $\alpha$, the metric $g_\beta(0)$ is sufficiently close to the metric $g_\alpha(0)$ ( in the $C^3$ topology). Then by Lemma 4.2 we can choose $\beta\in
(\alpha,1]\subset A_1$ sufficiently close to $\alpha$ so that the maximal time $T_\beta$ of $g_\beta(t)$ satisfies $T_\beta>\tilde{t}
+ (T_\alpha - \tilde{t})/2$; moreover, by the continuous dependence of the initial metric, the curvature operator of $g_\beta(t)$ is also positive at the time $t = \tilde{t}$. Hence by Theorem 1 in [@BW] we know that $g_\beta(0)$ will shrink to a round point at the maximal time $T_\beta<+\infty$, which contradicts with $\beta\in
(\alpha,1] \subset A_1$. Therefore we have proved the Claim 1. 0.1cm [**Claim 2:**]{} *$\psi_\alpha(t)$ exactly has two bumps as long as the solution exists.*
Indeed, since $g_\alpha(t)$ is a rotationally symmetric solution of the Ricci flow on $S^{n+1}$, we know that at the poles of the right and left polar caps $\psi_\alpha(t)=0$ for any time $0\leq t <
T_\alpha$, so there always exists one bump. By the standard Sturmian comparison in [@An] we know that the number of the bumps is nonincreasing in time. Suppose at some time $t_0 \in (0,T_\alpha)$ such that the right-most bump disappeared, then $\psi_\alpha(t_0)$ has only one bump. Thus by Lemma 4.1 and our assumption, it shrinks to a round point. Particularly, there exists a time $t_0<\tilde{t}<T_\alpha$ such that the curvature is positive for all times $t\geq \tilde{t}$. By the same argument as above, we can choose $\beta\in (\alpha,1]\subset A_1$ sufficiently close to $\alpha$ so that the maximal time $T_\beta$ of $g_\beta(t)$ is greater than $\tilde{t}$ and the curvature of $g_\beta(t)$ at the time $t = \tilde{t}$ is also positive. By applying Theorem 1 in [@BW] again we know that $g_\beta(0)$ will shrink to a round point at the maximal time $T_\beta<+\infty$, which also contradicts with $\beta\in (\alpha,1]\subset A_1$. Thus we have proved the Claim 2.
So in the following we always assume that $\psi_\alpha(t)$ has two bumps for all times $t\in [0,T_\alpha)$.
Since at the maximal time $T_\alpha$, the solution $g_\beta(t)$ does not develop a neckpinch. In views of Angenent-Knopf’s result [@AK], the smaller polar cap must collapse. So, without loss of generality, we may assume that singularity occurs on the right polar cap. Similarly as in Lemma 4.1, we first take the maximal points $(\tilde{P}_m,t_m)$ on $S^{n+1}$ (i.e., $|Rm(\tilde{P}_m,t_m)|=\sup_{t\leq t_m,Q\in S^{n+1}}|Rm(Q,t)|$). We then take the nearest point $P_m'$ to the pole $P$ on the right polar cap, such that $(\psi_\alpha)_s(P_m',t_m)=-\frac{1}{2}$. If $d_{t_m}(P_m',P)>d_{t_m}(\tilde{P}_m,P)$, then we set $P_m=\tilde{P}_m$; otherwise we set $P_m=P_m'$. Clearly in the region between $P_m$ and $P$, we have $|(\psi_\alpha)_s|\geq\frac{1}{2}$. 0.1cm Define $$g_{ij}^{(m)}(\cdot,t)=R^{(\alpha)}(P_m,t_m)(g_\alpha)_{ij}(\cdot,t_m+\frac{t}
{R^{(\alpha)}(P_m,t_m)})$$ for $t\in
[-t_mR^{(\alpha)}(P_m,t_m),0]$.
0.1cm [**Claim 3:**]{} *A subsequence of $g_{ij}^{(m)}(\cdot,t)$ around the point $P$ will converge to a nonflat complete ancient $\kappa$-solution on a smooth manifold $M$, where $\kappa$ is some positive constant depending only on the initial metric $g_\alpha(0)$.*
Indeed, if the maximal point $(\tilde{P}_m,t_m)$ is on the right polar cap, then by the estimate in Proposition 2.1(2) and by the condition that $\psi_\alpha$ is nonincreasing on the right polar cap and by the choice of the point $P_m$, we know that $$|Rm(\tilde{P}_m,t_m)|\leq
\frac{C}{\psi_\alpha^2(\tilde{P}_m,t_m)}\leq
\frac{C}{\psi_\alpha^2(P_m,t_m)}.$$ On the other hand $$K_1(P_m,t_m)=\frac{1-(\psi_\alpha)_s^2(P_m,t_m)}
{\psi_\alpha^2(P_m,t_m)}=\frac{3}{4\psi_\alpha^2(P_m,t_m)}.$$ So $$|Rm(P_m,t_m)|\geq K_1(P_m,t_m)=
\frac{3}{4\psi_\alpha^2(P_m,t_m)}\geq
\frac{3}{4C}|Rm(\tilde{P}_m,t_m)|.$$ Obviously since $(\tilde{P}_m,t_m)$ is the maximal point, we have $$|Rm(P_m,t_m)|\leq
|Rm(\tilde{P}_m,t_m)|.$$ So the curvatures at $(P_m,t_m)$ and $(\tilde{P}_m,t_m)$ are comparable, where we used the definition of $|Rm|$ to be the largest absolute value of the eigenvalues of the curvature operator $Rm$. So by repeating (part of) the argument as in case 1 in Lemma 4.1, we know that a subsequence of $g_{ij}^{(m)}(\cdot,t)$ around the point $P$ will converge to a nonflat complete ancient $\kappa$-solution on a smooth manifold $M$ for some positive constant $\kappa$ depending only on the initial metric $g_\alpha(0)$.
We remain to consider the case that the maximal point $(\tilde{P}_m,t_m)$ does not lie on the right polar cap, then it must lie in the region between the two bumps.
In this case, we first prove the following assertion:
*For any $A<+\infty$, there exists a positive constant $C(A)$ such that the curvatures of $g_{ij}^{(m)}(\cdot,t)$ at the new time $t=0$ satisfy the estimate$$|Rm^{(m)}(y,0)|\leq C(A)$$ whenever $d_{g^{(m)}(\cdot,0)}(y,P_m)\leq A$ and $m\geq 1$, where $Rm^{(m)}$ denotes the curvature of the metric $g_{ij}^{(m)}$.*
This assertion in the three-dimensional Ricci flow has been verified by Perelman in his proof of the Theorem 12.1 in [@P1] (the first detailed exposition of this part of Perelman’s argument appeared in the first version of Kleiner-Lott [@KL]), where the only three-dimension features he used are the Hamliton-Ivey curvature pinching estimate and the canonical neighborhood condition of an ancient $\kappa$-solution consisting the $\varepsilon$-necks and $\varepsilon$-caps. In our case, by noting that the metric is rotationally symmetric, the canonical neighborhood condition can be easily obtained as pointed out before, and the pinching estimate has already given in Proposition 2.1(4). So by some slight modifications, Perelman’s argument also works for our case. In the following we only give the details for the modified parts. For the complete details, one can compare with [@KL] (from page 85 to 87) or [@CZ] (from page 400 to 402).
For each $\rho \geq 0$, set $$M(\rho)=\sup\{R^{(m)}(x,0)\ |\ m\geq 1, x\in S^{n+1} \ \mbox{ with } \
d_{0}(x,P_m)\leq \rho\}$$ and $$\rho_0=\sup\{\rho \geq 0\ |\ \ M(\rho)<+\infty \}.$$ By the pinching estimate in Proposition 2.1(4), it suffices to show $\rho_0=+\infty.$
We need to adapt Perelman’s argument to show that $\rho_0>0$.
For arbitrary fixed small $\varepsilon >0$, by Proposition 4.3, we know that there exists $K=K(\varepsilon,
g_\alpha(0))>\max\{2\varepsilon^{-1}, Q(\frac{3}{4}T_\alpha)\}>0$, where $Q(\frac{3}{4}T_\alpha)$ denotes the upper bound of the curvature for the times $t\leq\frac{3}{4}T_\alpha$, such that for any point $(x_0,t_0)$ with $t_0\geq\frac{3}{4}T_\alpha$ and $Q=R^{(\alpha)}(x_0,t_0)\geq K$, the solution in $\{(y,t)|d_{t_0}^2(y,x_0)<\varepsilon^{-2}Q^{-1},
t_0-\varepsilon^{-2}Q^{-1}\leq t \leq t_0\}$ is, after scaling by the factor $Q$, $\varepsilon$-close to the corresponding subset of some orientable ancient $\kappa$-solution for some positive constant $\kappa$ depending only on the initial metric $g_\alpha(0)$. Consequently we have the gradient estimate in the region$$|\nabla(R^{-\frac{1}{2}})|\leq \eta \ \ \mbox{ and } \ \
|\frac{\partial}{\partial t}(R^{-1})|\leq \eta \eqno(4.6)$$ for some constant $\eta=\eta(\kappa)>0.$
If $R^{(\alpha)}(P_m,t_m)\geq K$, then by the above gradient estimate (4.6), we know that there exists some constant $c=c(\eta)>0$ such that $$R^{(\alpha)}(x,t_m)\leq
2R^{(\alpha)}(P_m,t_m)$$ for any point $x\in
B_{t_m}(P_m,c(R^{(\alpha)}(P_m,t_m))^{-\frac{1}{2}})$. Hence in this case we have $\rho_0\geq c>0$.
If $R^{(\alpha)}(P_m,t_m)< K$, then we prove that $\rho_0\geq
\tilde{c}$ for some constant $\tilde{c}=\tilde{c}(c,K,\bar{c})$, where $\bar{c}$ is the positive lower bound of the scalar curvature $R^{(\alpha)}$ on $S^{n+1}$ at initial time. In fact, consider the points $x\in
B_{t_m}(P_m,\frac{c}{2}(R^{(\alpha)}(P_m,t_m))^{-\frac{1}{2}})$, if $R^{(\alpha)}(x,t_m)< K$ for all points $x$, then $\rho_0\geq
\frac{c}{2}>0$ (since $R^{(\alpha)}(P_m,t_m)\geq \bar{c})$; if $R^{(\alpha)}(x,t_m)\geq K$ for some point $x$, consider the nearest point $y_0 \in
B_{t_m}(P_m,\frac{c}{2}(R^{(\alpha)}(P_m,t_m))^{-\frac{1}{2}})$ to $P_m$ such that $R^{(\alpha)}(y_0,t_m)= K$, then by Proposition 4.3 and the gradient estimate (4.6), we know that $$R^{(\alpha)}(y,t_m)\leq 2R^{(\alpha)}(y_0,t_m)=2K \eqno(4.7)$$ for any point $y\in
B_{t_m}(y_0,c(R^{(\alpha)}(y_0,t_m))^{-\frac{1}{2}})$. Since the scalar curvature has a positive lower bound $\bar{c}$ by our assumption, we know that there exists $c'=cK^{-\frac{1}{2}}\bar{c}^{\frac{1}{2}}>0$ such that $$B_{t_m}(y_0,c(K)^{-\frac{1}{2}})\supset
B_{t_m}(P_m,c'(R^{(\alpha)}(P_m,t_m))^{-\frac{1}{2}}).\eqno(4.8)$$ By (4.7) and (4.8) we know that for any point $z\in
B_{t_m}(P_m,c'(R^{(\alpha)}(P_m,t_m))^{-\frac{1}{2}})$ we have $$R^{(\alpha)}\leq 2K.$$ Then $\rho_0\geq c'>0$. Set $\tilde{c}=\min\{\frac{c}{2},c'\}$, then in this case we have $\rho_0\geq \tilde{c}>0$.
Hence we have proved $\rho_0>0$.
In the rest, we can apply the same argument of Perelman [@P1] (see also [@KL] and [@CZ] for details) to obtain that $\rho_0=+\infty$. That is, the curvatures of $g_{ij}^{(m)}(\cdot,t)$ at the new times $t=0$ stay uniformly bounded at bounded distances from $P_m$ for all $m$. Furthermore, by the estimate in Proposition 2.1(2) and using $|(\psi_\alpha)_s|\geq\frac{1}{2}$ in the region between $P_m$ and $P$, we know that the distance from $P_m$ and $P$ measured in the rescaled metric $g_{ij}^{(m)}(\cdot,0)$ is bounded. So we obtained that the curvature of $g_{ij}^{(m)}(\cdot,t)$ at the new times $t=0$ stay uniformly bounded at bounded distances from $P$ for all $m$. This completes the proof of the assertion.
By the gradient estimate in Proposition 4.3 and Shi’s local derivative estimate in [@Sh89] and Hamilton’s compactness theorem in [@Ha95], we can take a $C_{loc}^\infty$ subsequent limit to obtain $(M,g_\infty(\cdot,t),P)$ which is complete, $\kappa$-noncollapsed on all scales and is defined on a space-time open subset of $M\times (-\infty,0]$ containing the time slice $M\times \{0\}$. Clearly it follows from the pinching estimate in Proposition 2.1(4) that the limit $(M,g_\infty(\cdot,t),P)$ has nonnegative curvature operator. Then exactly as Perelman’s argument in Theorem 12.1 of [@P1] (see also [@KL] and [@CZ] for details), we can get that the curvature of the limit $g_\infty(\cdot,t)$ at $t=0$ has bounded curvature and also that the limit $g_\infty(\cdot,t)$ can be defined on $(-\infty,0]$. So we have proved that $g_\infty(\cdot,t)$ is an ancient $\kappa$-solution on $M$ and Claim 3 holds.
Since by our assumption that the singularity is not of Type II. Then there exists some constant $\tilde{C}>0$ such that $$0\leq R(P_m,t)\leq
\frac{\tilde{C}}{T_\alpha -t}.$$ Then by Claim 3 we know that a subsequence of $g_{ij}^{(m)}(\cdot,t)$ around $P$ converges to a nonflat ancient $\kappa$-solution $g_{ij}$ on $M$. Then by the same proof as in the case 1 in Lemma 4.1, we obtain that there exists a sequence of times $t_k\rightarrow -\infty$ such that the scaling of $g_{ij}(\cdot,t)$ around $P$ with the factor $|t_k|^{-1}$ and with the times $t_k$ shifting to the new time zero converge to a nonflat gradient shrinking soliton in $C_{loc}^\infty$ topology. If the nonflat gradient shrinking soliton is noncompact, Proposition 3.1 gives us that it is $R\times S^n$. But since the limit is taking around the pole $P$ and the metric is rotationally symmetric, it can not be $R\times S^n$. So this contradiction implies that the nonflat gradient shrinking soliton is compact. By Proposition 3.1 again, we know that it is the round $S^{n+1}$. Consequently the curvature of the original solution becomes positive as the time $t$ close to the maximal time $T_\alpha$. Then repeat the same proof as in Claim 1, we can choose $\beta\in (\alpha,1]\subset A_1$ sufficiently close to $\alpha$ such that $g_\beta(0)$ will also shrink to a round point at the maximal time $T_\beta<+\infty$, which contradicts with $\beta\in
(\alpha,1] \subset A_1$. So the singularity must be of Type II. Therefore we have proved our theorem 1.2.$$\eqno \#$$
0.2cm[**Remark 1.**]{} During the proof of the main theorem, we actually proved the existence of Type II singularities on noncompact manifolds. More precisely, we proved that for each $n\geq2$, there exists complete and rotationally symmetric metrics on $R^{n+1}$ with bounded curvatures such that the Ricci flow starting at the metrics develop Type II singularities at some times $T<+\infty$. In particular, we can take the initial metrics on $R^{n+1}$ to be the complete and rotationally symmetric, with nonnegative sectional curvature and positive scalar curvature, and asymptotic to the round cylinder of scalar curvature 1 at infinity. 0.2cm[**Remark 2.**]{} In the unpublished preprint [@BR], Robert Bryant proved the existence of the nontrivial steady Ricci solitons on $R^{n}$ by solving certain nonlinear ODE system. These steady Ricci solitons are complete, rotationally symmetric with positive curvatures. By combining with the work of Hamilton [@Ha93], this paper gives another proof for the existence of the nontrivial steady Ricci solitons on $R^n$ for all dimensions $n\geq 3$, which are also complete, rotationally symmetric and have positive curvatures.
[99]{}
S. Altschuler, S. Angenent, S. Giga, [ *Mean curvature flow through singularities for surfaces of rotation.*]{}, The Journal of Geometric Analysis, [**5**]{} no.3, (1995), 293-358.
S. Angenent, [*The zero set of a solution of a parabolic equation*]{}, J. reine angew. Math. [**390**]{} (1988), 79-96.
S. Angenent, [*Shrinking doughnuts*]{}, Proc. Conf. Elliptic Parabolic Equations, Greynog, Wales, 1989.
S. Angenent, J.J.L. Vel$\acute{a}$zquez, [*Degenerate neckpinches in mean curvature flow*]{}, J.Reine Angew.Math. [**482**]{} (1997), 15-66.
S. Angenent, and D. Knopf, [*An example of neckpinching for Ricci flow on $S^{n+1}$*]{}, Math. Res. Lett. [**11**]{} (2004), no. 4, 493-518.
S. Angenent, and D. Knopf, [*Precise asymptotics of the Ricci flow neckpinch*]{}, arXiv:math.DG/0511247 v1 9 Nov 2005.
C. B$\ddot{o}$hm, and B. Wilking, [*Manifolds with positive curvature operators are space forms*]{}, arXiv:math.DG/0606187 June 2006.
D. Burago, Y. Burago, and S. Ivanov, [*A Course in Metric Geometry,*]{} Graduate Studies in Mathematics, Volume **33**, Amer. Math. Soc. Providence,Rhode Island.
R. Bryant, [*Local existence of gradient Ricci solitons*]{}.
H. D. Cao, and X. P. Zhu, [*A complete proof of the Poincar$\acute{e}$ and geometrization conjecture – application of the Hamilton-Perelman theory of the Ricci flow,*]{} Asian J. Math. [**10**]{} (2006), no. 2, 165-492.
J. Cheeger, and D. Ebin, [*Comparison theorems in Riemannian geometry,*]{} North-Holland (1975).
B. L. Chen, and X. P. Zhu, [*Ricci Flow with Surgery on Four-manifolds with Positive Isotropic Curvature,*]{} J. Differential Geometry, **74** (2006), 177-264.
B. Chow, [*The Ricci flow on 2-sphere*]{}, J. Diff. Geom. [**33**]{} (1991) 325-334.
B. Chow, and D. Knopf, [*The Ricci flow: An introduction*]{}, Mathematical Surveys and Monographs, Amer. Math. Soc., Providence, RI, 2004.
P. Daskalopoulos, and R. S. Hamilton, [*Geometric estimates for the logarithmic fast diffusion equation*]{}, Communications in Analysis and Geometry, vol. 12, no. 1-2, pp. 143-164, 2004.
M. Feldman, T. Ilmanen, D. Knopf, [*Rotationally symmetric shrinking and expanding gradient K$\ddot{a}$hler-Ricci solitons*]{}, J. Geom. Anal. [**65**]{} (2003), 169-209.
M. Grayson, [*A short note on the evolution of a surface by its mean curvature*]{}, Duke Math. J. [**58**]{}, (1989) 555-558.
D. Gromoll, and W. Meyer, [*On complete open manifolds of positive curvature*]{}, Ann. of Math., **90** (1969), 95-90.
R. S. Hamilton, [*Three manifolds with positive Ricci curvature* ]{}, J. Diff. Geom. [**17**]{} (1982), 255-306.
R. S. Hamilton, [*Four–manifolds with positive curvature operator*]{}, J. Differential Geom. [**24**]{} (1986), 153-179.
R. S. Hamilton, [*The Ricci flow on surfaces*]{}, Contemporary Mathematics [**71**]{} (1988) 237-261.
R. S. Hamilton, [*Eternal solutions to the Ricci flow*]{}, J. Differential Geom. [**38**]{} (1993) 1-11.
R. S. Hamilton, [*A compactness property for solution of the Ricci flow*]{}, Amer. J. Math. [**117**]{} (1995), 545-572.
R. S. Hamilton, [*The formation of singularities in the Ricci flow*]{}, Surveys in Differential Geometry (Cambridge, MA, 1993), [**2**]{}, 7-136, International Press, Combridge, MA,1995.
R. S. Hamilton, [*Four manifolds with positive isotropic curvature*]{}, Commu. in Analysis and Geometry,[**5**]{}(1997),1-92.
B. Kleiner and J. Lott, [*Notes on Perelman’s papers*]{}, arXiv:math.DG/0605667 v1 25 May 2006.
G. Perelman, [*The entropy formula for the Ricci flow and its geometric applications*]{}, arXiv:math.DG/0211159 v1 November 11, 2002.preprint.
G. Perelman, [*Ricci flow with surgery on three manifolds*]{}, arXiv:math. DG/0303109 v1 March 10, 2003. prepeint.
P. Petersen, [*Riemannian Geometry*]{}, Berlin-Heidelberg: Springer-Verlag, 1997.
W. X. Shi, [*Deforming the metric on complete Riemannian manifold*]{}, J. Differential Geometry [**30**]{} (1989), 223-301.
M. Simon. [*A class of Riemannian manifolds that pinch when evolved by Ricci flow*]{}, Manuscripta Math. [**101**]{} (2000), no. 1, 89-114.
P. Topping, [*Lectures on the Ricci flow*]{}, L.M.S. Cambridge University Press (12 Oct 2006), ISBN: 0521689473. http://www.maths.warwick.ac.uk/ topping/RFnotes.html.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give a new way of looking at the Cho–Faddeev–Niemi (CFN) decomposition of the Yang-Mills theory to answer how the enlarged local gauge symmetry respected by the CFN variables is restricted to obtain another Yang-Mills theory with the same local and global gauge symmetries as the original Yang-Mills theory. This may shed new light on the fundamental issue of the discrepancy between two theories for independent degrees of freedom and the role of the Maximal Abelian gauge in Yang-Mills theory. As a byproduct, this consideration gives new insight into the meaning of the gauge invariance and the observables, e.g., a gauge-invariant mass term and vacuum condensates of mass dimension two. We point out the implications for the Skyrme–Faddeev model.'
---
------------------------------------------------------------------------
------------------------------------------------------------------------
1.0cm
**Yang-Mills Theory**
0.4cm
**Constructed from**
0.4cm
**Cho–Faddeev–Niemi Decomposition**
0.5cm
[ **Kei-Ichi Kondo,$^{\dagger,\ddagger,{1}}$ Takeharu Murakami$^{\ddagger,{2}}$ and Toru Shinohara$^{\ddagger,{3}}$** ]{}
0.5cm
*${}^{\dagger}$Department of Physics, Faculty of Science, Chiba University, Chiba 263-8522, Japan*
0.3cm
*${}^{\ddagger}$Graduate School of Science and Technology, Chiba University, Chiba 263-8522, Japan*
Key words: Cho-Faddeev-Niemi decomposition, magnetic condensation, Abelian dominance, monopole condensation, quark confinement
PACS: 12.38.Aw, 12.38.Lg
------------------------------------------------------------------------
0.1cm ${}^1$ E-mail: [[email protected]]{}
${}^2$ E-mail: [[email protected]]{}
${}^3$ E-mail: [[email protected]]{}
0.5cm
Introduction
============
In understanding a non-perturbative feature of quantum field theory, it is quite important to extract the dynamical (local) or topological (global) degrees of freedom which are most relevant to the physics in question. For example, it is widely accepted that the magnetic monopole is responsible for quark confinement and the Yang-Mills instanton for chiral symmetry breaking.
From this viewpoint, the decomposition or the change of variables of the Yang-Mills gauge field proposed by Cho [@Cho80], Faddeev and Niemi [@FN99] (CFN) is very interesting, since it enables us to extract explicitly certain types of topological configurations in Yang-Mills theory, especially, the magnetic monopole of Wu–Yang type and a multi-instanton of Witten type (see e.g., [@TTF00] for the mutual dependence). A characteristic feature of the [*CFN decomposition*]{} for the SU(2) Yang-Mills gauge field $\mathscr{A}_\mu(x)$ is to introduce a three-dimensional unit vector field $\bm{n}(x)=(n^1(x), n^2(x), n^3(x))$ satisfying $\bm{n}(x) \cdot \bm{n}(x) :=n^A(x) n^A(x) = 1$; that is, there are two extra degrees of freedom, in addition to the two vector fields: $C_\mu(x)\bm{n}(x)$, parallel to $\bm{n}(x)$, and $\mathbb{X}_\mu(x)$, perpendicular to $\bm{n}(x)$ \[and hence $\bm{n}(x) \cdot \mathbb{X}_\mu(x)=0$\]. The $\bm{n}$ field represents the color direction (in a gauge invariant way) and plays the distinguished role of expressing the topological configurations mentioned above. An on-shell decomposition of the gauge field is also given in Faddeev–Niemi [@FN99], in agreement with the on-shell counting of the degrees of freedom in the original Yang-Mills field. This is very useful for the purpose of finding the classical solution of the equation of motion. The CFN decomposition is further developed in [@Shabanov99a; @Shabanov99b; @Gies01].
The CFN decomposition has enormous potential which has enabled us to discover novel non-perturbative features previously overlooked in Yang-Mills theory. For example, the Skyrme–Faddeev model [@FN97] describing a glueball as a knot soliton solution can be deduced from the Yang-Mills theory by way of the CFN decomposition of the Yang-Mills theory [@LN99; @Gies01; @BCK01; @FN02; @Cho03; @Kondo04; @KKMSS05; @KKMSSI05] (see, e.g., [@Hirayama] for exact solutions). Moreover, the stability recovery of the Savvidy vacuum [@Savvidy77] through the elimination of a tachyon mode [@NO78] has been shown [@Cho03; @Kondo04; @KKMSS05] by using the CFN decomposition. (See [@Cho03] for treatment of a massless gluon and [@Kondo04; @KKMSS05] for treatment of a massive gluon caused by novel magnetic condensation.) Also, numerical simulations on a lattice can be performed based on the CFN decomposition [@KKMSS05; @KKMSSI05].
As mentioned above, however, the CFN change of variables introduces two extra degrees of freedom. At the level of classical Yang-Mills theory, this does not cause any subtle problems. However, a question arises when we consider the quantization of the Yang-Mills theory written in terms of the CFN variables, which we call the [*CFN-Yang-Mills theory*]{} (the extended Yang-Mills theory), because the reparametrization seems to increase the number of dynamical degrees of freedom in the Yang-Mills theory. How do we deal with the two extra degrees of freedom introduced by $\bm{n}$? This question has caused much controversy concerning the treatment and interpretation of $\bm{n}$.
A partial answer was given by Shabanov [@Shabanov99b]: To obtain the same degrees of freedom as in the original Yang-Mills theory, two constraint conditions, $\bm{\chi}(x)=0$, must be imposed, e.g., using the Lie-algebra valued functional $\bm{\chi}(x):=\chi^A(x)T^A$ subject to $\bm{n}(x) \cdot \bm{\chi}(x)=0$. In fact, the integration measure for the CFN variables in the framework of the functional integration method has been constructed so that it is completely equivalent to the standard integration measure of the original Yang-Mills theory, provided that the two conditions $\bm{\chi}(x)=0$ are imposed on the measure in accordance with the above viewpoint.
Some subtle aspects of this issue have been clarified and resolved by Cho and his collaborators [@BCK01]. They pointed out that it is meaningless to eliminate the two extra degrees of freedom created by $\bm{n}$ by adding two extra constraints. The constraint should be regarded as a gauge fixing condition or a consistency condition. Moreover, the topological field $\bm{n}$ becomes dynamical with the gauge fixing, increasing the number of dynamical degrees of freedom. They argued that the extended Yang-Mills theory is modified in a subtle but important way and they have proposed three quantization schemes based on different choices of the decomposition (or reparametrization), but neither their uniqueness nor equivalence has been demonstrated. This clearly shows that there remain the unresolved questions concerning the CFN decompositions. The lack of a complete understanding of the CFN decomposition is merely an obstacle to utilizing this machinery.
The purpose of this paper is to reconsider the meaning of the CFN decomposition from a different viewpoint which could give a thorough and unambiguous understanding of the CFN-Yang-Mills theory. In particular, we explicitly specify [*the gauge group for the enlarged gauge symmetry of the CFN-Yang–Mills theory*]{}, rather than merely counting the degrees of freedom. Then we can answer the following questions.
1. What gauge symmetry does the CFN-Yang-Mills theory possess? Which part of the enlarged gauge symmetry of the CFN-Yang–Mills theory must be constrained to reproduce the gauge theory with the same gauge symmetry as the original Yang-Mills theory? How do we fix the enlarged gauge symmetry to meet this requirement?
2. What are the gauge invariant observables written in terms of the CFN variables?
3. How do we define the CFN decomposition on a lattice, and how do we perform numerical simulations of the lattice CFN-Yang–Mills theory [@KKMSS05; @KKMSSI05]?
4. What is the correct form of the Faddeev–Popov ghost term in the BRST quantization in the continuum formulation [@KMS05b]?
These are advantages of our way of interpreting the CFN variables from a new viewpoint. In this paper, we discuss the first two issues, while other issues are reported elsewhere [@KKMSS05; @KKMSSI05; @KMS05b].
Yang-Mills theory in the CFN decomposition
==========================================
Local gauge symmetry in terms of the CFN variables
--------------------------------------------------
The Cho–Faddeev-Niemi (CFN) decomposition (or change of variables) of the original Yang-Mills gauge field $\mathscr{A}_\mu(x)$ is performed as follows. We restrict our consideration to the gauge group $G=SU(2)$. First of all, we introduce a unit vector field $\bm{n}(x)$ as $$\begin{aligned}
\bm{n}(x) \cdot \bm{n}(x) := n^A(x) n^A(x) = 1 \quad (A=1,2,3) .
\label{nn=1}\end{aligned}$$ Then the off-shell CFN decomposition is written in the form $$\mathscr{A}_\mu(x)
=c_\mu(x) \bm{n}(x)
+g^{-1}\partial_\mu \bm{n}(x)\times \bm{n}(x)
+\mathbb X_\mu(x) ,
\label{CFN}$$ where $\mathbb{X}_\mu(x)$ is perpendicular to $\bm{n}$: $$\begin{aligned}
\bm{n}(x) \cdot \mathbb{X}_\mu(x) = 0 .
\label{nX=0}\end{aligned}$$ The first term on the right-hand side of (\[CFN\]) is denoted by $
\mathbb{C}_\mu(x) := c_\mu(x){\bm n}(x) .
$ It is parallel to $\bm{n}(x)$ and is called the restricted potential. The second term is denoted $
\mathbb{B}_\mu(x) := g^{-1}\partial_\mu{\bm n}(x)\times{\bm n}(x) .
$ It is perpendicular to $\bm{n}(x)$ and is called the magnetic potential. For later convenience, we define the sum of $\mathbb{C}_\mu(x)$ and $\mathbb{B}_\mu(x)$: $$\begin{aligned}
\mathbb{V}_\mu(x) := \mathbb{C}_\mu(x) + \mathbb{B}_\mu(x)
= c_\mu(x) \bm{n}(x)
+g^{-1}\partial_\mu \bm{n}(x)\times \bm{n}(x) .\end{aligned}$$ The form of $\mathbb{V}_\mu(x)$ is determined by the requirement that the field $\bm{n}(x)$ be a covariant constant in the background field $\mathbb{V}(x)$: $$\begin{aligned}
D_\mu[\mathbb{V}] \bm{n} := \partial_\mu \bm{n} + g\mathbb{V}_\mu \times \bm{n} = 0 .
\label{D[V]n=0}\end{aligned}$$ Here, only the perpendicular part, $\mathbb{B}_\mu$, is uniquely determined, while the parallel component $c_\mu$ is not determined uniquely [@Manton77]. (In fact, any four-vector is allowed.)
As pointed out in [@Shabanov99b], the restricted potential $c_\mu$ and the gauge covariant potential $\mathbb{X}_\mu$ are specified by $\bm{n}$ and $\mathscr{A}_\mu$ as $$\begin{aligned}
c_\mu(x)
&={\bm n}(x)\cdot\mathscr{A}_\mu(x),
\quad
\mathbb X_\mu(x)
=g^{-1}{\bm n}(x)\times D_\mu[\mathscr{A}]{\bm n}(x) .
\label{def:X}\end{aligned}$$ The first equation is obtained from (\[nX=0\]) and (\[nn=1\]), which yield $\bm{n}(x) \cdot \partial_\mu \bm{n}(x)=0$, while the second equation is obtained by making use of the fact (\[D\[V\]n=0\]), which yields $$\begin{aligned}
D_\mu[\mathscr{A}] \bm{n}
:= \partial_\mu \bm{n} + g\mathscr{A}_\mu \times \bm{n}
= D_\mu[\mathbb{V}] \bm{n} + g\mathbb{X}_\mu \times \bm{n}
= g\mathbb{X}_\mu \times \bm{n} .
\label{XnA}\end{aligned}$$ More explicitly, the Yang-Mills gauge field $\mathscr{A}_\mu(x)$ can be cast into the equivalent form $$\begin{aligned}
\mathscr{A}_\mu
=& (\bm{n} \cdot \mathscr{A}_\mu)\bm{n} + \mathscr{A}_\mu - (\bm{n} \cdot \mathscr{A}_\mu) \bm{n}
\nonumber\\
=& (\bm{n} \cdot \mathscr{A}_\mu)\bm{n} + (\bm{n}\cdot\bm{n}) \mathscr{A}_\mu - (\bm{n} \cdot \mathscr{A}_\mu) \bm{n}
\nonumber\\
=& (\bm{n} \cdot \mathscr{A}_\mu)\bm{n} + \bm{n} \times (\mathscr{A}_\mu \times \bm{n})
\nonumber\\
=& (\bm{n} \cdot \mathscr{A}_\mu)\bm{n} - g^{-1}\bm{n} \times \partial_\mu \bm{n} + g^{-1}\bm{n} \times (\partial_\mu \bm{n} + \mathscr{A}_\mu \times \bm{n})
\nonumber\\
=& (\bm{n} \cdot \mathscr{A}_\mu)\bm{n} + g^{-1} \partial_\mu \bm{n} \times \bm{n} + g^{-1}\bm{n} \times D_\mu[\mathscr{A}]\bm{n} ,
\label{CFNform}\end{aligned}$$ where we have used only the relation $\bm{n}\cdot\bm{n}=1$ in the second equality.
An important observation from (\[def:X\]) is that [*the local gauge transformations $\delta c_\mu$ and $\delta\mathbb X_\mu$ are uniquely determined once the transformations $\delta{\bm n}$ and $\delta\mathscr{A}_\mu$ are specified*]{}.[^1]
Now we consider the local gauge symmetry possessed by the Yang-Mills theory written in terms of CFN variables, which we call [*CFN–Yang-Mills theory*]{}.
- The invariance of the Lagrangian is guaranteed by the usual gauge transformation: $$\delta\mathscr{A}_\mu(x)
=D_\mu[\mathscr{A}]{\bm\omega}(x) .
\label{gtA}$$ This symmetry is the local $G=SU(2)$ gauge symmetry and denoted $SU(2)_{local}^{\omega}$.
- The gauge transformation of $\bm{n}$ is nothing but the map from $S^2$ to $S^2$ at each spacetime point, because $\bm{n}$ is always defined to be a three-dimensional unit vector field, i.e., ${\bm n}(x)^2=1$. Therefore, it is expressed as a local rotation by an angle ${\bm\theta}(x)$:[^2] $$\delta{\bm n}(x)
=g{\bm n}(x) \times {\bm\theta}(x)
=g{\bm n}(x) \times {\bm\theta}_\perp(x) ,
\label{gtn}$$ where ${\bm\theta}_\perp(x)$ is perpendicular to $\bm{n}(x)$ \[i.e., $\bm n(x)\cdot{\bm\theta}_\perp(x)=0$\] and has two independent components. For the parallel component, ${\bm\theta}_\parallel(x)=\theta_\parallel(x)\bm n(x)$, the vector field $\bm{n}(x)$ is invariant under this transformation \[a rotation about the axis of $\bm{n}(x)$\]. Therefore, it is a [*redundant*]{} symmetry, say the U(1)$^{\theta}$ symmetry, of the CFN-Yang-Mills theory, as $c_\mu(x)$ and $\mathbb{X}_\mu(x)$ are also unchanged for a given $\mathscr{A}_\mu(x)$. (Note that $S^2 \simeq SU(2)/U(1)$.) Therefore, this symmetry is the local SU(2)/U(1) symmetry and denoted $[SU(2)/U(1)]_{local}^{\theta}$.
Note that ${\bm\omega}(x)$ and ${\bm\theta}(x)$ are independent, since the original Yang-Mills Lagrangian does not depend on the choice of ${\bm\theta}(x)$. For later convenience, we distinguish between the above transformations by $\delta_\theta$ and $\delta_\omega$ as follows: $$\begin{aligned}
{\bm\theta}={\bm0} , {\bm\omega}\ne{\bm0}:
\Longrightarrow &
\delta_\omega{\bm n}(x)={\bm0},
\quad
\delta_\omega\mathscr{A}_\mu(x)
=D_\mu[\mathscr{A}]{\bm\omega}(x) ,
\\
{\bm\theta}\ne{\bm0} (\bm{n}\cdot{\bm\theta}=0), {\bm\omega}={\bm0}:
\Longrightarrow &
\delta_\theta{\bm n}(x)
=g{\bm n}(x)\times{\bm\theta}(x),
\quad
\delta_\theta\mathscr{A}_\mu(x)={\bm0}.\end{aligned}$$ The general local gauge transformation in the CFN–Yang-Mills theory is obtained by combining $\delta_\theta$ and $\delta_\omega$. Thus, [*the CFN–Yang-Mills theory has the local gauge symmetry $$\tilde{G}^{\omega,\theta}_{local} := SU(2)_{local}^{\omega} \times [SU(2)/U(1)]_{local}^{\theta} ,$$ i.e., the direct product of $SU(2)_{local}^{\omega}$ and $[SU(2)/U(1)]_{local}^{\theta} $, which is larger than the local $G=SU(2)_{local}^{\omega}$ symmetry of the original Yang-Mills theory*]{} (see Fig. \[fig:sym-cfn-ym\]).
![ The relationship between the CFN-Yang–Mills theory and the original Yang-Mills theory. The CFN-Yang–Mills theory obtained through the CFN decomposition has a larger (local and global) gauge symmetry than the original Yang-Mills theory and becomes equivalent to the original Yang-Mills theory after the new MAG is imposed. []{data-label="fig:sym-cfn-ym"}](CFN-symmetry.eps){height="6.5cm"}
In the papers [@BCK01; @Cho03; @Kondo04], two local gauge transformations are introduced by decomposing the original gauge transformation, $
\delta_\omega \mathscr{A}_\mu(x) = D_\mu[\mathscr{A}] \bm{\omega}(x) .
$
:
$$\begin{aligned}
\delta_\omega \bm{n} =& 0 ,
\\
\delta_\omega c_\mu =& \bm{n} \cdot D_\mu[\mathscr{A}] \bm{\omega} ,
\\
\delta_\omega \mathbb{X}_\mu =& D_\mu[\mathscr{A}] \bm{\omega} - \bm{n}( \bm{n} \cdot D_\mu[\mathscr{A}] \bm{\omega}) ,
$$
:
$$\begin{aligned}
\delta_{\omega}' \bm{n} =& g \bm{n} \times \bm{\omega'} ,
\\
\delta_{\omega}' c_\mu =& \bm{n} \cdot \partial_\mu \bm{\omega'} ,
\\
\delta_{\omega}' \mathbb{X}_\mu =& g \mathbb{X}_\mu \times \bm{\omega'} ,
$$
The gauge transformation I has been called the passive or quantum gauge transformation, while II has been called the active or background gauge transformation. However, this classification is not necessarily independent, and it leads to sometimes confusing and misleading results. The local gauge transformation I defined in the previous paper [@Kondo04] is identical to $\delta_\omega$. In order to see how the gauge transformation II defined in [@Kondo04] is reproduced, we apply the gauge transformations (\[gtA\]) and (\[gtn\]) to (\[def:X\]). Then we can show that the gauge transformation of other CFN variables are given by [^3] $$\begin{aligned}
\delta c_\mu(x)
= g(\bm{n}(x) \times \mathscr{A}_\mu(x)) \cdot (\bm{\omega}_\perp(x) - \bm{\theta}_\perp(x)) + \bm{n}(x) \cdot \partial_\mu \bm{\omega}(x) ,
\label{gtc}
\\
\delta \mathbb{X}_\mu(x)
= g \mathbb{X}_\mu(x) \times (\bm{\omega}_\parallel(x)+\bm{\theta}_\perp(x)) + D_\mu[\mathbb{V}](\bm{\omega}_\perp(x)-\bm{\theta}_\perp(x)) .
\label{gtX}\end{aligned}$$ If $\bm\omega_\perp(x)=\bm\theta_\perp(x)$, the transformations (\[gtc\]) and (\[gtX\]) reduce to the gauge transformation II with the parameter $\bm\omega'(x)=(\bm\omega_\parallel(x),\bm\omega_\perp(x)=\bm\theta_\perp(x))$. Therefore, the gauge transformation II corresponds to the special case $\bm\omega_\perp(x)=\bm\theta_\perp(x)$.
A new viewpoint for the CFN-Yang–Mills theory
---------------------------------------------
The CFN-Yang-Mills theory has the local gauge symmetry $\tilde{G}^{\omega,\theta}_{local}$ which is larger than that of the original Yang-Mills theory, because we can rotate the CFN variable $\bm{n}(x)$ by an angle $\bm{\theta}^{\perp}(x)$ independently of the gauge transformation parameter $\bm{\omega}(x)$ of $\mathscr{A}_\mu(x)$. In order to obtain the gauge theory with the same local gauge symmetry as the original Yang-Mills theory, therefore, we proceed to impose a gauge fixing condition by which $\tilde{G}^{\omega,\theta}_{local}$ is broken down to $SU(2)$, a subgroup of $\tilde{G}^{\omega,\theta}_{local}$ (see Fig. \[fig:sym-cfn-ym\]).
We have found that one way of imposing such a gauge fixing condition is to impose the minimizing condition $$\begin{aligned}
0 = \delta\int d^4x\frac12\mathbb X_\mu^2 ,
\label{MAGcond}\end{aligned}$$ [*with respect to the enlarged gauge transformation*]{} $(\omega, \theta)$, which we call the new maximal Abelian gauge (nMAG). This is done as follows. Because the relationship (\[XnA\]) leads to $$\begin{aligned}
g^2 \mathbb X_\mu^2
= ({\bm n}\times D_\mu[\mathscr{A}]{\bm n})^2
= \left\{
(D_\mu[\mathscr{A}]{\bm n})^2
-({\bm n}\cdot D_\mu[\mathscr{A}]{\bm n})^2
\right\}
= (D_\mu[\mathscr{A}]{\bm n})^2 ,
\label{X2}\end{aligned}$$ the local gauge transformation of $\mathbb X^2$ is calculated as $$\begin{aligned}
\delta\frac12\mathbb X_\mu^2
&=g^{-2}
(D_\mu[\mathscr{A}]{\bm n}) \cdot \delta(D_\mu[\mathscr{A}]{\bm n})
\nonumber \\
&=g^{-2}
(D_\mu[\mathscr{A}]{\bm n}) \cdot
(D_\mu[\mathscr{A}]\delta{\bm n}
+g\delta\mathscr{A}_\mu\times\bm n)
\nonumber \\
&=g^{-2}
(D_\mu[\mathscr{A}]{\bm n}) \cdot
\{gD_\mu[\mathscr{A}](\bm n\times\bm\theta_\perp)
+(gD_\mu[\mathscr{A}]\bm\omega)\times\bm n\}
\nonumber \\
&=g^{-2}
(D_\mu[\mathscr{A}]{\bm n}) \cdot
\{(gD_\mu[\mathscr{A}] \bm n) \times\bm\theta_\perp
+{\bm n}\times (gD_\mu[\mathscr{A}]\bm\theta_\perp)
-{\bm n} \times (gD_\mu[\mathscr{A}]\bm\omega) \}
\nonumber \\
&=g^{-1}
(D_\mu[\mathscr{A}]{\bm n}) \cdot
\{D_\mu[\mathscr{A}](\bm\omega-\bm\theta_\perp) \times \bm n \}
\nonumber \\
&= g^{-1}
(D_\mu[\mathscr{A}]{\bm n}) \cdot
\{D_\mu[\mathscr{A}](\bm\omega_\perp - \bm\theta_\perp) \times \bm n \} ,
\label{eq:dX^2}\end{aligned}$$ where we have used (\[gtn\]) and (\[gtA\]) in the third equality, and in the last equality we have decomposed $\bm\omega-\bm\theta_\perp$ into the parallel component $\bm\omega_\parallel= \omega_\parallel \bm n$ and perpendicular component $\bm\omega_\perp-\bm\theta_\perp$ and used the fact that the parallel part does not contribute, because $
D_\mu[\mathscr{A}] \bm\omega_\parallel \times \bm n
= D_\mu[\mathscr{A}]( \omega_\parallel \bm n) \times \bm n
= \{ \bm n \partial_\mu \omega_\parallel + \omega_\parallel D_\mu[\mathscr{A}]\bm n \} \times \bm n
= \omega_\parallel (D_\mu[\mathscr{A}]\bm n ) \times \bm n .
$ Therefore, the local gauge transformation II does not change $\mathbb X^2$.
Then the average over the spacetime of (\[eq:dX\^2\]) reads $$\begin{aligned}
\delta\int d^4x\frac12\mathbb X_\mu^2
&=g^{-1} \int d^4x
(D_\mu[\mathscr{A}]{\bm n}) \cdot
\{ D_\mu[\mathscr{A}](\bm\omega_\perp-\bm\theta_\perp) \times \bm n\}
\nonumber \\
&= \int d^4x
\mathbb X_\mu\cdot
D_\mu[\mathscr{A}](\bm\omega_\perp-\bm\theta_\perp)
\nonumber \\
&=- \int d^4x
(\bm\omega_\perp-\bm\theta_\perp)\cdot
D_\mu[\mathscr{A}]\mathbb X_\mu
\nonumber \\
&=- \int d^4x
(\bm\omega_\perp-\bm\theta_\perp)\cdot
D_\mu[\mathbb V]\mathbb X_\mu ,
\label{minX2}\end{aligned}$$ where we have used (\[def:X\]) in the second equality and integration by parts in the third equality. Hence, the minimizing condition (\[MAGcond\]) for arbitrary $\bm\omega_\perp$ and $\bm\theta_\perp$ yields a gauge-fixing condition in differential form:[^4] $$\mathbb F_{\rm MA} = \bm{\chi}
:=D_\mu[\mathbb V]\mathbb X_\mu
\label{dMAG}
\equiv0 .$$ Note that (\[dMAG\]) denotes two conditions, since $\bm{n} \cdot \bm{\chi} =0$, which follows from the identity $ \bm{n} \cdot D_\mu[\mathbb V]\mathbb X_\mu = 0$. Therefore, the minimization condition (\[MAGcond\]) works as a gauge fixing condition, except in the case of the gauge transformation II, i.e., $\bm\omega_\perp(x)=\bm\theta_\perp(x)$. In fact, the condition (\[dMAG\]) does not transform covariantly, except in the case of the gauge transformation II, because the gauge transformation of the condition (\[dMAG\]) reads $$\begin{aligned}
\delta \bm{\chi}
= g \bm{\chi} \times (\bm{\omega}_\parallel +\bm{\theta}_\perp )
- g^2 \mathbb{X}_\mu \times [\mathbb{X}_\mu \times (\bm{\omega}_\perp -\bm{\theta}_\perp )]
+ D_\mu[\mathbb{V}]D_\mu[\mathbb{V}](\bm{\omega}_\perp -\bm{\theta}_\perp ) .
\label{gtF}\end{aligned}$$ For $\bm\omega_\perp(x)=\bm\theta_\perp(x)$, the condition (\[dMAG\]) transforms covariantly, because $
\delta \bm{\chi}
= g \bm{\chi} \times (\bm{\omega}_\parallel +\bm{\omega}_\perp )
= g \bm{\chi} \times \bm{\omega}.
$ Here the local rotation of $\bm{n}$, i.e., $
\delta{\bm n}(x) =g{\bm n}(x) \times {\bm\theta}_\perp(x)
$, leads to $\delta \bm{\chi}=0$ on $\bm{\chi}=0$. Moreover, the U(1)$_{local}^{\omega}$ part in $G=SU(2)_{local}^{\omega}$ is not affected by this condition. Hence, the gauge symmetry corresponding to $\bm\omega_\parallel(x)$ remains unbroken.
Therefore, if we impose the condition (\[MAGcond\]) on the CFN-Yang–Mills theory, we have a gauge theory with the local gauge symmetry $G'=SU(2)_{local}^{\omega=\theta}$ corresponding to the gauge transformation parameter $\bm\omega(x)=(\bm\omega_\parallel(x),\bm\omega_\perp(x)=\bm\theta_\perp(x))$, which is the diagonal SU(2) part $\tilde{G}^{\omega=\theta}_{local}$ of the original $\tilde{G}^{\omega,\theta}_{local}$. The local gauge symmetry $G'=SU(2)_{local}^{\omega=\theta}$ is the same as the gauge symmetry II.
The form of the condition (\[dMAG\]) is identical to that of the MAG fixing condition for the CFN variables (see, e.g., [@Kondo04]). However, (\[dMAG\]) and (\[MAGcond\]) are completely different from the conventional MAG fixing condition [@tHooft81; @KLSW87], which has been used to this time to fix the off-diagonal part of the local gauge symmetry SU(2) of the original Yang-Mills theory (based on the Cartan decomposition) keeping the U(1) part intact, because the MAG introduced in this paper plays the role of eliminating the extra gauge symmetry generated by using the CFN variables and leaves the full SU(2) local gauge symmetry. Therefore, we call (\[MAGcond\]) \[and (\[dMAG\])\] the [*new MAG*]{} (the differential form). [^5] The new Yang-Mills theory obtained by imposing the nMAG on the CFN-Yang-Mills theory is called [*Yang–Mills theory II*]{} hereafter. Among the three gauge degrees of freedom $\bm{\omega}=(\bm{\omega}_\perp,\bm{\omega}_\parallel)$ and two degrees of freedom $\bm{\theta}_\perp$ in the CFN-Yang–Mills theory, two extra gauge degrees of freedom were eliminated by imposing the two conditions expressed by the nMAG, $\mathbb F_{\rm MA}=0$, and then the remaining degrees of freedom in Yang–Mills theory II are those in the same as the original Yang-Mills theory I. In the Yang–Mills theory II, we can impose any further gauge-fixing condition for fixing the diagonal SU(2) after the nMAG is imposed, e.g., instead of Landau in Fig. \[fig:sym-cfn-ym\]. In fact, we can furthermore impose the conventional MAG, if desired. This leads to the possibility of examining the gauge invariance even after the nMAG. In the previous approach, the MAG is one of the gauge fixings and there is no specific reason to take the MAG (except for the coincidence of the degrees of freedom). But in our approach the MAG plays a different and distinguished role. Even after imposing the nMAG, Yang–Mills theory II has the full SU(2) symmetry.
Our viewpoint for the CFN-Yang–Mills theory resolves in a natural way the crucial issue of the discrepancy in the independent degrees of freedom between the two theories, i.e., the original Yang-Mills theory and the CFN-Yang–Mills theory. Moreover, it reveals the necessity of adopting the nMAG in the CFN-Yang–Mills theory, although the conventional MAG is merely one choice of the gauge fixings. To the best of our knowledge to this time, this point had not been correctly understood.
In the paper [@KKMSS05; @KKMSSI05], we performed Monte Carlo simulations of the CFN-Yang-Mills theory for the first time, by imposing the nMAG and the Lattice–Landau gauge (LLG) simultaneously.[^6] Here, the LLG fixes the local gauge symmetry $G'=SU(2)_{local}^{\omega=\theta}$ and the MAG imposed in Refs.[@KKMSS05; @KKMSSI05] is the nMAG mentioned above, not the conventional MAG. In general, we can impose any gauge fixing condition instead of LLG, in addition to the nMAG in numerical simulations. This is an advantage of our viewpoint for the CFN-Yang–Mills theory. Yang–Mills theory II was constructed on a vacuum selected in a gauge invariant way among the possible vacua of the CFN-Yang–Mills theory, since the nMAG is satisfied for the CFN field configurations realizing the minimum of the functional $\int d^4x\frac12\mathbb X_\mu^2$, and the minimum $$\begin{aligned}
\min_{\bm\omega,\bm\theta} \int d^4x\frac12\mathbb X_\mu^2 \end{aligned}$$ is gauge invariant in the sense that it does no longer change the value with respect to the enlarged local gauge transformation. Therefore, [*the nMAG is a gauge-invariant criterion for choosing a vacuum on which Yang–Mills theory II is defined from the vacua of the CFN-Yang–Mills theory*]{}, although the nMAG is not necessarily a unique prescription for selecting the gauge-invariant vacuum. This demonstrates the quite different role played by the nMAG compared with the conventional MAG. The original Yang-Mills theory with the local gauge symmetry $G=SU(2)_{local}^{\omega}$, i.e., Yang-Mills theory I, is reproduced from the CFN-Yang–Mills theory by fixing the field variable $\bm{n}(x)$ as $\bm{n}(x) \equiv \bm{n}_{\infty}:=(0,0,1)$ at the all spacetime points.
Independent variables of the respective theory
----------------------------------------------
In order to clarify which variables are independent variables in the respective theory, we write the partition function of the respective theory with the integration measure, up to the gauge fixing term and the associated Faddeev–Popov ghost term to be investigated in [@KMS05b].
The partition function of the original Yang–Mills theory (Yang–Mills I) is in the Euclidean formulation: $$\begin{aligned}
Z_{{\rm YM}} = \int \mathcal{D}\mathscr{A}_\mu \exp (-S_{{\rm YM}}[\mathscr{A}]) .\end{aligned}$$ By introducing the auxiliary color field $\bm{n}(x)$, the CFN-Yang–Mills theory is first defined by a partition function written in terms of $\bm{n}(x)$ and $\mathscr{A}_\mu(x)$, $$\begin{aligned}
\tilde{Z}_{{\rm YM}} = \int \mathcal{D}\bm{n} \delta(\bm n\cdot\bm n-1)
\int \mathcal{D}\mathscr{A}_\mu \exp (-S_{{\rm YM}}[\mathscr{A}]) ,
\label{Z}\end{aligned}$$ and then it is rewritten in terms of the CFN variables $
(\bm n,c_\mu,\mathbb X_\mu)
$ as $$\begin{aligned}
\tilde{Z}_{{\rm YM}}
= \int \mathcal{D}\bm{n} \delta(\bm n\cdot\bm n-1)
\int \mathcal{D}c_\mu \int
\mathcal{D}\mathbb{X}_\mu \delta(\bm n\cdot\mathbb X_\mu)
J
\exp (-\tilde S_{\rm YM}[\bm n,c,\mathbb X]) , \end{aligned}$$ where $J$ is the Jacobian associated with the change of variables from $
(\bm n, \mathscr A_\mu)
$ to $
(\bm n,c_\mu,\mathbb X_\mu)
$ and the action $\tilde S_{\rm YM}[\bm n,c,\mathbb X]$ is obtained by substituting the CFN decomposition of $\mathscr A_\mu$ into $S_{{\rm YM}}[\mathscr{A}]$: $$\begin{aligned}
\tilde S_{\rm YM}[\bm n,c,\mathbb X]
=S_{\rm YM}[\mathscr A] .\end{aligned}$$
In order to fix the enlarged symmetry in the CFN-Yang–Mills theory and retain only the gauge symmetry II, we impose the constraint $\bm{\chi}[ \mathscr{A},\bm{n}]=0$ (the new maximal Abelian gauge). Then, we write unity in the form $$\begin{aligned}
1 = \int \mathcal{D} \bm{\chi}^\theta \delta(\bm{\chi}^\theta)
= \int\! \mathcal{D}\bm\theta\delta(\bm\chi^\theta)
\det\left(\frac{\delta\bm\chi^\theta}{\delta{\bm\theta}}\right) ,\end{aligned}$$ where $\bm{\chi}^\theta$ is the constraint written in terms of the gauge-transformed variable, i.e., $\bm{\chi}^\theta:=\bm{\chi}[ \mathscr{A},\bm{n}^\theta]$, and then we insert this into the functional integral (\[Z\]). This yields $$\begin{aligned}
Z_{{\rm YM}} = \int \mathcal{D}\bm{n} \delta(\bm n\cdot\bm n-1) \int \mathcal{D}\mathscr{A}_\mu \int\! \mathcal{D}\bm\theta\delta(\bm\chi^\theta)
\det\left(\frac{\delta\bm\chi^\theta}{\delta{\bm\theta}}\right)
\exp (-S_{{\rm YM}}[\mathscr{A}]) .\end{aligned}$$ Then we cast the partition function of the CFN-Yang–Mills theory II into the form[^7] $$\begin{aligned}
\tilde{Z}_{{\rm YM}}
=& \int \mathcal{D}\bm{n} \delta(\bm n\cdot\bm n-1)
\int \mathcal{D}c_\mu \int
\mathcal{D}\mathbb{X}_\mu \delta(\bm n\cdot\mathbb X_\mu)
J
\nonumber\\
& \times \int\! \mathcal{D}\bm\theta\delta(\bm\chi^\theta)
\det\left(\frac{\delta\bm\chi^\theta}{\delta{\bm\theta}}\right)
\exp (-\tilde S_{\rm YM}[\bm n,c,\mathbb X]) . \end{aligned}$$ We next perform the change of variables $\bm{n} \rightarrow \bm{n}^{\theta}$ obtained through a local rotation by the angle $\theta$ and the corresponding gauge transformations II for the other CFN variables $c_\mu$ and $\mathbb X_\mu$: $c_\mu, \mathbb{X}_\mu \rightarrow c_\mu^{\theta}, \mathbb{X}_\mu^{\theta}$. From the gauge invariance II of the action $\tilde S_{\rm YM}[\bm n,c,\mathbb X]$ and the measure $
\mathcal{D}\bm{n} \delta(\bm n\cdot\bm n-1)
\mathcal{D}c_\mu
\mathcal{D}\mathbb{X}_\mu \delta(\bm n\cdot\mathbb X_\mu)
$, we can rename the dummy integration variables $\bm{n}^{\theta}, c_\mu^{\theta}, \mathbb{X}_\mu^{\theta}$ as $\bm{n}, c_\mu, \mathbb{X}_\mu$. Thus the integrand does not depend on $\theta$, and the gauge volume $\int\! \mathcal{D}\bm\theta$ can be removed: $$\begin{aligned}
\tilde{Z}_{{\rm YM}}
=& \int\! \mathcal{D}\bm\theta
\int \mathcal{D}\bm{n} \delta(\bm n\cdot\bm n-1)
\int \mathcal{D}c_\mu \int
\mathcal{D}\mathbb{X}_\mu \delta(\bm n\cdot\mathbb X_\mu)
J
\nonumber\\
& \times \delta(\bm\chi)
\det\left(\frac{\delta\bm\chi}{\delta{\bm\theta}}\right)
\exp (-\tilde S_{\rm YM}[\bm n,c,\mathbb X]) .\end{aligned}$$ Note that the Faddeev–Popov determinant $\det\left(\frac{\delta\bm\chi}{\delta{\bm\theta}}\right)$ can be rewritten into another form, $
\Delta_{FP}^{nMAG}
:= \det\left(\frac{\delta\bm\chi}{\delta{\bm\theta}}\right)_{\bm{\chi}=0}
= \det\left(\frac{\delta\bm\chi}{\delta\bm n^\theta}\right)_{\bm{\chi}=0}.
$ This is the same as the determinant called the Shabanov determinant [@Shabanov99b], $
\Delta_{S}[\mathscr{A}_\mu, \bm{n}]
:= \det \left|{\delta \bm{\chi} \over \delta \bm{n}} \right|_{\bm{\chi}=0} ,
$ which gurantees the equivalence of Yang-Mills theories I and II. From our viewpoint, therefore, the Shabanov determinant is simply the Faddeev–Popov determinant associated with the nMAG. Thus, the the partition function of Yang-Mills theory II is given by $$\begin{aligned}
Z_{{\rm YM}}^\prime
=& \int \mathcal{D}\bm{n} \delta(\bm n\cdot\bm n-1)
\int \mathcal{D}c_\mu \int
\mathcal{D}\mathbb{X}_\mu \delta(\bm n\cdot\mathbb X_\mu)
J
\nonumber\\ & \times
\delta(\tilde{\bm\chi})
\Delta_{FP}^{nMAG}
\exp (-\tilde S_{\rm YM}[\bm n,c,\mathbb X]) , \end{aligned}$$ where the constraint is written in terms of the CFN variables: $$\tilde{\bm\chi}
:=\tilde{\bm\chi} [\bm n, c,\mathbb X]
:=D^\mu[\mathbb V ]\mathbb X_\mu, \quad
\mathbb V_\mu \equiv c_\mu \bm{n}
+g^{-1}\partial_\mu \bm{n} \times \bm{n} .$$ In Yang–Mills theory II, the independent variables are regarded as $\bm{n}(x)$, $c_\mu(x)$ and $\mathbb{X}_\mu(x)$.
In order to obtain a completely gauge-fixed theory, we must repeat the gauge-fixing procedures after imposing both the nMAG and the gauge fixing condition for SU(2) symmetry, e.g., the Landau gauge $\partial^\mu \mathscr{A}_\mu(x)=0$. According to the clarification of the symmetry in the CFN-Yang–Mills theory explained above, we can obtain the unique Faddeev-Popov ghost terms associated with the gauge fixing conditions adopted in quantization. This is another advantage of our viewpoint for the CFN-Yang–Mills theory. The explicit derivation of the FP ghost term has been worked out in a separate paper [@KMS05b].
Gauge invariance and observables
--------------------------------
The above consideration shows that the gauge invariant quantities in Yang–Mills theory II must be regarded as those which are invariant under the gauge transformation II. In this sense, $\mathbb{X}_\mu(x)^2$ is a gauge invariant quantity and must have a definite physical meaning. Therefore, the vacuum condensation $\left< \mathbb{X}_\mu(x)^2 \right>$ could be an important physical quantity. In fact, recalling that the Skyrme–Faddeev model [@FN97; @MS04] is derived from this vacuum condensation as pointed out in [@Kondo04], the vacuum condensation of mass dimension two, $\left< \mathbb{X}_\mu(x)^2 \right>$, could be related to this observable [@GSZ01; @Kondo01; @Slavnov04; @Slavnov05; @Kondo05].
Surprisingly, we can write the gauge-invariant mass term using the CFN variables, $$\begin{aligned}
\mathscr{L}_{M}' := \frac{1}{2} M^2 \mathbb{X}_\mu(x)^2 , \end{aligned}$$ in Yang–Mills theory II with the gauge symmetry II.
In the CFN-Yang–Mills theory with the enlarged gauge symmetry, this term can be identified with the kinetic term of the scalar field $\bm{\phi}(x)$ in the adjoint representation (\[X2\]), $$\begin{aligned}
\tilde{\mathscr{L}}_{M} = \frac{1}{2} (D_\mu[\mathscr{A}] \bm{\phi}(x))^2 ,
\quad \bm{\phi}(x) := g^{-1}M \bm{n}(x) .\end{aligned}$$ This can also be rewritten as a the gauge-invariant mass term similar to the Stückelberg type, $$\begin{aligned}
\tilde{\mathscr{L}}_{M} =&
\frac{1}{2} M^2 (\mathscr{A}_\mu(x) - \mathbb{V}_\mu(x))^2 ,
\nonumber\\
\mathbb{V}_\mu(x)
=& \bm{n}(x) [{\bm n}(x)\cdot\mathscr{A}_\mu(x)]
+g^{-1}\partial_\mu \bm{n}(x)\times \bm{n}(x) ,\end{aligned}$$ where $\mathbb{V}_\mu(x)$ is written in terms of $\bm{n}(x)$ and the original Yang-Mills field $\mathscr{A}_\mu(x)$. The $\bm{n}(x)$ field is written as a quite complicated composite field of the original Yang-Mills field $\mathscr{A}_\mu(x)$, after the gauge fixing (see [@KKMSS05; @KKMSSI05]). Further details will be discussed elsewhere.
Global gauge symmetry
---------------------
We have imposed the nMAG to fix the off-diagonal symmetry of the local gauge symmetry $\tilde{G}^{\omega,\theta}_{local}$ and to keep the local gauge symmetry $SU(2)_{local}^{\omega=\theta}$. Moreover, the nMAG also breaks the global gauge symmetry $\tilde{G}^{\omega,\theta}_{global}:=SU(2)_{global}^{\omega} \times [SU(2)/U(1)]_{global}^{\theta}$ into $SU(2)_{global}^{\omega=\theta}$. In fact, the global parameters $\bm\theta$ and $\bm\omega$ yield the change $$\begin{aligned}
\delta\frac12\mathbb X_\mu^2(x)
&=
(D_\mu[\mathscr{A}]{\bm n}(x)) \cdot
\{ [ \mathscr{A}_\mu(x) \times (\bm\omega^\perp - \bm\theta^\perp)] \times \bm n(x) \}
\nonumber\\
&=
(D_\mu[\mathscr{A}]{\bm n}(x)) \cdot (\bm\omega^\perp - \bm\theta^\perp)
[{\bm n}(x) \cdot \mathscr{A}_\mu(x)]
,
\label{eq:dX^2b}\end{aligned}$$ and the right-hand side is nonzero in general, since $D_\mu[\mathscr{A}]{\bm n}(x)$ is perpendicular to ${\bm n}(x)$. Therefore, Yang-Mills theory II, i.e., the CFN-Yang-Mills theory with the nMAG, has the local gauge symmetry $SU(2)_{local}^{\omega=\theta}$ as well as the global gauge symmetry $SU(2)_{global}^{\omega=\theta}$. These are the same local and global gauge symmetries as in the original Yang-Mills theory.
Moreover, we can impose one more gauge fixing condition to fix the remaining local gauge symmetry $SU(2)_{local}^{\omega=\theta}$, so that it maintains the global symmetry $SU(2)_{global}^{\omega=\theta}$, e.g., the Landau gauge. After imposing the nMAG and one more gauge fixing condition, the original local gauge symmetry, $\tilde{G}^{\omega,\theta}_{local}$, is completely fixed, while the global symmetry, $SU(2)_{global}^{\omega=\theta}$, is left intact.
We must focus on the quantities which are invariant under $SU(2)_{global}^{\omega=\theta}$ from the viewpoint of color confinement. In other words, only the $SU(2)_{global}^{\omega=\theta}$ singlet in CFN-Yang–Mills theory can have physical meaning. In fact, we have measured only the $SU(2)_{global}$ invariant quantities in the numerical simulations based on a new algorithm preserving $SU(2)_{global}^{\omega=\theta}$ in Refs.[@KKMSS05; @KKMSSI05]. In general, the spontaneous breakdown of the color symmetry $SU(2)_{global}^{\omega=\theta}$ could occur. In this case, we do not know what happens in the CFN-Yang–Mills theory and how the equivalence between the two theories is modified. This issue should be investigated in subsequent works.
Conclusion and discussion
=========================
We have shown that the CFN-Yang-Mills theory (the Yang-Mills theory written in terms of the CFN variables $\bm{n},c_\mu$ and $\mathbb{X}_\mu$) has the local gauge symmetry $\tilde{G}=SU(2)_{local}^{\omega} \times [SU(2)/U(1)]_{local}^{\theta}$, which is larger than the local gauge symmetry $G=SU(2)_{local}$ of the original Yang-Mills theory. We have imposed the new MAG to reduce the local gauge symmetry to the diagonal part, i.e., $G'=SU(2)_{local}^{\omega=\theta}$. This procedure explicitly breaks the global gauge symmetry $SU(2)_{global}^{\omega} \times [SU(2)/U(1)]_{global}^{\theta}$ to $SU(2)_{global}^{\omega=\theta}$ simultaneously. Then Yang-Mills theory II, i.e., the CFN-Yang-Mills theory at the new MAG, has the same local and global gauge symmetries as the original Yang-Mills theory, before the conventional gauge fixing is imposed. The new MAG is used as a criterion for choosing the vacuum of Yang-Mills theory II in a gauge invariant way from the vacua of the CFN-Yang–Mills theory.
The local gauge symmetry $G'=SU(2)^{\omega=\theta}$ of Yang-Mills theory II is identical to the gauge symmetry II defined in [@Kondo04], i.e., a local rotation of $\bm{n}$, $\mathbb{X}_\mu$ and an Abelian-like gauge transformation of $c_\mu$. Therefore, the transformation properties of the CFN variables under the gauge symmetry II can lead to a new set of gauge invariant operators and observables (vacuum condensates) which were previously unexpected, e.g., a gauge-invariant mass term $
\frac{1}{2} M^2 \mathbb{X}_\mu(x)^2
$ and a composite operator of mass dimension two, $\mathbb{X}_\mu(x)^2$, and its condensate, $\left< \mathbb{X}_\mu(x)^2 \right>$.
To quantize Yang-Mills theory II, we must introduce an appropriate gauge fixing condition to fix the gauge symmetry II, $G'=SU(2)_{local}^{\omega=\theta}$, in the conventional sense. We can definitely obtain the gauge-fixing term and the associated Faddeev–Popov ghost term. For example, we can choose the Landau gauge as the gauge-fixing condition that keeps the global gauge symmetry $SU(2)_{global}^{\omega=\theta}$ unbroken. In fact, we have tested this framework by Monte Carlo simulations on a lattice [@KKMSS05; @KKMSSI05] and succeeded in extracting novel nonperturbative features of the Yang-Mills theory. The conventional MAG could have been used at this stage breaking the global gauge symmetry. Therefore, the new MAG introduced above is completely different from the conventional MAG. The new MAG is a logical necessity, while the conventional MAG is merely one of a number of possible gauge-fixing choices. This may shed new light on the role of MAG in Yang-Mills theory.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by a Grant-in-Aid for Scientific Research (C)14540243 from the Japan Society for the Promotion of Science (JSPS), and in part by a Grant-in-Aid for Scientific Research on Priority Areas (B)13135203 from the Ministry of Education, Culture, Sports, Science and Technology (MEXT).
Local Gauge Transformations I and II {#Appendix:gt}
====================================
For the CFN decomposition of the gauge field (\[CFN\]), the non-Abelian field strength $\mathscr{F}_{\mu\nu}(x)$ is decomposed as $$\begin{aligned}
\mathscr{F}_{\mu\nu} [\mathscr{A}]
:=& \partial_\mu \mathscr{A}_\nu - \partial_\nu \mathscr{A}_\mu + g \mathscr{A}_\mu \times \mathscr{A}_\nu
\nonumber\\
=& \mathscr{F}_{\mu\nu} [\mathbb{V}] + \mathscr{F}_{\mu\nu} [\mathbb{X}] + g \mathbb{V}_\mu \times \mathbb{X}_\nu + g \mathbb{X}_\mu \times \mathbb{V}_\nu
\nonumber\\
=& \mathscr{F}_{\mu\nu} [\mathbb{V}] + D_\mu[\mathbb{V}] \mathbb{X}_\nu - D_\nu[\mathbb{V}] \mathbb{X}_\mu + g \mathbb{X}_\mu \times \mathbb{X}_\nu ,\end{aligned}$$ where $
D_\mu[\mathbb{V}] := \partial_\mu + g \mathbb{V}_\mu \times
$ is the covariant derivative in the background field $\mathbb{V}_\mu$.
The field strength $\mathscr{F}_{\mu\nu} [\mathbb{V}]$ is further decomposed as $$\begin{aligned}
\mathscr{F}_{\mu\nu} [\mathbb{V}]
=& \mathscr{F}_{\mu\nu} [\mathbb{B}] + \mathscr{F}_{\mu\nu} [\mathbb{C}] + g \mathbb{B}_\mu \times \mathbb{C}_\nu + g \mathbb{C}_\mu \times \mathbb{B}_\nu
\nonumber\\
:=& \mathbb{H}_{\mu\nu} + \mathbb{E}_{\mu\nu} := \mathbb{G}_{\mu\nu} , \end{aligned}$$ where the two kinds of field strength are defined by $$\begin{aligned}
\mathbb{H}_{\mu\nu} :=& \mathscr{F}_{\mu\nu} [\mathbb{B}]
= \partial_\mu \mathbb{B}_\nu - \partial_\nu \mathbb{B}_\mu + g \mathbb{B}_\mu \times \mathbb{B}_\nu
\\
\mathbb{E}_{\mu\nu} :=& \mathscr{F}_{\mu\nu} [\mathbb{C}] + g \mathbb{B}_\mu \times \mathbb{C}_\nu + g \mathbb{C}_\mu \times \mathbb{B}_\nu
.\end{aligned}$$ Due to the special definition of $\mathbb{B}_\mu$, the ’magnetic field’ strength $\mathbb{H}_{\mu\nu}$ is rewritten as $$\begin{aligned}
\mathbb{H}_{\mu\nu}
&= - g \mathbb{B}_\mu \times \mathbb{B}_\nu
= - g^{-1} (\partial_\mu \bm{n} \times \partial_\nu \bm{n})
= H_{\mu\nu} \bm{n},
\\
H_{\mu\nu} &:= - g^{-1} \bm{n} \cdot (\partial_\mu \bm{n} \times \partial_\nu \bm{n}) ,\end{aligned}$$ where we have used the fact that $\mathbb{H}_{\mu\nu}$ is parallel to $\bm{n}$. Similarly, the ’electric field’ strength $\mathbb{E}_{\mu\nu}$ is parallel to $\bm{n}$: $$\begin{aligned}
\mathbb{E}_{\mu\nu} = E_{\mu\nu} \bm{n}, \quad
E_{\mu\nu} := \partial_\mu c_\nu - \partial_\nu c_\mu .\end{aligned}$$
The gauge transformations of the CFN variables are given as follows.
(the passive or quantum gauge transformation):
$$\begin{aligned}
\delta_\omega \bm{n} =& 0 ,
\\
\delta_\omega c_\mu =& \bm{n} \cdot D_\mu[\mathscr{A}] \bm{\omega} ,
\\
\delta_\omega \mathbb{X}_\mu =& D_\mu[\mathscr{A}] \bm{\omega} - \bm{n}( \bm{n} \cdot D_\mu[\mathscr{A}] \bm{\omega}) ,
\\
\Longrightarrow & \delta_\omega \mathbb{B}_\mu = 0 ,
\quad
\delta_\omega \mathbb{V}_\mu
= \bm{n}( \bm{n} \cdot D_\mu[\mathscr{A}] \bm{\omega}) . \end{aligned}$$
The gauge transformation for the field strength can be obtained in the similar way: $$\begin{aligned}
\delta_\omega \mathbb{E}_{\mu\nu}
=& \bm{n} \delta_\omega {E}_{\mu\nu}
= \bm{n} \{ \partial_\mu (\bm{n} \cdot D_\nu[\mathscr{A}] \bm{\omega}) - \partial_\nu (\bm{n} \cdot D_\mu[\mathscr{A}] \bm{\omega}) \} ,
\\
\delta_\omega \mathbb{H}_{\mu\nu}
=& \bm{n} \delta_\omega {H}_{\mu\nu} = 0 . \end{aligned}$$
(the active or background gauge transformation):
$$\begin{aligned}
\delta_{\omega}' \bm{n} =& g \bm{n} \times \bm{\omega'} ,
\\
\delta_{\omega}' c_\mu =& \bm{n} \cdot \partial_\mu \bm{\omega'} ,
\\
\delta_{\omega}' \mathbb{X}_\mu =& g \mathbb{X}_\mu \times \bm{\omega'} ,
\\
\Longrightarrow & \delta_{\omega}' \mathbb{B}_\mu
= D_\mu[\mathbb{B}] \bm{\omega'} - (\bm{n} \cdot \partial_\mu \bm{\omega'}) \bm{n} ,
\quad
\delta_{\omega}' \mathbb{V}_\mu = D_\mu[\mathbb{V}] \bm{\omega'} . \end{aligned}$$
The gauge transformation for the field strength can be obtained in a similar way. It is easy to show that $\mathbb{G}_{\mu\nu}:=\mathscr{F}_{\mu\nu} [\mathbb{V}]$ (the sum $\mathbb{E}_{\mu\nu}+\mathbb{H}_{\mu\nu}$) is subject to the adjoint rotation $$\begin{aligned}
\delta_\omega' \mathbb{G}_{\mu\nu}
=& g \mathbb{G}_{\mu\nu} \times \bm{\omega}' ,\end{aligned}$$ while this is not the case for the individual quantities, $E_{\mu\nu}$ and $H_{\mu\nu}$: $$\begin{aligned}
\delta_\omega' \mathbb{E}_{\mu\nu}
=& g \mathbb{E}_{\mu\nu} \times \bm{\omega}' + \bm{n} \{ \partial_\mu (\bm{n} \cdot \partial_\nu \bm{\omega}') - \partial_\nu (\bm{n} \cdot \partial_\mu \bm{\omega}') \} ,
\\
\delta_\omega' \mathbb{H}_{\mu\nu} =& g \mathbb{H}_{\mu\nu} \times \bm{\omega}' - \bm{n} \{ \partial_\mu (\bm{n} \cdot \partial_\nu \bm{\omega}') - \partial_\nu (\bm{n} \cdot \partial_\mu \bm{\omega}') \} .
$$ Hence, the squared field strength has the SU(2)$_{\rm II}$ invariance $$\begin{aligned}
\delta_\omega' \mathbb{G}_{\mu\nu}^2
=& 0 .\end{aligned}$$ The inner product of $\mathbb{G}_{\mu\nu}$ with $\bm{n}$ is also SU(2)$_{\rm II}$ invariant: $$\begin{aligned}
\delta_\omega'( \bm{n} \cdot \mathbb{G}_{\mu\nu})
& \equiv \delta_\omega' G_{\mu\nu}
= 0 ,
\\
G_{\mu\nu} &= \partial_\mu c_\nu - \partial_\nu c_\mu - g^{-1} \bm{n} \cdot (\partial_\mu \bm{n} \times \partial_\nu \bm{n}), \quad
c_\mu
={\bm n} \cdot \mathscr{A}_\mu .\end{aligned}$$ This is not the case for the individual quantities, $E_{\mu\nu}$ and $H_{\mu\nu}$.
Moreover, we can show that $$\begin{aligned}
\delta_\omega' ( D_\mu[\mathbb{V}] \mathbb{X}_\nu - D_\nu[\mathbb{V}] \mathbb{X}_\mu)
=& g (D_\mu[\mathbb{V}] \mathbb{X}_\nu - D_\nu[\mathbb{V}] \mathbb{X}_\mu) \times \bm{\omega}' ,
\\
\delta_\omega' (\mathbb{X}_\mu \times \mathbb{X}_\nu) =& g (\mathbb{X}_\mu \times \mathbb{X}_\nu) \times \bm{\omega}' ,\end{aligned}$$ Therefore, all the inner products among $$\begin{aligned}
\bm{n}, \quad
\mathbb{G}_{\mu\nu}:=\mathbb{E}_{\mu\nu}+\mathbb{H}_{\mu\nu}, \quad
\mathbb{X}_\mu \times \mathbb{X}_\nu
&\leftarrow \text{parallel~to~} \bm{n}
\\
\mathbb{X}_\mu, \quad
( D_\mu[\mathbb{V}] \mathbb{X}_\nu - D_\nu[\mathbb{V}] \mathbb{X}_\mu),
&\leftarrow \text{perpendicular~to~} \bm{n}\end{aligned}$$
possess SU(2)$_{\rm II}$ gauge invariance. For example, we have $$\begin{aligned}
\delta_\omega' (D_\mu[\mathbb{V}] \mathbb{X}_\nu - D_\nu[\mathbb{V}] \mathbb{X}_\mu)^2
= 0 ,
\quad
\delta_\omega' (\mathbb{X}_\mu \times \mathbb{X}_\nu)^2 = 0 ,
\\
\delta_\omega'[\mathbb{G}_{\mu\nu} \cdot (g \mathbb{X}_\mu \times \mathbb{X}_\nu)]
= 0. \end{aligned}$$
In particular, when $\bm{\omega}'(x)$ is parallel to $\bm{n}$, i.e., $\bm{\omega}'(x)=\theta'(x) \bm{n}(x)$, we obtain :
$$\begin{aligned}
\delta_{\theta}' \bm{n} =& 0 ,
\\
\delta_{\theta}' c_\mu =& \partial_\mu \theta' ,
\\
\delta_{\theta}' \mathbb{X}_\mu =& g \mathbb{X}_\mu \times \theta' \bm{n} ,
\\
\Longrightarrow & \delta_{\theta}' \mathbb{B}_\mu
= 0 ,
\quad
\delta_{\theta}' \mathbb{V}_\mu = \bm{n} \partial_\mu \theta' . \end{aligned}$$
Note that $\bm{n}$ and $\mathbb{B}_\mu$ are invariant under the U(1)$_{\rm II}$ gauge transformation, while $c_\mu$ transforms as the U(1)$_{\rm II}$ gauge field. It is easy to show the local U(1)$_{\rm II}$ gauge invariance of the field strengths, i.e., $$\begin{aligned}
\delta_{\theta}' \mathbb{E}_{\mu\nu} = 0 ,
\quad
\delta_{\theta}' \mathbb{H}_{\mu\nu} = 0 ,\end{aligned}$$ which is also consistent with the initial definitions: $
\mathbb{E}_{\mu\nu} = \bm{n}(\partial_\mu c_\nu - \partial_\nu c_\mu) ,
$ $
\mathbb{H}_{\mu\nu} = - g \mathbb{B}_\mu \times \mathbb{B}_\nu .
$ Therefore, the dimension two composite operators $\mathbb{B}_\mu^2$ and $\sqrt{\mathbb{H}_{\mu\nu}^2}$, and the dimension four operators $\mathbb{H}_{\mu\nu}^2$ are gauge invariant under the local U(1)$_{\rm II}$ gauge transformation [@Kondo04; @KKMSS05].
[999]{} Y.M. Cho, Phys. Rev. D 21, 1080-1088 (1980).\
Y.M. Cho, Phys. Rev. D 23, 2415–2426 (1981).
L. Faddeev and A.J. Niemi, \[hep-th/9807069\], Phys. Rev. Lett. 82, 1624-1627 (1999).
T. Tsurumaru, I. Tsutsui and A. Fujii, \[hep-th/0005064\], Nucl. Phys. B589, 659-668 (2000).
S.V. Shabanov, \[hep-th/9903223\], Phys. Lett. B 458, 322-330 (1999).
S.V. Shabanov, \[hep-th/9907182\], Phys. Lett. B 463, 263-272 (1999).
H. Gies, \[hep-th/0102026\], Phys. Rev. D 63, 125023 (2001).
L. Faddeev and A.J. Niemi, \[hep-th/9610193\], Nature 387, 58 (1997).
E. Langmann and A.J. Niemi, \[hep-th/9905147\], Phys. Lett. B 463, 252–256 (1999).
L. Faddeev and A.J. Niemi, \[hep-th/0101078\], Phys. Lett. B 525, 195-200 (2002).
W.S. Bae, Y.M. Cho and S.W. Kimm, \[hep-th/0105163\], Phys. Rev. D 65, 025005 (2001).
G.K. Savvidy, Phys. Lett. B 71, 133-134 (1977).
N.K. Nielsen and P. Olesen, Nucl. Phys. B 144, 376-396 (1978).
Y.M. Cho, hep-th/0301013.\
Y.M. Cho, M.L. Walker and D.G. Pak, \[hep-th/0209208\], JHEP 05, 073 (2004).\
Y.M. Cho and D.G. Pak, \[hep-th/0201179\], Phys. Rev. D65, 074027 (2002).
K.-I. Kondo, \[hep-th/0404252\], Phys.Lett. B 600, 287–296 (2004).\
K.-I. Kondo, \[hep-th/0410024\], Intern. J. Mod. Phys. A 20, 4609-4614 (2005).
S. Kato, K.-I. Kondo, T. Murakami, A. Shibata and T. Shinohara, hep-ph/0504054.
S. Kato, K.-I. Kondo, T. Murakami, A. Shibata, T. Shinohara and S. Ito, \[hep-lat/0509069\], Phys. Lett. B 632, 326–332 (2006).
M. Hirayama and C.-G. Shi, \[hep-th/0310042\], Phys. Rev. D69, 045001 (2004).
S.V. Shabanov, \[hep-lat/0110065\], Phys. Lett. B 522, 201-209 (2001).
L. Dittmann, T. Heinzl and A. Wipf, \[hep-lat/0210021\], JHEP 0212, 014 (2002).
K.-I. Kondo, T. Murakami and T. Shinohara, \[hep-th/0504198\], Eur. Phys. J. C 42, 475–481 (2005).
N.S. Manton, Nucl. Phys. B 126, 525–541 (1977).
G. ’t Hooft, Nucl.Phys. B 190 \[FS3\], 455-478 (1981).
A. Kronfeld, M. Laursen, G. Schierholz and U.-J. Wiese, Phys. Lett. B 198, 516-520 (1987).
N. Manton and P. Sutcliffe, Topological solitons (Cambridge Univ. Press, 2004).
F.V. Gubarev, L. Stodolsky and V.I. Zakharov, \[hep-ph/0010057\], Phys. Rev. Lett. 86, 2220–2222 (2001).\
F.V. Gubarev and V.I. Zakharov, \[hep-ph/0010096\], Phys. Lett. B 501, 28–36 (2001).
K.-I. Kondo, \[hep-th/0105299\], Phys. Lett. B 514, 335–345 (2001).\
K.-I. Kondo, \[hep-th/0306195\], Phys. Lett. B 572, 210-215 (2003).\
K.-I. Kondo, T. Murakami, T. Shinohara and T. Imai, \[hep-th/0111256\], Phys. Rev. D 65, 085034 (2002).
A.A. Slavnov, hep-th/0407194.
A.A. Slavnov, Phys. Lett. B 608, 171-176 (2005).
K.-I. Kondo, \[hep-th/0504088\], Phys. Lett. B 619, 377–386 (2005).
[^1]: We emphasize this fact, which is pointed out in Ref.[@Shabanov99b], because it is important from our viewpoint.
[^2]: Shabanov[@Shabanov99b] argued that it is possible to consider a more general transformation of the field ${\bm n}(x)$, even a nonlocal one, keeping the condition ${\bm n}(x)^2=1$. However, it is unrealistic to consider an explicit transformation other than the local rotation treated in this paper.
[^3]: This transformation law was obtained by Shabanov [@Shabanov99b]. In it, $\delta \bm{n}(x)$ is not specified and is left undetermined on the right-hand side, based on the viewpoint that $\delta \bm{n}(x)$ should be determined by the choice of the constraint condition $\bm{\chi}(\bm{n},\mathscr{A}) \equiv \bm{\chi}(\bm{n},c,\mathbb{X})=0$, which reduces the degrees of freedom to the original ones (by solving $\delta \bm{\chi}=0$ on the hypersurface $\bm{\chi}=0$). The new MAG defined below is consistent with the local rotation of $\bm{n}$, as a part of the gauge transformation II. Therefore, our result is in agreement with the claim given in [@Shabanov99b] \[see (\[dMAG\]) and (\[gtF\])\].
[^4]: Of course, it is trivial that two constraints are necessary and sufficient to eliminate the two extra degrees introduced by the CFN decomposition. Indeed, the form of the new MAG condition is the same as that given in Ref.[@Shabanov99b]. However, there is no argument to identify which part of the enlarged gauge symmetry is fixed by this constraint, even in the case of a local rotation for $\bm{n}$. This is important to avoid the misunderstanding that appears in the literature. It is not the naively expected $[SU(2)/U(1)]^{\theta}_{local}$ symmetry that is fixed by two constraint conditions, as is clear from our argument. The correct identification of the gauge symmetry influences the explicit form of the BRST transformation [@KMS05b], and finding the correct way of implementing the CFN decomposition on a lattice [@KKMSS05; @KKMSSI05].
[^5]: Note that (\[MAGcond\]) is more general than (\[dMAG\]), since (\[dMAG\]) is the differential form, which is valid only in the absence of Gribov copies. The condition (\[MAGcond\]) is the most general MAG condition, which can be used also in numerical simulations on a lattice and works even if Gribov copies exist and leads to the true minimum, while (\[dMAG\]) leads only to the local minimum along the gauge orbit.
[^6]: A possible algorithm for the numerical simulation was proposed in [@Shabanov01], and actual simulations were first attempted in [@DHW02]. However, from our point of view, they cannot be identified with the CFN-Yang-Mills theory. Here we point out that only the field $\bm{n}$ was constructed in these works, and the simulation results show the breaking of the global SU(2) invariance, even in the Landau gauge, which cannot be regarded as the correct implementation of the CFN decomposition on a lattice. It is the essence to preserve the color symmetry. Details are given in Refs.[@KKMSS05; @KKMSSI05].
[^7]: It is not difficult to show that the Jacobian $J$ for the change of variables from ${\bm n}^A$ and $\mathscr{A}_\mu^A$ to ${\bm n}^A, c_\mu$ and $\mathbb{X}_\mu^A$ is equal to 1, if the integration measures $\mathcal{D}\bm{n}$ and $\mathcal{D}\mathbb{X}_\mu$ are understood to be written in terms of independet degrees of freedom by taking into account the constraints $\bm n\cdot\bm n=1$ and $\bm n\cdot\mathbb X_\mu=0$. In fact, only the independent degrees of freedom have been used to calculate the explicit form of the effective potential [@Cho03; @Kondo04]. Therefore, we do not pay special attention to the Jacobian $J$ in what follows.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A closed form expression for the three-electron Hylleraas integral involving the inverse quadratic power of one inter-particle coordinate is obtained, and recursion relations are derived for positive powers of other coordinates. This result is suited for high precision calculations of relativistic effects in lithium and light lithium-like ions.'
author:
- Krzysztof Pachucki
- Mariusz Puchalski
title: 'Extended Hylleraas three-electron integral'
---
Introduction
============
The subject of this work is the extended three-electron Hylleraas integrals involving $1/r_{ij}^2$ and $1/r_i^2$ terms, which appear in matrix elements of relativistic operators in the Breit-Pauli Hamiltonian. These integrals are defined as $$\begin{aligned}
f(n_1,n_2,n_3,n_4,n_5,n_6) &=& \int\frac{d^3\,r_1}{4\,\pi}\int\frac{d^3\,r_2}{4\,\pi}
\int\frac{d^3\,r_3}{4\,\pi}\,e^{-w_1\,r_1-w_2\,r_2-w_3\,r_3}\nonumber \\ &&
r_{23}^{n_1-1}\,r_{31}^{n_2-1}\,r_{12}^{n_3-1}\,
r_{1}^{n_4-1}\,r_{2}^{n_5-1}\,r_{3}^{n_6-1} ,
\label{01}\end{aligned}$$ with one of $n_i$ equal to $-1$, and all other $n_i$ are nonnegative integers. They have been studied in detail in a series of papers by King [@king_lit1; @king_lit2; @king_lit3] and by Yan [*et al.*]{} [@yan_lit]. Their approach is based on the expansion of $1/r_{ij}^n$ in an infinite series of some orthogonal polynomials. The resulting multiple summation is performed with the help of convergence accelerators. In some special cases [@king_lit1] this expansion can be avoided and Eq. (\[01\]) can be expressed in terms of the two-electron Hylleraas integral. Using these methods leading relativistic and QED corrections to energies [@king_rev; @hyl_lit], isotope-shift [@lit_iso], and $g$-factors [@lit_g] have been calculated to a high degree of accuracy. The analytic approach to Hylleraas integral has been so far much less successful. Fromm and Hill in [@fh] were able to obtain an analytic expression for a more general integral with exponents in $r_i$, as well as in $r_{ij}$. However, their expression is quite complicated, involves multivalued dilogarithmic functions, and thus is of limited use. In parallel Remiddi in [@remiddi] obtained a simple analytic expression for the Hylleraas integral with $n_i=0$. Recently, together with Remiddi we derived recursion relations [@rec] for Hylleraas integrals for arbitrary large powers of $r_i$ and $r_{ij}$. This result allows for a convenient calculation of the nonrelativistic wave function of the lithium atom and light lithium-like ions.
In this work we present an analytic approach which allows for a fast and high precision calculation of extended 3-electron Hylleraas integrals. This sets the ground for improving the precision of theoretical energy levels by including higher order relativistic corrections [@fw]. We obtain here the closed form expression for the master integral $f(-1,0,0,0,0,0)$ and derive recursion relations for increasing values of arguments $n_i$ of $f(-1,n_2,n_3,n_4,n_5,n_6)$. The other integrals, such us $f(n_1,n_2,n_3,-1,n_5,n_6)$, can be obtained by using already derived recursion relations for $f(n_1,n_2,n_3,n_4,n_5,n_6)$ with nonnegative $n_i$ followed by a one dimensional numerical integration with respect to the corresponding parameter $w_i$.
Recursion relations for $r_{23}^{-2}$ integral
==============================================
Our derivation is based on integration by parts identities which are commonly used for the calculations of multi-loop Feynman diagrams [@parts]. We follow here the former work [@rec] and first consider the integral $G$ $$\begin{aligned}
G(m_1,m_2,m_3;m_4,m_5,m_6) &=& \frac{1}{8\,\pi^6}\,\int d^3k_1\int d^3k_2\int d^3k_3\,
(k_1^2+u_1^2)^{-m_1}\,(k_2^2+u_2^2)^{-m_2} \nonumber\\ &&(k_3^2+u_3^2)^{-m_3}\,
(k_{32}^2+w_1^2)^{-m_4}\,(k_{13}^2+w_2^2)^{-m_5}\,(k_{21}^2+w_3^2)^{-m_6}, \label{02}\end{aligned}$$ which is related to $f$ by: $f(0,0,0,0,0,0,0) = G(1,1,1,1,1,1)|_{u_1=u_2=u_3=0}$. The following 9 integration by part identities are valid because the integral of the derivative of a function vanishing at infinity vanishes, $$\begin{aligned}
&&0 \equiv {\rm id}(i,j) =
\int d^3k_1\int d^3k_2\int d^3k_3\,\frac{\partial}{\partial\,{\vec k_i}}
\Bigl[ \vec k_j\,(k_1^2+u_1^2)^{-1}
\nonumber \\ &&
(k_2^2+u_2^2)^{-1}\,(k_3^2+u_3^2)^{-1}
(k_{32}^2+w_1^2)^{-1}\,(k_{13}^2+w_2^2)^{-1}\,(k_{21}^2+w_3^2)^{-1}
\Bigr] ,
\label{03}\end{aligned}$$ where $i,j=1,2,3$. The reduction of the scalar products from the numerator leads to the identities for the linear combination of the $G$-function. If any of the arguments is equal to 0, then $G$ becomes a known two-electron Hylleraas type integral, Eq. (\[B6\]). The explicit form of all 9 identities is presented in Eq. (\[A1\]) and the whole derivation presented below is performed with the help of [*MATHEMATICA*]{} program for symbolic computation.
In the first step of deriving recursion relations we take the difference ${\rm id}(3,2)-{\rm id}(2,2)$ and use it as an equation for $G(1, 2, 1, 1, 1, 1)$, $$\begin{aligned}
G(1, 2, 1, 1, 1, 1)\,(u_2^2 - u_3^2 + w_1^2) &=&
G(1, 1, 1, 0, 1, 2) - G(1, 1, 1, 1, 0, 2)
\nonumber \\ &&
+ G(1, 1, 1, 1, 1, 1) - G(1, 2, 0, 1, 1, 1)
\nonumber \\ &&
+ G(1, 2, 1, 0, 1, 1) - 2\,G(1, 1, 1, 2, 1, 1)\,w_1^2
\nonumber \\ &&
+ G(1, 1, 1, 1, 1, 2)\,(w_2^2 - w_1^2 - w_3^2) .
\label{04}\end{aligned}$$ Similarly, the difference ${\rm id}(2,3)-{\rm id}(3,3)$ is used to obtain $G(1, 1, 2, 1, 1, 1)$, $$\begin{aligned}
G(1, 1, 2, 1, 1, 1)\,(u_2^2 - u_3^2 - w_1^2) &=&
G(1, 0, 2, 1, 1, 1) - G(1, 1, 1, 0, 2, 1)
\nonumber \\ &&
- G(1, 1, 1, 1, 1, 1) + G(1, 1, 1, 1, 2, 0)
\nonumber \\ &&
- G(1, 1, 2, 0, 1, 1) + 2\,G(1, 1, 1, 2, 1, 1)\,w_1^2
\nonumber \\ &&
+ G(1, 1, 1, 1, 2,1)\,(w_1^2 +w_2^2 -w_3^2) .
\label{05}\end{aligned}$$ These two equations are used now to derive recursions in $n_2$ and $n_3$. With the help of the trivial formula $$\int_0^\infty du\,\frac{e^{-u\,r}}{r} = \frac{1}{r^2} ,
\label{06}$$ one integrates with respect to $u_1$, which lowers the first argument $n_1$ to $-1$. Next, one differentiates with respect to $u_2,u_3,w_1,w_2,w_3$ at $u_2=u_3=0$ to generate arbitrary powers of $r_{13}, r_{12}, r_1, r_2, r_3$ and obtains quite long recursion relations for $n_2$ and $n_3$, $$\begin{aligned}
f(-1,n_2+2,n_3,n_4,n_5,n_6) &=&
\frac{1}{w_1^2\,w_3}\,\biggl[
n_4\,(n_4-1)\,(n_2+1)\,f(-1,n_2,n_3,n_4-2,n_5,n_6+1)
\nonumber \\ &&
-n_5\,(n_5-1)\,(n_2+1)\,f(-1,n_2,n_3,n_4,n_5-2,n_6+1)
\nonumber \\ &&
+n_6\,(1+n_2+2\,n_4+n_6)\,(n_2+1)\,f(-1,n_2,n_3,n_4,n_5,n_6-1)
\nonumber \\ &&
-n_3\,(n_3-1)\,n_6\,f(-1,n_2+2,n_3-2,n_4,n_5,n_6-1)
\nonumber \\ &&
+n_4\,(n_4-1)\,n_6\,f(-1,n_2+2,n_3,n_4-2,n_5,n_6-1)
\nonumber \\ &&
-2\,n_4\,(n_2+1)\,w_1\,f(-1,n_2,n_3,n_4-1,n_5,n_6+1)
\nonumber \\ &&
-2\,n_6\,(n_2+1)\,w_1\,f(-1,n_2,n_3,n_4+1,n_5,n_6-1)
\nonumber \\ &&
-2\,n_4\,n_6\,w_1\,f(-1,n_2+2,n_3,n_4-1,n_5,n_6-1)
\nonumber \\ &&
+n_6\,w_1^2\,f(-1,n_2+2,n_3,n_4,n_5,n_6-1)
\nonumber \\ &&
+2\,n_5\,(n_2+1)\,w_2\,f(-1,n_2,n_3,n_4,n_5-1,n_6+1)
\nonumber \\ &&
-(2+n_2+2\,n_4+2\,n_6)\,(n_2+1)\,w_3\,f(-1,n_2,n_3,n_4,n_5,n_6)
\nonumber \\ &&
+(n_3-1)\,n_3\,w_3\,f(-1,n_2+2,n_3-2,n_4,n_5,n_6)
\nonumber \\ &&
-n_4\,(n_4-1)\,w_3\,f(-1,2+n_2,n_3,-2+n_4,n_5,n_6)
\nonumber \\ &&
+2\,(n_2+1)\,w_1\,w_3\,f(-1,n_2,n_3,n_4+1,n_5,n_6)\,w_1\,w_3
\nonumber \\ &&
+2\,n_4\,w_1\,w_3\,f(-1,n_2+2,n_3,n_4-1,n_5,n_6)
\nonumber \\ &&
+(n_2+1)\,(w_1^2-w_2^2+w_3^2)\,f(-1,n_2,n_3,n_4,n_5,n_6+1)
\nonumber \\ &&
+\delta_{n_4}\,w_3\,\Gamma(n_3+n_5-1,n_2+n_6+1,-1;w_2,w_3,0)
\nonumber \\ &&
-\delta_{n_3}\,w_3\,\Gamma(-1+n_4+n_5,n_6,n_2;w_1+w_2,w_3,0)
\nonumber \\ &&
-(n_2+n_6+1)\,\delta_{n_4}\,\Gamma(n_3+n_5-1,n_2+n_6,-1;w_2,w_3,0)
\nonumber \\ &&
+n_6\,\delta_{n_3}\,\Gamma(n_4+n_5-1,n_6-1,n_2;w_1+w_2,w_3,0)
\nonumber \\ &&
+(n_2+1)\,\delta_{n_5}\,\Gamma(n_6-1,n_3+n_4-1,n_2;w_3,w_1,0)
\biggr] ,
\label{07}\\
f(-1,n_2,n_3+2,n_4,n_5,n_6) &=&\frac{1}{w_1^2\,w_2}\,\biggl[
-n_2\,(n_2-1)\,n_5\,f(-1,-n_2-2,n_3+2,n_4,n_5-1,n_6)
\nonumber \\ &&
+(n_3+1)\,n_4\,(n_4-1)\,f(-1,n_2,n_3,n_4-2,n_5+1,n_6)
\nonumber \\ &&
+(n_3+1)\,n_5\,(1+n_3+2\,n_4+n_5)\,f(-1,n_2,n_3,n_4,n_5-1,n_6)
\nonumber \\ &&
-(n_3+1)\,n_6\,(n_6-1)\,f(-1,n_2,n_3,n_4,n_5+1,n_6-2)
\nonumber \\ &&
+n_4\,(n_4-1)\,n_5\,f(-1,n_2,n_3+2,n_4-2,n_5-1,n_6)
\nonumber \\ &&
-2\,(n_3+1)\,n_4\,w_1\,f(-1,n_2,n_3,n_4-1,n_5+1,n_6)
\nonumber \\ &&
-2\,(n_3+1)n_5\,w_1\,f(-1,n_2,n_3,n_4+1,n_5-1,n_6)
\nonumber \\ &&
-2\,n_4\,n_5\,w_1\,f(-1,n_2,2+n_3,n_4-1,n_5-1,n_6)
\nonumber \\ &&
+(n_3+1)\,(w_1^2+w_2^2-w_3^2)\,f(-1,n_2,n_3,n_4,n_5+1,n_6)
\nonumber \\ &&
+n_5\,w_1^2\,f(-1,n_2,n_3+2,n_4,n_5-1,n_6)
\nonumber \\ &&
+n_2\,(n_2-1)\,w_2\,f(-1,n_2-2,n_3+2,n_4,n_5,n_6)
\nonumber \\ &&
-(n_3+1)\,(n_3+2\,n_4+2\,n_5+2)\,w_2\,f(-1,n_2,n_3,n_4,n_5,n_6)
\nonumber \\ &&
-n_4\,(n_4-1)\,w_2\,f(-1,n_2,2+n_3,n_4-2,n_5,n_6)
\nonumber \\ &&
+2\,(n_3+1)\,w_1\,w_2\,f(-1,n_2,n_3,n_4+1,n_5,n_6)
\nonumber \\ &&
+2\,n_4\,w_1\,w_2\,f(-1,n_2,n_3+2,n_4-1,n_5,n_6)
\nonumber \\ &&
+2\,(n_3+1)\,n_6\,w_3\,f(-1,n_2,n_3,n_4,n_5+1,n_6-1)
\nonumber \\ &&
+\delta_{n_4}\,w_2\,\Gamma(n_3+n_5+1,n_2+n_6-1,-1;w_2,w_3,0)
\nonumber \\ &&
-\delta_{n_2}\,w_2\,\Gamma(n_4+n_6-1,n_5,n_3;w_1+w_3,w_2,0)
\nonumber \\ &&
+(n_3+1)\delta_{n_6}\,\Gamma(n_2+n_4-1,n_5-1,n_3;w_1,w_2,0)
\nonumber \\ &&
-(1+n_3+n_5)\,\delta_{n_4}\,\Gamma(n_3+n_5,n_2+n_6-1,-1;w_2,w_3,0)
\nonumber \\ &&
+n_5\,\delta_{n_2}\,\Gamma(n_4+n_6-1,n_5,n_3-1;w_1+w_3,w_2,0)
\biggr] ,
\label{08}\end{aligned}$$ where $\delta_n$ denotes Kronecker delta $\delta_{n,0}$ and $\Gamma$ is a two-electron Hylleraas integral, which is defined in Eq. (\[B1\]). These recursions assume that the values of $f(-1,0,0,n_4,n_5,n_6)$, $f(-1,1,0,n_4,n_5,n_6)$, $f(-1,0,1,n_4,n_5,n_6)$ and $f(-1,1,1,n_4,n_5,n_6)$ are known. We calculate master integrals for the last three cases explicitly and express them in terms of two-electron Hylleraas integrals as in [@king_lit1], $$\begin{aligned}
f(-1,1,1,0,0,0) &=& \int\frac{d^3\,r_1}{4\,\pi}\int\frac{d^3\,r_2}{4\,\pi}
\int\frac{d^3\,r_3}{4\,\pi}\frac{e^{-w_1\,r_1-w_2\,r_2-w_3\,r_3}}{r_{23}^2\,r_1\,r_2\,r_3}
\nonumber \\ &=& \frac{1}{w_1^2\,(w_2^2-w_3^2)}\,\ln\frac{w_2}{w_3} =
\frac{1}{w_1^2}\,\Gamma(0,0,-1;w_2,w_3,0) ,\label{09}\\ \nonumber \\
f(-1,0,1,0,0,0) &=& \int\frac{d^3\,r_1}{4\,\pi}\int\frac{d^3\,r_2}{4\,\pi}
\int\frac{d^3\,r_3}{4\,\pi}\frac{e^{-w_1\,r_1-w_2\,r_2-w_3\,r_3}}{r_{23}^2\,r_{13}\,r_1\,r_2\,r_3}
\nonumber \\ &=& \frac{1}{4\,w_1^2\,w_2}\,
\biggl[\ln^2\biggl(\frac{w_2}{w_2 + w_3}\biggr) - \ln^2\biggl(\frac{w_2}{w_1 + w_2 + w_3}\biggr) +
2\,{\rm Li}_2\biggl(\frac{w_2}{w_2 + w_3}\biggr) \nonumber \\ &&
-2\,{\rm Li}_2\biggl(\frac{w_2}{w_1 + w_2 + w_3}\biggr) +
2\,{\rm Li}_2\biggl(1 - \frac{w_3}{w_2}\biggr) -
2\,{\rm Li}_2\biggl(1 - \frac{w_1 + w_3}{w_2}\biggr)\biggr]\nonumber \\ &=&
\frac{1}{w_1^2}\,\Bigl[\Gamma(-1,0,-1;w_3,w_2,0)-\Gamma(-1,0,-1;w_1+w_3,w_2,0)\Bigr] ,
\label{10}\\
\nonumber \\
f(-1,1,0,0,0,0) &=& \int\frac{d^3\,r_1}{4\,\pi}\int\frac{d^3\,r_2}{4\,\pi}
\int\frac{d^3\,r_3}{4\,\pi}\frac{e^{-w_1\,r_1-w_2\,r_2-w_3\,r_3}}{r_{23}^2\,r_{12}\,r_1\,r_2\,r_3}
\nonumber \\ &=&
\frac{1}{w_1^2}\,\Bigl[\Gamma(-1,0,-1;w_2,w_3,0)-\Gamma(-1,0,-1;w_1+w_2,w_3,0)\Bigr] ,
\label{11}\end{aligned}$$ where Li$_2$ is a dilogarithmic function, Eq. (\[C8\]). The recursion relations in $n_4, n_5, n_6$ are obtained by differentiation with respect to $w_1, w_2$ and $w_3$ $$\begin{aligned}
f(-1,n_2,n_3,n_4,n_5,n_6) &=& \frac{1}{w_1^2}\,\biggl[
-n_4\,(n_4-1)\,f(-1,n_2,n_3,n_4-2,n_5,n_6)
\nonumber \\ &&
+2\,n_4\,w_1\,f(-1,n_2,n_3,n_4-1,n_5,n_6)
\nonumber \\ &&
+\delta_{n_4}\,\Gamma(n_5+n_3-1,n_6+n_2-1,-1;w_2,w_3,0)
\nonumber \\ &&
-\delta_{n_3}\,\Gamma(n_4+n_5-1,n_6,-1;w_1+w_2,w_3,0)
\nonumber \\ &&
-\delta_{n_2}\,\Gamma(n_4+n_6-1,n_5,-1;w_1+w_3,w_2,0)\biggr] ,
\label{12}\end{aligned}$$ for $(n_2,n_3) \in \{(1,1), (1,0),(0,1)\}$.
The calculation of the integral $f(-1,0,0,n_4,n_5,n_6)$ is much more elaborate and we have to return to integration by parts identities, see Appendix A. These are 9 equations, which we solve against the following $X_{i=1,9}$ unknowns at $u_2=u_3=0$, $$\begin{aligned}
X_1 &=& G(1,2,1,1,1,1)\, u_1 , \nonumber \\
X_2 &=& G(1,1,2,1,1,1)\, u_1 , \nonumber \\
X_3 &=& G(1,1,1,1,2,1)\, u_1 , \nonumber \\
X_4 &=& G(1,1,1,1,1,2)\, u_1 , \nonumber \\
X_5 &=& G(1,2,1,1,1,1) , \nonumber \\
X_6 &=& G(1,1,2,1,1,1) , \nonumber \\
X_7 &=& G(1,1,1,1,2,1) , \nonumber \\
X_8 &=& G(1,1,1,1,1,2) , \nonumber \\
X_9 &=& G(2,1,1,1,1,1) .\label{13}\end{aligned}$$ The solution for $X_7$ and $X_8$ is $$\begin{aligned}
G(1, 1, 1, 1, 2, 1) &=& -G(2, 1, 1, 1, 1, 1)\,\frac{u_1^2}{2\,w^2_2} -
G(1, 1, 1, 2, 1, 1)\,\frac{w^2_1 + w^2_2 - w^2_3}{2\,w^2_2}
\nonumber \\ &&
+G(1, 1, 1, 1, 1, 1)\,\frac{3\,w^2_1 + w^2_2 - w^2_3}{4\,w^2_1\,w^2_2} +
\frac{F(u_1)}{4\,w_1^2\,w_2^2} ,\label{14}
\\ \nonumber \\
G(1, 1, 1, 1, 1, 2) &=& -G(2, 1, 1, 1, 1, 1)\,\frac{u_1^2}{2\,w^2_3} -
G(1, 1, 1, 2, 1, 1)\,\frac{w^2_1 - w^2_2 + w^2_3}{2\,w^2_3}
\nonumber \\ &&
+ G(1, 1, 1, 1, 1, 1)\,\frac{3\,w^2_1 - w^2_2 + w^2_3}{4\,w^2_1\,w^2_3} -
\frac{F(u_1)}{4\,w_1^2\,w_3^2} ,\label{15}\end{aligned}$$ where $$\begin{aligned}
F(u_1) &=& 2\,\biggl[G(1, 1, 1, 0, 2, 1)\,w_2^2 - G(1, 1, 1, 1, 2, 0)\,w_2^2
\nonumber \\ &&
+ G(2, 0, 1, 1, 1, 1)\,w_2^2 - G(2, 1, 1, 1, 1, 0)\,w_2^2
\nonumber \\ &&
- G(1, 1, 1, 0, 1, 2)\,w^2_3 + G(1, 1, 1, 1, 0, 2)\,w^2_3
\nonumber \\ &&
- G(2, 1, 0, 1, 1, 1)\,w^2_3 + G(2, 1, 1, 1, 0, 1)\,w^2_3\biggr] .\label{16}\end{aligned}$$ We now use explicit form of two-electron integrals in Eq. (\[B6\]) and integrate both equations with respect to $u_1$, $$\begin{aligned}
f(-1,0,0,0,1,0) &=& \frac{1}{2\,w_1^2\,w_2}
\Bigl[F + (2\,w_1^2 + w_2^2 - w_3^2)\,f(-1, 0, 0, 0, 0, 0)
\nonumber \\ &&
- w_1\,(w_1^2 + w_2^2 - w_3^2)\,f(-1, 0, 0, 1, 0, 0)\Bigr] ,\label{17}\\
f(-1,0,0,0,0,1) &=& \frac{1}{2\,w_1^2\,w_3}
\Bigl[-F + (2\,w_1^2 - w_2^2 + w_3^2)\,f(-1, 0, 0, 0, 0, 0)
\nonumber \\ &&
- w_1\,(w_1^2 - w_2^2 + w_3^2)\,f(-1, 0, 0, 1, 0, 0)\Bigr] ,\label{18}\end{aligned}$$ where $$\begin{aligned}
F &=& \int_0^\infty du_1\,F(u_1) \nonumber \\ &=&\frac{1}{2}\biggl[
2\,{\rm Li}_2\biggl( -\frac{w_2}{w_1}\biggr) - {\rm Li}_2\biggl(1 - \frac{w_2}{w_3}\biggr) +
{\rm Li}_2\biggl(1 - \frac{w_1 + w_2}{w_3}\biggr) - 2\,{\rm Li}_2\biggl(-\frac{w_3}{w_1}\biggr)
\nonumber \\ &&
+ {\rm Li}_2\biggl(\frac{w_2}{w_2 + w_3}\biggr) - {\rm Li}_2\biggl(\frac{w_3}{w_2 + w_3}\biggr) +
{\rm Li}_2\biggl(\frac{w_2}{w_1 + w_2 + w_3}\biggr) -
{\rm Li}_2\biggl(\frac{w_3}{w_1 + w_2 + w_3}\biggr)
\nonumber \\ &&
+ {\rm Li}_2\biggl(1 - \frac{w_3}{w_2}\biggr) - {\rm Li}_2\biggl(1 -
\frac{w_1 + w_3}{w_2}\biggr)
+\ln\biggl(\frac{w_2}{w_3}\biggr)\,\ln\biggl(\frac{w_1 + w_2 + w_3}{w_2 +
w_3}\biggr)\biggr]\nonumber \\
&=& \Gamma(0,-1,-2;w_2,w_1+w_3,0)- \Gamma(0,-1,-2;w_3,w_1+w_2,0)
\nonumber \\ &&
+ \Gamma(0,-1,-2,0;w_1,w_2) - \Gamma(0,-1,-2,0;w_1,w_3)
\nonumber \\ &&
+ w_2 \,\Gamma(-1,0,-1;w_1,0,w_2) - w_3 \,\Gamma(-1,0,-1;w_1,0,w_3)
\nonumber \\ &&
+ w_2 \,\Gamma(-1,0,-1;0,w_2,w_3) - w_3 \,\Gamma(-1,0,-1;0,w_3,w_2)\,.\label{19}\end{aligned}$$ Next, we multiply both equations by powers of $w_i$ to eliminate any $w_i$ from the denominator and differentiate with respect to $w_1, w_2$ and $w_3$. This leads to the following recursion relations in $n_5$ and $n_6$ of the $f$-function $$\begin{aligned}
f(-1,0,0,n_4,n_5+1,n_6) &=& \frac{1}{2\,w_1^2\,w_2}\Bigl[
n_4\,(n_4-1)\,(n_4 + 2\,n_5)\,f(-1,0,0,n_4-2,n_5,n_6)
\nonumber \\ &&
-2\,n_4\,(n_4-1)\,w_2\,f(-1,0,0,n_4-2,n_5+1,n_6)
\nonumber \\ &&
-n_4\,(3\,n_4 + 4\,n_5+1)\,w_1\,f(-1,0,0,n_4-1,n_5,n_6)
\nonumber \\ &&
+4\,n_4\,w_1\,w_2\,f(-1,0,0,n_4-1,n_5+1,n_6)
\nonumber \\ &&
+(n_4+1)\,(n_5-1)\,n_5\,f(-1,0,0,n_4,n_5-2,n_6)
\nonumber \\ &&
- 2\,(n_4+1)\,n_5\,w_2\,f(-1,0,0,n_4,n_5-1,n_6)
\nonumber \\ &&
- (n_4+1)\,(n_6-1)\,n_6\,f(-1,0,0,n_4,n_5,n_6-2)
\nonumber \\ &&
+ 2\,(n_4+1)\,n_6\,w_3\,f(-1,0,0,n_4,n_5,n_6-1)
\nonumber \\ &&
+ [(3\,n_4 + 2\,n_5+2)\,w_1^2 + (n_4+1)\,w_2^2 - (n_4+1)\,w_3^2]
\nonumber \\ &&
\times f(-1,0,0,n_4,n_5,n_6)
\nonumber \\ &&
-(n_5-1)\,n_5\,w_1\,f(-1,0,0,n_4+1,n_5-2,n_6)
\nonumber \\ &&
+2\,n_5\,w_1\,w_2\,f(-1,0,0,n_4+1,n_5-1,n_6)
\nonumber \\ &&
+(n_6-1)\,n_6\,w_1\,f(-1,0,0,n_4+1,n_5,n_6-2)
\nonumber \\ &&
-2\,n_6\,w_1\,w_3\,f(-1,0,0,n_4+1,n_5,n_6-1)
\nonumber \\ &&
- w_1\,(w_1^2 + w_2^2 - w_3^2)\,f(-1,0,0,n_4+1,n_5,n_6)
\nonumber \\ &&
+ F(n_4,n_5,n_6)\Bigr] ,\label{20}\\
f(-1,0,0,n_4,n_5,n_6+1) &=& \frac{1}{2\,w_1^2\,w_3}\,\Bigl[
n_4\,(n_4-1)\,(n_4 + 2\,n_6)\,f(-1,0,0,n_4-2,n_5,n_6)
\nonumber \\ &&
-2\,(n_4-1)\,n_4\,w_3\,f(-1,0,0,n_4-2,n_5,n_6+1)
\nonumber \\ &&
-n_4\,(3\,n_4 + 4\,n_6 + 1)\,w_1\,f(-1,0,0,n_4-1,n_5,n_6)
\nonumber \\ &&
+4\,n_4\,w_1\,w_3\,f(-1,0,0,n_4-1,n_5,n_6+1)
\nonumber \\ &&
- (n_4+1)\,(n_5-1)\,n_5\,f(-1,0,0,n_4,n_5-2,n_6)
\nonumber \\ &&
+ 2\,(n_4+1)\,n_5\,w_2\,f(-1,0,0,n_4,n_5-1,n_6)
\nonumber \\ &&
+ (n_4+1)\,(n_6-1)\,n_6\,f(-1,0,0,n_4,n_5,n_6-2)
\nonumber \\ &&
- 2\,(n_4+1)\,n_6\,w_3\,f(-1,0,0,n_4,n_5,n_6-1)
\nonumber \\ &&
+ [(3\,n_4 + 2\,n_6 + 2)\,w_1^2 - (n_4+1)\,w_2^2 + (n_4+1)\,w_3^2]
\nonumber \\ &&
\times f(-1,0,0,n_4,n_5,n_6)
\nonumber \\ &&
+(n_5-1)\,n_5\,w_1\,f(-1,0,0,n_4+1,n_5-2,n_6)
\nonumber \\ &&
-2\,n_5\,w_1\,w_2\,f(-1,0,0,n_4+1,n_5-1,n_6)
\nonumber \\ &&
-(n_6-1)\,n_6\,w_1\,f(-1,0,0,n_4+1,n_5,n_6-2)
\nonumber \\ &&
+2\,n_6\,w_1\,w_3\,f(-1,0,0,n_4+1,n_5,n_6-1)
\nonumber \\ &&
- w_1\,(w_1^2 - w_2^2 + w_3^2)\,f(-1,0,0,n_4+1,n_5,n_6)
\nonumber \\ &&
- F(n_4,n_5,n_6)\Bigr] ,\label{21}\end{aligned}$$ where $$\begin{aligned}
F(n_4,n_5,n_6) &=&
(-\partial_{w_1})^{n_4}\,(-\partial_{w_2})^{n_5}\,(-\partial_{w_3})^{n_6}\,F
\nonumber \\
&=&
(n_6 - n_5)\,\delta_{n_4}\,\Gamma(-1,n_5-1,n_6-1;0,w_2,w_3)
\nonumber \\ &&
+ w_2\,\delta_{n_4}\,\Gamma(-1,n_5,n_6-1;0,w_2,w_3)
\nonumber \\ &&
-w_3\,\delta_{n_4}\,\Gamma(-1,n_6,n_5-1;0,w_3,w_2)
\nonumber \\ &&
- (n_5-1)\,\delta_{n_6}\,\Gamma(n_4-1,0,n_5-2;w_1,0,w_2)
\nonumber \\ &&
+w_2\,\delta_{n_6}\,\Gamma(n_4-1,0,n_5-1;w_1,0,w_2)
\nonumber \\ &&
+ (n_6-1)\,\delta_{n_5}\,\Gamma(n_4-1,0,n_6-2;w_1,0,w_3)
\nonumber \\ &&
-w_3\,\delta_{n_5}\,\Gamma(-1+n_4,0,n_6-1;w_1,0,w_3)
\nonumber \\ &&
+ \Gamma(n_5,n_4+n_6-1,-2;w_2,w_1+w_3,0)
\nonumber \\ &&
-\Gamma(n_6,n_4+n_5-1,-2;w_3,w_1+w_2,0) .\label{22}\end{aligned}$$ What remains is the calculation of $f(-1,0,0,n_4,0,0)$. In the following we derive a differential equation for $h(w_1) \equiv f(-1,0,0,0,0,0)$, from which we obtain $f(-1,0,0,n_4,0,0)$. The solutions for $$G(2,1,1,1,1,1)=X_9 ,\label{23}$$ and for the difference $$X_1\,u_1^{-2}-X_5=0 ,\label{24}$$ form two algebraic equations, which however are too long to be written here. They involve the terms $G(2, 1, 1, 1, 1, 1)$, $G(2, 1, 1, 1, 1, 1) u_1^2$, $G(1, 1, 1, 1, 1, 1)$, $G(1, 1, 1, 1, 1, 1)\,u_1^{-2}$, $G(1, 1, 1, 2, 1, 1)$, $G(1,1, 1, 2, 1, 1)\,u_1^{-2}$, and the known two-electron terms, where one of the arguments of $G$-function is equal to 0. We integrate both equations in $u_1$ from $\epsilon$ to $\infty$, approach the limit $\epsilon \rightarrow 0$ and drop $\ln\epsilon$ $$\begin{aligned}
\int_\epsilon^\infty du_1\, G(2, 1, 1, 1, 1, 1) &=& -\frac{g(w_1)}{2} ,\nonumber \\
\int_\epsilon^\infty du_1\, u_1^2\,G(2, 1, 1, 1, 1, 1) &=& \frac{h(w_1)}{2} ,\nonumber \\
\int_\epsilon^\infty du_1\, G(1, 1, 1, 1, 1, 1) &=& h(w_1) ,\nonumber \\
\int_\epsilon^\infty du_1\, u_1^{-2}\,G(1, 1, 1, 1, 1, 1) &=&
g(w_1)-f(1,0,0,0,0,0) , \nonumber \\
\int_\epsilon^\infty du_1\, G(1, 1, 1, 2, 1, 1) &=& -\frac{h'(w_1)}{2} ,\nonumber \\
\int_\epsilon^\infty du_1\, u_1^{-2}\, G(1,1, 1, 2, 1, 1) &=&
-\frac{g'(w_1)}{2} +\frac{1}{2}\,\frac{\partial f(1,0,0,0,0,0)}{\partial w_1} ,
\label{25}\end{aligned}$$ where $$\begin{aligned}
h(w_1) &=& f(-1,0,0,0,0,0) ,\nonumber \\
g(w_1) &=& \int\frac{d^3\,r_1}{4\,\pi}\int\frac{d^3\,r_2}{4\,\pi}
\int\frac{d^3\,r_3}{4\,\pi}\,e^{-w_1\,r_1-w_2\,r_2-w_3\,r_3}
(\ln r_{23}+\gamma)\,r_{31}^{-1}\,r_{12}^{-1}\,
r_{1}^{-1}\,r_{2}^{-1}\,r_{3}^{-1} ,\label{26}\end{aligned}$$ and $$f(1,0,0,0,0,0) = -\frac{1}{w_2^2\,w_3^2}\,
\ln\biggl[\frac{w_1\,(w_1+w_2+w_3)}{(w_1+w_2)\,(w_1+w_3)}\biggr] .\label{27}$$ The set of both equations forms first order differential equations for $h(w_1)$ and $g(w_1)$. We eliminate $g(w_1)$ to obtain the following second order differential equation for $h(w_1)$ $$\begin{aligned}
&& w\,w_1^2\,h''(w_1) +
w_1\,\bigl[4\,w_1^2\,(w_1^2 - w_2^2 - w_3^2)+w\bigr]\,h'(w_1)
\nonumber \\
&&+\bigl[w_1^4 + 4\,w_1^2\,(w_1^2 - w_2^2 - w_3^2)-w\bigr]\,h(w_1) = R(w_1),
\label{28}\end{aligned}$$ where $$\begin{aligned}
R(w_1) &=& w_1\,w_2\,\ln\biggl(1 + \frac{w_1}{w_2}\biggr) +
w_1\,w_3\,\ln\biggl(1 + \frac{w_1}{w_3}\biggr) +
(w_2^2 - w_3^2)\,\ln\biggl(\frac{w_1 + w_3}{w_1 + w_2}\biggr)
\nonumber \\ &&
+2\,w_1^2\,\ln\biggl(\frac{w_1\,(w_1 + w_2 + w_3)}{(w_1 + w_2)\,(w_1 + w_3)}\biggr) +
(w_2^2 - w_3^2)\,F ,
\label{29}\end{aligned}$$ and $$\begin{aligned}
w &=& w_1^4+w_2^4+w_3^4-2\, w_1^2\, w_2^2 -2\, w_2^2\, w_3^2 -2\, w_1^2\, w_3^2\nonumber \\
&=& -(-w_1 + w_2 + w_3) (w_1 - w_2 + w_3) (w_1 + w_2 - w_3) (w_1 + w_2 + w_3) .
\label{30}\end{aligned}$$ Two linearly independent solutions of the homogenous equation are: $$\begin{aligned}
h_1(w_1) &=& \frac{1}{w_1\,\sqrt{w_2\,w_3}}\,{\rm
K}\biggl[\frac{(w_1+w_2-w_3)(w_1-w_2+w_3)}{4\,w_2\,w_3}\biggr] ,\nonumber \\
h_2(w_1) &=& \frac{1}{w_1\,\sqrt{w_2\,w_3}}\,{\rm
K}\biggl[\frac{(-w_1+w_2+w_3)(w_1+w_2+w_3)}{4\,w_2\,w_3}\biggr] ,
\label{31}\end{aligned}$$ where $K$ is a complete elliptic integral of the first kind as defined in Eq. (\[C1\]), and the Wronskian $W$ is $$W = h_1(w_1)\,h_2'(w_1) - h_1'(w_1)\,h_2(w_1) =
\frac{2\,\pi}{w\,w_1} ,\label{32}$$ where $w$ is defined in Eq. (\[30\]). The solution in Eq. (\[31\]) is valid for $w_1$ in the range $|w_2-w_3|<w_1<w_2+w_3$, because the elliptic integral $K$ has a branch cut for arguments exceeding 1. We use the identity in Eq. (\[D1\]), to obtain solution $h_1$ of the homogenous equation for $w_1>w_2+w_3$ $$h_1(w_1) = \frac{2}{w_1\,\sqrt{(w_1+w_2-w_3)(w_1-w_2+w_3)}}\,
K\biggl[\frac{4\,w_2\,w_3}{(w_1+w_2-w_3)(w_1-w_2+w_3)}\biggr] ,\label{35}$$ and $h_2$ for $w_1<|w_2-w_3|$ $$h_2(w_1) = \frac{2}{w_1\,\sqrt{(-w_1+w_2+w_3)(w_1+w_2+w_3)}}\,
K\biggl[\frac{4\,w_2\,w_3}{(-w_1+w_2+w_3)(w_1+w_2+w_3)}\biggr].\label{36}$$ The solution of the inhomogeneous equation is obtained by Euler’s method of variation of the constant, $$h(w_1) = \frac{h_1(w_1)}{2\,\pi}\,\int_{w_1}^{w_2+w_3}
dw'\,\frac{R(w')\,h_2(w')}{w'} +
\frac{h_2(w_1)}{2\,\pi}\,\int_{|w_2-w_3|}^{w_1}
dw'\,\frac{R(w')\,h_1(w')}{w'} .\label{33}$$ There is no additional term being a solution of homogenous equation, because $h(w_1)$ is finite for all values of $w_1$, but not $h_1(w_1)$ and $h_2(w_1)$. Therefore this is the right solution. Having obtained $f(-1,0,0,0,0,0)\equiv h(w_1)$ and $f(-1,0,0,1,0,0) = -h'(w_1)$ we calculate $f(-1,0,0,n_4,0,0)=h(w_1,n_4) =
(-\partial_{w_1})^{n_4}h(w_1)$ recursively. The inhomogeneous differential equation (\[28\]) is differentiated $n$-times with respect to $w_1$, to obtain $$\begin{aligned}
h(w_1, n+2) &=& \frac{1}{w\,w_1^2}\Bigl\{-(n-3)\,(n-2)^3\,(n-1)\,n\,h(w_1,n-4)
\nonumber \\ &&
+(n-2)\,(n-1)\,n\,(13 - 17\,n + 6\,n^2)\,w_1\,h(w_1,n-3)
\nonumber \\ &&
-(n-1)\,n\,\Bigl[(14 - 25\,n + 15\,n^2)\,w_1^2 - 2\,(n-1)^2\,w_s\Bigr]\,h(w_1,n-2)
\nonumber \\ &&
+2\,n\,w_1\,\Bigl[(3 - 5\,n + 10\,n^2)\,w_1^2 + (-1 + 3\,n - 4\,n^2)\,w_s\Bigr]\,h(w_1,n-1)
\nonumber \\ &&
+ \Bigl[-(4 + 10\,n + 15\,n^2)\,w_1^4 +
(1 - n^2)\,w_p^2 + 2\,(1 + 3\,n + 6\,n^2)\,w_1^2\,w_s\Bigr]\,h(w_1, n)
\nonumber \\ &&
+ w_1\,\Bigl[(5 + 6\,n)\,w_1^4 + (1 + 2\,n)\,w_p^2 - 2\,(3 + 4\,n)\,w_1^2\,w_s\Bigr]\,
h(w_1,n+1)
\nonumber \\ &&
+ R(w_1,n)\Bigr\} ,\label{39}\end{aligned}$$ where $w_s = w_2^2+w_3^2$, $w_p = w_2^2-w_3^2$ and $$\begin{aligned}
R(w_1,n) &=& (-\partial_{w_1})^n\,R(w_1)\nonumber \\
&=& -\frac{4\,(n-3)!}{w_1^{n-2}}+\frac{5\,w_2\,(n-2)!}{(w_1+w_2)^{n-1}}
+\frac{4\,(n-3)!}{(w_1+w_2)^{n-2}}+
\frac{5\,w_3\,(n-2)!}{(w_1+w_3)^{n-1}}
\nonumber \\ &&
+\frac{4\,(n-3)!}{(w_1+w_3)^{n-2}}
- \frac{2\,(w_2+w_3)^2\,(n-1)!}{(w_1+w_2+w_3)^n}
- \frac{4\,(w_2+w_3)\,(n-2)!}{(w_1+w_2+w_3)^{n-1}}
\nonumber \\ &&
-\frac{4\,(n-3)!}{(w_1+w_2+w_3)^{n-2}}
+\frac{(4\,w_2^2-w_3^2)\,(n-1)!}{(w_1+w_2)^n}
+\frac{(4\,w_3^2-w_2^2)\,(n-1)!}{(w_1+w_3)^n}
\nonumber \\ &&
+(w_2^2-w_3^2)\,F(n,0,0),\;\;\;\mbox{\rm for}\; n>2,
\nonumber \\
R(w_1,0) &=& R(w_1) , \nonumber \\
R(w_1,1) &=& -(w_2 + w_3)+\frac{4\,w_2^2 - w_3^2}{w_1+w_2} - 2\,\frac{(w_2+w_3)^2)}{w_1+w_2+w_3}
+\frac{4\,w_3^2-w_2^2}{w_1+w_3} - w_2\,\ln\biggl(1+\frac{w_1}{w_2}\biggr)
\nonumber \\ &&
- w_3\,\ln\biggl(1+\frac{w_1}{w_3}\biggr) -
4\,w_1\,\ln\biggl[\frac{w_1\,(w_1+w_2+w_3)}{(w_1+w_2)\,(w_1+w_3)}\biggr]
+(w_2^2-w_3^2)\,F(1,0,0) ,
\nonumber \\
R(w_1,2) &=& \frac{5\,w_2}{w_1+w_2}+\frac{5\,w_3}{w_1+w_3} - \frac{2\,(w_2+w_3)^2}{(w_1+w_2+w_3)^2}
- \frac{4\,(w_2+w_3)}{w_1+w_2+w_3}
\nonumber \\ &&
+\frac{4\,w_2^2-w_3^2}{(w_1+w_2)^2} +\frac{4\,w_3^2-w_2^2}{(w_1+w_3)^2}
+4\,\ln\biggl[\frac{w_1\,(w_1+w_2+w_3)}{(w_1+w_2)\,(w_1+w_3)}\biggr]
\nonumber \\ &&
+(w_2^2-w_3^2)\,F(2,0,0) .\label{40}\end{aligned}$$ In the case $w_1\approx w_{\rm sing} = w_2+w_3$ or $|w_2-w_3|$, the recursion in Eq. (\[39\]) is not numerically stable. Therefore, one instead of this recursion, calculates the recursion exactly at $w_1=w_{\rm sing}$, what corresponds to setting $w=0$ in Eq. (\[39\]), $$\begin{aligned}
h(w_1,n+1) &=&
\frac{-1}{w_1\,\Bigl[(5 + 6\,n)\,w_1^4 + (1 + 2\,n)\,w_p^2 - 2\,(3 + 4\,n)\,w_1^2\,w_s\Bigr]}
\nonumber \\ &&
\Bigl\{-(n-3)\,(n-2)^3\,(n-1)\,n\,h(w_1,n-4)
\nonumber \\ &&
+(n-2)\,(n-1)\,n\,(13 - 17\,n + 6\,n^2)\,w_1\,h(w_1,n-3)
\nonumber \\ &&
-(n-1)\,n\,\Bigl[(14 - 25\,n + 15\,n^2)\,w_1^2 - 2\,(n-1)^2\,w_s\Bigr]\,h(w_1,n-2)
\nonumber \\ &&
+2\,n\,w_1\,\Bigl[(3 - 5\,n + 10\,n^2)\,w_1^2 + (-1 + 3\,n - 4\,n^2)\,w_s\Bigr]\,h(w_1,n-1)
\nonumber \\ &&
+ \Bigl[-(4 + 10\,n + 15\,n^2)\,w_1^4 +
(1 - n^2)\,w_p^2 + 2\,(1 + 3\,n + 6\,n^2)\,w_1^2\,w_s\Bigr]\,h(w_1, n)
\nonumber \\ &&
+ R(w_1,n)\Bigr\} ,\label{41}\end{aligned}$$ where $w_1=w_{\rm sing}$. This completes the recursion relations for the extended Hylleraas integral with $r_{23}^{-2}$.
Numerical evaluation
====================
We pass now to numerical implementation of recursions and integration of the master integral in Eq. (\[33\]). All the computation is performed with extended precision arithmetic, namely quad and sextuple precision. Even higher precision, the octuple one, is used for checking numerical accuracy. The starting point is the master integral. One needs to calculate it with the highest possible accuracy, because the recursions depend most significantly on the value of initial terms. The integrand in Eq. (\[33\]) is a product of the function $R$ and the complete elliptic integral $K$. The function $R$ defined in Eq. (\[29\]) has singularities only at $w_i=0$ and $w_i=\infty$, and the complete elliptic integral has logarithmic singularities at $w_1=w_{\rm sing} \equiv |w_2-w_3|$ or $w_2+w_3$ which correspond to zeros of $w$ in Eq. (\[30\]). In the following we assume that $w_2-w_3\neq 0$. When $|w_2-w_3|+\epsilon_1<w_1<w_2+w_3-\epsilon_2$, the integral in Eq. (\[33\]) can be performed by the Gauss-Legendre quadrature [@nr]. We have verified that for $\epsilon_i\approx 0.2$ the integration with 100 points gives the accuracy of at least 32 digits if not more. For the cases $w_1>w_2+w_3+\epsilon_2$ and $w_1<|w_2-w_3|-\epsilon_1$, the integration contour is deformed on the complex plane to avoid the singularity. So this contour consists of two lines on a real axis, joined by a half-circle with origin at the singular point and integration is performed independently on each part using the Gauss-Legendre quadrature. When $w_1$ is close to the singular point, we first obtain $h(w_{\rm sing}\pm\epsilon)$, next we calculate $h(w_{\rm sing})$ by matching the Taylor expansion from recursion in Eq. (\[41\]) with $h(w_{\rm sing}\pm\epsilon)$ and in the last step we again use recursion in Eq. (\[41\]) to obtain $h(w_1)$. For $w_2-w_3 = 0$ we separate $R$ in Eq. (\[29\]) into the part which is free of $\ln w_1$ and the part which is proportional to $\ln w_1$. The first part is integrated using the Gauss-Legendre quadrature an the second one is integrated using the Gauss quadrature adapted for the logarithmic weight function. Several numerical results for some selected $w_i$ are presented in Table I.
$w_1$ $w_2$ $w_3$ $f(-1,0,0,0,0,0)$
------- ------- ------- ----------------------------------------
4.0 1.0 0.5 1.243735828073620173310981564244\[-1\]
4.0 1.0 1.0 9.855133136060504470218647797889\[-2\]
4.0 1.0 1.5 8.181412007841597436460514476518\[-2\]
4.0 1.0 2.0 6.983588391604680181982031823035\[-2\]
4.0 1.0 2.5 6.077218287692100226048417176715\[-2\]
4.0 1.0 3.0 5.365400720042544709716176264176\[-2\]
4.0 1.0 3.5 4.791010346652078406517300908585\[-2\]
4.0 1.0 4.0 4.317729831064450749511048756748\[-2\]
4.0 1.0 4.5 3.921185585221614693378573393156\[-2\]
4.0 1.0 5.0 3.584332630993527980351431968712\[-2\]
4.0 1.0 5.5 3.294856745699432037984459599008\[-2\]
: Values of the master integral for selected $w_1,w_2$, and $w_3$[]{data-label="table1"}
Considering recursions, all but one involve denominators limited from below. Only that in Eq. (\[39\]) for increasing $n_4$ has a denominator which can be arbitrarily close to zero. Therefore, if $w_1\approx w_{\rm sing}$, one instead of recursion in Eq. (\[39\]), uses the recursion in Eq. (\[41\]) and calculates $h(w_1,n)$ from Taylor expansion at $w_1=w_{\rm sing}$. All other recursions are calculated directly as in Eqs. (\[07\], \[08\], \[12\], \[20\], \[21\]). They involve two-electron integrals $\Gamma$. The calculation of $\Gamma$ including singular cases has recently been described in detail in Refs. [@kor1] and [@harris], and it does not pose any problem. Finally, several numerical results for three-electron integral involving powers of $r_i$ and $r_{ij}$ are presented in Table II. For comparison with the former results obtained in Ref. [@king_lit3], we have taken the same $n_i$ and $w_i$ as in Table II of this Ref. Our results agree up to the precision achieved in Ref. [@king_lit3] with the one correction. In the fifth position instead of $I(2,1,1,3,3,-2,4.338,4.338,7.384)$, it should be $I(1,1,2,3,3,-2,4.338,4.338,7.384)$.
$w_1$ $w_2$ $w_3$
------------------- ------- ------- ------- ----------------------------------------
$f(-1,2,2,2,2,2)$ 2.700 2.700 2.700 3.622072193238069065841911460566\[-3\]
$f(-1,2,4,1,1,1)$ 2.700 2.900 0.650 2.044941897990188175637070889313\[-1\]
$f(-1,0,2,1,1,1)$ 2.700 2.900 0.650 8.560152684198427372519849562718\[-3\]
$f(-1,0,0,1,1,1)$ 2.700 2.900 0.650 7.695548443927856456193296733495\[-3\]
$f(-1,4,4,3,2,2)$ 7.384 4.338 4.338 2.516457130304929175434829560592\[-6\]
$f(-1,0,0,2,2,2)$ 3.000 2.000 1.000 7.759319533814226728190558692235\[-3\]
$f(-1,0,0,1,3,2)$ 3.000 1.000 2.000 1.528428874506937507531543743291\[-2\]
$f(-1,0,4,2,4,3)$ 2.000 3.000 4.000 6.208037315282433323108011184899\[-3\]
$f(-1,2,2,1,1,1)$ 2.700 2.900 0.650 4.036629272285446411970138933470\[-2\]
$f(-1,2,2,1,1,1)$ 2.500 2.500 0.600 1.025702855657754018359340659240\[-1\]
$f(-1,2,2,0,0,0)$ 2.700 2.900 0.650 3.674068373009625515617159197784\[-2\]
$f(-1,2,2,1,2,3)$ 1.000 1.000 1.000 4.576295463451984514935097879411\[2\]
$f(-1,2,2,2,2,2)$ 1.000 1.000 1.000 5.436536048634697021325813246683\[2\]
$f(-1,2,4,4,3,1)$ 3.000 2.000 1.000 6.126463692215932446059888955061\[0\]
$f(-1,2,4,1,0,0)$ 1.000 1.000 1.000 4.219398540932754336898663066822\[2\]
$f(-1,4,4,1,1,1)$ 2.700 2.900 0.650 3.826213635276544192395399453200\[0\]
$f(-1,4,6,1,1,1)$ 2.700 2.900 0.650 4.206326264336604338380540655410\[1\]
$f(-1,6,6,1,1,1)$ 2.700 2.900 0.650 1.886948258407236970462772961418\[3\]
: Three-electron Hylleraas integral involving $1/r_{12}^2$. $n_i$ and $w_i$ are from Table II of Ref. [@king_lit3]. Function $I$ from this Ref. should be divided by $(4\,\pi)^3$ for comparison with our function $f$.[]{data-label="table2"}
Considering the extended Hylleraas integral with $r_1^{-2}$, we calculate it by numerical integration with respect to $w_1$ $$f(n_1,n_2,n_3,-1,n_5,n_6) = \int_{w_1}^\infty dw_1\,f(n_1,n_2,n_3,0,n_5,n_6) .
\label{42}$$ The recursion relations for $f(n_1,n_2,n_3,n_4,n_5,n_6)$ with nonnegative $n_i$ have been derived previously [@rec], and they seem to be stable enough to perform this integration numerically. Since, we have not found in the literature the method of integration, which is adapted to two weight functions: the constant and the logarithmic one, we use the standard Gauss-Legendre quadrature. In Table III we present several numerical results. It is observed the significant loss of precision due to the presence of $\ln w_1/w_1^n$ at large $w_1$ asymptotic. Therefore, precise integration requires in some cases, the subtraction of these terms.
------------------- ----------------------------
$f(0,0,0,-1,0,0)$ 5.112034507187907543\[-2\]
$f(0,1,0,-1,0,0)$ 1.376985263507039164\[-2\]
$f(0,2,0,-1,0,0)$ 6.942269369095712105\[-3\]
$f(0,3,0,-1,0,0)$ 5.403223451815895118\[-3\]
$f(0,4,0,-1,0,0)$ 5.907661306554417555\[-3\]
$f(0,5,0,-1,0,0)$ 8.587459945883427557\[-3\]
$f(0,6,0,-1,0,0)$ 1.598496287482975980\[-2\]
$f(0,7,0,-1,0,0)$ 3.698745219132481190\[-2\]
$f(0,8,0,-1,0,0)$ 1.036442843454920448\[-1\]
$f(0,9,0,-1,0,0)$ 3.434721508609856189\[-1\]
------------------- ----------------------------
: Three-electron Hylleraas integral involving $1/r_{1}^2$ at $w_1=2, w_2=3, w_3=4$[]{data-label="table3"}
Summary
=======
An analytic approach is presented for the calculation of three-electron Hylleraas integrals involving one inverse quadratic power of inter-particle coordinate. This approach is based on exact recursion relations in powers of coordinates. These recursions involve initial terms and two-electron integrals. For the initial term $f(-1,0,0,0,0,0)$ as a function of $w_1$, one constructs a linear second order differential equation. Its solution is expressed as a one-dimensional integral over dilogarithmic and elliptic function $K$, which can be obtained numerically with arbitrary high precision. The two-electron Hylleraas integrals have already been derived in the literature and they also can be obtained with arbitrary high precision.
These extended Hylleraas integrals are necessary for the calculation of relativistic effects in lithium and light lithium-like ions [@king_rev; @hyl_lit; @lit_iso; @lit_g]. One interesting goal is the high precision calculation of the lithium hyperfine splitting [@lit_hfs], which can serve as a benchmark result for other less accurate methods. Moreover, it has recently become possible to derive nuclear parameters such as the charge radius from the measurement of the isotope shift [@lit_iso]. The hyperfine splitting is sensitive to the distribution of magnetic moment. Therefore the measurement of hfs in various odd isotopes of Li or light lithium-like ions may lead to the determination of the so called magnetic radius, which is very difficult to access experimentally. Even more interesting is the possible extension of this analytic method to beryllium and beryllium-like ions, the 4-electron systems. The use of Hylleraas basis set will allow for a high precision calculation of the wave function and, for example, various transition rates which are of astrophysical relevance.
Acknowledgments
===============
We are grateful to Vladimir Korobov for his source code of the fast multi-precision arithmetics and the dilogarithmic function. We wish to thank Krzysztof Meissner for help in solving differential equation and Frederick King for presenting us his numerical results for some selected values of extended Hylleraas integral and for useful comments. This work was supported by EU grant HPRI-CT-2001-50034.
[99]{} F.W. King, Phys. Rev. A [**44**]{}, 7108 (1991); F.W. King, K.J. Dykema and A.D. Lund, Phys. Rev. A [**46**]{}, 5406 (1992). I. Porras and F.W. King, Phys. Rev. A [**49**]{}, 1637 (1994); P.J. Pelzl and F.W. King, Phys. Rev. E [**57**]{}, 7268 (1998); F.W. King, Adv. At. Mol. Opt. Phys. [**40**]{}, 57 (1999). P.J. Pelzl, G.J. Smethells, and F.W. King, Phys. Rev. E [**65**]{}, 036707 (2002). Z.-C. Yan, G.W.F. Drake, J. Phys. B [**30**]{}, 4723 (1997); Z.-C. Yan, J. Phys. B [**33**]{}, 2437 (2000). F.W. King, J. Mol. Struct. (Theochem) [**400**]{}, 7 (1997); F.W. King, [*et al*]{}, Phys. Rev. A [**58**]{}, 3597 (1998). G.W.F. Drake and Z.-C. Yan, Phys. Rev. A [**46**]{}, 2378 (1992); Z.-C. Yan and G. W. F. Drake, Phys. Rev. A [**52**]{}, 3711 (1995); Phys. Rev. Lett. [**81**]{}, 774 (1998); Phys. Rev. A [**61**]{}, 022504 (2000); Phys. Rev. A [**66**]{}, 042504 (2002); Phys. Rev. Lett. [**91**]{}, 113004 (2003). Z.-C. Yan and G. W. F. Drake, Phys. Rev. A [**66**]{}, 042504 (2002). Z.-C. Yan, J. Phys. B [**35**]{}, 1885 (2002). D. M. Fromm and R. N. Hill, [Phys. Rev. A [**36**]{}, 1013 (1987).]{} E. Remiddi, [Phys. Rev. A [**44**]{}, 5492 (1991).]{} K. Pachucki, M. Puchalski and E. Remiddi, Phys. Rev. A. [**70**]{}, 032502 (2004). K. Pachucki, Phys. Rev. A [*to be published*]{}. F.V. Tkachov, Phys. Lett. B [**100**]{}, 65 (1981); K.G. Chetyrkin and F.V. Tkachov, Nucl. Phys. B [**192**]{}, 159 (1981). W. H. Press, S.A. Teukolsky, W.T. Vetterling and B.F. Flannery, [*Numerical Recipes in FORTRAN 77*]{}, 2nd ed. (Cambridge University Press, Cambridge, 1992). R.A. Sack, C.C.J Roothaan and W. Kołos, J. Math. Phys. [**8**]{}, 1093 (1967). V.I. Korobov, J. Phys. B [**35**]{}, 1959 (2002). F.E Harris, A.M. Frolov and V.S. Smith, Jr., J. Chem. Phys [**121**]{}, 6323 (2004). K. Pachucki, Phys. Rev. A [**66**]{}, 062501 (2002). M. Abramowitz and I. A. Stegun (1974), [ *Handbook of Mathematical Functions*]{} (Dover, New York, 1964).
Integration by parts identities
===============================
The complete set of recursion relations as obtained from integration by parts identities is presented below. Function $G$ is defined in Eq. (\[02\]) and id$(i,j)= 0$ for $i,j=1,2,3$, $$\begin{aligned}
{\rm id}(1,1)&=&
-G(0,1,1,1,1,2) - G(0,1,1,1,2,1) + G(1,0,1,1,1,2) + G(1,1,0,1,2,1)
\nonumber \\ &&
- G(1,1,1,1,1,1) + 2\,G(2,1,1,1,1,1)\,u_1^2 +G(1,1,1,1,2,1)\,(u_1^2 - u_3^2 + w_2^2)
\nonumber \\ &&
+ G(1,1,1,1,1,2)\,(u_1^2 - u_2^2 + w_3^2) ,
\nonumber \\
{\rm id}(2,1)&=&
-G(0,1,1,1,1,2) - G(0,1,1,1,2,1) + G(1,0,1,1,1,2) + G(1,1,0,1,2,1)
\nonumber \\ &&
- G(1,1,1,0,2,1) + G(1,1,1,1,2,0) - G(2,0,1,1,1,1) + G(2,1,1,1,1,0)
\nonumber \\ &&
+ G(1,1,1,1,1,2)\,(u_1^2 - u_2^2 - w_3^2) +
G(2,1,1,1,1,1)\,(u_1^2 + u_2^2 - w_3^2)
\nonumber \\ &&
+ G(1,1,1,1,2,1)\,(u_1^2 - u_3^2+ w_1^2 - w_3^2) ,
\nonumber \\
{\rm id}(3,1)&=&
-G(0,1,1,1,1,2) -G(0,1,1,1,2,1) + G(1,0,1,1,1,2) + G(1,1,0,1,2,1)
\nonumber \\ &&
- G(1,1,1,0,1,2) + G(1,1,1,1,0,2) - G(2,1,0,1,1,1) + G(2,1,1,1,0,1)
\nonumber \\ &&
+ G(1,1,1,1,2,1)\,(u_1^2 - u_3^2 - w_2^2) + G(2,1,1,1,1,1)\,(u_1^2 + u_3^2 - w_2^2)
\nonumber \\ &&
+ G(1,1,1,1,1,2)\,(u_1^2 - u_2^2 + w_1^2 - w_2^2) ,
\nonumber \\
{\rm id}(2,2)&=&
G(0,1,1,1,1,2) - G(1,0,1,1,1,2) - G(1,0,1,2,1,1) + G(1,1,0,2,1,1)
\nonumber \\ &&
- G(1,1,1,1,1,1) + 2\,G(1,2,1,1,1,1)\,u_2^2 + G(1,1,1,2,1,1)\,(u_2^2 - u_3^2 + w_1^2)
\nonumber \\ &&
+ G(1,1,1,1,1,2)\,(-u_1^2 + u_2^2 + w_3^2) ,
\nonumber \\
{\rm id}(1,2)&=&
G(0,1,1,1,1,2) - G(0,2,1,1,1,1) - G(1,0,1,1,1,2) - G(1,0,1,2,1,1)
\nonumber \\ &&
+ G(1,1,0,2,1,1) - G(1,1,1,2,0,1) + G(1,1,1,2,1,0) + G(1,2,1,1,1,0)
\nonumber \\ &&
+ G(1,1,1,1,1,2)\,(-u_1^2 + u_2^2 - w_3^2) + G(1,2,1,1,1,1)\,(u_1^2 + u_2^2 -w_3^2)
\nonumber \\ &&
+ G(1,1,1,2,1,1)\,(u_2^2 - u_3^2 + w_2^2 - w_3^2) ,
\nonumber \\
{\rm id}(3,2)&=&
G(0,1,1,1,1,2) - G(1,0,1,1,1,2) - G(1,0,1,2,1,1) + G(1,1,0,2,1,1)
\nonumber \\ &&
+G(1,1,1,0,1,2) - G(1,1,1,1,0,2) - G(1,2,0,1,1,1) + G(1,2,1,0,1,1)
\nonumber \\ &&
+ G(1,1,1,2,1,1)\,(u_2^2 - u_3^2 - w_1^2) + G(1,2,1,1,1,1)\,(u_2^2 + u_3^2 - w_1^2)
\nonumber \\ &&
+ G(1,1,1,1,1,2)\,(-u_1^2 + u_2^2 - w_1^2 + w_2^2) ,
\nonumber \\
{\rm id}(3,3)&=&
G(0,1,1,1,2,1) + G(1,0,1,2,1,1) - G(1,1,0,1,2,1) - G(1,1,0,2,1,1)
\nonumber \\ &&
- G(1,1,1,1,1,1) + 2\,G(1,1,2,1,1,1)\,u_3^2 + G(1,1,1,2,1,1)\,(-u_2^2 + u_3^2 + w_1^2)
\nonumber \\ &&
+ G(1,1,1,1,2,1)\,(-u_1^2 + u_3^2+ w_2^2) ,
\nonumber \\
{\rm id}(2,3)&=&
G(0,1,1,1,2,1) + G(1,0,1,2,1,1) - G(1,0,2,1,1,1) - G(1,1,0,1,2,1)
\nonumber \\ &&
- G(1,1,0,2,1,1) + G(1,1,1,0,2,1) - G(1,1,1,1,2,0) + G(1,1,2,0,1,1)
\nonumber \\ &&
+ G(1,1,1,2,1,1)\,(-u_2^2 + u_3^2 - w_1^2) + G(1,1,2,1,1,1)\,(u_2^2 + u_3^2 - w_1^2)
\nonumber \\ &&
+ G(1,1,1,1,2,1)\,(-u_1^2 + u_3^2- w_1^2 + w_3^2) , \nonumber \\
{\rm id}(1,3)&=&
G(0,1,1,1,2,1) - G(0,1,2,1,1,1) + G(1,0,1,2,1,1) - G(1,1,0,1,2,1)
\nonumber \\ &&
- G(1,1,0,2,1,1) + G(1,1,1,2,0,1) - G(1,1,1,2,1,0) + G(1,1,2,1,0,1)
\nonumber \\ &&
+ G(1,1,1,1,2,1)\,(-u_1^2 + u_3^2 - w_2^2) + G(1,1,2,1,1,1)\,(u_1^2 + u_3^2 - w_2^2)
\nonumber \\ &&
+ G(1,1,1,2,1,1)\,(-u_2^2 + u_3^2 - w_2^2 + w_3^2) . \nonumber \\
\label{A1}\end{aligned}$$
Two-electron integrals
======================
The two-electron integral $\Gamma$ is defined by $$\Gamma(n_1,n_2,n_3;\alpha_1,\alpha_2,\alpha_3) = \int\frac{d^3\,r_1}{4\,\pi}
\int\frac{d^3\,r_2}{4\,\pi}\,e^{-\alpha_1\,r_1-\alpha_2\,r_2-\alpha_3\,r_{12}}\,
r_{1}^{n_1-1}\,r_{2}^{n_2-1}\,r_{12}^{n_3-1} .\label{B1}$$ In the simplest case of $n_1=n_2=n_3=0$ it is $$\Gamma(0,0,0,;\alpha_1,\alpha_2,\alpha_3) =
\frac{1}{(\alpha_1+\alpha_2)(\alpha_2+\alpha_3)(\alpha_3+\alpha_1)} .\label{B2}$$ In the general case of $n_i\geq0$ $$\Gamma(n_1,n_2,n_3;\alpha_1,\alpha_2,\alpha_3) =
\biggl(-\frac{\rm d}{{\rm d}\alpha_1}\biggr)^{n_1}\,
\biggl(-\frac{\rm d}{{\rm d}\alpha_2}\biggr)^{n_2}\,
\biggl(-\frac{\rm d}{{\rm d}\alpha_3}\biggr)^{n_3}\,
\frac{1}{(\alpha_1+\alpha_2)(\alpha_2+\alpha_3)(\alpha_3+\alpha_1)},\label{B3}$$ and recursions relations for its evaluations have been derived in Refs. [@kolos]. The two-electron integral $\Gamma$ for $n_i<0$ can be obtained by the integration over $\alpha_i$. Typical examples are $$\begin{aligned}
\Gamma(-1,n_2,n_3;\alpha_1,\alpha_2,\alpha_3) &=&
\biggl(-\frac{\rm d}{{\rm d}\alpha_2}\biggr)^{n_2}\,
\biggl(-\frac{\rm d}{{\rm d}\alpha_3}\biggr)^{n_3}\,
\frac{\ln(\alpha_1+\alpha_2)-\ln(\alpha_1+\alpha_3)}
{(\alpha_2-\alpha_3)(\alpha_2+\alpha_3)} , \label{B4}\\
\Gamma(-1,n_2,-1;\alpha_1,\alpha_2,\alpha_3) &=&
\biggl(-\frac{\rm d}{{\rm d}\alpha_2}\biggr)^{n_2}\,
\frac{1}{2\,\alpha_2}\,\bigg[\frac{\pi^2}{6}+\frac{1}{2}\,
\ln^2\biggl(\frac{\alpha_1+\alpha_2}{\alpha_2+\alpha_3}\biggr)
\nonumber \\ &&
+{\rm Li}_2\biggl(1-\frac{\alpha_1+\alpha_3}{\alpha_1+\alpha_2}\biggr)
+{\rm Li}_2\biggl(1-\frac{\alpha_1+\alpha_3}{\alpha_2+\alpha_3}\biggr)
\biggr] .\label{B5}\end{aligned}$$ Other examples together with recursion relations to calculate derivatives may be found in Refs. [@kor1; @harris]. Some three-electron integrals $G$ can be expressed in terms of $\Gamma$. It is when any of argument is equal to zero. Complete list of all cases is: $$\begin{aligned}
G(0,1,1,1,1,1) &=& \Gamma(-1,0,-1;w_2+w_3,w_1,u_2+u_3) , \nonumber\\
G(1,0,1,1,1,1) &=& \Gamma(-1,0,-1;w_1+w_3,w_2,u_1+u_3) , \nonumber\\
G(1,1,0,1,1,1) &=& \Gamma(-1,0,-1;w_1+w_2,w_3,u_1+u_2) , \nonumber\\
G(1,1,1,0,1,1) &=& \Gamma(-1,0,-1;w_2+u_3,u_1,w_3+u_2) , \nonumber\\
G(1,1,1,1,0,1) &=& \Gamma(-1,0,-1;w_1+u_3,u_2;w_3+u_1) , \nonumber\\
G(1,1,1,1,1,0) &=& \Gamma(-1,0,-1;w_1+u_2,u_3,w_2+u_1) . \label{B6}\end{aligned}$$
Special functions
=================
The complete elliptic integral of the first and second kind, $K$ and $E$ respectively, are defined according to [@abram] as $$\begin{aligned}
K(m) &=& \int_0^1 dt\,(1-t^2)^{-1/2}\,(1-m\,t^2)^{-1/2} ,\label{C1}\\
E(m) &=& \int_0^1 dt\,(1-t^2)^{-1/2}\,(1-m\,t^2)^{1/2} .\label{C2}\end{aligned}$$ They are related to a hypergeometric function $$\begin{aligned}
K(m) &=& \frac{\pi}{2}\,_2F_1(1/2,1/2;1;m) ,\label{C3}\\
E(m) &=& \frac{\pi}{2}\,_2F_1(-1/2,1/2;1;m) ,\label{C4}\end{aligned}$$ and fulfill the Legendre’s relation $$E(m)\,K(1-m)+E(1-m)\,K(m)-K(m)\,K(1-m) = \frac{\pi}{2} .\label{C5}$$ Their first derivatives are $$\begin{aligned}
K'(m) &=& \frac{E(m)}{2\,m\,(1-m)} - \frac{K(m)}{2\,m} ,\label{C6}\\
E'(m) &=& \frac{E(m)-K(m)}{2\,m} .\label{C7}\end{aligned}$$ Elliptic functions for $|m| \le 1 $ can be conveniently calculated numerically as described in [@nr]. For $m>1$ one uses the identity [@abram] $$\begin{aligned}
K(m\pm i\,\epsilon) &=& \frac{1}{\sqrt{m}}\,K\biggl(\frac{1}{m}\biggr)
\pm i\,K\bigl(1-m\bigr) , \label{D1} \\
E(m\pm i\,\epsilon) &=&
\sqrt{m}\,E\biggl(\frac{1}{m}\biggr)+\frac{1-m}{\sqrt{m}}\,K\biggl(\frac{1}{m}\biggr)
\pm i\,\bigl[K(1-m)-E(1-m)\bigr] ,\end{aligned}$$ and for $m<-1$ $$\begin{aligned}
K(m) &=&\frac{1}{\sqrt{1-m}}\,K\biggl(\frac{m}{m-1}\biggr) \label{D2}\\
E(m) &=&{\sqrt{1-m}}\,E\biggl(\frac{m}{m-1}\biggr) . \end{aligned}$$ The Laurent expansion near the singularity $m=1$ is $$K(m) =
\sum_{n=0}^\infty\,\biggl[\frac{1}{n!}\,\biggl(\frac{1}{2}\biggr)_n\biggr]^2\,
\biggl[\psi(n+1) - \psi(n+1/2)-
\frac{1}{2}\,\ln(1-m)\biggr]\,(1-m)^n , \label{expansion}$$ where $\psi$ is a logarithmic derivative of the Euler Gamma function.
The dilogarithmic function Li$_2$ is defined by $${\rm Li}_2(z) = -\int_0^z \frac{\ln(1-z)}{z} .\label{C8}$$ Taylor expansion around origin $${\rm Li}_2(z) = \sum_{i=1}^\infty \frac{z^i}{i^2} , \label{C9}$$ is convergent for $|z| \leq 1$. Two useful relations $$\begin{aligned}
{\rm Li}_2(-x) + {\rm Li}_2\bigg(-\frac{1}{x}\biggr) &=&
-\frac{\pi^2}{6}-\frac{\ln^2 x}{2} ,\label{C10}\\
{\rm Li}_2(x) + {\rm Li}_2(1-x) &=&
\frac{\pi^2}{6}-\ln x\,\ln(1-x) ,\label{C11}\end{aligned}$$ are used for simplification of result of integrations in Eqs.(\[10\], \[19\]). Further formulas may be found in [@abram] and an efficient numerical evaluation is described among others in [@kor1].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a new variant of the Växjö interpretation: contextualistic statistical realistic. Basic ideas of the Växjö interpretation-2001 are essentially clarified. We also discuss applications to biology, psychology, sociology, economy,...'
author:
- |
Andrei Khrennikov[^1]\
International Center for Mathematical\
Modeling in Physics and Cognitive Sciences,\
MSI, University of Växjö, S-35195, Sweden\
Email: [email protected]
title: 'Växjö interpretation-2003: realism of contexts'
---
The first version of the [*Växjö interpretation*]{} of quantum mechanics was presented in \[1\], see also \[2\], after the conference “Quantum Theory: Reconsideration of foundations”, Växjö, June-2001, on the basis of numerous exciting discussions with participants. I was really surprised, that in spite of the formal acceptance of the official [*Copenhagen interpretation,*]{} many people (having top-qualification in quantum physics) still have doubts of various kinds and many of them are still looking for a [*realistic interpretation.*]{} This dream about [*quantum realism*]{} was the main stimulus for my attempt to present a new version of the realistic interpretation of quantum mechanics. The main problem was to create such an interpretation in which realism would coexist with a rather strange (people like to say “nonclassical”) behaviour of quantum probabilities – [*Born’s rule and interference of probabilities.*]{}[^2]
In 2000 it was demonstrated \[5\], see also \[6\]-\[10\], that Born’s rule and interference of probabilities can be (quite easily) derived in the realistic framework.[^3] The only thing that should not be neglected is [*contextuality of probabilities*]{} – dependence of probabilities on complexes of physical conditions (physical contexts). My investigations were induced by interest to frequency probability theory, see R. von Mises \[20\], see also \[21\]. In the frequency approach probabilities directly depend on collectives (random sequences) which are associated with concrete complexes of physical conditions. This probability theory is contextual from the very beginning.
Since 2001 I organized a series of conferences on foundations of probability and quantum mechanics [^4] and through intensive discussions (and, in particular, hard critique of some my colleagues, see \[25\], \[26\]) my ideas on interpretations of quantum mechanics are now essentially clearer.
First of all I understood the difference between my contextualism and Bohr’s contextualism, see Remark 1.
Then I understood the difference between my contextualism and contextualism of operational (empirisist) interpretation.
Another new issue is understanding of the role of special [*reference observables*]{} which are used in a concrete model for probabilistic representation contexts (e.g., in classical and quantum physical models we use the [*position*]{} and [*momentum*]{} observables).
Finally, it became clear that, in fact, I discussed \[1\], \[2\] not an interpretation of quantum mechanics, but a model – statistical and contextual – of physical reality. The corresponding interpretation of quantum mechanics is obtained automatically for those models in which statistical data can be represented by complex amplitudes. This is the Växjö interpretation of quantum mechanics: [*contextualistic statistical realistic interpretation.*]{}
By starting not from the formalism of quantum mechanics (calculus of probabilities in a complex Hilbert space), but from the general [*contextual statistical model of reality,*]{} we get the possibility to apply contextual statistical methods in many domains of science: [*biology, psychology, sociology, economy,...*]{} In some special cases we can use even the quantum probabilistic formalism. Such new applications of powerful mathematical methods developed in quantum theory can induce revolutionary changes in many sciences. But for us the quantum formalism is not the starting point. We should start with the general Växjö model of reality (physical, biological, psychological, social,...) and then test statistical data to find an appropriative mathematical formalism. In general there are no reasons to hope to obtain the complex [*quantum-like representation.*]{} For example, there might appear models in which data cannot be represented by complex amplitudes, but by hyperbolic ones. In this case we should use the formalism of [*hyperbolic quantum mechanics,*]{} \[10\]. [^5]
**1. Realism of contexts**
We start with the basic definition:
[**Definition 1.**]{} [*A physical [*context*]{} $C$ is a complex of physical conditions.*]{}
In principle, the notion of context can be considered as a generalization of a widely used notion of [*preparation procedure*]{} \[15\]. I prefer to use contextualistic and not preparation terminology. By using the preparation terminology we presuppose the presence of an experimenter preparing physical systems for a measurement. By using the contextualistic terminology we need not appeal to experimental preparations, experimenter should appear only on the stage of a measurement. Moreover, context need not be macroscopic. Of course, there exist [*experimental contexts*]{} – preparation procedures. However, in general contexts are not coupled to preparation procedures. I consider contexts as [*elements of physical reality*]{} which exist independently of experimenters[^6] This is the cornerstone of my contextualistic viewpoint to physics (quantum as well as classical):
[**Contexts are elements of reality**]{}
To construct a concrete model $M$ of reality, we should fix some set of contexts ${\cal C},$ see definition 2.
[**Remark 1.**]{} (Copenhagen and Växjö contextualisms) [Bohr’s interpretation of quantum mechanics is considered as contextualistic, see \[28\] for detailed analysis. However, we should sharply distinguish two types of contextualism: Copenhagen and Växjö contextualisms. For N. Bohr “context” had the meaning “context of a measurement”. For example, in his answer to the EPR challenge N. Bohr pointed out that position can be determined only in [*context of position measurement.*]{} For me “context” has the meaning a complex of physical conditions. As was underlined, a context is an element of physical reality and it has no direct relation to measurements (or existence of experimenters at all).[^7] For example, there exist contextualistic statistical models which cannot be represented in a complex Hilbert space – so called hyperbolic quantum-like models, \[10\].]{} Moreover, a Bohrian measurement context is always [*macroscopic*]{}, our context – a complex of physical conditions – need not be macroscopic.
**2. Observables**
Suppose that there is fixed a set of observables[^8] ${\cal O}$ such that any observable $a\in {\cal O}$ can be measured under a complex of physical conditions $C$ for any $C\in {\cal C}.$
There can be in principle defined other observables on contexts ${\cal C}$ which do not belong to the system ${\cal O},$ but they will define another contextual model of reality, see definition 2.
We remark that our general Växjö-representation of reality does not contain physical systems, cf. footnote 5. At the moment we do not (and need not) consider observables as observables on physical systems. It is only supposed that if a context $C$ is fixed then for any instant of time $t$ we can perform a measurement of any observable $a \in {\cal O}.$
We do not assume that all these observables can be measured simultaneously; so they need not be compatible. The sets of observables $ {\cal O}$ and contexts ${\cal C}$ are coupled through
[**Axiom 1:**]{} [*For any observable $a \in {\cal O},$ there are well defined contexts $C_\alpha$ corresponding to $\alpha$-filtrations: if we perform a measurement of $a$ under the complex of physical conditions $C_\alpha,$ then we obtain the value $a=\alpha$ with probability 1. It is supposed that the set of contexts ${\cal C}$ contains filtration-contexts $C_\alpha$ for all observables $a\in {\cal O}.$*]{}
**3. Probabilistic representation of contexts**
[**Axiom 2:**]{} [*There are defined contextual probabilities ${\bf P}(a=\alpha/C)$ for any context $C \in {\cal C}$ and any observable $a \in {\it O}.$*]{}
At the moment we do not fix a definition of probability. Depending on a choice of probability theory we can obtain different models. For any $C\in {\cal C},$ there is defined the set of probabilities: $$E({\cal O}, C)= \{ {\bf P}(a=\alpha/C): a \in {\cal O}\}$$ We complete this probabilistic data by $C_\alpha$-contextual probabilities: $$D({\cal O}, C)= \{ {\bf P}(a=\alpha/C),...,
{\bf P}(a=\alpha/C_\beta), {\bf P}(b=\beta/C_\alpha),...: a,b,... \in {\cal O}\}$$ (we remark that $D({\cal O}, C)$ does not contain the simultaneous probability distribution of observables ${\cal O}).$ Data $D({\cal O}, C)$ gives a probabilistic image of the context $C$ through the system of observables ${\cal O}.$ Probabilities ${\bf P}(a=\alpha/C_\beta),...$ play the role of [*structural constants*]{} of a model. We denote by the symbol ${\cal D}({\cal O}, {\cal C})$ the collection of probabilistic data $D({\cal O}, C)$ for all contexts $C\in {\cal C}.$ There is defined the map: $$\label{MP}
\pi :{\cal C} \to {\cal D}({\cal O}, {\cal C}), \; \; \pi(C)= D({\cal O}, C).$$ In general this map is not one-to-one. Thus the $\pi$-image of contextualistic reality is very rough: [*not all contexts can be distinguished with the aid of probabilistic data produced by the class of observables ${\cal O}.$* ]{}
Mathematically such probabilistic data can be represented in various ways. In some special cases it is possible to represent data by complex amplitudes. A complex amplitude (wave function) $\phi\equiv
\phi_{D({\cal O}, C)}$ is constructed by using a formula of total probability with $\cos$-interference term, see \[6\]-\[10\] for extended exposition. In this way we obtain the probabilistic formalism of quantum mechanics. In other cases it is possible to represent data by hyperbolic amplitudes[^9] and we obtain the probabilistic formalism of “hyperbolic quantum mechanics," \[9\], \[10\].
**4. Contextualistic statistical model (Växjö model)**
[**Definition 2.**]{} [*A contextualistic statistical model of reality is a triple $$\label{VM}
M =({\cal C}, {\cal O}, {\cal D}({\cal O}, {\cal C}))$$ where ${\cal C}$ is a set of contexts and ${\cal O}$ is a set of observables which satisfy to axioms 1,2, and ${\cal D}({\cal O}, {\cal C})$ is probabilistic data about contexts ${\cal C}$ obtained with the aid of observables ${\cal O}.$*]{}
We call observables belonging the set ${\cal O}\equiv {\cal O}(M)$ [*reference of observables.*]{} Inside of a model $M$ observables belonging ${\cal O}$ give the only possible references about a context $C\in {\cal C}.$
**5. Realistic interpretation of reference observables**
Our general model can (but, in principle, need not) be completed by some interpretation of reference observables $a\in {\cal O}.$ By the Växjö interpretation reference observables are interpreted as [*properties of contexts:*]{}
“If an observation of $a$ under a complex of physical conditions $C \in {\cal C}$ gives the result $a=\alpha,$ then this value is interpreted as the objective property of the context $C$ (at the moment of the observation).”
As always, a model is not sensitive to interpretation. Therefore, instead of the realistic Växjö interpretation, we might use the Bohrian measurement-contextualistic interpretation, see Remark 1. However, by assuming the reality of contexts it would be natural to assume also the reality of observables which are used for the statistical representation of contexts. Thus we use the realistic interpretation both for contexts and reference observables. This is Växjö realism.
**6. On the role of reference observables**
Reader has already paid attention that reference observables play the special role in our model. I interpret the set ${\cal O}$ as a family of observables which represent some fixed class of properties of contexts belonging ${\cal C}.$ For example, such a family can be chosen by some class of cognitive systems ${\it Z}_{\rm{cogn}}$ – “observers” – which were interested only in the ${\cal O}$-properties of contexts ${\cal C}$ (and in the process of evolution they developed the ability to “feel” these and only these properties of contexts). The latter does not mean that observables ${\cal O}$ are not realistic. I would like just to say that observers $\tau \in {\it Z}_{\rm{cogn}}$ use only observables ${\cal O}.$
We remark again that there can exist other properties of contexts ${\cal C}$ which are not represented by observables ${\cal O}.$ The same set of contexts ${\cal C}$ can be the basis of various models of contextual reality: $M_i= ({\cal C}, {\cal O}_i, {\cal D}({\cal O}_i, {\cal C})), i=1,2,....$ For example, such models can be created by various classes of cognitive systems ${\it Z}_{\rm{cogn},i}.$
Moreover, we may exclude the spiritual element from observables. By considering “observation” as “feeling”of a context $C$ by some system $\tau$ we need not presuppose that $\tau$ is a cognitive system. Such a $\tau$ can be, e.g., a physical system (e.g. an electron) which “feel” a context $C$ (e.g., electromagnetic-context).
[**Remark 2.**]{} (Number of reference observables) In both most important physical models – in classical and quantum models – the set ${\cal O}$ of reference observables consists of [**two observables:**]{} [*position and momentum.*]{} I think that this number “two” of reference observables plays the crucial role (at least in the quantum model).
**7. Växjö model outside physics**
Our contextual statistical realistic models of reality can be used not only in physics, but in any domain of natural and social sciences. Instead of complexes of physical conditions, we can consider complexes of biological, social, economic,... conditions – contexts – as elements of reality. Such elements of reality are represented by probabilistic data obtained with the aid of reference observables (biological, social, economic,...).
In the same way as in physics in some special cases it is possible to encode such data by complex amplitudes. In this way we obtain representations of some biological, social, economic,.... models in complex Hilbert spaces. We call them [*complex quantum-like models.*]{} These models describe the usual $\cos$-interference of probabilities.
Thus, when we speak, e.g., about a quantum-like mental model, this has nothing to do with quantum mechanics for electrons, photons, ... contained in the brain, see \[29\] for detail. A quantum-like mental model is a contextualistic probabilistic model of brain and nothing more, \[\]. There were found (at least preliminary) experimental evidences that in psychology there can be obtained quantum-like (i.e., represented by complex probability amplitudes) statistical data, see \[30\]; such data also can be generated by some games, \[31\] (which have been called “quantum-like games” in \[31\]).
**8. Choice of a probability model**
As was mentioned, any Växjö model $M$ should be combined on some concrete probabilistic model describing probabilistic data ${\cal D}({\cal O}, {\cal C}).$ Of course, the Kolmogorov measure-theoretical model dominates in modern physics. However, this is not the only possible model for probability, see \[21\]. In particular, I strongly support using of the frequency model \[20\], \[21\]. Here we shall use this model to describe probabilistic data. It does not mean that other models which are used in physics cannot be combined with some Växjö models. Of course, such a combination is not straightforward, see \[8\] on the use of the contextual extension of the Kolmogorov model. We now present the frequency probabilistic description of data ${\cal D}({\cal O}, C)$ for some $C \in {\cal C}.$
**9. Frequency description of probability distributions**
By taking into account Remark 2, we consider a set of reference observables ${\cal O}= \{ a, b \}$ consisting of two observables $a$ and $b.$ We denotes the sets of values (“spectra”) of the reference observables by symbols $X_a$ and $X_b,$ respectively.
Let $C$ be some context. In a series of observations of $b$ (which can be infinite in a mathematical model) we obtain a sequence of values of $b:$ $$\label{KOL1}
x\equiv x(b/C) = (x_1, x_2,..., x_N,...), \;\; x_j\in X_b.$$ In a series of observations of $a$ we obtain a sequence of values of $a:$ $$\label{KOL2}
y\equiv y(a/C) = (y_1, y_2,..., y_N,...), \;\; y_j\in X_a.$$ We suppose that the principle of statistical stabilization for relative frequencies holds true and the frequency probabilities are well defined: $$\label{KOL3}
p^b(\beta) \equiv {\bf P}_x( b=\beta)= \lim_{N\to \infty} \nu_N(\beta; x), \;\; \beta \in X_b;$$ $$\label{KOL3a}
p^a(\alpha) \equiv {\bf P}_y( a=\alpha)= \lim_{N\to \infty} \nu_N(\alpha; y), \;\; \alpha\in X_a.$$ Here $\nu_N(\beta; x)$ and $ \nu_N(\alpha; y)$ are frequencies of observations of values $b=\beta$ and $a=\alpha,$ respectively (under the complex of conditions $C).$
Let $C_{\alpha}, \alpha\in X_a,$ be contexts corresponding to $\alpha$-filtrations, see Axiom 1. By observation of $b$ under the context $C_\alpha$ we obtain a sequence: $$\label{KOL4}
x^{\alpha} \equiv x(b/C_\alpha) = (x_1, x_2,..., x_{N},...), \;\; x_j \in X_b.$$ It is also assumed that for sequences of observations $x^{\alpha}, \alpha\in X_a,$ the principle of statistical stabilization for relative frequencies holds true and the frequency probabilities are well defined: $$\label{KOL5}
p^{b/a}(\beta/\alpha) \equiv {\bf P}_{x^{\alpha}}(b=\beta)= \lim_{N \to \infty} \nu_{N}(\beta; x^{\alpha}), \;\;
\beta \in X_b.$$ Here $\nu_N(\beta; x^\alpha), \alpha\in X_a,$ are frequencies of observations of value $b=\beta$ under the complex of conditions $C_\alpha.$ We obtain probability distributions: $$\label{KKK4}
{\bf P}_x(\beta), \;\; {\bf P}_y (\alpha), \;
{\bf P}_{x^{\alpha}}(\beta),\;\;\alpha\in X_a, \beta \in X_b.$$ We can repeat all previous considerations by changing $b/a$-conditioning to $a/b$-conditioning. We consider contexts $C_\beta, \beta \in X_b,$ corresponding to selections with respect to values of the observable $b$ and the corresponding collectives $y^{\beta}\equiv y(a/C_\beta)$ induced by observations of $a$ in contexts $C_\beta.$ There can be defined probabilities $p^{a/b}(\alpha/\beta)\equiv {\bf P}_{y^{\beta}}(\alpha).$ Combining these data with data (\[KKK4\]) we obtain $$D({\cal O}, C)= \{ p^a(\alpha), p^b(\beta), p^{b/a}(\beta/\alpha), p^{a/b}(\alpha/\beta): \alpha\in X_a, \beta \in X_b\}$$
This data gives a statistical contextual image of reality based on reference observables $a$ and $b.$ As was remarked, there exist various mathematical methods for encoding of data $D({\cal O}, C),$ e.g., in some cases by complex amplitudes – complex quantum-like representations.
**10. Representation in a complex Hilbert space**
Let $M$ be a contextualistic statistical model such that ${\cal O}$ contains only two observables $a$ and $b.$ For any context $C\in {\cal C},$ by using statistical data $D(a,b, C)$ we can compute a quantity $\lambda(\beta/\alpha, C), \alpha \in X_a, \beta\in X_b,$ see \[6\]-\[10\]. This quantity was called in \[6\]-\[10\] a [*measure of statistical disturbance*]{} (of the $b$-observable by the $a$-observations under the context $C).$ If $$\label{EN}
\vert \lambda(\beta/\alpha, C)\vert \leq 1$$ for all $\alpha \in X_a, \beta\in X_b,$ then data $D({\cal O}, C)$ can be represented (by using the formula of total probability with interference term) by a complex amplitude $\phi_C$ or in the abstract framework by an element the unit sphere $U_1$ of the complex Hilbert space $H.$ Denote the family of all contexts which satisfy to (\[EN\]) by the symbol ${\cal C}^{\rm{tr}}.$ We have the map: $$\label{M1}
J : {\cal C}^{\rm{tr}} \to U_1$$ We emphasize that $J$ is determined by the reference observables $a$ and $b.$ Thus (\[M1\]) is a Hilbert space representation of contexts determined by these concrete reference observables. The map $J$ is not one to one. Thus by representing contexts by complex amplitudes we lose a lot of information about contexts.
The map (\[M1\]) induces \[8\] a map: $$\label{M2}
L: {\cal O} \to L (H),$$ where $L (H)$ is the set of self-adjoint operators. Probability distributions of operators $\hat{a}= L(a)$ and $\hat{b}= L(b)$ (calculated by using quantum Hilbert space framework) in the state $\phi_C$ coincide with $ p^a(\alpha)$ and $p^b(\beta).$
If for a context $C$ we find that $$\label{EN1}
\vert \lambda(\beta/\alpha, C)\vert \geq 1$$ then $C$ can be represented by a hyperbolic amplitude
**11. Systems, ensemble representation**
We now complete the contextualistic statistical model by considering systems $\omega$ (e.g., physical or cognitive, or social,..) Systems are also [**elements of realty.** ]{} In our model a context $C \in {\cal C}$ is represented by an ensemble $S_C$ of systems which have been interacted with $C.$ For such systems we shall use notation: $$\omega \hookleftarrow C$$ The set of all (e.g., physical or cognitive, or social) systems which are used to represent all contexts $C\in {\cal C}$ is denoted by the symbol $\Omega\equiv \Omega({\cal C}).$ Thus we have a map: $$\label{VMM}
C \to S_C=\{ \omega\in \Omega: \omega \hookleftarrow C \}.$$ This is the ensemble representation of contexts. We set $${\cal S}\equiv {\cal S}({\cal C})=\{S: S=S_C, C \in {\cal C}\}.$$ The ensemble representation of contexts is given by the map (\[VMM\]) $$I: {\cal C} \to {\cal S}$$ Reference observables ${\cal O}$ are now interpreted as observables on systems $\omega\in \Omega.$ In principle, we can interpret values of observables as [*objective properties*]{} of systems. Oppositely to the very common opinion, such models (with realistic observables) can have nontrivial quantum-like representations (in complex and hyperbolic Hilbert spaces) which are based on the formula of total probability with interference terms.
Probabilities are defined as ensemble probabilities, see \[21\].
[**Definition 3.**]{} [*The ensemble representation of a contextualistic statistical model $M =({\cal C}, {\cal O}, {\cal D}({\cal O}, {\cal C}))$ is a triple $$\label{VM1}
S(M) =({\cal S}, {\cal O}, {\cal D}({\cal O}, {\cal C}))$$ where ${\cal S}$ is a set of ensembles representing contexts ${\cal C}$, ${\cal O}$ is a set of observables, and ${\cal D}({\cal O}, {\cal C})$ is probabilistic data about ensembles ${\cal S}$ obtained with the aid of observables ${\cal O}.$*]{}
[**12. Algebraic structure on the set of reference observables**]{}
We do not assume the presence of any algebraic structure on ${\cal O}.$ Even if these observables take values in some set endowed with an algebraic structure, e.g., in ${\bf R},$ we do not assume that this structure induces (in the standard way) the corresponding algebraic structure on ${\cal O}.$ If $a,b \in {\cal O}$ and take values in ${\bf R}$ it does not imply that $d=a+b$ is well defined as observable on every context $C\in {\cal C}.$ In the general contextual approach it is very clear why we cannot do this. If $a$ and $b$ are not compatible, then we cannot measure they simultaneously under a context $C$ at the fixed instant of time and form $d=a+b.$ But a reader may say:
“You use the realistic interpretation of the reference observables in a model $M.$ Thus one can form the sum $d=a+b.$”
a). By the realistic contextualistic interpretation, $a(t)$ and $b(t)$ are objective properties of a context $C$ at the instant of time $t.$ There is defined $d(t)=a(t) + b(t)$.
b). By the realistic interpretation of the model with systems, $a(\omega)$ and $b(\omega)$ are objective properties of a system $\omega.$ There is defined $d(\omega)=a(\omega) + b(\omega)$.
However, this is the ontic or “hidden sum” and the representation (\[M2\]) cannot be extended to such sums. Quantum theory cannot say us anything about $d=a+b$ as pointwise observable. Of course, we can define the sum of operators $\hat{d}= \hat{a}+\hat{b},$ but in general this operator would represent not the ontic observable $d,$ but another observable $d_{\rm{quant}}.$ Observables $d$ and $d_{\rm{quant}}$ can have different probability distributions! (see \[8\]). Nevertheless (and this seems to be crucial in using of quantum theory), averages of these observables coincide: $$\label{SUM}
<d_{\rm{quant}}>\equiv <\hat{d}>= <d>$$ This is a consequence of linearity of both quantum (Hilbert space) and classical probabilistic averages and the coincidence of probability distributions of reference observables and they representatives in the Hilbert space.
**13. Realist and empirisist interpretations of quantum mechanics**
We emphasize again that up to now we have not been considering an interpretation of quantum mechanics. There was proposed a contextualistic statistical model of physical reality. Sometimes this model can be mathematically described by using the formalism of classical mechanics, sometimes quantum, sometimes hyperbolic and so on. However, it is useful to discuss relation of our model to models of physical reality corresponding to various interpretations of quantum mechanics. Here we follow to P. Busch, M. Grabowski, P. J. Lahti \[15\], de Muynck, De Baere, and Martens \[32\] and L. Ballentine \[33\].
[**13.1. Empirisist interpretation.**]{} In this interpretation the formalism of quantum mechanics does not describe reality as such. It only serves to calculate probabilities (relative frequencies) of certain phenomena that can be interpreted as corresponding to the results of a quantum measurement. The probabilities are conditioned on certain procedures, to be interpreted as quantum mechanical preparation procedures. Thus, the wave function or density operator can be interpreted as symbolizing a preparation procedure; in the same way a hermitian operator describes symbolically a quantum mechanical measurement. Wave function and hermitian operator are not thought to correspond to something existing in microscopic reality. They are just labels of (macroscopic) instruments that can be found in the laboratory. QM is thought to describe only (cor)relations of preparation acts and measurement phenomena, It is also important for our further considerations to underline that in an empirisist interpretation of QM the eigenvalues of the hermitian operator do not play a significant role, because these eigenvalues do not correspond to properties of the microscopic object. The empirisist interpretation has achieved great popularity because its antimetaphysical flavor: physics must be about observables only, and about nothing else. Hence, in this interpretation neither the wave function nor the observable must be taken as a property of the microscopic object system.
[**13.2. Realist interpretation.**]{} By this interpretation values of physical observables are considered as objective properties – properties of objects (physical systems).
[**13.3. Växjö interpretation: realism of contexts.**]{} In the Växjö approach quantum mechanics (as a physical theory) is a particular contextualistic statistical model of reality in which the probabilistic data ${\cal D}({\cal O}, {\cal C})$ can be encoded by complex amplitudes. This point of view to quantum formalism induces the Växjö interpretation of quantum mechanics. This is a [*contextualistic statistical realistic interpretation*]{} of quantum mechanics. And Växjö realism is realism of contexts and reference observables.
The Växjö interpretation of quantum mechanics is quite close to the empirisist interpretation. The crucial difference is that by the Växjö interpretation quantum mechanics is about reality – reality of contexts, and not about preparation and measurement procedures. Contexts exist independently of our measurement activity and values of reference observables $a\in {\cal O}$ are objective properties of contexts. The space-scale does not play any role, because the description of reality is purely probabilistic. Quantum probabilistic behaviour is a consequence of complementarity of information for reference observables. Such complementarity of information can take place at microscopic as well as macroscopic scales and, moreover, not only in physics, but in any domain of natural and social sciences.
We remark that by considering context as an element of reality we eliminated the important difference between realist and empirisist interpretations – [*the wave function is considered as a description of the result of preparation rather than as a symbolic representation of the preparation itself.*]{} If a model $M$ has a quantum(-like) complex representation then the wave function represents context – a complex of physical (or biological,...) conditions.
[**13.4. Växjö interpretation: realism of contexts, systems and observables.**]{} Let us now consider the completed Växjö model which contains physical systems, contexts are represented by ensembles of systems. Physical observables are considered as objective properties of systems. As well as for general contextualistic model, quantum mechanics (as a physical theory) is about a rather special class of contexts ${\cal C}^{\rm{tr}}$ such the probabilistic data ${\cal D}({\cal O}, {\cal C}^{\rm{tr}})$ can be encoded by complex amplitudes. The only difference is that probabilities are defined as ensemble probabilities. This interpretation of quantum mechanics is very close to the well known ensemble interpretation which was strongly supported by A. Einstein, see introduction; L. Ballentine called it the statistical interpretation, see \[33\]. A difference is that in our model we start with reality of contexts which can be (but need not be) represented by ensembles.
But this is not the main difference. The main difference is that we did not start at all with an interpretation of one special mathematical formalism, calculus of probabilities in complex Hilbert spaces. We started with a general contextual statistical model of reality and then demonstrated that some special contexts can be represented by quantum-like complex amplitudes. Interpretation of such amplitudes follows automatically from the basic contextual statistical model.
[**References**]{}
1\. A. Yu. Khrennikov, Växjö interpretation of quantum mechanics, http://xxx.lanl.gov/abs/quant-ph/0202107.
2\. A. Yu. Khrennikov, On foundations of quantum theory. Proc. Conf. [*Quantum Theory: Reconsideration of Foundations,*]{} ed. A. Yu. Khrennikov. Ser. Math. Modelling, [**2**]{}, 163-196,Växjö Univ. Press (2002).
3\. A. Fine, [*The Shaky game.*]{} Univ. Chicago Press, Chicago/London (1988).
4\. M. Lockwood, What Schrödinger should have learned from his cat? In [*Erwin Schrödinger: Philosophy and the birth of quantum mechanics,*]{} eds. M. Bitbol, O. Darrigol. Editions Frontieres, Gif-sur Yvette (1992).’
5\. A. Yu. Khrennikov, Origin of quantum probabilities. Proc. Conf. [*Foundations of Probability and Physics.*]{} In: [*Q. Probability and White Noise Analysis*]{}, [**13**]{}, 180-200, WSP, Singapore (2001).
6\. A. Yu. Khrennikov, [*Linear representations of probabilistic transformations induced by context transitions.*]{} [*J. Phys. A: Math. Gen.,*]{} [**34**]{}, 9965-9981 (2001); http://xxx.lanl.gov/abs/quant-ph/0105059
7\. A. Yu. Khrennikov, Contextual viewpoint to quantum stochastics. [*J. Math. Phys.*]{}, [**44**]{}, 2471- 2478 (2003).
8\. A. Yu. Khrennikov, Representation of the Kolmogorov model having all distinguishing features of quantum probabilistic model. [*Phys. Lett. A*]{}, [**316**]{}, 279-296 (2003).
9\. A. Yu. Khrennikov, Hyperbolic quantum mechanics. [*Advances in Applied Clifford Algebras,*]{} [**13**]{}(1), 1-9 (2003).
10\. A. Yu. Khrennikov, Interference of probabilities and number field structure of quantum models. [*Annalen der Physik,*]{} [**12**]{}, 575-585 (2003).
11\. G. W. Mackey, [*Mathematical foundations of quantum mechanics.*]{} W. A. Benjamin INc, New York (1963).
12\. A. Lande, [*Foundations of quantum theory.*]{} Yale Univ. Press (1955).
A. Lande, [*New foundations of quantum mechanics*]{} Cambridge Univ. Press, Cambridge (1968).
13\. G. Ludwig, [*Foundations of quantum mechanics.*]{} Springer, Berlin (1983).
14\. K. Kraus, [*States, effects and operations.*]{} Springer-Verlag, Berlin (1983).
15\. P. Busch, M. Grabowski, P. Lahti, Operational Quantum Physics, Springer Verlag (1995).
16\. S. P. Gudder, [*Axiomatic quantum mechanics and generalized probability theory.*]{} Academic Press, New York (1970).
17\. A. S. Holevo, [*Probabilistic and statistical aspects of quantum theory.*]{} North-Holland, Amsterdam (1982).
A. S. Holevo, [*Statistical structure of quantum theory.*]{} Springer, Berlin-Heidelberg (2001).
18\. P. A. M. Dirac, [*The Principles of Quantum Mechanics.*]{} Oxford Univ. Press (1930).
19\. R. Feynman and A. Hibbs, [*Quantum Mechanics and Path Integrals.*]{} McGraw-Hill, New-York (1965).
20\. R. Von Mises, [*The mathematical theory of probability and statistics.*]{} Academic, London (1964).
21\. A. Yu. Khrennikov, [*Interpretations of Probability.*]{} VSP Int. Sc. Publishers, Utrecht/Tokyo (1999).
22\. A. Yu. Khrennikov, ed., Proc. Int. Conf. [*Quantum Theory: Reconsideration of Foundations.*]{} Ser. Math. Modelling, [**2,**]{} Växjö Univ. Press (2002).
23\. A. Yu. Khrennikov, ed., Proc. Conf. [*Foundations of Probability and Physics-2,*]{} Ser. Math. Modelling, [**5,**]{} Växjö Univ. Press (2002).
24\. A. Yu. Khrennikov, ed., Proc. Int. Conf. [*Quantum Theory: Reconsideration of Foundations-2.*]{} Ser. Math. Modelling, [**10,**]{} Växjö Univ. Press (2004).
25\. C. Fuchs, The anti-Växjö interpretation of quantum mechanics. Proc. Int. Conf. [*Quantum Theory: Reconsideration of Foundations.*]{} ed. A. Yu. Khrennikov, Ser. Math. Modelling, [**2**]{}, 99-116, Växjö Univ. Press, 2002; http://www.msi.vxu.se/forskn/quantum.pdf
26\. A. Plotnitsky, The spirit and the letter of Copenhagen: a response to Andrei Khrennikov, http://xxx.lanl.gov/abs/quant-ph/0206026.
27\. A. Einstein, B. Podolsky, N. Rosen, [*Phys. Rev.*]{}, [**47**]{}, 777–780 (1935).
28\. A. Plotnitsky, Quantum atomicity and quantum information: Bohr, Heisenberg, and quantum mechanics as an information theory, Proc. Conf. [*Quantum theory: reconsideration of foundations,*]{} ed: A. Yu. Khrennikov, Ser. Math. Modelling, [**2**]{}, 309-343, Växjö Univ. Press (2002).
A. Plotnitsky, Reading Bohr: Complementarity, Epistemology, Entanglement, and Decoherence. Proc. NATO Workshop ”Decoherence and its Implications for Quantum Computations”, Eds. A.Gonis and P.Turchi, p.3–37, IOS Press, Amsterdam, 2001.
29\. A. Yu. Khrennikov, On cognitive experiments to test quantum-like behaviour of mind. [*Rep. Växjö Univ.: Math. Nat. Sc. Tech.,*]{} N 7 (2002); http://xxx.lanl.gov/abs/quant-ph/0205092.
A. Yu. Khrennikov, Quantum-like formalism for cognitive measurements. [*Biosystems,*]{} [**70**]{}, 211-233 (2003).
30\. E. Conte, O. Todarello, A. Federici, T. Vitiello, M. Lopane, A. Yu. Khrennikov A Preliminar Evidence of Quantum Like Behavior in Measurements of Mental States. [*Reports from MSI, Växjö Univ.,*]{} N. 03090, 2003;\
http://xxx.lanl.gov/abs/quant-ph/0307201.
31\. A. Grib, A. Khrennikov, K.Starkov, Probability amplitude in quantum like games. [*Reports from MSI, Växjö Univ.,*]{} N. 03088, 2003;\
http://xxx.lanl.gov/abs/quant-ph/0308074.
32\. W. M. de Muynck, W. De Baere, and H. Martens, Interpretations of quantum mechanics, joint measurement of incompatible observables, counterfactual definiteness. [*Found. Physics,*]{} [**24**]{}, N. 12 (1994).
33\. L. E. Ballentine, [*Rev. Mod. Phys.*]{}, [**42**]{}, 358 (1970).
L. E. Ballentine, [*Quantum mechanics.*]{} Englewood Cliffs, New Jersey (1989).
[^1]: Supported in part by the EU Human Potential Programme, contact HPRN–CT–2002–00279 (Network on Quantum Probability and Applications) and Profile Math. Modelling in Physics and Cogn. Sc. of Växjö University.
[^2]: This problem was well known to founders of quantum theory, see, for example, the correspondence between A. Einstein and E. Schrödinger (see \[3\] for English translation and comments, see also \[4\]). E. Schrödinger did not like the Copenhagen interpretation; in particular, he created his cat just to demonstrate absurdness of this interpretation. Unfortunately, people practically forgot about this, see \[4\] for detail. We also recall that Schrödinger’s cat was just a modification of Einstein’s example “involving a charge of gunpowder in a state of unstable chemical equilibrium”, see \[4\] (letter of Einstein to Schrödinger, 8 August 1935, see \[3\], p.78). But neither Einstein nor Schrödinger could combine a realistic ensemble model with quantum statistics. In particular, Schrödinger wrote to Einstein that he will accept the realistic statistical interpretation of quantum mechanics if the interference of probabilities would be explained, see \[4\]. Consequently the Copenhagen interpretation preserved its status of the official quantum ideology until present time.
[^3]: And this is one of advantages of my contextual statistical approach, cf. G. Mackey \[11\], A. Lande \[12\], G. Ludwig \[13\], K. Kraus \[14\], see also P. Busch, M. Grabowski, P. Lahti \[15\], S. Gudder \[16\], and A. Holevo \[17\]. In our approach everything is trivial: complex amplitudes are constructed automatically on the basis of the formula of total probability with interference term, see \[6\]-\[10\]. Moreover, besides the ordinary complex Hilbert space representation, there exists the hyperbolic one and mixed hyper-trigonometric, see \[10\]. I emphasize that the study of former approaches, especially investigations of A. Lande \[12\] and G. Mackey \[11\], were very important for me. However, the starting point were the books of P. Dirac \[18\] and R. Feynman \[19\] in which they payed attention to the mystery of quantum interference of probabilities.
[^4]: See http://www.msi.vxu.se/aktuellt/konferens/index.html and \[22\]–\[24\].
[^5]: In particular, our approach implies that quantum mechanics is not complete.
[^6]: We use the notion “elements of physical reality” in common sense. There is no direct coupling with the EPR sufficient condition for values of physical observables to be elements of physical reality, see \[27\]. Moreover, in general the Växjö model need not contain physical systems. Thus even the formulation of the question: “Can values of observables be considered as objective properties of physical systems?” – is in general meaningless. We shall come to the problem of reality of quantum observables as observables on physical systems (i.e., classical or EPR reality) in section 11. There we shall present the Växjö model completed by physical systems.
[^7]: We remark that so far we do not speak about an interpretation of quantum mechanics. We are presenting an approach to modeling of physical reality. The quantum representation is possible only for some class of models, ${\cal M}_{\rm{quantum}}.$ The class ${\cal M}_{\rm{quantum}}$ is a very special subclass of the class of contextualistic statistical models.
[^8]: We shall denote observables by Latin letters, $a,b,...,$ and their values by Greek letters, $\alpha, \beta,...$
[^9]: Such amplitudes are constructed by using a formula of total probability with $\cosh$-interference term (“hyperbolic interference”), see \[10\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider various possible scenarios to explain the recent observation of a claimed broad H$\alpha$ absorption in our Galactic halo with peak optical depth $\tau \simeq 0.01$ and equivalent width $W \simeq 0.17 \, \rm
\AA$. We show that the absorbed feature cannot arise from the circumgalactic and ISM H$\alpha$ absorption. As the observed absorption feature is quite broad $\Delta\lambda \simeq 30 \, \rm \AA$, we also consider CNO lines that lie close to H$\alpha$ as possible alternatives to explain the feature. We show such lines could also not account for the observed feature. Instead, we suggest that it can arise from diffuse interstellar bands (DIBs) carriers or polyaromatic hydrocarbons (PAHs) absorption. While we identify several such lines close to the H$\alpha$ transition, we are unable to determine the molecule responsible for the observed feature, partly because of selection effects that prevents us from identifying DIBs/PAHs features close to H$\alpha$ using local observations. Deep integration on a few extragalactic sources with high spectral resolution might allow us to distinguish between different possible explanations.
author:
- 'Shiv K. Sethi$^1$, Yuri Shchekinov$^{1,2}$, Biman B. Nath$^1$'
title: 'The mysterious $6565\,\AA$ absorption feature of the Galactic halo'
---
Introduction
============
A recent observation of claimed H$\alpha$ absorption along the lines of sight to SDSS galaxies has put the physical state of hydrogen atoms in the Galactic halo in a spotlight. [@zhang2017] \[ZZ17 hereafter\] stacked the spectra of more than 700,000 galaxies from the Sloan Digital Sky Survey. Focussing on the wavelength range of $6340\hbox{--}6790\, \rm \AA$, they detected a broad H$\alpha$ absorption line with peak optical depth $\tau \simeq 0.01$ and equivalent width $\sim 0.17\, \rm \AA$. The correlation between the features and Galactic longitude is consistent with the absorbing systems being located at rest in our Galactic halo within $L\sim 100$ kpc.
The measured width of the line corresponds to a line of sight velocity dispersion of order $\pm 700 \, \rm km \, s^{-1}$, but there is significant uncertainty in this interpretation, and the velocity spread can be of order $\pm 390 \, \rm km \, s^{-1}$ (ZZ17). The absorption line is consistent with an average column density of hydrogen atoms in $n=2$ state: $N_2\approx(7.34 \pm 0.04) \times 10^{11}$ cm$^{-2}$. The mean absorption map suggests that the absorption is isotropic and is prevalent across most lines of sights.
In the next section, we discuss possible implications of interpreting the observed feature as H$\alpha$ absorption; we discuss direct constraints and possible mechanisms to populate the $2p$ and $2s$ states. In section \[sec:dibpah\], we consider physical processes that might explain the observation while obviating the direct constraints. We consider trapped Ly-$\alpha$ lines, metal lines, and complex molecules (DIB carriers and PAH) as possible candidates.
Maintaining H$\alpha$ absorbing circumgalactic gas {#sec:halpha}
==================================================
The crux of the problem is to maintain a large amount of H$\alpha$ absorbing gas in the halo, which can explain the observed value of column density $N_2$. There are two aspects to consider in order to understand this observation: pumping HI($n=2$) states to explain the observed column density, and constraining emission from the excited states.
The observed absorption could be caused by either $2p$ or $2s$ state. The H$\alpha$ transition could arise from multiple transitions: two transitions originating from the $2p$ level, $2p \hbox{--} 3d$ and $2p \hbox{--} 3s$, or by a $2s\hbox{--}3p$ transition[^1]. ZZ17 reported an absorption trough with peak optical depth $\tau \simeq 0.01$ and equivalent width $W = 0.170 \pm 0.001$Å, yielding a column density $N_2 = 1.13
\times 10^{20} W/(f\lambda^2) \, \rm \AA \, cm^{-2}$ with an oscillator strength $f = 0.6410$. In principle, we could derive multiple column densities depending on oscillator strengths of these transitions but that does not change the conclusions we reach in the paper.
One can readily rule out the $2p$ level if the absorption originates in a medium optically thin to Ly-$\alpha$ photons. The corresponding Ly-$\alpha$ photon energy density in the halo due to radiative de-excitation of $2p$ level would be $N_2 A_{2p,1s} (10.2 \, {\rm eV})/c \approx
10^{11} \, \rm eV \, cm^{-3}$, which is more than a factor $\simeq 10^{12}$ larger than the observed value of UV background in the solar neighbourhood [see, e.g. chapter 12 in @2011piim.book.....D]. Therefore, we consider the $2s$ state. The decay of $2s$ state gives two photons with a broad energy distribution [@sptz]. As the A-coefficient of this decay is nearly 8 orders of magnitude smaller than Ly-$\alpha$ transition it causes a corresponding decrease in the photon energy density. Even this decay gives a photon number density which larger than the observed value by a factor $10^4$, but we consider it as a plausible mechanism to underline the possible physical processes that can populate the first excited state of hydrogen in the ISM.
2s state
--------
As the $2s$ state is metastable with a two-photon decay rate $A_{2s,1s}=8.227$ s$^{-1}$ [@sptz], keeping the excited state sufficiently populated is much easier in this case as compared to the $2p$ state. There are several mechanisms of populating the $2s$ state: radiative recombination of ionized hydrogen to the $2s$ state, collisional excitations by thermal electrons (contributions from protons and other ions is negligible) and radiative excitations of higher energy levels $n\geq 3$ by Lyman-series photons followed by radiative decays to $2s$ state.
The equation of the balance of 2$s$ level is given by: \[pop2s\] n\_[2s]{}=[\_2(T)A\_[2s,1s]{}]{}n\_e\^2+[\_[1s,2s]{}+C\_[1s,2s]{}(T) n\_eA\_[2s,1s]{}]{}n\_[1s]{}, where $C_{1s,2s}(T)$, $\alpha_2(T)$ is the coefficient of collisional excitation [e.g. @janev; @flin; @mclaugh] and recombination to the $2s$ state, respectively[^2]. $\Gamma_{1s,2s}$ is the rate at which Lyman series photons can populate the $2s$ state; $\Gamma_{1s,2s} = \sum_{i>2}\Gamma_{1s,i}P_{i2}$. Here $\Gamma_{1s,i}$ gives the rate at which absorbed photons populate states with $i=n>2$ and $P_{i2}$ is the probability that through radiative decay the atom returns to the $2s$ state, e.g. $P_{32} = A_{32}/(A_{31}+A_{32})$. The absorption of photons by the $3p$ state dominates this process.
### Recombination
This process will dominate if the medium is highly ionized or $n_e \simeq n_{\rm H}$, where $n_H$ is the total hydrogen density. Assuming a photoionized region with $T= 10^4 \, \rm K$, this gives: \[rec\] n\_[2s]{}10\^[-12]{} ([n\_e10 cm\^[-3]{}]{})\^2 [cm\^[-3]{}]{}, To satisfy the observation, $n_{2s} \simeq N_2/L$, we require the size of the region $L\simeq 100 \, \rm kpc$ for $n_e \simeq 10 \, \rm cm^{-3}$. This results in an emission measure (EM): EMA\_[2s,1s]{}N\_2/\_2(T)10\^7 [ cm\^[-6]{} pc]{}. For comparison the Orion nebula has an emission measure $\simeq 5 \times 10^6 \, \rm cm^{-6}\, pc$. For ionized gas of such $EM$ along every line of the sight through the halo, as the observation suggests, the free-free opacity exceeds unity for $\nu \le 1 \, \rm GHz$. This means extragalactic sources would not be observable for radio frequencies below $1 \, \rm GHz$. Thus, we can rule out this physical process as being responsible for populating the $2s$ state. This inference cannot be changed by clumping the gas along the line of sight as both $n_{2s}$ and free-free absorption depend upon the square of $n_e$.
### Collisions
Collisional excitation can contribute to populating the $2s$ state in partially ionized gas at $T \simeq 10^4 \, \rm K$. This gives us: \[popco\] n\_[2s]{} 410\^[-15]{} ([n\_H10 cm\^[-3]{}]{})\^2 [cm\^[-3]{}]{} For computing the electron density we assume the medium to be collisionally ionized for $T = 10^4 \, \rm K$. Using $L \simeq N_2/n_{2s}$, we get the size of the absorbing column, $L\simeq 50 \, \rm Mpc$ for $n_H \simeq 10 \, \rm cm^{-3}$. Clearly this case results in even more unrealistic values of densities and the absorbing column. We have checked the whole range of $T$ and $n_{\rm H}$ and could not come up with a single case that gives reasonable values such as $n_H \simlt 1 \, \rm cm^{-3}$, as observed in the local ISM and in the halo within $L \leq 100 \, \rm kpc$.
### Lyman-series photons
If the claimed H$\alpha$ absorption originates in an extended galactic halo and/or in surrounding IGM gas of the Local group as suggested by (ZZ17), one can assume that absorbing gas is exposed to the extragalactic UV with a photon flux at the Ly-continuum edge $F_\nu^c=10^6~J_{21}$ phot cm$^{-2}$ s$^{-1}$ [see also more recent measurements at $\lambda\approx 2000$ Å in @frances]. We further assume photon flux at $\lambda\approx 1000 \, \rm \AA$ is of the same order of magnitude $F_{1000}\sim F_\nu^c$. We use this value to compute $\Gamma_{1s,2s} \simeq \sigma_{1s,3p}(T) n_\gamma c (A_{32}/(A_{31}+A_{32})$, where $\sigma_{1s,3p}$ is the cross-section of photon absorption from $1s$ to $3p$ level at the line center and $n_\gamma \simeq F_\nu/c$ is the number density of photons that cause the transition[^3]. This gives us: n\_[2s]{} =[\_[1s,2s]{} n\_[1s]{} A\_[2s,1s]{}]{} 510\^[-9]{} ([n\_H10 cm\^[-3]{}]{}) [cm\^[-3]{}]{} For our estimate, we use $T= 5000 \, \rm K$, as expected for the Warm Neutral Medium (WNM) of the ISM and use the absorption cross-section at the line center. In this case, we get $L \simeq 400 \, \rm pc$ for $n_H \simeq 1 \, \rm cm^{-3}$. Clearly this case gives a more believable picture: a single region of a fraction of the size of the galactic halo with hydrogen number density and UV flux expected in the ISM.
There are two related issues which need to be discussed before this estimate becomes reliable. First, the observed line width, which could be due to turbulent motion, is nearly 50 times larger than the thermal line width we assumed here. If we had used the value of absorption line-center cross section suitable for the observed line then the corresponding numbers for $L$ or $n_H$ would be higher by a factor of 50.
Trapped Ly-$\alpha$, metal contamination, DIB, and PAH {#sec:dibpah}
======================================================
There are at least two possible ways to circumvent the constraints in the foregoing. discussion. The first would be to assume that the $2p$ emission originates from a region that is optically thick to Ly-$\alpha$, the trapping of photons in the region reduces the effective decay time of the $2p$ which also diminishes the luminosity in the line. The second approach would be to posit the observed absorption is caused by a transition from the ground state of an element other than hydrogen or it arises from electronic transitions of more complex molecules (e.g. DIB carriers or PAH).
Ly-$\alpha$ from optically thick regions
----------------------------------------
If the Ly-$\alpha$ emission emerges from a region optically thick to this photon, then the coupled problem of solving level populations along with the evolution of the radiative intensity can be greatly simplified under the condition of large line center optical depth. In this case the effective $A$-coefficient is replaced by $A_{2p,1s}/(0.5\tau)$, where $\tau$ is the optical depth at the line center (e.g. see [@2011piim.book.....D], chapter 19; for a more recent discussion see [@2016ApJ...820...10D]). In the case, the occupancy of the $2p$ state increases by $\tau$ and the luminosity in the Ly-$\alpha$ decreases by the same factor. To model this case, we need to ensure that the occupancy of $2p$ far exceeds the occupancy of the $2s$ state and that the luminosity of the line satisfies the constraints on sky brightness. The former can be achieved by exciting the line with collisions, a case already discussed above. Any transition to states $n >2$ yields a Ly-$\alpha$ photon after a few scattering and therefore the occupation of $2s$ state is suppressed with respect to $2p$ state which gets populated by Ly-$\alpha$ photons for which the number of scatterings: $N_{\rm scat} \simeq \tau$ [^4]. To achieve the latter, we need an optical depth $\tau > 10^{10}$, which puts strong constraints on the HI column. This gives us[^5]: $$n_{2p} \simeq { 0.5 C_{1s,2p}(T) n_e n_{1s} \tau \over A_{2p,1s}}$$ As $n_{2p} \ll n_{1s}$, $n_{1s} \simeq n_H$ in this case. $\tau \simeq \sigma_\alpha(0) n_H L$, where $\sigma_\alpha(0)$ is the cross section for Ly-$\alpha$ at the line center; we assume $T = 10^{4} \, \rm K$ . However, $L \simeq N_2/n_{2p}$ to satisfy the observation of H$\alpha$ absorption, which gives: $$\begin{aligned}
n_{2p} & \simeq & \left ({0.5 C_{1s,2s}(T) n_e n_{H}^2 N_2 \sigma_\alpha(0) \over A_{2p,1s}} \right )^{1/2} \nonumber \\
& \simeq & 1.3 \times 10^{-10} \left ({n_H \over 10^2} \right)^{3/2} \, \rm cm^{-3} \end{aligned}$$ For $n_H = 10^2 \, \rm cm^{-3}$, the size of the region, $L \simeq 1.7 \, \rm kpc$ and $\tau \simeq 5\times 10^{10}$ and HI column density, $N_{\rm HI} \simeq 5.2 \times 10^{23} \, \rm cm^{-2}$. These numbers are clearly unrealistic. A change in hydrogen density does not alleviate this problem. Also, the observed velocity width decreases the line-center cross section by a factor of 50 which makes the scenario delineated here even less plausible.
Metal lines
-----------
Given that the width of the reported line is close to 30 Å, it is worthwhile to investigate whether this line could arise from metal lines that lie close to the H$\alpha$ transition. We consider carbon, nitrogen, and oxygen as they are the most abundant elements in the ISM following hydrogen and helium. We neglect the lines arising from transitions between two excited states of the atom to avoid facing the same issues we discussed in Section \[sec:halpha\], e.g., CII line at $\lambda \simeq 6578 \, \AA$ which causes a transition between two excited states.
Two lines of singly ionized nitrogen (NII) — $3P_1$–$1D_2$ ($6549.9 \, \AA$) and $3P_2$–$1D_2$ ($6585.3 \, \AA$) — lie close to H$\alpha$ emission. One distinct advantage in this case is that the absorption is caused, unlike H$\alpha$ absorption, by an element in its ground state, which means there arise no constraints from the decay of the line.
The abundance of nitrogen in the local ISM is $7 \times 10^{-5}$ (e.g. [@2011piim.book.....D]). For simplicity we assume that all the nitrogen is singly ionized. For $T\simeq 10^{4} \, \rm K$, the cross-section at the line center for these two transitions, $\sigma_{\rm NII} \simeq 2 \times 10^{-22} \, \rm cm^2$. To achieve the observed optical depth $\simeq 0.01$, with an average $n_e \simeq 10^{-2}$, we require the absorbing column to be $L \simeq 30~\rm Mpc$, which is far in the access of the halo size of the Milky way. Therefore, this can be ruled out as a plausible mechanism to achieve the observed optical depth.
Diffuse interstellar bands
--------------------------
If the observed features in absorption around the H$\alpha$ line cannot be modelled as either transition of hydrogen or metal then it is conceivable that they arise from more complex molecules. In this section we explore this possibility.
There are a plethora of spectral DIB features. These are presumably caused by carbon chains, PAH and hydrogenated carbons. However, their origin remains highly uncertain. Typical DIBs spectral width FWHM$\sim 0.5\hbox{--}3$Å [@hobbs08], is larger than the Doppler width for interstellar gas [see Chapter 6 in @2010pcim.book.....T], and it might mimic the observed wide Doppler width of the 6565Å feature.
An ongoing ESO survey seeks to detect and characterize DIB for $\lambda=$305–1042 nm with unprecedented spectral resolution and signal-to-noise ratio, along more than 100 sight-lines [@2017arXiv170801429C]. The first results of this survey show a range of spectral features in the wavelength band of interest to us [for results on DIB, see also @tieliau; @2010pcim.book.....T]. It should be pointed out that most of observational data for diffuse interstellar bands (and their templates) in the ISM are obtained from observation of absorption towards nearby stars in the galactic plane. Some of these lines of sights show strong H$\alpha$ absorption features that arise from the stellar atmosphere of the target star, e.g. Figure B1 of [@2017arXiv170801429C].
@Lan15 stacked SDSS stars, galaxies, and QSOs to detect DIBs in the Milky way. Their composite absorption spectra, based on 40000 stellar spectra and obtained after subtracting stellar SEDs, show a discernible absorption feature at H$\alpha$ frequency, which they attribute to stellar absorption residual. Even though 95% of the stars they use have temperatures in the range $4500\hbox{--}7000 \, \rm K$, whose spectra are not expected to show prominent H$\alpha$ absorption, a small level of residual absorption might remain if the SED of the star is corrected for. We note that if the feature they observe (peak optical depth $\tau \simeq 0.002$ and $\Delta\lambda \simeq 10 \, \rm
\AA$) is attributed to absorption by ISM, it might be compatible with the results of (ZZ17). It is because stars compiled by [@Lan15] lie at typical distance of $2\hbox{--}3 \, \rm kpc$ while the lines of sights analysed by (ZZ17) traverse the galactic halo which could be $50 \hbox{--}100 \, \rm kpc$. However, the errors incurred in subtracting stellar SED might be large enough to remove this weak feature expected from ISM.
This discussion shows that there exists an observational selection bias against the detection of DIB close to H$\alpha$ transition. This is supported by observations of emission lines in the Red Rectangle nebula corresponding to DIB absorption lines, from regions excluding the central star, particularly the strong lines at $\approx 6560$ and $6570\,\AA$ [@sarre; @scar]. Further analysis has shown the presence of two relatively strong features at $\lambda=6552.4$Å and $6563.4$Å without contamination from H-$\alpha$ line [@vanwin].
This selection effect is avoided in the observation of absorption from extragalactic sources, e.g. SDSS galaxies, for which intrinsic H$\alpha$ absorption from stellar atmospheres is redshifted, e.g. the work of (ZZ17) or for late type stars without strong H$\alpha$ absorption in their spectra. We should also consider laboratory measurements of the spectra of PAH and complex molecules.
[PAH lines near $\lambda=6565$Å]{}
-----------------------------------
From laboratory measurements of the spectra of complex molecules, we present, as an example, a tentative list of molecular compounds and the corresponding lines what are reasonably close (within $30\hbox{--}40$Å) to the 6565Å and which can in principle mimic H$\alpha$ absorption.
1. [Naphthalene (Np)]{} cation C$_{10}$H$_8^+$: a feature at $\lambda\simeq 6520$Å ($f=10^{-4}$, $\Delta\nu=100$ cm$^{-1}=42$Å), and a weaker feature at $\lambda\simeq 6600$Å [Table III and Fig. 3a in @salama], vibronic transition from the ground state $^2$B$_{3g}$(D$_2$)$\leftarrow$X$^2$A$_u$(D$_0$).
2. A hydrogenated form H$_n$-HC$_{42}$H$_{16}^+$: a feature at 6550Å [($f\approx 0.03$ to $\approx 0.05$)]{}, [see Fig 5d, e,f in @hamm].
3. Protonated pyrene 2H-Py$^+$: a relatively strong feature at 6550Å (theoretical oscillator strengths lie around $f\sim 0.03-0.05$) [Fig 4 and Table 2 in @chin].
4. [\[FePAH\]$^+$ complexes with a band at 6600Å ($f\approx 0.002$), [Table 3 and Fig 2, Fig 3 in @lanza]]{}.
### Optical depth of DIB/PAH absorbing halo gas
Let us assume, for a conservative estimate, the oscillator strength of the line is $f_{6565}\simeq 10^{-3}$, the abundance of the carrying compound $\chi_{6565}\simeq 10^{-8}$ [see, e.g. p. 218 in @2010pcim.book.....T] and the mass of the compound is $\mu_{6565}$. Then the line-center optical depth is: \[pah\] \_[6565]{}810\^[-4]{}\_[6565]{}\^[1/2]{}T\_4\^[-1/2]{} N\_[20]{} Here $N_{20} \equiv N_{HI}/(10^{20} \rm cm^{-2})$.
Similar estimates of typical optical depth for a DIB compound gives, \[taudib\] \_[DIB]{}=[e\^2fmc\_[D]{}]{}\_[DIB]{}N([H]{})0.08\_[DIB]{}\^[1/2]{}T\_4\^[-1/2]{}N\_[20]{}, \[dib\] where the abundance $\chi_{\rm DIB}\simeq 3\times 10^{-10}/f$, with $f$ being the oscillator strength is assumed following @tieliau.
From Eqs. (\[pah\]) and (\[dib\]) we can readily obtain the observed optical depth $\simeq 0.005\hbox{--}0.01$, for an acceptable range of parameters. These estimates are also consistent with constraints from FIR emission [e.g. @2010pcim.book.....T]. The main reason this constitutes a plausible mechanism as compared to metals is that the absorption cross section is larger for PAH/DIB as compared to forbidden lines that lie close the ground state for metals (e.g. NII).
Summary {#sec:sumcon}
=======
In this paper, we attempt to explain a recent observation that report the detection of a broad absorption feature ($\Delta \lambda \simeq 30 \, \rm \AA$) at $\lambda \simeq 6565 \, \rm \AA$ in the galactic halo (ZZ17). Even though the feature corresponds to the H$\alpha$ wavelength, we argue that it could not arise from such a transition.
First, the decay of the excited state through the spontaneous de-excitation of (optically thin) $2p$ and $2s$ levels gives a sky brightness which is incompatible with UV observations. Second, we show that the observed absorption feature cannot be modelled using $2s$ transition for parameters expected of the Galactic halo. We investigate the possibility of Ly-$\alpha$ photons trapped in an optically thick region. In this case, the sky brightness constraint can be overcome only for unrealistically large optical depths.
We next consider CNO metal lines close to H$\alpha$ transition. All transitions that connect two excited states can be ruled out (e.g. CII) for the same reason as listed above. A doublet of singly ionized nitrogen (NII) has transition frequencies within $20 \, \rm \AA$ of H$\alpha$ transition and the transitions connect the ground state with an excited state, thereby obviating the sky brightness constraint. However, even in this case, we fail to find parameters that are compatible with expected properties of ISM and Galactic halo.
Finally, we consider DIB and PAH to explain the observation. There are a number of transitions of such complex molecules in the frequency range of interest. We show that known models of DIB and PAH, based on optical and UV absorption of FIR/NIR emission, might explain the absorption features, even though we are not able to identify the molecule responsible for this line. This could partly be owing to the fact that H$\alpha$ absorption in the local stars used to identify these line could constitute a selection bias in this case.
Deep observation of a few bright extragalactic sources (for a range of galactic longitudes) with high spectral resolution (SDSS spectral resolution $\lambda/\Delta\lambda\sim 1500\hbox{--}2500$ could have caused blending of lines) and signal-to-noise ratio might reveal the nature of this mysterious absorption feature.
We thank an anonymous referee for valuable suggestions. YS acknowledges support from RFBR (project code 15-02-08293).
Chin, C. H., Lin, S. H., 2016, PCCP, 18, 14569
Cox, N., Cami, J., Farhang, A., et al. 2017, arXiv:1708.01429
Dijkstra, M., Sethi, S., & Loeb, A. 2016, , 820, 10
Draine, B. T., Lazarian, A., 1998, ApJ, 494, L19
Draine, B. T. 2011, Physics of the Interstellar and Intergalactic Medium by Bruce T. Draine. Princeton University Press, 2011. ISBN: 978-0-691-12214-4,
Franceschini, A., Rodighiero, G., 2017 [[](http://adsabs.harvard.edu/abs/2017A&A...603A..34)]{} [[](https://doi.org/10.1051/0004-6361/201629684)]{}
Fritsch, W., & Lin, C. D., 1982, Phys. Rev. A 26, 762
Hammonds, M., Pathak, A., Sarre, P. J., 2009, Phys. Chem. Chem. Phys., 11, 4458
Hobbs, L. M., York, D. G., Snow, T. P., 2008, ApJ, 680, 1256 [[](http://adsabs.harvard.edu/abs/2008ApJ...680.1256H)]{} [[](https://doi/10.1086/587930)]{}
Janev, R. K., Langer, W. D., Evans, K., Jr., Post, D. E., Jr. 1987, Elementary Processes in Hydrogen-Helium Plasmas, Springer-Verlag, 321 pp.
Lan, T.-W., Ménard, B., Zhu,. G., 2015, MNRAS, 452, 3629 [[](http://adsabs.harvard.edu/abs/2015MNRAS.452.3629L)]{} [[](https://doi/10.1093/mnras/stv1519)]{}
Lanza, M., Simon, A., Amor, M. B., 2015, JPCA, 119, 6123
McLaughlin, B. M., Winter, T. G., McCann, J. F., 1997, J. Phys. B, 30, 1043
Salama, F., Allamandola, L. J., 1991, J. Chem. Phys., 94, 6964
Sarre, P. J., 1991, Nat., 351, 356
Scarrott, S. M., Watkin, S., Miles, J. R., Sarre, P. J., 1992, MNRAS, 255, 11p
Spitzer, L., Greenstein, J. L., 1951, ApJ, 114, 4075 [[](http://adsabs.harvard.edu/abs/1951ApJ...114..407S)]{}, [[](http://dx.doi.org/10.1086/145480)]{}
Tielens, A. G. G. M. 2010, The Physics and Chemistry of the Interstellar Medium, by A. G. G. M. Tielens, Cambridge, UK: Cambridge University Press, 2010,
Tielens, A. G. G. M., 2013, in: The Diffuse Interstellar Bands, IAU Symp 297, J. Cami & N. L. J. Cox, eds, p. 399
van Winckel, H., Cohen, M., & Gull, T. R., 2002, A&A, 390, 147
Zhang, H, and Zaritsky, D., 2017 [[](http://adsabs.harvard.edu/abs/2017NatAs...1E.103Z)]{}, [[](http://dx.doi.org/10.1038/s41550-017-0103)]{}
[^1]: https://physics.nist.gov/\
PhysRefData/Handbook/Tables/hydrogentable3.htm
[^2]: collisional excitation/recombination to states $n \ge 3$ followed by radiative decay to $2s$ state have negligible contribution.
[^3]: in this case each atom that makes a transition to the $2s$ state is accompanied by H$\alpha$ emission. However this emission is isotropic and the flux of these photons along the line of sight is proportional to the solid angle of the observation (roughly square of an arcsecond) which is negligible.
[^4]: the ratio of the occupancy of $2s$ to $2p$ states scales as $A_{2s, 1s}/(\tau A_{2p,1s})$, see e.g. [@2016ApJ...820...10D]
[^5]: For large optical depth, the level population of the $1s$ and $2p$ states can thermalize or $n_{2p} = 3 n_{1s} \exp(-h\nu_\alpha/(kT))$. The equivalent condition is $n_{2p} C_{2p,1s} \simeq A_{2p,1s}/\tau$ and it is not reached for the range of parameters relevant to the paper.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study collaborative machine learning (ML) across wireless devices, each with its own local dataset. Offloading these datasets to a cloud or an edge server to implement powerful ML solutions is often not feasible due to latency, bandwidth and privacy constraints. Instead, we consider federated edge learning (FEEL), where the devices share local updates on the model parameters rather than their datasets. We consider a heterogeneous cellular network (HCN), where small cell base stations (SBSs) orchestrate FL among the mobile users (MUs) within their cells, and periodically exchange model updates with the macro base station (MBS) for global consensus. We employ gradient sparsification and periodic averaging to increase the communication efficiency of this hierarchical federated learning (FL) framework. We then show using CIFAR-10 dataset that the proposed hierarchical learning solution can significantly reduce the communication latency without sacrificing the model accuracy.'
author:
- 'Mehdi Salehi Heydar Abad and Emre Ozfatura and Deniz Gunduz and Ozgur Ercetin [^1] [^2] [^3]'
bibliography:
- 'IEEEabrv.bib'
- 'ref.bib'
title: Hierarchical Federated Learning Across Heterogeneous Cellular Networks
---
Introduction
============
Vast amount of data is generated today by mobile devices, from smart phones to autonomous vehicles, drones, and various Internet-of-things (IoT) devices, such as wearable sensors, smart meters, and surveillance cameras. Machine learning (ML) is key to exploit these massive datasets to make intelligent inferences and predictions. Most ML solutions are centralized; that is, they assume that all the data collected from numerous devices in a distributed manner is available at a central server, where a powerful model is trained on the data. However, offloading these huge datasets to an edge or cloud server over wireless links is often not feasible due to latency and bandwidth constraints. Moreover, in many applications dataset reveal sensitive personal information about their owners, which adds privacy as another concern against offloading data to a centralized server. A recently proposed alternative approach is federated edge learning (FEEL) [@ML_overair2; @ML_overair3; @ML_overair4; @int_edge1], which enables ML at the network edge without offloading any data.
[*Federated learning (FL)* ]{}is a collaborative ML framework [@fedlearn1; @fedlearn2], where random subsets of devices are selected in an offline manner to update model parameters based on locally available data. Local models are periodically averaged and exchanged among participating devices. This can either be done with the help of a parameter server, which collects the local model updates and shares the updated global model with the devices; or, in a fully distributed manner, where the devices taking part in the collaborative training process seek consensus on the global model through device-to-device communications.
Although the communication bottleneck of FL has been acknowledged in the ML literature, and various communication-efficient distributed learning techniques have been introduced, implementation of these techniques on wireless networks, particularly in heterogeneous cellular networks (HCNs), and the successful orchestration of the large scale learning problem have not been fully addressed. To this end, there are some very recent works that focus on the distributed machine learning framework with a particular focus on wireless communications [@ML_overair1; @ML_overair2; @ML_overair3; @ML_overair4; @ML_overair5; @ML_overair6; @ML_overair7; @ML_overair8; @FL_wireless1; @FL_wireless2]. Most of these works propose new communication-efficient learning strategies, specific to wireless networks, which is called the [*over-the-air aggregation*]{} [@ML_overair1; @ML_overair2; @ML_overair3; @ML_overair4; @ML_overair5; @ML_overair6; @ML_overair7; @ML_overair8]. In this approach, mobile computing nodes are synchronised for concurrent transmission of their local gradient computations or model updates over the wireless channel, and the parameter server receives the noisy version of the gradient sum via utilizing the superposition property of the wireless channel. Although, over-the-air aggregation is a promising solution to mitigate the communication bottleneck in the future communication networks, it imposes stringent synchronization requirements and very accurate power alignment, or, alternatively, the use of a very large number of antennas [@ML_overair8].
In this paper, we focus on FEEL across HCNs, and introduce a communication-efficient hierarchical ML framework. In this framework mobile users (MUs) with local datasets are clustered around small-cell base stations (SBSs) to perform distributed stochastic gradient descent (SGD) with decentralized datasets, and these SBSs communicate with a macro-cell base station (MBS) periodically to seek consensus on the shared model of the corresponding ML problem. In order to further reduce the communication latency of this hierarchical framework, we utilize gradient sparsification, and introduce an optimal resource allocation scheme for synchronous gradient updates.
Distributed hierarchical SGD framework has been recently studied in [@sgd_local4; @sgd_local5], and hierarchical FL is considered in [@fedlearn8]. However, only periodic averaging strategy is employed in these works, and the wireless nature of the communication medium is not taken into account. Our contributions in this paper can be summarized as follows:
- We introduce a hierarchical FL framework for HCNs and provide a holistic approach for the communication latency with a rigorous end-to-end latency analysis.
- We employ communication efficient distributed learning techniques, in particular, sparsification and periodic averaging, jointly, and design a resource allocation strategy to minimize the end-to-end latency.
- Finally, focusing on the distributed image classification problem using popular dataset CIFAR-10, we demonstrate that, with the proposed approach, communication latency in a large scale FEEL framework, implemented over HCNs, can be reduced dramatically without sacrificing the accuracy much.
System Model {#s:System_Model}
============
Consider a HCN with a single MBS and $N$ SBSs. In this cellular setting, $K$ MUs collaborate to jointly solve an optimization problem of the form $$\begin{aligned}
\min_{\mathbf{w}\in \mathbb{R}^Q} \frac{1}{L} \sum^{L}_{i=1} f_i(\mathbf{w}),\end{aligned}$$ where $\mathbf{w}$ is a vector of size $Q\times 1$ denoting the parameters of the model to be learned and $f_i$ is the training loss associated with the $i$th training data sample. We assume that MU $k$ has a training data set $\mathcal{D}^k$ that is not shared with any other MUs due to bandwidth and privacy constraints. Note that calculating the loss over all of the dataset is time consuming and in some cases not feasible since it cannot fit in the memory. Thus, we employ minibatch SGD in which MU $k$ uses a subset $\mathcal{I}^k\in \mathds{D}^k$ of its dataset to calculate the loss. We assume that the batch size $|\mathcal{I}^{k}|=I$ for all $k=1,\ldots,K$. In distributed learning, each MU calculates the loss gradients with respect to its own data set. Then, the gradients are shared with other MUs through either peer-to-peer links, or using a central entity (MBS in this work). The MBS collects the gradients, aggregates them by taking the average, and eventually transmits the average gradient to all the MUs [^4]. Each MU upon receiving the average gradient applies a gradient descend step as follows: $$\mathbf{w}_{t+1} = \mathbf{w}_{t} - \eta_t \left[ \frac{1}{K}\sum^{K}_{k=1} \frac{1}{|\mathcal{I}^{k}|}\sum_{i\in\mathcal{I}^{k}}\nabla f_i(\mathbf{w}_t) \right],$$ where $\mathcal{I}^{k}\subseteq \mathcal{D}^k$ is the mini-batch randomly chosen from the data samples of MU $k$, and $\eta_t$ is the learning rate. The generic federated learning (FL) algorithm is described in Algorithm \[alg:dist\]. The Algorithm \[alg:dist\] is synchronous in the sense that MBS waits until the gradients from all the MUs are received.
Initialize $\mathcal{D}_k$, learning rate $\eta$, $\mathbf{w}_k=\mathbf{w}_0$ Each MU $k$ does the following Randomly select a mini-batch $\mathcal{I}_k\subseteq\mathcal{D}_k$ Calculate $\mathbf{g}_{k,t}=\frac{1}{|\mathcal{I}^{k}|}\sum_{i\in\mathcal{I}^{k}}\nabla f_i(\mathbf{w}_t)$ Transmit $\mathbf{g}_{k,t}$ to MBS \[alg:Tx\] Receive $\mathbf{g}_{t} = \frac{1}{K}\sum_k \mathbf{g}_{k,t}$ \[alg:Rx\] $\mathbf{w}_{t+1} = \mathbf{w}_t - \eta_t \mathbf{g}_{t}$
Note that transmitting local gradients and receiving the average gradients (i.e., lines \[alg:Tx\] and \[alg:Rx\]) introduce latency to the training time specially considering the deep neural networks with tens of millions of parameters. Hence, an efficient communication protocol is required for this purpose considering the synchronous nature of Algorithm \[alg:dist\]. We assume that the bandwith available for communication is $B$ Hz. We employ an orthogonal access scheme with OFDM, and assign distinct sub-carriers to MUs . Denote by $M=\frac{B}{B_0}$ the number of sub-carriers, where $B_0$ is sub-carrier spacing. We denote the channel gain between MU $k$ and MBS on sub-carrier $m$ by $\gamma_{k,m} = |h_{k,m}|^2$, where $h_{k,m} $ is the complex channel coefficient. The distance of MU $k$ to MBS is denoted by $d_k$, and the path loss exponent by $\alpha$.
Uplink Latency
--------------
For the latency analysis, we consider the fixed-rate transmission policy with sub-optimal power allocation introduced in [@goldsmith], which is simple to implement and performs closely to the optimal water-filling power allocation algorithm. The power allocation policy is truncated channel inversion, which only allocates power if the channel gain is above a threshold, otherwise does not use that subcarrier.
Let $p_{k,m}$ denote the power allocated to sub-carrier $m$ by MU $k$ based on the observed channel gain $\gamma_{k,m}$, and let $\mathcal{M}_k$ be the set of uplink (UL) sub-carriers assigned to MU $k$. We should satisfy the average power constraint: $$\begin{aligned}
\mathbb{E}\bigg[ \sum_{m\in\mathcal{M}_k} p_{k,m} \bigg] \leq P_{max},\end{aligned}$$ where the expectation is with respect to the probability density function (pdf) of the channel gain, $f(\gamma_{k,m})$. Since the channel gain is i.i.d over sub-carriers the power constraint becomes $$\begin{aligned}
\mathbb{E}\left[ p_{k,m} \right] \leq \frac{P_{max}}{|\mathcal{M}_k|},\ m\in \mathcal{M}_k . \label{eq:power-const1}\end{aligned}$$ According to the truncated channel inversion policy, the allocated power by MU $k$ on sub-carrier $m\in\mathcal{M}_k$ becomes $$p_{k,m} = \begin{dcases}
\frac{\rho_{k,m}}{\tilde{\gamma}_{k,m}}, & \gamma_{k,m}\geq \gamma^{th}_{k,m}\\
0 & \gamma_{k,m}< \gamma^{th}_{k,m}
\end{dcases},
\label{eq:power-allocation1}$$ where $\rho_{k,m}$ ensures that the power constraint in is met and,
$$\tilde{\gamma}_{k,m}=\frac{\gamma_{k,m}}{N_0 B_0 d_{k}^{\alpha}}$$
is the normalized channel gain and $N_0 B_0$ is the AWGN noise power on a single sub-carrier. The average power constraint in results in [@goldsmith]: $$\begin{aligned}
\rho_{k,m}(\gamma^{th}_{k,m}) = \frac{P_{max}}{|\mathcal{M}_k| N_0 B_0 d_k^{\alpha} {\mkern 1.5mu\overline{\mkern-1.5mu[\frac{1}{\gamma_{k,m}}]\mkern-1.5mu}\mkern 1.5mu}_{\gamma^{th}_{k,m}}},\end{aligned}$$ where $$\begin{aligned}
{\mkern 1.5mu\overline{\mkern-1.5mu\left[\frac{1}{\gamma_{k,m}}\right]\mkern-1.5mu}\mkern 1.5mu}_{\gamma^{th}_{k,m}} \overset{\Delta}{=} \int^{\infty}_{\gamma^{th}_{k,m}} \frac{f(\gamma_{k,m})}{\gamma_{k,m}} d\gamma_{k,m}.\end{aligned}$$ Rather than Shannon capacity, we consider a practical approach where the bits are modulated using $M$-ary QAM. For a given target [*bit error rate*]{} (BER) requirement, the instantaneous UL rate of MU $k$ to the MBS on sub-carrier $m$ becomes [@goldsmith]: $$\begin{aligned}
U_{k,m} = B_0 \log_2\left( 1 + \frac{1.5}{-\ln(5 BER)}\frac{p_{k,m} \gamma_{k,m}}{N_0B_0 d_{k}^{\alpha}}\right). \label{eq: inst-rate1}\end{aligned}$$ By substituting into , the instantaneous rate becomes:
$$\begin{aligned}
U_{k,m} =
B_0 \log_2\left( 1 + \frac{1.5\rho_{k,m}(\gamma^{th}_{k,m})}{-\ln(5 BER)} \right)\mathds{1}_{\gamma_{k,m} \geq \gamma^{th}_{k,m}}, \end{aligned}$$
where $\mathds{1}_{x}=1$ if the argument $x$ is true and $0$ otherwise. For the maximum expected transmission rate, we have
$$\begin{aligned}
\label{eq:ULrate}
\bar{U}_{k,m} = \max_{\gamma^{th}_{k,m}} B_0 \log_2\left( 1 + \frac{1.5\rho_{k,m}(\gamma^{th}_{k,m})}{-\ln(5 BER)}\right)\mathbb{P}(\gamma_{k,m} \geq \gamma^{th}_{k,m})\end{aligned}$$
The average UL rate of MU $k$ for gradient aggregation: $$\begin{aligned}
\bar{U}_k = \sum_{m\in \mathcal{M}_k}\bar{U}_{k,m} \label{eq:UL-Worker-CH}\end{aligned}$$ Each MU uses $\hat{Q}$ bits to quantize each element of its gradient vector. Since the model has $Q$ parameters, each MU needs to send $Q\cdot \hat{Q}$ bits in total to the MBS at each iteration. To minimize the latency of uploading the gradients to the MBS, we should allocate the sub-carriers so that the minimum average UL rate among MUs is maximized. Hence, we perform the following optimization problem:
$$\begin{aligned}
\label{opt-ul}
&\max_{\mathcal{M}_k, \gamma^{th}_{k,m}}\min\left(\sum_{m\in \mathcal{M}_k}\bar{U}_{k,m}\right)^K_{k=1} \nonumber\\
&\text{s.t.} \bigcup^{K}_{k=1} \mathcal{M}_k = \mathcal{M} \end{aligned}$$
For the given solution of , i.e., the optimal sub-carrier allocation, $\left\{\mathcal{M}^{\star}_{k}\right\}^K_{k=1}$, the uplink latency of MU $k$ on average is $$\begin{aligned}
T^{UL}_{k} = \frac{Q\Hat{Q}}{\bar{U}^\star_k},\end{aligned}$$ where $\bar{U}^\star_k = \sum_{m\in \mathcal{M}^{\star}_k}\bar{U}_{k,m}$. Accordingly, the uplink latency in aggregating the gradients of MUs is equal to $$T^{U}=\max_k(T^{U}_{k}).$$
Downlink Latency
----------------
After all the MUs transmit their gradients to the MBS, the average gradient $\mathbf{g} = \sum_{k=1}^{K} \mathbf{g}_k$ is calculated and MBS is required to transmit the average value back to the MUs. However, since the all workers share a common message, we employ a broadcast policy for this case. We assume that the MBS employ a rateless coding scheme that is adapted to the worst instantaneous signal-to-noise ratio (SNR) on each subcarrier. We assume that the MBS allcoates its available power uniformely over all subcarriers. Specifically, let $SNR_{k,m}(t)$ be the SNR of worker $k$ on subcarrier $m$. Then, the instantaneous broadcast rate on subcarrier $m$ becomes:
$$\begin{aligned}
R_m(t) = \min_{k} B_0 \log_2( 1 + SNR_{k,m}(t)),\end{aligned}$$
where, $$\begin{aligned}
SNR_{k,m}(t) = \frac{P_{max} \gamma_{k,m}(t)}{M N_0 B_0 d^{\alpha}_k}\end{aligned}$$
The broadcast will end when all $Q\cdot\hat{Q}$ parameters are received by the workers. The broadcast latency, $T^{DL}$, can be computed as follows: $$\begin{aligned}
T^{DL} = \mathds{E}\min \{t | T_s\sum^{t}_{\tau=1}\sum^{M}_{m=1}R_{m}(\tau)\geq Q\hat{Q} \} ,\end{aligned}$$ where the expectation is with respect to the PDF of the channel gain. Per iteration, the end-to-end latency of the FL protocol is given by $T^{FL} = T^{UL} + T^{DL}$
Sub-carier Allocation Policy
----------------------------
Initialize (number of sub-carriers $M^{U}$, number of MUs $K$) Initialize $M^{U}_k=1$ for all $k$\[alg:sub\_allocL2\] optimize $\gamma^{th}_{k,m}$ in $k^{\star} = \arg\min \bar{U}_{k}$ $M_{k^{\star}}\leftarrow M_{k^{\star}} + 1$
The optimal sub-carrier allocation problem in is presented in Algorithm \[alg:sub\_alloc\]. It starts by assigning a single sub-carrier for each MU. Then with a single sub-carrier, each MU $k$ optimizes the threshold in . Then the algorithm looks for a MU with minimum average UL rate, i.e., $k^{\star} = \arg\min \bar{U}_{k}$, and allocates a single carrier to that MU. Then MU $k^{\star}$ updates its threshold and $\bar{U}^\star_k$ value. This procedure continues until all available sub-carriers are allocated. The following theorem establishes the optimality of the proposed policy.
The sub-carrier allocation policy in Algorithm \[alg:sub\_alloc\] is optimal.
The proof is by induction. Let the number of sub-carriers be $M^{U}= K + 1$. Then, a single sub-carrier is allocated to each MU first (line \[alg:sub\_allocL2\]), since otherwise there will be a single MU with a rate of zero. The optimal choice is to allocate the remaining single sub-carrier to $k^{\star} = \arg\min \bar{U}_{k}$. To see why, let $\bar{U}_{k^{\star}(m_k)}$ denote the rate of MU $k^{\star}$ when $m_k$ sub-carriers are allocated. Now consider an alternative policy that allocates the remaining sub-carrier to a different MU $k\neq k^{\star}$. Denote the rates achieved under this alternative policy by $\bar{U}^{\ast}_{k}$. It is obvious that $\bar{U}_{k^\star}(2)>\bar{U}_{k^\star}(1) = \min_{k} \bar{U}^{\ast}_{k}$. Thus, the optimal policy allocates the remaining sub-carrier to MU $k^{\star}$. Now, assume that Algorithm \[alg:sub\_alloc\] allocates $M^{U}=K + m$ sub-carrier optimally. We need to prove the optimality of the algorithm for $M^{U}=K + m + 1$. Consider that $K + m$ sub-carriers are allocated and we need to allocate the last sub-carrier. The last sub-carrier is allocated to $k^{\star}$ so that number of its sub-carrier becomes $m_k+1$. The alternative policy allocates the last sub-carrier to another MU $k\neq k^{\star}$. It is clear that $\bar{U}_{k^\star}(m_k+1)>\bar{U}_{k^\star}(m_k) = \min_{k} \bar{U}^{\ast}_{k}$. Hence, the alternative policy is sub-optimal. This concludes the proof.
Distributed Hierarchical Federated Learning
===========================================
In centralized FL [@fedlearn6], MUs transmit their computation results (local gradient estimates) to the parameter server (MBS) for aggregation at each iteration. However, in large scale networks, this centralized framework may result in high communication latency and thus increases the convergence time. To this end, we introduce [*hierarchical federated learning*]{}, in which multiple parameter servers are employed to reduce the communication latency. In the proposed hierarchical framework, MUs are clustered according to their locations. In each cluster a small cell base station (SBS) is tasked with being the parameter server. At each iteration, MUs in a cluster send their local gradient estimates to the SBS for aggregation, instead of the MBS. Then, the SBSs compute the average gradient estimates and transmit the results back to their associated MUs to update their model accordingly.
In this framework, gradient communication is limited to clusters, particularly between the MUs and the corresponding SBSs. This not only reduces the communication distance, communication latency, but also allows the spatial reuse of the available communication resources. On the other hand, limiting the gradient communications within clusters may prevent convergence to a single parameter model (i.e., global consensus).
To this end, we combine aforementioned intra-cluster gradient aggregation method with inter-cluster model averaging strategy, such that after every $H$ consecutive intra-cluster iterations, SBSs send their local model updates to the MBS to establish a global consensus. The overall procedure is illustrated in Figure \[fig:HFL\].
[1]{} {width=".8\linewidth"}
[1]{} {width=".8\linewidth"}
Denote by $\mathcal{C}_n$ the set of MUs belonging to cluster $n=1,\ldots,N$, with $N$ being the number of clusters. During each consecutive $H$ intra-cluster iterations, the local gradient estimates of the MUs are aggregated within the clusters. For example, at iteration $t$ each MU $k$, $k\in\mathcal{C}_n$ for $n=1,\ldots,N$ computes the local gradient estimate, denoted by $\mathbf{g}_{n,k,t}=\frac{1}{|\mathcal{I}^{k}|}\sum_{i\in\mathcal{I}^{k}}\nabla f_i(\mathbf{w}_{n,t})$, and transmits it to the SBS in cluster $n$. Then, the SBS $n$ aggregates the gradients simply by taking the average, $$\begin{aligned}
\mathbf{g}_{n,t} = \frac{ \sum_{k\in\mathcal{C}_n} \mathbf{g}_{n,k,t}}{|\mathcal{C}_n|}. \label{eq:avg-grad-cluster}\end{aligned}$$ This average is then sent back by the SBS to the MUs in its cluster, and the model at cluster $n=1,\ldots,N$ is updated as $$\begin{aligned}
\mathbf{w}_{n, t+1} = \mathbf{w}_{n,t} - \eta_t \mathbf{g}_{n,t}\end{aligned}$$ After $H$ iterations, all SBSs transmit their models to the MBS through UL fronthaul links. The MBS calculates the model average $\mathbf{w} = \sum^N_{n=1}\frac{\mathbf{w}_{n}}{N}$, and transmits it to the SBSs over the DL fronthaul links. Upon receiving the model update, the SBSs share it with the MUs in their cluster. Hence, after $H$ iterations all the MUs share a common model parameters, globally. The HFL algorithm is presented in Algorithm \[alg:dist-hier\].
Initialize $\mathcal{D}_k$, learning rate $\eta$, $\mathbf{w}_k=\mathbf{w}_0$, $H$ Each MU $k$ does the following Randomly select a mini-batch $\mathcal{I}_k\subseteq\mathcal{D}_k$ Calculate $\mathbf{g}_{k,t}=\frac{1}{|\mathcal{I}^{k}|}\sum_{i\in\mathcal{I}^{k}}\nabla f_i(\mathbf{w}_t)$ Transmit $\mathbf{g}_{k,t}$ to SBS Receive $\mathbf{g}_{n,t} = \frac{1}{K}\sum_k \mathbf{g}_{k,t}$ \[alg:Rx\] $\mathbf{w}_{n, t+1} = \mathbf{w}_{n,t} - \eta_t \mathbf{g}_{n,t}$ $n$-th SBS, $n=1,\ldots,N$ send $\mathbf{w}_n$ to MBS $\mathbf{w}\leftarrow \sum^{N}_{n=1}\frac{\mathbf{w}_n}{N}$ MBS transmit $\mathbf{w}$ to all SBSs SBSs transmit $\mathbf{w}$ to their MUs $\mathbf{w}_n=\mathbf{w}$
Communication Latency analysis
------------------------------
In the hierarchical scheme, after clustering the MUs, clusters are colored so that any two clusters with the same color are separated by at least distance $D_{th}$ to minimize interference between clusters. For simplicity, we assume that there is zero interference on receivers located beyond $D_{th}$. If $N_{c}$ colors are used in total, the available OFDM sub-carriers are divided into $N_{c}$ groups, and the sub-carriers in each group are allocated to clusters with a particular color. Consequently, in each cluster the number of available OFDM sub-carriers is proportional to $1/N_{c}$. Before the delay analysis, we have the following assumptions regarding the location of the MUs.
The MUs are uniformly distributed and each cluster contains equal number of MUs.
The SBSs are located at the origin of the corresponding clusters.
In the local gradient update step of HFL (see Figure \[fig:HFLl\]) communication latency analysis is similar to the FL. The only difference is the number of sub-carriers inside the clusters which is $M/N_{c}$. Moreover, the MUs transmit to the SBSs instead of MBS. Denote by $\bar{U}^{*n}_{k}$ the maximum average UL rate of MU $k\in\mathcal{C}_n$. The UL latency of gradient aggregation in cluster $n$ is denoted by $\Gamma^{U}_{n}=\max_k \frac{Q\hat{Q}}{\bar{U}^{*n}_{k}}$, $\forall k\in \mathcal{C}_n$. Similarly, let $\bar{R}^{*n}_{k}$ the maximum average DL rate of MU $k\in\mathcal{C}_n$. The DL latency of gradient aggregation in cluster $n$ is denoted by $\Gamma^{D}_{n}=\max_k \frac{Q\hat{Q}}{\bar{R}^{*n}_{k}}$, $\forall k\in \mathcal{C}_n$.
After $H$ iterations, SBSs send the model to the MBS for the purpose of averaging the clusters model. Let $U^{SBS}$, $R^{SBS}$ be the UL, DL rate of SBSs to the MBS, respectively. The UL, DL latency at each period of $H$ iterations become $\Theta^{U}=\frac{Q\hat{Q}}{U^{SBS}}$ and $\Theta^{D}=\frac{Q\hat{Q}}{R^{SBS}}$, respectively. There is also the latency of transmitting the average model by SBSs to their associated MUs. The average latency associated with a period of hierarchical distributed SGD becomes.
$$\begin{aligned}
\Gamma^{period} = \max_{n\in \mathcal{N} }\left(\sum^{H}_{i=1} \Gamma^{U}_{n}(i)+\Gamma^{D}_{n}(i)\right)+\Theta^{U}+\Theta^{D} + \max_n(\Gamma^D_n),\end{aligned}$$
where $T^{U}_{n}(i)$ and $T^{D}_{n}(i)$ is the latency of $i$-th iteration of UL and DL aggregation in cluster $n$, respectively. The average per iteration latency of HFL becomes $\Gamma^{HFL} = \frac{\Gamma^{period}}{H}$.
Sparse Communications
=====================
Initialize $\mathcal{D}_k$, learning rate $\eta$, $\mathbf{w}_k=\mathbf{w}_0$, sparse factor $\phi^{ul}_{MU}$ Each MU $k$ does the following Randomly select a mini-batch $\mathcal{I}_k\subseteq\mathcal{D}_k$ Calculate $\mathbf{g}_{k,t}=\frac{1}{|\mathcal{I}^{k}|}\sum_{i\in\mathcal{I}^{k}}\nabla f_i(\mathbf{w}_t)$ $\mathbf{u}_{k,t} = \sigma \mathbf{u}_{k,t-1} + \mathbf{g}_{k,t}$ $\mathbf{v}_{k,t} = \mathbf{v}_{k,t-1} + \mathbf{u}_{k,t}$ $g_{th} \leftarrow \phi^{ul}_{MU}$ of $|\mathbf{v}_{k,t}|$ $mask \leftarrow |\mathbf{v}_{k,t}|\geq g_{th}$ $\hat{\mathbf{g}}_{k,t} = \mathbf{v}_{k,t} \odot mask$ $\mathbf{u}_{k,t} = \mathbf{u}_{k,t} \odot \neg mask$ $\mathbf{v}_{k,t} = \mathbf{v}_{k,t} \odot \neg mask$ send $\hat{\mathbf{g}}_{k,t}$ to MBS receive $\mathbf{g}_{t} = \sum^{K}_{k=1} \hat{\mathbf{g}}_{k,t}$ $\mathbf{w}_{t+1} = \mathbf{w}_t - \eta_t \mathbf{g}_{t}$
The trend of going deeper in the *depth* of the neural networks for increasing the accuracy have resulted in NNs with tens of millions of parameters. The amount of data required to be communicated is challenging even for cable connections let alone the wireless links. On top of performing periodical parameter averaging, sparse communication can be used to significantly improve the latency. In sparsification the fact that the gradient vector is sparse is used to transmit only a fraction (i.e., $1-\phi$ ) of parameters and considerably reduce latency.
To make sure that all gradients are transmitted eventually, a separate parameter vector, $\mathbf{v}$ is used to accumulate the error associated with the gradients that are not transmitted. The gradients that are not transmitted will grow in time, as recorded by $\mathbf{v}$, and eventually will be transmitted. More specifically the error buffer is calculated as $$\begin{aligned}
\mathbf{v}_{k,t} = \mathbf{v}_{k-1,t} + \mathbf{g}_{k,t}\end{aligned}$$ Now each MU $k$, instead of transmitting $\mathbf{g}_{k,t}$ transmits $sparse(\mathbf{v}_{k,t})$ to be aggregated at the MBS (or SBSs in the clusters).
Note that the vanilla SGD used an Algorithm \[alg:dist\] and \[alg:dist-hier\] is the simplest form of a optimizer and its performance is quite poor in large scale optimization problems. An efficient way of accelerating the performance of vanilla SGD is to apply momentum method. In momentum method the parameters are updated as following: $$\begin{aligned}
\label{eq:momsgd}
\mathbf{u}_t = \sigma \mathbf{u}_{t-1} + \mathbf{g}_t \nonumber\\
\mathbf{w}_{t+1} = \mathbf{w}_{t} - \eta_t \mathbf{u}_t,\end{aligned}$$ where $\sigma$ is the momentum and $\mathbf{g}$ is the aggregated gradient.
Directly applying momentum to the sparsed gradients will result in a poor performance and momentum correction is required. Here, we directly employ the method preoposed in [@SGD_sparse1]
$$\begin{aligned}
&\mathbf{u}_{k,t} = \sigma \mathbf{u}_{k,t-1} + \mathbf{g}_{k,t}\\
&\mathbf{v}_{k,t} = \mathbf{v}_{k-1,t} + \mathbf{u}_{k,t}\\
&\mathbf{w}_{t+1} = \mathbf{w}_{t} - \frac{\eta}{K}\sum^{K}_{k=1}sparse(\mathbf{v}_{k,t})\end{aligned}$$
Sparsification delays transmitting gradients that are too small. When they are finally transmitted they become stale and slow down the convergence. To combat the staleness [@SGD_sparse1], we apply the inverted sparsification to both accumulated gradients and momentum factor as follows: $$\begin{aligned}
&mask \leftarrow |\mathbf{v}_{k,t}|\geq g_{th}\\
&\mathbf{u}_{k,t} = \mathbf{u}_{k,t} \odot \neg mask\\
&\mathbf{v}_{k,t} = \mathbf{v}_{k,t} \odot \neg mask\end{aligned}$$ The mask simply prevents the stale momentums to be applied. The detailed algorithms for sparse federated SGD is represented in Algorithm \[alg:noc\].
Parameter Definition
--------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------
$\phi^{ul}_{MU}$, $\phi^{dl}_{SBS}$, $\phi^{ul}_{SBS}$, $\phi^{dl}_{MBS}$ Sparsification parameters for uplink from MU to SBS, downlink from SBS to MU, uplink from SBS to MBS and downlink from MBS to SBS.
$\mathbf{e}$, $\mathbf{e}_{n}$, $\mathbf{\epsilon}_{n}$ Model errors due to sparsification before downlink from MBS to SBSs, downlink from $SBS_{n}$ to MU and uplink from $SBS_{n}$ to MBS respectively.
$\mathbf{w}_{k}$, $\mathbf{W}_{n}$ and $\mathbf{W}$ Parameter model at $k$th MU, $n$th SBS and MBS respectively.
$\Delta_{\mathbf{W}}$, $\Delta_{\mathbf{W}_n}$ and $\delta\mathbf{w}_{n}$ Model difference send to SBSs from MBS, to MBS from $SBS_{n}$ and to MUs from $SBS_{N}$ respectively.
: Parameters
Sparse Communication and Error Accumulation
-------------------------------------------
Our proposed HFL framework consists of 4 communication steps: uplink from MU to SBS, downlink from SBS to MU, uplink from SBS to MBS and downlink from MBS to SBS. For each communication step, we employ different sparsification parameters, $\phi^{ul}_{MU}$, $\phi^{dl}_{SBS}$, $\phi^{ul}_{SBS}$ and $\phi^{dl}_{MBS}$ respectively, to speed up the communication. We introduce the function $\Omega(\mathbf{V},\phi):\mathcal{R}^d \rightarrow \mathcal{R}^d$, which maps a $d$ dimensional vector to its sparse form where only $1-\phi$ portion of the indicies have non-zero values.\
The sparsification procedure in each step leads to an error in the parameter model and thus slows down the convergence. To overcome this issue we employ the discounted error accumulation technique, similar to [@SGD_q4; @fedlearn7], which uses the discounted version of the error for the next model update. Before the details of the error accumulation strategy, we want to introduce the following parameters $\mathbf{w}_{n}$ and $\mathbf{W}$ which denotes the parameter model at $n$th SBS and MBS respectively. We note that to employ sparsification for model averaging it is more convenient to transmit the model difference rather than the model. To this end, we introduce the reference models $\tilde{\mathbf{W}}_{n}$ and $\tilde{\mathbf{W}}$ for SBSs and MBS respectively so that each SBS sends the model difference based on $\tilde{\mathbf{W}}$ to the the MBS and based on $\tilde{\mathbf{W}}_{n}$ to the corresponding MUs. In the proposed HFL framework, we use $\mathbf{e}$, $\mathbf{e}_{n}$, and $\mathbf{\epsilon}_{n}$ to keep the model errors due to sparsification before downlink from MBS to SBSs, downlink from $SBS_{n}$ to MU and uplink from $SBS_{n}$ to MBS respectively. The overall procedure for the HFL framework is illustrated in Algorithm \[alg:c\], where error accumulation strategy is employed at lines 21, 28, 34 and parameters $\beta_{m}$ and $\beta_{s}$ are the discount factors for the error accumulation.
Initialize $\mathbf{W}$ and $\tilde{\mathbf{W}}$ Initialize $\mathbf{W}_{n}$ and $\tilde{\mathbf{W}_{n}}$ Initialize $\mathbf{w}_{k}$ **Computation and Uplink**: Randomly select a mini-batch $\mathcal{I}_k\subseteq\mathcal{D}_k$ Calculate $\mathbf{g}_{k,t}=\frac{1}{|\mathcal{I}^{k}|}\sum_{i\in\mathcal{I}^{k}}\nabla f_i(\mathbf{w}_t)$ $\mathbf{u}_{k,t} = \sigma \mathbf{u}_{k,t-1} + \mathbf{g}_{k,t}$ $\mathbf{v}_{k,t} = \mathbf{v}_{k,t-1} + \mathbf{u}_{k,t}$ $g_{th} \leftarrow \phi^{ul}_{MU}$ of $|\mathbf{v}_{k,t}|$ $mask \leftarrow |\mathbf{v}_{k,t}|\geq g_{th}$ $\hat{\mathbf{g}}_{k,t} = \mathbf{v}_{k,t} \odot mask$ $\mathbf{u}_{k,t} = \mathbf{u}_{k,t} \odot \neg mask$ $\mathbf{v}_{k,t} = \mathbf{v}_{k,t} \odot \neg mask$ send $\hat{\mathbf{g}}_{k,t}$ to associated $SBS$ **Model Average**: Update $\mathbf{W}_n(t+1) = \tilde{\mathbf{W}}_n(t) -\eta \hat{\mathbf{g}}_{n} + \beta_{s} \mathbf{\epsilon}_{n}(t)$ $\Delta_{\mathbf{W}_n}(t+1) = \mathbf{W}_n(t+1)- \tilde{\mathbf{W}}(h)$ send $\Omega(\Delta_{\mathbf{W}_n}(t+1),\phi^{ul}_{SBS})$ to MBS $\mathbf{e}_n(t+1)=$ $\Delta_{\mathbf{W}_n}(t+1) -\Omega(\Delta_{\mathbf{W}_n}(t+1),\phi^{ul}_{SBS})$ $\Delta_{\mathbf{W}} = \sum_{n} \Omega(\Delta_{\mathbf{W}_n}(t+1),\phi^{ul}_{SBS})+\beta_{m}e$ MBS transmit $\Omega(\Delta_{\mathbf{w}},\phi^{dl}_{MBS})$ to all SBSs $\mathbf{e}= \Delta_\mathbf{W}-\Omega(\Delta_{\mathbf{W}},\phi^{dl}_{MBS})$ $\tilde{\mathbf{W}}(h+1) = \tilde{\mathbf{W}}(h) + \Omega(\Delta_{\mathbf{W}},\phi^{dl}_{MBS})$ $\mathbf{W}_n(t+1) =$ $\tilde{\mathbf{W}}(h) + \Omega(\Delta_\mathbf{W},\phi^{dl}_{MBS})+\mathbf{e}_n(t+1)/N$ $\delta_{\mathbf{W}_n}(t+1)=\mathbf{W}_n(t+1)-\tilde{W}_n(t)$ $SBS_{n}$ sends $\Omega(\delta_{\mathbf{W}_n}(t+1),\phi^{dl}_{SBS})$ to MUs $\tilde{\mathbf{W}}_n(t+1)=\tilde{\mathbf{W}}_n(t)+ \Omega(\delta_{\mathbf{W}_n}(t+1),\phi^{dl}_{SBS})$ $\mathbf{\epsilon}_{n}(t+1)= \delta_{\mathbf{W}_n}(t+1)-\Omega(\delta_{\mathbf{W}_n}(t+1),\phi^{dl}_{SBS})$ **Update**: $\mathbf{w}_{k}(t+1)=\tilde{\mathbf{W}}_{n}(t+1)$
Numerical Results
=================
Number of sub-carriers $M=600$
------------------------ -----------
Sub-carrier spacing $30KHz$
Noise power $-150$dB
MBS Tx power $20$W
SBS Tx Power $6.3$W
MU Tx power $0.2$W
Path-loss exponent 2.8
BER $10^{-3}$
: Simulation Parameters
Network topology
----------------
We consider a circular area with radius $750$ meters where users are generated uniformly randomly inside it. We consider hexagonal clusters where the diameter of circle inscribed is $500$ meters. The SBSs are exactly resided in the center of the hexagons. To mitigate the interference between the clusters, we use a simple reuse pattern [@reuse] as shown in Figure \[fig:cluster\]. We assume that the fronthaul link is $100$ times faster than the UL, DL between MUs and SBSs [^5]. Total number of clusters are $7$.
![Clustering layout. Frequency reuse pattern is one. Each color illustrates distinct set of subcarriers.[]{data-label="fig:cluster"}](clusters.png)
There are $300$ sub carriers with sub carrier spacing of $30$ KHz. The maximum transmit powers of MBS and SSBs are $20$ and $6.3$W, respectively and maximum transmit power of MUs is $0.2$W [@earth].
Implementation guideline
------------------------
In our numerical analysis, we consider the CIFAR-10 dataset for image classification problem with 10 different image classes [@dataset2] and train the ResNet18 architecture [@training2]. For further details regarding the trained NN structures please refer to [@nn].\
For the simulations, we also utilize some large batch training tricks such as scaling the learning rate $\eta$ and employing a warm-up phase [@training1]. In all simulations, data sets are divided among the MUs without any shuffling and through the iterations MUs train the same subset of the dataset as in the FL framework and we set the batch size for training to $\beta=64$. In general, batch size $K\times \beta=128$ is accepted as the baseline batch size with the corresponding learning rate $\eta=0.1$ and the initial learning rate is adjusted according to the cumulative batch size $K\times\beta$ accordingly [@training1; @training2]. Hence, we set the initial learning to $0.25$, also we consider the first $5$ epoch as the gradual warm-up phase where training starts with a small learning rate and then increased linearly at each iteration until it reaches the value of the given initial learning rate. For the network training, we follow the guideline in [@sgd_local4], we train the network for 300 epochs, and at the end of 150th epoch we drop the initial learning rate by factor of 10, and similarly end of the 225th epoch we drop the learning rate by factor of 10 again. Further, for the weight decay factor[^6] and the momentum term we use $w=0.0001$ and $\sigma=0.9$ respectively in all simulations. Finally, for the discounted error accumulation we set $\beta_{m}=0.2$ and $\beta_{s}=0.5$.
Results
-------
We first study the amount of speed up in latency achieved by HFL versus FL. We measure the speed by comparing the per iteration latency of HFL i.e., $\Gamma^{HFL}$ and FL, $T^{FL}$. More specifically, speed up $=\frac{T^{FL}}{\Gamma^{HFL}}$. By varying the number of MUs in each cluster, and for different periods of $H=2,4,6$, we plot the speed up achieved by HFL, when sparsity parameters $\phi^{ul}_{MU}=0.99$, $\phi^{dl}_{SBS}=0.9$, $\phi^{ul}_{SBS}=0.9$, $\phi^{dl}_{MBS}=0.9$ are used, in Figure \[fig:latency\_comp\]. We observe that HFL achieves good latency speed up with respect to FL and it improves when the period increases.
![Latency Speed Up HFL versus FL.[]{data-label="fig:latency_comp"}](latency_comp.pdf)
Clustering reduces the communication distance and as a result improves the SNR. The amount of improvement depends on the amount of reduction in path-loss. In Figure \[fig:latency\_comp\_pathloss\], we illustrate the amount of latency speed up due to clustering as a function of the path-loss exponent, $\alpha$. When the path-loss exponent is increased, the SNR in centralized scheme is punished more severely than clustering scheme due to longer communication paths. Thus, the latency speed up improves when the path-loss is more severe.
![Latency speed Up HFL versus FL as a function of path-loss exponent $\alpha$.[]{data-label="fig:latency_comp_pathloss"}](pathloss_comp.pdf)
To see the importance of sparsification, we compare the HFL and FL with sparse HFL and sparse FL in Figure \[fig:sparse\_comp\_c\] and \[fig:sparse\_comp\_noc\], respectively. For both FL and HFL the sparsification provides a significant improvment. However, the latency improvement in HFL is more robust with respect to increasing the number of MUs. This is due to the fact that MBS is serving more number of MUs than a single MBS, and hence the scarcity of resources in macro cell has more impact to the latency than small cells.
The Top-1 accuracy achieved by FL and HFL algorithms are illustrated in Figure \[fig:resnetacc\] for CIFAR-10 data set trained with ResNet 18. We observe that latency speed up delivered by HFL over FL schemes does not compromise the accuracy of the ML model. In fact, a closer look at the accuracy (average over $5$ runs) presented in Table \[tab:acc\] show that HFL is able to achieve a better accuracy than FL in all situations. The $\emph{\text{mean}}\pm \emph{\text{standard error mean}}$ results for the last epoch is reprted in Table \[tab:acc\], where the *Baseline* result is obtained by training a single MU on the whole training set. We observe a small degradation in the accuracy for our introduced hierarchical distributed learning strategy. We conjecture that this degradation is mainly due to the use of momentum SGD at each MU instead of a global momentum and due to sparsification. We also observe that FL, based on [@SGD_sparse1], performs poorly compared to HFL. We believe that this poor performance is mainly due to the downlink sparsification that we consider.\
We believe that using an additional global momentum term [@sgd_local4] or utilizing momentum averaging strategy [@sgd_local3] accuracy of the HFL can be improved further.
[.55]{} ![Latency speed up due to sparsification[]{data-label="fig:sparse_comp"}](sparse_comp_c.pdf "fig:"){width=".8\linewidth"}
[.55]{} ![Latency speed up due to sparsification[]{data-label="fig:sparse_comp"}](sparse_comp_noc.pdf "fig:"){width=".8\linewidth"}
![Top-1 accuracy []{data-label="fig:resnetacc"}](resnet.pdf)
CIFAR-10 [@dataset2] / ResNet18 [@training2] Baseline $92.48\pm 0.13$ -
---------------------------------------------- ------------ ------------------ -------------------------------
CIFAR-10 [@dataset2] / ResNet18 [@training2] FL $89.23\pm 0.42$ 28 MUs
CIFAR-10 [@dataset2] / ResNet18 [@training2] HFL, $H=2$ $90.27\pm 0.11$ $7$ clusters $\times$ $4$ MUs
CIFAR-10 [@dataset2] / ResNet18 [@training2] HFL, $H=4$ $90.474\pm 0.20$ $7$ clusters $\times$ $4$ MUs
CIFAR-10 [@dataset2] / ResNet18 [@training2] HFL, $H=6$ $91.03\pm 0.19$ $7$ clusters $\times$ $4$ MUs
Discussions
-----------
As mentioned above, a momentum correction strategy for the clusters can help to improve the accuracy as weel as increasing the convergence speed We will consider this as a future work.\
In addition, one of the recent research area regarding the federated learning framework is the non-IID data distribution among the MUs [@fedlearn5; @fedlearn7; @fedlearn8; @fedlearn9] hence, we are planning to extend our study to the non-IID distribution scenario as well. Finally, another research direction that we are planning to investigate is the optimal batchsize for MUs. In [@training3], it has been shown that learning speed increases with the batchsize until certain point, hence we believe that a good strategy for federated learning is to adjust the batchsize according to the number users. We are planing to extend our analysis in this direction and do an extended analysis on training time including the computation time of the MUs.
[^1]: Emre Ozfatura and Deniz G[ü]{}nd[ü]{}z are with Information Processing and Communications Lab, Department of Electrical and Electronic Engineering, Imperial College London Email: {m.ozfatura, d.gunduz} @imperial.ac.uk.
[^2]: Mehdi Salehi Heydar Abad and Ozgur Ercetin are with the Faculty of Engineering and Natural Sciences, Sabanci University Email: {mehdis, oercetin} @sabanciuniv.edu
[^3]: This work was supported in part by the Marie Sklodowska-Curie Action SCAVENGE (grant agreement no. 675891), and by the European Research Council (ERC) Starting Grant BEACON (grant agreement no. 677854).
[^4]: Note that the MBS can update the model itself and transmit the updated model instead of the average gradients, thus avoiding replicating the update at $K$ MUs. However, it is possible to apply sparsification on the average gradient to improve latency in this manner.
[^5]: For a $8\times8$ MIMO the fornthaul rate estimate is $8$ Gbps with 3GPP split option 2 and $67$ Gbps with 3GPP split option 7 [@5gfronthaul].
[^6]: We dont apply weight decay to batch normalization parameters
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The paper studies the problem of making Getz’s bicycle model traverse a strictly convex Jordan curve with bounded roll angle and bounded speed. The approach to solving this problem is based on the virtual holonomic constraint (VHC) method. Specifically, a VHC is enforced making the roll angle of the bicycle become a function of the bicycle’s position along the curve. It is shown that the VHC can be automatically generated as a periodic solution of a scalar periodic differential equation, which we call virtual constraint generator. Finally, it is shown that if the curve is sufficiently long as compared to the height of the bicycle’s centre of mass and its wheel base, then the enforcement of a suitable VHC makes the bicycle traverse the curve with a steady-state speed profile which is periodic and independent of initial conditions. An outcome of this work is a proof that the constrained dynamics of a Lagrangian control system subject to a VHC are generally not Lagrangian.'
author:
- 'Luca Consolini, Manfredi Maggiore,[^1]'
bibliography:
- 'citazioni.bib'
title: Control of a Bicycle Using Virtual Holonomic Constraints
---
Introduction
============
This paper investigates the problem of maneuvering a bicycle along a closed Jordan curve $\C$ in the horizontal plane in such a way that the bicycle does not fall over and its velocity is bounded. The simplified bicycle model we use in this paper, developed by Neil Getz [@Get94; @GetMar95], views the bicycle as a point mass with a side slip velocity constraint, and models its roll dynamics as those of an inverted pendulum, see Figure \[fig:bicycle\]. The model neglects, among other things, the steering kinematics and the wheels dynamics with the associated gyroscopic effect. In [@HauSacFre04], Hauser-Saccon-Frezza investigate the maneuvering problem for Getz’s bicycle using a dynamic inversion approach to determine bounded roll trajectories. They constrain the bicycle on the curve and, given a desired velocity signal $v(t)$, they find a trajectory with the property that the velocity of the bicycle is $v(t)$ and its roll angle $\varphi$ is in the interval $(-\pi/2,\pi/2)$, i.e, the bicycle doesn’t fall over. In [@HauSac06], Hauser-Saccon develop an algorithm to compute the minimum-time speed profile for a point-mass motorcycle compatible with the constraint that the lateral and longitudinal accelerations do not make the tires slip, and apply their algorithm to Getz’s bicycle model.
The problem of maneuvering Getz’s bicycle along a closed curve is equivalent to moving the pivot point of an inverted pendulum around the curve without making the pendulum fall over. On the other hand, the seemingly different problem of maneuvering Hauser’s PVTOL aircraft [@HauSasMey92] along a closed curve in the vertical plane can be viewed as the problem of moving the pivot of an inverted pendulum around the curve without making the pendulum fall over. The two problems are, therefore, closely related, the main difference being the fact that in the former case the pendulum lies on a plane which is orthogonal to the plane of the curve, while in the latter case it lies on the same plane. In [@ConMagNieTos10], the path following problem for the PVTOL was solved by enforcing a virtual holonomic constraint (VHC) which specifies the roll angle of the PVTOL as a function of its position on the curve. In this paper we follow a similar approach for the bicycle model, and impose a VHC relating the bicycle’s roll angle to its position along the curve. However, rather than finding one VHC, as we did in [@ConMagNieTos10], we show how to automatically generate VHCs as periodic solutions of a scalar periodic differential equation which we call the [*virtual constraint generator*]{}. We show that if the path is sufficiently long compared to the height of the bicycle’s centre of mass and the wheel base, then the VHC can be chosen so that on the constraint manifold the bicycle traverses the entire curve with bounded speed, and its speed profile is periodic in steady-state. Finally, we design a controller that enforces the VHC, and recovers the asymptotic properties of the bicycle on the constraint manifold.
![Getz’s bicycle model.[]{data-label="fig:bicycle"}](FIGURES/bicycle){width=".5\textwidth"}
The concept of VHC is a promising paradigm for motion control. It is one of the central ideas in the work of Grizzle and collaborators on biped locomotion (e.g., [@PleGriWesAbb03] and [@WesGriKod03]), where VHCs are used to encode different walking gaits. The work of Shiriaev and collaborators in [@Can04; @ShiPerWit05; @FreRobShiJoh08] investigates VHCs for Lagrangian control systems, i.e., systems of the form [@Blo03] $$\label{eq:lagrangian}
\frac{d}{dt} \frac{\partial L}{\partial \dot q} - \frac{\partial
L}{\partial q} = G u,$$ with control input $u$ and smooth Lagrangian $L(q,\dot q) = (1/2) \dot
q\trans D(q) \dot q - V(q)$, with $D = D\trans$ positive definite. In [@ShiPerWit05], the authors consider systems of the form with degree of underactuation one. They find an integral of motion for the constrained dynamics, and use it to select a desired closed orbit on the constraint manifold. This orbit is then stabilized by linearizing the control system along it, and designing a time-varying controller for the linearization. In [@ShiFreGus10], these ideas are extended to systems with degree of underactuation greater than one, and in [@ShiFreRobJohSan07] they are applied to the stabilization of oscillations in the Furuta pendulum. In [@MagCon13], we investigated VHCs for Lagrangian control systems with degree of underactuation one. We introduced and characterized a notion of regularity of VHCs, and we presented sufficient conditions under which the reduced dynamics on the constraint manifold (described by a second-order unforced system) are Lagrangian (i.e., they satisfy the Euler-Lagrange equations, which have the form with zero right-hand side). An outcome of this paper (Proposition \[prop:internal\_dynamics\_general\_properties\]) is a simple sufficient condition under which the reduced dynamics are [ *not*]{} Lagrangian. We refer the reader to Remarks \[rem:shiriaev\], \[rem:Lagrange\], and \[rem:not\_EL\] for more details.
This paper is organized as follows. In Section \[sec:problem\] we present Getz’s bicycle model and we formulate the maneuvering problem investigated in this paper. In Section \[sec:vcg\] we present the virtual constraint generator idea. The main result is Proposition \[prop:generator:periodic\_solutions\] which gives a constructive methodology to find VHCs for Getz’s bicycle that meet the requirements of the maneuvering problem. In Section \[sec:constrained\_motion\] we analyze the motion of Getz’s bicycle on the virtual constraint manifold. In Proposition \[prop:internal\_dynamics\_general\_properties\] we provide a general result with sufficient conditions under which an unforced second-order system of a certain form possesses an asymptotically stable closed orbit. In Proposition \[prop:internal\_dynamics\_bicycle\] we apply this result to Getz’s bicycle model. In Section \[sec:solution\] we bring these results together and solve the maneuvering problem. Finally, in Section \[sec:numerics\] we make remarks on numerical implementation of the proposed controller.
[**Notation.**]{} Throughout this paper we use the following notation. If $x$ is a real number and $T>0$, the number $x$ modulo $T$ is denoted by $[x]_T$. We let $[\Re]_T : = \{[x]_T: x \in \Re\}$. The set $[\Re]_T$ is diffeomorphic to the unit circle. We let $\pi: \Re
\to [\Re]_T$ be defined as $\pi(t) = [t]_T$. Then, $\pi$ is a smooth covering map (see [@Lee13 p.91]). If $M$ is a smooth manifold, and $h: [\Re]_T \to M$ is a smooth function, we define $\tilde h:=h
\circ \pi : \Re \to M$. This is a $T$-periodic function. Moreover, by [@Lee13 Theorem 4.29], $\tilde h$ is smooth if and only if $h$ is smooth. Finally, $\image(h)$ denotes the image of a function $h$.
Problem formulation {#sec:problem}
===================
Consider the bicycle model depicted in Figure \[fig:bicycle\], with the following variable conventions (taken from [@HauSacFre04]):
- $(x,y)$ - coordinates of the point of contact of the rear wheel
- $\varphi$ - roll angle (a positive $\dot \varphi$ implies that the bicycle leans to the right)
- $\psi$ - yaw angle (a positive $\dot \psi$ means that the bicycle turns right)
- $\alpha$ - projected steering angle on the $(x,y)$ plane
- $b$ - distance between the projection of the centre of mass and the point of contact of the rear wheel
- $p$ - wheel base
- $h$ - pendulum length
- $v$ - forward linear velocity of the bicycle
- $f$ - thrust force.
We denote $\bar \kappa = (\tan \alpha) / p = \dot \psi / v$. For a given velocity signal $v(t)$ and steering angle signal $\alpha(t)$, $\bar \kappa(t)$ represents the curvature of the path $(x(t),y(t))$ traced by the point of contact of the rear wheel. In [@GetMar95], the bicycle of Figure \[fig:bicycle\] was modelled by writing the Lagrangian of the unconstrained bicycle, incorporating the nonholonomic constraints that the wheels roll without slipping in the Lagrangian, and then extracting the model through the Lagrange-d’Alembert equations as in [@Blo03 Section 5.2]. The resulting model, which we’ll refer to as [*Getz’s bicycle model*]{}, reads as $$\begin{aligned}
&\dot {\bar \kappa} = \tau \nonumber\\
&M \begin{bmatrix} \ddot \varphi \\ \dot v
\end{bmatrix} =
F + B \begin{bmatrix}
\tau \\ f
\end{bmatrix}, \label{eq:original_system}\end{aligned}$$ where $\tau$, the time derivative of the [curvature $\bar
\kappa(t)$]{}, and $f$ are the control inputs and, denoting ${s_\varphi}=
\sin \varphi$, ${c_\varphi}=\cos \varphi$, $$\begin{aligned}
& M = \begin{bmatrix} h^2 & b h {c_\varphi}\bar \kappa \\
b h {c_\varphi}\bar \kappa & 1 + (b^2+h^2 {s_\varphi}^2) \bar \kappa^2 - 2 h {\bar \kappa} {s_\varphi}\end{bmatrix}, \\
& F= \begin{bmatrix} g h {s_\varphi}- (1 - h
{\bar \kappa} {s_\varphi}) h {c_\varphi}{\bar \kappa} v^2 \\ (1 - h {\bar \kappa} {s_\varphi}) 2 h {c_\varphi}{\bar \kappa}
v \dot \varphi + b h {\bar \kappa} {s_\varphi}\dot \varphi^2
\end{bmatrix}, \\
& B =\begin{bmatrix} -b h {c_\varphi}v & 0 \\ -(b^2 {\bar \kappa} -
h {s_\varphi}(1-h {\bar \kappa} {s_\varphi})) v & 1/m
\end{bmatrix}.
\end{aligned}$$ In this model, $M$ is the inertia matrix, $F$ represents the sum of Coriolis, centrifugal and conservative forces, and $B$ is the input matrix. Now consider a $C^3$ closed Jordan curve $\C$ in the $(x,y)$ plane with regular parametrization $\sigma: [\Re]_T \to \Re^2$, not necessarily unit speed. Let $\kappa:[\Re]_T \to \Re$ denote the signed curvature of $\C$. Throughout this paper, we assume the following.
\[assump\_convex\] The curve $\C$ is strictly convex, i.e., $\kappa(s) > 0$ for all $s
\in [\Re]_T$.
In this paper we investigate the dynamics of the bicycle when the point $(x,y)$ is made to move along the curve $\C$ by an appropriate choice of the steering angle. In order to derive the constrained dynamics, suppose that $( x(0), y(0) ) \in \C$, i.e., $(x(0),y(0)) =
\sigma(s_0)$, for some $s_0 \in [\Re]_T$. A point $\sigma(s(t))$ moving on $\C$ has linear velocity $$\label{eqn_for_v}
v(t) = \|\sigma'(s(t))\| \dot s(t)$$ and acceleration $$\label{eqn_for_v_dot}
\dot v(t) = \|\sigma'(s(t))\| \ddot s(t) + \frac{\dot s^2(t)}{
\|\sigma'(s(t))\|} \sigma'(s(t))\trans \sigma''(s(t)).$$
For an arbitrary input signal $f(t)$, $(x(t), y(t))$ traverses $\C$ with velocity $v(t)$ if and only if $(x(0),y(0)) \in \C$, $(\dot
x(0),\dot y(0))$ is tangent to $\C$, $\bar \kappa(0) = \kappa(s_0)$, and the input signal $\tau(t)$ is chosen to be $\tau(t) = \kappa'(s(t)) \dot s(t)$, where $\kappa'(s(t)) =
\frac{d\kappa} {ds} (s(t))$. With this choice, we obtain $$\label{eqn_for_bar_kappa}
{\bar \kappa}(t) = \kappa(s(t)),$$ where $s(t)$ and $\dot s(t)$ are solutions of a differential equation to be specified later. The motion of the bicycle on the curve $\C$ is now found by substituting (\[eqn\_for\_v\]), (\[eqn\_for\_v\_dot\]), (\[eqn\_for\_bar\_kappa\]) and $\tau = \kappa'(s) \dot s$ in (\[eq:original\_system\]): $$\label{eq:constrained_system}
\bar{M} \begin{bmatrix} \ddot \varphi \\ \|\sigma^\prime\| \ddot {s} +
\frac{(\sigma')\trans \sigma''}{\|\sigma'\|} \dot{s}^2
\end{bmatrix} =
\bar F + \begin{bmatrix} 0 \\1/m \end{bmatrix} f\;,$$ where $\bar M=M|_{\bar \kappa =\kappa(s)}$ and $$\bar F = \left(F+B \begin{bmatrix} 1 \\ 0
\end{bmatrix}\kappa'(s) \dot s
\right)\bigg|_{\bar \kappa =\kappa(s), v=\|\sigma'(s)\|\dot s}.$$ Note that we have used the control input $\tau$ to make the curve $\C$ invariant, so in we are left with one control input, the thrust force $f$. One can check that system is a Lagrangian control system with input $f$, i.e., it has the form with $q=
(\varphi,s)\in \cQ =S^1 \times [\Re]_T$, $G = [0 \ \ 1/m]\trans$, and $$L(q,\dot q) = \frac 1 2 [\dot
\varphi \ \ \|\sigma^\prime(s)\| \dot{s}] \bar{M}(q) \begin{bmatrix}
\dot \varphi \\ \|\sigma^\prime(s)\| \dot{s}
\end{bmatrix}- g h \cos \varphi.$$ Since the control force $f$ enters nonsingularly in the $\ddot s$ equation, we can define a feedback transformation for $f$ in such that $\ddot s=u$, where $u$ is the new control input. With this transformation, the motion of the bicycle when its rear wheel is made by feedback control to follow $\C$ reads as $$\label{eq:sys}
\begin{aligned}
& \ddot \varphi = h^{-1}g {s_\varphi}- h^{-1}\Big[ (1 - h \kappa(s)
{s_\varphi})\kappa(s) \|\sigma'(s)\| + b \kappa'(s) \\
& +\frac{b \kappa(s)}{\|\sigma'(s)\|^2} {\sigma'(s)}\trans
\sigma''(s)\Big]{c_\varphi}\|\sigma'(s)\| \dot s^2 - a(s) {c_\varphi}u \\
& \ddot s = u,
\end{aligned}$$ where $a(s) = b h^{-1} \kappa(s) \|\sigma'(s)\|$. In the above equation, $u$ is the new control input, and it represents the acceleration of the curve parameter $s$. We will denote $\cX := \{
(q,\dot q) \in S^1 \times [\Re]_T \times \Re^2\}$ the state space of .
[If $\sigma(s)$ is a unit speed parametrization of $\C$ (i.e. it satisfies $\|\sigma'(s)\|=1$), then the first differential equation in reduces to $h \ddot \varphi = g {s_\varphi}- \Big[ (1 - h \kappa(s) {s_\varphi})\kappa(s) + b
\kappa'(s)\Big]{c_\varphi}\dot s^2 - b \kappa(s) {c_\varphi}u$. ]{}
[Maneuvering Problem]{} Find a feedback $u(q,\dot q)$ for system such that there exists a set of initial conditions $\Omega$ with the property that, for all $( q(0),\dot q(0) ) \in \Omega$, the bicycle does not overturn, i.e., $|\varphi(t)| < \pi/2$ for all $t \geq 0$, and traverses the entire curve $\C$ in one direction, i.e., there exists $\bar t>0$ such that $| \dot s(t) | >0$ for all $ t \geq \bar t$. Moreover, the speed $\dot s(t)$ of the bicycle on $\C$ should remain bounded.
Our solution of this problem relies on the notion of VHC.
A [**virtual holonomic constraint (VHC)**]{} for system is a relation $\varphi = \Phi(s)$, where $\Phi: [\Re]_T \to S^1$ is smooth and the set $\Gamma = \{(q,\dot q)\in \cX: \varphi = \Phi(s), \ \dot \varphi = \Phi'(s)
\dot s\}$ is controlled invariant. That is, there exist a smooth feedback $u(q,\dot q)$ such that $\Gamma$ is invariant for the closed-loop system. The set $\Gamma$ is called the [**constraint manifold**]{} associated with the VHC $\varphi = \Phi(s)$.
The definition above formalizes the notion of VHC used in [@WesGriKod03] in the context of biped locomotion. The constraint manifold $\Gamma$ is a two-dimensional submanifold of $\cX$ parametrized by $(s,\dot s)$, and therefore diffeomorphic to the cylinder $[\Re]_T \times \Re$. It is the collection of all those phase curves of that can be made to satisfy the constraint $\varphi = \Phi(s)$ via feedback control. In order to solve the maneuvering problem, we look for VHCs $\varphi = \Phi(s)$ such that $| \Phi (s) | < \pi/2$ for all $s \in [\Re]_T$, and then stabilize the associated virtual constraint manifold. In [@HauSacFre04], Hauser-Saccon-Frezza find “bounded roll trajectories,” i.e., controlled trajectories of along which the roll angle $\varphi$ is bounded in the interval $(-\pi/2,\pi/2)$. In our context, each VHC $\varphi = \Phi(s)$ provides a [*family*]{} of bounded roll trajectories. Once $\Gamma$ has been made invariant via feedback control, bounded roll trajectories can be obtained by picking arbitrary $(s_0,\dot s_0) \in
[\Re]_T \times \Re$, and picking as initial condition in , $(\varphi(0),s(0)) = (\Phi(s_0),s_0)$, $(\dot
\varphi(0),\dot s(0)) = (\Phi'(s_0)\dot s_0,\dot s_0)$. The resulting solution $(q(t),\dot q(t))$ will satisfy $\varphi(t) = \Phi(s(t))$, implying that the roll angle trajectory $\varphi(t)$ is bounded in the interval $(-\pi/2,\pi/2)$.
The virtual constraint generator {#sec:vcg}
================================
In this section we show that VHCs for can be generated as solutions of a first-order differential equation, which we call the [*VHC generator.*]{} This idea was first presented in our previous work [@ConMag10_1]. We begin with a sufficient condition for a relation $\Phi$ to be a VHC for .
\[lem:feasibility\] A relation $\varphi = \Phi(s)$, where $\Phi:[\Re]_T \to S^1$ is smooth, is a VHC for system if $$\label{eq:feasibility}
(\forall s \in [\Re]_T) \ \Phi'(s) + a(s) \cos \Phi(s) \neq 0.$$
Letting $H(\varphi,s):=\varphi - \Phi(s)$, condition is the requirement that system with output $H(\varphi,s)$ has a well-defined uniform relative degree 2 on the set $\{(q,\dot q)\in \cX: \varphi =
\Phi(s)\}$. The associated zero dynamics manifold is precisely $\Gamma$, and it is controlled invariant.
The foregoing lemma inspires the following observation. Instead of guessing a relation $\varphi = \Phi(s)$ and checking whether it is a VHC for , as in Lemma \[lem:feasibility\], we will assign a nonzero right-hand side to , view the resulting identity as an ODE, and [*generate*]{} VHCs by solving the ODE. More precisely, recall the notation $\tilde h:=h \circ \pi$, and consider the scalar $T$-periodic differential equation $$\label{eq:constraint_generator}
\dot x = - \tilde a(t)\cos x + \tilde \delta(t),$$ where $\tilde \delta:\Re \to \Re \backslash \{0\}$ is a $T$-periodic function to be assigned. Then, $T$-periodic solutions of give rise to VHCs.
\[lem:vcg\] Suppose that $x(t)$ is a $T$-periodic solution of , where $\tilde \delta :\Re \to \Re
\backslash \{0\}$ is smooth and $T$-periodic. Then, the relation $\varphi = \Phi(s)$, with $\Phi:=x \circ \pi^{-1}:[\Re]_T \to \Re$, is a VHC for .
Let $x(t)$ be a $T$-periodic solution of . Then, $x(t)$ is smooth because the right-hand side of is smooth. Since $x(t)$ is $T$-periodic, $\Phi:=x \circ \pi^{-1}$ is a well-defined function $[\Re]_T \to \Re$, and by [@Lee13 Theorem 4.29] it is smooth. Since $x = \Phi \circ \pi$, we have $\dot x(t) =
\Phi'(\pi(t)) \dot \pi(t) = \Phi'(\pi(t))$, so that $\Phi'(s) =\dot
x(t)|_{t = \pi^{-1}(s)}$. Using , we get $$\begin{aligned}
\Phi'(s) & = -\tilde a(\pi^{-1}(s)) \cos \Phi(s) + \tilde
\delta(\pi^{-1}(s)) \nonumber \\
&=- a(s) \cos \Phi(s) +
\delta(s). \label{eq:Phiprime}\end{aligned}$$ Since $\delta(s) \neq 0$ for all $s \in [\Re]_T$, by Lemma \[lem:feasibility\] the relation $\varphi = \Phi(s)$ is a VHC for system .
In light of Lemma \[lem:vcg\], we call the differential equation a [*VHC generator*]{}, for which one is to pick a $T$-periodic $\tilde \delta:\Re \to \Re
\backslash \{0\}$ yielding a $T$-periodic solution. The next proposition shows how to pick $\tilde \delta$.
\[prop:generator:periodic\_solutions\] Consider the smooth and $T$-periodic differential equation , and set $\tilde \delta(t) =
\epsilon \tilde \mu(t)$, where $\tilde \mu: \Re \to (0,\infty)$ is smooth and $T$-periodic. Let $K^+ = \max_{t \in [0, T]}[ \tilde
\mu(t)/\tilde a(t)]$, $K^- = \min_{t \in [0,T]} [\tilde
\mu(t)/\tilde a(t)]$. Then, for any $x_0 \in (0,\pi/2)$ and $t_0
\in \Re$, the following two properties hold:
(i) there exists a unique $\epsilon \in [\epsilon^-,\epsilon^+] = [
( c_{x_0}) / K^+,$ $( c_{x_0}) / K^- ]$, such that the solution $x(t)$ of with initial condition $x(t_0) = x_0$ is $T$-periodic, and setting $\Phi=x \circ \pi^{-1}$, the relation $\varphi = \Phi(s)$ is a VHC for .
(ii) If $\tilde\mu:\Re \to (0,\infty)$ is chosen so that $K^+ / K^- <
(\cos x_0)^{-1}$, then the VHC $\varphi=\Phi(s)$ in part (i) satisfies $\image(\Phi) \subset (\Phi^-,\Phi^+) \subset (0,\pi/2)$, where $\Phi^-=\cos^{-1}\big(\frac{K^+}{K^-} c_{x_0}\big)$ and $\Phi^+=\cos^{-1}\big(\frac{K^-}{K^+} c_{x_0}\big)$.
\[rem:mu\] [By choosing $\tilde \mu(t) \equiv \tilde a(t)$, we have $K^+ = K^- =1$, $\epsilon^+ = \epsilon^- = c_{x_0}$. Proposition \[prop:generator:periodic\_solutions\] implies that, for all $x_0 \in (0,\pi/2)$, setting $\tilde \delta(t) = \cos (x_0)
\tilde a(t)$, the VHC generator produces a $T$-periodic solution $x(t)$ with image contained in $(0,\pi/2)$. As a matter of fact, the solution in question is $x(t) \equiv x_0$, so that the relation $\varphi =x_0$ is a VHC for . The bicycle subject to this VHC has a constant roll angle as it travels around $\C$. The proposition provides great flexibility in finding VHCs with the property that the roll angle is confined within the interval $(0,\pi/2)$. All such constraints are compatible with the maneuvering problem.]{}
Pick an arbitrary $x_0 \in (0,\pi/2)$ and $t_0 \in \Re$, and consider the VHC generator $$\label{eq:constraint_generator:proof}
\dot x = - \tilde a(t) \cos x + \epsilon \tilde \mu(t),$$ where $\tilde \mu(\cdot)>0$. Let $\cS(\epsilon)$ denote the solution of at time $t_0+T$ with initial condition $x(t_0) = x_0$. Since the right-hand side of is $T$-periodic, the solution of with initial condition $x(t_0) =x_0$ is $T$-periodic if and only if $\cS(\epsilon) = x_0$. For all $\epsilon > \epsilon^+$ we have $$\begin{aligned}
\dot x & > -\tilde a(t) \cos x + \tilde \mu(t) \cos(x_0) / K^-\\
& > -\tilde a(t) \left[ \cos x - \cos x_0 (\tilde \mu(t) / \tilde
a(t) ) / K^- \right]] \\
& > -\tilde a(t) (\cos x - \cos x_0).
\end{aligned}$$ The solution of $\dot x = -\tilde a(t) (\cos x - \cos x_0)$ with initial condition $x(t_0) = x_0$ is $x(t) \equiv x_0$. Therefore, by the comparison lemma ([@Kha02 Lemma 3.4]), for all $\epsilon>
\epsilon^+$ the solution of with initial condition $x(t_0) = x_0$ satisfies $x(t)> x_0$ for all $t >
t_0$, so that $\cS(\epsilon)> x_0$ for all $\epsilon> \epsilon^+$. A similar argument can be used to show that for all $\epsilon<
\epsilon^-$ the solution of satisfies $x(t) < x_0$ for all $t>t_0$, so that $\cS(\epsilon)< x_0$ for all $\epsilon<\epsilon^-$. By continuity of solutions with respect to parameters, the map $\cS$ is continuous, and therefore there exists $\epsilon \in [\epsilon^-,\epsilon^+]$ such that $\cS(\epsilon) =
x_0$. The corresponding solution $x(t)$ is $T$-periodic. The same comparison argument shows that if $0<\epsilon_1 < \epsilon_2$, and for $i=1,2$, $x_i(t)$ is the solution of with $\epsilon = \epsilon_i$, then $x_1(t_0+T) < x_2(t_0+T)$, so that $\cS(\epsilon_1) <
\cS(\epsilon_2)$. In other words, $\cS(\epsilon)$ is a monotonically increasing function, and so the value of $\epsilon$ above is unique. This concludes the proof of part (i). Now suppose that $K^+ /
K^- < (\cos x_0)^{-1}$, and let $\epsilon \in [\epsilon^-,\epsilon^+]$ be such that the solution $x(t)$ of with initial condition $x(t_0)=x_0$ is $T$-periodic, as in part (i). Let $x^+ =
\cos^{-1}\big(\frac{K^-}{K^+} c_{x_0}\big)$, and consider the subset of the extended phase space $S^+= \{(x,t): x > x^+\}$. This subset is positively invariant since $\tilde a>0$ and $$\begin{aligned}
\dot x\big|_{x=x^+} &= -\tilde a(t) \frac{K^-}{K^+} \cos x_0 +
\epsilon \tilde \mu(t) \\
&\hspace*{-5ex}\geq -\tilde a(t)
\frac{K^-}{K^+} \cos x_0 + (\cos x_0 / K^+) \tilde \mu(t)\\
&\hspace*{-5ex}\geq \frac{\tilde a(t)}{K^+} \cos x_0 \left( -K^- +
\tilde \mu(t)/\tilde a(t)\right)\\
&\hspace*{-5ex}\geq 0.
\end{aligned}$$ Similarly, letting $x^- = \cos^{-1}\big(\frac{K^+}{K^-} c_{x_0}\big)$, the subset $S^-= \{(x,t): x < x^-\}$ is positively invariant. Since $\{ (x,t): x= x_0\}$ has empty intersection with $S^+ \cup S^-$, and since $S^+ \cup S^-$ is positively invariant, it follows that the $T$-periodic solution $x(t)$ with initial condition $x(t_0)=x_0$ must be contained in the complement of $S^+ \cup S^-$, i.e., for all $t\in\Re$, $x^- \leq x(t) \leq x^+$. For, if for some $\bar t \in
\Re$, $x(\bar t) \in S^+ \cup S^-$, then the fact that $x(t) \in S^+
\cup S^-$ for all $t\geq \bar t$ would contradict the periodicity of $x(t)$.
\[ex:ellipse\]
Suppose $\C$ is an ellipse with major semiaxis $A$, minor semiaxis $B$, and $2 \pi$-periodic parametrization $\sigma(s) = (A
\cos s, B \sin s)$, with $A=15$, $B=10$. The curvature is [$\kappa(s) = AB / ( A^2 \sin^2 s + B^2 \cos^2 s
)^{3/2}$]{}. For the initial condition of the VHC generator , we pick $x(0) = \pi/8$. Following Proposition \[prop:generator:periodic\_solutions\], we need to choose a $2 \pi$-periodic function $\tilde \mu(t)>0$, set $\tilde \delta(t) = \epsilon \tilde \mu(t)$, and find the unique value of $\epsilon>0$ guaranteeing that the solution with initial condition $x(0) = \pi/8$ is $2\pi$-periodic. There is much freedom in the choice of $\tilde \mu(t)$. For instance, picking $\tilde
\mu(t) = 1$, we numerically find [$\epsilon \approx
0.6482$]{}. The corresponding VHC is depicted in Figure \[fig:ellipse:virtual\_constraint\]. The condition, in Proposition \[prop:generator:periodic\_solutions\](ii), that $K^+/
K^- < (\cos x_0)^{-1}$ is conservative. Indeed, with our choice of $\tilde \mu$ we have $K^+ =3/2$, $K^-=2/3$, and thus the condition is violated. Yet, we see in Figure \[fig:ellipse:virtual\_constraint\] that $ \Phi(s) \in (0,
\pi/2)$ for all $s \in [\Re]_T$.
![VHC for the ellipse in Example \[ex:ellipse\].[]{data-label="fig:ellipse:virtual_constraint"}](FIGURES/ellipse_constraint_pend "fig:"){width=".495\textwidth"} ![VHC for the ellipse in Example \[ex:ellipse\].[]{data-label="fig:ellipse:virtual_constraint"}](FIGURES/ellipse_constraint_fun "fig:"){width=".495\textwidth"}
Motion on the constraint manifold {#sec:constrained_motion}
=================================
Having chosen $\tilde \delta(t)= \epsilon \tilde \mu(t)$ as in Proposition \[prop:generator:periodic\_solutions\](ii), and having obtained a VHC $\varphi= \Phi(s)$, the next step is to analyse the [*reduced dynamics*]{} of on the constraint manifold $\Gamma = \{ (q,\dot q) \in \cX : \varphi = \Phi(s), \ \dot \varphi =
\Phi'(s) \dot s\}$. These are the zero dynamics of with output function $H(\varphi,s)=\varphi - \Phi(s)$. To this end, we impose that $[(d/dt) (\Phi'(s) \dot s)] |_\Gamma = \ddot
\varphi|_\Gamma$. Expanding both sides of the equation above, using the expression of $\ddot \varphi$ in , identity , and the fact that $\delta \neq 0$, we obtain the feedback making $\Gamma$ invariant $$\label{eqn_inv_feedback}
\begin{aligned}
& u = \frac{h^{-1}g \sin \Phi}{\delta} - \frac{ \dot s^2
}{\delta} \Big[ \Phi'' + \frac 1 h ( (1-h \kappa \sin \Phi) \kappa
\|\sigma'\| \\
& + b \kappa' +b\kappa {\sigma'}\trans\sigma'' / \|\sigma'\|^2)
\|\sigma'\| \cos \Phi \Big].
\end{aligned}$$ Substituting feedback in the $s$ dynamics, we get the reduced dynamics on $\Gamma$ $$\label{eq:motion_on_Gamma}
\ddot s = \Psi_1(s) + \Psi_2(s) \dot s^2,$$ with $$\label{eq:Psi}
\begin{aligned}
&\Psi_1 = \frac{h^{-1}g s_{\Phi}}{\delta} \\
& \Psi_2= - \frac{1}{\delta} \Big[ \Phi'' + \frac 1 h ( (1-h
\kappa s_{\Phi}) \kappa\|\sigma'\| \\ & + b \kappa'
+ b \kappa \sigma'{}\trans \sigma'' / \|\sigma'\|^2 )
c_{\Phi} \|\sigma'\| \Big].
\end{aligned}$$ System describes the motion of system on the constraint manifold $\Gamma$ in the following sense. For a given initial condition $(s(0),\dot s(0)) =
(s_0,\dot s_0) \in [\Re]_T \times \Re$, let $(s(t),\dot s(t))$ be the corresponding solution of , and let $(\varphi(t),\dot \varphi(t), s_1(t),\dot s_1(t))$ be the solution of with initial condition $(\varphi(0),\dot
\varphi(0),s(0),\dot s(0)) = ( \Phi(s_0), \Phi'(s_0) \dot s_0,s_0,\dot
s_0) \in \Gamma$. Then, for all $t \geq 0$ $(s_1(t),\dot s_1 (t)) =
(s(t),\dot s(t))$, and $(\varphi(t),\dot \varphi(t)) =$ $(\Phi(s(t)),$ $\Phi'(s(t)) \dot s(t))$. In order for the VHC $\varphi=\Phi(s)$ to be compatible with the maneuvering problem, we need to verify whether or not the rear wheel of the bicycle traverses the entire curve $\C$ with bounded speed, i.e., we need to show that for any solution $(s(t),\dot s(t))$ of , there exist $\bar
t>0$ and $\epsilon>0$ such that $\dot s(t)>\epsilon > 0$ for all $ t
\geq \bar t$, and $\lim\sup_{t \to \infty} \dot s(t) < \infty$. The next result explores general properties of systems of the form .
\[prop:internal\_dynamics\_general\_properties\] Consider a differential equation of the form with state space $\{(s,\dot s) \in
[\Re]_T \times \Re\}$. Assume that $\Psi_1, \Psi_2:[\Re]_T \to \Re$ are smooth functions such that $\Psi_1(s) > 0$ for all $s$ and $\int_0^T \tilde \Psi_2(\tau) d\tau <0$. Then, there exists a smooth function $\nu:[\Re]_T \to (0,\infty)$ such that the closed orbit $\cR
= \{(s,\dot s) \in [\Re]_T \times \Re: \dot s = \nu(s)\}$ is exponentially stable for , with domain of attraction containing the set $\cD = \{(s,\dot s) \in [\Re]_T \times
\Re: \dot s \geq 0\}$. [Thus,]{} for all initial conditions in $\cD$, the solution $(s(t),\dot s(t))$ of converges to the unique asymptotically stable limit cycle $\cR$.
\[rem:doa\] [It can be shown that the domain of attraction of the limit cycle $\cR$ is $\{(s,\dot s) : \dot s > -
\nu(s)\}$.]{}
\[rem:shiriaev\] [Proposition \[prop:internal\_dynamics\_general\_properties\] has general implications. Consider a Lagrangian control system of the form with configuration vector $q=
(q_1,\ldots, q_n)$ and degree of underactuation one. Consider a VHC of the form $q_i = \Phi_i(q_n)$, $i=1,\ldots,n-1$, and suppose $q_n
\in [\Re]_T$. For this system, it is shown in [@WesGriKod03; @ShiPerWit05; @MagCon13] that the reduced dynamics have the form , with $s$ replaced by $q_n$. In this context, Proposition \[prop:internal\_dynamics\_general\_properties\] gives sufficient conditions under which the reduced dynamics have an asymptotically stable closed orbit, implying that they are not Lagrangian. On the other hand, in Theorem 2 of [@Shiriaev2006900] (see also [@ShiPerWit05]) it is shown that for any $(s_0,\dot s_0)$, system possesses an “integral of motion”[^2] $I(s,\dot s) = \dot s^2 +2 V(s)/ M(s) - \dot
s_0^2/M(s)$, where $ M(s) = \exp \left\{ -2 \int_{s_0}^s
\Psi_2(\tau) d \tau \right\}$, $V(s) = -\int_{s_0}^s \Psi_1(\mu)
M(\mu) d\mu$. This integral of motion predicted by Theorem 2 of [@Shiriaev2006900] seemingly contradicts Proposition \[prop:internal\_dynamics\_general\_properties\]. Indeed, if $M(s)$ and $V(s)$ are well-defined functions, letting $L(s,\dot
s) = (1/2) M(s) \dot s^2 - V(s)$ one has that $\frac{d}{dt}
\frac{\partial L}{\partial \dot s} - \frac{\partial L}{\partial s} =
0$, and hence the reduced dynamics are Lagrangian. This fact rules out the existence of isolated closed orbits for predicted by Proposition \[prop:internal\_dynamics\_general\_properties\]. This contradiction is due to the fact that Theorem 2 in [@Shiriaev2006900] implicitly assumes that the functions $M(s)$ and $V(s)$ are well-defined. This is always true if $s \in
\Re$. However, when $s \in [\Re]_T$, the functions $M(s)$ and $V(s)$ may be multi-valued, in which case the Lagrangian $L(s,\dot s)$ is undefined. Indeed, the assumptions of Proposition \[prop:internal\_dynamics\_general\_properties\] imply that both $M(s)$ and $V(s)$ are multi-valued. We refer the reader to Section IV of [@MagCon13] for a detailed discussion on this subject, and to [@MohMagCon13] for necessary and sufficient conditions under which is Lagrangian.]{}
\[rem:Lagrange\] [In Lagrangian mechanics, the Lagrange-d’Alembert principle implies that the enforcement of an ideal holonomic constraint (i.e., a constraint for which the constraint forces do not produce work) on a Lagrangian system gives rise to a reduced Lagrangian system. Proposition \[prop:internal\_dynamics\_general\_properties\] shows that this result is not true when enforcing VHCs on Lagrangian control systems. There is, therefore, a sharp difference between [*ideal*]{} and [*virtual*]{} holonomic constraints when it comes to the reduced motion they induce. This difference is due to the fact that VHCs are enforced through forces produced by feedback control. For underactuated systems such as the constrained Getz’s bicycle model in , the control forces produce work. ]{}
[In the context of walking robots, [@WesGriKod03] proved that under certain conditions, the hybrid zero dynamics subject to a VHC exhibit an exponentially stable hybrid limit cycle. The mechanism enabling this exponential stability property is the jump map representing the impact of the robot’s foot with the ground. Proposition \[prop:internal\_dynamics\_general\_properties\] shows that VHCs may induce stable limit cycles even when the control system has no jumps.]{}
Consider the differential equation $$\label{eq:constrained_system_lifted}
\ddot x = \tilde \Psi_1 (x) + \tilde \Psi_2(x) \dot x^2.$$ Let $\bar \pi: \Re \times \Re \to [\Re]_T \times \Re$ be defined as $\bar \pi(x,\dot x) = (\pi(x),\dot x)$. Then, it is readily seen that the vector fields in and are $\bar
\pi$-related [@Lee13]. It follows that if $(x(t),\dot x(t))$ is a solution of , then $(s(t),\dot
s(t)) = (\pi(x(t)),\dot x(t))$ is a solution of . We will show that there exists a smooth $T$-periodic function $\tilde \nu : \Re \to (0,+\infty)$ such that the set $\tilde \cR = \{(x,\dot x) \in \Re^2: \dot x = \tilde
\nu(x)\}$ is an exponentially stable orbit of , with domain of attraction containing $\tilde \cD = \{(x,\dot x)\in \Re^2: \dot x\geq 0\}$. Then, setting $\nu = \tilde \nu \circ \pi^{-1}$, by [@Lee13 Theorem 4.29] we will obtain a smooth function $[\Re]_T \to (0,+\infty)$, and the set $\cR= \{(s,\dot s) \in [\Re]_T
\times \Re: \dot s = \nu(s)\}$ will be an exponentially stable closed-orbit of with domain of attraction containing $\cD$, proving the proposition. The set $\tilde \cD$ is positively invariant for because, by assumption, $\ddot x|_{\dot x=0} = \tilde \Psi_1(x) >0$ for all $x\in \Re$. In the rest of the proof we will restrict initial conditions to lie in $\tilde \cD$. Letting[^3] $z = \dot
x^2$, we have $\dot z = 2\dot x (\tilde \Psi_1(x) + \tilde \Psi_2(x)
z)$. For all $(x,\dot x) \in \tilde \cD$, we have $\dot x>0$, so we can use $x$ as a time variable: $$\label{eq:core_subsystem}
\frac { d z } {dx} = 2\tilde \Psi_1(x) + 2\tilde \Psi_2(x) z.$$ The above is a scalar linear $T$-periodic system. Letting $\phi(x) =
\exp(2\int_0^x \tilde \Psi_2(\tau) d \tau)$, the solution of with initial condition $z(x_0)=z_0$ is $$z(x) = \phi(x) \phi(x_0)^{-1} z_0 + 2\int_{x_0}^x \phi(x) \phi^{-1}
(\tau) \tilde \Psi_1(\tau) d \tau.$$ Note that, for any integer $k$, $\phi(x+k T) = \phi(T)^k \phi(x)$. System has a $T$-periodic solution if and only if there exists $\bar z_0$ such that the solution $z(x)$ of with initial condition $z(x_0) = \bar
z_0$ satisfies, $z(x_0+T) = \bar z_0$, or $$\label{eq:z0}
\bar z_0 = \phi(T) \bar z_0 + 2 \int_{x_0}^{x_0+T}\phi(x_0+T)
\phi^{-1}(\tau) \tilde \Psi_1(\tau) d \tau.$$ By assumption, $0< \phi(T) <1$, so there is a unique $\bar z_0>0$ solving , implying that there is a unique $T$-periodic solution $\bar z:\Re \to \Re$ of . Since for all $x\in \Re$, $z=0 \implies (d z/ dx) = 2\tilde \Psi_1(x)>0$, the set $\{(z,x): z>0\}$ is positively invariant for , and therefore the $T$-periodic function $\bar z$ satisfies $\image(\bar z) \subset (0,\infty)$. Let $z:\Re
\to \Re$ be any other arbitrary solution of (therefore, not necessarily $T$-periodic). Let $k$ be a nonnegative integer and denote $z_k:=z(x_0+k T)$, $\phi_k:=\phi(x_0+kT)$, $x_k:=x_0+kT$. Then, $$\begin{aligned}
&z_{k+1} = \phi_{k+1} \phi_k^{-1} z_k + 2\int_{x_k}^{x_{k+1}}
\phi_{k+1} \phi^{-1}(\tau) \tilde \Psi_1(\tau) d \tau \\
& = \phi(T) z_k +2\int_{x_0}^{x_1} \phi_1 \phi(T)^k
\phi^{-1} (\tau + kT) \tilde \Psi_1(\tau + kT) d \tau \\
& = \phi(T) z_k +2\int_{x_0}^{x_1} \phi_1
\phi^{-1}(\tau) \tilde \Psi_1(\tau) d \tau \; \text{($\tilde
\Psi_1$ is $T$-periodic)}\\
& = \phi(T) ( z_k - \bar z_0) + \bar z_0. \quad
\text{(by~\eqref{eq:z0})}
\end{aligned}$$ In light of the above, letting $\tilde z_k = z_k -\bar z_0$, we have $\tilde z_{k+1} = \phi(T) \tilde z_k$. Since $0<\phi(T)<1$, the origin of this discrete-time system is globally exponentially stable, proving that the $T$-periodic solution $\bar z:\Re\to(0,\infty)$ is globally exponentially stable for . Define $\tilde \nu:\Re \to (0,\infty)$ as $\nu:= \sqrt{\bar z}$, and return to system . For an arbitrary initial condition $(x(0),\dot x(0)) \in \tilde \cD$, the solution $(x(t),\dot
x(t))$ satisfies $\dot x(t) > 0$ for all $t >0$, and $\dot x(t)
=\sqrt{ z(x(t))}$, where $z:\Re\to\Re$ is the solution of with initial condition $z(x(0)) =(\dot
x(0))^2$. Since the solution $\bar z:\Re \to(0,\infty)$ is globally exponentially stable for , the set $\tilde
\cR$ is exponentially stable for with domain of attraction containing $\tilde \cD$. Going back to $(s,\dot s)$ coordinates, this implies that $\cR = \{(s,\dot s) \in
[\Re]_T \times \Re: \dot s = \nu(s)\}$ is exponentially stable for , and its domain of attraction contains $\cD$. Note that $\cR$ is a simple closed curve in $[\Re]_T \times
\Re$. We are left to show that $\cR$ is a closed orbit of . The set $\cR$ is closed, invariant, and no proper subset of it has these properties. For, if there existed a closed and invariant proper subset $\cR' \subset \cR$, then any solution $(s(t),\dot s(t))$ originating in $\cR'$ would not traverse the entire curve $\cR$, contradicting the fact that, for all $(s,\dot
s)\in\cR$, $\dot s >0$. $\cR$ is, therefore, a minimal set for . By [@Har02 Theorem 12.1], $\cR$ is a closed orbit.
We now show that if $\C$ is sufficiently long as compared to $b$ and $h$, the bicycle satisfies the hypotheses of Proposition \[prop:internal\_dynamics\_general\_properties\], and so the reduced motion satisfies the requirements of the maneuvering problem.
\[prop:internal\_dynamics\_bicycle\] Assume that the curvature of $\C$ satisfies the inequality $$\label{eq:curvature_bound}
\frac 1 L \int_0^T \tilde \kappa(\tau) \|\tilde \sigma'(\tau)\| d\tau
< \frac{h}{b^2+h^2},$$ where $L=\int_0^T \|\tilde \sigma'(\tau)\| d \tau$ is the length of $\C$. Then, any VHC $\varphi = \Phi(s)$ with $\image(\Phi) \subset
(0,\pi/2)$ yields functions $\Psi_1,\Psi_2:[\Re]_T \to \Re$ in that satisfy the hypotheses of Proposition \[prop:internal\_dynamics\_general\_properties\]. Hence, in particular, the reduced dynamics have an exponentially stable closed orbit.
[The integral $(1/L)\int_0^T \tilde \kappa(\tau) \|\tilde
\sigma'(\tau)\| d\tau$ is equal to ([*turning number*]{} of $\C)
\times 2\pi/{L}$. [The turning number of $\C$ is the number of counterclockwise revolutions that the tangent vector to $\C$ makes as its base point is moved once around $\C$ in a way consistent with the orientation of $\C$. The turning number of a Jordan curve is $\pm 1$, and for curves satisfying Assumption \[assump\_convex\] it is always $1$. Thus, condition can be written as $L > 2 \pi
(b^2+h^2)/h$. It requires the curve to be sufficiently long as compared to the bicycle parameters $b$ and $h$.]{}]{}
\[rem:not\_EL\] [Proposition \[prop:internal\_dynamics\_bicycle\] implies that when a VHC $\varphi = \Phi(s)$ with $\image(\Phi) \subset (0,\pi/2)$ is enforced on the Lagrangian control system , the resulting reduced dynamics are not Lagrangian.]{}
The VHC $\varphi = \Phi(s)$ arising from Proposition \[prop:generator:periodic\_solutions\], part (ii), satisfies $\Phi(s) \in (0, \pi/2)$ for all $s\in [\Re]_T$, so that $\sin \Phi(s) >0$. Recall that $\Phi$ satisfies with $\delta(s) =\epsilon \mu(s)>0$ for all $s \in [\Re]_T$. This implies that $\Psi_1(s)>0$. Moreover, $\Phi'' = \delta' -r \kappa'
\|\sigma'\| c_\Phi - (r \kappa\|\sigma'\|)^2 s_\Phi c_\Phi +
r \kappa \|\sigma'\|\delta s_\Phi -r \kappa {\sigma'}\trans
\sigma'' / \|\sigma'\|$. Substituting this expression in , we get $$\begin{aligned}
\Psi_2 &= -\frac{\delta'}{\delta} - \frac {\kappa\|\sigma'\|} {\delta
h} \Big[ b \delta s_\Phi - \|\sigma'\|c_\Phi \big( \kappa s_\Phi
\frac{b^2+h^2} h -1 \big) \Big] \\
&\leq -\frac{\delta'}{\delta} + \frac{\kappa\|\sigma'\|^2} {\delta h}
c_\Phi \big( \kappa s_\Phi \frac{b^2+h^2} h -1 \big) .
\end{aligned}$$ Since $\int_0^T \tilde \delta'(\tau)/\tilde \delta(\tau) = \ln \tilde
\delta(T) - \ln \tilde \delta(0) =0$, using the fact that $\image(\tilde \Phi) \subset (\Phi^-,\Phi^+) \subset (0,\pi/2)$, we have $$\begin{aligned}
&\int_0^T \tilde \Psi_2(\tau) d\tau \leq \int_0^T \frac{\tilde
\kappa\|\tilde \sigma'\|^2} {\tilde \delta h} c_{\tilde \Phi}
\left( \tilde \kappa s_{\tilde \Phi} \frac{b^2+h^2} h -1 \right)
d\tau \\
&\leq \max_t \left(\frac{\tilde \kappa\|\tilde \sigma'\|} {\tilde
\delta h}\right) c_{\Phi^-} \int_0^T \left(\tilde \kappa
\frac{b^2+h^2} h -1 \right) \|\tilde \sigma'\| d\tau \\
&\leq \max_t \left(\frac{\tilde \kappa\|\tilde \sigma'\|^2} {\tilde
\delta h}\right) c_{\Phi^-} \left( \frac{b^2+h^2} h \int_0^T
\tilde \kappa \|\tilde \sigma'\| d\tau -L \right)\\
&<0.
\end{aligned}$$
We return to the ellipse of Example \[ex:ellipse\] and the VHC displayed in Figure \[fig:ellipse:virtual\_constraint\]. For this example, $(1/2\pi)\int_0^{2\pi} \tilde \kappa(\tau) d\tau \approx 0.0792$, and $h/(b^2+h^2)=0.6711$ and thus is satisfied. Indeed, one can numerically check that $\int_0^{2\pi} \tilde
\Psi_2(\tau) d\tau \approx -105.1<0$, and Proposition \[prop:internal\_dynamics\_general\_properties\] applies. The phase portrait of the dynamics on the constraint manifold is displayed in Figure \[fig:ellipse:phase\_portrait\]. The figure illustrates the set $\cR$, corresponding to the steady-state velocity profile of the bicycle on $\Gamma$. The domain of attraction of $\cR$, shaded in the figure, is the set $\{(s,\dot s):
\dot s > -\nu(s)\}$, as pointed out in Remark \[rem:doa\].
![Phase portrait of the dynamics on $\Gamma$ and set $\cR$ for the ellipse in Example \[ex:ellipse\]. The shaded region is the domain of attraction of $\cR$. Since $s \in [\Re]_{2\pi}$, the $s$ axis wraps around, and points on the lines $s = 0$ and $s = 2\pi$ are identified in the figure, from which it follows that $\cR$ is a closed orbit.[]{data-label="fig:ellipse:phase_portrait"}](FIGURES/ellipse_phase_portrait){width=".5\textwidth"}
[Let now $\C$ be a circle of radius $R$ with parametrization $\sigma(s) =(R \cos s,R \sin s)$. We have $T=2\pi$, $\|\sigma'(s)\|
\equiv R$ and $\kappa(s) \equiv 1/R$, and so $a(s) \equiv
bh^{-1}$. For any $x_0 \in (0,\pi/2)$, picking $\tilde \delta= bh^{-1}
\cos x_0$, as in Remark \[rem:mu\], we obtain the constant VHC $\varphi=x_0$. Equation becomes $$\frac{dz}{dx} =\frac{2g}{b} \tan x_0 - \frac{2}{bR} ( 1-
(h/R)\sin x_0) z.$$ The above is a linear time-invariant system with constant input which is stable if $R > h \sin x_0$. The periodic solution $\bar z(t)$ in this case is simply the equilibrium of the system above, $\bar z= g
R^2 \tan x_0 / (R - h \sin x_0) $, and thus the asymptotic velocity of the bicycle on $\Gamma$ is constant, and reads as $\nu = R \sqrt{g
\tan x_0 / (R - h \sin \Phi_0)}$. It can be verified that $\nu$ is an increasing function of $x_0$. The conclusion is that the bicycle can travel around the circle with any constant roll angle in the interval $(0,\pi/2)$. In steady-state, its speed is constant. The larger is the roll angle $x_0$, the higher is the asymptotic speed of the bicycle. ]{}
Solution of the maneuvering problem {#sec:solution}
===================================
\[thm:solution\] Suppose that the curvature of $\C$ satisfies inequality . If $\varphi=\Phi(s)$ is a VHC such that $\Phi(s) \in
(0,\pi/2)$ for all $ s \in [\Re]_T$, then the feedback $$\label{eq:stabilizing_feedback}
\begin{aligned}
u& =\frac{1}{\Delta(q)}\Big( \frac 1 h g {s_\varphi}- \Big( \Phi'' + \frac 1
h\Big( (1-h \kappa {s_\varphi}) \kappa \|\sigma'\|\\
&+ b \kappa'+\frac{b \kappa {\sigma'}\trans \sigma''}{\|\sigma'\|^2}\Big) {c_\varphi}\|\sigma'\|\Big) \dot s^2 + K_1 e + K_2 \dot e\Big),
\end{aligned}$$ where $\Delta(q) = \Phi'(s) + a(s) {c_\varphi}$, $e= \varphi - \Phi(s)$, $\dot e =\dot \varphi - \Phi'(s)\dot s$, and $K_1$, $K_2$ are positive design parameters, solves the maneuvering problem and has the following properties:
(i) The constraint manifold $\Gamma$ is invariant and locally exponentially stable for the closed-loop system , .
(ii) There exists a $C^1$ function $\nu: [\Re]_T
\to (0,\infty)$ such that the closed orbit $\bar \cR =
\{(q,\dot q) \in \Gamma: \dot s=\nu(s)\}$ is asymptotically stable for the closed-loop system, and its domain of attraction is a neighbourhood of the set $\{(q,\dot q) \in \Gamma : \dot s>0\}$.
(iii) For initial conditions in the domain of attraction of $\bar
\cR$, the rear wheel of the bicycle traverses the entire curve $\C$, and its speed is periodic in steady-state.
[Control law (\[eq:stabilizing\_feedback\]) is [an input-output linearizing feedback for system with output $H(\varphi,s)=\varphi-\Phi(s)$. Letting $e= \varphi - \Phi(s)$, the closed-loop system satisfies $\ddot{e}=-K_1 e - K_2 \dot e$, and it exponentially stabilizes $\Gamma = \{(q,\dot q): e=\dot
e=0\}$. Substituting $e=\dot e=0$ in (\[eq:stabilizing\_feedback\]) we recover the feedback in (\[eqn\_inv\_feedback\]) rendering $\Gamma$ invariant.]{}]{}
The map $(\varphi,\dot \varphi,s,\dot s) \mapsto (e,\dot e,s,\dot s)$ is a diffeomorphism, and the image of $\Gamma$ under this map is the set $\tilde \Gamma :=\{(e,\dot e,s,\dot s): e=\dot e=0\}$. The feedback is smooth in a neighbourhood of $\Gamma$, and it yields $\ddot e + K_1 e + K_2 =0$, with $K_1,K_2>0$. Therefore, the set $\tilde \Gamma$ is exponentially stable, implying that $\Gamma$ is exponentially stable as well, and proving part (i). As for part (ii), by Propositions \[prop:internal\_dynamics\_general\_properties\] and \[prop:internal\_dynamics\_bicycle\], there exists a smooth function $\nu:[\Re]_T \to (0,\infty)$ such that the closed orbit $\cR
= \{(s,\dot s): \dot s = \nu(s)\}$ is exponentially stable for system . Therefore, $\bar \cR$ is asymptotically stable relative to $\Gamma$ (i.e., asymptotically stable when the initial conditions are restricted to lie on $\Gamma$). In order to prove that $\bar \cR$ is asymptotically stable for initial conditions in a neighbourhood of $\Gamma$, note that $\bar
\cR$ is a closed curve, and hence a compact set. Owing to the reduction principle for asymptotic stability of compact sets (see [@SeiFlo95] and [@ElHMag13 Theorem 10]), the asymptotic stability of $\bar \cR$ relative to $\Gamma$, together with the asymptotic stability of $\Gamma$, imply that $\bar \cR$ is asymptotically stable for . By Propositions \[prop:internal\_dynamics\_general\_properties\] and \[prop:internal\_dynamics\_bicycle\], its domain of attraction contains the set $\{(q,\dot q) \in \Gamma : \dot s>0\}$. Since the domain of attraction of a closed set is an open set, the domain of attraction of $\Gamma$ is a neighbourhood of $\{(q,\dot q) \in \Gamma
: \dot s>0\}$. Finally, concerning part (iii), for all $(q,\dot q)
\in\bar \cR$ we have $\dot s = \nu(s)>0$. Hence, for all initial conditions in the domain of attraction of $\bar \cR$ there exist $\bar
t>0$ and $\epsilon >0$ such that $\dot s(t)>\epsilon>0$ for all $ t
\geq \bar t$, and hence the bicycle traverses the entire curve $\C$. Since $\bar \cR$ is diffeomorphic to $S^1$, since it is asymptotically stable, and solutions originating on it are periodic, $\bar \cR$ is a stable limit cycle of the closed-loop system. Therefore, solutions in the domain of attraction of $\bar \cR$ converge asymptotically to a periodic orbit.
Numerical implementation {#sec:numerics}
========================
To implement the solution presented in Theorem \[thm:solution\] one needs an analytical expression of the function $\Phi: [\Re]_T \to
\Re$. This function is found through numerical integration of the virtual constraint generator using a numerical procedure. We will now outline this procedure and make informal deductions about the impact of the approximation. We begin our design by picking a function $\tilde \mu:\Re \to (0,\infty)$ according to Proposition \[prop:generator:periodic\_solutions\], part (ii), and an initial condition $x_0$. Using a one-dimensional search, we find the unique value of $\epsilon$ such that the solution of is $T$-periodic with a desired accuracy. The function $\Phi:[\Re]_T \to \Re$ is determined as a spline interpolation from the samples of the solution of over a period. With a spline interpolation at hand, one can compute $\Phi', \Phi''$ analytically and implement . The numerical approximation process used to find $\Phi$ introduces a bounded error in the feedback controller which can be made arbitrarily small. The effect of this error is to perturb the constraint manifold $\Gamma$ without affecting its exponential stability. The dynamics on the perturbed manifold are still governed by , where the functions $\Psi_1,\Psi_2$ are affected by small perturbations. Such perturbations have no effect on the hypotheses of Proposition \[prop:internal\_dynamics\_general\_properties\], because they involve strict inequalities, $\Psi_1 >0$, $\int_0^T \tilde \Psi_2
d \tau <0$. Thus, the conclusion of Theorem \[thm:solution\] still holds. The approximation of $\Phi$ perturbs the constraint manifold $\Gamma$ and the asymptotically stable limit cycle $\bar \cR$.
We return to example of the ellipse, with the VHC depicted in Figure \[fig:ellipse:virtual\_constraint\]. The simulation results for the closed-loop system with controller and $K_1=5$, $K_2=2$ are shown in Figures \[fig:ellipse:closed\_loop2\], \[fig:ellipse:closed\_loop1\] for the initial condition $(\varphi(0),\dot \varphi(0),s(0),\dot
s(0)) = (0.1,0,0,0.2)$. Figure \[fig:ellipse:closed\_loop2\] illustrates the exponential convergence of $\varphi(t)- \Phi(s(t))$ to zero. Figure \[fig:ellipse:closed\_loop1\] displays the projection of the phase curve on the $(s,\dot s)$ plane and its convergence to the submanifold $\cR$.
![Simulation of the closed-loop system for the ellipse example. The solution converges to the constraint manifold $\Gamma$.[]{data-label="fig:ellipse:closed_loop2"}](FIGURES/ellipse_closed_loop2bis){width=".5\textwidth"}
![Simulation of the closed-loop system for the ellipse example. The solution converges to the submanifold $\cR \subset \Gamma$.[]{data-label="fig:ellipse:closed_loop1"}](FIGURES/ellipse_closed_loop1){width=".5\textwidth"}
\[example\_non\_convex\]
In this second example, we consider the $2 \pi$-periodic curve shown in Figure \[fig:non\_convex\], parametrized as $(5+1.5 \cos
(2 t)) (\cos t, \sin t)$, and then reparameterized with respect to the arc-length $s$. The curve has length $L=39.129$. Since $\sigma:[\Re]_L \to \Re^2$ is not convex, Assumption \[assump\_convex\] is not satisfied and the results presented in the paper cannot be applied directly. Nevertheless, the maneuvering problem can still be solved applying the VHC method. For the virtual constraint generator (\[eq:constraint\_generator\]), we pick $\tilde \delta(t)\equiv \epsilon$, and choose $x(0)=0.35$. Numerically, we find that setting $\epsilon=0.1194$ the solution of is $L$-periodic, thus giving rise to a valid VHC.
![The non-convex curve considered in example \[example\_non\_convex\].[]{data-label="fig:non_convex"}](FIGURES/non_convex){width=".5\textwidth"}
Again, the simulation results for the closed-loop system with controller and $K_1=5$, $K_2=2$ are shown in Figures \[fig:non\_convex:closed\_loop2\], \[fig:non\_convex:closed\_loop1\] for the initial condition $(\varphi(0),\dot \varphi(0),s(0),\dot s(0))
= (0,0,0,2)$.
![Simulation of the closed-loop system for the curve considered in example \[example\_non\_convex\]. The solution converges to the constraint manifold $\Gamma$.[]{data-label="fig:non_convex:closed_loop2"}](FIGURES/non_convex_closed_loop2bis){width=".5\textwidth"}
![Simulation of the closed-loop system for the curve considered in example \[example\_non\_convex\]. The solution converges to the submanifold $\cR \subset \Gamma$.[]{data-label="fig:non_convex:closed_loop1"}](FIGURES/non_convex_closed_loop1bis){width=".5\textwidth"}
Acknowledgements {#acknowledgements .unnumbered}
================
We are grateful to Alireza Mohammadi for his helpful comments on this paper.
[^1]: L. Consolini is with the Department of Information Engineering, University of Parma, Viale Usberti 181/A, Parma, 43124 Italy. E-mail: [[email protected]]{}. M. Maggiore is with the Department of Electrical and Computer Engineering, University of Toronto, 10 King’s College Road, Toronto, Ontario, M5S 3G4, Canada. E-mail: [[email protected]]{}.
[^2]: An integral of motion of , or first integral, is a real-valued function of the state whose value is constant along all solutions of (see, e.g., [@Har02]). The function $I(s,\dot s)$ is only constant along the solution of with initial condition $(s(0),\dot
s(0)) = (s_0,\dot s_0)$. It is not constant along other solutions, because the Lie derivative of the function $I(s,\dot s)$ along the vector field of is not zero. In our previous work [@MagCon13] we have shown that if $M$ and $V$ are well-defined, then the correct integral of motion of is $E(s,\dot s) = (1/2) M(s) \dot
s^2 + V(s)$.
[^3]: This substitution is standard. See, for instance, [@Pol03 Section 2.9.3-2, item 25]. It is also used in [@Shiriaev2006900].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Hendrik Ranocha
date: 28th August 2017
title: Comparison of some Entropy Conservative Numerical Fluxes for the Euler Equations
---
Acknowledgements {#acknowledgements .unnumbered}
================
The author would like to thank the anonymous reviewers for their helpful comments.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The strong evidence of new physics coming from atmospheric neutrino experiments has motivated a series of critical studies to test the robustness of the available flux calculations. In view of a more precise determination of the parameters of new physics, new and more refined flux calculations are in progress. Here we review the most important sources of theoretical uncertainties which affect these computations, and the attempts currently under way to improve them.'
author:
- 'G. Battistoni'
title: Uncertainties on Atmospheric Neutrino Flux Calculations
---
=-0.3cm
2[m\^2]{} 2[sin\^2(2)]{}
\
Introduction {#sec:introduction}
============
The evidence for new neutrino physics beyond the standard model, as is emerging from the results of Super–Kamiokande[@sk] and other atmospheric neutrino experiments[@macro; @soudan2], is now considered robust. Soon after the first announcement of Super–Kamiokande in 1998, many efforts have been devoted to the examinations of theoretical uncertainties in the knowledge of atmospheric neutrino fluxes. None of these sources of uncertainty resulted so critical to vanish the crucial qualitative feature of atmospheric neutrinos: the up–down symmetry of fluxes in absence of oscillations (or other possible mechanisms invoked to explain the observed “anomaly”). Even in absence of oscillations, there exist recognized violations of this absolute symmetry, as those due to the geomagnetic cutoff of primary cosmic rays, but they can be treated as additional corrections. These perturbations are found to be significant mostly in the Sub-GeV region. A new phase of the experimentation on atmospheric $\nu$’s has started, with the primary goal to improve and constraint as much as possible the parameters of the proposed new physics (essentially $sin^2 2\Theta_{atm}$ and $\Delta m^2_{atm}$ for the 2–family oscillation scenario). This is not only a goal for the existing experiments, but also the motivation for the proposal of new generation high precision detectors, such as ICARUS[@icarus]. For this purpose, the theoretical error has to be reduced as much as possible and new refined calculations are necessary. As a first step we need to improve our quantitative understanding of all the factors which affect fundamental quantities like symmetry, flavor ratio, the absolute flux value, the spectral index and the details of angular distributions. In the following sections we intend to review the major sources of uncertainties, with some emphasis on the hadronic interaction sector. The work is still in progress, as outlined in the conclusions.
The status of present flux calculations {#sec:present}
=======================================
The present relevant experiments are making reference, for their analyses, mainly to the neutrino flux calculations from Honda et al. (HKKM)[@honda] and the Bartol group[@bartol]. These works have been recognized as the most accurate and their authors introduced for the first time important ingredients in the simulation. For instance the back–tracing technique for the evaluation of geomagnetic cutoff[@honda] and the effects of muon polarization[@bartol]. These calculations have in common the 1-dimensional calculation approach, in which all secondary particles in the showers, neutrinos included, are considered collinear with the primary cosmic rays. This has been found to be a non correct approximation, at least in principle. A new set of calculation, based on the full FLUKA simulation code[@fluka], has been recently presented[@3d]. There it has been realized that a correct 3-dimensional approach in the earth’s spherical geometry leads to different results for the angular distribution at low energy. A more didactic explanation of this is given in ref.[@lipari-geo]. However, the new FLUKA calculation was essentially motivated by the emerging need of a more accurate description of particle production in hadron and nuclear interactions. In fact, there are reasons to consider more critically the standard references. For instance, it has been noticed how they obtain very close final results for the $\nu$-fluxes, although starting from different particle production models and different primary spectra. Another calculation appeared with a reference to FLUKA[@walt], but there the authors made use of one the hadronic interfaces extracted from the GEANT package v.3.21 called FLUKA. Actually, that package is only a limited and obsolete part of the hadronic model contained in the real FLUKA code used for the present work. It gives results which can well be different and less reliable when compared with experimental data with respect to FLUKA. Significant differences exist at all energies, but they are particularly striking for hadron energies below a few GeV.
Sources of uncertainty {#sec:uncertainty}
======================
Attempting a review of all possible sources of systematic uncertainties in flux calculations, the following items must be considered: primary spectra (fluxes, nuclear component, isotropy and its breaking), geomagnetic description, atmosphere models, the geometry of calculations, other minor details in the modeling (detector altitude, mountain profiles, etc. ) and particle production in hadronic interactions. In the recent past there have been other discussion (at least in part) of these topics. Beyond the already quoted references [@3d; @lipari-geo], important discussions are also in [@lipari-ven99; @lipari-ew; @bartol-had; @gaisser-comp].
Primary Spectrum {#sec:uncertainty.pri}
----------------
On of the most relevant achievements in the measurement of primary cosmic ray spectra is the fact that BESS[@bess] and AMS[@ams] particularly succeeded in producing results in very good agreement one to the other, in particular for the proton component. The scientific community has the attitude of considering these last results as the most reliable and therefore the uncertainty of the primary flux value is probably smaller with respect to the estimates of few years ago, although one should not forget other different data sets, like those of CAPRICE[@caprice] until the topic is definitively settled down. A summary of some of the recent proton measurements is shown in Fig. \[fig:pr\_spec\], together with the lines showing the solar minimum fits used in [@honda] and [@bartol]. It has to be noticed how the input primary spectrum used by Bartol (and later by FLUKA) is in very good agreement with BESS data, while HKKM made use of a parameterization (based on the old compilation of ref.[@webber]) which has a significantly higher normalization above 20$\div$30 GeV. It is therefore natural to ask how the eventual result from the HKKM calculation would change if they used the same input spectrum as Bartol. This is one of the reasons why it is important to analyze in depth the relevance of the particle production model.
As far as the Helium component is concerned, the latest AMS and BESS measurements are now converging[@bess2; @ams2]. The heavier nuclei have less relevance for low energy neutrinos; in any case, further results from AMS will hopefully clarify the picture.
The arguments exposed here are relevant for the energy range contributing to contained events in Super–Kamiokande. In case of higher energy neutrinos, like those measured through the detection of up-going muons in MACRO and Super–Kamiokande, the uncertainties on the relevant energy of the spectrum (up to tens of TeV) are still large (up to 20%).
Geomagnetic description {#sec:uncertainty.mag}
-----------------------
The effect of geomagnetic field is recognized as the most important source of up–down symmetry breaking in the atmospheric neutrino flux. At present, the confidence in the accuracy of IGRF models is rather strong[@igrf], and the technique of anti-proton back–tracing is now accepted as the standard procedure to be adopted to evaluate the correct cutoff for primary cosmic rays arriving to the earth. Solar modulation has to be considered as well. Algorithm relating primary flux to data from neutron monitors exist, and at present their quality is considered satisfactory. Recently, there has been discussion about two items: i) the role of recirculating sub-cutoff particles as pointed out by AMS data, and ii) the anysotropy (far from earth) related to solar wind effects at the GeV scale. In both cases it can be demonstrated (see [@lipari-nu2000]) that both phenomena are of small relevance, since at most they affect neutrino rates by less than 1%.
Geometry of calculations {#sec:uncertainty.geo}
------------------------
As already reported in the introduction, one of the most interesting outcomes in last two years is the realization of the importance of 3–Dimensional computations in a spherical geometry[@3d; @lipari-geo]. In summary, the net eventual result of the correct geometrical description of neutrino production around the earth is a modification of both angular distribution and of normalization in the Sub-GeV region (where. $<\theta_{\nu - p}>$ is significant), with respect to the collinear approximation. The impact of this on the determination of oscillation parameters is still under study. In fact, a final reliable 3–Dimensional calculation of atmospheric neutrino flux is still missing. The computations by the FLUKA group, reported in [@flukanu], must still be considered as preliminary, since bending of charged particles in the geomagnetic fiels has not yet been introduced. As discussed in [@lipari-ew], this effect should play a non negligible role. However, a detailed calculation on the whole earth sphere taking into account also the whole B–field map introduces technical complexities in the computation, since spherical symmetry is lost. In fact, in absence of B–field, any point on the earth’s surface is equivalent to another, and this allows to make use all generated events even for a specifc detector location, provided that a proper rotation of trajectory parameters to the geografical coordinates of interest is performed. This problem has not yet been solved completely. In [@lipari-ew] a simplified solution was proposed for the first studies, while the FLUKA group is designing a new dedicated simulation in which specific weighting algorithms have to be introduced.
Atmosphere Description and other details {#sec:uncertainty.atm}
----------------------------------------
It is practically impossible to introduce a realistic description of the atmosphere all around the earth, valid at all altitudes and for all weather conditions. This remains an irreducible source of systematics, although probably small. The fact that neutrino experiments last a considerable time and detect neutrinos produced all over the earth positions, gives some confidence on the essential validity of average atmosphere models. A comparison between the performance of different codes, carried on by the Bartol and FLUKA groups, is showing that having considered, or not, ingredients in particle transport like energy loss fluctuation, multiple scattering, etc., may affect the final results, in normalization and flavor ratio, at the level of percent. Other important inputs in the simulation for a specific detector site concern the introduction of detector altitude and possible rock overburden. This last element, for instance, can introduce a difference in the $\nu_e/\nu_\mu$ ratio, again at the level of some percent.
Particle Production in Hadronic Interactions {#sec:uncertainty.had}
--------------------------------------------
For given shower model and primary spectrum, the most important source of theoretical uncertainty comes from the hadronic interaction model. Since QCD does not allow to compute the bulk properties of particle production in the non perturbative regime, the available models are based either on phenomenological models, possibly inspired by partonic concepts, constrained by accelerator data, or directly on the parameterizations of these experimental results. As an example, the Bartol and FLUKA groups have used the same input primary spectrum and, with different models, have obtained different fluxes. The current comparison of angle integrated fluxes is reported (for the Super–Kamiokande site) in Fig.\[fig:fluxes\], where also the differences between the 1D and 3D approach are shown for FLUKA.
The two groups have started a comparison of their hadronic models (FLUKA and TARGET), both in single interaction and within the same shower code, in order to understand the impact of different choices. For the time being, this comparison is limited to an energy region useful for contained and partially contained events in Super–Kamiokande. The two codes are constructed in very different ways, and the net final result is that there is a $\sim$20% asymptotic difference in the neutrino flux normalization from the two models. In reality the difference is energy dependent: it is larger and reversed at low energy. The resulting spectral index is somewhat different, FLUKA being a little harder than the Bartol one. A comprehensive discussion of the matter should requires a dedicated paper, and here we can only summarize a few of the crucial conclusions.
As far as neutrinos up to few tens of GeV are concerned, pion production in nucleon–Nucleus interaction is the process that mostly contributes to the yield. For known kinematical reasons, Kaon production becomes relevant at higher energy, for instance in the region from which upgoing through-muons as detected by MACRO and Super–Kamiokande are produced. At first order, we are interested in both total multiplicity and in the shape of the energy fraction distribution of these secondary particles, or of Feynman–$X$ and similar longitudinal variables. Unfortunately, although there exist reports on the total charged multiplicity, mostly obtained in emulsion experiments[@fred], there are not enough data on the $X$ distribution of pions in the interactions with light nuclei. At present the most valuable data set are those of ref. [@eichten] and [@abbot], covering different region of phase space. A direct comparison of these data with the predictions from FLUKA and TARGET shows that, very probably, the Bartol model produces too many pions at low $X$. This difference directly reflect in the neutrino yield. The average number of muon neutrinos for vertical proton showers, as a function of primary energy, is shown in Fig.\[fig:yield\] for the two models. It must be remembered that from the experimental point of view, event rates should be considered, that is after the convolution with neutrino interaction cross sections.
Although the present data do not allow to give a reliable estimate of the overall systematic error associated to hadronic interactions, in the author’s opinion the existing difference between FLUKA and TARGET ($\sim$ 20%) should not be considered as a measurement of the real error. The different capability of TARGET and FLUKA model to reproduce accelerator data suggests that the actual theoretical uncertainty is likely to be definitively smaller than 20%.
A preliminary comparison, at the level of single interaction features, with the models adopted in HKKM calculations, gives indications that, if they had used the same primary spectrum of Bartol and FLUKA, also in their case the normalization would have been lower. The yield difference between FLUKA and TARGET is larger at lower nucleon energy. This has some direct consequence in the prediction of up-down symmetry of fluxes in different geographical sites. For instance, at Super–Kamiokande, where the cutoff from the above direction is rather high (around 10 GeV), the yield enhancement of TARGET is ineffective, and the FLUKA/TARGET ratio in the Sub-GeV region is reversed (see also Fig.\[fig:fluxes\]) with respect to the situation of Soudan, which is instead a low–cutoff site. Therefore, since the analysis of Soudan data are based upon the Bartol predictions, they should expect a lower asymmetry in the Sub-GeV region: see Fig.\[fig:asym\]
-1. cm
Another way of looking at this is given in Fig.\[fig:asym2\], where the ratio of FLUKA (1D) to TARGET fluxes are shown as a function of energy for the different laboratories.
[cc]{} &\
\
-1. cm
The two interaction models also differ in the $\pi^+/\pi^-$ ratio, which affects the $\nu/\nubar$ ratio. Due to the different interaction cross section this is a significant parameter. Considering the proper weighting with the $X$ distribution, FLUKA has a larger $\pi^+/\pi^-$ ratio by an amount which is around 7% for protons at 10 GeV. For increasing energy the difference becomes smaller, as expected. FLUKA is able to satisfactorily reproduce the charge ratio measured by many experiments[@ferrari].
We have found instead that the $({{\nu_e}}+ {\bar\nu_e})/({{\nu_\mu}}+ {\bar\nu_\mu})$ ratio has not a relevant dependence on the hadronic interaction model.
Oscillation analysis of neutrino events requires the knowledge of production height, which depends on the longitudinal development of showers. This, on turn, depends on the point of first interaction, determined by total inelastic cross sections, and by the following development driven by the energy fraction carried away by leading nucleons in each interaction. Total cross sections are eventually determined by nucleon–nucleon cross sections, which are well constrained by existing data compilation, to which all groups make strong reference. Of course, the energy fraction taken by pions (including $\pi_0$, feeding the e.m. component of showers) is not independent from that taken by nucleons. FLUKA and TARGET have different elasticities and therefore give rise to some differences in the longitudinal development of cascades. Again, the comparison with existing data reinforces some confidence on the FLUKA model. More complete experimental data on particle production on light nuclei would be fundamental to minimize the theoretical uncertainties and constraint the existing models. In order to be significant for this purpose, a new experiment must explore a range of beam energies from few GeV up to at least 30 $\div$ 50 GeV, with targets of different atomic number, from Be up to at least Al, in order to study the dependence on the number of elementary collisions (which in the Glauber approach scales as $A^{1/3}$[@glauber]). Secondary particles must be measured in a wide solid angle to cover as much as possible the available phase space. For these reasons the scientific community welcomes the HARP experiment[@harp], proposed to perform a dedicated study on these subjects.
Conclusions {#sec:conclusions}
===========
The current understanding of the sources of uncertainty in atmospheric neutrino flux calculations is still improving. A better confidence on the primary cosmic ray spectra is the first necessary condition. After that, the most important source of uncertainty is that due to the particle production model. There is discussion upon which kind of experiment or study can help in achieving better constraining of simulations. Data on muon fluxes in atmosphere can help, and they are a useful benchmark tool. However, the connection between muon and neutrino yield at different altitudes and energy is still a rather indirect process, and it is not yet clear if muon balloon experiments can guarantee a level of systematics below $\sim$10%. In the author’s opinion, the impact of new data from accelerators could be more relevant on model building, provided that in this case systematics is kept under better control. In this will be the case, there are reason to believe that, although the theoretical error cannot be erased, it could be reduced to the 10% level or even less. However, one of the most serious problem could stay not only in the flux calculations, but also in the knowledge of absolute neutrino cross sections with nuclei, expecially at low energy for quasi–elastic scattering and resonance production. As a matter of fact, the experimentalists are more interested in the eventual event rates than in the flux itself. There is the serious risk that there will be no experimental answer to clarify this aspect.
Other improvements in these computations are needed, but they will require more and more efforts. As an example, the FLUKA group is planning to make use of further developments in particle production models. In particular we are aiming to study the effect of nuclear projectiles instead of recurring to the usual incoherent addition of nucleons (superposition model) adopted so far also in the other standard references. The impact of more precise calculations on the measurement of the parameters of “new physics” is still to be understood in detail. In our opinion, this can be reliably done only introducing the correct simulations of the actual experiments, since detector sources of systematics and resolutions are likely to be of an importance comparable to that of theoretical factors.
Acknowledgments {#sec:acknowledgments}
===============
The author is indebted to the other authors of the FLUKA–neutrino calculation, A. Ferrari, T. Montaruli, and. P.R. Sala, for the help received in preparing this review. This work has been made possible thanks to the collaboration with the Bartol group: R. Engel, T.K. Gaisser, P. Lipari and T. Stanev. They have provided the TARGET model and a lot of essential data and suggestions.
[9]{} Y. Fukuda et al. (Super–Kamiokande Coll.), Phys. Rev. Lett. [**81**]{} (1998) 1562; T. Kajita, Proc. of NOW2000, Sept. 2000, Otranto, Italy. M. Ambrosio et al. (MACRO Coll.), Phys. Lett. [**B434**]{} (1998) 451; F. Ronga, Proc. of NOW2000, Sept. 2000, Otranto, Italy. W.W.M. Allison et al. (Soudan-2 Coll.), Phys. Lett. [**B449**]{} (1998) 137; A. Mann (Soudan-2 Collab.), Proc. of Neutrino 2000 Conference, Sudbury (Canada), June 16, 2000. F. Arneodo et al. (ICARUS and NOE Coll.), LNGS-P21/99, INFN/AE-99-17, CERN/SPSC 99-25, SPSC/P314; A. Rubbia (ICARUS Coll.), hep-ex/0001052. M. Honda et al., Phys. Lett. [**B 248**]{} (1990) 193. G. Barr, T.K. Gaisser and T. Stanev, Phys. Rev. [**D 39**]{} (1989) 3532; V. Agrawal, et al., Phys.Rev. [**D53**]{}, 1314 (1996). A. Fass' o, A. Ferrari, J. Ranft and P.R. Sala. See http://www.cern.ch/fluka and references therein. G. Battistoni et al., Astrop. Phys. [**12**]{} (2000) 315. P. Lipari, Astrop. Phys. [**14**]{} (2000) 153. Y. Tserkovnyak et al., hep-ph/9907450. P. Lipari, Proc. of the VIII Int. Workshop on Neutrino Telescopes, Venezia, February 1999; also in hep-ph/9905506. P. Lipari, Astrop. Phys. [**14**]{} (2000) 188. R. Engel et al., Phys. Lett. [**472**]{} (2000) 113. T.K. Gaisser et al., Phys. Rev. [**D54**]{} (1996) 5578. T. Sanuky, astro-ph/0002481, to appear in Ap.J. J. Alcaraz et al, Phys. Lett. [**B490**]{} (2000) 27, Phys. Lett. [**472**]{} (2000) 215. M. Boezio at al., Astrophys. Journal [**429**]{} (1994) 736. W.R. Webber, R.L. Golden & S.A. Stephens, Proc. 20th ICRC (Moscow) vol. [**1**]{} (1987) 325. T. Sanuki (BESS Coll.), Proc. of NOW2000, Sept. 2000, Otranto, Italy. B. Bertucci (AMS Coll.), Proc. of NOW2000, Sept. 2000, Otranto, Italy. Fundamental references and models can be found in http://nssdc.gsfc.nasa.gov/. P. Lipari, Proc. of Neutrino 2000 Conference, Sudbury (Canada), June 16, 2000. FLUKA flux tables are available in http://www.mi.infn.it/battist/neutrino.html. S. Fredriksson et al., Phys. Rep. [**144**]{} (1987) 107; A. Tufail et al., Phys. Rev. [**D42**]{} (1990) 2187. T. Eichten et al., Nucl. Phys. [**B44**]{} (1972) 333. T. Abbott et al., Phys, Rev [**D45**]{} No. 11 (1992) 3906. Proc. of the Workshop on Calorimetry, Annecy, October 2000. M.G. Catanesi et al., (HARP Collaboration), CERN-SPSC/99-35, SPSC/P315. R.J. Glauber, Phys. Rev. [**100**]{} (1955) 242; R.J. Glauber and G. Matthiae, Nucl. Phys. [**B21**]{} (1970) 135.
[ ]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We exhibit a model structure on , obtained by transfer from $\sset$ across the adjunction $C_2 \circ Sd^2 \hspace{0.25em}
\dashv \hspace{0.25em} Ex^2 \circ N_2$.
author:
- 'K. Worytkiewicz, K. Hess, P.E.Parent,A.Tonks'
bibliography:
- 'ProcessSemantics.bib'
- 'beke.bib'
- 'benabou.bib'
- 'cat.bib'
- 'catta.bib'
- 'cisinski.bib'
- 'gab.bib'
- 'god.bib'
- 'goub.bib'
- 'gray.bib'
- 'hermida.bib'
- 'kw.bib'
- 'milner.bib'
- 'modelcats.bib'
- 'street.bib'
- 'winskel.bib'
- 'wolff.bib'
title: A Model Structure à la Thomason on
---
Introduction
============
intro
\[sec:2-Categories\]2-Categories and 2-Nerves
=============================================
m2-cat1
\[sec:Model-Category-Theory\]Model Category Theory
==================================================
mmodelcats1
\[sec:modelstr\]A Model Structure à la Thomason
===============================================
mms
\[sec:Homotopy\]Homotopy
========================
homo
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Ito equations are derived for simple stochastic cellular automaton with parameters and compared with results obtained from the histogram method. Good agreement for various parameters supports wide applicability of the Ito equation as a macroscopic model.
[keywords]{}: [stochastic processes, cellular automata, avalanches, discrete solvable models, ultra-discrete equations, time series]{} 02.50.Ey, 05.10.Gg, 05.45.-a, 45.70.Ht;
address:
- 'Z. Czechowski, Institute of Geophysics, Polish Academy of Sciences, ul. Ks. Janusza 64, Warszawa, Poland'
- 'M. Bia[ł]{}ecki, Institute of Geophysics, Polish Academy of Sciences, ul. Ks. Janusza 64, Warszawa, Poland'
author:
- 'Zbigniew Czechowski and Mariusz Bia[ł]{}ecki'
title: Ito equations out of domino cellular automaton with efficiency parameters
---
Introduction
============
The Ito equation describes evolution of a stochastic diffusion Markov process (of order 1). Its parameters can be nonlinear functions of the process. Therefore, the Ito equation is a good candidate for a nonlinear model of phenomenon which manifests non-regular, random behaviour. Quite often observed variables have macroscopic character, and hence the Ito equation can be considered as a macroscopic model of the complex system, in which microscopic collective interactions have been averaged to an adequate form of terms in the equation.
Various complex processes, geophysical, biological, economical etc., in spite of apparent random appearance, can lead to some regular behaviours or patterns. Registered time series are investigated by different, less or more sophisticated methods, which help to find these dependencies. Some interesting regularities may constitute of new macroscopic laws of complex systems.
Time series analysis is well developed branch of science. In the linear case, well known procedures were elaborated (ARMA etc.). However, effective methods of construction of nonlinear models from time series data still are not very satisfactory and require further improvements. Using the Ito equation may provide some progress in the field.
First attempts of determination of the Ito equation were proposed by Haken and Borland (Haken 1988, Borland and Haken 1992a, b, 1993, Borland 1996)). Their procedure (SEQUIN) uses the knowledge of some moments of the joined distribution function, which form constraints for the Maximum Information Principle. The SEQUIN method works sufficiently well for the case of a weak noise. However, it was shown (Rozmarynowska 2009) that the procedure fails in the case of strong multiplicative noise, when long-tail distributions appear.
A purely numerical procedure of construction of the Ito equation from time series data was proposed by Siegert et. al. (1998). This direct procedure, based on the histogram of the joined distribution function, always produces an output (i.e. clouds of points) It gives some approximation of terms in the Ito equation. It is effective for strong noise too, but fitting proper functions to scattered clouds of points is a difficult task. Moreover, it should be underlined that this method may be fallacious, when the time series can not be approximated by a diffusion Markov process. It is possible to determine the Markov order of a given time series (Racca et al. 2007) but there is no method to verify if it is diffusion process or not. The only inspection we may perform is a comparison of the input time series with that generated by the reconstructed Ito equation.
An adequacy of approximation of a natural phenomena by nonlinear models is of a great importance. To touch the topic in the context of the Ito equation we propose as follows. First, we replace the natural phenomenon by the domino cellular automaton, which can be fully monitored. Due to a complete knowledge of the three level hierarchy of the model, relations between the macroscopic (the Ito equation, moment equations) and the microscopic (rules of the automaton) description are clear and understood. Then, we construct the Ito equation in two independent ways. The standard histogram method is compared with results derived analytically. Derivation of the Ito equation out of a cellular automaton is a crucial step in this approach. To this aim a special stochastic cellular automaton with avalanches was introduced (Bia[ł]{}ecki and Czechowski 2010b, Czechowski and Bia[ł]{}ecki 2010a). The defining rules of the domino automaton were chosen to satisfy two opposite requirements: complex behavior and relatively simple mathematical structure. Introducing efficiency parameters in the domino automaton leads to a big diversity of states, which covers wide range of avalanche sizes. The goal of this paper is an analytical derivation of the Ito equations for a domino cellular automaton with efficiency parameters and a comparison with results of the histogram method.
In Section \[sec:domino\] we shortly introduce the domino cellular automaton with efficiency parameters and present equations describing its behaviour in equilibrium state. Properties of the automaton for various efficiency parameters are analysed. In Section \[sec:fluct\] we consider fluctuations around the equilibrium and propose suitable approximation formulas. In Section \[sec:ito\] the Ito equation is derived. The comparison of histogram procedure results and analytical formulas are presented. We conclude with short remarks in Section \[sec:concl\].
Domino cellular automaton with efficiency parameters {#sec:domino}
====================================================
![Domino cellular automaton at the quasi-equilibrium state: left –- time series for density $\rho(i)$, right –- time series for avalanche sizes $w(i)$. Three examples for different cases: $\eta = 4$ (upper diagrams), $\eta = 1$ (middle) and $\eta = 0.25$ (lower).[]{data-label="fig:Fig1"}](fig1a "fig:"){height="4cm" width="12cm"}\
![Domino cellular automaton at the quasi-equilibrium state: left –- time series for density $\rho(i)$, right –- time series for avalanche sizes $w(i)$. Three examples for different cases: $\eta = 4$ (upper diagrams), $\eta = 1$ (middle) and $\eta = 0.25$ (lower).[]{data-label="fig:Fig1"}](fig1b "fig:"){height="4cm" width="12cm"}\
![Domino cellular automaton at the quasi-equilibrium state: left –- time series for density $\rho(i)$, right –- time series for avalanche sizes $w(i)$. Three examples for different cases: $\eta = 4$ (upper diagrams), $\eta = 1$ (middle) and $\eta = 0.25$ (lower).[]{data-label="fig:Fig1"}](fig1c "fig:"){height="4cm" width="12cm"}
We present shortly rules governing the automaton and some relations derived in Bia[ł]{}ecki and Czechowski (2010a, b). Consider 1-dimensional grid of cells. There are only two states of a cell: empty or occupied (by a single ball). The evolution rules are as follows. In a time step a single new ball is added to the system to the randomly chosen cell.
- If the new ball hits an empty cell it becomes occupied with probability $\nu$, or the ball rebounds from the grid with probability $(1-\nu)$.
- If the new ball hits an occupied cell it rebounds with probability $(1 - \mu)$, or with probability $\mu$ it knocks out balls from the cell and other adjacent occupied cells from the whole cluster (they become empty –- an avalanche is triggered).
An important macroscopic variable, which describes the state of the automaton is the density $\rho$ i.e. the rate of number of occupied cells to the grid size. We monitor two time series generated by the cellular automaton: the avalanche sizes $w(i)$ and the density $\rho(i)$ computed after each avalanche. Figure \[fig:Fig1\] presents three examples of such time series for three chosen values of the ratio $\mu/\nu = 4, 1,$ and $0.25$. An increase of the average density, of the average avalanche size and of the range of fluctuations of $\rho$ with decreasing $\mu/\nu$ is evident.
A state of the automaton can be characterized by the cluster size distribution $n_i$, i.e. by numbers of clusters of size $i$ (normalized by the size of the grid). The following relations, which are exact in the equilibrium state, were derived: $$\begin{aligned}
\quad \quad n_1 &=& \frac{1}{\eta+2} (1-\rho) - \frac{2}{3} n + \frac{1}{3}n_1^0 \label{eq:n1}\\
\quad \quad n_2 &=& \frac{2}{2\eta+2} \left( 1- \frac{n_1^0}{n} \right) n_1 \label{eq:n2} \\
\quad \quad n_i &=& \frac{1}{\eta i+2}
\left(
2 n_{i-1} \left( 1- \frac{n_1^0}{n} \right)
+ n_1^0 \sum_{k=1}^{i-2} \frac{n_k n_{i-1-k}}{n^2}
\right) \quad \text{for} \quad i\geq 3. \label{eq:ni}\end{aligned}$$ Here $\eta$ is the rate $\mu / \nu$, $n$ is the number of all clusters, $n= \sum_{i\geq1} n_i$, and $\rho$ is the density of occupied cells, $\rho= \sum_{i\geq1} i n_i$. The $n_1^0$ is the number of single empty cells and is given by the following formula $n_1^0 = \frac{ 2 n}{\left(3 + \frac{2\eta\rho}{n} \right)}.$
The following simple balance equations for moments of $n_i$ were also derived: $$\begin{aligned}
2m_0 + (1+\eta) m_1 &=& 1, \label{eq:moments1} \\
m_1 + \eta m_2 &=& 1, \label{eq:moments2} \end{aligned}$$ where: $m_0 = n$, $m_1 = \rho$ and $m_2=\sum_{i\geq 1} n_i i^2$. The average cluster size $<i>$, the average avalanche size $<w>$ and the average square deviation from the average cluster size $<T>$ (i.e. the analogue of temperature) are expressed by these moments: $$\begin{aligned}
<i> &=& \frac{m_1}{m_0} =\frac{2\rho}{1-(1+\eta)\rho}\label{eq:avri} \\
<w> &=& \frac{m_2}{m_1} = \frac{1-\rho}{\eta\rho} \label{eq:avrw} \\
<T> &=& \sum_{i\geq 1} n_i (i-<i>)^2 = \rho (<w>-<i>)\label{eq:avrT}. \end{aligned}$$ The equilibrium value of the density $\rho_e$ can be found by numerical solution of implicit algebraic equation $$\rho= \sum_{i\geq1} i n_i,
\label{eq:defrho}$$ where $n_i$ are given by equations -.
---------------- ------------ ------------ ------------ ------------
analytical simulation analytical simulation
$\eta=4$ 0.1501 0.1498 1.4147 1.4159
$\eta=1$ 0.3075 0.3068 2.2520 2.2421
$\eta={1}/{4}$ 0.4815 0.4801 4.3074 4.3382
---------------- ------------ ------------ ------------ ------------
: Comparison of values of equilibrium density $\rho_e$ and average avalanche size $<w>$ calculated from equations with those obtained from simulations.[]{data-label="tab:Table1"}
![Dependences of equilibrium variables: $\rho_e$, $n$, $<i>$, $<w>$ and $<T>$ on the ratio of efficiency parameters $\eta =\mu / \nu $.[]{data-label="fig:Fig2"}](fig2){height="10cm" width="12cm"}
Figure \[fig:Fig2\] shows dependencies of equilibrium variables: $\rho_e$, $n$, $<i>$, $<w>$ and $<T>$, on the ratio $\eta$. As it was expected, they decrease with increasing $\eta$, because greater probability $\mu$ (relevant [****]{} to $\nu$) of triggering the avalanche decrease the number of occupied cells on the grid. Only the total number of clusters $n$, has a maximum for $\eta$ around $1/4$. Table \[tab:Table1\] display a comparison of values of $\rho_e$ and $<w>$ calculated from equations with respective values taken from simulation.
Formulas for fluctuations around the equilibrium {#sec:fluct}
================================================
For deviations from the equilibrium state $\rho_e$ formulas from Section \[sec:domino\] are not valid. Therefore, we use the approximation introduced in our parallel paper (Czechowski and Bia[ł]{}ecki, 2010b). It was based on the assumption that the following geometric-like form for $n_i(\rho)$ is maintained: $$\begin{aligned}
n_1(\rho) &=& (1-\rho)^2 a_1(\rho) = (1-\rho)^2 a_1 \rho, \label{eq:an1} \\
n_k(\rho) &=& n_{k-1}(\rho)a_k(\rho) \quad \text{for} \quad 4 \geq k \geq 2, \\
n_k(\rho) &=& n_{k-1}(\rho)(a_5(\rho))^{k-4} \quad \text{for} \quad k \geq 5, \label{eq:ank}\end{aligned}$$ where $a_k(\rho)$, $k = 1, \ldots, 5$ are linear functions of $\rho$. The forms of the formulas - extend the 1-D percolation geometric cluster size distribution, $n_k = \rho^k(1-\rho)^2$. Numerical simulations of the domino automaton confirm a validity of this approximation.
Linear functions $a_k(\rho)$, $k = 2, \ldots, 5$, are found from the Taylor expansion of ratios: $$n_k(\rho)/n_{k-1}(\rho)$$ around $\rho_e$ to the order 1. Using the procedure presented in Czechowski and Bia[ł]{}ecki (2010b) numerical formulas for $a_k(\rho)$, $k = 1,\ldots, 5$, for chosen values of the parameter $\eta$ were found and they are presented in Table \[tab:Table2\]. As a check for these expressions, we we plot the first moment of $n_k(\rho)$ (which should be equal to the density $\rho$) as a function of $\rho$ for chosen parameters $\eta$. Figure \[fig:Fig3\] shows satisfactory fit in the full range of variability of $\rho$ for all three values of $\eta$.
$\eta=4$ $\eta=1$ $\eta=0.25$
------------- ------------------------ ------------------------ -----------------------
$a_1(\rho)$ $0.9528 \rho$ $0.8365 \rho$ $0.7348\rho$
$a_2(\rho)$ $0.6449 \rho + 0.0715$ $0.7032 \rho + 0.1223$ $0.5669\rho + 0.1470$
$a_3(\rho)$ $2.4095 \rho - 0.1846$ $1.4316 \rho - 0.0484$ $1.3742\rho - 0.0546$
$a_4(\rho)$ $0.5163 \rho + 0.0989$ $0.5651 \rho + 0.2273$ $0.4908\rho + 0.3761$
$a_5(\rho)$ $0.5139 \rho + 0.0994$ $0.5569 \rho + 0.2328$ $0.4772\rho + 0.4054$
: Numerical values of functions $a_k$ for expressions -.[]{data-label="tab:Table2"}
![Sum $\sum_{i\geq1} i n_i$ computed by using formulas - versus $\rho$ for three cases: $\eta = 4$ (left graph), $\eta = 1$ (middle), $\eta = 0.25$ (right).[]{data-label="fig:Fig3"}](fig3){height="3.5cm" width="12cm"}
Derivation of the Ito equation {#sec:ito}
==============================
First of all, we should decide which time series, of the density $\rho(i)$ or of the avalanche size $w(i)$, may be treated as a realization of the Markov diffusion process. Because the avalanche time series is more complex and not Markovian of order 1, we use the time series for density. An easy check ensures that automaton rules provide the required 1st order Markov property for $\rho(i)$. We emphasize that the following convention is assumed: the state of the grid (and the size of the avalanche) is monitored after each avalanche only. In this way we avoid less interesting stairs-like increase of $\rho(i)$ in periods between avalanches.
In order to derive the Ito equation $$d \rho = a (\rho) dt + \sqrt{b(\rho)} dW(t),
\label{eq:ito}$$ we need the transition probability function $P(\rho + \Delta\rho, t+ \Delta t \ | \ \rho,t)$ (see Risken 1996). Functions $a(\rho)$ and $b(\rho)$ are known to be the drift and diffusion coefficients respectively, and $W(t)$ is the Wiener process. Following our parallel paper (Czechowski and Bia[ł]{}ecki, 2010b, in which the case $\eta= 1$ was considered) the probability of effective gain $EG(k)$ (an increase of $\rho(i)$ by $k$ occupied cells) and effective loss $EL(k)$ are given by the formulas: $$\begin{aligned}
EG(k) &\equiv& P \left( \rho_i+ \frac{k}{N}, i+1 | \rho_i,i \right) = \nonumber \\
&&= \sum_{s=k+1} (\nu(1-\rho))^s \mu \rho w_{s-k}(\rho) \quad \text{for} \quad k \geq 0, \label{eq:eg}\\
EL(k) &\equiv& P \left( \rho_i - \frac{k}{N}, i+1 | \rho_i,i \right) = \nonumber \\
&&= \sum_{s=k} (\nu(1-\rho))^{s-k} \mu \rho w_{s}(\rho) \quad \text{for} \quad k \geq 1, \label{eq:el}\end{aligned}$$ where $w_s=\frac{n_s s}{\rho}$ is the probability that the occupied cell is a part of the cluster of size $s$ (which also corresponds to the probability of avalanche of size $s$) and $N$ is the grid size.
Balls rebounded off the grid do not trigger avalanches, they set the relative efficiency for triggering an avalanche with respect the efficiency of occupation of an empty cell. Hence, for time series analysis, one can consider only active (not rebounded) balls introducing appropriate corrective coefficient. It follows, in expressions and parameters $\mu$ and $\nu$ should be replaced by effective parameters: $$\begin{aligned}
\mu_e &=& \frac{\mu}{\mu\rho+\nu(1-\rho)} = \frac{\eta}{\eta\rho+1-\rho}, \\
\nu_e &=& \frac{\nu}{\mu\rho+\nu(1-\rho)} = \frac{1}{\eta\rho+1-\rho}. \end{aligned}$$ As a result, probabilities $EG(k)$ and $EL(k)$ satisfy the normalization condition.
The drift and diffusion forces in the Ito equation correspond (see Risken 1996) to the first and the second moment of the transition probability: $$\begin{aligned}
a(\rho) &\propto& \sum_{k\geq1} k \left( EG(k)-EL(k) \right), \\
b(\rho) &\propto& \sum_{k\geq1} k^2 \left( EG(k)+EL(k) \right).\end{aligned}$$
![Illustration of functions $a(\rho)$ and $b(\rho)$ in Ito equations for three values of parameters: $\eta = 4$ (upper diagrams), $\eta = 1$ (middle), $\eta = 0.25$ (lower). Dots represent values reconstructed by the histogram procedure from time series generated by the domino cellular automaton (see Figure \[fig:Fig1\]), lines –- calculated from formulas and using approximation - and parameters presented in Table \[tab:Table2\].[]{data-label="fig:Fig4"}](fig4a "fig:"){height="4cm" width="12cm"}\
![Illustration of functions $a(\rho)$ and $b(\rho)$ in Ito equations for three values of parameters: $\eta = 4$ (upper diagrams), $\eta = 1$ (middle), $\eta = 0.25$ (lower). Dots represent values reconstructed by the histogram procedure from time series generated by the domino cellular automaton (see Figure \[fig:Fig1\]), lines –- calculated from formulas and using approximation - and parameters presented in Table \[tab:Table2\].[]{data-label="fig:Fig4"}](fig4b "fig:"){height="4cm" width="12cm"}\
![Illustration of functions $a(\rho)$ and $b(\rho)$ in Ito equations for three values of parameters: $\eta = 4$ (upper diagrams), $\eta = 1$ (middle), $\eta = 0.25$ (lower). Dots represent values reconstructed by the histogram procedure from time series generated by the domino cellular automaton (see Figure \[fig:Fig1\]), lines –- calculated from formulas and using approximation - and parameters presented in Table \[tab:Table2\].[]{data-label="fig:Fig4"}](fig4c "fig:"){height="4cm" width="12cm"}
Tedious transformations on sums (see Appendix in Czechowski and Bia[ł]{}ecki, 2010b) lead to expressions for $a(\rho)$ and $b(\rho)$ in terms of the first, the second and the third moment of $n_k(\rho)$: $$\begin{aligned}
a(\rho) &\propto& \frac{\nu_e(1-\rho)}{1-\nu_e(1-\rho)} - \frac{1}{\rho} \sum_{k\geq1} k^2 n_k, \label{eq:arhoex}\\
b(\rho) &\propto& \frac{\nu_e(1-\rho)(1+\nu_e(1-\rho))}{(1-\nu_e(1-\rho))^2} + \label{eq:brhoex}\\
&& - \frac{2\nu_e(1-\rho)}{\rho(1-\nu_e(1-\rho))} \sum_{k\geq1} k^2 n_k
+ \frac{1}{\rho} \sum_{k\geq1} k^3 n_k. \nonumber\end{aligned}$$ Note, the dependence on $\nu_e$ is not essential. The above formulas can be also expressed in terms of $\mu_e$; in fact they depends on $\eta$ and $\rho$ only. The second and the third moments itself can be calculated using approximate formulas - from Section \[sec:fluct\].
Conclusions {#sec:concl}
===========
Introduction of efficiency parameters $\mu$ and $\nu$ essentially enriched the automaton. By changing the ratio $\eta=\mu/\nu$ the domino cellular automaton can attain a wide range of values for important variables: $\rho_e$, $n$, $<i>$, $<w>$ and $<T>$ (see Figure \[fig:Fig2\]), which characterize a macroscopic behaviour. The time evolution of the model can be described by the Ito equation, therefore, we derive analytically a satisfactory approximation of Ito terms $a(\rho)$ and $b(\rho)$ as functions of the ratio $\eta=\mu/\nu$. For three chosen values of $\eta$ analytical and simulation results were compared. Figure \[fig:Fig4\] shows a good fit for these three cases. Thus, we have a simple, fully elaborated model which can manifest a wide range of behaviours.
Moreover, we show that the histogram method of reconstruction of Ito equation from time series works well in the case of domino cellular automaton. The automaton plays a role of a complex system, in which energy is slowly accumulated and rapidly released in avalanches. Therefore, this is possible that the histogram procedure can offer an adequate method of nonlinear modelling of similar natural phenomena basing on the time series data. The method fills a gap between linear stochastic models (ARMA etc.) and nonlinear deterministic models (Takens method, see Takens 1981) because it is both stochastic and nonlinear.
Acknowledgement {#acknowledgement .unnumbered}
===============
References {#references .unnumbered}
==========
Bia[ł]{}ecki M. and Czechowski Z. (2010a) [*Analytic approach to stochastic cellular automata: exponential and inverse power distributions out of Random Domino Automaton*]{}, arXiv:1009.4609 \[nlin.CG\]
Bia[ł]{}ecki M. and Czechowski Z. (2010b) [*On a simple stochastic cellular automaton with avalanches: simulation and analytical results*]{}, Chapter 5 in V. De Rubeis, Z. Czechowski, and R. Teisseyre, editors, [*Synchronization and triggering: from fracture to earthquake processes*]{}, Springer, pages 63–75.
Borland L. (1996), [*Simultaneous modelling of nonlinear deterministic and stochastic dynamics*]{}, Physica D, 90, 175-190.
Borland L. and H. Haken (1992a), [*Unbiased determination of forces causing observed processes. The case of additive and weak multiplicative noise*]{}, Z. Phys. B, 88, 95-103.
Borland L. and H. Haken (1992b), [*Unbiased estimate of forces from measured correlation funtions, including the case of strong multiplicative noise*]{}, Ann. Physik, 1, 452-460.
Borland L. and H. Haken (1993), [*On the constraints necessary for macroscopic prediction of stochastkic time-dependent processes*]{}, Rep. Math. Phys., 33, 35-42.
Czechowski Z. and M. Bia[ł]{}ecki (2010a), [*Ito equations as macroscopic stochastic models of geophysical phenomena - construction of the models on a base of time series and analytical derivation*]{}, Chapter 6 in V. De Rubeis, Z. Czechowski and R. Teisseyre (Eds.) [*Synchronization and triggering: from fracture to earthquake processes*]{}, Springer, pages 77-96.
Czechowski Z. and Bia[ł]{}ecki M. (2010b), [*Three-level description of the domino cellular automaton*]{}, arXiv:1012.5902 \[nlin.CG\]
Haken H. (1988), [*Information and Self-Organization*]{}, Springer.
Racca E., Laio F., Poggi D. anf Ridolfi L. (2007), [*Test to determine the Markov order of time series*]{}, Phys. Rev. E 75, 011126-1-6.
Risken H. (1996), [*The Fokker-Planck Equation*]{}, Springer.
Rozmarynowska A. (2009) [*On the reconstruction of Ito models on the base of time series with long-tail distributions*]{}, Acta Geophysica, 57, no 2, 311-329.
Siegert S., Friedrich R. and Peinke J. (1998) [*Analysis of data sets of stochastic systems*]{}, Phys. Lett. A, 243, 275-280.
Takens F. (1981), [*Detecting strange attractors in turbulence*]{}, In D.A. Rand and L.S. Young (eds.) [*Dynamical Systems and Turbulence*]{}, Lecture Notes in Mathematics, vol. 898, Springer.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A space $X$ has the Rothberger property in all finite powers if, and only if, its collection of $\omega$-covers has Ramseyan properties.'
author:
- by Marion Scheepers
title: 'Rothberger’s property in all finite powers [^1] '
---
For $s\in\,[\mathbb{N}]^{<\aleph_0}$ and for $B\in[\mathbb{N}]^{\aleph_0}$ use $s< B$ to denote that $s=\emptyset$ or $\max(s) < \min(B)$. For $s < B$ define $
[s,B] = \{s\cup C\in[\mathbb{N}]^{\aleph_0}: \, s < C\subseteq B\}.
$ The family $\{[s,B]:\, s\subset\mathbb{N} \mbox{ finite and } s < B\in[\mathbb{N}]^{\aleph_0}\}$ forms a basis for a topology on $[\mathbb{N}]^{\aleph_0}$. This is the *Ellentuck topology* on $[\mathbb{N}]^{\aleph_0}$ and was introduced in [@El].
Recall that a subset $N$ of a topological space is *nowhere dense* if there is for each nonempty open set $U$ of the space a nonempty open subset $V\subset U$ such that $N\cap V = \emptyset$. And $N$ is said to be *meager* if it is a union of countably many nowhere dense sets. A subset of a topological space is said to have the *Baire property* if it is of the form $(U\setminus M)\bigcup(M\setminus U)$ for some open set $U$ and some meager set $M$.
\[galvinprikry\] For a set $R\subset [\mathbb{N}]^{\aleph_0}$ the following are equivalent:
1. [$R$ has the Baire property in the Ellentuck topology.]{}
2. [For each finite set $s\subset{{\mathbb N}}$ and for each infinite set $S\subset \mathbb{N}$ with $s< S$ there is an infinite set $T\subset S$ such that either $[s,T]\subset R$, or else $[s,T]\cap R = \emptyset$.]{}
The proof of $(1)\Rightarrow(2)$ is nontrivial but uses only the techniques of Galvin and Prikry [@G-P]. Galvin and Prikry proved a precursor of Theorem \[galvinprikry\]: If $R$ is a Borel set in the topology inherited from $2^{{{\mathbb N}}}$ via representing sets by their characteristic functions, then $R$ has property (2) in Theorem \[galvinprikry\]. Silver and Mathias subsequently gave metamathematical proofs that analytic sets (in the $2^{{{\mathbb N}}}$-topology) have this property. Theorem \[galvinprikry\] at once yields all these prior results. The original papers [@El] and [@G-P] give a nice overview of these facts, and more.
When a subset $\mathcal{X}$ of $[{{\mathbb N}}]^{\aleph_0}$ inherits the Ellentuck topology from $[{{\mathbb N}}]^{\aleph_0}$, we shall speak of “$\mathcal{X}$ with the Ellentuck topology". For $A$ an abstract countably infinite set define the Ellentuck topology on $[A]^{\aleph_0}$ by fixing a bijective enumeration $(a_n:n\in{{\mathbb N}})$ of $A$ and by defining for $s$ and $T$ nonempty subsets of $A$: $$s< T \mbox{ if: }a_n\in s \mbox{ and }a_m\in T \Rightarrow n < m.$$ With the relation $s< T$ defined, define the Ellentuck topology on $[A]^{\aleph_0}$ as above. For $B\subseteq A$ and for finite set $s\subseteq A$ we write $B|s$ for $\{a_n\in B: s< \{a_n\}\}$.
For families $\mathcal{A}$ and $\mathcal{B}$ we now define a sequence of statements:
> ${{\sf E}}(\mathcal{A},\mathcal{B})$: For each countably infinite $A\in\mathcal{A}$ and for each set $R\subset [A]^{\aleph_0}\cap\mathcal{B}$ the implication (1)$\Rightarrow$(2) holds, where:
>
> 1. [$R$ has the Baire property in the Ellentuck topology on $[A]^{\aleph_0}\cap\mathcal{B}$.]{}
>
> 2. [For each $S\subset A$ with $S\in\mathcal{A}$ and each finite subset $s$ of $A$, there is an infinite $B\subset S|s$ with $B\in\mathcal{B}$ such that $[s,B]\cap\mathcal{B}\subseteq R$ or $[s,B]\cap\mathcal{B}\cap R = \emptyset$.]{}
>
Thus, ${\sf E}([{{\mathbb N}}]^{\aleph_0},[{{\mathbb N}}]^{\aleph_0})$ is Ellentuck’s theorem.
> ${\sf GP}(\mathcal{A},\mathcal{B})$: For each countably infinite $A\in\mathcal{A}$ and each $R\subset [A]^{\aleph_0}\cap \mathcal{B}$ the implication $(1)\Rightarrow(2)$ holds:
>
> 1. [$R$ is open in the $2^{{{\mathbb N}}}$ topology on $[A]^{\aleph_0}\cap\mathcal{B}$.]{}
>
> 2. [For each $S\in[A]^{\aleph_0}\cap\mathcal{A}$ there is a set $B\in[S]^{\aleph_0}\cap\mathcal{B}$ such that either $([B]^{\aleph_0}\cap\mathcal{B})\subseteq R$, or else $[B]^{\aleph_0}\cap\mathcal{B}\cap R = \emptyset$.]{}
>
Thus, ${\sf GP}([{{\mathbb N}}]^{\aleph_0},[{{\mathbb N}}]^{\aleph_0})$ is part of the Galvin-Prikry theorem.
A subset $\mathcal{S}$ of $[A]^{<\aleph_0}$ is:
1. [dense if for each $B\in[A]^{\aleph_0}\cap \mathcal{A}$, $\mathcal{S}\cap[B]^{<\aleph_0} \neq \emptyset$.]{}
2. [thin if no element of $\mathcal{S}$ is an initial segment of another element of $\mathcal{S}$.]{}
The following is an abstract formulation of Galvin’s generalization of Ramsey’s Theorem, announced in [@Galvinnotices] and in [@G-P] derived from Theorem 1 there:
> ${\sf FG}(\mathcal{A},\mathcal{B})$: For each countably infinite $A\in\mathcal{A}$ and for each dense set $\mathcal{S}\subset [A]^{<\aleph_0}$ there is a $B\in[A]^{\aleph_0}\cap\mathcal{B}$ such that each $C\in[B]^{\aleph_0}\cap\mathcal{B}$ has an initial segment in $\mathcal{S}$.
In this notation Galvin’s generalization of Ramsey’s theorem reads that ${\sf FG}([{{\mathbb N}}]^{\aleph_0},[{{\mathbb N}}]^{\aleph_0})$. Similarly, the following is an abstract formulation of Nash-Williams’ theorem:
> ${\sf NW}(\mathcal{A},\mathcal{B})$: For each countably infinite $A\in\mathcal{A}$ and for each thin family $\mathcal{T}\subset [A]^{<\aleph_0}$ and for each $n$, and each partition $\mathcal{T} = \mathcal{T}_1 \cup \mathcal{T}_2 \cup \cdots \cup \mathcal{T}_n$ there is a $B\in[A]^{\aleph_0}\cap \mathcal{B}$ and an $i\in\{1,\cdots,n\}$ such that $[B]^{<\aleph_0}\cap\mathcal{T} \subseteq \mathcal{T}_i$.
In this notation Nash-Williams’ theorem reads that ${\sf NW}([{{\mathbb N}}]^{\aleph_0},[{{\mathbb N}}]^{\aleph_0})$.
> $\mathcal{A}\longrightarrow(\mathcal{B})^n_k$: For positive integers $n$ and $k$ and for each countable $A\in\mathcal{A}$ and for each function $f:[A]^n\rightarrow\{1,\cdots,\, k\}$ there is a $B\in[A]^{\aleph_0}\cap\mathcal{B}$ and an $i\in\{1,\cdots,k\}$ such that $f$ has value $i$ on $[B]^n$.
In this notation Ramsey’s theorem reads: For each $n$ and $k$, $[{{\mathbb N}}]^{\aleph_0}\longrightarrow([{{\mathbb N}}]^{\aleph_0})^n_k$.
An open cover $\mathcal{U}$ of a topological space $X$ is said to be an $\omega$-cover if $X\not\in\mathcal{U}$, but there is for each finite set $F\subset X$ a $U\in\mathcal{U}$ with $F\subseteq U$. The symbol $\Omega_X$ denotes the collection of $\omega$-covers of $X$. The symbol $\mathcal{O}_X$ denotes the collection of open covers of $X$. In [@Ro38] Rothberger introduced the following covering property: For each sequence $(\mathcal{U}_n:n\in{{\mathbb N}})$ of open covers of $X$ there is a sequence $(U_n:n\in{{\mathbb N}})$ such that each $U_n\in\mathcal{U}_n$, and $\{U_n:n\in{{\mathbb N}}\}$ is a cover of $X$. The symbol ${{\sf S}_1}(\mathcal{O}_X,\mathcal{O}_X)$ denotes this statement. The corresponding statement for $\omega$-covers of $X$, ${{\sf S}_1}(\Omega_X,\Omega_X)$, was introduced in [@Sakai] by Sakai. It states: For each sequence $(\mathcal{U}_n:n\in{{\mathbb N}})$ of $\omega$-covers of $X$ there is a sequence $(U_n:n\in{{\mathbb N}})$ such that each $U_n\in\mathcal{U}_n$, and $\{U_n:n\in{{\mathbb N}}\}$ is an $\omega$ cover for $X$. Sakai proved that $X$ has ${{\sf S}_1}(\Omega_X,\Omega_X)$ if, and only if, all finite powers of $X$ have ${{\sf S}_1}(\mathcal{O}_X,\mathcal{O}_X)$. According to Gerlits and Nagy [@GN] a space is said to be an $\epsilon$-*space* if each $\omega$-cover contains a countable subset which still is an $\omega$-cover. A space is an $\epsilon$-space if and only if it has the Lindelöf property in all finite powers - see [@GN] for details. In this paper we prove:
\[egprothberger\] For an $\epsilon$-space $X$, the following are equivalent:
1. [${{\sf S}_1}(\Omega_X,\Omega_X)$.]{}
2. [${{\sf E}}(\Omega_X,\Omega_X)$.]{}
3. [${\sf GP}(\Omega_X,\Omega_X)$.]{}
4. [${\sf FG}(\Omega_X,\Omega_X)$.]{}
5. [${\sf NW}(\Omega_X,\Omega_X)$.]{}
6. [For all $n$ and $k$, $\Omega_X\rightarrow(\Omega_X)^n_k$.]{}
7. [$\Omega_X\rightarrow(\Omega_X)^2_2$.]{}
The proof of ${{\sf S}_1}(\Omega_X,\Omega_X)\,\Rightarrow\,{{\sf E}}(\Omega_X,\Omega_X)$:
=========================================================================================
Assume that $X$ has property ${{\sf S}_1}(\Omega_X,\Omega_X)$. Fix a countable $A\in\Omega_X$ and fix a set $R\subset [A]^{\aleph_0}\cap\Omega_X$. For the remainder of the argument, fix a bijective enumeration of $A$, say $(a_n:n\in{{\mathbb N}})$. Sets of the form $[s,C] = \{D: s<C \mbox{ and }s \subset D \subseteq s\cup C\}$ constitute a basis for the Ellentuck topology on $[A]^{\aleph_0}$.
For a finite set $s\subset A$ and for $B\in[A|s]^{\aleph_0}\cap\Omega_X$:
1. [$B$ accepts $s$ if $[s,B]\cap\Omega_X\subseteq R$.]{}
2. [$B$ rejects $s$ if no $C\in[B]^{\aleph_0}\cap\Omega_X$ accepts $s$.]{}
Lemma \[acceptreject\] will be used without special reference:
\[acceptreject\] Let a finite set $s\subset A$ and a set $B\in[A|s]^{\aleph_0}\cap\Omega_X$ be given:
1. [$B$ accepts $s$ if, and only if, each $C\in[B]^{\aleph_0}\cap\Omega_X$ accepts $s$.]{}
2. [$B$ rejects $s$ if, and only if, each $C\in[B]^{\aleph_0}\cap\Omega_X$ rejects $s$.]{}
\[decideoneset\] For each finite set $s\subset A$, there is a $B\in[A|s]^{\aleph_0}\cap \Omega_X$ such that $B$ accepts $s$ or $B$ rejects $s$.
${\bf Proof:}$ If $A|s$ does not reject $s$, choose a $B\in[A|s]^{\aleph_0}\cap\Omega_X$ accepting $s$. ${\Box\vspace{0.15in}}$
\[acceptstrongreject\] Let $t\subset A$ be a finite set. Let $B\in[A]^{\aleph_0}\cap\Omega_X$ be such that for each finite set $s\subset (t\cup B)$, $B|s$ accepts $s$ or $B|s$ rejects $s$. If $B|t$ rejects $t$ then $C =\{u\in B: B|(t\cup\{u\}) \mbox{ rejects } t\cup\{u\}\}$ is a member of $\Omega_X$.
${\bf Proof:}$ Suppose not. Then $D = t\cup(B\setminus C)\in\Omega_X$, and for each $u\in D|t$, $B|(t\cup\{u\})$ accepts $t\cup\{u\}$. Thus for each $u\in D|t$, $D|(t\cup\{u\})$ accepts $t\cup\{u\}$. This means that $[t,D|t] = \cup_{u\in D}[t\cup\{u\},D|(t\cup\{u\})]\subseteq R$, and so $D|t$ accepts $t$. This contradicts Lemma \[acceptreject\] (2) since $D\in [B]^{\aleph_0}\cap\Omega_X$ and $B|t$ rejects $t$. ${\Box\vspace{0.15in}}$
$\omega$-covers accepting or rejecting all finite subsets. {#omega-covers-accepting-or-rejecting-all-finite-subsets. .unnumbered}
==========================================================
The game ${{\sf G}_1}(\Omega_X,\Omega_X)$ is played as follows: Players ONE and TWO play an inning per positive integer. In the $n$-th inning ONE first chooses an $O_n\in\Omega_X$; TWO responds with a $T_n\in O_n$. A play $O_1,\, T_1,\, \cdots,\, O_n,\, T_n,\, \cdots$ is won by TWO if $\{T_n:n\in{{\mathbb N}}\}\in \Omega_X$; else, ONE wins. It was shown in [@coc3] that
\[coc3gone\] For a topological space $Y$ the following are equivalent:
1. [$Y$ has property ${{\sf S}_1}(\Omega_Y,\Omega_Y)$.]{}
2. [ONE has no winning strategy in ${{\sf G}_1}(\Omega_Y,\Omega_Y)$.]{}
\[decidedfin\] If $Y$ has property ${{\sf S}_1}(\Omega_Y,\Omega_Y)$, then for each finite set $t\subset A$ and for each $B\in[A|t]^{\aleph_0}\cap\Omega_Y$ there is a $C\in[B]^{\aleph_0}\cap\Omega_Y$ such that for each finite set $s\subset t\cup C$, $C|s$ accepts $s$ or $C|s$ rejects $s$.
${\bf Proof:}$ Let $t$ and $B\in[A|t]^{\aleph_0}\cap\Omega_Y$ be given. Define a strategy $\sigma$ for ONE of ${{\sf G}_1}(\Omega_Y,\Omega_Y)$ as follows:
Enumerate the set of all subsets of $t$ as $\{t_1,\cdots,t_n\}$. Using Lemma \[decideoneset\] recursively choose $B_1\supset B_2\supset\cdots \supset B_n$ in $[B]^{\aleph_0}\cap \Omega_Y$ such that for each $i$, $B_i$ accepts $t_i$ or $B_i$ rejects $t_i$. Then define: $$\sigma(\emptyset) = B_n.$$
If TWO now chooses $T_1\in\sigma(\emptyset)$ then use Lemma \[decideoneset\] in the same way to choose $$\sigma(T_1) \in[\sigma(\emptyset)|\{T_1\}]^{\aleph_0}\cap\Omega_Y$$ such that for each set $F\subset t\cup\{T_1\}$, $\sigma(T_1)$ accepts $F$, or rejects $F$.
When TWO responds with $T_2\in F(T_1)$, enumerate the subsets of $t\cup \{T_1, T_2\}$ as $(t_1,\cdots,t_n)$ say, and choose by Lemma \[decideoneset\] sets $B_1,\cdots B_n\in[\sigma(T_1)|\{T_2\}]^{\aleph_0}\cap \Omega_X$ such that $B_j$ accepts $t_j$ or $B_j$ rejects $t_j$ for $1\le j\le n$ and $B_j\subset B_{j-1}$. Finally put $$\sigma(T_1,T_2) = B_n.$$ Note that for each finite subset $F$ of $t\cup\{T_1, T_2\}$, $\sigma(T_1,T_2)$ accepts $F$ or rejects it.
It is clear how player ONE’s strategy is defined. By Theorem \[coc3gone\] $\sigma$ is not a winning strategy for ONE. Consider a $\sigma$-play lost by ONE, say $$\sigma(\emptyset), \, T_1,\, \sigma(T_1),\, T_2,\, \sigma(T_1,T_2),\, \cdots,\, T_n,\, \sigma(T_1,\cdots,T_n),\, \cdots$$ Then $C = t\cup \{T_n:\, n\in{{\mathbb N}}\} \subset B$ is an element of $\Omega_Y$.
We claim that for each finite subset $s$ of $t\cup C$, $C|s$ accepts $s$ or $C|s$ rejects $s$. For consider such a $s$. If $s\subseteq t$, then as $C\subset F(\emptyset)$ and $F(\emptyset)$ accepts or rejects $s$, also $C$ does. If $s\not\subseteq t$, then put $n = \max\{m:T_m\in s\}$. Then $s$ is a subset of $t\cup\{T_1,\cdots,T_n\}$, so that $s$ is accepted or rejected by $\sigma(T_1,\cdots,T_n)$. But $C|s \subseteq \sigma(T_1,\cdots,T_n)$, and so $C|s$ accepts or rejects $s$. ${\Box\vspace{0.15in}}$
Completely Ramsey sets {#completely-ramsey-sets .unnumbered}
======================
The subset $R$ of $[A]^{\aleph_0}\cap \Omega_X$ is said to be *completely Ramsey* if there is for each finite set $s\subset A$ and for each $B\in[A|s]^{\aleph_0}\cap\Omega_X$ a set $C\in[B]^{\aleph_0}\cap\Omega_X$ such that
1. [either $([s,C]\cap\Omega_X)\subseteq R$,]{}
2. [or else $([s,C]\cap\Omega_X)\cap R = \emptyset$.]{}
\[complramseyunions\] If $R$ and $S$ are completely Ramsey subsets of $[A]^{\aleph_0}\cap \Omega_X$, then so is $R\bigcup S$.
${\bf Proof:}$ Let a finite set $s\subset A$ and $B\in[A|s]^{\aleph_0}\cap\Omega_X$ be given. Since $R$ is completely Ramsey, choose $C\in[B]^{\aleph_0}\cap\Omega_X$ such that $([s,C]\cap\Omega_X)\subset R$, or $([s,C]\cap\Omega_X)\cap R = \emptyset$. If the former hold we are done. In the latter case, since $S$ is completely Ramsey, choose $D\in[C]^{\aleph_0}\cap\Omega_X$ such that $([s,D]\cap\Omega_X)\subseteq S$, or $([s,D]\cap\Omega_X)\cap S = \emptyset$. In either case the proof is complete. ${\Box\vspace{0.15in}}$
The following Lemma is obviously true.
\[complramseycomplements\] If $R$ is completely Ramsey, then so is $([A]^{\aleph_0}\cap\Omega_X)\setminus R$.
\[intersections\] If $R$ and $S$ are completely Ramsey subsets of $[A]^{\aleph_0}\cap \Omega_X$, then so is $R\bigcap S$.
${\bf Proof:}$ Lemmas \[complramseyunions\] and \[complramseycomplements\], and De Morgan’s laws.${\Box\vspace{0.15in}}$
Open sets in the Ellentuck topology {#open-sets-in-the-ellentuck-topology .unnumbered}
===================================
We are still subject to the hypothesis that $X$ satisfies ${{\sf S}_1}(\Omega_X,\Omega_X)$.
\[openprecursor\] For each finite set $t\subset A$ and for each $B\in[A|t]^{\aleph_0}\cap\Omega_X$ such that for each finite subset $F$ of $t\cup B$, $B|F$ accepts, or rejects $F$ the following holds: For each finite set $s\subset t\cup B$ such that $B|s$ rejects $s$, there is a $C\in[B|s]^{\aleph_0}\cap\Omega_X$ such that for each finite set $F\subset C$, $C|F$ rejects $s\cup F$.
${\bf Proof:}$ Fix $B$ and $s$ as in the hypotheses. Define a strategy $\sigma$ for ONE in ${{\sf G}_1}(\Omega_X,\Omega_X)$ as follows: By Lemma \[acceptstrongreject\] $$\sigma(\emptyset) =\{U\in B:\,s<\{U\}\mbox{ and $B|\{U\}$ rejects }s\cup\{U\}\} \in\Omega_X.$$ Notice that $\sigma(\emptyset)$ accepts or rejects each of its finite subsets, it rejects $s$, and for each $U\in \sigma(\emptyset)$, $\sigma(\emptyset)|\{U\}$ rejects $s\cup\{U\}$.
If TWO now chooses $T_1\in \sigma(\emptyset)$, then by Lemma \[acceptstrongreject\] $$\sigma(T_1)=\{U\in \sigma(\emptyset)\setminus\{T_1\}: \sigma(\emptyset)|F \mbox{ rejects }s\cup F \mbox{ for each finite }F\subset\{T_1,U\}\}$$ is in $\Omega_X$. As before, $\sigma(T_1)$ accepts or rejects each of its finite subsets, and for any $U\in \sigma(T_1)$, for each finite subset $F$ of $\{U,T_1\}$, $\sigma(T_1)|F$ rejects $s\cup F$.
If next TWO chooses $T_2\in \sigma(T_1)$, then by Lemma \[acceptstrongreject\] $$\sigma(T_1,T_2) = \{U\in \sigma(T_1)\setminus\{T_2\}: \sigma(T_1)|F \mbox{ rejects } s\cup F \mbox{ for any finite } F\subset \{T_1,T_2,U\}\}$$ is an element of $\Omega_X$.
Continuing in this way we define a strategy $\sigma$ for ONE in ${{\sf G}_1}(\Omega_X,\Omega_X)$. Since $X$ satisfies ${{\sf S}_1}(\Omega_X,\Omega_X)$, $\sigma$ is not a winning strategy for ONE. Consider a $\sigma$-play lost by ONE, say: $$\sigma(\emptyset), \, T_1,\, \sigma(T_1),\, T_2,\, \sigma(T_1,T_2),\, T_3,\, \sigma(T_1,T_2,T_3),\, \cdots$$ Put $C = \{T_n:n\in{{\mathbb N}}\}$. Then $C\in [B|s]^{\aleph_0}\cap\Omega_X$. We claim that for each finite set $F\subset C$, $C|F$ rejects $s\cup F$.
For choose a finite set $F\subset C$. Then $F\cap s = \emptyset$. Fix $n=\max\{m:T_m\in F\}$. Then $C|F \subset \sigma(T_1,\cdots,T_n)$, and the latter rejects $s\cup F$ for all finite subsets $F$ of $\{T_1,\cdots,T_n\}$. Thus $C|F$ rejects $s\cup F$.${\Box\vspace{0.15in}}$
\[opencomplRamsey\] If $X$ has property ${{\sf S}_1}(\Omega_X,\Omega_X)$, then every open subset of $[A]^{\aleph_0}\cap\Omega_X$ is completely Ramsey.
${\bf Proof:}$ Let $R\subset[A]^{\aleph_0}\cap\Omega_X$ be open in this subspace. Consider a finite set $s\subset A$ and a $B \in [A|s]^{\aleph_0}\cap\Omega_X$. Since $(X,d)\models{{\sf S}_1}(\Omega_X,\Omega_X)$, choose by Theorem \[decidedfin\] a $C\in[B]^{\aleph_0}\cap\Omega_X$ such that for each finite set $F\subset(s\cup C)$, $C|F$ accepts or rejects $F$.
If $C$ accepts $s$ then we have $[s,C]\cap\Omega_X\subseteq R$, and we are done. Thus, assume that $C$ does not accept $s$. Then $C$ rejects $s$, and we choose by Lemma \[openprecursor\] a $D\in[C|s]^{\aleph_0}\cap\Omega_X$ such that for each finite subset $F$ of $D$, $D|F$ rejects $s\cup F$.
We claim that $([s,D]\cap\Omega_X)\cap R = \emptyset$. For suppose not. Choose $E\in[s,D] \cap\Omega_X \cap R$. Since $R$ is open, choose an Ellentuck neighborhood of $E$ contained in $R$, say $[t,K]\cap \Omega_X$. Then we have $s\subset E \subset s\cup D$ and $t\subset E \subset t\cup K$. But then $s\cup t\subset E \subset t \cup K$ and $[s\cup t,K|s]\subset R$, whence also $[s\cup t,E|(s\cup t)]\subset R$. But then $E|(s\cup t)$ accepts $s\cup t$ where $t$ is a finite subset of $s\cup D$, and $E|(s\cup t)\subset D|t$, and $D|t$ rejects $s\cup t$, a contradiction. ${\Box\vspace{0.15in}}$
Meager subsets in the Ellentuck topology {#meager-subsets-in-the-ellentuck-topology .unnumbered}
========================================
If the subset $R$ of $[A]^{\aleph_0}\cap\Omega_X$ is nowhere dense in the topology, then for each $B\in[A]^{\aleph_0}\cap\Omega_X$ and for each finite set $s\subset A$, $B|s$ rejects $s$. We now examine the meager subsets of $[A]^{\aleph_0}\cap \Omega_X$.
\[nwdfin\] If $R$ is nowhere dense, then there is for each $B\in[A]^{\aleph_0}\cap\Omega_X$ and each finite set $t\subset A$ a set $C\in[B|t]^{\aleph_0}\cap\Omega_X$ such that for each finite set $s\subset t\cup C$, $C|s$ rejects $s$.
${\bf Proof:}$ Since $R$ is nowhere dense, no $\omega$-cover contained in $A$ can accept a finite set. Thus each $\omega$-cover contained in $A$ rejects each finite subset of $A$. ${\Box\vspace{0.15in}}$
\[clnwdcomplramsey\] Assume ${{\sf S}_1}(\Omega_X,\Omega_X)$. If $R$ is a closed nowhere dense subset of $[A]^{\aleph_0}\cap\Omega_X$ then there is for each finite subset $s\subset A$ and for each $B\in[A|s]^{\aleph_0}\cap\Omega_X$ a $C\in[B]^{\aleph_0}\cap\Omega_X$ such that $[s,C]\cap R = \emptyset$.
${\bf Proof:}$ First, note that closed nowhere dense subsets are complements of open dense sets. By Theorem \[opencomplRamsey\], each open set is completely Ramsey. By Lemma \[complramseycomplements\] each closed, nowhere dense set is completely Ramsey. By Lemma \[nwdfin\] the rest of the statement follows. ${\Box\vspace{0.15in}}$
By taking closures, the preceding lemma implies:
\[nwdcomplramsey\] Assume ${{\sf S}_1}(\Omega_X,\Omega_X)$. If $R$ is a nowhere dense subset of $[A]^{\aleph_0}\cap\Omega_X$ then there is for each finite subset $s\subset A$ and for each $B\in[A|s]^{\aleph_0}\cap\Omega_X$ a $C\in[B]^{\aleph_0}\cap\Omega_X$ such that $[s,C]\cap R = \emptyset$.
And now we prove:
\[meagerellentuck\] Assume ${{\sf S}_1}(\Omega_X,\Omega_X)$. For a subset $N$ of $[A]^{\aleph_0}\cap\Omega_X$ the following are equivalent:
1. [$N$ is nowhere dense.]{}
2. [$N$ is meager.]{}
${\bf Proof:}$ We must show that (2)$\Rightarrow$(1). Thus, assume that $N$ is meager and write $N=\bigcup_{n\in{{\mathbb N}}}N_n$, where for each $n$ we have $N_n\subseteq N_{n+1}$, and $N_n$ is nowhere dense in $[A]^{\aleph_0}\cap \Omega_X$. Consider any basic open set $[s,B]$ of $[A]^{\aleph_0}\cap\Omega_X$. Define a strategy $\sigma$ for ONE in the game ${{\sf G}_1}(\Omega_X,\Omega_X)$ as follows:
Since $N_1$ is nowhere dense, choose by Corollary \[nwdcomplramsey\] an $O_1\in[B]^{\aleph_0}\cap \Omega_X$ with $[s,O_1]\cap N_1 = \emptyset$. Define $\sigma(\emptyset) = O_1$.
When TWO chooses $T_1\in \sigma(\emptyset)$ choose by Corollary \[nwdcomplramsey\] an $O_2\in[\sigma(\emptyset)|\{T_1\}]^{\aleph_0}\cap\Omega_X$ with $[s,O_2]\cap N_2 = \emptyset$, and define $\sigma(T_1) = O_2$.
Now when TWO chooses $T_2\in \sigma(T_1)$, find by Corollary \[nwdcomplramsey\] an $O_3\in[\sigma(T_1)|\{T_2\}]^{\aleph_0}\cap\Omega_X$ with $[s,O_3]\cap N_3 = \emptyset$, and define $\sigma(T_1,T_2) = O_3$.
It is clear how to define ONE’s strategy $\sigma$. By Theorem \[coc3gone\] $F$ is not a winning strategy for ONE. Consider a play $$\sigma(\emptyset),\, T_1,\, \sigma(T_1),\, T_2,\, \cdots,\, \sigma(T_1,\cdots,T_n),\, T_{n+1},\, \cdots$$ lost by ONE. Put $C = \{T_n:n\in{{\mathbb N}}\}$. Then $C\in[B]^{\aleph_0}\cap\Omega_X$. Observe that by the definition of $\sigma$ we have for each $k$ and each finite set $F\subset \{T_1,\cdots,T_k\}$ that $[s\cup F,\sigma(T_1,\cdots,T_k)]\cap N_k = \emptyset$.\
$[s,C]\cap N = \emptyset$.\
For suppose that instead $[s,C]\cap N \neq \emptyset$. Choose $V\in [s,C]\cap N$, and then choose $m$ so that $V\in N_m$. Choose the least $k>m$ with $T_k\in V| s$. This is possible because $s$ is finite. Observe also that $s\subseteq V\subseteq s\cup C = s\cup\{T_j:j\in{{\mathbb N}}\}$. Put $F = V\cap \{T_1,\cdots,T_k\}$. Thus we have that $[s\cup F,V|F]\cap N_k \neq \emptyset$, which contradicts the fact that $V|F\subset \sigma(T_1,\cdots,T_k)$, and $[s\cup F,\sigma(T_1,\cdots,T_k)]\cap N_k = \emptyset$. This completes the proof of the claim. ${\Box\vspace{0.15in}}$
Using Lemmas \[complramseyunions\] and \[complramseycomplements\] and Corollary \[intersections\] we have:
\[ellentuck\] Suppose $X$ satisfies ${{\sf S}_1}(\Omega_X,\Omega_X)$. Then for each $A\in\Omega_X$, every subset of $[A]^{\aleph_0}\cap\Omega_X$ which has the Baire property is completely Ramsey.
The proof of ${{\sf E}}(\Omega_X,\Omega_X)\,\Rightarrow\,{{\sf S}_1}(\Omega_X,\Omega_X)$:
=========================================================================================
Note that a set open in the $2^{{{\mathbb N}}}$ topology is also open in the Ellentuck topology. The implication $(2)\Rightarrow(3)$ of Theorem \[egprothberger\] follows from this remark. Now we start with $(3)$.
\[densetoinitsegm\] Assume ${\sf GP}(\Omega_X,\Omega_X)$. Then ${\sf FG}(\Omega_X,\Omega_X)$ holds.
${\bf Proof:}$ Let $\mathcal{S}\subset[A]^{<\aleph_0}$ be dense and define $\mathcal{I}$ to be the set $\{D\in[A]^{\aleph_0}\cap\Omega_X:\, D \mbox{ has an initial segment in }\mathcal{S}\}$. Then we have: $$\mathcal{I} = \cup\{[s,D|s]:s\in \mathcal{S},\, D\in[A]^{\aleph_0}\cap\Omega_X \mbox{ and s an initial segment of }D\}$$ is a $2^{{{\mathbb N}}}$-open subset of $[A]^{\aleph_0}\cap\Omega_X$. Choose a $B\in[A]^{\aleph_0}\cap\Omega_X$ such that $[B]^{\aleph_0}\cap\Omega_X\subset\mathcal{I}$, or $[B]^{\aleph_0}\cap\Omega_X\cap\mathcal{I} = \emptyset$. But the second alternative implies the contradiction that $[B]^{<\aleph_0}\cap\mathcal{S} = \emptyset$. It follows that the first alternative holds.${\Box\vspace{0.15in}}$
\[thintohomogeneous\] Assume ${\sf FG}(\Omega_X,\Omega_X)$. Then ${\sf NW}(\Omega_X,\Omega_X)$ holds.
${\bf Proof:}$ Fix a thin family $\mathcal{T}\subset [A]^{<\aleph_0}$ and positive integer $n$, and a partition $\mathcal{T} = \mathcal{T}_1 \cup \mathcal{T}_2 \cup \cdots \cup \mathcal{T}_n$. We may assume $n=2$. If $\mathcal{T}_1$ is not dense, we can choose $B\in[A]^{\aleph_0}\cap \Omega_X$ such that $[B]^{<\aleph_0}\cap\mathcal{T} \subseteq \mathcal{T}_2$. Thus, assume $\mathcal{T}_1$ is dense. Choose, by the hypothesis, a $B\in[A]^{\aleph_0}\cap\Omega_X$ such that for each $C\in[B]^{\aleph_0}\cap\Omega_X$, some initial segment of $C$ is in $\mathcal{T}_1$.
Consider any $s\in\mathcal{T}\cap[B]^{<\aleph_0}$, and put $D = s\cup (B|s)$. Then $s$ is an initial segment of $D$, and $D\in[B]^{\aleph_0}\cap\Omega_X$, and so some initial segment of $D$, say $t$, is in $\mathcal{T}_1$. Since both $t$ and $s$ are initial segments of $D$ and are both in $\mathcal{T}$, and since $\mathcal{T}$ is thin, we have $s=t$, and so $s\in\mathcal{T}_1$. Consequently we have $[B]^{<\aleph_0}\cap\mathcal{T}\subseteq \mathcal{T}_1$. ${\Box\vspace{0.15in}}$
\[ramsey\] Assume that ${\sf NW}(\Omega_X,\Omega_X)$ holds. Then: For each $n$ and $k$ we have $\Omega_X\rightarrow(\Omega_X)^n_k$.
${\bf Proof:}$ Let $A\in\Omega_X$ be countable. Let positive integers $n$ and $k$ be given. Put $\mathcal{T}=[A]^n$. Then $\mathcal{T}$ is thin. Apply the hypothesis. ${\Box\vspace{0.15in}}$
The following theorem was proven in [@coc2] (Theorem 6.1) and [@coc1] (Theorem 24)[^2]. It, together with the above sequence of implications, completes the proof of Theorem \[egprothberger\].
\[soneramsey\] The following are equivalent:
1. [For each $n$ and $k$, $\Omega_X\longrightarrow(\Omega_X)^n_k$]{}
2. [$X\models{{\sf S}_1}(\Omega_X,\Omega_X)$.]{}
Remarks
=======
The results above are given for $\Omega$, but a study of the proofs will reveal that these equivalences hold for several other families $\mathcal{A}$. The main requirements on $\mathcal{A}$ are that each element of $\mathcal{A}$ has a countable subset in $\mathcal{A}$, that for each $k$ $\mathcal{A}\rightarrow (\mathcal{A})^1_k$ holds, and that ${{\sf S}_1}(\mathcal{A},\mathcal{A})$ is equivalent to ONE not having a winning strategy in ${{\sf G}_1}(\mathcal{A},\mathcal{A})$, and that this is equivalent to $\mathcal{A}\rightarrow(\mathcal{A})^2_2$. Though this general treatment can be given without much additional effort, I preferred to illustrate the equivalences using a well-known concrete example, because of the connections of this example with forcing (pointed out below) and with the famous Borel Conjecture. Here are a few examples of such families $\mathcal{A}$:
For a topological space $X$ and an element $x\in X$, define $\Omega_x = \{A\subset X\setminus\{x\}:\, x\in\overline{A}\}$. According to [@Sakai] $X$ has strong countable fan tightness at $x$ if the selection principle ${{\sf S}_1}(\Omega_x,\Omega_x)$ holds. Consider for a Tychonoff space $X$ the subspace of the Tychonoff product $\Pi_{x\in X}\mathbb{R}$ consisting of the continuous functions from $X$ to $\mathbb{R}$. The symbol ${\sf C}_p(X)$ denotes this subspace with the inherited topology. Since ${\sf C}_p(X)$ is homogeneous, the truth of ${{\sf S}_1}(\Omega_f,\Omega_f)$ at some point $f$ implies the truth of ${{\sf S}_1}(\Omega_f,\Omega_f)$ at any point $f$. Thus we may confine attention to $\Omega_{\bf o}$, where ${\bf o}$ is the function which is zero on $X$. Using the techniques above one can prove:
\[strfantight\] For a Tychonoff space $X$ the following are equivalent for ${\sf C}_p(X)$:
1. [${{\sf S}_1}(\Omega_{\bf o},\Omega_{\bf o})$.]{}
2. [${{\sf E}}(\Omega_{\bf o},\Omega_{\bf o})$.]{}
3. [${\sf GP}(\Omega_{\bf o},\Omega_{\bf o})$.]{}
4. [${\sf FG}(\Omega_{\bf o},\Omega_{\bf o})$.]{}
5. [${\sf NW}(\Omega_{\bf o},\Omega_{\bf o})$.]{}
6. [For all $n$ and $k$, $\Omega_{\bf o}\rightarrow(\Omega_{\bf o})^n_k$.]{}
For a topological space $X$ let $\mathcal{D}$ denote the collection whose members are of the form $\mathcal{U}$, a family of open subsets of $X$, such that no element of $\mathcal{U}$ is dense in $X$, but $\cup\mathcal{U}$ is dense in $X$. And let $\mathcal{D}_{\Omega}$ be the set of $\mathcal{U}\in\mathcal{D}$ such that for each finite family $\mathcal{F}$ of nonempty open subsets of $X$ there is a $U\in\mathcal{U}$ with $U\cap F\neq\emptyset$ for each $F\in\mathcal{F}$. The families $\mathcal{D}$ and $\mathcal{D}_{\Omega}$ were considered in [@coc5] where it was proved that for $X$ a set of real numbers, and ${\sf PR}(X)$ the Pixley-Roy space over $X$, the following holds:
\[prreals\] If $X$ is a set of real numbers, the following are equivalent for ${\sf PR}(X)$:
1. [${{\sf S}_1}(\mathcal{D}_{\Omega},\mathcal{D}_{\Omega})$.]{}
2. [ONE has no winning strategy in the game ${{\sf G}_1}(\mathcal{D}_{\Omega},\mathcal{D}_{\Omega})$.]{}
3. [For each $n$ and $k$ $\mathcal{D}_{\Omega}\rightarrow(\mathcal{D}_{\Omega})^n_k$.]{}
Each of these statements is equivalent to $X$ having ${{\sf S}_1}(\Omega_X,\Omega_X)$.
Using the techniques above one can prove:
\[prRamsey\] For a set $X$ of reals the following are equivalent for ${\sf PR}(X)$:
1. [${{\sf S}_1}(\mathcal{D}_{\Omega},\mathcal{D}_{\Omega})$.]{}
2. [${{\sf E}}(\mathcal{D}_{\Omega},\mathcal{D}_{\Omega})$.]{}
3. [${\sf GP}(\mathcal{D}_{\Omega},\mathcal{D}_{\Omega})$.]{}
4. [${\sf FG}(\mathcal{D}_{\Omega},\mathcal{D}_{\Omega})$.]{}
5. [${\sf NW}(\mathcal{D}_{\Omega},\mathcal{D}_{\Omega})$.]{}
6. [For all $n$ and $k$, $\mathcal{D}_{\Omega}\rightarrow(\mathcal{D}_{\Omega})^n_k$.]{}
For a non-compact topological space $X$ call an open cover $\mathcal{U}$ a [k]{}-cover if there is for each compact $C\subset X$ a $U\in\mathcal{U}$ such that $C\subseteq U$, and if $X\not\in\mathcal{U}$. Let $\mathcal{K}$ denote the collection of [k]{}-covers of such an $X$. If $X$ is a separable metric space then each member of $\mathcal{K}$ has a countable subset which still is a member of $\mathcal{K}$. Using the techniques above one can prove:
\[kcovRamsey\] For separable metric spaces $X$ the following are equivalent:
1. [ONE has no winning strategy in ${{\sf G}_1}(\mathcal{K},\mathcal{K})$.]{}
2. [${{\sf S}_1}(\mathcal{K},\mathcal{K})$.]{}
3. [${{\sf E}}(\mathcal{K},\mathcal{K})$.]{}
4. [${\sf GP}(\mathcal{K},\mathcal{K})$.]{}
5. [${\sf FG}(\mathcal{K},\mathcal{K})$.]{}
6. [${\sf NW}(\mathcal{K},\mathcal{K})$.]{}
7. [For all $n$ and $k$, $\mathcal{K}\rightarrow(\mathcal{K})^n_k$.]{}
The equivalence of (2) and (7) for n=2 and k=2 is Theorem 8 of [@dkm]. The equivalence of (1) and (2) is a result of [@ST]. The remaining equivalences are then derived as was done above for $\Omega$.
A collection $\mathcal{C}$ of subsets of a set $S$ is said to be a *combinatorial* $\omega$-*cover* of $S$ if $S\not\in\mathcal{C}$, but for each finite subset $F$ of $S$ there is a $C\in\mathcal{C}$ with $F\subseteq C$. For an infinite cardinal number $\kappa$ let $\Omega_{\kappa}$ be the set of *countable* combinatorial $\omega$-covers of $\kappa$. Let ${\sf cov}(\mathcal{M})$ be the least infinite cardinal number $\kappa$ such that the real line is a union of $\kappa$ first category sets. By the Baire Category Theorem ${\sf cov}(\mathcal{M})$ is uncountable. Using the techniques of this paper one can prove:
\[cardinality\] For an infinite cardinal number $\kappa$ the following are equivalent:
1. [$\kappa<{\sf cov}(\mathcal{M})$.]{}
2. [${{\sf S}_1}(\Omega_{\kappa},\Omega_{\kappa})$.]{}
3. [${{\sf E}}(\Omega_{\kappa},\Omega_{\kappa})$.]{}
4. [${\sf GP}(\Omega_{\kappa},\Omega_{\kappa})$.]{}
5. [${\sf FG}(\Omega_{\kappa},\Omega_{\kappa})$.]{}
6. [${\sf NW}(\Omega_{\kappa},\Omega_{\kappa})$.]{}
7. [For all positive integers $n$ and $k$, $\Omega_{\kappa}\rightarrow(\Omega_{\kappa})^n_k$.]{}
Rothberger’s property and forcing
=================================
Now we explore the connections between forcing and Rothberger’s property. Much of this part of the paper is inspired by Theorem 9.3 of [@JEB].
We begin by defining the following version of the well-known *Mathias reals* partially ordered set. Fix as before a countable $\omega$-cover $A$ of $X$, and enumerate it bijectively as $(a_n:n\in{{\mathbb N}})$. For $s\subset A$ finite, and $C\subset A|s$ with $C\in\Omega_X$, define: $$\mathcal{M}_A :=\{(s,C): s\in[A]^{<\aleph_0} \mbox{ and }C\subset A|s \mbox{ and }C\in\Omega_X\}.$$ For $(s_1,C_1)$ and $(s_2,C_2)$ elements of $\mathcal{M}_A$, we define $(s_1,C_1)\prec (s_2,C_2)$ if: $s_2\subset s_1$ and $C_1\subset C_2$ and $s_1\setminus s_2\subset C_2|s_2$.
Now $(\mathcal{M}_A,\prec)$ is a partially ordered set. Its combinatorial and forcing properties are related to the combinatorial properties of $\omega$-covers of $X$. In this section we will show (see Theorem 9.3 of [@JEB]):
\[truthlemma\] The following are equivalent:
1. [${{\sf S}_1}(\Omega_X,\Omega_X)$ holds.]{}
2. [For each countable $A\in\Omega_X$, for each sentence $\psi$ in the $\mathcal{M}_A$-forcing language, and for each $(s,B)\in\mathcal{M}_A$, there is a $C\subset B$ with $C\in\Omega_X$ such that $(s,C){\mathrel{\|}\joinrel\mathrel{-}}\psi$, or $(s,C){\mathrel{\|}\joinrel\mathrel{-}}\neg\psi$]{}
Proof of $(1)\Rightarrow (2)$:
------------------------------
Fix a sentence $\psi$ of the $\mathcal{M}_A$-forcing language and fix $(s,B)\in \mathcal{M}_A$. Define the subsets $$\mathcal{W} = \cup\{[t,C]: (t,C)\in\mathcal{M}_A \mbox{ and } (t,C){\mathrel{\|}\joinrel\mathrel{-}}\psi\}$$ and $$\mathcal{D} = \cup\{[t,C]: (t,C)\in\mathcal{M}_A \mbox{ and } (t,C){\mathrel{\|}\joinrel\mathrel{-}}\neg\psi\}.$$ Then $\mathcal{W}$ and $\mathcal{D}$ are open sets in the Ellentuck topology on $[A]^{\aleph_0}\cap\Omega_X$. Moreover, by Corollary VII.3.7(a) of [@kunen], $\mathcal{R} = \mathcal{W}\bigcup \mathcal{D}$ is dense. By Theorem \[ellentuck\], $\mathcal{R}$, $\mathcal{W}$ and $\mathcal{D}$ are completely Ramsey. Thus, for the given $(s,B)\in\mathcal{M}_A$ there is a $B_1\in[B]^{\aleph_0}\cap\Omega_X$ such that $[s,B_1]\subset \mathcal{R}$, or $[s,B_1]\cap \mathcal{R}=\emptyset$; since $\mathcal{R}$ is dense and $[s,B_1]$ is nonempty and open we have $[s,B_1]\subset\mathcal{R}$. But now $\mathcal{W}$ is completely Ramsey and so there is a $B_2\in[B_1]^{\aleph_0}\cap\Omega_X$ with $[s,B_2]\subset \mathcal{W}$, or $[s,B_2]\cap\mathcal{W}=\emptyset$. Since $[s,B_2]\subset[s,B_1]$, we have that $[s,B_2]\subset \mathcal{W}$ or $[s,B_2]\subset\mathcal{D}$. In either case we have $(s,B_2){\mathrel{\|}\joinrel\mathrel{-}}\psi$, or $(s,B_2){\mathrel{\|}\joinrel\mathrel{-}}\neg\psi$.
Proof of $(2)\Rightarrow (1)$:
------------------------------
This proof takes more work. We show that in fact $(2)$ implies that $\Omega_X\rightarrow(\Omega_X)^2_2$ holds. To see this, assume on the contrary that $\Omega_X\rightarrow(\Omega_X)^2_2$ fails. Choose a countable $A\in\Omega_X$ and a function $f:[A]^2\rightarrow \{0,1\}$ which witness this failure. Enumerate $A$ bijectively as $(a_n:n\in{{\mathbb N}})$ and build the following corresponding partition tree:\
$T_{\emptyset} = A$. $T_{(i)}:=\{a_n:n>1 \mbox{ and }f(\{a_1,a_n\}) = i\}$. For $\sigma\in^{<\omega}\{0,1\}$ of length $m$ for which $T_{\sigma}\in \Omega_X$, $T_{\sigma\frown(i)}:=\{a_n\in T_{\sigma}:n>m \mbox{ and }f(\{a_m,a_n\}) = i\}$.
Observe that for each $\sigma$ with $T_{\sigma}\in\Omega_X$ we have $T_{\sigma\frown(0)}\in\Omega_X$ or $T_{\sigma\frown(1)}\in\Omega_X$. For each $n$, define $\mathcal{T}_n:=\{T_{\sigma}\in\Omega_X: length(\sigma)=n\}$. Then we have from the definitions that:
1. [For each $B\in[A]^{\aleph_0}\cap\Omega_X$ and for each $n$ there is a $T\in\mathcal{T}_n$ with $B\cap T\in\Omega_X$.]{}
2. [For each $n$, for each $T\in\mathcal{T}_{n+1}$ there is a unique $T^{\prime}\in\mathcal{T}_n$ with $T\subset T^{\prime}$.]{}
3. [For each $n$ and $\sigma$, if $a_n\in T_{\sigma}$, then $n> m = length(\sigma)$.]{}
If there is a $B\in[A]^{\aleph_0}\cap\Omega_X$ such that for each $n$ there is a $T\in\mathcal{T}_n$ with $B\setminus\{a_j:j\leq n\}\subset T$, then there is a $C\in[A]^{\aleph_0}\cap\Omega_X$ such that $f$ is constant on $[C]^2$.\
For let such a $B$ be given. Since the elements of $\mathcal{T}_1$ are pairwise disjoint, choose the unique $i_1\in\{0,1\}$ with $B\setminus\{a_1\} \subset T_{(i_1)}$. Letting $T$ be the unique element of $\mathcal{T}_2$ with $B\setminus\{a_1,a_2\}\subset T$, we see that $T\subset T_{(i_1)}$, and so for a unique $i_2\in\{0,1\}$, $B\setminus\{a_1,a_2\}\subset T_{(i_1,i_2)}$. Arguing like this we find an infinite sequence $(i_j:j<\infty)$ in $^{{{\mathbb N}}}\{0,1\}$ such that for each $m$, $B\setminus\{a_1,\cdots,a_m\}\subset T_{(i_1,\cdots,i_m)}$.
Write $B =\{a_{n_j}:j\in{{\mathbb N}}\}$ where $n_i<n_j$ whenever $i<j$. Put $B_1=\{a_{n_j}: i_{n_j}=1\}$ and $B_0=\{a_{n_j}: i_{n_j}=0\}$. Then $B_0\in\Omega_X$, or $B_1\in\Omega_X$. In the former case $f$ is constant of value $0$ on $[B_0]^2$, and in the latter case $f$ is constant of value $1$ on $[B_1]^2$. This completes the proof of Claim 1.
Note that the conclusion of Claim 1 holds also if instead we hypothesize that $B\in[A]^{\aleph_0}\cap\Omega_X$ is such that for each $n$ with $a_n\in B$ there is a $T\in \mathcal{T}_{n}$ with $B\setminus\{a_j:j\le n\}\subset T$.
Since we are assuming that there is no $B\in[A]^{\aleph_0}\cap\Omega_X$ with $f$ constant on $[B]^2$, we get: There is no $B\in[A]^{\aleph_0}\cap\Omega_X$ such that for each $n$ with $a_n\in B$ there is a $T\in\mathcal{T}_n$ with $B\setminus\{a_j:j\leq n\}\subset T$. Indeed, this is equivalent to:
> For each $B\in[A]^{\aleph_0}\cap\Omega_X$ there is an $n$ with $a_n\in B$ but for each $T\in\mathcal{T}_n$ we have $B\setminus\{a_j:j\leq n\}\not\subset T$.
In what follows we will use ${\stackrel{\bullet}{a}}$ to denote the canonical name of the ground model object $a$ in the forcing language. Define the $\mathcal{M}_A$-name $$\Gamma:=\{({\stackrel{\bullet}{a}}_n,(s,B)): (s,B)\in \mathcal{M}_A \mbox{ and } a_n\in s\}.$$ Then for each $\mathcal{M}_A$-generic filter $G$ we have $$\Gamma_G = \cup\{s\in[A]^{<\aleph_0}:(\exists B\in[A]^{\aleph_0}\cap\Omega_X)((s,B)\in G)\}.$$
$(\emptyset,A){\mathrel{\|}\joinrel\mathrel{-}}``(\exists n)(\forall T\in{\stackrel{\bullet}{\mathcal{T}}_n})(\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le n\}\not\subseteq T)"$.\
For suppose that on the contrary $(s,B){\mathrel{\|}\joinrel\mathrel{-}}``(\forall n)(\exists T\in{\stackrel{\bullet}{\mathcal{T}}}_n)(\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le n\}\subseteq T)"$. Since we have $B\in[A]^{\aleph_0}\cap\Omega_X$, choose an $n_1$ so that $B\setminus\{a_j:j\le n_1\}$ is not a subset of any $T\in\mathcal{T}_n$. Then choose a $T_{n_1}\in\mathcal{T}_{n_1}$ so that $B\cap T_{n_1}\in\Omega_X$. Also choose $a_m\in B\setminus(T_{n_1}\cup\{a_j:j\le n_1\})$. Put $B^{\prime} = B|\{a_j:j\le m_1\}$ and put $t = s\cup\{a_m\}$. Then as $(s,B){\mathrel{\|}\joinrel\mathrel{-}}``\stackrel{\bullet}{\mathcal{T}}_{n_1}\mbox{ is a disjoint family}"$ and $(t, B^{\prime}\cap T_{n_1}){\mathrel{\|}\joinrel\mathrel{-}}``(\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le \stackrel{\bullet}{n}_1\})\cap \stackrel{\bullet}{T}_{n_1} \neq\emptyset"$. $$\label{force1}
(t, B^{\prime}\cap T_{n_1}){\mathrel{\|}\joinrel\mathrel{-}}``\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le \stackrel{\bullet}{n}_1\}\subset\stackrel{\bullet}{T}_{n_1}".$$ But evidently we also have $$\label{force2}
(t, B^{\prime}\cap T_{n_1}){\mathrel{\|}\joinrel\mathrel{-}}``\stackrel{\bullet}{a}_m\in(\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le \stackrel{\bullet}{n}_1\})\setminus \stackrel{\bullet}{T}_{n_1}".$$ Thus we have a condition forcing contradictory statements, a contradiction. It follows that Claim 2 holds.
Now we construct a sentence $\Psi(\Gamma)$ in the forcing language: $$``\Gamma\cap\{\stackrel{\bullet}{a}_j:j<n\} \mbox{ is even for the least }n \mbox{ with } \stackrel{\bullet}{a}_n\in\Gamma \mbox{ and for all }T \in \stackrel{\bullet}{\mathcal{T}}_n \Gamma\setminus \{\stackrel{\bullet}{a}_j:j\le n\}\not\subset T,\, "$$ By hypothesis 2 of the theorem, choose a $B\in[A]^{\aleph_0}\cap \Omega_X$ such that $(\emptyset,B)$ decides $\Psi(\Gamma)$.
Choose $k_1$ minimal so that $a_{k_1}\in B$ and for each $T\in\mathcal{T}_{k_1}$ we have $B\setminus\{a_j:j\leq k_1\}\not\subseteq T$. Put $B_1 = B\setminus\{a_j:j\leq k_1\}$ and choose $T_{k_1}\in\mathcal{T}_{k_1}$ so that $C_1:=B_1\cap T_{k_1}\in\Omega_X$. Choose $\ell_1$ so that $a_{\ell_1}\in B_1\setminus T_{k_1}$. Then $(\{a_{k_1},a_{\ell_1}\}, C_1) < (\emptyset,B)$ and so also $(\{a_{k_1},a_{\ell_1}\}, C_1)$ decides $\Psi(\Gamma)$.
By the construction of $C_1$ we see that for $T^{\prime}\in\mathcal{T}_{k_1}\setminus\{T_{k_1}\}$, also $(\{a_{k_1},a_{\ell_1}\}, C_1){\mathrel{\|}\joinrel\mathrel{-}}``\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le \stackrel{\bullet}{k}_1\}\not\subseteq \stackrel{\bullet}{T^{\prime}}"$. And since $a_{\ell_1}\not\in T_{k_1}$ we also have $(\{a_{k_1},a_{\ell_1}\}, C_1){\mathrel{\|}\joinrel\mathrel{-}}``\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le \stackrel{\bullet}{k}_1\}\not\subseteq \stackrel{\bullet}{T}_{k_1}"$. Moreover, $(\{a_{k_1},a_{\ell_1}\}, C_1){\mathrel{\|}\joinrel\mathrel{-}}``\Gamma\cap\{\stackrel{\bullet}{a}_j:j< \stackrel{\bullet}{k}_1\}=\emptyset"$. Since $k_1$ was chosen minimal and $a_{k_1}\in B$, the least $n$ having the properties of $k_1$ is $k_1$. It follows that $(\{a_{k_1},a_{\ell_1}\}, C_1){\mathrel{\|}\joinrel\mathrel{-}}\Psi(\Gamma)$, and as $(\emptyset,B)$ already decides $\Psi(\Gamma)$, we have $$\label{positiveforce}
(\emptyset,B){\mathrel{\|}\joinrel\mathrel{-}}\Psi(\Gamma).$$
Now repeat the previous construction starting with $C_1$ in place of $B$. Choose $k_2$ minimal so that $a_{k_2}\in C_1$ and for each $T\in\mathcal{T}_{k_2}$ we have $C_1\setminus\{a_j:j\leq k_2\}\not\subseteq T$. Since $a_{k_1}\not\in C_1$, we have $k_2>k_1$. Put $B_2 = C_1\setminus\{a_j:j\leq k_2\}$ and choose $T_{k_2}\in\mathcal{T}_{k_2}$ so that $C_2:=B_2\cap T_{k_2}\in\Omega_X$. Choose $\ell_2$ so that $a_{\ell_2}\in B_2\setminus T_{k_2}$. Then $(\{a_{k_1}, a_{k_2}, a_{\ell_2}\}, C_2) < (\emptyset,B)$ and so also $(\{a_{k_1}, a_{k_2}, a_{\ell_2}\}, C_2)$ decides $\Psi(\Gamma)$. By the construction of $C_2$ we see that for $T^{\prime}\in\mathcal{T}_{k_2}\setminus\{T_{k_2}\}$, also $(\{a_{k_1}, a_{k_2}, a_{\ell_2}\}, C_2){\mathrel{\|}\joinrel\mathrel{-}}``\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le \stackrel{\bullet}{k}_2\}\not\subseteq \stackrel{\bullet}{T^{\prime}}"$. And since $a_{\ell_2}\not\in T_{k_2}$ we also have $(\{a_{k_1}, a_{k_2}, a_{\ell_2}\}, C_2){\mathrel{\|}\joinrel\mathrel{-}}``\Gamma\setminus\{\stackrel{\bullet}{a}_j:j\le \stackrel{\bullet}{k}_2\}\not\subseteq \stackrel{\bullet}{T}_{k_2}"$. By minimality of $k_2$ and the fact that $a_{k_2}\in C_2$, we get that the minimal $n$ with these properties of $k_2$ is $k_2$: However, $(\{a_{k_1}, a_{k_2}, a_{\ell_2}\}, C_2){\mathrel{\|}\joinrel\mathrel{-}}``\Gamma\cap \{\stackrel{\bullet}{a}_j:j\le \stackrel{\bullet}{k}_2\}=\{\stackrel{\bullet}{a}_{k_1}\}"$. This means that $(\{a_{k_1}, a_{k_2}, a_{\ell_2}\}, C_2){\mathrel{\|}\joinrel\mathrel{-}}\neg\Psi(\Gamma)$. Since $(\{a_{k_1}, a_{k_2}, a_{\ell_2}\}, C_2) < (\emptyset,B)$ and $(\emptyset,B)$ already decides $\Psi(\Gamma)$, we find that
$$\label{negativeforce}
(\emptyset,B){\mathrel{\|}\joinrel\mathrel{-}}\neg\Psi(\Gamma).$$
Since (\[positiveforce\]) and (\[negativeforce\]) yield a contradiction, the hypothesis that $\Omega\rightarrow(\Omega)^2_2$ fails is false. This completes the proof of $(2)\Rightarrow(1)$ of Theorem \[truthlemma\]. ${\Box\vspace{0.15in}}$
The above result is again given for $\Omega$, but a study of the proofs will reveal that these equivalences hold for several other families $\mathcal{A}$, including the examples mentioned earlier. Theorem \[truthlemma\] has several consequences that will be explored elsewhere. One of the mentionable consequences is that forcing with $\mathcal{M}_A$ preserves cardinals, and in the generic extension the only groundmodel sets of reals having ${{\sf S}_1}(\Omega,\Omega)$ are the countable sets. And a countable support iteration of length $\aleph_2$ over a ground model satisfying the Continuum Hypothesis gives a model of Borel’s Conjecture, just like the usual Mathias reals iteration does - [@JEB].
In closing: Analogous results can be proved for the selection principle ${{\sf S}_{fin}}(\mathcal{A},\mathcal{A})$ and its relatives. These will be reported elsewhere.
[**Appendix A: Regarding Theorem \[soneramsey\]:**]{}
Strictly speaking, the only equivalence that has been explicitly proved in the literature is the equivalence of ${{\sf S}_1}(\Omega,\Omega)$ with $\Omega\longrightarrow(\Omega)^2_2$, with the remark that the techniques used to prove this case yield by an induction the full version that for all finite $n$ and $k$ we have $\Omega\longrightarrow(\Omega)^n_k$. It is perhaps worth putting down the main elements of such an argument explicitly for future reference. The only implication we need to prove is the implication that $\Omega\longrightarrow(\Omega)^2_2$ implies that for all $n$ and $k$ $\Omega\longrightarrow(\Omega)^n_k$. $\Omega\longrightarrow(\Omega)^2_2$ implies that for each $k>1$, $\Omega\longrightarrow(\Omega)^2_k$.\
This can be done by induction on $k+1$. For $k=1$ this is the hypothesis. Assuming we have proven the implication for $j\le k$, consider a countable $\omega$-cover $\mathcal{U}$ of $X$ and a coloring $f:[\mathcal{U}]^2\longrightarrow\{1,\cdots,k+1,k+2\}$. Define a new coloring $g$ so that $$g(\{U,V\}) = \left\{\begin{tabular}{ll}
f(\{U,V\}) & \mbox{ if }f(\{U,V\})$<$ k+1.\\
k+1 & \mbox{ otherwise}
\end{tabular}\right.$$ Applying the induction hypothesis we find an $\omega$-cover $\mathcal{V}\subset\mathcal{U}$ and an $i\in\{1,\cdots,k+1\}$ such that $g(\{U,V\})=i$ for all $\{U,V\}\in\mathcal{V}^2$. If $i<k+1$ then indeed $\mathcal{V}$ works for $f$. Else, $\mathcal{V}$ is an $\omega$-cover on whose pairs $f$ takes values $k+1$ or $k+2$, and now apply $\Omega\longrightarrow(\Omega)^2_2$.\
For $n>2$ and $k>1$, $\Omega\longrightarrow(\Omega)^n_k$ implies $\Omega\longrightarrow(\Omega)^2_k$.\
This can be done by starting with a countable $\omega$-cover $\mathcal{U}$ and a coloring $f:[\mathcal{U}]^2\longrightarrow\{1,\cdots,k\}$. Enumerate $\mathcal{U}$ bijectively as $\{U_m:m\in{{\mathbb N}}\}$. Define $g:[\mathcal{U}\}]^n\longrightarrow\{1,\cdots,k\}$ by $$g(\{U_{i_1}\cdots,U_{i_n}\}) = f(\{U_{i_1},U_{i_2}\}),$$ where we list the $n$-tuples according to increasing index in the chosen enumeration. Apply $\Omega\longrightarrow(\Omega)^n_k$.\
For $n>1$ and $k>1$, $\Omega\longrightarrow(\Omega)^n_k$ implies $\Omega\longrightarrow(\Omega)^{n+1}_k$.\
To prove this we use the fact that For $n>1$ and $k>1$, $\Omega\longrightarrow(\Omega)^n_k$ implies $\Omega\longrightarrow(\Omega)^2_2$, which in turn implies that ONE has no winning strategy in the game ${{\sf G}_1}(\Omega,\Omega)$.
Let a countable $\omega$-cover $\mathcal{U}$ be given, as well as $f:[\mathcal{U}]^{n+1}\longrightarrow\{1,\cdots,k\}$. Enumerate $\mathcal{U}$ bijectively as $\{U_m:m\in{{\mathbb N}}\}$. Define a strategy $F$ for ONE in the game ${{\sf G}_1}(\Omega,\Omega)$ as follows:\
Fix $U_1$ and define $$g_1:[\mathcal{U}\setminus\{U_1\}]^n\longrightarrow\{1,\cdots,k\}$$ by $g_1(\mathcal{V}) = f(\{U_1\}\bigcup\mathcal{V})$. Using $\Omega\longrightarrow(\Omega)^n_k$, fix an $i_1\in\{1,\cdots,k\}$ and an $\omega$-cover $\mathcal{U}_1\subset \mathcal{U}$ such that $g_1(\mathcal{V})=i_1$ for each $\mathcal{V}\in[\mathcal{U}_1]^n$. Declare ONE’s move to be $F(\emptyset) = \mathcal{U}_1$.
When TWO responds with $T_1 = U_{n_1}\in F(\emptyset)$, ONE first defines $$g_2:[\mathcal{U}_1\setminus\{U_j:j\le n_1\}]^n\longrightarrow\{1,\cdots,k\}$$ by $g_2(\mathcal{V}) = f(\{U_{n_1}\}\bigcup\mathcal{V})$. Then, using $\Omega\longrightarrow(\Omega)^n_k$, fix an $i_{n_1}\in\{1,\cdots,k\}$ and an $\omega$-cover $\mathcal{U}_2\subset \mathcal{U}_1\setminus\{U_j:j\le n_1\}$ such that $g_2(\mathcal{V})=i_{n_1}$ for each $\mathcal{V}\in[\mathcal{U}_2]^n$. Declare ONE’s move to be $F(T_1) = \mathcal{U}_2$.
When TWO responds with $T_2 = U_{n_2}\in F(T_1)$, ONE first defines $$g_3:[\mathcal{U}_2\setminus\{U_j:j\le n_2\}]^n\longrightarrow\{1,\cdots,k\}$$ by $g_3(\mathcal{V}) = f(\{U_{n_2}\}\bigcup\mathcal{V})$. Then, using $\Omega\longrightarrow(\Omega)^n_k$, fix an $i_{n_2}\in\{1,\cdots,k\}$ and an $\omega$-cover $\mathcal{U}_3\subset \mathcal{U}_2\setminus\{U_j:j\le n_2\}$ such that $g_3(\mathcal{V})=i_{n_2}$ for each $\mathcal{V}\in[\mathcal{U}_3]^n$. Declare ONE’s move to be $F(T_1,T_2) = \mathcal{U}_3$.
This describes ONE’s strategy in this game. Since it is not winning for ONE, we find a play $F(\emptyset), T_1, F(T_1), T_2, F(T_1,T_2), T_3, \dots$ which is lost by ONE. Associated with this play we have an increasing infinite sequence $n_1< n_2<\cdots <n_k<\cdots$ for which $T_k = U_{n_k}$, all $k$, and a sequence $i_{n_k}, k\in {{\mathbb N}}$ of elements of $\{1,\cdots,k\}$, and a sequence $\mathcal{U}_n,\, n\in{{\mathbb N}}$, of $\omega$-covers such that:
1. [For each $m$, $T_m=U_{n_m}\in\mathcal{U}_m\subset\mathcal{U}_{m-1}\setminus\{U_{n_j}:j\le m-1\}$.]{}
2. [For each $m$, $f(\{T_m\}\bigcup\mathcal{V}) = i_{n_m}$ whenever $\mathcal{V}\in[\mathcal{U}_{m+1}]^n$.]{}
3. [$\{T_m:m\in{{\mathbb N}}\}\subset\mathcal{U}$ is an $\omega$-cover.]{}
Fix an $i$ such that $\mathcal{W} = \{T_m:i_{n_m}=i \mbox{ and }m>n\}$ is an $\omega$-cover. Then for each $\mathcal{V}\in[\mathcal{W}]^{n+1}$ we have $f(\mathcal{V})=i$.
[**Acknowledgements**]{}
I thank the referee for a careful reading of the paper, and for very useful suggestions. I also thank the Boise Set Theory seminar for useful suggestions and remarks during a series of lectures on these results. And finally, I thank the organizing committee of the third Workshop on Coverings, Selections and Games in Topology, hosted in April 2007 in Vrnjacka Banja, Serbia, for the opportunity to present some of these results at the workshop.
J.E. Baumgartner, *Iterated forcing*, [**London Mathematical Society Lecture Note Series**]{} 87 (1983), 1 - 59. G. di Maio, Lj.D.R. Kočinac and E. Meccariello, *Applications of k-covers*, [**Acta Mathematica Sinica**]{} (English Series) 22 (2006), 1151 - 1160. E. Ellentuck, *A new proof that analytic sets are Ramsey*, [**The Journal of Symbolic Logic**]{} 39:1 (1974), 163 - 165. F. Galvin, *A generalization of Ramsey’s Theorem*, [**Notices of the American Mathematical Society**]{} 15 (1968), p. 548. Abstract 68T-368. F. Galvin and K. Prikry, *Borel sets and Ramsey’s Theorem*, [**The Journal of Symbolic Logic**]{} 38:2 (1973), 193 - 198. J. Gerlits and Zs. Nagy, *Some properties of C(X), I*, [**Topology and its Applications**]{} 14 (1982), 151 – 161. W. Just, A.W. Miller, M. Scheepers and P.J. Szeptycki, [*Combinatorics of open covers (II)*]{}, [**Topology and its Applications**]{} 73 (1996), 241 - 266. K. Kunen, *Set Theory: An introduction to independence proofs*, [**Studies in Logic and the Foundations of Mathematics**]{} 102 (1980). F. Rothberger, [*Eine Verschärfung der Eigenschaft [C]{}*]{}, [**Fundamenta Mathematicae**]{} 30 (1938), 50 - 55. M. Sakai, [*Property [C]{}” and function spaces*]{}, [**Proceedings of the American Mathematical Society**]{} 104 (1988), 917 - 919. N. Samet and B. Tsaban, *Ramsey theory of open covers: Lecture 3*, preprint. M. Scheepers, *The least cardinal for which the Baire category theorem fails*, [**Proceedings of the American Mathematical Society**]{} 125:2 (1997), 579 - 585. M. Scheepers, [*Combinatorics of open covers (I): Ramsey Theory*]{}, [**Topology and its Applications**]{} 69 (1996), 31-62. M. Scheepers, [*Combinatorics of open covers (III): games, ${\sf C}_p(X)$*]{}, [**Fundamenta Mathematicae**]{} 152 (1997), 231 - 254. M. Scheepers, *Combinatorics of open covers (V): Pixley-Roy spaces of sets of reals and $\omega$-covers*, [**Topology and its Applications**]{}102 (2000), 13 - 31. M. Scheepers, *${{\sf S}_1}(\mathcal{A},\mathcal{B})$ in distributive lattices*, [**Quaderni Mathematicae**]{} (to appear)
Address:\
Department of Mathematics\
Boise State University\
Boise, ID 83725.\
e-mail: [email protected]
[^1]:
[^2]: See Appendix A
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce an intrinsic [Ly$\alpha$]{} emission line profile reconstruction method for high-$z$ quasars (QSOs). This approach utilises a covariance matrix of emission line properties obtained from a large, moderate-$z$ ($2 \leq z \leq 2.5$), high signal to noise (S/N$>15$) sample of BOSS QSOs. For each QSO, we complete a Monte Carlo Markov Chain fitting of the continuum and emission line properties and perform a visual quality assessment to construct a large database of robustly fit spectra. With this dataset, we construct a covariance matrix to describe the correlations between the high ionisation emission lines [Ly$\alpha$]{}, [C[IV]{}]{}, [Si[IV]{}+O[IV\]]{}]{} and [C[III\]]{}]{}, and find it to be well approximated by an $N$-dimensional Gaussian distribution. This covariance matrix characterises the correlations between the line width, peak height and velocity offset from systemic while also allowing for the existence of broad and narrow line components for [Ly$\alpha$]{} and [C[IV]{}]{}. We illustrate how this covariance matrix allows us to statistically characterise the intrinsic [Ly$\alpha$]{} line solely from the observed spectrum redward of 1275Å. This procedure can be used to reconstruct the intrinsic [Ly$\alpha$]{} line emission profile in cases where [Ly$\alpha$]{} may otherwise be obscured. Applying this reconstruction method to our sample of QSOs, we recovered the [Ly$\alpha$]{} line flux to within 15 per cent of the measured flux at 1205Å (1220Å) $\sim85$ (90) per cent of the time.'
author:
- |
Bradley Greig$^{1}$[^1], Andrei Mesinger$^{1}$, Ian D. McGreer$^{2}$, Simona Gallerani$^{1}$, Zolt[á]{}n Haiman$^{3}$\
$^1$Scuola Normale Superiore, Piazza dei Cavalieri 7, I-56126 Pisa, Italy\
$^2$Steward Observatory, The University of Arizona, 933 North Cherry Avenue, Tucson, AZ 85721-0065, USA\
$^3$Department of Astronomy, Columbia University, 550 West 120th Street, New York, NY 10027, USA
bibliography:
- 'Papers.bib'
title: '[Ly$\alpha$]{} emission line reconstruction for high-$z$ QSOs'
---
quasars: emission lines – quasars: absorption lines – quasars: general – cosmology: observations – cosmology: theory
Introduction
============
The ability to measure the intrinsic spectrum of quasars (QSOs) plays an important role in astrophysics. Resolving individual emission line profiles can enable the detailed study of both the central supermassive black hole (SMBH) and the internal structure within the SMBH powered accretion disk (through reverberation mapping e.g. @Blandford:1982 [@Peterson:1993] or the virial method e.g. @Kaspi:2000 [@Peterson:2004; @Vestergaard:2006]). On cosmological scales, the relative strength (or lack thereof) of the emission profiles and spectral slopes of the QSO continuum can be used to yield insights into the thermal and ionisation state of the intergalactic medium (IGM) through studies of the QSO proximity zone [e.g. @Mesinger:2004p3625; @Wyithe:2004; @Fan:2006p4005; @Bolton:2007p3623; @Mesinger:2007p855; @Carilli:2010p1; @Calverley:2011; @Wyithe:2011; @Schroeder:2013p919].
The key ingredient for using QSOs to explore the IGM is observing the intrinsic ultraviolet (UV) emission. Thermal emission from the accretion disk peaks in UV, which then interacts with surrounding neutral hydrogen and recombines two thirds of the time into [Ly$\alpha$]{} photons (rest frame $\lambda_{{\ifmmode\mathrm{Ly}\alpha\else{}Ly$\alpha$\fi}{}}=1215.67$Å) to produce a prominent [Ly$\alpha$]{} emission profile. However, owing to the large scattering cross-section of [Ly$\alpha$]{} photons, neutral hydrogen column densities of $N_{{\ifmmode\mathrm{H\,{\scriptscriptstyle I}}\else{}H\,{\scriptsize I}\fi}{}} > 10^{18} {\rm cm}^{-2}$ are sufficient for [Ly$\alpha$]{} to enter into the strong absorption regime. In the case of the IGM, even minute traces ($\sim$ few per cent) of intervening neutral hydrogen along the line-of-sight are capable of scattering [Ly$\alpha$]{} photons away from the observer.
At $z\lesssim3$, the IGM is on average very highly ionised except in dense self shielded clumps [e.g. @Fan:2006p4005; @McGreer:2015p3668; @Collaboration:2015p4320; @Collaboration:2016; @Collaboration:2016p5913]. Typically, at these redshifts the resonant scattering and absorption of [Ly$\alpha$]{} photons occurs only due to diffuse amounts of neutral hydrogen which appear blueward of [Ly$\alpha$]{} ($\lambda<\lambda_{{\ifmmode\mathrm{Ly}\alpha\else{}Ly$\alpha$\fi}{}}$) as a series of discrete narrow absorption features, known as the [Ly$\alpha$]{} forest [@Rauch:1998]. More problematic for measuring the intrinsic [Ly$\alpha$]{} emission line are larger neutral column density absorbers such as Lyman-limit systems and damped [Ly$\alpha$]{} absorbers (DLAs). In the case of DLAs, with column densities of $N_{{\ifmmode\mathrm{H\,{\scriptscriptstyle I}}\else{}H\,{\scriptsize I}\fi}{}} > 10^{20} {\rm cm}^{-2}$ not only do these systems lead to completely saturated absorption, but they are also sufficiently dense to allow [Ly$\alpha$]{} absorption within the Lorentzian wings. If these DLAs are located sufficiently close to the source QSO, absorption in the wings can significantly affect the intrinsic [Ly$\alpha$]{} emission profile. Furthermore, associated strong absorbers intrinsic to the host QSO environment itself can additionally impact the [Ly$\alpha$]{} line emission [@Shen:2012].
At $z>6$, the IGM becomes increasingly neutral, and the [Ly$\alpha$]{} forest gives way to completely dark (absorbed) patches [e.g. @Barkana:2002; @Gallerani:2008; @Mesinger:2010p6068; @McGreer:2015p3668]. Once the IGM itself obtains a sufficiently large column density, it too begins to absorb [Ly$\alpha$]{} in the Lorentzian wings of the scattering cross-section, referred to as IGM damping wing absorption [@MiraldaEscude:1998p1041]. In principle, through the detection of the IGM damping wing imprint from the [Ly$\alpha$]{} emission spectra of high-$z$ QSOs one is able to provide a direct measurement of the neutral fraction of the IGM during the reionisation epoch. Importantly, this approach requires an estimate of the intrinsic [Ly$\alpha$]{} emission line in order to determine the contribution from the smooth component absorption of the IGM damping wing. Using this technique, limits on the IGM neutral fraction at $z>6$ have been obtained [e.g. @Mesinger:2007p855; @Bolton:2011p1063; @Schroeder:2013p919].
In the absence of a viable method to recover the intrinsic [Ly$\alpha$]{} emission profile, one can attempt to use an emission template constructed from a QSO composite spectrum. Numerous composite spectra exist, constructed by averaging over a vastly different number of QSOs [e.g. @Francis:1991p5112; @Brotherton:2001p1; @VandenBerk:2001p3887; @Telfer:2002p5713; @Shull:2012p5716; @Stevans:2014p5726; @Harris:2016p5028]. However, by construction, these composite spectra only describe the average properties of QSOs, not the intrinsic variations of individual QSOs. Not accounting for these intrinsic properties can bias estimates of the IGM damping wing [@Mortlock:2011p1049; @Bosman:2015p5005].
An alternative method is to reconstruct a template using a principle component analysis (PCA) [e.g. @Boroson:1992p4641; @Francis:1992p5021; @Suzuki:2005p5157; @Suzuki:2006p4770; @Lee:2011p1738; @Paris:2011p4774]. An improvement over the use of a composite spectrum, this approach aims to use the minimal subset of eigenvectors to characterise the QSO emission profile. For example, @Francis:1992p5021 find 3 eigenvectors are sufficient to describe 75 per cent of the observed profile variation, and 95 per cent if 10 eigenvectors are used. Beyond 10, the eigenvectors obtain little information and become rather noisy [@Suzuki:2005p5157]. However, these eigenvectors are extracted from a fit to the original QSO spectrum. Therefore, given that the [Ly$\alpha$]{} line cannot be directly recovered at high redshifts, in order to obtain an estimate of the intrinsic [Ly$\alpha$]{} profile some form of reconstruction/extrapolation of the relevant eigenvectors that describe [Ly$\alpha$]{} would be required.
Conveniently, studies of these QSO composites and PCA approaches have revealed a wealth of information regarding correlations amongst various emission lines and other observable properties of the source QSO. For example, the first eigenvector of @Boroson:1992p4641 showed that in the H$\beta$ region a strong anti-correlation existed between \[O\] and Fe. Furthermore, @Hewett:2010p6194 performed an in-depth exploration to establish relationships between emission line profiles and the source systemic redshift to reduce the scatter and biases in redshift determination. More directly relevant for this work, both @Shang:2007p4862 and @Kramer:2009p920 observe a strong correlation between the [Ly$\alpha$]{} and [C[IV]{}]{} peak blue shifts.
Motivated by the existence of correlations amongst the various emission lines properties, and the lack of either a robust physical model of quasar emission regions [e.g. @Baldwin:1995] or a method to recover the intrinsic [Ly$\alpha$]{} profile within high-$z$ or heavily [Ly$\alpha$]{} obscured QSOs, we propose a new method to reconstruct the intrinsic [Ly$\alpha$]{} emission line profile for any high-$z$ or [Ly$\alpha$]{} obscured QSO. In this work, we develop our reconstruction method using QSOs selected from the Baryon Oscillation Spectroscopic Survey (BOSS; @Dawson:2013p5160), a component of SDSS-III [@Eisenstein:2011p5159]. In summary, our [Ly$\alpha$]{} reconstruction method is as follows:
- Perform a Monte Carlo Markov Chain (MCMC) fit to a subset of emission lines for each of our selected QSOs.
- Construct a covariance matrix which describes all correlations amongst the QSO emission lines.
- Assume a $N$-dimensional Gaussian likelihood function to describe the covariance matrix.
- MCMC fit a high-$z$ or [Ly$\alpha$]{} obscured QSO characterising all lines except [Ly$\alpha$]{}.
- Use the recovered emission line information from the QSO to statistically characterise the intrinsic [Ly$\alpha$]{} line.
The remainder of this paper is organised as follows. In Section \[sec:Data\], we discuss the observational data used within this work, and the selection criteria we apply to construct our sample of QSOs. In Section \[sec:Fitting\] we describe our MCMC fitting procedure, and how we model the QSO continuum, emission line features and other components to aid the estimation of the observed QSO flux. With all the data obtained from fitting our entire QSO sample, in Section \[sec:Covariance\] we construct our covariance matrix, and discuss the major correlations and recovered features. In Section \[sec:Recon\] we outline our [Ly$\alpha$]{} reconstruction method, and highlight the performance of this approach. Following this, in Section \[sec:Discussions\] we provide a discussion of the potential applications for both the MCMC fitting algorithm and the [Ly$\alpha$]{} reconstruction pipeline. Finally, in Section \[sec:Conclusions\] we finish with our closing remarks.
Data {#sec:Data}
====
In this work, we select our QSOs from Data Release 12 (DR12) [@Alam:2015p5162] of the large-scale SDSS-III observational programme BOSS [@Dawson:2013p5160]. The full details of the SDSS telescope are available in @Gunn:2006p1 and the details of the upgraded SDSS/BOSS spectrographs may be obtained in @Smee:2013p1. For reference, the wavelength coverage of the BOSS spectrograph is $3,\!600{\rm \AA} < \lambda < 10,\!400{\rm \AA}$ in the observed frame, with a resolution of $R\sim2000-2500$ corresponding to pixel resolution of $\sim120-150$ km/s. The QSO target selection for BOSS consisted of a variety of schemes, including colour and variability selection, counterparts to radio and X-ray sources, and previously known QSOs [@Bovy:2011p1; @Kirkpatrick:2011p1; @Ross:2012p1]. The candidate QSO spectra were then visually inspected following the procedure outlined in @Paris:2016p1, with updated lists of confirmed QSOs for DR12 available online[^2]. Furthermore, we use the publicly available flux calibration model of @Margala:2015p1 to perform the spectrophotometric corrections of the DR12 QSO fluxes.
In total, DR12, contains 294,512 uniquely identified QSOs, 158,917 of which are observed within the redshift range $2.15 < z < 3.5$. The principal science goal of the BOSS observational programme was the detection of the baryon acoustic oscillation scale from the [Ly$\alpha$]{} forest. To perform this, only relatively low to moderate S/N QSOs are required [e.g. @White:2003p578; @McDonald:2007p6752]. However, within this work, we aim to statistically characterise the correlations between numerous QSO emission lines, which require much higher S/N.
To construct the QSO sample used within this work, we preferentially selected QSOs with a median S/N across all filters ($ugriz$) of S/N $>$ 15 (‘snMedian $>15$’). This choice is arbitrary, and in principle we could go lower, however for decreasing S/N the weaker emission lines become more difficult to differentiate from the noise, reducing potential correlations. We selected QSOs containing broad-line emission using the ‘BROADLINE’ flag and removed all sources visually confirmed to contain broad absorption lines (BALs) by using the ‘QSO’ selection flag. Furthermore, only QSOs with ‘ZWARNING’ set to zero are retained, where the redshifts were recovered with high confidence from the BOSS pipeline. Finally, we restrict our QSO redshift range to $2.08 < z < 2.5$[^3], selecting QSOs by their BOSS pipeline redshift[^4] (e.g. @Bolton:2012SDSS; see Appendix \[sec:systemic\_redshift\] for a more in-depth discussion on our adopted choice of BOSS redshift). Following these selection cuts, we recover a total QSO sample of 3,862[^5].
Though the total QSO number of 3,862 appears small, it should be more than sufficient to elucidate any statistically significant correlations amongst the emission line parameters[^6]. Furthermore, to properly assess the MCMC fitting to be discussed in Section \[sec:Fitting\] a detailed visual inspection will be required, therefore it is preferential to restrict the sample size. In selecting our redshift range of $2.08 < z < 2.5$, we are assuming there is no strong variation in the emission line profiles of QSOs across the age of the Universe. That is, the covariance matrix recovered from this QSO sample will always be representative of the underlying QSO population. In @Becker:2013p1008, these authors constructed 26 QSO composite spectra between rest-frame $1040 < \lambda < 1550$Å across the redshift range of $2 < z < 5$. They found for [Ly$\alpha$]{} and the emission lines redward, no significant variation with redshift, lending confidence to our assumption.
Note, throughout this work, we do not deredden our QSOs to account for interstellar dust extinction within the Milky Way [e.g. @Fitzpatrick:1999p5034]. The dust extinction curve varies strongly in the UV and ultimately will impact the [Ly$\alpha$]{} line more than the [C[IV]{}]{} line. However, this will not greatly impact our results as we are attempting to reconstruct the [Ly$\alpha$]{} emission line using a covariance matrix of correlations from other emission lines. Applying the extinction curve will only impact slightly on the peak of the emission lines, not the other characteristics of the line profile. When highlighting the performance of the reconstruction process in Section \[sec:Recon\], we see that this slight overestimate should be well within the errors of the reconstructed profile.
MCMC QSO Fitting {#sec:Fitting}
================
We now introduce our MCMC fitting procedure. In Section \[sec:lines\], we outline and justify the selection of the emission lines considered. In Section \[sec:template\] we then detail the construction of the QSO template. In order to improve the ability to recover the intrinsic emission line profile and QSO continuum we outline the treatment of ‘absorption’ features in Section \[sec:absorb\]. In Section \[sec:MCMC\] the iterative MCMC fitting procedure is discussed and an example is presented. Finally, in Section \[sec:QA\] we perform a visual quality assessment of our entire sample of fit QSOs to remove contaminants that could impact the observed emission line correlations.
Emission line selection {#sec:lines}
-----------------------
Owing to the limited wavelength coverage of BOSS, the selection of available emission lines to be used is limited. Due to the complexity in observing high-$z$ ($z\gtrsim6$) QSOs, where [Ly$\alpha$]{} is redshifted into the near infrared (IR), the emission lines we use from the BOSS sample needs to be consistent with what is detectable with near-IR instruments such as Keck/MOSFIRE [@McLean:2010; @McLean:2012] and VLT/X-Shooter [@Vernet:2011].
The strongest emission lines, especially [Ly$\alpha$]{}, [C[IV]{}]{} and [Mg[II]{}]{}, are known to contain a broad and narrow component [e.g. @Wills:1993; @Baldwin:1996; @VandenBerk:2001p3887; @Shang:2007p4862; @Kramer:2009p920; @Shen:2011p4583]. These different components are thought to arise from moving clouds of gas above and below the accretion disk known as the broad and narrow line regions. Note that, for our purposes, the physical origins of these components is irrelevant; a double gaussian merely provides us a flexible basis set in which to characterise the line profile. Throughout this work, we will approximate all emission lines to have a Gaussian profile. While emission line profiles are known to be Lorentzian [e.g. @Peebles1993], the fitting of a Lorentzian profile is complicated by the requisite identification of the broader wings relative to the uncertainty in the QSO continuum[^7]. Ultimately, this is a subtle difference as we are only interested in characterising the total line profile.
In order to decide which strong emission line profiles should be fit with a single or double component Gaussian[^8], we perform a simple test analysing the Bayes information criteria (BIC; @Schwarz:1978p1 [@Liddle:2004p5730]) of each of the individual line profiles in Appendix \[sec:line\_component\]. Below we summarise our findings:
- [Ly$\alpha$]{}: The strongest observed UV emission line. In order to better characterise the fit to the line profile, a two component Gaussian is preferred. This is well known and consistent with other works.
- [Si[IV]{}+O[IV\]]{}]{}: This line complex consists of both the [Si[IV]{}]{} ($\lambda$1396.76Å) and [O[IV\]]{}]{} ($\lambda$1402.06Å) line centres. Unfortunately, these individual line profiles are intrinsically broad, preventing our ability to distinguish between them. Therefore these two lines appear as a single, blended line. In Appendix \[sec:line\_component\] we find a single Gaussian component is sufficient to characterise the line profile.
- [C[IV]{}]{}: This line profile is a doublet ($\lambda$$\lambda$1548.20, 1550.78Å). Again, both components being intrinsically broad prevents individual detection, therefore the line profile is observed as a single, strong line. For the [C[IV]{}]{} line, we find a two component Gaussian to be preferable to characterise the line.
- [C[III\]]{}]{}: For the [C[III\]]{}]{} line ($\lambda$1908.73Å) we find a single component Gaussian to be sufficient to characterise the line. For almost all of the BOSS QSOs in our sample, the nearby, weak Si line ($\lambda$1892.03Å), which would appear on top of the much broader [C[III\]]{}]{} line is not resolvable[^9]. Furthermore, the [C[III\]]{}]{} line may also be contaminated by a continuum of low ionisation [Fe]{} lines, however, unlike [Mg[II]{}]{} (see next) the relative impact on the [C[III\]]{}]{} line is minor.
- [Mg[II]{}]{}: While the [Mg[II]{}]{} line ($\lambda$2798.75Å) is observable within the wavelength range of our BOSS QSO sample, we choose not to fit this emission profile. This is because of the contamination from the [Fe]{} pseudo-continuum. Though [Fe]{} emission templates exist [e.g. @Vestergaard:2001p3921], the fitting procedure is complicated. One must simultaneously fit several spectral regions in order to calibrate the true flux level of the [Fe]{} pseudo-continuum in order to remove it and obtain an estimate of the [Mg[II]{}]{} emission line. Since we aim to simultaneously fit the full QSO spectrum, this approach would add degeneracies to the model (e.g. between the true continuum, and the [Fe]{} pseudo-continuum). While this can be overcome, it requires first fitting the QSO continuum blueward of [Mg[II]{}]{}, before fitting the [Fe]{} pseudo-continuum, which reduces the flexibility of our model.
In addition to these lines above, we also consider several other lines all modelled by a single Gaussian. These include: (i) [N[V]{}]{} ($\lambda$$\lambda$1238.8, 1242.8Å), which is an important line for characterising the [Ly$\alpha$]{} line profile as it can be degenerate with the broad [Ly$\alpha$]{} line component (ii) [Si[II]{}]{} ($\lambda$1262.59Å), (iii) the O[I]{}/Si[II]{} blended complex ($\lambda$1304.35, $\lambda$1306.82Å), (iv) C[II]{} ($\lambda$1335.30Å), (v) [He[II]{}]{} ($\lambda$1640.42Å) (vi) O[III]{} ($\lambda$1663.48Å) and (vii) Al[III]{} ($\lambda$1857.40Å).
Continuum and emission line template {#sec:template}
------------------------------------
In this work, we fit the rest-frame wavelength range ($1180{\rm \AA} < \lambda < 2300{\rm \AA}$) with a single power-law for the QSO continuum, $$\begin{aligned}
\label{eq:continuum}
f_{\lambda} = f_{1450}\left(\frac{\lambda}{1450{\rm \AA}}\right)^{\alpha_{\lambda}} {\rm erg\,cm^{-2}\,s^{-1}\,\AA^{-1}},\end{aligned}$$ where $\alpha_{\lambda}$ describes the spectral slope of the continuum and $f_{1450}$ is the normalisation of the QSO flux which we choose to measure at 1450Å. While we normalise the QSO flux at 1450Å, we allow this quantity to vary within our MCMC fitting algorithm (by adding a small, variable perturbation). While this quantity does not depart greatly from the original normalised value, it allows us to compensate for situations where the nearby region around 1450Å might be impacted by line absorption or a noise feature from the spectrograph.
Secondary (broken) power-law continua have been fit to QSO spectra at $\lambda > 4500$Å for an independent red continuum slope [e.g. @VandenBerk:2001p3887; @Shang:2007p4862], however, this is beyond our QSO sample wavelength coverage. Typically, a broken power-law continuum is also adopted blueward of [Ly$\alpha$]{}, visible in low-$z$ HST spectra [e.g @Telfer:2002p5713; @Shull:2012p5716]. In this work, we do not consider a different slope near [Ly$\alpha$]{} as we only fit down to 1180Å, therefore there is insufficient information to include a secondary component. By only considering a single power-law through the [Ly$\alpha$]{} line profile, we may bias our results slightly on fitting the [Ly$\alpha$]{} broad component. However, provided we are consistent in our usage of this QSO continuum between this fitting approach and the reconstruction method, this should not impact our results.
As mentioned in the previous section, we model each emission line component with a Gaussian profile. Following @Kramer:2009p920 the total flux for each component can be defined as, $$\begin{aligned}
\label{eq:flux}
F_{i} = a_{i}\,{\rm exp}\left[ - \frac{(\lambda - \mu_{i})^{2}}{2\sigma^{2}_{i}}\right],\end{aligned}$$ where $a_{i}$ describes the amplitude of the line peak, $\mu_{i}$ is the location of the line centre in Å, $\sigma_{i}$ is the width of the line in Å and the subscript ‘$i$’ denotes the specific line species (e.g. [Ly$\alpha$]{}). Note that within this work, the peak amplitude is always normalised by the continuum flux at $1450$Å, $f_{1450}$, therefore it is always a dimensionless quantity. More intuitively, the line centre location, $\mu_{i}$, can be written in terms of a velocity offset relative to the systemic line centre, $$\begin{aligned}
v_{{\rm shift},i} = c\,\frac{\mu_{i} - \lambda_{i}}{\lambda_{i}} \,{\rm km/s},\end{aligned}$$ and the line width can be expressed as $$\begin{aligned}
\label{eq:width}
\sigma_{i} = \lambda_{i}\left(\frac{{\rm LW}}{c}\right)\,{\rm \AA},\end{aligned}$$ where LW is the line width measured in km/s. Throughout Equations \[eq:flux\]-\[eq:width\], both $\lambda$ and $\lambda_{i}$ are measured in the rest frame.
Each Gaussian line component can therefore be fully described by its three component parameters, the line width, peak amplitude and velocity offset. In total, we fit each QSO with two continuum parameters, two double component Gaussians ([Ly$\alpha$]{} and [C[IV]{}]{}), and 9 single component Gaussian profiles, resulting in a total of 41 continuum and emission line parameters.
Identifying absorption features {#sec:absorb}
-------------------------------
Intervening diffuse neutral hydrogen or larger column density absorbers (smaller than DLAs) along the line of sight can produce narrow absorption features. Furthermore, metal pollution in stronger absorption systems can additionally result in narrow absorption features appearing in the observed profiles of the emission lines.
These narrow features, if not measured and accounted for, can artificially bias the shape and peak amplitude of the emission line components. Therefore, in this section we outline our approach for identifying these features in a clean and automated manner:
- Identify all flux pixels that are a local minima within a 2Å region surrounding the central pixel in question. This choice of 2Å is arbitrary, but is selected to be sufficiently broad to ignore features that might arise from noise fluctuations.
- Construct a horizontal line of constant flux that begins from the global flux minimum.
- Incrementally increase this line of constant flux, recording the depth and width of each absorption feature enclosed by the line of constant flux.
- If the depth becomes larger than 3$\sigma$ (5$\sigma$ in the vicinity of [Ly$\alpha$]{}) of the observed error in the flux at that pixel and the candidate absorption line has remained isolated (i.e. not overlapped with a nearby feature) it is then classified as an absorption line.
Our adopted choice of 3$\sigma$ (5$\sigma$ near [Ly$\alpha$]{}) arises after rigorously testing this pipeline against the absorption features that were visually identified, until the vast majority of the absorption lines could be robustly located by this procedure. Across the full QSO sample, we found a broad range in the number of features identified within each individual QSO, from only a few up to $\sim40$. For each identified absorption feature, a single Gaussian (described by its own three parameters) is assigned and is simultaneously fit with the continuum and emission line profiles outlined previously. Therefore, the total number of parameters to be fit per QSO varies.
MCMC sampling the QSO template {#sec:MCMC}
------------------------------
We fit each QSO within a Bayesian MCMC framework, using the $\chi^{2}$ likelihood function to determine the maximum likelihood fit to the QSO spectrum. This choice enables us to fully characterise any potential model degeneracies between our model parameters, while also providing the individual probability distribution functions (PDFs) for each model parameter. In this work, we utilise the publicly available MCMC python code <span style="font-variant:small-caps;">CosmoHammer</span> [@Akeret:2012p842] built upon <span style="font-variant:small-caps;">EMCEE</span> [@ForemanMackey:2013p823] which is based on the affine invariant MCMC sampler [@Goodman:2010p843].
Assuming flat priors across our $>50$ model parameters and fitting the full high-resolution BOSS spectrum within the wavelength range ($1180{\rm \AA} < \lambda < 2300{\rm \AA}$) simultaneously is computationally inefficient. Instead, we perform an iterative procedure to boost the computational efficiency which we outline below:
- Normalise the full QSO spectrum at 1450Å and then fit the two continuum parameters, $f_{1450}$ and $\alpha_{\lambda}$ within a set of selected wavelength ranges of the full spectrum which are minimally contaminated by emission lines. We choose to fit the QSO continuum within the regions \[1275, 1295\], \[1315, 1330\], \[1351, 1362\], \[1452, 1520\], \[1680, 1735\], \[1786, 1834\], \[1970, 2040\] and \[2148, 2243\] Å.
- Break the full QSO spectrum up into wavelength regions centred around the emission lines. We choose \[1180, 1350\] Å centred on [Ly$\alpha$]{}, \[1350, 1450\] Å centred on [Si[IV]{}+O[IV\]]{}]{}, \[1450, 1700\] Å centred on [C[IV]{}]{} and \[1700, 1960\] Å centred on [C[III\]]{}]{}.
- Within each of these four regions, we take the continuum estimated from above, and then fit all emission lines and any absorption features which fall within the respective wavelength ranges. For each of these regions we then perform an MCMC fit.
- From the individual PDFs for each parameter we construct a flat prior across a much narrower allowed range, driven by the width of the individual distributions from the fitting above.
- Finally, we fit the entire QSO spectrum using the entire model parameter set and recover a maximum likelihood model which describes the full spectrum.
In Figure \[fig:QSOexample\], we provide an example of one of the BOSS QSOs from our full sample. In this figure we provide zoomed in panels of both the QSO continuum (red dashed curve, top panel) and the various emission lines which are simultaneously fit within our MCMC approach across all other panels. This figure highlights that any notable absorption features have been identified by our pipeline and accurately characterised (e.g. the [Ly$\alpha$]{} line profile in the left panel of the middle row). From Figure \[fig:QSOexample\] we note that this approach is able to fit the full spectrum. The maximum likelihood we obtained was $\chi^{2} = 3110$, for which we had 2899 bins (from the raw spectrum) and 71 free parameters, corresponding to a reduced $\chi^{2}$ of 1.1. For reference, this took $\sim$1 hour on a single processing core, which can be rapidly improved if a binned spectrum is used.
QSO fitting quality assessment {#sec:QA}
------------------------------
After fitting the entire sample with our full MCMC QSO fitting pipeline, we are in a position to construct a higher fidelity sample of QSOs for constructing our covariance matrix and investigating the correlations amongst the emission line parameters. This is performed by visually inspecting our QSO sample, applying a simple selection criteria. Following the completion of this selection process, we produce two separate QSO samples, one classified as ‘good’ and another classified as ‘conservative’, the details of which we discuss below (in Appendix \[sec:QSO\_QA\] we provide a few select examples to visually highlight this subjective process). The criteria are outlined as follows:
- We remove all QSOs with a poor characterisation of the continuum. These include QSOs with a positive spectral index, or a clear departure from a single power-law continuum. Only a handful of QSOs exhibit this behaviour. See Figure \[fig:QA\_RemovedContinuum\] for some examples.
- We further remove any QSOs which have either (i) missing sections of flux that overlap with any of the emission lines (ii) a sufficient number of absorption features which cause a loss of confidence in the fitting of the [Ly$\alpha$]{} profile (iii) either an intervening dense neutral absorber or sufficiently strong/broad absorption blueward of line centre which can impact the broad component of either the [Ly$\alpha$]{} or the [C[IV]{}]{} lines (iv) a [Ly$\alpha$]{} line region that is not well fit or characterised by our double component Gaussian, which might arise from numerous absorption profiles or the lack of a prominent [Ly$\alpha$]{} peak.
- Following the removal of these contaminants from our sample, we are left with 2,653 QSOs, which we call our ‘conservative’ sample[^10]. These can still include QSOs which might have absorption features centred on the [Ly$\alpha$]{} line centre, or absorption line complexes which might contaminate large sections of the [Ly$\alpha$]{} line profile (see Figure \[fig:QA\_Conservative\] for examples).
- We then apply a secondary criteria, which preferentially selects QSOs which contain (i) fewer absorption features and (ii) no absorption features on the line centre. This results in a final sample of 1673 QSOs, which we refer to as our ‘good’ sample (see Figure \[fig:QA\_Good\] for examples).
The major difference between the ‘good’ and ‘conservative’ sample is that the ‘conservative’ sample will contain QSO spectra for which we have less confidence in either the identification and fitting of one, or several of the emission lines due to the presence of absorption features. Ultimately, if our claim that any correlations amongst the emission lines are a universal property of the QSOs then the covariance matrices of the two should be almost identical, with the ‘conservative’ sample containing additional scatter (slightly weaker correlations). In the next section, we construct our covariance matrices and investigate this further.
Data Sample Covariance {#sec:Covariance}
======================
With the refined, quality assessed QSO spectra from the previous section, we now construct our covariance matrix to characterise all correlations amongst the various emission lines.
The covariance matrix {#sec:covariance}
---------------------
In performing the quality assessment of our QSO spectra, we note that a number of the weaker emission lines are not always well characterised or resolved. This becomes more prevalent for the QSOs nearer to our sample limit of S/N $=15$, which correspond to the highest density of QSOs in our sample. It would be interesting to investigate correlations between the strong high ionisation lines and the weaker low ionisation lines, as well as correlations amongst these two classifications. However, since these lines are not always readily available in our QSOs we refrain from doing so. Regardless, by still attempting to fit these weaker lines we retain the flexibility of the MCMC approach, and more importantly this enables the QSO continuum to be estimated to a higher accuracy.
In light of this, we construct our covariance matrix from a subset of these emission lines: the most prominent emission lines that should always be resolvable in a lower S/N or lower resolution QSO spectrum (as in the near-IR for $z>6$ QSOs). The lines we identify are [Ly$\alpha$]{}, [Si[IV]{}+O[IV\]]{}]{}, [C[IV]{}]{} and [C[III\]]{}]{}. Since for [Ly$\alpha$]{} and [C[IV]{}]{} we allow a broad and narrow component, we recover an 18$\times$18 covariance matrix. We choose to exclude the [N[V]{}]{} line from our covariance matrix for two reasons: (i) a clear, identifiable [N[V]{}]{} line is not always present in our QSO spectra, therefore including it would artificially reduce any visible correlation and (ii) the ultimate goal of this work is the reconstruction of the [Ly$\alpha$]{} line whereby including the [N[V]{}]{} line to the reconstruction process would only increases the complexity (the [N[V]{}]{} line should in principle be recoverable from the high-$z$ or [Ly$\alpha$]{} obscured QSO).
In Figure \[fig:CovarianceMatrix\], we present the correlation coefficient matrix, which is obtained from the full covariance matrix from our ‘good’ sample of 1673 QSOs. In constructing this, we assume the standard format for the covariance matrix, $\bmath{\Sigma}_{ij}$, given by, $$\begin{aligned}
\bmath{\Sigma}_{ij} = \frac{1}{N-1}\sum^{N}_{i}(\textbfss{X}_{i} - \bmath{\mu}_{i})(\textbfss{X}_{j} - \bmath{\mu}_{j}),\end{aligned}$$ where $\textbfss{X}_{i}$ is the data vector for the full QSO sample and $\bmath{\mu}$ is the mean data vector for the $i$th parameter. For this data vector we use the values from the parameter set which provide the maximum likelihood fit to each QSO. The correlation coefficient matrix, $\textbfss{R}_{ij}$, is then defined in the standard way, $$\begin{aligned}
\textbfss{R}_{ij} = \frac{\textbfss{C}_{ij}}{\sqrt{\textbfss{C}_{ii}\textbfss{C}_{jj}}},\end{aligned}$$ where each diagonal entry, $\textbfss{R}_{ij}$, corresponds to the correlation coefficient between the $i$th and $j$th model parameters.
Note, for the covariance matrix, we do not include the two continuum parameters. Firstly, for the parameters defining the emission line peak we define this parameter as the normalised peak value, where it is normalised by the continuum flux at $f_{1450}$. Therefore, if included, this would be completely degenerate with the peak amplitudes. Some correlation with continuum spectral index is expected, given that external (to the broad-line region) reddening will simultaneously weaken the bluer UV lines and redden the continuum spectral index. However, mild reddening is seen in only $\sim10-20$ per cent of SDSS quasars [e.g. @Richards:2003p4639; @Hopkins:2004p4640] so this has little effect on the correlations (we find very weak correlations on the order of 10 per cent). Therefore we do not report them, as these provide no additional information with respect to the individual line correlations.
Interpreting the covariance matrix {#sec:interpretation}
----------------------------------
In order to aid the interpretation of the correlation matrix, we divide Figure \[fig:CovarianceMatrix\] into the four emission line species (denoted by black curves), while the narrow and broad line species are further separated by black dashed lines. Positive correlations are represented by a decreasing (weakening) shading of red, with white representing no correlation and an increasing blue scaling denotes a strengthening of the anti-correlation. Using the upper half of the correlation matrix enables a faster analysis of the correlation patterns amongst the emission line parameters, and the lower half reports the numerical value of the correlation (or anti correlation). For the most part, each 3$\times$3 sub-matrix returns the same correlations and anti-correlations amongst the peaks, widths and velocity offsets, but with varying degrees of strength.
Firstly, we note that with this correlation matrix it is straightforward to notice that all pairs of peak height parameters are positively correlated. This is to be expected as the peak amplitude is positively correlated with the QSO luminosity. This correlation between the peak heights is the strongest of the trends. Most importantly for this work, the correlations amongst the [Ly$\alpha$]{} peak height parameters are the strongest. For example, we find the strongest correlation ($\rho = 0.8$) between the peak height of the [Ly$\alpha$]{} narrow component and the associated peak height of the [C[IV]{}]{} narrow component. In Figure \[fig:PeakCorrelation\], we provide the 2D scatter plot for this strong correlation (central panel) and the 1D marginalised PDFs for the [Ly$\alpha$]{} (top) and [C[IV]{}]{} (right). The green solid and dashed contours in the central panel denote the 68 and 95 per cent 2D marginalised joint likelihood contours, which describe the relative scatter amongst our sample of QSOs. In the 1D marginalised PDFs, the solid black curves are the histograms of the sample of QSOs, while the blue solid curves are an approximated 1D Gaussian for that associated parameter. Note in this figure, and in subsequent figures, these blue curves are not a fit directly to the raw data, rather instead they are approximations of a Gaussian PDF with 1$\sigma$ scatter equivalent to the raw data.
Returning to the correlation matrix, we additionally recover a relatively strong trend (anti-correlation) between the peak height and the line width. In Figure \[fig:PeakWidthCorrelation\], we provide the strongest of these anti-correlations ($\rho=-0.74$) which is between the [Ly$\alpha$]{} peak amplitude of the narrow line component and the width of the single component [C[III\]]{}]{} line. We observe, that for an increasing peak amplitude for the [Ly$\alpha$]{} narrow component, a decreasing of the [C[III\]]{}]{} line width. This behaviour could be inferred as a result of the Baldwin effect [@Baldwin:1977p5910]. For an increasing peak amplitude (i.e. QSO luminosity), we expect a weaker broad-line emission. Physically, a plausible scenario to describe this could be the over-ionisation of the inner broad-line region by the continuum in high luminosity QSOs, resulting in carbon being ionised into higher order species (i.e. no [C[IV]{}]{}/[C[III\]]{}]{}). For these high luminosity QSOs, the high ionisation lines may then predominantly arise at larger radii where the velocity dispersion is lower, producing smaller line widths [e.g. @Richards:2011p6031]. Alternately, it could arise from systematics from our line fitting. As a result of the decreasing equivalent widths with increasing QSO luminosity, it could be that the single component Gaussian for the [C[III\]]{}]{} does not characterise the line profile as accurately. With a less prominent broad-line component, the emission from the wings could be underestimated, producing a narrower [C[III\]]{}]{} line.
We additionally recover a trend for a positive correlation between the widths of the various line species, though this is a relatively weaker trend. However, for the broad and narrow components for the same line species, we find more moderate correlations ($\rho=0.54$ for [Ly$\alpha$]{} and $\rho=0.62$ for [C[IV]{}]{}). It is difficult to interpret these correlations though, as these broad and narrow lines are simultaneously fit, and therefore in principle could be degenerate.
@Shang:2007p4862 investigated correlations amongst the [Ly$\alpha$]{}, [C[IV]{}]{} and [C[III\]]{}]{} lines for a significantly smaller sample of 22 QSOs. These authors recover moderate to strong correlations between a few of their emission line width full width half maxima (FWHMs). They find [Ly$\alpha$]{}–[C[IV]{}]{} ($\rho=0.81$) and [C[IV]{}]{}–[C[III\]]{}]{} ($\rho=0.53$). If we equate this FWHM as the width of the narrow component of our emission line profiles we find correlations between the same species of [Ly$\alpha$]{}–[C[IV]{}]{} ($\rho=0.35$) and [C[IV]{}]{}–[C[III\]]{}]{} ($\rho=0.46$)[^11]. While we recover an equivalent correlation for the [C[IV]{}]{}–[C[III\]]{}]{} lines, we find a significant discrepancy for the [Ly$\alpha$]{}–[C[IV]{}]{} lines. This potentially can be explained as an artificially large correlation owing to their small number of objects (22 compared to our 1683 QSOs)[^12]. Supporting this hypothesis, in @Corbin:1996p5640 for a larger sample of 44 QSOs, this is reduced to $\rho=0.68$ for [Ly$\alpha$]{}–[C[IV]{}]{}.
Additionally, @Kramer:2009p920 recovered a moderate positive correlation ($\rho=0.68$) between the velocity offsets for the [Ly$\alpha$]{}–[C[IV]{}]{} components. For the same combination, @Shang:2007p4862 found a stronger correlation of $\rho=0.81$. In Figure \[fig:VelocityCorrelation\], we provide our weaker results between the narrow components of the [Ly$\alpha$]{}–[C[IV]{}]{} lines for which we recover a moderate correlation of $\rho=0.35$. Once again, this lack of an equivalently strong correlation in our sample could arise purely from the fact that our sample contains two orders of magnitude more QSOs. @Shang:2007p4862 equivalently quote correlations of $\rho=0.45$ ([Ly$\alpha$]{}–[C[III\]]{}]{}) and $\rho=0.39$ ([C[IV]{}]{}–[C[III\]]{}]{}). Using the narrow lines for [Ly$\alpha$]{} and [C[IV]{}]{}, we find similarly moderate, but slightly stronger correlations of $\rho=0.54$ ([Ly$\alpha$]{}–[C[III\]]{}]{}) and $\rho=0.48$ ([C[IV]{}]{}–[C[III\]]{}]{}).
In the construction of the covariance matrix, and associated correlation matrix we have only used the maximum likelihood values, and ignored the relative errors for each parameter from the marginalised 1D PDFs. Our reasoning for this is that the amplitude of the error on the individual parameters is small with respect to the scatter that arises for each parameter across the full QSO sample. In principle however, we could perform an MCMC sampling of the full covariance matrix, allowing the means and all correlation coefficients to be free parameters. This would allow a more accurate characterisation of both the individual errors for each parameter and the total scatter across the full QSO sample, potentially tightening the correlations amongst the emission line parameters. However, this would require fitting $\frac{1}{2}N(N+1)$ parameters encompassing both the means of the individual parameters and all correlation coefficients (in our case, 171 free parameters).
Correlation of the QSO continuum parameters
-------------------------------------------
Though the continuum parameters are not included in the covariance matrix in Figure \[fig:CovarianceMatrix\], in Figure \[fig:Continuum\] we provide the 2D scatter between these two parameters. In our ‘good’ sample of QSOs, we recover only a relatively weak anti-correlation ($\rho=-0.22$) between the QSO continuum spectral index, $\alpha_{\lambda}$ and the normalisation at 1450 Å ($f_{1450}$). This weak anti-correlation likely arises due to dust reddening, which causes a drop in the normalisation, $f_{1450}$ and shallower spectral slopes. While both of these QSO parameters are well characterised by a Gaussian, the respective scatter in each parameter is considerable.
For our sample of ‘good’ (‘conservative’) QSOs, we recover a median QSO spectral index of $\alpha_{\lambda} = -1.30\pm0.37$ ($ -1.28\pm0.38$). In contrast, @Harris:2016p5028 construct a QSO composite spectrum with a wavelength coverage of ($800{\rm \AA} < \lambda < 3300{\rm \AA}$) from the same BOSS DR12 sample. Using $\sim100,000$ QSOs, these authors find a median spectral index for their QSO sample of $\alpha_{\lambda} = -1.46$, consistent with our results within the large scatter. However, these authors only fit the QSO continuum between 1440-1480Å and 2160-2230Å. Not fitting to the same spectral region likely results in a different recovered spectral slope. Furthermore, our estimate of $\alpha_{\lambda} = -1.30$ ($\alpha_{\nu} = -0.70$) is also consistent with the lower redshift samples of @Scott:2004p5709 ($\alpha_{\nu}=-0.56\substack{+0.38 \\ -0.28}$), @Shull:2012p5716 ($\alpha_{\nu}=-0.68\pm0.14$) and @Stevans:2014p5726 ($\alpha_{\nu}=-0.83\pm0.09$). Note, within all these works, considerable scatter in the spectral index is also prevalent.
Potential sample bias {#sec:Bias}
---------------------
In Sections \[sec:covariance\] and \[sec:interpretation\], we presented the correlation matrix and discussions on the relative trends between the emission lines and their relative strengths. However, these results were drawn from our refined, quality assessed ‘good’ QSO sample. In order to guard against a potential bias which may have arisen following our specific selection process, we additionally construct the correlation matrix for our ‘conservative’ QSO sample. This ‘conservative’ sample contains $\sim1000$ additional QSOs which are defined to be less robust than required for our ‘good’ sample. Therefore, this ‘conservative’ sample should contain more scatter amongst the recovered parameters, which would notably degrade the strength of the correlations relative to the ‘good’ sample if we have artificially biased our results.
In Figure \[fig:CovMatDiff\] we provide the matrix of the relative difference in the correlation coefficients between the ‘good’ and ‘conservative’ QSO samples. Here, we show the amplitude of the change in correlation coefficient between the two QSO samples. A positive (red) difference is indicative of the ‘good’ sample having a stronger correlation (either positive or anti-correlation), while a negative (blue) difference is indicative of the ‘conservative’ sample having a stronger correlation. Note, squares marked with a dot-dashed cross indicate a change between a positive and anti-correlation, which arises when the correlations are close to zero in either sample.
For the most part, the relative change in the correlation is minor, of the order of $|\Delta\rho| < 0.04$. More importantly, for any of the strong correlations and notable trends we discussed in the previous section, we observe no sizeable differences, with the ‘good’ sample providing on average slightly stronger correlations (as indicated by the prevalence of red squares). For example, the [Ly$\alpha$]{}–[C[IV]{}]{} peak heights of the narrow component from our ‘good’ sample was $\rho=0.80$, whereas for the ‘conservative sample it is $\rho=0.79$. Therefore, in the absence of any drastic differences in the strong and notable trends discussed previously between the two samples, it is clear that we have not biased our QSO covariance matrix in the construction of the ‘good’ sample. Given the relative amplitude of these differences between the two QSO samples is small, and tend to be slightly weaker for the ‘conservative’ sample, the reconstructed profile recovered from the covariance matrix should recover effectively the same best-fit [Ly$\alpha$]{} profile, with slightly broader errors owing to the increased scatter (reduced correlations).
Equally, to be prudent, we considered several other ways to divide our QSO sample. First, we performed the same analysis, comparing instead the ‘conservative’ minus ‘good’ sample (the $\sim1000$ QSOs from the ‘conservative’ sample not classified as ‘good’) and secondly, constructing two equally sized random samples from the ‘conservative’ sample. In both instances, we find the same strong correlations as in Figure \[fig:CovarianceMatrix\] with similar amplitude variations between the respective correlation matrices as shown in Figure \[fig:CovMatDiff\].
Finally, our results on line parameter correlations are sensitive to the choice of redshift estimate and any inherent biases in those estimates. We discuss this issue in detail in Appendix \[sec:systemic\_redshift\], but, in summary, we find that while different redshift estimates do result in slightly different covariance matrices, our general conclusions are robust and the redshift estimate we have chosen (the BOSS pipeline redshift, $z_{\rm pipeline}$) performs the best at the [Ly$\alpha$]{} profile reconstruction (see the next section).
[Ly$\alpha$]{} Reconstruction {#sec:Recon}
=============================
We now use this covariance matrix to reconstruct the [Ly$\alpha$]{} line profile. In this section we outline the reconstruction pipeline, and associated assumptions followed by an example of this approach.
Reconstruction method {#sec:Reconstruction}
---------------------
Within our reconstruction approach, we approximate the distribution of the emission line parameters (across the entire data-sample) as a Gaussian[^13]. In Figures \[fig:PeakCorrelation\]-\[fig:VelocityCorrelation\], we observe that for these six parameters shown, this approximation is well motivated. Clearly, showing the 18 individual 1D PDFs for each parameters would be uninformative, however, we confirm by eye that this approximation is valid for all model parameters. Note, that in some of these cases, a Gaussian approximation holds only after taking the logarithm, especially for the normalised peak amplitude (normalised by $f_{1450}$) and the emission line width. Importantly, if anything, these Gaussian approximations have a tendency to slightly overestimate the relative scatter within each parameter (see Figure \[fig:VelocityCorrelation\] for example), therefore, this assumption in fact turns out to be a conservative estimate of the true scatter. The Gaussian nature of the scatter in Figures \[fig:PeakCorrelation\]-\[fig:VelocityCorrelation\] and the conservative overestimation of the errors by our Gaussian approach lends confidence that our approach should not significantly underestimate the errors that one might recover from a fully Bayesian approach.
In order to perform the reconstruction, we assume that the QSO can be fit following the same procedure outlined in Section \[sec:Fitting\], except now we only fit the QSO red-ward of 1275 Å. Our choice of $\lambda > 1275$ Åis conservatively selected to be both close enough to [Ly$\alpha$]{} as possible, but with minimal to no contamination from emission line wings namely [Ly$\alpha$]{}, N[V]{} or Si[II]{}). Furthermore, given that this approach is best suited for recovering the intrinsic [Ly$\alpha$]{} profile from a [Ly$\alpha$]{} obscured or high-$z$ QSO, it is best to be sufficiently far from any possible contamination of the [Ly$\alpha$]{} line region.
We can then define the $N$ dimensional parameter space (i.e. our 18 emission line parameters outlined previously) as an $N$ dimensional likelihood distribution given by, $$\begin{aligned}
\label{eq:ML}
\mathcal{L} = \frac{1}{(2\pi)^{N/2}|\bmath{\Sigma}|}{\rm exp}\left[\frac{1}{2}(\bmath{x}-\bmath{\mu})^{\mathsf{T}}\bmath{\Sigma}^{-1}(\bmath{x}-\bmath{\mu})\right].\end{aligned}$$ Here, $\bmath{\Sigma}$ is the recovered QSO covariance matrix (Section \[sec:covariance\]), $\bmath{\mu}$ is the data vector of the means obtained from the full QSO sample for each of the individual line profile parameters and $\bmath{x}$ is the data vector measured from our MCMC fitting algorithm for the individual obscured QSO spectrum.
After the [Ly$\alpha$]{} obscured or high-$z$ QSO has been fit following our fitting procedure, the recovered best-fit values for the unobscured emission line parameters are folded into Equation \[eq:ML\]. That is, we recover the best-fit estimates of the [Si[IV]{}+O[IV\]]{}]{}, [C[IV]{}]{} and [C[III\]]{}]{} emission lines and evaluate Equation \[eq:ML\] to collapse the 18 dimensional likelihood function into a simple, six dimensional likelihood function describing the six unknown [Ly$\alpha$]{} emission line parameters (two Gaussian components each defined by three parameters). The maximum likelihood of this six dimensional function then describes the best-fit reconstructed profile for the [Ly$\alpha$]{} emission line, while the full six dimensional matrix contains the correlated uncertainty.
Reconstruction example {#sec:ReconExample}
----------------------
We now present an example to highlight the performance of our approach. In order to do this, we choose the same QSO we showed in Figure \[fig:QSOexample\], fitting for $\lambda > 1275$ Åand recovering the six dimensional estimate for the [Ly$\alpha$]{} line profile. In restricting our fitting algorithm to $\lambda > 1275$ Å we are not accessing all the information that was used by the full fit to estimate the QSO continuum. While, this does not affect the recovery of the [Ly$\alpha$]{} peak profile itself, the [Ly$\alpha$]{} profile plus continuum could be affected. In Appendix \[sec:continuum\_comparison\] we test this assumption, finding that the QSO continuum parameters can be recovered equivalently from these two approaches, with a small amount of scatter in the QSO spectral index.
Before providing the full reconstructed profile, we first recover the individual marginalised 1D PDFs for each of the six [Ly$\alpha$]{} emission line parameters to better visualise the relative size of the errors. In order to obtain the recovered 1D PDFs for the [Ly$\alpha$]{} profile parameters, we marginalise the six dimensional likelihood function over the remaining five [Ly$\alpha$]{} parameters. In Figure \[fig:PDFs\_Recon\] we present these recovered 1D marginalised PDFs, showing in the top row the [Ly$\alpha$]{} broad line component and in the bottom row, the narrow line component. The vertical black dashed lines represent the recovered values from fitting the full QSO in Figure \[fig:QSOexample\], and the coloured curves represent the recovered 1D marginalised PDFs for each [Ly$\alpha$]{} parameter for the ‘good’ sample (red) and ‘conservative’ sample (blue). It is clear that both QSO samples recover almost identical best-fit values for each [Ly$\alpha$]{} parameter, highlighting that both samples are characterised by the same correlations within the covariance matrix. Furthermore, the choice in constructing a ‘good’ sample is more evident here as consistently the ‘good’ sample provides marginally narrower constraints. For the most part, the six reconstructed [Ly$\alpha$]{} parameters are recovered within the 68 percentile limits of the original fit to the QSO, with the exception being the velocity offset of the broad [Ly$\alpha$]{} component which is only slightly beyond this limit. Otherwise, the five remaining [Ly$\alpha$]{} parameters are effectively centred around the expected value from the full fit to the same QSO (Figure \[fig:QSOexample\]) which should enable a relatively robust recovery of the full [Ly$\alpha$]{} line profile.
In Figure \[fig:Profile\_Recon\] we provide the full reconstructed [Ly$\alpha$]{} line profile. In the top left panel we show the reconstructed [Ly$\alpha$]{} line profile obtained from the ‘good’ QSO sample, whereas in the top right panel is the reconstructed profile from the ‘conservative QSO sample. In all figures, we present 100 reconstructed [Ly$\alpha$]{} profiles denoted by the thin grey curves which are randomly drawn from the full posterior distribution. This small subset of reconstructed profiles highlight the relative scale of the variations in the total [Ly$\alpha$]{} line profile peak height, width and position. The black curve is the raw data from the observed QSO whereas the white curve is the original fit to the full QSO spectrum as shown in Figure \[fig:QSOexample\].
In both, we find the total shape of the reconstructed [Ly$\alpha$]{} line profile to match extremely well with the original full fit to the same QSO, highlighting the strength and utility of this covariance matrix reconstruction method. Both the ‘good’ and ‘conservative’ samples recover almost the identical reconstructed profile, though the ‘conservative’ QSO sample provides slightly broader errors. Note that for both these QSO samples, the reconstructed [Ly$\alpha$]{} line profile is systematically below the original fit to the same QSO, indicated by the original QSO fit (white curve) being above the highest density of line profiles[^14]. However, this systematic offset is only minor (less than 10 per cent in the normalised flux), well within the errors of the reconstruction. Referring back to Figure \[fig:PDFs\_Recon\], we can see that this underestimation appears due to the narrow component of the [Ly$\alpha$]{} peak amplitude (bottom, central panel).
Improving the reconstruction with priors {#sec:fluxprior}
----------------------------------------
We presented in Figure \[fig:Profile\_Recon\] our best-fit reconstructed [Ly$\alpha$]{} profiles to a representative QSO drawn from our full sample. However, note that in this reconstruction method we have only used information from the QSO spectrum above $\lambda > 1275$Å. In doing this, for the case of our example QSO we found our maximum likelihood estimates to slightly underestimate (by less than 10 per cent) the original [Ly$\alpha$]{} line profile. At the same time, the total 68 per cent marginalised likelihoods of the reconstructed profiles are relatively broad. Motivated by this, we investigate whether we can provide an additional prior on the [Ly$\alpha$]{} reconstruction profile to further improve the robustness of the recovered profile and to reduce the relative scatter.
In the top panels of Figure \[fig:Profile\_Recon\], redward of [Ly$\alpha$]{} we see that the reconstructed [Ly$\alpha$]{} line profile drops well below the observed flux of the original QSO. Given we are only reconstructing the [Ly$\alpha$]{} line, this is to be expected as we are not recovering or fitting the [N[V]{}]{} emission line. However, in both [Ly$\alpha$]{} obscured and high-$z$ QSOs, the [N[V]{}]{} emission line should be relatively unobscured, therefore, the observed flux within this region could be used as a relative prior on the overall [Ly$\alpha$]{} flux amplitude.
We therefore include a flux prior[^15] into our [Ly$\alpha$]{} reconstruction by performing the following steps:
- As before, we fit the QSO at $\lambda > 1275$Å, recovering the QSO continuum and all emission line profiles necessary for our covariance matrix approach.
- Using these estimates, we collapse the 18-dimensional covariance matrix into a six dimensional estimate of the intrinsic [Ly$\alpha$]{} emission line profile.
- We then jointly MCMC sample the observed QSO spectrum within the range $1230 < \lambda < 1275$Å. We fit the [N[V]{}]{} and [Si[II]{}]{} lines at the same time sampling from our six dimensional reconstructed [Ly$\alpha$]{} likelihood function obtained from the $\lambda > 1275$Åfit. Fitting to the observed QSO flux, and using the observed noise in the spectrum, we obtain a maximum likelihood for the reconstructed profile. In other words, we require the reconstructed [Ly$\alpha$]{} line profiles to fit the observed spectrum over the range $1230 < \lambda < 1275$Å.
Implementing this prior on the observed flux closer to $\lambda = 1230$Å[^16] accesses additional information on the [Ly$\alpha$]{} profile that was not available through our original $\lambda > 1275$Å reconstruction method. Near $\lambda = 1230$Å, there should be a contribution from the [Ly$\alpha$]{} broad line component, which is somewhat degenerate with the [N[V]{}]{} line (as can be seen in Figure \[fig:QSOexample\]). By simultaneously fitting the [N[V]{}]{} line and the [Ly$\alpha$]{} likelihood function, we use this additional information to place a prior on the [Ly$\alpha$]{} broad line component, which should then reduce the overall scatter in the six dimensional [Ly$\alpha$]{} likelihood function.
Referring back to Figure \[fig:PDFs\_Recon\] we provide an example of this prior applied to the same QSO fit and reconstructed previously. In Figure \[fig:PDFs\_Recon\], the yellow curves represent the recovered 1D marginalised PDFs for each of the six [Ly$\alpha$]{} emission line parameters. Immediately, it is clear that the application of this prior further reduces the relative error in the recovered PDFs. Furthermore, these PDFs remain centred on the originally recovered values, highlighting we have not biased our reconstruction method.
In the bottom panel of Figure \[fig:Profile\_Recon\] we present the full reconstructed [Ly$\alpha$]{} profile after the addition of this prior. The red curve indicates the fit to the QSO within $1230 < \lambda < 1275$Å, which we have used as our prior to improve the reconstruction of the [Ly$\alpha$]{} line profile. At $\lambda < 1230$Å, we then have the same 100 thin grey curves representing the full posterior distribution of the reconstructed [Ly$\alpha$]{} likelihood profiles. By applying this additional prior on the total observed flux, we reduce the overall scatter in the reconstructed profile. The maximum likelihood profiles (thin grey curves) now provide a more robust match to the observed QSO.
Statistical performance of the reconstruction pipeline
------------------------------------------------------
Thus far we have only applied our reconstruction pipeline to recover the [Ly$\alpha$]{} line profile from a single, test example selectively drawn from our ‘good’ sample of 1673 QSOs. In order to statistically characterise the performance of the [Ly$\alpha$]{} reconstruction pipeline across the QSO sample, in Figure \[fig:Recon\_All\] we present 2D scatter plots of the maximum a posteriori (MAP) estimates[^17] of the reconstructed [Ly$\alpha$]{} emission line parameters compared to the originally obtained values from the full MCMC fit to the QSO (without masking out the [Ly$\alpha$]{} line). For each of the six [Ly$\alpha$]{} emission line parameters, we highlight the 68 and 95 percentiles of the joint marginalised likelihoods for the distributions by the green solid and dashed contours, respectively. Additionally, the red dashed curve demarcates the one-to-one line, along which all QSOs would sit if the reconstruction profile worked idealistically.
Across the six [Ly$\alpha$]{} panels, we find strong agreement ($\rho > 0.7$) amongst half of our [Ly$\alpha$]{} line parameters, those being the peak amplitudes of both the [Ly$\alpha$]{} broad and narrow components and the velocity offset for the [Ly$\alpha$]{} narrow line component. For the remainder of the parameters we find moderate to weaker recovery of the original line parameters. However, note that in order to keep this figure as clear as possible we are only providing the MAP estimates. Within Figures \[fig:PDFs\_Recon\] and \[fig:Profile\_Recon\] we found that the relative scatter on the reconstructed [Ly$\alpha$]{} line profile parameters was notable, and therefore the majority of the reconstructed parameters are within the 68 per cent marginalised errors.
The reconstructed [Ly$\alpha$]{} profile parameters highlighted here reflect the correlations recovered from the covariance matrix in Figure \[fig:CovarianceMatrix\]. We found strong correlations in the peak amplitudes of the [Ly$\alpha$]{} profiles, and the narrow component velocity offset. The lack of a strong correlation for the width of the [Ly$\alpha$]{} line in our covariance matrix, translates to weaker recovery of these parameters. In principle, these weaker correlations could be further strengthened by adding an appropriate prior on the line widths motivated by the statistical distributions recovered from the full sample, or other line properties such as correlations between the equivalent widths.
In Figure \[fig:Recon\_All\], there is also slight evidence for a bias in the reconstructed parameters, as highlighted by the orientation of the green contours (68 and 95 percentiles of the reconstructed parameter distributions) relative to the reference one-to-one line. However, this could artificially arise as the increase/decrease in any one of these [Ly$\alpha$]{} line parameters could be compensated for by respective changes in others (i.e. model degeneracies), whereas the full six dimensional [Ly$\alpha$]{} likelihood function takes these model degeneracies into account when estimating the full reconstructed [Ly$\alpha$]{} profile.
In order to better illustrate the full reconstruction of the joint [Ly$\alpha$]{} parameter likelihoods, in Figure \[fig:FluxDistribution\] we show the information from the six individual [Ly$\alpha$]{} line parameters as a single, total measured [Ly$\alpha$]{} line flux at two arbitrarily defined locations blue (1205Å) and redward (1220Å) of the [Ly$\alpha$]{} line centre. We compare the total reconstructed [Ly$\alpha$]{} line flux against the measured [Ly$\alpha$]{} line flux from our full fit to the same QSO, providing the reference one-to-one relation as the red dashed curve, and the grey shaded region encompasses the region in which the reconstructed [Ly$\alpha$]{} line flux is within 15 per cent of the measured flux. Immediately obvious from this figure is that there are no apparent biases in the reconstruction process, i.e. we neither systematically over nor underestimate the reconstructed [Ly$\alpha$]{} line profile. This figure highlights the strength of the [Ly$\alpha$]{} reconstruction process. Redward of [Ly$\alpha$]{} (at 1220Å), closer to our flux prior at 1230Å, we find that $\sim90$ per cent of all reconstructed [Ly$\alpha$]{} line profiles have a recovered flux within 15 per cent. As one would expect, the scatter increases blueward of [Ly$\alpha$]{} (at 1205Å), however we still find the reconstructed [Ly$\alpha$]{} line flux to be within 15 per cent for $\sim85$ per cent of our sampled QSOs. This highlights statistically, that the reconstruction process performs an excellent job of recovering the full [Ly$\alpha$]{} line profile.
Potential Applications {#sec:Discussions}
======================
In this work, we have developed an MCMC fitting algorithm for the sole purpose of characterising the QSO continuum and the emission line profiles within the range $1180{\rm \AA} < \lambda < 2300{\rm \AA}$. Our goal was the construction of a covariance matrix to reconstruct the [Ly$\alpha$]{} line profile. However, due to the flexibility of the MCMC approach, many other applications could benefit from such a pipeline.
Firstly, we were only interested in correlations amongst the strongest, high ionisation emission line parameters for our covariance matrix. However, various properties of QSOs can be extracted from accurate recovery of the line widths and ratios. For example, the QSO metallicity has been estimated from measuring the [N[V]{}]{}/[C[IV]{}]{}, [N[V]{}]{}/He[II]{} and the [Si[IV]{}+O[IV\]]{}]{}/[C[IV]{}]{} line ratios in samples of QSOs from @Nagao:2006p4776 ($2 < z < 4.5$) and @Juarez:2009p4775 ($4 < z < 6.4$). Using the same line ratios, we could recover estimates for the metallicities for all QSOs within our measured sample.
Several other emission line ratios (e.g. the R23 parameter, \[O[III]{}\]/\[O[II]{}\], [C[IV]{}]{}/He[II]{}, N-based ratios) can additionally be used as proxies for QSO metallicity [e.g @Nagao:2006; @Matsuoka:2009; @Batra:2014]. By extending our QSO MCMC framework, many other emission lines can be obtained and characterised to improve the QSO metallicity estimates. Crucially, this MCMC approach, would enable a large dataset of QSOs to be rapidly explored. The extension to measuring the metallicities of the QSOs would enable quantities such as the SMBH mass to be recovered through existing correlations between the emission line FWHMs.
In Section \[sec:absorb\] we have outlined our method to identify and fit absorption features (see e.g. @Zhu:2013 for a more robust approach). While in this work these have been considered contaminants, the prevalence and measurement of these lines could be used to infer properties of the metallicities within intervening absorbers [e.g @RyanWeber:2006; @Becker:2009; @DOdorico:2013] and to reveal the presence of massive outflows of ionised gas from their nuclei [e.g. @Crenshaw:2003 and reference therein]. For example, we have thrown out any QSOs with strong intervening absorption from systems such as DLAs, however, analysing these sources with our MCMC fitting algorithm could yield measurements on the internal properties of these absorption systems.
Assuming the input quasars to our MCMC are representative of the quasar population as a whole, and further that the UV spectral properties of quasars do not evolve with redshift, our model can be used to predict the intrinsic distribution of quasar spectra at redshifts beyond those from which the model was calibrated ($z\sim2.5$). A similar procedure has been often employed to characterise the colour selection efficiency of quasar surveys [e.g @Fan:1999], although the correlations obtained through our MCMC approach provide more detailed reconstruction of quasar emission features and hence more reliable colour models. Particularly since the [Ly$\alpha$]{} line plays such a key role in the selection of high-$z$ quasars, our model could be used to identify quasars missing from current surveys due to selection effects and provide more robust statistics for high-$z$ quasar luminosity functions.
In addition, by recovering an estimate for the intrinsic [Ly$\alpha$]{} emission line profile one can investigate the QSO proximity effect. This approach requires an estimate of the intrinsic QSO luminosity, coupled with the modelling of the [Ly$\alpha$]{} forest. Within the sphere of influence of the QSO, the photoionisation background is higher than the mean background permeating the IGM. Modelling the transition between the mean IGM background and the drop-off in QSO luminosity has been used by several authors to recover estimates of the photoionisation background in the IGM [e.g @Bolton:2005p6088; @Bolton:2007p3273; @Calverley:2011]. At the redshifts where these studies have been performed (e.g. $2 < z < 6$), the [Ly$\alpha$]{} line profile is not obscured or attenuated by a neutral IGM. Therefore, one could in principle push the flux prior we used in this work much closer to the [Ly$\alpha$]{} line centre, to substantially reduce the errors on the intrinsic [Ly$\alpha$]{} emission line profile blueward of [Ly$\alpha$]{}.
At $z>6$, for an increasingly neutral IGM, the intrinsic [Ly$\alpha$]{} emission line can become increasingly attenuated by the Gunn-Peterson IGM damping wing. While existing methods have been developed to access information on the red-side of [Ly$\alpha$]{} [e.g. @Kramer:2009p920], for the $z=7.1$ QSO ULASJ1120+0641 [@Mortlock:2011p1049], evidence suggests that the IGM damping wing imprint extends further redward [e.g @Mortlock:2011p1049; @Bolton:2011p1063], limiting the effectiveness of these approaches. The approach developed in this work should be unaffected by this as here we do not fit the [Ly$\alpha$]{} line, our reconstruction method therefore is perfectly suited for exploring the potential imprint of the IGM damping wing on the $z=7.1$ QSO [e.g. @Greig:2016p1], along with other future $z>6$ QSOs.
Conclusion {#sec:Conclusions}
==========
Characterising the continuum and emission lines properties of QSOs provides a wealth of information on the internal properties of the AGN, such as the mass of the SMBH, the QSO metallicity, star formation rates of the host galaxy, nuclear outflows and winds etc. Furthermore, the intrinsic [Ly$\alpha$]{} line shape can be used to probe properties of the IGM, such as the mean photoionisation background and the abundance of neutral hydrogen in the IGM at $z>6$.
Motivated by correlations amongst QSO emission lines [e.g. @Boroson:1992p4641; @Sulentic:2000; @Shen:2011p4583], in this work, we developed a new reconstruction method to recover the intrinsic [Ly$\alpha$]{} profile. This method is based on the construction of a covariance matrix built from a large sample of moderate-$z$ ($2.0 < z < 2.5$), high S/N (S/N $> 15$) QSOs from the BOSS observational programme. We use this moderate-$z$ sample to characterise the intrinsic [Ly$\alpha$]{} line profile, where it should be relatively unaffected by intervening neutral hydrogen in the IGM. In order to characterise each QSO within our sample we developed an MCMC fitting algorithm to jointly fit the QSO continuum, the emission lines and any absorption features that could contaminate or bias the fitting of the QSO. We modelled the QSO continuum as a single two parameter power-law ($\propto \lambda^{\alpha_{\lambda}}$), and each emission line is modelled as a Gaussian defined by three parameters, its width, peak amplitude and velocity offset from systemic.
We constructed our covariance matrix from a refined sample of 1673 QSOs, using the high ionisation emission lines [Ly$\alpha$]{}, [Si[IV]{}+O[IV\]]{}]{}, [C[IV]{}]{} and [C[III\]]{}]{}. Owing to the flexibility of the MCMC framework, we explored various combinations of single and double component Gaussians to characterise these lines. For [Ly$\alpha$]{} and [C[IV]{}]{} we settled on a double component Gaussian, to model the presence of a broad and narrow component. For the remaining two lines, we considered a single component. This resulted in an $18\times18$ covariance matrix, which we used to investigate new and existing correlations amongst the line profiles. We identified several strong trends from our covariance matrix, most notably the strong positive correlation between the peak amplitudes of the [Ly$\alpha$]{} and [C[IV]{}]{} narrow components ($\rho=0.8$) and the strong anti-correlation between the [Ly$\alpha$]{} narrow component peak amplitude and the width of the [C[III\]]{}]{} line ($\rho=-0.74$). These two were the strongest examples of a consistent trend of a positive correlation in the peak amplitudes across all the emission line species and the anti-correlation between the peak amplitudes and line widths.
Using this covariance matrix, we constructed an $N$-dimensional Gaussian likelihood function from which we are able to recover our reconstructed [Ly$\alpha$]{} line profile. The reconstruction method works as follows:
- Fit a QSO with our MCMC pipeline within the range $1275{\rm \AA} < \lambda < 2300$Å, recovering the parameters defining the QSO continuum, and the [Si[IV]{}+O[IV\]]{}]{}, [C[IV]{}]{} and [C[III\]]{}]{} lines.
- Obtain a six dimensional estimate of the reconstructed [Ly$\alpha$]{} line profile (modelled as a two component Gaussian) which provides the best-fit profile and correlated uncertainties, by evaluating the $N$-dimensional likelihood function describing our full covariance matrix including a prior on the reconstructed [Ly$\alpha$]{} line using the observed QSO flux within the range $1230 < \lambda < 1275$Å.
To visually demonstrate the performance of this reconstruction method, we applied it to a randomly selected QSO from our full data set. Finally, we quantitatively assessed its performance by applying it to the full QSO sample, and compared the reconstructed [Ly$\alpha$]{} profile parameters to those recovered from the original full MCMC fit of the same QSO. We found that estimates for both the [Ly$\alpha$]{} peak amplitudes are recovered strongly, as is the velocity offset of the narrow line component. For both of the line widths and the broad component velocity offset we find moderate agreement. We additionally explored the total reconstructed [Ly$\alpha$]{} flux (rather than individual parameters) relative to the original full MCMC fit at two distinct wavelengths blueward (1205Å) and redward (1220Å) of [Ly$\alpha$]{}. Our reconstruction method recovered the [Ly$\alpha$]{} line flux to within 15 per cent of the measured flux at 1205Å (1220Å) $\sim85$ ($\sim90$) per cent of the time.
There are several potential applications for both the MCMC fitting method and the [Ly$\alpha$]{} reconstruction pipeline. The MCMC fitting could be easily modified to measure any emission or absorption feature within a QSO spectrum. With this, many properties of the source QSO could be extracted, for example QSO metallicities. The ability to reconstruct the intrinsic [Ly$\alpha$]{} line profile could have important cosmological consequences such as improving estimates of the IGM photoionisation background or recovering estimates of the IGM neutral fraction.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the anonymous referee for their helpful suggestions. AM and BG acknowledge funding support from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 638809 – AIDA – PI: AM). ZH is supported by NASA grant NNX15AB19G.
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/.
SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University.
Selecting the systemic redshift for our BOSS sample {#sec:systemic_redshift}
===================================================
The reconstruction method discussed within this work is strongly dependant on having an accurate systemic redshift for the QSO sample. In the absence of knowing the true source redshift, systematic errors in its estimation will filter into the recovered line profile parameters, most notably the velocity offset of the line centre. As a result, this will affect any reported line profile correlations which are necessary for the [Ly$\alpha$]{} line profile reconstruction. Within this appendix, we discuss our attempts to determine the most robust and accurate redshift to be adopted within this work.
The BOSS DR12Q database [@Paris:2016p1] provides six available redshift estimates; a visually inspected redshift, $z_{\rm VI}$, the BOSS pipeline redshift based on a decomposed eigenvalue training set, $z_{\rm pipeline}$ (see e.g. @Bolton:2012SDSS), an automated PCA based approach, $z_{\rm PCA}$, and three other estimates determined from the location of the peak emission for the individual line profiles, $z_{{\ifmmode\mathrm{C\,{\scriptscriptstyle IV}}\else{}C\,{\scriptsize IV}\fi}{}}$, $z_{{\ifmmode\mathrm{C\,{\scriptscriptstyle III]}}\else{}C\,{\scriptsize III]}\fi}{}}$ and $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$. Using the earlier DR9Q database, @Font-Ribera:2013 performed a detailed analysis on each of these redshift estimates. These authors found $z_{\rm PCA}$ to be the optimal redshift estimator, returning both the smallest dispersion and bias amongst these six available measurements.
However, in adopting $z_{\rm PCA}$ as the optimal redshift estimator we find a bias within our selected sample of QSOs. In Figure \[fig:z\_PCA\], we show the velocity offset of the narrow component of the [C[IV]{}]{} line (measured with $z_{\rm PCA}$) as a function of the estimated redshift from the PCA approach using our ‘good’ sample of 1673 QSOs. In the right most panel, we show the one-dimensional histogram for the narrow component of [C[IV]{}]{}. Immediately obvious from this figure, is the non-Gaussian (bi-modal) nature of the distribution. Contrast this with the equivalent figure of the narrow component of [C[IV]{}]{} for the $z_{\rm pipeline}$ redshift estimator shown in Figure \[fig:z\_pipe\], which shows no obvious departure from Gaussianity. We have additionally tested this for the other redshift estimators (e.g. $z_{\rm VI}$ and $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$) and found no evidence for the same bi-modal distribution as shown by $z_{\rm PCA}$.
In the central panel of Figure \[fig:z\_PCA\], we show the two-dimensional histogram, with red signifying the highest density of QSOs. Presenting the data in this manner immediately highlights the source of error. The two peaks in the [C[IV]{}]{} velocity offset distribution are driven by two separate, discrete regions within the [C[IV]{}]{}–$z_{\rm PCA}$ parameter space: the primary peak in the [C[IV]{}]{} velocity offset at $\sim-250$ km/s (for all redshifts except the ‘hole’ at $z\sim2.3$) and the secondary peak at $\sim-1250$ km/s (at $z\sim2.3$). This discrete ‘overdensity’ of QSOs at $z\sim2.3$ is most likely an unphysical artefact. The likely cause of this discrete transition near $z\sim2.3$ is the redshifting of the [Mg[II]{}]{} line between strong OH sky line features. Typically, the [Mg[II]{}]{} emission line is one of the strongest anchors in which the source redshift can be determined. At $2.25 \lesssim z \lesssim 2.33$, [Mg[II]{}]{} is observed within a region that has a relatively low sky background with few sky lines, resulting in a higher [Mg[II]{}]{} S/N. However, at $z\lesssim2.25$ and $z\gtrsim2.33$, the [Mg[II]{}]{} feature appears amongst OH forest regions with very strong sky lines, resulting both in a noisier signal and it being strongly affected by systematics owing to the sky line residuals. Therefore, the weakening of the [Mg[II]{}]{} information in determining the QSO redshift within the PCA approach at $z\gtrsim2.33$ is the likely cause of this transition. Due to this bi-modal nature, which appears to be intrinsic only to the $z_{\rm PCA}$ redshift estimator, and since throughout this work we have assumed the distributions of line profiles can all be represented by a Gaussian distribution, we refrain from adopting the PCA redshift within this analysis.
The next three best candidates after $z_{\rm PCA}$ from the @Font-Ribera:2013 analysis are $z_{\rm VI}$, $z_{\rm pipeline}$ and $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$[^18]. Of these three, $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ has the smallest bias, followed by $z_{\rm pipeline}$ and $z_{\rm VI}$. In contrast, $z_{\rm VI}$ has the smallest dispersion, followed by $z_{\rm pipeline}$ and $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$. In this work, we do not consider using the visually inspected redshift as: (i) it has the strongest systematic bias of these choices (ii) it is the least objective measure (not entirely automated), requiring intervention to improve the redshift.
The QSOs within our ‘good’ sample are selected within $2.08 \leq z \leq 2.5$. By adopting, $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ instead of $z_{\rm pipeline}$, we lose all QSOs above $z>2.4$, where no $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ redshift is reported for any of the sources in DR12Q. Ultimately, the loss of sources reduces the total number of QSOs within our sample by $\sim25$ per cent. Despite the reduction in sample size, the $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ and $z_{\rm pipeline}$ samples return almost identical covariance matrices, signifying consistency in both redshift estimates. The major difference between the two samples is the notably stronger correlation in the velocity offsets for the emission line parameters within the $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$. However, this is to be expected owing to the differences in how the redshifts are estimated. For $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$, a single line profile is used to estimate the redshift, whereas in $z_{\rm pipeline}$ several line profiles are compared to a training set to estimate the best candidate redshift. By using several line profiles to determine the source redshift, intrinsic scatter within the source spectrum can cause the redshift estimates to deviate compared to those from a single line measurement such as $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$. Despite $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ providing the stronger line profile components for the velocity offsets than the $z_{\rm pipeline}$, we refrain from using the [Mg[II]{}]{} redshifts as these correlations may be artificially high owing to their estimation from only a single line profile. We deem the correlations from $z_{\rm pipeline}$ sample to be more conservative, while also benefitting from the addition of maintaining a larger statistical sample.
Finally, $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ and $z_{\rm pipeline}$ return relatively similar dispersions, with $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ being less biased. With the vast majority of the QSOs in our sample having both redshifts, in principle we can attempt to calibrate our $z_{\rm pipeline}$ sample by the $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ sample. In Figure \[fig:z\_offset\] we provide the $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$–$z_{\rm pipeline}$ parameter space, with the red dashed curve denoting the one-to-one relation, from which one can see a small offset. We recalibrate our sample for this offset by creating a pipeline corrected redshift, $$\begin{aligned}
z_{\rm pipeline\_corr} = z_{\rm pipeline} + \langle z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}} - z_{\rm pipeline} \rangle,\end{aligned}$$ for which we recover a small offset of $\sim2.77\times10^{-3}$ ($\sim361$ km/s at $z\sim2.3$).
Employing this BOSS pipeline corrected redshift, we perform the full analysis as performed within Sections \[sec:Covariance\] and \[sec:Recon\]. With respect to the covariance matrices, we recover effectively identical correlations amongst our line profile parameters. In order to better gauge the performance of these two possible samples, we recover the fraction of QSOs with a reconstructed flux recovered to within 15 per cent at both 1205Å and 1220Å (see e.g. Figure \[fig:FluxDistribution\]). Though the differences are minor, we find that the original BOSS pipeline redshift ($z_{\rm pipeline}$) performs marginally better than $z_{\rm pipeline\_corr}$. At 1205Å, we recover the QSO flux to within 85 (82) per cent for $z_{\rm pipeline}$ ($z_{\rm pipeline\_corr}$) and 90 (88) per cent at 1220Å[^19]. Given that the primary aim of this work is to reconstruct [Ly$\alpha$]{} emission profiles, then the most natural choice for a redshift estimate is the one which performs best at this reconstruction; the line parameter correlations are a secondary product. Therefore, we conclude that owing to the marginally improved statistics on the recovery of the QSO flux from the $z_{\rm pipeline}$ sample relative to $z_{\rm pipeline\_corr}$, and the reasons mentioned previously for not considering other redshift estimates, the best redshift to adopt within this work is $z_{\rm pipeline}$.
Preference for the number of Gaussian components for a line profile {#sec:line_component}
===================================================================
In the construction of our covariance matrix in Section \[sec:covariance\], we model the [Ly$\alpha$]{} and [C[IV]{}]{} emission lines as a double component Gaussian whereas for the [Si[IV]{}+O[IV\]]{}]{} and [C[III\]]{}]{} lines we consider only a single component Gaussian. In this appendix we detail quantitatively the reasoning for our choices. To objectively differentiate between the various combinations we use the Bayes Information Criterion (BIC; @Schwarz:1978p1 [@Liddle:2004p5730]). The BIC can be computed using, $$\begin{aligned}
\label{eq:BIC}
{\rm BIC} = -2\,{\rm ln}(\mathcal{L}) + k\,{\rm ln}(N),\end{aligned}$$ where $\mathcal{L}$ is recovered from the $\chi^{2}$ fit to the individual emission line profile, $k$ is the number of free parameters and $N$ is the number of available data points. A lower recovered BIC for an individual fit to the emission line profile is considered to be a better characterisation of the line.
The comparison between two possible sets of parameters to describe the emission line profile can be quantitatively described by the difference in the two BICs, $\Delta$BIC. A recovered $\Delta$BIC $> 10$, provides very strong evidence for the model with the lower BIC, whereas for $6 < \Delta$BIC $< 10$ strong evidence exists. Throughout this appendix we discuss each emission line profile individually, and recover the $\Delta$BIC for each model to distinguish between our preferred choices.
-------------------------------------------------------------- -------------------------------------------------------------- ------------------------------- ---------------------------------- -- -- -- --
Line fitting type Line fitting type QSOs with $\Delta$BIC $ > 10$ QSOs with $6 < \Delta$BIC $< 10$
(preferred option) (compared against) (per cent) (per cent)
[Ly$\alpha$]{} double component, [N[V]{}]{} and [Si[II]{}]{} [Ly$\alpha$]{} double component and [N[V]{}]{} 60.9 1.1
[Ly$\alpha$]{} double component, [N[V]{}]{} and [Si[II]{}]{} [Ly$\alpha$]{} single component, [N[V]{}]{} and [Si[II]{}]{} 79.4 0.7
[Ly$\alpha$]{} double component, [N[V]{}]{} and [Si[II]{}]{} [Ly$\alpha$]{} single component and [N[V]{}]{} 93.0 0.4
[Si[IV]{}+O[IV\]]{}]{} single component [Si[IV]{}+O[IV\]]{}]{} double component 64.7 6.9
[C[IV]{}]{} double component [C[IV]{}]{} single component 94.0 0.3
[C[III\]]{}]{} single component and [Al[III]{}]{} Only [C[III\]]{}]{} double component 53.8 2.9
[C[III\]]{}]{} single component and [Al[III]{}]{} [C[III\]]{}]{} double component and [Al[III]{}]{} 45.7 6.3
[C[III\]]{}]{} single component and [Al[III]{}]{} Only [C[III\]]{}]{} single component 92.2 0.9
-------------------------------------------------------------- -------------------------------------------------------------- ------------------------------- ---------------------------------- -- -- -- --
\[tab:BIC\]
The [Ly$\alpha$]{} profile
--------------------------
For the [Ly$\alpha$]{} emission line profile, we consider the full emission line complex within the rest-frame wavelength range $\lambda = [1170,1280]$ Å. This provides three different potential line species to be fit, [Ly$\alpha$]{}, [N[V]{}]{} and the [Si[II]{}]{} line. Given the importance of the [N[V]{}]{} line for the full line complex, we always fit the [N[V]{}]{} line in addition to the [Ly$\alpha$]{} line. Therefore, we consider four different combinations to potentially characterise the full [Ly$\alpha$]{} line complex. Firstly, we consider the [Ly$\alpha$]{} line as a double component Gaussian, with a single component Gaussian each for the [N[V]{}]{} and [Si[II]{}]{} emission lines. Next, the same, but only a single component [Ly$\alpha$]{} component. Finally, we consider these same two combinations excluding the [Si[II]{}]{} line.
In Table \[tab:BIC\], we report the $\Delta$BIC for these four combinations of emission lines. We compute the BIC as outlined in Equation \[eq:BIC\], recovering the $\chi^{2}$ within the wavelength range $\lambda = [1170,1280]$ Å. In the construction of Table \[tab:BIC\], we use the ‘conservative’ sample consisting of 2653 QSOs. We then recover the $\Delta$BIC from the difference between a model with a double component Gaussian for [Ly$\alpha$]{} and the single components for [N[V]{}]{} and [Si[II]{}]{}. It is clear from this table that this is the preferred choice for modelling the [Ly$\alpha$]{} line complex. Importantly, we find strong evidence for a double component [Ly$\alpha$]{} emission line in over 80 per cent of the QSOs we consider. In principle, we could have explored whether further components for the [Ly$\alpha$]{} line are preferred, however, for the sake of simplicity we have refrained from doing so. Furthermore, while the [Si[II]{}]{} emission line is typically only a weak ionisation line, and in some cases not resolved within individual QSOs, we still find strong evidence for its inclusion in the QSO fitting (preferred in $\sim62$ per cent of our sample).
The [Si[IV]{}+O[IV\]]{}]{} profile
----------------------------------
Unlike the [Ly$\alpha$]{} line considered above, the blended [Si[IV]{}+O[IV\]]{}]{} doublet is a strong, isolated emission line. We therefore consider the simplified choice of whether a single or double component Gaussian is preferred for this line within the wavelength range $\lambda = [1360,1440]$ Å. From Table \[tab:BIC\], we find that a single component Gaussian is very strongly preferred at $\sim65$ per cent, which becomes $\sim72$ per cent when the criteria is lowered to only strong evidence.
The [C[IV]{}]{} profile
-----------------------
For the [C[IV]{}]{} doublet, we consider the emission line complex within the wavelength range, $\lambda = [1500,1600]$ Å. Again, the [C[IV]{}]{} emission line is an isolated, very strong high ionisation line, therefore we only need to consider whether a single or double component Gaussian is preferred. In Table \[tab:BIC\], we recover very strong evidence for the choice of a double component Gaussian, with this being preferred 94 per cent of the time.
The [C[III\]]{}]{} profile
--------------------------
The [C[III\]]{}]{} emission line is not an isolated line, instead it is the strongest line in a complicated emission line profile [e.g. @VandenBerk:2001p3887]. Surrounding the [C[III\]]{}]{} line are contributions from the Fe pseudo-continuum, as well as contributions from the [Al[III]{}]{} and Si[III\]]{} line. In this work we consider the contribution of the Fe pseudo-continuum to be negligible to the [C[III\]]{}]{} line, and for all intents and purposes we additionally consider the Si[III\]]{} contribution negligible. This latter point arises from the insufficient S/N of these BOSS QSOs to fully resolve and separate out this line contribution. While it may be present in the highest S/N QSOs in our sample, for the vast majority it cannot be easily distinguished. Furthermore, from @VandenBerk:2001p3887, the relative line fluxes differ by over a factor of a hundred.
We consider four scenarios to characterise the [C[III\]]{}]{} line complex. A double component [C[III\]]{}]{} line in combination with a single component Gaussian for [Al[III]{}]{}. A single component [C[III\]]{}]{} and [Al[III]{}]{} line, and then just a single or double component Gaussian for the [C[III\]]{}]{} line, without considering the [Al[III]{}]{} line. In Table \[tab:BIC\], we find just a single [C[III\]]{}]{} emission line to be very strongly disfavoured with respect to a single component [C[III\]]{}]{} and [Al[III]{}]{} ($\sim92$ per cent). However, in the case of the remaining scenarios it is not as clear.
Comparing the single [C[III\]]{}]{} and [Al[III]{}]{} components to a purely double component [C[III\]]{}]{} (no [Al[III]{}]{}), we find marginally stronger evidence for the single [C[III\]]{}]{} scenario (preferred $\sim54$ per cent of the time). However, note that in allowing for a double component [C[III\]]{}]{}, these two scenarios are effectively the same. For example, the broad component can just fit the [Al[III]{}]{} line, leaving a single [C[III\]]{}]{} line. This degeneracy therefore tends to draw these two scenarios to be more equally preferred. As a result, the single [C[III\]]{}]{} and [Al[III]{}]{} scenario is considered to be preferable.
In the case of comparing a single or double component [C[III\]]{}]{} line in combination with [Al[III]{}]{}, we find comparable results for the $\Delta$BIC. Between these two scenarios we find a single component very strongly preferred in $\sim46$ per cent of the QSOs sampled, which increases to $\sim52$ per cent if we lower the criteria to just strong evidence. It is therefore, not immediately clear from the $\Delta$BIC which is the preferred fitting scenario to consider for the [C[III\]]{}]{} line complex. As a result, we choose to consider a single [C[III\]]{}]{} component with a [Al[III]{}]{} contribution purely on the grounds of there being less model parameters, removing potential degeneracies between a broad and narrow component [C[III\]]{}]{} from our covariance matrix.
QSO quality assessment examples {#sec:QSO_QA}
===============================
In Section \[sec:QA\], we briefly outlined the visual quality assessment we performed to obtain the two QSO samples for the covariance matrix. Within this appendix we provide a more detailed explanation of the respective criteria we consider, as well as providing a select few examples which best indicate the various decisions which we have made.
Removal of QSOs
---------------
The first step of our quality assessment was the removal of any QSOs from which we recovered a poor characterisation of the QSO continuum. In Figure \[fig:QA\_RemovedContinuum\] we provide two examples of QSOs which were removed failing this criteria. In the top panel, the recovered QSO returns a positive spectral index, $\alpha_{\lambda}$, which is counter to the negative spectral indices recovered by the vast majority of our QSO sample. The recovery of a positive spectral index is unphysical, and likely arises from flux calibration errors. Only a few QSOs within our sample are found to have a positive spectral index. In the bottom panel of Figure \[fig:QA\_RemovedContinuum\], we provide an example of a QSO where a single power-law is insufficient to characterise the QSO continuum within our fitting region. To adequately characterise this QSO, a double component continuum pivoting on [C[IV]{}]{} would be required. The rarity in recovering such an object likely points to this being an error in the data reduction pipeline, or the flux calibration.
The next step in our visual QSO quality assessment is the removal of any QSO spectra which have either (i) missing sections of the observed QSO flux which overlap with our chosen emission lines (ii) a large number of absorption features (iii) either an intervening dense neutral absorber or sufficiently strong/broad absorption blueward of [C[IV]{}]{} (iv) a [Ly$\alpha$]{} line region that is not well fit or characterised by our MCMC pipeline. In Figure \[fig:QA\_RemovedLya\] we present nine examples of QSOs removed from our sample. We discuss each one of these QSOs below[^20]:
- A clear, dense neutral absorber in the direct vicinity of the QSO. Presence of the neutral absorber completely removes any ability to recover a [Ly$\alpha$]{} broad component.
- Two strong absorption features near [Ly$\alpha$]{}. The first, blue ward of [Ly$\alpha$]{} removes almost half of the [Ly$\alpha$]{} line. Even in the presence of this absorber, the intrinsic [Ly$\alpha$]{} line profile recovery is still reasonably good. However, it still under predicts the true profile. Furthermore, absorption on the [N[V]{}]{} line centre removes any ability to recover the [N[V]{}]{} line, resulting in a much broader [Ly$\alpha$]{} component than should be recovered.
- Strong [Ly$\alpha$]{} absorption at the line centre. Without a sufficient fraction of the [Ly$\alpha$]{} line centre, the intrinsic [Ly$\alpha$]{} line cannot be recovered. While a broad [Ly$\alpha$]{} component can still be estimated, no narrow component can be reliably obtained.
- Another, even more extreme example of strong absorption on [Ly$\alpha$]{} line centre. Presence of several strong [Ly$\alpha$]{} absorption features completely removes any information to obtain the [Ly$\alpha$]{} narrow line component.
- A [Ly$\alpha$]{} line profile with no prominent narrow line component. As a result, the two Gaussian components both prefer to be broad, and are therefore degenerate. The absorption on line centre likely contaminated the [Ly$\alpha$]{} peak determination leading to the two preferred degenerate broad components.
- A large, connected series of narrow absorption features centred on [Ly$\alpha$]{}. Since the entire [Ly$\alpha$]{} profile is contaminated by absorption, measuring the [Ly$\alpha$]{} narrow component peak amplitude becomes degenerate with the depth of the absorption features.
- The MCMC pipeline struggles to recover the extremely narrow, sharp peak around [Ly$\alpha$]{} line centre. Evidence of two small absorption features (which are not identified by our algorithm) on either side of line centre appear to cause this peak to become too narrow to fit.
- Significant blue ward absorption features near the [C[IV]{}]{} line contaminates the recovery of the broad component. Not a BAL QSO, however the absorption features are still sufficient to contaminate the recovery of the full line profile.
- Deep, narrow absorption features on the [C[IV]{}]{} line profile. The unfortunate location of these features at the wing of the [C[IV]{}]{} narrow component causes a narrower [C[IV]{}]{} line than should be expected. The depth of these features causes this behaviour, weaker absorption would not impact the narrow component recovery.
Examples of QSOs in the ‘conservative’ sample
---------------------------------------------
In Figure \[fig:QA\_Conservative\] we provide three examples of QSOs which are classified as ‘conservative’ only (not considered for the ‘good’ sample). In the first example (left panel), several strong narrow absorption features are detected and fit. These are offset from the line centre of [Ly$\alpha$]{}, allowing the peak height of [Ly$\alpha$]{} to still be reliably estimated. Though the intrinsic profile appears to be well recovered, the uncertainties in the peak amplitude owing to potential degeneracies with the relative strengths of the absorption features and the number of identified absorption features causes this to be only characterised as ‘conservative’.
The second example (middle panel) has a strong absorption feature on [Ly$\alpha$]{} line centre. However, the relative strength of this feature is not sufficient to completely contaminate the peak of the [Ly$\alpha$]{} line (unlike similar examples in Figure \[fig:QA\_RemovedLya\]). It appears the full [Ly$\alpha$]{} line profile is recovered well, despite numerous narrow absorption features and the strong central absorption. The uncertainties surrounding the narrow [Ly$\alpha$]{} component determination from the strong absorption results in its assessment as ‘conservative’ only. Finally, in the last example (right panel), the full [Ly$\alpha$]{} line profile is contaminated by a large number of narrow absorption features. None of these are located at line centre, nor do they appear to impact the recovery of the [Ly$\alpha$]{} line profile. However, the sheer number of features causes this QSO to be classified as conservative.
Examples of QSOs in the ‘good’ sample
-------------------------------------
Finally, in Figure \[fig:QA\_Good\] we present three examples of QSOs explicitly selected for our refined ‘good’ sample (all QSOs in the ‘good’ sample appear in the ‘conservative’ sample). In the first example (left panel), there are only a few very small absorption features. In the absence of anything stronger, these QSOs are classified as ‘good’. The second example (middle panel), once again has very few absorption features. Despite being strong, narrow absorption lines, these are well characterised and do not impact the recovery of the intrinsic [Ly$\alpha$]{} line at all. In the final example (right panel), a larger number of absorption lines are present. Importantly though, these are located offset from the [Ly$\alpha$]{} line centre, and therefore we are confident that the true profile is recovered. Note, this profile is similar to the first example in Figure \[fig:QA\_Conservative\]. The reason this QSO is classified as ‘good’ and the previous was ‘conservative’ arises from the number of absorption lines. In this example, only two narrow features impact the [Ly$\alpha$]{} line, whereas in the previous case, several lines have resulted in a smaller fraction of the total [Ly$\alpha$]{} profile being unaffected by the absorption features. Ultimately, this is only a minor difference that does not change any of the conclusions of this work, since in Section \[sec:Bias\] we found similar correlation matrices between the two samples.
Recovery of continuum parameters from $\lambda>1275{\rm \AA}$ {#sec:continuum_comparison}
=============================================================
In Section \[sec:ReconExample\], we alluded to an assumption regarding the usage of the QSO continuum between the full MCMC fitting and the reconstruction profile fitting. Specifically, we assume that the continuum measured from a [Ly$\alpha$]{} obscured or high-$z$ QSO in the region ($\lambda > 1275$Å) will be equivalent to the continuum recovered from fitting the full QSO within the range $1180{\rm \AA} < \lambda < 2300$Å. However, by not fitting in the $1180{\rm \AA} < \lambda < 1275$Åregion, we could be losing information from the [Ly$\alpha$]{} line profile (and surrounding region) that could be otherwise used to refine the estimate of the QSO continuum. Of course, this cannot be tested on an actual QSO that requires reconstruction, but instead we can statistically test this assumption across the full QSO sample we have used within this work.
In Figure \[fig:ContinuumComparison\], we present the 2D scatter of the two QSO continuum parameters, the continuum flux amplitude normalised to $1450$ Å ($f_{1450}$, left panel) and the spectral index, $\alpha_{\lambda}$ (right panel). To facilitate the comparison, the red dashed curve corresponds to the one-to-one relation. Note that while the flux amplitude is normalised at $1450$Å, which should be unaffected by the different fitting regions, we allow a perturbation on this flux to refine its true normalisation. In the case of this continuum normalisation, we find very little scatter, and all QSOs lie along the one-to-one relation indicative of their being no adverse effects from this assumption.
In the case of the QSO spectral index, we again find all QSOs to essentially lie perfectly along the one-to-one relation. However, in several instances, there are QSOs which deviate away from this line. This small scatter indicates that for a select few QSOs, fitting in the $1180{\rm \AA} < \lambda < 1275$Åregion is important for the recovery of the true continuum spectral index. Importantly, around ${\ifmmode\mathrm{Ly}\alpha\else{}Ly$\alpha$\fi}{}$ even for the extreme QSOs in this 2D plot, the level of discrepancy in the continuum flux will be at most 10 per cent. Furthermore, this is purely for the continuum flux, and not the total line flux. Referring to Figure \[fig:Profile\_Recon\], a 10 per cent error on the continuum flux at [Ly$\alpha$]{} will fall well within the errors of the reconstructed [Ly$\alpha$]{} emission line profile. Therefore, within the model uncertainties of the [Ly$\alpha$]{} reconstruction pipeline, this scatter is not significant to drastically affect the reconstructed [Ly$\alpha$]{} line profile and hence the assumption on using the same QSO continuum is justified.
[^1]: E-mail: [email protected] (BG),
[^2]: http://www.sdss.org/dr12/algorithms/boss-dr12-quasar-catalog/
[^3]: This choice of redshift range was a trade off between the wavelength coverage of the spectrograph and requiring that we were sufficiently blueward of rest-frame [Ly$\alpha$]{} to characterise the line profile (arbitrarily chosen to be 1180Å, corresponding to $\sim3600$Å at $z=2.08$) while also including the [Mg[II]{}]{} emission line ($\lambda=2798.75$Å, corresponding to $\lambda = 10,000$Å at $z=2.5$).
[^4]: Note that the r.m.s of the quasar redshift error distribution for the BOSS pipeline redshift from DR9 was determined to be $550$ km/s with a mean offset of the quasar redshift (bias) of $\sim-150$ km/s [@Font-Ribera:2013].
[^5]: We computed the absolute AB magnitude from the QSO continuum at 1450Å($M_{1450}$) for this sample of QSOs and recovered a median of $M_{1450} = -26.1$ with an interquartile range of 0.7.
[^6]: The errors and biases when attempting to accurately measure the covariance matrix of a set of parameters ($N_{\rm par}$) from a total number of sources ($N_{\rm src}$) scales approximately as $N_{\rm par}$/$N_{\rm src}$ [e.g @Dodelson:2013; @Taylor:2014; @Petri:2016]. Within this work, we ultimately construct an 18 parameter covariance matrix from a subsample of 1673 QSOs, corresponding to an expected error of $\sim1$ per cent.
[^7]: In practice, thermal (or other) broadening will make the Lorentzian wings irrelevant in the case of AGN emission lines. For even modest thermal broadening of $\sim10$ km/s only the Gaussian core can be detected [see e.g. figure 4 of @Dijkstra:2014p1].
[^8]: In principle, we could allow for $N>2$ components to our line fitting procedure. For example, several authors have fitted three Gaussian components to the [C[IV]{}]{} emission line profile [e.g. @Dietrich:2002p5912; @Shen:2008p5975; @Shen:2011p4583]. However, in this work, we attempt to avoid adding too much complexity to the fitting method.
[^9]: In principle, by ignoring the Si feature, the centroid of the measured [C[III\]]{}]{} line can be shifted bluer than its true value. Quantitatively, we determine the extent of this blueshift by fitting all QSOs within our sample that have a clear, discernible Si feature. We find that by excluding Si the average recovered [C[III\]]{}]{} emission line centre is blue shifted by $\sim310$ km/s (with the associated line width broadened by $\sim315$ km/s). This sample, however, only contains 57 (out of 1673) QSOs, corresponding to only $\sim3.4$ per cent of the full ‘good’ sample affected by this blueshift. For the vast majority of QSOs in our sample the S/N around [C[III\]]{}]{} is insufficient to tease out the individual emission component of the Si feature. Therefore, owing to the small resultant blueshift in the line centre estimate (compared to the average dispersion of $\sim550$ km/s in the redshift estimate of the chosen BOSS pipeline redshift) and the corresponding low number of sources potentially affected by this systematic shift, we deem the exclusion of Si a valid simplification. As a final point, owing to the weak total flux expected in the Si emission line, including this in the MCMC fitting for the full QSO sample (where Si is not clearly present) would cause a strong degeneracy with the [C[III\]]{}]{} line unless extremely strong priors on the line shape are imposed.
[^10]: Our naming convention of ’good’ and ’conservative’ refers to the tightness of constraints recovered from the reconstruction procedure. The ’conservative’ sample, with less assumptions on data quality, produces slightly broader errors relative to the ’good’ sample.
[^11]: By converting our fitted emission line parameters (both broad and narrow) into a line profile FWHM width, we find correlations similar in strength to those recovered purely from the narrow line component.
[^12]: Note that, if we adopt a systemic redshift measured purely from a single ionisation line (such as the [Mg[II]{}]{} redshift) rather than the pipeline redshift used throughout this work, we can recover a significantly stronger correlation of [Ly$\alpha$]{}–[C[IV]{}]{} ($\rho=0.7$).
[^13]: The choice in adopting a Gaussian covariance matrix is driven by the large computational burden required to perform a full end-to-end Bayesian approach folding in all modelling uncertainties.
[^14]: This is more prevalent when averaging over the full distribution of reconstructed [Ly$\alpha$]{} profiles.
[^15]: Note that this choice in phrasing does not imply a ‘prior’ in the Bayesian statistical sense, rather it is used to highlight that we are including additional information into our reconstruction procedure compared to the reconstruction method discussed in Section \[sec:Reconstruction\].
[^16]: This choice of 1230Å is purely arbitrary, and is chosen based on the assumption that a damped absorption signal would not extend this far redward of [Ly$\alpha$]{}. However, this choice is flexible, and can be adjusted on a case-by-case basis if evidence for stronger attenuation beyond 1230Å is present.
[^17]: Note that throughout this work, the recovered MAP estimates from the full posterior distribution do not differ significantly from the peak of the associated marginalised PDFs.
[^18]: We do not consider either of $z_{{\ifmmode\mathrm{C\,{\scriptscriptstyle IV}}\else{}C\,{\scriptsize IV}\fi}{}}$ or $z_{{\ifmmode\mathrm{C\,{\scriptscriptstyle III]}}\else{}C\,{\scriptsize III]}\fi}{}}$ as these have the largest dispersion and bias.
[^19]: For completeness, for $z_{{\ifmmode\mathrm{Mg\,{\scriptscriptstyle II}}\else{}Mg\,{\scriptsize II}\fi}{}}$ we recover the QSO flux to within 15 per cent for 80 per cent of the QSOs at 1205Å and 85 per cent at 1220Å.
[^20]: Note that in Figures \[fig:QA\_RemovedLya\] – \[fig:QA\_Good\] we represent the corresponding fitted absorption profiles as being subtracted from the continuum flux level rather than the true flux level. As a result, in some cases it appears that the absorption extends to negative flux. However, this choice is purely to aid the visual representation of the full emission profile fits.
| {
"pile_set_name": "ArXiv"
} |
[**DISCO: AN OBJECT-ORIENTED SYSTEM FOR**]{}\
[**MUSIC COMPOSITION AND SOUND DESIGN**]{}
Hans G. Kaper,$^1$ Sever Tipei,$^2$ and Jeff M. Wright$^3$
$^1$ MCS Division, Argonne National Laboratory, Argonne, IL 60439, USA (`[email protected]`)\
$^2$ School of Music, University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA (`[email protected]`)\
$^3$ University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA (`[email protected]`)
> [**Abstract.**]{} This paper describes an object-oriented approach to music composition and sound design. The approach unifies the processes of music making and instrument building by using similar logic, objects, and procedures. The composition modules use an abstract representation of musical data, which can be easily mapped onto different synthesis languages or a traditionally notated score. An abstract base class is used to derive classes on different time scales. Objects can be related to act across time scales, as well as across an entire piece, and relationships between similar objects can replicate traditional music operations or introduce new ones. The DISCO (Digital Instrument for Sonification and Composition) system is an open-ended work in progress.
[2]{}
### 1. INTRODUCTION {#introduction .unnumbered}
The compositional process is based on the assumption that aural events can be ordered in time: a musical composition represents a trajectory in sound space. The composer controls the structure, if not the details, of the trajectory and thus the nature of the composition. The control takes the form of an algorithm—a set of rules governing the nature of the objects, their evolution, and their interrelations—which defines the musical composition. Composing thus means defining objects and relating those attributes that yield a desired trajectory in sound space.
The object-oriented paradigm and the software implementation we describe here reflect this point of view. They also provide a way of merging two activities which, traditionally, have been considered separate: writing music and building instruments. With the exception of Harry Partch \[Partch, 1960\], who built actual instruments responsive to his music’s structure (based on ratios), and Xenakis \[Xenakis, 1993\], who used stochastic distributions to generate the structure of computer-generated sounds as well as large scale textures, few composers have shown an interest in combining these two areas. The system presented here treats both activities in a uniform way by using similar logic, objects, and procedures. The software modules for music composition and sound design are consistently and comprehensively interconnected. The resulting code, currently referred to as DISCO (Digital Instrument for Sonification and Composition), is a work in progress. The system was used recently by one of the authors for the composition of a piece for violin and computer-generated tape \[Tipei, 2000\].
### 2. OBJECTS AND PROPERTIES {#objects-and-properties .unnumbered}
The composition modules use an abstract representation of musical data, which can easily be mapped onto different synthesis languages or, as the case may be, a traditionally notated score. This is achieved by defining “Instrument” and “Property” classes in response to the requirements of the target output.
The Instrument class is essentially a collection of properties that define all of an instrument’s control parameters. A very simple instrument might be defined by the properties “Start Time,” “Duration” and “Pitch.” Each property is stored in a table, which is indexed by a string identifier. The Instrument class includes the methods describing the manner in which the instrument’s output is to be generated. Note that the Instrument class does not necessarily correspond to any actual instrument, but serves rather as an abstraction for defining the properties of a given sound object.
The Property class enables us to easily classify the different properties of a sound object. A composition would likely contain a number of sound objects sharing certain properties, such as “Start Time.” In this case, the advantages of the polymorphic nature of the system become evident, as one can work with these properties without knowing the type of instrument. The Property class also incorporates methods to check for the correct type of input data. For example, many instruments share the property “Pitch,” which may be represented as a floating-point frequency value, as an integer that indexes a tuning table, or as a string spelling the name of a note.
### 3. TIME SCALES {#time-scales .unnumbered}
The perception of aural events and their organization in larger structures points to the existence of time scales associated with particular objects. We mention, in order of increasing magnitude, the time scales of audio frequencies and of frequency and amplitude modulations, which affect partials and sounds; the time scales associated with melodic phrases, chordal aggregates and more complex textures; the time scales of larger formal units, such as sections and movements; and the time scale associated with an entire piece \[Kaper, 1999a\].
An abstract base class, “Event,” is used to derive classes on different time scales. The Event class has a relatively simple structure, which is defined by three attributes: start time, duration and name. Subclasses are derived from it in response to particular needs.
An event may contain other events and thus become a “Compound Event.” An entire piece is the most inclusive compound event. At the other extreme are the “Atomic Events,” which do not include other events. Partials in a sound or the graphic symbol of a note in a printed score are examples of atomic events. “Sections,” “Phrases,” “Motives,” “Chords” and “Aggreggates” are compound events which contain events of shorter or equal duration and may be themselves part of larger structures—of other compound events.
Besides the three inherited attributes (start time, duration and name), the derived classes have the property that they can be related to other similar classes or to classes of finer or coarser granularity. The type of a class, as well as its potential relations to other classes, are reflected in the class’s name.
Relationships or associations can act across time scales. An example is the congregation of partials into sounds, of sounds into chords or melodic gestures, and of sections into a composition. Also, more sophisticated relationships can be established between objects at different time scales and/or different locations in the piece. For example, the presence of a sound with a particular spectral envelope may trigger the assignment of a specific chord in a remote section of the piece.
Relationships between similar objects can replicate traditional music operations, such as transposition, inversion, and retrograde of a group of sounds, augmentation/diminution of durations or pitch intervals, chord inversion or other rearrangements of sounds in a chord, etc.
### 4. HIGHER LEVEL OF ABSTRACTION {#higher-level-of-abstraction .unnumbered}
“Generator” classes provide the composer with the ability to generate events based on some specific algorithm. They are designed to serve across time scales and can be of a generalized or specific type. For example, a simple random generator can create “NumberProperty” objects, which can be assigned any property of an instrument or event that is derived from the NumberProperty class. A specific generator to create only events of a certain type can be obtained by combining several simple generators into an “Event Generator.” One such utility, already in place, is designed to select the number of partials within a sound, the number of sounds in a cluster, or the number of sections in a piece according to a selected probability distribution. Another utility, the “Envelope” class, also in place, reads an envelope and interpolates values as necessary, thus giving the user control over the shape of events on various time scales. Still other classes enable the user to assign values manually from a list of possibilities or by using a script.
We intend to design a number of common algorithmic composition techniques as Generator classes to implement customized algorithmic techniques of the composer’s design. These classes will be extendible and can be used by themselves, as well as in combination.
### 5. METHODS AND APPLICATIONS {#methods-and-applications .unnumbered}
The type of classes and the methods to relate them are determined by the type of music the user wishes to compose. Objects like “Melody,” “Chord” and “Rhythm,” and methods such as “Canon” and “Chorale” anticipate a traditional composition; “Markov,” “Stochastic” and “Heterophony” show a different bend. While the initial emphasis was on less-than-traditional modes of composing, the system has acquired a much wider scope and now supports traditional, as well as nontraditional thinking. In addition, it supports sound design for scientific sonification—the faithful rendition of complex data sets in sounds \[Kaper, 1999b\]. The DISCO system is a truly open-ended work in progress, which is continuously being enriched with new classes and methods.
Among the first utilities developed for the DISCO system was the “Matrix” class. It was designed to enable the choice of a start time and a duration for each section in a Manifold Composition according to certain probability distributions. A Manifold Composition is essentially a collection of variants of one and the same piece, differing in details but with a similar overall structure \[Tipei, 1989\]. The differences between the variants are the result of stochastic choices. We briefly explain how the Matrix class was used to construct the probability matrices for the choice of start times and durations.
Suppose there are $n+1$ time marks in the piece (including the start time and end time). The start time of each section is supposed to coincide with one of the time marks. The end time of the piece cannot be the start time of a section, so there are $n$ possible start times; we label them $t_1$ through $t_n$. Each time mark $t_j$ has a certain weight $q_j$ associated with it, which measures the likelihood of the time mark becoming the start time of a section. Suppose there are $m$ possible sections, labeled $s_1$ through $s_m$. Each section $s_i$ has a certain (relative) weight $w_i$ associated with it; furthermore, $s_i$ has a certain probability $p_{ij}$ to become active at the time mark $t_j$. Using the Matrix class, a probability matrix $P$ is constructed with $m$ rows and $n$ columns. The elements of $P$ are $$P_{ij}
=
\frac{\sum_{k=1}^{i}\sum_{l=1}^{j} w_k p_{kl} q_l}
{\sum_{k=1}^{m}\sum_{l=1}^{n} w_k p_{kl} q_l} , \quad
\begin{array}{l}
i = 1, \ldots\,,m , \\
j = 1, \ldots\,,n .
\end{array}$$ Then $P_{ij}$ is the probability that section $s_i$ will start at the time mark $t_j$. Once the start times have been chosen, the duration $d_i$ of each section $s_i$ is determined from a probability matrix $Q$, which is constructed in a similar manner.
The matrices $P$ and $Q$ are dynamically adjusted. Once a start time and a duration have been assigned to a particular section, adjustments are made to diminish the probability that any other section is selected during the same time interval or at nearby times.
The Matrix class enables the assignment of events in any order, not necessarily as they appear in the piece—a reflection of the way most human composers work. The class has the potential of correlating various rationales leading to a particular selection, and its methods can be used in connection with any parameter values and intervals of any event. Not only sections in a piece can be defined this way, but also sounds in a section, chords and motives in a section, etc. A logical step will be to combine the matrices for the selection of start times and durations into one three-dimensional matrix and, eventually, to include all parameters in a single multidimensional matrix. Any one choice will then be the result of a combination of all available criteria and will determine all aspects of an event. Finding the appropriate data representation for such a multidimensional matrix, however, is not trivial—especially in C++.
### 6. INTERFACES {#interfaces .unnumbered}
All the basic classes described here have been implemented in C++. However, even for experienced programmers, C++ is a difficult language, and although some composers are excellent programmers, we cannot assume that all composers are willing to spend the time and effort to become proficient in C++. For this reason, most C++ classes have an analagous interface in Python, an interpreted high-level object-oriented language that is considerably easier to learn than C++ \[Lutz, 1996; Beazley, 1999\]. The choice of language is left to the user.
The wrapper code that allows the C++ classes to be used as Python classes is generated by SWIG \[Beazley, 1996\], which automates the process of combining C and C++ code with higher-level languages such as Python, Perl and Tcl. Although Python is currently the only language supported by the system, it is relatively simple to generate wrappers for Perl and Tcl.
### 7. CONCLUSION {#conclusion .unnumbered}
In this paper we have described an object-oriented system for music composition and sound design. The object-oriented approach has the advantage that one can easily add different classes and/or methods taylored to a particular composition or aesthetic. The code is like an open-ended work in progress, which invites the creation of structures and relationships between sounds not yet employed.
### ACKNOWLEDGMENTS {#acknowledgments .unnumbered}
This work was supported by the Mathematical, Information, and Computational Sciences Division subprogram of the Office of Advanced Scientific Computing Research, U.S. Department of Energy, under Contract W-31-109-Eng-38.
### REFERENCES {#references .unnumbered}
Partch, H. 1960. *Genesis of a Music; An Account of a Creative Work, Its Roots and Its Fulfillments*, New York, Da Capo Press, Second Edition (1974).
Xenakis, I. 1992. *Formalized Music, Thought and Mathematics in Music*, Revised Edition, Pendragon Press, pp. 289–293.
Tipei, S. 2000. “AntiPhan” for Violin and Computer-Generated Tape (unpublished).
Kaper, H. G. and Tipei, S. 1999a. “Formalizing the Concept of Sound,” Proc. Int’l. Computer Music Conference, Beijing, China, pp. 387–390.
Kaper, H. G., Tipei, S., and Wiebel, E. 1999b. “Data Sonification and Sound Visualization,” *Computing in Science and Engineering*, Vol. 1, No. 4, pp. 48–58.
Tipei, S. 1989. “Manifold Compositions: A (Super)Computer-Assisted Composition Experiment in Progress,” Proc. Int’l. Computer Music Conference, Columbus, Ohio, pp. 324–327.
Lutz, M. 1996. *Programming Python*, O’Reilly & Associates.
Beazley, D. 1999. *Python Essential Reference*, New Riders.
Beazley, D. 1996. “SWIG: An Easy to Use Tool for Integrating Scripting Languages with C and C++,” Presented at the 4th Annual Tcl/Tk Workshop, Monterey, Cal. (http://www.swig.org/papers/Tcl96/tcl96.html)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the mixed twistor $\nbigd$-modules associated to meromorphic functions. In particular, we describe their push-forward and specialization under some situations. We apply the results to study the twistor property of a type of better behaved GKZ-hypergeometric systems, and to study their specializations. As a result, we obtain some isomorphisms of mixed TEP-structures in the local mirror symmetry.
Keywords: mixed twistor $\nbigd$-module, generalized Hodge structure, polarization, GKZ-hypergeometric system, local mirror symmetry.
MSC2010: 14F10 32C38 32S35
author:
- Takuro Mochizuki
title: 'Twistor property of GKZ-hypergeometric systems'
---
Introduction
============
$\nbigd$-modules associated to meromorphic functions {#section;15.11.22.10}
====================================================
Mixed twistor $\nbigd$-modules associated to meromorphic functions
==================================================================
Graded sesqui-linear dualities
==============================
Comparisons of polarizations {#section;14.12.12.1}
============================
Mixed twistor $\nbigd$-modules and GKZ-hypergeometric systems {#section;14.12.8.11}
=============================================================
Quantum $\nbigd$-modules
========================
Duality of meromorphic flat bundles {#section;14.12.26.10}
===================================
Pairings and their functoriality {#section;14.12.26.11}
================================
Some functoriality of $\nbigrtilde$-modules {#section;14.12.26.12}
===========================================
Comparison of some $\nbigrtilde$-modules {#section;14.12.27.1}
========================================
Better behaved GKZ-systems and de Rham complexes {#section;15.5.2.2}
================================================
[*Address\
Research Institute for Mathematical Sciences, Kyoto University, Kyoto 606-8502, Japan\
[email protected]* ]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Advanced gravitational wave interferometric detectors will operate at their design sensitivity with nearly $\sim$1MW of laser power stored in the arm cavities. Such large power may lead to the uncontrolled growth of acoustic modes in the test masses due to the transfer of optical energy to the mechanical modes of the arm cavity mirrors. These parametric instabilities have the potential of significantly compromising the detector performance and control. Here we present the design of “acoustic mode dampers" that use the piezoelectric effect to reduce the coupling of optical to mechanical energy. Experimental measurements carried on an Advanced LIGO-like test mass shown a 10-fold reduction in the amplitude of several mechanical modes, thus suggesting that this technique can greatly mitigate the impact of parametric instabilities in advanced detectors.'
author:
- 'S. Gras'
- 'P. Fritschel'
- 'L. Barsotti'
- 'M. Evans'
bibliography:
- 'AMDbib.bib'
date: today
title: Resonant Dampers for Parametric Instabilities in Gravitational Wave Detectors
---
Introduction
============
The network of advanced gravitational wave detectors currently under construction (two LIGO [@0264-9381-27-8-084006] detectors in the USA, the Advanced Virgo [@AdvancedVIRGO] detector in Italy, and Kagra [@0264-9381-29-12-124007] in Japan) promises to open the new window of gravitational wave astronomy within this decade.
Sensitivity to gravitational wave strains of order $10^{-24}$ requires high optical power circulating in the arm cavities of these detectors. For instance, up to 750 kW of optical power will be sustained in the steady-state regime inside the Advanced LIGO arm cavities.
It has been experimentally observed that the stored energy in a resonant cavity can leak from the optical modes to the mechanical modes of the cavity mirrors via a 3-mode interaction [@PhysRevA.78.023807]. Given sufficiently high circulating optical power, and mirror materials with very low mechanical loss as required to avoid thermal noise, the uncontrolled growth of test mass acoustic modes can lead to Parametric Instabilities (PI) [@Evans2010665; @0264-9381-27-20-205019]. If left unaddressed PI will prevent high power operation, and thus limit the astrophysical output of gravitational wave detectors. While Advanced LIGO will serve as the primary example in this paper, all advanced gravitational wave detectors are susceptible to these instabilities.
Several schemes have been proposed to damp PI [@PhysRevLett.94.121102; @PhysRevA.81.013822]. In particular, solutions directly applicable to Advanced LIGO involve active damping of acoustic modes by means of the test mass electro-static drive actuators [@Miller2011788], and thermal tuning of the optical modes using the test mass ring heaters [@CQG_0264-9381-26-13-135012]. A significant constraint on any technique is that it must preserve the inherently low mechanical loss of the test mass in the gravitational wave frequency band to maintain a low level of thermal noise.
Here we present a novel method to passively control PI by reducing the Q-factor of the test mass acoustic modes with small resonant dampers. These “acoustic mode dampers" (AMD) dissipate the strain energy of the test mass mode through a resistive element after converting it to electrical energy via the piezoelectric effect (see figure \[fig:spring-mass\]).
The resonant nature of the AMD allows it to effectively damp test mass acoustic modes without introducing significant mechanical loss at lower frequencies where thermal noise can limit detector performance. With respect to other proposed solutions, this approach has the advantage of being simple, self-contained, and completely passive. Models indicate that AMDs can provide a broadband reduction in the Q of mechanical modes relevant to PI, and are therefore particularly beneficial in the presence of a large number of unstable modes.
The structure of this paper is as follows. In section \[sec:PI\] we set the stage by giving a brief overview of parametric instabilities, including equations of particular relevance to evaluating AMD performance. Section \[sec:ds\] presents a simple 1-dimensional model of the AMD which highlights the principles of AMD operation. This is followed by a description of the detailed finite element model (FEM) used to analyze the AMDs ability to suppress PI when attached to an Advanced LIGO test mass. The FEM predictions are compared with experimental results from a full-scale prototype in section \[sec:er\]. Finally, in section \[sec:tn\] we discuss an AMD design that will provide Advanced LIGO with protection from instability, without significantly increasing test mass thermal noise.
Parametric Instabilities (PI) {#sec:PI}
=============================
The acousto-optic interactions responsible for parametric instabilities have been extensively studied [@Braginsky2001331; @PhysRevLett.94.121102; @Evans2010665]. They consist of a scattering process and radiation pressure operating together in an optical cavity in a closed-loop manner. The graphical representation of this process is shown in Fig.\[fig:ShearPZT\].
The e-folding growth time, or “ring-up time”, of an acoustic mode in the presence of a 3-mode interaction is $\tau =2Q_{m}/(\omega_{m}(R-1))$ [@Ju2006360], where $Q_m$ and $\omega_m$ are the Q-factor and angular frequency of the mode, respectively, and $R$ is the parametric gain. When $R > 1$ the amplitude increases exponentially until a saturation point is reached [@arXiv:1303.4561v2]. The parametric gain $R$ can be defined as $$R = 4 \pi^2 P_{c}\Lambda Q_{m}\times\frac{\nu_{o}}{\delta\nu_{hom}}\times\frac{\nu_{o}} {\delta\nu_{00}}\times\Gamma(\Delta\omega) \label{eqn:R}$$ where $P_{c}$ is the optical power circulating in the arm cavity, $\nu_{o}$ is the optical frequency of the light, and $\delta\nu_{00}$ and $\delta\nu_{hom}$ are the cavity linewidths (full width, half maximum) for the fundamental and the higher-order optical mode, respectively. The parameter $\Lambda$ measures the spatial overlap between the acoustic and the higher-order optical modes; $\Gamma$ is representative of the interferometer optical configuration and is a function of the 3-mode interaction tuning $\Delta\omega = \omega_{m} - 2\pi \Delta\nu_{hom}$, where $\Delta\nu_{hom}$ is the frequency difference between the fundamental and higher-order optical modes. For $\Delta\omega\rightarrow 0$, the parameter $\Gamma$ reaches its maximum (see [@Strigin200710] for a more detailed description).
Unstable acoustic modes with parametric gain up to $R \simeq 100$ may arise in Advanced LIGO in the 10-90 kHz band [@Evans2010665]. To prevent these instabilities, a damping mechanism must be introduced to reduce the Q-factor of all unstable acoustic modes in this frequency band [*without introducing excess thermal fluctuation in the detection band*]{} of 10 Hz to 1 kHz.
Model of the Acoustic Mode Damper (AMD) {#sec:ds}
=======================================
In order to reduce the Q test mass acoustic modes we designed a resonant AMD which can be attached to the test mass and provide dissipation via the piezo-electric effect.
In this section we first describe the interaction between the AMD and the test mass with a simple 1-D model, then we present a complete Finite Element Model of the entire system.
Simplified 1-D Model {#sec:1D}
--------------------
The AMD and test-mass system can be described as a pair of coupled oscillators with a large mass ratio. The AMD mass $m$ is attached to the much more massive test mass via piezo electric shear plates, which are modeled as a lossy spring with complex spring constant of magnitude $k$ and loss angle $\eta$.
The test mass acoustic mode for which we would like to estimate the impact of the AMD is simplified in this model to a mass $M$, equal to the modal mass of the acoustic mode, attached to a fixed reference by a lossless spring $K$. The coupled systems is then excited by the radiation pressure force $F$ applied to the TM mode, as shown in Fig. \[fig:spring-mass\].
At frequencies near the resonance of the AMD, the lossy spring produced by the piezoelectric material and resistive load will dissipate the energy of the excited acoustic mode, as seen in Fig. \[fig:PZTloss\].
For this system of coupled oscillators, the amplitude $A(\omega)$ of the acoustic mode driven by force $F$ at angular frequency $\omega$ is $$\begin{gathered}
A(\omega) = \frac{F}{M \omega^2} \sqrt{
\frac{\epsilon^2 + \eta^2} {(\delta \epsilon + \mu)^2 + \eta^2 (\delta + \mu)^2 }} \\
\mbox{where}~~ \delta = 1 - \omega_0^2 / \omega^2, ~
\epsilon = 1 - \omega^2 / \omega_D^2 \label{eq:ampl} \nonumber \\
\quad \omega_0^2 = \frac{K}{M}, ~ \omega_D^2 = \frac{k}{m} , ~
\mbox{and} ~ \mu = \frac{m}{M}. \nonumber \\\end{gathered}$$
The resulting effective Q-factor is $$\qeff= \frac{\max ( A(\omega) )}{ A(\omega=0)} \simeq \frac{\eta^2 + (1 - \rho)^2}{\eta \mu \rho}
\label{eq:qeff}$$ where $\rho = \omega_{0} / \omega_{1}$, and we assume $\mu \ll 1$.
When the acoustic mode resonance is near that of the AMD, $\eta \gg |1 - \rho|$, the acoustic mode Q is reduced to $\qeff \simeq \eta / \mu$. When the acoustic mode resonance is well above the AMD resonance, $\qeff \simeq \rho / \eta \mu$, and when it is well below the AMD resonance, $\qeff \simeq 1 / \eta \mu \rho$, assuming $\eta^2 \ll 1$.
To suppress PIs, the test mass acoustic mode Q-factors only need to be reduced from $\approx 10^7$ to $10^5-10^6$. Using this simple model we can estimate that a $3{\rm g}$ AMD with $\eta = 0.1$ on a $10{\rm kg}$ test mass, can give $\qeff \lesssim 10^5$ for resonances with $\tfrac{1}{3} < \rho < 3$.
However, this model ignores a number of important factors. One of these is the location of the AMD relative to the nodes and anti-nodes of each test mass acoustic mode. Quantitatively speaking, the modal mass $M$ of a given mode should be increased in this model by the ratio of the displacement at the AMD location to that of the mode’s anti-node squared $M' = M (x_{\rm max} / x_{\rm AMD})^2$. Thus, an AMD located near a node will have a reduced value of $\mu$, and will provide little damping.
Other important factors include the multiple coupled degrees of freedom of the AMD and the directional nature of the piezo damping material, both of which are covered in the follow section.
.
--------------- ----------------------------------- ---------------------------------- -------------------------------------
Component Material Dimenssions Loss angle
Resistor Ceramic, surface mount 0.7$\times$1.2$\times$0.5 mm$^3$ 10 k$\Omega$
Reaction mass Tungsten 0.7$\times$1.2$\times$0.5 mm$^3$ 4e-5
Epoxy conductive TruDuct 2902 25 $\mu m$ 0.15
Piezo shear plate TRS 200HD 0.7$\times$1.2$\times$0.5 mm$^3$ 0.014
Base fused silica + gold coating 0.7$\times$1.2$\times$0.5 mm$^3$ 7.6$\cdot10^{-12}\cdot$f $ ^{0.77}$
Epoxy non-conductive (EP30, MasterBond) 25 $\mu$m 0.1
\[1ex\]
--------------- ----------------------------------- ---------------------------------- -------------------------------------
Finite Element Model (FEM) {#sec:FEM}
--------------------------
The simple 1-D model introduced in the previous section is useful to provide an intuitive understanding of the AMD damping mechanism. However, it is inadequate to represent the details of the interaction between the AMD and the test mass acoustic modes (TMAMs). Each AMD has at least six resonant modes and hundreds of TMAMs are present in the frequency band of interest; a Finite Element Model is required in order to properly reproduce these modes and calculate the mode overlap between them.
A FEM of the Advanced LIGO test mass with two attached AMDs was constructed with the ANSYS program [@ANSYS]. The AMD model corresponds to the parameters of our prototype AMD (see Table \[table:AMDcomp\]), and the test mass model parameters are reported in Table \[table:TM\]. All dissipation mechanisms in the test mass substrate, coating and bonds were included in the FEM, along with all the losses related to the acoustic mode damper structure (see table \[table:TNres\] for a full list).
A piezoelectric material (PZT), for which energy dissipation can be easily controlled, is ideal for AMD construction. The AMD design modeled here has 2 PZTs sandwiched between a reaction mass and the interferometer test mass (see figure \[fig:spring-mass\]). The PZTs operate in shear, and are poled orthogonally to ensure that all but 1 of the 6 lowest frequency AMD modes are damped. See table \[tab:AMDreso\] for a list of modes; the compression mode is not damped by the shear plates.
For the shear resonant damper, the spring constant can be associated with the shear deformation of the piezomaterial $$k(1+i\eta) = Re(c^{su}_{55})\left(1 + i \eta^{pzt}\right)\frac{S}{h},$$ where $c_{55}$ is the shear stiffness matrix element, $S$ is the area, and $h$ is the height of the shear plate, respectively.\
The loss factor $\eta^{pzt}$ is induced by shunting the shear plate with a resistor. The active stiffness component $c_{55}$ becomes a complex quantity with a nonzero imaginary stiffness due to the shunt. The magnitude of $Im(c_{55})$ is strictly related to the impedance of the shear plate-resistor circuit. The loss factor of the shunted shear plate can be defined as $$\eta^{pzt} = \frac{Im(c^{su}_{55})}{Re(c^{su}_{55})}.$$
Because the impedance of any PZT is capacitive, the loss factor $\eta^{pzt}$ is frequency dependent. As such, careful selection of the shunting resistor and PZT dimensions can be used to maximize loss in the band of interest, as shown in figure \[fig:PZTloss\]. A more detailed discussion of the piezoelectric loss angle can be found in Appendix \[a:pzt\].
We validated the FEM of the AMD by computing its principle resonances, and comparing them with direct measurements performed on a prototype AMD. A total of five principle resonances were identified, with three types of modes: two flag, two anti-flag, and one rotation mode (see Table \[tab:AMDreso\]). All these modes are characterized by large shear deformation for the double piezo configuration in the AMD. The sixth compression mode was not measured as it does not involve shear of the piezo plate. Table \[tab:AMDreso\] shows the good agreement between the output of the model and measurements on the AMD prototype.
------------------ ------------- ------------------
Mode type FEM \[kHz\] Measured \[kHz\]
\[0.5ex\]Fy-mode 9.96 9.77
Fx-mode 12.87 12.61
R-mode 23.28 24.13
aFy-mode 38.36 37.39
aFx-mode 50.51 48.86
\[1ex\]
------------------ ------------- ------------------
: Verification of the finite element model for AMD. Five principle resonances (two flag, two anti-flag and one rotational mode) obtained with the model are compared to the measured values. The x, y suffix corresponds to the shear plate polarization direction. The principle resonances of the AMD were measured with a capacitive bridge circuit where one of the matching capacitors was the AMD prototype. The difference between flag and anti-flag modes corresponds to the location of the rotation axis about which reaction mass rocks. For the flag pole the rotation axis is at the bottom surface of the shear plate whereas for the anti-flag mode the rotation axis is at the height of mass center of the reaction mass.
\[tab:AMDreso\]
Harmonic analysis (finite element analysis with an excitation at a fixed frequency) was conducted to estimate the Q-factor of each of the TMAMs. This approach allows us to include frequency dependent variables such as the shear plate stiffness $\Re(c_{55}^{su})$, and its loss angle (see Appendix \[a:pzt\]).
The modal Q-factor of each resonant mode of the system was calculated as $$Q(f_n) = \frac{\sum_{m}E_m(f_n)}{\sum_{m} E_m(f_n) \, tan(\phi_{m}) }\label{eq:eff_loss},$$ where $f_n$ is the frequency of the $n^{\rm th}$ acoustic mode (see figure \[fig:modes\]). The loss associated with each structural component is treated separately; $E_m$ is the modulus of the strain energy in the $m^{\rm th}$ component and $\tan(\phi_{m})$ is its loss factor (see tables \[table:AMDcomp\] and \[table:TM\]). The modeled resonant frequencies of 12 TMAMs were then compared to the measured resonance frequencies (see table \[table:TMresults\]), obtained according to methods described in section \[sec:er\]. The front face displacement of the modes are shown in Fig. \[fig:modes\].
Note that the agreement between calculated and measured Q-values increases with frequency. This may indicate an additional dissipation process which is missing in the model. Nevertheless, a small relative frequency offset $\Delta f$ below 1% in Table \[table:TMresults\] indicates a good agreement of the FEM with measured values.
------------ -------------- -------------- ------------- ----- ---------- -------
Mode \# Freq. \[Hz\] Freq. \[Hz\] $\Delta$f % Q-factor
\[0.5ex\] FEM Measured FEM Measured Ratio
\[0.5ex\]1 8128.3 8150.9 0.3 37M 1.9M 19.5
2 10391.1 10418.1 0.3 63M 14M 4.5
3 12999.1 12984.7 0.1 29M 15M 1.9
4 15101.4 15047.2 0.4 56M 17M 3.3
5 15151.0 15539.1 2.5 55M 16M 3.4
6 19487.0 19544.9 0.3 30M 7.0M 4.3
7 20113.6 20185.5 0.3 27M 13M 2.1
8 24824.0 24901.8 0.3 32M 16M 2.0
9 26504.3 26681.2 0.7 48M 18M 2.7
10 29767.4 29699.5 1.0 18M 15M 1.1
11 30912.1 31003.3 0.3 18M 12M 1.5
12 32664.1 32743.2 0.2 14M 13M 1.1
\[1ex\]
------------ -------------- -------------- ------------- ----- ---------- -------
: The Q-factor of the test mass modes without AMDs installed. The 4th and 7th column correspond to the relative frequency noise $\Delta$f% and Q-factor ratio between calculated and measured value, respectively.
\[table:TMresults\]
Experimental Results {#sec:er}
====================
Several AMD prototypes were constructed, each consisting of six components (see Table \[table:AMDcomp\]):
[*Reaction Mass*]{}: A 10 g tungsten cylinder which tunes the AMD principle resonances to frequencies above 10 kHz is located on top of the shear plates.
[*Shear Plate*]{}: Two piezoelectric shear plates, oriented with perpendicular polarizations are glued to the reaction mass and base with conductive epoxy Tra-Duct2902. The epoxy serves to electrically connect the PZT electrodes to the reaction mass and base.
[*Base*]{}: The interface between flat shear plates and curved barrel of the test mass. The top flat surface is gold coated with separate sections to which shear plates are glued. The bottom surface is curved and match to the test mass barrel curvature. The base is made from fused silica and glued to the test mass with nonconductive epoxy.
[*Shunting circuit*]{}: One 10 k$\Omega$ resistor for each shear plate is glued to the reaction mass with conductive epoxy. The circuit is closed with 100 $\mu$m diameter silver coated copper wire, which is soldered to the resistor and gold coated surface of the base.\
To measure Q-factors of the test mass modes, both with and without AMDs attached, a 16 m optical cavity was used, with the test mass forming the end mirror of the cavity. The optical cavity supports a 2 mm diameter Gaussian mode; the resonant beam probes the motion in the center of the test mass face, so that modes with an anti-node in the center of the mass are easily measured. A total of 12 modes were identified (Fig. \[fig:modes\]) and measured, with the results shown in Table \[table:TMresults\].
In order to measure the impact of the AMD on the Q of the TMAMs, several modes of the test mass were excited using electrostatic actuators [@0264-9381-29-11-115005] and their ring-down times observed. This measurement was repeated in three configurations: with no AMD, with 2 AMDs, with 2 AMDs which lacked resistive damping (shunt wires cut).
Each acoustic mode was detected in the cavity locking error signal as a peak in the Fourier domain. After being excited, each mode amplitude was recorded as a function of time to estimate the decay time $\tau$. The Q-factor was determined from the decay time $\tau$, according to $$Q = \pi f_{0} \tau \label{eq:q} .$$ where $f_{0}$ is the resonant frequency in Hz and $\tau$ is the exponential decay time constant.
Measurements of the test mass mode Q-factor performed after installing two AMDs clearly show the substantial damping capability of the AMD prototype, as reported in Table \[table:TMresults\] (see also Table \[table:AMDresults\] in Appendix for additional details).
The results indicate that out of 12 modes, 11 are suppressed by at least factor of 10 and in some cases by more than two orders of magnitude. The relatively large discrepancy between model and measurement for mode \#2 can be explained by the AMD and TM interaction condition for this particular mode. The FEM predicts that mode \#2 will be within 500Hz of the AMD Fy-mode at 10 kHz, while the measured values give a separation of 800Hz (mostly due to the AMD resonance being off), reducing the interaction between the AMD and TM mode.
In the off-resonance interactions, which are more common and set the lower limit to AMD performance, the discrepancy is generally less than a factor of a few.
Surprisingly, modes \#2 and \#6 show a counterintuitive behavior; opening the resistive circuit of the AMDs to decreases the TMAM Q-factor. However, since electrical circuit of AMD affects mechanical stiffness of the shear plates it is expected that principle resonances also changed when the circuit is opened. Also, if the AMD resonance is close to the TMAM frequency, equation \[eq:qeff\] indicates that the mode Q can be decreased by *lowering* the AMD loss.
The large Q reduction for modes \#3 and \#4 is due the on-resonance interaction between TMAM and AMD, whereas for modes \#5 and \#12 the large Q reduction is due to the anti-node AMD location on the TMAM.
----------- -------------- --------- --------------------
Mode \# Freq. \[Hz\] Damping Resistive
\[0.5ex\] factor contribution \[%\]
1 8150.9 32.2 12
2 10418.1 31.8 -57
3 12984.7 441.2 21
4 15047.2 81.0 9
5 15539.1 $>$320 $>$71
6 19544.9 15.6 -2
7 20185.5 13.8 41
8 24901.8 5.2 4
9 26681.2 $>$360 $>$0
10 29699.5 23.8 43
11 31003.3 307.7 68
12 32743.2 $>$260 $>$4
\[1ex\]
----------- -------------- --------- --------------------
: Test mass mode suppression obtained with two AMDs.
\[table:results\]
Analysis of the AMD thermal noise {#sec:tn}
=================================
The AMD is designed to increase the mechanical damping of the test mass acoustic modes above 10 kHz. At the same time, the AMD must introduce minimal additional mechanical loss in the 10-1000 Hz band, where low mechanical loss is required to keep test mass thermal noise small [@FDT]. Thus it is critical to calculate the thermal noise resulting from the AMDs in our overall evaluation of this PI mitigation technique.
We used the FEM described above to calculate the thermal noise resulting from our experimental test of two AMDs attached to a test mass. The AMD thermal noise was calculated numerically at 100 Hz, the most sensitive part of the detection band, using Levin’s approach [@PhysRevD.57.659]. The energy dissipation per cycle was computed using Eqn. \[eq:losseff\] for a pressure profile corresponding to the Advanced LIGO geometry (laser beam radius $\omega_{0} = 5.5$ cm, incident on the front face of the TM). Results are shown in Table \[table:TNres\].
The test mass thermal noise level of $5.2 \times 10^{-21}{\rm m/\sqrt{Hz}}$ corresponds to the design level for Advanced LIGO, and is dominated by optical coating loss [@0264-9381-19-5-305; @Evans2008]. As Table \[table:TNres\] shows, while our *prototype* AMD would contribute significantly to thermal noise it is not orders of magnitude above the more fundamental sources of thermal noise.
[l c]{} Component & Thermal noise @ 100 Hz\
& $[10^{-21} m / \sqrt{Hz}]$\
\
Substrate & 0.8\
Optical coating & 5.1\
Suspension ears & 0.0\
Ears bond & 0.6\
Total per TM & [**[5.2]{}**]{}\
\
Reaction mass (RM)& 0.3\
Epoxy (RM-PZT)& 4.8\
Shear plate (PZT-X)& 5.5\
Shear plate (PZT-Y)& 5.5\
Epoxy Base-PZT & 7.2\
Base & 0.0\
Epoxy Base-TM & 6.0\
Total per AMD & [**[13]{}**]{}\
\[table:TNres\]
Relative to the prototype device, our model points to several design and material improvements that can be made to significantly reduce the thermal noise impact. The major AMD thermal noise contributors are the epoxies used to bond the AMD elements, and structural loss in the piezo shear plates. The former can be improved with lower loss epoxy and thinner bond layers. The latter can be improved with a more suitable choice of piezo material.
Other design elements can also be altered. The mass of the reaction mass can be reduced to lower the thermal noise without significantly affecting the acoustic mode damping performance. Another modification would be to avoid alignment of the piezo shear plate polarization with the laser beam axis, to minimize the contribution of resistive loss to the thermal noise. These and other design optimizations will be explored in a future paper.
Conclusion {#sec:con}
==========
Acoustic mode dampers represent a simple yet effective approach to damping parametric instabilities. The great advantage of this approach over active damping [@Miller2011788] is that many test mass acoustic modes are effected simultaneously, and no further intervention is necessary. This is likely to be a critical feature in instruments that would otherwise suffer from multiple acoustic modes simultaneously excited by parametric instabilities.
The investigation presented here involved modeling and construction of a prototype AMD, which was shown to effectively damp test mass acoustic modes. The thermal noise associated with this prototype AMD was also computed. Though the prototype AMD does not meet the stringent thermal noise requirements of gravitational wave detectors, several design elements were identified for improvement, making this approach a viable solution to parametric instabilities.
The authors gratefully acknowledge the support of the National Science Foundation and the LIGO Laboratory, operating under cooperative Agreement No. PHY-0757058 This paper has been assigned LIGO Document No. LIGO-P1400257.
\[a:pzt\] Piezoelectric material
================================
Any piezoelectric material is strictly characterized by electromechanical properties. The fact that both electrical and mechanical properties are inseparable allows us to represent the dissipation process in a shunted piezoelectric material either as a Jonson heat or mechanical loss. For the purpose of this work we focus on disspation process in terms of the mechanical loss.\
From the stress-charge of the piezoelectric constitutive equation it is straightforward to derive the total induced stress in the shunted piezoelectric material [@Hagood1991243]. Assuming no external current plugged to the piezoelement electrode we get $$\boldsymbol{\sigma}_{(6\times 1)} = \left(\boldsymbol{c}^{E}_{(6\times
6)}-i\omega\frac{\boldsymbol{e}^{t}_{(6\times
3)}\boldsymbol{Z}^{TOT}_{(3\times 3)}\boldsymbol{A}_{(3\times
3)}\boldsymbol{e}_{(3\times 6)}}{\boldsymbol{L}_{(3\times
3)}}\right)\boldsymbol{S}_{(6\times 1)} \label{eqA:stress},$$ where $\boldsymbol{c}^{E}$ is the mechanical stiffness matrix under constant electric field, $\boldsymbol{e}^{t}$ and $\boldsymbol{e}$ are the piezoelectric stress constant. The upper script $t$ refers to transpose operator [@26560]. Electrode area is represented by matrix $\boldsymbol{A}$, whereas thickness of the piezomaterial between electrodes by matrix $\boldsymbol{L}$. The strain $\boldsymbol{S}$ is a product of the acting stress $\boldsymbol{\sigma}$ on piezo-element and charge accumulation in the piezo-element. Each bracket corresponds to the matrix dimenssion. It is assumed that piezoelectric element has a brick shape thus matrices $\boldsymbol{Z}^{TOT}$,$\boldsymbol{A}$, and $\boldsymbol{L}$ are diagonal.\
The total impedance in Eqn.\[eqA:stress\] is inverse sum of the piezoelement admittance under constant electric field $\boldsymbol{Y}^E$ and the admittance of the external circuit $\boldsymbol{Y}^{SU}$ connected to the piezoelement electrodes, thus $$\boldsymbol{Z^{TOT}} =
\left(\boldsymbol{Y}^E+\boldsymbol{Y}^{SU}\right)^{-1}
\label{eqA:impedance}.$$ The admittance of piezelement $\boldsymbol{Y}^E$ is assumed to be exclusively capacitive. Hence, $$\boldsymbol{Y}^E =
i\omega\boldsymbol{A}\boldsymbol{\epsilon}^S\boldsymbol{L}^{-1}=i\omega\boldsymbol{C}^S=i\omega\boldsymbol{C}^T\left(\boldsymbol{\epsilon}^T\right)^{-1}\boldsymbol{\epsilon}^S
\label{eqA:piezoadmitance},$$ where $\boldsymbol{C}$, $\boldsymbol{\epsilon}$ is the capacitance and the dielectric constant matrices under constant strain $S$ and constant stress $T$, respectively. It is convenient to operate with $\boldsymbol{C}^T$ since this quantity can be easily measured at stress free conditions and no shortened piezelement electrodes.\
According to Eqn.\[eqA:stress\], if the piezoelement is integrated with a nonzero impedance electric circuit, the imaginary part stiffness tensor arises. The imaginary part can be interpreted as a dissipative component of the stiffness tensor. The total shunted stiffness matrix is $$\boldsymbol{c}^{SU} =
\boldsymbol{c}^{E}-i\omega\boldsymbol{e}^{t}\boldsymbol{Z}^{TOT}\boldsymbol{A}\boldsymbol{e}\boldsymbol{L}^{-1}
\label{eqA:comp} .$$ Because matrices $\boldsymbol{A}$, $\boldsymbol{Z}^{TOT}$, and $\boldsymbol{L}$ are diagonal and knowing that $\boldsymbol{A}\boldsymbol{L}^{-1} =
\boldsymbol{C}^{T}(\boldsymbol{\epsilon}^T)^{-1}$ Eqn. \[eqA:comp\] can be written as $$\boldsymbol{c}^{SU} =
\boldsymbol{c}^{E}-\boldsymbol{e}^{t}\boldsymbol{e}\boldsymbol{\zhat}\left(\boldsymbol{\epsilon}^S\right)^{-1}
\label{eq:comp_final},$$ where $\boldsymbol{\widehat{Z}}$ is the nondemensional impedance $$\boldsymbol{\zhat} = \boldsymbol{Z}^{TOT}\boldsymbol{Y}^E \label{eqA:imp}.$$ Note, for zero impedance shunted circuit sets $\boldsymbol{\zhat}$ to be a unity matrix $\boldsymbol{\zhat}=\boldsymbol{I}$ whereas for the nonzero impedance shunting $\boldsymbol{\zhat}$ is a complex quantity. Using indexing Eqn.\[eqA:comp\] can be written as $$c_{ij}^{su} = c_{ij}^{E}- \frac{e_{ki} e_{kj}\zhat_{kk}}{\epsilon_{kk}^S}
\label{eqA:general} .$$ where $k = 1,2,3$ and corresponds to the electrode position, $\boldsymbol{\epsilon}^S$ is the dielectric permittivity of the piezoelectric material.
From Eqn.\[eqA:piezoadmitance\] and \[eqA:general\] becomes clear that loss can be control with the shunting circuit. Moreover, such loss is in fact a frequency dependent quantity with arbitrary shape of the loss curve dependent on shunting circuit.\
For the shear plate with the single pair of electrodes only a stiffness matrix element $c_{55}^{su}$ is affected by shunting resistor, thus $$c_{55}^{su} = c_{55}^{E}- \frac{e_{15} e_{15}\zhat_{11}}{\epsilon_{11}^S}
\label{eqn:55} .$$ and the nondemensional complex impedance $\zhat_{kk}$ has a single nonunity element $$\begin{aligned}
\zhat_{11} = \frac{i\omega\epsilon_{11}^S RC^T}{i\omega\epsilon_{11}^S
RC^T+\epsilon_{11}^T} = \nonumber\\
\frac{\left(\omega\epsilon_{11}^S
RC^T\right)^2}{\left(\omega\epsilon_{11}^S
RC^T\right)^2+\left(\epsilon_{11}^T\right)^2}+\nonumber\\
i\frac{\omega\epsilon_{11}^S\epsilon_{11}^T
RC^T}{\left(\omega\epsilon_{11}^S
RC^T\right)^2+\left(\epsilon_{11}^T\right)^2} \label{eqn:z11}.\end{aligned}$$ Since the stiffness matrix $\bm{c}^{su}$ is a complex quantity we can define the loss factor $\bm{\eta}$ as $$\bm{\eta} = \frac{Im(\boldsymbol{c}^{su})}{Re(\boldsymbol{c}^{su}) }
\label{eqA:loss1},$$ and thus $$\eta_{kij} = \frac{Im(\zhat_{kk})\chi_{kij}}{Re(\zhat_{kk})\chi_{kij}+1}
\label{eq:loss2},$$ where $\chi_{kij}= e_{ki} e_{kj} (c_{ij}^{E}\epsilon_{kk}^S)^{-1}$. Index $k$ is associated with the electrodes orientation, whereas indices $i$ and $j$ correspond to stress-strain directions in the stiffness matrix $\boldsymbol{c}^{su}$. The shunting loss factor of the shear plate becomes $$\eta_{155} = \frac{\Im(\zhat_{11})\chi_{155}}{\Re(\zhat_{11})\chi_{155}+1}\label{eq:loss3}.$$ This is the main loss mechanism based on which shear AMD operates. Note that according to Eqn.\[eqn:55\], the material stiffness (the real part of $c_{55}^{su}$) is also reduced and should be included in analysis.\
Total mechanical loss angle
---------------------------
An assumption of the stiffness matrix being real in the absence of shunt is not sufficient for accurate estimation of energy dissipation. It becomes especially important in the thermal noise analyisis, see Sec. \[sec:tn\].\
For known material loss angle of the nonshunted piezoelement, the total loss factor of piezoelectric material can be defined as $$\eta_{kij}^{tot} =
\frac{tan(\phi_{ij}^{mat})+Im(\zhat_{ii})\chi_{kij}}{1+Re(\zhat_{ii})\chi_{kij}}
\label{eqA:losstot},$$ where $\tan(\phi_{ij}^{mat})$ is the material loss factor matrix of piezo-element. The loss factor $\tan(\phi_{ij}^{mat})$ of $\boldsymbol{{c}^E}$ can be easily computed using Eqn. \[eqA:loss1\].
All piezoelectric material has anisotropic structure what implies that strain energy dissipation in such material must depend on the piezoelement geometric shape. It is more convenient to use effective loss angle $\phi_{eff}$ rather than the loss angle $\phi$ for each stiffness matrix component. We define the energy dissipation per cycle in the piezo-element $$\begin{aligned}
W_{dis} &=& 2 \pi W_{st} \eta_{eff} \\
&=& \int{Re(S_{i}S_{j}^{*}) Re(c_{ji}^{su})\eta_{ji}^{tot}}dV \label{eq:losseff}\end{aligned}$$ where $W_{st}$ is the stored strain energy such that $2\pi W_{st} =
\int{Re(S_{i}T_{j}^{*})}dV$, where S, T are the complex strain and stress, respectively and V is the volume of the piezo-element.\
Since loss factor of the piezoelement is inverse of its Q-factor or equally a ratio of dissipated energy per cycle $W_{dis}$ to energy stored $W_{st}$ in the piezoelement, we can defined the effective noise fuctor $\eta_{eff}$ $$\eta_{eff} = \frac{W_{dis}}{2\pi W_{st}}=\frac{\int Re(S_{i}S_{j}^*)
Re(c_{ji}^{su})\eta_{ji}^{tot}dV}{\int Re(S_{i}T_{j}^{*})dV}
\label{eqA:losseff},$$ where $S$, $T$ are the complex strain and stress, respectively and $V$ is the volume of the piezo-element. This is a key equation in the finite element analysis which properly estimate contribution of piezoelemnt in the strain energy dissipation in the test mass.
In our analysis we assumed constant material intrinsic loss factor for all $\boldsymbol{c}^E$ elements equal to $\tan(\phi_{ij}^{mat})$=0.014 [^1], what leads to $$\eta_{155}^{tot} =
\frac{0.014+Im(\zhat_{11})\chi_{155}}{1+Re(\zhat_{11})\chi_{155}}
\label{eq:losstot} .$$ Note, the remaining elements of shunting induced loss $\eta_{kij}^{tot}$ are equal to the material structural loss factor. For the shear plate configuration $\chi_{155}$ is simply a function of electromechanical coupling coefficient $k_{15}$ and is equal to $\chi_{155} =
\frac{k_{15}^2}{1-k_{15}^2}$. The total loss angle for the $c_{55}^E$ stiffness element is shown in Fig.\[fig:PZTloss\].
Additional Tables
=================
[lll]{} [***Optical parameters:***]{}&&\
End test mass transmissivity &5.0 ppm&\
\
&Ta$_2$O$_5$&SiO$_2$ Ear bond\
Young’s modulus&140 GPa&70 GPa 7.2 GPa\
Poisson ratio&0.23&0.170.17\
Density&8300 kg/m$^3$&2201 kg/m$^3$2202 kg/m$^3$\
Refractive index&2.06539&1.45-\
Loss angle&2.4$\cdot10^{-4}$+f$\cdot$1.8$\cdot10^{-9}$ &0.4$\cdot10^{-4}$+f$\cdot$1.1$\cdot10^{-9}$0.1\
&&\
&&\
Radius&0.17 m&\
Thickness&0.2 m&\
Flats width&0.095 m&\
Wedge angle &0.07 deg&\
Mass&40 kg&\
Loss angle&7.6$\cdot10^{-12}\cdot$f $ ^{0.77}$ &\
Material& fused silica\
\
------------------ ------ ---------- ------ ----------
\[0.5ex\]Mode \#
FEM Measured FEM Measured
1 52k 59k 70k 67k
2 7.9k 440k 9.9k 280k
3 23k 34k 23k 43k
4 420k 210k 510k 230k
5 2.4M $<$50k 2.2M 170k
6 1.9M 450k 4.8M 440k
7 1.1M 940k 1.6M 1.6M
8 6.7M 3.1M 9.5M 3.2M
9 49k $<$50k 64k $<$50k
10 1.9M 630k 3.3M 1.1M
11 61k 39k 116k 120k
12 25k $<$50k 260k 52k
\[1ex\]
------------------ ------ ---------- ------ ----------
: Results of the Q-factor measurement with attached AMDs for shunted and non-shunted cases.
\[table:AMDresults\]
[^1]: The information about the imaginary part of the stiffness matrix is not available. We assumed that the material loss factor corresponds to the Q-factor of the piezomatrial provided by the manufacturer. In our opinion low Q PZT should have fairly constant loss for all stiffness components
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Computer vision has achieved impressive progress in recent years. Meanwhile, mobile phones have become the primary computing platforms for millions of people. In addition to mobile phones, many autonomous systems rely on visual data for making decisions and some of these systems have limited energy (such as unmanned aerial vehicles also called drones and mobile robots). These systems rely on batteries and energy efficiency is critical. This article serves two main purposes: (1) Examine the state-of-the-art for low-power solutions to detect objects in images. Since 2015, the IEEE Annual International Low-Power Image Recognition Challenge (LPIRC) has been held to identify the most energy-efficient computer vision solutions. This article summarizes 2018 winners’ solutions. (2) Suggest directions for research as well as opportunities for low-power computer vision.'
author:
- 'Sergei Alyamkin, Matthew Ardi, Alexander C. Berg, Achille Brighton, Bo Chen, Yiran Chen, Hsin-Pai Cheng, Zichen Fan, Chen Feng, Bo Fu, Kent Gauen, Abhinav Goel, Alexander Goncharenko, Xuyang Guo, Soonhoi Ha, Andrew Howard, Xiao Hu, Yuanjun Huang, Donghyun Kang, Jaeyoun Kim, Jong Gook Ko, Alexander Kondratyev, Junhyeok Lee, Seungjae Lee, Suwoong Lee, Zichao Li, Zhiyu Liang, Juzheng Liu, Xin Liu, Yang Lu, Yung-Hsiang Lu[^1], Deeptanshu Malik, Hong Hanh Nguyen, Eunbyung Park, Denis Repin, Liang Shen, Tao Sheng, Fei Sun, David Svitov, George K. Thiruvathukal, Baiwu Zhang, Jingchi Zhang, Xiaopeng Zhang, Shaojie Zhuo'
bibliography:
- 'ms.bib'
title: 'Low-Power Computer Vision: Status, Challenges, Opportunities'
---
[Author ]{}
Computer Vision, Low-Power Electronics, Object Detection, Machine Intelligence
Introduction
============
Competitions are an effective way of promoting innovation and system integration. “Grand Challenges” can push the boundaries of technologies and open new directions for research and development. The DARPA Grand Challenge in 2004 opened the era of autonomous vehicles. Ansari X Prize started a new era of space tourism. Since 2010, the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) has become a popular benchmark in computer vision for detecting objects in images. ILSVRC is an online competition: contestants submit their solutions through a website. The only factor for comparison is the accuracy and there is no strict time limit.
Phones with cameras appeared as early as year 2000. Since 2011, smartphones outsell personal computers and become the primary computing platforms for millions of people. Computer vision technologies have become widely used on smartphones. For example, people use smartphones for comparison shopping by taking photographs of interested products and search for reviews, similar products, and prices. Face detection has become a standard feature on digital cameras. Better computer vision technologies, however, are not the “most desired” features for future smartphones. Instead, longer battery life has consistently ranked as one of the most desired features. Recognizing the need for energy-efficient computer vision technologies, the IEEE Rebooting Computing Initiative started the first IEEE Low-Power Image Recognition Challenge (LPIRC) in 2015. LPIRC aims to identify energy-efficient solutions for computer vision. These solutions have a wide range of applications in mobile phones, drones, autonomous robots, or any intelligent systems equipped with digital cameras carrying limited energy.
LPIRC is an annual competition identifying the best system-level solutions for detecting objects in images while using as little energy as possible [@7858303; @7372672; @8342099; @AIM2018; @Lu2019]. Although many competitions are held every year, LPIRC is the only one integrating both image recognition and low power. In LPIRC, a contestant’s system is connected to the referee system through an intranet (wired or wireless). There is no restriction on software or hardware. The contestant’s system issues HTTP GET commands to retrieve images from the referee system and issues HTTP POST commands to send the answers. To reduce the overhead of network communication, the referee system allows retrieving 100 images at once as a zip file. Also, one HTTP POST command can send multiple detected objects from many images. Each solution has 10 minutes to process all images (5,000 images in 2015 and 20,000 images since 2016). The scoring function is the ratio of accuracy (measured by [*mean average precision*]{}, mAP) and the energy consumption (measured by Watt-Hour). Table \[table:scores2015-2018\] shows the champions’ scores. As can be seen in this table, the scores improve 24 times from 2015 to 2018. The best mAP in LPIRC is lower than the best score in the ImageNet Challenge (ILSVRC). The winner of the ImageNet Challenge (in 2017) achieved mAP of 0.731. The ImageNet Challenge has no time restriction, nor is energy considered.
[|llllr|]{} Year & Accuracy & Energy & Score & Ratio\
2015 & 0.02971 & 1.634 & 0.0182 & 1.0\
2016 & 0.03469 & 0.789 & 0.0440 & 2.4\
2017 & 0.24838 & 2.082 & 0.1193 & 6.6\
2018 (Track 2) & 0.38981 & 1.540 & 0.2531 & 13.9\
2018 (Track 3) & 0.18318 & 0.412 & 0.4446 & 24.5\
\
\
\
\
This paper has two major contributions: (1) It describes an international competition in low-power image recognition and the winners have demonstrated significant progress. The importance of this competition is reflected by the list of sponsors as well as the profiles of the teams. (2) The paper provides an overview of different methods that won the competition. The winners explain the design decisions and the effects of these factors. This paper extends an earlier online version that focuses on the 2018 LPIRC [@LPIRC2018arXiv] by adding the discussion of future research directions for low-power computer vision. This paper is organized as follows. Section 2 describes the different tracks in LPIRC. Section 3 summarizes the scores. Section 4 explains the 2018 winners’ solutions. Section 5 describes industry support for low-power computer vision. Section 6 discusses challenges and opportunities for low-power computer vision. Section 7 concludes the paper.
Tracks of LPIRC
===============
In the past four years, LPIRC has experimented with different tracks with different rules and restrictions. In 2015, offloading of processing was allowed but only one team participated. In 2016, a track displayed images on a computer screen and a constantant’s system used a camera (instead of using the network) to capture the images but only one team participated in this track. These two tracks are no longer offered. Since 2015, LPIRC has been an onsite competition: contestants bring their systems to conferences (Design Automation Conference for 2015 and 2016, International Conference on Computer Vision and Pattern Recognition for 2017 or 2018). Onsite competitions are necessary since LPIRC has no restrictions on hardware or software. Even though this gives contestants the most flexibility, the need to travel to the conferences with hardware potentially discourages some participants. In 2018, two new tracks were created to encourage participation. These two tracks allowed online submissions and contestants did not have to travel. The third track for 2018 LPIRC was the original track always been offered: no restriction on hardware or software. The following sections describe the tracks in 2018 LPIRC.
Track 1: TfLite Model on Mobile Phones
--------------------------------------
![Relationship between MACs and latency [@ChenGilbert20180420]. []{data-label="fig:maclatency"}](track1-latency-vs-mac.jpg){width="3in"}
This new track, also known as the On-device Visual Intelligence Competition, focused on model architecture: contestants submitted their inference neural network models in TfLite format using Tensorflow. The models were benchmarked on a fixed inference engine (TfLite) and hardware model (Pixel 2 XL phone). The task is ImageNet classification. The submissions should strive to classify correctly as many images as possible given a time budget of 30 ms per image. The submissions were evaluated on a single core with a batch-size of one to mimic realistic use cases on mobile devices. Pixel 2 (and 2XL) are mobile phones running Android 8.0. Both use Qualcomm Snapdragon 835 processors and have 4GB memory.
This track provides a benchmarking platform for repeatable measurements and fair comparison of mobile model architectures. The flexibility of submitting just the model and not the full system allows this track to be an online competition. This convenience helps to boost the submission count to 128 models within just 2 weeks. Although the scoring metric is based on inference latency rather than energy consumption, the two are usually correlated when the same benchmarks are used on the same hardware. Track 1’s latency-based metric is critical to accelerating the development of mobile model architectures. Prior to this metric, there was no common, relevant, and verifiable metric to evaluate the inference speed of mobile model architectures for vision tasks. Many papers characterize inference efficiency using unspecified benchmarking environments or latency on desktop as a proxy. Even the most commonly used metric, MACs (multiply-add count) does not correlate well with inference latency in the real-time (under 30 ms) range [@ChenGilbert20180420], as shown in Figure \[fig:maclatency\]. The high participation of this track established the new state-of-the-art in mobile model architecture design. Within the 30 ms latency range, the best of track 1 submissions outperformed the previous state-of-the-art based on quantized MobileNet V1, by 3.1% on the holdout test set, as shown in Figure \[fig:track1holdout\].
![Latency and scores in 2018 Track 1 for the holdout dataset. []{data-label="fig:track1holdout"}](track1-Accuracy-vs-mac.png){width="3in"}
The submissions have demonstrated a considerable amount of diversity. Slightly over half (51.7%) of the solutions are quantized. Most of the architectures (74.1%) are variations of the existing Mobilenet model family, namely quantized V2 (22.4%), quantized V1 (24.1%) and float V2 (27.6%). The predominant dependence on Mobilenets is not surprising, considering their exceptional performance on-device and the convenient support by TfLite. To sustain future research on mobile computer vision, Track 1 should also reward previously under- or un-explored model architectures. In future LPIRC, Track 1 may consider methods for facilitating the discovery of novel architectures and making the platform more inclusive to general vision researchers.
Track 2: Caffe 2 and TX 2
-------------------------
To make online submission possible, Track 2 uses Nvidia TX2 as the hardware platform and Caffe2 as the software platform so that the organizers may execute contestants’ solutions and calculate the scores. To further assist contestants, a software development kit (SDK) is available [@IEEELowPowera]. TX2 has NVIDIA Pascal GPU with 256 CUDA cores. The processor is HMP Dual Denver 2/2 MB L2 and Quad ARM® A57/2 MB L2. It has 8GB 128bit LPDDR4 memory with 59.7 GB/s data rate. Caffe2 is a framework for deep learning supporting GPUs and cloud through the cross-platform libraries. The other parts of Track 2 are identical to Track 3.
Track 3: Onsite, No Restriction
-------------------------------
This is the original track always offered since 2015. This track has no restriction in hardware or software and gives contestants the most flexibility. Figure \[fig:track3\] illustrates the interactions between a contestant’s system and the referee system. The two systems are connected through a network router; both wireless and wired networks are available. A system’s energy consumption is measured using a high-speed power meter, Yokogawa WT310 Digital Power Meter. It can measure AC or DC and can synchronize with the referee system. Each team has 10 minutes to process all the images. There are 200 classes of objects in the competition (same as ImageNet). The training set is not restricted to ImageNet and contestants can use any datasets for training. Test images are stored in the referee system and retrieved through the router. The results are uploaded to the referee system through the same network. After the contestant’s system logs in, the power meter starts measuring the power consumption. The power meter stops after 10 minutes or when a contestant’s system logs out. This allows a team that finishes all images within 10 minutes to reduce their energy consumption. Additional information about Track 3 can be obtained from prior publications [@7858303; @7372672; @8342099; @AIM2018].
To win Track 2 or 3, a solution must be able to detect objects in images and mark the objects’ locations in the images by [*bounding boxes*]{}. A successful object detection must identify the category correctly and the bounding box must overlap with the correct solution (also called “ground truth”, marked by the LPIRC organizers) by at least 50%. A superior solution must be able to detect objects and their locations in the images as fast as possible, while consuming as little energy as possible.
![Interactions between a contestant’s system (the black box) and the referee system. []{data-label="fig:track3"}](track-2-system.png){width="3in"}
Training and Testing Data
-------------------------
LPIRC uses the same training images as ILSVRC. For Track 1, the data from localization and classification task (LOC-CLS) is used. It consists of around 1.2 million photographs, collected from Flickr and other search engines, hand-labeled with the presence or absence of 1,000 object categories. Bounding boxes for the objects are available, although Track 1 only considers classification accuracy as a performance metric. For tracks 2 and 3, the data from object detection is used. It has around 550,000 images and bounding boxes of labels for all 200 categories.
For testing data, Track 1 used newly created data with the ILSVRC image crawling tool to collect 100 images for each category from Flickr. When crawling the images, synonyms are used so that relevant images are classified into the corresponding categories. The competition uses only the images uploaded after June 2017. Thumbnail images are used to remove duplicates by resizing the images to 30 $\times$ 30, and calculating the L2-norm of differences with images in the previous dataset. Ten representative images are manually chosen from each category. Representative images refer to the images that can be identified as one of the 1000 ImageNet classification categories without ambiguity. Figure \[fig:track1test\] shows samples test images for Track 1. Tracks 2 and 3 also use the crawling tool to obtain images from Flickr with the context query phrases (e.g., “cat in the kitchen”) instead of the label names (e.g., cat) as the keywords. The tool retrieves images with various types of objects. The annotations (bounding boxes) are created by using Amazon Mechanical Turk.
![Sample test images for Track 1. []{data-label="fig:track1test"}](track2-examples.jpg "fig:"){width="3in"} ![Sample test images for Track 1. []{data-label="fig:track1test"}](track2-example2.jpg "fig:"){width="3in"}
Other Datasets for Image Recognition
------------------------------------
For reference, this section describes some popular datasets used in computer vision: PASCAL VOC, COCO (Common Objects in COntext), SUN (Scene UNderstanding), INRIA Pedestrian Dataset, KITTI Vision Benchmark Suite, and Caltech Pedestrian Dataset. Over time the image datasets have improved in two primary ways. First, the quality of the image annotations has significantly improved due to more sophisticated methods for crowdsourcing. Second, the variety of the dataset has increased, in both content and annotations, by increasing the number of classes.
PASCAL VOC started its first challenge in 2005 for object detection and classification of four classes. In 2012, the final year of the competition, the PASCAL VOC training and validation datasets consisted of 27,450 objects in 11,530 images with 20 classes. From 2009 - 2005 the overall classification average precision improved from 0.65 - 0.82 and detection average precision improved from 0.28 - 0.41 [@Everingham:2015:PVO:2725268.2725369]. The COCO competition continues to be held annually with a primary focus of correctness. COCO contains over 2.5 million labeled instances in over 382,000 images with 80 common objects for instance segmentation [@10.1007/978-3-319-10602-1_48]. The performance of a model on COCO has improved for bounding-box object detection from 0.373 to 0.525 during 2015 to 2017.
2018 LPIRC Scores
=================
Track 1 Scores
--------------
Table \[table:track1winner\] shows the score of the winner of Track 1. It uses ImageNet for validation. The holdout set is freshly collected for the purpose of this competition. The terms are defined as
- Latency (ms): single-threaded, non-batch runtime measured on a single Pixel 2 big core of classifying one image
- Test metric (main metric): total number of images corrected in a wall time of 30 ms divided by the total number of test images
- Accuracy on Classified: accuracy in \[0, 1\] computed based only on the images classified within the wall-time
- Accuracy/Time: ratio of the accuracy and either the total inference time or the wall-time, whichever is longer
- \# Classified: number of images classified within the wall-time
Validation Set Holdout Set
------------------------ ---------------- -------------
Latency 28.0 27.0
Test Metric 0.64705 0.72673
Accuracy on Classified 0.64705 0.72673
Accuracy / Time 1.08 E-06 2.22 E-06
\# Classified 20000 10927
: Track 1 Winner’s Score[]{data-label="table:track1winner"}
Track 1 received 128 submissions. Figure \[fig:track1scores\] represents a total of 121 valid submissions (submissions that passed the bazel test and successfully evaluated). Among them, 56 submissions received test metric scores between 0.59 and 0.65. The mean is 0.4610; the median is 0.5654; the mode is 0.6347; the standard deviation is 0.2119.
Some duplicate submissions that are successfully evaluated have differences with the “Test Metrics” score due to changes in the evaluation server. These changes are made throughout the progress of the competition. Files that have the same [md5]{} hash information are considered duplicates. Duplicates may be caused by submissions by the same team using different accounts. After eliminating duplicates, there are 97 unique submissions.
Prize mAP Energy Score
-------- -------- -------- --------
Winner 0.3898 1.540 0.2531
Second 0.1646 2.6697 0.0616
Third 0.0315 1.2348 0.0255
: Track 2 Winners[]{data-label="table:track2winners"}
Prize mAP Energy Score
-------- -------- -------- ---------
Winner 0.1832 0.4120 0.44462
Second 0.2119 0.5338 0.39701
Second 0.3753 0.9463 0.39664
Third 0.2235 1.5355 0.14556
: Track 3 Winners[]{data-label="table:track3winners"}
Tracks 2 and 3 Scores
---------------------
Table \[table:track2winners\] shows the scores of Track 2’s winners. The champion achieved very high accuracy and finished recognizing all 20,000 images in less than 6 minutes. The 2018 winner’s score is twice as high as the 2017 best score. Table \[table:track3winners\] shows the scores for Track 3’s winners. Two teams’ scores are very close and both teams win the second prize. Also, the top three scores outperform that of the 2017 champion. Two teams’ energy consumption is much lower than previous years’ champion’s. It should be noted that the champion in Track 3 has lower mAP with much lower energy consumption, thus a higher score.
2018 Winners’ Solutions
=======================
This section describes the winners’ solutions. Some winners decided not to share their solutions in this paper.
First Prize of Track 1
----------------------
The Qualcomm team wins the first prize of Track 1. Qualcomm provides edge AI platforms with Snapdragon (including the Google Pixel 2 phone used in this competition). Accurate and fast image recognition on edge devices requires several steps. First, a neural network model needs to be built and trained to identify and classify images Then, the model should run as accurate and fast as possible. Most neural networks are trained on floating-point models and usually need to be converted to fixed-point to efficiently run on edge devices.
Neural network architecture Input image resolution Data type Accuracy (%) Google Pixel-2 inference time (ms) Accuracy/ inference time (%/ms)
----------------------------- ------------------------ ----------- -------------- ------------------------------------ ---------------------------------
mobilenet v1 224x224 float32 70.2 81.5 0.86
mobilenet v1 224x224 uint8 65.5 68.0 0.96
mobilenet v1 128x128 uint8 64.1 28.0 2.28
mobilenet v2 150x150 uint8 64.4 36.6 1.75
mobilenet v2 132x132 uint8 62.7 31.8 1.97
mobilenet v2 130x130 uint8 59.9 31.2 1.91
For this competition, the model is based on MobileNet V2, but is modified to be quantization-friendly. Although Google’s MobileNet models successfully reduce parameter sizes and computation latency due to the use of separable convolution, directly quantizing a pre-trained MobileNet v2 model can cause large precision loss. The team analyzes and identifies the root cause of accuracy loss due to quantization in such separable convolution networks, and solves it properly without utilizing quantization-aware re-training. In separable convolutions, depthwise convolution is applied on each channel independently. However, the min and max values used for weights quantization are taken collectively from all channels. An outlier in one channel may cause a huge quantization loss for the whole model due to an enlarged data range. Without correlation crossing channels, depthwise convolution may be prone to produce all-zero values of weights in one channel. This is commonly observed in both MobileNet v1 and v2 models. All-zero values in one channel means small variance. A large “scale” value for that specific channel would be expected while applying batch normalization transform directly after depthwise convolution. This hurts the representation power of the whole model.
As a solution, the team proposes an effective quantization-friendly separable convolution architecture, where the nonlinear operations (both batch normalization and ReLU6) between depthwise and pointwise convolution layers are all removed, letting the network learn proper weights to handle the batch normalization transform directly. In addition, ReLU6 is replaced with ReLU in all pointwise convolution layers. From various experiments in MobileNet v1 and v2 models, this architecture shows a significant accuracy boost in the 8-bit quantized pipeline.\
Using fixed-point inferencing while preserving a high level of accuracy is the key to enable deep learning use cases on low power edge devices. The team identifies the industry-wide quantization issue, analyzes the root cause, and solves it on MobielNets efficiently. The quantized modified MobileNet\_v2\_1.0\_128 model can achieve 28 milliseconds per inference with high accuracy (64.7% on ImageNet validation dataset) on a single ARM CPU of Pixel 2. More details are described in the paper [@ShengFeng2018MobileNets].
Third Prize of Track 1
----------------------
The Expasoft team wins the third prize of Track 1. The team builds a neural network architecture that gives the high accuracy and inference time equal to 30 ms on Google Pixel 2. The team chooses two most promising architectures MobileNet and MobileNet-v2. Running MobileNet-v1\_224 with float-32 on Pixel-2 phone gives 70% accuracy and inference time of 81.5 ms. The team chooses two main directions to accelerate the neural network: quantization and reducing input image resolution. Both methods lead to accuracy reduction and the team finds trade-off for accuracy- speed relation. Tests of MobileNet and MobileNet-v2 architectures suggest quantizing the neural networks into uint8 data format. The team’s evaluation shows final score equal to 68.8% accuracy and 29ms inference time. Table \[table:track1thirdprize\] compares the accuracy and time for different configurations.
Quantization to uint8 reduces inference time from 81ms to 68ms but leads to significant accuracy drops. During standard quantization process in Tensorflow, it is required to start from full-precision trained model and learn quantization parameters (min and max values). Instead of joint training of neural network and tuning quantization parameters, the Expasoft team proposes another approach: tuning quantization parameters using Stochastic Gradient Descent approach with State Through Estimator [@DBLP:journals/corr/BengioLC13] of gradients of discrete functions (round, clip) without updating weights of the neural networks. Loss function for this process is L2 for embedding layers of full-precision and quantized networks. Mobilenet-v1 224 $\times$ 224 quantized such way shows 65.3% accuracy. The proposed method requires more detailed research but has two significant advantages: (1) The method doesn’t require labeled data for tuning quantization parameters. (2) Training is faster because there is no need to train neural network and only tuning is needed for quantizatizing parameters.
First Prize of Track 2
----------------------
The Seoul National University team wins the first prize in Track 2. Optimization is needed to balance speed and accuracy for existing deep learning algorithms originally designed for fast servers. The team discovers that optimization efficiency differs from network to network. For this reason, the team compares various object detection networks on the Jetson TX2 board to find the best solution.
The team compares five one-stage object detectors: YOLOv2, YOLOv3, TinyYOLO, SSD, and RetinaNet. Two-stage detectors such as Faster R-CNN are excluded because they are slightly better in accuracy than the one-stage detectors but much slower in speed. The baseline for each network is selected as the authors’ published form. For the comparison, the networks are improved with several software techniques briefly introduced as follows.
1. Pipelining: An object detection network can be split into three stages: input image pre-processing, a convolutional network body, and post-processing with output results of the body. Since the network body is typically run on the GPU and the others are on the CPU, they can run concurrently with the pipelining technique.
2. Tucker decomposition: It is one of the low-rank approximation techniques. As described in [@Tucker1966], 2-D Tucker decomposition is used in the network comparison.
3. Quantization: For the Jetson TX2 device, only 16-bit quantization is allowed and it is applied to the networks.
4. Merging batch normalization layer into weights: Since the batch normalization layer is a linear transformation, it can be integrated into the previous convolutional layer by adjusting the filter weights before running the inference.
5. Input size reduction: This is an easy solution to enhance the network throughput. It was also observed in experiments that the effect of this technique depends on the networks.\
By comparing the networks, the team finds that *YOLOv2* outperforms the other networks for the on-device object detection with the Jetson TX2. Table \[table:track2winner1\] shows how much the YOLOv2 network is enhanced by the series of the improvements. Since the total energy consumption is inversely proportional to the network speed, the score (mAP/Wh) can be estimated as mAP $\times$ speed. This optimized *YOLOv2* network is selected for the LPIRC Track 2 competition.
Description mAP(A) FPS(B) A x B Normalized score
-------------- -------- -------- ------- ------------------
Baseline 51.1 7.97 407 1.00
Pipelining 51.1 8.85 452 1.11
Tucker 50.2 15.1 758 1.86
Quantization 50.2 19.9 999 2.45
256 x 256 43.0 32.5 1640 4.03
Batch = 16 43.0 90.3 3880 9.54
: Different improvements and their scores by the Seoul National University team. The baseline is 416x416. FPS: frames per second.[]{data-label="table:track2winner1"}
YOLOv2 is tested on the Darknet framework in the experiments and it needs to be translated to Caffe2 framework for Track 2. The team implements custom Caffe2 operators to support Darknet-specific operators and optimization techniques such as pipelining and 16-bit quantization. Additionally, various batch sizes for the network are tested to find the best batch size for the submission. Through the steps illustrated above, the estimated score for the YOLOv2 has increased about 9.54 times compared with the baseline and this result surpassed the other object detection networks on the Jetson TX2.
Third Prize of Track 2
----------------------
The team’s members are from Tsinghua University, University of Science and Technology of China, and Nanjing University. The team evaluates several mainstream object detection neural models and picks the most efficient one. The selected model is then fine-tuned with sufficient dataset before being quantized into 16-bit float datatype in order to achieve better power-efficiency and time-efficiency in the exchange of minimal accuracy loss.
The team explores popular object detection neural architectures such as YOLO, RCNN and their variants. Among these architectures, YOLO V3 achieves the best balance between accuracy and computational cost. However, considering the difficulty of engineering implementation in a short amount of time, the team chooses faster RCNN as the base structure and then quantizes the parameters in order to shorten the inference time and reduce power consumption.
Faster RCNN consists of 2 sub-networks: the feature extraction network and the detection network. While the latter doesn’t seem to have much space for altering, there are many options for the former, such as VGG and MobileNet. The MobileNet family is known for their lower computational cost for achieving equivalent accuracy compared with traditional feature extraction architectures. Although MobileNets are reported with good classification results, the mAP for object detection seems low. The team decides to choose VGG-16.
The overview of the software optimization methods can be seen in Figure \[fig:track2third\]. The team reshapes the images on the CPU and then conducts the inference on the GPU. After the inference, the images are reshaped again to obtain the coordinates of the bounding boxes. The team applies three different techniques to accelerate the image recognition process: TensorRT-based inference, 16-bit Quantization, and CPU multithreading.
This implements the well-trained Fast-RCNN model used for image recognition on TensorRT. NVIDIA TensorRT is a high-performance deep learning inference optimizer and runtime system that delivers low latency and high-throughput for deep learning inference applications. The inference time becomes much shorter by using TensorRT and therefore saves energy. TensorRT provides quantization operation for the GPU inference engine. The computation latency is shortened because of less floating point arithmetic. In order not to reduce the mAP of the model, the team quantizes the weight to 16-bit while doing inference.
The method reshapes the images before and after the inference process. Since reshaping on the CPU is costly, the method uses CPU multithreading to accelerate the process. The method generates two threads on each CPU core and each thread processes one batch of data. TX 2 has 6 CPU cores and the method creates 12 threads. The inference on the GPU can only work in a single thread; thus the method takes the inference as a mutual process and different threads need to compete for the GPU. Applying multithreading method results in a 40% speeding up (the fps without multithreading is 3.17 while the fps with multithreading is 4.41), not as high as expected. The reason is that competition for the mutual recourse among the threads restricts acceleration. Different threads need to wait for the inference process because it needs mutual lock to avoid the hazard.
First and Second Prizes of Track 3
----------------------------------
The ETRI and KPST team wins both the first and the second prizes of Track 3. The performance of object detection can be evaluated using accuracy, speed, and memory. The accuracy is a common measurement and it has been widely used in comparing different object detection architectures. However, its performance is dependent upon speed and memory as well as accuracy for the given environments and applications. In this competition, the performance is measured using accuracy (mAP) and energy consumption (WH). The accuracy is influenced by the detection architectures, an input resolution and a base network for feature extraction. The more complex structures with a high-resolution image may achieve higher accuracy at higher energy consumption. To obtain a high score, it is necessary to balance accuracy and energy consumption.
The team examines the score function and finds the most important factor is energy consumption. At the same accuracy, the score is higher at lower energy consumption. Moreover, the accuracy is weighted by the processed images ratio within 10 minutes. This means that the detection architecture should be fast and light and the trade-off between accuracy and energy consumption. Accordingly, single stage detectors such as SSD [@10.1007/978-3-319-46448-0_2] and YOLO [@8100173] are considered in the pipeline. The team selects SSD as the detection framework due to its simplicity and stable accuracy among feature extractors [@8099834]. To obtain the best score, the team performs three optimization steps: (1) detection structure optimization, (2) feature extraction optimization, (3) system and parameters optimization.
For detection structure optimization, the original SSD is based on VGG and its performance is well balanced in accuracy and speed. Its speed and memory can be improved for low-power and real-time environments. To speed up, the team proposes efficient SSD (eSSD) by adding additional feature extraction layers and prediction layers in SSD. Table \[table:track31st2nd\] shows the comparison of SSD, SSDLite [@Sandler2018CVPR], and eSSD. In SSD, additional feature extraction is computed by 1x1 conv and 3x3 conv with stride 2 and prediction uses 3x3 conv. The SSDLite replaces all 3x3 conv with depthwise conv and 1x1 conv. The eSSD extracts additional features with depthwise conv and 1x1 conv and predicts classification and bounding box of an object with 1x1 conv. This reduces memory and computational resources.
Type Additional feature extraction Prediction
--------- --------------------------------------- -------------------------
SSD 1x1 conv / 3x3 conv-s2 3x3 conv
SSDLite 1x1 conv / 3x3 conv(dw)-s2 / 1x1 conv 3x3 conv(dw) / 1x1 conv
eSSD 3x3 conv(dw)-s2 / 1x1 conv 1x1 conv
: The comparison of feature extraction and prediction in SSD, SSDLite, and eSSD.[]{data-label="table:track31st2nd"}
Table \[table:track31st2ndbase\] shows accuracy, speed, and memory comparison of SSD variants in VOC0712 dataset. In this experiment, a simple prediction layer such as 3 by 3 or 1 by 1 is applied and an inference time (forward time) is measured in a single Titan XP (Pascal) GPU. SSDLite is more efficient in memory usage than eSSD, but eSSD shows better performance in speed than SSDLite.
Base Network (300x300) Feature extraction / Prediction mAP Speed Memory
------------------------ --------------------------------- ------ ------- --------
MobileNetV1 SSD / 3x3 68.6 8.05 34.8
MobileNetV1 SSD / 1x1 67.8 6.19 23.1
MobileNetV1 SSDLite /1x1 67.8 5.91 16.8
MobileNetV1 eSSD / 1x1 67.9 5.61 17.4
MobileNetV1 (C=0.75) eSSD / 1x1 65.8 5.20 11.1
MobileNetV1 (C=0.75) eSSD(L=5) /1x1 65.8 4.62 10.9
VGG SSD / 3x3 77.7 12.43 105.2
: The accuracy (mAP), speed (ms), and memory (MB) for different feature extraction and prediction architectures in VOC 0712 train and VOC 2007 test dataset.[]{data-label="table:track31st2ndbase"}
For feature extraction optimization, the base model of feature extractor, MobilNetV1 [@Howard2017arxiv] is used and feature extraction layers of MobileNetV1 is optimized in eSSD. To improve memory usage and computational complexity, the team uses 75% weight filter channels (C=0.75). Although this drops accuracy, energy consumption is greatly reduced. The team uses five additional layers and modified anchors for a low resolution image. It generates a small number of candidate boxes and improves detection speed. After detection structures are optimized, the team also modifies MobileNetV1 by applying early down-sampling and weight filter channel reduction in earlier layers and trained the base model (MobileNetV1+) from scratch. All models are trained with ImageNet dataset and Caffe framework. In training, batch normalization is used and trained weights are merged into final model as introduced in [@Fu2017arxiv].
After feature extraction optimization, scores of proposed models are measured. Furthermore, the team compares these scores with scores of YOLOv3-tiny [@redmon2018yolov3]. Comparison results are presented in Table \[table:track3\_yoloCompare\]. Scores of proposed models are located in the range between 0.18 and 0.21. In contrast, scores of models from YOLOv3-tiny have a peak in the range of input image resolution between 256 and 224, and the score drops rapidly when the resolution decreases to 160. Scores of all three proposed models are higher than the best score from YOLOv3-tiny. In detail, comparing the proposed model of input resolution 224 and YOLOv3-tiny model of input resolution of 256, both have similar accuracy (25.7 vs 25.5) but the proposed model consumes less power about 17.5%.
Models mAP sec Energy Score
-------------------------------------------- ------ ----- -------- --------
eSSD-MobileNetV1+ (C=0.75, I=224, th=0.01) 25.8 476 1.244 0.2074
eSSD-MobileNetV1+ (C=0.75, I=192, th=0.01) 23.3 445 1.170 0.1991
eSSD-MobileNetV1+ (C=0.75, I=160, th=0.01) 19.3 381 1.049 0.1840
Yolov3-tiny (I=256, th=0.01) 25.5 428 1.508 0.1691
Yolov3-tiny (I=224, th=0.01) 22.5 411 1.313 0.1714
Yolov3-tiny (I=160, th=0.01) 15.7 330 1.082 0.1451
: Score of proposed models and YOLOv3-tiny. *C, I, th* represent channel reduction, input resolution, Confidence threshold, respectively. The time is measured for processing 20,000 images. []{data-label="table:track3_yoloCompare"}
For system and parameters optimization, after training the models, the system is set up and the trained models are ported to NVIDIA TX2. In object detection, multiple duplicate results are obtained and Non-Maximal Suppression (NMS) with thresholding is important. The team tunes the NMS process between CPU and GPU to reduce computational complexity and adjust the confidence threshold to decrease result file size for the network bandwidth. Then batch size modification and queuing are applied to maximize speed in detection and to increase the efficiency of network bandwidth. After tuning the speed, to minimize energy consumption, the team uses the low power mode (max-Q mode) in NVIDIA Jetson TX2.
Scores of the proposed models after system and parameters optimization are described in Table \[table:track3\_batchExp\]. The team achieves 1.5x–2x higher score after optimization, but the order is reversed, i.e. after optimization, the model of input resolution 160 has the highest score which has the lowest score before optimization. Batch processing increases the throughput of the detection process. Accordingly, it improves the score. However, batching process with larger input resolution needs more memory access so it decreases the score.
Models mAP sec Energy Score
------------------------------------------------------- ------ ----- -------- --------
eSSD-MobileNetV1+ (C=0.75, 224, th=0.05, batch=64) 24.7 448 0.780 0.3167
eSSD-MobileNetV1+ (C=0.75, 192, th=0.05, batch=64) 22.2 415 0.617 0.3598
eSSD-MobileNetV1+ (C=0.75, I=160, th=0.05, batch= 96) 18.5 305 0.512 0.3613
: Score improvement after system and parameters optimization. Batch represents batch size. []{data-label="table:track3_batchExp"}
Table \[table:track3final\] shows final model specifications at on-site challenge. Scores of both models are further increased by 10–23%. The result shows similar accuracy with in-house measurement described in Table \[table:track3\_batchExp\], but the power consumption is significantly decreased. The team hypothesizes the difference is caused by the network condition and test dataset.
Models mAP sec Energy Score
------------------------------------------------------- -------- ----- -------- --------------
eSSD-MobileNetV1+ (C=0.75, I=160, th=0.05, batch= 96) 18.318 300 0.4119 0.4446 (1st)
eSSD-MobileNetV1+ (C=0.75, 192, th=0.05, batch=64) 21.192 362 0.5338 0.3970 (2nd)
: Proposed (eSSD-MobileNetV1+) Model specifications of on-site challenge . []{data-label="table:track3final"}
Second Prize of Track 3
-----------------------
The Seoul National University team shares the second prize of Track 3 because the scores are very close. The team chooses the TX2 board as the hardware platform because of the GPU-based deep learning application. The object detection network is an improved *YOLOv2* network. The *C-GOOD framework* [@Kang:2018:CCG:3240765.3240786] is used. Derived from Darknet, this framework helps explore a design space of deep learning algorithms and software optimization options. The team reduces the energy consumption by managing the operating frequencies of the processing elements. The team discovers the optimal combination of CPU and GPU frequencies.
Low-Power Vision Solutions in Literature
----------------------------------------
MobileNet [@Sandler2018CVPR] and SqueezeNet [@Iandola2016] are network micro-architectures that set the benchmark accuracy with low energy consumption. Without any optimization, MobileNet requires $81.5$ ms to classify a single image, higher than the $30$ ms wall-time requirement for Track 1. However, speeding up the network by quantizing the parameters of MobileNet leads to large precision loss. In order to further the gains of the micro-architecture solution, the Track 1 winning team develop a method to use quantization on MobileNet that can perform inference in $28$ ms, without much loss of accuracy. CondenseNet [@Huang2017] and ShuffleNet [@Zhang2018] are two neural network architectures that use group convolutions to reuse feature maps in order to improve parameter efficiency. These architectures can classify a single image in about $52.8$ ms. Faster inference and lower memory requirements are obtained with Binarized Neural Networks [@XNOR][@Cour2016], where each parameter is represented with a single bit. The accuracy of such networks is lower, making them unsuitable for some applications. LightNN [@Ding2017] and CompactNet [@Goel2018] bridge the gap between Binarized Neural Networks and architectures like CondenseNet, by using a variable quantization scheme.
Some studies design hardware for neural networks. For example, Lee et al. [@8481682] design an accelerator with precision using different numbers of bits. The system can achieve 35.6, 14.5, and 7.6 frames per second for running VGG-16 when using 1, 4, and 8 bits. It achieves 3.08 TOPS/W at 16-bit precision and 50.6 TOPS/W at 1-bit precision. Abtahi et al. [@8392465] compare specially-designed hardware for CNN with Jetson TX1 and Cortex A53, Zynq FPGA and demonstrate higher throughput and better energy efficiency. Gholami et al. [@8575377] reduce the sizes of neural networks for mobile hardware with comparable or better accuracy with prior work. The smaller networks are faster and consume less energy. Even though these studies use ImageNet, they do not have the same settings as LPIRC. As a result, it is not possible making direct comparison with the winners’ solutions.
Single stage object detection techniques, like SSD, Yolo v2, Yolo v3, and TinyYolo have an inherent advantage in terms of energy consumption over multiple stage techniques, like Faster RCNN. However, when using multiple stages it may be possible to obtain significantly higher accuracy. The winners of Track 2 improve the single stage detector by using quantization and pipelining. Their technique outperforms all previous solutions on the score. Some other teams also use efficient network architectures in the two stage Faster RCNN object detector to improve its execution speed and memory requirement.
Industry Support for Low-Power Computer Vision
==============================================
Google Tensorflow Models
------------------------
Google provided technical support to Track 1 in two aspects. First, Google open-sources TfLite, a mobile-centric inference engine for Tensorflow models. TfLite encapsulates implementations of standard operations such as convolutions and depthwise-convolutions that shield the contestants from such concerns. This allows more people to participate in the competition. Second, Google provides the mobile benchmarking system and it allows repeatable measurement of the performance metric. The system comes in two flavors: a stand-alone App that the contestants can run on their local phones, and a cloud-based service where the submissions will be automatically dispatched to a Pixel 2 XL phone and benchmarked using the standardized environment. The App and a validation tool are provided to facilitate rapid development, exploring novel models and catching runtime bugs. Once the model is sufficiently polished, it will be available as cloud service for refinements and verification.
Facebook AI Performance Evaluation Platform
-------------------------------------------
Machine learning is a rapidly evolving area: new and existing framework enhancements, new hardware solutions, new software backends, and new models. With so many moving parts, it is difficult to quickly evaluate the performance of a machine learning model. However, such evaluation is important in guiding resource allocation in (1) development of the frameworks, (2) optimization of the software backends, (3) selection of the hardware solutions, and (4) iteration of the machine learning models. Because of this need, Facebook has developed an AI performance evaluation platform (FAI-PEP) to provide a unified and standardized AI benchmarking methodology. FAI-PEP supports Caffe2 and TFLite frameworks, the iOS, Android, Linux, and Windows operating systems. The FAI-PEP is modularly composed so that new frameworks and backends can be added easily. The built-in metrics collected by FAI-PEP are: latency, accuracy, power, and energy. It also supports reporting arbitrary metrics that the user desires to collect. With FAI-PEP, the benchmark runtime condition can be specified precisely, and the ML workload can be benchmarked repeatedly with low variance.
Future of Low Power Computer Vision
===================================
In 2018 CVPR, LPIRC invites three speakers from Google and Facebook sharing their experience on building energy-efficient computer vision. More than 100 people attended the workshop. The panel after the speeches answers many attendees’ questions. The high participation suggests that there is strong interest, in both academia and industry, to create datasets and common platforms (both hardware and software) for benchmarking different solutions. Readers interested in contributing to future low-power computer vision are encouraged to contact the LPIRC organizers for further discussion. Over four years, LPIRC has witnessed impressive improvement of the champions’ scores by 24 times. The tasks in LPIRC–detecting object in images– marks a major triumph in the vision technologies. However, many more challenges are still yet to be conquered.
Future Vision Challenges: Action, Intention, Prediction, Emotion, Implication
-----------------------------------------------------------------------------
Visual data (image or video) has rich information that is often difficult to express by words. Here, the authors suggest possible topics for future competitions in low-power computer vision. More specifically, we suggest research directions in computer vision beyond processing pixels: action, intention, prediction, emotion, and implication.
Consider Figure \[fig:oscarsselfie\] as an example. This single image can provide rich information beyond detecting faces and recognizing names, even though both are important tasks in computer vision. The action here can be described as [*laughing*]{}, [*gathering*]{}, [*sharing*]{}. What is their intention? Perhaps [*celebrating*]{}, [*enjoying*]{}. This photograph was considered one of the most retweeted postings. Would it be possible to [*predict*]{} that this image would be popular becasue of the celebrities? The emotion expressed in this photograph can be described as [*happy*]{}. Answering these questions would need to draw deeper knowledge beyond the pixels. Even if a computer vision program can recognize the faces and know these are actors and actresses, would it be capable of [*predicting*]{} that this image would be shared widely?
Next, consider the image shown in Figure \[fig:1999soccer\]. A vision task may be able to detect people but a more interesting (and challenging) task is to describe the emotion: [*happiness*]{}, [*joy*]{}, [*relieved*]{}, [*excited*]{}, and [*victorious*]{}. This can be considerably difficult because the face of Chastain alone, shown in Figure \[fig:1999soccer\] (b), may be insufficient to decide the emotion. Adding the running team members at the background might be helpful inferring the emotion. An even more difficult question is the [*implication*]{}: this team won and the other team lost. To obtain this information, it would be necessary to draw the knowledge about the [*context*]{}: it is a soccer game. This may be possible from time and location when this image was taken. Many digital cameras (in particular smartphones) are equipped with GPS (global positioning systems) now and visual data can include the time and location. Such information can be used to look up other information, such as the event (e.g., a soccer game) and the participants (e.g., the teams).
Creating a competition for evaluating computer vision’s capability for understanding action, intention, prediction, emotion, or implication can be extremely difficult for three reasons. First, the “correct” answers can be subjective or even dependent in culture, time, location, or other conditions. As a result, it would not be easy creating datasets for training or evaluation. Second, even if the correct answers can be determined and agreed upon, labeling large amounts of data would require significant amounts of efforts. Third, learning-based vision technologies need training examples; thus, vision technologies may be inadequate handling [*exceptional*]{} situations that rarely or have never occurred. In some cases (such as a natural disaster), data may have never been seen before. New thinking would be need to [*expedite*]{} training of unseen data for rescue operations.
Low-Power Systems for Computer Vision
-------------------------------------
The embedded computing systems for computer vision can be divided into three categories: (1) *General-purpose systems:* CPU and GPU are popular for computer vision tasks on embedded and mobile systems. Many low-power versions of CPU and GPU have been designed (e.g., NVIDIA’s TX2) for this purpose. These designs often feature conventional low-power functions such as dynamic voltage and frequency scaling. Recently, they start to support functions in machine learning such as low-precision arithmetics. (2) *Reconfigurable systems:* FPGA is also a popular computing platform for computer vision because it can be dynamically reconfigured to support different structures and precision [@DBLP:journals/corr/abs-1809-00110]. FPGA has been deployed in data centers (e.g, Microsoft Azure) to support cloud tasks. Comprehensive software-hardware co-design flows have been also developed by leading FPGA companies to support machine learning tasks. (3) *Application-specific integrated circuit (ASIC):* ASIC offers the highest possible efficiency among all platforms but it needs the highest development cost and the longest design time. If the volume is large, ASIC can become cost attractive. Huawei recently proposes Kirin 980 chipset to support AI tasks on their phones. Other examples include Google’s Tensor Processing Unit (TPU) and Intel Nervana Neural Network Processor. Processors optimized for machine learning have been introduced by several vendors recently and these new systems may appear in future LPIRC.
Among the above three technologies, GPU offers the most user friendly interface and comprehensive commercial support; FPGA offers the most flexible hardware reconfigurablity but requires complex hardware programming; ASIC offers the highest computing efficiency but also the longest development cycle. A major challenge in future low-power vision processing system is the co-design between software and hardware, e.g., how to prune and quantify the network so that it can be efficiently mapped onto the hardware under certain constraints such as available memory capacity etc.
Conclusion
==========
This paper explains the three tracks of the 2018 Low-Power Image Recognition Challenge. The winners describe the key improvements in their solutions. As computer vision is widely used in many battery-powered systems (such as drones and mobile phones), the need for low-power computer vision will become increasingly important. The initial success of the new tracks in 2018 also suggests the advantages of making focused advances on specific components of the vision system, as well as lowering the entry barrier to be inclusive of the general vision and machine learning communities.
Acknowledgments {#acknowledgments .unnumbered}
===============
IEEE Rebooting Computing is the founding sponsor of LPIRC. The sponsors since 2015 include Google, Facebook, Nvidia, Xilinx, Mediatek, IEEE Circuits and Systems Society, IEEE Council on Electronic Design Automation, IEEE GreenICT, IEEE Council on Superconductivity. ETRI and KPST’s work is supported by ETRI grant funded by the Korean government (2018-0-00198, Object information extraction and real-to-virtual mapping based AR technology).
Authors’ Affiliations {#authors-affiliations .unnumbered}
=====================
The authors are ordered alphabetically: Sergei Alyamkin (Expasoft), Matthew Ardi (Purdue), Alexander C. Berg (University of North Carolina at Chapel Hill), Achille Brighton (Google), Bo Chen (Google), Yiran Chen (Duke), Hsin-Pai Cheng (Duke), Zichen Fan (Tsinghua University), Chen Feng (Qualcomm), Bo Fu (Purdue, Google), Kent Gauen (Purdue), Abhinav Goel (Purdue), Alexander Goncharenko (Expasoft), Xuyang Guo (Tsinghua University), Soonhoi Ha (Seoul National University), Andrew Howard (Google), Xiao Hu (Purdue), Yuanjun Huang (University of Science and Technology of China), Donghyun Kang (Seoul National University), Jaeyoun Kim (Google), Jong Gook Ko (ETRI), Alexander Kondratyev (Expasoft), Junhyeok Lee (KPST), Seungjae Lee (ETRI), Suwoong Lee (ETRI), Zhiyu Liang (Qualcomm), Zichao Li (Nanjing University), Xin Liu (Duke), Juzheng Liu (Tsinghua University), Yang Lu (Facebook), Yung-Hsiang Lu (Purdue), Deeptanshu Malik (Purdue), Hong Hanh Nguyen (KPST), Eunbyung Park (University of North Carolina at Chapel Hill), Denis Repin (Expasoft), Tao Sheng (Qualcomm), Liang Shen (Qualcomm), Fei Sun (Facebook), David Svitov (Expasoft), George K Thiruvathukal (Loyola University Chicago and Argonne National Laboratory), Jingchi Zhang (Duke), Baiwu Zhang (Qualcomm), Xiaopeng Zhang (Qualcomm), Shaojie Zhuo (Qualcomm).
[^1]: Corresponding Author: Yung-Hsiang Lu, Email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'M. G. Campo,[@inst1][^1]'
title: 'Structural and dynamic properties of SPC/E water'
---
Introduction
============
Water is the subject of numerous studies due to its biological significance and its universal presence \[1–3\]. The thermodynamic behavior of water presents important differences compared with those of the other substances, and many of the characteristics of such behavior are often attributed to the existence of hydrogen bonds between water molecules. Scientists have found that the water structure produced by the hydrogen bonds is peculiar as compared to that of other liquids. Then, the advances in the knowledge of hydrogen bond behavior are crucial to understanding water properties.
The method of molecular dynamics (MD) allows to analyze the structure and dynamics of water at the microscopic level and hence to complement experimental techniques in which these properties can be interpreted only in a qualitative way (infra-red absorption and Raman scattering [@ref04], depolarized light scattering [@ref05; @ref06], neutron scattering [@ref07], femtosecond spectroscopy \[8–11\] and other techniques \[12–14\].
Among the usual methods to study the short range order in MD simulations of water are the calculus of radial distribution functions, hydrogen bond distributions and order parameters. The orientational order parameter *Q* measures the tendency of the system to adopt a tetrahedral configuration considering the water oxygen atom as vertices of a tetrahedron, whereas the translational order parameter $\tau $ quantifies the deviation of the pair correlation function from the uniform value of unity seen in an ideal gas [@ref15; @ref16]. The order parameters are used to construct an order map, in which different states of a system are mapped onto a plane $\tau $-*Q*. The order parameters are, in general, independent, but they are linearly correlated in the region in which the water behaves anomalously [@ref16a].
The dynamics of water can by characterized by the bond lifetime, $\tau _{HB}$, associated to the process of rupturing and forming of hydrogen bonds between water molecules which occurs at very short time scale \[9, 18, 19, 21–23\]. $\tau _{HB}$ is obtained in MD using the history-dependent bond correlation function $P(t)$, which represents the probability that an hydrogen bond formed at time $t=0$ remained continuously unbroken and breaks at time $t$ [@ref23; @ref24].
Also, the dynamics of water can be studied by analyzing the mean-square displacement time series $M(t)$. In addition to the diffusion coefficient calculation at long times in which $M(t)\propto t$, in the supercooled region of temperatures and at intermediate times $M(t)\propto t^{\alpha}$ ($0<\alpha <1$). This behavior of $M(t)$ is associated to the subdiffusive movement of the water molecules, caused by the *caging* effect in which a water molecule is temporarily trapped by its neighbors and then moves in short bursts due to nearby cooperative motion. A time $t^{\ast}$ characterizes this *caging* effect (see Sec. II for more details) [@ref24a; @ref24b].
In a previous work, we found a $q$-exponential behavior in $P(t)$, in which $q$ increases with $T^{-1}$ approximately below 300 K. $q(T)$ is also correlated with the probability of occurrence of four hydrogen bonds, and the subdiffusive motion of the water molecules [@ref25].
The relationship between dynamics and structural properties of water has not been clearly established to date. In this paper, I explore whether the effect that temperature has on the water dynamics reflects a more general connection between the structure and the dynamics of this substance.
Theory and method
=================
I have performed molecular dynamic simulations of SPC/E water model using the GROMACS package [@ref26; @ref27], simulating fourteen similar systems of 1185 molecules at 1 bar of pressure in a range of temperatures from 213 K to 360 K. I initialized the system at 360 K using an aleatory configuration of water molecules, assigning velocities to the molecules according to a Boltzmann’s distribution at this temperature. For stabilization, I applied Berendsen’s thermal and hydrostatic baths at the same temperature and 1 bar of pressure [@ref28]. Then, I ran an additional MD obtaining an isobaric-isothermal ensemble. I obtained the other systems in a similar procedure, but using as initial configuration that of the system of the preceding higher temperature and cooling it at the slow rate of 30 K ns$^{-1}$ [@ref16a]. Stabilization and sampling periods for the systems at different temperatures are indicated in Table \[tab01\]. Simulation and sampling time steps were 2 fs and 10 fs, respectively. The sampling time step was shorter than the typical time during which a hydrogen bond can be destroyed by libration movements.
Temp. range (K) $t_{est}$ (ns) $t_{MD}$ (ns)
----------------- ---------------- ---------------
213 - 243 20.0 10.0
253 - 273 16.0 10.0
283 - 360 16.0 8.0
: Details of the simulation procedure. Duration of the stabilization period ($t_{est}$) and the MD sampling ($t_{MD}$) in the different ranges of temperatures[]{data-label="tab01"}
![Hydrogen bonds distribution functions $f(n)$ ($n=0,...,5$) versus $T$. The zones A, B and C correspond to ranges of temperatures in which occur different relationships between $f(4), f(3)$ and $f(2)$. Note the reciprocal scale for the temperatures. See the text for details.[]{data-label="figure01"}](hbonds){width="7.4cm"}
I calculated the hydrogen bond distribution functions $f(n)$ ($n=0,1,...,5$), which is the probability of occurrence of $n$ hydrogen bonds by molecule, considering a geometric definition of hydrogen bond [@refhb01]. As parameters for this calculation, I used a maximum distance between oxygen atoms of 3.5 $\mathring{A}$ and a minimum angle between the atoms O$_{donor}$–H–O$_{acceptor}$ of 145$^{\circ}$.
The radial distribution function (RDF) is a standard tool used in experiments, theories, and simulations to characterize the structure of condensed matter. Using RDFs, I obtained the average number, $N$, of water molecules in the first hydration layer (the hydration number)
$$N= 4 \pi \rho \int_{0} ^{r_{min}} g(r) r^{2} dr
\label{ec01}$$
where $\rho$ is the number density.
![Oxygen-oxygen radial distribution functions for the systems at 213 K (continuous line), 293 K (dashed line), and 360 K (dotted line). Inset: The hydration number $N$ vs. $T^{5}$.[]{data-label="figure02"}](gr){width="7.4cm"}
The translational order parameter, $\tau$, is defined in Ref. [@ref16] as
$$\tau\equiv\int_{0} ^{S_{c}} |g(s)-1|ds
\label{ec02}$$
where the dimensionless variable $s\equiv rn^{1/3}$ is the radial distance r scaled by the mean intermolecular distance $n^{1/3}$, and $S_{c}$ corresponds to half of the simulation box size.
The orientational order parameter $Q$ is defined as [@ref15] $$Q= \left \langle 1- \frac{3}{8} \sum_{i=1}^N \sum_{j=1}^4 \sum_{l=j+1}^4 \left [ cos\theta_{jik} + \frac{1}{3} \right ]^2 \right \rangle
\label{ec03}$$where $\theta_{jik}$ is the angle formed by the atoms O$_{j}$–O$_{i}$–O$_{k}$. Here, O$_{i}$ is the reference oxygen atom, and O$_{j}$ and O$_{k}$ are two of its four nearest neighbors. $Q$=1 in an ideal configuration in which the oxygen atoms would be located in the vertices of a tetrahedron.
I obtained the bond correlation function $P(t)$ from the simulations by building a histogram of the hydrogen bonds lifetimes for each configuration. Then, I fitted this function with a Tsallis distribution of the form
$$\exp _{q}(t)=\left[ 1+\left( 1-q\right) t\right] ^{1/\left( 1-q\right) }
\label{ec04}$$
being $t$ the hydrogen bond lifetime and $q$ the nonextensivity parameter [@ref25; @ref29]. If $q = 1$, Eq. (\[ec04\]) reduces to an exponential, whereas if $q > 1$, $P(t)$ decays more slowly than an exponential. This last behavior occurs when long lasting hydrogen bonds increase their frequency of occurrence.
The subdiffusive movement of water occurs when the displacement of the molecules obeys a non-Gaussian statistics. This behavior is characterized by $t^{*}$, the time in which the non-Gaussian parameter $\alpha_{2}(t)$ reaches a maximum \[see Eq. (\[ec05\])\]. Then, $t^{*}$ is the parameter associated to the average time during which a water molecule is trapped by its environment (*caging* effect), and this prevents it from reaching the diffusive state [@ref24a; @ref24b].
$$\alpha_{2}(t)=\frac{3 \langle r^{4}(t) \rangle}{5 \langle r^{2}(t) \rangle} -1
\label{ec05}$$
Results and discussion
======================
Three zones or ranges of temperatures can be distinguished in the graph of the hydrogen bond distributions $f(n)$ vs. $T$ (see Fig. \[figure01\]). Zone A ($T$ $>$ 350 K) in which $
f(3)>f(2)>f(4)$, zone B (293 K $> T >$ 350 K) in which $f(3)>f(4)>f(2)$, and zone C ($T$ $<$ 293 K) in which $f(4)>f(3)>f(2)$. These results indicate a predominant structure of three and two hydrogen bonds (3HB-2HB) in zone A, 3HB-4HB in zone B, and 4HB-3HB in zone C, respectively. $f(4)\propto T^{-1}$ in all ranges of temperatures, showing that the tetrahedral structure of water decreases with the increase of this variable.$f(3)$ increases with $T$ up to 293 K, and then remains approximately constant ($\sim 0.4$) up to 360 K. $f(2)$ also increases with ${T}$ in all range of temperatures, but only overcomes $f(4)$ at $T >$ 350 K.
![ Position of the first minimum of the oxygen-oxygen radial distribution function vs $T^{4}$, associated to the size of the first hydration layer.[]{data-label="rmin"}](rmin){width="7.4cm"}
Fig. \[figure02\] shows the oxygen-oxygen RDFs corresponding to the systems at 213 K, 293 K and 360 K. When the temperature decreases, the minimum and maximum tend to be more defined. This being associated with an increasing order in the system. The position of the first minimum moves closer to the origin decreasing the size of the first hydration layer ($\propto T^{4}$ see Fig. \[rmin\]). Both facts can be associated to the decrease of the hydration number from $N \sim$ 5 to $N \sim$ 4 (see inset, Fig. \[figure02\]).
![Order map with the values of the order parameters corresponding to the simulated systems. Note the change in the slope of the line at $T\sim$ 273 K[]{data-label="figure03"}](ordermap){width="7.4cm"}
The simultaneous behavior of *Q* and $\tau$ is shown in the order map of Fig. \[figure03\], in which the location of the values corresponding to 293 K are indicated by an arrow. The order parameters present similar behaviors with the temperature. Upon cooling, these parameters are linearly correlated and move in the order map along a line of increasing values, up to reaching maximum values at 213 K. The slope of the line increases a little for $T >$ 293 K, indicating that $\tau$ has a response to the increase of $T$ slightly higher than *Q* in this range of temperatures. The positive values of the slopes indicate an increasing order of the system when the temperature decrease.
The $f(n)$ functions allow to obtain a more detailed picture of the structural orientational changes at shorter ranges between water molecules than the orientational order parameter. While a small change at 293 K occurs in the order map, the structures of two, three and four hydrogen bonds are alternated in importance when the temperature changes. The ability of $f(n)$ to more reliably describe the structure of the water occurs because the calculation of the hydrogen bond distributions includes the location of the hydrogen atoms, whereas *Q* only quantifies the changes in the average angle between neighbor oxygen atoms. Although the behavior of $f(4)$ and *Q* are correlated (see Fig. \[figure03b\]), $f(4)$ shows a greater response to the temperature than *Q*, indicating that the main change in the tetrahedral structure with the decrease of the temperature occurs mainly in the orientation of the bonds between water molecules. The approximately linear correlation between both variables also indicates a similar dependence with the temperature ($\propto T^{-1}$).
![ *Q* vs $f(4)$. The change in $f(4)$ is higher than that of *Q* in the range of temperatures studied. See the text for details.[]{data-label="figure03b"}](qvsf4){width="7.4cm"}
![(a) Semilog plot of $t^{*}$ vs. *Q*. (b) $q$ vs. *Q*. See the text for details.[]{data-label="figure05"}](tesyqvsq){width="7.4cm"}
Figure \[figure05\] shows the behavior of the dynamical parameters $t^{*}$ and $q$ with *Q*. The characteristic time $t^{*}$ has an exponential response to $Q \geq 0.58$ (T $\leq$ 360 K), but the slope of the semilog plot of $t^{*}$ vs. *Q* increases significantly for $Q \geq 0.67$. A similar change occurs for $Q \geq 0.67$ in the linear correlation between $q$ and $Q$. Then, the values $Q \approx 0.67$ and $\tau \approx 1.1$ of the order map can be associated to changes in the dynamics of the system. The transition of $q\approx 1$ to $q>1$ indicates the increase of the probability of two water molecules remaining bonded by a hydrogen bond during an unusual long time, whereas the increase of $t^{*}$ is associated to the increase of the time during which the molecules remain in a subdiffusive regime.
However, only the analysis of the $f(n)$ functions reveals the structural modification that explains the structural and dynamic changes in the system. The changes in the increase of the order map, $t^{*}(Q)$ and $q(Q)$ occur below 293 K, in the range of temperatures in which prevail a structure of four hydrogen bonds in the system.
Conclusions
===========
The molecular dynamic method allows to study the structure and dynamics of the SPC/E model of water in the range of temperatures from 213 K to 360 K.
Lowering the temperature of the system from 360 K to 213 K, the number of water molecules in the first hydration layer decreases from $N \sim 5$ to $N \sim 4$, along with a decrease in size. The increase of the tetrahedral structure of the system is also characterized by a growth of the percentage of occurrence of four hydrogen bonds and the orientational order parameter $Q$. However, only the analysis of the behavior of the hydrogen bond distribution allows to deduce that, when a tetrahedral structure associated to the percentage of four hydrogen bonds predominates, the behavior of the dynamical variables $P(t)$ and $t^{*}$ show the occurrence of long lasting hydrogen bonds and *caging* effect between the molecules of the system.
I am grateful for the financial support by PICTO UNLPAM 2005 30807 and Facultad de Ciencias Exactas y Naturales (UNLPam).
[33]{} D Eisenberg, W Kauzmann, *The structure and properties of water*, Clarendon Press, Oxford (1969). F H Stillinger, *Water revisited*, Science **209**, 451 (1980). O Mishima, H E Stanley, *The relationship between liquid, supercooled and glassy mater*, Nature **396**, 329 (1998). E W Castner, Y J Chang, Y C Chu, G E Walrafen, *The intermolecular dynamics of liquid water*, J. Chem. Phys. **102**, 653 (1995). C J Montrose, J A Bucaro, J Marshall-Coakley, T A Litovitz, *Depolarized Rayleigh scattering and hydrogen bonding in liquid water*, J. Chem. Phys. **60**, 5025 (1974). W Danninger, G Zundel, *Intense depolarized Rayleigh scattering in Raman spectra of acids caused by large proton polarizabilities of hydrogen bonds*, J. Chem. Phys. **74**, 2779 (1981). J Yeixeira, S H Chen, M-C Bellissent-Funel, *Molecular dynamics of liquid water probed by neutron scattering*, J. Mol. Liq., **48**, 111 (1991). R Laenen, C Rauscher, A Laubereau, *Dynamics of local substructures in water observed by ultrafast infrared hole burning*, Phys. Rev. Lett. **80**, 2622 (1998). S Woutersen, U Emmerichs, H J Bakker, *Femtosecond mid-IR pump-probe spectroscopy of liquid water: Evidence for a two-component structure*, Science **278**, 658 (1997). H K Nienhuys, S Woutersen, R A van Santen, H J Bakker, *Mechanism for vibrational relaxation in water investigated by femtosecond infrared spectroscopy*, J. Chem. Phys. **111**, 1494, (1999). G M Gale, G Gallot, F Hache, N Lascoux, S Bratos, J C Leickman, *Femtosecond dynamics of hydrogen bonds in liquid water a real time study*, Phys. Rev. Lett. **82**, 1068, (1999). A H Narten, M D Danford, H A Levy, *X-ray diffraction study of liquid water in the temperature range 4 – 200* , Faraday Discuss. **43**, 97 (1967). A K Soper, F Bruni, M A Ricci, *Site-site pair correlation functions of water from 25 to 400 : Revised analysis of new and old diffraction data*, J. Chem. Phys. **106**, 247 (1997). K Modig, B G Pfrommer, B Halle, *Temperature-dependent hydrogen-bond geometry in liquid water*, Phys. Rev. Lett. **90**, 075502 (2003). H Tanaka, *Simple physical explanation of the unusual thermodynamic behavior of liquid water*, Phys. Rev. Lett. **80**, 5750 (1998). J R Errington, P G Debenedetti, *Relationship between structural order and the anomalies of liquid water*, Nature **409**, 318 (2001). N Giovambattista, P G Debenedetti, F Sciortino, H E Stanley, *Structural order in glassy water*, Phys. Rev. E **71**, 061505 (2005). C A Angell, V Rodgers, *Near infrared spectra and the disrupted network model of normal and supercooled water*, J. Chem. Phys. **80**, 6245 (1984). J D Cruzan, L B Braly, K Liu, M G Brown, J G Loeser, R J Saykally, *Quantifying hydrogen bond coopertivity in water: VRT spectroscopy of the water tetramer*, Science **271**, 59 (1996). D. C. Rapaport, *Hydrogen bonds in water*, Mol. Phys. **50**, 1151 (1983). A Luzar, *Resolving the hydrogen bond dynamics conundrum*, J. Chem. Phys. **113**, 10663 (2000). F Mallamace, M Broccio, C Corsaro, A Faraone, U Wandrlingh, L Liu, C Mou, S H Chen, *The fragile-to-strong dynamics crossover transition in confined water: nuclear magnetic resonance results*, J. Chem. Phys. **124**, 124 (2006). C J Montrose, J A Búcaro, J Marshall-Coakley, T A Litovitz, *Depolarizated light-scattering and hydrogen bonding in liquid water*, J. Chem. Phys. **60**, 5025 (1974). F W Starr, J K Nielsen, H E Stanley, *Fast and slow dynamics of hydrogen bonds in liquid water*, Phys. Rev. Lett. **82**, 2294 (1999). F W Starr, J K Nielsen, H E Stanley, *Hydrogen-bond dynamics for the extended simple point charge model of water*, Phys. Rev. E. **62**, 579 (2000). S Chatterjee, P G Debenedetti, F H Stillinger, R M Lynden-bell, *A computational investigation of thermodynamics, structure, dynamics and solvation behavior in modified water models*, J. Chem. Phys. **128**, 124511 (2008). M G Mazza, N Giovambattista, H E Stanley, F W Starr, *Connection of translational and rotational dynamical heterogeneities with the breakdown of the Stokes-Einstein and Stokes-Einstein-Debye relations in water*, Phys. Rev. E **76**, 031203 (2007). M G Campo, G L Ferri, G B Roston, *$q$-exponential distribution in time correlation function of water hydrogen bonds*, Braz. J. Phys. **39**, 439 (2009).
H J C Berendsen, D van der Spoel, R V Drunen, *GROMACS: a message passing parallel molecular dynamics implementation*, Comp. Phys. Comm. **91**, 43 (1995). H J C Berendsen, J R Grigera, T P Straatsma, *The missing term in effective pair potentials*, J. Phys. Chem. **91**, 6269 (1987). H J C Berendsen, J Postma, W van Gunsteren, A Di Nola, J Haak, *Molecular dynamics with coupling to an external bath*, J. Chem. Phys. **81**, 3684 (1984). C Tsallis, *Possible generalization of Boltzmann-Gibbs statistics*, J. Stat. Phys. **52**, 479 (1988). L A Báez, P Clancy, *Existence of a density maximum in extended simple point charge water*, J. Chem. Phys. **101**, 9837 (1994).
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'J. E. Trümper, A. Zezas, Ü. Ertan ,'
- 'N. D. Kylafis'
date: 'Received ; accepted '
title: 'The energy spectrum of anomalous X-ray pulsars and soft gamma-ray repeaters'
---
[Anomalous X-ray pulsars (AXPs) and soft gamma-ray repeaters (SGRs) exhibit characteristic X-ray luminosities (both soft and hard) of around $10^{35}$ erg s$^{-1}$ and characteristic power-law, hard X-ray spectra extending to about 200 keV. Two AXPs also exhibit pulsed radio emission.]{} [Assuming that AXPs and SGRs accrete matter from a fallback disk, we attempt to explain both the soft and the hard X-ray emission as the result of the accretion process. We also attempt to explain their radio emission or the lack of it.]{} [We test the hypothesis that the power-law, hard X-ray spectra are produced in the accretion flow mainly by bulk-motion Comptonization of soft photons emitted at the neutron star surface. Fallback disk models invoke surface dipole magnetic fields of $10^{12} - 10^{13}$ G, which is what we assume here.]{} [Unlike normal X-ray pulsars, for which the accretion rate is highly super-Eddington, the accretion rate is approximately Eddington in AXPs and SGRs and thus the bulk-motion Comptonization operates efficiently. As an illustrative example we reproduce both the hard and the soft X-ray spectra of AXP 4U 0142+61 well using the XSPEC package compTB.]{} [Our model seems to explain both the hard and the soft X-ray spectra of AXPs and SGRs, as well as their radio emission or the lack of it, in a natural way. It might also explain the short bursts observed in these sources. On the other hand, it cannot explain the giant X-ray outbursts observed in SGRs, which may result from the conversion of magnetic energy in local multipole fields.]{}
Introduction
============
Anomalous X-ray pulsars (AXPs) and soft gamma-ray repeaters (SGRs) constitute a special population of young neutron stars distinguished by much higher X-ray luminosities than their rotational powers and spin periods clustered in a narrow range (2 - 12 s). The AXPs and SGRs are now believed to belong to the same class of objects, since short bursts that were once believed to be a distinctive property of the SGRs were observed from some of the AXPs as well (Gavriil et al. 2002; Kaspi et al. 2003). They are all spinning down with spin-period derivatives in the $10^{-13} - 10^{-11}$ s s$^{-1}$ range (Woods & Thompson 2006; Kaspi 2007; Mereghetti 2008 for recent reviews on AXPs and SGRs).
In addition to these main properties, broad-band observations have revealed many other peculiarities of these sources, which provide constraints for the models. In particular, some of these sources are persistent with soft X-ray luminosities of $L_{\mathrm{x,soft}} \sim 10^{34} - 10^{36}$ erg s$^{-1}$, while others, discovered in recent years, are transients with $L_{\mathrm{x,soft}} \sim ~10^{33}$ erg s$^{-1}$ in quiescence (see Table 1 of Mereghetti 2008). Both transient and persistent sources show occasional X-ray enhancements lasting from months to more than years and are correlated, in the long term, with infrared (IR) luminosities (Tam et al. 2004). The transient AXPs have been discovered in such X-ray outbursts (enhancements) when their $L_{\mathrm{x,soft}}$ levels were about two orders of magnitude higher than when in quiescence (Torii et al. 1998; Kouveliotou et al. 2003; Gotthelf et al. 2004; Ibrahim et al. 2004; Mereghetti et al. 2006, Israel et al. 2007; Muno et al. 2007). There are both similarities and systematic differences in the X-ray outburst light-curve morphologies of transient and persistent sources. Other constraints on models come from the properties of AXPs and SGRs in the optical to mid-IR bands and their relations with the X-ray luminosities.
These sources are widely believed to be magnetars mainly 1) because they account for the energetics of the super-Eddington soft gamma-ray bursts and 2) because the $P$ and $\dot{P}$ measurements lead to large magnetic dipole strengths based on the assumption that these sources spin down by magnetic dipole radiation. Magnetar models (Duncan & Thompson 1992; Thompson & Duncan 1993; 1995) do explain the super-Eddington bursts of SGRs. However, the quantitative explanation of the persistent soft X-ray luminosity by magnetic field decay, the optical and IR properties during persistent states, the X-ray and the accompanying IR enhancements, and the period clustering of AXP and SGRs all seem to meet difficulties within the original frame of the magnetar model.
On the other hand, fallback disk models (Chatterjee et al. 2000; Alpar 2001) can account for these observational facts. They have been developed in a series of connected, self-consistent ideas by means of detailed quantitative models (Ekşi & Alpar 2003; Ertan & Alpar 2003; Ertan & Cheng 2004; Ertan et al. 2006; Ertan & Çal[i]{}şkan 2006; Ertan et al. 2007; Ertan & Erkut 2008, Ertan et al. 2009). These models do not explain the super-Eddington bursts. Nevertheless, they are compatible with the presence of magnetar fields provided that these fields are in higher multipoles rather than in the dipole component. In these models, the strength of the dipole magnetic field is found to be less than $\sim 10^{13}$ G to account for the observations. The first direct observational support for the presence of fallback disks around these systems came from SPITZER observations of AXP 4U 0142+61 in mid-IR bands by Wang et al. (2006). The same source has also been detected in optical and near-IR bands. Through model fits, Wang et al. (2006) show that the mid-IR data can be reproduced by an irradiated disk model. These authors propose that the near-IR and the optical luminosities have a magnetic origin, while the mid-IR flux originates in a passive and irradiated disk. Later it was shown that all the data sets from optical to mid-IR can be accounted for by a single disk model that is active and irradiated by X-rays, which themselves are produced by accretion of disk matter onto the neutron star (Ertan et al. 2007). Recently, AXP 1E 2259+586 was also detected in the SPITZER bands. The combined overall spectrum, including the earlier IR detections of this source, is similar to that of AXP 4U 0142+61 (Kaplan et al. 2009).
In recent years, some of the AXPs, namely AXP 1E 1841–045 (Molkov et al. 2004; Kuiper et al. 2004), AXP 1RXS J1708–40 (Revnivtsev et al. 2004; den Hartog et al. 2008a) and AXP 4U 0142+61 (Kuiper et al. 2006; den Hartog et al. 2008b), have been found to emit pulsed hard X-rays up to 150 keV or more with isotropic luminosities close to the soft X-ray luminosities (see den Hartog et al. 2008b and Mereghetti 2008).
Detailed discussions of the magnetar models trying to explain the hard X-ray emission properties of AXPs can be found in (Heyl & Hernquist 2005; Heyl 2007; Beloborodov & Thompson 2007; Baring & Harding 2007). The advantages and disadvantages of the different magnetar models are summarized in den Hartog et al. (2008b).
The main aim of this paper is to discuss how the hard and the soft X-ray components can be produced in the framework of accretion from a fallback disk. In doing so, we treat AXPs and SGRs as a class with the following representative properties:
1\) Luminosities are $\sim 10^{35}$ erg s$^{-1}$.
2\) Soft ($< 10$ keV) and hard ($> 10$ keV) luminosities are similar.
3\) Dipolar magnetic field strengths are $10^{12} - 10^{13}$ G.
4\) Rotational periods are $\sim 5$ s.
We find that we can explain the observations very naturally by considering the bulk-motion Comptonization (BMC) that takes place in the accretion flow above the polar cap. The seed photons for this Comptonization are provided by the polar cap, while the observed soft X-ray emission comes from both the polar cap and an extended region around it.
As an illustrative example, we fit the observed X-ray spectrum (both soft and hard) of AXP 4U 0142+61 using the XSPEC package compTB (Farinelli et al. 2008). The fit is extremely good and the resulting parameters are very reasonable.
The BMC model is presented in § 2. In § 3 we present the X-ray data of AXP 4U 0142+61 and the fit to them using the XSPEC package compTB. In § 4 we discuss our model, and in § 5 we present the relation of AXPs and SGRs to other source classes. Finally in § 6 we give our conclusions.
Bulk-motion Comptonization model
================================
Soft X-ray emission
-------------------
The soft X-ray spectra of AXPs below 10 keV can be fitted by two-component models composed of two blackbody spectra or a blackbody plus a steep power-law spectrum (Mereghetti 2008). As pointed out by Gotthelf $\&$ Halpern (2005), the double blackbodies are more physically motivated. The hotter component may represent photospheric radiation from a small hot polar cap (area $A_{\rm hot}$), while the cooler component is photospheric emission from a large fraction of the neutron star surface (area $A_{\rm cool}$). Actually, such a combination of spectra is also found for young energetic radio pulsars and for isolated neutron stars showing purely thermal emission.
For an accreting neutron star, whose magnetic axis is inclined with respect to the fallback disk, the emitting region $A_{\rm hot}$ at the base of the accretion flow will not have radial symmetry around the magnetic pole, but will assume the bow-shaped configuration discussed by Basko & Sunyaev (1976) and confirmed by the MHD calculations of Romanova et al. (2004; see also Bachetti et al. 2009). This region is heated by the infall of accreting matter and the produced soft X-ray photons either escape unscattered, and are observed as such, or get upscattered in the accretion flow and produce the hard X-ray spectrum.
Hard X-ray emission
-------------------
The equation describing upscattering of soft photons in a converging, ionized, fluid flow was first introduced and solved by Blandford & Payne (1981a;b) and Payne & Blandford (1981). Their Comptonization equation (bulk-motion Comptonization) was solved in the case of spherical accretion onto a neutron star by Mastichiadis & Kylafis (1992), who assumed a reflective neutron-star surface. This last work was generalized by Titarchuk et al. (1996, 1997), who examined the general case of a partially reflecting inner boundary and also included a second order term in the flow velocity (see also Psaltis & Lamb 1997; Psaltis 2001).
The idea in the bulk-motion Comptonization (BMC) model, as applied to the AXP X-ray spectra, is the following. Soft X-ray photons, emitted by the polar cap region, where the accretion occurs, find themselves in the accretion flow. For accretion rates comparable to the Eddington rate, the optical depth to electron scattering in the accretion flow is close to unity or more. Thus, a fraction of the emitted soft photons get trapped in the flow and after several, nearly head-on collisions with the accreting electrons acquire significant amounts of energy. As a result, a power-law, hard X-ray spectrum is produced.
Torrejon et al. (2004) have applied the BMC model to the wind-accreting, slowly spinning, neutron star 4U 2206+54, which has similar properties to our sources ($L_x \sim 10^{35}$ erg s$^{-1}$, $L_{\rm soft} \sim L_{\rm hard}$, $kT_{\rm bb} \sim 1.2$ keV, and a hard spectral tail extending up to $\buildrel > \over \sim 90$ keV). The BMC model was also used to fit hard X-ray tails observed in neutron star sources by Paizis et al. (2006), and it was significantly improved by Farinelli et al. (2008).
Application to AXP 4U 0142+61
=============================
The AXP 4U 0142+61 is one of the brightest known AXPs. Therefore its spectrum can provide one of the most stringent tests of the applicability of the BMC model to the X-ray spectra of AXPs. It has been extensively observed with the [*Chandra*]{}, XMM-[*Newton*]{}, and INTEGRAL X-ray missions.
To constrain the low-energy (0.5 - 10.0 keV) spectrum of 4U 0142+61, we opted for the available [*Chandra*]{} high-energy transmission grating (HETG) spectra. These data are least affected by pile-up, while they do cover the desired energy band. There are also [*Chandra*]{} data obtained in continuous clocking mode (previously analyzed by Patel et al. 2003); however, the calibration of this mode for spectral analysis is still uncertain, as is indicated by the differences between the published analysis of the HETG and the continuous clocking mode data (Juett et al. 2006; Patel et al. 2003). For the high-energy spectrum of 4U 0142+61, we use the extensive ISGRI INTEGRAL data, which cover the 20 - 200 keV band. The spectral analysis has been performed with the XSPEC v12.0 package (Arnaud 1996). All cited errors are at the 90% confidence level for one interesting parameter unless otherwise specified.
[*Chandra*]{} data
------------------
4U 0142+61 has been observed with the [*Chandra*]{} High-Energy Transmission Grating for 25 ksec in May 2001 (OBSID 1018; PI C. Canizares). We used the reduced spectra available from the [*Chandra*]{} Grating Data Archive and Catalog (TGCat; http://tgcat.mit.edu/). Since even the 0th order spectrum is affected by pile up, we used the positive and negative 1st order of the Medium Energy Transmission Grating and the High Energy Transmission Grating. The individual spectra were binned in order to have at least 50 counts per bin. The four spectra were fitted simultaneously using the relevant response matrices (rmfs) and ancillary responce matrices (arfs) for each order.
A fit of the HETG data with a two component blackbody and power-law model, affected by photoelectric absorption by cold gas ( model in XPSEC) gives a blackbody temperature of $0.42^{+0.03}_{-0.02}$ keV, a power-law slope of $\Gamma=3.66^{+0.25}_{-0.29}$, and an absorbing $\rm{H}{\textsc i}$ column density of $0.95^{+0.10}_{-0.12}\times10^{22}$ cm$^{-2}$. These parameters are almost identical to the parameters derived by Juet et al. (2002) from the analysis of the same data ($kT=0.418 \pm 0.013$ keV; $\Gamma=3.3 \pm 0.4$, $N_{\rm H} = (0.88 \pm 0.13) \times 10^{22}$ cm$^{-2}$). However, they are slightly different from the parameters derived by Patel et al. ($kT=0.470 \pm 0.008$ keV; $\Gamma = 3.40 \pm 0.06$, $N_{\rm H}= (0.93 \pm 0.02)\times10^{22}$ cm$^{-2}$) from the spectral fits of the continuous clocking mode data. Given the uncertainties in the calibration of the latter mode, we consider that the HETG data provide a more accurate representation of the low-energy spectrum of 4U 0142+61.
INTEGRAL ISGRI data
-------------------
4U 0142+61 has been extensively observed by INTEGRAL. We downloaded all public pointed observations from the INTEGRAL archive, which are longer than 3 ksec and the source falls within 10 degrees from the pointing direction. The data were analyzed with the OSA v8.0[^1] data analysis software provided by the Integral Science Data Analysis Center. We used version 8.0.1 of the instrument characteristics database (which provides the latest calibration data) and version 30.0 of the reference catalog (which provides a list of known sources used for the source detection and spectral extraction processes). First, we produced an image of the observed region in the 4 standard ISGRI bands (20 - 40 keV, 40 - 60 keV, 80 - 100 keV, 100 - 200 keV) following the standard procedures described in the IBIS analysis user manual[^2]. 4U 0142+61 has been clearly detected in all four bands. Next we extracted a spectrum of 4U 0142+61, again following the standard procedures for IBIS data analysis. As recommended, we added 2% systematic errors to the spectrum to account for calibration uncertainties.
We fitted the spectrum with a power-law model, using the latest arf and rmf files. We find an energy slope of $\Gamma=0.94\pm0.10$, consistent with the slope of $\Gamma=1.05\pm0.11$ reported by Kuiper et al. (2006). We also fitted the spectrum with older arf files available in the instrument characteristics database in order to assess the effect of different calibration data on the measured spectral parameters. We did not find any statistically significant difference between the estimated spectral parameters.
Joint [*Chandra*]{} and INTEGRAL ISGRI fits
-------------------------------------------
In the previous paragraphs, we have shown that our analysis of the archival [*Chandra*]{} and INTEGRAL ISGRI data of 4U 0142+61 nicely reproduces the canonical model consisting of an absorbed blackbody and power-law model in the soft X-ray band (0.5-10.0 keV) and a hard power-law above 20.0 keV. Next we investigated whether the same spectrum can be reproduced with a bulk-motion Comptonization model. For this reason we used the model of Farinelli et al. (2008) provided as an external XSPEC model (model [^3]). This model includes a self-consistent treatment of the seed blackbody spectrum and the thermal and/or bulk-motion Comptonization of its photons. Therefore we did not include an ad-hoc blackbody component. The seed spectrum is described by a modified blackbody function $S(E) \propto
E^{\gamma-1} / [exp(E/kT_{s}) -1]$, where $kT_{s}$ is the characteristic temperature of the blackbody, and $E^{\gamma}$ is a power-law component that modifies the blackbody. For $\gamma=3$ this component simplifies to a pure blackbody. Since this model includes both bulk-motion and thermal Comptonization, an important parameter is the relative efficiency of the two components defined as $\delta = <E_{\rm bulk}>) / <E_{\rm th}>$. For $\delta=0$ we have a pure thermal Comptonization spectrum. Other parameters of this model are the energy index of the Comptonization spectrum ($\alpha$; for more details see Farinelli et al. 2008), the temperature of the Comptonizing electrons ($kT_{e}$), a factor describing the ratio between the observed Compton scattered spectrum and the observed seed blackbody spectrum ($A$), and the normalization of the seed photon spectrum ($C_{N}$). To model photoelectric absorption by cold gas, we included the XSPEC model component.
The results of this fit are presented in Table \[Table:Spec\] and Figures \[Fig:Spec1\] and \[Fig:Spec2\]. The first figure shows the data and the best-fit model (top panel). It also shows the ratio of the data to the best-fit model (bottom panel). Figure \[Fig:Spec2\] shows the model and the data corrected for instrumental effects (unfolded spectrum) in $E^{2}f(E)$ space. It is evident from this figure that the model gives a very good fit to the broad band spectrum of 4U 0142+61 ($\chi_{\nu}^{2}=166.7/248$). The temperature of the seed photons ($kT=0.72$ keV) is slightly higher than the temperature estimated from the blackbody fit of the [*Chandra*]{} data alone. However, the parameter $\gamma=0.95$, which modifies the seed blackbody spectrum indicates that it is not a pure blackbody. Furthermore, we find that the bulk-motion Comptonization dominates over thermal Comptonization ($\delta=2.32$) and that only 0.3% of the observed photons have been Compton scattered \[$\log(A)=-2.3$\]. These results show that the spectrum of 4U 0142+61 can be reproduced equally well with a self-consistent BMC model with very reasonable parameters.
Parameter Value
--------------------------------------------------- ---------------------
Blackbody temperature ($kT_{s}$) 0.72 keV
Index of Seed photons ($\gamma$) 0.95
Energy index of Compton spec. ($\alpha$) 0.36
Efficiency of bulk over thermal Compt. ($\delta$) 2.32
Electron Temperature ($kT_{e}$) 24.6 keV
Ilumination factor (log$(A)$) -2.3
Normalization ($C_{N}$)$\dagger$ $1.18\times10^{-2}$
$\chi^{2}$/d.o.f. 166.74 / 248
: Fits of the ISGRI and [*Chandra*]{} data with the model[]{data-label="Table:Spec"}
$\dagger$ Normalization of the seed-photon spectrum defined as $L_{39}/D_{10}^2$, where $L_{39}$ is its luminosity in units of $10^{39}\,\rm{erg\,s^{-1}}$, and $D_{10}$ is the distance to the source in units of 10 kpc
Discussion
==========
The advantage of the bulk-motion Comptonization model is that it uses a proven concept, which has been applied to many cases including the slowly pulsating source 4U 2206+54, which in turn shows a low-luminosity ($10^{35}$ erg s$^{-1}$) and a hard X-ray tail like the AXPs (Torrejon et al. 2004). Another advantage is that the whole spectrum, from the soft thermal component to the hard X-ray tail, is explained by a single, well-understood mechanism using as few parameters as possible. The accreting matter both heats the polar cap, which emits the seed photons for Comptonization, and provides the high-energy electrons, which do the Comptonization. Depending on detailed conditions, both the soft and the hard X-ray components have luminosities that are within the same order of magnitude as observed. Since the soft and hard components come from the same place at the neutron star surface, their maxima occur at basically the same rotational phase. This agrees well with the observational fact that these sources show a very slow change in pulse profile with energy (Kaspi 2007). On the other hand, there are fine structures in the energy-dependent pulse profiles, which can be explained by cyclotron resonance effects, as in the case of normal X-ray pulsars having dipole magnetic fields of similar strengths to the ones postulated by our model for AXPs and SGRs.
The radio emission observed in two of the AXPs \[1E 1547-54 (Camilo et al. 2007) and XTE J1810-197 (Halpern et al. 2005; Camilo et al. 2006)\] can also be explained by the BMC model. Usually, it is thought that the radio emission of a neutron star is quenched by the accretion, but there may be little or no overlap in our model between the circumpolar zone of radio emission and the bow-shaped accretion zone (Romanova et al. 2004). Indeed, according to the canonical expression, the size of the radio-active polar cap is $r_{\rm pc} = 10^4 P^{-1/2} \sim 5 \times 10^3$ cm. It seems likely that the bow-shaped accretion region is farther from the magnetic pole. Therefore, both radio and X-ray emission may coexist.
A necessary requirement for a detectable radio flux is that the magnetic dipole power $\dot E$ is strong enough. Comparing $\dot E = (1/6 c^3) B^2 R^6 (2\pi/P)^4$, where $B$ is the surface magnetic field, $R$ the radius of the neutron star, $P$ the rotational period and $c$ the speed of light, with the observed radio luminosities of the two pulsars, we estimate the radio emission efficiencies to be $\sim 10^{-4}$ and $\buildrel > \over \sim 0.03$ for 1E 1547-54 and XTE J1810-197, respectively, when assuming a polar magnetic field of $10^{13}$ G. A lower magnetic field strength is also allowed in our model, because the observed radio luminosity is not isotropic. These efficiencies are quite large, but comparable to those of the most efficient normal radio pulsars with periods $> 2$ s, taken from the ATNF catalog http://www.atnf.csiro.au/research/pulsar/psrcat/ (Manchester et al. 2005).
The fact that most AXPs and SGRs have remained undetected in the radio band despite deep searches (Burgay et al. 2007) may be due to their long periods, which may place them beyond the death line. Another reason could be that long-period radio pulsars have narrow beams with widths scaling with $P^{-1/3}$ (Lyne & Manchester 1988). For a period of 6 s, the full beam is then only $\sim 7$ degrees. Therefore the probability of detecting the radio beam of an AXP is only a few percent.
Relation to other source classes
================================
The main point of this paper is to explain the hard X-ray spectra of AXPs and SGRs as a result of accretion from a fallback disk. In addition, we attribute the soft X-ray spectra of AXPs and SGRs to thermal, photospheric radiation from the polar cap (hotter component) and from the bulk surface of the neutron star (cooler component). This latter point leads us to the following thoughts:
The class of isolated neutron stars showing purely thermal emission that was discovered by ROSAT (XDINS, “Magnificent Seven”) exhibits blackbody emission as well, though at lower temperatures (by a factor $\sim 10$) and lower luminosities (by $\sim 10^4$) compared with AXPs. Like the latter, the XDINs have long periods and strong magnetic fields ($\ge 10^{13}$ G, Haberl 2007). Their period derivatives and proper motion data indicate ages of a few hundred thousand years, and they are not associated with supernova remnants. All these facts show that XDINs are older than AXPs, supporting the idea that some (or possibly all) XDINs have been AXPs before they exhausted their residual disk and have subsequently cooled down.
Another young neutron-star population, namely the central compact objects (CCOs, see review by Pavlov et al. 2004), are probably related to AXPs, SGRs, and XDINs as well. It was originally proposed by Alpar (2001) that similarities and differences in these neutron star classes could be explained if fallback disks with different properties are included in the initial parameters of the evolution models in addition to initial spin period and the magnetic moment of the neutron star.
Conclusions
===========
Contrary to common belief, AXPs and SGRs may not be that different from normal accreting X-ray pulsars. The major differences between the two classes of X-ray sources are the extent of the hard X-ray spectrum and the X-ray luminosity, which is lower for AXPs/SGRs by two to three orders of magnitude. The hard X-ray power-law spectral index is essentially the same for both classes, but the spectrum of normal X-ray pulsars extends to about 20 keV, while that of the anomalous X-ray pulsars extends to about 200 keV. In our model, this difference is caused by the normalized (to the Eddington value) accretion rate being near unity in AXPs/SGRs, and bulk-motion Comptonization takes place and produces the hard X-ray tail, which extends to $\sim 200$ keV.
Our model naturally explains the $\sim 100$ % pulsations observed at $\sim 100$ keV, because they come from the accretion flow above the polar cap. It also explains the similarity between AXPs/SGRs and XDINS, which has often been noted. It may be that XDINS are evolved AXPs with no remaining accretion disk. The absence of cyclotron lines in the observed X-ray spectra is a natural consequence of our model. For a dipole magnetic field of $\sim 10^{13}$ G, the electron cyclotron line would appear at $E > 100$ keV, where the photon statistics are not good enough. On the other hand, a proton cyclotron line would appear at $E \sim 0.05$ keV, where interstellar absorption is huge.
Our model could in principle explain the short bursts observed from AXPs and SGRs as the result of infalling lumps of matter with appropriate mass. We defer an analysis of this process to another paper. On the other hand, our model does not explain the giant bursts observed from AXPs and SGRs. These may indeed require very strong, magnetar-type magnetic fields. Our model does not exclude the existence of such super strong magnetic fields if they reside in multipole components.
Last but not least, we feel that the detection of a fallback disk around the neutron stars in AXP 4U 0142+61 and AXP 1E 2259+586 gives strong observational support to our model.
Ü.E. and J.T. thank the Astrophysics Group of the University of Crete and FORTH for their hospitality. This research has been supported in part by EU Marie Curie project no. 39965, EU REGPOT project number 206469 and by EU FPG Marie Curie Transfer of Knowledge Project ASTRONS, MKTD-CT-2006-042722. Ü.E. acknowledges research support from TÜB[İ]{}TAK (The Scientific and Technical Research Council of Turkey) through grant 107T013 and support from the Sabanci University Astrophysics and Space Forum.
Alpar, M. A. 2001, ApJ, 554, 1245
Arnaud, K.A., 1996, Astronomical Data Analysis Software and Systems V, eds. Jacoby G. and Barnes J., p17, ASP Conf. Series volume 101.
Bachetti, M., Romanova, M., Kulkarni, A., Burderi, L., di Salvo, T. 2009, arXiv:0911.4493
Baring, M. G., & Harding, A. K. 2007, Ap&SS, 308, 109
Basko, M. M., & Sunyaev, R. A. 1976, MNRAS, 175, 395
Beloborodov, A. M., & Thompson, C. 2007, ApJ, 657, 967
Blandford, R. D. & Payne, D. G. 1981a, MNRAS, 194, 1033
Blandford, R. D. & Payne, D. G. 1981b, MNRAS, 194, 1041
Burgay, M., Rea, N., Israel , G., et al, 2007, Ap&SS 308, 531B
Camilo, F., Ransom, S. M., Halpern, J. P., et al. 2006, Nature, 442, 892
Camilo, F., Ransom, S. M., Halpern, J. P. & Reynolds, J. 2007, ApJ, 666, L93
Chatterjee, P., Hernquist, L., & Narayan, R. 2000, ApJ, 534, 373
den Hartog, P. R., Kuiper, L., & Hermsen, W. 2008a, A&A, 489, 263 den Hartog, P. R., Kuiper, L., Hermsen, W., et al. 2008b, A&A, 489, 245
Duncan, R. A., & Thompson, C. 1992, ApJ, 392, 9
Ekşi, K. Y., & Alpar, M. A. 2003, ApJ, 599, 450
Ertan, Ü., & Alpar, M. A. 2003, ApJ, 593, L93
Ertan, Ü., & Çal[i]{}şkan, Ş. 2006, ApJ, 649, L87
Ertan, Ü, & Cheng, K. S. 2004, ApJ, 605, 840
Ertan, Ü., Ekşi, K. Y., Erkut, M. H., & Alpar, M. A. 2009, ApJ, 702, 1309
Ertan, Ü., Erkut, M. H., Ekşi, K. Y., & Alpar, M. A. 2007, ApJ, 657, 441
Ertan, Ü., & Erkut, M. H. 2008, ApJ, 673, 1062
Ertan, Ü., Göğüş, E., & Alpar, M.A. 2006, ApJ, 640, 435
Farinelli, R., Titarchuk, L., Paizis, A., & Frontera, F. 2008, , 680, 602
Gavriil, F. P., Kaspi, V. M. & Woods, P.M. 2002, Nature, 419, 142
Gotthelf, E. V., & Halpern, J. P. 2005, ApJ, 632, 1075
Gotthelf, E. V., Halpern, J. P., Buxton, M., & Bailyn, C. 2004, ApJ, 605, 368
Haberl, F., 2007, Ap&SS, 308, 181
Halpern, J. P., Gotthelf, E.V., Becker, R. H., Helfand, D. J., & White, R. L. 2005 ApJ, 632, L29
Heyl, J. S. 2007, Ap&SS, 308, 101
Heyl, J. S., & Hernquist, L. 2005, ApJ, 618, 463
Ibrahim, A. I., Markwardt, C. B., Swank, J. H. , et al. 2004, ApJ, 609, 21
Israel , G. L., Campana, S., Dall’Osso, et al. 2007, ApJ, 664, 448
Juett, A. M., Marshall, H. L., Chakrabarty, D., & Schulz, N. S. 2002, , 568, L31
Kaplan, D. L., Chakrabarty, D., Wang, Z., & Wachter, S. 2009, ApJ, 700, 149
Kaspi, V. M., Gavriil, F. P., Woods, P.M., et al. 2003, ApJ, 588, L93
Kaspi, V. M., 2007, , 308, 1
Kouveliotou, C., et al. 2003, ApJ, 596, 79
Kuiper, L., Hermsen, W., & Mendez, M. 2004, ApJ, 613, 1173
Kuiper, L., Hermsen, W., den Hartog, R. P. & Collmar, W. 2006, ApJ, 645, 556
Lyne, A. G., & Manchester, R. N. 1988, MNRAS, 234, 477
Manchester, R. N., Hobbs, G. B., Teoh, A. & Hobbs, M. 2005, AJ, 129, 1993
Mastichiadis, A. & Kylafis, N. D. 1992, ApJ, 384, 136
Mereghetti, S. 2008, A&Arv, 15, 225
Mereghetti, S., et al. 2006, A&A, 450, 759
Molkov, S. V., Cherepashchuk, A. M., Lutovinov, A. A., et al. 2004, Astron. Lett., 30, 534
Muno, M. P., Gaensler, B. M., Clark, J. S., de Grijs, R., Pooley, D., Stevens, I. R., & Portegies Zwart, S. F. 2007, MNRAS, 378, L44
Paizis, A. et al. 2006, A&A, 459, 187
Patel, S. K., et al. 2003, , 587, 367
Pavlov, G. G., Sanwal, D., & Teter, M. A., 2004, in Young Neutron Stars and Their Environments, IAU Symposium, Vol. 218, F. Camilo and B. M. Gaensler, eds.(astro-ph/0311526)
Payne, D. G. & Blandford, R. D. 1981, MNRAS, 196, 781
Psaltis, D. 2001, ApJ, 555, 786
Psaltis, D. & Lamb, F. K. 1997, ApJ, 488, 881 Revnitsev , M. G., Sunyaev, R. A., Varshalovich, D. A., et al. 2004, Astron. Lett., 30, 382
Romanova, M. M., Ustyugova, G.V., Koldoba, A.V., & Lovelace, R.V.E. 2004, ApJ 610, 920
Tam, C. R., Kaspi, V.M., van Kerkwijk, M.H., $\&$ Durant, M. 2004, ApJ, 617, L53
Thompson C., & Duncan, R. C. 1993, ApJ, 408, 194
Thompson C., & Duncan, R. C. 1995, MNRAS, 275, 255
Titarchuk, L., Mastichiadis, A., & Kylafis, N. D. 1996, A&AS, 120, C171
Titarchuk, L., Mastichiadis, A., & Kylafis, N. D. 1997, ApJ, 487, 834
Torii, K. et al. 1998, ApJ, 503, 843
Torrejon, J. M., Kreykenbohm, I., Orr, A., Titarchuk, L., & Negueruela, I. 2004, A&A, 423, 301
Wang, Z., Chakrabarty, D., & Kaplan, D. L. 2006, Nature, 440, 772
Woods, P. M., & Thompson, C. 2006, in “Compact Stellar X-ray Sources”, eds. W.H.G. Lewin and M. van der Klis, Cambridge Univ. Press.(astro-ph/0406133)
[^1]: http://www.isdc.unige.ch/integral/download/osa\_sw
[^2]: http://isdcul3.unige.ch/Soft/download/osa/osa\_doc/osa\_doc-8.0/osa\_um\_ibis-7.0.pdf
[^3]: http://heasarc.gsfc.nasa.gov/xanadu/xspec/models/comptb.html
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Radio pulsars are believed to have their emission powered by the loss of rotational kinetic energy. By contrast, magnetars show intense X-ray and $\gamma$-ray radiation whose luminosity greatly exceeds that due to spin-down and is believed to be powered by intense internal magnetic fields. A basic prediction of this picture is that radio pulsars of high magnetic field should show magnetar-like emission. Here we report on a magnetar-like X-ray outburst from the radio pulsar PSR J1119$-$6127, heralded by two short bright X-ray bursts on 2016 July 27 and 28 [@klm+16; @ykr16]. Using Target-of-Opportunity data from the [*Swift*]{} X-ray Telescope and [*NuSTAR*]{}, we show that this pulsar’s flux has brightened by a factor of $>160$ in the 0.5–10 keV band, and its previously soft X-ray spectrum has undergone a strong hardening, with strong pulsations appearing for the first time above 2.5keV, with phase-averaged emission detectable up to 25 keV. By comparing [*Swift*]{}-XRT and [*NuSTAR*]{} timing data with a pre-outburst ephemeris derived from [*Fermi*]{} Large Area Telescope data, we find that the source has contemporaneously undergone a large spin-up glitch of amplitude $\Delta\nu/\nu = 5.74(8) \times 10^{-6}$. The collection of phenomena observed thus far in this outburst strongly mirrors those in most magnetar outbursts and provides an unambiguous connection between the radio pulsar and magnetar populations.'
author:
- 'R. F. Archibald, V. M. Kaspi, S. P. Tendulkar, and P. Scholz'
title: 'A Magnetar-like Outburst from a High-B Radio Pulsar'
---
Introduction {#sec:intro}
============
PSR J1119$-$6127 is a radio pulsar having spin period $P=0.407$ s, discovered in the Parkes multibeam 1.4-GHz survey [@ckl+00]. The pulsar’s $P$ and spin-down rate $\dot{P} = 4.0 \times 10^{-12}$ imply a characteristic age $\tau < 2$ kyr, making this object one of the youngest pulsars in the Galaxy, consistent with an association with the supernova remnant G292.2$-$0.5 [@cgk+01] at a distance of 8.4kpc [@cmc04]. Those same spin parameters, under the assumption of vacuum dipole spin-down, imply a dipolar surface magnetic field $B = 4.1 \times 10^{13}$ G, among the highest known among radio pulsars. The pulsar’s spin-down luminosity is $\dot{E} = 2.3 \times 10^{36}$ erg s$^{-1}$.
Past X-ray observations of the source [@gs03b; @sk08; @nkh+12] have shown it to be a soft X-ray pulsar, with strong pulsations below 2.5 keV, and none seen above this energy. This emission was well described by a two-component model consisting of a power law of index $\sim$2.1, with a hot thermal component of blackbody temperature $kT \simeq 0.2$ keV, high compared to lower-field radio pulsars of comparable age [see also @km05; @zkm+11; @ozv+13]. Radio pulse profile changes, short radio bursts and unusual timing recoveries were observed near epochs of glitches in this source [@wje11; @awe+15], reminiscent of radio radiative behavior [e.g. @crh+06] and glitch recoveries [@dk14] following magnetar outbursts. The pulsar is also among the highest-B sources to have been detected in $\gamma$-rays by [*Fermi*]{} [@pkh+11].
The [*Fermi*]{} Gamma-ray Burst Monitor (GBM) and [*Swift*]{} Burst Alert Telescope (BAT) both reported short magnetar-like bursts from PSR J1119$-$6127, on 2016 July 27 [UT 13:02:08; @ykr16] and 2016 July 28 [UT 01:27:51; @klm+16], respectively. Immediately following the BAT burst, the [*Swift*]{} X-ray Telescope (XRT) found a bright X-ray source at the position of PSR J1119$-$6127 [@klm+16] with pulsations at the rotational period [@ave16]. This suggests that this radio pulsar has had a magnetar-like outburst, similar to the 2006 transition of rotation-powered, but radio-quiet, pulsar PSR J1846$-$0258 [@ggg+08]. Interestingly, the radio pulsations from [PSR J1119–6127]{} have disappeared [@bpk+16].
We report here on our Target-of-Opportunity (ToO) X-ray observations of PSR J1119$-$6127 with the [*Swift*]{}/XRT and [*NuSTAR*]{} during the first few days of the transition, as well as on pre-outburst [*Fermi*]{}/LAT timing data.
Observations & Analysis
=======================
[*Swift*]{}-XRT & [*NuSTAR*]{} Observations
-------------------------------------------
The [*Swift*]{}-XRT [@bhn+05] slewed to observe [PSR J1119–6127]{} 62.8s after the BAT trigger [@klm+16]. XRT was operated in Photon Counting (PC) mode for this 2.2-ks observation (ObsID 00706396000, spanning 2016 July 28 01:28 – 02:07 UT), and in Windowed-Timing (WT) mode for the follow-up observations (ObsID 00034632001/2), spanning July 28 17:20 to July 29 03:11 UT and July 31 04:20 to August 1 20:37 UT for exposures of 9.9 and 4.8ks, respectively.
As the time resolution of PC mode is 2.5s (longer than the period of the pulsar), only WT mode observations, with a time resolution of $1.76\,$ms, were used in the timing analysis.
[*NuSTAR*]{} [@hcc+13] began ToO observations of [PSR J1119–6127]{} at 2016 July 28, 23:05:12 UT yielding a total exposure time of 54.5ks (ObsID 80102048002) partially overlapping with XRT observation 00034632001. The data from the two focal plane modules of [*NuSTAR*]{} are referred to hereafter as FPMA and FPMB.
The [*Swift*]{}-XRT and [*NuSTAR*]{} data were processed with the standard `xrtpipeline` and `nupipeline` scripts, respectively, using *HEASOFT* v6.17 and time corrected to the Solar System barycenter from the [*Chandra*]{} location of [PSR J1119–6127]{} [@gs03b].
For [*Swift*]{}, we selected only Grade 0 events for spectral fitting as other event grades are more likely to be caused by background events [@bhn+05]. [*Swift*]{} spectra were extracted from the selected regions using [extractor]{}. Source photons were extracted from a 10-pixel radius circular region centered on [PSR J1119–6127]{} with an annular background region with an inner and outer radius of 75 and 125 pixels, respectively.
The WT observations had multiple soft X-ray bursts which appear in both the source and background regions that, from past experience, seem to be instrumental in origin. Hence, for the WT mode data, we excluded all photons below 0.7keV from our analysis.
For [*NuSTAR*]{} the source events were extracted within a 30-pixel (72) radius around the centroid. Appropriate background regions were selected to be on the same detector as the source location. Spectra were extracted using the `nuproducts` script.
Using `grppha`, channels 0–70 ($<0.7$keV) and 700–1023 ($>7\,$keV) for [*Swift*]{} data and channels 0–35 ($<3$keV) and 1935–4095 ($>79\,$keV) for [*NuSTAR*]{} data were ignored and all good channels were binned to have a minimum of one count per energy bin.
[*Fermi*]{} Large Area Telescope Observations
---------------------------------------------
We downloaded Pass 8R2 events of [*Fermi*]{} Large Area Telescope (LAT) [@aaa+09a] all-sky survey observations from 2008 August 4 to 2016 July 30 from a one degree radius surrounding the [*Chandra*]{} position of [PSR J1119–6127]{} and applied the recommended event selection. In the timing analysis, we used only photons having energy greater that 500 MeV based on the $\gamma$-ray pulse profile of [PSR J1119–6127]{} [@pkh+11]. We corrected the arrival times of each photon to the Solar System barycenter using the [tempo2]{} [fermi]{} plug-in [@rkp+11].
Results
=======
Timing Analysis {#sec:timing}
---------------
Times-of-arrivals (TOAs) of $\gamma$-ray pulses were extracted using a maximum likelihood (ML) method, described in [@lrc+09]. We extracted a TOA from photons collected in every 100 day span as a trade-off between TOA spacing and precision. We extended the ephemeris presented by @awe+15 using the LAT detected photons until the GBM-detected burst [@ykr16]. We present a phase-coherent ephemeris in Table \[tab:timing\] and the timing residuals in the left panel of Figure \[fig:resplot\].
To determine an ephemeris for the post-outburst [*Swift*]{} and [*NuSTAR*]{} observations, we folded the soft X-ray photons ($<10$keV) from each observation, starting with the ephemeris from the LAT observations and extracted TOAs from each orbit using the ML method. As there is no apparent evolution in the pulse profiles over the $<10\,$keV energy band, the offset between TOAs from both telescopes should be minimal. We then used the [tempo2]{} timing software package [@hem06] to fit the TOAs.
It is apparent that the LAT ephemeris did not accurately describe the post-outburst TOAs and requires a change in the spin frequency. Due to the long integration times required to extract a TOA from LAT, we are unable to constrain the exact glitch epoch; for this analysis, we have fixed the glitch epoch to the time of the first GBM-detected burst [@ykr16] and fitted for a glitch in spin frequency and frequency derivative.
We measure a spin-up glitch with $\Delta\nu=1.40(2)\times10^{-5}$Hz and $\Delta\dot{\nu}=-1.9(5)\times10^{-12}\,$Hz s$^{-1}$. We caution that this represents a snapshot of the frequency evolution, and that following glitches in magnetars, complex recoveries are often observed [e.g. @dkg08; @dk14].
{width="\textwidth"}
[ll]{}\
Dates (MJD) & 54732.82–57544.08\
Dates & 2008 Sept 23 – 2016 June 5\
Epoch (MJD) & 56264.00000\
$\nu\;$ (s$^{-1}$) & 2.442 579 294 0(9)\
$\dot{\nu}\;$ (s$^{-2}$) &$-$2.390 210(4)$\times 10^{-11}$\
$\ddot{\nu}\;$ (s$^{-3}$) & $5.68(3)\times 10^{-22}$\
$\dddot{\nu}\;$(s$^{-4}$) & $ -1.46(9)\times 10^{-30}$\
$\nu^{(4)}$$\;$(s$^{-5}$) & $ -3.1(7)\times 10^{-38}$\
$\nu^{(5)}$$\;$(s$^{-6}$) & $ 1.8(1)\times 10^{-45}$\
$\nu^{(6)}$$\;$(s$^{-7}$) & $ 3(1)\times 10^{-53}$\
RMS residual (ms) & 15.5\
RMS residual (phase) & 0.037\
$\chi^2_\nu$/dof & 1.08/20\
\
Dates (MJD) & 57597.72–57601.85\
Dates & 28 July –1 Aug 2016\
Epoch (MJD) & 57600.\
$\nu\;$ (s$^{-1}$) & 2.439 837 34(8)\
$\dot{\nu}\;$ (s$^{-2}$) &$-$2.57(5)$\times 10^{-11}$\
RMS residual (ms) & 4.22\
RMS residual (phase) & 0.001\
$\chi^2_\nu$/dof & 0.74/46\
\
Glitch Epoch (MJD), fixed & 57596.547\
$\Delta\nu\;$ (s$^{-1}$) & $1.40(2)\times 10^{-5}$\
$\Delta\dot{\nu}\;$ (s$^{-2}$) & $-1.9(5)\times 10^{-12}$\
Note: Figures in parentheses are the 1$\sigma$ <span style="font-variant:small-caps;">tempo2</span> uncertainties in the least-significant digits quoted. The source location was fixed at the [*Chandra*]{} position.
![Pre- and post-outburst X-ray pulse profiles of [PSR J1119–6127]{} from [*Swift*]{}-XRT (00034632001, red) and 2011 [*XMM-Newton*]{} data [@nkh+12] (black) in the soft (0.7–2.5keV, top panel) and hard (2.5–10keV, bottom panel). The profiles have been arbitrarily offset vertically for clarity and aligned in phase based on the soft X-ray profile.[]{data-label="fig:profs"}](Prof_compare){width="48.00000%"}
In Figure \[fig:profs\] we show the soft (0.7–2.5keV), and hard (2.5–10keV) X-ray pulse profiles from the radio pulsar state in 2011 [@nkh+12], and the magnetar-like state in 2016. In the soft band, the root-mean-squared (RMS) pulse fraction has increased from 38(3)% to 71(4)%, while the pulse shape has remained similar. In the hard band, the RMS pulse fraction went from $<10\%$ to 56(3)% pulsed.
X-ray Spectroscopy {#sec:spectroscopy}
------------------
All X-ray spectra were fit using `XSPEC` v12.9.0 [@arn96] with a common value for hydrogen column density ($N_\mathrm{H}$). Magnetar spectra are typically described with an absorbed blackbody plus power-law model, which we use here. However, the independent [*Swift*]{}-XRT observations 00706396000 and 00034632002 were fit with a simple absorbed blackbody model between 0.7–7keV as there is no constraint on the power law without the [*NuSTAR*]{} spectrum and there is little power-law contribution below 7keV.
We used Cash statistics [@cas79a] for fitting and parameter estimation of the unbinned data. $N_\mathrm{H}$ was fit using `wilm` abundances and `vern` photoelectric cross-sections. The normalizations of [*NuSTAR*]{} FPMB and [*Swift*]{}-XRT spectra were allowed to vary with respect to that of the [*NuSTAR*]{} FPMA spectrum.
The hard X-ray tail seen above 8keV in the PC mode observation (00706396000) may be caused by contamination due to short temporally unresolved X-ray bursts. Magnetar-like bursts are intrinsically harder than the average spectrum and the high count-rate leads to pile-up effects within the 2.5-s CCD readout time [e.g. @sk11]. These pile-up effects are not mitigated by standard techniques such as the removal of the central bright region, as they are bunched temporally rather than spatially. If the hard X-ray tail were indeed real, it would have needed to fade by $\sim2$ orders of magnitude in the day before the [*NuSTAR*]{} pointing to be consistent with the measured hard X-ray flux, while the blackbody temperatures measured on the two epochs are consistent with slow cooling. Hence, to fit the average spectra, we truncate the 00706396000 spectra above 7keV.
![X-ray spectra of [PSR J1119–6127]{}. The data are as follows: pre-outburst [*XMM-Newton*]{} spectrum (inverted black triangles), [*Swift*]{}-XRT spectrum at burst (red squares), [*Swift*]{}-XRT and [*NuSTAR*]{} spectrum one day after the burst (green and blue circles, respectively), and a [*Swift*]{}-XRT spectrum three days after the burst (purple triangles). The flux increase above 8keV in the [*Swift*]{}-XRT burst spectrum is likely due to pileup (see Section \[sec:spectroscopy\] for details). For clarity, the [*NuSTAR*]{} spectra from FPMA and FPMB are combined.[]{data-label="fig:1119_all_spec"}](1119_all_spec){width="48.00000%"}
Figure \[fig:1119\_all\_spec\] shows the current [*Swift*]{}+[*NuSTAR*]{} spectral fits in comparison with a pre-burst [*XMM-Newton*]{} spectrum [from @nkh+12]. Table \[tab:nustar\_spectra\] details the parameter values with 90% confidence error bars.
Immediately after the burst, we measure a blackbody temperature $kT=1.10(6)$keV, slightly decreasing to $0.96(1)$keV and $0.93(1)$keV in the follow-up spectra. This is substantially higher than the pre-burst blackbody temperature of 0.21(4)keV [@nkh+12]. In the [*NuSTAR*]{} spectra, we also measure a hard power law with photon index $\Gamma=1.2(2)$ that is marginally harder than the pre-burst value $\Gamma_\mathrm{PSR}=2.1(8)$ [@nkh+12]. We also note that @nkh+12 and @sk08 measured the power law emission from the pulsar wind nebula (PWN) around [PSR J1119–6127]{} to have $\Gamma_\mathrm{PWN} = 1.1-1.4$, close to the current hard power law index, but with a flux of $\sim2 \times 10^{-14}\,\mathrm{erg\,cm^{-2}\,s^{-1}}$, almost three orders of magnitude fainter.
[llc]{} `tbabs` & $N_\mathrm{H}$ ($10^{22}\,\mathrm{cm^{-2}}$) & $1.2\pm0.1$\
`bbody` & $kT_\mathrm{BB}$ (keV) & $1.10\pm0.06$\
$\mathrm{C-Stat}/\mathrm{dof}$ & & 333.82/413\
goodness & & $12\%$\
Flux (0.5–10keV) & & $4.1\pm0.1$\
$L_X$ (0.5–10keV) & & 3.5\
`const` & $C_{\mathrm{FPMB}}$ & $1.01\pm0.02$\
& $C_\mathrm{XRT}$ & $0.94\pm0.04$\
`bbody` & $kT_\mathrm{BB}$ (keV) & $0.96\pm0.01$\
`powerlaw` & $\Gamma$ & $1.2\pm0.2$\
$\mathrm{C-Stat}/\mathrm{dof}$ & & 2133.1/2327\
goodness & & $20\%$\
Flux (0.5–10keV) & & $2.7\pm0.1$\
$L_X$ (0.5–10keV) & & 2.3\
Flux (3–79keV) & & $1.9\pm0.1$\
$L_X$ (3–79keV) & & 1.6\
`bbody` & $kT_\mathrm{BB}$ (keV) & $0.93\pm0.06$\
$\mathrm{C-Stat}/\mathrm{dof}$ & & 304.7/355\
goodness & & $34\%$\
Flux (0.5–10keV) & & $2.1\pm0.2$\
$L_X$ (0.5–10keV) & & 1.8\
Discussion {#sec:discussion}
==========
The outburst from PSR J1119$-$6127 is observationally very similar to those seen in magnetars. The phenomenology of magnetar outbursts is rich [see @re11 for a review], but with established commonalities, practically all of which are observed in the PSR J1119$-$6127 event. A hallmark of magnetar outbursts are short-duration ($<1$-s) hard X-ray bursts, as reported for PSR J1119$-$6127 [@klm+16; @ykr16]. The large flux enhancement, here by a factor of $>160$, is commonly seen in magnetar outbursts, notably those in which the quiescent luminosity is below $\sim 10^{33}$ erg s$^{-1}$ [e.g. @sk11; @kkp+12]. Moreover, the spectral hardening we report is classic for magnetar outbursts [@re11], as are timing anomalies, most often spin-up glitches [@dkg08; @dk14]. Overall, the PSR J1119$-$6127 event is clearly magnetar-like.
Most similar to the PSR J1119$-$6127 event is the 2006 magnetar-like outburst of PSR J1846$-$0258 [@ggg+08]. The latter is also a young ($\tau < 1000$ yr), high-B ($B=5 \times 10^{13}$ G) rotation-powered pulsar, albeit radio undetected [@aklm08]. Next we compare this object’s 2006 outburst with the event studied here.
One difference between the outbursts is their energetics. For PSR J1119$-$6127, the 0.5–10-keV flux as measured in the joint [*Swift*]{} and [*NuSTAR*]{} observation was $4.1(1) \times 10^{-11}$ erg s$^{-1}$ cm$^{-2}$. For a distance of 8.4 kpc, and assuming isotropic emission, this implies a luminosity of $3.5 \times 10^{35}$ erg s$^{-1}$, or 0.1$\dot{E}$. This represents an increase over the quiescent value in this band of a factor of over $\sim160$. Including the flux from 10 keV extrapolated to the top of the [*NuSTAR*]{} band increases this value by $\sim$20%. The efficiency for conversion of $\dot{E}$ to [*Fermi*]{}-band $\gamma$-ray emission, at least in quiescence, was estimated by @pkh+11 to be 0.23. With a comparable amount of energy suddenly appearing in X-rays, the [*Fermi*]{}-band emission may have been affected during this outburst. However, the normally low [*Fermi*]{}/LAT count rate (see §\[sec:timing\]) requires multiple weeks of integration for a detection, hence a short-term anomaly may be undetectable.
By contrast, for PSR J1846$-$0258, in [*Chandra*]{} observations made within one week of its first detected magnetar-like burst, the unabsorbed flux in the 0.5–10-keV band was $4.0^{+1.6}_{-0.9} \times 10^{-11}$ erg s$^{-1}$ cm$^{-2}$ [@ggg+08]. For isotropic emission and a distance of 6 kpc [@lt08], this implies a luminosity of $1.7 \times 10^{35}$ erg s$^{-1}$, or 0.02$\dot{E}$. This represented an increase of a factor of $7.7^{+3.9}_{-1.8}$ over the quiescent value. As this outburst had no realtime trigger, it is possible that initially the pulsar brightened more. However the pulsed flux measured by [*RXTE*]{} at the initial burst epoch was comparable a week later. Hence the increase in flux was probably not much larger than a factor of $\sim$10, never exceeding a few percent of $\dot{E}$. Thus energetically, the PSR J1119$-$6127 outburst is far larger than that in PSR J1846$-$0258.
The spectral evolution in the two events also differed. In quiescence, PSR J1846$-$0258’s X-ray spectrum is well described by a simple power law of index 1.1(1) [@ggg+08], very different from the soft quiescent spectrum PSR J1119$-$6127. The latter, well described by a blackbody of $kT = 0.21(4)$ keV and power law of index 2.1(8) [@sk08; @nkh+12], made this pulsar the youngest with detected thermal emission. It was also one of the hottest and most luminous thermal emitters even among rotation-powered pulsars of similar age [@ozv+13]. This emission was also noted to be unusual for its high pulsed fraction [@nkh+12].
During outburst, PSR J1846$-$0258 developed a bright soft component, whereas the harder power-law spectrum remained unchanged, apart from an increase in normalization by $\sim$35% above 10 keV [@kh09]. The soft component was described as having a power-law spectrum of index 1.9(1) [@ks08b; @ggg+08]. By contrast, the spectrum of PSR J1119$-$6127 has undergone a radical hardening, with $kT$ increasing from 0.21(4)keV to 1.10(6)keV. Nevertheless, the net effect in both outbursts was a transition to a spectrum very similar to those of bright magnetars.
It is also interesting to compare the outburst timing anomalies. In PSR J1846$-$0258, it suffered a sudden spin-up having $\Delta\nu / \nu \simeq 3 \times 10^{-6}$, followed by a large increase in $\dot{\nu}$ yielding a strong over-recovery of the glitch [@lkg10; @kh09; @lnk+11]. The net long-term effect was a spin-[*down*]{}, accompanied by a change in braking index and a long-term enhancement in timing noise. While it is too early to know the post-outburst timing evolution in PSR J1119$-$6127, from our timing analysis, we find that the pulsar had a similar-sized spin-up glitch with $\Delta\nu / \nu \simeq 5.8 \times 10^{-6}$. Presently any increase in spin-down rate is modest compared to some glitch recoveries in magnetars and certainly compared to that following the 2006 PSR J1846$-$0258 glitch. However, greater evolution may yet be detected.
In young radio pulsars like PSR J1119$-$6127, hard X-ray emission is thought to arise in the context of outer gap models [e.g. @wtk13] from synchrotron radiation from secondary electron/positron pairs produced by inward propagating curvature radiation $\gamma$-rays. As discussed by @pkh+11, in PSR J1119$-$6127, the X-ray/$\gamma$-ray phase offset, together with the single-peak morphology of the $\gamma$-ray pulse, are well explained in outer gap models. The luminosity of both the X-ray and $\gamma$-ray emission in this picture must be bounded by the spin-down power.
The increase in X-ray luminosity particularly in the hard X-ray range during the outburst of PSR J1846$-$0258 was argued by @kh09 to be plausibly due to the above-described rotation-powered outer-gap emission, enhanced by particle injection due to perhaps to crust cracking that occurred at the glitch, reasonable given the lack of evidence for the hard X-ray luminosity exceeding more than a few percent of $\dot{E}$.
The new hard X-ray emission component in PSR J1119$-$6127 could have an outer-gap origin as well, but the large luminosity rise to within 0.1$\dot{E}$ in the X-ray band alone may be difficult to accommodate in such a picture, and might require a commensurate increase in $\gamma$-ray luminosity, impossible given the available $\dot{E}$ energy budget.
Alternatively, the hard X-ray emission may be magnetar-like. The origin of the bright hard X-ray component in magnetars has been argued to be a decelerating electron/positron flow in the closed magnetosphere, in the higher altitude regions of large magnetic loops [@bel13]. This emission is powered ultimately by the internal stellar magnetic field and is not limited by $\dot{E}$. In this interpretation, the pulsar suffered an instability such that a significant twist in its field lines occurred, with highly relativistic particles ($\gamma >> 10$) injected near the star where $B >> B_{QED} = 4.4 \times 10^{13}$ G. If this is origin of the hard X-rays in PSR J1119$-$6127, then the true internal field of this pulsar is far higher than is inferred from its dipole component. This would support the argument for additional non-dipolar field components in apparently low-magnetic-field magnetars [@ret+10; @skc14]. One way to test this explanation is through modelling of the phase-resolved hard X-ray spectrum. This can yield constraints on the geometry of the emission region [e.g. @hbd14]. Such constraints could then be compared with similar ones from radio polarimetry [@wje11] and/or modelling of the $\gamma$-ray light curve [@pkh+11].
Importantly, PSRs J1846$-$0258 and now J1119$-$6127 are the only two rotation-powered pulsars to have exhibited radiative changes at glitch epochs; this must be a consequence of their high spin-inferred $B$. Indeed no X-ray enhancement was seen in [*Chandra*]{} observations made 3.5 days following a large ($\Delta\nu/\nu = 3 \times 10^{-6}$) spin-up glitch in the lower-field ($B=3.4 \times 10^{12}$ G) Vela radio pulsar [@hgh01]. On the other hand, there was no evidence for an X-ray enhancement in PSR J1846$-$0258 near the epoch of a much smaller glitch having $\Delta\nu/\nu = 2.5 \times 10^{-9}$ [@lkgk06], nor in previous glitches ($\Delta\nu/\nu = 2.9 \times 10^{-7}$ and $4.1 \times 10^{-6}$ in 2004 and 2007, respectively) in PSR J1119$-$6127 [@wje11; @awe+15], although prompt X-ray observations were not performed in those cases. Moreover, multiple sizeable glitches in [*bona fide*]{} magnetars have been unaccompanied by radiative changes [@sak+14; @dk14]. This may indicate that an independent parameter such as the crustal depth of the glitch location plays a role in the radiative detectability of high-B neutron-stars at glitch epochs [see, e.g., @ec89a; @let02].
The possibility of a magnetar-like outburst from a high-B radio pulsar was discussed by @km05, who also suggested the possibility of radio emission from magnetars prior to its discovery by @crh+06. @pp11a and @pp11b provided theoretical groundwork for the hypothesis, and magnetothermal modelling such as that by @vrp+13 have further developed these ideas, which are now on solid observational ground. Other high-B radio pulsars like PSRs J1718$-$3718 [@zkm+11] and J1734$-$3333 [@ozv+13] seem likely to also undergo a magnetar-like transition in coming years.
[*Acknowledgements:*]{} The authors thank the operations teams of [*NuSTAR*]{}, particularly Karl Forster, and [*Swift*]{} for their speed and flexibility scheduling these observations. We thank [*Fermi*]{}-LAT Collaboration for the public data and tools used in this work. This work made use of data from the [*NuSTAR*]{} mission, a project led by the California Institute of Technology, managed by the Jet Propulsion Laboratory, and funded by the NASA. We acknowledge the use of public data from the [*Swift*]{} data archive. R.F.A. acknowledges support from an NSERC CGSD. V.M.K. receives support from an NSERC Discovery Grant, an Accelerator Supplement and from the Gerhard Herzberg Award, an R. Howard Webster Foundation Fellowship from the Canadian Institute for Advanced Study, the Canada Research Chairs Program, and the Lorne Trottier Chair in Astrophysics and Cosmology. S.P.T acknowledges support from a McGill Astrophysics postdoctoral fellowship. P.S. acknowledges support from a Schulich Graduate Fellowship.
natexlab\#1[\#1]{}
, D., [Vasilopoulos]{}, G., & Espinoza, C. M. 2016, ATel, 9282
, D., [Weltevrede]{}, P., [Espinoza]{}, C. M., [et al.]{} 2015, , 447, 3924
, A. M., [Kaspi]{}, V. M., [Livingstone]{}, M. A., & [McLaughlin]{}, M. A. 2008, , 688, 550
Arnaud, K. A. 1996, in Astronomical Data Analaysis Software and Systems V, ed. G. Jacoby & J. Barnes, Vol. 101 (San Fransisco: ASP), 17
, W. B., [Abdo]{}, A. A., [Ackermann]{}, M., [et al.]{} 2009, ApJ, 697, 1071
, A. M. 2013, , 762, 13
, M., [Possenti]{}, A., [Kerr]{}, M., [et al.]{} 2016, ATel, 9286
, D. N., [Hill]{}, J. E., [Nousek]{}, J. A., [et al.]{} 2005, Space Sci. Rev., 120, 165
Camilo, F., Kaspi, V. M., Lyne, A. G., [et al.]{} 2000, ApJ, 541, 367
, F., [Ransom]{}, S., [Halpern]{}, J., [et al.]{} 2006, Nature, 442, 892
, W. 1979, ApJ, 228, 939
, J. L., [McClure-Griffiths]{}, N. M., & [Cheung]{}, M. C. M. 2004, MNRAS, 352, 1405
Crawford, F., Gaensler, B. M., Kaspi, V. M., [et al.]{} 2001, ApJ, 554, 152
, R., & [Kaspi]{}, V. M. 2014, , 784, 37
, R., [Kaspi]{}, V. M., & [Gavriil]{}, F. P. 2008, ApJ, 673, 1044
, D., & [Cheng]{}, A. F. 1989, , 336, 360
, F. P., [Gonzalez]{}, M. E., [Gotthelf]{}, E. V., [et al.]{} 2008, Science, 319, 1802
Gonzalez, M., & Safi-Harb, S. 2003, ApJ, 591, L143
, F. A., [Craig]{}, W. W., [Christensen]{}, F. E., [et al.]{} 2013, ApJ, 770, 103
, R., [Beloborodov]{}, A. M., & [den Hartog]{}, P. R. 2014, , 786, L1
, D. J., [Gotthelf]{}, E. V., & [Halpern]{}, J. P. 2001, ApJ, 556, 380
Hobbs, G. M., Edwards, R. T., & Manchester, R. N. 2006, MNRAS, 369, 655
, O., [Kouveliotou]{}, C., [Pavlov]{}, G. G., [et al.]{} 2012, ApJ, 748, 26
, V. M., & [McLaughlin]{}, M. A. 2005, ApJ, 618
Kennea, J. A., Lien, A. Y., Marshall, F. E., [et al.]{} 2016, GCN, 19735
, L., & [Hermsen]{}, W. 2009, , 501, 1031
, H. S., & [Safi-Harb]{}, S. 2008, ApJ, 678, L43
, D. A., & [Tian]{}, W. W. 2008, , 480, L25
, M. A., [Kaspi]{}, V. M., & [Gavriil]{}, F. P. 2010, ApJ, 710, 1710
, M. A., [Kaspi]{}, V. M., [Gotthelf]{}, E. V., & [Kuiper]{}, L. 2006, ApJ, 647, 1286
, M. A., [Ng]{}, C.-Y., [Kaspi]{}, V. M., [Gavriil]{}, F. P., & [Gotthelf]{}, E. V. 2011, , 730, 66
, M. A., [Ransom]{}, S. M., [Camilo]{}, F., [et al.]{} 2009, ApJ, 706, 1163
, Y., [Eichler]{}, D., & [Thompson]{}, C. 2002, ApJ, 580, L69
, C.-Y., [Kaspi]{}, V. M., [Ho]{}, W. C. G., [et al.]{} 2012, , 761, 65
, S. A., [Zhu]{}, W. W., [Vogel]{}, J. K., [et al.]{} 2013, , 764, 1
, D., [Kerr]{}, M., [den Hartog]{}, P. R., [et al.]{} 2011, ApJ, 743
, R., & [Pons]{}, J. A. 2011, , 727, L51
, J. A., & [Perna]{}, R. 2011, , 741, 123
, P. S., [Kerr]{}, M., [Parent]{}, D., [et al.]{} 2011, ApJS, 194, 17
, N., & [Esposito]{}, P. 2011, Astrophysics and Space Science Proceedings, 21, 247
, N., [Esposito]{}, P., [Turolla]{}, R., [et al.]{} 2010, Science, 330, 944
, S., & [Kumar]{}, H. S. 2008, , 684, 532
, P., [Archibald]{}, R. F., [Kaspi]{}, V. M., [et al.]{} 2014, , 783, 99
, P., & [Kaspi]{}, V. M. 2011, ApJ, 739, 94
, P., [Kaspi]{}, V. M., & [Cumming]{}, A. 2014, , 786, 62
, D., [Rea]{}, N., [Pons]{}, J. A., [et al.]{} 2013, , 434, 123
, Y., [Takata]{}, J., & [Cheng]{}, K. S. 2013, , 764, 51
, P., [Johnston]{}, S., & [Espinoza]{}, C. M. 2011, , 411, 1917
Younes, G., Kouveliotou, C., & Roberts, O. 2016, GCN, 19736
, W. W., [Kaspi]{}, V. M., [McLaughlin]{}, M. A., [et al.]{} 2011, , 734, 44
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we study planar morphs between straight-line planar grid drawings of trees. A morph consists of a sequence of morphing steps, where in a morphing step vertices move along straight-line trajectories at constant speed. We show how to construct planar morphs that simultaneously achieve a reduced number of morphing steps and a polynomially-bounded resolution. We assume that both the initial and final drawings lie on the grid and we ensure that each morphing step produces a grid drawing; further, we consider both upward drawings of rooted trees and drawings of arbitrary trees.'
author:
- 'Fidel [Barrera-Cruz]{}$^1$, Manuel Borrazzo$^2$, Giordano [Da Lozzo]{}$^2$, Giuseppe $^2$, Fabrizio Frati$^2$, Maurizio Patrignani$^2$, and [Vincenzo Roselli]{}$^2$ [^1]'
bibliography:
- 'bibliography.bib'
title: How to Morph a Tree on a Small Grid
---
Introduction {#se:introduction}
============
The problem of morphing combinatorial structures is a consolidated research topic with important applications in several areas of Computer Science such as Computational Geometry, Computer Graphics, Modeling, and Animation. The structures of interest typically are drawings of graphs; a *morph* between two drawings $\Gamma_0$ and $\Gamma_1$ of the same graph $G$ is defined as a continuously changing family of drawings $\{\Gamma_t\}$ of $G$ indexed by time $t \in [0,1]$, such that the drawing at time $t=0$ is $\Gamma_0$ and the drawing at time $t=1$ is $\Gamma_1$. A morph is usually required to preserve a certain drawing standard and pursues certain qualities.
The *drawing standard* is the set of the geometric properties that are maintained at any time during the morph. For example, if both $\Gamma_0$ and $\Gamma_1$ are planar drawings, then the drawing standard might require that all the drawings of the morph are planar. Other properties that might required to be preserved are the convexity of the faces, or the fact that the edges are straight-line segments, or polylines, etc.
Regarding the *qualities* of the morph, the research up to now mainly focused on limiting the number of *morphing steps*, where in a morphing step vertices move along straight-line trajectories at constant speed. A morph $\mathcal{M}$ can then be described as a sequence of drawings $\mathcal M=\langle \Gamma_0=\Delta_0,\Delta_1,\dots,\Delta_{k}=\Gamma_1\rangle$ where the morph $\langle \Delta_{i-1}, \Delta_i \rangle$, for $i=1, \dots, k$, is a morphing step. Following the pioneeristic works of Cairns and Thomassen [@c-dprc-44; @t-dpg-83], most of the literature focused on the straight-line planar drawing standard. A sequence of recent results in [@DBLP:journals/siamcomp/AlamdariABCLBFH17; @DBLP:conf/soda/AlamdariACBFLPRSW13; @DBLP:conf/icalp/AngeliniLBFPR14; @DBLP:conf/compgeom/AngeliniLFLPR15; @DBLP:conf/gd/AngeliniFPR13] proved that a linear number of morphing steps suffices, and is sometimes necessary, to construct a morph between any two straight-line planar drawings of a graph.
Although the results mentioned in the previous paragraph establish strong theoretical foundations for the topic of morphing graph drawings, they produce morphs that are not appealing from a visualization perspective. Namely, such algorithms produce drawings that have poor *resolution*, i.e., they may have an exponential ratio of the distances between the farthest and closest pairs of geometric objects (points representing nodes or segments representing edges), even if the same ratio is polynomially bounded in the initial and final drawings. Indeed, most of the above cited papers mention the problem of constructing morphs with bounded resolution as the main challenge in this research area.
The only paper we are aware of where the resolution problem has been successfully addressed is the one by Barrera-Cruz et al. [@DBLP:conf/gd/Barrera-CruzHL14], who showed how to construct a morph with polynomially-bounded resolution between two *Schnyder drawings* $\Gamma_0$ and $\Gamma_1$ of the same planar triangulation. The model they use in order to ensure a bound on the resolution requires that $\Gamma_0=\Delta_0,\Delta_1,\dots,\Delta_{k}=\Gamma_1$ are *grid drawings*, i.e., vertices have integer coordinates, and the resolution is measured by comparing the area of $\Gamma_0$ and $\Gamma_1$ with the area of the $\Delta_i$’s. We remark that morphs between planar orthogonal drawings of maximum-degree-$4$ planar graphs, like those in [@DBLP:journals/talg/BiedlLPS13; @DBLP:conf/compgeom/GoethemV18], inherently have polynomial resolution.
In this paper we show how to construct morphs of tree drawings that simultaneously achieve a reduced number of morphing steps and a polynomially-bounded resolution. Adopting the setting of [@DBLP:conf/gd/Barrera-CruzHL14], we assume that $\Gamma_0$ and $\Gamma_1$ are grid drawings and we ensure that each morphing step produces a grid drawing.
We present three algorithms. The first two algorithms construct morphs between any two strictly-upward straight-line planar grid drawings $\Gamma_0$ and $\Gamma_1$ of $n$-node rooted trees; *strictly-upward* drawings are such that each node lies above its children. Both algorithms construct morphs in which each intermediate grid drawing has linear width and height, where the input size is measured by $n$ and by the width and the height of $\Gamma_0$ and $\Gamma_1$. The first algorithm employs $\Theta(n)$ morphing steps. The second algorithm employs $\Theta(1)$ morphing steps, however it only applies to binary trees. The third algorithm allows us to achieve our main result, namely that for any two straight-line planar grid drawings $\Gamma_0$ and $\Gamma_1$ of an $n$-node tree, there is a planar morph with $\Theta(n)$ morphing steps between $\Gamma_0$ and $\Gamma_1$ such that each intermediate grid drawing has polynomial area, where the input size is again measured by $n$ and by the width and the height of $\Gamma_0$ and $\Gamma_1$.
The first algorithm uses recursion; namely, it eliminates a leaf in the tree, it recursively morphs the drawings of the remaining tree and it then reintroduces the removed leaf in suitable positions during the morph. The second algorithm morphs the given drawings by independently changing their $x$- and $y$-coordinates; this technique is reminiscent of a recent paper by Da Lozzo et al. [@DBLP:conf/gd/LozzoBFPR18]. Finally, the third algorithm scales the given drawings up in order to make room for a bottom-up modification of each drawing into a “canonical” drawing of the tree.
We remark that, although tree drawing algorithms are well investigated in Graph Drawing, morphs of tree drawings have not been the subject of research until now, with the exception of the recent work by Arseneva et al. [@DBLP:conf/gd/ArsenevaBCDDFLT18], who showed how to construct a three-dimensional crossing-free morph between two straight-line planar drawings of an $n$-node tree in $O(\log n)$ morphing steps.
The rest of the paper is organized as follows. In \[se:preliminaries\] we present some definitions and preliminaries. In \[se:upward\] we present our results on small-area upward planar morphs between strictly-upward straight-line planar grid drawings of rooted trees. In \[se:general\] we present our main result on small-area planar morphs between straight-line planar grid drawings of trees. Finally, in \[se:conclusions\] we conclude, present some open problems, and argue about the generality of the model adopted in this paper. Namely, we prove that the problem of constructing a planar morph with polynomial resolution between two planar straight-line drawings of the same graph can be reduced to the problem of constructing a planar morph with polynomial area between two planar straight-line grid drawings of the same graph.
Preliminaries {#se:preliminaries}
=============
In this section we introduce some definitions and preliminaries; see also [@DETT].
#### Trees.
The node and edge sets of a tree $T$ are denoted by $V(T)$ and $E(T)$, respectively. The *degree* $\deg(v)$ of a node $v$ of $T$ is the number of its neighbors. In an *ordered* tree, a counter-clockwise order of the edges incident to each node is specified.
A *rooted tree* $T$ is a tree with one distinguished node, which is called *root* and is denoted by $r(T)$. For any node $u \in V(T)$ with $u\neq r(T)$, consider the unique path from $u$ to $r(T)$ in $T$; the *ancestors* of $u$ are the nodes of such a path, the *proper ancestors* of $u$ are the ancestors different from $u$ itself, and the *parent* $p(u)$ of $u$ is the proper ancestor of $u$ which is adjacent to $u$. For any two nodes $u,v\in V(T)$, the *lowest common ancestor* is the ancestor of $u$ and $v$ whose graph-theoretic distance from $r(T)$ is maximum. For any node $u\in V(T)$ with $u\neq r(T)$, the *children* of $u$ are the neighbors of $u$ different from $p(u)$; the *children* of $r(T)$ are all its neighbors. The nodes that have children are called *internal*; a non-internal node is a *leaf*. For any node $u\in V(T)$ with $u\neq r(T)$, the *subtree* $T_u$ of $T$ rooted at $u$ is defined as follows: remove from $T$ the edge $(u,p(u))$, thus separating $T$ in two trees; the one containing $u$ is the subtree of $T$ rooted at $u$. If each node of $T$ has at most two children, then $T$ is a *binary tree*.
An *ordered rooted tree* is a tree that is rooted and ordered. In an ordered rooted tree $T$, for each node $u\in V(T)$, a *left-to-right* (linear) order $u_1,\dots,u_k$ of the children of $u$ is specified. If $T$ is binary then the first (second) child in the left-to-right order of the children of any node $u$ is the *left* (*right*) *child* of $u$, and the subtree rooted at the left (right) child of $u$ is the *left* (*right*) *subtree* of $u$.
#### Tree drawings.
In a *straight-line drawing* $\Gamma$ of a tree $T$ each node $u$ is represented by a point of the plane (whose coordinates are denoted by $x_{\Gamma}(u)$ and $y_{\Gamma}(u)$) and each edge is represented by a straight-line segment between its end-points. All the drawings considered in this paper are straight-line, even when not specified. In a *planar* drawing no two edges intersect except, possibly, at common end-points. For a rooted tree $T$, a *strictly-upward* drawing $\Gamma$ is such that each edge $(u,p(u))\in E(T)$ is represented by a curve monotonically increasing in the $y$-direction from $u$ to $p(u)$; if $\Gamma$ is a straight-line drawing, this is equivalent to requiring that $y_{\Gamma}(u)<y_{\Gamma}(p(u))$. For an ordered tree $T$, an *order-preserving* drawing $\Gamma$ is such that, for each node $u\in V(T)$, the counter-clockwise order of the edges incident to $u$ in $\Gamma$ is the same as the order associated to $u$ in $T$. Note that a strictly-upward drawing $\Gamma$ of an ordered rooted tree $T$ is order-preserving if and only if, for each node $u\in V(T)$, the edges from $u$ to its children enter $u$ in the left-to-right order associated with $u$.
The *bounding box* of a drawing $\Gamma$ is the smallest axis-parallel rectangle enclosing $\Gamma$. In a *grid* drawing $\Gamma$ each node has integer coordinates; then the *width* and the *height* of $\Gamma$, denoted by $w(\Gamma)$ and $h(\Gamma)$, respectively, are the number of grid columns and rows intersecting the bounding box of $\Gamma$, while the *area* of $\Gamma$ is its width times its height. For a node $v$ in a drawing $\Gamma$, an *$\ell$-box centered at $v$* is the convex hull of the square whose corners are $(x_\Gamma(v) \pm \frac{\ell}{2}, y_\Gamma(v) \pm \frac{\ell}{2})$.
#### Morphs.
A *morph* between two straight-line drawings $\Gamma_0$ and $\Gamma_1$ of a graph $G$ is a continuously changing family of drawings $\{\Gamma_t\}$ of $G$ indexed by time $t\in [0,1]$, such that the drawing at time $t=0$ is $\Gamma_0$ and the drawing at time $t=1$ is $\Gamma_1$. A morph is *planar* if all its intermediate drawings are planar. A morph between two strictly-upward drawings of a rooted tree is *upward* if all its intermediate drawings are strictly-upward. A morph is *linear* if each node moves along a straight-line trajectory at constant speed. Whenever the linear morph between two straight-line planar drawings $\Gamma_0$ and $\Gamma_1$ of a graph $G$ is not planar, one is usually interested in the construction of a piecewise-linear morph with small complexity between $\Gamma_0$ and $\Gamma_1$. This is formalized by defining a *morph* between $\Gamma_0$ and $\Gamma_1$ as a sequence $\langle \Gamma_0=\Delta_0,\Delta_1,\dots,\Delta_k=\Gamma_1 \rangle$ of drawings of $G$ such that the linear morph $\langle \Delta_{i-1},\Delta_{i} \rangle$ is planar, for $i=1,\dots,k$; each linear morph $\langle \Delta_{i-1},\Delta_{i} \rangle$ is called a *morphing step* or simply a *step*.
The *width* $w(\mathcal M)$ of a morph $\mathcal M=\langle \Delta_0,\Delta_1,\dots,\Delta_k \rangle$, where $\Delta_i$ is a grid drawing, for $i=0,1,\dots,k$, is equal to $\max \{w(\Delta_0),w(\Delta_1),\dots,w(\Delta_k)\}$. The *height* $h(\mathcal M)$ of $\mathcal M$ and the area of $\mathcal M$ are defined analogously.
The algorithms we design in this paper receive in input two order-preserving straight-line planar grid drawings $\Gamma_0$ and $\Gamma_1$ of an ordered tree and construct morphs $\langle \Gamma_0=\Delta_0,\Delta_1,\dots,\Delta_k=\Gamma_1 \rangle$ with few steps and small area.
A necessary and sufficient condition for the existence of a planar morph between two straight-line planar drawings $\Gamma_0$ and $\Gamma_1$ of a tree $T$ is that they are “topologically-equivalent”, i.e., the counter-clockwise order of the edges incident to each node $u\in V(T)$ is the same in $\Gamma_0$ and $\Gamma_1$. In order to better exploit standard terminology about tree drawings, we ensure that $\Gamma_0$ and $\Gamma_1$ are topologically-equivalent by assuming that $T$ is ordered and that $\Gamma_0$ and $\Gamma_1$ are order-preserving drawings; hence, dealing with ordered trees and with order-preserving drawings is not a loss of generality.
The width and the height of the morphs we construct are expressed not only in terms of the number of nodes of the input tree $T$, but also in terms of the width and the height of the input drawings $\Gamma_0$ and $\Gamma_1$ of $T$; this is necessary, given that $\max\{w(\Gamma_0),w(\Gamma_1)\}$ and $\max\{h(\Gamma_0),h(\Gamma_1)\}$ are obvious lower bounds for the width and the height of any morph between $\Gamma_0$ and $\Gamma_1$, respectively.
The morphs $\langle \Delta_0,\Delta_1,\dots,\Delta_k \rangle$ we construct in this paper are such that $\Delta_0,\Delta_1,\dots,\Delta_k$ are *grid* drawings, even when not explicitly specified.
In the following we introduce two tools we are going to use later. First, we observe that whether a linear morph is upward only depends on the upwardness of the initial and final drawings of the morph.
\[obs:upward-stays-upward\] Let $\Gamma_0$ and $\Gamma_1$ be two strictly-upward straight-line drawings of a rooted tree $T$. Then the linear morph $\langle \Gamma_0,\Gamma_1 \rangle$ is upward.
Assume that the morph $\langle \Gamma_0,\Gamma_1 \rangle$ happens between the time instants $t=0$ and $t=1$. For any $t\in [0,1]$, denote by $\Gamma_t$ the drawing of $T$ in $\langle \Gamma_0,\Gamma_1 \rangle$ at time $t$.
Consider any edge $(p(v),v)$ of $T$. Since $\Gamma_0$ and $\Gamma_1$ are strictly-upward, it follows that $y_{\Gamma_0}(p(v))>y_{\Gamma_0}(v)$ and $y_{\Gamma_1}(p(v))>y_{\Gamma_1}(v)$. Hence, at any time instant $t\in [0,1]$ of the morph $\langle \Gamma_0,\Gamma_1 \rangle$, we have $y_{\Gamma_t}(p(v))=(1-t)\cdot y_{\Gamma_0}(p(v)) + t \cdot y_{\Gamma_1}(p(v)) > (1-t)\cdot y_{\Gamma_0}(v) + t \cdot y_{\Gamma_1}(v) = y_{\Gamma_t}(v)$. It follows that the drawing $\Gamma_t$ is strictly-upward, and hence that $\langle \Gamma_0,\Gamma_1 \rangle$ is an upward morph.
The planarity of a morph can not be ensured as simply as its upwardness. However, if the initial and final drawings of the morph satisfy some further conditions, it turns out that planarity is actually guaranteed; this is similar to a lemma by Da Lozzo [*et al.*]{} [@DBLP:conf/gd/LozzoBFPR18].
\[le:unidirectional-morph\] Let $\Gamma_0$ and $\Gamma_1$ be two order-preserving strictly-upward straight-line planar drawings of a rooted ordered tree $T$. Suppose that, for each node $v\in V(T)$, we have $y_{\Gamma_0}(v)=y_{\Gamma_1}(v)$. Then the linear morph $\langle \Gamma_0,\Gamma_1 \rangle$ is planar.
The proof exploits Corollary 7.2 in [@DBLP:journals/siamcomp/AlamdariABCLBFH17], which we introduce in the following.
Assume that the morph $\langle \Gamma_0,\Gamma_1 \rangle$ happens between the time instants $t=0$ and $t=1$. For any $t\in [0,1]$, denote by $\Gamma_t$ the drawing of $T$ in $\langle \Gamma_0,\Gamma_1 \rangle$ at time $t$ and denote by $u_t$ the position of a node $u$ at time $t$, where $u_t=(1-t) \cdot u_0 + t \cdot u_1$.
Now, consider a point $q_0$ *of* $\Gamma_0$, that is a point that represents a node of $T$ or that belongs to a segment representing an edge of $T$ in $\Gamma_0$. The morph $\langle \Gamma_0,\Gamma_1 \rangle$ moves $q_0$ to a point $q_1$ of $\Gamma_1$. This is evident if $q_0$ represents a node of $T$ in $\Gamma_0$; if $q_0$ is a point of the segment representing an edge $(u,v)$ in $\Gamma_0$, then $q_0=(1-\gamma) \cdot u_0+ \gamma \cdot v_0$, for some value $0<\gamma<1$, and the point $q_1=(1-\gamma) \cdot u_1 + \gamma \cdot v_1$ indeed belongs to the segment representing $(u,v)$ in $\Gamma_1$. More in general, for any $t\in [0,1]$, the point $q_t=(1-t) \cdot q_0 + t \cdot q_1$ belongs to the segment representing $(u,v)$ in $\Gamma_t$.
Now consider any three points $q_0$, $r_0$, and $s_0$ of $\Gamma_0$, which are moved to three points $q_1$, $r_1$, and $s_1$ of $\Gamma_1$ by $\langle \Gamma_0,\Gamma_1 \rangle$, respectively. Suppose that $q_i$ is to the left (to the right) of the line through $r_i$ and $s_i$, for each $i=0,1$. Then Corollary 7.2 in [@DBLP:journals/siamcomp/AlamdariABCLBFH17] ensures that $q_t$ is to the left (resp. to the right) of the line through $r_t$ and $s_t$, for any $t\in [0,1]$. The applicability of Corollary 7.2 from [@DBLP:journals/siamcomp/AlamdariABCLBFH17] to the morph $\langle \Gamma_0,\Gamma_1 \rangle$ is a consequence of the fact that $\langle \Gamma_0,\Gamma_1 \rangle$ is *unidirectional*, that is, all the node trajectories are parallel (indeed, each node $v\in V(T)$ moves along a horizontal line, given that $y_{\Gamma_0}(v)=y_{\Gamma_1}(v)$).
We prove the planarity of $\langle \Gamma_0,\Gamma_1 \rangle$. Suppose, for a contradiction, that two edges $(u,p(u))$ and $(v,p(v))$ of $T$ cross during $\langle \Gamma_0,\Gamma_1 \rangle$. For the sake of the simplicity of notation, let $u'=p(u)$ and $v'=p(v)$. Then there exists a point $q_0$ that belongs the segment $\overline{u_0u'_0}$ and such that, for some $t\in (0,1)$, the point $q_{t}=(1-{t}) \cdot q_0 + {t} \cdot q_1$ is a point of the segment $\overline{u_{t}u'_{t}}$ and also a point of the segment $\overline{v_{t} v'_{t}}$. Since every node of $T$ only moves horizontally in $\langle \Gamma_0,\Gamma_1 \rangle$, we have that $q_0$ ($q_1$) lies in the strip delimited by the horizontal lines through $v_0$ and $v'_0$ (resp. through $v_1$ and $v'_1$). Since $\Gamma_0$ and $\Gamma_1$ are planar, $q_0$ and $q_1$ do not belong to the segments $\overline{v_0v'_0}$ and $\overline{v_1v'_1}$, respectively. It follows that $q_0$ is either to the left or to the right of the line $\ell_0^v$ that passes through $v_0$ and $v'_0$; analogously, $q_1$ is either to the left or to the right of the line $\ell_1^v$ that passes through $v_1$ and $v'_1$. Assume that $q_0$ is to the left of $\ell_0^v$; we claim that this implies that $q_1$ is to the left of $\ell_1^v$. The claim follows from the fact that, both in $\Gamma_0$ and in $\Gamma_1$, the two paths of $T$ from $u$ and $v$ to their lowest common ancestor $w$ are monotone in the $y$-direction (since $\Gamma_0$ and $\Gamma_1$ are strictly-upward), do not cross each other (since $\Gamma_0$ and $\Gamma_1$ are planar), and enter $w$ in the same left-to-right order (since $\Gamma_0$ and $\Gamma_1$ are order-preserving). Now since $q_0$ is to the left of $\ell_0^v$ and $q_1$ is to the left of $\ell_1^v$, Corollary 7.2 from [@DBLP:journals/siamcomp/AlamdariABCLBFH17] implies that $q_{t}$ is to the left of the line that passes through $v_{t}$ and $v'_{t}$, while $q_{t}$ is supposed to be a point of $\overline{v_{t} v'_{t}}$. This contradiction proves the lemma.
Upward planar morphs between strictly-upward drawings of rooted ordered trees maintain the drawing order-preserving at all times, as proved in the following.
\[le:upwardplanar-implies-order\] Let $\Gamma_0$ and $\Gamma_1$ be two order-preserving strictly-upward straight-line planar drawings of a rooted ordered tree $T$. Let $\mathcal M$ be any upward planar morph between $\Gamma_0$ and $\Gamma_1$. Then any intermediate drawing of $\mathcal M$ is order-preserving.
Assume that the morph $\mathcal M$ happens between the time instants $t=0$ and $t=1$. For any $t\in [0,1]$, denote by $\Gamma_t$ the drawing of $T$ in $\mathcal M$ at time $t$. Since $\mathcal M$ is upward, the drawing $\Gamma_t$ is strictly-upward, for any $t\in [0,1]$. Hence, it suffices to prove that, for any internal node $v$ of $T$, the edges from $v$ to its children enter $v$ in $\Gamma_t$ in the left-to-right order associated with $v$; this is indeed true for $t=0$ and $t=1$. Suppose that, for some $t\in (0,1)$, the left-to-right order in which two edges $(u_1,v)$ and $(u_2,v)$ enter $v$ in $\Gamma_t$ is different than in $\Gamma_0$. Since such edges are represented by curves monotonically increasing in the $y$-direction from $u_1$ and $u_2$ to $v$ throughout $\mathcal M$, it follows that there is a time $t^*\in (0,t)$ such that the edges $(u_1,v)$ and $(u_2,v)$ overlap in $\Gamma_{t^*}$. However, this cannot happen due to the planarity of $\mathcal M$.
Upward Planar Morphs of Rooted-Tree Drawings {#se:upward}
============================================
In this section we show how to construct small-area morphs between order-preserving strictly-upward straight-line planar grid drawings of rooted ordered trees.
Our first result shows that such morphs can always be constructed consisting of a linear number of steps. This is obtained via an inductive algorithm which is described in the following. Let $T$ be an $n$-node rooted ordered tree. The *rightmost path* of $T$ is the maximal path $(s_0,\dots,s_m)$ such that $s_0=r(T)$ and $s_{i}$ is the rightmost child of $s_{i-1}$, for $i=1,\dots,m$. Note that $s_m$ is a leaf, which is called the *rightmost leaf* $l^{\rightarrow}_T$ of $T$. For a straight-line grid drawing $\Gamma$, denote by $\ell_{\Gamma}$ the rightmost vertical line intersecting $\Gamma$; note that $\ell_{\Gamma}$ is a grid column.
Let $\Gamma_0$ and $\Gamma_1$ be two [order-preserving strictly-upward straight-line planar grid]{}drawings of $T$. We inductively construct a morph $\mathcal M$ from $\Gamma_0$ to $\Gamma_1$ as follows.
In the base case $n=1$; then $\mathcal M$ is the linear morph $\langle \Gamma_0,\Gamma_1\rangle$.
In the inductive case $n>1$. Let $l=l^{\rightarrow}_T$ be the rightmost leaf of $T$. Let $\pi=p(l)$ be the parent of $l$. Let $T'$ be the $(n-1)$-node tree obtained from $T$ by removing the node $l$ and the edge $(\pi,l)$. Let $\Gamma'_0$ and $\Gamma'_1$ be the drawings of $T'$ obtained from $\Gamma_0$ and $\Gamma_1$, respectively, by removing the node $l$ and the edge $(\pi,l)$. Inductively compute a $k$-step upward planar morph $\mathcal M'=\langle \Gamma'_0=\Delta'_1,\Delta'_2,\dots,\Delta'_{k}=\Gamma'_1\rangle$.
We now construct a morph $\mathcal M=\langle \Gamma_0,\Delta_1,\Delta_2,\dots,\Delta_k,\Gamma_1\rangle$. For each $i=2,3,\dots,k-1$, we define $\Delta_i$ as the drawing obtained from $\Delta'_i$ by placing $l$ one unit below $\pi$ and one unit to the right of $\ell_{\Delta'_i}$. Further, we define $\Delta_1$ ($\Delta_{k}$) as the drawing obtained from $\Delta'_1$ (resp. from $\Delta'_k$) by placing $l$ one unit below $\pi$ and one unit to the right of $\ell_{\Gamma_0}$ (resp. $\ell_{\Gamma_1}$). Note that the point at which $l$ is placed in $\Delta_1$ (in $\Delta_k$) is one unit to the right of $\ell_{\Delta'_1}$ (resp. $\ell_{\Delta'_k}$), similarly as in $\Delta_2, \Delta_3, \dots, \Delta_{k-1}$, except if $l$ is to the right of every other node of $\Gamma_0$ (of $\Gamma_1$); in that case $l$ might be several units to the right of $\ell_{\Delta'_1}$ (resp. $\ell_{\Delta'_k}$). This completes the construction of $\mathcal M$. We get the following.
\[th:nary-trees-linearsteps-polyarea\] Let $T$ be an $n$-node rooted ordered tree, and let $\Gamma_0$ and $\Gamma_1$ be two [order-preserving strictly-upward straight-line planar grid]{}drawings of $T$. There exists a $(2n-1)$-step upward planar morph $\mathcal M$ from $\Gamma_0$ to $\Gamma_1$ with $h(\mathcal M) = \max\{h(\Gamma_0),h(\Gamma_1)\}$ and $w(\mathcal M) = \max\{w(\Gamma_0),w(\Gamma_1)\} +n-1$.
For a drawing $\Gamma$ of $T$ and for any two nodes $s$ and $t$ of $T$, denote by $d^v_\Gamma(s,t)$ the vertical distance between $s$ and $t$ in $\Gamma$ (that is, the absolute value of the difference between their $y$-coordinates).
We claim that the morph $\mathcal M=\langle \Gamma_0,\Delta_1,\Delta_2,\dots,\Delta_k,\Gamma_1\rangle$ defined before the statement of the theorem satisfies the requirements of the theorem and also satisfies the property that $d^v_{\Delta_i}(u,r(T))\leq \max \{d^v_{\Gamma_0}(u,r(T)),d^v_{\Gamma_1}(u,r(T))\}$, for any $i=1,2,\dots,k$ and for each node $u$ of $T$.
First, note that $\mathcal M$ has $k+1=2n-1$ steps. This is trivially true if $n=1$ and it is true by induction if $n>1$ since $\mathcal M$ has $2$ steps more than $\mathcal M'$.
For each $i=1,2,\dots,k$, the drawing $\Delta_i$ is straight-line by construction; further, $\Delta_i$ is strictly-upward since $\Delta'_i$ is strictly-upward and since $l$ lies one unit below $\pi$ in $\Delta_i$. By \[obs:upward-stays-upward\], the morph $\mathcal M$ is upward. Since $\ell_{\Gamma_0}, \ell_{\Delta'_2},\ell_{\Delta'_3},\dots,\ell_{\Delta'_{k-1}},\ell_{\Gamma_1}$ are grid columns and since $\pi$ is placed at a grid point in each of $\Delta'_1,\Delta'_2,\dots,\Delta'_k$, we have that $l$ is placed at a grid point in each of $\Delta_1,\Delta_2,\dots,\Delta_k$, hence each of such drawings is a grid drawing.
We now analyze $w(\mathcal M)$. If $n=1$, then $\mathcal M=\langle \Gamma_0,\Gamma_1\rangle$, hence $w(\mathcal M)=\max\{w(\Gamma_0),w(\Gamma_1)\} +n-1$. Assume next that $n\geq 2$. Consider any $i\in \{2,3,\dots,k-1\}$. By construction $\Delta_i$ occupies one more grid column than $\Delta'_i$. By induction $w(\Delta'_i)\leq \max\{w(\Gamma'_0),w(\Gamma'_1)\} +n-2$; further, $w(\Gamma'_0)\leq w(\Gamma_0)$ and $w(\Gamma'_1)\leq w(\Gamma_1)$. Hence, $w(\Delta_i)\leq \max\{w(\Gamma_0),w(\Gamma_1)\} +n-1$, as required. In order to bound $w(\Delta_1)$, the argument is the same as the one above if $\ell_{\Gamma_0}$ contains a node of $\Delta'_1$. Otherwise, $\ell_{\Gamma_0}$ contains $l$ and no other node of $\Gamma_0$; then, by construction, we have $w(\Delta_1)= w(\Gamma_0)+1\leq \max\{w(\Gamma_0),w(\Gamma_1)\} +n-1$, as required. The proof that $w(\Delta_k)\leq \max\{w(\Gamma_0),w(\Gamma_1)\} +n-1$ is analogous.
Next, consider any index $i\in \{1,2,\dots,k\}$ and any node $u$ of $T$. We prove that $d^v_{\Delta_i}(u,r(T))\leq \max \{d^v_{\Gamma_0}(u,r(T)),d^v_{\Gamma_1}(u,r(T))\}$.
- If $u\neq l$, then we have $d^v_{\Delta_i}(u,r(T))= d^v_{\Delta'_i}(u,r(T'))$. By induction, we have $d^v_{\Delta'_i}(u,r(T'))\leq \max \{d^v_{\Gamma'_0}(u,r(T')),d^v_{\Gamma'_1}(u,r(T'))\}$. Since $d^v_{\Gamma'_0}(u,r(T'))= d^v_{\Gamma_0}(u,r(T))$ and $d^v_{\Gamma'_1}(u,r(T'))= d^v_{\Gamma_1}(u,r(T))$, we have $d^v_{\Delta_i}(u,r(T))\leq \max \{d^v_{\Gamma_0}(u,r(T)),d^v_{\Gamma_1}(u,r(T))\}$ as required.
- By construction, we have $d^v_{\Delta_i}(l,r(T))= d^v_{\Delta_i}(\pi,r(T))+1$. Further, $d^v_{\Delta_i}(\pi,r(T))=d^v_{\Delta'_i}(\pi,r(T'))$. By induction we have $d^v_{\Delta'_i}(\pi,r(T'))\leq \max \{d^v_{\Gamma'_0}(\pi,r(T')),d^v_{\Gamma'_1}(\pi,r(T'))\}$. Since $d^v_{\Gamma'_0}(\pi,r(T'))=d^v_{\Gamma_0}(\pi,r(T))$ and $d^v_{\Gamma'_1}(\pi,r(T'))=d^v_{\Gamma_1}(\pi,r(T))$, we have $d^v_{\Delta_i}(l,r(T)) \leq \max \{d^v_{\Gamma_0}(\pi,r(T)),d^v_{\Gamma_1}(\pi,r(T))\}+1$. Since $\Gamma_0$ and $\Gamma_1$ are strictly-upward, we have $d^v_{\Gamma_0}(l,r(T))\geq d^v_{\Gamma_0}(\pi,r(T))+1$ and $d^v_{\Gamma_1}(l,r(T))\geq d^v_{\Gamma_1}(\pi,r(T))+1$, hence $d^v_{\Delta_i}(l,r(T)) \leq \max \{d^v_{\Gamma_0}(l,r(T)),d^v_{\Gamma_1}(l,r(T))\}$ as required.
Since $d^v_{\Delta_i}(u,r(T))\leq \max \{d^v_{\Gamma_0}(u,r(T)),d^v_{\Gamma_1}(u,r(T))\}$, for any index $i\in \{1,2,\dots,k\}$ and any node $u$ of $T$, we directly get $h(\Delta_i)\leq \max\{h(\Gamma_0),h(\Gamma_1)\}$ and hence $h(\mathcal M) = \max\{h(\Gamma_0),h(\Gamma_1)\}$.
It remains to prove the planarity of $\mathcal M$; then Lemma \[le:upwardplanar-implies-order\] implies that the drawing of $T$ is order-preserving throughout $\mathcal M$. The following property is useful (refer to \[fig:upward-linearsteps-polyarea-property\]).
\
\[pr:empty-right\] Consider any [order-preserving strictly-upward straight-line planar grid]{}drawing $\Gamma$ of an ordered rooted tree $S$. Let $h^1_{\Gamma}$ and $h^2_{\Gamma}$ be the horizontal half-lines starting at the root $r(S)$ and at the rightmost leaf $l^{\rightarrow}_S$ of $S$, respectively, and directed rightwards. Let $\mathcal R_{\Gamma}$ be the region of the plane which is delimited by the rightmost path $R_S$ of $S$ from the left and by $h^1_{\Gamma}$ and $h^2_{\Gamma}$ from above and below, respectively. Then no node or edge of $S$, other than those of $R_S$, intersects $\mathcal R_{\Gamma}$.
We first argue about possible intersections with the interior of $\mathcal R_{\Gamma}$.
- Suppose, for a contradiction, that an edge of $S$ intersects the interior of $\mathcal R_{\Gamma}$. Since $S$ is connected, it follows that there is an edge $e$ of $S$ that intersects the interior of $\mathcal R_{\Gamma}$ [*and*]{} its boundary. Since $\Gamma$ is planar, the edge $e$ does not cross $R_S$. Since $R_S$ is the rightmost path of $S$, the edge $e$ does not share a node with $R_S$. Since $\Gamma$ is strictly-upward and $r(S)$ is the root of $S$, we have that $e$ does not intersect $h^1_{\Gamma}$ other than, possibly, at $r(S)$; however, $r(S)$ belongs to $R_S$, and we already ruled out the possibility that $e$ shares a node with $R_S$. Finally, if $e$ intersects $\mathcal R_{\Gamma}$ and $h^2_{\Gamma}$, then, since $\Gamma$ is strictly-upward, the path from $e$ to $r(S)$ intersects $R_S$, hence there is an edge $e'$ of $S$ that intersects $\mathcal R_{\Gamma}$ and $R_S$, a case which we already ruled out; we thus get a contradiction.
- Suppose, for a contradiction, that a node of $S$ lies in the interior of $\mathcal R_{\Gamma}$. Since $S$ is connected, it follows that an edge of $S$ also intersects the interior of $\mathcal R_{\Gamma}$, a case which we already ruled out; we thus get a contradiction.
We now argue about possible intersections with the boundary of $\mathcal R_{\Gamma}$.
- Suppose, for a contradiction, that a node $u$ not in $R_S$ lies on the boundary of $\mathcal R_{\Gamma}$. Since $\Gamma$ is planar, the node $u$ does not lie on $R_S$. Since $\Gamma$ is strictly-upward and the root $r(S)$ of $S$ is a node of $R_S$, the node $u$ does not lie on $h^1_{\Gamma}$. Finally, if $u$ lies on $h^2_{\Gamma}$, then since $\Gamma$ is strictly-upward, the edge from $u$ to its parent intersects the interior of $\mathcal R_{\Gamma}$, a case which we already ruled out; we thus get a contradiction.
- Suppose, for a contradiction, that the interior of an edge $e$ not in $R_S$ intersects the boundary $\mathcal R_{\Gamma}$. Since $\Gamma$ is planar, the interior of $e$ does not intersect $R_S$. If the interior of $e$ intersects $h^1_{\Gamma}$ or $h^2_{\Gamma}$, then, since $\Gamma$ is strictly-upward, it intersects the interior of $\mathcal R_{\Gamma}$ as well, a case which we already ruled out; we thus get a contradiction.
This concludes the proof.
We now exploit \[pr:empty-right\] to prove the planarity of $\mathcal M$. Since $\mathcal M'$ is planar, by induction, and since $\mathcal M$ coincides with $\mathcal M'$ when restricted to the nodes and edges of $T'$, it follows that $\mathcal M$ is planar, as long as the edge $(\pi,l)$ does not intersect any edge, other than at a common end-point, throughout $\mathcal M$. We now argue about the possible intersections of the edge $(\pi,l)$ in $\mathcal M$.
First, we deal with the morph $\langle \Gamma_0,\Delta_1 \rangle$, in which only the node $l$ moves. By \[pr:empty-right\] applied to $T$ and $\Gamma_0$, no node or edge of $T$, other than those of the righmost path $R_T$ of $T$, intersects $\mathcal R_{\Gamma_0}$. Hence, it suffices to prove that the edge $(\pi,l)$ lies in $\mathcal R_{\Gamma_0}$ throughout $\langle \Gamma_0,\Delta_1 \rangle$; refer to \[fig:upward-linearsteps-polyarea-planar\]. Since $\Gamma_0$ is a strictly-upward grid drawing, we have that $l$ lies at least one unit below $\pi$ in $\Gamma_0$; further, by construction, $l$ lies one unit below $\pi$ in $\Delta_1$. Moreover, by construction, the position of $l$ in $\Delta_1$ is at least one unit to the right of the positions of $\pi$ and $l$ in $\Gamma_0$. It follows that the point at which $l$ is placed in $\Delta_1$ lies in the part $\Pi_{\Gamma_0}$ of $\mathcal R_{\Gamma_0}$ that is delimited by the representation of the edge $(\pi,l)$ in $\Gamma_0$ from the left and by the horizontal lines through the positions of $\pi$ and $l$ in $\Gamma_0$ from above and from below, respectively. The convexity of $\Pi_{\Gamma_0}$ implies that the edge $(\pi,l)$ lies in $\mathcal R_{\Gamma_0}$ throughout $\langle \Gamma_0,\Delta_1 \rangle$.
The proof that the morph $\langle \Delta_k,\Gamma_1 \rangle$ is planar is symmetric.
We now prove that the drawing $\Delta_i$ is planar, for each $i=0,\dots,k$; refer to \[fig:upward-linearsteps-intermediate\]. Since $\Delta'_i$ is planar, by induction, any crossing in $\Delta_i$ involves the edge $(\pi,l)$. Suppose, for a contradiction, that an edge $(u,v)$ crosses $(\pi,l)$. If $(u,v)$ is an edge of $R_{T}\setminus \{(\pi,l)\}$, then $(u,v)$ and $(\pi,l)$ are separated by the horizontal line through $\pi$ and thus do not cross. Assume hence that $(u,v)$ is not an edge of $R_{T}\setminus \{(\pi,l)\}$. The intersection between $(u,v)$ and $(\pi,l)$ in $\Delta_i$ has to happen at an interior point $c$ of both edges. Indeed:
- $l$ is to the right of both $u$ and $v$, hence it is not on $(u,v)$;
- $\pi$ is not in the interior of $(u,v)$, since $\Delta'_i$ is planar;
- $u$ and $v$ are not in the interior of $(\pi,l)$, given that $y_{\Delta_i}(l)=y_{\Delta_i}(\pi)-1$ and hence there is no grid point in the interior of $(\pi,l)$;
- if one of $u$ and $v$, say $u$, overlaps with $\pi$, then the planarity of $\Delta'_i$ implies that $u$ is the same node as $\pi$; since $v$ is not in the interior of $(\pi,l)$ and $l$ is not on $(u,v)$, it follows that $(u,v)$
Since the interior of $(u,v)$ intersects the interior of $(\pi,l)$, it follows that $(u,v)$ intersects the interior of the triangle $D$ which is delimited by $(\pi,l)$, by the horizontal line $\ell_\pi$ through $\pi$, and by $\ell_{\Delta'_i}$. Since $y_{\Delta_i}(l)=y_{\Delta_i}(\pi)-1$, it follows that $D$ contains no grid point in its interior. Hence, either $(u,v)$ intersects $\ell_{\pi}$ to the right of $\pi$, thus contradicting \[pr:empty-right\] for $\Delta'_i$, or $(u,v)$ intersects $\ell_{\Delta'_i}$ below $\ell_{\pi}$, thus implying that $u$ or $v$ is to the right of $\ell_{\Delta'_i}$; this contradicts the definition of $\ell_{\Delta'_i}$ and hence proves that $\Delta_i$ is planar.
Finally, we prove that, for each $i=1,\dots,k-1$, the morph $\langle \Delta_i,\Delta_{i+1} \rangle$ is planar. Suppose, for a contradiction, that an edge $(u,v)$ of $T$ crosses the edge $(\pi,l)$ during $\langle \Delta_i,\Delta_{i+1} \rangle$. Consider the first drawing during $\langle \Delta_i,\Delta_{i+1} \rangle$ in which $(u,v)$ and $(\pi,l)$ cross; denote such a drawing by $\Gamma$ and recall that $\Gamma\neq \Delta_i,\Delta_{i+1}$. Since any drawing before $\Gamma$ in $\langle \Delta_i,\Delta_{i+1} \rangle$ is planar, it follows that in $\Gamma$ an end-point of one of $(u,v)$ and $(\pi,l)$ lies on the other edge. Since $l$ is to the right of $u$ and $v$ both in $\Delta_i$ and in $\Delta_{i+1}$, it follows that $l$ is to the right of $u$ and $v$ throughout $\langle \Delta_i,\Delta_{i+1} \rangle$, hence it does not lie on $(u,v)$ in $\Gamma$. Since $\langle \Delta'_i,\Delta'_{i+1} \rangle$ is planar, $\pi$ does not lie on $(u,v)$ (except if it is the same node as $u$ or $v$, which however does not cause the crossing between $(u,v)$ and $(\pi,l)$). Hence, one of $u$ and $v$, say $u$, lies in the interior of $(\pi,l)$ in $\Gamma$.
Assume that $\pi$ does not move during $\langle \Delta_i,\Delta_{i+1} \rangle$; this is not a loss of generality, as a planar linear morph between two drawings remains planar if one of the two drawings is translated by an arbitrary vector (see, e.g., [@DBLP:journals/siamcomp/AlamdariABCLBFH17]). Refer to \[fig:upward-linearsteps-intermediate-new\]. Since $l$ is one unit below the horizontal line $\ell_{\pi}$ through $\pi$ both in $\Delta_i$ and in $\Delta_{i+1}$, it follows that $l$ moves horizontally in $\langle \Delta_i,\Delta_{i+1} \rangle$. By assumption, the straight-line segment representing the trajectory of $u$ in $\langle \Delta_i,\Delta_{i+1} \rangle$ crosses the triangle whose vertices are $\pi$ and the positions of $l$ in $\Delta_i$ and $\Delta_{i+1}$. Note that $\langle \Delta_i,\Delta_{i+1} \rangle$ moves $u$ from a grid point in $\Delta_i$ to a grid point in $\Delta_{i+1}$. However, since there are no grid points below $\ell_{\pi}$ and above the horizontal line through $l$, it follows that during $\langle \Delta_i,\Delta_{i+1} \rangle$ either $u$ crosses $R_{T'}$, a contradiction to the planarity of $\langle \Delta'_i,\Delta'_{i+1} \rangle$, or it crosses the horizontal line through $r(T)$, a contradiction to the fact that $\langle \Delta_i,\Delta_{i+1} \rangle$ is an upward morph, or it crosses the vertical line through $l$, a contradiction to the fact that $l$ is to the right of $u$ throughout $\langle \Delta_i,\Delta_{i+1} \rangle$, or it lies in $\mathcal R_{\Delta_i}$ in $\Delta_i$, or it lies in $\mathcal R_{\Delta_{i+1}}$ in $\Delta_{i+1}$; both the last two possibilities contradict \[pr:empty-right\]. This concludes the proof of the theorem.
In view of \[th:nary-trees-linearsteps-polyarea\], it is natural to ask whether a sub-linear number of steps suffices to construct a small-area morph between any two [order-preserving strictly-upward straight-line planar grid]{}drawings of a rooted ordered tree. In the following we prove that this is indeed the case for binary trees, for which just three morphing steps are sufficient.
Our algorithm borrows ideas from a recent paper by Da Lozzo [*et al.*]{} [@DBLP:conf/gd/LozzoBFPR18], which deals with upward planar morphs of *upward plane graphs*.
Consider any two order-preserving strictly-upward straight-line planar grid drawings $\Gamma_0$ and $\Gamma_1$ of an $n$-node rooted ordered binary tree $T$. We define two order-preserving strictly-upward straight-line planar grid drawings $\Gamma'_0$ and $\Gamma'_1$ of $T$ such that the $3$-step morph $\langle \Gamma_0,\Gamma'_0,\Gamma'_1,\Gamma_1 \rangle$ is upward and planar.
![The $3$-step morph $\langle \Gamma_0,\Gamma'_0,\Gamma'_1,\Gamma_1 \rangle$.[]{data-label="fig:binary-upward-morph"}](figures/Binary1.pdf)
For $i=0,1$, we define $\Gamma'_i$ recursively as follows; refer to \[fig:binary-upward-morph\]. Let $x_{\Gamma'_i}(r(T))=0$ and let $y_{\Gamma'_i}(r(T))=y_{\Gamma_i}(r(T))$. If the left subtree $L$ of $r(T)$ is non-empty, then recursively construct a drawing of it. Let $x_M$ be the maximum $x$-coordinate of a node in the constructed drawing of $L$; horizontally translate such a drawing by subtracting $x_M+1$ from the $x$-coordinate of every node in $L$, so that the maximum $x$-coordinate of any node in $L$ is now $-1$. Symmetrically, if the right subtree $R$ of $r(T)$ is non-empty, then recursively construct a drawing of it. Let $x_m$ be the minimum $x$-coordinate of a node in the constructed drawing of $R$; horizontally translate such a drawing by subtracting $x_m-1$ from the $x$-coordinate of every node in $R$, so that the minimum $x$-coordinate of any node in $R$ is now $1$.
\[th:binary-trees-upward-lineararea\] Let $T$ be an $n$-node rooted ordered binary tree, and let $\Gamma_0$ and $\Gamma_1$ be two [order-preserving strictly-upward straight-line planar grid]{}drawings of $T$. There exists a $3$-step upward planar morph $\mathcal M$ from $\Gamma_0$ to $\Gamma_1$ with $h(\mathcal M) = \max\{h(\Gamma_0),h(\Gamma_1)\}$ and $w(\mathcal M) = \max\{w(\Gamma_0),w(\Gamma_1),n\}$.
We prove that the morph $\mathcal M=\langle \Gamma_0,\Gamma'_0,\Gamma'_1,\Gamma_1\rangle$, where $\Gamma'_0$ and $\Gamma'_1$ are the drawings defined before the statement of the theorem,
Consider any $i\in \{1,2\}$. By construction, we have $y_{\Gamma'_i}(r(T))=y_{\Gamma_i}(r(T))$; since the drawings of $L$ and $R$ are constructed recursively and then translated horizontally, we have $y_{\Gamma'_i}(v)=y_{\Gamma_i}(v)$ for each node $v\in V(T)$. This has two implications. First, we have that $h(\Gamma'_i)=h(\Gamma_i)$ and hence that $h(\mathcal M) = \max\{h(\Gamma_0),h(\Gamma_1)\}$. Second, we have that $\Gamma'_i$ is strictly-upward, given that $\Gamma_i$ is strictly-upward, and hence that $\mathcal M$ is upward, by \[obs:upward-stays-upward\].
An easy inductive argument shows that no two nodes have the same $x$-coordinate in $\Gamma'_i$ and that every grid column intersecting $\Gamma'_i$ contains a node of $T$, hence $w(\Gamma'_i)=n$ and $w(\mathcal M) = \max\{w(\Gamma_0),w(\Gamma_1),n\}$.
It remains to prove that $\mathcal M$ is planar; then Lemma \[le:upwardplanar-implies-order\] implies that the drawing of $T$ is order-preserving throughout $\mathcal M$.
We first deal with the planarity of the drawings $\Gamma'_0$ and $\Gamma'_1$ and of the morph $\langle \Gamma'_0,\Gamma'_1\rangle$. Note that the assignment of $x$-coordinates to the nodes of $T$ in $\Gamma'_0$ and in $\Gamma'_1$ depends on $T$, and not on $\Gamma_0$ or $\Gamma_1$. It follows that $x_{\Gamma'_0}(v)=x_{\Gamma'_1}(v)$, for each node $v\in V(T)$. Hence, each node moves along a vertical line in $\langle \Gamma'_0,\Gamma'_1\rangle$. We now prove that any two distinct edges $(u,p(u))$ and $(v,p(v))$ of $T$ do not cross in $\langle \Gamma'_0,\Gamma'_1\rangle$. If $u=p(v)$, then the edges $(u,p(u))$ and $(v,p(v))$ are separated by the horizontal line through $u$ throughout $\langle \Gamma'_0,\Gamma'_1\rangle$, given that such a morph is upward, hence they do not intersect, except at $u$; similarly, if $u$ is a proper ancestor of $p(v)$, or if $v=p(u)$, or if $v$ is a proper ancestor of $p(u)$, then the edges $(u,p(u))$ and $(v,p(v))$ do not cross. Now suppose that $u$ is not an ancestor of $p(v)$ and $v$ is not an ancestor of $p(u)$. Then $u$ and $v$ are respectively in the left subtree and in the right subtree of their lowest common ancestor $w$ (up to renaming $u$ with $v$). By construction, the drawing of the left subtree of $r(T)$ lies to the left of the vertical line $\ell_w$ through $w$ both in $\Gamma'_0$ and in $\Gamma'_1$, hence it lies to the left of $\ell_w$ throughout $\langle \Gamma'_0,\Gamma'_1\rangle$. Analogously, the drawing of the right subtree of $r(T)$ lies to the right of $\ell_w$ throughout $\langle \Gamma'_0,\Gamma'_1\rangle$. It follows that $(u,p(u))$ and $(v,p(v))$ are separated by $\ell_w$ throughout $\langle \Gamma'_0,\Gamma'_1\rangle$, hence they do not cross.
Since $\Gamma'_0$ and $\Gamma'_1$ are order-preserving strictly-upward straight-line planar drawings of $T$ and since $y_{\Gamma'_0}(v)=y_{\Gamma_0}(v)$ and $y_{\Gamma'_1}(v)=y_{\Gamma_1}(v)$ for each node $v\in V(T)$, \[le:unidirectional-morph\] applies twice to ensure the planarity of the morphs $\langle \Gamma_0,\Gamma'_0\rangle$ and $\langle \Gamma'_1,\Gamma_1\rangle$. This concludes the proof.
Planar Morphs of Tree Drawings {#se:general}
==============================
In this section we show how to construct small-area morphs between straight-line planar grid drawings of trees. In particular, we prove the following result.
\[th:morph-straight-line-planar-drawings\] Let $T$ be an $n$-node ordered tree and let $\Gamma_0$ and $\Gamma_1$ be two order-preserving straight-line planar grid drawings of $T$. There exists an $O(n)$-step planar morph $\mathcal M$ from $\Gamma_0$ to $\Gamma_1$ with $h(\mathcal M) \in O(D^3n \cdot H)$ and $w(\mathcal M) \in O(D^3 n \cdot W)$, where $H = \max\{h(\Gamma_0),h(\Gamma_1)\}$, $W = \max\{w(\Gamma_0),w(\Gamma_1)\}$, and $D = \max\{H,W\}$.
The rest of this section is devoted to the proof of \[th:morph-straight-line-planar-drawings\]. We are going to use the following definition (see \[fig:canonical\]).
\[def:canonical\] An *upward canonical drawing* of a rooted ordered tree $T$ is an order-preserving strictly-upward straight-line planar grid drawing $\Gamma$ of $T$ satisfying the following properties:
- if $|V(T)|=1$, then $\Gamma$ is a grid point in the plane, representing $r(T)$;
- otherwise, let $\Gamma_1,\dots,\Gamma_k$ be upward canonical drawings of the subtrees $T_1,\dots,T_k$ of $r(T)$ (in their left-to-right order), respectively; then $\Gamma$ is such that:
- $r(T)$ is one unit to the left and one unit above the top-left corner of the bounding box of $\Gamma_1$;
- the top sides of the bounding boxes of $\Gamma_1, \dots, \Gamma_k$ have the same $y$-coordinate; and
- the right side of the bounding box of $\Gamma_i$ is one unit to the left of the left side of the bounding box of $\Gamma_{i+1}$, for $i=1, \dots, k-1$.
By counter-clockwise rotating an upward canonical drawing of $T$ by $\frac{\pi}{2}$, $\pi$, and $\frac{3\pi}{2}$ radians, we obtain a *leftward*, a *downward*, and a *rightward canonical drawing* of $T$, respectively. A *canonical drawing* of $T$ is an upward, leftward, downward, or rightward canonical drawing of $T$. Note that, in an upward, leftward, downward, or rightward canonical drawing $\Gamma$ of $T$, $r(T)$ is placed at the top-left, bottom-left, bottom-right, and top-right corner of the bounding box of $\Gamma$, respectively.
\[remark:canonical-area\] If $T$ has $n$ nodes, then a canonical drawing of $T$ lies in the $2n$-box centered at $r(T)$.
We prove the statement for an upward canonical drawing $\Gamma$; the proof for the other types of canonical drawings is analogous. First, we have $w(\Gamma)=n$, since each node has a distinct $x$-coordinate and, for each node $u$ of $T$, either there exists a node $v$ with $x(v)=x(u)+1$, or no node $v$ exists with $x(v)>x(u)$. Further, $h(\Gamma) \leq n$, since each node is vertically spaced from $r(T)$ by its graph-theoretic distance from it.
The following lemma allows us to morph one canonical drawing into another in a constant number of morphing steps.
\[lemma:rotate\] Let $\Gamma$ and $\Gamma'$ be two canonical drawings of a rooted ordered tree $T$, where $r(T)$ is at the same point in $\Gamma$ and $\Gamma'$. If $\Gamma$ and $\Gamma'$ are
upward and leftward, or
leftward and downward, or
downward and rightward, or
rightward and upward, respectively,
then the morph $\langle \Gamma, \Gamma' \rangle$ is planar and lies in the interior of the right, top, left, or bottom half of the $2n$-box centered at $r(T)$, respectively.
We prove the statement in the case in which $\Gamma$ is upward and $\Gamma'$ is leftward; the other cases are symmetric. Without loss of generality, up to translation of the Cartesian axes, we assume that $r(T)$ lies at the point $o = (0,0)$.
We insert in $\Gamma$ and in $\Gamma'$ the drawings $B$ and $B'$, respectively, of a $4$-cycle $(r(T), a, b, c)$ as follows. The coordinates of $a$, $b$, and $c$ in $B$ are $a(0) = (0,-n)$, $b(0) = (n,-n)$, and $c(0) = (n,0)$, respectively. The coordinates of $a$, $b$, and $c$ in $B'$ are $a(1) = (n,0)$, $b(1) = (n,n)$, and $c(1) = (0,n)$, respectively. Note that the drawing of $T$ in $\Gamma$ ($\Gamma'$) lies strictly inside $B$ ($B'$), except for $r(T)$. For each node $u$ of $T$ denote by $\omega_u$, $\alpha_u$, $\beta_u$, and $\gamma_u$ non-negative reals such that $u(0) =\omega_u \cdot o + \alpha_u \cdot a(0) + \beta_u \cdot b(0) + \gamma_u \cdot c(0)$, where $u(0)$ is the position of $u$ in $\Gamma$ and $\omega_u + \alpha_u + \beta_u + \gamma_u = 1$.
Consider the morph $\langle B,B'\rangle$ and assume it happens within the time interval $[0,1]$. Let $B(t)$ be the drawing of the cycle $(r(T), a, b, c)$ at time $t \in [0,1]$, where $B(0)=B$ and $B(1)=B'$. Note that the coordinates of $r(T)$, $a$, $b$, and $c$ in $B(t)$ are $o = (0,0)$, $a(t)=(tn,(t-1)n)$, $b(t)=(n,(2t-1)n)$, and $c(t)=((1-t)n,tn)$, respectively. For any $t \in [0,1]$, the affine transformation $\psi_t(\vec{x}):= \boldsymbol{A_t}\cdot\vec{x}$ with $\boldsymbol{A_t} =
\begin{pmatrix}
1-t && -t \\
t && 1-t
\end{pmatrix}
$ turns $B$ into $B(t)$. Let $\Gamma(t)$ be the drawing of $T$ obtained by applying $\psi_t$ to $\Gamma$. Since $\psi_t$ is an affine transformation and $det(\boldsymbol{A_t}) = (1-t)^2+t^2 \neq 0$, for any real value of $t$, we get that:
for each node $u$ of $T$, the position $u(t)$ of $u$ in $\Gamma(t)$ is $u(t) =\omega_u \cdot o + \alpha_u \cdot a(t) + \beta_u \cdot b(t) + \gamma_u \cdot c(t)$, that is, the coefficients of the convex combination expressing the placement of $u$ in $\Gamma(t)$ with respect to the placement of $r(T)$, $a$, $b$, and $c$ are the same as in $\Gamma$; and
$\Gamma(t)$ is an order-preserving straight-line planar drawing of $T$.
It remains to prove that $\Gamma(t)$ is the drawing at time $t$ of the morph $\langle \Gamma, \Gamma' \rangle$, for any $t \in [0,1]$. First, $\Gamma = \Gamma(0)$ since $\boldsymbol{A_0}$ is the identity matrix. Second, since a leftward canonical drawing of $T$ is obtained by counter-clockwise rotating an upward canonical drawing of $T$ by $\frac{\pi}{2}$ and since $\boldsymbol{A_1}=
\begin{pmatrix}
0 && -1 \\
1 && 0
\end{pmatrix}$ defines the same rotation, we have that $\Gamma'=\Gamma(1)$. Finally, for any $0<t<1$ and for each node $u$ of $T$, the position of $u$ at time $t$ of the morph $\langle \Gamma, \Gamma' \rangle$ is [ $$\begin{aligned}
&& (1-t)[\omega_u \cdot o + \alpha_u \cdot a(0) + \beta_u \cdot b(0) + \gamma_u \cdot c(0)] +t[\omega_u \cdot o + \alpha_u \cdot a(1) + \beta_u \cdot b(1) + \gamma_u \cdot c(1)] =\\
& = &\omega_u \cdot o + \alpha_u [a(0)\cdot (1-t) + a(1)\cdot t] + \beta_u [b(0)\cdot(1-t) + b(1)\cdot t] + \gamma_u [c(0)\cdot(1-t) + c(1)\cdot t] = \\
& = &\omega_u \cdot o + \alpha_u \cdot a(t) + \beta_u \cdot b(t) + \gamma_u \cdot c(t) = u(t).\end{aligned}$$ ]{} This concludes the proof that $\langle \Gamma, \Gamma' \rangle$ is planar.
Observe that, for each node $u$ of $T$, both $u(0)$ and $u(1)$ have their $x$-coordinates in the interval $[0,n-1]$ and their $y$-coordinates in the interval $[-n+1,n-1]$. This implies that, for any $t \in [0,1]$, the same holds for the coordinates of $u(t)$. Therefore, the morph $\langle \Gamma, \Gamma' \rangle$ lies in the right half of the $2n$-box centered at $r(T)$.
We now describe a proof of Theorem \[th:morph-straight-line-planar-drawings\]. Let $T$ be an $n$-node ordered tree and let $\Gamma_0$ and $\Gamma_1$ be two order-preserving straight-line planar grid drawings of $T$. In order to compute a morph $\cal{M}$ from $\Gamma_0$ to $\Gamma_1$, we root $T$ at any leaf $r(T)$. Since $T$ is ordered, this determines a left-to-right order
We are going to construct three morphs: a morph $\mathcal M^0$ from $\Gamma_0$ to a canonical drawing $\Gamma_0^*$ of $T$, a morph $\mathcal M^1$ from $\Gamma_1$ to a canonical drawing $\Gamma^*_1$ of $T$, and a morph $\mathcal M^{0,1}$ from $\Gamma_0^*$ to $\Gamma^*_1$. The morph $\mathcal M$ is then obtained by composing $\mathcal M^0$, $\mathcal M^{0,1}$, and the reverse of $\mathcal M^1$. The morph $\mathcal M^{0,1}$ consists of $O(1)$ steps and can be constructed by applying \[lemma:rotate\]. We describe below how to construct $\mathcal M^0$; the construction of $\mathcal{M}^1$ is analogous. However, before describing the construction of $\mathcal M^0$, we introduce a labeling of the nodes of $T$ and the concept of “partially-canonical drawing”.
Let $T[0]$ be the tree $T$ together with a labeling of each of the $k$ internal nodes of $T$ as `unvisited` and of each leaf as `visited`. We perform a bottom-up visit of $T$, labeling one-by-one the internal nodes of $T$ as `visited`. We label a node $v$ as `visited` only after all of its children have been labeled as `visited`. For $i=0,\dots,k$, we denote by $T[i]$ the tree $T$ once $i$ of its internal nodes have been labeled as `visited`.
The outline of our algorithm for constructing $\mathcal M^0$ is as follows. In a first morphing step, we scale $\Gamma_0$ up in order to make some free room around each node. Then we process the nodes of $T$ and label them as `visited` one by one, as described above. When we label a node $v$ as `visited`, we morph the current drawing into one in which $T_v$ is upward or downward canonical; this is accomplished by only moving the subtrees rooted at the children of $v$. Note that, when $v$ is labeled as `visited`, all the children of $v$ are already labeled as `visited`, hence the drawings of the subtrees rooted at them are upward or downward canonical. Thus, the movement of such drawings only consists of translations and rotations to bring such drawings where they need to be in the upward or downward canonical drawing of $T_v$. The initial scaling ensures that there is enough room around $v$ so that an upward or downward canonical of $T_v$ does not intersect the rest of the drawing. In order to formalize this process, we need to describe the properties of the drawing that we obtain after a number of nodes of $T$ have been labeled as `visited`; we call *partially-canonical* such a drawing.
Let $D_0 = \max\{w(\Gamma_0),h(\Gamma_0)\}$. Let $\Gamma$ be a drawing of $T$ and let $v$ be a node of $T$. We denote by ${\ensuremath{Large}({v})}$, ${\ensuremath{Med}({v})}$, and ${\ensuremath{Small({v})}}$ the $(\ell_0+4n)$-box, the $\ell_0$-box, and the $2n$-box centered at $v$ in $\Gamma$, respectively, where $\ell_0 = k_0 D_0^2n$ for some constant $k_0 > 1$ to be determined later. We have the following definition.
![A partially-canonical drawing $\Gamma$ of $T[i]$. The illustration focuses the attention on an `unvisited` $v$ whose children are all `visited`.[]{data-label="fig:partially-canonical"}](boxed.pdf){width=".5\textwidth"}
\[def:partially-canonical\] An order-preserving straight-line planar grid drawing $\Gamma$ of $T$ is a *partially-canonical drawing* of $T[i]$ if it satisfies the following properties (refer to \[fig:partially-canonical\]):
(a) \[partially-canonical-a\] for each `visited` node $u$ of $T$, the drawing $\Gamma_u$ of $T_u$ in $\Gamma$ is upward canonical or downward canonical; further, if $u\neq r(T)$, then $\Gamma_u$ is upward canonical, if $y_\Gamma(u) \le y_\Gamma(p(u))$, or downward canonical, if $y_\Gamma(u) > y_\Gamma(p(u))$;
(b) \[partially-canonical-c\] for each edge $e=(v,u)$ of $T$, where $v$ is the parent of $u$ and $v$ is `unvisited`, there exists a sector $S_e$ of a circumference centered at $v$ such that:
1. $S_e$ encloses ${\ensuremath{Small({u})}}$;
2. $S_e$ contains no node with the exception of $v$ and of, possibly, the nodes of $T_u$, and no edge with the exception of $e$ and of, possibly, the edges of $T_u$;
3. the intersection between $S_e$ and ${\ensuremath{Med}({v})}$ contains a $2n$-box $B_u$ whose corners have integer coordinates and whose center $c_u$ is such that $y_\Gamma(c_u) \leq y_\Gamma(v)$ if and only if $y_\Gamma(u) \leq y_\Gamma(v)$; and
4. for any edge $e' \neq e$ incident to $v$, the sectors $S_e$ and $S_{e'}$ are internally disjoint;
(c) for any two `unvisited` nodes $v$ and $w$, it holds ${\ensuremath{Large}({v})} \cap {\ensuremath{Large}({w})} = \emptyset$; and
(d) \[partially-canonical-b\] for any `unvisited` node $v$ of $T$, ${\ensuremath{Large}({v})}$ contains no node different from $v$, and any edge $e$ or any sector $S_e$ intersecting ${\ensuremath{Large}({v})}$ is such that $e$ is incident to $v$.
Note that, by Property (a), a partially-canonical drawing of $T[k]$ is a canonical drawing of $T$.
The algorithm to construct $\mathcal M^0$ is as follows. First, we scale $\Gamma_0$ up by a factor in $O(D_0^3n)$ so that the resulting drawing $\Delta_0$ is a partially-canonical drawing of $T[0]$ (see \[lemma:scaling\]). Clearly, the morph $\mathcal M_0 = \langle \Gamma_0, \Delta_0 \rangle$ is planar, $w(\mathcal M_0) = w(\Delta_0)$, and $h(\mathcal M_0) = h(\Delta_0)$.
For $i=1,\dots, k$, let $v_i$ be the node that is labeled as `visited` at the $i$-th step of the bottom-up visit of $T$. Starting from a partially-canonical drawing $\Delta_{i-1}$ of $T[i-1]$, we construct a partially-canonical drawing $\Delta_{i}$ of $T[i]$ and a morph $\mathcal M_{i-1,i}$ from $\Delta_{i-1}$ to $\Delta_{i}$ with $O(\deg(v_i))$ steps, with $w(\mathcal M_{i-1,i}) \leq w(\Delta_0) + \ell_0 + 4n$ and $h(\mathcal M_{i-1,i}) \leq h(\Delta_0) + \ell_0 + 4n$ (see \[lemma:partially-canonical-from-partially-canonical\]).
Composing the morphs $\mathcal M_0, \mathcal M_{0,1},\mathcal M_{1,2},\dots,\mathcal M_{k-1,k}$ yields the desired morph $\mathcal M^0$ from $\Gamma_0$ to a canonical drawing $\Delta_k=\Gamma^*_0$ of $T$. The morph has $O(\sum_{i=1}^k deg(v_i))\subseteq O(n)$ steps (by \[lemma:partially-canonical-from-partially-canonical\]). Further, $w(\mathcal M^0) \leq w(\Delta_0) + \ell_0 + 4n$ and $h(\mathcal M^0) \leq h(\Delta_0) + \ell_0 + 4n$ (again by \[lemma:partially-canonical-from-partially-canonical\]), hence $w(\mathcal M^0) \in O(D^3_0n \cdot w(\Gamma_0))$ and $h(\mathcal M^0) \in O(D^3_0n \cdot h(\Gamma_0))$ (by \[lemma:scaling\]).
It remains to show how to construct the partially-canonical drawing $\Delta_i$, for $i=0,\dots,k$, and the morphs $\mathcal M_{i-1,i}$, for $i=1,\dots,k$. Specifically, in the following we prove the next two lemmas.
\[lemma:scaling\] There is an integer $B_0 \in O(D^3_0n)$ such that the drawing $\Delta_0$ obtained by scaling the drawing $\Gamma_0$ of $T$ up by $B_0$ is a partially-canonical drawing of $T[0]$.
\[lemma:partially-canonical-from-partially-canonical\] For any $i\in \{1,\dots,k\}$, let $\Delta_{i-1}$ be a partially-canonical drawing of $T[i-1]$. There exists a partially-canonical drawing $\Delta_i$ of $T[i]$ and an $O(\deg(v_i))$-step planar morph $\mathcal M_{i-1,i}$ from $\Delta_{i-1}$ to $\Delta_{i}$ such that $w(\mathcal M_{i-1,i}) \leq w(\Delta_0) + \ell_0 + 4n$ and $h(\mathcal M_{i-1,i}) \leq h(\Delta_0) + \ell_0 + 4n$.
Proof of \[lemma:partially-canonical-from-partially-canonical\] {#sse:lemma-new-canonical}
---------------------------------------------------------------
We denote by $T^*$ the tree obtained by removing $T_{v_i}$ from $T$. Let $\Delta_i$ be the drawing of $T$ obtained from $\Delta_{i-1}$ by redrawing $T_{v_i}$ so that it is upward canonical, if $v_i=r(T)$ or if $v_i\neq r(T)$ and $y_{\Delta_{i-1}}(v_i) \leq y_{\Delta_{i-1}}(p(v_i))$, or downward canonical, otherwise, while keeping the placement of $v_i$ and of every node of $T^*$ unchanged. We have the following.
\[lem:partially-canonical-delta-i\] The drawing $\Delta_i$ is a partially-canonical drawing of $T[i]$.
By construction, the drawing $\Delta_i$ is such that:
(1) the drawing of $T^*$ is the same both in $\Delta_i$ and in $\Delta_{i-1}$,
(2) $v_i$ is in the same position in $\Delta_i$ and in $\Delta_{i-1}$, and
(3) the drawing of $T_{v_i}$ in $\Delta_i$ is upward canonical, if $v_i=r(T)$ or if $v_i\neq r(T)$ and $y_{\Delta_{i-1}}(v_i) \leq y_{\Delta_{i-1}}(p(v_i))$, or downward canonical, otherwise.
The drawing $\Delta_i$ is straight-line by construction and it is an order-preserving grid drawing since $\Delta_{i-1}$ is. The planarity of $\Delta_i$ can be proved as follows. First, the drawing of $T^*+(v_i,p(v_i))$ in $\Delta_i$ is planar since it coincides with the drawing of $T^*+(v_i,p(v_i))$ in $\Delta_{i-1}$, by (1) and (2), and since $\Delta_{i-1}$ is planar. Second, the drawing of $T_{v_i}$ in $\Delta_i$ is planar since it is a canonical drawing, by construction. Third, the drawing of $T_{v_i}$ in $\Delta_i$ does not intersect any node or edge of $T^*$; namely, the drawing of $T_{v_i}$ in $\Delta_i$ is contained in ${\ensuremath{Small({v_i})}}$, by \[remark:canonical-area\]; further ${\ensuremath{Small({v_i})}}$ is contained in ${\ensuremath{Large}({v_i})}$, by definition; finally, ${\ensuremath{Large}({v_i})}$ contains no node or edge of $T^*$, by Property (d) of $\Delta_{i-1}$. Fourth, the drawing of $T_{v_i}$ in $\Delta_i$ does not intersect the edge $(v_i,p(v_i))$, if such an edge exists, by (3).
We now show that $\Delta_i$ satisfies Properties (a)–(d) of \[def:partially-canonical\].
We start with Property (a). Consider any `visited` node $u$. If $u$ is not a node of $T_{v_i}$, then $T_u \subset T^*$. Thus, Property (a) holds for $u$ due to (1) and to the fact that $\Delta_{i-1}$ satisfies Property (a). If $u$ is a node of $T_{v_i}$, then the property holds due to (3).
We show that $\Delta_i$ satisfies Properties (b.i)–(b.iv) of \[def:partially-canonical\] by defining the sector $S_e$, for each edge $e=(u,v)$ where $v$ is the parent of $u$ and $v$ is `unvisited`, as the one defined for the edge $e$ in $\Delta_{i-1}$. Note that any `unvisited` node $v$ of $T[i]$ is also `unvisited` in $T[i-1]$.
Property (b.i) holds since the `unvisited` nodes of $T[i]$, as well as their children, have the same placement in $\Delta_i$ and in $\Delta_{i-1}$ (by (1) and (2)) and since $\Delta_{i-1}$ satisfies Property (b.i).
We show that Property (b.ii) holds. Consider any edge $e=(u,v)$ of $T[i]$, where $v$ is the parent of $u$ and $v$ is `unvisited`; note that $v$ is a node of $T^*$. By (1) and since $\Delta_{i-1}$ satisfies Property (b.ii), we have that $S_e$ contains no node of $T^*$ with the exception of $v$ and of, possibly, the nodes of $T_u$, and contains no edge of $T^*$ with the exception of $(u,v)$ and of, possibly, the edges of $T_u$. If $u=v_i$ or if $u$ is a proper ancestor of $v_i$, then this completes the proof. Otherwise, it remains to argue that $S_e$ contains no node or edge of $T_{v_i}$; this follows from the fact that $S_e$ does not intersect ${\ensuremath{Large}({v_i})}$, by Property (d) of $\Delta_{i-1}$ and by (1) and (2), and from the fact that the drawing of $T_{v_i}$ in $\Delta_i$ is canonical, hence by \[remark:canonical-area\] it is contained in ${\ensuremath{Small({v_i})}}$ and thus in ${\ensuremath{Large}({v_i})}$.
The drawing $\Delta_{i}$ trivially satisfies Properties (b.iii) and (b.iv), given that $\Delta_{i-1}$ satisfies the same properties, and given that the sector $S_e$ of any edge $e=(u,v)$ of $T[i]$, where $v$ is `unvisited` and $u$ is a child of $v$, is the same both in $\Delta_{i-1}$ and in $\Delta_{i}$, by (1) and (2) and by construction.
We show that $\Delta_i$ satisfies Property (c) of \[def:partially-canonical\]. First, for any two `unvisited` nodes $v$ and $w$ of $T[i-1]$, it holds ${\ensuremath{Large}({v})} \cap {\ensuremath{Large}({w})} = \emptyset$ in $\Delta_{i-1}$, since $\Delta_{i-1}$ satisfies Property (c) of \[def:partially-canonical\]. Further, any node that is `unvisited` in $T[i]$ is also `unvisited` in $T[i-1]$. Finally, by (1) the `unvisited` nodes of $T[i]$ have the same placement in $\Delta_i$ as in $\Delta_{i-1}$. Thus, Property (c) of \[def:partially-canonical\] holds for $\Delta_i$.
We show that $\Delta_i$ satisfies Property (d) of \[def:partially-canonical\]. Consider any `unvisited` node $v$ of $T[i]$ and note that $v$ is in $T^*$. By (1) the node $v$ has the same placement in $\Delta_i$ as in $\Delta_{i-1}$, hence the box ${\ensuremath{Large}({v})}$ is the same both in $\Delta_i$ and in $\Delta_{i-1}$. By (1) and (2) and since $\Delta_{i-1}$ satisfies Property (d) of \[def:partially-canonical\], it follows that ${\ensuremath{Large}({v})}$ contains no node of $T^*+(v_i,p(v_i))$ different from $v$, and any edge $e$ of $T^*+(v_i,p(v_i))$ or any sector $S_e$ of an edge $e$ of $T^*+(v_i,p(v_i))$ intersecting ${\ensuremath{Large}({v})}$ is such that $e$ is incident to $v$. Further, ${\ensuremath{Large}({v})}$ contains no node or edge of $T_{v_i}$ since the drawing of $T_{v_i}$ in $\Delta_i$ is contained in ${\ensuremath{Small({v_i})}}$, by \[remark:canonical-area\], since ${\ensuremath{Small({v_i})}}$ is contained in ${\ensuremath{Large}({v_i})}$, by definition, and since ${\ensuremath{Large}({v})}\cap {\ensuremath{Large}({v_i})}=\emptyset$, by Property (c) of $\Delta_{i-1}$. Finally, note that no sector is defined for any edge of $T_{v_i}$, as all the nodes of $T_{v_i}$ are `visited` in $T[i]$.
In order to prove \[lemma:partially-canonical-from-partially-canonical\], it remains to show how to construct an $O(\deg(v_i))$-step planar morph $\mathcal M_{i-1,i}$ from $\Delta_{i-1}$ to $\Delta_{i}$ such that $w(\mathcal M_{i-1,i}) \leq w(\Delta_0) + \ell_0 + 4n$ and $h(\mathcal M_{i-1,i}) \leq h(\Delta_0) + \ell_0 + 4n$. This is done in several phases as follows. Since the drawing of $T^*$ stays unchanged during $\mathcal M_{i-1,i}$, no two edges in $T^*$ cross during such a morph. Thus, we will omit to repeat this in the proofs of the planarity of the morphing steps that compose $\mathcal M_{i-1,i}$.
First, consider the drawing $\Delta'$ of $T$ obtained as described next; refer to \[fig:partially-canonical-first-morph\]. Initialize $\Delta'=\Delta_{i-1}$. Then, for each child $u$ of $v_i$, translate the drawing of $T_{u}$ (which is an upward or downward canonical drawing, by Property (a) of $\Delta_{i-1}$) so that $u$ is at the center of a $2n$-box $B_u$ that lies in the intersection between $S_e$ and ${\ensuremath{Med}({v_i})}$, whose corners have integer coordinates, and whose center $c_u$ is such that $y_{\Delta_{i-1}}(c_u) \leq y_{\Delta_{i-1}}(v_i)$ if and only if $y_{\Delta_{i-1}}(u) \leq y_{\Delta_{i-1}}(v_i)$; such a box exists by Property (b.iii) of $\Delta_{i-1}$.
\[claim:m-prime\] The morph $\langle \Delta_{i-1},\Delta'\rangle$ is planar.
First, for each child $u$ of $v_i$, the drawing of $T_u$ lies in the interior of the sector $S_e$ with $e = (v_i,u)$ both in $\Delta_{i-1}$ and in $\Delta'$; the former follows by Property (b.i) of $\Delta_{i-1}$ and by \[remark:canonical-area\], while the latter follows by Property (b.iii) of $\Delta_{i-1}$, by \[remark:canonical-area\], and by construction. It follows that the drawing of $T_u+(v_i,u)$ lies in the interior of $S_e$ throughout the morph $\langle \Delta_{i-1},\Delta'\rangle$. Thus, by Properties (b.ii) and (b.iv) of $\Delta_{i-1}$, the drawing of $T_u+(v_i,u)$ crosses neither the drawing of $T^*+(v_i,p(v_i))$ nor the drawing of $T_w+(v_i,w)$ during $\langle \Delta_{i-1},\Delta'\rangle$, where $w\neq u$ is a child of $v_i$.
It remains to prove that no two edges in $T_u + (v_i,u)$ cross each other during $\langle \Delta_{i-1},\Delta'\rangle$, for each child $u$ of $v_i$. By construction, the drawing of $T_u$ in $\Delta'$ is a translation of the drawing of $T_u$ in $\Delta_{i-1}$, hence no two edges of $T_u$ cross during the morph. Further, by construction, we have $y_{\Delta_{i-1}}(c_u) \leq y_{\Delta_{i-1}}(v_i)$ if and only if $y_{\Delta_{i-1}}(u) \leq y_{\Delta_{i-1}}(v_i)$. This implies that $v_i$ lies above $u$ in $\Delta'$ if and only if it lies above $u$ in $\Delta_{i-1}$. Therefore, the edge $(v_i,u)$ does not cross any edge of $T_u$ during the morph. This concludes the proof that $\langle \Delta_{i-1},\Delta'\rangle$ is planar.
![Regions for $v_i$.[]{data-label="fig:regions"}](boxed.pdf)
Second, we show how to move the subtrees rooted at the children of $v_i$ in the interior of ${\ensuremath{Large}({v_i})}$, so that they land in the position they have in $\Delta_i$. By Property (d) of $\Delta_{i-1}$, no node or edge of $T^*$ intersects ${\ensuremath{Large}({v_i})}$. Since the drawing of $T^*+(v_i,p(v_i))$ stays unchanged throughout the morph from $\Delta'$ to $\Delta_i$, no node or edge of $T^*$ crosses any node or edge of $T_{v_i}+(v_i,p(v_i))$ during such a morph. Thus, in the proofs of the planarity of the morphing steps that compose the morph from $\Delta'$ to $\Delta_i$, we will only need to prove that the subtrees rooted at the children of $v_i$ do not cross each other and do not cross the edges incident to $v_i$.
The way we move the subtrees rooted at the children of $v_i$ depends on their placement with respect to $v_i$ and to the drawing of the edge $(v_i,p(v_i))$. We assume that, if $v_i\neq r(T)$, then $y(p(v_i))\geq y(v_i)$ and $x(p(v_i)) \geq x(v_i)$; the other cases can be treated similarly. We distinguish four regions $\mathcal R_1$, $\mathcal R_2$, $\mathcal R_3$, and $\mathcal R_4$ defined as follows; refer to \[fig:regions\]. Let $h_\rightarrow(v)$ and $h_\leftarrow(v)$ be the horizontal rays originating at a node $v$ and directed rightward and leftward, respectively. Further, let $h_\uparrow(v)$ and $h_\downarrow(v)$ be the vertical rays originating at a node $v$ and directed upward and downward, respectively.
Region $\mathbf{\mathcal R_1}$
: is defined as follows. If $v_i=r(T)$, then $\mathcal R_1$ is the intersection of ${\ensuremath{Med}({v_i})}$ with the wedge centered at $v_i$ obtained by clockwise rotating $h_\uparrow(v_i)$ until it coincides with $h_\rightarrow(v_i)$. Otherwise, $\mathcal R_1$ is the intersection of ${\ensuremath{Med}({v_i})}$ with the wedge centered at $v_i$ obtained by counter-clockwise rotating $h_\rightarrow(v_i)$ until it passes through $p(v_i)$; note that, if $(v_i,p(v_i))$ is a horizontal segment, then $\mathcal R_1=\emptyset$.
Region $\mathbf{\mathcal R_2}$
: is the rectangular region that is the lower half of ${\ensuremath{Med}({v_i})}$;
Region $\mathbf{\mathcal R_3}$
: is the intersection of ${\ensuremath{Med}({v_i})}$ with the wedge centered at $v_i$ obtained by clockwise rotating $h_\leftarrow(v_i)$ until it coincides with $h_\uparrow(v_i)$; and
Region $\mathbf{\mathcal R_4}$
: is defined as follows. If $v_i=r(T)$, then $\mathcal R_4=\emptyset$. Otherwise, $\mathcal R_4$ is the intersection of ${\ensuremath{Med}({v_i})}$ with the wedge centered at $v_i$ obtained by clockwise rotating $h_\uparrow(v_i)$ until it passes through $p(v_i)$; note that, if $(v_i,p(v_i))$ is a vertical segment, then $\mathcal R_4=\emptyset$.
Note that ${\ensuremath{Med}({v_i})}=\mathcal R_1 \cup \mathcal R_2 \cup \mathcal R_3 \cup \mathcal R_4$.
\
We define two more regions, which will be exploited as “buffer regions” to allow rotations of subtrees via \[lemma:rotate\] in a safe way; refer again to \[fig:regions\]. Let $\mathcal S_L$ and $\mathcal S_R$ be the rectangular regions in $\Delta'$ containing all the points in ${\ensuremath{Large}({v_i})} - {\ensuremath{Med}({v_i})}$ to the left of the left side of ${\ensuremath{Med}({v_i})}$ and to the right of the right side of ${\ensuremath{Med}({v_i})}$, respectively. By Properties (b.iii) and (d) of the partially-canonical drawing $\Delta_{i-1}$, by the construction of $\Delta'$, and by the assumptions that, if $v_i\neq r(T)$, then $y(p(v_i))\geq y(v_i)$ and $x(p(v_i)) \geq x(v_i)$, we have that $\mathcal S_L$ is empty, while $\mathcal S_R$ may only contain the drawing of the part of the edge $(v_i, p(v_i))$ that possibly traverses such a region.
We start by dealing with the children $u_j$ of $v_i$ that lie in the interior of $\mathcal R_2$; refer to \[m:lower\]. Consider the edges $(v_i,u_j)$ in the order $(v_i,u_1),(v_i,u_2),\dots, (v_i,u_m)$ in which such edges are encountered while clockwise rotating $h_\rightarrow(v_i)$; see \[m:lower-a\]. Let $\Psi_1$ be the drawing obtained from $\Delta'$ by translating the drawing of the tree $T_{u_1}$ so that $u_1$ lies one unit below $v_i$ and so that the right side of the bounding box of the drawing of $T_{u_1}$ lies upon the right side of ${\ensuremath{Large}({v_i})}$; refer to \[m:lower-b\].
The morph $\langle \Delta',\Psi_1\rangle$ is planar.
By \[remark:canonical-area\], we have that the drawing of $T_{u_1}$ in $\Psi_1$ lies in $\mathcal S_R$, hence the drawing of $T_{u_1}$ is in ${\ensuremath{Large}({v_i})}$ throughout $\langle \Delta',\Psi_1\rangle$. Note that only $T_{u_1}+(v_i,u_1)$ moves during $\langle \Delta',\Psi_1\rangle$, hence any crossing during such a morph involves an edge of $T_{u_1}+(v_i,u_1)$ and an edge of $T_{v_i}+(v_i,p(v_i))$.
First, no two edges of $T_{u_1}$ cross each other during $\langle \Delta',\Psi_1\rangle$ since the drawing of $T_{u_1}$ in $\Psi_1$ is a translation of the drawing of $T_{u_1}$ in $\Delta'$.
Second, since the edge $(v_i,u_1)$ lies above the drawing of $T_{u_1}$ both in $\Delta'$ and in $\Psi_1$, it lies above the drawing of $T_{u_1}$ throughout $\langle \Delta',\Psi_1\rangle$, hence it does not cross any edge of $T_{u_1}$.
It remains to prove that no edge of $T_{u_1}+(v_i,u_1)$ crosses an edge of $T_{v_i}+(v_i,p(v_i))$ that is not in $T_{u_1}+(v_i,u_1)$. Consider the region $R(u_1)$ which is the intersection of ${\ensuremath{Large}({v_i})}$ and the wedge obtained by clockwise rotating the ray $h_\rightarrow(v_i)$ around $v_i$ until it passes through both rays delimiting the sector $S_{v_i,u_1}$. We have that $T_{u_1}$ moves in the interior of $R(u_1)$ during the morph. Further, by Properties (b.iii) and (b.iv) of the partially-canonical drawing $\Delta_{i-1}$, and by the assumptions that $y(p(v_i))\geq y(v_i)$, that $x(p(v_i))\geq x(v_i)$, and that $(v_i,u_1)$ is the first edge incident to $v_i$ that is encountered while clockwise rotating $h_\rightarrow(v_i)$, it follows that $R(u_1)$ does not contain any edge of $T_{v_i}+(v_i,p(v_i))$ that is not in $T_{u_1}+(v_i,u_1)$ throughout the morph $\langle \Delta',\Psi_1\rangle$. Hence, no edge of $T_{u_1}+(v_i,u_1)$ crosses an edge of $T_{v_i}+(v_i,p(v_i))$ that is not in $T_{u_1}+(v_i,u_1)$.
For $j=2,\dots,m$, let $\Psi_j$ be the drawing obtained from $\Psi_{j-1}$ by translating the drawing of the tree $T_{u_j}$ so that $u_j$ lies one unit below $v_i$ and so that the right side of the bounding box of the drawing of $T_{u_j}$ lies one unit to the left of $u_{j-1}$; refer to \[m:lower-c\].
For $j=2,\dots,m$, the morph $\langle \Psi_{j-1},\Psi_j\rangle$ is planar.
By \[remark:canonical-area\], we have that the drawing of $T_{u_j}$ in $\Psi_j$ lies in $\mathcal S_R$, hence the drawing of $T_{u_j}$ is in ${\ensuremath{Large}({v_i})}$ throughout $\langle \Psi_{j-1},\Psi_j\rangle$. Note that only $T_{u_j}+(v_i,u_j)$ moves during $\langle \Psi_{j-1},\Psi_j\rangle$, hence any crossing during such a morph involves an edge of $T_{u_j}+(v_i,u_j)$ and an edge of $T_{v_i}+(v_i,p(v_i))$.
First, no two edges of $T_{u_j}$ cross each other during $\langle \Psi_{j-1},\Psi_j\rangle$ since the drawing of $T_{u_j}$ in $\Psi_j$ is a translation of the drawing of $T_{u_j}$ in $\Psi_{j-1}$.
Second, since the edge $(v_i,u_j)$ lies above the drawing of $T_{u_j}$ both in $\Psi_{j-1}$ and in $\Psi_j$, it lies above the drawing of $T_{u_j}$ throughout $\langle \Psi_{j-1},\Psi_j\rangle$, hence it does not cross any edge of $T_{u_j}$.
It remains to prove that no edge of $T_{u_j}+(v_i,u_j)$ crosses an edge of $T_{v_i}+(v_i,p(v_i))$ that is not in $T_{u_j}+(v_i,u_j)$. First, by the assumption that $y(p(v_i))\geq y(v_i)$ and since the drawing of $T_{u_j}+(v_i,u_j)$ stays in the lower half of ${\ensuremath{Large}({v_i})}$ throughout $\langle \Psi_{j-1},\Psi_j\rangle$, we have that the drawing of $T_{u_j}+(v_i,u_j)$ does not cross the edge $(v_i,p(v_i))$, if such an edge exists. Second, we deal with possible crossings between $T_{u_j}+(v_i,u_j)$ and $T_{u_l}+(v_i,u_l)$, for any $l=j+1,\dots,m$. Consider the region $R(u_j)$ which is the intersection of ${\ensuremath{Large}({v_i})}$ and the wedge obtained by clockwise rotating the ray $h_\rightarrow(v_i)$ around $v_i$ until it passes through both rays delimiting the sector $S_{v_i,u_j}$. We have that $T_{u_j}$ moves in the interior of $R(u_j)$ during the morph. By Properties (b.iii) and (b.iv) of the partially-canonical drawing $\Delta_{i-1}$, the region $R(u_j)$ does not contain any edge of $T_{u_l}+(v_i,u_l)$ throughout $\langle \Psi_{j-1},\Psi_j\rangle$, hence no edge of $T_{u_j}+(v_i,u_j)$ crosses any edge of $T_{u_l}+(v_i,u_l)$. Third, we deal with possible crossings between $T_{u_j}+(v_i,u_j)$ and $T_{u_l}+(v_i,u_l)$, for any $l=1,\dots,j-1$. The edge $(v_i,u_l)$ is above the drawing of $T_{u_j}+(v_i,u_j)$ throughout $\langle \Psi_{j-1},\Psi_j\rangle$, hence $(v_i,u_l)$ does not cross any edge of $T_{u_j}+(v_i,u_j)$. Finally, the drawing of $T_{u_l}$ is to the right of the drawing of $T_{u_j}+(v_i,u_j)$ throughout $\langle \Psi_{j-1},\Psi_j\rangle$, hence no edge of $T_{u_l}$ crosses any edge of $T_{u_j}+(v_i,u_j)$.
Let $\Delta^+$ be the drawing obtained from $\Psi_m$ by horizontally translating every subtree $T_{u_j}$ so that $u_j$ lands at the position it has in $\Delta_i$, for $j=1,2,\dots,m$; see \[m:lower-c,m:lower-d\].
\[claim:horizontal-shift\] The morph $\langle \Psi_m,\Delta^+\rangle$ is planar.
Let $T_2$ be the rooted ordered tree composed of the node $v_i$ and of $T_{u_j}+(v_i,u_j)$, for $j=1,\dots,m$; that is, $T_2$ is the tree obtained from $T_{v_i}$ by removing the nodes of the subtrees rooted at the children of $v_i$ in $\mathcal R_1$, $\mathcal R_3$, and $\mathcal R_4$, and their incident edges. By \[remark:canonical-area\], we have that the drawing of $T_2$ in $\Delta^+$ lies in ${\ensuremath{Small({v_i})}}$, hence the drawing of $T_2$ is in ${\ensuremath{Large}({v_i})}$ throughout $\langle \Psi_m,\Delta^+\rangle$. Note that only the nodes of $T_2$ move during $\langle \Psi_m,\Delta^+\rangle$, hence any crossing during such a morph involves an edge of $T_2$ and an edge of $T_{v_i}+(v_i,p(v_i))$.
By the assumption that $y(p(v_i))\geq y(v_i)$ and since the drawing of $T_2$ stays in the lower half of ${\ensuremath{Large}({v_i})}$ throughout $\langle \Psi_m,\Delta^+\rangle$, we have that the drawing of $T_2$ does not cross the edge $(v_i,p(v_i))$, if such an edge exists, and does not cross any edge of a subtree of $T_{v_i}$ rooted at a child of $v_i$ that lies in $\mathcal R_1$, $\mathcal R_3$, or $\mathcal R_4$. Finally, the drawing of $T_2$ is order-preserving, strictly-upward, straight-line, and planar both in $\Psi_m$ and in $\Delta^+$; further, by construction, we have $y_{\Psi_m}(v)=y_{\Delta^+}(v)$, for each node $v$ of $T_2$. Thus, the linear morph $\langle \Psi_m,\Delta^+\rangle$ is planar, by \[le:unidirectional-morph\].
\
Next, we deal with the children $w_j$ of $v_i$ that lie in the interior of $\mathcal R_1$. Consider the edges $(v_i,w_j)$ in the order $(v_i,w_1),(v_i,w_2),\dots, (v_i,w_\ell)$ in which such edges are encountered while counter-clockwise rotating $h_{\rightarrow}(v_i)$ around $v_i$; refer to \[fig:r-two\]. We are going to move the subtrees rooted at the children of $v_i$ in $\mathcal R_1$, one by one in the order $T_{w_1},T_{w_2},\dots,T_{w_\ell}$, so that they land in the position that they have in $\Delta_i$. Such a movement consists of four phases. First, we rotate the drawing of $T_{w_j}$ so that it becomes leftward canonical (see \[fig:r-two-b\]). Second, we translate the drawing of $T_{w_j}$ so that $w_j$ lies in the interior of $S_R$ and one unit below $v_i$ (see \[fig:r-two-c\]). Third, we rotate the drawing of $T_{w_j}$ so that it becomes upward canonical (see \[fig:r-two-d\]). Finally, we horizontally translate the drawing of $T_{w_j}$ to its final position in $\Delta_i$ (see \[fig:r-two-e\]). We now provide the details of the above transformations.
For $j=1,\dots,\ell$, let $\xi_{j-1}^{4}$ be a drawing of $T$ with the following properties, where $\xi_0^{4} = \Delta^+$ (refer to \[fig:r-two-a,fig:r-two-e\], showing $\xi_{0}^{4}$ and $\xi_{1}^{4}$, respectively):
1. \[item:props-four-ii\] the drawing of $T^*$ is the same as in $\Delta_i$;
2. \[item:props-four-iii\] $v_i$ lies at the same point as in $\Delta_i$;
3. \[item:props-four-iv\] the drawing of the subtrees $T_{u_1},T_{u_2}, \dots,T_{u_m}$ and $T_{w_1},T_{w_2}, \dots,T_{w_{j-1}}$ is the same as in $\Delta_i$;
4. \[item:props-four-i\] the drawing of the subtrees $T_{w_{j}},T_{w_{j+1}}, \dots, T_{w_\ell}$ is the same as in $\Delta^+$; and
5. \[item:props-four-v\] the drawing of the subtrees of $T_{v_i}$ rooted at the children of $v_i$ that lie in the interior of $\mathcal R_3$ and $\mathcal R_4$ is the same as in $\Delta^+$.
For $j=1,\dots,\ell$, we construct a drawing $\xi_j^{1}$ from $\xi_{j-1}^{4}$ by rotating $T_{w_j}$ so that it is leftward canonical in $\xi_j^{1}$ and by leaving the position of the nodes not in $T_{w_j}$ unaltered. This rotation can be accomplished via a linear morph $\langle \xi_{j-1}^{4}, \xi_j^{1}\rangle$ by \[lemma:rotate\].
For $j=1,\dots,\ell$, the morph $\langle \xi_{j-1}^{4}, \xi_j^{1}\rangle$ is planar.
By \[remark:canonical-area\], we have that the drawing of $T_{w_j}$ in $\xi_j^{1}$ lies in ${\ensuremath{Small({w_j})}}$, hence the drawing of $T_{w_j}$ is in ${\ensuremath{Large}({v_i})}$ throughout $\langle \xi_{j-1}^{4}, \xi_j^{1}\rangle$. Note that only $T_{w_j}+(v_i,w_j)$ moves during $\langle \xi_{j-1}^{4}, \xi_j^{1}\rangle$, hence any crossing during such a morph involves an edge of $T_{w_j}+(v_i,w_j)$ and an edge of $T_{v_i}+(v_i,p(v_i))$.
First, no two edges of $T_{w_j}$ cross each other during $\langle \xi_{j-1}^{4}, \xi_j^{1}\rangle$, by \[lemma:rotate\].
Second, since the edge $(v_i,w_j)$ lies below the drawing of $T_{w_j}$ both in $\xi_{j-1}^{4}$ and in $\xi_j^{1}$, it lies below the drawing of $T_{w_j}$ throughout $\langle \xi_{j-1}^{4}, \xi_j^{1}\rangle$, hence it does not cross any edge of $T_{w_j}$.
It remains to prove that no edge of $T_{w_j}+(v_i,w_j)$ crosses an edge of $T_{v_i}+(v_i,p(v_i))$ that is not in $T_{w_j}+(v_i,w_j)$. No edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of $T_{u_h}+(v_i,u_h)$, for any $h\in \{1,\dots,m\}$, and any edge of $T_{w_h}+(v_i,w_h)$, for any $h\in \{1,\dots,j-1\}$, by Property (P3). Further, no edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of $T_{w_h}+(v_i,w_h)$, for any $h\in \{j+1,\dots,\ell\}$, by Properties (P4) and (P5) and since the sector $S_{(v_i,w_j)}$, which contains ${\ensuremath{Small({w_j})}}$ (by Property (b.i) of $\Delta_{i-1}$) and hence the drawing of $T_{w_j}+(v_i,w_j)$ throughout the morph $\langle \xi_{j-1}^{4}, \xi_j^{1}\rangle$, does not intersect the sector $S_{(v_i,w_h)}$ (by Property (b.iv) of $\Delta_{i-1}$), which contains ${\ensuremath{Small({w_h})}}$ (by Property (b.i) of $\Delta_{i-1}$) and hence the drawing of $T_{w_h}+(v_i,w_h)$ throughout the morph $\langle \xi_{j-1}^{4}, \xi_j^{1}\rangle$. The same arguments prove that no edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of $T_{w_h}+(v_i,w_h)$, for any $h\in \{j+1,\dots,\ell\}$, crosses any edge of a subtree of $T_{v_i}$ rooted at a child of $v_i$ lying in $\mathcal R_3$ or in $\mathcal R_4$.
For $j=1,\dots,\ell$, let $\xi_j^{2}$ be the drawing obtained from $\xi_j^{1}$ by translating the drawing of $T_{w_j}$ in such a way that $w_j$ lies one unit below $v_i$ and so that the right side of ${\ensuremath{Small({w_j})}}$ lies upon the right side of ${\ensuremath{Large}({v_i})}$.
For $j=1,\dots,\ell$, the morph $\langle \xi^1_j,\xi^2_j\rangle$ is planar.
By \[remark:canonical-area\], we have that the drawing of $T_{w_j}$ in $\xi_j^{1}$ lies in ${\ensuremath{Small({w_j})}}$, hence the drawing of $T_{w_j}$ is in ${\ensuremath{Large}({v_i})}$ throughout $\langle \xi_{j}^{1}, \xi_j^{2}\rangle$. Note that only $T_{w_j}+(v_i,w_j)$ moves during $\langle \xi_{j}^{1}, \xi_j^{2}\rangle$, hence any crossing during such a morph involves an edge of $T_{w_j}+(v_i,w_j)$ and an edge of $T_{v_i}+(v_i,p(v_i))$.
First, no two edges of $T_{w_j}$ cross each other during $\langle \xi_{j}^{1}, \xi_j^{2}\rangle$ since the drawing of $T_{w_j}$ in $\xi_j^{2}$ is a translation of the drawing of $T_{w_j}$ in $\xi_{j}^{1}$.
Second, since the edge $(v_i,w_j)$ lies to the left of the drawing of $T_{w_j}$ both in $\xi_{j}^{1}$ and in $\xi_j^{2}$, it lies to the left of the drawing of $T_{w_j}$ throughout $\langle \xi_{j}^{1}, \xi_j^{2}\rangle$, hence it does not cross any edge of $T_{w_j}$.
It remains to prove that no edge of $T_{w_j}+(v_i,w_j)$ crosses an edge of $T_{v_i}+(v_i,p(v_i))$ that is not in $T_{w_j}+(v_i,w_j)$. First, consider the region $R(w_j)$ which is the intersection of ${\ensuremath{Large}({w_j})}$ and the wedge obtained by counter-clockwise rotating the ray $h_\downarrow(w_j)$ around $w_j$ until it passes through both rays delimiting the sector $S_{v_i,w_j}$. We have that $T_{w_j}$ moves in the interior of $R(w_j)$ during the morph. Further, by Property (P4), the region $R(w_j)$ does not contain any edge of $T_{w_h}+(v_i,w_h)$, for any $h\in \{j+1,\dots,\ell\}$, hence no edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of $T_{w_h}+(v_i,w_h)$, for any $h\in \{j+1,\dots,\ell\}$. Similarly, no edge of $T_{w_j}+(v_i,w_j)$ crosses $(v_i,p(v_i))$ and, by Property (P5), no edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of a subtree of $T_{v_i}$ rooted at a child of $v_i$ lying in $\mathcal R_3$ or in $\mathcal R_4$. Next, we argue about the absence of crossings between the edges of $T_{w_j}+(v_i,w_j)$ and the edges of $T_{w_h}+(v_i,w_h)$, for any $h\in \{1,\dots,j-1\}$. The edge $(v_i,w_j)$ lies above all the edges of $T_{w_h}+(v_i,w_h)$ throughout $\langle \xi_{j}^{1}, \xi_j^{2}\rangle$. Further, $T_{w_j}$ lies above $T_{w_h}$ in every drawing of the morph $\langle \xi_{j}^{1}, \xi_j^{2}\rangle$, except for $\xi_j^{2}$, in which $T_{w_j}$ is to the right of $T_{w_h}$. Finally, $T_{w_j}$ lies above the line through $(v_i,w_h)$ in $\xi_{j}^{1}$ and in $\xi_j^{2}$, and hence throughout $\langle \xi_{j}^{1}, \xi_j^{2}\rangle$. It follows that no edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of $T_{w_h}+(v_i,w_h)$, for any $h\in \{1,\dots,j-1\}$. An analogous proof shows that no edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of $T_{u_h}+(v_i,u_h)$, for any $h\in \{1,\dots,m\}$.
For $j=1,\dots,\ell$, we construct a drawing $\xi_j^{3}$ from $\xi_j^{2}$ by rotating $T_{w_j}$ so that it is upward canonical in $\xi_j^{3}$ and by leaving the position of the nodes not in $T_{w_j}$ unaltered. This rotation can be accomplished via a linear morph $\langle \xi_j^{2}, \xi_j^{3}\rangle$ by \[lemma:rotate\].
For $j=1,\dots,\ell$, the morph $\langle \xi_{j}^{2}, \xi_j^{3}\rangle$ is planar.
By \[remark:canonical-area\], we have that the drawing of $T_{w_j}$ lies in ${\ensuremath{Small({w_j})}}$ both in $\xi_j^{2}$ and in $\xi_j^{3}$, and hence throughout $\langle \xi_{j}^{2}, \xi_j^{3}\rangle$. This implies that the drawing of $T_{w_j}$ lies in ${\ensuremath{Large}({v_i})}$, and in particular in $\mathcal S_R$, throughout $\langle \xi_{j}^{2}, \xi_j^{3}\rangle$. Note that only $T_{w_j}$ moves during $\langle \xi_{j}^{2}, \xi_j^{3}\rangle$, hence any crossing during such a morph involves an edge of $T_{w_j}$ and an edge of $T_{v_i}+(v_i,p(v_i))$.
First, no two edges of $T_{w_j}$ cross each other during $\langle \xi_{j}^{2}, \xi_j^{3}\rangle$, by \[lemma:rotate\].
Second, since the edge $(v_i,w_j)$ lies to the left of the drawing of $T_{w_j}$ both in $\xi_{j}^{2}$ and in $\xi_j^{3}$, it lies to the left of the drawing of $T_{w_j}$ throughout $\langle \xi_{j}^{2}, \xi_j^{3}\rangle$, hence it does not cross any edge of $T_{w_j}$.
Finally, no edge of $T_{w_j}$ crosses any edge of $T_{v_i}+(v_i,p(v_i))$ that is not in $T_{w_j}+(v_i,w_j)$, as throughout $\langle \xi_{j}^{2}, \xi_j^{3}\rangle$ the former lies in $\mathcal S_R$, while the latter lies in ${\ensuremath{Med}({v_i})}$.
For $j=1,\dots,\ell$, let $\xi_j^{4}$ be the drawing obtained from $\xi_j^{3}$ by translating the drawing of $T_{w_j}$ in such a way that $w_j$ lands at the position it has in $\Delta_i$ (that is, one unit below $v_i$ and one unit to the right of the rightmost node in $T_{w_{j-1}}$, if $j\geq 2$, or one unit to the right of the rightmost node in $T_{u_m}$, if $j=1$ and $v_i$ has children in $\mathcal R_2$ in $\Delta'$, or one unit to the right of $v_i$ if $j=1$ and $v_i$ has no child in $\mathcal R_2$ in $\Delta'$).
For $j=1,\dots,\ell$, the morph $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$ is planar.
By \[remark:canonical-area\], we have that the drawing of $T_{w_j}$ in $\xi_j^{3}$ lies in ${\ensuremath{Small({w_j})}}$, hence the drawing of $T_{w_j}$ is in ${\ensuremath{Large}({v_i})}$ throughout $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$. Note that only $T_{w_j}+(v_i,w_j)$ moves during $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$, hence any crossing during such a morph involves an edge of $T_{w_j}+(v_i,w_j)$ and an edge of $T_{v_i}+(v_i,p(v_i))$.
First, no two edges of $T_{w_j}$ cross each other during $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$ since the drawing of $T_{w_j}$ in $\xi_j^{4}$ is a translation of the drawing of $T_{w_j}$ in $\xi_{j}^{3}$.
Second, since the edge $(v_i,w_j)$ lies above the drawing of $T_{w_j}$ both in $\xi_{j}^{3}$ and in $\xi_j^{4}$, it lies above the drawing of $T_{w_j}$ throughout $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$, hence it does not cross any edge of $T_{w_j}$.
It remains to prove that no edge of $T_{w_j}+(v_i,w_j)$ crosses an edge of $T_{v_i}+(v_i,p(v_i))$ that is not in $T_{w_j}+(v_i,w_j)$. First, the edges of $T_{w_j}+(v_i,w_j)$ lie below the horizontal line through $v_i$ throughout $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$. Hence, no edge of $T_{w_j}+(v_i,w_j)$ crosses the edge $(v_i,p(v_i))$, or the edges of $T_{w_h}+(v_i,w_h)$, for any $h\in \{j+1,\dots,\ell\}$, or the edges of the subtrees of $T_{v_i}$ rooted at the children of $v_i$ lying in $\mathcal R_3$ or in $\mathcal R_4$, as all such edges lie above the horizontal line through $v_i$ throughout $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$. Next, we argue about the absence of crossings between the edges of $T_{w_j}+(v_i,w_j)$ and the edges of $T_{w_h}+(v_i,w_h)$, for any $h\in \{1,\dots,j-1\}$. The edge $(v_i,w_j)$ lies above all the edges of $T_{w_h}+(v_i,w_h)$ throughout $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$. Further, $T_{w_j}$ lies to the right of $T_{w_h}+(v_i,w_h)$ throughout $\langle \xi_{j}^{3}, \xi_j^{4}\rangle$. It follows that no edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of $T_{w_h}+(v_i,w_h)$, for any $h\in \{1,\dots,j-1\}$. An analogous proof shows that no edge of $T_{w_j}+(v_i,w_j)$ crosses any edge of $T_{u_h}+(v_i,u_h)$, for any $h\in \{1,\dots,m\}$.
Note that $\xi_j^{4}$ satisfies Properties (P1)–(P5), given that $\xi_{j-1}^{4}$ satisfies the same properties and given that during the morph $\langle \xi_{j-1}^{4}, \xi_{j}^{1}, \xi_{j}^{2}, \xi_{j}^{3}, \xi_j^{4}\rangle$ only the nodes of $T_{w_j}$ move, from their position in $\Delta^+$ to their position in $\Delta_i$. Eventually, the drawing $\xi_{\ell}^{4}$ coincides with $\Delta_{i}$, except for the drawing of the subtrees lying in the interior of $\mathcal R_3$ and $\mathcal R_4$.
Subtrees in $\mathcal R_3$ are treated symmetrically to the ones in $\mathcal R_1$. In particular, the subtrees rooted at the children of $v_i$ that lie in $\mathcal R_3$ are processed according to the clockwise order of the edges from $v_i$ to their roots, while the role played by $\mathcal S_R$ is now assumed by $\mathcal S_L$.
The treatment of the subtrees in $\mathcal R_4$ is similar to the one of the subtrees in $\mathcal R_3$. However, when a subtree is considered, it is first horizontally translated in the interior of $\mathcal R_3$ and then processed according to the rules for such a region.
Altogether, we have described a morph $\mathcal M_{i-1,i}$ from the partially-canonical drawing $\Delta_{i-1}$ of $T[i-1]$ to $\Delta_i$, which is a partially-canonical drawing of $T[i]$ by \[lem:partially-canonical-delta-i\]. Next, we argue about the properties of $\mathcal M_{i-1,i}$.
We first deal with the space requirements of $\mathcal M_{i-1,i}$. Consider the drawing $\Delta_0$ and place the boxes ${\ensuremath{Large}({v})}$ around the nodes $v$ of $T$; the bounding box of the arrangement of such boxes has width $w(\Delta_0)+\ell_0+4n$ and height $h(\Delta_0)+\ell_0+4n$. We claim that the drawings of $\mathcal M_{i-1,i}$ lie inside such a bounding box. Assume this is true for $\Delta_{i-1}$ (this is indeed the case when $i=1$); all subsequent drawings of $\mathcal M_{i-1,i}$ coincide with $\Delta_{i-1}$, except for the placement of the subtrees rooted at the children of $v_i$, which however lie inside ${\ensuremath{Large}({v_i})}$ in each of such drawings. Since $v_i$ has the same position in $\Delta_{i}$ as in $\Delta_0$ and since ${\ensuremath{Large}({v_i})}$ has width and height equal to $\ell_0+4n$, the claim follows.
Finally, we deal with the number of linear morphs composing $\mathcal M_{i-1,i}$. The morph $\mathcal M_{i-1,i}$ consists of the morph $\langle \Delta_{i-1},\Delta'\rangle$, followed by the morphs needed to move the subtrees rooted at the children of $v_i$ to their final positions in $\Delta_i$. Since the number of morphing steps needed to deal with each of such subtrees is constant, we conclude that $M_{i-1,i}$ consists of $O(\deg(v_i))$ linear morphing steps. This concludes the proof of \[lemma:partially-canonical-from-partially-canonical\].
Conclusions and Open Problems {#se:conclusions}
=============================
We presented an algorithm that, given any two order-preserving straight-line planar grid drawings $\Gamma_0$ and $\Gamma_1$ of an $n$-node ordered tree $T$, constructs a morph $\langle \Gamma_0=\Delta_0,\Delta_1,\dots,\Delta_{k}=\Gamma_1\rangle$ such that $k$ is in $O(n)$ and such that the area of each intermediate drawing $\Delta_i$ is polynomial in $n$ and in the area of $\Gamma_0$ and $\Gamma_1$. Better bounds can be achieved if $T$ is rooted and $\Gamma_0$ and $\Gamma_1$ are also strictly-upward drawings, especially in the case in which $T$ is a binary tree.
We make two remarks about the generality of the model that we adopted. Both observations apply not only to tree drawings but, more in general, to planar graph drawings.
First, our assumption that $\Gamma_0$ and $\Gamma_1$ are grid drawings seems restrictive, and it seems more general to consider drawings that have bounded resolution, where the [*resolution*]{} of a drawing is the ratio between the largest and the smallest distance between a pair of geometric objects in the drawing (points representing nodes or segments representing edges). However, by using an observation from [@cegl-dgppof-12], one can argue that two morphing steps suffice to transform a drawing with resolution $r$ in a grid drawing whose area is polynomial in $r$. This is formalized in the following.
\[le:from-resolution-to-grid\] Let $\Gamma$ be a planar straight-line drawing of a planar graph $G$ and let $r$ be the resolution of $\Gamma$. There exists a $2$-step planar morph $\langle \Gamma, \Gamma', \Gamma'' \rangle$ such that the resolution of $\Gamma'$ is $r$ and such that $\Gamma''$ is a planar straight-line grid drawing lying on an $O(r)\times O(r)$ grid.
First, we construct $\Gamma'$ by scaling $\Gamma$ in such a way that the smallest distance between any pair of geometric objects is $2$. Clearly, $\langle \Gamma, \Gamma' \rangle$ is a planar linear morph and the resolution of $\Gamma'$ is the same as the one of $\Gamma$; hence, the largest distance between any pair of geometric objects in $\Gamma'$ is in $O(r)$. Second, we construct $\Gamma''$ from $\Gamma'$ by moving each node to the nearest grid point; the linear morph $\langle \Gamma', \Gamma'' \rangle$ is planar since each node moves by at most $\sqrt 2/2$, hence this motion brings any two geometric objects closer by at most $\sqrt 2$, while their distance is at least $2$. Thus, $\Gamma''$ lies on an $O(r)\times O(r)$ grid.
Second, our model deals with morphs that consist of sequences of drawings $\Delta_0,\Delta_1,\dots,\Delta_k$ lying on polynomial-size grids. However, no bound on the resolution is explicitly required for the drawings visualized during these morphs, i.e., for the drawings intermediate to each linear morph $\langle \Delta_i,\Delta_{i+1} \rangle$. Thus, one might wonder whether the resolution becomes arbitrarily large in some drawing of such morphs. The next lemma proves that this is not the case.
\[le:resolution\] Let $\Gamma_0$ and $\Gamma_1$ be two planar straight-line grid drawings of the same graph $G$ lying on $\ell\times \ell$ grids, for some value $\ell>0$. Suppose that the linear morph $\langle \Gamma_0,\Gamma_1 \rangle$ is planar. Then the maximum resolution of any drawing of $\mathcal M$ is in $O(\ell^4)$.
Assume w.l.o.g. that the morph $\langle \Gamma_0,\Gamma_1 \rangle$ happens between times $t=0$ and $t=1$; then, for every $t\in [0,1]$, denote by $\Gamma_t$ the drawing of $\langle \Gamma_0,\Gamma_1 \rangle$ at time $t$.
Since each of $\Gamma_0$ and $\Gamma_1$ lies on a $\ell\times \ell$ grid, it follows that $\Gamma_t$ lies on a $\ell\times \ell$ grid, for every $t\in [0,1]$. Hence, the largest distance between a pair of geometric objects in $\Gamma_t$ is in $O(\ell)$. It remains to prove that the smallest distance between a pair of geometric objects in $\Gamma_t$ is in $\Omega(\frac{1}{\ell^3})$, for every $t\in [0,1]$. Such a smallest distance occurs either between two vertices of $G$ or between a vertex and an edge of $G$. Indeed, the distance between two edges always coincides with the distance between one of the edges and an end-vertex of the other edge.
- First, consider any two vertices $u$ and $v$ of $G$. We are going to prove that, for any $t\in [0,1]$, the distance between $u$ and $v$ in $\Gamma_t$ is in $\Omega(\frac{1}{\ell})$. In order to do so, it is convenient to translate $\Gamma_1$ by a vector $\vec v:= (x_{\Gamma_0}(v)-x_{\Gamma_1}(v),y_{\Gamma_0}(v)-y_{\Gamma_1}(v))$; this defines a drawing $\Gamma'_1$ of $G$ in which $v$ is placed at the same point as in $\Gamma_0$. Note that the morph $\langle \Gamma_0,\Gamma'_1 \rangle$ is planar, as for any $t\in [0,1]$ the drawing $\Gamma'_t$ at time $t$ of $\langle \Gamma_0,\Gamma'_1 \rangle$ coincides with the drawing $\Gamma_t$ translated by the vector $t \cdot \vec v$. Hence, the distance between $u$ and $v$ in $\Gamma'_t$ is the same as in $\Gamma_t$. Further, since $v$ does not move during $\langle \Gamma_0,\Gamma'_1 \rangle$, the minimum distance between $u$ and $v$ during $\langle \Gamma_0,\Gamma'_1 \rangle$ coincides with the distance between $v$ and the straight-line segment $s$ whose end-points are the positions of $u$ in $\Gamma_0$ and in $\Gamma'_1$. Since both $\Gamma_0$ and $\Gamma'_1$ lie inside a box with side-length $2\ell$ centered at $v$, it follows by Lemma \[claim:minimum-grid-distance\] that the distance between $v$ and $s$ is in $\Omega(\frac{1}{\ell})$.
- Second, consider any vertex $u$ and any edge $(v,w)$ of $G$, where $u\neq v,w$. In order to compute the minimum distance between $u$ and $(v,w)$ during $\langle \Gamma_0,\Gamma_1 \rangle$, we translate $\Gamma_0$ by a vector $\vec v:= (-x_{\Gamma_0}(v),-y_{\Gamma_0}(v))$, thus obtaining a drawing $\Gamma'_0$ of $G$ in which $v$ lies at $(0,0)$, and we translate $\Gamma_1$ by a vector $\vec v':= (-x_{\Gamma_1}(v),-y_{\Gamma_1}(v))$, thus obtaining a drawing $\Gamma'_1$ of $G$ in which $v$ also lies at $(0,0)$. Note that the morph $\langle \Gamma'_0,\Gamma'_1 \rangle$ is planar, as for any $t\in [0,1]$ the drawing $\Gamma'_t$ at time $t$ of $\langle \Gamma'_0,\Gamma'_1 \rangle$ coincides with the drawing $\Gamma_t$ translated by the vector $(1-t) \cdot \vec v + t \cdot \vec v'$. Hence, the distance between $u$ and $(v,w)$ in $\Gamma'_t$ is the same as in $\Gamma_t$. Note that $x_{\Gamma'_t}(u)=(1-t)x_{\Gamma'_0}(u)+t \cdot x_{\Gamma'_1}(u)$, and similar for $y_{\Gamma'_t}(u)$, $x_{\Gamma'_t}(w)$, and $y_{\Gamma'_t}(w)$; further, $x_{\Gamma'_t}(v)=y_{\Gamma'_t}(v)=0$.
Let $t^*\in [0,1]$ be the time of $\langle \Gamma'_0,\Gamma'_1 \rangle$ in which the distance $d$ between $u$ and $(v,w)$ is minimum; we need to prove that $d\in \Omega(\frac{1}{\ell^3})$. Let $l^{vw}_{t^*}$ be the line through $v$ and $w$ in $\Gamma'_{t^*}$ and let $d^*$ be the distance between $u$ and $l^{vw}_{t^*}$; since $(v,w)$ is part of $l^{vw}_{t^*}$ in $\Gamma'_{t^*}$, we have $d\geq d^*$.
- We can assume that $d^*>0$. Indeed, if $d^*=0$, we have that $u$ lies on $l^{vw}_{t^*}$. By the planarity of $\langle \Gamma'_0,\Gamma'_1 \rangle$, we have that $u$ does not lie on the edge $(v,w)$, hence $d$ is equal to the distance between $u$ and one of the vertices $v$ and $w$, which is in $\Omega(\frac{1}{\ell})$ as proved above.
- We can also assume that $t^*\in (0,1)$; indeed, if $t^*=0$ or $t^*=1$, then $d,d^*\in \Omega(\frac{1}{\ell})$, by \[claim:minimum-grid-distance\].
The line $l^{vw}_{t^*}$ has equation $(y-y_{\Gamma'_{t^*}}(v))/(y_{\Gamma'_{t^*}}(w)-y_{\Gamma'_{t^*}}(v))=(x-x_{\Gamma'_{t^*}}(v))/(x_{\Gamma'_{t^*}}(w)-x_{\Gamma'_{t^*}}(v))$, which is $y_{\Gamma'_{t^*}}(w) \cdot x - x_{\Gamma'_{t^*}}(w) \cdot y = 0$. Then we have
$$\label{le:distance}
d^*=\frac{|y_{\Gamma'_{t^*}}(w) \cdot x_{\Gamma'_{t^*}}(u)- x_{\Gamma'_{t^*}}(w) \cdot y_{\Gamma'_{t^*}}(u)|}{\sqrt{(y_{\Gamma'_{t^*}}(w))^2+(x_{\Gamma'_{t^*}}(w))^2}}.$$
Since $\Gamma'_{t^*}$ lies inside a box with side-length $2\ell$ centered at $v\equiv (0,0)$, it follows that $|y_{\Gamma'_{t^*}}(w)|\leq \ell$ and $|x_{\Gamma'_{t^*}}(w)|\leq \ell$, hence the denominator of \[le:distance\] is at most $\sqrt{2\ell^2}\in O(\ell)$. It remains to prove that the numerator of \[le:distance\] is in $\Omega(\frac{1}{\ell^2})$.
The numerator of \[le:distance\] is the absolute value of a second-degree polynomial $\mathcal P(t)$, namely $$\begin{aligned}
\mathcal P(t) &:=& y_{\Gamma'_{t^*}}(w) \cdot x_{\Gamma'_{t^*}}(u)- x_{\Gamma'_{t^*}}(w) \cdot y_{\Gamma'_{t^*}}(u)=\\
&=& (((1-t)y_{\Gamma'_0}(w)+t \cdot y_{\Gamma'_1}(w))\cdot ((1-t)x_{\Gamma'_0}(u)+t \cdot x_{\Gamma'_1}(u))) -\\ & & (((1-t)x_{\Gamma'_0}(w)+t \cdot x_{\Gamma'_1}(w))\cdot ((1-t)y_{\Gamma'_0}(u)+t \cdot y_{\Gamma'_1}(u)))=\\
&=& A\cdot t^2 + B\cdot t + C,
\end{aligned}$$
where each of $A$, $B$, and $C$ is the algebraic sum of a constant number of terms, each of which is either an element of $\mathcal S=\{y_{\Gamma'_0}(w),y_{\Gamma'_1}(w),x_{\Gamma'_0}(w),x_{\Gamma'_1}(w),y_{\Gamma'_0}(u),y_{\Gamma'_1}(u),x_{\Gamma'_0}(u),x_{\Gamma'_1}(u)\}$, or the product of two elements in $\cal S$; note that each element in $\mathcal S$ is an integer in $[-\ell,\ell]$. Since $d^*>0$ and since the function $z=\mathcal P(t)$ is continuous, we have that $\mathcal P(t)$ is either positive over the entire interval $[0,1]$ or negative over the entire interval $[0,1]$. Assume the former, as in the latter case the absolute value in the numerator of \[le:distance\] reverts the signs of the terms of $\mathcal P(t)$ and our arguments are the same. We then only need to prove that $\mathcal P(t^*)\in \Omega(\frac{1}{\ell^2})$.
- We can assume that the minimum value of $\mathcal P(t)$ over the interval $[0,1]$ is achieved at a time $t^m\in (0,1)$. Indeed, if the minimum value of $\mathcal P(t)$ over the interval $[0,1]$ is achieved at $t=0$ or $t=1$, then we have $\mathcal P(t^*)\geq A + B + C\geq 1$ or $\mathcal P(t^*)\geq C\geq 1$ (recall that $A$, $B$, and $C$ are integers and that $\mathcal P(t)$ is positive over the entire interval $[0,1]$), respectively.
Since $\mathcal P(t)$ has a minimum at $t^m\in (0,1)$, we have $A>0$ and the derivative $\frac{\partial \mathcal P(t)}{\partial t}=2A\cdot t+B$ has a zero in $t^m=-B/2A$. Then we have $\mathcal P(t^*)\geq \mathcal P(t^m) = A (\frac{-B}{2A})^2 + B (\frac{-B}{2A}) + C = \frac{-B^2+4AC}{4A}$. The numerator of the previous fraction is greater than or equal to $1$ (given that $A$, $B$, and $C$ are integers and that $\mathcal P(t)$ is positive over the entire interval $[0,1]$), while the denominator is in $O(\ell^2)$ (given that $A$ is the algebraic sum of a constant number of terms, each of which is either an integer in $[-\ell,\ell]$, or the product of two integers in $[-\ell,\ell]$). Hence, $\mathcal P(t^m)\in \Omega(\frac{1}{\ell^2})$ and thus $\mathcal P(t^*)\in \Omega(\frac{1}{\ell^2})$.
This concludes the proof of the lemma.
\[le:from-resolution-to-grid,le:resolution\] imply that the problem of constructing planar morphs with polynomial resolution between two planar straight-line drawings of the same planar graph can indeed be reduced to the problem of constructing planar morphs with polynomial area between two planar straight-line grid drawings of the same planar graph. For example, it follows by \[le:from-resolution-to-grid,le:resolution\] and by \[th:morph-straight-line-planar-drawings\] that, given an $n$-node ordered tree $T$ and given two order-preserving straight-line planar drawings $\Gamma_0$ and $\Gamma_1$ of $T$ with maximum resolution $r$, there exists an $O(n)$-step planar morph $\cal M$ from $\Gamma_0$ to $\Gamma_1$ such that the resolution of any intermediate drawing of $\cal M$ is a polynomial function of $r$.
Several problems are left open by our research. Is it possible to generalize our results to graph classes richer than trees? How about outerplanar graphs? Is it possible to improve our area bounds for morphs of straight-line planar grid drawings of trees or even just of paths? Is there a trade-off between the number of steps and the area required by a morph? Is it possible to construct upward planar morphs with a constant number of steps between any two order-preserving strictly-upward straight-line planar grid drawings of an $n$-node rooted ordered tree? Are there other relevant tree drawing standards for which it makes sense to consider the morphing problem?
[^1]: This research was supported in part by MIUR Project “MODE” under PRIN 20157EFM5C, by MIUR Project “AHeAD” under PRIN 20174LF3T8, by H2020-MSCA-RISE project 734922 – “CONNECT”, and by MIUR-DAAD JMP N$^\circ$ 34120.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Using wide-field photometric data from the Sloan Digital Sky Survey (SDSS) we recently showed that the Galactic globular cluster Palomar 5 is in the process of being tidally disrupted. Its tidal tails were initially detected in a 2.5 degree wide band along the celestial equator. A new analysis of SDSS data for a larger field now reveals that the tails of Pal5 have a much larger spatial extent and can be traced over an arc of 10$^\circ$ on the sky, corresponding to a projected length of 4 kpc at the distance of the cluster. The tail that trails behind the Galactic motion of the cluster fades into the field at an angular distance of $6\fdg 5$ from the cluster center but shows a pronounced density maximum between $2^\circ$ and $4^\circ$ from the center. The leading tail of length $3\fdg 5$ extends down to the border of the available field and thus presumably continues beyond it. The projected width of these tails is small and almost constant (FWHM $\sim$ 120 pc), which implies that they form a dynamically cold and hence long-lived structure. The number of former cluster stars found in the tails adds up to about 1.2 times the number of stars in the cluster, i.e. the tails are more massive than the cluster in its present state. The radial profile of stellar surface density in the tails follows approximately a power law $r^{\gamma}$ with $-1.5 \le \gamma \le -1.2$.
The stream of debris from Pal5 is significantly curved, which demonstrates its acceleration by the Galactic potential. The stream sets tight constraints on the geometry of the cluster’s Galactic orbit. We conclude that the cluster is presently near the apocenter but has repeatedly undergone disk crossings in the inner part of the Galaxy leading to strong tidal shocks. Using the spatial offset between the tails and the cluster’s orbit we estimate the mean drift rate of the tidal debris and thus the mean mass loss rate of the cluster. Our results suggest that the observed debris originates mostly from mass loss within the last 2 Gyrs. The cluster is likely to be destroyed after the next disk crossing, which will happen in about 100 Myr. There is strong evidence against the suggestion that Pal5 might be associated with the Sgr dwarf galaxy.
author:
- 'Michael Odenkirchen, Eva K. Grebel, Walter Dehnen, Hans-Walter Rix, Brian Yanny Heidi Newberg, Constance M. Rockosi, David Martinez-Delgado, Jon Brinkmann, Jeffrey R. Pier'
title: 'The extended tails of Palomar 5: A ten degree arc of globular cluster tidal debris'
---
Introduction
============
Globular clusters are the oldest stellar systems commonly found in the Milky Way, having typical ages of 12 to 15 Gyr. They thus represent fossil relics from the early formation history of the Galaxy. However, the globular clusters we see today are probably not representative of the system of Galactic globular clusters at early stages. They may in fact be the selected ’survivors’ of an initially much more abundant population. Analytic estimates and numerical experiments predict that on time scales of Gigayears globular clusters undergo external and internal dynamical evolution, by which they may suffer a permanent loss of members, and eventually dissolve.
One of the major factors governing the dynamical evolution of those clusters is the Galactic tidal field. The tidal field has two important effects: (1) It creates drains through which stars are carried away from the outer part of the cluster, and hereby truncates the bound part of the cluster to a certain spatial limit (von Hoerner 1957, King 1962). (2) It feeds energy into the cluster through so-called tidal shocks, i.e., rapid variations of the strength of the external forces which occur during crossings of the Galactic disk or close passages of the Galactic bulge (Ostriker, Spitzer, & Chevalier 1972, Kundic & Ostriker 1995). Detailed simulations of globular cluster dynamics for a variety of masses and internal structures and different types of orbits in different Galactic model potentials have shown that tidal shocks accelerate the dynamical evolution of globular clusters and enhance their mass loss in such a way that 50% or more of the present-day globulars will be destroyed within the next Hubble time (Gnedin & Ostriker 1997, including disk-shocks; Baumgardt & Makino 2002, using a model without disk). Similarly, it has been shown that the present sample of globular clusters is most likely the remainder of an initially much more numerous system of clusters, many of which are meanwhile dissolved (Murali & Weinberg 1997, Fall & Zhang 2001). In this way, the spatial distribution, kinematics, and mass function of the globular cluster system may have changed a lot. Also, the shape of the stellar mass function of individual clusters may have changed considerably since their formation because of preferential depletion of low-mass stars as a result of mass segregation (Baumgardt & Makino 2002).
Observations suggest that the Milky Way globular clusters are indeed spatially truncated by the Galactic tidal field. Measurements of the radial surface density profiles or surface brightness profiles of globular clusters (e.g., King et al. 1968; Trager, King & Djorgovski 1995) showed that many globulars have profiles that decline more steeply than a power law and, by extrapolation, suggest the existence of a finite boundary. Their profiles are often well fit by King (1966) models, which are of finite size. The estimated limiting radii obtained by extrapolation with King models were found to correlate with the clusters’ galactocentric distances, and the way in which they correlate agrees to what would be expected for tidal radii in a Galaxy with a flat rotation curve (Djorgovski 1995). Nevertheless, a firm observational proof for the predicted mass loss and dissolution of globular clusters in the Galactic tidal field has been missing until recently.
Using color-magnitude selected star counts, Grillmair et al. (1995) measured the stellar surface densities of a number of globular clusters to lower levels and hence larger radii than earlier studies. They then found that at very low levels (typically more than four orders of magnitude below the central surface density) the observed profiles frequently exceed the prediction from King models and extend beyond the limiting radius of these models. Similar results were obtained by Lehmann & Scholz (1997), Testa et al. (2000), Leon et al. (2000), and Siegel et al. (2001). The observed departures from King models are in some cases associated with a break in the logarithmic slope of the profile, which resembles the results of numerical simulations of globular clusters, where a break in the radial surface density profile marks the transition between the bound part and the unbound part of the cluster population (e.g., Oh & Lin 1992, Johnston et al. 1999a). Hence, these observations suggest that many clusters are surrounded by weak haloes or tails of unbound stars that might result from tidal stripping. On the other hand, the two dimensional surface density distributions obtained by Grillmair et al. (1995), Testa et al. (2000), and Leon et al. (2000) did not clearly confirm this suggestion and left doubts about the reality of the observed structures because these were in most cases too complex and diffuse to be unambiguously identifiable as tails of tidal debris. In fact, contamination by galaxy clustering in the background or by variable extinction across the field may have lead to spurious detections of such tails. The latter was recently demonstrated by Law et al. (2003) for the low-latitude cluster $\omega$ Cen.
Among the much bigger and more massive dwarf spheroidal (dSph) satellites in the Milky Way halo, there has been growing evidence that at least the Sagittarius dSph, which is the nearest of these systems, is subject to very substantial mass loss and produces strong tails of tidal debris. The stellar stream from this dSph has meanwhile been detected around the whole celestial sphere (see Majewski et al. 2003 and references therein).
Turning back to the globular clusters, deep small-field studies with the Hubble Space Telescope (HST) revealed that several of the Galactic globulars have luminosity functions that are unusually flat or even decreasing towards the low-luminosity end (Piotto, Cool, & King 1997; De Marchi et al. 1999; Piotto & Zoccali 1999; Grillmair & Smith 2001). This deficiency in low-mass stars could be an indication of tidal mass loss (when combined with mass segregation in the cluster) and has frequently been interpreted in this sense. However this is not by itself a proof of tidal mass loss because (1) the observations do not necessarily represent the overall luminosity function of the cluster (spatial variations due to mass segregation, either dynamical or primordial), and (2) it might be possible that intrinsic differences exist between the overall luminosity functions of different clusters.
The evolution of a cluster depends on its internal parameters and its orbit. Ostriker & Gnedin (1997) presented so-called ’vitality diagrams’ for globular clusters in the parameter space of half-mass radius, mass, and Galactocentric distance, showing in which region of this space a cluster should lie in order to survive more than 10 Gyr. Clusters that do not lie in this region, and hence are expected to dissolve due to disk- and bulge shocks, are those with large half-mass radius and low mass. An extreme example for such an object is the sparse cluster Palomar5, which has a mass of less than $10^4 M_\odot$ and a half mass radius of about 20 pc. Since Pal5 is also one of the clusters that were found to have an atypically flat luminosity function (Smith et al. 1986, Grillmair & Smith 2001) it presents a particularly interesting test case for tidally-induced mass loss.
The commissioning of the Sloan Digital Sky Survey (SDSS; York et al. 2000) provided deep multi-color CCD imaging in $2.5^\circ$ wide stripe across Pal5. This allowed a wide-field search for cluster stars in the surroundings of Pal5 (Odenkirchen et al. 2001a, hereafter Paper I). A previous investigation of Pal5 using photographic plates (Leon et al. 2000) had been compromised by severe problems with contamination from background galaxies. The SDSS observations, however, enabled excellent separation between stars and background galaxies and an efficient selection of cluster stars by color and magnitude. We found strong evidence for two massive tails of tidal debris emerging from Pal5. These tails showed a well-defined characteristic shape and were found to contain about half as much mass as the cluster.
The detection of such tails with clearly identifiable structure has two important aspects: (1) It provides conclusive, direct proof for on-going tidal mass loss in a globular cluster. (2) It reveals unique information on the orbit of the cluster and opens a very promising way for investigating the gravitational potential in the Galactic halo (e.g., Murali & Dubinski 1999, Johnston et al. 1999b). In the present paper we describe the analysis of further SDSS data for a more than five times larger field around Pal5, which have become available since Paper I. The goal of this study is to trace the tidal debris of Pal5 to larger distances from the cluster in order to obtain a more complete census of its mass loss and to constrain the basic properties of the distribution of the debris such as its shape and density profile. As we will show, Pal5 is the first globular cluster that exhibits fully-fledged tidal tails with a total angular extent of $10^\circ$ on the sky.
In §2 we provide details about the observations and the photometric data derived from them. §3 describes the methods used to analyse the data. In §4 we present the resulting surface density distribution of cluster stars and describe the basic features of the tidal tails. §5 deals with the determination of the cluster’s local orbit and its extrapolation to a global scale. In §6 we derive estimates of the mass loss rate and the total mass loss of the cluster. The results are discussed and summarized in §7, and a brief outlook on future work is given in §8.
Observations
============
The SDSS is a large deep CCD survey designed to cover 10,000 square degrees of sky by imaging in five optical passbands, and by spectroscopy. The imaging data are obtained in great-circle drift scans using a large mosaic camera on a dedicated 2.5m telescope at Apache Point Observatory, Arizona. (For further information on the survey and its technical details see York et al. 2000, Gunn et al. 1998, Fukugita et al. 1996, Hogg et al. 2001, Smith et al. 2002, and Pier et al. 2002).
The data that we use in this study stem from the SDSS imaging runs 745, 752, 1458, 1478, 2190, 2327, 2334, and 2379, carried out between March 1999 and June 2001. The various strips of sky scanned in these runs yield complete coverage of a $6\fdg5$ to $8^\circ$ wide zone along the equator in the right ascension range from $224^\circ$ to $236^\circ$. Hereby we have multi-color photometry for Pal5 and its surroundings in a contiguous, trapezium-shaped field with an area of 87 square degrees. The observations reach down to an average magnitude limit of about 23.0 mag in $i^*$ and 23.5 mag in $r^*$ (approximate limits of 90% incompleteness). Photometric and astrometric data reduction and object classification were done by the standard SDSS image processing pipeline (see Lupton et al. 2001 and Pier et al. 2002 for different parts of the pipeline). The photometry used here is from before the public data release DR1 and hence does not precisely match the final SDSS photometric system.[^1] However, the preliminary photometric calibration of the data is spatially uniform to about 3% (Stoughton et al. 2002). The lack of the final calibration does not affect our study since we use the photometry in a purely empirical and differential way.
Our investigation is restricted to objects classified as unresolved sources (thus eliminating a large number of background galaxies that would otherwise contaminate the star counts) and uses object magnitudes derived by point-spread function (PSF) fitting. The median internal errors of the magnitudes in $g^*$, $r^*$, and $i^*$ are 0.015 mag or better for stars brighter than 18.0 mag (in $g^*,r^*,i^*$ respectively), rise to values between 0.023 and 0.035 mag at magnitude 20.0, and reach the level of 0.10 to 0.17 mag at magnitude 22.0. We confirmed these errors by analysing repeated measurements in overlapping scans. The median differences between magnitudes from independent observations are between $1.0\times$ and $1.2\times$ the median internal errors, showing that the quoted median errors provide reliable estimates of the photometric accuracy of these data.
According to the dust maps of Schlegel, Finkbeiner & Davis (1998) the southern and south-eastern part of the field is affected by a considerable amount of interstellar extinction while in the northwestern part the extinction is much lower. More specifically, the extinction in the $g$ band varies from 0.15 mag at the northern edge to 0.75 mag at the southern edge of the field. This corresponds to variations in the reddening of the color index $g-i$ in the range $0.07 \le E(g-i) \le 0.34$. To remove these variations from the observed magnitudes we applied individual extinction corrections derived from the local $E(B-V)$ reddening given by Schlegel, Finkbeiner & Davis. The resulting dereddened magnitudes should properly be named $g_0^*$ etc., however, for simplicity, we will suppress the index 0 here. Since the reddening data from Schlegel et al. represent the integrated extinction along the entire line of sight, the magnitudes of stars that are in front of the bulk of intervening material would in this way become overcorrected. However this is unlikely to happen for stars belonging to Pal5, which are located more than 20 kpc from the sun and seen far behind the northern part of the Galactic bulge. Color-magnitude diagrams for different parts of the field with different amounts of extinction show, that there is no sign of overcorrection in the blue edge (main-sequence turn-off) of the halo field star population. In the case of significant overcorrection this edge would be inclined to the blue with increasing brightness, which is not observed.
Due to variations in the observing conditions the completeness of object detection at faint magnitudes is somewhat different from run to run. This causes artificial inhomogeneities in the stellar surface density of the faintest stars. In order to avoid such effects it was necessary to cut the sample at $i^* = 21.8$ mag. At the bright end we chose a threshold of $i^* = 15.0$ mag because brighter stars risk to have saturated images and because none of the giants in Pal5 is brighter than this limit. The resulting data set contains about 940,000 point sources.
Photometric object filtering
============================
Unfiltered sample
-----------------
The full set of SDSS point sources with magnitudes in the range $15.0
\le i^* \le 21.8$ has an average surface density of 3.0 stars per arcmin$^2$ and a large-scale surface density gradient of 0.13 stars per arcmin$^2$ per degree in the direction of decreasing galactic latitude. Figure 1 shows a map of the stellar surface density derived by source counts in pixels of $3'\times 3'$. There are two strong local density enhancements in this field, at positions ($229\fdg0$,$-0\fdg1$) and (2296,$+2\fdg1$) in right ascension, declination (J2000). The first one represents the remote cluster Pal5 while the second one is due to the much closer and much richer globular M5. The latter is not relevant for this paper, except that one must avoid this region when investigating the properties of the general field around Pal5. The peak surface density of cluster stars in the center of Pal5 is 25.2 arcmin$^{-2}$, i.e., $8.4\times$ the mean density of the surrounding field.
Figure 1 also reveals an arc of very weak overdensity extending northeast and southwest of Pal5. The results that we will present in §4.1 confirm that these are rudimentary traces of extended debris from Pal5. This is remarkable since it means that weak traces of the cluster’s debris are visible even without any particular photometric filtering. However, the surface density of these features is only on the level of $1\sigma$ (rms of surface density in individual pixels) above the local mean surface density of the field.
In order to get detailed information on the distribution of stars from Pal5 one needs to enhance the contrast between the cluster and the field, in particular the dominating Galactic foreground. To first order, this can be achieved through simple cuts in color and magnitude. A more efficient variant of this method is to use an appropriately shaped polygonal mask in color-magnitude (hereafter c-m) space. This approach was taken, e.g., by Grillmair et al. (1995), Leon et al. (2000), and in Paper I in the context of globular clusters, and also by Majewski et al. (2000) and Palma et al. (2003) in the context of dwarf spheroidal (dSph) galaxies. However, sharply defined cuts or windows in c-m space do still not provide the optimal method to map out the spatial distribution of the stellar population of Pal5 because each star is treated as either a definite member or a definite non-member of the cluster population. This does not completely exploit the available information. Since the photometry actually allows us to derive smoothly varying membership probabilities as a function of color and magnitude one can optimize the object filtering by using these probabilities.
Optimal contrast filtering
--------------------------
A straight-forward way to make comprehensive use of the photometric information is to construct empirical c-m density distribution templates $f_C(m,c)$, $f_F(m,c)$ for the cluster and the field ($m$ and $c$ denoting magnitude and color index), and to use these to determine the surface density $\Sigma_C$ of the cluster population for each position in the field by a weighted least-squares adjustment. This kind of approach was, e.g., described by Kuhn, Smith, & Hawley (1996) in a study of the Carina dSph, and recently discussed in more detail by Rockosi et al. (2002), who used it for an analysis of a smaller SDSS data set on Pal5. The adjustment is done such that the weighted sum of field stars plus cluster stars yields the best approximation of the observed total c-m distribution. In the present study we have applied this method of optimal data filtering in the following way:
Using the magnitudes $g^*$, $r^*$, and $i^*$ (which provide higher accuracies than $u^*$ and $z^*$), we first defined orthogonal color indices $c_1$ and $c_2$ as in Paper I:
$$\begin{aligned}
c_1 & = & \ 0.907 (g^* - r^*) + 0.421 (r^* - i^*)\ \ \\
c_2 & = & -0.421 (g^* - r^*) + 0.907 (r^* - i^*)\ \\end{aligned}$$
The choice of the indices is such that the main axis of the almost one-dimensional locus of Pal5 stars in the $(g-r)$ versus $(r-i)$ color-color plane lies along the $c_1$-axis while the $c_2$ axis is perpendicular to it. We then preselected the sample in $c_2$ by discarding all objects with $|c_2| > 2\sigma_{c_2}(i^*)$, where $\sigma_{c_2}(i^*)$ is the rms dispersion in $c_2$ for stars of magnitude $i^*$ in Pal5. Stars with these $c_2$ colors are unlikely to be from the cluster. We also preselected in $c_1$ by restricting the sample to the range $0.0 \le c_1 \le 1.0$ because one expects very few stars of the cluster population outside this range.
Next we constructed c-m density diagrams (Hess diagrams) for the cluster and the field by sampling the stars that lie within $12'$ from the center of Pal5 (cluster diagram) and those that are more than $1^\circ$ away from the location of the stream and outside M5 (field diagram) on a grid in the plane of $c_1$ and $i^*$. Bins of 0.01mag$\times$0.05mag were used and the counts were smoothed with a parabolic kernel of radius 3 pixels.[^2] The cluster c-m distribution was corrected for the presence of field stars in the $12'$ circle around the cluster center by subtracting the field c-m distribution in appropriate proportion. The resulting diagrams of the normalized c-m densities $f_C$, $f_F$ of cluster stars and field stars are shown in Figures 2a and 2b. The cluster members are concentrated along well-defined branches (giant branch, horizontal branch, subgiant branch and main sequence) while the field star distribution is more diffuse, showing local maxima along $c_1 \approx 0.4$, which can be attributed to the turn-off region of thick disk and halo.[^3] By comparing the field star c-m distribution in the region northwest of the cluster to that in the region southeast of the cluster it appears that $f_F$ is not strictly constant over the field. However, the differences are not dramatic because the deviations from the mean distribution are mostly below 10%. Since a more local estimate of $f_F$ can only be obtained at the cost of higher noise or lower c-m resolution we preferred to neglect the spatial variations and to work with the mean distribution shown in Figure 2b.
In order to derive the surface density distribution of cluster stars on the sky one needs a mathematical model that provides a link to the observed distributions. The general ansatz for the stellar density in the hyperspace spanned by the celestial sphere and the c-m plane is a sum of densities $S_C$ and $S_F$ for the cluster and the field (i.e., non-cluster stars)
$$\begin{aligned}
S(\alpha,\delta,m,c)
& = & S_C + S_F \\
S_C & = & \Sigma_C(\alpha,\delta)\Phi_C(m,c) \\
S_F & = & \Sigma_F(\alpha,\delta)\Phi_F(\alpha,\delta,m,c)\end{aligned}$$
where each component can be split up into a product of a surface density $\Sigma$ on the sphere and a position-dependent normalized c-m density $\Phi$. ($\alpha,\delta$ denote coordinates on the celestial sphere, $m$ and $c$ denote magnitude and color index.) For the cluster component as a sample of stars of common origin we assume that (1) it is everywhere composed of the same mix of stellar types and that (2) all stars are at practically the same distance from the observer. This implies that $\Phi_C$ does in fact not depend on position ($\alpha,\delta$) (as in equation 2b). In contrast to this the field component is an inhomogeneous sample, i.e., its composition by stars of different types and its density distribution along the line of sight are spatially variable, so that $\Phi_F$ must in principle vary with position on the sky (equation 2c).
Let $j$ be an index labelling the pixels of a grid in the c-m plane and $k$ be an index labelling the positions $(\alpha_k,\delta_k)$ of a grid on the sky, then the number $\nu(k,j)$ of stars lying in a solid angle $\Omega_k$ centered on $(\alpha_k,\delta_k)$ and with magnitude and color falling on the pixel $j$ of area $P_j$ is obtained by integrating equation (2a) over $\Omega_k$ and $P_j$, i.e.
$$\begin{aligned}
\nu(k,j) &=& \nu_C(k) f_C(j) + \nu_F(k) f_F(k,j) \quad\\
\nonumber\\
\mathrm{where} \nonumber\\
\nu_{\,C,F}(k)&=&\int_{\Omega_k} S_{\,C,F}\ d\Omega \nonumber\\
f_C(j)&=&\int_{P_j} \Phi_C\ dm\,dc \nonumber\\
f_F(k,j)&=&\frac{1}{\nu_F(k)}\int_{\Omega_k} \int_{P_j}
\Phi_F\ dm\,dc\,d\Omega \nonumber\end{aligned}$$
Although $f_F$ does in principle depend on position index k, this dependence is in our case not important because we can assume that substantial changes in the characteristics of the field population occur only on larger scales and that hence (as the observations show) $f_F$ is approximately constant within the chosen field. [^4] The distributions $f_C$ and $f_F$ in the model of equation (3) can thus be represented by the above normalized average c-m distributions that have been drawn from the observations. $\nu_C(k)$ and $\nu_F(k)$ are the numbers of cluster stars and field stars in $\Omega_k$, the former of them being the target of our analysis. Apart from observational noise (and apart from small deviations due to spatial variations in the c-m distribution of the field stars, which the model neglects), the left hand side of equation (3) should correspond to the observed star counts $n(k,j)$ in $\Omega_k \times P_j$. Thus one can plug in $n(k,j)$ for the expected number $\nu(k,j)$ in equation (3). Equation (3) then does not have an exact solution. However, we can determine a least-squares solution for $\nu_C(k)$ by demanding that the square sum of the noise-weighted deviations between the observed number $n(k,m)$ and the expected number $\nu(k,m)$ given by the right hand side of equation (3), summed up over the c-m grid, is minimal. Since the contribution of the cluster population to the total counts is small (outside the cluster) we assume the noise to be dominated by the field stars, i.e., we expect $\sigma_n^2(k,m) = \nu_F(k) f_F(m)$. The sum of weighted squares to be minimized thus is:
$$\begin{aligned}
\chi^2(k) = \hspace*{5.5cm}\\
\quad \sum_j \frac{\left(n(k,j) - \nu_C(k) f_C (j) -
\nu_F(k) f_F(j)\right)^2}{\nu_F(k) f_F(j)} \nonumber\end{aligned}$$
It is straightforward to calculate (via $d\chi^2/d\nu_C = 0$) that the least-squares solution for $\nu_C(k)$, which we call $n_C(k)$, and its variance $\sigma^2_{n_C}(k)$ are:
$$\begin{aligned}
n_C(k) = \frac{\displaystyle \left(\sum_j n(k,j)\frac{f_C(j)}{f_F(j)}
\right) - n_F(k)}{\displaystyle \sum_j\frac{f_C^2(j)}{f_F(j)}}\\
\sigma^2_{n_C}(k) = \quad n_F(k) \bigg/ \sum_j
\frac{f^2_C(j)}{f_F(j)} \qquad\end{aligned}$$
In principle, one could determine both $n_C$ and $n_F$ (i.e., the best estimate for $\nu_F$) in this way by minimizing the $\chi^2$ of equation (4). However, for $\nu_F$ we already know (or can safely assume) that it must be a smoothly varying function of position that can be described by a simple (polynomial) model. Thus we preferred to use this constraint to determine $n_F$ externally (for practical details see §4.1) and not in a simultaneous least-squares adjustment with $\nu_C$. This makes the solution for $\nu_C$ more robust.
Equation (5a) allows the following interpretation: One finds the least-squares solution $n_C(k)$ by weighting each star in the solid angle $\Omega_k$ by the quotient $w(j)=f_C(j)/f_F(j)$ according to its position in the c-m plane, summing up the weights of all stars, and dividing this sum by the factor $a = \sum_j (f_C^2/f_F)$. This yields the estimated number of cluster stars $n_C$ plus a term $n_F/a$, i.e., the estimated number of field stars attenuated by $a$. By subtracting this residual field star contribution one obtains $n_C$. Equation (5b) shows that the variance of $n_C$ is reduced to $1/a$ times the variance of the field star counts. In other words, the noise in the determination of the surface density of cluster stars decreases by the factor $\sqrt{a}$. The weight function $w=f_C/f_F$ is shown by the contour plot in Figure 2c. We obtain an attenuation factor $a=5.1$ and hence a noise reduction of $\sqrt{a} = 2.3$ with respect to the unfiltered, but preselected sample. In total, i.e., in comparison with the full sample, the noise level is reduced by a factor of 4.3.\
The tidal stream
================
Surface density map
-------------------
We constructed a map of the stellar surface density of Pal5 stars by applying the above method of least squares estimation on a grid with pixels of $3'\times 3'$ in the plane of the sky. The residual contribution from field stars was determined by fitting a bi-linear background model to the weighted counts in those pixels that are at least $1^\circ$ away from the cluster and the tails, and also away from M5. After pre-selection and weighting, the mean surface density of the field stars is 0.16 arcmin$^{-2}$ and the surface density gradient across the field is about $5\times 10^{-3}$ arcmin$^{-2}$ per degree. By subtracting the best-fit bi-linear background model the density distribution in the field becomes essentially flat. For further reduction of the noise the counts were smoothed by weighted averaging with a parabolic kernel of radius 4 pixels. In regions with surface density above the 5$\sigma$ level we used a kernel with a smaller radius to preserve higher resolution. Figure 3 presents the resulting surface density map as a grey scale and contour plot in equatorial celestial coordinates (i.e., right ascension and declination).
This map shows a striking coherent structure that is spatially connected to the main body of Pal5 and has stellar surface densities varying from $1.5 \sigma$ up to $5\sigma$ ($\sim$ 0.12 arcmin$^{-2}$) and higher. The geometry relative to the cluster clearly identifies this structure as debris from Pal5. The debris forms two long narrow tails on opposite sides of the cluster and extends over an arc of about 10$^\circ$. This corresponds to a projected length of 4 kpc at the distance of the cluster. The tails have a width of about $0\fdg 7$ in projection on the plane of the sky. Apart from small scale variations the width of the tails does not systematically change with angular distance from the cluster. The northern tail, which - as will be shown later in §5.2 - is trailing behind the cluster, is traced out to an angular distance of at least $5\fdg 8$ from the center of the cluster, but possibly out to $6\fdg 5$, and appears slightly curved. The maximum surface density of stars in this tail is about 0.2 arcmin$^{-2}$. It occurs at angular distances between $2\fdg 2$ and $3\fdg 7$ from the cluster center and was hence not covered by the initial detection of the tails in Paper I.
The southern tail, which is the one that leads the motion of the cluster (see §5.2), is seen over $3.6^\circ$ and reaches down to the edge of the currently available field. This suggests that the tail continues beyond this limit, as one would indeed expect when assuming approximate point symmetry in the distribution of debris with respect to the cluster. The southern tail exhibits density maxima at angular distances of $1\fdg 6$ and $3\fdg 5$ from the cluster center, which are, however, less pronounced than in the northern (trailing) tail. The transition between the cluster and the tails is not straight. Instead we see a characteristic S-shape in the distribution of stars, which closely resembles the structures seen in simulations of disrupting globular clusters and satellites (e.g., Combes et al. 1999, Johnston et al. 2002). This S-shape feature clearly indicates that the stars are stripped from the outer part of the cluster by the Galactic tidal field dragging them away in the direction towards the Galactic center and anticenter.
Besides the two tails of Pal5 and a spurious patch of high stellar density left over from the cluster M5, the map shows a number of small isolated patches with densities on the level of $2\sigma$ above zero, which are dispersed over the field. These are most likely not traces of the population of Pal5 but the result of random fluctuations in the distribution of field stars. To test this we generated random fields with the same mean density as the observed residual field star density and sampled these artificial fields in exactly the same way as done with the observations. This Monte Carlo experiment showed that random fields yield $2\sigma$-patches of the same size and with very similar number densities as in Figure 3. As another statistical test we resampled the observations on a grid of non-overlapping pixels with a size of $9' \times 9'$ and determined the frequency distribution of pixel counts in those regions that lie outside the clusters M5 and Pal5 and the tails of Pal5. This distribution closely agrees with the expected Poisson distribution for a random field of the given mean density (see similar tests of fluctuations described in Odenkirchen et al. 2001b). Both tests reveal that the isolated patches in the map of Figure 3 provide no evidence for further significant local overdensities.
In order to estimate the fraction of stars in the tails and in the cluster, we integrated the surface density of Pal5 stars in a $42'$ wide band covering the tails (2.3 times the FWHM of the tails, see §6.1) and in a circle of radius $12'$ around the center of the cluster. A somewhat smaller radius than the cluster’s limiting (or tidal) radius of $16'$ (Odenkirchen et al. 2002, hereafter Paper II) was used because the bound and unbound part of the cluster overlap in projection on the sky and cannot be strictly separated. We find that the number of stars in the tails is about 1650 while the number of stars in the cluster is about 1350. This yields a number ratio between tails and cluster of $\beta = 1.22$. The number of stars seen in the southern (leading) tail is about half the number found in the northern (trailing) tail (i.e., there are about 1100 stars in the northern and about 550 stars in the southern tail). To check these number ratios we also took an alternative approach and performed integrated number counts on a variety of samples defined by different choices of cut-off lines in the c-m plane. Hereby we obtained values for the ratio between the number of stars in the tails and in the cluster in the range from 1.18 to 1.31. It thus appears that $\beta = 1.25\pm 0.06$ is a robust estimate for the observed field. Since the tails may easily extend beyond the area currently covered by the SDSS, the ’true’ ratio is likely to be higher. In any case, we can safely conclude that the tails contain more stellar mass than the cluster.
Density profile along the tail
------------------------------
Since the debris forms a long and relatively thin structure, it makes sense to treat it as a one-dimensional object and to characterize it by its distribution of linear density. To determine this linear density we modelled the central line of each tidal tail by a sequence of short straight-line segments (fitting by eye) and projected all stars within $0\fdg 35$ from the central line perpendicularly onto it. We then performed weighted star counts as described in equation (4a) in bins of the arc length parameter $\lambda$ along the central line, using a bin size of $0\fdg 25$. The field star contribution was determined using the bilinear background model from Section 4.1, and was subtracted from the counts. We thus obtained the density distributions shown in Figure 4 (statistical uncertainty of the number counts indicated by error bars).
The density curve for the northern (trailing) tail (Fig.4a) shows three pronounced maxima, which correspond to extended density clumps around RA $230\fdg 9$, $231\fdg 8$, and $233\fdg 3$ in the map of Figure 3. The linear density at these maxima is about two times as high as it is on average. Apart from those local variations there is a general decline in the density with increasing $\lambda$. The mean density level decreases by roughly a factor of 3 when comparing $\lambda\sim 0\fdg 5$ to $\lambda\sim 6^\circ$.
In order to judge the significance of the observed variations we approximate the general trend in the data by fitting a straight line to the innermost five and the outermost five data points (dashed line in Fig.4a). Comparing the data with this line shows that the smaller amplitude variations in the counts lie within the error bars and hence are likely of statistical nature whereas the strong maxima at the above given locations exceed the straight-line model by about 3 times the error bar. Therefore, these maxima are statistically significant and present real substructure in the tail. Two of the clumps may in fact be part of one broad density enhancement because their separation by only one bin of lower density could be the result of statistical fluctuations.
Along the southern (leading) tail (Fig.4b) the linear density is generally lower than in the corresponding part of the northern (trailing) tail. Again, there are local variations which reflect the presence of density clumps in Figure 3. However, these variations have lower amplitude than those occuring in the northern tail, and the deviations from the general trend of the data (fit of straight line to entire set of data points) are thus not highly significant. In particular, the southern (leading) tail shows no obvious counterpart to the broad density enhancement in the northern (trailing) tail. Since the data for the southern (leading) tail cover a smaller range in $\lambda$ there is less information on the large-scale trend of the density. With the exception of the outermost bin the data points seem to suggest a weak outward decrease. On the other hand, taking into account the error bars the counts are also consistent with the assumption of a constant mean density level. In any case, the density curve leaves no doubt that the southern tail must continue beyond the border of the field. The steep rise of the counts in the outermost bin, be it a statistical fluctuation or due to a real clump, shows that the mean density does not drop to zero at this point. Whether or not the linear density at higher $\lambda$ declines in a similar way as seen in the outer part of the northern (trailing) tail is an interesting question that can presently not be answered.
Radial profile of the surface density
-------------------------------------
Another way to describe the tidal debris is by determining the radial profile of the surface density, i.e., the azimuthally averaged surface density as a function of distance from the cluster center. This description disregards the fact that tidal tails are not a circularly symmetric structure, but has the advantage to provide a uniform view of both the cluster and the debris. Therefore, observational studies of globular clusters and local dwarf galaxies are often judging the existence of tidal debris in this way, and results of theoretical studies are also frequently presented in this form (e.g., Johnston et al. 1999a, Johnston et al. 2002).
We derived the radial profile of the cluster and the two tails through weighted number counts in sectors of concentric rings. Out to $r = 15'$ we divided each ring into its northern and southern half. At larger radii we used progressively narrower sectors to bracket the tails and to minimize the influence of the field, but referred the (background corrected) counts to the full area of the corresponding half ring. This yields the profiles shown in Figure 5. For comparison we also show an analogous profile obtained in two cones away from the tidal tails, i.e., at position angles $100^\circ\pm 35^\circ$ and $280^\circ\pm 35^\circ$.
It is clearly visible that the tidal debris is distinguished from the cluster by a characteristic break in the slope of the logarithmically plotted radial profile. Outside the cluster’s core region, i.e., at radii $r > 3'$, the surface density first decreases steeply as a power law $r^\gamma$ with exponent $\gamma = -3$. Between $15'$ and $20'$ there is a transition region where the profile becomes less steep, and from $20'$ outwards the decline of the density is similar to a power law with an exponent in the range $-1.5 < \gamma < -1.2$. The comparison profile, which has been measured perpendicularly to the tails and should thus not be affected by tidal debris, shows the same $r^{-3}$ power law decline between $3'$ and $10'$ but falls off more steeply at $r > 10'$. This shows that perpendicular to the tails the cluster has a well-defined radial limit. A fit of a King profile to these counts suggests a limiting (or tidal) radius of approximately $16'$ (see Paper II). This is near to the radius where the overall radial profile shows the break. By comparing the different radial profiles the tidal perturbation of the cluster is noticable from about $r = 12'$ outwards.
To determine the power law exponent for the outer part of the radial profile we made a weighted least-squares fit to the data points at $r \ge 20'$. For the southern (leading) tail this fit yields $\gamma = -1.25 \pm 0.06$. For the northern (trailing) tail the use of all data points results in a poor fit with $\gamma = -1.36$. When leaving out the three most discrepant data points, which describe the strong local density maximum in the range $140' < r < 220'$, we obtain an acceptable fit and $\gamma = -1.46 \pm 0.06$. The overall decline of the radial surface density profile of the northern (trailing) tail is thus somewhat steeper than for the southern (leading) tail. For both tails we find power law exponents $\gamma < -1$, which means that the decline is steeper than it would be for a stream of constant linear density (having a radial profile $\propto r^{-1}$ because the area of the averaging annuli increases proportional to $r$). This confirms that the linear density of the stream is decreasing with angular distance from the cluster as stated in §4.2. On the other hand, it also reveals that the decrease in linear density is distinctly less steep than $1/r$ because we find $\gamma \ge -1.5$.
Distances
---------
It is important to recall that our mapping of the tidal debris is built on the assumption that the debris is located at the same heliocentric distance as the cluster (at least within the limits of the photometric accuracy and the natural photometric dispersions). For the immediate vicinity of the cluster this necessarily holds true. With increasing angular distance from the cluster the heliocentric distances might however increasingly deviate, depending on how much the tidal stream is inclined against the plane of the sky. If, for example, this inclination were $\ge 50^\circ$ the distances should differ by $\ge 10\%$ over an angle of $5^\circ$, resulting in shifts of $\pm 0.2$mag or more in apparent magnitude. One might suspect that shifts of this size, if true, could affect our measurements of the stellar surface density along the tails. On the other hand if such shifts in apparent magnitude were detectable, this would also provide interesting constraints on the extent of the tidal debris and the cluster’s orbit in the third dimension .
Unfortunately, the stars that we have access to in the tails are not well suited as precise distance indicators. In order to measure small distance effects we would ideally need stars with characteristic luminosities such as horizontal branch (HB) stars. These are not very numerous, even in the main body of the cluster ($\approx$ 30 HB candidates within $12'$ from the center including variables), and occur mostly on the red side of the HB. In the tails an occasional red HB star from Pal5 would (in the absence of kinematic information) be indistinguishable from Galactic field stars. The subgiant branch is also not sufficiently well populated to allow such cluster members to be recognized on a purely statistical basis. Therefore, one has to rely on stars near and below the main-sequence turn-off, whose luminosities cover a wider range. Even for stars of this type one needs to integrate over a substantial part of the tails in order to be able to identify their location in the c-m plane. Therefore, distance variations can only be investigated at low angular resolution.
In Figure 6 we present Hess diagrams for the outer parts of the two tails, obtained by sampling stars in two $18'$ wide bands ($\approx$ the FWHM of the tails, see §6.1) along the ridge lines of the tails. Panel (a) of this figure shows the integrated c-m distribution in the northern (trailing) tail between 35 and 56 from the center of Pal5, while panel (b) shows the same for the southern (leading) tail from 15 to its outer end. The two samples, which have almost the same size, are thus spatially separated by an angle of at least $5^\circ$. For comparison with the cluster, the ridge line of the cluster’s c-m distribution as derived from Figure 2a is overplotted (middle dot-dashed line) and repeated with magnitude offsets of $-0.2$ and +0.2 mag (upper and lower dot-dashed line, respectively), corresponding to a 10% smaller or larger mean distance of the stars. The location of the density maxima in these diagrams reveal that the outer part of the northern tail is centered on the same distance as the cluster while the outer part of the southern tail appears to be about 0.1 mag brighter. Hence its mean distance may be about 5% smaller. The fact that the c-m distribution of the southern tail sample extends to brighter magnitudes also appears to be influenced by field stars with $c_1 \approx 0.35$, which are seen to be more abundant than in the northern sample and spread into the locus of the cluster members. Thus the mean distance of the southern (leading) tail sample is probably not smaller than that of the cluster by as much as 10% (i.e., $-0.2$ mag) or more. Accepting a relative difference of 5% between the mean distances of the northern (trailing) and the southern (leading) tail sample and considering that the mean angular separation between those two samples is 71, the inclination of the tidal stream against the plane of the sky may be of the order of $22^\circ$. Since the data shown in Figure 6 do not support a difference $\ge 10\%$ in the mean distances of the two samples an inclination of $\ge 38^\circ$ can be excluded.
To determine how variations in heliocentric distance along the southern (leading) tail might influence the determination of the stellar surface density in the stream we shifted the color-magnitude distribution $f_C$ of the cluster by $-$0.1 mag and $-$0.2 mag, recomputed the weight function, and rederived the least-squares solution for the surface density. Figure 7 shows the resulting linear density profiles along the southern (leading) tail. One can see that the above magnitude shifts in the cluster template lead to slightly lower densities in most of the bins. The general trend of the data with arc length along the tail as determined by the best-fit straight line (dashed lines in Fig.7) does not change significantly. Only in the outermost bin (at $\lambda \sim 3\fdg6$) magnitude shifts of $-0.1$ and $-0.2$ mag produce an increase in the number density of stars such that the measured density exceeds the general trend by two times the statistical error. The general conclusion from this experiment thus is that despite a possible decrease of the distance along the southern (leading) tail of up to 10% (out to the tip of the tail) the assumption of constant distance as used in the previous sections does not cause a significant underestimation of the stellar surface density in the outer part of this tail.
Luminosity functions
--------------------
Another assumption in the filtering method described in §3.2 is that the tidal debris has the same luminosity function and c-m distribution as the cluster. This is not necessarily the case because there could be mass segregation effects (see §7.4). However, using star counts in a narrowly confined band containing the tails it can a posteriori be shown that down to our magnitude limit of $i^*=21.8$, which comprises only a small range in stellar mass, the assumption holds true. In Figure 8 we present the luminosity function of the stellar population in the tidal tails and compare it to the luminosity function of the stars in the cluster itself. These luminosity functions were obtained by restricting the star counts to an appropriate window around the loci of the giant branch, subgiant branch, and upper main sequence of Pal5 in the c-m plane. The luminosity function of the cluster was derived by counting stars within $r \le 6'$ from the cluster center. The luminosity function of the tails was obtained by counts within $\pm0\fdg25$ angular distance from the central line through the tails (see §4.2). Possible variations in the line-of-sight distances of the stars along the tails were neglected since their effect in apparent magnitude is small. The statistical contamination by intervening field stars was determined with counts in neighboring zones and subtracted after proper scaling with the respective areas. Finally, the luminosity function of the tails was renormalized in order to bring it to the same level as the cluster’s luminosity function in the magnitude bins centered on $i^* = 19.0$ and $i^*=19.5$ (renormalization factor of 100). Figure 8 shows that the two luminosity functions are almost identical from $i^*=19.0$ down to the faintest bin. At magnitudes brighter than 18.5 the number of stars in the tails is too small to decide whether or not the surface density is higher than in the field. However, within the statistical errors the counts agree with the number of giants in the cluster, which is also quite small. Anyway, stars of this brightness are unimportant in the filtering process. The agreement between the two luminosity functions proves that the filtering method is based on firm grounds.
Implications on the orbit of Pal5
=================================
Numerical simulations of globular clusters in external potentials demonstrate that stars tidally stripped off from such systems remain closely aligned with the orbit of the cluster (e.g., Combes et al. 1999). This happens because the stars have only small differential velocities and small spatial offsets when they become unbound from the cluster. N-body simulations for dSph satellites have shown that even debris from such more massive systems may remain on approximately the same orbit as the parent object over long time-scales (e.g., Johnston et al. 1996, Zhao et al. 1999). The stream of debris from Pal5 thus provides a unique tool for tracing the orbital path of this cluster and subsequently also its orbital kinematics. This offers an exceptional opportunity to probe the Milky Way’s potential with observations of a Galactic orbit. So far, only the Sgr dSph galaxy with its global tidal stream has allowed to constrain the Galactic potential in a similar way (see Ibata et al. 2001, Majewski et al. 2003). Classically, determinations of the potential of the Galactic halo have been based on statistical investigations of the bulk properties (spatial distribution and average velocities) of object samples, either halo stars, globular clusters, or dSph galaxies (e.g., Zaritsky et al. 1989, Kulessa & Lynden-Bell 1992, Wilkinson & Evans 1999). The availability of traces of individual orbits as for Pal5 and the Sgr dSph is clearly an advantage over the statistical approach because this allows to obtain information on the potential without assumptions on, e.g., steady state or velocity anisotropy and without dependence on proper motion measures, which are often unreliable.
Isochrone approximation
-----------------------
In order to make the connection between the cluster’s Galactic orbit and its tidal tails more transparent we outline a simple analytic approach. Herein we use a spherical logarithmic halo as a model of the Galactic potential and describe the orbits of the cluster and the debris stars with the so-called isochrone approximation (Dehnen 1999). In contrast to classical epicycle theory, whose application is limited to almost circular orbits, this method allows an accurate approximation of substantially eccentric orbits. A detailed description for the special case of a logarithmic potential is given in Appendix A. The key point is that using appropriate transformations of the radial coordinate $R$ (galactocentric distance) and the time parameter $t$, the radial motion can be described by a harmonic oscillation whose period $T_R$ is proportional to the radius $R_E$ of a circular orbit of equal energy. Furthermore, the eccentricity $e$ of the orbit is a function of the quotient $L/R_E$, where $L$ is the angular momentum.
Let us approximate the orbit of the cluster in this way. If at an arbitrary point on the orbit (say at $t = t_0$) we shift one star from the cluster radially from $R$ to $R' = \alpha R$ ($\alpha > 0$), and release it as an independent test particle having the same instantaneous space velocity vector as the cluster, then it follows from Eqs. (A13) to (A19) that the eccentricity of its orbit does not change and that the new orbital path of this particle is simply a scaled copy of the orbital path of the cluster, i.e., $R'(\varphi) = \alpha R(\varphi)$, $\varphi$ being the azimuth or phase angle. The motion along this path is characterized by a scaling relation
$$\begin{aligned}
t'(\varphi_2) - t'(\varphi_1) = \alpha (t(\varphi_2) - t(\varphi_1)) \end{aligned}$$
between the time parameter $t$ for the cluster and $t'$ for the shifted particle, or by the equivalent relation $T' = \alpha T$ for the periods of the radial oscillation (e.g., the time intervals between successive apocenter or pericenter passages). A star shifted outward ($\alpha > 1$) thus has a proportionally longer period and trails behind the cluster while a star shifted inward has a proportionally shorter period and leads the cluster.
We now apply this model to the tidal debris of a globular cluster: Stars leave the cluster by passing near the inner or the outer Lagrange point on the line connecting the cluster to the Galactic center, i.e., those points where a force balance between the internal field of the cluster and the external tidal field exists. They are likely to pass these points with small relative velocity because the internal velocity dispersion in the cluster is low, in particular in low-mass clusters like Pal5 ($\sigma_{los} < 0.7$, Paper II). Subsequently, the debris is decoupled from the cluster and behaves like a swarm of test particles that are radially offset from the cluster and released with almost the same galactocentric velocity vector as the cluster. In the framework of the above model and the ideal case of zero velocity dispersion this means that the cluster and its debris are on confocal orbits that are equal up to radial scaling but have different angular velocities and thus exhibit azimuthal shear.
If the separation $\delta\varphi$ between the azimuth angles of the shifted and the unshifted particle (i.e., between a debris star and the cluster) is small, the relation between $\delta\varphi$ and the time $\Delta t$ since the release of the shifted particle can be expressed in a simple formula. Provided that $\delta\varphi$ is small enough to ensure that the galactocentric distance $\alpha R$ along the orbit of the shifted particle and hence its angular velocity, which is $L / \alpha R^2$, can be considered as being approximately constant over $\delta\varphi$, it follows that
$$\begin{aligned}
\delta\varphi \approx \frac{L}{\alpha R^2} \delta t =
\frac{(\alpha-1)}{\alpha} \Delta t \frac{L}{R^2}\ .\end{aligned}$$
Here, $\delta t$ means the time lag that corresponds to $\delta\varphi$, for which equation (6) yields $\delta t = (\alpha - 1)\Delta t$. Note that equation (7) is independent of the value of the circular velocity $v_c$ of the potential. The relation shown in equation (7) is very useful because it provides a key for estimating the mass loss rate of the cluster (see §6).
Local orbit and tangential velocity
-----------------------------------
We now describe what observational constraints we have on the cluster’s local orbit, i.e., its orbit near the present position of the cluster and the location of its tails. Adopting $d=23.2$ kpc (Harris 1996) for the heliocentric distance of Pal5 and $R_\odot = 8.0$ kpc for the distance of the Sun from the Galactic center we derive the position of Pal5 in the Galaxy as $(x,y,z) =
(8.2,0.2,16.6)$ kpc. Here, $x,y,z$ denote right-handed galactocentric cartesian coordinates, with $y$ being parallel to the Galactic rotation of the local standard of rest (LSR) and $z$ pointing in the direction of the northern Galactic pole. In other words, the Sun has coordinates ($-$8.0,0.0,0.0) in this system. From the above position of Pal5 it follows that the inclination between the line of sight and the orbital plane of the cluster must be $\le 18^\circ$. On the other hand, our view of the orbital plane cannot be entirely edge-on because Figure 3 clearly shows the S-shape bending of the tidal debris near the cluster. This feature obviously reflects the opposite radial offsets between the two tails and the cluster. Considering the orientation of this S feature and the perspective of the observer, we infer that the orbit of the cluster (in projection on the plane of the sky) must be located east of the northern tail and west of the southern tail (referring to the equatorial coordinate system used in Fig.3).
The simple model from §5.1 tells us that the tidal debris should be on similar orbits as the cluster if velocity differences can be neglected. Taking into account the local symmetry of the tidal field, the limited range in azimuth angle $\varphi$ covered by the observations, and the relatively small angle between the orbital plane and the line of sight, one thus expects the offsets between the tails and the orbit of the cluster in projection on the tangential plane of the observer to be constant and of equal size on both sides of the cluster. An additional argument for this assumption is that the tails show a constant width, i.e., the projection does not reveal that they become wider as a function of angular distance from the cluster. If the mean (projected) separation between the tidal debris and the orbit of the cluster were increasing with angular distance from the cluster one should expect to see the tails to become wider, which is not the case. Therefore we continue the analysis under the assumption that the cluster’s projected orbit runs parallel to the two tails.
First of all, this sets a tight constraint on the direction of the cluster’s velocity vector in the tangential plane. The tails imply that the tangential motion of the cluster has a position angle of $231^\circ\pm 2^\circ$ with respect to the direction pointing to the northern equatorial pole, and $280^\circ\pm 2^\circ$ with respect to Galactic North (see Fig. 8). The orientation of this angle (i.e., $PA = 280^\circ$ and not $PA = 100^\circ$) follows when taking into account the direction to the Galactic center. Figure 9 shows the surface density map of the tails on a grid of galactic celestial coordinates $l\cos b,b$, where $l$ means galactic longitude and $b$ galactic latitude. Since the Galactic center ($l=0,b=0$) lies to the bottom of this plot, the tail that points to the right (also called the southern tail), must be the one at smaller galactocentric distance, which is thus leading, and the tail that points to the left (also called northern tail) be the more distant one, which trails behind. This means that the cluster is in prograde rotation about the Galaxy, in agreement with indications from different measurements of its absolute proper motion (Schweitzer, Cudworth & Majewski 1993; Scholz et al. 1999; Cudworth 1998 unpublished, cited in Dinescu et al. 1999).
Next we consider whether the observed part of the stream is long enough to see a deviation from straight line motion. Figure 9 demonstrates that this curvature is indeed detectable. The long dashed line shows the projection of a straight line in space (with position angle $280^\circ$) plotted over the surface density contours of the debris. The curvature of this line (due to projection onto the $(l\cos b,b)$ coordinate grid) is obviously too small to fit the tidal stream. Thus the curvature of the stream provides clear, direct evidence that the motion of the cluster is accelerated.
To further constrain the orbit, the radial velocity of the cluster in the Galactic frame is needed and an assumption on the acceleration field near the cluster has to be made. The cluster’s heliocentric radial velocity of $-58.7
\pm 0.2$ (Paper II), combined with a solar motion of $(U,V,W)_\odot =(10.0,5.3,7.2)$ (Dehnen & Binney 1998b, velocity components in our $x,y,z$ system), and $v_{\mathrm{LSR}} = 220$ for the rotation velocity of the local standard of rest yields a Galactic rest frame radial velocity of $-44.3$ (observer at rest at the present location of the Sun). The cluster’s absolute proper motion is not yet measured with comparable accuracy. However, the existing measurements (see the above references) can be used to derive rough limits for the tangential velocity. According to Cudworth’s measurement, which we consider to be the most reliable one, the absolute proper motion of Pal5 lies between 2.6 and 3.8 mas/yr (3-$\sigma$ limits). Using the above values for the cluster’s distance and the motion of the Sun and assuming the direction of the tangential velocity to be near $PA =
280^\circ$, this yields a lower and upper limit for $v_t$ (i.e., tangential velocity as seen by observer at rest at present location of the Sun) of 60 and 195 , respectively. To determine possible space velocity vectors for the cluster we thus combined the radial velocity with tangential velocities in the range from 50 to 200 and the direction $PA =280^\circ$.
The simplest way to model the local acceleration field near the position of the cluster is a spherically symmetric field with constant acceleration. Assuming that the circular velocity in the Galactic halo is between 150 and 250 , it follows that plausible values for the acceleration $a$ range from $(150\,{\mbox{km\,s$^{-1}$}})^2/R_{cl}$ to $(250\,{\mbox{km\,s$^{-1}$}})^2/R_{cl}$, with $R_{cl}$ being the cluster’s galactocentric distance. We thus adopted $a = (220\,{\mbox{km\,s$^{-1}$}})^2/18.5\,\mathrm{kpc}$ and used this to integrate the orbit locally for a sequence of tangential velocities covering the interval from 40 to 180 in steps of 5 to 10. Each orbit (more precisely its projection on the plane of the sky) was then compared with the tidal tails. It turned out that an orbit with a good fit to the geometry of the tails is obtained when using $v_t = 95\,{\mbox{km\,s$^{-1}$}}$. This orbit is shown by the solid line in Figure 9. Changing $v_t$ by $\pm 15\,{\mbox{km\,s$^{-1}$}}$ leads to orbits with significantly different curvature (short-dashed and dot-dashed line in Fig. 9) which fit the tidal tails less well. These cases are considered as the limits of the range of acceptable orbits. For other values of $a$ in the above range one can obtain orbits with essentially the same projected path when increasing or decreasing $v_t$ accordingly. This means that fitting a local orbit to the tidal tails sets up a relation between the acceleration $a$ and the cluster’s tangential velocity $v_t$. The value of $a$ can however not be determined in this way from the current data unless the velocity $v_t$ is known independently.
A more realistic model for the Galactic field near the cluster is given by the spherical logarithmic potential $\Phi = v_c^2 \ln R$, which yields a constant circular velocity $v_c$ rather than constant acceleration. We repeated the orbit integration using this model and again looked for the best accordance with the geometry of the tidal tails. It turns out that for similar sets of parameter values the orbits obtained with this potential have practically the same projected paths as those obtained with the $a=const$ model. This implies that local variations in the size of the acceleration vector have little influence on the determination of the cluster’s local orbit. Hence it does not matter which of the two models we use.
The relation between the circular velocity $v_c$ of the logarithmic potential (or the local acceleration $a=v+c^2/18.5$kpc) and the tangential velocity $v_t$ of the cluster, for which one obtains a local orbit with a projected path identical to the solid line in Figure 9, is shown in Figure 10. This relation is linear with a slope $dv_t/dv_c = 0.43$. A straightforward way to determine the parameter $v_c$ of the potential would be to obtain $v_t$ through a precise astrometric measurement of the cluster’s absolute proper motion. An accuracy of 5 in $v_c$ would require an accuracy of 2 in $v_t$, which at $d=23.2$ kpc corresponds to a proper motion error of 18 $\mu$as/y. This level of accuracy can be achieved in future astrometric space missions like SIM and GAIA . On the other hand, from the currently available proper motion measurements for Pal5 one can clearly not derive a meaningful constraint on $v_c$. The proper motion obtained by Cudworth yields $v_t = 135$ and would thus suggest $v_c > 300$ , which is uncomfortably high. The quoted error of this proper motion of 0.17 mas/y per component results in a typical uncertainty of 19 in $v_t$, so that $v_c$ cannot be determined to better than $\pm$47 from this data. A more comprehensive way of constraining the Galactic potential however consists in gathering kinematic data all along the tails and not only for the cluster. This will allow to use more complex models for the potential and to determine more than one parameter (e.g., the size and the direction of the local acceleration vector).
In spherical potentials, the simplest motion would be that on a circular orbit. However, there is no such solution in the above sequence of orbits because for the given position, radial velocity, and direction of tangential motion a circular orbit would require a tangential velocity of $v_t =885.5\,{\mbox{km\,s$^{-1}$}}$, the total velocity on this orbit then being $v_c=886.6\,{\mbox{km\,s$^{-1}$}}$. Apart from the fact that velocities of this size are very far from realistic, it turns out that the projection of the resulting orbit would be similar to that of motion on a straight line (see Fig.9) and hence not yield a good fit to the tidal tails. In order to allow a circular orbit with a velocity of 220 or less one needs to increase the position angle of the tangential velocity to $PA \ge 310^\circ$. However, the orbit would then deviate from the direction of the tails by at least $30^\circ$. A circular or nearly circular orbit is thus not an option.
We also checked for possible effects from a flattening of the Galactic potential. The potential in the Galactic halo is in fact likely to be flattened because of the influence of the disk, and may have additional flattening due to a flattened distribution of mass in the halo. We thus constructed a model composed of (a) an exponential disk with scale length $h_r = 3\,\mathrm{kpc}$, scale height $h_z = 0.3\,\mathrm{kpc}$, and mass density $0.12\,M_\odot\mathrm{pc}^{-3}$ near the solar circle, and (b) a modified logarithmic halo potential $\Phi_h = v_c^2 \ln ((r^2+(z/q)^2)^{1/2})$ with $v_c = 180\,{\mbox{km\,s$^{-1}$}}$. Here, $r=(x^2+y^2)^{-1/2}$ denotes the cylindrical radius. Together, the two components yield a flat rotation curve (in the Galactic plane) with a velocity of about 220 . In the region near the cluster the contribution of the disk to the total potential can be described with only a monopole and a quadrupole term. For the halo component two cases were considered, a spherical halo ($q=1.0$) and a flattened halo with $q=0.8$. Again, we find that these potentials lead to orbits whose local projections are practically indistiguishable from those obtained with the previous models. For $q=1.0$ the best-fit projected local orbit (in the sense of the solid line in Figure 9) requires $v_t = 90\,{\mbox{km\,s$^{-1}$}}$. In the case of $q=0.8$ an equivalent orbit, i.e., with the same local projected path, can be obtained by increasing the tangential velocity to $v_t = 105\,{\mbox{km\,s$^{-1}$}}$. This demonstrates that over the currently known angular extent of the tidal tails the projected orbital path of the cluster is not sensitive to the flattening of the Galactic potential.
In conclusion, the fit of the cluster’s projected orbital path to the geometry of the tidal tails does not depend on a particular model for the Galactic field. As long as there are no additional kinematic constraints from radial velocities or proper motions along the tails the result of the fit is compatible with a variety of orbits for different fields, which locally project onto the same path on the sky. Those orbits of course differ from each other along the line-of-sight. However, within the region where we see the tidal tails the differences are not substantial. This allows us to estimate how much the distance varies along the cluster’s orbit over the $10^\circ$ arc of the stream. It turns out that the end of the leading tail at $l \cos b = -3\fdg2$ lies nearest to the observer, in accordance with the negative radial velocity of the cluster. The different orbits obtained with the above selection of Galactic models put this point at a galactocentric distance $R$ between 17.6 and 17.9 kpc, and at a heliocentric distance $d$ of 22.1 to 22.4 kpc. For the opposite end of the trailing tail at $l \cos b = +7\fdg0$ these orbits predict galactocentric distances between 18.6 and 19.0 kpc, and heliocentric distances of 23.4 to 23.9 kpc. The maximum of $d$ lies between 23.6 and 23.9 kpc. The distance between the cluster’s orbit and the observer thus varies by up to 1.8 kpc over the length of both tails. This is a variation of +3% and $-$5% relative to the present distance of the cluster. Observationally, this corresponds to a magnitude difference of +0.06 mag and $-$0.1 mag, respectively, which is completely consistent with the results presented in §4.4.
The maximum value of $R$ along the local orbit lies in the range from 18.7 to 19.4 kpc (see upper panel of Fig.10). All solutions for the local orbit place its apocenter between 41 and 66 in $l \cos b$. This reveals that the cluster’s present position must be close to the apocenter of its orbit. The different orbit solutions suggest that the cluster passed this apogalactic point between 13 and 27 Myrs ago. The cluster’s proximity to an apogalactic point implies, that the variation of the galactocentric distance $R$ along the local orbit is small. Indeed, in the region of the trailing tail the variation of R along the orbit is at most 0.6 kpc or 3%, and in the region of the leading tail this variation is 0.9 kpc or 5% (see the above minimum and maximum values).
Is the tail aligned with the orbit?
-----------------------------------
With a model of the cluster’s orbit at hand, we can a posteriori test and validate our working hypothesis that the tidal tails lie parallel to the local orbital path of the cluster even if the stars do not have exactly the same velocity as the cluster when they decouple from the cluster. To this end we simulated a sample of test particles in a spherical logarithmic potential with $v_c=220$ . The particles were released over the time interval from $-$2.0 Gyr to present at equal time steps of 20 Myr. We emphasize that this experiment is not meant to provide a realistic model of the mass loss history of the cluster, but just serves to reveal the geometry of the tails. The particles were released from the cluster with a radial offset, i.e., either in the direction of the Galactic center or in the opposite direction, and with small velocity offsets. The size of the radial offset was chosen to be $3 \times r_L$, with $r_L$ the distance of the Lagrange point of local force balance between the cluster and the Galactic potential, i.e.,
$$\begin{aligned}
r_L^3 = G M_{cl} \frac{R^2}{v_c^2} \end{aligned}$$
For the present position of the cluster equation (8) yields $r_L = 57$ pc, using $M_{cl} = 6 \times 10^3 M_{\odot}$ as the mass of the cluster. The velocities of the particles were offset from the velocity of the cluster by a velocity vector with a size of 1.0 (i.e., about $2\times$ the dynamical velocity dispersion of the cluster, see Paper II), pointing either in the radial direction (i.e., towards the Galactic center on the inner side and away from it on the outer side) or at $45^\circ$ from this direction.
Figure 11 shows the distribution of this sample of test particles in the plane of the sky at $t=0$. The plot covers the same field as Figure 9 and uses the same (galactic) coordinates. The solid line shows the path of the cluster’s orbit, same as in Figure 9. It can be seen that the stream of test particles has approximately the same width as the observed tails, and that it is well aligned to the cluster’s projected orbital path. In other words, the relatively small peculiar velocities that stars may have when escaping from Pal5 do not have a significant impact on the mean location of the tidal debris with respect to the cluster’s orbit, at least not in projection onto the plane of the sky. Thus our assumption, that the orbit of the cluster must be fit such that it is parallel to the observed tidal tails proves to be entirely valid.
Global orbit
------------
We saw that the apogalactic distance $R_{max}$ of Pal5 can be derived from the local orbit and hence does not strongly depend on specific assumptions on the Galactic field. However, the determination of other characteristic parameters of the cluster’s orbit such as the perigalactic distance $R_{min}$ or the distance $R(z=0)$ at which the cluster crosses the Galactic disk requires extensive extrapolation beyond the region of the tails and thus depends on a global Galactic mass model. In order to estimate these and other orbital parameters we made use of Model 2 from the series of Galactic models developed by Dehnen & Binney (1998a). This mass model consists of three exponential disks representing the stellar thin and thick disk and the interstellar material, and of a bulge and a halo component. The parameters of the model are chosen such that the model accomodates a variety of observational constraints, e.g., the Milky Way’s rotation curve, the local vertical force, the local surface density of the disk etc. (for details see Dehnen & Binney 1998a).
The tangential velocity of Pal5 was set to $v_t = 90$. Hereby the above model provides an orbit whose path locally coincides with the solid line in Figure 9 and hence meets the condition of a good fit to the tidal tails . The equations of motion were integrated over the time interval from $-$1 Gyr to 1 Gyr. Part of the resulting orbit is shown in Figure 12abc. We find that the orbit has perigalactic distances in the range from 6.7 to 5.7 kpc.[^5] This reveals that the cluster penetrates deeply into the inner part of the Milky Way. The typical time scales of the orbit are $<T_R> = 291$Myr (mean period of radial oscillation) and $<T_{\psi}> = 443$Myr (mean period of rotation around the Galactic z-axis). The local constraints from the tidal tails allow us to vary $v_t$ by about $\pm 10$. When doing so the perigalactic distances of the orbit change by about $\pm 0.8$kpc, and the eccentricity thus varies by $\pm 0.05$. The periods $<T_R>$ and $<T_\psi>$ change only slightly, i.e., by $\pm 3\%$ and $\pm 1\%$.
From $-$1 Gyr to present the orbit makes five disk crossings. Three of them happen near or inside the solar circle, at distances of 6.7, 6.8, and 8.3 kpc, while two are at much larger distances of 14 to 18 kpc. When changing $v_t$ by $\pm 10$ the galactocentric distances of the crossings of the inner disk vary by typically $\pm 0.7$kpc but occasionally up to $\pm 1.1$kpc. Interestingly, the predicted location of the next future disk crossing is at an even lower galactocentric distance of $5.9 \pm 0.8$kpc, which is very close to the next perigalacticon (see Figs.12a and 12b). This disk crossing is predicted to happen in $+110 \pm 2$Myr from present.
Besides the models of Dehnen & Binney there exist a variety of other Milky Way mass models from other authors, e.g., Pacynski (1990), Allen & Santillan (1991), Johnston et al. (1995), Flynn et al. (1996). To test in how far our conclusions on the orbit of Pal5 depend on the particular model we repeated the integration of the orbit using the same initial velocities and the Milky Way mass model of Allen & Santillan. This is a three-component model with bulge, disk, and halo, where the disk potential is of the Miyamoto-Nagai form (Miyamoto & Nagai 1975). The corresponding orbit is shown in Figure 12def. The general characteristics of the orbit are very similar to the one found with the Dehnen & Binney potential. In particular, we find small pericentric distances down to 5.5 kpc. This confirms that the orbit of the cluster leads through the inner part of the Milky Way. The sequence of near and far disk crossings and apogalactic and perigalactic passages is the same as with the other model, but the associated time scales are somewhat shorter (e.g., $<T_R> = 275$ Myr, $<T_\psi> = 412$ Myr). Again, the orbit predicts an exceptionally small galactocentric distance, namely of 5.7kpc for the next crossing of the disk (in about 107 Myrs from present). When using a simple spherical logarithmic potential with $v_c$ in the range from 150 to 260 the resulting orbits yield even lower values for the galactocentric distance of this disk passage. It thus appears that an upper distance limit of $R \le 6$kpc for the next disk crossing is a safe prediction.
Clues on the mass loss history
==============================
Mean mass loss rate
-------------------
Using the results from §5 we can translate the amount of mass that is observed in the tails of Pal5 into a rough estimate of the mean mass loss rate. It was shown that the variation of $R$ along the local orbit is small, in particular in the region of the trailing tail. This means that the variation of the angular velocity $\dot{\varphi} = L / R^2$ along the local orbit is also small. Assuming $L=const$, the relative change in $\dot{\varphi}$ is twice the relative change in $R$ and thus $\le 6\%$ for the trailing tail and $\le 10\%$ for the leading tail. Therefore, it is justified to estimate the time scale of the angular drift between the debris and the cluster in the way described by equation (7).
The key parameter is the relative radial offset $\alpha$ from the orbit of the cluster. In reality, individual stars do not escape from the cluster under exactly the same conditions and hence do not settle on orbits with the same radial offset. Their orbits will not even be strictly confocal because they do not escape with exactly the same velocity. Hence stars at a certain azimuthal distance from the cluster will have taken different intervals of time to drift to this place. However we assume, that we can estimate the mean time scale $\Delta t$ of this drift by applying equation (7) to the mean value of $\alpha$.
To determine the mean offset we measured for each star its rectangular separation from the solid line of Figure 9 in the plane of the sky. We then counted the weighted number of stars in $2'$ wide bins of this rectangular separation, using the same weighting scheme as described in §3.2. Separate counts were made for the leading and the trailing tail. The resulting star count histograms are plotted in Figure 13. Each tail shows up as a symmetric peak on top of a constant background. We determined the center and the width of each peak by fitting a Gaussian plus a constant to the counts (see dashed lines in Fig. 13). For the trailing tail we thus measure a mean rectangular separation of 118$\pm$05 from the orbit and a FWHM of 184$\pm$12. For the leading tail we find a mean separation of 101$\pm$08 and a FWHM of 172$\pm$19. [^6] From the mean angular separations as seen in projection we reconstructed the mean radial distance between the debris and the orbit of the cluster in the orbital plane. This was done in the following way: We increased and decreased the length of the galactocentric radius vector of the cluster by 200pc, determined the positions of the endpoints of these vectors on the sky as observed from the Sun, and then computed the rectangular separation of these points from the projected orbit in the same way as done for the stars. This yields separations of 90 and 91, respectively, i.e., 0.76 and 0.90 times the observed separations. This implies that the observed separations correspond to mean radial distances between the tails and the orbit of 263pc for the trailing tail and 222pc for the leading tail.
We first focus on the trailing tail, which is better covered by the observations and which is most suited for applying equation (7). Using $v_t=95$ the angular momentum of the cluster’s orbit is $L= 1814$kpc . From $R_{max}= 19.0$kpc we derive the angular velocity at the apogalactic point as $\dot\varphi = 5.14$/kpc or, in other units, $\dot\varphi = 0\fdg295$/Myr. Comparing the mean radial offset of the tail of 263pc and the apogalactic distance of 19.0kpc we have $\alpha = 1.014$. The material in the trailing tail is basically spread over an arc of $6^\circ$ on the sky. Along this arc, the orbit of the cluster subtends an azimuth angle of 77 in the orbital plane.
Putting these numbers into equation (7) we obtain $\Delta t = 1.94$Gyr. This is the typical time it has taken debris stars to drift from the cluster center to the “tip” of the tail. The trailing tail contains about 0.8 times as many stars as the cluster (see §4.1). If we assume that the tail has the same mass function as the cluster, the mass in the tail should be $0.8 \times M_{cl}$, where $M_{cl}$ denotes the present mass of the cluster. With regard to the mass function, this is a lower limit, because the tail is likely to contain a larger fraction of low-mass stars than the cluster. The cluster is known to be underabundant in low-mass stars and may have lost them through mass segregation followed by tidal stripping from the outer part of the cluster. Since low-mass stars are not represented in our sample such a difference in the mass function would result in a somewhat higher total mass of the tail. Using the above values, and taking into account that equal amounts of mass are lost on both sides of the cluster, we finally obtain an estimate of the mean mass loss rate of $-\dot{M}/M_{cl} = 0.82$/Gyr. Multiplying by the present mass of the cluster, which has recently been estimated to be $-M_{cl}
= 6 \times 10^3 M_{\odot}$ (Paper II), we get $\dot{M} = 4.9 M_{\odot}$/Myr.
To check this result we do an analogous calculation for the leading tail using a mean distance of $R=18.2$kpc (see §5.2). The mean angular velocity then is $\dot\varphi = 0.321^\circ$/Myr, and the mean radial offset of the tail from the orbit of the cluster yields $\alpha =0.988$. Furthermore we have $\delta\varphi = 4\fdg8$ as the azimuth angle of the orbit along the leading tail (seen from the Galactic center), and $M_{tail} = 0.4 M_{cl}$ as an estimate of its mass (see §4.1). Equation (7) thus yields $\Delta t = 1.26$Gyr, and this leads to a mean mass loss rate of $-\dot{M}/M_{cl} = 0.63$/Gyr or $-\dot{M} = 3.8 M_{\odot}$/Myr. This rate is somewhat lower than the one obtained from the trailing tail because the mean drift rate along the leading tail is only slightly higher and cannot compensate the lower surface density in the leading tail.
The accuracy of these estimates is limited by a number of potential sources of errors, the most important of which are: (1) The uncertainty in the determination of the offset between the tails and the orbit. (2) The uncertainty of the angular velocity resulting from errors in the distance and the tangential velocity of the cluster. (3) Deviations of the motion of individual stars from the drift rate at the mean radial offset due to a spread in their initial positions and velocities. While the fitting of the histograms of Figure 13 yields a formal uncertainties of 05 and 08 for the mean angular separation, we expect that the true error of this quantity is more like $1'$ to 15, in particular because the exact location of the cluster’s orbit is not known. The relative error in the mean angular separation and hence in the mean radial offset is thus assumed to be between 10% and 15%, producing a relative error in the factor $(\alpha-1)/\alpha$ of approximately the same size. The heliocentric distance and the estimate of the tangential velocity of the cluster are both thought to have a relative uncertainty of 10%. This translates into relative errors of 12% and 10% for the corresponding galactocentric quantities and thus results in a combined error for the angular velocity $\dot\varphi$ of 16%. Hereby, the relative error in the fractional mass loss rate would be about 22%. How much the above estimate of the mass loss rate is biassed by the spread in the drift motions of individual stars needs to be investigated with forthcoming detailed N-body simulations of the tidal disruption Pal5. However, we expect that this may also contribute a relative uncertainty of about 20%. The total error is thus believed to be of the order of 30%. In conclusion, our result for the mean mass loss rate of the cluster with respect to its present mass is $-\dot{M}/M_{cl} = 0.7 \pm 0.2$/Gyr or $-\dot{M}=4.3\pm1.3 M_\odot$/Myr (average from both tails).
It is interesting to compare this mass loss rate with predictions from approximate formulae devised by Johnston et al. 2002 (hereafter JCG02) for calculating mass loss rates of satellites from radial surface density profiles. These formulae (see equations (6) and (7) of JCG02) relate the mass loss rate to the break radius of the surface density profile, the surface density outside this radius, and the time period for orbital or circular motion, assuming that the radial profile decreases as either $r^{-1}$ or $r^{-2}$. Since the observed profiles for Pal5 are characterized by $-1.5 \le \gamma \le -1.2$ they lie in between those two cases. We calculated the predicted mass loss rates with two different radii, $r=20'$ and $r=60'$. Using $r_{break}=16'$, average surface densities as provided in Figure 5, and the time scales $T_{orb} = 356$ Myr, $T_{circ} =
528$ Myr, equation (6) of JCG02 predicts mass loss rates of $6.2 M_\odot$/Myr ($with r=20'$) and $3.7 M_\odot$/Myr (with $r=60'$) while equation (7) of JCG02 predicts mass loss rates of $2.1 M_\odot$/Myr and $3.8 M_\odot$/Myr, respectively. The predictions are hence within 50% of our above detailed estimate of the mean mass loss rate of the cluster.
Total mass loss
---------------
According to numerical simulations the mean rate of tidal mass loss of a globular cluster orbiting in a stationary Galactic potential remains approximately constant over most of its lifetime, i.e., the cluster’s mass decreases approximately linear with time (see, e.g., Gnedin et al. 1999, Johnston et al. 1999a, Baumgardt & Makino 2002). From the observations it is not entirely clear whether this really holds for Pal5 because the outward decrease of the linear density of the tails could be interpreted as an indication for a secular increase in the mean mass loss with time. However, if we assume that the mean mass loss rate was constant, we can use the measured mean rate of recent mass loss to get an idea of the cluster’s mass at earlier epochs. An important condition is that the Galactic potential must have been essentially the same over the time interval spanned by the extrapolation. Extrapolating over the cluster’s entire age of 12 Gyr thus poses a problem because disk shocks seem to be an important driver of the mass loss of Pal5, and the age distribution of disk stars suggests that the Galactic disk was not in place at very early epochs. However, if the disk was less massive or non-existent early on, the initial rate of the cluster’s tidal mass loss was presumably lower than the present rate. Therefore, an extrapolation with the present mass loss rate may overestimate the cluster’s initial mass but can put an upper limit on it.
Our measurements suggest a mean mass loss rate of $-\dot{M}/M_{cl} = 0.7$/Gyr. Multiplying this with 12 Gyr and adding the cluster’s present mass, we obtain an upper limit for the cluster’s initial mass of $M = 9.4 M_{cl}$. Pal5 may thus have started with roughly ten times as much mass as it has today. If one restricts the extrapolation to a time interval of about 8 Gyr, for which the existence of a massive Galactic disk is likely, one obtains a cluster mass of $6.6 M_{cl}$. Since the cluster is certainly older than 8 Gyr, this presents a lower limit to its initial mass, provided that the cluster has been on its present orbit during the entire period. Of course, these limits vary as a function of the error of the mass loss rate and provide not more than a rough guide line.
Discussion and Summary
======================
Compelling evidence for tidal disruption
----------------------------------------
Our analysis of an enlarged set of SDSS data reveals that the cluster Pal5 is connected to a long stream of tidal debris that contains at least 1.2 times as much stellar mass as the cluster itself contains in its present state. This confirms and extends the results of Paper I, which presented first evidence for tidal tails from a study of a smaller field. The most basic aspect of these results is that they provide the first stringent observational proof that globular clusters in the Milky Way’s halo may be subject to significant tidal mass loss, by which they eventually dissolve. The detection of fully-fledged tidal tails, which is unique so far, makes Pal5 the prototype of such tidally-disrupting globular clusters. We showed that the stream of debris is thin and maintains its small width over a length of several kiloparsecs, suggesting that it is a kinematically cold system. This is very much consistent the low velocity dispersion inside the cluster, which is only about 0.5 (Paper II). The numerical experiment described in §5.3 and Figure 11 lends further support to this view. It is obvious that the relatively small transverse spread of the debris on the sky has strongly favored the detection of the stream. In the case of a massive cluster, having a much higher internal velocity dispersion, the debris would probably spread out in a much wider stream and thus be more difficult to detect. Another favorable circumstance is the fact that Pal5 is presently located close to its apogalacticon. This means that the distribution of the debris has the smallest possible angular dispersion along the orbit and therefore shows a relatively high density.
The tails and the orbit of the cluster
--------------------------------------
It was demonstrated that the arc over which the tidal debris has been traced is now sufficiently long to recognize the intrinsic curvature of the stream. Since there is very good reason to assume that the tails are closely aligned with the orbit of the cluster the curvature of the stream reveals the local curvature of the cluster’s orbit. This is a remarkable point because curvature means acceleration. Direct measurements of the Galactic gravitational acceleration of an individual halo star or star cluster in the sense of observing a non-linear change in position or a change in velocity over time are for technical reasons totally out of reach. Therefore the curvature, and also the bipolar and S-shaped structure of Pal5’s stream of debris provides one of the first occasions where the acceleration of a halo object by the gravitational field of the Galaxy is directly visible. A considerable drawback however is that the tidal stream is known only in projection on the plane of the sky. This projected view provides clear evidence that the acceleration of the cluster is non-zero, but does not yet allow to derive a useful constraint on its specific value. To achieve this the observations would need to cover a major part of the cluster’s orbit or one would require precise information on the tangential velocity of the cluster or on the variation of the distance or the kinematics along the tails. As an example, we determined the best-fitting orbits in a spherical logarithmic potential and showed that measuring the absolute proper motion of the cluster on the accuracy level of 10 $\mu$as/y would allow to estimate the local circular velocity to about 2%. This is similar to the conclusion drawn by Johnston et al. (1999b) from a study of simulated tidal streams.
Using different Galactic models and selecting orbits that optimally fit the tidal tails, we found that the cluster must be near its local maximum distance from the Galactic center and that therefore, the galactocentric as well as the heliocentric distance along its orbit varies by no more than a few percent over the length of the tails. This is in full agreement with the photometry, because color-magnitude diagrams show that there is no systematic difference in apparent brightness between the tidal debris and the cluster of more than 0.1 mag. On the other hand, it turned out that the pericenters of the cluster’s orbit must lie in the inner Galaxy, namely at $R \le 7$kpc. This means that its orbit is rather eccentric and that the cluster must have repeatedly crossed the Galactic disk near or inside the solar circle. Since such disk crossings at small galactocentric radii lead to strong tidal shocks this provides a convincing explanation for Pal5’s heavy mass loss.
Of particular interest is the fact that the next disk crossing, which will happen in about 100 Myr, is predicted to take place at a galactocentric distance $\le 6$kpc, causing a very strong tidal shock. Given the small amount of mass that is left in the cluster and its low spatial concentration, one must suspect that this event will trigger the total disruption of the cluster. Dinescu et al. (1999), using formulae for the destruction rates due to disk- and bulge shocks as developed by Gnedin & Ostriker (1997) and orbital parameters determined from measured proper motions, derived a theoretical estimate of the destruction time of 0.1 Gyr for Pal5. This is identical to our estimate for the time until the next disk crossing. On the other hand, the original paper by Gnedin & Ostriker (1997) gave estimates of the destruction time for Pal5 of either 1.1 Gyr or 46 Gyr, depending on the model for the Galactic potential. These time scales are too long in view of what is known about Pal5 now, and must have resulted from a very different model for the cluster’s orbit.[^7] This demonstrates the importance of reliable individual orbital data for clusters like Pal5 since such data can place much stronger constraints on the dynamical evolution of these systems than purely theoretical analyses.
The surface density in the tails
--------------------------------
The surface density distribution of the stars in the tails shows some important peculiarities:\
(1) The distribution is clumpy, i.e., there are a number of density maxima and minima in each tail. While the density variations in the leading tail are moderate and not necessarily significant (i.e., they could still be the result of statistical fluctuations) the trailing tail shows at least two strong density maxima and one gap, where the local density deviates significantly from the mean density level. Although disk shocks, which are believed to be the primary cause of Pal5’s mass loss, are likely to modulate the instantaneous mass loss rate, it is implausible that the observed clumps are due to this modulation because the density variations would then need to occur symmetrically in both tails. Also, such variations should preferentially be visible near the cluster while at larger angular distances they should be washed out by the differential drift between stars of different orbital energy.
The region of maximal stellar density in the trailing tail lies at arc lengths between 23 and 37 from the center of Pal5. According to the results from §6.1 the mean rate of apparent drift of the tidal debris is 031/100 Myr. Thus the stars that form this broad density clump are expected to have escaped from the cluster in the interval between 740 and 1190 Myr ago. In §5.4 we found that during the last 1 Gyr the cluster crossed the Galactic disk five times, and three times thereof at small distances from the Galactic center. The latter occurred 140 Myr, 480 Myr, and 740 Myr before present. If these inner disk crossings produce overdensities that can be observed as distinct clumps in projection on the sky, then one would expect to see one such clump close to the cluster ($\lambda \approx 0\fdg3$) and two clumps - or perhaps one broad clump from the merging of the two - at arc lengths between 15 and 23. This does not correspond to the observations. In order to achieve some kind of agreement one would need to assume that the drift of the debris is in fact 50% faster than derived from the radial offset between the tails and the orbit of the cluster.
It may seem intriguing that the major density enhancement in the stream is near the apogalactic point of the cluster’s orbit. Nevertheless, the fact that a stellar stream gets compressed near the apocenter (because the angular velocity is minimal at this point) cannot explain this local enhancement because the angular scale of the observed feature (less than 25 when viewed from the galactic center) is much too small. One could speculate that the density variations in the tails might come from scattering by small-scale perturbations of the Galactic potential as produced by spiral arms, molecular clouds, dark matter clumps etc. However it has not yet been demonstrated that such perturbations can indeed generate the observed features. We note that such perturbations would certainly need to be much weaker than those in a dark halo of massive ($\approx 10^5 M_\odot$) black holes because encounters with such massive compact objects would have destroyed the cluster - and most likely also its tails - on a very short time scale (Moore 1993).
One might also think that the broad density enhancement in the trailing tail could be related to the cluster M5 since it is uncomfortably close to it on the sky. But a dedicated search for debris from M5 (using a proper color-magnitude filter for this cluster) yields absolutely no evidence for an extended distribution of M5 stars that could overlap with the tail of Pal5.
\(2) Apart from local variations, the radial profile of the stellar surface density at $r \le 20'$ declines like a power law with an exponent of $-$1.2 to $-$1.5, which means that the linear density along the tails decreases slowly with increasing distance from the cluster. How does this compare to the results of N-body simulations of stellar systems in Milky Way like potentials? Combes et al. (1999) show density profiles for two of their simulated clusters and find that at radii $r > r_t$ the volume density of debris stars decreases like $r^{-4}$, hence the surface density decreases like $r^{-3}$. This is clearly a much steeper decline than the one we observe. In contrast to this Johnston et al. (1999a) show surface density profiles of simulated globular cluster-like systems, in which the unbound outer part has a much shallower decline that is almost like $1/r$. As these authors point out, such a $1/r$ decline can easily be explained if one considers the very simplified case of a cluster on a circular orbit, which looses mass at a constant rate and with constant energy offsets. Our observations are not extremely far from the simple $1/r$ case but do not match it exactly. More recently Johnston et al. (2002), presented a series of simulations, in which the obtained radial surface density profiles that show a wide variety of logarithmic slopes in the outer part. They report values for the power-law exponent $\gamma$ between $-$1 and $-$4, and show that the result depends to some extent on the parameters of the orbit and the orbital phase. Even for an almost circular orbit they find $\gamma$ to scatter between $-$1 and $-$3. For very eccentric orbits their simulations seem to predict that $\gamma$ is mostly below $-3$, in particular at the apocenter. However, the system parameters (mass,radius) and the orbits that were used in these recent simulations are actually more representative of dwarf satellites than of clusters like Pal5. The conclusion from this comparison is that different simulations predict a wide range of possible power law exponents for the debris and that our observed values of $\gamma$ lie in the upper part of this range.
The fact that in both tails the observed surface density profile is somewhat steeper than the simple $1/r$ may indicate that either the angular velocity along the orbit increases with increasing angular distance from the cluster or the (orbit-averaged) mass loss rate has undergone a secular change. The variation of the orbital angular velocity must certainly enter the game on larger scales but cannot have a significant impact on the present results because the arc length of the tails is too short. We are thus left with the possibility that the mean mass loss rate may have steadily increased or that the mass loss process may have suddenly set in not much longer than 2 Gyrs ago. In the latter case an outward decrease of the linear density of the tails could result from the fact that the tip of each tail would only contain stars with the highest energy offset from the cluster, which should occur in low numbers, while at smaller angular distance from the cluster one would find both, stars with smaller energy offset, which should be more numerous, and also stars with higher energy offset, which were released from the cluster more recently.
Indeed, the traces of the trailing tail disappear before the tail would reach the edge of the observed field, and it is unknown whether or not the stream continues and reappears with significant density farther out along the orbit. If so, there would be a gap or a section of very low density with an arc length $\le 1^\circ$, that would need to be explained. If not, this would imply that the mass loss history of Pal5 underwent an abrupt change about 2 Gyr ago and that the mass loss rate in the earlier phase was much lower than in the recent phase or even zero. This would certainly require a fundamental change of the cluster’s orbit.
One possibility would then be that Pal5 came from outside the Galaxy and was accreted by it in a merger event with a smaller galaxy. This merger event would need to have happened fairly recently, and the cluster would have been accreted as a system with a mass of only $1.5\times 10^4 M_\odot$. As we will discuss in §7.7, an association of Pal5 with the Sagittarius (Sgr) dSph galaxy is very unlikely. From statistical studies there is evidence for further halo substructure, i.e., possible great circle streams of outer halo satellites (e.g., Lynden-Bell & Lynden-Bell 1995, Palma et al. 2002), which might be a hint on former merger events. However, based on sky position, distance, and radial velocity, Pal5 has not qualified as a possible member of any hypothetical stream in these studies. Since not all of the satellites may in fact belong to such a stream, an association with an individual satellite needs not necessarily be obvious through great circle alignments. In this sense the question of whether or not Pal5 is likely to be associated with one of the dSph satellites remains open because there is currently no well-constrained orbital information on most of these objects.
As another possibility we mention that the orbit of Pal5 could have changed drastically through a close encounter with one of the Milky Way satellites. Since the present orbit of Pal5 is unlikely to reach beyond galactocentric distances of 20 kpc, the only known candidate for such an encounter would be the Sgr dwarf spheroidal. However such scenarios are highly speculative. It is thus very important to learn more about the spatial extent of the stream.
\(3) Overall, the stellar surface density in the leading tail is on a lower level than in the trailing tail. This is surprising because the symmetry of the tidal force field suggests that the distribution of tidal debris should - at least in the vicinity of the cluster - be symmetric with respect to the cluster center. Both, orbit calculations and the analysis of the photometric data have shown that variations in the line-of-sight distances cannot be responsible for this effect. One may thus wonder if the star count results could be influenced by variable interstellar extinction. In Figure 14 we show the surface density contours of Pal5 and its tidal stream overplotted on a grey-scale map of interstellar extinction derived from the reddening data of Schlegel, Finkbeiner & Davis (1998). Although integrated Galactic foreground extinction corrections do not necessarily hold for stars within the Milky Way because they are integrated along the entire line of sight, we believe that this only leads to small errors for a cluster as distant and as far above the plane as Pal5. Another possible effect is the dependence of extinction on stellar temperature (and thus color) as discussed in Grebel & Roberts (1995). Considering our filter choice and color range, this should amount to at most 0.02 to 0.03 mag uncertainty, which is negligible for our purposes.
Figure 14 reveals that there is indeed significant variation in interstellar extinction over the field, as was already mentioned in §2. One notices that the leading tail is located close to the region of enhanced extinction but fortunately does not run across this region. This shows that there is little reason to assume a significant impact of extinction on the measurement of the stellar surface density in the tails. If enhanced extinction played a role, it would result in a loss of faint stars close to the detection limit. While the definition of our sample already involves a magnitude cut-off at $i^*=21.8$ well above the limit, we tentatively increased the cut-off by an additional 0.5 mag, thus restricting the sample to $i \le 21.3$. It turned out that the resulting surface density distribution is similar to the one for the larger sample and maintains the same overall imbalance between the leading and the trailing tail. It is therefore clear that the overall difference between the number counts in the two tails is not the result of variable extinction.
The luminosity functions
------------------------
We showed that the luminosity function of the stars in the tidal tails is in very good agreement with the stellar luminosity function in the cluster. This result is not self-evident, although it nicely fits the idea that the tails consist of stars from the same stellar population as the cluster. As briefly noted in previous sections, a deep study of the core of Pal5 using the HST (Grillmair & Smith 2001) has revealed that the luminosity function of Pal5 is relatively flat. This means that there is a strong deficiency in low-mass stars, at least in the core of the cluster. If this deficiency is a consequence of tidal mass loss (in combination with mass segregation), as Grillmair & Smith suggested, one would expect to find a corresponding overabundance of low-mass stars in the cluster’s debris. The luminosity functions should then actually be different. However, the flattening of the luminosity function shown by Grillmair & Smith becomes effective at absolute magnitudes $M_I > 5$mag which is below the limit of our analysis of SDSS data. Therefore, even though we find that the luminosity functions of the tails and the cluster agree down to the limit of this study, it seems likely that they will diverge when probing the stellar content of the tails at fainter magnitudes. If so, the higher fraction of low-mass stars in the tidal debris will contribute additional mass to the tails. This means that our current estimate of the mass ratio between the tails and the cluster (§4.1) provides a lower limit while the true mass ratio may be somewhat higher.
The mass loss rate
------------------
We showed that the transverse offset between the tails and the cluster can be used to estimate the time needed by debris stars to drift away from the cluster by a certain angle. This leads to the conclusion that the tails in their currently known extent represent the cluster’s mass loss from essentially the last 2 Gyrs. The result of our numerical experiment on the drift of debris along the cluster’s orbit, which is shown in Figure 11, confirms this time scale. It follows that the mean mass loss rate of Pal5 in this period was about 0.7($\pm$ 0.2) times the present mass of the cluster per Gyr. However, these estimates involve a number of approximations and simplifying assumptions. In particular, it remains to be investigated whether the fact that stars are actually distributed over a range of velocities and positions when escaping from the cluster leads to a substantial bias in the above estimates. For more definitive results accurate modelling of the details of the mass loss process are needed. Thus N-body simulations of the system under realistic conditions need to be performed.
A simple extrapolation with the present mean mass loss rate shows that the cluster may initially have had a total mass of between 6 and 10 times its present mass. Obviously, such an estimate depends on whether the cluster maintained the same mean (i.e., orbit-averaged) mass loss rate over most of its lifetime. For time intervals in which (1) the Galactic potential was the same as today and in which (2) the cluster followed the same orbit as it is today, one can indeed expect this to hold true. However, the observed decline of the surface density of the tails with increasing distance from the cluster warns us that the mean mass loss rate might in earlier times have been lower than it presently is. For clarifying this issue it is important to find out whether or not the tails extend to larger distances from the cluster.
Tangential velocity versus proper motion
----------------------------------------
The condition that the local orbit of the cluster needs to fit the location and curvature of the tidal tails allowed us to determine the vector of the tangential velocity of the cluster in a completely new manner. It is interesting to see whether the tangential velocity obtained in this way is consistent with the measured absolute proper motion of Pal5. Dinescu et al. 1999 report the proper motion of Pal5 measured by Cudworth (1998, unpublished) as $\mu_\alpha \cos\delta = -2.55\pm0.17$mas/y and $\mu_\delta = -1.93\pm0.17$mas/y. The quoted proper motion error corresponds to 19 per component at the distance of the cluster. If one transforms the above proper motion into the galactic rest frame using $d=23.2$ kpc and the velocity components of the Sun specified in §5.2 one obtains a tangential velocity of 137 with position angle $PA = 300^\circ$. Thus the direction of the cluster’s tangential motion derived from the measured proper motion differs from the direction given by the tidal tails by $20^\circ$. The absolute velocities show a difference of about 40 . Alternatively, one can do the inverse transformation, assuming a tangential velocity of $v_t = 90$ or $v_t = 95$ with position angle $PA = 280^\circ$ as implied by the fit of the local orbit. This yields predicted proper motions of $\mu_\alpha \cos\delta = -2.01$mas/y, $\mu_\delta = -2.03$mas/y, or $\mu_\alpha \cos\delta = -2.05$mas/y, $\mu_\delta = -2.06$mas/y, respectively. Comparison with the measured values shows that the declination component agrees well with our predictions within the quoted error while the right ascension component deviates from the prediction by about $3\sigma$ or 0.5 mas/y. Although this looks like a significant difference between the two completely independent determinations of the cluster’s tangential motion, a proper motion difference of 0.5 mas/y is actually not unreasonably large and could be explained by an underestimation of the measuring error of the proper motions.\
A former member of the Sgr dwarf?
---------------------------------
Based on position, radial velocity and rough proper motion data, some authors have argued for a possible association of Pal5 with the Sgr dwarf galaxy (Lin 1996; Palma et al. 2002, Bellazzini et al. 2003). Palma et al. classified it as a possible but unlikely member of Sgr, because the pole families were similar but orbital energy and angular momentum were found to be different. Bellazzini et al. argued that Pal5 lies relatively close to the orbit of Sgr as given by the model of Ibata & Lewis (1998), both in position $(x,y,z)$ and in the plane of $v_r$ vs. $R$. The impression of a good agreement in radial velocity $v_r$ is however fake because one gets this only by associating the cluster with a wrong position along the orbit of Sgr. In reality, the radial velocity of the orbit of Sgr near the position of Pal5 differs from the radial velocity of Pal5 by about 100 (adopting the model of Ibata & Lewis). Thus there is some accordance in position but not in kinematics.
The tidal tails and the cluster’s local orbit that we derive from it allow a much more robust comparison and strengthen the evidence against a former membership to Sgr. While the orbit of Sgr is almost polar, that of Pal5 is clearly not so. Hence Pal5 does certainly not orbit in the plane of the Sgr stream. The local orbit of the cluster crosses the orbit of Sgr at a large angle (compare our Fig. 12 with Fig. 1 of Bellazzini et al., but note that the orientation of the y-axis is inverted). Thus not only the radial motion, but also the tangential motion of the cluster is clearly discordant with the orbit of Sgr. The space velocity vectors of Pal5 and of the orbit of Sgr enclose an angle of $108^\circ$. Another argument is the strong difference in apogalactic distance. We showed that Pal5 is almost at its apogalacticon and does certainly not reach galactocentric distances of more than 20 kpc. The orbit of the Sgr dwarf however is thought to have apogalactica between 50 and 60 kpc (see, e.g., Dinescu et al. 2000, Bellazzini et al. 2002). This reflects a large difference in specific orbital energy. Thus, former membership in Sgr would require that the cluster lost much of its orbital energy after the departure from its host, which would be difficult to explain. This shows that apart from an approximate positional correlation with the orbit of Sgr, which can be a coincidence, there is little reason to assume a connection between Pal5 and the Sgr dwarf. On the contrary, our results on the cluster’s orbit make it rather unlikely that Pal5 originates from this dwarf. This also holds if one considers other models for the orbit of the Sgr dwarf, e.g., by Helmi & White (2001), Johnston et al. (1999c), or Gomez-Flechoso, Fux & Martinet (1999), since all of them are polar. However, this does not exclude the possibility of a close encounter of the two systems by which the orbit of Pal5 might have been deflected (see §7.3).
Outlook
=======
The tidal stream of Pal5 opens a new and promising way to constrain the gravitational potential in the Galactic halo. Detailed information on the orbits of individual halo objects like Pal5 or the Sgr dSph from their tidal debris can in principle produce much more powerful constraints on the Galactic potential than classic statistical approaches. We showed that with the current positional data for the tails of Pal5, which cover a limited range of orbital phase angles, the orbit of the cluster is not yet uniquely determined by the observations alone, and conclusions on the Galactic potential can therefore not yet be drawn. However, this will change drastically when kinematic data for the tails are added to the analysis (see, e.g., Murali & Dubinski 1999). Precise proper motions could be very useful, but will remain unavailable until future astrometric space missions like SIM or GAIA are flown because very high accuracy is required. On the other hand, precise radial velocities for giants stars associated with Pal5 are within reach of today’s 8-10m class telescopes. Another approach is to measure main-sequence turn-off stars, which are more numerous, but much fainter and hence can only be observed at lower spectral resolution. An observing program to obtain radial velocities of candidate giants along the tails of Pal5 has been started on the VLT. We expect that the results of this program will allow us to break the degeneracy in the determination of the cluster’s local orbit and allow a direct measurement of the gravitational acceleration in the Milky Way halo at a galactocentric radius of 18 to 19 kpc.
A major open question is how far the tidal stream continues and what the full time span of the mass loss history of Pal5 thus is. This can be clarified with targeted searches for further tidal debris from Pal5 along the arc outlined by our model of the cluster’s orbit. If the stream can be traced farther out, one should at some point also discover a substantial variation in the heliocentric distance of the stars. Hence such detections would not only provide information on the mass loss history, but also provide further important constraints on the orbit.
The fact that we see Pal5 while it is only about 100 Myrs away from its complete disruption, makes it very likely that there have been more clusters of similar type, which dissolved during the last few Gyrs. This provides observational support for the common conjecture that the Milky Way’s globular cluster system was originally much richer in low-mass cluster than it is today (see Fall & Zhang 2002 and references therein). Since the tidal stream of Pal5 is at least about 2 Gyrs old, tidal streams from other low-mass clusters that dissolved recently may also still exist and be observable. Without the presence of a parent object such streams are of course more difficult to find. On the other hand, the detection of anonymous streams that are left-overs from globular clusters would provide important information on the evolution of the globular cluster system of the Milky Way and also provide further possibilities to probe the Galactic potential. The SDSS presents an excellent data base to search for such cluster remnants.
Funding for the creation and distribution of the SDSS Archive has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Aeronautics and Space Administration, the National Science Foundation, the U.S. Department of Energy, the Japanese Monbukagakusho, and the Max Planck Society. The SDSS Web site is http://www.sdss.org/. The SDSS is managed by the Astrophysical Research Consortium (ARC) for the Participating Institutions. The Participating Institutions are The University of Chicago, Fermilab, the Institute for Advanced Study, the Japan Participation Group, The Johns Hopkins University, Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, University of Pittsburgh, Princeton University, the United States Naval Observatory, and the University of Washington.\
M.O. thanks Andi Burkert for fruitful discussions and Don Schneider and the referee for comments that helped to improve the manuscript.
Isochrone epicycle approximation for the logarithmic potential
==============================================================
Following Dehnen (1999) the isochrone approximation for the motion of a particle in a spherically symmetric potential $\Phi$ is obtained by transforming from radius $R$ and time parameter $t$ to a new radial coordinate $x = \sqrt{R^2 + b^2}$ and a new parameter $\eta$ with $dt/{d\eta} = x$. Hereby, the equation of motion changes from $$\begin{aligned}
\frac{d^2 R}{dt^2} &=& \frac{d}{dR}\,Y \qquad\mbox{for}\quad
Y := \left(E - \Phi (R)\right) - \frac{L^2}{2R^2} \end{aligned}$$ to $$\begin{aligned}
\frac{d^2 x}{d\eta^2} &=& \frac{d}{dx}\,\tilde Y\qquad\mbox{for}\quad
\tilde Y := R^2 Y \quad .\end{aligned}$$ The quantity $\tilde Y$ as a function of $x$ is expanded into a Taylor series about its maximum using the first, second, and third derivative. For the first derivative we have $$\begin{aligned}
\frac{d\tilde Y}{dx} &=& x \left(2\left(E-\Phi(R)\right) -
R \frac{d\Phi}{dR}\right) \quad .\end{aligned}$$ Since the radius $R_E$ of a circular orbit with energy $E$ is defined by the equation $$\begin{aligned}
2\left(E-\Phi(R_E)\right) &=& \frac{L^2}{R_E^2} =
R_E \left(\frac{d\Phi}{dR}\right)_{R_E} \end{aligned}$$ it is evident that the maximum of $\tilde Y$ lies at the value of $x$ that corresponds to $R_E$, i.e., $$\begin{aligned}
\frac{d\tilde Y}{dx} &=& 0 \qquad\mbox{at}\quad x=x_E=\sqrt{R_E^2+b^2} .\end{aligned}$$ Here, we focus on the special case of the logarithmic potential $\Phi(R) = v_c^2 \ln (R/R_0)$. Eqn. (A3) then reads $$\begin{aligned}
\frac{d\tilde Y}{dx} &=& x \left(2\left(E-v_c^2 \ln (R/R_0)\right) -
v_c^2\right) \quad .\end{aligned}$$ The second and third derivative of $\tilde Y$ are then $$\begin{aligned}
\frac{d^2\tilde Y}{dx^2} &=& \left(2\left(E-v_c^2 \ln (R/R_0)\right) -
v_c^2\right) - 2 v_c^2 \frac{x^2}{R^2} \\
\frac{d^3\tilde Y}{dx^3} &=& -2 v_c^2 \frac{x}{R^2}\left(3 -
2\frac{x^2}{R^2}\right) \quad .\end{aligned}$$ The parameter $b$ can be chosen such that the third derivative vanishes in $x=x_E$. This requires $$\begin{aligned}
b^2 = \frac{1}{2} R_E^2 .\end{aligned}$$ With this choice the 3rd order in the Taylor expansion of $\tilde Y$ about $x_E$ vanishes and the error introduced by truncating after the quadratic term is only of order 4. The location of the maximum of $\tilde Y$ then is $$\begin{aligned}
x_E &=& \sqrt{3/2}\,R_E\end{aligned}$$ and the second derivative of $\tilde Y$ at $x = x_E$ reads $$\begin{aligned}
\left(\frac{d^2\tilde Y}{dx^2}\right)_{x_E} &=& -2 v_c^2
\left( 1 + \frac{b^2}{R_E^2} \right) = -3 v_c^2 \quad . \end{aligned}$$ The expansion of $\tilde Y$ (up to third order) yields the approximate equation of motion $$\begin{aligned}
\frac{d^2 x}{d\eta^2} &=& \frac{d}{dx}\left(
-\frac{3}{2}v_c^2 \left( x - x_E \right)^2\right),\end{aligned}$$ which is solved by a harmonic oscillation $$\begin{aligned}
x(\eta) &=& x_E \left(1 + e \cos \left(\sqrt{3}\,v_c\,\eta\right)\right) \quad . \end{aligned}$$ Here, the zero point of the parameter $\eta$ is (without loss of generality) chosen such that it coincides with the apocenter $x=x_{max}$ For simplicity, we absorb the factor $\sqrt{3}\,v_c$ by setting $\tilde\eta := \eta \sqrt{3}\,v_c$ Integrating $dt/d\tilde\eta = x/(\sqrt{3}\,v_c)$ from $\tilde\eta=0$ to $\tilde\eta = 2\pi$ one finds the period of the oscillation to be $$\begin{aligned}
T_R &=& \sqrt{2}\,\pi \frac{R_E}{v_c} \end{aligned}$$ Evaluation of $\dot{R}$ at $R=R_E$ yields $\dot{R}\,(R=R_E) = \pm 3 v_c e / \sqrt{2}$ and, by combination with the equation of energy conservation, provides the expression $$\begin{aligned}
e &=& \frac{\sqrt{2}}{3} \sqrt{1-\frac{L^2}{v_c^2 R_E^2}} \quad . \end{aligned}$$ for the eccentricity parameter $e$. By integration of $\dot{\varphi} = L/R^2$ the azimuth angle $\varphi$ is $$\begin{aligned}
\varphi(\tilde\eta_2) &=&
\sqrt{\frac{2}{3}}\frac{L}{R_E}\int_{\tilde\eta_1}^{\tilde\eta_2}
\frac{\left(1+e\cos\tilde\eta \right)}{\left(1+e\cos\tilde\eta
\right)^2-\frac{1}{3}} d\tilde\eta + \varphi(\tilde\eta_1) \quad .\end{aligned}$$ If radius $R$ and absolute velocity $v$ are given for an arbitrary instant $t_1$ the equation of conservation of orbital energy in the logarithmic potential yields the parameter $R_E$ as $$\begin{aligned}
R_E &=& R(t_1) \exp \left(\frac{1}{2}\left(
\frac{v(t_1)^2}{v_c^2} - 1 \right) \right) \quad .\end{aligned}$$ Using this and knowing also the angular momentum $L = R(t_1) v_{\perp}(t_1)$ the parameter $e$ can be obtained from Eqn. (A15). The value of $\tilde\eta$ that corresponds to $t_1$ then follows from Eq. (A13) as $$\begin{aligned}
\tilde\eta_1 = \arccos \left(\frac{1}{e} \left(\sqrt{\frac{2}{3}
\frac{R(t_1)^2}{R_E^2} +\frac{1}{3}} - 1 \right)\right) \quad .\end{aligned}$$ Finally, by integration of $dt/d\tilde\eta = x/(\sqrt{3}\,v_c)$ the time $t$ for arbitrary $\tilde\eta$ is $$\begin{aligned}
t &=& \frac{x_E}{\sqrt{3}\,v_c} \Big[\tilde\eta + e \sin \tilde\eta
\Big]_{\tilde\eta_1}^{\tilde\eta} + t_1 \quad .\end{aligned}$$ In this way the radial component of the orbit and the time parameter $t$ are completely and explicitly determined as functions of $\tilde\eta$ for any given set of initial conditions.\
Baumgardt, H., Makino, J., 2002, /astro-ph/0211471 Bellazzini M., Ferraro R.F., Ibata R., 2003, , 125, 188 Binney, J., Tremaine, S., 1987, Galactic Dynamics. Princeton University Press, Princeton, New Jersey Chen, B., Stoughton, C., Smith, J.A., Uomoto, A., Pier, J.R, et al., 2001, , 553, 184 Dehnen, W., 1999, , 118, 1201 Dehnen, W., Binney, J., 1998a, MNRAS 294, 429 Dehnen, W., Binney, J., 1998b, MNRAS, 298, 387 De Marchi, G., Leibundgut, B., Paresce, F., Pulone, L., 1999, , 343, L9 Dinescu D.I., Majewski, S.R., Girard, T.M., Cudworth, K.M., 2000, , 120, 1892 Djorgovski, S.G., 1995, IAU Symp. 174, 9 Duquennoy, A., Mayor, M., 1991, A&A, 248, 485 Fall, S.M., Zhang, Q., 2001, , 561, 751 Flynn, C., Sommer-Larsen, J., Christensen, P.R., 1996, MNRAS, 281, 1027 Fukugita, M., Ichikawa, T., Gunn, J.E., Doi, M., Shimasaku, K., & Schneider D.P. 1996, , 111, 1748 Gnedin, O.Y., Ostriker, J.P., 1997, , 474, 223 Gnedin, O.Y., Lee, H.M., Ostriker, J.P., 1999, , 522, 935 Gomez-Flechoso, M. A., Fux, R., Martinet, L., 1999, A&A, 347, 77 Grebel, E.K. Roberts, W.J., 1996. A&AS, 109, 293 Grillmair, C.J., Smith, G.H., 2001, , 122, 3231 Gunn, J.E., et al. 1998, , 116, 3040 Harris, W.E., 1996, AJ 112, 1487 Helmi, A., White, S.D.M., 2001, MNRAS, 323, 529 Hogg, D.W., Schlegel, D.J., Finkbeiner, D.P., & Gunn, J.E. 2001, , 122, 2129 Ibata R.A., Wyse R.F.G., Gilmore, G., Irwin M.J., Suntzeff, N.B., 1997, , 113, 634 Ibata, R.A., Lewis, G., 1998, , 500, 575 Ibata, R.A., Lewis, G.F., Irwin, M., Totten, E., Quinn, T., 2001, , 551, 294 Johnston, K.V., Spergel, D., Hernquist, L., 1995, , 451, 598 Johnston, K.V., Hernquist, L., Bolte, M., 1996, , 465, 278 Johnston, K.V., Sigurdsson S., Hernquist, L., 1999a, MNRAS, 302, 771 Johnston, K.V., Zhao, H., Spergel, D., Hernquist, L., 1999b, , 512, L109 Johnston, K.V., Majewski, S. R., Siegel, M. H., Reid, I. N., Kunkel, W. E., 1999c, , 118, 1719 Johnston, K.V., Choi, P.I., Guhathakurta, P., 2002, , 124, 127 King, I.R., 1962, , 67, 471 King, I.R., 1966, , 71, 64 Kuhn, J.R., Smith, H.A., Hawley, S.L., 1996, , 469, L93 Kulessa, A.S., Lynden-Bell, D., 1992, MNRAS, 255, 105 Law, D.R., Majewski, S.R., Skrutskie, M.F., Carpenter J.M., Ayub, H.F., 2003, astro-ph/0305385 Lehmann, I., Scholz, R.D., 1997, , 320, 776 Lin, D.N.C., 1996, in: O. Lahav, E. Terlevich, R.J. Terlevich (eds.), Gravitational Dynamics, Cambridge University Press, Cambridge, p.15 Lupton, R., Gunn, J. E., Ivezić, Z., Knapp, G. R., Kent, S., & Yasuda, N. 2001, in ASP Conf. Ser. 238, Astronomical Data Analysis Software and Systems X, ed. F. R. Harnden, Jr., F. A. Primini, and H. E. Payne (San Francisco: Astr. Soc. Pac.) Lynden-Bell, D., Lynden-Bell, R.M., 1995, MNRAS, 275, 429 Mandushev, G., Spassova, N., Staneva, A., 1991, A&A, 252, 94 Majewski, S.R., Ostheimer J.C., Patterson, R.J., Kunkel, W.E., Johnston, K.V., Geisler, D., 2000, , 119, 760 Majewski, S.R., Skrutskie, M.,F., Weinberg, M.D., Ostheimer, J.C., 2003, astro-ph/0304198 Miyamoto, M., Nagai, R., 1975, PASJ, 27 ,533 Moore, B., 1993, , 413, L93 Murali, C., Dubinski, J., 1999, , 118, 911 Murali, C., Weinberg, M., 1997, MNRAS, 291, 717 Odenkirchen, M., Grebel, E.K., Rockosi, C.M., et al., 2001a, , 548, L165 Odenkirchen, M. et al. 2001b, , 122, 2538 Odenkirchen, M., Grebel, E.K., Dehnen, W., Rix, H.W., Cudworth, K.M., 2002, , 124, 1497 Oh, K.S., Lin, D.N.C., 1992, , 386, 519 Ostriker, J.P., Spitzer, L., Chevalier, R., 1972, , 576, L51 Pacynski, B., 1990, , 348, 485 Palma, C., Majewski, S.R., Johnston, K.V., 2002, , 564, 763 Palma, C., Majewski, S.R., Siegel, M.H., Patterson, R.J., Ostheimer, J.C., Link, R., 2003, , 125, 1352 Pier, J.R., Munn, J.A., Hindsley, R.B., Hennessy, G.S., Kent, S.M., Lupton, R.H., and Ivezic, Z. 2003, AJ, 125, 1559 Piotto, G., Cool, A.M., King, I.R., 1997, , 113, 1345 Piotto, G., Zoccali, M., 1999, A&A, 345, 485 Rockosi, C.M., Odenkirchen, M., Grebel, E., et al. 2002, , 124, 349 Schlegel, D.J,, Finkbeiner, D.P., Davis, M. 1998, , 500, 525 Siegel, M.H., Majewski, S.R., Cudworth, K.M., Takamiya, M., 2001, , 121, 935 Smith, G.H., McClure, R.D., Stetson P.B., Hesser, J.E, 1986, , 91, 842 Smith, J.A, Tucker, D.L., Kent S., et al. 2002, , 123, 2121 Stoughton, C., et al., 2002, , 123, 485 Testa, V., Zaggia, S.R., Andreon, S., Longo, G., Scaramella, R., Djorgovski, S.G., de Carvalho, R., 2000, , 356, 127 Trager, S.C., King, I., Djorgovski, S., 1995, , 109, 218 von Hoerner, S., 1957, , 125, 451 Wilkinson, M.I., Evans, N.W., 1999, MNRAS, 310, 645 York, D.G., et al., 2000, , 120, 1579 Zaritsky, D., Olszewski, E.W., Schommer, R., Peterson, R.C., Aaronson, M., 1989, , 345, 759 Zhao, H., Johnston , K.V., Hernquist, L., Spergel, D.N., 1999, A&A, 348, L49
[^1]: By convention, the magnitudes in the preliminary system are quoted using asterisks.
[^2]: In order to treat the bins at the borders of the c-m domain in the same way as in the interior the grid counts were actually extended somewhat beyond the c-m limits specified above.
[^3]: Note that a substantial fraction of field stars actually lies at $c_1 > 1.0$ and was already eliminated by the preselection in $c_1$ and $c_2$
[^4]: This assumption is of course violated at the location of the foreground cluster M5.
[^5]: Note that for an orbit in a flattened potential perigalactic passages do in general not occur all at the same distance.
[^6]: When cutting the tails into two parts of equal length we get the same FWHM for each part within the errors of the fit. The width of the tails can thus be regarded as constant.
[^7]: Note that these authors used random values for the cluster’s velocity components perpendicular to the line of sight.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A strong laser field and the Coulomb field of a nucleus can produce [$e^+e^-~$]{}pairs. It is shown for the first time that there is a large probability that electrons and positrons created in this process collide after one or several oscillations of the laser field. These collisions can take place at high energy resulting in several phenomena. The quasielastic collision [$e^+e^-~$]{}$\rightarrow$ [$e^+e^-~$]{}allows acceleration of leptons in the laser field to higher energies. The inelastic collisions allow production of high energy photons $e^+e^-\rightarrow 2\gamma$ and muons $e^+e^-\rightarrow \mu^+\mu^-$. The yield of high-energy photons and muons produced via this mechanism exceeds exponentially their production through conventional direct creation in laser and Coulomb fields. A relation of the phenomena considered with the antenna-mechanism of multiphoton absorption in atoms is discussed.'
author:
- 'M. Yu. Kuchiev'
title: 'Production of high energy particles in laser and Coulomb fields and [$e^+e^-~$]{}antenna'
---
Creation of [$e^+e^-~$]{}pairs in a laser field has similarity with the Schwinger mechanism of pair production by a static electric field [@Sch51]. The [$e^+e^-~$]{}production by laser and Coulomb fields may also be compared with Ref.[@BH34], being sometimes called the nonlinear Bethe-Heitler process. This mechanism of [$e^+e^-~$]{}production is considered in literature as a possible candidate for experimental studies for a new generation of lasers [@Rin01]. Various theoretical aspects of this process were discussed in Refs. [@Yak66; @Mit87; @MVG03; @MVG032; @MVG04; @DP98; @AAMS-03; @MMHJK06; @KR07]. A popular approach to the problem uses Volkov wave functions for leptons in a laser field [@Vol35], which is close to the Keldysh approximation [@Kel65] for multiphoton phenomena in atoms [@NR66; @PPT66; @Fai73; @Rei80; @GK97]. The recent Ref. [@KR07] simplified the formalism for the [$e^+e^-~$]{}production and derived several important characteristics of the problem in simple analytical form.
The leptons $e^+$ and $e^-$ initially created in the vicinity of the nucleus, propagate out of the Coulomb center along classical trajectories, which exhibit wiggling in the laser field. Obviously, $e^+$ and $e^-$ follow different trajectories. However, we will see that the wiggling can force the leptons to collide after one or several periods of laser oscillations. The probability of this collision proves surprisingly large. At the moment of collision the energy of the [$e^+e^-~$]{}pair can be high, which results in several interesting phenomena. The quasielastic collision [$e^+e^-~$]{}$\rightarrow$ [$e^+e^-~$]{}can make energy of a lepton higher than its initial energy. There is also a possibility for inelastic collisions, which result in production of high-energy particles. (The term high-energy in this work is applied energies above the electron mass.) This includes production of high-energy photons via annihilation [$e^+e^-~$]{}$\rightarrow 2\gamma$, and production of heavy particles, such as muons, [$e^+e^-~$]{}$\rightarrow \mu^+\mu^-$.
In the phenomena considered the [$e^+e^-~$]{}pair fulfills two major functions. First, it accumulates energy from the laser field, and then it transfers this energy to other high-energy particles. This reminds closely the work of conventional aerials in radio devices. Moreover, physical reasons, which ensure effective accumulation of energy by the [$e^+e^-~$]{}pair and by aerials are very similar. In both cases it is vital that leptons propagate large distances while exhibiting oscillations in an external electromagnetic field. This similarity makes it convenient to think about the [$e^+e^-~$]{}pair in the phenomena discussed as if it was a particular [$e^+e^-~$]{}antenna. A similar idea was put forward in relation to multiphoton processes in atoms [@kuchiev-87; @Corkum-93; @Schafer-93], where it was called atomic antenna, or rescattering mechanism. Its validity was confirmed experimentally, see e.g. [@Weckenbrock-etal-03; @Pedregosa-Gutierrez-etal-04] and references therein.
Consider a linearly polarized laser field, which propagates along the $z$-axis with electric and magnetic fields stretched along the $x$ and $y$ axes. Describe this field by the vector-potential $e \mathbf{A}(\varphi)= - m \,\xi\, (\sin\varphi,0,0)$, where $\varphi=\omega(t-z)$, $\omega$ is the laser frequency, $\xi=eE/m\omega$ is an adiabatic parameter, $e>0$ and $m$ are the absolute value of the electron charge and electron mass ($\hbar=c$). Consider the tunneling regime presuming that the laser field is strong, though below the critical value $E_\mathrm{c} =m^2/e$, and $\omega$ is low $$\label{tun}
{\mathcal{F}}\equiv E/E_\mathrm{c}=eE/m^2\ll 1, \quad\quad \xi=eE/m\omega\gg 1~.$$ We will base our discussion on the differential rate of the [$e^+e^-~$]{}production $dW_{e^+e^-}$ found in [@KR07] $$\begin{aligned}
\label{dW}
dW_{e^+e^-}=&B \exp \bigg(-\frac{ 2\sqrt{3} } {{\mathcal{F}}} \, \,\frac{\cal{A}}{m^2} \,\bigg) \,\,\frac{ d^3 p_+\, d^3 p_-}{(2\pi)^6}~,
\\ \label{A}
\cal{A} = &
\frac{(k_+ - k_-)^2}{8\xi^2}+
\frac{(k_+ + k_-)^2}{4}
+\frac{\kappa_+^2+\kappa_-^2}{2}
\\ \nonumber
&\quad \quad + \frac{ 7 (\delta \eta_+^2+\delta \eta_-^2)+ 2\delta \eta_{+} \delta \eta_{-} }{12}~.\end{aligned}$$ Here the states of the positron and electron are described by momenta, which in the Cartesian coordinates read $\mathbf{p}_\pm=(k_\pm,\kappa_\pm,q_\pm)$. The subscripts $\pm$ mark variables of the positron and electron. It is useful to express the longitudinal components of momenta $q_\pm=(m^2+k_\pm^2+\kappa_\pm^2-\eta_\pm^2)/(2\eta_\pm)$ via the variables $\eta_\pm$. The $\eta_\pm$ variables have mean values $\bar {\eta} =m/\sqrt{2}$ [@KR07], deviations from them are called $\delta \eta_\pm= \eta_\pm-\bar {\eta}$. The convenience of $\eta_\pm$ prompts one to use the definition $d^3 p_\pm\equiv dk_\pm
d\kappa_\pm d\eta_\pm$ for integrations over momenta. The constant $B$ in Eq.(\[dW\]) was calculated in [@KR07]. Integrating Eq.(\[dW\]) with this constant one reproduces the rate of the [$e^+e^-~$]{}production $$\label{W}
W_{e^+e^-}=\frac{ Z^2\alpha^2 m }{ \sqrt{2}\,\,\pi }\bigg( \frac{ {\mathcal{F}}}{ 2\sqrt{3} }\bigg)^3 \exp
\bigg( -\frac{ 2\sqrt{3} }{ {\mathcal{F}}} \,\bigg)~.$$ Here $Z$ is the charge of the nucleus, $\alpha=e^2\simeq1/137$. Eq.(\[W\]) complies with the results obtained previously (see [@KR07; @MMHJK06] for discussion and references), which confirms the validity of Eq.(\[dW\]).
Eqs.(\[dW\]),(\[A\]) show that the momenta distribution of the [$e^+e^-~$]{}pair has a maximum when $\mathbf{p}_+=\mathbf{p}_-=\big(0,0,m/(2\sqrt2)\,\big)$. Deviations of momenta from these values are small, proportional to $\sqrt{{\mathcal{F}}}\,m \ll m$. An important exception is the case when momenta of both leptons satisfy $$\label{soft}
k_++k_-=0~, \quad \kappa_\pm=\delta \eta_\pm=0~.$$ Then only the first term in Eq.(\[A\]) is present. A large factor $\xi^2$ in its denominator makes possible relatively large variations of momenta $|\delta(k_+-k_-)|=2|\delta k_-|$, $|\delta k_-|\sim \xi \sqrt{{\mathcal{F}}}\,m\gg \sqrt{{\mathcal{F}}}\,m$. Consequently, we find that the integration in Eq.(\[dW\]) is saturated in the close vicinity of the region, where momenta satisfy Eq.(\[soft\]). We derive a simple and important conclusion that leptons are predominantly created with momenta which satisfy Eq.(\[soft\]).
An [$e^+e^-~$]{}pair is initially created in the close vicinity of the nucleus $r\le m^{-1}$ [@KR07]. However, after a short period of time the leptons $e^+$ and $e^-$ go away from the Coulomb center into the region where the laser field is stronger than the Coulomb one. Further propagation of leptons is guided mostly by the laser field. The semiclassical nature of this field allows one to base analysis on the classical trajectories of leptons $\mathbf{r}=\mathbf{r}(\varphi)$, which can be written in a simple form (compare e.g. [@LL02-87-sh]) $$\label{r}
\mathbf{r}(\varphi)=
\frac{1} {\omega \eta } \int_0^ \varphi \mathbf{P}(\varphi')\,d\varphi'~.$$ Here $\mathbf{P}(\varphi)=(K(\varphi), \kappa, Q(\varphi))$ is the instant momentum, $$\begin{aligned}
\label{PT}
&K(\varphi)=k\mp eA(\varphi)~,
\\\label{PL}
&Q(\varphi)=( m^2+K^2(\varphi)+\kappa^2-\eta^2 )/(2\eta)~.\end{aligned}$$ The instant energy of a lepton is $\varepsilon(\varphi)=Q(\varphi)+\eta$. In Eqs.(\[r\])-(\[PL\]) the subscripts $\pm$ are suppressed, the upper (lower) sign in Eq.(\[PT\]) corresponds to the case of positron (electron). One can anticipate that the [$e^+e^-~$]{}pair is created at the moment of time when the laser field is the strongest, i.e. at $t=0$ or $t=\pi/\omega$. Detailed calculations, see e.g. [@KR07], support this expectation. Without loss of generality we can presume that the pair is created at $t=0$; we also assume that the nucleus is located at the origin. This implies the initial condition $\mathbf{r}_\pm(0)=0$. Additionally, the initial momenta of the created pair should satisfy $\mathbf{P}_\pm(0)=\mathbf{p}_\pm$. These initial conditions are accounted for in Eqs.(\[r\])-(\[PL\]) by the chosen constants of integration.
We know that leptons of an [$e^+e^-~$]{}pair are predominantly created with momenta, which satisfy Eq.(\[soft\]). Let us examine a possibility of their collision in this case. Eq.(\[soft\]) implies that trajectories of leptons are characterized by one parameter, which can be taken as the $x$-component of the electron momentum $k\equiv k_-$. Other components of momenta are $\boldsymbol{p}_\pm=(\mp k,0,q)$, where $q=(m^2+k^2-\bar{\eta}^2)/(2\bar{\eta})$ and $\eta_\pm=\bar{\eta}\equiv m/\sqrt{2}$. Using these facts and Eqs.(\[PT\]),(\[PL\]) one finds that the longitudinal momenta and energies of both leptons are equal, $P_+(\varphi)=P_-(\varphi)$, $\varepsilon_+(\varphi)=\varepsilon_-(\varphi)$. Eq.(\[r\]) ensures then that the $z$-components of the radius-vectors of leptons are same for any $\varphi$, while the $y$-components remain zero, $$\begin{aligned}
\label{yz}
y_+(\varphi)=y_-(\varphi)=0~,\quad \quad ~ z_+(\varphi)=z_-(\varphi)~.\end{aligned}$$ Thus, the only condition, which needs to be satisfied to ensure that the collision takes place is given by the $x$-components of coordinates, $x_+(\varphi)=x_-(\varphi)$. Using Eqs.(\[r\]),(\[PT\]) one rewrites it as an equation on $\varphi$ $$\label{k}
k=-\frac{1}{\varphi}\int_0^\varphi eA(\varphi') \,d\varphi'=m\xi\,\frac{1-\cos\varphi}{\varphi}
~.$$ Here $k=k_-$, $\varphi=\omega(t-z)$ with $t>0$ and $z<t$, which implies $\varphi>0$. Eq.(\[k\]) shows that for $k=0$ an infinite number of collisions with $\varphi=2\pi n$, $n=1,2, \dots$ are possible; with the increase of $k$ the number of possible collisions decreases; collisions do not take place when either $k > 0.725\, m\xi$, or $k<0$, where $0.725 = \max\big((1-\cos\varphi)/\varphi\big)$.
We conclude that if momenta of leptons satisfy Eq.(\[soft\]) then the leptons definitely collide provided $0\le k\le 0.725\,m\xi$. The total instant energy $\varepsilon(\varphi) =2\varepsilon_-(\varphi)$ of colliding leptons equals $$\begin{aligned}
\label{etot}
\varepsilon(\varphi)
=\sqrt{2}\,m \big[\,\xi^2\,\big(\,(1-\cos\varphi)/\varphi- \sin\varphi\big)^2+3/2\,\big] ~.\end{aligned}$$ We take into account here that $\varepsilon_-(\varphi)=Q_-( \varphi )+\bar{\eta}$, use Eq.(\[PL\]), and express $k$ via $\varphi$ using Eq.(\[k\]). We see that at the moment of collision the energy of the pair can be high, $\varepsilon(\varphi) \propto \xi^2 m\gg m$. The highest energy $\varepsilon(\varphi) = 2.21\, \xi^2 m$ is reached when $\varphi = 1.409\,\pi$ and $k= 0.290\, m\xi$. For $k=0$ collisions take place at much lower energy $\varepsilon(2\pi n)=3m/\sqrt{2}$.
Let us find the rate $W_\mathrm{c}$ with which the collision events take place. With this purpose we need to multiply the [$e^+e^-~$]{}creation probability by a factor $f$, which accounts for the fact that the pair would eventually collide, $dW_\mathrm{c}=f\,dW_{e^+e^-}$. Clearly, $f$ is proportional to the cross section $\tilde \sigma$ of a collision process considered. At the moment of collision the impact parameter $b$ of the [$e^+e^-~$]{}pair is small, $b^2\sim \tilde \sigma$. One deduces from this that the necessary factor reads $ f=\tilde {\sigma} \delta^{(2)}
(\mathbf{b})$, which can also be conveniently rewritten as $f=\int_0^{\infty} \!\tilde \sigma \,v\,
\delta \big(\,\mathbf{r}_{+}(\varphi)-\mathbf{r}_{-}(\varphi)\,\big)dt$. Here $v=|\mathbf{v}|=|\mathbf{v}_+-\mathbf{v}_-|$ is the collision velocity of the [$e^+e^-~$]{}pair, a two-dimensional vector $\mathbf{b}$, which is orthogonal to the collision velocity, $\mathbf{v}\cdot \mathbf{b}=0$, represents an impact parameter for the [$e^+e^-~$]{}collision. The integration over the moment of time of collision $t$ is introduced in $f$ to simplify calculations. Calculating $W_\mathrm{c}=\int f\,dW_{e^+e^-}$ using Eq.(\[dW\]) one can integrate over electron momenta $d^3p_-$ with the help of the delta-function from $f$, which gives the determinant $\det\big( \partial \boldsymbol{p}_-/ \partial \boldsymbol{r}_{-} \big)$. It is convenient to rewrite the integral over $dt$ in terms of integration over the electron momentum $dk_-$, which brings in the factor $\partial t/\partial k_-$. This determinant and derivative are calculated directly from the classical trajectory Eqs.(\[r\])-(\[PL\]). The final result reads $$\begin{aligned}
\label{fin}
W_\mathrm{c}=B\int
\! \exp \bigg(\!-\frac{ 2\sqrt{3}}{{\mathcal{F}}}\frac{\bar{\cal{A}}}{m^2 }\bigg)\,
\frac{\sigma\, \varepsilon \,m\, w}{(t-z)^2\,v}
\, \frac{d^3 p_{+}dk_{-}\!}{(2\pi)^6}~.\end{aligned}$$ Here $t=t(\varphi)$ and $z=z(\varphi)$ are the moment of time and $z$-coordinate of [$e^+e^-~$]{}collision, $\varepsilon=\varepsilon(\varphi)$ is the instant energy of the [$e^+e^-~$]{}pair at the moment of collision, $\sigma$ is the cross section in the center of mass reference frame. The transition from the cross section in the laboratory reference frame $\tilde \sigma$ to $\sigma$ prompts the appearance of a modified velocity $w=[v^2-(\mathbf{v}_+ \times\mathbf{v}_-)^2]^{1/2}$, see e.g. [@LL02-87-sh].
The integration in Eq.(\[fin\]) runs over all possible momenta that lead to a collision. This includes any configuration of momenta, which satisfy Eqs.(\[soft\]),(\[k\]). Moreover, using the classical trajectories Eq.(\[r\])-(\[PL\]) it it easy to verify that if electron and positron momenta are close to a configuration specified by these equations (which is very probable, as Eqs.(\[dW\]) shows), then a collision can definitely take place provided the momenta of the [$e^+e^-~$]{}pair satisfy $$\begin{aligned}
\label{ke}
\kappa_-= \kappa_+~,\quad \delta\eta_-=\delta\eta_+~. \end{aligned}$$ The factor $\bar{\cal{A}}$ in Eq.(\[fin\]) equals $\cal{A}$ in Eq.(\[A\]) evaluated for these particular momenta. This discussion ensures that there is no additional suppression in the integral over $d^3p_+dk_-$ in Eq.(\[fin\]), compared to the integral over same four variables in Eq.(\[dW\]). Both these integrals are saturated in approximately same integration volume. As a result, we can find an estimate for the ratio $R$ of the number of collisions of [$e^+e^-~$]{}pairs to the total number of created [$e^+e^-~$]{}pairs simply by comparing the integrands in Eqs.(\[dW\]),(\[fin\]) for this 4D integration $$\begin{aligned}
\label{R}
R = \frac{ W_\mathrm{c} }{ W_{ e^+e^- }}\approx
\frac{ \sigma\, \varepsilon \,m\,w}{(t-z)^2\,v} \frac{1}{{\mathcal{F}}m^2}\approx
\frac{ \sigma \, \omega \,m}{\xi}~.\end{aligned}$$ The factor ${\mathcal{F}}m^2$ in the denominator of the middle expression here arises from an integration over $\kappa_-$ and $\eta_-$ in Eq.(\[dW\]). The last equality in Eq.(\[R\]) takes into account that $t\sim z\sim \xi^2/\omega$, $t-z\sim 1/\omega$, as Eqs.(\[r\])-(\[PT\]) show, and that according to Eq.(\[etot\]) $\varepsilon \sim \xi^2m$, which also implies $w\simeq v\simeq 1$. In Eq.(\[R\]) $\sigma$ describes a collision process considered, see examples in (\[qe\])-(\[mumu\]) below.
A simple estimation $R_0$ for a relative number of [$e^+e^-~$]{}collisions compares $\sigma$ with a distance $z$ to the point of collision, giving $R_0 = \sigma /z^2\approx \sigma
\omega^2/\xi^4$, which is much smaller than Eq.(\[R\]) is predicting. Thus, Eq.(\[R\]) shows that there exists a large enhancing factor ${\cal K}$, which makes collisions more probable $$\begin{aligned}
\label{K}
R \approx {\cal K}\,R_0~,\quad\quad {\cal K}=\xi^4/{\mathcal{F}}\gg 1~.\end{aligned}$$ An analysis of Eq.(\[R\]) reveals two physical reasons, which combine to produce the enhancement. Eq.(\[dW\]) ensures that momenta of the leptons created are necessarily close to a configuration given by Eq.(\[soft\]), in which case a collision is almost imminent (since the necessary additional condition Eq.(\[k\]) is satisfied very easily). There is also a relativistic factor, which arises due to stretching of lepton trajectories along the $z$-axis that makes collisions of electrons and positron even more probable.
Collisions of [$e^+e^-~$]{}pairs bring interesting implications. First of all there are quasielastic collisions $$\begin{aligned}
\label{qe}
e^+ + e^-\rightarrow e'^{\,+} + e'^{\,-}~. \end{aligned}$$ Any collision process of this type takes only a small fraction of the period of laser oscillations. Therefore the conservation laws in a collision process are applied to the instant energy and instant momentum. A redistribution of these quantities in the reaction (\[qe\]) may lead to a strong increase of a lepton quasienergy $ \bar {\varepsilon}_\pm $, which is defined as the energy averaged over the period of laser oscillation, $\bar {\varepsilon}_\pm=\int_0^{2\pi}\varepsilon_\pm(\varphi)\,d\varphi/(2\pi)$. To illustrate this statement assume that the momenta of a pair of leptons satisfy Eq.(\[soft\]). Then using kinematics arguments one finds that the quasienergy of the lepton in the final state can be as large as $\bar{\varepsilon}_\pm'=\varepsilon(\varphi)+\xi^2 k^2/(\sqrt{2}m)$. Here $\varepsilon(\varphi)$ is the instant energy of the initially created [$e^+e^-~$]{}pair. The term $\propto \xi^2 k^2$ in the quasienergy can become parametrically large provided the frequency of the laser field is sufficiently low $\omega \ll (E/E_c)^{3/2}\,m$. However, even if this condition is not satisfied, the increase of the quasienergy of the lepton can be quite sizable because the instant energy can be higher than the quasienergy.
There are also inelastic collisions. An important example is the creation of $\gamma$-quanta $$\begin{aligned}
\label{2g}
e^+ + e^-\rightarrow 2\,\gamma~. \end{aligned}$$ The energy and momentum of photons created in this reaction equal the instant energy and momentum of the [$e^+e^-~$]{}pair at the moment of collision, which can be large prompting the energy of $\gamma$-quanta to be large as well. Eq.(\[etot\]) shows that the highest possible instant energy of the pair is $\varepsilon(\varphi)=\ 2.21 \xi^2 m$, which gives the upper limit for the energy of each photon $\varepsilon_{\gamma,\,\mathrm{max}}=1.10\, \xi^2 m$.
It is important that (\[2g\]) provides the most probable way for production of high-energy photons by a laser and Coulomb fields. Indeed, the high-energy photons cannot be produced in the region of small distances. Their production there can be considered as an effective addition to the mass of leptons, $m\rightarrow m+\delta m$, which inevitably results in exponential decrease of the probability Eq.(\[W\]). The high-energy $\gamma$-quanta also cannot be produced during propagation of a lepton in the laser field. The adiabatic nature of this propagation ensures exponential suppression of this process. In contrast, the photon yield through reaction (\[2g\]) is relatively large, as Eq.(\[R\]) shows.
Another interesting example is the creation of heavy particles, for example muons in the process $$\begin{aligned}
\label{mumu}
{\varepsilon}^+ + e^-\rightarrow \mu^+ + \mu^-.\end{aligned}$$ To make this process possible the instant energy of the [$e^+e^-~$]{}pair in the center of mass reference frame should exceed the threshold of the [$\mu^+\mu^-~$]{}production. The latter is given by the minimum of the instant energy of muons, which equals $2m_\mu$. From Eqs.(\[k\])-(\[etot\]) one deduces that the center of mass instant energy of the [$e^+e^-~$]{}pair is $\varepsilon_\mathrm{scm}(\varphi)=2(m^2+k^2)^{1/2}$, which is reduced to $2 k$ provided $k\gg m$. Thus the threshold condition for [$\mu^+\mu^-~$]{}production restricts $k$ from below $k\ge m_\mu$. At the same time Eqs.(\[dW\]),(\[A\]) restrict it from above, $k\lesssim m\,\xi\, \sqrt{{\mathcal{F}}}$. Combining both conditions, one concludes that the muon production via collisions of [$e^+e^-~$]{}pairs is probable provided $ m_\mu/m \lesssim \xi \sqrt {\mathcal{F}}$. This condition can be understood as a restriction on the laser frequency $$\begin{aligned}
\label{om}
\omega ~\lesssim~ (E/E_c)^{3/2}\,m^2/m_\mu\ll m~.\end{aligned}$$ Energy absorbed from the laser field when the [$e^+e^-~$]{}pair is initially created is typically $\varepsilon \simeq \xi^2 m/\sqrt{2}$ [@KR07]. Meanwhile, the energy of particles in the final state of all phenomena considered in (\[qe\])-(\[mumu\]) can be significantly higher. Thus, collisions of leptons of the [$e^+e^-~$]{}pair result in additional absorption of energy from the field.
This fact reminds us again that in the processes discussed the [$e^+e^-~$]{}pair acts as an antenna. We verified that it can efficiently accumulate high energy from a laser field and then transfer it to other particles. As a result, creation of muons and high-energy photons becomes possible, which is notable since their production through the conventional mechanism is suppressed exponentially. The [$e^+e^-~$]{}antenna removes this exponential suppression, making creation of high-energy particles much more probable.
I am thankful to D.J.Robinson for discussions. This work was supported by the Australian Research Council.
[25]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
, ****, (), .
, ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
, , , , , ****, ().
, ().
, ****, ().
, ****, (), .
, ****, (), .
, , , ****, (), .
, ****, ().
, ****, ().
, ****, ().
, ****, (), .
, ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ** (, , ).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper introduces an Algebraic MultiScale method for simulation of flow in heterogeneous porous media with embedded discrete Fractures (F-AMS). First, multiscale coarse grids are independently constructed for both porous matrix and fracture networks. Then, a map between coarse- and fine-scale is obtained by algebraically computing basis functions with local support. In order to extend the localization assumption to the fractured media, four types of basis functions are investigated: (1) Decoupled-AMS, in which the two media are completely decoupled, (2) Frac-AMS and (3) Rock-AMS, which take into account only one-way transmissibilities, and (4) Coupled-AMS, in which the matrix and fracture interpolators are fully coupled. In order to ensure scalability, the F-AMS framework permits full flexibility in terms of the resolution of the fracture coarse grids. Numerical results are presented for two- and three-dimensional heterogeneous test cases. During these experiments, the performance of F-AMS, paired with ILU(0) as second-stage smoother in a convergent iterative procedure, is studied by monitoring CPU times and convergence rates. Finally, in order to investigate the scalability of the method, an extensive benchmark study is conducted, where a commercial algebraic multigrid solver is used as reference. The results show that, given an appropriate coarsening strategy, F-AMS is insensitive to severe fracture and matrix conductivity contrasts, as well as the length of the fracture networks. Its unique feature is that a fine-scale mass conservative flux field can be reconstructed after any iteration, providing efficient approximate solutions in time-dependent simulations.'
address:
- |
Department of Geoscience and Engineering, Delft University of Technology,\
P.O. Box 5048, 2600 GA Delft, The Netherlands.
- 'The Petroleum Institute, P.O. Box: 2533, Abu Dhabi, United Arab Emirates.'
author:
- Matei Ţene
- Mohammed Saad Al Kobaisi
- Hadi Hajibeygi
title: 'Algebraic multiscale method for flow in heterogeneous porous media with embedded discrete fractures (F-AMS)'
---
algebraic multiscale methods,flow in porous media,naturally fractured porous rock,heterogeneous permeability,scalable linear solvers.
Introduction {#sec:intro}
============
In many geoscience applications, including hydrocarbon production and geothermal energy exploitation, the target formations are naturally fractured. These formations are often highly heterogeneous matrix rock is crossed by several networks of lower-dimensional highly-conductive fractures at multiple length scales [@Berkowitz1]. This raises important challenges for flow simulation, motivating the development of advanced modelling and numerical solution techniques.
Among the proposed methods, the hierarchical fracture modelling approach allows for avoiding complexities associated with the discretization and dynamic nature of fracture geometries [@SeongFrac2]. In this approach, small-scale fractures (below the matrix grid resolution) are homogenized within the matrix rock, altering its effective permeability [@Sander]. The remaining fractures are then represented as explicit control volumes [@SeongFrac1; @hadi-frac-jcp]. If the fracture and matrix grids are generated independently, then the formation is said to be discretized according to the Embedded Discrete Fracture Model (EDFM) [@Moinfar1]. Alternatively, the fracture cells can be constrained to lie at the interfaces of matrix cells, i.e. by employing Discrete Fracture Modelling methods (DFM), which require unstructured grids [@karimi-fard]. Both DFM and EDFM have been applied to reservoirs with complex fracture geometries [@stephanM1] and fluid physics [@rainer1; @Moinfar2]. Recent developments include higher-order approximation schemes within finite-volume [@dfm-Ahmed2015] and finite-element [@Geiger09] methods.
Note that the total number of degrees of freedom (DOF), even after homogenizing small-scale fractures, is beyond the scope of classical simulation methods. This motivates the development of an efficient multiscale method for heterogeneous fractured porous media.
Multiscale finite element (MSFE) [@Hou97] and finite volume (MSFV) methods [@Jenny03] have been introduced and evolved mainly for heterogeneous, but non-fractured, porous media (see [@Kippe06] for a comparison). Recent developments include efficient solution of the pressure equation for multi-component displacements within sequentially- [@SEON-BO; @hadi-compositional-spej] and fully-implicit [@cusini; @ADM] frameworks. They have also been extended to capture complex wells [@Patrick-well2; @Wolfsteiner-well] and to the transport equations [@zhou-trans]. In addition, enriched multiscale methods are targeted at media with high conductivity contrasts [@Yalchin-enriched1; @Yalchin-enriched2; @Davide14], with modifications to maintain their monotonicity [@Hesse08a; @yixuan-monotonemsfv]. More recently, a multiscale formulation has been devised to support unstructured grids [@MsRSB-olav].
The first application of MSFV methods to fractured reservoirs was developed on the basis of coupling the matrix pressure to the average pressure in each fracture network, at coarse scale [@hadi-frac-jcp]. Based on their results for two-dimensional (2D) problems, this method was efficient for media with highly conductive fracture networks which span short spatial length scales (relative to that of the domain). However, convergence was observed to degrade for test cases with significant variations in the pressure distribution along the fracture network. In combination with streamline-based mixed formulations, multiscale methods have also been employed to 2D fractured reservoirs [@Natvig09]. More recently, a multiscale approach was developed for 2D reservoirs which assigned one coarse node at each fracture intersection only, with no coarse nodes in the matrix [@Sandve]. Note that none of these methods include 3D heterogeneous reservoirs nor has their performance been benchmarked against a commercial linear solver. More importantly, the literature is lacking a multiscale method which allows for flexible coarse grids inside the matrix as well as its embedded fractures and, thus, able to accommodate heterogeneous cases with fracture networks of different length scales.
This paper presents the development of an Algebraic MultiScale method for heterogeneous Fractured porous media (F-AMS) using EDFM. Given a partition of the fine-scale cells into primal and dual coarse blocks for both the matrix and fracture networks, the algorithm algebraically constructs the multiscale prolongation (mapping coarse- to fine-scale) and restriction (mapping fine- to coarse-scale) operators. The prolongation operator columns are the local basis functions, solved on dual-coarse cells, for both matrix and fractures.
F-AMS supports four different matrix-fracture coupling strategies, at the coarse-scale. First, the Decoupled-AMS basis functions are defined by neglecting the contribution of a medium’s coarse solution (e.g., fractures) in the interpolated solution in the other (e.g., matrix), thus preserving sparsity in the resulting coarse-scale system. Then, two semi-coupled (one-way) strategies, Rock-AMS and Frac-AMS, are considered. The Rock-AMS approach constructs a prolongation operator in which the matrix coarse solutions also contribute in computing the interpolated fine-scale solution in neighbouring fractures. Similarly, Frac-AMS considers the influence of the fracture coarse solution when interpolating the pressure inside the surrounding porous rock. Finally, the fully coupled strategy, Coupled-AMS, is devised, where coarse-scale solutions from both media play a role in finding the fine-scale solution of each other. This last approach, although allowing for full fracture-matrix coupling, leads to a dense coarse-scale system with additional overhead during the associated algebraic (matrix-vector, matrix-matrix) operations. As such, for practical applications, the Coupled-AMS prolongation operator may require tuning via truncation, where values below a specified threshold are algebraically deleted, followed by a rescaling step, to maintain partition of unity. This option is also investigated in the paper.
To summarize, F-AMS allows for arbitrary coarse grid resolutions in both fractures and matrix, as well as all possible coarse-scale coupling between them. Furthermore, once these coarse grids are defined, the F-AMS procedure is formulated and implemented in algebraic form, in line with the previously published formulations of incompressible (AMS) [@yixuan-ams] and compressible (C-AMS) [@Tene-cams] flows. In the limit, if the Frac-AMS coupling strategy and only one coarse node per fracture network is used, F-AMS automatically reduces to the method described in [@hadi-frac-jcp]. However, this setup proves inefficient for many of the test cases in this paper. From a bottom-up perspective, F-AMS extends the AMS prolongation operator, as previously described in [@yixuan-ams], with additional columns. Some of these columns correspond to the enriched fracture coarse domain, as explained above. The remainder represent local well basis functions for Peaceman wells [@Patrick-well2].
In order to test F-AMS method, a proof-of-concept implementation is developed with a focus on reservoirs defined on 3D structured grids with embedded vertical fracture plates (for the challenges associated with unstructured multiscale simulation see [@MsRSB-olav]). For the presented experiments, a finite-element (FE) restriction operator is employed to obtain a symmetric-positive-definite coarse system. If approximate (non-converged) F-AMS solutions are used, in the last iteration step, a finite-volume (FV) restriction operator is employed followed by mass-conservative reconstruction of fine-scale flux field. The performance of this in-house object oriented serial-processing simulator was measured based on CPU times, as well as convergence rates. Numerical test cases are considered in order to study the effect of the different components of the algorithm, namely, the coarsening ratios and basis function coupling strategies. F-AMS is developed as an accurate multiscale (approximate) pressure solver. In order to assess its scalability, however, its performance (based on CPU times) is compared against the commercial Algebraic MultiGrid (AMG) solver, SAMG [@SAMG], as a preconditioner. Results of these systematic studies show that only few DOF per fracture network are necessary to obtain a good trade-off between convergence rate and computational expense. In conclusion, F-AMS is found efficient and scalable for solving flow in heterogeneous and naturally fractured porous media. Its development marks an important step forward towards the integration multiscale methods as “black-box” pressure solvers within existing reservoir simulators, with the possibility of extension to more complex physics and scenarios.
The paper is structured as follows. First, the EDFM fine-scale discrete system is described in Section \[sec:equations\]. Then, the components of the F-AMS algorithm are detailed in Section \[sec:fams\]. Section \[sec:results\] consists of numerical results for both 2D and 3D test cases. Finally, conclusions and remarks make the subject of Section \[sec:conclusions\].
Fine-scale discretized system {#sec:equations}
=============================
The mass-conservation equations for single-phase flow in fractured media, using Darcy’s law, can be written as $$\begin{aligned}
\bigg[\frac{\partial (\phi \rho)}{\partial t}-\nabla \cdot (\rho \lambda \cdot \nabla p)\bigg]^m &= [\rho q]^{mw} + [\rho Q]^{m} + [\rho q]^{mf} \hspace{5mm} \text{on} \hspace{2mm} \Omega^m \subset R^n \label{pm}\end{aligned}$$ for the matrix (superscript $^m$) and $$\begin{aligned}
\bigg[\frac{\partial (\phi \rho)}{\partial t}-\nabla \cdot (\rho \lambda \cdot \nabla p)\bigg]^f &= [\rho q]^{fw}+ [\rho Q]^{f} + [\rho q]^{fm} \hspace{5mm} \text{on} \hspace{2mm} \Omega^f \subset R^{n-1}\label{pf}\end{aligned}$$ for the fracture (superscript $^f$) spatial domains. Here, the phase mobility $\lambda = k^{*}/\mu$ consists of fluid viscosity ($\mu$) and effective fracture ($k^{*} = k^f$) or rock ($k^{*} = k^m$) permeability. Note that the latter can also account for the homogenized small-scale fractures, as described in the hierarchical fracture model [@SeongFrac1; @Sander]. Also, $\rho$ and $\phi$ are the fluid density and rock porosity, respectively. Here, for simplicity, fracture porosity is always considered to be $1$. The $q^{mw}$ and $q^{fw}$ denote the matrix and fracture external source terms, respectively, i.e. from injection/production wells. For a perforated matrix control volume $V$, it reads $$\begin{aligned}
q^{mw}_V~ =~ PI~ \lambda^m~ (p^w - p^m) / V~ \equiv~ \beta^m~ (p^w - p^m),\end{aligned}$$ where $\beta^m = PI~ \lambda^m / V$, $q^{mw}_{tot} = \int_V q^{mw}_V dV$ is the total injection rate, and $PI$ is the productivity index [@Peaceman_well]. Similarly, the fracture-matrix coupling terms are modeled such that for a matrix volume $V$ intersecting with a fracture surface $A$ one obtains $$\begin{aligned}
q^{mf}_V~ =~ CI~ \lambda^{f-m}~ (p^f - p^m) / V~ \equiv~ \eta^m~ (p^f - p^m)\end{aligned}$$ and $$\begin{aligned}
q^{fm}_A~ =~ CI~ \lambda^{f-m}~ (p^m - p^f) / A~ \equiv~ \eta^f~ (p^m - p^f),\end{aligned}$$ where $\eta^m = CI~ \lambda^{f-m} / V$ and $\eta^f = CI~ \lambda^{f-m} / A$. This ensures the total flux between a fracture element of area $A$ and a matrix element of volume $V$ is equal [@hadi-frac-jcp], i.e., $$\begin{aligned}
\int_{V} {q^{mf}_V} dV = - \int_{A} q^{fm}_A dA.\end{aligned}$$ The $\lambda^{f-m}$ is the effective mobility at the interface between the fractures and their surrounding matrix. The $CI$ is the connectivity index, defined on a discrete system as $$\begin{aligned}
CI_{ij} = \frac{A_{ij}}{\langle d \rangle_{ij}},\end{aligned}$$ where $A_{ij}$ is the area fraction of fracture element $i$ overlapping with the matrix element $j$, and $\langle d \rangle_{ij}$ is the average distance of the two elements [@hadi-frac-jcp]. Finally, the $Q^m$ and $Q^f$ terms describe other external source terms for matrix and fractures (e.g., gravity terms).
These equations are to be solved for matrix and fracture pressures, $p^m$ and $p^f$, on the matrix $\Omega^m$ and fracture $\Omega^f$ domains, as depicted in Fig. \[fig:fineGrid\]. Note that a fracture network can consist of several fractures, which are represented in a lower dimensional space, i.e. $\Omega^f \subset R^{n-1}$, than the matrix (reservoir rock) $\Omega^m \subset R^{n}$. The main advantage of this type of formulation is that the matrix and fracture grids are independent and, thus, can be freely adapted to accommodate the appropriate physics for each medium. This is especially important in highly fractured reservoirs or when fractures are generated (and closed) during simulation, e.g., in geothermal formations [@KarvounisGeothermal].
The incompressible single-phase pressure solution obtained using the EDFM approach for a 2D fractured reservoir model, shown in Fig. \[fig:fine2d\], is provided in Fig. \[fig:pres2d\_homo\]. Two pressure-constrained wells are placed on the East and West boundaries, and the reservoir rock is homogeneous.
When non-linearities are present (e.g., compressible flows), the flow equations need to first be linearized, i.e., $$\begin{aligned}
\label{dis-mf}
\bm A^{\nu} p^{\nu+1} \equiv {\left[\begin{array}{ccc}\bm A^{mm} & \bm A^{mf} & \bm A^{mw} \\ \bm A^{fm} & \bm A^{ff} & \bm A^{fw} \\ \bm A^{wm} & \bm A^{wf} & \bm A^{ww} \end{array}\right]^{\nu}}
\left[\begin{array}{c}p^m \\p^f\\p^w\end{array}\right]^{\nu+1}
=
\left[\begin{array}{c}q^m \\q^f \\ q^{w} \end{array}\right]^{\nu}
\equiv q^{\nu},\end{aligned}$$ and then iteratively solved, in a Newton-Raphson loop, until the converged solution is achieved. Note that this system shows an implicit treatment of the coupling between fracture and matrix through the $\bm A^{fm}$ entries, and that $\bm A$ can be non-symmetric, due to the compressibility effects [@Tene-cams; @hadi-comp-jcp].
Developing an efficient solution strategy for the linearized system is quite challenging for several reasons. On the one hand, the size of this system can exceed several millions of unknowns for realistic test cases. On the other hand, the value of the condition numbers for the system matrix is worsened by high contrasts between reservoir properties (matrix permeability is highly heterogeneous over large scales, fractures are typically much more conductive than the matrix, etc.).
Clearly a classical upscaling method cannot be employed here due to the highly resolved fractures, which play an important role in mass transport. This creates a niche for conservative multiscale methods, which have the important advantage of solving coarse-scale problems while honouring fine-scale data [@imsfv-jcp; @Zhou] in an iterative error reduction strategy [@giuseppe-ib08; @hadi-aimsfv-jcp; @SeongTrJcp] which allows for mass-conservative flux reconstruction at any stage [@Jenny06]. Next, the development of the F-AMS method is presented.
Algebraic Multiscale Method for Heterogeneous Porous Media with Embedded Discrete Fractures (F-AMS) {#sec:fams}
===================================================================================================
This section describes the F-AMS procedure, an efficient multiscale solution strategy for Eq. . Given a computational domain with $N_f$ fracture networks and $N_w$ wells, F-AMS first superimposes two coarse grids on top of both the matrix and fracture domains. The primal-coarse grid is a non-overlapping decomposition of the domain, inside which a fine-scale cell is selected as coarse node (Fig. \[fig:primal2d\] for 2D, and \[fig:primal3d\] and \[fig:primal3dfrac\] for 3D cases). By connecting the coarse nodes, a secondary overlapping coarse grid is obtained, which is called the dual coarse grid (Figs. \[fig:dual2d\], \[fig:dual3d\] and \[fig:dual3dfrac\]). There exist $N_{cm}$ and $N_{dm}$ matrix primal-coarse and dual-coarse blocks and, similarly, each fracture network $f_i$ contains $N_{cf_i}$ and $N_{df_i}$ fracture primal-coarse and dual-coarse blocks. Note that $N_w$ (injection or production) wells exist in the domain, as driving forces for the flow.
\
\
F-AMS approximates the solution to Eq. , $p$, as a superposition of coarse-scale solutions ($\breve p$) using locally computed basis functions ($\Phi$), i.e. $$\begin{aligned}
\label{pm-ms}
p^m \approx {p'^m} = \sum_{i=1}^{N_{cm}} \Phi_i^{mm} \breve p_i^m + \sum_{i=1}^{N_{f}} \sum_{j=1}^{N_{cf_i}} \Phi_j^{mf_i} \breve p_j^{f_i} + \sum_{k=1}^{N_{w}} \Phi_j^{mw} \breve p_k^{w},\end{aligned}$$ for the matrix and $$\begin{aligned}
\label{ms-f}
p^f \approx {p'^f} = \sum_{i=1}^{N_{cm}} \Phi_i^{fm} \breve p_i^m + \sum_{i=1}^{N_{f}} \sum_{j=1}^{N_{c{f_i}}} \Phi_j^{ff_i} \breve p_j^{f_i} + \sum_{k=1}^{N_{w}} \Phi_j^{fw} \breve p_k^{w}\end{aligned}$$ for the fractures, respectively. The basis functions associated with matrix coarse cells (i.e., $\Phi^{m*}$) are $\Phi_i^{mm}$ for matrix-matrix effects, $\Phi_j^{mf}$ for the matrix-fracture coupling, and $\Phi_j^{mw}$ matrix-well interactions. These basis functions (interpolators) are employed in order to capture the effects of all the important factors (matrix, fractures, and wells) in the construction of a good approximation for the matrix pressure field $p^m$. Similarly for fractures, $\Phi^{f*}$ consists of the contributions from the matrix $\Phi^{fm}$, fractures $\Phi^{ff}$, and wells $\Phi^{fw}$, if present.
One of the novel aspects of this work is that the pressure field inside fractures, $p^f$, is included explicitly in the multiscale formulation (Eq. ). This means that the fracture grid cells are also decomposed into primal and dual coarse blocks (Fig. \[fig:coarseGrids\]), similar to the matrix. Their solutions are also mapped to the coarse scale and back to the original resolution, again, similar to the matrix. More specifically, each fracture network $f_i$ is decomposed into $N_{cf_i}$ primal-coarse grid blocks, for which sets of basis functions are calculated. One could employ the same formulation for wells, i.e., discretize them into several fine-scale cells which can then be coarsened on the superimposed multiscale grids. However, for the sake of simplicity, in the experiments presented in this work, each well is assigned a single fine-scale DOF, which is mapped to the coarse-scale using the identity restriction operator, i.e. $$\begin{aligned}
p^{w_i} = p'^{w_i} = \breve p^{w_i} \hspace{1cm} \forall i \in \{1, \cdots, N_w\}.\end{aligned}$$
In algebraic notation, the superpositions and can be expressed as $$\begin{aligned}
\label{pm-ams}
p^m \approx {p'}^m = \bm{\mathcal{P}}^m \breve p \equiv [\bm{\mathcal{P}}^{mm} ~ ~ \bm{\mathcal{P}}^{mf} ~ ~ \bm{\mathcal{P}}^{mw}] ~ [\breve p^m ~ ~ \breve p^f ~ ~ \breve p^w]^T\end{aligned}$$ and $$\begin{aligned}
p^f \approx {p'}^f = \bm{\mathcal{P}}^f \breve p \equiv [\bm{\mathcal{P}}^{fm} ~ ~ \bm{\mathcal{P}}^{ff} ~ ~ \bm{\mathcal{P}}^{fw}] ~ [\breve p^m ~ ~ \breve p^f ~ ~ \breve p^w]^T,\end{aligned}$$ respectively.
The basis functions are assembled in the columns of the multiscale *prolongation operator*, $\bm{\mathcal{P}}$, with the dimension of $N_{fine} \times N_{coarse}$, where $N_{fine}$ and $N_{coarse}$ are the total number of fine- and coarse-scale control volumes, respectively. The part of $\bm{\mathcal{P}}$ corresponding to the matrix fine-cells is defined as $$\begin{aligned}
\label{pm-ams-pro}
\bm{\mathcal{P}}^m =
\left[\begin{array}{ccc|ccccccc|ccc}
\vdots & \cdots & \vdots & \vdots & \cdots & \vdots& \cdots & \vdots& \cdots&\vdots&\vdots & \cdots & \vdots\\
\Phi^{mm}_1 & \cdots & \Phi^{mm}_{N_{cm}} & \Phi^{mf_1}_1 & \cdots & \Phi^{mf_1}_{N_{cf_1}}& \cdots & \Phi^{mf_{N_f}}_1& \cdots& \Phi^{mf_{N_f}}_{N_{cf_{N_f}}} &\Phi^{mw}_1 & \cdots & \Phi^{mw}_{N_w}\\
\vdots & \cdots & \vdots & \vdots & \cdots & \vdots & \cdots & \vdots& \cdots&\vdots&\vdots & \cdots & \vdots
\end{array}\right].\end{aligned}$$ Notice the three sub-blocks which represent matrix-matrix, matrix-fracture, and matrix-well coupling. Similarly, the prolongation operator for fractures can be stated as $$\begin{aligned}
\bm{\mathcal{P}}^f =
\left[\begin{array}{ccc|ccccccc|ccc}
\vdots & \cdots & \vdots & \vdots & \cdots & \vdots& \cdots & \vdots& \cdots&\vdots&\vdots & \cdots & \vdots\\
\Phi^{fm}_1 & \cdots & \Phi^{fm}_{N_{cm}} & \Phi^{ff_1}_1 & \cdots & \Phi^{ff_1}_{N_{cf_1}}& \cdots & \Phi^{ff_{N_f}}_1& \cdots& \Phi^{ff_{N_f}}_{N_{cf_{N_f}}}&\Phi^{fw}_1 & \cdots & \Phi^{fw}_{N_w}\\
\vdots & \cdots & \vdots & \vdots & \cdots & \vdots & \cdots & \vdots& \cdots&\vdots& \vdots& \cdots&\vdots
\end{array}\right].\end{aligned}$$
Algebraically, the complete F-AMS prolongation operator reads $$\begin{aligned}
\label{prolongation_single}
\bm{\mathcal{P}} =
\left[\begin{array}{c}
\bm{\mathcal{P}}^{m} \\ \bm{\mathcal{P}}^{f} \\ \bm{\mathcal{P}}^{w}
\end{array} \right]
=
\left[\begin{array}{ccc}
\bm{\mathcal{P}}^{mm} & \bm{\mathcal{P}}^{mf} & \bm{\mathcal{P}}^{mw} \\
\bm{\mathcal{P}}^{fm} & \bm{\mathcal{P}}^{ff} & \bm{\mathcal{P}}^{fw}\\
\bm{\mathcal{P}}^{wm} & \bm{\mathcal{P}}^{wf} & \bm{\mathcal{P}}^{ww}\\
\end{array} \right],\end{aligned}$$ where $\bm{\mathcal{P}}^{wm}$ and $\bm{\mathcal{P}}^{wf}$ are set to zero, while $\bm{\mathcal{P}}^{ww}$ is the identity matrix.
Note that the prolongation operator, as described in Eq. , allows full flexibility in consideration of the fracture-matrix coupling in the interpolated solution, i.e. via the values in $\bm{\mathcal{P}}^{mf}$ and $\bm{\mathcal{P}}^{fm}$. This leads to the definition of four operators, differentiated by the coupling strategy they employ:
1. **Decoupled-AMS:** $ \bm{\mathcal{P}}^{mf} = 0$ and $\bm{\mathcal{P}}^{fm} = 0$
2. **Frac-AMS:** only $\bm{\mathcal{P}}^{fm} = 0$.
3. **Rock-AMS:** only $\bm{\mathcal{P}}^{mf} = 0$.
4. **Coupled-AMS:** $\bm{\mathcal{P}}^{mf}$ and $\bm{\mathcal{P}}^{fm}$ both non-zero.
The first option, i.e., Decoupled-AMS, constructs the most sparse $\bm{\mathcal{P}}$ and thus has an efficient setup phase. The fourth option, i.e., Coupled-AMS, can lead to more accurate multiscale simulations, however, it can severely increase the density of the multiscale operators. In such cases, one may be able to obtain a trade-off between the quality of the prolongation operator and its sparsity via truncation, followed by a rescaling of the rows to ensure partition of unity. A CPU-based study considering the overhead introduced by the density of $\bm{\mathcal{P}}$ is presented in Section \[sec:results\].
In order to construct the coarse-scale system, F-AMS also needs the specification of a restriction operator, which is a map from fine- to coarse-scale (dimension $N_{coarse} \times N_{fine}$). Due to its algebraic formulation, F-AMS can accommodate multiscale finite volumes (MSFV), multiscale finite elements (MSFE) or even a hybrid multiscale finite elements and volumes restriction (MSMIX). More specifically, the MSMIX employs a FV-based restriction for part of the domain (e.g., fractures or wells), and FE for the rest (e.g., matrix rock). It is important to note that, after any MSFV stage, it is possible to construct a mass-conservative flux field for both matrix and fractures. As such, in multiphase simulations, if iterations for pressure Eq. are stopped before full convergence is achieved, MSFV needs to be employed before solving transport equations. On the other hand, MSFE leads to a symmetric-positive-definite (SPD) coarse system if the fine-scale system matrix is also SPD [@yixuan-ams; @Tene-cams], and is the option used during all numerical experiments presented in this work. Note that MSMIX can be tweaked to achieve the desired compromise between MSFV and MSFE.
In the following sub-sections, first, the formulation of the local basis functions is explained. Then, the F-AMS system, and finally the simulation strategy is described in detail.
Basis function formulations {#sec:basis}
---------------------------
As stated before, F-AMS constructs a non-overlapping partition on the given fine-scale computational domain for both matrix rock and fracture cells, i.e. the primal-coarse grid. Then, by connecting the coarse nodes, the overlapping decomposition of the domain, i.e., dual-coarse grid, is obtained. Following the original description of the MSFV basis functions [@Jenny03; @Jenny06] and its algebraic description [@Zhou-tams; @yixuan-ams; @Tene-cams], local basis functions are calculated for each coarse node $i$, corresponding to each dual block, by respecting the *wirebasket* hierarchy [@wirebasket]. First the pressure in the coarse nodes (also called vertices, shown in orange in Fig. \[fig:dual2d\]) is set to $$\begin{aligned}
\delta_{ij} =\begin{cases}
1, & \text{ if } i=j \\
0, & \text{ if } i \neq j
\end{cases}.
\label{kronecker-basis}\end{aligned}$$ Then the dual blocks in the neighbourhood of node $i$ are resolved, in sequence, as follows: first all the 1D dual blocks (or edges, shown in green in Fig. \[fig:dual2d\]), followed by the 2D (or faces, shown in blue in Fig. \[fig:dual2d\]) and, finally, if applicable, the 3D dual blocks (or interiors, shown in purple in Fig. \[fig:dual3d\]). The fact that each dual block (e.g. edge) neglects the transmissibilities to neighbouring cells belonging to blocks of inferior rank in the wirebasket hierarchy (i.e. faces and interiors), constitutes the *localization assumption* [@yixuan-ams], which ensures that each basis function has a limited support.
The pressure values obtained in the manner described above, for each coarse node $i$, are assembled in column $i$ of the prolongation operator to form basis function $\Phi_i$.
It is important to note that by having independent fine-scale grids for each media, a matrix cell (say from a face block) can be directly connected to fracture cells belonging to dual blocks of any type (vertex, edge or face). This is an important difference from non-fractured media, where the two-point flux approximation (TPFA) stencil on structured grids ensured that any dual block would have connections only to blocks of directly superior or inferior rank in the wirebasket hierarchy (e.g., the external neighbours of a face cell are either edge or interior cells). As such, the multscale localization assumption needs to be extended to account for the connection between the two media. In the F-AMS framework, this leads to definition of basis functions which account for different degrees of coupling between the matrix and its perforating fractures.
In order to provide a compact definition of the various basis functions (matrix, fracture, well), paired with one of the four different coupling strategies considered (Decoupled-AMS, Rock-AMS, Frac-AMS, Coupled-AMS), the following “skeleton” is introduced $$\begin{aligned}
- \nabla \cdot (\lambda^* \cdot \nabla \Phi^{*\bullet})+ \sum_{j \in \text{conn}_{mf}^*} \eta^*_j~ \xi(\Phi^{*\bullet})+ \sum_{j \in \text{perf}_w^*} \beta^*_j~ (\Phi^{*\bullet} - \Phi^{w\bullet}) = 0, \label{skeleton-basis}\end{aligned}$$ which is solved for all basis functions $\forall \Phi^{*\bullet} \in \{\Phi^{mm},\Phi^{mf},\Phi^{mw},\Phi^{fm},\Phi^{ff},\Phi^{fw}\}$, subject to the localization assumption within each domain. Recall that $\bm{\mathcal{P}}^{wm}$ and $\bm{\mathcal{P}}^{wf}$ are zero, while $\bm{\mathcal{P}}^{ww}$ is the identity matrix. The skeleton expression is based on the incompressible pressure equation, since it was found computationally efficient, even when compressibility is involved [@Tene-cams]. In its definition, $\text{perf}_w^*$ represents the set of (matrix or fracture) cells perforated by wells. Moreover, $\text{conn}_{mf}^*$ is the set of all cells with cross-media connectivities from the corresponding (matrix or fracture) domain. Finally, the $\xi(\Phi^{*\bullet})$ function gives the type of matrix-fracture coupling captured by the basis function, and will be specified separately for each strategy, as follows:
1. **Decoupled-AMS:** all basis functions have no-flow boundary conditions between the matrix and fracture domains, i.e., $$\begin{aligned}
\xi(\Phi^{*\bullet}) = 0 \hspace{5mm} \forall \Phi^{*\bullet} \in \{\Phi^{mm},\Phi^{mf},\Phi^{mw},\Phi^{fm},\Phi^{ff},\Phi^{fw}\},\end{aligned}$$ which means that the fracture-matrix coupling term is completely omitted in Eq . The prolongation is solved algebraically, as described in Appendix \[sec:app-decoupledAMS\]. Figure \[fig:decoupledAMS\] illustrates the step-by-step procedure for a fracture and matrix basis function belonging to the 2D reservoir from Fig. \[fig:fineGrid\]. Note that the support of each of the interpolators is restricted to their containing medium. Finally, the Decoupled-AMS approach can be seen as applying the original AMS to separate sub-domains (i.e., matrix and fractures), having them coupled only at the coarse-scale system.
(15,0.5) (5.4,0.2)[Decoupled-AMS]{}
\
\
2. **Frac-AMS:** the fracture basis functions in the fracture domain, $\Phi^{ff}$, are first computed subject to no-flow conditions towards the matrix, the same as in the Decoupled-AMS (Fig. \[fig:decoupledAMS\_frac2\_vertices\] and \[fig:decoupledAMS\_frac2\]), i.e. by substituting $$\begin{aligned}
\xi(\Phi^{ff}) = 0\end{aligned}$$ in Eq. . Then, the obtained values are fixed and used as Dirichlet boundary conditions while computing $\Phi^{mf}$, for which the matrix-fracture transmissibility is taken into account, i.e., $$\begin{aligned}
\xi(\Phi^{mf}) = \Phi^{mf} - \Phi^{ff}.\end{aligned}$$
On the other hand, the matrix basis functions, $\Phi^{mm}$, are solved by setting $\Phi^{fm} = 0$ as Dirichlet boundary condition, i.e., $$\begin{aligned}
\xi(\Phi^{mm}) = \Phi^{mm}.\end{aligned}$$
This procedure is performed algebraically, as described in Appendix \[sec:app-fracAMS\]. Note from Fig. \[fig:fracAMS\] that, after this computation, the fracture functions have non-zero values in the matrix, while the support of the matrix basis functions is restricted to the rock domain.
(15,0.5) (5.6,0.2)[Frac-AMS]{}
\
\
3. **Rock-AMS:** First, $\Phi^{mm}$ is computed with no-flow to the fractures, as with Decoupled-AMS (Figs. \[fig:decoupledAMS\_rock\_vertices\]-\[fig:decoupledAMS\_rock\]), i.e., $$\begin{aligned}
\xi(\Phi^{mm}) = 0.\end{aligned}$$
Then, the values are fixed and used as Dirichlet boundaries while solving $\Phi^{fm}$, for which the fracture-matrix connections are taken into account in Eq. \[skeleton-basis\], i.e., $$\begin{aligned}
\xi(\Phi^{fm}) = \Phi^{fm} - \Phi^{mm}.\end{aligned}$$
For the fracture functions, $\Phi^{mf} = 0$ which is used as Dirichlet condition to compute $\Phi^{ff}$, i.e., $$\begin{aligned}
\xi(\Phi^{ff}) = \Phi^{ff}.\end{aligned}$$
Appendix \[sec:app-rockAMS\] presents the algebraic procedure corresponding to this coupling strategy. Note from Fig. \[fig:rockAMS\] that, in Rock-AMS, the matrix basis functions have non-zero values in the fractures, while the opposite does not hold.
(15,0.5) (5.6,0.2)[Rock-AMS]{}
\
\
4. **Coupled-AMS:** In order to preserve the two-way coupling between fractures and matrix, adjacent dual blocks of the same type are merged (e.g. fracture edges with the matrix edges they perforate), as shown in Fig. \[fig:mergedDuals\]. To clarify, two blocks are considered adjacent if there is a non-zero transmissibility between a cell from one of them and a cell from the other.
(15,0.7) (5.4,0.2)[Coupled-AMS]{}
\
\
On the resulting dual-coarse grid, computation of the basis functions follow the usual wirebasket hierarchy, with full consideration of the coupling, i.e., using $$\begin{aligned}
\xi(\Phi^{mm}) = \Phi^{mm} - \Phi^{fm}, & & \xi(\Phi^{fm}) = \Phi^{fm} - \Phi^{mm}, \nonumber \\
\xi(\Phi^{ff}) = \Phi^{ff} - \Phi^{mf} & \text{\hspace{9mm}and} & \xi(\Phi^{mf}) = \Phi^{mf} - \Phi^{ff}, \end{aligned}$$ for matrix and fracture cells, respectively (see Appendix \[sec:app-coupledAMS\] for the detailed algebraic procedure). Figure \[fig:coupledAMS\] presents an illustration of the Coupled-AMS basis functions. Note that fracture functions have non-zero values in the matrix, and vice versa, the matrix basis functions also have non-zero values inside fractures.
It is worth mentioning that, by construction, all four coupling strategies result in basis functions which form a partition of unity.
The consideration of wells is similar to what was described in the literature for 2D problems [@Patrick-well2], but extended here to 3D problems and integrated within the F-AMS framework. Each well is represented as a single coarse node and a well function is computed locally. The resulting values are assembled in the designated column of the prolongation operator. Note that, even for well perforations in the matrix, the corresponding well functions can have non-zero values inside fractures as well if either Coupled-AMS or Rock-AMS are employed.
As mentioned before, the basis functions have local support. However, in the Coupled-AMS case, the merging of dual blocks (Fig. \[fig:mergedDuals\]) can increase this support substantially in cases with a high density of interconnected (or long) fractures. This, in turn, can lead to a dense prolongation operator, with a severe impact on computational performance. To overcome this, one can impose a limiting criterion on the merging of the dual blocks, for example a maximum number of fine grid cells. Alternatively, or in combination with the previous method, one can discard the non-zeros from $\bm{\mathcal{P}}$ which lie below a specified threshold, and rescale the rows accordingly to preserve the partition of unity. The latter choice is studied in detail in Section \[sec:results\].
It is important to note that F-AMS basis functions are computed at the beginning of a time-dependent simulation, and adaptively updated only if the fine-scale properties change beyond a threshold value [@hadi-aimsfv-jcp]. Next, the F-AMS solution algorithm is described.
The F-AMS solution algorithm {#sec:algorithm}
----------------------------
In addition to the prolongation, in order to obtain the coarse-scale pressure system, the restriction operator $\bm{\mathcal{R}}$ (i.e., map from fine to coarse scale) is now defined. As previously described, F-AMS identifies three distinct types of features in the domain, i.e., matrix, fractures, and wells, therefore, $\bm{\mathcal{R}}$ can be defined in a much more general form than in the previous studies [@Zhou-tams; @yixuan-ams; @Tene-cams].
The first option is to apply a FV-based restriction to all domains, i.e. the MSFV restriction operator, $\bm{\mathcal{R}}^{FV}$, where the entry at row $i$ and column $j$ is $1$ only if the fine-scale cell $j$ (either from the matrix, fractures, or wells) belongs to primal-coarse block $i$. MSFV ensures mass conservation, at the coarse-scale, thus allowing the reconstruction of a fine-scale conservative flux field. However, in previous works [@yixuan-ams; @yixuan-monotonemsfv], it has been found sensitive to the heterogeneity contrast in the domain. Alternatively, one can construct a Galerkin-FE-based restriction operator, as $\bm{\mathcal{R}}^{FE} = \bm{\mathcal{P}}^T$, traditionally called MSFE, which leads to a symmetric-positive-definite (SPD) coarse linear system, if the fine-scale system is also SPD. Finally, one can consider a third option, where some of the features (e.g., matrix) are restricted according to FE, while, for the rest, FV is used. This will be referred to as the MSMIX restriction operator. Note that, for multiphase flow problems, if the pressure system is not solved to machine accuracy, a final iteration with $\bm{\mathcal{R}}^{FV}$ needs to be employed, followed by a conservative flux reconstruction stage, in order to facilitate the solution of the transport equations [@hadi-comp-jcp; @zhou-trans].
Although F-AMS can be used as a single-sweep multiscale solver, where the approximate solution, $p'$, is used with no iterations, previous studies have shown that an iterative procedure is needed for highly-heterogeneous reservoirs [@imsfv-jcp]. To this end, one needs to pair the F-AMS multiscale step with a fine-scale smoother, which ensures error reduction to any desired level. The F-AMS algorithm can now be compiled, as shown in Table \[fams-algorithm\].
------------------------------------------------------------------------------------------------------------------------------------------------------
Repeat the following steps until convergence to the desired accuracy is reached:
1\. **Initialize:** $p^\nu \leftarrow p^{\nu+1}$
2\. **Update linear system entries:** $\bm A^{\nu}$ and $q^\nu$ in
3\. **Update residual:** $r^\nu = q^\nu - \bm A^\nu p^\nu$
4\. **Compute (or adaptively update) $\bm{\mathcal{P}}$**: follow either coupling strategy from
Subsection \[sec:basis\].
5\. **Multiscale Stage:** $\delta p^ {\nu +1/2} = \bm{\mathcal{P}} (\bm{\mathcal{R}} \bm A^{\nu} \bm{\mathcal{P}})^{-1} \bm{\mathcal{R}} \ r^{\nu} $
6\. **Update residual** $r^{\nu+1/2} = r^{\nu} - \bm A^{\nu}\delta p^{\nu+1/2}$
7\. **Smoothing Stage:** $\delta p^ {\nu +2/2} = \bm M^{-1}_S \ r^{\nu+1/2}$
8\. **Update solution:** $p^{\nu+1} = (p^\nu + \delta p^{\nu +1/2} + \delta p^{\nu +2/2})$
------------------------------------------------------------------------------------------------------------------------------------------------------
: F-AMS solution algorithm
\[fams-algorithm\]
The smoothing operator, $\bm M^{-1}_S$, approximates the inverse of the complete fine-scale linear operator, $\bm A^\nu$, via ILU(0) decomposition [@Saad]. Note that the contrast between the matrix and fracture transmissibility values is usually severe, leading to a high condition number in Eq. . In such cases, F-AMS can easily be extended to include another smoothing stage, which employs iterations on the sub-block systems corresponding to each media, i.e. $\bm A^{mm}$ and $\bm A^{ff}$ from . A detailed study of the impact of such a smoothing stage is beyond the scope of this paper, and makes the object of future research.
Next, numerical results are presented in order to study the effect of each component on the performance of the F-AMS algorithm. Then, the scalability of F-AMS is studied, as linear solver, in a CPU benchmark, where the SAMG commercial solver [@SAMG] is used as reference.
Numerical Results {#sec:results}
=================
The aim of this section is to investigate the performance of F-AMS while simulating flow through fractured porous media. First, a 2D reservoir with heterogeneous matrix rock and a relatively complex fracture network is considered. A distance-based graph algorithm is introduced, in order to automate the fracture coarsening process. The convergence behaviour of F-AMS is studied, considering each of the four different coupling strategies introduced in the previous section. Then, the same fracture network is extruded along the Z axis and embedded in a 3D heterogeneous domain, for which simulations are performed considering different coarsening strategies, fracture-matrix conductivity ratios, fracture densities and domain sizes. CPU times are measured in detail for both the setup and solution stages in all test cases, and compared to those obtained using the industrial-grade SAMG solver [@SAMG]. Finally, the same 3D reservoir is used to investigate the effect of heterogeneous fracture conductivities, spanning several orders of magnitude.
During the upcoming simulations, special attention is given to the conductivity contrast between the matrix and the fracture domains. The transmissibility ratio $T_{ratio}$ is introduced as $$\begin{aligned}
T_{ratio} = \frac{\langle T \rangle_{frac}}{\langle T \rangle_{rock}},
\label{tratio}\end{aligned}$$ i.e. the ratio between the average fracture $\langle T \rangle_{frac}$ and matrix $ \langle T \rangle_{rock}$ transmissibility values, respectively.
It is important to note that, in all test cases, F-AMS employs a FE restriction operator. Furthermore, the coarse-scale linear system and the basis functions in each dual block are all solved using a direct solver, based on LU decomposition, from the PETSc package [@petsc].
For some experiments, a detailed breakdown of the CPU time spent in each stage of the F-AMS algorithm is presented. In the legends of the corresponding bar plots, the “Initialization” refers to the time spent on allocation of memory for the various data structures, the “Operators” represents the computation of basis functions and construction of the restriction and prolongation matrices (Step 4 in Table \[fams-algorithm\]). Also, “Fine linsys. constr.” denotes computation of the transmissibility values and fine-scale linear system assembly. In addition, the matrix multiplications resulting in the coarse-scale system are labelled as “Coarse linsys, constr.”, while “Solution” stands for the solution of the coarse system followed by the interpolation (Step 5 in Table \[fams-algorithm\]). Finally, “Smoother” accounts for Step 7.
Distance-based fracture coarsening
----------------------------------
A fracture network can be represented as a graph, in which the fracture lines (plates) are the arcs, while the nodes are the locations at which fractures intersect. This leads to a quasi-unstructured grid, where the complexity mostly revolves around the representation of the intersections. In this work, each intersection is assigned a pressure value, which is explicitly represented in the fine-scale linear system via an equation describing the conservation of flux coming from/going into the fracture control volumes it connects.
As previously described, F-AMS requires primal- and dual-coarse grids in the (quasi-unstructured) fracture domain. In order to hide this complexity from the user, this paper introduces a distance-based algorithm for the automatic coarsening of fracture networks, as described in Table \[fractureCoarsening\]. The network’s graph is traversed in a breadth-first order such that a distance of at least $d_{min}$ cells is guaranteed between each pair of resulting coarse nodes. As such, $d_{min}$ can be seen as a fracture coarsening factor. Note that choosing $d_{min} = \infty$ results in a single coarse node, as shown in Fig. \[fig:1dof\]. Moreover, Fig. \[fig:15dof\] depicts the result of the coarsening algorithm for $d_{min} = 20$ cells on a fairly complex fracture network.
--------------------------------------------------------------------------------------------------------
Repeat the following for each fracture network, $f_i$, which has $N_{cells}$ fine cells:
1\. Choose $d_{min}$, the minimum distance between two fracture coarse nodes.
2\. Initialize three empty queues: $Q_{vertex}, Q_1, Q_2$.
3\. Initialize two vectors of length $N_{cells}$: $level$ (set to $\infty$) and $primal$ (set to $0$).
4\. $N_{primal} \leftarrow 0$
5\. Choose a starting cell from $f_i$ and add it to $Q_{vertex}$.
Repeat until $Q_{vertex}$ is empty:
6\. $vertex \leftarrow$ extract top of $Q_{vertex}$
7\. $N_{primal} \leftarrow N_{primal} + 1$
8\. Create primal block number $N_{primal}$ with $vertex$ as its coarse node.
9\. $primal[vertex] \leftarrow N_{primal}$ and $level[vertex] \leftarrow 0$.
10\. Add $vertex$ to $Q_1$.
For $dist$ from $1$ up to and including $d_{min}$:
Repeat until $Q_1$ is empty:
11\. $cell \leftarrow$ extract top of $Q_1$
For each neighbour of $cell$, $neigh_j$, with $level[neigh_j] > dist$:
12\. remove $neigh_j$ from $Q_{vertex}$, if it is present (i.e. $level[neigh_j]$ = $d_{min}$)
13\. $level[neigh_j] \leftarrow dist$
14\. $primal[neigh_j] \leftarrow N_{primal}$
15\. add $neigh_j$ to $Q_2$.
16\. swap $Q_1$ and $Q_2$.
17\. empty $Q_1$ into $Q_{vertex}$.
At this point, $primal[i]$ gives the index of fine cell $i$’s primal block, while $level[i]$ is
the distance from cell $i$ to the nearest vertex. The fine cells which were not marked
as vertices will form edges on the dual-coarse grid.
--------------------------------------------------------------------------------------------------------
: Distance-based fracture coarsening algorithm.
\[fractureCoarsening\]
F-AMS convergence
-----------------
The fracture network from Fig. \[fig:fracCoarsening\] was embedded into a heterogeneous (patchy) matrix rock with two pressure-constrained wells added on the West and East boundaries. This 2D test case, depicted in Fig. \[fig:2dcase\], was used to study the convergence properties of F-AMS, with the four coupling strategies presented before.
In order to test the accuracy of the basis functions as pressure interpolators, F-AMS was stopped after Step 5 of its first iteration (see Table \[fams-algorithm\]). The solution is shown in Figs. \[fig:1MSiter\_1dof\] and \[fig:1MSiter\_15dof\]. Using a single fracture coarse DOF leads to a poor approximation of the pressure distribution, especially for Decoupled- and Frac-AMS. Figure \[fig:fracAMS\_1dof\] depicts the results of Frac-AMS with 1 fracture coarse DOF. In this setup, F-AMS treats fractures similar to [@hadi-frac-jcp]. Note that, due to the large length scale of the network, having a constant interpolator for the pressure along the network results in an initial solution which lacks a lot of the fine-scale features. In contrast, the Rock-AMS and Coupled-AMS (Figs. \[fig:rockAMS\_1dof\] and \[fig:coupledAMS\_1dof\]) place a lot more emphasis on the matrix basis functions and, since, in this test case, the rock heterogeneity is the main source of approximation error, their results are more accurate. It may seem unexpected that the Coupled-AMS performs slightly worse than Rock-AMS. This can be attributed to the fact that a single fracture DOF is not sufficient to accurately capture the pressure distribution in the large fracture network, especially under the localization assumption. However, when only few additional coarse DOF are added in the fracture domain (as shown in Fig. \[fig:1MSiter\_15dof\]), the situation improves dramatically for Decoupled-AMS, Frac-AMS and Coupled-AMS. Note that Rock-AMS, on the other hand, is insensitive to this change.
For the results in Fig. \[fig:conv\], as well as all subsequent experiments in this paper, F-AMS was iterated until converged to a residual 2-norm of $10^{-6}$. It is clear that Rock-AMS shows a good convergence rate on this particular 2D test case, regardless of the fracture coarsening factor. Also, the other strategies reach a similar behaviour when the fracture network is enriched with only few additional coarse-scale DOF.
\
\
\
Basis function truncation
-------------------------
In order to get an idea of the performance of F-AMS on realistic fractured reservoirs, Fig. \[fig:3dcase\] introduces a 3D scenario, where the fracture network from Fig. \[fig:2dcase\] was extruded and discretized along the Z axis. Two pressure-constrained horizontal wells are placed on opposite edges of the domain boundary. Figure \[fig:pres3d\] shows the fine-scale pressure solution obtained on the heterogeneous (patchy) matrix permeability field shown in Fig. \[fig:perm3d\]. Note that, even though the matrix-fracture conductivity contrast is of only two orders of magnitude, this is enough to make the pressure distribution in the fracture network insensitive to the matrix heterogeneity (see the approximately constant pressure in Fig. \[fig:fracPres3d\]).
\
The procedure described in Table \[fractureCoarsening\] is followed to determine the fracture coarse nodes along the projection of the network on the X-Y plane. Then, the resulting coarse grids are extruded along the Z-axis uniformly, with the vertical distribution of the coarse nodes honouring the user-specified coarsening ratio (see Fig. \[fig:coarseGrids\]).
The increased number of cells in both the matrix ($64 \times 64 \times 64$), as well as the fracture network ($575 \times 64$), compared to the 2D case, can lead to a larger number of non-zeros for the basis functions which take into account the coupling between the two media. This can be particularly severe for Coupled-AMS (Fig. \[fig:coupledAMS\]), since the high density of fractures can cause a large number of dual blocks to be merged. The resulting basis functions have a wider support and can potentially lead to more accurate interpolations, however, the added density to the prolongation operator will also increase the computational effort necessary to construct and solve the coarse-scale linear system (i.e., Step 5 in Table \[fams-algorithm\]).
One can limit the density of $\bm{\mathcal{P}}$ by truncating basis function values below a specified threshold, $\alpha \in [0,1)$. However, in order to preserve partition of unity, the affected rows in $\bm{\mathcal{P}}$ need to be rescaled by dividing the remaining values by the row sum. Figure \[fig:trunc\] shows the CPU time spent by F-AMS on the 3D test case, while varying the value of $\alpha$. Notice that the very restrictive value of $\alpha = 10^{-1}$ leads to an increase in the number of overall iterations, because the smoother needs to compensate for the induced inaccuracy of the basis functions. However, starting with $\alpha=10^{-2}$, the convergence is no longer much affected and the algorithm gains efficiency from the reduced number of FLOPS necessary to perform $\bm{\mathcal{R}} \bm A^\nu \bm{\mathcal{P}}$ and invert the result. The truncation has the biggest impact on the Coupled-AMS strategy, which experiences a speed-up factor of 2, compared to the un-truncated case (last bar in Fig. \[fig:cpu\_trunc\_coupledAMS\], where $\alpha=0$). Figure \[fig:basis\_support\] shows that, for this coupling strategy, when only a single DOF is considered for fracture network, the support of basis functions can be as big as the span of the fracture network. Also, this figure shows that after the truncation stage the locality of the basis function support can be maintained. As conclusion to this study, the subsequent experiments will use a value of $\alpha = 10^{-2}$, regardless of the choice of basis function coupling strategy.
(15,0.4) (2.4,0)[Decoupled-AMS]{} (9.9,0)[Frac-AMS]{}
\
(1.2,1.5) (0.0,0.3) (0.9,-0.1)[0]{} (0.9,1.7)[8]{} (0.7,3.3)[16]{}
(0.5,1.5) (0.2,-0.1)[0]{} (0.2,1.7)[8]{} (0.0,3.3)[16]{}
\
(15,0.7) (1.3,0.0) (2.0,0.0) (2.8,0.0) (3.55,0.0) (4.3,0.0) (5.05,0.0) (6.1,0.0) (8.2,0.0) (8.9,0.0) (9.7,0.0) (10.45,0.0) (11.2,0.0) (11.95,0.0) (13.0,0.0)
\
(15,0.2) (4.0,0.0)[$\alpha$]{} (11.0,0.0)[$\alpha$]{}
\
(15,0.7) (3.0,0)[Rock-AMS]{} (9.5,0)[Coupled-AMS]{}
\
(1.2,1.5) (0.0,0.3) (0.9,-0.1)[0]{} (0.9,1.7)[8]{} (0.7,3.3)[16]{}
(0.5,1.5) (0.2,-0.1)[0]{} (0.2,1.7)[8]{} (0.0,3.3)[16]{}
\
(15,0.7) (1.3,0.0) (2.0,0.0) (2.8,0.0) (3.55,0.0) (4.3,0.0) (5.05,0.0) (6.1,0.0) (8.2,0.0) (8.9,0.0) (9.7,0.0) (10.45,0.0) (11.2,0.0) (11.95,0.0) (13.0,0.0)
\
(15,0.2) (4.0,0.0)[$\alpha$]{} (11.0,0.0)[$\alpha$]{}
\
Sensitivity to the coarsening factor {#sec:coarsening_ratio}
------------------------------------
(15,1.5) (5.4,1.3)[Rock coarsening]{} (1.9,0.65)[$16 \times 16 \times 16$ $8 \times 8 \times 8$ $4 \times 4 \times 4$]{} (2.15,0.1)[(64 DOF) (512 DOF) (4096 DOF)]{}
\
(1.3,1.5) (0,1.0) (0.9,-0.1)[0]{} (0.75,1.1)[25]{} (0.75,2.3)[50]{} (0.75,3.5)[75]{} (0.6,4.7)[100]{}
(1.3,1.5) (0.2,1.8) (0.65,1.6)
\
(1.3,1.5) (0,1.0) (0.9,-0.1)[0]{} (0.75,1.1)[25]{} (0.75,2.3)[50]{} (0.75,3.5)[75]{} (0.6,4.7)[100]{}
(1.3,3.0) (0.2,2.0) (0.65,1.5) (1.3,0.2)
\
(1.3,1.5) (0,1.0) (0.9,-0.1)[0]{} (0.75,1.1)[25]{} (0.75,2.3)[50]{} (0.75,3.5)[75]{} (0.6,4.7)[100]{}
(1.3,1.5) (0.2,2.0) (0.65,1.4)
\
(15,2.3) (0,0.0) (0.8,0.0) (1.7,0.0) (2.5,0.0) (3.9,0.0) (4.7,0.0) (5.6,0.0) (6.4,0.0) (7.8,0.0) (8.6,0.0) (9.4,0.0) (10.3,0.0)
\
The sensitivity of F-AMS to the number of coarse DOF in the fracture network, as well as the matrix coarsening ratio is studied for the 3D test case shown in Fig. \[fig:3dcase\]. The coarsening factor is defined as the average number of fine cells contained in one (matrix or fracture) primal-coarse block, along each axis. Recall that, along the fracture length, this is given by the $d_{min}$ parameter in Table \[fractureCoarsening\]. The experiment designed for this purpose is focused on “isotropic” coarsening factors, mainly due to the point-wise nature of ILU(0), which was chosen as global smoothing stage for the implementation of F-AMS used here (Table \[fams-algorithm\]).
Figure \[fig:dof\] shows the F-AMS CPU times obtained with three different coarsening factors for the matrix, as well as the fracture network. From this figure, the Coupled-AMS is found to automatically adapt itself to the coarsening ratio. Its convergence rate is surpassed only in cases where there is a large discrepancy between the rock and fracture coarsening ratios. However, this comes with the additional computational cost of having basis functions with wider support. Still, due to the truncation factor $\alpha = 10^{-2}$, the construction and solution of coarse system for Coupled-AMS is not significantly higher than that of the alternative strategies. In addition, in each row, the optimum simulation results are obtained when fracture and matrix coarsening ratios are similar.
Based on this study and unless otherwise stated, the coarsening ratio of $8$ in each direction for both matrix and fracture media is employed in the experiments presented in the following subsections. Note that this option leads to more efficient coarse-scale systems than the alternative option of using coarsening ratios of $4$. In addition, the number of linear iterations can be significantly reduced when F-AMS is employed as preconditioner for GMRES [@Saad].
Sensitivity to the transmissibility ratio
-----------------------------------------
(15,0.2) (2.6,0)[Decoupled-AMS]{} (9.9,0)[Frac-AMS]{}
\
(1.2,1.5) (0.0,0.3) (0.9,-0.1)[0]{} (0.9,1.7)[6]{} (0.7,3.3)[12]{}
(0.5,1.5) (0.2,-0.1)[0]{} (0.2,1.7)[6]{} (0.0,3.3)[12]{}
\
(15,0.2) (1.7,0.0)[$10^0$]{} (2.6,0.0)[$10^1$]{} (3.45,0.0)[$10^2$]{} (4.3,0.0)[$10^3$]{} (5.15,0.0)[$10^4$]{} (6.05,0.0)[$10^5$]{} (8.6,0.0)[$10^0$]{} (9.5,0.0)[$10^1$]{} (10.35,0.0)[$10^2$]{} (11.2,0.0)[$10^3$]{} (12.05,0.0)[$10^4$]{} (12.95,0.0)[$10^5$]{}
\
(15,0.2) (3.8,.0)[$T_{ratio}$]{} (10.7,.0)[$T_{ratio}$]{}
\
(15,0.9) (3.2,0)[Rock-AMS]{} (9.4,0)[Coupled-AMS]{}
\
(1.2,1.5) (0.0,0.3) (0.9,-0.1)[0]{} (0.9,1.7)[6]{} (0.7,3.3)[12]{}
(0.5,1.5) (0.2,-0.1)[0]{} (0.2,1.7)[6]{} (0.0,3.3)[12]{}
\
(15,0.2) (1.7,0.0)[$10^0$]{} (2.6,0.0)[$10^1$]{} (3.45,0.0)[$10^2$]{} (4.3,0.0)[$10^3$]{} (5.15,0.0)[$10^4$]{} (6.05,0.0)[$10^5$]{} (8.6,0.0)[$10^0$]{} (9.5,0.0)[$10^1$]{} (10.35,0.0)[$10^2$]{} (11.2,0.0)[$10^3$]{} (12.05,0.0)[$10^4$]{} (12.95,0.0)[$10^5$]{}
\
(15,0.2) (3.8,.0)[$T_{ratio}$]{} (10.7,.0)[$T_{ratio}$]{}
\
![Convergence history of F-AMS and SAMG on the 3D test case with a fracture/matrix transmissibility ratio of $T_{ratio}=10^6$. Notice that neither method can converge when iterated in a Richardson’s loop. Instead, as preconditioners to GMRES, both methods converge after a few iterations.[]{data-label="fig:conv_perm1e6"}](Fig/transm_ratio/conv_perm1e6.pdf){width="80.00000%" height="40.00000%"}
The next set of experiments aims to investigate the sensitivity of F-AMS to the conductivity contrast between the matrix and fractures. The transmissibility ratio $T_{ratio}$, as defined in Eq. , is varied over several orders of magnitude (Fig. \[fig:transm\_ratio\]) while measuring CPU times and number of linear iterations performed by the Richardson loop. A coarsening factor of $8$ was chosen for both media, based on the results from the previous subsection. As the network becomes more conductive, the influence of the matrix heterogeneity on the fracture pressure decreases (see Figs. \[fig:fracPres3d\] and \[fig:pres3d\]). As such, Rock-AMS exhibits a degradation in performance for higher $T_{ratio}$ (Fig. \[fig:cpu\_transm\_rockAMS\]), while the reverse is true for Frac-AMS (Fig. \[fig:cpu\_transm\_fracAMS\]). On the other hand, by automatically adapting to the change, the Coupled-AMS strategy remains relatively insensitive to $T_{ratio}$, as shown in Fig. \[fig:cpu\_transm\_coupledAMS\]. Finally, Decoupled-AMS requires the most number of iterations when the fracture and matrix transmissibility values are close (Fig. \[fig:cpu\_transm\_decoupledAMS\]), since in this case the two-way coupling between the media is the most pronounced.
Note that the solver could not converge to the chosen tolerance, of $10^{-6}$ residual norm, when F-AMS was iterated in a Richardson’s loop for $T_{ratio} \ge 10^6$ and the same holds for SAMG [@SAMG]. However, as shown in Fig. \[fig:conv\_perm1e6\], both methods converge successfully when employed as preconditioners for GMRES [@Saad].
CPU benchmark study
-------------------
This final subsection presents the results of a benchmark study between F-AMS and SAMG [@SAMG] on 3D heterogeneous fractured reservoirs. Both methods are employed as preconditioners to GMRES [@Saad] and iterated until converged with a residual 2-norm below $10^{-6}$.
Unlike the Richardson loop, similar performance was observed for all experiments when F-AMS is used as preconditioner to GMRES, regardless of which coupling strategy was chosen. Therefore, the presentation of the results is restricted to Decoupled-AMS, for conciseness. Note that the Coupled-AMS strategy provides a more general framework, however, with a more complex implementation.
At each GMRES iteration, SAMG employs a single V-cycle. It is important to note that SAMG is a commercial black-box package. Thus, it is not possible to measure its CPU breakdown as accurately as for F-AMS. Instead, the time spent on its first iteration is considered as “Initialization”, while subsequent iterations were labelled as “Solution”. Finally, for both SAMG and F-AMS, the setup and construction of the operators are performed only once, at the beginning of the iteration procedure.
This study is aimed only to demonstrate the scalability of the F-AMS method. Note that a unique advantage of F-AMS over SAMG is that a fine-scale mass conservative flux field can be reconstructed after any iteration stage, once the coarse-scale system with $\bm{\mathcal{R}}^{FV}$ restriction operator is solved.
### Transmissibility contrast
The test case from Fig. \[fig:3dcase\] is used, with different values of fracture-matrix transmissibility contrasts, i.e., $T_{ratio}$ in Eq. . As can be seen in Fig. \[fig:bench\], the number of iterations and the CPU time for both F-AMS and SAMG is insensitive to the contrast. This is a significant achievement for F-AMS, compared to [@hadi-frac-jcp].
(15,0.8)(0.2,0.2)[$T_{ratio}$]{}(2.4,0.2)[$10^1$]{}(5.7,0.2)[$10^2$]{}(9.0,0.2)[$10^4$]{}(12.2,0.2)[$10^8$]{}
\
(1.0,3.5) (0.0,0.3) (0.75,-0.1)[0]{} (0.7,1.7)[2]{} (0.7,3.3)[4]{}
(0.6,3.5) (0.3,-0.1)[0]{} (0.3,1.7)[2]{} (0.3,3.3)[4]{}
(0.6,3.5) (0.3,-0.1)[0]{} (0.3,1.7)[2]{} (0.3,3.3)[4]{}
(0.6,3.5) (0.3,-0.1)[0]{} (0.3,1.7)[2]{} (0.3,3.3)[4]{}
\
(15,1.0)(1.2,0.0)(2.2,0.0)(4.4,0.0)(5.4,0.0)(7.75,0.0)(8.75,0.0)(11.00,0.0)(12.00,0.0)
\
### Fracture density
What follows is a study of the scalability of F-AMS when faced with a dynamic fracture network, where the number of fracture plates is increased step by step. The 3D fracture map shown in Fig. \[fig:fracs3d\] is considered, where the network is now created through 4 phases, as presented in Fig. \[fig:slicedFracs\]. Note that, as new fractures are added, not only the number of DOF increases, but also the pressure variation along the network can increase. The detailed description of the CPU times obtained using F-AMS in these four cases are depicted in Fig. \[fig:nfracs\]. It is clear that, by maintaining the prescribed fracture coarsening factor of $8 \times 8$, F-AMS maintains virtually the same convergence rate. The slight increase in CPU time is mainly due to computation of extra fracture basis functions, as well as the construction and solution of a slightly larger coarse-scale linear system.
In consequence, by having multiple coarse-scale DOF in each fracture network, F-AMS can automatically scale with fracture length and density. This is in contrast to [@hadi-frac-jcp], where the use of a single fracture basis function would lead to a drastic deterioration of the multiscale convergence for test cases containing fracture networks with large length scales.
(15,0.8)(0.2,0.2)[\#fracs]{}(2.4,0.2)[$27$]{}(5.7,0.2)[$76$]{}(8.9,0.2)[$96$]{}(12.1,0.2)[$127$]{}
\
(1.0,3.5) (0.0,0.3) (0.75,-0.1)[0]{} (0.7,1.7)[2]{} (0.7,3.3)[4]{}
(0.6,3.5) (0.3,-0.1)[0]{} (0.3,1.7)[2]{} (0.3,3.3)[4]{}
(0.6,3.5) (0.3,-0.1)[0]{} (0.3,1.7)[2]{} (0.3,3.3)[4]{}
(0.6,3.5) (0.3,-0.1)[0]{} (0.3,1.7)[2]{} (0.3,3.3)[4]{}
\
(15,1.0)(1.2,0.0)(2.2,0.0)(4.4,0.0)(5.4,0.0)(7.75,0.0)(8.75,0.0)(11.00,0.0)(12.00,0.0)
\
### Domain scale {#sec:domain_scale}
The scalability of F-AMS, benchmarked with SAMG, is investigated for heterogeneous (patchy) reservoir of increasing size. To this end, both the matrix and fracture fine-scale grid resolution is varied from $32^3$ matrix and $320\times32$ fractures (smallest) up to $256^3$ matrix and $2117\times256$ fracture cells (see Fig. \[fig:scaledFracs\]). The transmissibility ratio between the two media is $T_{ratio} = 10^2$. Figure \[fig:scale\] shows the obtained CPU times. F-AMS and SAMG both maintain their convergence rates and experience a similar level of scalability, in terms of CPU time, i.e., they grow linearly with the problem size. During these experiments, and as previously studied in Subsection \[sec:coarsening\_ratio\], F-AMS was found very sensitive to the coarsening strategy used. The reported results use the optimum coarsening ratio found during repeated experiments (some of which are detailed in Table \[tab:coarsening\_ratio\]). Note that SAMG uses adaptive coarsening at each coarse level in its V-cycles.
(15,0.8)(0.0,0.2)[Scale]{}(2.4,0.2)[$32^3$]{}(5.6,0.2)[$64^3$]{}(8.9,0.2)[$128^3$]{}(12.4,0.2)[$256^3$]{}
\
(1.2,1.5) (0.0,0.3) (0.95,-0.1)[0]{} (0.55,1.7)[0.5]{} (0.9,3.3)[1]{}
(0.6,1.5) (0.3,-0.1)[0]{} (0.3,1.7)[2]{} (0.3,3.3)[4]{}
(0.7,1.5) (0.4,-0.1)[0]{} (0.2,1.7)[12]{} (0.2,3.3)[24]{}
(0.8,1.5) (0.5,-0.1)[0]{} (0.3,1.7)[90]{} (0.1,3.3)[180]{}
\
(15,1.0) (1.2,0.0) (2.2,0.0) (4.4,0.0) (5.4,0.0) (7.75,0.0) (8.75,0.0) (11.25,0.0) (12.25,0.0)
\
### Heterogeneous fractures
Finally, the sensitivity of the F-AMS method (benchmarked with SAMG) to strongly heterogeneous fracture properties is investigated. The permeability of of the $127$ fracture plates from Fig. \[fig:fracs3d\] is randomly perturbed to a span $6$ orders of magnitude. Figure \[fig:hetero\] shows that, if an appropriate coarsening ratio is chosen –in this case $6 \times 6 \times 8$ in the matrix and $4 \times 8$ in the fractures–, then the F-AMS and SAMG performances are comparable. In addition to the coarsening ratio, multiscale methods are also sensitive to the heterogeneity contrasts (here, in both fractures and matrix). Improvements can be achieved by adapting the coarse grid geometry to follow the fracture and matrix conductivity distribution, or by enriching the prolongation operator with additional basis functions [@Yalchin-enriched1; @Yalchin-enriched2; @Davide14]. These are subjects of future studies.
![F-AMS (Decoupled-AMS) performance compared with SAMG on a test case with heterogeneous fracture permeability. The logarithm of the fracture permeability is shown on the left. The number of iterations to reach $10^{-6}$ residual 2-norm is given on top of each bar. For these experiments, both methods were employed as preconditioners for GMRES.[]{data-label="fig:hetero"}](Fig/hetero/perm_hetero.pdf "fig:"){height="80.00000%" width="80.00000%"}\[fig:perm\_hetero\]\
(5,0.5) (3.2,0.2)[$log_{10}(k^f)$]{}
(1.0,3.5) (0.0,0.3) (0.75,-0.1)[0]{} (0.7,1.7)[2]{} (0.7,3.3)[4]{}
\[fig:cpu\_hetero\]\
(5,1.2) (1.1,0.0) (2.1,0.0)
\
The results of the experiments presented in this section show that the performance and scalability of F-AMS is comparable to that of SAMG. As such, even in its current proof of concept implementation, F-AMS is found a promising multiscale method for naturally fractured porous media. Note that, for realistic multiphase test cases, simulations can be further accelerated by employing only few iterations of F-AMS, followed by a flux reconstruction stage, leading to efficient approximate solutions [@hadi-erest-spej].
Conclusions {#sec:conclusions}
===========
In this paper, a novel general multiscale framework, F-AMS, was devised for efficient and accurate simulation of flow through heterogeneous porous media with embedded fractures of various length scales. For the first time in the multiscale community, the possibility to prescribe an arbitrary coarse grid in each fracture network was presented. Then, for each coarse node (from both matrix and fractures), a locally-supported basis function was defined, by considering one of the four cross-media coupling strategies (Decoupled-AMS, Frac-AMS, Rock-AMS or Coupled-AMS). All of these flexibilities allow the user to tweak the trade-off between the computational budget of the setup stage and the convergence rate.
Aligned with the EDFM approach of having independent grids for fracture and matrix [@SeongFrac2], this work also introduced a distance-based automatic coarsening algorithm for the fracture domain. This allows the user to specify the desired (uniform) coarsening factor for the fracture domain, in similar fashion to the matrix. In addition, the effect of truncating small non-zeros from the prolongation operator was studied, in order to maintain efficiency, especially for the Coupled-AMS strategy. For all test cases considered, the truncation value of $\alpha = 10^{-2}$ was found optimum.
The numerical results illustrate that F-AMS (similar as SAMG) is insensitive to the fracture-matrix conductivity contrast, and - importantly - shares the same scalability with respect to the fracture density, domain scale and heterogeneous properties. However, the performance of F-AMS was found to dramatically degrade if a sub-optimal coarsening strategy is chosen for challenging test cases. The method can be further extended to address this, e.g. by considering enhanced coarsening strategies, different choices for the smoother or employment of enriched prolongation operators [@Yalchin-enriched1; @Yalchin-enriched2].
Another important finding of this paper was that all basis function coupling strategies perform similar when F-AMS is used as preconditioner to GMRES. This recommends the Decoupled-AMS approach for commercial reservoir simulation, due to its convenient implementation, an attractive feature for real-field applications.
In summary, it is concluded that F-AMS is an important multiscale development for flow in heterogeneous media with embedded fractures. It was shown that only few fracture coarse nodes are required to deliver good approximate pressure solutions, at the original fine-scale resolution. Future developments of F-AMS will include consideration of complex physics such as stress-dependent fracture networks and also capillarity and gravity effects for multiphase flow simulations.
Acknowledgements {#sec:acknowledgements .unnumbered}
================
Financial support of PI/ADNOC is acknowledged. Also, many thanks are due towards the members of the DARSim research group of TU Delft, for the useful discussions during the development of F-AMS method.
Algebraic Formulation of the F-AMS prolongation operators
=========================================================
Consider the fine-scale system for matrix and fractures, $\bm{A} p = q$, i.e., $$\begin{aligned}
\label{A-dis-mf}
\underbrace{\left[\begin{array}{cc} \bm A^{mm} & \bm A^{mf} \\ \bm A^{fm} & \bm A^{ff} \end{array}\right]}_{\bm A}
\underbrace{\left[\begin{array}{c}p^m \\p^f\end{array}\right]}_{p}
=
\underbrace{\left[\begin{array}{c}q^m \\q^f \end{array}\right]}_{q}.\end{aligned}$$ The permutation operator $\bm\wp$ containing matrix and fracture permutation block operators $\bm\wp^{m}$ and $\bm\wp^{f}$, respectively, $$\begin{aligned}
{\bm \wp} \equiv {\left[\begin{array}{cc}\bm\wp^{m} & 0 \\0 & \bm\wp^{f} \end{array}\right]},\end{aligned}$$ is defined such that it reorders the linear system based on the wirebasket ordering [@NordbottenBjorstad07; @wirebasket; @yixuan-ams] of Internal (I), Face (F), Edge (E) and Vertex (V) for both matrix (superscript $^m$) and fracture (superscript $^f$) unknowns, i.e., $$\begin{aligned}
\label{A-reordered}
\underbrace{\left[\begin{array}{cccc|ccc}
\bm A^{{I^m}{I^m}} & \bm A^{{I^m}{F^m}} & 0 & 0 &
\bm A^{{I^m}{F^f}} & \bm A^{{I^m}{E^f}} & \bm A^{{I^m}{V^f}} \\
\bm A^{{F^m}{I^m}} & \bm A^{{F^m}{F^m}} & \bm A^{{F^m}{E^m}} & 0&
\bm A^{{F^m}{F^f}} & \bm A^{{F^m}{E^f}} & \bm A^{{F^m}{V^f}} \\
0 & \bm A^{{E^m}{F^m}} & \bm A^{{E^m}{E^m}} & \bm A^{{E^m}{V^m}} &
\bm A^{{E^m}{F^f}} & \bm A^{{E^m}{E^f}} & \bm A^{{E^m}{V^f}} \\
0 & 0 & \bm A^{{V^m}{E^m}} & \bm A^{{V^m}{V^m}} &
\bm A^{{V^m}{F^f}} & \bm A^{{V^m}{E^f}} & \bm A^{{V^m}{V^f}} \\
\hline
\bm A^{{F^f}{I^m}} & \bm A^{{F^f}{F^m}} & \bm A^{{F^f}{E^m}} & \bm A^{{F^f}{V^m}} &
\bm A^{{F^f}{F^f}} & \bm A^{{F^f}{E^f}} & 0 \\
\bm A^{{E^f}{I^m}} & \bm A^{{E^f}{F^m}} & \bm A^{{E^f}{E^m}} & \bm A^{{E^f}{V^m}} &
\bm A^{{E^f}{F^f}} & \bm A^{{E^f}{E^f}} & \bm A^{{E^f}{V^f}} \\
\bm A^{{V^f}{I^m}} & \bm A^{{V^f}{F^m}} & \bm A^{{V^f}{E^m}} & \bm A^{{V^f}{V^m}} &
0 & \bm A^{{V^f}{E^f}} & \bm A^{{V^f}{V^f}}
\end{array}\right]}_{\bm{\wp A \wp^T}}
\underbrace{\left[\begin{array}{c}
p^{I^m}\\
p^{F^m}\\
p^{E^m}\\
p^{V^m}\\
\hline
p^{F^f}\\
p^{E^f}\\
p^{V^f}\\
\end{array}\right]}_{\bm\wp p}
=
\underbrace{\left[\begin{array}{c}
q^{I^m}\\
q^{F^m}\\
q^{E^m}\\
q^{V^m}\\
\hline
q^{F^f}\\
q^{E^f}\\
q^{V^f}\\
\end{array}\right]}_{\bm\wp q}.\end{aligned}$$ Note that fractures have only Face, Edge, and Vertex cells, since they are represented in a lower-dimensional space than the matrix. Also, according to the two-point flux approximation (TPFA) stencil for structured grids, $\bm A^{{I^m}{E^m}}$, $\bm A^{{E^m}{I^m}}$, $\bm A^{{I^m}{V^m}}$, $\bm A^{{V^m}{I^m}}$, $\bm A^{{V^m}{F^m}}$, $\bm A^{{F^m}{V^m}}$, $\bm A^{{F^f}{V^f}}$, $\bm A^{{V^f}{F^f}}$ are zero. More importantly, for media with embedded fractures, the coupling off-diagonal blocks $\bm A^{mf}$ and $\bm A^{fm}$ are full, i.e., each matrix cell may overlap with fracture cells of any type (F, E, or V). This is the main reason behind the consideration of the four types of basis functions, each with a different level of matrix-fracture coupling, as previously discussed in this paper.
The algebraic construction of the prolongation operator for each strategy is described next.
Decoupled-AMS {#sec:app-decoupledAMS}
-------------
In the Decoupled-AMS prolongation operator, the matrix-fracture coupling terms are completely neglected. To this end, all off-diagonal block matrix entries (belonging to $\bm A^{mf}$ and $\bm A^{fm}$) are set to zero. In addition, similar to the AMS [@yixuan-ams] and C-AMS [@Tene-cams] methods, the linear system is further simplified to account for the localization boundary condition within each medium (by neglecting connectivity between each cell and its lower-ranked neighbours in the wirebasket hierarchy). This leads to the following approximate linear system: $$\begin{aligned}
\label{DecoupledAMS-approxSys}
\underbrace{\left[\begin{array}{cccc|ccc}
\underline{\bm A}^{{I^m}{I^m}} & \bm A^{{I^m}{F^m}} & 0 & 0 &
0 & 0 & 0 \\
0 & \overline{\underline{\bm{A}}}^{{F^m}{F^m}} & \bm A^{{F^m}{E^m}} & 0&
0 & 0 & 0 \\
0 & 0 & \overline{\underline{\bm{A}}}^{{E^m}{E^m}} & \bm A^{{E^m}{V^m}} &
0 & 0 & 0 \\
0 & 0 & 0 & \breve{\bm{A}}^{mm} &
0 & 0 & \breve{\bm{A}}^{mf} \\
\hline
0 & 0 & 0 & 0 &
\underline{\bm A}^{{F^f}{F^f}} & \bm A^{{F^f}{E^f}} & 0 \\
0 & 0 & 0 & 0 &
0 & \overline{\underline{\bm{A}}}^{{E^f}{E^f}} & \bm A^{{E^f}{V^f}} \\
0 & 0 & 0 & \breve{\bm{A}}^{fm} &
0 & 0 & \breve{\bm{A}}^{ff}
\end{array}\right]}_{\bm{\wp A' \wp^T}} \underbrace{\left[\begin{array}{c}
p'^{I^m}\\
p'^{F^m}\\
p'^{E^m}\\
p'^{V^m}\\
\hline
p'^{F^f}\\
p'^{E^f}\\
p'^{V^f}\\
\end{array}\right]}_{\bm\wp p'}
=
\underbrace{\left[\begin{array}{c}
0\\
0\\
0\\
\breve{q}^m\\
\hline
0\\
0\\
\breve{q}^f\\
\end{array}\right]}_{\bm\wp q'}.\end{aligned}$$ Here, the diagonal blocks marked as $\overline{\bm{A}}$ indicate that the matrix-matrix and fracture-fracture transmissibilities, neglected due to this localization assumption ($\bm A^{{F^m}{I^m}}$, $\bm A^{{E^m}{F^m}}$, $\bm A^{{V^m}{E^m}}$ and $\bm A^{{E^f}{F^f}}$, $\bm A^{{V^f}{E^f}}$, respectively) have also been removed from the diagonal term. At the same time, the notation $\underline{\bm A}$ indicates diagonal blocks where the neglected matrix-fracture transmissibilities have been removed from the diagonal term. Finally, $$\begin{aligned}
\breve{\bm{A}} \equiv \left[\begin{array}{cc}
\breve{\bm{A}}^{mm} & \breve{\bm{A}}^{mf} \\
\breve{\bm{A}}^{fm} & \breve{\bm{A}}^{ff} \end{array}\right] = \bm{(\mathcal{R} A \mathcal{P})}, & &\breve{p} \equiv \left[\begin{array}{c} p'^{V^m} \\
p'^{V^f} \end{array}\right], & &\breve q \equiv \left[\begin{array}{c}
\breve{q}^m \\
\breve{q}^f
\end{array}\right] = \bm{\mathcal{R}} q, $$ are the components of the coarse-scale system.
After solving for the coarse-scale pressures, $\breve p = {\breve{\bm{A}}}^{-1} \breve q$, the approximate system can be inverted algebraically, due to its upper-triangular structure. Consequently, the prolongation operator, which satisfies $p' = \bm{\mathcal{P}} \breve p$, reads $$\begin{aligned}
\label{DecoupledAMS-prlg}
\bm{\mathcal{P}} = {\bm \wp^T}{\left[\begin{array}{c|c}
- {\underline{\bm{A}}^{{I^m}{I^m}}}^{-1} \bm{A}^{{I^m}{F^m}} \bm{\mathcal{P}}^{F^m V^m} & 0\\
-{\overline{\underline{\bm{A}}}^{{F^m}{F^m}}}^{-1} \bm{A}^{{F^m}{E^m}} \bm{\mathcal{P}}^{E^m V^m} & 0\\
-{\overline{\underline{\bm{A}}}^{{E^m}{E^m}}}^{-1} \bm{A}^{{E^m}{V^m}} & 0\\
\bm{I}^{V^m V^m}& 0\\
\hline
0 &
-{\underline{\bm{A}}^{{F^f}{F^f}}}^{-1} \bm{A}^{{F^f}{E^f}} \bm{\mathcal{P}}^{E^f V^f}\\
0 &
-{\overline{\underline{\bm{A}}}^{{E^f}{E^f}}}^{-1} \bm{A}^{{E^f}{V^f}} \\
0 &
\bm{I}^{V^f V^f}
\end{array}\right]},\end{aligned}$$ where $\bm{I}$ is the identity matrix and the transpose operator ${\bm \wp^T}$ back-transforms the wirebasket ordering into the natural ordering. Also, $\bm{\mathcal{P}}^{E^m V^m}$, $\bm{\mathcal{P}}^{F^m V^m}$ and $\bm{\mathcal{P}}^{E^f V^f}$ are sub-blocks of the prolongation with the corresponding rows and columns given in the superscripts. For example, $\bm{\mathcal{P}}^{E^m V^m} = -{\overline{\underline{\bm{A}}}^{{E^m}{E^m}}}^{-1} \bm{A}^{{E^m}{V^m}}$ and $ \bm{\mathcal{P}}^{F^m V^m} = -{\overline{\underline{\bm{A}}}^{{F^m}{F^m}}}^{-1} \bm{A}^{{F^m}{E^m}} \bm{\mathcal{P}}^{E^m V^m} $. Note that, once computed, the higher-rank sub-blocks of $\bm{\mathcal{P}}$ become boundary conditions for the values of basis functions in lower-rank cells, in accordance to the localization assumption (e.g. the values obtained for matrix edges, $\bm{\mathcal{P}}^{E_m V_m}$, are used to compute the prolongation in adjacent faces, $\bm{\mathcal{P}}^{F^m V^m}$).
Frac-AMS {#sec:app-fracAMS}
--------
The Frac-AMS approach considers the effect of the $\bm A^{mf}$ transmissibilities when computing basis functions. This leads to the following approximate system operator $$\begin{aligned}
\label{FracAMS-approxSys}
\bm{\wp A' \wp^T} = {\left[\begin{array}{cccc|ccc}
\bm{A}^{{I^m}{I^m}} & \bm{A}^{{I^m}{F^m}} & 0 & 0 &
\bm{A}^{{I^m}{F^f}} & \bm{A}^{{I^m}{E^f}} & \bm{A}^{{I^m}{V^f}} \\
0 & \overline{\bm{A}}^{{F^m}{F^m}} & \bm{A}^{{F^m}{E^m}} & 0&
\bm{A}^{{F^m}{F^f}} & \bm{A}^{{F^m}{E^f}} & \bm{A}^{{F^m}{V^f}} \\
0 & 0 & \overline{\bm{A}}^{{E^m}{E^m}} & \bm{A}^{{E^m}{V^m}} &
\bm{A}^{{E^m}{F^f}} & \bm{A}^{{E^m}{E^f}} & \bm{A}^{{E^m}{V^f}} \\
0 & 0 & 0 & \breve{\bm{A}}^{mm} &
0 & 0 & \breve{\bm{A}}^{mf} \\
\hline
0 & 0 & 0 & 0 &
\underline{\bm{A}}^{{F^f}{F^f}} & \bm{A}^{{F^f}{E^f}} & 0 \\
0 & 0 & 0 & 0 &
0 & \overline{\underline{\bm{A}}}^{{E^f}{E^f}} & \bm{A}^{{E^f}{V^f}} \\
0 & 0 & 0 & \breve{\bm{A}}^{fm} &
0 & 0 & \breve{\bm{A}}^{ff}
\end{array}\right]},\end{aligned}$$ where $\bm A^{{F^m}{I^m}}$, $\bm A^{{E^m}{F^m}}$, $\bm A^{{V^m}{E^m}}$, $\bm A^{{V^m}{F^f}}$ and $\bm A^{{V^m}{E^f}}$ are set to zero due to localization boundary condition corresponding to Frac-AMS coupling for matrix, while, at the same time, the $\bm A^{{E^f}{F^f}}$, $\bm A^{{V^f}{E^f}}$ are zero in the fracture equations.
The Frac-AMS prolongation operator reads $$\begin{aligned}
\label{FracAMS-prlg}
\bm{\mathcal{P}} = {\bm \wp^T}{\left[\begin{array}{c|c}
- {\bm{A}^{{I^m}{I^m}}}^{-1} \bm{A}^{{I^m}{F^m}} \bm{\mathcal{P}}^{F^m V^m} &
- {{\bm{A}}^{{I^m}{I^m}}}^{-1} (\bm{A}^{{I^m}{F^m}} \bm{\mathcal{P}}^{F^m V^f} + \bm{A}^{{I^m}{f}} \bm{\mathcal{P}}^{f V^f}) \\
-{\overline{\bm{A}}^{{F^m}{F^m}}}^{-1} \bm{A}^{{F^m}{E^m}} \bm{\mathcal{P}}^{E^m V^m} &
- {\overline{\bm{A}}^{{F^m}{F^m}}}^{-1} (\bm{A}^{{F^m}{E^m}} \bm{\mathcal{P}}^{E^m V^f} + \bm{A}^{{F^m}{f}} \bm{\mathcal{P}}^{f V^f}) \\
-{\overline{\bm{A}}^{{E^m}{E^m}}}^{-1} \bm{A}^{{E^m}{V^m}} &
- {\overline{\bm{A}}^{{E^m}{E^m}}}^{-1} \bm{A}^{{E^m}{f}} \bm{\mathcal{P}}^{f V^f} \\
\bm{I}^{V^m V^m}&
0\\
\hline
0 &
- {\underline{\bm{A}}^{{F^f}{F^f}}}^{-1} \bm{A}^{{F^f}{E^f}} \bm{\mathcal{P}}^{E^f V^f}\\
0 &
-{\overline{\underline{\bm{A}}}^{{E^f}{E^f}}}^{-1} \bm{A}^{{E^f}{V^f}}\\
0 &
\bm{I}^{V^f V^f}\\
\end{array}\right]},\end{aligned}$$ where the superscript $^f$ (e.g. from $\bm{\mathcal{P}}^{fV^f}$) corresponds to all the fracture cells, regardless of their containing dual block. Similar as in the previous case, the sub-blocks of prolongation operator $\bm{\mathcal{P}}^{F^m V^m}$, $\bm{\mathcal{P}}^{F^m V^f}$, $\bm{\mathcal{P}}^{f V^f}$, $\bm{\mathcal{P}}^{E^m V^m}$, $\bm{\mathcal{P}}^{E^m V^f}$, $\bm{\mathcal{P}}^{f V^f}$ and $\bm{\mathcal{P}}^{E^f V^f}$ represent the corresponding rows and columns given in their superscripts. For example, $\bm{\mathcal{P}}^{E^m V^f} = - {\overline{\bm{A}}^{{E^m}{E^m}}}^{-1} \bm{A}^{{E^m}{f}} \bm{\mathcal{P}}^{f V^f}$ and, specially, $$\begin{aligned}
\label{pfVf}
\bm{\mathcal{P}}^{f V^f} = {\left[ \begin{array}{c}
- {\underline{\bm{A}}^{{F^f}{F^f}}}^{-1} \bm{A}^{{F^f}{E^f}} \bm{\mathcal{P}}^{E^f V^f}\\
-{\overline{\underline{\bm{A}}}^{{E^f}{E^f}}}^{-1} \bm{A}^{{E^f}{V^f}}\\
\bm{I}^{V^f V^f}\\
\end{array} \right]}.\end{aligned}$$
Rock-AMS {#sec:app-rockAMS}
--------
For Rock-AMS, the $\bm A^{mf}$ transmissibilities are set to zero,$$\begin{aligned}
\label{RockAMS-approxSys}\bm{\wp A' \wp^T} = {\left[\begin{array}{cccc|ccc}\underline{\bm{A}}^{{I^m}{I^m}} & \bm{A}^{{I^m}{F^m}} & 0 & 0 &0 & 0 &0 \\
0 & \overline{\underline{\bm{A}}}^{{F^m}{F^m}} & \bm{A}^{{F^m}{E^m}} & 0&0 & 0 &0 \\
0 & 0 & \overline{\underline{\bm{A}}}^{{E^m}{E^m}} & \bm{A}^{{E^m}{V^m}} &0 & 0 &0 \\
0 & 0 & 0 & \breve{\bm{A}}^{mm} &0 & 0 & \breve{\bm{A}}^{mf} \\
\hline\bm{A}^{{F^f}{I^m}} & \bm{A}^{{F^f}{F^m}} & \bm{A}^{{F^f}{E^m}} & \bm{A}^{{F^f}{V^m}} &\bm{A}^{{F^f}{F^f}} & \bm{A}^{{F^f}{E^f}} & 0 \\
\bm{A}^{{E^f}{I^m}} & \bm{A}^{{E^f}{F^m}} & \bm{A}^{{E^f}{E^m}} & \bm{A}^{{E^f}{V^m}} &0 & \overline{\bm{A}}^{{E^f}{E^f}} & \bm{A}^{{E^f}{V^f}} \\
0 & 0 & 0 & \breve{\bm{A}}^{fm} &0 & 0 & \breve{\bm{A}}^{ff}\end{array}\right]},$$where, the localization boundary condition was also applied.\
The Rock-AMS prolongation operator reads$$\begin{aligned}
\label{RockAMS-prlg}\bm{\mathcal{P}} = {\bm \wp^T}{\left[\begin{array}{c|c}- {\underline{\bm{A}}^{{I^m}{I^m}}}^{-1} \bm{A}^{{I^m}{F^m}} \bm{\mathcal{P}}^{F^m V^m} &0 \\
-{\overline{\underline{\bm{A}}}^{{F^m}{F^m}}}^{-1} \bm{A}^{{F^m}{E^m}} \bm{\mathcal{P}}^{E^m V^m} &0 \\
-{\overline{\underline{\bm{A}}}^{{E^m}{E^m}}}^{-1} \bm{A}^{{E^m}{V^m}} & 0 \\
\bm{I}^{V^m V^m}& 0\\
\hline -{\bm{A}^{{F^f}{F^f}}}^{-1} (\bm{A}^{{F^f}{E^f}} \bm{\mathcal{P}}^{E^f V^m} + \bm{A}^{{F^f}{m}} \bm{\mathcal{P}}^{m V^m}) & - {\bm{A}^{{F^f}{F^f}}}^{-1} \bm{A}^{{F^f}{E^f}} \bm{\mathcal{P}}^{E^f V^f}\\
- {\overline{\bm{A}}^{{E^f}{E^f}}}^{-1} \bm{A}^{{E^f}{m}} \bm{\mathcal{P}}^{m V^m} & -{\overline{\bm{A}}^{{E^f}{E^f}}}^{-1} \bm{A}^{{E^f}{V^f}}\\
0 & \bm{I}^{V^f V^f}\\
\end{array}\right]},$$where the superscript $^m$ (e.g. from $\bm{\mathcal{P}}^{mV^m}$) corresponds to all the matrix cells, regardless of their containing dual block. The sub-blocks of the prolongation operator, i.e.,$\bm{\mathcal{P}}^{F^m V^m}$, $\bm{\mathcal{P}}^{E^m V^m}$, $\bm{\mathcal{P}}^{E^f V^m} $, $\bm{\mathcal{P}}^{m V^m} $ and $\bm{\mathcal{P}}^{E^f V^f}$, are defined similar to the previous cases in the sense that they represent the corresponding rows and columns given in their superscripts. For example, $ \bm{\mathcal{P}}^{E^m V^m} = -{\overline{\underline{\bm{A}}}^{{E^m}{E^m}}}^{-1} \bm{A}^{{E^m}{V^m}}$ and, specially, $$\begin{aligned}
\label{PmVm}
\bm{\mathcal{P}}^{m V^m} =
\left[
\begin{array}{c}- {\underline{\bm{A}}^{{I^m}{I^m}}}^{-1} \bm{A}^{{I^m}{F^m}} \bm{\mathcal{P}}^{F^m V^m}\\
-{\overline{\underline{\bm{A}}}^{{F^m}{F^m}}}^{-1} \bm{A}^{{F^m}{E^m}} \bm{\mathcal{P}}^{E^m V^m}\\
-{\overline{\underline{\bm{A}}}^{{E^m}{E^m}}}^{-1} \bm{A}^{{E^m}{V^m}} \\
\bm{I}^{V^m V^m}\\
\end{array}\right].\end{aligned}$$
Coupled-AMS {#sec:app-coupledAMS}
-----------
In the Coupled-AMS approach, all adjacent Face and Edge blocks are merged between the media (Fig. \[fig:mergedDuals\]), i.e. $F = F^m \bigcup F^f$ and $E = E^m \bigcup E^f$. Also, let $V$ denote the set of coarse nodes, irrespective of their location. In this new setting, the $\bm\wp$-reordered approximate linear system is defined as$$\begin{aligned}
\label{CoupledAMS-approxSys}\bm{\wp A' \wp^T} = {\left[\begin{array}{cccc}\bm{A}^{I I} & \bm{A}^{I F} & \bm{A}^{I E} & \bm{A}^{I V} \\
0 & \overline{\bm{A}}^{F F} & \bm{A}^{F E} & \bm{A}^{F V} \\
0 & 0 & \overline{\bm{A}}^{E E} & \bm{A}^{E V} \\
0 & 0 & 0 & \breve{\bm{A}}\end{array}\right]},$$where the localization boundary condition was appropriately employed.\
Then, the Coupled-AMS prolongation operator reads $$\begin{aligned}
\label{Coupled-AMS_prolong_1}
\bm{\mathcal{P}} = {\bm \wp^T}{\left[\begin{array}{c}
-{\bm{A}^{I I}}^{-1} (A^{I F} \bm{\mathcal{P}}^{F V} + \bm{A}^{I E} \bm{\mathcal{P}}^{E V} + \bm{A}^{I V})\\
- {\overline{\bm{A}}^{F F}}^{-1} (\bm{A}^{F E} \bm{\mathcal{P}}^{E V} + \bm{A}^{F V})\\
-{\overline{\bm{A}}^{E E}}^{-1} \bm{A}^{E V}\\
\bm{I}^{V V}
\end{array}\right]}.\end{aligned}$$
Coarsening ratios used during scale sensitivity test
====================================================
The scale sensitivity experiment revealed that the performance of F-AMS is highly dependent on the coarsening ratios used. Only the optimum configuration was featured in the plots presented in the body of the manuscript. Table \[tab:coarsening\_ratio\] lists experimental results obtained when using primal grids with more refined and more coarse blocks, respectively, for comparison purposes.
[|c|c|c|c|]{} Scale &
-----------------
ratio
(matrix, fracs)
-----------------
: Performance of F-AMS during the scale sensitivity test cases, when using different coarsening factors. The middle row for each test (shown in bold) is the optimum configuration, whose results were presented in the body of the manuscript.[]{data-label="tab:coarsening_ratio"}
&
----------
CPU time
(sec)
----------
: Performance of F-AMS during the scale sensitivity test cases, when using different coarsening factors. The middle row for each test (shown in bold) is the optimum configuration, whose results were presented in the body of the manuscript.[]{data-label="tab:coarsening_ratio"}
& \# iterations\
& $2 \times 2 \times 2$, $2 \times 2$ & $1.904$ & $11$\
& $\bm{6 \times 6 \times 6}$, $\bm{6 \times 6}$ & $\bm{0.330}$ & $\bm{23}$\
& $8 \times 8 \times 8$, $8 \times 8$ & $0.351$ & $29$\
& $4 \times 4 \times 4$, $4 \times 4$ & $3.270$ & $13$\
& $\bm{6 \times 6 \times 6}$, $\bm{6 \times 6}$ & $\bm{2.186}$ & $\bm{19}$\
& $9 \times 9 \times 9$, $9 \times 9$ & $2.371$ & $27$\
& $6 \times 6 \times 6$, $6 \times 6$ & $21.790$ & $17$\
& $\bm{8 \times 8 \times 8}$, $\bm{8 \times 8}$ & $\bm{17.620}$ &$\bm{22}$\
& $11 \times 11 \times 11$, $11 \times 11$ & $21.350$ & $31$\
& $8 \times 8 \times 8$, $8 \times 8$ & $164.600$ & $18$\
& $\bm{10 \times 10 \times 10}$, $\bm{10 \times 10}$ & $\bm{150.400}$ & $\bm{23}$\
& $17 \times 17 \times 17$, $17 \times 17$ & $252.100$ & $40$\
| {
"pile_set_name": "ArXiv"
} |
---
title: New features in curvaton model
---
Introduction
============
The inflationary paradigm [@Guth:1980zm] has become an important ingredient of modern cosmology. Inflation provides a natural explanation for the production of the first density perturbations in the early universe which seeded the formation of the large scale structure (LSS) in the distribution of galaxies and the temperature anisotropies in the cosmic microwave background radiation (CMBR) [@Guth:1982ec]. However, the precise details of the mechanism for generating the primordial curvature perturbation is not fully established. The standard mechanism is via the quantum fluctuations of the inflaton field. An alternative scenario which frees the inflaton from the job of generating perturbations, besides giving rise to inflation, is the curvaton scenario [@Enqvist:2001zp; @Lyth:2001nq; @Moroi:2001ct]. The curvaton is assumed to be a light scalar field which begins evolving at the end of inflation. Its energy density is assumed to be subdominant during inflation, but it can share a significant part of the total energy in the universe before its decay. The entropy perturbations caused by the curvaton field finally get converted into adiabatic perturbations.
A large number of light scalar fields are expected to be present in any fundamental theory that goes beyond the standard model of particle physics. During the inflationary era these fields would have had the same amplitude of quantum fluctuations. It is plausible that at least some of them played important roles in the early universe, for example, as the curvaton field. An important distinguishing property of the curvaton scenario as the generating mechanism for primordial perturbations, from the standard single slow-rolling field picture, is the possibility for the primordial perturbations to have large deviations from Gaussian distribution. This property becomes very attractive in the light of the recent result from WMAP which suggests that primordial non-Gaussianity may be large [@Komatsu:2010fb]. The non-Gaussianity generated in the curvaton scenario must have a local shape because it is generated on superhorizon scales. Then the curvature perturbation can be expanded at the same spatial point to non-linear orders, as, ([**x**]{})=\_g([**x**]{})+[35]{}f\_[NL]{}$\zeta_g^2({\bf
x})-\langle \zeta_g^2\rangle $+[925]{} g\_[NL]{}$\zeta_g^3({\bf
x})-3\langle
\zeta_g^2\rangle \zeta_g$+... , \[zetafg\] where $f_{NL}$ and $g_{NL}$ are the so-called non-Gaussianity parameters. The WMAP 7yr result implies a constraint on the size of local form bispectrum as $f_{NL}=32\pm21$ at $1-\sigma$ level. The limits on $g_{NL}$ are $-3\times 10^{-5} < g_{NL} < 8\times 10^{-5}$ from LSS [@Desjacques:2009jb] and similar limits are obtained from CMB data from WMAP 5yr data [@Smidt:2010sv] as well. A convincing detection of the local form non-Gaussianity will rule out all single-field inflation in a model-independent way.
In the simplest case the curvaton potential is assumed to have a quadratic form and the typical size of the bispectrum is bounded by the tensor-scalar ratio, as, $f_{NL}<10^3
r^{1/4}$, [@Huang:2008ze]. Since the curvaton field evolves linearly in this case, the size of the trispectrum is linearly related to that of the bispectrum, as, g\_[NL]{}-[103]{}f\_[NL]{}. \[eq:gnlfnl\] There is, however, no reason for the above relation to hold in general from the viewpoint of fundamental theory. Apart from the quadratic potential, all the curvaton models that have been considered so far in the literature focus on potentials which deviate from the quadratic form at large field values but tend to the quadratic form at small field values. The predictions of such models, particularly the level of non-Gaussianity, are then compared with those from the quadratic potential so as to understand their distinguishing features. Clearly, the distinction becomes more prominent as the initial curvaton field value becomes larger and larger. A distinct signature of departure of the curvaton potential from quadratic form is a breakdown of the relation (\[eq:gnlfnl\]). If the curvaton self-interaction term becomes dominant, giving rise to higher order corrections in the curvaton potential, the order of magnitude of $g_{NL}$ can be ${\cal O}(f_{NL}^2)$ [@Huang:2008zj; @Enqvist:2009zf; @Enqvist:2009eq; @Enqvist:2009ww]. The predictions of curvaton model with nearly quadratic potential are investigated in [@Enqvist:2005pg; @Sasaki:2006kq; @Enqvist:2008gk; @Huang:2008bg], where the non-linear evolution of curvaton after inflation but prior to its oscillation is taken into account. Another promising curvaton candidate is the pseudo-Nambu-Goldstone boson – axion, whose potential significantly deviates from quadratic form around the top of its potential. A numerical analysis of the axion-type curvaton model is discussed in [@Kawasaki:2008mc; @Chingangbam:2009xi]. From the viewpoint of fundamental theory, one can generically expect multi curvatons models and such a model is investigated in [@Assadullahi:2007uw; @Huang:2008rj]. While the discussion this far has ignored scale dependence of the bispectrum and the trispectrum, it is possible that in the future such scale dependences may become accessible to experimental observation and hence important [@Byrnes:2009pe]. Other papers of relevance are [@Kawakami:2009iu; @Matsuda:2009kp; @Takahashi:2009cx; @Nakayama:2009ce; @Kawasaki:2009hp; @Cai:2010rt].
We investigate two new curvaton models different from the ones described above. The first is a potential which has tiny oscillations superimposed on the quadratic form. The resulting effect on the curvaton evolution is that it experiences the small bumps of the oscillations in the potentials during the stage of it evolution when it undergoes oscillations about the minimum of the potential. As a consequence, the curvaton evolution during this stage is non-linear (the curvaton equation of motion is not that of a damped simple harmonic oscillator), making it significantly different from the curvaton oscillation about the minimum of a quadratic potential. Our goal is to calculate the non-linear curvature perturbation up to cubic order and obtain the predictions for non-Gaussianity from such a model. We find very interesting new implications for the non-linearity parameters $f_{NL}$ and $g_{NL}$ arising in this model. First, $f_{NL}$ is no longer restricted to have positive values. Depending on the amplitude and the frequency of the superimposed oscillations on the potential, it can take a wide range of both positive and negative, with a switch of sign from positive to negative. $g_{NL}$ , on the other hand, remains negative and can take large negative values. The sign switch of $f_{NL}$ brings up the possibility that the most important contribution to primordial non-Gaussianity could come from the $g_{NL}$ term, with $f_{NL}$ being negligibly small.
The second model we discuss is a class of potentials characterized by a single [*feature*]{} separating two quadratic regimes with different mass scales. The feature depends on a single parameter and depending on the sign of the parameter, it can be either a single bump or a flattening of the slope of the quadratic potential at some characteristic scale. We find that the effect of the feature on $f_{NL}$ and $g_{NL}$ is rather dramatic, causing them to oscillate with increasing amplitude as the strength of the feature increases.
This paper is organized as follows: in section 2, we briefly summarize the method for computation of the non-linear curvature perturbation using the $\delta N$ formalism and the curvaton equation of motion. In section 3, we describe the specific forms of the curvaton potentials we are considering here and display our results for the non-linear corrections to the curvature perturbations. In section 3.1, we discuss the case of the washboard potential, while in section 3.2, we discuss the single feature potential. We end with a summary of our results and comments in section 4. A brief description of the curvaton with quadratic potential is given in the appendix to highlight the differences from our study and novelty of our results.
The non-linear curvature perturbation
=====================================
On sufficiently large scales, the curvature perturbation on the uniform density slicing can be calculated by using the so-called $\delta N$ formalism [@Starobinsky:1986fxa; @Sasaki:1995aw; @Sasaki:1998ug; @Lyth:2004gb; @Lyth:2005fi]. Starting from any initial flat slice at time $t_{ini}$, on the uniform density slicing, the curvature perturbation is (t,[**x**]{})=NN(t,[**x**]{})-N\_0(t), where $N(t,{\bf x})=\ln a(t,{\bf x})/a(t_{ini})$ describes the local expansion of our universe, and $N_0(t)=\ln
a(t)/a(t_{ini})$ is the unperturbed amount of expansion. In curvaton model, the difference between local expansion and the unperturbed expansion is caused by the quantum fluctuations of curvaton field during inflation. Therefore =N\_[,]{}+N\_[,]{}\^2+[16]{} N\_[,]{} \^3+... , where $N_{,\sigma}=dN/d\sigma$, $N_{,\sigma\sigma}=d^2N/d\sigma^2$ and $N_{,\sigma\sigma\sigma}=d^3N/d\sigma^3$. Considering $\delta\sigma={H_*}/2\pi$, the amplitude of the power spectrum generated by curvaton field is P\_[\_]{}=N\_[,]{}\^2 ${H_*\over 2\pi} $\^2, and the non-Gaussianity parameters are given by f\_[NL]{}&=& [56]{}[N\_[,]{}N\_[,]{}\^2]{},\
g\_[NL]{}&=& [2554]{} [N\_[,]{}N\_[,]{}\^3]{}, where $H_*$ is the Hubble parameter during inflation. On the other hand, the amplitude of the tensor perturbation only depends on the inflation scale, namely P\_T=[H\_\*\^2/M\_p\^2 \^2/2]{}. Thus the tensor-scalar ratio $r$ is given by rP\_T/P\_[\_]{}=[8N\_[,]{}\^2 M\_p\^2]{}. Here we consider the simplest version of curvaton scenario where the quantum fluctuations of convaton field contribute the total curvature perturbation.
After inflation, the equations of motion are H\^2&=&[13M\_p\^2]{}(\_r+\_),\
\_r&+&4H\_r=0,\
\_&=&\^2+V(),\
&+&3H+[dV()d]{}=0, where $\rho_r$ and $\rho_\sigma$ are the energy densities of radiation and curvaton respectively, and $V(\sigma)$ is curvaton potential. In order to numerically solve the above differential equations, we define the reduced curvaton field $\tilde \sigma$ and reduced curvaton potential $V(\tilde \sigma)$ as follows &=& /\_\*,\
V()&=& [V()m\^2\_\*\^2]{}, where $\sigma_*$ is the vacuum expectation value (VEV) of curvaton field in the inflationary era. Now the equations of motion can be simplified to be N’&=&$$\alpha e^{-4N}+{\sigma_*^2\over 3M_p^2}\(\half {\tilde
\sigma}'^2 +V({\tilde \sigma})\)$$\^, \[eqn:nprime\]\
”&+&3N’’+[dV() d]{}=0, \[eqn:tsigma\] where $N(x)=\ln a(t)$, $\alpha={\rho_{r,ini}\over
3M_p^2m^2}={H_{ini}^2/m^2}$, and the prime denotes the derivative with respect to dimensionless time coordinate $x\equiv mt$, and the Hubble parameter becomes H=mN’. The solution for the subdominant curvaton with quadratic potential is analytically discussed in the appendix.
The scale factor can be rescaled to satisfy $a(t_{ini})=1$, or equivalently $N(t_{ini})=0$. For numerical calculation, we also need to input the value of $\alpha$. If the vacuum energy of inflaton suddenly decays into radiation, a reasonable choice is $\alpha=H_{inf}^2/m^2$ which is much larger than one. However we don’t know its value exactly. But as long as $\alpha$ is large enough it does not affect our numerical result because the curvaton field almost does not move when the Hubble parameter is much larger than its mass. For example, it is reasonable to assume that the Hubble parameter at the inflationary era is one order of magnitude larger than the curvaton mass at least and then we set $\alpha=10^2$ in this paper.
The models
==========
In this section we consider two new curvaton models which have some small features around the exactly quadratic form of the curvaton potential. We can expect that these features will introduce non-linear effects to the oscillating curvaton field and consequently affect the non-Gaussianity parameters. Our aim is to calculate the precise effects. Note that these effects are different from what was considered in [@Enqvist:2005pg; @Sasaki:2006kq; @Enqvist:2008gk; @Huang:2008bg] where the non-linear evolution of curvaton after inflation but prior to its oscillation was considered. Since the non-linear nature of the curvaton motion makes analytic solutions extremely difficult to obtain, we rely on numerical methods to get our results. We solve the Eqs. (\[eqn:nprime\]) and (\[eqn:tsigma\]) as a coupled set of differential equations for each potential under consideration.
Washboard curvaton model
------------------------
Let us consider a curvaton potential which has tiny oscillations superimposed on the exactly quadratic form. We call it the [*washboard model*]{} and it takes the following explicit form V()=m\^2 \^2+ V\_0 $1-\cos ({\sigma\over F})$, where $V_0\ll V_*=\half m^2\sigma_*^2$. The reduced potential of $\tilde \sigma$ is V()=\^2+$1-\cos({\tilde
\sigma/\delta})$, \[eqn:pot\_wb\] where =[V\_0m\^2\_\*\^2]{}, =[F\_\*]{}.
![The washboard curvaton potential given by Eq. (3.2) is shown on the left panel for visual comparision with the corresponding quadratic one. The parameter values are $\epsilon=5\times 10^{-4}$ and $\delta=10^{-2}$. We have chosen a large value of $\epsilon$ in order to make the oscillations clearly visible. The right panel shows the curvaton oscillations about the potential minimum for the quadratic and washboard cases, with the same initial field value given by $\sigma_*/M_p=0.1$. The parameter values for this plot are $\epsilon=10^{-4}$ and $\delta=10^{-2}$[]{data-label="fig:pwb"}](pot_washboard.ps "fig:"){height="6.5cm" width="6.cm"} ![The washboard curvaton potential given by Eq. (3.2) is shown on the left panel for visual comparision with the corresponding quadratic one. The parameter values are $\epsilon=5\times 10^{-4}$ and $\delta=10^{-2}$. We have chosen a large value of $\epsilon$ in order to make the oscillations clearly visible. The right panel shows the curvaton oscillations about the potential minimum for the quadratic and washboard cases, with the same initial field value given by $\sigma_*/M_p=0.1$. The parameter values for this plot are $\epsilon=10^{-4}$ and $\delta=10^{-2}$[]{data-label="fig:pwb"}](washboard_curvaton_motion_alpha100.ps "fig:"){height="6.5cm" width="9.cm"}
Here $\epsilon$ measures the size of the correction and $\delta$ characterizes the period of oscillation of the correction term in the washboard potential. The reduced potential is shown in the left panel of Fig. \[fig:pwb\] for easy visualization. When ${\tilde \sigma}\gg \sqrt{\epsilon}$, the potential is almost quadratic. If ${\tilde \sigma}\ll \delta$, then $V(\tilde
\sigma)\simeq \half (1+{\epsilon/ \delta^2}) {\tilde
\sigma}^2$. Then the curvaton potential is roughly quadratic as well, but has a deformed mass.
The dynamics of curvaton field after inflation is governed by ”+[32x]{}’++ ${\tilde \sigma/
\delta} $=0. Even though the correction to the potential is small, the dynamics of curvaton field becomes significantly non-linear if the period of the correction term is small enough. Here we consider the case in which the dynamics of curvaton is dominated by the mass term in the beginning, which implies $\epsilon/\delta <1$. Once the amplitude of the curvaton oscillation drops below $\epsilon/\delta$, the curvaton evolves non-linearly. On the right panel of Fig. \[fig:pwb\] we have plotted the oscillation of the curvaton field about the minimum of the potential for the quadratic and the washboard potential cases, for the same initial field value given by $\sigma_*/M_p=0.1$. We can see that the amplitude of oscillation in the washboard case decreases faster than the quadratic case. Moreover, the frequency of oscillation for the washboard curvaton is time dependent, it oscilates about the constant frequency of the quadratic case.
![$N_{,\sigma},\ N_{,\sigma\sigma} ,\
N_{,\sigma\sigma\sigma}$, $f_{NL}$ and $g_{NL}$ are shown as functions of $\epsilon$ for the washboard potential, for fixed values of $\delta,\ \Gamma/m$ and $\sigma_*/M_p$. We have shown plots for two different values of $\Gamma/m$ in order to demonstrate the systematic variation as we change $\Gamma/m$. []{data-label="fig:varygamma"}](efold_fnl_varygamma_alpha100.ps "fig:"){height="9.5cm" width="7.5cm"} ![$N_{,\sigma},\ N_{,\sigma\sigma} ,\
N_{,\sigma\sigma\sigma}$, $f_{NL}$ and $g_{NL}$ are shown as functions of $\epsilon$ for the washboard potential, for fixed values of $\delta,\ \Gamma/m$ and $\sigma_*/M_p$. We have shown plots for two different values of $\Gamma/m$ in order to demonstrate the systematic variation as we change $\Gamma/m$. []{data-label="fig:varygamma"}](efold_gnl_varygamma_alpha100.ps "fig:"){height="9.5cm" width="7.5cm"}
We now illustrate how the small features in the curvaton potential play an important role for the non-Gaussianity parameters. First we solve for $N_{,\sigma}$, $N_{,\sigma\sigma}$ and $N_{,\sigma\sigma\sigma}$ and then obtain $f_{NL}$ and $g_{NL}$ from them. In general, we need to scan four independent parameters, namely, $\Gamma/m, \ \sigma_*/M_p, \ \epsilon$ and $\delta$ in order to satisfy observational constraints such as amplitude of perturbations and the limits on $f_{NL}$ and $g_{NL}$. Our strategy here is to fix $\Gamma/m, \ \sigma_*/M_p$ and $\delta$ and obtain $f_{NL}$ and $g_{NL}$ as functions of $\epsilon$. Our results are obtained for two values of $\Gamma/m$ and $\delta$ each, to understand how these parameters systematically affect $f_{NL}$ and $g_{NL}$.
In the case of the quadratic potential $\Gamma/m$ is typically required to be of the order of $10^{-8}$ for the amplitude of perturbations to be COBE normalized. Evolving the equations numerically till the energy density of the curvaton decreases to such small value is prohibitively time consuming. Moreover, for the purpose of capturing the essential features of the dependence of $f_{NL}$ and $g_{NL}$ on $\epsilon$ and $\delta$, it is enough to fix $\Gamma/m$ at a relatively larger value. To demonstrate this point, in Fig. [\[fig:varygamma\]]{} we have plotted for $\Gamma/m = 10^{-2}$ and $2\times 10^{-2}$, how $N_{\sigma},\ N_{\sigma\sigma},\
N_{\sigma\sigma\sigma}$, $f_{NL}$ and $g_{NL}$ vary as functions of $\epsilon$, for fixed values of $\delta=10^{-2}$ and $\sigma_*/M_p=0.1$. We can see that $\Gamma/m$ systematically changes the amplitudes of $f_{NL}$ and $g_{NL}$, but does not alter the essential functional shapes. The correctness of the numerical calculations are tested by ensuring that in the limit $\epsilon \rightarrow 0$, $f_{NL}$ and $g_{NL}$ tend to their analytically expected values for the quadratic potential, as clearly seen in the figure. As $\epsilon$ increases, $f_{NL}$ and $g_{NL}$ get strongly affected and deviates from their expectation from quadratic potential. $f_{NL}$ crosses over from positive to increasingly negative values as $\epsilon$ increases. On the other hand, $g_{NL}$ remains negative throughout but its magnitude becomes very large as $\epsilon$ increases. The inset figures in the panels showing $N_{\sigma\sigma\sigma}$ and $g_{NL}$ zoom in on the $\epsilon \rightarrow 0$ region to show them approaching the negative values expected from quadratic potential.
Next, in Fig. \[fig:varydelta\] we have plotted $N_{\sigma},\
N_{\sigma\sigma},\ N_{\sigma\sigma\sigma}$, $f_{NL}$ and $g_{NL}$ as functions of $\epsilon$, for two different values of $\delta$. We have chosen $\delta=10^{-2}$ and $1.8\times 10^{-2}$ and fixed $\Gamma/m = 10^{-2}$ and $\sigma_*/M_p=0.1$. We see that for very small $\epsilon$, varying $\delta$ has little effect on the behavior of $f_{NL}$ and $g_{NL}$. This can be explained by the fact that $\epsilon \rightarrow 0$ kills off the oscillations superimposed on the potential, regardless of the frequency of oscillations which is controlled by $\delta$. At relatively larger values of $\epsilon$, the effect of $\delta$ becomes prominent. As the deviation of $f_{NL}$ and $g_{NL}$ from the quadratic potential behavior increases as $\delta$ decreases, due to the increase in the frequency of the oscillations in the potential. As in Fig. \[fig:varygamma\] the inset figures in the panels showing $N_{\sigma\sigma\sigma}$ and $g_{NL}$ zoom in on the $\epsilon
\rightarrow 0$ region to show them approaching the negative values expected from quadratic potential.
![Same as Fig. 2 but for two different values of $\delta$, with $\Gamma/m$ and $\sigma_*/M_p$ kept fixed. []{data-label="fig:varydelta"}](efold_fnl_varydelta_alpha100.ps "fig:"){height="9.5cm" width="7.5cm"} ![Same as Fig. 2 but for two different values of $\delta$, with $\Gamma/m$ and $\sigma_*/M_p$ kept fixed. []{data-label="fig:varydelta"}](efold_gnl_varydelta_alpha100.ps "fig:"){height="9.5cm" width="7.5cm"}
Single-feature curvaton model
-----------------------------
In this subsection we consider a curvaton model with the potential given by V()=m\^2 \^2 $1+{c\over 1+(\sigma/M)^{2n}}$, where $n>0$, and $M$ is an energy scale which measures the position of the feature. In the regime $\sigma \gg M$ or $\sigma\ll M$, the curvaton potential has a quadratic form, but around $\sigma\sim M$ the potential deviates from quadratic form. As in the previous subsection, we define a reduced potential, as follows V()=\^2$1+{c\over 1+({\tilde
\sigma}/d)^{2n}}$, \[eqn:psb\] where d=[M\_\*]{}.
![ The single feature curvaton potential given by Eq. (3.7) is shown on the left panel for comparision with the corresponding quadratic one. $d$ is fixed to be $0.1$ and we have plotted for $c=2$ and $-1$ to show how the nature of the feature changes with the sign of $c$. The right panel shows the corresponding curvaton oscillations about the potential minimum, with the same initial field value given by $\sigma_*/M_p=0.1$. []{data-label="fig:psb"}](pot_singlebump.ps "fig:"){height="6.5cm" width="6.cm"} ![ The single feature curvaton potential given by Eq. (3.7) is shown on the left panel for comparision with the corresponding quadratic one. $d$ is fixed to be $0.1$ and we have plotted for $c=2$ and $-1$ to show how the nature of the feature changes with the sign of $c$. The right panel shows the corresponding curvaton oscillations about the potential minimum, with the same initial field value given by $\sigma_*/M_p=0.1$. []{data-label="fig:psb"}](singlebump_curvaton_motion.ps "fig:"){height="6.5cm" width="9.cm"}
The reduced potential given by Eq. (\[eqn:psb\]) is shown in Fig. \[fig:psb\] for $n=2$ and $d=0.1$. The nature of the feature depends on the sign of $c$. If $c$ is positive, then there is a bump, whereas, a negative $c$ changes the slope of the potential to make it flatter around some scale set by the parameter $d$.
The equation of motion for the reduced curvaton field becomes ”&+&3N’’+$1+c{1-(n-1)({\tilde
\sigma}/d)^{2n} \over (1+({\tilde \sigma}/d)^{2n})^2} $ [ ]{}=0. We restrict our analysis here to $n=2$. If $d\ll 1$ and the initial curvaton field value is large enough, then the curvaton evolves linearly prior to its oscillation. We choose $d=0.1$. As in the washboard curvaton model, we choose $\Gamma_\sigma/m$ to be $10^{-2}$ and $\sigma_*/M_p=10^{-1}$. Then we solve for $f_{NL}$, scanning the parameter $c$.
![ $N_{,\sigma},\ N_{,\sigma\sigma} ,\
N_{,\sigma\sigma\sigma}$, $f_{NL}$ and $g_{NL}$ are shown as functions of $c$ for the single-feature potential, with $d,\
\Gamma/m$ and $\sigma_*/M_p$ kept fixed. []{data-label="fig:sovv"}](fnl_singlebump.ps "fig:"){height="9.5cm" width="7.5cm"} ![ $N_{,\sigma},\ N_{,\sigma\sigma} ,\
N_{,\sigma\sigma\sigma}$, $f_{NL}$ and $g_{NL}$ are shown as functions of $c$ for the single-feature potential, with $d,\
\Gamma/m$ and $\sigma_*/M_p$ kept fixed. []{data-label="fig:sovv"}](gnl_singlebump.ps "fig:"){height="9.5cm" width="7.5cm"}
On the left panel of Fig. \[fig:sovv\], we have plotted $N_{,\sigma}$, $N_{,\sigma\sigma}$ and $f_{NL}$ for the single-feature potential. As shown in the figure, $f_{NL}$ oscillates about zero with increasing amplitude as ${\lvertc\rvert}$ increases. We can see $f_{NL}$ flattening out and approaching the expected value from the quadratic potential as ${\lvertc\rvert}\rightarrow 0$. On the right panel of the same fugure we have plotted $N_{,\sigma}$, $N_{,\sigma\sigma\sigma}$ and $g_{NL}$. We again obtain oscillatory behavior of $g_{NL}$ as $c$ varies. Similar to $f_{NL}$, we can see the curve flattening out near $c=0$ for $g_{NL}$ to assumes the value expected from quadratic potential, as shown in the inset figure on the right bottom panel.
Conclusion and discussion
=========================
We have studied two new curvaton models in this paper. The first is the washboard model where the potential has tiny oscillations superimposed on the quadratic form, and the second one has a potential with two quadratic regimes having different mass scales separated by either a bump or a flattening of the potential. For the washboard model we have investigated in detail how the two parameters that control the oscillations, namely, the amplitude and the frequency, affect the non-linear corrections to the curvature perturbation via their effect on $f_{NL}$ and $g_{NL}$. We have shown that the relation $g_{NL} \propto -f_{NL}$, which holds for the quadratic potential, is no longer valid in this case. We also found that there is a wide range of both positive and negative values for $f_{NL}$, while $g_{NL}$ remains negative but its magnitude can be very large depending on the model parameters. In comparision, the quadratic potential restricts $f_{NL}$ to be positive and $g_{NL}$ to be negative. For the single-feature model we have again calculated $f_{NL}$ and $g_{NL}$, and demonstrated that they strongly depend on the strength of the feature and oscillate as the strength increases.
What is new in the models considered here is that the curvation motion as it oscillates about the potential minimum is non-linear, unlike other models that have been considered so far in the literature. The results that we have found have interesting implications for searches for non-Gaussianity in observational data. The fact that $f_{NL}$ can switch sign at some parameter values implies that it is possible that the non-linear contributions to the curvature perturbation could be coming from $g_{NL}$ alone, with $f_{NL}$ being close to zero. Similar result was obtained in [@Enqvist:2005pg] in the context of curvaton potential with non-linear corrections to the quadratic term. It is also possible that both $f_{NL}$ and $g_{NL}$ contribute comparably with the same or opposite signs. The present work thus throws up the need to understand different sources of primordial non-Gaussianity and how they can be distinguished in the observational data. It is important to devise observables which can distinguish them. Such studies have been initiated in [@Chingangbam:2009vi; @Matsubara:2010te]. We also want to mention that $f_{NL}$ and $g_{NL}$ are controlled by two independent geometric quantities which characterize the hyper-surface in field space on which multi-field inflation ends. This has the implication that all of the possible results for $f_{NL}$ and $g_{NL}$ in curvaton models can be realized by tuning these two independent geometric quantities, as shown in [@Huang:2009vk; @Sasaki:2008uc; @Huang:2009xa].
In principle, the three free parameters in the washboard model, $\Gamma/m, \epsilon$ and $\delta$, can be constrained by using the observational constraints on $f_{NL}$, $g_{NL}$ and the amplitude of perturbations. The parameters $c$ and $d$ in the single-feature model can be similarly constrained. However, such a full scan of the parameter space is beyond the scope of the present analysis. Our purpose in this paper has been to understand the systematic behaviors of $f_{NL}$ and $g_{NL}$ as functions of the model parameters. We will tackle the problem of scanning the parameter space in a future work.
[**Acknowledgments**]{}
P.C. is supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MEST) (No. 2009-0062868). QGH is supported by the project of Knowledge Innovation Program of Chinese Academy of Science. The numerical calculation in this work was carried out on the QUEST cluster computing facility at Korea Institute for Advanced Study.
Curvaton model with quadratic potential
=======================================
For the curvaton model with quadratic potential, from [@Lyth:2005fi], we have f\_[NL]{}&=& [54f\_D]{}-[53]{}-[56]{}f\_D,\
g\_[NL]{}&=& -[256f\_D]{}+[25108]{}+[12527]{}f\_D+[2518]{}f\_D\^2, where f\_D=[3\_[,D]{}4-\_[,D]{}]{}. After inflation the universe is dominated by radiation and the Hubble parameter is related to the cosmic time $t$ by $H={1\over 2t}$. The equation of motion of curvaton field with quadratic potential becomes ”+[32x]{}’+=0, whose solution is =2\^[1/4]{}(5/4) x\^[-1/4]{} J\_[1/4]{}(x), where $J_{\nu}(x)$ is the Bessel function of the first kind. Therefore the energy density of curvaton is given by \_=m\^2\_\*\^2 ${\tilde \sigma}^2+({d{\tilde \sigma}\over dx})^2$ =[\^2(5/4)]{} m\^2\_\*\^2 x\^[-1/2]{}$J_{1/4}^2(x)+J_{5/4}^2(x)$. Adopting the sudden decay approximation, we have \_[,D]{}=[\_(x\_D)3M\_p\^2\_\^2]{} 0.35[\_\*\^2M\_p\^2]{}, in the limit of $x_D=\half {m\over
\Gamma_\sigma}\gg 1$. In the literatures, $\Omega_{\sigma,D}={\sigma_*^2\over 6M_p^2}\sqrt{m\over
\Gamma_\sigma}$ which is roughly half of our exact retult.
[99]{}
A. H. Guth, “The Inflationary Universe: A Possible Solution To The Horizon And Flatness Problems,” Phys. Rev. D [**23**]{}, 347 (1981).
A. H. Guth and S. Y. Pi, “Fluctuations In The New Inflationary Universe,” Phys. Rev. Lett. [**49**]{}, 1110 (1982).
K. Enqvist and M. S. Sloth, “Adiabatic CMB perturbations in pre big bang string cosmology,” Nucl. Phys. B [**626**]{}, 395 (2002) \[arXiv:hep-ph/0109214\].
D. H. Lyth and D. Wands, “Generating the curvature perturbation without an inflaton,” Phys. Lett. B [**524**]{}, 5 (2002) \[arXiv:hep-ph/0110002\].
T. Moroi and T. Takahashi, “Effects of cosmological moduli fields on cosmic microwave background,” Phys. Lett. B [**522**]{}, 215 (2001) \[Erratum-ibid. B [**539**]{}, 303 (2002)\] \[arXiv:hep-ph/0110096\].
E. Komatsu [*et al.*]{}, “Seven-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Cosmological Interpretation,” arXiv:1001.4538 \[astro-ph.CO\].
V. Desjacques and U. Seljak, “Signature of primordial non-Gaussianity of $\phi^3$-type in the mass function and bias of dark matter haloes,” Phys. Rev. D [**81**]{}, 023006 (2010) \[arXiv:0907.2257 \[astro-ph.CO\]\]. J. Smidt, A. Amblard, A. Cooray, A. Heavens, D. Munshi and P. Serra, “A Measurement of Cubic-Order Primordial Non-Gaussianity ($g_{NL}$ and $\tau_{NL}$) With WMAP 5-Year Data,” arXiv:1001.5026 \[astro-ph.CO\]. Q. G. Huang, “Large Non-Gaussianity Implication for Curvaton Scenario,” Phys. Lett. B [**669**]{}, 260 (2008) \[arXiv:0801.0467 \[hep-th\]\].
Q. G. Huang, “A Curvaton with a Polynomial Potential,” JCAP [**0811**]{}, 005 (2008) \[arXiv:0808.1793 \[hep-th\]\].
K. Enqvist, S. Nurmi, G. Rigopoulos, O. Taanila and T. Takahashi, “The Subdominant Curvaton,” JCAP [**0911**]{}, 003 (2009) \[arXiv:0906.3126 \[astro-ph.CO\]\].
K. Enqvist and T. Takahashi, “Effect of Background Evolution on the Curvaton Non-Gaussianity,” JCAP [**0912**]{}, 001 (2009) \[arXiv:0909.5362 \[astro-ph.CO\]\].
K. Enqvist, S. Nurmi, O. Taanila and T. Takahashi, “Non-Gaussian Fingerprints of Self-Interacting Curvaton,” JCAP [**1004**]{}, 009 (2010) \[arXiv:0912.4657 \[astro-ph.CO\]\].
K. Enqvist and S. Nurmi, “Non-gaussianity in curvaton models with nearly quadratic potential,” JCAP [**0510**]{}, 013 (2005) \[arXiv:astro-ph/0508573\].
M. Sasaki, J. Valiviita and D. Wands, “Non-gaussianity of the primordial perturbation in the curvaton model,” Phys. Rev. D [**74**]{}, 103003 (2006) \[arXiv:astro-ph/0607627\].
K. Enqvist and T. Takahashi, “Signatures of Non-Gaussianity in the Curvaton Model,” JCAP [**0809**]{}, 012 (2008) \[arXiv:0807.3069 \[astro-ph\]\].
Q. G. Huang and Y. Wang, “Curvaton Dynamics and the Non-Linearity Parameters in Curvaton Model,” JCAP [**0809**]{}, 025 (2008) \[arXiv:0808.1168 \[hep-th\]\].
M. Kawasaki, K. Nakayama and F. Takahashi, “Hilltop Non-Gaussianity,” JCAP [**0901**]{}, 026 (2009) \[arXiv:0810.1585 \[hep-ph\]\].
P. Chingangbam and Q. G. Huang, “The Curvature Perturbation in the Axion-type Curvaton Model,” JCAP [**0904**]{}, 031 (2009) \[arXiv:0902.2619 \[astro-ph.CO\]\].
H. Assadullahi, J. Valiviita and D. Wands, “Primordial non-Gaussianity from two curvaton decays,” Phys. Rev. D [**76**]{}, 103003 (2007) \[arXiv:0708.0223 \[hep-ph\]\].
Q. G. Huang, “The N-vaton,” JCAP [**0809**]{}, 017 (2008) \[arXiv:0807.1567 \[hep-th\]\].
C. T. Byrnes, S. Nurmi, G. Tasinato and D. Wands, “Scale dependence of local $f_{NL}$,” JCAP [**1002**]{}, 034 (2010) \[arXiv:0911.2780 \[astro-ph.CO\]\].
E. Kawakami, M. Kawasaki, K. Nakayama and F. Takahashi, “Non-Gaussianity from Isocurvature Perturbations : Analysis of Trispectrum,” JCAP [**0909**]{}, 002 (2009) \[arXiv:0905.1552 \[astro-ph.CO\]\].
T. Matsuda, “Delta-N formalism for the evolution of the curvature perturbations in generalized multi-field inflation,” Phys. Lett. B [**682**]{}, 163 (2009) \[arXiv:0906.2525 \[hep-th\]\].
T. Takahashi, M. Yamaguchi and S. Yokoyama, “Primordial Non-Gaussianity in Models with Dark Matter Isocurvature Fluctuations,” Phys. Rev. D [**80**]{}, 063524 (2009) \[arXiv:0907.3052 \[astro-ph.CO\]\].
K. Nakayama and J. Yokoyama, “Gravitational Wave Background and Non-Gaussianity as a Probe of the Curvaton Scenario,” JCAP [**1001**]{}, 010 (2010) \[arXiv:0910.0715 \[astro-ph.CO\]\].
M. Kawasaki, T. Takahashi and S. Yokoyama, “Density Fluctuations in Thermal Inflation and Non-Gaussianity,” JCAP [**0912**]{}, 012 (2009) \[arXiv:0910.3053 \[hep-th\]\].
Y. F. Cai and Y. Wang, “Large Nonlocal Non-Gaussianity from a Curvaton Brane,” arXiv:1005.0127 \[hep-th\].
A. A. Starobinsky, “Multicomponent de Sitter (Inflationary) Stages and the Generation of Perturbations,” JETP Lett. [**42**]{} (1985) 152.
M. Sasaki and E. D. Stewart, “A General Analytic Formula For The Spectral Index Of The Density Perturbations Produced During Inflation,” Prog. Theor. Phys. [**95**]{}, 71 (1996) \[arXiv:astro-ph/9507001\].
M. Sasaki and T. Tanaka, “Super-horizon scale dynamics of multi-scalar inflation,” Prog. Theor. Phys. [**99**]{}, 763 (1998) \[arXiv:gr-qc/9801017\].
D. H. Lyth, K. A. Malik and M. Sasaki, “A general proof of the conservation of the curvature perturbation,” JCAP [**0505**]{}, 004 (2005) \[arXiv:astro-ph/0411220\].
D. H. Lyth and Y. Rodriguez, “The inflationary prediction for primordial non-gaussianity,” Phys. Rev. Lett. [**95**]{}, 121302 (2005) \[arXiv:astro-ph/0504045\].
P. Chingangbam and C. Park, “Statistical nature of non-Gaussianity from cubic order primordial perturbations: CMB map simulations and genus statistic,” JCAP [**0912**]{}, 019 (2009) \[arXiv:0908.1696 \[astro-ph.CO\]\]. T. Matsubara, “Analytic Minkowski Functionals of the Cosmic Microwave Background: Second-order Non-Gaussianity with Bispectrum and Trispectrum,” Phys. Rev. D [**81**]{}, 083505 (2010) \[arXiv:1001.2321 \[astro-ph.CO\]\].
Q. G. Huang, “A geometric description of the non-Gaussianity generated at the end of multi-field inflation,” JCAP [**0906**]{}, 035 (2009) \[arXiv:0904.2649 \[hep-th\]\].
M. Sasaki, “Multi-brid inflation and non-Gaussianity,” Prog. Theor. Phys. [**120**]{}, 159 (2008) \[arXiv:0805.0974 \[astro-ph\]\].
Q. G. Huang, “The Trispectrum in the Multi-brid Inflation,” JCAP [**0905**]{}, 005 (2009) \[arXiv:0903.1542 \[hep-th\]\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We report a search for a dark vector gauge boson $U^\prime$ that couples to quarks in the decay chain $D^{*+} \to D^0 \pi^+, D^0 \to K^0_S \eta, \eta \to
U^\prime \gamma$, $U^\prime \to \pi^+ \pi^-$. No signal is found and we set a mass-dependent limit on the baryonic fine structure constant of $10^{-3} - 10^{-2}$ in the $U^\prime$ mass range of 290 to 520 MeV/$c^2$. This analysis is based on a data sample of 976 fb$^{-1}$ collected by the Belle experiment at the KEKB asymmetric-energy $e^+e^-$ collider.
author:
- 'E. Won'
- 'I. Adachi'
- 'H. Aihara'
- 'S. Al Said'
- 'D. M. Asner'
- 'T. Aushev'
- 'R. Ayad'
- 'I. Badhrees'
- 'A. M. Bakich'
- 'V. Bansal'
- 'E. Barberio'
- 'P. Behera'
- 'B. Bhuyan'
- 'J. Biswal'
- 'A. Bobrov'
- 'A. Bozek'
- 'M. Bračko'
- 'D. Červenkov'
- 'V. Chekelian'
- 'A. Chen'
- 'B. G. Cheon'
- 'K. Chilikin'
- 'R. Chistov'
- 'K. Cho'
- 'V. Chobanova'
- 'Y. Choi'
- 'D. Cinabro'
- 'N. Dash'
- 'S. Di Carlo'
- 'Z. Doležal'
- 'Z. Drásal'
- 'D. Dutta'
- 'S. Eidelman'
- 'D. Epifanov'
- 'H. Farhat'
- 'J. E. Fast'
- 'T. Ferber'
- 'B. G. Fulsom'
- 'V. Gaur'
- 'N. Gabyshev'
- 'A. Garmash'
- 'R. Gillard'
- 'P. Goldenzweig'
- 'D. Greenwald'
- 'J. Haba'
- 'K. Hayasaka'
- 'H. Hayashii'
- 'W.-S. Hou'
- 'T. Iijima'
- 'K. Inami'
- 'G. Inguglia'
- 'A. Ishikawa'
- 'R. Itoh'
- 'Y. Iwasaki'
- 'I. Jaegle'
- 'H. B. Jeon'
- 'D. Joffe'
- 'K. K. Joo'
- 'T. Julius'
- 'K. H. Kang'
- 'T. Kawasaki'
- 'D. Y. Kim'
- 'J. B. Kim'
- 'K. T. Kim'
- 'M. J. Kim'
- 'S. H. Kim'
- 'Y. J. Kim'
- 'K. Kinoshita'
- 'P. Kodyš'
- 'P. Križan'
- 'P. Krokovny'
- 'T. Kuhr'
- 'R. Kulasiri'
- 'Y.-J. Kwon'
- 'J. S. Lange'
- 'I. S. Lee'
- 'C. H. Li'
- 'L. Li'
- 'Y. Li'
- 'L. Li Gioi'
- 'J. Libby'
- 'D. Liventsev'
- 'T. Luo'
- 'M. Masuda'
- 'T. Matsuda'
- 'D. Matvienko'
- 'K. Miyabayashi'
- 'H. Miyata'
- 'R. Mizuk'
- 'G. B. Mohanty'
- 'E. Nakano'
- 'M. Nakao'
- 'H. Nakazawa'
- 'T. Nanut'
- 'K. J. Nath'
- 'Z. Natkaniec'
- 'M. Nayak'
- 'S. Nishida'
- 'S. Ogawa'
- 'S. Okuno'
- 'P. Pakhlov'
- 'B. Pal'
- 'C.-S. Park'
- 'S. Paul'
- 'T. K. Pedlar'
- 'L. E. Piilonen'
- 'C. Pulvermacher'
- 'J. Rauch'
- 'M. Ritter'
- 'H. Sahoo'
- 'Y. Sakai'
- 'S. Sandilya'
- 'L. Santelj'
- 'T. Sanuki'
- 'Y. Sato'
- 'V. Savinov'
- 'T. Schlüter'
- 'O. Schneider'
- 'G. Schnell'
- 'C. Schwanda'
- 'Y. Seino'
- 'D. Semmler'
- 'K. Senyo'
- 'O. Seon'
- 'I. S. Seong'
- 'V. Shebalin'
- 'C. P. Shen'
- 'T.-A. Shibata'
- 'J.-G. Shiu'
- 'F. Simon'
- 'M. Starič'
- 'T. Sumiyoshi'
- 'M. Takizawa'
- 'U. Tamponi'
- 'F. Tenchini'
- 'K. Trabelsi'
- 'M. Uchida'
- 'S. Uehara'
- 'T. Uglov'
- 'Y. Unno'
- 'S. Uno'
- 'P. Urquijo'
- 'Y. Usov'
- 'C. Van Hulse'
- 'G. Varner'
- 'K. E. Varvell'
- 'V. Vorobyev'
- 'C. H. Wang'
- 'M.-Z. Wang'
- 'M. Watanabe'
- 'Y. Watanabe'
- 'E. Widmann'
- 'J. Yamaoka'
- 'H. Ye'
- 'Y. Yook'
- 'C. Z. Yuan'
- 'Y. Yusa'
- 'Z. P. Zhang'
- 'V. Zhilich'
- 'V. Zhukova'
- 'V. Zhulanov'
- 'A. Zupanc'
title: |
\
[****]{}
---
The Standard Model (SM) of particle physics cannot explain the nature of dark matter that is understood to have mostly gravitational effects on visible matter, on radiation, and on the large-scale structure of the universe [@ref:zwicky; @ref:clowe; @ref:komatsu; @ref:blumenthal]. The dark matter can be naturally explained by the introduction of a weakly interacting particle predicted in the supersymmetric extension of the SM [@ref:susy]. The absence of observation of any supersymmetric particles in hadron collider experiments [@ref:lhcsusy] motivates studies of new classes of models, commonly referred to as dark models, which introduce new gauge symmetries [@ref:dark_photon] and predict the existence of new particles that couple weakly to SM particles. Most accelerator-based experiments have focused on the dark photon or dark particles coupling to the SM photons [@ref:belledp], though many dark models suggest a new gauge boson that could couple predominantly to quarks [@ref:bboson; @ref:tulin]. This new dark boson (hereinafter referred to as the $U^\prime$ boson, instead of $B$ as is originally proposed in Ref. [@ref:bboson], to avoid confusion with the SM $B$ meson) can be produced from light SM meson decays through $P \rightarrow U^\prime \gamma$ or $V \rightarrow U^\prime P$, where $P$ refers to a pseudoscalar meson (e.g., $\pi^0, \eta, \eta^\prime$) and $V$ to a vector meson (e.g., $\omega, \phi$). Two recent experimental limits on searches for a dark photon $A^\prime$ via $\pi^0 \to A^\prime
\gamma, A^\prime \to e^+ e^-$ [@ref:wasa] and $\phi \to A^\prime \gamma, A^\prime \to e^+ e^-$ [@ref:kloe] can be applied to the $U^\prime$ boson search in a model-dependent way to constrain the baryonic fine structure constant $\alpha_{U^\prime} \equiv g^2_{U^\prime}/(4\pi)$, where $g_{U^\prime}$ is the universal gauge coupling between the $U^\prime$ boson and the quarks [@ref:tulin]. There are also limits from $\eta \rightarrow \pi^0 \gamma \gamma$ and $\phi \rightarrow \eta \pi^0 \gamma$ decays based on their total rate, as well as from the analysis of hadronic $\Upsilon (1S)$ decays [@ref:tulin].
We search for $U^\prime$ bosons decaying to $\pi^+ \pi^-$ pairs using $\eta \rightarrow \pi^+ \pi^- \gamma$ decays, where $\eta$ is produced in the decay chain $D^{*+} \to D^0 \pi^+, D^0 \to K^0_S \eta$ [@ref:sign]. The kinematics here allows us to suppress the combinatorial background significantly. The decay $U^\prime \rightarrow \pi^+ \pi^-$ is expected to have a relatively small branching fraction of 2-4% [@ref:tulin] but nevertheless provides a very clean signature for a possible dark vector gauge boson. The dominant decay modes are $\pi^0 \gamma$ at low $U^\prime$ mass and $\pi^+\pi^-\pi^0$ at higher $U^\prime$ mass, however they suffer from higher combinatorial background and therefore are not used in the analysis. We use the decay $\eta \rightarrow \pi^+ \pi^- \pi^0$ to validate our event reconstruction by measuring the branching fraction of $\eta \rightarrow \pi^+ \pi^- \gamma$ relative to that of $\eta \rightarrow \pi^+ \pi^- \pi^0$.
The data used in this analysis were recorded at the $\Upsilon(nS)$ resonances ($n=1,\hdots,5$) and 60 MeV below the $\Upsilon(4S)$ resonance with the Belle detector [@ref:belle] at the $e^+ e^-$ asymmetric-energy collider KEKB [@ref:kekb]. The sample corresponds to an integrated luminosity of 976 fb$^{-1}$. We generated two million Monte Carlo (MC) events [@ref:MC] each for $\eta\to\pi^+ \pi^-\gamma$, $\eta\to \pi^+ \pi^-\pi^0$, and $\eta\to U^\prime \gamma\to \pi^+\pi^-\gamma$ at a particular $U^\prime$ mass selected in the range from 280 to 540 MeV/$c^2$ in steps of 10 MeV/$c^2$ (*i.e.,* 58 million events in all). The lifetime of the $U^\prime$ is assumed to be negligible. The $U^\prime$ samples are used to determine the $M(\pi^+\pi^-)$ resolution. The $U^\prime$ signal shape parameters for intermediate $U^\prime$ mass values are determined using spline interpolation.
Except for tracks from $K^0_S$ decays, we require that the charged tracks originate from the vicinity of the interaction point (IP) with impact parameters along the beam direction ($z$ axis) and perpendicular to it of less than 4 cm and 2 cm, respectively. All such charged tracks are required to have at least two associated hits in the silicon vertex detector (SVD), both in the $z$ and perpendicular directions. Such charged tracks are identified as pions or kaons by requiring that the ratio of particle identification likelihoods, $\mathcal{L}_K/
(\mathcal{L}_K+\mathcal{L}_\pi)$, constructed using information from the central drift chamber (CDC), time-of-flight scintillation counters, and aerogel threshold Cherenkov counters, be larger or smaller than 0.6, respectively. For both kaons and pions, the efficiencies and misidentification probabilities are 86% and 14%, respectively.
For photon selection, we require the energy of the candidate photon to be greater than 60 MeV (100 MeV) when the candidate photon is reconstructed in the barrel (endcap) calorimeter that covers $32^\circ < \theta < 130^\circ$ ( $12^\circ < \theta < 32^\circ$ or $130^\circ < \theta < 157^\circ$) in the polar angle $\theta$ with respect to the $+z$ axis. To reject neutral hadrons, the ratio of the energy deposited by a photon candidate in the $3\times 3$ and $5 \times 5$ calorimeter arrays centered on the crystal with the largest signal is required to exceed 0.85.
Candidate $\pi^0$ mesons are reconstructed from pairs of $\gamma$ candidates; we require $M_{\gamma \gamma}$ $\in [120,150]$ MeV$/c^2$ and refit $\gamma$ momenta with the $\pi^0$ mass constraint.
Candidate $K^0_S \to \pi^+ \pi^-$ mesons are reconstructed from two tracks, assumed to be pions, using a neural network (NN) technique [@Feindt:2006pm] that uses the following information: the $K^0_S$ momentum in the laboratory frame; the distance along $z$ between the two track helices at their closest approach; the $K_S^0$ flight length in the transverse plane; the angle between the $K^0_S$ momentum and the vector joining the $K^0_S$ decay vertex to the IP; the angles between the pion momenta and the laboratory-frame direction in the $K^0_S$ rest frame; the distances of closest approach in the transverse plane between the IP and the two pion helices; and the pion hit information in the SVD and CDC. We also require that the $\pi^+ \pi^-$ invariant mass be within $\pm 9$ MeV/$c^2$ (about 3$\sigma$ in resolution [@ref:wonbr]) of the nominal $K^0_S$ mass [@PDG].
For the $\eta \rightarrow \pi^+\pi^- \gamma$ candidates, we require that the photon not be associated with a $\pi^0$ candidate and its transverse momentum be greater than 200 MeV/$c$ to remove $D^{*+} \rightarrow D^+ (\rightarrow K^0_S \pi^+ \pi^- \pi^+) \gamma$ background. For both $\eta \to \pi^+ \pi^- \gamma$ and $\eta \to \pi^+ \pi^- \pi^0$ candidates, we perform a vertex fit with the two charged pions and require the reduced $\chi^2$ to be less than 10. The efficiency of this requirement is 94%. We require the reconstructed mass of each $\eta$ candidate to be in the range \[500,600\] MeV/$c^2$ and refit momenta of its daughters with the constraint of the nominal $\eta$ mass.
Combinations of a $K^0_S$ candidate and $\eta$ candidate are fit to a common vertex and their invariant mass is required to be within $\pm40$ MeV/$c^2$ of the nominal $D^0$ mass. The $D^0$ and $\pi^+$ combinations are fitted to the IP, and the mass difference $\Delta
M_{D^*}=M(K^0_S\eta\pi^+)-M(K^0_S\eta)$ is required to satisfy $\Delta M_{D^*} \in [143,148]$ MeV/$c^2$. To remove the combinatorial background, the momentum of the $D^{*+}$ candidates, measured in the center-of-mass system, is required to be greater than 2.5, 2.6, and 3.0 GeV/$c$ for the data taken below, at, and above the $\Upsilon(4S)$ resonance, respectively. Figure \[fig:data\_mass\] shows the invariant mass of the $K^0_S\eta$ combinations (left) and the mass difference (right) for $\eta \to \pi^+ \pi^- \gamma$ decays after applying all selection criteria described above, except the mass requirements themselves. Figure \[fig:data\_eta\_B\_gauss\] shows the invariant mass of the $\pi^+\pi^-\gamma$ combinations after all requirements. There are clear peaks of signal events in all distributions; the increase of the background at low masses in the $M(\pi^+\pi^-\gamma)$ distribution is due to the feed-down from the $\eta\to\pi^+\pi^-\pi^0$ decays when a photon from $\pi^0$ is not reconstructed.
![ Invariant mass of the $K^0_S\eta$ combinations (left) and the $D^*$–$D^0$ mass difference (right) for $\eta \to \pi^+ \pi^- \gamma$ decays. []{data-label="fig:data_mass"}](data_mass_gamma.eps){width="50.00000%"}
To extract the signal yield, we perform a binned maximum likelihood fit to the $M(\pi^+\pi^-\gamma)$ distribution. The fit function is the sum of the signal, the combinatorial background and the feed-down background components. The signal probability density function (PDF) is the sum of a Gaussian and a bifurcated Gaussian with the ratios of widths fixed from the MC simulation. A linear function is used for the combinatorial background PDF. The feed-down contribution is described by a Gaussian with shape parameters fixed from the MC simulation. The confidence level (p-value) of the fit is 12% and the $\eta\to\pi^+\pi^-\gamma$ signal yield is $N_{\eta} = 2974 \pm 90$ events. The feed-down yield agrees well with the expectation.
As a cross-check, we measure the ratio of branching fractions ${\mathcal{B}(\eta \rightarrow \pi^+\pi^- \gamma)}/
{\mathcal{B}(\eta \rightarrow \pi^+\pi^- \pi^0)}$. The fit to the $\pi^+\pi^-\pi^0$ invariant mass distribution is similar to the one described above, except that the combinatorial background is described by a second-order polynomial and there is no feed-down background. The reconstruction efficiencies, determined from the MC simulation, are $\varepsilon (\pi^+\pi^-\gamma) = $ 5.1% and $\varepsilon (\pi^+\pi^-\pi^0) = $ 4.8%. The measured ratio of branching fractions, $0.185\pm0.007$, where the uncertainty is statistical only, is in good agreement with the world-average value of $0.184\pm0.004$ [@PDG].
We define the $\eta$ signal region as $M(\pi^+\pi^-\gamma) \in [535.5,560.5]$ MeV/$c^2$, and the sideband regions used for background subtraction as $M(\pi^+\pi^-\gamma) \in [520.0,532.5]$ or $[563.5,576.0]$ MeV/$c^2$. The $M(\pi^+\pi^-)$ distribution for the background-subtracted $\eta$ signal is shown in Fig. \[fig:data\_B\].
To describe the $M(\pi^+\pi^-)$ distribution, we use an expression of the differential decay rate based on low-energy quantum chromodynamics (QCD) phenomenology [@ref:stollenwerk; @ref:adlarson] using a combination of chiral perturbation theory and dispersive analysis, $$\begin{aligned}
\frac{d\Gamma}{d s}
\propto
|P(s) F_V(s)|^2 (m_\eta^2 - s)^3 s (1-4 m^2_\pi/s)^{3/2},
\label{eq:dgdx}\end{aligned}$$ where $s\equiv M(\pi^+\pi^-)^2$, $P(s)$ is a reaction-specific perturbative part, and $F_V(s)$ is the pion vector form factor. We use $|P(s)| = 1+ (1.89\pm 0.64) s$ [@ref:adlarson] and $|F_V(s)| = 1+ (2.12\pm0.01) s + (2.13\pm 0.01) s^2
+ (13.80\pm 0.14) s^3$ [@ref:stollenwerk] ($s$ in GeV$^2/c^4$). The numerical values and the uncertainties of the expansion coefficients of $|P(s)|$ and $|F_V(s)|$ are taken from fits to data of $\eta^{(\prime)} \to \pi^+ \pi^- \gamma$ decays. We multiply the $d\Gamma/ds$ expression from Eq. (\[eq:dgdx\]) by the reconstruction efficiency. The efficiency as a function of $M(\pi^+ \pi^-)$ is approximately flat but drops to zero at the kinematic limit of $m_\eta$. The fit results are presented in Fig. \[fig:data\_B\]. Equation (\[eq:dgdx\]) describes the $M(\pi^+\pi^-)$ distribution well, and the confidence level of the fit is 95%. We add the $U^\prime$ signal to the above fit function and perform fits while fixing the $U^\prime$ mass at a value between 290 and 520 MeV/$c^2$ in steps of 1 MeV/$c^2$. The $U^\prime$ signal is described by the sum of two Gaussians. The signal resolution of the core Gaussian is about 1 MeV/$c^2$ near the $2 m_\pi$ threshold and 2 MeV/$c^2$ at the $m_\eta$ kinematic limit. An example of the $U^\prime$ signal with the mass of 400 MeV/$c^2$ and arbitrary normalization is shown in Fig. \[fig:data\_B\]. We do not find a significant $U^\prime$ signal at any mass value. The typical uncertainty in the $U^\prime$ yield $N_{U^\prime}$ is $\mathcal{O}(1-10)$ events.
We express the baryonic fine structure constant $\alpha_{U^\prime}$ using the equation for the partial width ratio $\Gamma(\eta \rightarrow U^\prime \gamma)/\Gamma(\eta \rightarrow
\gamma \gamma)$ from Ref. [@ref:tulin] as: $$\begin{aligned}
\alpha_{U^\prime} &=&
\Bigg[
\frac{\alpha}{2}
\Bigg(
1-\frac{m^2_{U^\prime}}{m^2_\eta}
\Bigg)^{-3}
\Bigg|
\mathcal{F}(m^2_{U^\prime})
\Bigg|^{-2}
\frac{1}{\mathcal{B}(U^\prime \rightarrow \pi^+ \pi^-)}
\Bigg]
\nonumber \\
&\times&
\Bigg[
\frac{\Gamma(\eta \rightarrow \pi^+ \pi^- \gamma)}{\Gamma(\eta \rightarrow \gamma \gamma)}
\Bigg]
\nonumber \\
&\times&
\Bigg[
\frac{\Gamma(\eta \rightarrow U^\prime \gamma \rightarrow \pi^+ \pi^- \gamma)}
{\Gamma(\eta \rightarrow \pi^+ \pi^- \gamma)}
\Bigg],
\label{eq:alpha_B}\end{aligned}$$ where $\alpha$ is the electromagnetic fine structure constant. The first factor in Eq. (\[eq:alpha\_B\]), which is purely theoretical, contains the phase space, the form factor $\mathcal{F}(m^2_{U^\prime})$, and the branching fraction of $U^\prime \rightarrow \pi^+ \pi^-$ decay. The branching fraction is about 2-4%, as computed from formulae provided in Ref. [@ref:tulin] and references therein. The second factor is obtained from the latest measurements [@PDG]. The third factor is determined from the $\eta$ and $U^\prime$ yields and reconstruction efficiencies ${(N_{U^\prime}/\varepsilon(\eta \to U^\prime \gamma \to
\pi^+ \pi^- \gamma))}/{(N_{\eta}/\varepsilon(\eta \to \pi^+ \pi^- \gamma))}$.
To estimate the systematic uncertainties in the $\eta\to \pi^+\pi^-\gamma$ and $\eta\to U^\prime \gamma\to\pi^+\pi^-\gamma$ yields, we change the parameterization of the combinatorial background in the $M(\pi^+ \pi^- \gamma)$ fit from a first- to a second-order polynomial and account for the background non-linearity while subtracting the sidebands. The change in the $\eta$ yield is at the 1% level, while the change in the $U^\prime$ yield is negligible. The systematic effect due to the uncertainties of the expansion coefficients in $|P(s)|$ and $|F_V(s)|$ is negligible in the $U^\prime$ yield. The systematic uncertainty in the ratio of the reconstruction efficiencies $\varepsilon(\eta\to U^\prime \gamma\to\pi^+\pi^-\gamma)/
\varepsilon(\eta\to\pi^+\pi^-\gamma)$ is conservatively estimated to be 4% (1% per track and 3% per photon). The total systematic uncertainties are estimated by adding the above contributions in quadrature.
Using Eq. (\[eq:alpha\_B\]), we set a 95% confidence level upper limits on $\alpha_{U^\prime}$ using the Feldman-Cousins approach [@ref:fc], adding the statistical and systematic uncertainties in quadrature. The upper limit as a function of the $U^\prime$ boson mass is shown in Fig. \[fig:data\_limit\]. Considering other results in this mass region, we find that our limit is stronger than that from a model-dependent analysis [@ref:tulin] of the $\phi\to e^+e^-\gamma$ decays [@ref:kloe] for $m_{U^\prime}>450$ MeV/$c^2$, but weaker than the limit based on the $\eta\to\pi^0\gamma\gamma$ total rate [@ref:tulin]. During preparation of this manuscript, we learned that the data set in Ref. [@ref:babusci] contains many more $\eta \rightarrow \pi^+ \pi^- \gamma$ decays and can provide a more stringent limit on $\alpha_{U^\prime}$ in future.
![ Computed 95% upper limit on the baryonic fine structure constant $\alpha_{U^\prime}$ as a function of the unknown $U^\prime$ mass (solid curve). []{data-label="fig:data_limit"}](data_limit_zoom.eps){width="40.00000%"}
To conclude, we perform a search for a dark vector gauge boson $U^\prime$ that couples to quarks [@ref:tulin], using the decay chain $D^{*+} \to D^0 \pi^+, D^0 \to
K^0_S \eta, \eta \to U^\prime \gamma, U^\prime \to \pi^+ \pi^-$. Our results limit the baryonic fine structure constant $\alpha_{U^\prime}$ to below $10^{-3}-10^{-2}$ at 95% confidence level over the $U^\prime$ mass range 290 to 520 MeV/$c^2$. This is the first search for $U^\prime$ in the $\pi^+\pi^-$ mode. We find that our limit is stronger than that from a model-dependent analysis [@ref:tulin] of the $\phi\to e^+e^-\gamma$ decays [@ref:kloe] for $m_{U^\prime}>450$ MeV/$c^2$, but weaker than the limit based on the $\eta\to\pi^0\gamma\gamma$ total rate [@ref:tulin].\
We thank the KEKB group for the excellent operation of the accelerator; the KEK cryogenics group for the efficient operation of the solenoid; and the KEK computer group, the National Institute of Informatics, and the PNNL/EMSL computing group for valuable computing and SINET4 network support. We acknowledge support from the Ministry of Education, Culture, Sports, Science, and Technology (MEXT) of Japan, the Japan Society for the Promotion of Science (JSPS), and the Tau-Lepton Physics Research Center of Nagoya University; the Australian Research Council; Austrian Science Fund under Grant No. P 22742-N16 and P 26794-N20; the National Natural Science Foundation of China under Contracts No. 10575109, No. 10775142, No. 10875115, No. 11175187, No. 11475187 and No. 11575017; the Chinese Academy of Science Center for Excellence in Particle Physics; the Ministry of Education, Youth and Sports of the Czech Republic under Contract No. LG14034; the Carl Zeiss Foundation, the Deutsche Forschungsgemeinschaft, the Excellence Cluster Universe, and the VolkswagenStiftung; the Department of Science and Technology of India; the Istituto Nazionale di Fisica Nucleare of Italy; the WCU program of the Ministry of Education, National Research Foundation (NRF) of Korea Grants No. 2011-0029457, No. 2012-0008143, No. 2012R1A1A2008330, No. 2013R1A1A3007772, No. 2014R1A2A2A01005286, No. 2014R1A2A2A01002734, No. 2015R1A2A2A01003280 , No. 2015H1A2A1033649; the Basic Research Lab program under NRF Grant No. KRF-2011-0020333, Center for Korean J-PARC Users, No. NRF-2013K1A3A7A06056592; the Brain Korea 21-Plus program and Radiation Science Research Institute; the Polish Ministry of Science and Higher Education and the National Science Center; the Ministry of Education and Science of the Russian Federation and the Russian Foundation for Basic Research; the Slovenian Research Agency; Ikerbasque, Basque Foundation for Science and the Euskal Herriko Unibertsitatea (UPV/EHU) under program UFI 11/55 (Spain); the Swiss National Science Foundation; the Ministry of Education and the Ministry of Science and Technology of Taiwan; and the U.S. Department of Energy and the National Science Foundation. This work is supported by a Grant-in-Aid from MEXT for Science Research in a Priority Area (“New Development of Flavor Physics”) and from JSPS for Creative Scientific Research (“Evolution of Tau-lepton Physics”). EW acknowledges partially NRF grant of Korea Grants No. NRF-2011-0030865 and Korea University Future Research Grant, and thanks B. R. Ko for his suggestion on an inclusive version of this analysis.
[99]{}
F. Zwicky, Astrophys. J. [**86**]{}, 217 (1937).
D. Clowe, A. Gonzalez, and M. Markevitch, Astrophys. J. [**604**]{}, 596 (2004).
E. Komatsu [*et al*]{}., Astrophys. J. Suppl. [**180**]{}, 330 (2009).
G. R. Blumenthal [*et al*]{}., Nature [**311**]{}, 517 (1984).
B. Kane and M. Shifman, [*The Supersymmetric World*]{}, World Scientific, Singapore (2000).
G. Aad [*et al*]{}. (ATLAS Collaboration), J. High Energy Phys. [**10**]{}, 134 (2015); S. Chatrchyan [*et al*]{}. (CMS Collaboration), Phys. Rev. D [**88**]{}, 052017 (2013).
B. Holdom, Phys. Lett. B [**166**]{}, 196 (1986); P. Fayet, Phys. Rev. D [**75**]{}, 115017 (2007).
G. Agakishiev [*et al*]{}. (HADES Collaboration), Phys. Lett. B [**731**]{}, 265 (2014); I. Jaegle [*et al*]{}. (Belle Collaboration), Phys. Rev. Lett. [**114**]{}, 211801 (2015); J. P. Lees [*et al*]{}. (BaBar Collaboration), Phys. Rev. Lett. [**108**]{}, 211801 (2012); S. Giovannella [*et al*]{}. (KLOE Collaboration), J. Phys. Conf. Ser. [**335**]{}, 012067 (2011).
A. E. Nelson and N. Tetradis, Phys. Lett. B [**221**]{}, 80 (1989).
S. Tulin, Phys. Rev. D [**89**]{}, 114008 (2014).
P. Adlarson [*et al*]{}. (WASA-at-COSY Collaboration), Phys. Lett. B [**726**]{}, 187 (2013).
D. Babusci [*et al*]{}. (KLOE-2 Collaboration), Phys. Lett. B [**720**]{}, 111 (2013).
Throughout this paper, inclusion of the charge-conjugate decay is implied unless stated otherwise.
A. Abashian [*et al.*]{} (Belle Collaboration), Nucl. Instrum. Meth. Phys. Res., Sect. A [**479**]{}, 117 (2002); also see the detector section in J. Brodzicka [*et al.*]{}, Prog. Theor. Exp. Phys. [**2012**]{}, 04D001 (2012).
S. Kurokawa and E. Kikutani, Nucl. Instrum. Meth. Phys. Res., Sect. A [**499**]{}, 1 (2003), and other papers included in this volume; T. Abe [*et al.*]{}, Prog. Theor. Exp. Phys. [**2013**]{}, 03A001 (2013) and following articles up to 03A011.
$e^+e^- \rightarrow c\bar{c}$ events are generated with [Pythia]{} (T. Sjöstrand, Comput. Phys. Commun. [**135**]{}, 238 (2001)); particle decay is simulated with [EvtGen]{} (D. J. Lange, Nucl. Instrum. Methods Phys. Res., Sect. A [**462**]{}, 152 (2001)); the detector response is simulated with [Geant]{} 3.21 (R. Brun [*et al.*]{} GEANT 3.21, CERN Report DD/EE/84-1, 1984).
M. Feindt and U. Kerzel, Nucl. Instrum. Methods Phys. Res., Sect. A [**559**]{}, 190 (2006). E. Won [*et al*]{}. (Belle Collaboration), Phys. Rev. D [**80**]{}, 111101(R) (2009).
K. A. Olive [*et al.*]{} (Particle Data Group), Chin. Phys. C [**38**]{}, 090001 (2014).
P. Adlarson [*et al*]{}. (WASA-at-COSY Collaboration), Phys. Lett. B [**707**]{}, 243 (2012).
F. Stollenwerk [*et al*]{}., Phys. Lett. B [**707**]{}, 184 (2012).
G. J. Feldman and R. D. Cousins, Phys. Rev. D [**57**]{}, 3873 (1998).
D. Babusci [*et al*]{}. (KLOE/KLOE-2 Collaboration), Phys. Lett. B [**718**]{}, 910 (2013).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the weight and length of the minimum mean-weight cycle in the stochastic mean-field distance model, i.e., in the complete graph on [$\smash{n}$]{} vertices with edges weighted by independent exponential random variables. Mathieu and Wilson showed that the minimum mean-weight cycle exhibits one of two distinct behaviors, according to whether its mean weight is smaller or larger than [$\smash{1/(ne)}$]{}; and that both scenarios occur with positive probability in the limit [$\smash{n\to\infty}$]{}. If the mean weight is [$\smash{< 1/(ne)}$]{}, the length is of constant order. If the mean weight is [$\smash{> 1/(ne)}$]{}, it is concentrated just above [$\smash{1/(n{e})}$]{}, and the length diverges with [$\smash{n}$]{}. The analysis of Mathieu–Wilson gives a detailed characterization of the subcritical regime, including the (non-degenerate) limiting distributions of the weight and length, but leaves open the supercritical behavior. We determine the asymptotics for the supercritical regime, showing that with high probability, the minimum mean weight is [$\smash{(n{e})^{-1}[1 + \pi^2/(2 \log^2 n) + O((\log n)^{-3})]}$]{}, and the cycle achieving this minimum has length on the order of [$\smash{(\log n)^3}$]{}.'
address:
- |
Statistics Department\
University of Chicago\
Chicago, IL 60637
- |
Microsoft Research\
One Memorial Drive\
Cambridge, MA 02142 –and– Mathematics Department, Massachusetts Institute of Technology, Cambridge, MA 02139
- |
Microsoft Research\
One Microsoft Way\
Redmond, WA 98052
author:
- 'Jian Ding$^*$'
- 'Nike Sun$^\dagger$'
- 'David B. Wilson'
bibliography:
- 'cycles.bib'
title: 'Supercritical minimum mean-weight cycles'
---
[^1]
Introduction
============
Given a directed or undirected graph with edge weights, a *minimum mean-weight cycle* (<span style="font-variant:small-caps;">mmwc</span>) is any cycle that minimizes the mean weight (ratio of total weight to cycle length) over all cycles in the graph. Finding an <span style="font-variant:small-caps;">mmwc</span> is a fundamental subproblem to a wide variety of algorithms: for example, they have been used ([@goldberg-tarjan; @radzik-goldberg], building on [@klein]) to give a strongly polynomial-time algorithm for the minimum-cost circulation problem, of which the maximum flow problem is a special case. Other applications of <span style="font-variant:small-caps;">mmwc</span>’s include algorithms for multicommodity flow [@ouorou-mahey] and asymmetric travelling salesman tours [@kleinberg-williamson]; for further applications see the extensive discussion in [@dasdan].
Several <span style="font-variant:small-caps;">mmwc</span> algorithms are available (see [@dasdan; @georgiadis] and references therein); and it is of practical interest to understand their runtime in “average-case” settings, that is to say, on random inputs. Experimental studies [@DG98; @DG99; @dasdan; @georgiadis] suggest that an algorithm due to Young–Tarjan–Orlin [@YTO] (based on an improvement of a parametric shortest-path algorithm [@KO81]) has the best runtime in standard random graph ensembles, where it substantially outperforms its worst-case theoretical guarantees.
Motivated by the empirical studies, a natural direction is to understand the typical behavior of the <span style="font-variant:small-caps;">mmwc</span> in random graphs. With this in mind, Mathieu–Wilson [@mathieu-wilson] study the <span style="font-variant:small-caps;">mmwc</span> in the stochastic mean-field distance model: the complete graph, or complete digraph, with i.i.d. random edge weights. This is a canonical setting for the study of combinatorial optimization problems, both in the mathematics and physics literature: other examples include minimum spanning tree [@MR770868; @MR1054012], shortest path [@MR770869; @MR1723648; @MR2309622; @MR2433939], traveling salesman [@mezard-parisi-tsp; @MR2104159; @MR2600434], assignment [@mezard-parisi-replicas-opt; @MR1839499; @MR2036492; @MR2178256], spanners [@MR2551020], and Steiner tree [@MR2071332; @MR2927630]. The analysis of [@mathieu-wilson] uncovers an unusual dichotomy for the <span style="font-variant:small-caps;">mmwc</span> in the stochastic mean-field distance model, and characterizes the subcritical regime. In this work we complement their analysis by characterizing the supercritical regime.
Main result
-----------
Consider the stochastic mean-field distance model where each edge is independently weighted by an exponential random variable of unit rate. In this setting we study the minimum mean-weight cycle, which is unique with probability one. Denote its mean weight by [$\smash{{\mathscr{W}}_n}$]{}, and its length by [$\smash{{\mathscr{L}}_n}$]{}, so the cycle has total weight [$\smash{{\mathscr{L}}_n{\mathscr{W}}_n}$]{}. This model exhibits an unusual phase transition [@mathieu-wilson] at [$\smash{n{\mathscr{W}}_n=1/e}$]{}: In the *subcritical* regime [$\smash{n{\mathscr{W}}_n<1/e}$]{}, [$\smash{n{\mathscr{W}}_n}$]{} is non-concentrated, and [$\smash{{\mathscr{L}}_n}$]{} stays bounded; moreover the limiting distributions of [$\smash{n{\mathscr{W}}_n}$]{} and [$\smash{{\mathscr{L}}_n}$]{} are precisely characterized. In the *supercritical* regime [$\smash{n{\mathscr{W}}_n>1/e}$]{}, [$\smash{n{\mathscr{W}}_n}$]{} is very concentrated with [$\smash{n{\mathscr{W}}_n-1/e\to0}$]{} in probability, while [$\smash{{\mathscr{L}}_n}$]{} diverges with [$\smash{n}$]{}, at least on the order of [$\smash{(\log n)^2\log\log n}$]{}. The conclusions of [@mathieu-wilson] are much less precise in the supercritical case, leaving open the asymptotic order of [$\smash{n{\mathscr{W}}_n-1/e}$]{} and [$\smash{{\mathscr{L}}_n}$]{}.
In this paper we characterize the asymptotics of [$\smash{{\mathscr{W}}_n}$]{} and [$\smash{{\mathscr{L}}_n}$]{} in the supercritical regime. In particular, our results confirm simulations done by Uri Zwick and the last author which suggested that [$\smash{{\mathscr{L}}_n\asymp(\log n)^3}$]{}. Our main theorem is as follows:
\[t:main\] In the complete graph or complete digraph with i.i.d. unit-rate exponential edge weights, let [$\smash{{\mathscr{W}}_n}$]{} and [$\smash{{\mathscr{L}}_n}$]{} be the mean weight and length of the minimum mean-weight cycle, and write c\_(n) e\^[-1]{}\[1+\^2/\[2(n)\^2\]\]. For all [$\smash{\ep>0}$]{} there exists a constant [$\smash{C=C(\ep)>0}$]{} such that \[e:main.len.wt\] \_[n]{} ¶( .
[c]{} |n\_n-c\_(n)|(n)\^3 C\
1/C \_n/(n)\^3C\
| n\_n >[e]{}\^[-1]{} ) 1-. These bounds are optimal in the sense that for any interval [$\smash{I}$]{} with length [$\smash{|I|\leq 1/C}$]{}, \[e:main.nondegenerate\] \_[n]{} ¶( .
[c]{} \[n\_n-c\_(n)\](n)\^3 I\
\_n/(n)\^3 I
| n\_n >[e]{}\^[-1]{} ) .
Theorem \[t:main\] implies that in the supercritical regime [$\smash{n{\mathscr{W}}_n>1/e}$]{}, the random variables [$\smash{\log[{\mathscr{L}}_n/(\log n)^3]}$]{} and [$\smash{(\log n)^3[n{\mathscr{W}}_n-c_\star]}$]{} are tight, with non-degenerate distributions.
Proof ideas
-----------
In the remainder of this introductory section we highlight some of the main new ideas in our proof, which allow us to overcome obstacles in the analysis of [@mathieu-wilson]. We remark that very similar obstacles arose in an analysis [@ding] for a related model, *percolation of averages*, which asks for the longest path with mean weight below a parameter [$\smash{\lm}$]{}. While our work is carried out in the context of the <span style="font-variant:small-caps;">mmwc</span>, we expect our methods may be applied to improve results in [@ding].
In order to show that the supercritical <span style="font-variant:small-caps;">mmwc</span> has mean weight [$\smash{{\mathscr{W}}_n}$]{} in some interval [$\smash{[a,b]}$]{} and satisfies some property <span style="font-variant:small-caps;">p</span>, one must show (i) there are no cycles with mean weight in [$\smash{[1/(ne),a]}$]{}, (ii) there are no cycles violating <span style="font-variant:small-caps;">p</span> in [$\smash{[a,b]}$]{}, and (iii) there is at least one cycle (satisfying <span style="font-variant:small-caps;">p</span>) in [$\smash{[a,b]}$]{}. In particular, (ii) is usually done by first moment arguments, and is clearly easier for less restrictive properties <span style="font-variant:small-caps;">p</span>. On the other hand, a natural approach for (iii) is the second moment method, for which it is often advantageous to make <span style="font-variant:small-caps;">p</span> more restrictive.
Indeed, it is demonstrated in [@ding; @mathieu-wilson] that a straightforward second moment method on the number of cycles fails, due to an excessive contribution from atypically light cycles (or paths) — conditioned on finding one atypically light cycle [$\smash{{\mathscr{C}}}$]{}, it is very likely to have a large number of light cycles overlapping with [$\smash{{\mathscr{C}}}$]{}. To address this issue, these works developed a notion of *uniformity*: for a [$\smash{k}$]{}-cycle with edge weights [$\smash{w_1,\ldots,w_k}$]{} summing to [$\smash{k{\bar{c}}/n}$]{}, we say the cycle is [$\smash{A}$]{}-uniform if the process [$\smash{X}$]{} with increments [$\smash{X_{i+1}-X_i= (nw_i/{\bar{c}}-1)}$]{} has range at most [$\smash{A}$]{} — meaning the cycle has no excessively light subpath. The typical <span style="font-variant:small-caps;">mmwc</span> is uniform with high probability; but on the other hand the count of uniform cycles has small enough variance for the second moment method to go through.
However, this uniformity property is not sufficiently restrictive to yield accurate implications on the asymptotics of [$\smash{{\mathscr{W}}_n,{\mathscr{L}}_n}$]{}. Indeed, it is clear that the precision achievable for the window [$\smash{[a,b]}$]{} is dictated by the accuracy with which <span style="font-variant:small-caps;">p</span> captures the typical properties of the <span style="font-variant:small-caps;">mmwc</span>. In this work, we restrict further to uniform cycles with a *typical profile*. We defer the formal definitions to Section \[s:second.moment\]; roughly speaking, we restrict to cycles for which the associated process [$\smash{X}$]{} not only has range [$\smash{\le A}$]{}, but furthermore has *typical local times* within its range. The analysis of Sections \[s:rw\] and \[s:atypical\] will show that this restricted property is satisfied with high probability. On the other hand, in Section \[s:second.moment\] we show that the restriction captures the remaining variance, so that the number of cycles restricted in this manner is well-concentrated about its mean. Theorem \[t:main\] follows as a consequence.
A key technical ingredient in our proof is a collection of precise estimates for *exp-minus-one* random walks (that is, a random walk with increments distributed as [$\smash{E-1}$]{}, with [$\smash{E}$]{} a unit-rate exponential variable) conditioned to have restricted range. While our results are reminiscent of analogous estimates for simple random walk or Brownian motion, for general random walks under suitable moment assumptions there is no general theory yielding estimates to the level of accuracy needed for our <span style="font-variant:small-caps;">mmwc</span> analysis. Our estimates for the exp-minus-one walk are derived in Section \[s:rw\]. A crucial input to these estimates is a precise characterization of the principal eigenvalue and eigenfunction for the exp-minus-one walk in an interval with absorbing boundaries. This analysis may be of independent interest, and is presented in Section \[s:eig\].
Acknowledgements {#acknowledgements .unnumbered}
----------------
Computer experiments on the mean-field stochastic <span style="font-variant:small-caps;">mmwc</span> performed by Uri Zwick with D.B.W. provided valuable intuition at an early stage of this project. J.D. and N.S. thank the MSR Redmond Theory Group for its hospitality.
Preliminaries
=============
Notation {#notation .unnumbered}
--------
We write [$\smash{f_n \lesssim g_n}$]{} (or [$\smash{g_n \gtrsim f_n}$]{}) if there exists an absolute constant [$\smash{C>0}$]{} such that [$\smash{f_n \leq Cg_n}$]{} for all [$\smash{n\in \mathbb N}$]{}. We write [$\smash{\asymp}$]{} to indicate that [$\smash{\lesssim}$]{} and [$\smash{\gtrsim}$]{} both hold. For numerous parameters in the paper, we write for example [$\smash{C = C(\Delta, \ep)}$]{} to indicate that [$\smash{C>0}$]{} is a number depending only on [$\smash{\Delta}$]{} and [$\smash{\ep}$]{}.
Cycles and paths
----------------
For a cycle [$\smash{{\mathscr{C}}}$]{} we write [$\smash{\operatorname{len}({\mathscr{C}})}$]{} for the length (number of edges) of [$\smash{{\mathscr{C}}}$]{}; when [$\smash{\operatorname{len}({\mathscr{C}})=k}$]{} we refer to [$\smash{{\mathscr{C}}}$]{} as a [$\smash{k}$]{}-cycle. We write [$\smash{\operatorname{wgt}({\mathscr{C}})}$]{} for its total weight, and $$\label{e:c.bar}
{\bar{c}}({\mathscr{C}})\equiv n\operatorname{wgt}({\mathscr{C}})/\operatorname{len}({\mathscr{C}})$$ for its mean weight scaled by [$\smash{n}$]{}; we say [$\smash{{\mathscr{C}}}$]{} is *[$\smash{c}$]{}-light* if [$\smash{{\bar{c}}({\mathscr{C}})\le c}$]{}. We apply these terms to paths as well as cycles; note that a [$\smash{k}$]{}-cycle involves [$\smash{k}$]{} vertices while a [$\smash{k}$]{}-path involves [$\smash{k+1}$]{} vertices. To treat both undirected and directed random networks in a fairly unified manner, we will always take cycles and paths to be directed. In our random network, the weight of any given [$\smash{k}$]{}-cycle or [$\smash{k}$]{}-path is distributed as the sum of [$\smash{k}$]{} independent unit-rate exponential random variables: that is to say, a gamma random variable with shape parameter [$\smash{k}$]{}, with probability density $$\label{e:gamma.density}
f_k(x) = \f{e^{-x} x^{k-1} }{ (k-1)!},
\quad x\ge0.$$ We abbreviate this distribution as [$\smash{\text{Gam}(k)}$]{}. For the sake of review, we repeat the following calculation from [@mathieu-wilson]:
\[l:basic.first.moment\] Let [$\smash{Z^k_c}$]{} count all [$\smash{c}$]{}-light [$\smash{k}$]{}-cycles ([$\smash{k\ge2}$]{}), and let [$\smash{{\bar{Z}}^k_c}$]{} count all [$\smash{c}$]{}-light [$\smash{k}$]{}-paths ([$\smash{k\ge1}$]{}). For all [$\smash{c\lesssim1}$]{} and [$\smash{0<\delta<1}$]{} we have $$\begin{aligned}
\E [Z^k_c-Z^k_{c(1-\delta)}]
&\asymp \f{(n)_k}{n^k} \f{(ce)^k[1-(1-\delta)^k]}{k^{3/2}}
\le \f{(ce)^k}{k^{3/2}},\\
\E {\bar{Z}}^k_c
&\asymp \f{(n)_{k+1}}{n^k} \f{(ce)^k}{k^{1/2}}
\le n\f{(ce)^k}{k^{1/2}}.
\end{aligned}$$
In the complete graph on [$\smash{n}$]{} vertices, the number of (directed) [$\smash{k}$]{}-cycles ([$\smash{k\ge2}$]{}) is [$\smash{(n)_k/k}$]{}. The weight of any given [$\smash{k}$]{}-cycle is distributed as a [$\smash{\text{Gam}(k)}$]{} random variable, therefore $$\E Z^k_c
= \f{(n)_k}{k}
\P( \text{Gam}(k) \le ck/n )
= \f{(n)_k}{k}
\int_0^{ck/n}
\f{e^{-x} x^{k-1}}{(k-1)!} \,dx.$$ Since [$\smash{c\lesssim1}$]{} we have [$\smash{e^{-x}\asymp1}$]{} uniformly over the range of integration, therefore $$\begin{aligned}
\E [Z^k_c-Z^k_{c(1-\delta)}]
&\asymp
\f{(n)_k}{k}
\int_{c(1-\delta)k/n}^{ck/n}
\f{x^{k-1}}{(k-1)!} \,dx
\asymp
\f{(n)_k}{k}
\f{(ck/n)^k[1-(1-\delta)^k]}{k!}\\
&\asymp
\f{(n)_k}{n^k}
\f{(ce)^k[1-(1-\delta)^k]}{k^{3/2}}
\le \f{(ce)^k[1-(1-\delta)^k]}{k^{3/2}},
\end{aligned}$$ proving the estimate for [$\smash{c}$]{}-light [$\smash{k}$]{}-cycles. The estimate for [$\smash{c}$]{}-light [$\smash{k}$]{}-paths follows by noting that the number of [$\smash{k}$]{}-paths ([$\smash{k\ge1}$]{}) is [$\smash{(n)_{k+1}}$]{}.
An easy variation of the preceding calculation shows that in the targeted regime for [$\smash{{\mathscr{W}}_n}$]{}, there can be no cycles of length less than [$\smash{(\log n)^2}$]{}:
\[l:log.squared\] Given any [$\smash{c = 1/e[1 + O(1/(\log n)^2)]}$]{} and [$\smash{C>0}$]{}, the probability that there is a cycle [$\smash{{\mathscr{C}}}$]{} with [$\smash{1/e <{\bar{c}}({\mathscr{C}}) \le c}$]{} and [$\smash{\operatorname{len}({\mathscr{C}}) < C (\log n)^2}$]{} tends to zero in the limit [$\smash{n\to\infty}$]{}.
The expected number of directed [$\smash{k}$]{}-cycles that are [$\smash{c}$]{}-light but not [$\smash{(1/e)}$]{}-light is $$\E (Z^k_c-Z^k_{1/e})
=
\f{(n)_k}{k}
\int_{k/(en)}^{ck/n} \f{e^{-x}x^{k-1}}{(k-1)!} \,dx
\lesssim \f{(ce)^k-1}{k^{3/2}}
\lesssim \f{k/(\log n)^2}{k^{3/2}},$$ where the last bound used that [$\smash{k\lesssim(\log n)^2}$]{}. Summing over [$\smash{2\le k<(\log n)^2}$]{} and applying Markov’s inequality proves the claim.
Uniformity
----------
It was previously demonstrated in [@ding; @mathieu-wilson] that cycles which are uniform (in a sense defined formally below) have low variance. For a sequence of weights [$\smash{w_1,\ldots,w_k}$]{} summing to [$\smash{k{\bar{c}}/n}$]{}, let us define its *excedance relative to [$\smash{c}$]{}* (for short, *[$\smash{c}$]{}-excedance*) to be the quantity \[e:excedance\] \_i (nw\_i/c-1) = k([|[c]{}]{}/c-1).
\[d:unif\] Say that a cycle [$\smash{{\mathscr{C}}}$]{} is *[$\smash{(c,A)}$]{}-uniform* if it has no subpaths with [$\smash{c}$]{}-excedance outside [$\smash{[-A,A]}$]{}. We say simply that [$\smash{{\mathscr{C}}}$]{} is *[$\smash{A}$]{}-uniform* if it is [$\smash{({\bar{c}}({\mathscr{C}}),A)}$]{}-uniform. We apply the same terminology to paths as well as cycles.
\[d:bridge\] Given weights [$\smash{w_1,\ldots,w_k}$]{} with mean [$\smash{{\bar{c}}/n}$]{}, the *[$\smash{D}$]{}-tilted bridge* is the process $$W_j= \sum_{i=1}^j (n w_i-{\bar{c}}-D/k),
\quad\text{started from {$\smash{W_0=0}$}
and ending at {$\smash{W_k=-D}$}.}$$ We refer to the [$\smash{D=0}$]{} case as the *untilted* bridge: in particular, a cycle with mean weight [$\smash{{\bar{c}}/n}$]{} is [$\smash{A}$]{}-uniform if and only if its untilted bridge has range [$\smash{\le {\bar{c}}A}$]{}. Sometimes we say [$\smash{k}$]{}-bridge to emphasize that the bridge is defined on the time interval [$\smash{[0,k]}$]{}.
Exponential random walks {#ss:intro.exp.rw}
------------------------
The *exp-minus-one random walk* is the random walk on the real line with step distribution [$\smash{u-1}$]{}, where [$\smash{u}$]{} is a unit-rate exponential random variable. Observe that the cycle bridges (Definition \[d:bridge\]) can be rescaled to an exp-minus-one version: to see this, take any positive [$\smash{\mu}$]{} and consider the random vector [$\smash{{\underline{\smash{w}}}=(w_1,\ldots,w_k)}$]{} where the [$\smash{w_i}$]{} are independent exponential random variables with mean [$\smash{\mu}$]{}. Following the notation , suppose they sum to [$\smash{k{\bar{c}}/n}$]{}. Let [$\smash{\bar{s}}$]{} be any positive [$\smash{{\bar{c}}}$]{}-measurable random variable, and consider the process $$W_j = \sum_{i=1}^j (n w_i-\bar{s})
=\bar{s} \sum_{i=1}^j (n w_i/\bar{s}-1),
\quad 0\le j\le k.$$ Conditioned on [$\smash{{\bar{c}}}$]{}, the vector [$\smash{{\underline{\smash{w}}}}$]{} is distributed as a uniform sample from the space of all non-negative vectors in [$\smash{\R^k}$]{} with sum [$\smash{k{\bar{c}}/n}$]{}, regardless of the value of [$\smash{\mu}$]{}. In particular, taking [$\smash{\mu=\bar{s}/n}$]{} shows that $$\label{e:scale}
\begin{array}{l}
\text{the process {$\smash{W/\bar{s}}$}
is distributed as an
exp-minus-one random walk}\\
\text{started from the origin and
\emph{conditioned} to be at {$\smash{k({\bar{c}}/\bar{s}-1)}$}
at time {$\smash{k}$}.}\end{array}$$ For example, by taking [$\smash{\bar{s}={\bar{c}}}$]{}, we see that the process with increments [$\smash{(nw_i/{\bar{c}}-1)}$]{} is distributed simply as an exp-minus-one walk conditioned to return to the origin at time [$\smash{k}$]{}. We will apply with [$\smash{\bar{s}\ne{\bar{c}}}$]{} in the proof of Lemma \[l:unif.paths\].
In view of Definition \[d:bridge\], we study the exp-minus-walk with restricted range. In Section \[s:eig\] we give a precise computation of the principal eigenvalue [$\smash{\lm_A}$]{} of the exp-minus-walk with killing outside [$\smash{[0,A]}$]{}. In the limit of large [$\smash{A}$]{} it behaves as $$\label{e:lambda.A.asymptotic}
\lm_A=\exp\{ -\pi^2/(2A^2) + O(1/A^3) \}.$$ By comparison, for simple symmetric random walk on the integers killed outside [$\smash{\{1,\dots,A-1\}}$]{} (with [$\smash{A}$]{} integral), the principal eigenvalue is [$\smash{\lm_A^\textsc{srw}=\cos(\pi/A)}$]{}, which behaves in the limit of large [$\smash{A}$]{} as [$\smash{\exp\{ -\pi^2/(2A^2) + O(1/A^4) \}}$]{} (see e.g. [@kac]).
Proof overview
--------------
Having finished our preliminary calculations, we conclude this section by outlining the proof of our main result.
In Section \[s:rw\] we prove the necessary estimates for (range-restricted) exp-minus-one walks — taking as input the principal eigenvalue and eigenfunction for the walk with killing outside the interval [$\smash{[0,A]}$]{}, which will be computed in Section \[s:eig\]. The most important consequences of this section concern the *exp-minus-one random walk [$\smash{k}$]{}-bridge*, by which we mean an exp-minus-one random walk conditioned to return to the origin in [$\smash{k}$]{} steps. Lemma \[l:bridge\] computes the probability [$\smash{R^k_A}$]{} for this process to have range [$\smash{\le A}$]{}: $$R^k_A \asymp (\lm_A)^k \f{k^{3/2}}{A^3}
\quad\quad\text{for all }k\gtrsim A^2.$$ Lemma \[l:expected.profile\] shows that if we condition this process to have range [$\smash{\le A}$]{}, then its local times are comparable with those of the analogously range-restricted Brownian motion.
In Section \[s:atypical\] we apply the random walk estimates to rule out supercritical cycles which are atypically light or long. Recall that in Lemma \[l:basic.first.moment\] we computed the expectation of the number [$\smash{Z^k_c}$]{} of [$\smash{c}$]{}-light [$\smash{k}$]{}-cycles. Let [$\smash{Z^k_c(A)}$]{} count [$\smash{c}$]{}-light [$\smash{k}$]{}-cycles that are [$\smash{A}$]{}-uniform: in Lemma \[l:unif.cycles\] we apply the estimate on [$\smash{R^k_A}$]{} to prove $$\E Z^k_c(A)
= (\E Z^k_c) R^k_A
\asymp \f{(ce\lm_A)^k}{A^3}
\quad\quad\text{for all }k\gtrsim A^2.$$ If a [$\smash{c}$]{}-light cycle fails to be [$\smash{A}$]{}-uniform, then we can extract a subpath whose bridge decreases by [$\smash{-A+O(1)}$]{} and is [$\smash{(A+O(1))}$]{}-uniform. We let [$\smash{{\bar{Z}}^\ell_c(A)}$]{} count [$\smash{\ell}$]{}-paths of this type: in Lemma \[l:unif.paths\] we apply random walk estimates from Section \[s:rw\] to show that $$\E {\bar{Z}}^\ell_c(A)
\lesssim \f{(ce\lm_A)^\ell}{A^3} \f{n}{e^A}
\quad\quad
\text{for all }\ell\ge1.$$ Consequently, if [$\smash{Z^{\ge A^2}_c}$]{} counts all [$\smash{c}$]{}-light cycles of length [$\smash{\ge A^2}$]{}, we have $$\label{e:cycles.and.paths.lbd}
\P\big( Z^{\ge A^2}_c>0 \big)
\le \sum_{k\ge A^2} \E Z^k_c(A)
+ \sum_{\ell\ge1}\E {\bar{Z}}^\ell_c(A)
\lesssim \sum_{k\ge1} \f{(ce\lm_A)^k}{A^3}
\Big[1+\f{n}{e^A}\Big].$$ This calculation suggests that we take [$\smash{A=\log n+O(1)}$]{} and rule out values of [$\smash{c}$]{} that make [$\smash{ce\lm_A}$]{} too small. This leads to the definitions \[e:c.crit\] A\_n, c\_ = , applying . It suffices to prove Theorem \[t:main\] with [$\smash{c_\circ}$]{} in place of [$\smash{c_\star}$]{}. The lower bound on [$\smash{n{\mathscr{W}}_n}$]{} stated in the theorem is an easy consequence of Lemma \[l:log.squared\] and , and this is the first main consequence of Section \[s:atypical\].
The second main part of Section \[s:atypical\] is to rule out cycles that are much longer than [$\smash{(\log n)^3}$]{} in the regime [$\smash{|{\bar{c}}-c_\circ| \lesssim 1/(\log n)^3}$]{}. This argument is rather more involved, but it is guided by the same basic principle that the untilted bridge of a cycle either stays in a restricted range, or has sharp decreases over restricted ranges. If a cycle is very long, its bridge must either have many sharp decreases, or else stay in restricted ranges over long intervals. Both events impose a severe probability cost which can be used to rule out the presence of long cycles.
In Section \[s:second.moment\] we identify a subcollection of uniform cycles with a “typical local time profile,” and show that the number of such cycles is well concentrated about its mean. It follows that these cycles exist (with high probability) whenever their expected number is large. It follows from Lemma \[l:expected.profile\] that most of the contribution to [$\smash{\E Z^k_c(A)}$]{} comes from cycles having a typical local time profile — meaning that [$\smash{\E Z^k_c(A)}$]{} locates the sharp transition. Theorem \[t:main\] then follows in a straightforward manner.
Random walk estimates {#s:rw}
=====================
In this section we derive the necessary estimates for exp-minus-one walks subject to restrictions on the range of the walk. A few of the estimates require some understanding of the principal eigenvalue [$\smash{\lm_A}$]{}, and associated left eigenfunction [$\smash{\vph_A}$]{}, for the walk with killing outside the interval [$\smash{[0,A]}$]{}. These will be computed in Section \[s:eig\]; in the present section we shall require only the facts that $$\label{e:delta.A-1}
\lm_A
= \exp\{ -[1+O(1/A)]\pi^2/(2A^2) \}$$ and that when [$\smash{\vph_A}$]{} is normalized to be a probability density on [$\smash{[0,A]}$]{}, $$\label{e:delta.A-2}
\vph_A(x) \asymp \de_A(x)/A^2\,,\\
\text{ where }
\de_A(x)
\equiv \Ind{0\le x\le A}
[(x+1) \wedge (A-x+1)].$$ We will also make repeated use of a coupling of Komlós–Major–Tusnády [@KMT2 Theorem 1]: let [$\smash{(X_i)_{i\ge1}}$]{} be i.i.d. random variables with zero mean, unit variance, and finite exponential moments. For any [$\smash{\lm>0}$]{} there are constants [$\smash{K_1,K_2}$]{}, and a coupling of [$\smash{(X_i)_{i\ge1}}$]{} to i.i.d. standard Gaussian random variables [$\smash{(Y_i)_{i\ge1}}$]{}, such that $$\P\bigg( \max_{j\le k} \left|
\sum_{i=1}^j (X_i-Y_i)\right|
>K_1\log k+x \bigg) \le K_2 e^{-\lm x}.$$
Estimates for short time scales
-------------------------------
We begin with some estimates for exp-minus-one walks run for [$\smash{k}$]{} steps where [$\smash{k}$]{} is arbitrary. We will apply these estimates for the case [$\smash{k\lesssim A^2}$]{}; in the next subsection we derive better estimates for longer time scales [$\smash{k \gtrsim A^2}$]{}. The following lemma is well known, see e.g., [@pemantle-peres Lemma 3.3].
\[l:one.side\] Consider a random walk whose step distribution has zero mean and unit variance. The probability that the walk started from [$\smash{x}$]{} survives at least [$\smash{k}$]{} steps before going negative is [$\smash{\asymp (x+1)/(k+1)^{1/2}}$]{}, uniformly over [$\smash{k\ge0}$]{} and [$\smash{0\le x\lesssim k^{1/2}}$]{}.
Next we review an easy estimate for the probability the random walk stays confined in an interval; this will be substantially refined later for exp-minus-one walks.
\[l:exit.prob\] Consider a random walk whose step distribution has zero mean and unit variance. The probability the walk will survive for at least [$\smash{k}$]{} steps before exiting [$\smash{[0,A]}$]{} is [$\smash{\lesssim \exp\{ -\Om(k/A^2) \}}$]{} uniformly over all [$\smash{A\gtrsim1}$]{}, [$\smash{k\ge0}$]{}, [$\smash{X_0\in[0,A]}$]{}.
Divide the time interval [$\smash{[0,k]}$]{} into length-[$\smash{t}$]{} subintervals with [$\smash{t\asymp A^2}$]{}. It is then enough to note that the probability for the walk [$\smash{X}$]{} to survive over a single subinterval is bounded away from one, uniformly over [$\smash{A}$]{} and over the choice of the starting point [$\smash{x\in[0,A]}$]{}. Indeed, the survival probability is upper bounded by [$\smash{\P(|X_t-X_0|\le A)}$]{}, which is bounded away from one either by direct calculation with the gamma distribution, or by applying the central limit theorem for [$\smash{(X_t-t)/\sqrt{t}}$]{}.
From now on we restrict our attention to exp-minus-one random walks in intervals [$\smash{[0,A]}$]{} with absorbing boundaries. All our estimates hold also for one-minus-exp random walks (up to constant multiplicative error). Recall from the definition of [$\smash{\de_A}$]{}. We assume from now on that [$\smash{A}$]{} is at least of large constant size.
\[c:two.side\] There is an absolute constant [$\smash{\eta>0}$]{} such that for [$\smash{k\le (\eta A)^2}$]{}, the probability for the exp-minus-one walk started from [$\smash{x}$]{} to survive at least [$\smash{k}$]{} steps before exiting [$\smash{[0,A]}$]{} is [$\smash{\asymp \de_A(x)/(k+1)^{1/2}}$]{}, uniformly over [$\smash{0\le \de_A(x) \lesssim k^{1/2}}$]{}.
As before it suffices to consider [$\smash{k}$]{} exceeding any large fixed constant, so that [$\smash{k\asymp k+1}$]{}. Assume first that [$\smash{x\le A/2}$]{}, so [$\smash{\de_A(x)=x+1}$]{}; the case [$\smash{x\ge A/2}$]{} follows in a symmetric fashion. From the one-sided bound of Lemma \[l:one.side\], the probability for the walk started from [$\smash{x}$]{} to survive at least [$\smash{k}$]{} steps before going negative is [$\smash{\asymp (x+1)/k^{1/2}}$]{}, which trivially implies the upper bound. For the lower bound it suffices to subtract the probability that the walk exceeds [$\smash{A}$]{} before going negative, which has probability [$\smash{\lesssim (x+1)/A \le \eta(x+1)/k^{1/2}}$]{}. Taking sufficiently small [$\smash{\eta}$]{} gives the lower bound.
Let [$\smash{p^k_A(x,y)}$]{} denote the probability density for an exp-minus-one walk [$\smash{X}$]{} to go from [$\smash{x}$]{} to [$\smash{y}$]{} in [$\smash{k}$]{} time steps without exiting [$\smash{[0,A]}$]{}. Let [$\smash{\vph^k_A(x,y)}$]{} denote the density at [$\smash{X_k=y}$]{} conditioned on survival in [$\smash{[0,A]}$]{} for [$\smash{k}$]{} steps. $$\begin{array}{rl}
p^k_A(x,y)
\hspace{-6pt}&\equiv
\lim_{\ep\downarrow0} \ (2\ep)^{-1}
\P( |X_k-y|\le\ep; X_t \in[0,A]
\text{ for all } 0\le t\le k
\,|\, X_0=x ),\\
\vph^k_A(x,y)
\hspace{-6pt}&\equiv
\lim_{\ep\downarrow0} \
(2\ep)^{-1}
\P( |X_k-y|\le\ep \,|\,
X_t \in[0,A] \text{ for all } 0\le t\le k,
X_0=x
).
\end{array}$$
\[l:short.kernel\] It holds uniformly over [$\smash{A,k,x,y}$]{} that $$p^k_A(x,y)\lesssim
\f{\de_A(x) \de_A(y) }{ (k+1)^{3/2}}.$$
As before it suffices to consider [$\smash{k}$]{} exceeding any large fixed constant, so that [$\smash{k\asymp k+1}$]{}. Take exp-minus-one walks [$\smash{X}$]{} and [$\smash{Z}$]{} started from [$\smash{X_0=x}$]{} and [$\smash{Z_0=0}$]{} respectively, and take a one-minus-exp walk [$\smash{Y}$]{} started from [$\smash{Y_0=y}$]{}, with [$\smash{X,Y,Z}$]{} mutually independent. Fix also [$\smash{c\in(0,1/2)}$]{}. The probability for an exp-minus-one walk to go from [$\smash{x}$]{} to [$\smash{[y-\ep,y+\ep]}$]{} without exiting [$\smash{[0,A]}$]{} is upper bounded by the probability of the intersection of three events: $$\begin{array}{rl}
E_1 \hspace{-6pt}&=
\set{\text{{$\smash{X}$} survives at least
{$\smash{ck}$} steps
without exiting {$\smash{[0,A]}$}}},\\
E_2 \hspace{-6pt}
&= \set{\text{{$\smash{Y}$} survives at least
{$\smash{ck}$} steps
without exiting {$\smash{[0,A]}$}}},\\
E_3 \hspace{-6pt}&=
\set{\text{{$\smash{Z}$} goes from {$\smash{0}$}
to {$\smash{Y_{ck}-X_{ck} + [-\ep,\ep]}$}
in {$\smash{(1-2c)k}$} steps}}.
\end{array}$$ Now take [$\smash{c=1/3}$]{}, so the one-sided bound of Lemma \[l:one.side\] gives [$\smash{\P(E_1)\lesssim \de_A(x)/k^{1/2}}$]{} and similarly [$\smash{\P(E_2)\lesssim \de_A(y)/k^{1/2}}$]{}. The <span style="font-variant:small-caps;">clt</span> then gives [$\smash{\P(E_3 | E_1 \cap E_2, Y_{ck}-X_{ck})\lesssim \ep/k^{1/2}}$]{}, uniformly over all choices of [$\smash{Y_{ck}-X_{ck}}$]{}. Multiplying these probabilities together proves the bound.
\[c:apx.sine\] For [$\smash{k\asymp A^2}$]{}, it holds uniformly over [$\smash{x,y}$]{} that $$p^k_A(x,y)\asymp
\f{\de_A(x) \de_A(y) }{ A^3}
\quad\quad\text{and}\quad\quad
\vph^k_A(x,y) \asymp \f{ \de_A(y)}{A^2}.$$
Take the same notation as in the proof of Lemma \[l:short.kernel\], and take a constant [$\smash{c\in(0,1/2)}$]{} such that [$\smash{A^2\lesssim ck\le (\eta A)^2}$]{}. Corollary \[c:two.side\] then gives [$\smash{\P(E_1\cap E_2)\asymp \de_A(x)\de_A(y)/k}$]{}. Let $$E_4\equiv \left\{\hspace{-4pt}
\begin{array}{c}
\text{{$\smash{Z}$} goes from {$\smash{0}$}
to {$\smash{Y_{ck}-X_{ck} + [-\ep,\ep]}$}
in {$\smash{(1-2c)k}$} steps}\\
\text{without exiting the interval {$\smash{[-A/3,2A/3]}$}.}
\end{array}
\hspace{-4pt}\right\}$$ Conditioned on [$\smash{E_1\cap E_2}$]{}, the event [$\smash{E_5}$]{} that both [$\smash{X_{ck},Y_{ck}}$]{} are in [$\smash{[A/3, 2A/3]}$]{} occurs with probability [$\smash{\asymp1}$]{}: for [$\smash{\de_A(x),\de_A(y){\oldgg}\log k}$]{} this can be deduced directly from the KMT coupling; otherwise one can use the argument from Lemma \[l:one.side\] to reduce to the case [$\smash{\de_A(x),\de_A(y){\oldgg}\log k}$]{}. The functional <span style="font-variant:small-caps;">clt</span> gives $$\P(E_4 \,|\,
E_1 \cap E_2 \cap E_5) \asymp
\ep/(k-2ck)^{1/2}
\asymp \ep/k^{1/2},$$ so altogether [$\smash{p^k_A(x,y)\gtrsim \de_A(x)\de_A(y)/k^{3/2} \asymp \de_A(x)\de_A(y)/A^3}$]{}. Combining with the upper bound of Lemma \[l:short.kernel\] proves [$\smash{p^k_A(x,y) \asymp \de_A(x)\de_A(y)/A^3}$]{}. A final application of Corollary \[c:two.side\] proves the estimate on the conditional density, [$\smash{\vph^k_A(x,y)\asymp \de_A(y)/A^2}$]{}.
In preparation for the lemma that follows, observe that the [$\smash{\mathrm{Gam}(t)}$]{} density [$\smash{f_t}$]{} satisfies \[e:one.over.x.ubd\] f\_t(t+x) 1/|x+1| . For [$\smash{t=1}$]{} it is easy to verify that [$\smash{f_1(1+x)= e^{-1-x} \le 1/|x+1|}$]{}, uniformly over all [$\smash{x\ge -1}$]{}. For [$\smash{t\ge2}$]{}, note that [$\smash{f_t(t+x)}$]{} is unimodal in [$\smash{x}$]{}, with the unique mode at [$\smash{x=-1}$]{}. Since the density [$\smash{f_t}$]{} integrates to [$\smash{1}$]{}, for any [$\smash{x\ge-t}$]{} we have $$1\ge \int_0^{|x+1|} f_t(t-1+u)\,du
\ge |x+1| f_t(t+x),$$ and rearranging proves .
\[l:large.jump\] It holds uniformly over [$\smash{A,k,x,y}$]{} that $$p^k_A(x,y)\lesssim\left\{
\hspace{-4pt}
\begin{array}{rl}
(x+1)(A-y+1) / |x-y|^{3} &\text{for } x<y; \\
(A-x+1)(y+1) / |x-y|^{3} &\text{for } x>y. \\
\end{array}\right.$$
Assume [$\smash{0\le x < y \le A}$]{}; the case [$\smash{x>y}$]{} follows by a symmetric argument. Define [$\smash{x' \equiv x+(y-x)/4}$]{}, [$\smash{y' \equiv y-(y-x)/4}$]{}; note that we need only consider [$\smash{y-x}$]{} exceeding a large constant. Take exp-minus-one walks [$\smash{X}$]{} and [$\smash{Z}$]{} started from [$\smash{X_0=x}$]{} and [$\smash{Z_0=0}$]{} respectively, and take a one-minus-exp walk [$\smash{Y}$]{} started from [$\smash{Y_0=y}$]{}, with [$\smash{X,Y,Z}$]{} mutually independent. Define the stopping times [$\smash{\si\equiv\min\set{t\ge0 :X_t\notin[0,x']}}$]{} and [$\smash{\tau\equiv\min\set{t\ge0 :Y_t\notin [y',A]}}$]{}. The probability for an exp-minus-one random walk to go from [$\smash{x}$]{} to [$\smash{[y-\ep,y+\ep]}$]{} in [$\smash{k}$]{} steps without exiting [$\smash{[0,A]}$]{} can be upper bounded by the probability of the intersection of three events, $$\begin{array}{l}
E_1 = \set{ X_\si>x'}\,,\\
E_2 = \set{ Y_\tau < y'}\,,\\
E_3=\set{ \si+\tau \le k} \cap
\set{Z_{k-(\si+\tau)} \in
Y_\tau-X_\si + [-\ep,\ep] }\,.
\end{array}$$ It is clear from Lemma \[l:exit.prob\] that the stopping time [$\smash{\sigma}$]{} is integrable, so Wald’s identity gives [$\smash{x=\E X_\sigma}$]{}. Rearranging gives $$\P(E_1) =\f{x-a}{b-a}
\text{ where }
a\equiv \E[X_\sigma\,|\,X_\sigma<0]
\text{ and }
b\equiv \E[X_\sigma\,|\,X_\sigma>x'].$$ If [$\smash{X}$]{} is either the exp-minus-one or one-minus-exp walk then [$\smash{-1\le a\le0}$]{} and [$\smash{x'\le b\le x'+1}$]{}, so [$\smash{\P(E_1) \lesssim (x+1)/x'\lesssim (x+1)/(y-x)}$]{}, and similarly [$\smash{\P(E_2) \lesssim (A-y+1)/(y-x)}$]{}. Conditioned on [$\smash{E_1\cap E_2}$]{}, we have [$\smash{y'-1\le Y_\tau \le y'}$]{} a.s., while [$\smash{X_\si-x'}$]{} is a standard exponential random variable. For [$\smash{X_\si-x'}$]{} large we take the crude bound $$\P( E_3; X_\si-x' > (y-x)/4 )
\le
\max_{t\le k}
\max_{u\in\R}
\f{\P(Z_t \in u + [-\ep,\ep])}{e^{(y-x)/4}}
\lesssim \f{\ep}{ e^{(y-x)/4} }
\lesssim \f{\ep}{y-x}.$$ For smaller values of [$\smash{X_\si-x'}$]{} we instead bound $$\P( E_3; X_\si-x' \le (y-x)/4 )
\le \max_{t\le k}
\max_{ 0\le u\le 1+(y-x)/4 }
\P(Z_t \in (y'-x')-u
+ [-\ep,\ep]).$$ which is again [$\smash{\lesssim \ep/(y-x)}$]{} by . Multiplying the probabilities together gives the claimed bound.
Estimates for long time scales
------------------------------
Recall from and the definitions of [$\smash{\lm_A,\vph_A,\de_A}$]{}.
\[l:long.kernel\] For [$\smash{k\gtrsim A^2}$]{}, it holds uniformly over [$\smash{x,y}$]{} that $$p^k_A(x,y)\asymp (\lm_A)^k
\f{\de_A(x) \de_A(y) }{ A^3}
\quad\text{and}\quad
\vph^k_A(x,y) \asymp \f{\de_A(y)}{A^2}.$$
Take [$\smash{t\le k}$]{} with [$\smash{A^2\lesssim t\le (\eta A)^2}$]{}, and run the walk for [$\smash{t}$]{} time steps: by Corollary \[c:two.side\], the walk survives in [$\smash{[0,A]}$]{} up to time [$\smash{t}$]{} with probability [$\smash{\asymp \de_A(x)/t^{1/2} \asymp \de_A(x)/A}$]{}. By Corollary \[c:apx.sine\], the density at [$\smash{X_t=z}$]{} conditioned on survival is [$\smash{\vph^t_A(x,z) \asymp \de_A(z)/A^2}$]{}. In particular we can find a large absolute constant [$\smash{C}$]{} such that $$C^{-1} \vph^{}_A(z)
\le \vph^t_A(x,z)
\le C \vph^{}_A(z).$$ Therefore, if we start from the distribution [$\smash{\vph^t_A(x,\cdot)}$]{} and evolve the walk forward for [$\smash{k-t}$]{} steps, the terminal density at [$\smash{X_k=y}$]{} (conditioned on survival up to time [$\smash{t}$]{}) will be sandwiched between [$\smash{C^{-1} (\lm_A)^{k-t}\vph_A(y)}$]{} and [$\smash{C (\lm_A)^{k-t} \vph_A(y)}$]{}. Both lower and upper bounds agree up to constant factors with [$\smash{(\lm_A)^k \de_A(y)/A^2}$]{}, and multiplying with the probability of survival up to time [$\smash{t}$]{} proves the first estimate [$\smash{p^k_A(x,y)\asymp (\lm_A)^k\de_A(x)\de_A(y)/A^3}$]{}. Integrating over [$\smash{y}$]{} proves that the probability of survival up to time [$\smash{k}$]{} is [$\smash{\asymp (\lm_A)^k \de_A(x)/A}$]{}, therefore [$\smash{\vph^k_A(x,y) \asymp \de_A(y)/A^2}$]{}.
\[r:monotone\] Since [$\smash{p^k_A(A/2,A/2)}$]{} is nondecreasing in [$\smash{A}$]{}, an immediate consequence of Lemma \[l:long.kernel\] is that [$\smash{\lm_A}$]{} is nondecreasing in [$\smash{A}$]{}.
Recall that [$\smash{(\lm_A)^k\asymp1}$]{} for [$\smash{k\lesssim A^2}$]{}. It therefore holds uniformly over all [$\smash{A,k,x,y}$]{} that $$\label{e:all.times.kernel.ubd}
p^k_A(x,y)
\lesssim
\f{(\lm_A)^k \de_A(x)\de_A(y)}{
(A^2 \wedge (k+1))^{3/2} }\quad
\begin{array}{l}
\text{by Lemma~\ref{l:short.kernel}
for {$\smash{k\lesssim A^2}$},}\\
\text{and Lemma~\ref{l:long.kernel}
for {$\smash{k\gtrsim A^2}$}.}
\end{array}$$ For [$\smash{A-x-y\asymp A}$]{} we also have the bounds $$\label{e:all.times.large.jump}
\left.\begin{array}{l}
p^k_A(x,A-y)\\
p^k_A(A-x,y)
\end{array}\hspace{-6pt}\right\}
\lesssim
\f{(\lm_A)^k(x+1)(y+1)}{A^3}\quad
\begin{array}{l}
\text{by
Lemma~\ref{l:large.jump}
for {$\smash{k\lesssim A^2}$},}\\
\text{and Lemma~\ref{l:long.kernel}
for {$\smash{k\gtrsim A^2}$}.}
\end{array}$$ In both and the [$\smash{\lesssim}$]{} can be replaced with [$\smash{\asymp}$]{} in the regime [$\smash{k\gtrsim A^2}$]{}.
We conclude with our estimates for the exp-minus-one [$\smash{k}$]{}-bridge, which we recall is an exp-minus-one walk conditioned to return to the origin at time [$\smash{k}$]{}. We first estimate the probability [$\smash{R^k_A}$]{} that this process has range at most [$\smash{A}$]{}, $$R^k_A
\equiv
\lim_{\ep\downarrow0}
\f{\P( |\operatorname{range}(X)|\le A;
|X_k|\le\ep )}{\P( |X_k|\le\ep )},
\quad
\text{{$\smash{X}$} an exp-minus-one walk.}$$
\[l:bridge\] For [$\smash{k\gtrsim A^2}$]{}, the exp-minus-one [$\smash{k}$]{}-bridge has range at most [$\smash{A}$]{} with probability $$R^k_A
\asymp (\lm_A)^k \f{k^{3/2}}{A^3}.$$
Let [$\smash{X}$]{} be an exp-minus-one walk started from [$\smash{X_0=0}$]{}. Decompose [$\smash{R^k_A = R^k_{A/4} + R^k_{A,\ge}}$]{}, where [$\smash{R^k_{A/4}}$]{} is the contribution from the event that the range of [$\smash{X}$]{} is smaller than [$\smash{A/4}$]{}, while [$\smash{R^k_{A,\ge}}$]{} is the contribution from the event that the range is in [$\smash{[A/4,A]}$]{}. In order to have range less than [$\smash{A/4}$]{} the walk must certainly stay confined within distance [$\smash{A/4}$]{} of the origin, so $$R^k_{A/4}
\le \lim_{\ep\downarrow0}
\f{\P(
(X_{0:k})\subseteq [-A/4,A/4];
|X_k|\le\ep)}
{\P(|X_k|\le\ep)}
\asymp k^{1/2} p^k_{A/2}(A/4,A/4)
\lesssim k^{1/2} (\lm_{A/2})^k / A$$ by . For [$\smash{k\gtrsim A^2}$]{} the right-hand side is [$\smash{\lesssim k^{3/2} (\lm_A)^k/A^3}$]{}.
For larger [$\smash{H}$]{} we make a more precise calculation. Let [$\smash{S,T\in[0,k)}$]{} denote the times where the walk achieves its minimum and maximum respectively: the contribution to [$\smash{R^k_{A,\ge}}$]{} from [$\smash{S<T}$]{} is $$\begin{aligned}
&\asymp
\int_{A/4}^A
\lim_{\ep\downarrow 0}
\f{
\P(\operatorname{range}(X|_{[0,k]})\in dH;
S<T;
|X_k|\le\ep )
}{\P(|X_k|\le\ep)}
\,dH\\
&\asymp
k^{1/2}
\sum_{s,t=0}^{k-1}
\Ind{s<t}
\int_{A/4}^A
p^{t-s}_H(0,H)
p^{k-t+s}_H(H,0)
\,dH
\lesssim
k^{5/2} \int_0^{3A/4}
\f{(\lm_{A-\Delta})^k}{(A-\Delta)^6}
\,d\Delta
\end{aligned}$$ where in the final step we applied and made the change of variables [$\smash{H=A-\Delta}$]{}. From the expansion and the monotonicity of [$\smash{\lambda_A}$]{} (Remark \[r:monotone\]) we have \[e:change.in.lm.of.A\] {1,{ }}, so we find that the contribution to [$\smash{R^k_{A,\ge}}$]{} from [$\smash{S<T}$]{} is $$\lesssim k^{5/2}\f{(\lm_A)^k}{A^6}
\int_0^{3A/4}
\Big(\f{\lm_{A-\Delta}}{\lm_A}\Big)^k \,d\Delta
\asymp k^{3/2}\f{(\lm_A)^k}{A^3}.$$ The contribution to [$\smash{R^k_{A,\ge}}$]{} from [$\smash{S>T}$]{} has the same value, so the upper bound follows. The lower bound can be obtained in a similar manner, but summing only over pairs [$\smash{s<t}$]{} with [$\smash{t-s\gtrsim A^2}$]{} and [$\smash{k-t+s\gtrsim A^2}$]{} and applying the lower bound from Lemma \[l:long.kernel\] (see the comment below ).
We conclude with an estimate on the local time profile for the exp-minus-one bridge, which will be used to control the variance of “good” cycles in our second moment argument. Let [$\smash{X}$]{} be an exp-minus-one [$\smash{k}$]{}-bridge, and let [$\smash{\widetilde{X}}$]{} be [$\smash{X}$]{} with a constant shift that centers it at [$\smash{A/2}$]{}: \[e:X.recentered.around.half.A\] \_tX\_t +( A-X-X)/2 We then define the local times of the range-restricted [$\smash{k}$]{}-bridge by $$\ell^k_A(S)
\equiv
\E\bigg[
\sum_{t=1}^k
\Ind{\widetilde{X}_t \in S}
\,\Big|\,|\operatorname{range}(X)|\le A
\bigg]\quad\quad
\text{for {$\smash{S\subseteq[0,A]}$}.}$$
\[l:expected.profile\] For [$\smash{k\gtrsim A^2}$]{}, it holds uniformly over subintervals [$\smash{S\subseteq [0,A]}$]{} that the local time of a range-restricted $k$-bridge is bounded by $$\ell^k_A(S) \lesssim (1+|S|) \max_{x\in S} \de_A(x)^2.$$
Let [$\smash{X}$]{} denote an exp-minus-one walk, and [$\smash{H}$]{} its exact range, so that the recentered walk [$\smash{\widetilde{X}}$]{} is confined in the interval [$\smash{I_{A,H} \equiv [(A-H)/2,(A+H)/2]}$]{}. We will compute $$\ell^k_A(S)
=
\lim_{\ep\downarrow0}
\f{
\sum_{t=1}^k
\P( \widetilde{X}_t\in S,
H\le A;
|X_k|\le\ep) }
{\P( H\le A; |X_k|\le\ep)}.$$ Lemma \[l:bridge\] gives that the denominator is [$\smash{\asymp \ep (\lm_A)^k k/A^3}$]{} (since [$\smash{X}$]{} is a walk rather than a bridge, there is a factor of [$\smash{\ep/\sqrt{k}}$]{}). We decompose [$\smash{\ell^k_A = \ell^k_{A,<}+ \ell^k_{A,\ge}}$]{} by separating the numerator into the cases [$\smash{H< A/4}$]{} and [$\smash{A/4\le H\le A}$]{}. For [$\smash{H<A/4}$]{}, applying the bound gives \[e:loc.times.bound.very.small.range\] \^k\_[A,<]{}(S) \^k\_[A,<]{}(\[0,A\]) A\^2( \_[A/2]{} / \_A)\^k A\^2 \_[xS]{}\_A(x)\^2, where in the last bound we used that [$\smash{\widetilde{X}}$]{} is confined in [$\smash{[3A/8,5A/8]}$]{}, so in fact [$\smash{\ell^k_{A,<}(S)}$]{} is zero unless [$\smash{\de_A(x)\asymp A}$]{} for some [$\smash{x\in S}$]{}.
For larger [$\smash{H}$]{}, we sum over all possibilities [$\smash{s,t\in[1,k]}$]{} where the walk achieves its minimum and maximum respectively, as well as all times [$\smash{u\in[1,k]}$]{} where [$\smash{\widetilde{X}_u\in S}$]{}. For simplicity we consider only the contribution from times [$\smash{s<u<t}$]{}; the contribution from other permutations is calculated similarly and will be of the same asymptotic order (up to constants): thus $$\ell^k_{A,\ge}(S)
\lesssim
\sum_{s,u,t=1}^k
\f{\Ind{s<u<t}}
{(\lm_A)^k k/A^3}
\int_{A/4}^A
\int_{S \cap I_{A,H}}
p_H^{u-s}(0,x_H)
p_H^{t-u}(x_H,H)
p_H^{k-t+s}(H,0)
\,dx
\,dH$$ where [$\smash{x_H \equiv x-(A-H)/2}$]{}, so that for [$\smash{x\in I_{A,H}}$]{} we have [$\smash{x_H\in[0,H]}$]{} and [$\smash{\de_H(x_H)\le\de_A(x)}$]{}. By symmetry it suffices to consider [$\smash{x_H\le H/2}$]{}: then [$\smash{p_H^{u-s}(0,x_H)}$]{} can be bounded by , while [$\smash{p_H^{t-u}(x_H,H)}$]{} and [$\smash{p_H^{k-t+s}(H,0)}$]{} can be bounded by . The total contribution to [$\smash{\ell^k_{A,\ge}(S)}$]{} from times [$\smash{s<u<t}$]{} with [$\smash{u-s\le H^2}$]{} is $$\lesssim
\f{k^2 }{(\lm_A)^k k/A^3}
\int_{A/4}^A
\f{(\lm_H)^k}{H^6}
\left(\int_S
\de_A(x)^2 \,dx\right) \,dH
\lesssim
|S| \max_{x\in S}\de_A(x)^2,$$ while the total contribution from times [$\smash{s<u<t}$]{} with [$\smash{u-s\ge H^2}$]{} is $$\lesssim
\f{k^3 }{(\lm_A)^k k/A^3}
\int_{A/4}^A \f{(\lm_H)^k}{H^9}
\left(\int_S
\de_A(x)^2 \,dx\right) \,dH
\lesssim
(k/A^3)|S|\max_{x\in S}\de_A(x)^2.$$ Adding these together and combining with gives the stated bound.
Light cycles and long cycles {#s:atypical}
============================
Recall that for a given cycle [$\smash{{\mathscr{C}}}$]{} we use [$\smash{{\bar{c}}({\mathscr{C}}) \equiv n\operatorname{wgt}({\mathscr{C}})/\operatorname{len}(P)}$]{} to denote its mean weight scaled by [$\smash{n}$]{}. In this section we prove that w.h.p. no cycles in the supercritical regime [$\smash{{\bar{c}}>1/e}$]{} have [$\smash{c_\circ-{\bar{c}}{\oldgg}1/(\log n)^3}$]{}. We also show that in the regime [$\smash{|{\bar{c}}-c_\circ|\lesssim 1/(\log n)^3}$]{}, w.h.p. there are no cycles of length [$\smash{{\oldgg}(\log n)^3}$]{}. The formal statement is as follows:
\[t:first\] In the complete graph or complete digraph with i.i.d. unit-rate exponential edge weights, for all [$\smash{\ep>0}$]{} there exists a constant [$\smash{C=C(\ep)>0}$]{} such that $$\label{e:t.first.too.light}
\P\left(
\exists\text{ cycle }{\mathscr{C}}:1/e<{\bar{c}}({\mathscr{C}})\leq c_\circ-C/(\log n)^3
\right) \le\ep.$$ For all [$\smash{\ep>0}$]{} and all [$\smash{C_1>0}$]{} there exists a constant [$\smash{C_2=C_2(\ep,C_1)>0}$]{} such that $$\label{e:t.first.light.and.too.long}
\P\left(
\begin{array}{c}
\exists\text{ cycle }{\mathscr{C}}:1/e< {\bar{c}}({\mathscr{C}})
\le
c_\circ+C_1/(\log n)^3\\
\quad\quad\quad\text{and
{$\smash{\operatorname{len}({\mathscr{C}})\geq C_2(\log n)^3}$}}
\end{array}
\right) \le\ep.$$
First moment for uniform cycles and paths {#sec:uniform-first-moment}
-----------------------------------------
We begin with some first moment estimates for uniform cycles and paths. Let $$\label{e:z.uniform}
\begin{array}{rl}
Z^k_c(A) \equiv
&\hspace{-6pt}
\text{number of {$\smash{c}$}-light
{$\smash{A}$}-uniform {$\smash{k}$}-cycles;}\\
{\bar{Z}}^k_c(A)
\equiv &\hspace{-6pt}
\text{number of {$\smash{k}$}-paths
that, for some {$\smash{A''\in[A-2,A]}$},}\\
&\text{have {$\smash{c}$}-excedance {$\smash{-A''}$}
and are {$\smash{(c,A'')}$}-uniform.}
\end{array}$$ We will also denote $$Z^{\ge k}_c(A)
\equiv \sum_{\ell\ge k}Z^{\ell}_c(A),
\quad\quad
Z^{}_c(A)
\equiv Z^{\ge 2}_c(A),\quad\quad
{\bar{Z}}^{}_c(A)
\equiv \sum_{\ell\ge1}
{\bar{Z}}^\ell_c(A).$$ The purpose of defining [$\smash{{\bar{Z}}_c(A)}$]{} is to bound the number of light cycles which do not contribute to [$\smash{Z_c(A)}$]{}. From Definition \[d:unif\], if a cycle [$\smash{{\mathscr{C}}}$]{} is [$\smash{c}$]{}-light but fails to be [$\smash{A}$]{}-uniform, then it has a subpath [$\smash{P}$]{} with [$\smash{c}$]{}-excedance [$\smash{<-A}$]{}. The following lemma shows how to extract further subpaths from [$\smash{P}$]{} that are both light and uniform:
\[l:many.subpaths\] For [$\smash{2\le A'\le A}$]{}, given a path [$\smash{P}$]{} with [$\smash{c}$]{}-excedance [$\smash{<-A}$]{}, one can extract at least [$\smash{1+\lfloor A-A' \rfloor}$]{} distinct (though not necessarily disjoint) subpaths [$\smash{p \subseteq P}$]{}, each contributing to [$\smash{{\bar{Z}}_c(A')}$]{} as defined in .
By scaling we can assume [$\smash{c=1}$]{}. Suppose [$\smash{P}$]{} has length $k$ and edge weights [$\smash{w_1,\ldots,w_k}$]{}. By assumption, the process [$\smash{W_j=\sum_{i=1}^j(nw_i-1)}$]{} goes from [$\smash{W_0=0}$]{} to [$\smash{W_k<-A}$]{}. Let $$(x_i,y_i) = -(i,i+A'),
\quad\text{for }
0\le i\le \lfloor A-A' \rfloor.$$ Since [$\smash{A'\ge2}$]{}, [$\smash{y_i+1<x_i-1}$]{} for each [$\smash{i}$]{}. Since [$\smash{W}$]{} decreases by at most one at each step, for each [$\smash{i}$]{} there is some [$\smash{\eta_i}$]{} for which [$\smash{W_{\eta_i} \in (x_i-1,x_i]}$]{}, and some [$\smash{\tau_i>\eta_i}$]{} for which [$\smash{W_{\tau_i} \in (y_i,y_i+1]}$]{}. Let [$\smash{[\bar\eta_i,\bar\tau_i]}$]{} be any minimal subinterval of [$\smash{[0,k]}$]{} for which [$\smash{W_{\bar\eta_i} \in (x_i-1,x_i]}$]{} and [$\smash{W_{\bar\tau_i} \in (y_i,y_i+1]}$]{}, and let [$\smash{p_i}$]{} be the subpath of [$\smash{P}$]{} corresponding to the interval [$\smash{[\bar\eta_i,\bar\tau_i]}$]{}. Since [$\smash{[\bar\eta_i,\bar\tau_i]}$]{} is minimal, the subpath [$\smash{p_i}$]{} is [$\smash{(W_{\bar\eta_i}-W_{\bar\tau_i})}$]{}-uniform, where [$\smash{A'-2\leq W_{\bar\eta_i}-W_{\bar\tau_i} \leq A'}$]{}. Since the intervals [$\smash{(x_i-1,x_i]}$]{} are disjoint, the paths [$\smash{p_i}$]{} are distinct.
Applying Lemma \[l:many.subpaths\] with [$\smash{A'=A}$]{}, when [$\smash{A\geq 2}$]{}, we see that any [$\smash{c}$]{}-light cycle either contributes to [$\smash{Z_c(A)}$]{}, or has a subpath contributing to [$\smash{{\bar{Z}}_c(A)}$]{}.
\[l:unif.cycles\] For [$\smash{k\gtrsim A^2}$]{}, [$\smash{c\lesssim1}$]{}, and [$\smash{0<\delta<1}$]{}, $$\label{eq:E[Z_c^k(A)]}
\E [Z^k_c(A)-Z^k_{c(1-\delta)}(A)]
\asymp
\f{(n)_k}{n^k} \f{(ce\lm_A )^k
[1-(1-\delta)^k]}{A^3}.$$
Recalling Definition \[d:unif\], a [$\smash{k}$]{}-cycle with edge weights [$\smash{w_1,\ldots,w_k}$]{} summing to [$\smash{k{\bar{c}}/n}$]{} contributes to [$\smash{Z^k_c(A)}$]{} if and only if [$\smash{{\bar{c}}\le c}$]{}, and the untilted bridge with increments [$\smash{(nw_i/{\bar{c}}-1)}$]{} has range at most [$\smash{A}$]{}. Applying with [$\smash{\bar{s}={\bar{c}}}$]{}, we see that [$\smash{\E Z^k_c(A)= (\E Z^k_c) R^k_A}$]{}, and then follows from Lemmas \[l:basic.first.moment\] and \[l:bridge\].
\[l:unif.paths\] For [$\smash{A\geq2}$]{}, [$\smash{c\lesssim1}$]{}, and [$\smash{k\ge 1}$]{}, $$\label{eq:E[tZ_c^k(A)]}
\E {\bar{Z}}^k_c(A)
\lesssim
\f{(ce\lm_A)^k}{A^3} \f{n}{e^A}.$$
A [$\smash{k}$]{}-path with edge weights [$\smash{w_1,\ldots,w_k}$]{} summing to [$\smash{k{\bar{c}}/n}$]{} contributes to [$\smash{{\bar{Z}}^k_c(A)}$]{} if and only if, for some [$\smash{A''\in[A-2,A]}$]{}, the process with increments [$\smash{(nw_i/c-1)}$]{} goes from [$\smash{0}$]{} to [$\smash{-A''}$]{} without exiting [$\smash{[-A'',0]}$]{}. In particular this implies [$\smash{{\bar{c}}=c(1-A''/k)}$]{}. Making the change of variables [$\smash{A''=A-\Delta}$]{}, we have $$\E{\bar{Z}}^k_c(A)
= (n)_{k+1}
\int_0^2
f_k[( k-A+\Delta ) c/n]
\f{p^k_{A-\Delta}(A-\Delta,0)}
{f_k( k-A+\Delta )}\,\f{c\,d\Delta}{n}$$ where the first factor in the integrand is the probability density of the path weight, and the second factor is the probability that an exp-minus-walk, conditioned to go from [$\smash{A-\Delta}$]{} to [$\smash{0}$]{} in [$\smash{k}$]{} steps, does so without exiting [$\smash{[0,A-\Delta]}$]{}. Simplifying gives $$\E{\bar{Z}}^k_c(A)
\leq (n)_{k+1}\f{c^k}{n^k}
\int_0^2 e^{k-A+\Delta}
p^k_{A-\Delta}(A-\Delta,0)
\,d\Delta
\lesssim
\f{n}{e^A}
\f{(ce\lm_A)^k}{A^3}
\int_0^2
\Big(\f{\lm_{A-\Delta}}{\lm_A}\Big)^k \,d\Delta,$$ where the last bound follows from . The final integrand is $\leq1$, which gives .
Uniform cycles and light cycles {#sec:uniform-light}
-------------------------------
We now apply our first moment estimates to argue that in the supercritical regime [$\smash{c>1/e}$]{}, cycles cannot be too uniform or too light. Recall that we set [$\smash{A_\circ\equiv\log n}$]{} and [$\smash{c_\circ \equiv 1/(e\lm_{A_\circ})}$]{}. More generally, we define the relation $$\label{e:gamma}
c\,e\lm_A = \exp\{-\gm/A^3\}$$ — where we can think of [$\smash{c}$]{} as being determined by [$\smash{A}$]{} and [$\smash{\gm}$]{}. A first consequence of Lemma \[l:unif.cycles\] is that cycles of length at least [$\smash{(\log n)^2}$]{} cannot be too uniform in the targeted regime for the mean weight:
\[p:too.unif\] For all [$\smash{\ep,C>0}$]{} there exists [$\smash{\Delta=\Delta(\ep,C)>0}$]{} large enough that $$\P\left(\begin{array}{c}
\exists\text{ cycle }{\mathscr{C}}:
{\bar{c}}({\mathscr{C}})\le c_\circ+C/(\log n)^3,\\
\text{{$\smash{\operatorname{len}({\mathscr{C}})\ge A^2}$},
and {$\smash{{\mathscr{C}}}$} is {$\smash{A}$}-uniform}
\end{array}
\right)\le\ep
\quad\quad\text{for }
A\equiv \log n-\Delta.$$
Suppose holds with [$\smash{\gm>0}$]{}. Using Markov’s inequality with Lemma \[l:unif.cycles\] and , $$\P(Z^{\ge A^2}_c(A)>0)
\le\E Z^{\ge A^2}_c(A)
\lesssim A^{-3}\sum_{k\ge A^2}e^{-k\gm/A^3}
\lesssim 1/\gm\,,$$ so by taking [$\smash{\gm\asymp 1/\ep}$]{} we can ensure [$\smash{\P(Z^{\ge A^2}_c(A)>0)\le\ep}$]{}. Using the expansion for [$\smash{\lambda_A}$]{}, we see that by choosing [$\smash{A=\log n-\Delta}$]{} with [$\smash{\Delta\asymp C+\gamma}$]{}, where [$\smash{\gamma}$]{} and [$\smash{C}$]{} are fixed as [$\smash{n\to\infty}$]{}, we can ensure [$\smash{c=1/(e\lambda_A) \exp(-\gamma/A^3) \geq c_\circ + C/(\log n)^3}$]{}.
We next show how Lemma \[l:unif.cycles\] and Lemma \[l:unif.paths\] can be combined to rule out cycles that are lighter than the targeted regime for the mean weight.
\[p:too.light\] For all [$\smash{\ep>0}$]{} there exists [$\smash{C=C(\ep)>0}$]{} such that $$\P\left(
\begin{array}{c}
\exists\text{ cycle }{\mathscr{C}}: {\bar{c}}({\mathscr{C}}) \le c_\circ-C/(\log n)^3\\
\text{ and }
\operatorname{len}({\mathscr{C}})\ge (\log n)^2
\end{array}
\right) \le\ep.$$
From Lemma \[l:many.subpaths\], assuming [$\smash{A\geq 2}$]{}, a [$\smash{c}$]{}-light [$\smash{k}$]{}-cycle either contributes to [$\smash{Z^k_c(A)}$]{}, or has a subpath contributing to [$\smash{{\bar{Z}}^\ell_c(A)}$]{} for some [$\smash{\ell\le k}$]{}. By Lemmas \[l:unif.cycles\] and \[l:unif.paths\], if we take [$\smash{A=A_\circ}$]{} and [$\smash{\gm}$]{} a large positive constant, then $$\P(Z^{\ge A^2}_c>0)
\le
\E Z^{\ge A^2}_c(A)
+\E {\bar{Z}}^{}_c(A)
\lesssim \sum_{\ell\ge1}
\f{e^{-\gm \ell/A^3}}{A^3}
\asymp 1/\gm\,,$$ so by taking [$\smash{\gm\asymp 1/\ep}$]{} we can ensure [$\smash{\P(Z^{\ge A^2}_c>0)\le\ep}$]{}. Since [$\smash{c/c_\circ = \exp(-\gm/A^3)}$]{} and [$\smash{c_\circ\asymp1}$]{}, the proposition follows with [$\smash{C\asymp 1/\ep}$]{}.
Non-uniform cycles and long cycles {#sec:non-uniform-long}
----------------------------------
Recall that in Lemma \[p:too.unif\] above we ruled out cycles that are too uniform in the targeted regime for the mean weight. We now prove the complementary assertion that cycles cannot be too non-uniform in this regime:
\[c:too.non.unif\] For all [$\smash{\ep,C>0}$]{} there exists [$\smash{\Delta=\Delta(\ep,C)>0}$]{} large enough that $$\P\left(
\begin{array}{c}
\exists\text{ cycle }{\mathscr{C}}: {\bar{c}}({\mathscr{C}})\le
c_\circ + C/(\log n)^3\\
\text{ and {$\smash{{\mathscr{C}}}$}
is not {$\smash{(\log n+\Delta)}$}-uniform}
\end{array}
\right)
\le\ep.$$
We can set [$\smash{\gamma=1}$]{} and [$\smash{A'\equiv\log n-\Delta'}$]{} with [$\smash{\Delta'\asymp C}$]{} and have $c \equiv 1/(e \lambda_{A'}) \exp(-\gamma/(A')^3) \geq c_\circ + C/(\log n)^3$. Then Lemma \[l:unif.paths\] gives $$\E {\bar{Z}}_c(A')\lesssim \frac{n}{(A')^3 e^{A'}} \frac{1}{1-e^{-\gamma/(A')^3}} \asymp n/e^{A'} = e^{\Delta'}\,.$$ Let [$\smash{A=\log n+\Delta \ge A'}$]{}. By Definition \[d:unif\] and Lemma \[l:many.subpaths\] together with Markov’s inequality, $$\P(Z_c > Z_c(A))
\le \P \big({\bar{Z}}_c(A') \ge
1+\lfloor A-A' \rfloor \big)
\le \f{\E {\bar{Z}}_c(A')}
{1+\lfloor \Delta+\Delta' \rfloor}
\lesssim \f{e^{\Delta'}}{(\Delta+\Delta')}.$$ By taking [$\smash{\Delta \asymp e^{\Delta'}/\ep}$]{} we can ensure that this probability is sufficiently small.
In the remainder of this section we argue that there are no cycles in the targeted weight regime that have length much longer than [$\smash{(\log n)^3}$]{}. In view of Lemmas \[p:too.unif\] and \[c:too.non.unif\], it remains to consider cycles which are neither too uniform nor too non-uniform:
\[l:light.and.too.long\] For all [$\smash{\ep,C,\Delta>0}$]{} there exists [$\smash{B=B(\ep,C,\Delta)>0}$]{} large enough that $$\P\left(
\begin{array}{c}
\exists\text{ cycle }{\mathscr{C}}:
c_\circ - C/(\log n)^3 \le
{\bar{c}}({\mathscr{C}}) \le c_\circ + C/(\log n)^3,\\
\text{{$\smash{{\mathscr{C}}}$} is {$\smash{(\log n+\Delta)}$}-uniform
but not {$\smash{(\log n-\Delta)}$}-uniform},\\
\text{and }\operatorname{len}({\mathscr{C}})\ge B(\log n)^3
\end{array}
\right)
\le\ep.$$
We denote $$c^\pm \equiv c_\circ\pm C/(\log n)^3,\quad\quad
A^\pm \equiv \log n \pm \Delta\,.$$ Consider a [$\smash{k}$]{}-cycle with weights [$\smash{w_1,\ldots,w_k}$]{} summing to [$\smash{k{\bar{c}}/n}$]{}, where [$\smash{c^-\le{\bar{c}}\le c^+}$]{}, which is [$\smash{A^+}$]{}-uniform but not [$\smash{A^-}$]{}-uniform. Let [$\smash{X}$]{} denote the corresponding untilted bridge, with increments [$\smash{(nw_i/{\bar{c}}-1)}$]{}. For convenience we shift [$\smash{X}$]{} by a constant so that its minimum is zero. Then its maximum is in [$\smash{[A^-,A^+]}$]{}.
Let [$\smash{H}$]{} denote the ceiling of the maximum of [$\smash{X}$]{}. Write [$\smash{H\equiv \lceil A_\circ\rceil+\Delta_H}$]{} (so [$\smash{\Delta_H}$]{} is integer-valued). Let [$\smash{A' \equiv \lceil A_\circ\rceil-\Delta' }$]{} be the largest integer for which [$\smash{A' \le A^- -2}$]{} and \[e:condition.on.delta.prime\] c\^+ e \_A { -1/A\^3 } (n)/2 A A’+2. The constraint [$\smash{A'\leq A^- -2}$]{} ensures [$\smash{\Delta'+\Delta_H\geq 1}$]{}. It follows from that [$\smash{\Delta' \asymp\max(C,\Delta)}$]{}. Define $$\begin{array}{rl}
x \hspace{-6pt}&\equiv H-(A'-2)
=\Delta_H+\Delta'+2,\\
y \hspace{-6pt}&\equiv A'-2
= H-(\Delta_H+\Delta'+2).
\end{array}$$ For fixed [$\smash{C}$]{} and [$\smash{\Delta}$]{} when [$\smash{n}$]{} is big enough we have [$\smash{0<x<y<H}$]{}.
For a sequence of times [$\smash{0<t_1<\ldots<t_q\le k}$]{} with [$\smash{q}$]{} even, let [$\smash{E[t_{1:q}]}$]{} denote the event that the [$\smash{t_i}$]{} partition the bridge [$\smash{X}$]{} into alternating up-crossings and down-crossings: that is, $$E[t_{1:q}]
=
\set{U[t_{i-1},t_i]\text{ for all {$\smash{i}$} even}}
\cap
\set{D[t_{i-1},t_i]\text{ for all {$\smash{i}$} odd}}$$ where [$\smash{U[a,b]}$]{} and [$\smash{D[a,b]}$]{} indicate the events of up-crossing and down-crossing on [$\smash{[a,b]}$]{}: $$\begin{array}{rl}
U[a,b]
\hspace{-6pt}&\equiv
\set{ X_a<x\text{ and }b=\min\set{t>a : X_t>y} }\\
D[a,b]
\hspace{-6pt}&\equiv\set{ X_a>y\text{ and }b=\min\set{t>a: X_t<x} }.
\end{array}$$ (with cyclic time indexing). The proof idea is as follows: any cycle with a very large number of crossings will give an excessive contribution to [$\smash{{\bar{Z}}_c(A'')}$]{} where [$\smash{A'' \equiv y-x}$]{}, so we can assume a bounded number of crossings. We then argue that a long cycle with a bounded number of crossings is unlikely to occur because its range is effectively reduced from [$\smash{A}$]{} to [$\smash{A'}$]{}.
Each down-crossing corresponds to a subpath with [$\smash{c}$]{}-excedance [$\smash{<-A''}$]{}. By Lemma \[l:many.subpaths\], from each such down-crossing we can extract a contribution to [$\smash{{\bar{Z}}_{c^+}(A'')}$]{}. From Lemma \[l:unif.paths\] and , the expected number of cycles with at least [$\smash{Q/2}$]{} down-crossings is $$\leq \f{\E {\bar{Z}}_{c^+}(A'')}{Q/2}
\lesssim \f{n}{Q e^{A''}}
\sum_{\ell\ge1} \f{(c^+e\lm_{A''}
)^\ell}{(A'')^3}
\lesssim \f{n}{Q e^{A''}}
\lesssim \f{e^{\Delta+2\Delta'}}{Q},$$ which can be made arbitrarily small by taking [$\smash{Q}$]{} large.
It therefore remains to control the cycles with $q<Q$. To this end, we shall first bound the probability for an exp-minus-one walk [$\smash{Y}$]{} started at [$\smash{Y_0=X_0}$]{} — with no conditioning on the value of [$\smash{Y_k}$]{} — to belong to the event [$\smash{E[t_{1:q}]}$]{}.
1. First consider the probability of a length-[$\smash{t}$]{} down-crossing, [$\smash{D[a,a+t]}$]{}, conditioned on having just completed an up-crossing at time [$\smash{a}$]{}. This means that [$\smash{\lceil Y_a \rceil = y+h}$]{} for some integer [$\smash{1\le h\le H-y}$]{}. Conditional on [$\smash{h}$]{}, the probability to make a down-crossing is upper bounded by the probability that an exp-minus-one walk started exactly at [$\smash{y+h}$]{} will travel to [$\smash{[x-1,x+1]}$]{} in [$\smash{t}$]{} steps without exiting the interval [$\smash{[x-1,H+1]}$]{}. It follows that the probability of down-crossing [$\smash{D[a,a+t]}$]{}, conditioned on having just completed an up-crossing at time [$\smash{a}$]{}, is upper bounded by [$\smash{p_\text{dn}(t)}$]{} where $$p_\text{dn}(t)
= \max_{1\le h\le H-y}
\int_0^2
p^t_{A'}(y+h-(x-1),u) \,du.$$ Recalling [$\smash{H-y = \Delta_H+\Delta'+2}$]{} and applying gives $$p_\text{dn}(t)
\lesssim \f{(\lm_{A'})^t(\Delta_H+\Delta')}{(A')^3}.$$
2. Similarly, consider the probability of a length-[$\smash{t}$]{} up-crossing [$\smash{U[a,a+t]}$]{}, conditioned on having just completed a down-crossing at time [$\smash{a}$]{}. This is upper bounded by the probability [$\smash{p_\text{up}(t)}$]{} that an exp-minus-one walk started exactly at the integer [$\smash{x}$]{} will travel to [$\smash{[y,H+1]}$]{} in [$\smash{t}$]{} steps, and remain in the interval [$\smash{[0,y+1]\subset[0,A']}$]{} up to time [$\smash{t-1}$]{}. Conditioning on the position [$\smash{y+1-u}$]{} of the walk at time [$\smash{t-1}$]{} and the position [$\smash{z}$]{} at time [$\smash{t}$]{} gives $$p_\text{up}(t)
\le \int_y^{H+1}
\int_0^{y+1}
\f{p^{t-1}_{A'}(x,y+1-u)}
{ e^{(z-y)+u} }
\,du
\, dz
\lesssim
\int_0^{y+1}
\f{p^{t-1}_{A'}(x,y+1-u)}{e^u}
\,du.$$ If [$\smash{t\gtrsim (\log n)^2}$]{}, then gives $$p_\text{up}(t)
\lesssim
\f{(\lm_{A'})^t x}{(A')^3}
\int_0^{y+1} \f{u+1}{e^u} \,du
\lesssim
\f{(\lm_{A'})^t (\Delta_H+\Delta')}{(A')^3}$$ If [$\smash{t\lesssim (\log n)^2}$]{}, the contribution from [$\smash{u\le y/2}$]{} satisfies the same bound by applying in place of , while the contribution from [$\smash{u\ge y/2}$]{} is negligible in comparison.
Conditioned on all these crossings, the probability that [$\smash{|Y_k-Y_0|\le\ep}$]{} is [$\smash{\lesssim\ep}$]{}, whereas unconditionally we have [$\smash{|Y_k|\le\ep}$]{} with probability [$\smash{\asymp \ep/k^{1/2}}$]{}. Altogether we find $$\P(X\in E[t_{1:q}])
= \lim_{\ep\downarrow0}
\f{\P(Y\in E[t_{1:q}] ; |Y_k|\le\ep)}
{\P(|Y_k|\le\ep)}
\lesssim k^{1/2} \f{(\lm_{A'})^k (\Delta_H+\Delta')^{q} e^{O(q)}}{(A')^{3q}}.$$ Therefore, the probability to have a cycle [$\smash{{\mathscr{C}}}$]{} with [$\smash{|{\bar{c}}({\mathscr{C}})-c_\circ|\le C/(\log n)^3}$]{}, which is [$\smash{A^+}$]{}-uniformly light but not [$\smash{A^-}$]{}-uniformly light, has length [$\smash{\operatorname{len}({\mathscr{C}})\ge L=B(\log n)^3}$]{}, and [$\smash{q\le Q}$]{} is, using Markov’s inequality and Lemma \[l:basic.first.moment\], $$\lesssim
\sum_{k\ge L} \E Z^k_{c^+}
\sum_{q\le Q}
\sum_{0<t_1<\ldots<t_q\le k}
\P(X\in E[t_{1:q}])
\lesssim
\sum_{k\ge L}
\f{(c^+ e \lm_{A'})^k}{k^{3/2}}
\sum_{q\le Q}
\f{k^{1/2} [(\Delta_H+\Delta')k]^q e^{O(q)}}{(A')^{3q}}\,.$$ We can assume [$\smash{B\gtrsim 1/(\Delta_H+\Delta')}$]{} to make the summand increasing in $q$. Recalling , this in turn is $$\lesssim
\f{(A')^3 (\Delta_H+\Delta')^Q e^{O(Q)} }{B(\log n)^3}
\sum_{k\ge L}
\f{Q [k/(A')^3]^Q}{\exp\{ k/(A')^3 \}}
\f{1}{(A')^3}
\lesssim
\f{Q (\Delta_H+\Delta')^Q e^{O(Q)}}{B}
\int_{B}^\infty
\f{z^Q}{e^z} \,dz.$$ This can be made arbitrarily small by taking [$\smash{B}$]{} large, and the result follows.
\[p:light.and.too.long\] For all [$\smash{\ep,C>0}$]{} there exists [$\smash{B=B(\ep,C)>0}$]{} large enough that $$\P\left(
\begin{array}{c}
\exists\text{ cycle }{\mathscr{C}}:
c_\circ - C/(\log n)^3 \le
{\bar{c}}({\mathscr{C}}) \le c_\circ + C/(\log n)^3\\
\text{and }\operatorname{len}({\mathscr{C}})\ge B(\log n)^3
\end{array}
\right)
\le\ep.$$
Follows by combining Lemmas \[p:too.unif\], \[c:too.non.unif\], and \[l:light.and.too.long\].
Proposition \[p:too.light\] rules out light supercritical cycles of length [$\smash{\geq (\log n)^2}$]{}, and Lemma \[l:log.squared\] rules out the ones shorter than [$\smash{(\log n)^2}$]{}, which implies . Proposition \[p:light.and.too.long\] rules out cycles of length [$\smash{{\oldgg}(\log n)^3}$]{} in the regime [$\smash{|{\bar{c}}-c_\circ|\lesssim 1/(\log n)^3}$]{}, which together with implies .
Variance of typical-profile cycles {#s:second.moment}
==================================
In this section we complete the proof of our main result Theorem \[t:main\], the key ingredient of which is to demonstrate that cycles of length [$\smash{\asymp (\log n)^3}$]{} exist in the regime [$\smash{|c - c_\circ| \lesssim 1/(\log n)^3}$]{}. (We also argue that it is unlikely that much shorter cycles exist in this weight range.) We prove the existence of these cycles by a second-moment computation on a set of “good” cycles, as we now formally define.
Good cycles
-----------
Recall [$\smash{A_\circ\equiv\log n}$]{}, and let [$\smash{A\equiv \lceil A_\circ\rceil-\Delta}$]{} for [$\smash{\Delta}$]{} a large positive integer, with [$\smash{\Delta\le (\log n)/2}$]{}. Choose [$\smash{c\equiv c_A}$]{} such that [$\smash{ce\lm_A=1}$]{}.
\[d:good\] A [$\smash{k}$]{}-cycle [$\smash{{\mathscr{C}}}$]{}, with weights [$\smash{w_1,\ldots,w_k}$]{} summing to [$\smash{k{\bar{c}}/n}$]{}, is termed [$\smash{\Delta}$]{}-*good* if
(i) [$\smash{c(1-1/k) \le{\bar{c}}\le c}$]{} for [$\smash{c=c_A}$]{}, [$\smash{A= \lceil A_\circ\rceil-\Delta}$]{};
(ii) the process [$\smash{X}$]{} with increments [$\smash{(nw_i/c-1)}$]{} has range [$\smash{\le A-2}$]{},
(iii) the recentered process [$\smash{\widetilde{X}}$]{} (as in ) has [$\smash{\le \Delta \times \delta_A(x)^4}$]{} visits to [$\smash{(x-1,x]}$]{}, for each [$\smash{1\le x\le A}$]{}.
Let [$\smash{\bm{\Omega}_\Delta(k)}$]{} be the number of [$\smash{\Delta}$]{}-good [$\smash{k}$]{}-cycles [$\smash{{\mathscr{C}}}$]{}, and let $${\bm\Omega}_\Delta \equiv \sum_{k=A^3/\Delta}^{A^3\Delta} {\bm\Omega}_\Delta(k)\,.$$ We show by the moment method that for large [$\smash{\Delta}$]{}, [$\smash{\bm{\Omega}_\Delta}$]{} is positive with large probability. To this end we first argue that [$\smash{\bm{\Omega}_\Delta}$]{} is large in expectation:
\[l:large.moment\] For [$\smash{\Delta}$]{} large, [$\smash{\E\bm{\Omega}_\Delta \asymp\Delta}$]{}.
For [$\smash{A^2\lesssim k \lesssim n^{1/2}}$]{}, consider a [$\smash{k}$]{}-cycle [$\smash{{\mathscr{C}}}$]{} that satisfies properties (i) and (ii). The conditional probability that (iii) *fails* is, by taking a union bound over [$\smash{1\le x\le A}$]{} and applying Lemma \[l:expected.profile\] with Markov’s inequality, $$\lesssim \sum_{x=1}^A
\f{ 1}
{\Delta\de_A(x)^2}
\lesssim 1/\Delta.$$ It follows, assuming [$\smash{\Delta}$]{} large enough, that the expectation of [$\smash{\bm{\Omega}_\Delta(k)}$]{} is equal up to constants to the expected number of [$\smash{k}$]{}-cycles satisfying properties (i) and (ii) only. By Lemma \[l:unif.cycles\], $$\E \bm{\Omega}_\Delta(k)
\asymp
\E[Z^k_c(A)-Z^k_{c(1-1/k)}(A)]
\asymp
\f{(ce\lm_A)^k}{A^3} =\f{1}{A^3}.$$ Summing over [$\smash{\Delta^{-1}\le k/A^3 \le\Delta}$]{} proves the claim.
Variance bound
--------------
The crux of the proof is in the following lemma which argues that the count of good cycles has low variance.
\[l:conditional\] For any cycle [$\smash{{\mathscr{C}}}$]{}, $$\sum_{{\mathscr{C}}'\ne{\mathscr{C}}}
\P(\text{{$\smash{{\mathscr{C}}'}$} is {$\smash{\Delta}$}-good} \,|\,
\text{{$\smash{{\mathscr{C}}}$} is {$\smash{\Delta}$}-good})
\le \E\bm{\Omega}_\Delta+ o_\Delta(1)$$ where [$\smash{o_\Delta(1)}$]{} indicates an error tending to zero in the limit of [$\smash{\Delta\to\infty}$]{}.
Fix [$\smash{\Delta}$]{} large, and fix a [$\smash{k}$]{}-cycle [$\smash{{\mathscr{C}}}$]{}. Condition on the event that [$\smash{{\mathscr{C}}}$]{} is good (from here on, “good” means [$\smash{\Delta}$]{}-good), and write $$\bm{\Sigma}({\mathscr{C}})
\equiv
\sum_{{\mathscr{C}}'\ne{\mathscr{C}}}
\P(\text{{$\smash{{\mathscr{C}}'}$} is good} \,|\,
\text{{$\smash{{\mathscr{C}}}$} is good}).$$ If a cycle [$\smash{{\mathscr{C}}'}$]{} is disjoint from [$\smash{{\mathscr{C}}}$]{} then the events [$\smash{\set{\text{{$\smash{{\mathscr{C}}}$} is good}}}$]{} and [$\smash{\set{\text{{$\smash{{\mathscr{C}}'}$} is good}}}$]{} are independent: thus, the contribution to [$\smash{\bm{\Sigma}({\mathscr{C}})}$]{} from all [$\smash{{\mathscr{C}}'\cap{\mathscr{C}}=\emptyset}$]{} is at most [$\smash{\E\bm{\Omega}_\Delta}$]{}. It remains then to argue that the contribution from cycles intersecting with [$\smash{{\mathscr{C}}}$]{} is [$\smash{o_\Delta(1)}$]{}.
To this end, consider [$\smash{{\mathscr{C}}'\ne{\mathscr{C}}}$]{} such that [$\smash{{\mathscr{C}}\cap {\mathscr{C}}'}$]{} consists of [$\smash{q}$]{} shared segments, with [$\smash{q\ge1}$]{}. Let [$\smash{r}$]{} count the vertices in [$\smash{{\mathscr{C}}'\setminus{\mathscr{C}}}$]{}. The cycle [$\smash{{\mathscr{C}}'}$]{} is partitioned into alternating *in-segments* (contained in [$\smash{{\mathscr{C}}}$]{}) and *out-segments* (edge-disjoint from [$\smash{{\mathscr{C}}}$]{}); the combined length of all the out-segments is [$\smash{r+q}$]{}. Label the in-segments [$\smash{[u_j,v_j]}$]{}, [$\smash{1\le j\le q}$]{}, in order of their traversal by [$\smash{{\mathscr{C}}'}$]{} — [$\smash{{\mathscr{C}}}$]{} may traverse the segments in a different order. Let [$\smash{\mathbf{n}_\text{out}(r,q)}$]{} count the number of such cycles [$\smash{{\mathscr{C}}'}$]{} for a given tuple [$\smash{({\mathscr{C}},r,q,(u_j,v_j)_{1\le j\le q})}$]{}, $$\mathbf{n}_\text{out}(r,q)
\le n^r \binom{r+q-1}{q-1}
\le
n^r (r+q)^{q-1}$$ (since [$\smash{{\mathscr{C}}'}$]{} is given by choosing an ordered sequence of [$\smash{r}$]{} vertices, then dividing them in to [$\smash{q}$]{} groups).
Let [$\smash{X'}$]{} be the process with increments [$\smash{(nw'_i/c-1)}$]{}, where [$\smash{w'_i}$]{} are the edge weights on [$\smash{{\mathscr{C}}'}$]{}. The edge weights in [$\smash{{\mathscr{C}}'\setminus{\mathscr{C}}}$]{} are still i.i.d. unit-rate exponential random variables after the conditioning on [$\smash{{\mathscr{C}}}$]{}. For [$\smash{1\le j\le q}$]{} let integers [$\smash{y_j,z_j}$]{} be defined by $$\begin{array}{rll}
y_j\hspace{-6pt}&=\lceil X'_{v_j} \rceil
-\lceil X'_{u_j}\rceil
&\text{(discretized increment over {$\smash{j}$}-th in-segment);}\\
z_j\hspace{-6pt}&=\lceil X'_{u_{j+1}}\rceil
-\lceil X'_{v_j}\rceil
&\text{(discretized increment over {$\smash{j}$}-th out-segment).}
\end{array}$$ Write [$\smash{\bm{y}\equiv(y_1,\ldots,y_q)}$]{} and likewise [$\smash{\bm{z}\equiv(z_1,\ldots,z_q)}$]{}. Let [$\smash{\overline{y}}$]{} denote the average of the [$\smash{y_j}$]{}, and likewise [$\smash{\overline{z}}$]{} the average of the [$\smash{z_j}$]{}. For [$\smash{{\mathscr{C}}'}$]{} to satisfy property (i), we must have [$\smash{q(\overline{y}+\overline{z})\in\set{-1,0}}$]{}. For each [$\smash{j}$]{}, [$\smash{y_j+A\ge1}$]{} and [$\smash{-z_j+A\ge1}$]{}. Thus, given [$\smash{\overline{y}}$]{}, the number of choices for compatible [$\smash{(\bm{y},\bm{z})}$]{} is at most $$\mathbf{a}_q(\overline{y})
\equiv \binom{q(A+\overline{y})-1}{q-1}
\sum_{i=0}^1
\binom{q(A+\overline{y})+i-1}{q-1}
\le [e^{O(1)}(A+\overline{y})]^{2q}.$$ Next, by property (iii), the number of choices of [$\smash{(u_j,v_j)}$]{} consistent with [$\smash{\bm{y}}$]{} is at most $$\prod_{j=1}^q\Big\{
\Delta^2
\sum_x \delta_A(x)^4 \delta_A(x+|y_j|)^4\Big\}
\le
\Delta^{2q}
\prod_{j=1}^q\Big\{
(A-|y_j|)
\max_x
\big[
(x+1)^4 (A-|y_j|-x+1)^4
\big]\Big\},$$ where we have used that there are [$\smash{\le A-|y_j|}$]{} choices of [$\smash{x}$]{} for which the summand is positive. For each [$\smash{j}$]{} the maximum is achieved with [$\smash{x+1=(A-|y_j|)/2}$]{}, and combining with Jensen’s inequality the above is $$\le
\Delta^{2q}\prod_{j=1}^q (A-|y_j|)^9
\le\Delta^{2q} \Big(
A- q^{-1}\sum_j |y_j|
\Big)^{9q}
\le \Delta^{2q}(A+\overline{y})^{9q}
\equiv \mathbf{b}_q(\overline{y}).$$ Lastly, let [$\smash{\mathbf{p}(\bm{z})}$]{} be the probability that [$\smash{X'}$]{} has increments as specified by [$\smash{\bm{z}}$]{}, with range at most [$\smash{A}$]{} (otherwise [$\smash{{\mathscr{C}}'}$]{} would violate property (ii)). Writing [$\smash{\ell_j}$]{} for the length of the [$\smash{j}$]{}-th out-segment, a similar calculation as in the proof of Lemma \[l:unif.paths\] gives $$\mathbf{p}(\bm{z})
\le
\prod_{j=1}^q\bigg\{
O(c/n)
\f{f_{\ell_j}[ (\ell_j+z_j)c/n]}
{f_{\ell_j}(\ell+z_j)}
\max_x p_A^{\ell_j}(x,x+z_j)
\bigg\}.$$ Applying and making some straightforward manipulations (recalling [$\smash{ce\lm_A=1}$]{}), we arrive at the conclusion that for any [$\smash{\bm{z}}$]{} consistent with [$\smash{\bm{y}}$]{}, $$\mathbf{p}(\bm{z})\le
\Big(\f{ce\lm_A}{n}\Big)^{r}
\Big[
\f{e^{O(1)}
ce\lm_A}{ n e^{-\overline{z}} }
\f{(A-\overline{z})^2}{A^3}
\Big]^q \le e^{O(q)}
\mathbf{c}_{r,q}(\overline{y}),
\quad
\mathbf{c}_{r,q}(\overline{y})\equiv
\f{1}{n^r}
\Big[
\f{(A+\overline{y})^2}
{ n e^{\overline{y}} A^3}
\Big]^q.$$ Combining these estimates gives $$\bm{\Sigma}({\mathscr{C}})-\E\bm{\Omega}_\Delta
\le e^{O(q)} \sum_{q,r,\overline{y}}
\mathbf{n}_\textup{out}(r,q)
\mathbf{a}_q(\overline{y})
\mathbf{b}_q(\overline{y})
\mathbf{c}_{r,q}(\overline{y})
\le \sum_{q,r,\overline{y}}
\Big[
\f{e^{O(1)} \Delta^2
(A+\overline{y})^{13}
}{ ne^{\overline{y}} }
\Big]^{q}
\f{(r+q)^{q-1}}{A^{3q}}$$ where the sum is taken over [$\smash{q\ge1}$]{}, [$\smash{q\le r+q \le \Delta A^3}$]{}, and [$\smash{|\overline{y}|\le A-2}$]{} with [$\smash{q\overline{y}}$]{} integer-valued. Making the change of variables [$\smash{Y \equiv A+\overline{y}\ge2}$]{}, we find $$\bm{\Sigma}({\mathscr{C}})-\E\bm{\Omega}_\Delta
\le
\sum_{q, Y}
\Big[
\f{ \Delta^3
}{ n/e^A }
\f{Y^{13}}{e^Y}
\Big]^q
\le
\sum_{q, Y}
\Big[
\f{ \Delta^3
}{ e^\Delta }
\f{Y^{13}}{e^Y}
\Big]^q
\lesssim \f{ \Delta^3
}{ e^\Delta },$$ which tends to zero in the limit [$\smash{\Delta\to\infty}$]{} as claimed.
\[c:whp\] In the limit [$\smash{\Delta\to\infty}$]{}, [$\smash{\P(\bm{\Omega}_\Delta=0)\to0}$]{}.
By Chebychev’s inequality and Lemma \[l:conditional\], $$\P(\bm{\Omega}_\Delta=0)
\le\f{\text{Var}\,|\bm{\Omega}_\Delta|}{(\E|\bm{\Omega}_\Delta|)^2}
\le \f{\E|\bm{\Omega}_\Delta| + o_\Delta(1)}{(\E|\bm{\Omega}_\Delta|)^2}.$$ By Lem. \[l:large.moment\] the right-hand side tends to zero in the limit [$\smash{\Delta\to\infty}$]{}.
Conclusion
----------
We conclude this section with the proof of our main result.
Theorem \[t:first\] and Corollary \[c:whp\] together imply that we can choose [$\smash{C_1=C_1(\ep)}$]{} sufficiently large so that $$\P\left(\left.
\begin{array}{c}
c_\circ-C_1/(\log n)^3
\le n{\mathscr{W}}_n
\le c_\circ+C_1/(\log n)^3\\
\text{and
{$\smash{(\log n)^2/C_1\le {\mathscr{L}}_n
\le C_1(\log n)^3}$}}
\end{array}
\,\right|\,n{\mathscr{W}}_n>1/e
\right)\ge1-\ep.$$ Combining this with Lemma \[l:unif.cycles\] and Lemma \[c:too.non.unif\], we can choose [$\smash{C_2=C_2(\ep,C_1)}$]{} sufficiently large so that for any interval [$\smash{I\subseteq[0,C_1]}$]{} with [$\smash{|I| \le 1/C_2}$]{}, $$\P\left(\begin{array}{c}
\exists\text{ cycle }{\mathscr{C}}:
[{\bar{c}}-c_\circ](\log n)^3
\in I\\
\text{or
{$\smash{\operatorname{len}({\mathscr{C}})/(\log n)^3
\in I}$}}
\end{array}
\right)\le\ep.$$ Combining these proves and .
Eigenfunctions of the one-minus-exp walk {#s:eig}
========================================
Recall that the one-minus-exp random walk is the real-valued stochastic process [$\smash{(X_t)_{t\ge0}}$]{} whose jumps [$\smash{X_{t+1}-X_t}$]{} are independent and identically distributed as [$\smash{1-{\operatorname{\textsc{Exp}}}(1)}$]{}. In this section we study eigenfunctions of this walk with killing outside an interval [$\smash{[0,H]}$]{}.
Eigenfunctions and eigenvalues {#ss:eig.intro}
------------------------------
Consider the one-minus-exp walk killed outside [$\smash{[0,H]}$]{}, and suppose that [$\smash{g}$]{} is a left eigenfunction of this process with eigenvalue [$\smash{\lm}$]{}: this means (see Remark \[r:left.right\]) that [$\smash{g}$]{} is a function supported on [$\smash{[0,H]}$]{}, not identically zero, satisfying \[e:left.eigenvalue.eqn\] g(x) = e\^[x-1]{}\_0\^H g(u) e\^[-u]{} du, 0x H. As [$\smash{g}$]{} is not identically zero, [$\smash{\set{x : g(x)\ne0}}$]{} is a non-empty subset of [$\smash{[0,H]}$]{}, and we denote its infimum by [$\smash{x_{\min}}$]{}. We then have $$\lm g(x) e^{1-x} = \int_0^H g(u) e^{-u}\,du,
\quad 0\le x \le \min\set{H,1+x_{\min}}.$$ — including for some [$\smash{x_{\min} \le x \le \min\set{H,1+x_{\min}}}$]{} for which [$\smash{g(x)\ne0}$]{}. If the integral on the right-hand side vanishes, then [$\smash{\lm=0}$]{}; otherwise [$\smash{\lm\ne0}$]{}, [$\smash{x_{\min}=0}$]{}, and (by rescaling) [$\smash{g(x) = e^x}$]{} for [$\smash{0\le x\le \min\set{H,1}}$]{}. In particular, for [$\smash{H\le1}$]{} we conclude the process has a unique non-zero eigenvalue [$\smash{\lm=H/e}$]{} with associated left eigenvector [$\smash{g(x) = e^x \Ind{0\le x\le H}}$]{}.
We see from that [$\smash{g}$]{} is smooth on [$\smash{(0,H)\setminus\Z}$]{}, with continuous derivatives up to order [$\smash{k}$]{} on [$\smash{(0,H)\setminus\Z_{\le k}}$]{}. We can therefore differentiate to find \[eq:g-ODE\] g’(x) = g(x) - g(x-1),1<x<H. Take for the moment [$\smash{H=\infty}$]{}. We can solve for [$\smash{g}$]{} on the intervals [$\smash{[k,k+1]}$]{} ([$\smash{k\in\Z_{\ge1}}$]{}) one at a time, as follows: if [$\smash{g}$]{} satisfies , then [$\smash{h}$]{} satisfies [$\smash{h'(x) = g(x-1)/\lm}$]{} for all [$\smash{x>1}$]{}, where \[e:from.g.to.h\] h(x) -g(x) + \_1\^x g(u)du,x1. Knowing [$\smash{g(x)}$]{} for [$\smash{1\le x\le \overline{x}}$]{} determines [$\smash{h(x)}$]{} for [$\smash{1\le x\le \overline{x}}$]{}; the reverse also holds since $$0=-g(x) + \int_1^x g(u)\,du,
\quad 0\le x\le\overline{x}$$ has only the trivial solution [$\smash{g=0}$]{}. Suppose inductively that [$\smash{g}$]{} (hence [$\smash{h}$]{}) has been determined on [$\smash{[0,k]}$]{} for [$\smash{k\in\Z_{\ge1}}$]{}: then we can determine [$\smash{h}$]{} on the next interval [$\smash{[k,k+1]}$]{} by evaluating $$h(x) = h(k)+ \lm^{-1} \int_{k-1}^{x-1} g(u)\,du,
\quad
k\le x \le k+1,$$ which determines [$\smash{g}$]{} on [$\smash{[k,k+1]}$]{}. This proves that in the case [$\smash{H=\infty}$]{}, there is a unique continuous function [$\smash{g= g_\lm}$]{} supported on [$\smash{[0,\infty)}$]{} that satisfies [$\smash{g(x)=e^x}$]{} for [$\smash{0\le x\le1}$]{}, and satisfies for [$\smash{x>1}$]{}. It is straightforward to verify that \[eq:g\] g\_(x) e\^x \_[k=0]{}\^[x-1]{} ,x0 is such a function, so it must be the unique one.
![The left plot shows $g_1(x)$, and the right plot shows $g_{0.95}(x)$.[]{data-label="fig:g_lambda(x)"}](g-lambda=1 "fig:"){width="45.00000%"}![The left plot shows $g_1(x)$, and the right plot shows $g_{0.95}(x)$.[]{data-label="fig:g_lambda(x)"}](g-lambda=0-95 "fig:"){width="45.00000%"}
If for finite [$\smash{H}$]{} and [$\smash{\lm\ne0}$]{}, [$\smash{g}$]{} is a solution of , then [$\smash{g}$]{} satisfies , and it follows from the preceding discussion that [$\smash{g}$]{} must be the restriction of [$\smash{g_\lm}$]{} to [$\smash{[0,H]}$]{}. We calculate $$\f{1}{\lm e}
\int_a^b g_\lm(u) e^{-u}\,du
= \sum_{k=1}^{\lceil a\rceil}
\f{(a+1-k)^{k}}{(-\lm e)^k k!}
-\sum_{k=1}^{\lceil b\rceil}
\f{(b+1-k)^{k}}{(-\lm e)^k k!}
= \f{g_\lm(a+1)}{e^{a+1}}
-\f{g_\lm(b+1)}{e^{b+1}},$$ so equation with [$\smash{g=g_\lm}$]{} and [$\smash{1\le x\le H}$]{} simplifies to $$0
= \f{g_\lm(x)}{e^x}
-\f{1}{\lm e}\int_{x-1}^H g_\lm(u) e^{-u}\,du
= \f{g_\lm(H+1)}{e^{H+1}},$$ therefore [$\smash{g_\lm(H+1)=0}$]{} which is a polynomial equation in [$\smash{\lm}$]{} of degree [$\smash{\lceil H\rceil}$]{} (Fig. \[fig:lambdas(H)\]).
![The eigenvalues of the walk with two-sided killing as a function of [$\smash{H}$]{}.[]{data-label="fig:lambdas(H)"}](lambdas-of-H){width="50.00000%"}
\[r:left.right\] We comment briefly on left and right eigenfunctions of the process. Formally, the operator for the one-minus-exp walk [$\smash{X_t}$]{} killed outside [$\smash{[0,H]}$]{} is given by $$(Kf)(x)
\equiv \E_x[ f(X_1) \Ind{0\le X_1\le H}]
= \int_0^H f(y) p(x,y) \,dy,$$ where [$\smash{p}$]{} is the transition kernel from [$\smash{x}$]{} to [$\smash{y}$]{}. The left action of [$\smash{K}$]{} is the right action of the adjoint operator, [$\smash{(K^\star f)(y) = (fK)(y)}$]{}. In the current setting, [$\smash{p(x,y) = k(y-x)}$]{} where [$\smash{k}$]{} is the density function of the [$\smash{1-{\operatorname{\textsc{Exp}}}(1)}$]{} random variable. We therefore have $$(K^\star f)(x)
= \int_0^H f(y) k(x-y)\,dy
= \int_0^H f(H-y) k(y-(H-x))\,dy
= (R_H KR_H f)(x),$$ where [$\smash{R_H f(x) \equiv f(H-x)}$]{}. The reflection [$\smash{R_H}$]{} is involutive, and it relates the left and right eigenfunctions of [$\smash{K}$]{}: [$\smash{g}$]{} is a left eigenfunction of [$\smash{K}$]{} (i.e. a right eigenfunction of [$\smash{K^\star}$]{}) if and only if [$\smash{R_H g}$]{} is a right eigenfunction of [$\smash{K}$]{} with the same eigenvalue. If [$\smash{f_1,f_2}$]{} are right eigenfunctions of [$\smash{K}$]{} with non-zero eigenvalues [$\smash{\lm_1,\lm_2}$]{} then, writing [$\smash{\langle\cdot,\cdot\rangle}$]{} for the [$\smash{L^2[0,H]}$]{} inner product, $$\langle R_H f_1,f_2 \rangle
= (\lm_2)^{-1} \langle R_H f_1, K f_2 \rangle
= (\lm_2)^{-1} \langle K^\star R_H f_1, f_2 \rangle
= (\lm_1/\lm_2)\langle R_H f_1,f_2 \rangle.$$ Consequently, if [$\smash{\lm_1\ne\lm_2}$]{} then [$\smash{\langle R_H f_1,f_2 \rangle=0}$]{}.
Series expansion
----------------
We now review the solution obtained by Wright [@wright] for general homogeneous difference-differential equations with constant coefficients, \[e:Lambda.equals.zero\] (y)(x) \_[=0]{}\^m\_[=0]{}\^n a\_ y\^[()]{}(x+b\_)=0, a\_, 0=b\_0<b\_1<…<b\_m, where [$\smash{y^{(\nu)}}$]{} denotes the [$\smash{\nu}$]{}-th derivative of [$\smash{y}$]{}. Observe that $$\Lm(e^{sx})
= e^{sx} \tau(s) \text{ where }
\tau(s)\equiv\sum_{\mu=0}^m\sum_{\nu=0}^n
a_{\mu\nu} e^{b_\mu s} s^\nu,$$ so if [$\smash{\tau(s)=0}$]{} then [$\smash{y(x)=e^{s x}}$]{} solves [$\smash{\Lm(y)=0}$]{}. More generally, we can apply the formula [$\smash{(fg)^{(\nu)}
= \sum_{r=0}^\nu\binom{\nu}{r} f^{(r)} g^{(\nu-r)}}$]{} to calculate $$\Lm(x^\ell e^{sx})
= e^{sx}
\sum_{\mu=0}^m\sum_{\nu=0}^n
a_{\mu\nu}
\sum_{r\ge0}
\f{(\nu)_r (\ell)_r }{r!}
e^{b_\mu s}
s^{\nu-r}
(x+b_\mu)^{\ell-r},$$ where for [$\smash{r\in\Z_{\ge0}}$]{}, [$\smash{(j)_r}$]{} denotes the falling factorial [$\smash{j(j-1)\cdots(j-r+1)}$]{}, which is zero if [$\smash{r-j\in\Z_{\ge1}}$]{}. Expanding in powers of [$\smash{x}$]{} gives $$\Lm(x^\ell e^{sx})
= e^{sx}
\sum_{j=0}^\ell
\binom{\ell}{j}
x^j
\underbracket{
\sum_{\mu=0}^m\sum_{\nu=0}^n
a_{\mu\nu}
\sum_{r\ge0}
\f{(\nu)_r(\ell-j)_r}{r!}
e^{b_\mu s} s^{\nu-r} (b_\mu)^{\ell-j-r}
}_{\tau^{(\ell-j)}(s)}.$$ If [$\smash{s}$]{} is a root of [$\smash{\tau}$]{} of order [$\smash{\ell+1}$]{}, meaning [$\smash{0=\tau(s)=\tau'(s)=\ldots=\tau^{(\ell)}(s)}$]{}, then [$\smash{\Lm(x^j e^{sx})=0}$]{} for the integers [$\smash{0\le j\le \ell}$]{}.
Assume that [$\smash{m,n\ge1}$]{}, and that the [$\smash{m\times n}$]{} coefficient matrix [$\smash{a_{\mu\nu}}$]{} contains a non-zero entry in each of the first and last rows and columns, which eliminates pure difference equations and pure differential equations. Under these assumptions, it is shown [@wright] that the general solution to is given by a limit of linear combinations of the solutions [$\smash{x^j e^{s x}}$]{} described above; further, it is explained how to compute the coefficients of this linear combination given initial data [$\smash{y^{(\nu)}(0)}$]{} ([$\smash{0\le \nu<n}$]{}) and [$\smash{y^{(n)}(x)}$]{} ([$\smash{0\le x\le b_m}$]{}). Let $$\begin{aligned}
\bm{H}_1(s)
&\equiv \sum_{\mu=1}^m \sum_{\nu=0}^n
a_{\mu\nu} e^{b_\mu s}
\int_0^{b_\mu}
y^{(\nu)}(u) e^{-su}\,du,\\
\bm{H}(s)&\equiv \bm{H}_1(s)
+ \sum_{\mu=0}^m\sum_{\nu=1}^n
a_{\mu\nu} e^{b_\mu s}
\sum_{\iota=1}^{\nu}
s^{\nu-\iota}
y^{(\iota-1)}(0),
\end{aligned}$$ and let [$\smash{F_x(s) \equiv e^{sx}\bm{H}(s)/\tau(s)}$]{}, regarded as a function of [$\smash{s\in\C}$]{}. It is then shown that \[e:wright.formula\] y(x) = \_[s]{} \_[s]{} F\_x where [$\smash{\res_{s} F_x}$]{} denotes the complex residue of [$\smash{F_x}$]{} at a pole [$\smash{s\in\C}$]{}, and the sum is taken over all poles [$\smash{s}$]{}. To ensure convergence, the poles are arranged in an appropriate order [@wright] which is somewhat delicate in general; in our particular setting we will find below that the summation is absolutely convergent.
To avoid the singularity of [$\smash{g'}$]{} at zero, let [$\smash{y(x) \equiv g_\lm(x+\ep)}$]{}, so [$\smash{y}$]{} satisfies $$\lm^{-1}y(x) - y(x+1) + y'(x+1)=0,\quad
-\ep<x< \infty.$$ This clearly corresponds to with [$\smash{m=n=1}$]{}, \[eq:tau\] a\_ = , b\_= , (s)=1/- e\^s + s e\^s. Note that [$\smash{\tau'(s)=e^s s}$]{} and [$\smash{\tau''(s)=e^s(s+1)}$]{}, so if [$\smash{\tau(s)=0}$]{} then [$\smash{s}$]{} is a simple root unless [$\smash{s=0}$]{}, in which case [$\smash{\lm=1}$]{} and [$\smash{s=0}$]{} is a double root. From we have initial data $$y(x) = g_\lm(x+\ep)
= \left\{\hspace{-4pt}\begin{array}{ll}
e^{x+\ep}
&\text{for } x+\ep\in[0,1],\\
e^{x+\ep}[1-(x+\ep-1)/(\lm e)]
&\text{for } x+\ep\in[1,2],
\end{array}\right.$$ which we use to evaluate $$\bm{H}_1(s)
= e^s\int_0^1 [-y(u)+y'(u)] e^{-su}\,du
= \f{-(e^\ep-e^{\ep s})}{\lm(1-s)}, \quad
\bm{H}(s)
= \bm{H}_1(s) + e^{s+\ep},$$ where [$\smash{\bm{H}_1(s)|_{s=1}}$]{} is understood to be [$\smash{-\ep e^\ep}$]{}. If [$\smash{\tau(s)=0}$]{} then [$\smash{\bm{H}(s) = e^{(1+\ep)s}\ne0}$]{}, so the poles of [$\smash{F_x(s) \equiv e^{sx}\bm{H}(s)/\tau(s)}$]{} correspond precisely to the zeroes of [$\smash{\tau(s)}$]{}. We compute $$\res_s F_x
= \left\{\begin{array}{ll}
\displaystyle \lim_{z\to s} (z-s) F_x(z)
= \f{e^{sx} \bm{H}(s)}{\tau'(s)}
= \f{e^{(x+\ep)s}}{s}
&\text{if {$\smash{s}$} is simple pole,}\\
\displaystyle
\lim_{z\to 0} \f{d}{dz} [ z^2 F_x(z) ]
= 2(x+\ep+1/3)
&\text{if
{$\smash{\lm=1,s=0}$}}
\end{array}\right.$$ (recall from above that these are the only two cases for roots of [$\smash{\tau}$]{}). Applying gives \[e:g.lambda.sum.over.poles\] g\_(x) =\_ (0,1), g\_1(x) = 2(x+1/3) + \_ ,modulo issues of convergence to be addressed in the next subsection.
Roots of and Lambert W function
-------------------------------
The roots of [$\smash{\tau}$]{} can be expressed in terms of the Lambert [$\smash{W}$]{} function, which has been very well studied and which we now briefly review.[^2] The number of rooted spanning trees of the complete [$\smash{n}$]{}-vertex graph is [$\smash{n^{n-1}}$]{} (Cayley’s formula). The *tree function* [$\smash{T}$]{} is the associated exponential generating function, $$T(z)
\equiv \sum_{n=1}^\infty
\f{n^{n-1}}{n!} z^n.$$ From Stirling’s formula ([$\smash{n!\sim \sqrt{2\pi n}(n/e)^n}$]{}), the sum converges for all $|z|\leq 1/e$. It satisfies the relation [$\smash{T(z) = z\exp\{ T(z)\}}$]{} (see e.g. [@stanley Proposition 5.3.1]). The *Lambert [$\smash{W}$]{} function* is defined by the equation \[e:def.Lambert.W\] z = W(z) e\^[W(z)]{}. This is a multivalued function, with branches [$\smash{W_k}$]{} naturally indexed by the integers [$\smash{k\in\Z}$]{}; see [@Lambert-W §4] and Fig. \[f:lambert.w\].
![(This is a duplicate of [@Lambert-W Fig. 4].) The Lambert [$\smash{W}$]{} function is defined as the (multi-valued) inverse of the function [$\smash{f(w) = w e^w}$]{}. The curves in the figure show the preimage set [$\smash{f^{-1}(\R_{<0})}$]{} — they are a subset of [$\smash{\R\cup\set{w:\xi=-\eta\cot\eta}}$]{} where [$\smash{\xi\equiv\real w}$]{} and [$\smash{\eta\equiv\imag w}$]{}. The curves naturally partition the [$\smash{w}$]{}-plane into branches indexed by [$\smash{k\in\Z}$]{} (the [$\smash{k}$]{}-th branch is the image of [$\smash{W_k}$]{}). Following convention [@Lambert-W], branches are defined to be closed in the direction of increasing [$\smash{\imag w}$]{}.[]{data-label="f:lambert.w"}](wexpw-minusreal-preim){width="50.00000%"}
The principal branch [$\smash{W_0}$]{} satisfies [$\smash{W_0(z) = -T(-z)}$]{} for [$\smash{|z|\leq 1/e}$]{}, and can be defined elsewhere by analytic continuation. It is straightforward to check that [$\smash{T(1/e) = 1}$]{}. For [$\smash{\delta\in\C}$]{} with [$\smash{|e^{-\delta}|\le1}$]{}, we can use [$\smash{T(1/e) = 1}$]{} and the relation [$\smash{T(z) = z\exp\{ T(z)\}}$]{} to deduce that \[eq:treefunctionatcritical\] T(1/e\^[1+]{}) = 1 - + O(). The branch [$\smash{W_0}$]{} has a cut on [$\smash{z\in(-\infty,-1/{e})}$]{}; and in view of , it has a square-root-type singularity near [$\smash{z=-1/{e}}$]{}. Any other branch [$\smash{W_k}$]{} ([$\smash{k\ne0}$]{}) has a cut on [$\smash{z\in(-\infty,0)}$]{} with a logarithmic singularity near [$\smash{z=0}$]{}.
Recalling , the solutions of [$\smash{\tau(s)=0}$]{} are given by \[e:def.s.k.using.W.k\] s\_k=1+w\_k, w\_kW\_k(-1/([e]{})), k.For [$\smash{\lm\in(0,1]}$]{}, [$\smash{s_k}$]{} is obtained by evaluating [$\smash{W_k}$]{} precisely on the branch cut. Following the convention that branches are closed in the direction of increasing [$\smash{\imag w}$]{} ([@Lambert-W], see also Fig. \[f:lambert.w\]), we see that [$\smash{s_k}$]{} and [$\smash{s_{-k-1}}$]{} are complex conjugates for each [$\smash{k\ge0}$]{}. When [$\smash{\lm=1}$]{}, [$\smash{s_0}$]{} and [$\smash{s_{-1}}$]{} are obtained by evaluating [$\smash{W_0}$]{} and [$\smash{W_{-1}}$]{} at the branch point [$\smash{-1/e}$]{}, giving the double root [$\smash{s_0=s_{-1}=0}$]{}. There is a convergent series expansion [@Lambert-W eqn. 4.20] for each branch [$\smash{W_k}$]{}; truncating the series and evaluating at [$\smash{-1/({e}\lambda)}$]{} gives, for positive integers [$\smash{k}$]{}, \[eq:s\_k-asymptotic\]
[rl]{} s\_k &= - +i(4k+1)/2 + O(1+k\^[-1]{}[k]{}),\
s\_[-k-1]{} &= - -i(4k+1)/2 + O(1+k\^[-1]{}[k]{}).
Thus, for any [$\smash{x\in\R}$]{} there exists a finite constant [$\smash{C_x}$]{} such that [$\smash{|\exp\{s_k x\}/s_k| \le C_x/|k|^{1+x}}$]{} for all [$\smash{k\in\Z}$]{}, excluding the case [$\smash{s_0=s_{-1}=0}$]{}. It follows that for [$\smash{x>0}$]{}, the summations in are absolutely convergent as claimed.
![Complex roots of [$\smash{\tau(s)
=1/\lambda - e^s(1-s)}$]{}, shown for [$\smash{\lambda=0.95}$]{}.[]{data-label="fig:roots-alpha"}](alphas-0-95){width="50.00000%"}
![Taking [$\smash{\lm=0.95}$]{}, the figure shows [$\smash{g_{0.95}(x)}$]{} with its one-term and two-term approximations in the series expansion (i.e., using the rightmost or two rightmost conjugate pairs of [$\smash{s_k}$]{} from Fig. \[fig:roots-alpha\]). The three curves are nearly indistinguishable except for negative or very small $x$. []{data-label="fig:series-x"}](g-0-95-approx){width="50.00000%"}
Asymptotics for near 1
----------------------
We now extract the asymptotics of [$\smash{g_\lm(x)}$]{} when [$\smash{\lm\in(0,1]}$]{} is close to [$\smash{1}$]{}, and [$\smash{x}$]{} is large. Recalling that [$\smash{s_k}$]{} and [$\smash{s_{-k-1}}$]{} are complex conjugates for [$\smash{k\in\Z_{\ge0}}$]{}, we can re-express as \[eq:g-series\] g\_(x) = f\_(x) + 2\_[k1]{} (e\^[s\_k x]{}/s\_k), f\_(x) = {
[ll]{} 2(e\^[s\_0x]{}/s\_0) & (0,1),\
2(x+1/3) & =1.
. It is strongly suggested by and Fig. \[fig:roots-alpha\] that [$\smash{f_\lm(x)}$]{} is a good approximation to [$\smash{g_\lm(x)}$]{} in the limit of large positive [$\smash{x}$]{}. We shall prove for [$\smash{\lm}$]{} sufficiently near [$\smash{1}$]{} that this is indeed true, which amounts to proving that [$\smash{\real s_k}$]{} is strictly maximized over [$\smash{k\in\Z_{\ge0}}$]{} at [$\smash{k=0}$]{}. Recalling , it is clearly equivalent to prove that [$\smash{\real w_k}$]{} is strictly maximized over [$\smash{k\in\Z_{\ge0}}$]{} at [$\smash{k=0}$]{}. For [$\smash{\lm}$]{} near [$\smash{1}$]{}, [$\smash{w_0}$]{} lies near [$\smash{W(-1/e)=-1}$]{}. It is clear from the definition of the branch cuts ([@Lambert-W §4] and Fig. \[f:lambert.w\]) that for [$\smash{\lambda\in(0,1]}$]{} and [$\smash{k\ge0}$]{}, \[Im-W\_k-range\] w\_k \[2k,(2k+1)), so [$\smash{|w_k|\ge2\pi}$]{} for [$\smash{k\ge1}$]{}. Combining with gives $$\real w_k = -\log|e\lm|-\log|w_k|
\le -1-\log(2\pi)-\log|\lm|
\le -2
\text{ for }
|\lm|\text{ near }1\,.$$ Combining with , we see that there is a finite constant [$\smash{C}$]{} such that for all [$\smash{\lm}$]{} in a neighborhood of [$\smash{1}$]{}, [$\smash{|g_\lm(x) - f_\lm(x)|\le C e^{-2x}}$]{} for all [$\smash{x>0}$]{}.
We next approximate [$\smash{f_\lm(x)}$]{} for [$\smash{\lm = e^{-\delta}}$]{}, [$\smash{\delta}$]{} a small positive real. Taking [$\smash{z=-1/(e\lm)}$]{} and [$\smash{p=\sqrt{2(1+ez)}}$]{}, for [$\smash{|p|<\sqrt{2}}$]{} there is a convergent series expansion [@Lambert-W eqn. (4.22)] $$s_0=1+W_0(z)
= \sum_{\ell\ge1}\mu_\ell p^\ell
= p-\f13 \, p^2 + \f{11}{72} \, p^3 + O(p^4)\,.$$ Since [$\smash{p= i\sqrt{2(e^\delta-1)} = i \sqrt{2\delta}[1+O(\delta)]}$]{}, the first term in the expansion matches what we have already noted in . Therefore $$\real s_0 = (2\delta/3)[1+O(\delta)]\,,
\quad\quad
\imag s_0 = \sqrt{2\delta}[1+O(\delta)]\,,$$ and as a result, for [$\smash{\lm=e^{-\delta}}$]{} with [$\smash{\delta>0}$]{} small, $$\begin{array}{rl}
f_\lm(x)
\hspace{-6pt}&= 2\exp\{(\real s_0)x\}
\sin\Big[ (\imag s_0)x + \arctan[(\real s_0)/(\imag s_0)] \Big] \Big/ |s_0|
\\
\hspace{-6pt}&= 2\exp\{ (2\delta/3)x[1+O(\delta)]\}
\sin\Big[
[1+O(\delta)]
\sqrt{2\delta}(x+1/3)
\Big] \Big/ \sqrt{2\delta}\,.
\end{array}$$ In particular, taking [$\smash{\delta\downarrow0}$]{} with [$\smash{x}$]{} fixed, we recover [$\smash{f_1(x) = 2(x+1/3)}$]{}.
Lastly we identify the value of [$\smash{H}$]{} for which [$\smash{g_\lm}$]{}, [$\smash{\lm=e^{-\delta}}$]{}, gives the principal eigenfunction. Recalling the discussion of §\[ss:eig.intro\], [$\smash{H+1}$]{} corresponds to the smallest positive zero of [$\smash{g_\lm}$]{}. From the above, [$\smash{f_\lm(x_0)=0}$]{} for $$x_0 \equiv \f{\pi-\arctan[(\real s_0)/(\imag s_0)]}
{\imag s_0}
= \f{\pi}{\sqrt{2\delta}} - \f13
+ O(\sqrt{\delta})\,.$$ Recalling that [$\smash{f_\lm}$]{} approximates [$\smash{g_\lm}$]{}, we can write $$\begin{array}{rl}
g_\lm(x_0+x)
\hspace{-6pt}&= O(\exp\{-\Omega(x_0+x)\})
+ f_\lm(x_0+x)\\
\hspace{-6pt}&=O(\exp\{-\Omega(x_0+x)\})
+ 2\exp\{(\real s_0)(x_0+x)\}
\sin[ (\imag s_0) x ]\,.
\end{array}$$ From this we see that we can choose [$\smash{x_\star=x_0 + \exp\{-\Omega(x_0)\}}$]{} and a sufficiently large constant [$\smash{x'}$]{} such that [$\smash{g_\lm}$]{} has a root at [$\smash{x_\star}$]{}, and is non-vanishing between [$\smash{x'}$]{} and [$\smash{x_\star}$]{}. To rule out zeroes on [$\smash{[0,x']}$]{}, note that [$\smash{g_1}$]{} is non-zero on all of [$\smash{\R_{\ge0}}$]{}, since no finite [$\smash{H}$]{} has an eigenvalue of [$\smash{1}$]{}. It follows by continuity in [$\smash{g_\lm}$]{} that for [$\smash{\lm}$]{} sufficiently near [$\smash{1}$]{}, [$\smash{g_\lm}$]{} is non-vanishing on [$\smash{[0,x']}$]{}. Therefore $$H = x_\star-1
= \f{\pi}{\sqrt{2\delta}} - \f43
+ O(\sqrt{\delta})\,, \quad\quad
\delta=
\f{\pi^2}{2}\f{1+O(1/H^2)}{(H+4/3)^2}\,.$$
[^1]: Research supported in part by $^*$NSF grant DMS-1313596 and $^\dagger$NSF MSPRF grant DMS-1401123
[^2]: This function also arose in the analysis [@mathieu-wilson] of the subcritical regime of the minimum mean-weight cycle, for seemingly different reasons. It figures prominently in the analysis of random graphs near the phase transition, e.g. [@JKLP]; see [@Lambert-W §2] for additional applications. Some of the discussion in this section is adapted from [@mathieu-wilson].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $A_n$ be the alternating group of even permutations of $X:=\{1,2,\cdots,n\}$ and ${\mathcal E}_n$ the set of even derangements on $X.$ Denote by $A\T_n^q$ the tensor product of $q$ copies of $A\T_n,$ where the Cayley graph $A\T_n:=\T(A_n,{\mathcal E}_n)$ is called the even derangement graph. In this paper, we intensively investigate the properties of $A\T_n^q$ including connectedness, diameter, independence number, clique number, chromatic number and the maximum-size independent sets of $A\T_n^q.$ By using the result on the maximum-size independent sets $A\T_n^q$, we completely determine the full automorphism groups of $A\T_n^q.$'
author:
- |
Yun-Ping Deng$^{1}$, Fu-Ji Xie$^2$ and Xiao-Dong Zhang$^{1,\dagger}$\
[$^1$Department of Mathematics, Shanghai Jiao Tong University, Shanghai 200240, P.R.China]{}\
[$^2$ Antai College of Economics & Management Shanghai Jiao Tong University, Shanghai 200052, P.R.China]{}\
[Emails: [email protected], [email protected], [email protected]]{}\
title: ' Maximum-Size Independent Sets and Automorphism Groups of Tensor Powers of the Even Derangement Graphs [^1]'
---
\[section\] \[theorem\][Corollary]{} \[theorem\][Definition]{} \[theorem\][Conjecture]{} \[theorem\][Question]{} \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Example]{} \[theorem\][Problem]{}
c Ø ø $H$ ¶[$P$ ]{} ł
[[**Key words:**]{} Automorphism group; Cayley graph; tensor product; maximum-size independent sets; alternating group. ]{}\
[[**AMS Classifications:**]{} 05C25, 05C69]{} 0.5cm
Introduction
============
For a simple graph $\T,$ we use $V(\T), E(\T)$ and $\Aut(\T)$ to denote its vertex set, edge set and full automorphism group, respectively. We denote by $N_{\T}(v)$ the neighbourhood of a vertex $v$ in $\T.$ Let be a finite group and $S$ a subset of not containing the identity element $1$ with $S=S^{-1}.$ The [*Cayley graph*]{} $\T:=\T(G,S)$ on with respect to $S$ is defined by $$V(\T){=}G,~E(\T){=}\{(g,sg): g{\in} G,\ s{\in} S\}.$$ Clearly, $\T$ is a $|S|$-regular and vertex-transitive graph, since $\Aut(\T)$ contains the right regular representation $R(G)$ of . Moreover, $\T$ is connected if and only if is generated by $S.$
Let $S_n$ be the [*symmetric group*]{} and $A_n$ the [*alternating group*]{} on $X=\{1,2,\cdots,n\}.$ Let ${\mathcal
D}_n:=\{\o\in S_n:x^{\,\sigma}\neq x,\forall x\in X\}$ and ${\mathcal E}_n:={\mathcal D}_n\cap A_n$ denote the degrangements and the even derangements on respectively. Then the graph $\T_n:=\T(S_n,{\mathcal D}_n)$ and $A\T_n:=\T(A_n,{\mathcal E}_n)$ are called the [*derangement graph*]{} [@Paul] and the [*even derangement graph*]{} on respectively.
The [*tensor product*]{} $\T_1\otimes\T_2$ of two graphs $\T_1$ and $\T_2$ is the graph with vertex set $V(\T_1)\times V(\T_2)$ and edge set consisting of those pairs of vertices $(u_1,u_2),\,(v_1,v_2)$ where $u_1$ is adjacent to $v_1$ in $\T_1$ and $u_2$ is adjacent to $v_2$ in $\T_2$. A [*projection*]{} is a homomorphism $pr_{i,n}:\T^q\rightarrow\T$ given by $pr_{i,n}(x_1,x_2,\cdots,x_q)=x_i$ for some $i,$ where $\T^q$ is the tensor product of $q$ copies of a graph $\T.$ By the definition of tensor product, it is easy to see that $A\T_n^q$ is the Cayley graph $\T(A_n^q,{\mathcal E}_n^q),$ where $A_n^q$ is the direct product of $q$ copies of $A_n$ and ${\mathcal
E}_n^q:=\{(\sigma_1,\sigma_2,\cdots,\sigma_q):\sigma_i\in{\mathcal
E}_n,i=1,2,\cdots,q\}.$
A family $I\subseteq S_n$ is [*intersecting*]{} if any two elements have at least one common entry. It is easy to see that an intersecting family of maximal size in $S_n$ corresponds to a maximum-size independent set in $\T_n.$ In [@Cameron], Cameron and Ku showed that the only intersecting families of maximal size in $S_n$ are the cosets of point stabilizers. In [@Ku], Ku and Wong proved that analogous results hold for the alternating group and the direct product of symmetric groups, which equivalently shows that the structure of maximum-size independent sets of $A\T_n$ is as follows:
(Theorem 1.2 in ${{\fs\cite{Ku}}}$)\[pr-1.1\] All the maximum-size independent sets of $A\T_n~(n\geq5)$ are $B_{i,j}=\{\sigma\in A_n:\,i^{\sigma}=j\},\,i,j=1,2,\cdots,n.$ In particular, each $|B_{i,j}|=\frac{(n-1)!}{2}.$
In this paper, we prove that the result analogous to [@Cameron] holds for the direct product of the alternating groups, which can be equivalently stated as follows:
\[dl-1.2\] All the maximum-size independent sets of $A\T_n^q~(q\geq1,n\geq5)$ are $$B_{i,j}^{(k)}=\{(\sigma_1,\sigma_2,\cdots,\sigma_q)\in A_n^q:
\,i^{\sigma_k}=j\},\,i,j=1,2,\cdots,n;\,k=1,2\cdots,q.$$ In particular, the independence number of $A\T_n^q$ is $$|B_{i,j}^{(k)}|=\frac{(n-1)!n!^{q-1}}{2^q}.$$
[**Remark.**]{} Generally speaking, for a graph $\T,$ all maximum-size independent sets of $\T^q$ are not necessarily preimages of maximum-size independent sets of $\T$ under projections (see [@Ku2; @Larose2]). Theorem \[dl-1.2\] shows that all maximum-size independent sets of $A\T_n^q$ are preimages of maximum-size independent sets of $A\T_n$ under projections.
Many researchers (see [@Cameron; @deza1977; @Eggleton1985; @Larose; @Paul; @Rasmussen]) have studied the properties of $\T_n,$ such as the clique number, the chromatic number, the independence number, maximum-size independent sets and so on. Motivated by the nice structures of $\T_n,$ here we show that $A\T_n^q$ have the similar nice structures. For example, we obtain that the diameter $D(A\T_n^q)=2,$ the clique number $\omega(A\T_n^q)=n$ and the chromatic number $\chi(A\T_n^q)=n.$
Cayley graphs are of general interest in the field of Algebraic Graph Theory due to their good properties, especially their high symmetry. One difficult problem in Algebraic Graph Theory is to determine the automorphism groups of Cayley graphs. Although there are some nice results on the automorphism groups of Cayley graphs (see [@Fang; @Feng2; @Feng1; @Godsil2; @Huan; @Xu; @Z; @Zhang]), we still lack enough understanding on them. In this paper, we completely determine the automorphism groups of $A\T_n^q,$ which in fact gives a kind of method on the computation of automorphism group of Cayley graph by using the characterization of the maximum-size independent sets. Another main result of this paper is as follows:
\[dl-1.3\] Define the mapping $\varphi_k:\,A_n^q\rightarrow A_n^q$ as $(\sigma_1,\cdots,\sigma_{k-1},\sigma_k,\sigma_{k+1},\cdots,\sigma_q)^{\varphi_k}=(\sigma_1,\cdots,\sigma_{k-1},\sigma_k^{-1},\sigma_{k+1},\cdots,\sigma_q)$ for $k=1,2,\cdots,q.$ For $q\geq1$ and $n\geq 5,$ $$\Aut(A\T_n^q)=(R(A_n^q)\rtimes (\Inn(S_n)\wr S_q))\rtimes Z_2^q,$$ where $\Inn(S_n)\,(\,\cong S_n)$ is the inner automorphism group of $S_n,$ $Z_2^q=\langle\varphi_1\rangle\times\langle\varphi_2\rangle\times\cdots\langle\varphi_q\rangle$ and $\,\Inn(S_n)\wr S_q$ denotes the wreath product of $\,\Inn(S_n)$ and $S_q.$
[**Remark.**]{} Sanders and George [@Sanders] showed that for a graph $\T,$ $\Aut(\T^2)\geq\Aut(\T)\wr S_2,$ where $\wr$ denotes the wreath product, however, the equality cannot hold in most situations. Theorem \[dl-1.3\] implies that $\Aut(A\T_n^q)=\Aut(A\T_n)\wr S_q.$
The rest part of this paper is organized as follows. In Section 2, we give the connectedness and diameter of $A\T_n^q.$ In Section 3, we determine the independence number and the structure of maximum-size independent sets of $A\T_n^q,$ as its corollary, we obtain the clique number and chromatic number of $A\T_n^q.$ In section 4, we completely determine the full automorphism groups of $A\T_n^q.$
The connectedness and diameter
==============================
In this section, we give the connectedness and diameter of $A\T_n^q.$
For a group $G,$ we denote the automorphism group and the inner automorphism group of $G$ by $\Aut(G)$ and $\Inn(G),$ respectively. Next we need the following known result:
${{\fs\cite{Suzuki}\,[III,\,(2.17)-(2.20)]}}$ \[pr-2.1\] If $n\geq2$ and $n\neq6,$ then $\Aut(A_n)=\Inn(S_n).$ If $n=6,$ then $|\Aut(A_6):\Inn(S_6)|=2,$ and for each $\alpha\in
\Aut(A_6){\setminus}\Inn(S_6),$ $\alpha$ maps a $3$-cycle to a product of two disjoint $3$-cycles.
\[yl-2.2\] If $n\geq 5,$ then the even derangement graph $A\T_n$ is connected.
By Theorem 2.8 of page 293 in [@Suzuki], the alternating group $A_n~(n\geq 5)$ is generated by the totality of $3$-cycles. Clearly $(1\,2\,3)=(1\,2\,\cdots\,n)^2\cdot(n\,n-1\,\cdots\,1)^2(1\,2\,3)$ and $(1\,2\,\cdots\,n)^2,\,(n\,n-1\,\cdots\,1)^2(1\,2\,3)\in
{\mathcal E}_n$ by $n\geq 5.$
For any $3$-cycle $(i\,j\,k),$ there exists a $\phi\in \Inn(S_n)$ such that $(1\,2\,3)^{\phi}=(i\,j\,k).$ By Proposition \[pr-2.1\], we have $\Aut(A_n,{\mathcal E}_n)=\{\phi\in
\Aut(A_n):\,{\mathcal E}_n^{\,\phi}={\mathcal E}_n\}=\Inn(S_n).$ Thus
$$(i\,j\,k)=(1\,2\,3)^{\phi}=[(1\,2\,\cdots\,n)^2]^{\phi}\cdot[(n\,n-1\,\cdots\,1)^2(1\,2\,3)]^{\phi}$$ and $$[(1\,2\,\cdots\,n)^2]^{\phi},\,[(n\,n-1\,\cdots\,1)^2(1\,2\,3)]^{\phi}
\in {\mathcal E}_n.$$ So the alternating group $A_n~(n\geq 5)$ is generated by ${\mathcal E}_n,$ which implies that $A\T_n$ is connected.
[**Remark.**]{} If $n=3,$ clearly $A_3=\lg {\mathcal E}_3\rg,$ so $A\T_3$ is connected. If $n=4,$ then $A_4\neq\lg {\mathcal
E}_4\rg=\{1,\,(1\,2)(3\,4),\,(1\,3)(2\,4),\,(1\,4)(2\,3)\},$ so $A\T_4$ is not connected.
${}^{{\fs\cite{Graham}}}$\[yl-2.3\] (i) The tensor product of two connected graphs is bipartite if and only if at least one of them is bipartite.
\(ii) The tensor product of two connected graphs is disconnected if and only if both factors are bipartite.
\[dl-2.4\] $A\T_n^q$ is connected and non-bipartite for any $q\geq 1$ and $n\geq 5.$
Since $A\T_n^q=\underbrace{A\T_n\otimes\cdots\otimes A\T_n}_q$ and $A\T_n$ is connected and non-bipartite for $n\geq 5$ by Lemma \[yl-2.2\], the assertion holds by Lemma \[yl-2.3\].
\[yl-2.5\] For any $g_1,g_2\in A_n~(n\geq 5),$ there exists a $g\in A_n$ such that $g\in N_{A\T_n}(g_1)\cap N_{A\T_n}(g_2).$
If $n=5,$ we have $$(a_1,a_2,a_3,a_4,a_5)=(a_1,a_4,a_2,a_5,a_3)^2,$$ $$(a_1,a_2)(a_3,a_4)=(a_5,a_1,a_3,a_2,a_4)(a_1,a_5,a_3,a_2,a_4),$$ $$(a_1,a_2,a_3)=(a_1,a_5,a_3,a_4,a_2)(a_1,a_3,a_5,a_2,a_4),$$ $$1=(a_1,a_2,a_3,a_4,a_5)(a_5,a_4,a_3,a_2,a_1),$$ that is, for any $x\in A_5,$ there exist $s_1,s_2\in {\mathcal E}_5$ such that $x=s_1s_2.$ Now for $x=g_1g_2^{-1},$ we have $g_1g_2^{-1}=s_1s_2,\,s_1,s_2\in {\mathcal E}_5,$ i.e. $g_1=s_1s_2g_2.$ Set $g:=s_2g_2.$ Clearly $g\in N_{A\T_5}(g_1)\cap
N_{A\T_5}(g_2).$
If $n\geq 6,$ by proposition 6 in [@Cameron], for any $g_1,g_2\in A_n,$ there exists a $g\in S_n$ such that $g\in
N_{\T_n}(g_1)\cap N_{\T_n}(g_2).$ That is, there exist $s_1,\,s_2\in
{\mathcal D}_n$ such that $g=s_1g_1=s_2g_2.$ If $g\in A_n,$ then $s_1,\,s_2\in {\mathcal E}_n,$ so $g\in N_{A\T_n}(g_1)\cap
N_{A\T_n}(g_2)$ and the assertion holds. If $g\in S_n\setminus A_n,$ then $s_1,\,s_2\in {\mathcal D}_n\setminus {\mathcal E}_n.$ For any $i\in X=\{1,2,\cdots,n\},$ select a $j\in\{i,i^{s_1},i^{s_2},i^{s_1^{-1}},i^{s_2^{-1}}\}\neq
\emptyset~(n\geq 6).$ Set $$g':=(i\,j)g,\,s_1':=(i\,j)s_1,\,s_2':=(i\,j)s_2.$$ Thus $g'=s_1'g_1=s_2'g_2$ and $s_1',s_2'\in {\mathcal E}_n$ by $j\in
X\setminus\{i,i^{s_1},i^{s_2},i^{s_1^{-1}},i^{s_2^{-1}}\}.$ Hence $g'\in N_{A\T_n}(g_1)\cap N_{A\T_n}(g_2)$ and the assertion holds.
\[dl-2.6\] If $n\geq5,$ then $diam(A\T_n^q)=2,$ where $diam(A\T_n^q)$ is the diameter of $A\T_n^q.$
For any $(\sigma_1,\sigma_2,\cdots,\sigma_q),(\tau_1,\tau_2,\cdots,\tau_q)\in
A_n^q,$ by Lemma \[yl-2.5\], there exist $\varsigma_i\in
A_n~(i=1,2,\cdots,q)$ such that $\varsigma_i\in
N_{A\T_n}(\sigma_i)\cap N_{A\T_n}(\tau_i).$ So there exists a $(\varsigma_1,\varsigma_2,\cdots,\varsigma_q)\in A_n^q$ such that $$(\varsigma_1,\varsigma_2,\cdots,\varsigma_q)\in
N_{A\T_n^q}((\sigma_1,\sigma_2,\cdots,\sigma_q))\cap
N_{A\T_n^q}((\tau_1,\tau_2,\cdots,\tau_q)),$$ which implies that any two vertices in $A\T_n^q$ have at least a common neighbourhood. Hence $diam(A\T_n^q)=2.$
The stucture of maximum-size independent sets
=============================================
In this section we characterize the structure of maximum-size independent sets of $A\T_n^q$ for $q\geq 1,$ which is a generalization of Theorem 1.2 in [@Ku]. First we give the independence number of $A\T_n^q$ as follows:
\[yl-3.1\] For any $q\geq 1,\,n\geq 5,$ the independence number of $A\T_n^q$ is given by $$\alpha(A\T_n^q)=\frac{(n-1)!n!^{q-1}}{2^q}.$$
By Proposition 1.3 in [@Alon], we have
$$\frac{\alpha(A\T_n^q)}{|A_n^q|}=\frac{\alpha(A\T_n)}{|A_n|}
\Rightarrow\alpha(A\T_n^q)=\frac{\alpha(A\T_n)\cdot|A_n^q|}{|A_n|}.$$ Then by Proposition \[pr-1.1\], we obtain $$\alpha(A\T_n^q)=\frac{\frac{(n-1)!}{2}\cdot(\frac{n!}{2})^q}{\frac{n!}{2}}
=\frac{(n-1)!n!^{q-1}}{2^q}.$$ Thus the assertion holds.
For any two graphs $H_1$ and $H_2,$ a map $\phi$ from $V(H_1)$ to $V(H_2)$ is [*homomorphism*]{} if $\{u^{\phi},v^{\phi}\}\in E(H_2)$ whenever $\{u,v\}\in E(H_1),$ i.e. $\phi$ is a edge-preserving map. Next we need the following fundamental result of Albertson and Collins [@Albertson] which is also called ’No-Homomorphism Lemma’.
${{\fs\cite{Albertson}}}$\[yl-3.2\] Let $H_1$ and $H_2$ be graphs such that $H_2$ is vertex transitive and there exists a homomorphism $\phi:~V(H_1)\rightarrow V(H_2).$ Then $$\label{2}
\frac{\alpha(H_1)}{|V(H_1)|}\geq\frac{\alpha(H_2)}{|V(H_2)|}$$ Furthermore, if equality holds in (\[2\]), then for any independent set $I$ of cardinality $\alpha(H_2)$ in $H_2$, $I^{{\phi}^{-1}}$ is an independent set of cardinality $\alpha(H_1)$ in $H_1.$
\[yl-3.3\] Let $H=(V_1,V_2,E)$ be a $d$-regular bipartite graph whose partition has the parts $V_1$ and $V_2$ with $|V_1|=|V_2|.$ If $H$ is connected, then $|S|<|N_H(S)|$ for any $S\subsetneq V_1,$ where $N_H(S)$ is the neighborhood of $S$ in $H.$
Let $T=N_H(S)$ and $E(S,T)=\{(s,t)\in E:s\in S,t\in T\}.$ Then $$d|S|=|E(S,T)|\leq |E(V_1,T)|=d|T|.$$ If $|S|=|T|,$ then $E(S,T)|=|E(V_1,T)|,$ i.e. any vertex $u\in S\cup T$ is not adjacent to any vertex $v\not\in S\cup T,$ which contradicts the connectedness of $H.$ Thus $|S|<|T|=|N_H(S)|.$
\[yl-3.4\] All the maximum-size independent sets of $A\T_n^2~(n\geq7)$ are $$B_{i,j}^{(k)}=\{(g_1,g_2)\in A_n^2:
\,i^{g_k}=j\},\,i,j=1,2,\cdots,n;\,k=1,2.$$
Set ${\mathcal B}=\{B_{i,j}^{(k)}:i,j=1,2,\cdots,n;\,k=1,2\}.$ Clearly $|B_{i,j}^{(k)}|=\frac{(n-1)!n!}{4},$ which is equal to $\alpha(A\T_n^2)$ by Lemma \[yl-3.1\]. That is, $B_{i,j}^{(k)}$ is a maximum-size independent set of $A\T_n^2.$ Next for any maximum-size independent set $I$ of $A\T_n^2,$ it suffices to show that $I\in{\mathcal B}.$
Define a homomorphism $\phi$ from $A\T_n$ to $A\T_n^2$ as $g^{\phi}=(g,g).$ Without loss of generality, we may assume that the identity $Id=(id,id)\in I.$ By Proposition \[pr-1.1\], Lemmas \[yl-3.1\] and \[yl-3.2\], $I^{{\phi}^{-1}}$ is a maximum independent set of $A\T_n.$ So $I^{{\phi}^{-1}}=\{g\in
A_n:~i_0^g=j_0\}$ for some $i_0,\,j_0$ by Proposition \[pr-1.1\]. Since $id=(Id)^{{\phi}^{-1}}\in I^{\phi^{-1}},$ $I^{\phi^{-1}}=\{g\in A_n:~i_0^g=i_0\}$ for some $i_0.$ Therefore $I\supseteq I_0:=(I^{{\phi}^{-1}})^{\phi}=\{(g,g)\in
A_n^2:~i_0^g=i_0\}.$ Next we shall show that $I\in{\mathcal B}$ by the following four Claims:\
[**Claim 1.**]{} For any $(g_1,g_2)\in I,$ either $i_0^{g_1}=i_0$ or $i_0^{g_2}=i_0.$
Suppose on the contrary that $i_0^{g_1}\neq i_0$ and $i_0^{g_2}\neq
i_0.$ By Lemma \[yl-2.5\], there exists a $g\in A_n$ such that $g\in N_{A\T_n}(g_1)\cap N_{A\T_n}(g_2).$ That is, there exist $s_1,s_2\in {\mathcal E}_n$ such that $g=s_1g_1=s_2g_2.$
If $i_0^g=i_0,$ then $(g,g)\in I_0\subseteq I$ and $\{(g,g),(g_1,g_2)\}\in E(A\T_n^2),$ which contradicts the fact that $(g_1,g_2)\in I$ and $I$ is an independent set.
If $i_0^g\neq i_0,$ select a $j\in X\setminus
\{i_0,i_0^{g^{-1}},i_0^{s_1},i_0^{s_2},
i_0^{g^{-1}s_1^{-1}},i_0^{g^{-1}s_2^{-1}}\}\neq\emptyset~(n\geq 7).$ Set $$g'=(i_0,i_0^{g^{-1}},j)g,\,s_1'=(i_0,i_0^{g^{-1}},j)s_1,\,s_2'=(i_0,i_0^{g^{-1}},j)s_2.$$ Thus $i_0^{g'}= i_0,\,g'=s_1'g_1=s_2'g_2$ and $s_1',s_2'\in
{\mathcal E}_n$ by $j\in X\setminus
\{i_0,i_0^{g^{-1}},i_0^{s_1},i_0^{s_2},
i_0^{g^{-1}s_1^{-1}},i_0^{g^{-1}s_2^{-1}}\}.$ So $(g',g')\in
I_0\subseteq I$ and$\{(g',g'),(g_1,g_2)\}\in E(A\T_n^2),$ which as above yields a contradiction.
Hence Claim 1 holds.\
Set $$J_0=\{(g_1,g_2)\in A_n^2:i_0^{g_1}=i_0~and~i_0^{g_2}=i_0\},$$ $$J_1=\{(g_1,g_2)\in A_n^2:i_0^{g_1}=i_0~and~i_0^{g_2}\neq i_0\},$$ $$J_2=\{(g_1,g_2)\in A_n^2:i_0^{g_1}\neq i_0~and~i_0^{g_2}=i_0\}.$$
Clearly $|J_0|=\frac{(n-1)!^2}{4},|J_1|=|J_2|=\frac{(n-1)(n-1)!^2}{4}.$\
[**Claim 2.**]{} $A\T_n^2[J_1\cup J_2]$ is connected, where $A\T_n^2[J_1\cup J_2]$ denote the induced subgraph of $A\T_n^2$ by $J_1\cup J_2.$
For any $(\sigma_1,\sigma_2),(\tau_1,\tau_2)\in J_1,$ clearly they are not adjacent in $A\T_n^2.$ By Theorem \[dl-2.6\], there exists a $(\varsigma_1,\varsigma_2)\in A_n^2$ such that $\{(\sigma_1,\sigma_2),(\varsigma_1,\varsigma_2)\}$ and $\{(\tau_1,\tau_2),(\varsigma_1,\varsigma_2)\}\in E(A\T_n^2).$ That is, there exist $s_1,s_2,t_1,t_2\in {\mathcal E}_n$ such that $\varsigma_1=s_1\sigma_1=t_1\tau_1,\,\varsigma_2=s_2\sigma_2=t_2\tau_2.$ Clearly $i_0^{\varsigma_1}=i_0^{s_1\sigma_1}\neq i_0.$
If $i_0^{\varsigma_2}=i_0,$ then $(\varsigma_1,\varsigma_2)\in J_2.$
If $i_0^{\varsigma_2}\neq i_0,$ then select a $j\in X\setminus
\{i_0,i_0^{\varsigma_2^{-1}},i_0^{s_2},i_0^{t_2},
i_0^{\varsigma_2^{-1}s_2^{-1}},i_0^{\varsigma_2^{-1}t_2^{-1}}\}\neq\emptyset~(n\geq
7).$ Set $$\varsigma_2'=(i_0,i_0^{\varsigma_2^{-1}},j)\varsigma_2,\,s_2'=(i_0,i_0^{\varsigma_2^{-1}},j)s_2,\,t_2'=(i_0,i_0^{\varsigma_2^{-1}},j)t_2.$$ Thus and $i_0^{\varsigma_2'}=
i_0,\,\varsigma_2'=s_2'\sigma_2=t_2'\tau_2$ and $s_2',t_2'\in
{\mathcal E}_n$ by $j\in X\setminus
\{i_0,i_0^{\varsigma_2^{-1}},i_0^{s_2},i_0^{t_2},
i_0^{\varsigma_2^{-1}s_2^{-1}},i_0^{\varsigma_2^{-1}t_2^{-1}}\}.$ So $(\varsigma_1,\varsigma_2')\in J_2$ and $\{(\sigma_1,\sigma_2),(\varsigma_1,\varsigma_2')\},\{(\tau_1,\tau_2),(\varsigma_1,\varsigma_2')\}\in
E(A\T_n^2[J_1\cup J_2]).$
Similarly, for any $(\sigma_1,\sigma_2),(\tau_1,\tau_2)\in J_2,$ their exists $(\varsigma_1,\varsigma_2)\in J_1$ such that $\{(\sigma_1,\sigma_2),(\varsigma_1,\varsigma_2)\},\\
\{(\tau_1,\tau_2),(\varsigma_1,\varsigma_2)\}\in
E(A\T_n^2[J_1\cup J_2]).$
Hence Claim 2 holds.\
[**Claim 3.**]{} Either $I\cap J_1=\emptyset$ or $I\cap
J_2=\emptyset.$
Suppose on the contrary that $I\cap J_1\neq\emptyset$ and $I\cap
J_2\neq\emptyset,$ consider the following two possible cases:
[**Case 1.**]{} $I\cap J_1=J_1$ or $I\cap J_2=J_2.$
Since $I\cap(J_1\cup J_2)$ is an independent set, this case cannot happen.
[**Case 2.**]{} $I\cap J_1\subsetneq J_1$ and $I\cap J_2\subsetneq
J_2.$
It is easy to see that $A\T_n^2[J_1\cup J_2]$ is a regular bipartite graph whose partition has the parts $J_1$ and $J_2$ with $|J_1|=|J_2|.$ By Claim 2 and Lemma \[yl-3.3\], we have $$|I\cap J_1|<|N_{A\T_n^2[J_1\cup J_2]}(I\cap J_1)|.$$ Since $I\cap(J_1\cup J_2)$ is an independent set, we have $$\begin{aligned}
&&I\cap J_2\subseteq J_2\setminus N_{A\T_n^2[J_1\cup J_2]}(I\cap J_1)\\
&\Rightarrow&|N_{A\T_n^2[J_1\cup J_2]}(I\cap J_1)|+|I\cap J_2|\leq
|J_2|\\
&\Rightarrow&|I\cap J_1|+|I\cap J_2|<|J_2|\end{aligned}$$ By Claim 1, $I=\bigcup_{i=0}^2(I\cap J_i).$ Since $J_i~(i=0,1,2)$ are pairwise disjoint, we have $$\begin{aligned}
|I|&=&|I\cap J_0|+|I\cap J_1|+|I\cap J_2|\\
&<&|J_0|+|J_2|\\
&=&\frac{(n-1)!^2}{4}+\frac{(n-1)(n-1)!^2}{4}\\
&=&\frac{(n-1)!n!}{4}\end{aligned}$$ which is a contradiction, since $|I|=\frac{(n-1)!n!}{4}$ by Lemma \[yl-3.1\]. Hence Claim 3 holds.\
[**Claim 4.**]{} Either $I=J_0\cup J_1$ or $I=J_0\cup J_2.$
By Claim 3, either $I\cap J_1=\emptyset$ or $I\cap J_2=\emptyset.$ If $I\cap J_1=\emptyset,$ then we have $$\begin{aligned}
&&\frac{(n-1)!n!}{4}=|I|=|I\cap J_0|+|I\cap J_2|\leq |J_0|+|J_2|=\frac{(n-1)!n!}{4}\\
&\Rightarrow&I\cap J_0=J_0,\,I\cap J_2=J_2.\end{aligned}$$
So $I=\bigcup_{i=0}^2(I\cap J_i)=J_0\cup J_2.$
Similarly, if $I\cap J_2=\emptyset,$ then $I=J_0\cup J_1.$ Hence Claim 4 holds.\
By Claim 4, we have $I\in B,$ which conclude the proof.
${{\fs\cite{Alon}}}$\[yl-3.5\] Let $\T$ be a connected d-regular graph on n vertices and let $d=\mu_1\geq\mu_2\geq\cdots\mu_n$ be the eigenvalues of the adjacency matrix of $\T.$ If $$\frac{\alpha(\T)}{n}=\frac{-\mu_n}{d-\mu_n},$$ then for every integer $q\geq1,$ $$\frac{\alpha(\T^q)}{n^q}=\frac{-\mu_n}{d-\mu_n}.$$ Moreover, if $\T$ is also non-bipartite, and if $I$ is an independent set of size $\frac{-\mu_n}{d-\mu_n}n^q$ in $\T^q,$ then there exists a coordinate $i\in\{1,2,\cdots,q\}$ and a maximum-size independent set $J$ in $\T,$ such that $$I=\{(v_1,\cdots,v_q\in V(\T^q):v\in J\}.$$
${{\fs\cite{Ku2}}}$\[yl-3.6\] Let $\T$ be a connected, non-bipartite vertex-transitive graph. Suppose that the only independent sets of maximal cardinality in $H^2$ are the preimages of the independent sets of maximal cardinality in $\T$ under projections. Then the same holds for all powers of $\T.$
(of Theorem \[dl-1.2\]). For $n=5,6,$ it is easy to see that $A\T_n$ is connected, non-bipartite and $e(n)$-regular graph with $e(5)=24$ and $e(6)=130.$ Moreover, a [**Matlab**]{} computation shows that the least eigenvalue of the adjacency matrix of $A\T_5$ and $A\T_6$ are $-6$ and $-26,$ respectively. Thus the assertion holds by Lemmas \[yl-3.1\] and \[yl-3.5\].
For $n\geq 7,$ combining Proposition \[pr-1.1\], Lemma \[yl-3.4\] and \[yl-3.6\], the assertion holds.
\[co-3.7\] Let $\omega(A\T_n^q)$ and $\chi(A\T_n^q)$ denote the clique number and chromatic number of $A\T_n^q~(n\geq 5).$ Then we have $$\omega(A\T_n^q)=\chi(A\T_n^q)=n.$$
By [@Ku], we have $\omega(A\T_n)=n.$ Let $\{\sigma_1,\sigma_2,\cdots,\sigma_n\}$ be a clique of $A\T_n.$ Then clearly $\{(\sigma_1,\sigma_1,\cdots,\sigma_1),(\sigma_2,\sigma_2,\cdots,\sigma_2),\cdots,(\sigma_n,\sigma_n,\cdots,\sigma_n)\}$ is a clique of $A\T_n^q.$ So we have $\omega(A\T_n^q)\geq n.$ On the other hand, by Theorem \[dl-1.2\], we know that the independence number $\alpha(A\T_n^q)=\frac{(n-1)!n!^{q-1}}{2^q}.$ By Corollary 4 in [@Cameron], we have $\omega(A\T_n^q)\alpha(A\T_n^q)\leq
|V(A\T_n^q)|,$ that is $\omega(A\T_n^q)\cdot\frac{(n-1)!n!^{q-1}}{2^q}\leq
\frac{n!^q}{2^q},$ so $\omega(A\T_n^q)\leq n.$ Thus $\omega(A\T_n^q)=n.$
In addition, by Corollary 6.1.3 in [@Godsil], for any Cayley graph $\T:=\T(G,S),$ if $S$ is closed under conjugation and $\alpha(\T)\omega(\T)=|V(\T)|,$ then $\chi(\T)=\omega(\T).$ Note that for $A\T_n^q=\T(A_n^q,{\mathcal E}_n^q),\,{\mathcal E}_n^q$ is closed under conjugation and $\alpha(A\T_n^q)\omega(A\T_n^q)=|V(A\T_n^q)|.$ Hence $\chi(A\T_n^q)=\omega(A\T_n^q)=n.$
The automorphism group of $A\T_n^q$
===================================
In this section, we completely determine the full automorphism group of $A\T_n^q~(n\geq 5).$ First we introduce some definitions. Let $\Sym(\Omega)$ denote the set of all permutations of a set $\Omega.$ A [*permutation representation*]{} of a group is a homomorphism from into $\Sym(\Omega)$ for some set $\Omega.$ A permutation representation is also referred to as an action of on the set $\Omega,$ in which case we say that acts on $\Omega.$ Furthermore, if $\{g\in G:x^g=x,\,\forall x\in \Omega\}=1,$ we say the action of $G$ on $\Omega$ is [*faithful*]{}, or $G$ acts [*faithfully*]{} on $\Omega.$
Next we need the following known results:
${{\fs\cite{Johnson}}}$ \[pr-4.1\] Let $G^q=G\times G\times\cdots\times G$ be the external direct product of $q$ copies of the nontrivial group $G.$ If $G$ has the following properties:
\(i) the center $Z(G)$ of $G$ is trivial;
\(ii) G cannot be decomposed as a nontrivial direct product.\
Then $\Aut(G^q)=\Aut(G)\wr S_q.$
${{\fs\cite{Godsil3}}}$ \[pr-4.2\] Let $N_{Aut(\T(G,S)}(R(G))$ be the normalizer of $R(G)$ in $\Aut(\T(G,S)).$ Then $$N_{Aut(\T(G,S)}(R(G))=R(G)\rtimes \Aut(G,S)\leq \Aut(\T(G,S)),$$ where $\Aut(G,S)=\{\phi\in \Aut(G):\,S^{\phi}=S\}.$
\[yl-4.3\] Define the mapping $\varphi_k:\,A_n^q\rightarrow A_n^q$ as $(\sigma_1,\cdots,\sigma_{k-1},\sigma_k,\sigma_{k+1},\cdots,\sigma_q)^{\varphi_k}=(\sigma_1,\cdots,\sigma_{k-1},\sigma_k^{-1},
\sigma_{k+1},\cdots,\sigma_q)$ for $k=1,2,\cdots,q.$ For $n\geq 5,$ $$(R(A_n^q)\rtimes
(\Inn(S_n)\wr S_q))\rtimes Z_2^q\leq \Aut(A\T_n^q),$$ where $\Inn(S_n)~\cong S_n$ and $Z_2^q=\langle\varphi_1\rangle\times\langle\varphi_2\rangle\times\cdots\langle\varphi_q\rangle.$ In particular, $|\Aut(A\T_n^q)|\geq |(R(A_n^q)\rtimes (\Inn(S_n)\wr
S_q))\rtimes Z_2^q|=q!n!^{2q}.$
By Proposition \[pr-2.1\] and \[pr-4.1\], we have $$\begin{aligned}
\Aut(A_n^q,{\mathcal E}_n^q)&=&\{\phi\in \Aut(A_n^q):\,({\mathcal E}_n^q)^{\,\phi}={\mathcal E}_n^q\}\\
&=&\{\phi\in \Aut(A_n)\wr Sq:\,({\mathcal E}_n^q)^{\,\phi}={\mathcal E}_n^q\}\\
&=&\Inn(S_n)\wr Sq.\end{aligned}$$
Using Proposition \[pr-4.2\], we obtain $R(A_n^q)\rtimes
(\Inn(S_n)\wr S_q)\leq \Aut(A\T_n^q).$
Next we show that $\varphi_k$ is an automorphism of $A\T_n^q.$ $$\begin{aligned}
&&\{(\sigma_1,\cdots,\sigma_k,\cdots,\sigma_q),(\tau_1,\cdots,\tau_k,\cdots,\tau_q)\}\in E(A\T_n^q)\\
&\Leftrightarrow&\forall~i\in\{1,2,\cdots,n\},\forall~k\in\{1,2,\cdots,q\},\,i^{\sigma_k}\neq i^{\tau_k}\\
&\Leftrightarrow&\forall~i\in\{1,2,\cdots,n\},\forall~k\in\{1,2,\cdots,q\},\,(i^{{\sigma_k}^{-1}})^{\sigma_k}\neq (i^{{\sigma_k}^{-1}})^{\tau_k}\\
&\Leftrightarrow&\forall~i\in\{1,2,\cdots,n\},\forall~k\in\{1,2,\cdots,q\},\,i\neq i^{{\sigma_k}^{-1}\tau_k}\\
&\Leftrightarrow&\forall~i\in\{1,2,\cdots,n\},\forall~k\in\{1,2,\cdots,q\},\,i^{{\tau_k}^{-1}}\neq i^{{\sigma_k}^{-1}}\\
&\Leftrightarrow&\{(\sigma_1,\cdots,\sigma_k^{-1},\cdots,\sigma_q),(\tau_1,\cdots,\tau_k^{-1},\cdots,\tau_q)\}\in E(A\T_n^q)\\
&\Leftrightarrow&\{(\sigma_1,\cdots,\sigma_k,\cdots,\sigma_q)^{\varphi_k},(\tau_1,\cdots,\tau_k,\cdots,\tau_q)^{\varphi_k}\}\in
E(A\T_n^q).\end{aligned}$$ It is easy to see that $\varphi_k\not\in R(A_n^q)$ and $\varphi_k\not\in \Inn(S_n)\wr S_q.$ Hence $$(R(A_n^q)\rtimes
(\Inn(S_n)\wr S_q))\rtimes Z_2^q\leq \Aut(A\T_n^q),$$ where $Z_2^q=\langle\varphi_1\rangle\times\langle\varphi_2\rangle\times\cdots\langle\varphi_q\rangle.$ The assertion holds.
\[yl-4.4\] Let ${\mathcal
B}=\{B_{i,j}^{(k)},\,i,j=1,2,\cdots,n;\,k=1,2\cdots,q\}$, where $B_{i,j}^{(k)}=\{(\sigma_1,\sigma_2,\cdots,\sigma_q)\in
A_n^q:\,i^{\sigma_k}=j\}.$ Then the action of $\Aut(A\T_n^q)$ on ${\mathcal B}$ can be induced by the natural action of $\Aut(A\T_n^q)$ on $A_n^q$, and is faithful. Furthermore, any $\phi\in \Aut(A\T_n^q)$ is a permutation of ${\mathcal B}.$
Obviously, any $\phi\in\Aut(A\T_n^q)$ maps a maximum-size independent set of $A\T_n^q$ to a maximum-size independent set of $A\T_n^q.$ So by Theorem \[dl-1.2\], for any $B_{i,j}^{(k)}\in
{\mathcal B}$ and $\phi\in\Aut(\T_n^q),$ we have ${B_{i,j}^{(k)}}^{\,\phi}\in {\mathcal B}.$
Next we show that if $\phi\in\Aut(A\T_n^q)$ satisfies ${B_{i,j}^{(k)}}^{\,\phi}=B_{i,j}^{(k)}$ for each $B_{i,j}^{(k)}\in
{\mathcal B},$ then $\phi$ is the identity map. In fact, clearly, $$\forall\,(\sigma_1,\sigma_2,\cdots,\sigma_q)\in A_n^q,\,\{(\sigma_1,\sigma_2,\cdots,\sigma_q)\}=\bigcap_{k=1}^q\bigcap_{i=1}^n
B_{i,i^{\,\sigma_k}}^{(k)}.$$ So $$\begin{aligned}
\{(\sigma_1,\sigma_2,\cdots,\sigma_q)^{\,\phi}\}&=&(\bigcap_{k=1}^q\bigcap_{i=1}^n B_{i,i^{\,\sigma_k}}^{(k)})^{\,\phi}\\
&\subseteq& \bigcap_{k=1}^q\bigcap_{i=1}^n {B_{i,i^{\,\sigma_k}}^{(k)}}^{\phi}\\
&=&\bigcap_{k=1}^q\bigcap_{i=1}^nB_{i,i^{\,\sigma_k}}^{(k)}\\
&=&\{(\sigma_1,\sigma_2,\cdots,\sigma_q)\}.\end{aligned}$$ Thus $\phi$ is the identity map.
For any $B_{i,j}^{(k)},\,B_{i^{'},j^{'}}^{(k^{'})}\in {\mathcal B}$ and $\phi\in \Aut(A\T_n^q),$ we have $$\begin{aligned}
B_{i,j}^{(k)}\neq B_{i^{'},j^{'}}^{(k^{'})}&\Leftrightarrow&|B_{i,j}^{(k)}\cup B_{i^{'},j^{'}}^{(k^{'})}|>\frac{(n-1)!n!^{q-1}}{2^q}\\
&\Leftrightarrow&|(B_{i,j}^{(k)}\cup B_{i^{'},j^{'}}^{(k^{'})})^{\phi}|>\frac{(n-1)!n!^{q-1}}{2^q}\\
&\Leftrightarrow&|{B_{i,j}^{(k)}}^{\,\phi}\cup {B_{i^{'},j^{'}}^{(k^{'})}}^{\phi}|>\frac{(n-1)!n!^{q-1}}{2^q}\\
&\Leftrightarrow&{B_{i,j}^{(k)}}^{\,\phi}\neq
{B_{i^{'},j^{'}}^{(k^{'})}}^{\phi}.\end{aligned}$$ Thus $\phi$ is a permutation of ${\mathcal B}.$
\[yl-4.5\] $B_{i,j}^{(k)}\cap B_{i^{'},j^{'}}^{(k^{'})}=\emptyset$ if and only if $k=k^{'}$ and exactly one of $i=i^{'}$ and $j=j^{'}$ holds.
If $k=k^{'}$ and exactly one of $i=i^{'}$ and $j=j^{'}$ holds, then $B_{i,j}^{(k)}\cap B_{i^{'},j^{'}}^{(k^{'})}=\emptyset.$
If $k\neq k^{'},$ then $|B_{i,j}^{(k)}\cap
B_{i^{'},j^{'}}^{(k^{'})}|=|\{(\sigma_1,\sigma_2,\cdots,\sigma_q)\in
A_n^q:\,i^{\sigma_k}=j,\,{i^{'}}^{\sigma_{k^{'}}}=j^{'}\}|=\frac{(n-1)!^2n!^{q-2}}{2^q}.$
If $k=k^{'},\,i= i^{'}$ and $j= j^{'},$ then $B_{i,j}^{(k)}=
B_{i^{'},j^{'}}^{(k^{'})},$ so $B_{i,j}^{(k)}\cap
B_{i^{'},j^{'}}^{(k^{'})}\neq\emptyset.$
If $k=k^{'},\,i\neq i^{'}$ and $j\neq j^{'},$ then $$|B_{i,j}^{(k)}\cap
B_{i^{'},j^{'}}^{(k^{'})}|=\{(\sigma_1,\sigma_2,\cdots,\sigma_q)\in
A_n^q:\,i^{\sigma_k}=j,\,{i^{'}}^{\sigma_k}=j^{'}\}|=\frac{(n-1)!n!^{q-1}}{2^q}.$$
Thus the assertion holds.
\[yl-4.6\] Let ${\mathcal
B}^{(k)}=\{B_{i,j}^{(k)},\,i,j=1,2,\cdots,n\},\,k=1,2,\cdots,q.$ For any $\phi\in\Aut(A\T_n^q),$ There exists a $B_{i,j}^{(k)}\in
{\mathcal B}^{(k)}$ such that ${B_{i,j}^{(k)}}^{\,\phi}\in {\mathcal
B}^{(k^{'})}$ if and only if ${B_{i,j}^{(k)}}^{\,\phi}\in {\mathcal
B}^{(k^{'})}$ for any $B_{i,j}^{(k)}\in {\mathcal B}^{(k)}.$
Suppose on the contrary that there exist two distinct $B_{i,j}^{(k)},\,B_{i^{'},j^{'}}^{(k)}\in {\mathcal B}^{(k)}$ such that ${B_{i,j}^{(k)}}^{\,\phi}\in {\mathcal
B}^{(k^{'})},\,{B_{i^{'},j^{'}}^{(k)}}^{\phi}\in {\mathcal
B}^{(k^{''})}$ with $k^{'}\neq k^{''}.$
Since $B_{i,j}^{(k)}\neq B_{i^{'},j^{'}}^{(k)},$ we have $|B_{i,j}^{(k)}\cap
B_{i^{'},j^{'}}^{(k)}|=0~or~\frac{(n-2)!n!^{q-1}}{2^q}$ by using Lemma \[yl-4.5\] and its proof. So $$\begin{aligned}
|{B_{i,j}^{(k)}}^{\,\phi}\cup
{B_{i^{'},j^{'}}^{(k)}}^{\phi}|&=&|(B_{i,j}^{(k)}\cup
B_{i^{'},j^{'}}^{(k)})^{\,\phi}|=|B_{i,j}^{(k)}\cup
B_{i^{'},j^{'}}^{(k)}|\\
&=&\frac{2(n-1)!n!^{q-1}}{2^q}~or~\frac{2(n-1)!n!^{q-1}-(n-2)!n!^{q-1}}{2^q}.\end{aligned}$$ On the other hand, $$\begin{aligned}
{B_{i,j}^{(k)}}^{\,\phi}\in
B^{(k^{'})},{B_{i^{'},j^{'}}^{(k)}}^{\phi}\in B^{(k^{''})}\,
(k^{'}\neq k^{''})&\Rightarrow& |{B_{i,j}^{(k)}}^{\,\phi}\cap
{B_{i^{'},j^{'}}^{(k)}}^{\phi}|=\frac{(n-1)!^2n!^{q-2}}{2^q}\\
&\Rightarrow& |{B_{i,j}^{(k)}}^{\,\phi}\cup
{B_{i^{'},j^{'}}^{(k)}}^{\phi}|=\frac{2(n-1)!n!^{q-1}-(n-1)!^2n!^{q-2}}{2^q},\end{aligned}$$ which is a contradiction. Thus the assertion holds.
\[yl-4.7\] Let ${\mathcal
R}_i^{(k)}=\{B_{i,1}^{(k)},B_{i,2}^{(k)},\cdots,B_{i,n}^{(k)}\}$ and ${\mathcal
C}_j^{(k)}=\{B_{1,j}^{(k)},B_{2,j}^{(k)},\cdots,B_{n,j}^{(k)}\},\,k=1,2,\cdots,q.$ Then for any $x_1,x_2,\cdots,x_n\in {\mathcal B},$ we have $$x_1\cup x_2\cup\cdots\cup x_n=A_n^q$$ if and only if there exist some $k\in\{1,2,\cdots,q\}$ and some $i$ or $j\in \{1,2,\cdots,n\}$ such that $\{x_1,x_2,\cdots,x_n\}={\mathcal R}_i^{(k)}$ or ${\mathcal
C}_j^{(k)}.$
Clearly if $\{x_1,x_2,\cdots,x_n\}={\mathcal R}_i^{(k)}$ or ${\mathcal C}_j^{(k)}$ for some $k\in\{1,2,\cdots,q\}$ and some $i$ or $j\in \{1,2,\cdots,n\},$ then $x_1\cup x_2\cup\cdots\cup
x_n=A_n^q.$
Assume that $x_1\cup x_2\cup\cdots\cup x_n=A_n^q.$ Since $\forall\,i,\,|x_i|=\frac{(n-1)!n!^{q-1}}{2^q}$ and $|A_n^q|=\frac{n!^q}{2^q},$ we have $x_i\cap
x_j=\emptyset,\,\forall i,j,\,i\neq j.$ Applying Lemma \[yl-4.5\], we obtain $\{x_1,x_2,\cdots,x_n\}={\mathcal R}_i^{(k)}$ or ${\mathcal C}_j^{(k)}.$
\[yl-4.8\] Let $\Omega=\{{\mathcal C}_i^{(k)},{\mathcal
R}_j^{(k)},i,j=1,2,\cdots,n;\,k=1,2,\cdots,q\}.$ Then the action of $\Aut(A\T_n^q)$ on $\Omega$ can be induced by the action of $\Aut(A\T_n^q)$ on ${\mathcal B}$ in Lemma \[yl-4.4\], and it is faithful. Furthermore, any $\phi\in \Aut(A\T_n^q)$ is a permutation of $\Omega.$
First for any ${\mathcal R}_i^{(k)}\in \Omega$ and $\phi\in\Aut(A\T_n^q),$ we have $${B_{i,1}^{(k)}}^{\,\phi}\cup{B_{i,2}^{(k)}}^{\,\phi}\cup\cdots\cup{B_{i,n}^{(k)}}^{\,\phi}=
(B_{i,1}^{(k)}\cup B_{i,2}^{(k)}\cup\cdots\cup
B_{i,n}^{(k)})^{\,\phi}=(A_n^q)^{\,\phi}=A_n^q.$$ So by Lemma \[yl-4.7\], we have ${{\mathcal
R}_i^{(k)}}^{\,\phi}=\{{B_{i,1}^{(k)}}^{\,\phi},{B_{i,2}^{(k)}}^{\,\phi},\cdots,{B_{i,n}^{(k)}}^{\,\phi}\}\in\Omega.$
Similarly, for any ${\mathcal C}_j^{(k)}\in \Omega$ and $\phi\in\Aut(A\T_n^q),$ we have ${{\mathcal C}_j^{(k)}}^{\,\phi}\in
\Omega.$
Assume that $\phi\in \Aut(A\T_n^q)$ satisfies ${{\mathcal
R}_i^{(k)}}^{\,\phi}={\mathcal R}_i^{(k)}$ and ${{\mathcal
C}_j^{(k)}}^{\,\phi}={\mathcal C}_j^{(k)}$ for any $i,j\in
\{1,2,\cdots,n\}$ and $k\in \{1,2,\cdots,q\}.$ Then it suffices to show that $\phi$ is the identity map.
Since for any $B_{i,j}^{(k)}\in {\mathcal B},$ we have $\{{B_{i,j}^{(k)}}^{\,\phi}\}=({\mathcal R}_i^{(k)}\cap {\mathcal
C}_j^{(k)})^{\,\phi} \subseteq {{\mathcal R}_i^{(k)}}^{\,\phi}\cap
{{\mathcal C}_j^{(k)}}^{\,\phi} ={\mathcal R}_i^{(k)}\cap {\mathcal
C}_j^{(k)}=\{B_{i,j}^{(k)}\}.$ By Lemma \[yl-4.4\], the action of $\Aut(A\T_n^q)$ on ${\mathcal B}$ is faithful. Thus $\phi$ is the identity map.
For any $\omega_1,\omega_2\in \Omega$ and $\phi\in \Aut(A\T_n^q),$ $$\begin{aligned}
\omega_1\neq\omega_2&\Leftrightarrow&|\omega_1\cup\omega_2|>n\\
&\Leftrightarrow&|(\omega_1\cup\omega_2)^{\phi}|>n\\
&\Leftrightarrow&|\omega_1^{\phi}\cup\omega_2^{\phi}|>n\\
&\Leftrightarrow&\omega_1^{\phi}\neq\omega_2^{\phi}.\end{aligned}$$ Thus $\phi$ is a permutation of $\Omega.$
\[yl-4.9\] Let ${\mathcal R}^{(k)}=\{{\mathcal R}_1^{(k)},{\mathcal
R}_2^{(k)},\cdots,{\mathcal R}_n^{(k)}\},\,{\mathcal
C}^{(k)}=\{{\mathcal C}_1^{(k)},{\mathcal
C}_2^{(k)},\cdots,{\mathcal C}_n^{(k)}\}$ and $\Omega^{(k)}={\mathcal R}^{(k)}\cup {\mathcal
C}^{(k)},\,k=1,2,\cdots,q.$ For any $\phi\in \Aut(A\T_n^q),$ the following (i)-(iii) hold:
\(i) There exists a $\sigma\in S_q$ such that ${\Omega^{(k)}}^{\,\phi}=\Omega^{(k^{\sigma})},\,k=1,2,\cdots,q.$
\(ii) There exists some ${\mathcal R}_i^{(k)}\in {\mathcal R}^{(k)}$ such that ${{\mathcal R}_i^{(k)}}^{\,\phi}\in {\mathcal
R}^{(k^{'})}$ if and only if ${{\mathcal R}_i^{(k)}}^{\,\phi}\in
{\mathcal R}^{(k^{'})}$ for any ${\mathcal R}_i^{(k)}\in {\mathcal
R}^{(k)};$
\(iii) There exists some ${\mathcal R}_j^{(k)}\in {\mathcal R}^{(k)}$ such that ${{\mathcal R}_j^{(k)}}^{\,\phi}\in {\mathcal
C}^{(k^{'})}$ if and only if ${{\mathcal R}_j^{(k)}}^{\,\phi}\in
{\mathcal C}^{(k^{'})}$ for any ${\mathcal R}_j^{(k)}\in {\mathcal
R}^{(k)}.$
\(i) By Lemma \[yl-4.6\], for any $k\in \{1,2,\cdots,q\}$ there exists a $l\in \{1,2,\cdots,q\}$ such that ${{\mathcal
B}^{(k)}}^{\,\phi}={\mathcal B}^{(l)}.$ Moreover, if $k\neq k^{'},$ then by Lemma \[yl-4.4\], we have ${{\mathcal
B}^{(k)}}^{\,\phi}\neq{{\mathcal B}^{(k^{'})}}^{\,\phi}.$ Thus there exists a $\sigma\in S_q$ such that ${{\mathcal
B}^{(k)}}^{\,\phi}={\mathcal B}^{(k^{\sigma})},\,k=1,2,\cdots,q.$ By Lemma \[yl-4.8\], the assertion holds.
\(ii) First by (i), there exists some ${\mathcal R}_i^{(k)}\in
{\mathcal R}^{(k)}$ such that ${{\mathcal R}_i^{(k)}}^{\,\phi}\in
\Omega^{(k^{'})}$ if and only if ${{\mathcal R}_i^{(k)}}^{\,\phi}\in
\Omega^{(k^{'})}$ for any ${\mathcal R}_i^{(k)}\in {\mathcal
R}^{(k)}.$
Suppose on the contrary that there exist $i,j\,(\neq i)\in
\{1,2,\cdots,n\}$ such that ${{\mathcal R}_i^{(k)}}^{\,\phi}\in
{\mathcal R}^{(k^{'})}$ and ${{\mathcal R}_j^{(k)}}^{\,\phi}\in
{\mathcal C}^{(k^{'})}.$
Note that $${\mathcal
R}_i^{(k)}\cap {\mathcal R}_j^{(k)}=\emptyset~for~i\neq j,$$ $${\mathcal
R}_i^{(k)}\cap {\mathcal C}_j^{(k)}=\{B_{i,j}^{(k)}\}~for~any~i,j.$$ Then $$i\neq j\Rightarrow {\mathcal
R}_i^{(k)}\cap {\mathcal R}_j^{(k)}=\emptyset \Rightarrow |{\mathcal
R}_i^{(k)}\cup {\mathcal R}_j^{(k)}|=2n\Rightarrow |{{\mathcal
R}_i^{(k)}}^{\,\phi}\cup {{\mathcal
R}_j^{(k)}}^{\,\phi}|=|({\mathcal R}_i^{(k)}\cup {\mathcal
R}_j^{(k)})^{\,\phi}|=2n.$$ On the other hand, $${{\mathcal
R}_i^{(k)}}^{\,\phi}\in {\mathcal R}^{(k^{'})},\,{{\mathcal
R}_j^{(k)}}^{\,\phi}\in {\mathcal C}^{(k^{'})}\Rightarrow
|{{\mathcal R}_i^{(k)}}^{\,\phi}\cap {{\mathcal
R}_j^{(k)}}^{\,\phi}|=1\Rightarrow |{{\mathcal
R}_i^{(k)}}^{\,\phi}\cup {{\mathcal R}_j^{(k)}}^{\,\phi}|=2n-1,$$ which is a contradiction. Thus the assertion holds.
\(iii) The proof of (iii) is similar to that of (ii).
\[yl-4.10\] For $n\geq 5,$ we have $$|\Aut(A\T_n^q)|\leq q!n!^{2q}.$$
By (i) of Lemma \[yl-4.9\], for any $\phi\in \Aut(A\T_n^q),$ there exists a $\sigma\in S_q$ such that ${\Omega^{(k)}}^{\,\phi}=\Omega^{(k^{\sigma})}~(k=1,2,\cdots,q).$ Using (ii) and (iii) of Lemma \[yl-4.9\] we obtain the following disjoint alternatives:
\(i) ${{\mathcal R}^{(k)}}^{\,\phi}={\mathcal R}^{(k^{\sigma})}$ and ${{\mathcal C}^{(k)}}^{\,\phi}={\mathcal C}^{(k^{\sigma})};$
\(ii) ${{\mathcal R}^{(k)}}^{\,\phi}={\mathcal C}^{(k^{\sigma})}$ and ${{\mathcal C}^{(k)}}^{\,\phi}={\mathcal R}^{(k^{\sigma})}.$
So $\Aut(A\T_n^q)=\bigcup_{\sigma\in S_q}\{\phi\in
\Aut(A\T_n^q):{\Omega^{(k)}}^{\,\phi}=\Omega^{(k^{\sigma})},k=1,2,\cdots,q\}.$ Hence, if we can prove the last two inequalities, then we have $$\begin{aligned}
|\Aut(A\T_n^q)|&\leq&\Sigma_{\sigma\in S_q}|\{\phi\in
\Aut(A\T_n^q):{\Omega^{(k)}}^{\,\phi}=\Omega^{(k^{\sigma})},k=1,2,\cdots,q\}|\\
&\leq&\Sigma_{\sigma\in S_q}\Pi_{k=1}^q(|\{\phi\in
\Aut(A\T_n^q):{{\mathcal R}^{(k)}}^{\,\phi}={\mathcal
R}^{(k^{\sigma})},\,{{\mathcal C}^{(k)}}^{\,\phi}={\mathcal
C}^{(k^{\sigma})}\}|+\\
&&~~~~~~~~~~~~~~~~~|\{\phi\in \Aut(A\T_n^q):{{\mathcal
R}^{(k)}}^{\,\phi}={\mathcal C}^{(k^{\sigma})},\,{{\mathcal
C}^{(k)}}^{\,\phi}={\mathcal
R}^{(k^{\sigma})}\}|)\\
&\leq&\Sigma_{\sigma\in S_q}\Pi_{k=1}^q(\frac{n!^2}{2}+\frac{n!^2}{2})\\
&=&\Sigma_{\sigma\in S_q}n!^{2q}\\
&=&q!n!^{2q}.\end{aligned}$$
Now we show that $$|\{\phi\in \Aut(A\T_n^q):{{\mathcal
R}^{(k)}}^{\,\phi}={\mathcal R}^{(k^{\sigma})},\,{{\mathcal
C}^{(k)}}^{\,\phi}={\mathcal C}^{(k^{\sigma})}\}|\leq
\frac{n!^2}{2},$$ $$|\{\phi\in \Aut(A\T_n^q):{{\mathcal
R}^{(k)}}^{\,\phi}=C^{(k^{\sigma})},\,{C^{(k)}}^{\,\phi}={\mathcal
R}^{(k^{\sigma})}\}|\leq \frac{n!^2}{2}.$$
Indeed, for any $\phi\in \Aut(A\T_n^q)$ such that ${{\mathcal
R}^{(k)}}^{\,\phi}={\mathcal R}^{(k^{\sigma})},\,{{\mathcal
C}^{(k)}}^{\,\phi}={\mathcal C}^{(k^{\sigma})},$ define $\phi_1,\phi_2\in S_n$ as ${{\mathcal R}_i^{(k)}}^{\phi}={\mathcal
R}_{i^{\,\phi_1}}^{(k^{\sigma})},\,{{\mathcal
C}_j^{(k)}}^{\phi}={\mathcal C}_{j^{\,\phi_2}}^{(k^{\sigma})}.$
Since $\{{B_{ij}^{(k)}}^{\phi}\}=({\mathcal R}_i^{(k)}\cap {\mathcal
C}_j^{(k)})^{\phi}\subseteq {{\mathcal
R}_i^{(k)}}^{\phi}\cap{{\mathcal C}_j^{(k)}}^{\phi}={\mathcal
R}_{i^{\,\phi_1}}^{(k^{\sigma})}\cap {\mathcal
C}_{j^{\,\phi_2}}^{(k^{\sigma})}=\{B_{i^{\,\phi_1}j^{\,\phi_2}}^{(k^{\sigma})}\},$ we have $$\{(1,1,\cdots,1)^{\phi}\}\in (\bigcap_{i=1}^n B_{ii}^{(k)})^{\phi}
\subseteq\bigcap_{i=1}^n {B_{ii}^{(k)}}^{\phi} =\bigcap_{i=1}^n
B_{i^{\,\phi_1}i^{\,\phi_2}}^{(k^{\sigma})}
=\{(\tau_1,\tau_2,\cdots,\tau_q)\in
A_n^q:\tau_{k^{\sigma}}=\phi_1^{-1}\phi_2\}.$$ So $(1,1,\cdots,1)^{\phi}=(\tau_1,\cdots,\tau_{k^{\sigma}-1},\phi_1^{-1}\phi_2,\tau_{k^{\sigma}+1}\cdots,\tau_q)\in
A_n^q,$ which implies that $\phi_1^{-1}\phi_2\in A_n.$
Thus $$\begin{aligned}
&&|\{\phi\in \Aut(A\T_n^q):{{\mathcal
R}^{(k)}}^{\,\phi}={\mathcal R}^{(k^{\sigma})},\,{{\mathcal
C}^{(k)}}^{\,\phi}={\mathcal
C}^{(k^{\sigma})}\}|\\
&=&|\{\phi\in \Aut(A\T_n^q):{{\mathcal R}^{(k)}}^{\,\phi}={\mathcal
R}^{(k^{\sigma})},\,{{\mathcal C}^{(k)}}^{\,\phi}={\mathcal
C}^{(k^{\sigma})},\,\phi_1^{-1}\phi_2\in
A_n\}|\\
&\leq& \frac{n!^2}{2}.\end{aligned}$$
Similarly, $$|\{\phi\in \Aut(A\T_n^q):{{\mathcal
C}^{(k)}}^{\,\phi}={\mathcal R}^{(k^{\sigma})},\,{{\mathcal
C}^{(k)}}^{\,\phi}={\mathcal R}^{(k^{\sigma})}\}|\leq
\frac{n!^2}{2}.$$
Thus the assertion holds.
(of Theorem \[dl-1.3\]). By Lemma \[yl-4.3\], we have $|\Aut(A\T_n^q)|\geq q!n!^{2q}.$ On the other hand, by Lemma \[yl-4.10\], we obtain $|\Aut(A\T_n^q|\leq q!n!^{2q}.$ Hence $|\Aut(A\T_n^q)|=q!n!^{2q},$ and by Lemma \[yl-4.3\] again, the assertion holds.
Albertson M O, Collins K L. Homomorphisms of $3$-chromatic graphs. *Discrete Math*, [**54**]{}: 127–132 (1985)
Alon N, Dinur I, Friedgut E, et al. Graph products, fourier analysis and spectral techniques. *Geomatric And Functional Analysis*, [**14**]{}: 913–940 (2004)
Cameron P J, Ku C Y. Intersecting families of permutations. *European J Combin*, [**24**]{}: 881–890 (2003)
Deza M, Frank P. On the maximum number of permutations given maximal or minimal distance. *J Combin Theory Ser A*, [**22**]{}: 352–360 (1977)
Eggleton R B, Wallis W D. Problem 86: Solution I. *Math Mag*, [**58**]{}: 112–113 (1985)
Fang X G, Praeger C E, Wang J. On the automorphism groups of Cayley graphs of finite simple groups. *J London Math Soc*, [**(2)66**]{}: 563–578 (2002)
Feng Y Q. Automorphism groups of Cayley graphs on symmetric groups with generating transposition sets. *J Combin Theory Ser B*, [**96**]{}: 67–72 (2006)
Feng Y Q, Xu M Y. Automorphism groups of tetravalent Cayley graphs on regular p-groups. *Discrete Math*, [**305**]{}: 354–360 (2005)
Godsil C. Interesting Graphs and Their Colourings. http://www.math.ntu.edu.tw/$\sim$ gjchang/courses/2008-09-algebraic-graph-theory/Interesting%20Graphs%20and%20 their%20Colourings%20(Godsil).pdf
Godsil C D. The automorphism groups of some cubic Cayley graphs. *European J Combin*, [**4**]{}: 25–32 (1983)
Godsil C D. On the full automorphism group of a graph. *Combinatorica*, [**1**]{}: 243–256 (1981)
Graham R L, Grötschel M, Lovász L. Handbook of Combinatorics v II, Cambridge: The MIT Press (1995)
Huan H L, Liu H M, Xie W. Automorphism groups of a family of Cayley graphs on alternating groups. *J Syst Sci Inform*, [**5**]{}: 37–42 (2007)
Johnson E A. Automorphism groups of direct products of groups and their geometric realisations. *Math Ann*, [**263**]{}: 343–364 (1983)
Ku C Y, McMillan B. Independent sets of maximal size in tensor powers of vertex-transitive graphs. *J Graph Theory*, [**(4)60**]{}: 295–301 (2009)
Ku C Y, Wong T W H. Intersecting families in the alternating group and direct product of symmetric groups. *Electron J Combin*, [**14**]{}:\# R25 (2007)
Larose B, Malvenuto C. Stable sets of maximal size in Kneser-type graphs. *European J Combin*, [**14**]{}: 657–673 (2004)
Larose B, Tardif C. Projectivity and independent sets in powers of graphs. *J Graph Theory*, [**(3)40**]{}: 162–171 (2002)
Renteln P. On the spectrum of the derangement Graph. *Electron J Combin*, [**14**]{}: \# R82 (2007)
Rasmussen D J, Savage C D. Hamilton-connected derangement graphs on $S_n$. *Discrete Math*, [**133**]{}: 217–223 (1994)
Sanders R S, George J C. Basic results concerning the automorphism group of the tensor product of two graphs. *Utilitas Math*, 51–63 (June 1997)
Suzuki M. Group theory I, New York: Springer (1982)
Xu M Y. Automorphism groups and isomorphisms of Cayley digraphs. *Discrete Math*, [**182**]{}: 309–319 (1998)
Zhang Z, Huang Q X. Automorphism group of bubble-sort graphs and modified bubble-sort graphs. *Adv Math*, [**34**]{}: 441–447 (2005)
Zhang C, Zhou J X, Feng Y Q. Automorphisms of cubic Cayley graphs of order 2pq. *Discrete Math*, [**309**]{}: 2687–2695 (2009)
[^1]: This work is supported by National Natural Science Foundation of China (No:10971137), the National Basic Research Program (973) of China (No.2006CB805900), and a grant of Science and Technology Commission of Shanghai Municipality (STCSM, No: 09XD1402500) . $^{\dagger}$Correspondent author: Xiao-Dong Zhang (Email: [email protected])
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The problem of distinct value estimation has many applications. Being a critical component of query optimizers in databases, it also has high commercial impact. Many distinct value estimators have been proposed, using various statistical approaches. However, characterizing the errors incurred by these estimators is an open problem: existing analytical approaches are not powerful enough, and extensive empirical studies at large scale do not exist. We conduct an extensive large-scale empirical study of 11 distinct value estimators from four different approaches to the problem over families of Zipfian distributions whose parameters model real-world applications. Our study is the first that *scales to the size of a billion-rows* that today’s large commercial databases have to operate in. This allows us to characterize the error that is encountered in real-world applications of distinct value estimation. By mining the generated data, we show that estimator error depends on a key latent parameter — the average uniform class size — that has not been studied previously. This parameter also allows us to unearth error patterns that were previously unknown. Importantly, ours is the first approach that provides a framework for *visualizing the error patterns* in distinct value estimation, facilitating discussion of this problem in enterprise settings. Our characterization of errors can be used for several problems in distinct value estimation, such as the design of hybrid estimators. This work aims at the practitioner and the researcher alike, and addresses questions frequently asked by both audiences.'
author:
- 'Vinay Deolalikar[^1] Hernan Laffitte[^2]'
bibliography:
- 'uecbib.bib'
title: |
Extensive Large-Scale Study of Error in Samping-Based Distinct Value Estimators\
for Databases[^3]
---
=1
Dedication {#dedication .unnumbered}
==========
This study was carried out in 2011: the $40^{\textup{th}}$ year of the genocide of two million Hindus in 1971. This work is dedicated to their sacred memory, and especially to the women violated during that genocide; and also to Flt. Lt. Vijay Vasant Tambay.
[0.9]{}
[^1]: contact author, [[email protected]]{}, work done at HP Labs, Palo Alto.
[^2]: [[email protected]]{}
[^3]: This is the full-length version of a shorter published paper, and includes supplementary material for the published paper. Please cite as “Vinay Deolalikar and Hernan Laffitte: Extensive Large-Scale Study of Error in Samping-Based Distinct Value Estimators for Databases, IEEE Big Data Conference, Washington DC, December 2016."
| {
"pile_set_name": "ArXiv"
} |
The superconductor-insulator (SI) transition in ultrathin films is believed to be a continuous quantum phase transition [@Sondhi] occurring at $T =$ 0 as the quantum ground state of the system is tuned by varying disorder, film thickness, magnetic field, or carrier concentration through a critical value. Models of the SI transition can be grouped into two categories: ones that rely on fluctuations in the phase [@Fisher; @Fisher2; @Fisher3; @Cha] and others that rely on fluctuations in the amplitude [@Valles] of the order parameter to drive the transition. For models in the first category, Cooper pairs are assumed to be present on the superconducting side as well as on the insulating side of the transition. Based on this “dirty-boson” model, Fisher $\em{et}$ $\em{al.}$ [@Fisher2; @Fisher3] have proposed a phase diagram for a two-dimensional system as a function of disorder and magnetic field, with the transition being described in terms of interacting bosons (the Cooper pairs) in the presence of disorder. In the vicinity of the quantum critical point, the resistance is predicted to obey the scaling law [@Fisher2; @Fisher3]: $$R(\delta, T) \propto R_{c} f(\delta t),$$ where $t(T)$ $=$ $T^{-1/\nu z}$, and $\delta$ is the deviation of a tuning parameter from its critical value. For field-tuned transitions, $\delta = \vert B - B_{c} \vert$ with $B$ and $B_{c}$ being the applied magnetic field and the critical field, respectively. The critical resistance $R_{c}$ is predicted to have a universal value of $R_{Q} = h/4e^{2} \approx$ 6.5 k$\Omega/\Box$. Scaling arguments set a lower bound on the correlation length exponent $\nu$ $\ge$ 1 [@Sondhi] and give the value of the dynamical critical exponent $z$ $=$ 1 [@Fisher2; @Fisher3]. Experimentally, critical exponents consistent with these predictions have been found in scaling analyses of the field-tuned transitions in InO$_{x}$ [@Hebard] and MoGe [@Yazdani]. However, the predicted universal critical resistance is not observed in these experiments.
In this Letter, we report studies of the field-tuned SI transition in ultrathin Be films quench-condensed near 20 K. We describe scaling analyses of the field-tuned transition based on the “dirty boson” model. We have found that, for measurements performed in the zero-current limit, the scaling exponent product $\nu z$ $=$ 1.35 $\pm$ 0.10. This agrees well with results of the field-tuned transition in relatively thicker InO$_{x}$ [@Hebard] and MoGe [@Yazdani] films, but disagrees with a recent report [@Markovic; @Markovic2] of $\nu z$ $=$ 0.7 for the field-tuned transition in ultrathin Bi/Ge films. For our Be films having a robust superconducting phase, $R_{c}$ appears to be near 1.2$R_{Q}$ in the field-tuned transition. This agrees well with results from the Bi/Ge films [@Markovic; @Markovic2]. We have also carried out for the first time studies of the field-tuned transition in the presence of a dc bias current. The applied dc bias current should exert a force of $\bf{j} \times \bf{B}$ on the vortices in the direction perpendicular to both the applied magnetic field $\bf{B}$ and the current density $\bf{j}$. We have observed that the scaling exponent product $\nu z$ becomes 0.75 $\pm$ 0.10 in the presence of dc bias currents, suggesting that the field-tuned transitions with and without dc bias current belong to different universality classes.
In Fig. 1, we show, with zero bias current, the temperature dependence of the sheet resistance, $R_{\Box}$, measured at various field values for one of our Be films with the magnetic field applied perpendicular to the plane of the film. These Be films were quench-condensed onto bare glass substrates which were held near 20 K during evaporations, under UHV conditions inside a dilution refrigerator. This $\em{in}$ $\em{situ}$ progressive evaporation setup allowed for systematic studies of the SI transition as film thickness was varied. We deposited each set of films in fine steps. We carefully monitored $R_{\Box}$ during each evaporation step until a desirable value of $R_{\Box}$ was reached. The films were very close to 10 [Å]{} in thickness, however, our quartz thickness monitor was not sensitive enough to pick up the small thickness increments after each step. Film resistance was measured in a standard four-terminal geometry using a PAR-124A lock-in amplifier operating at 27 Hz. The ac probe current was fixed at 1 nA. At the finite measuring temperatures in the vicinity of the field-tuned transition, the [*I-V*]{} characteristics in this low-current regime were linear. More details regarding these Be films have been published elsewhere [@Bielejec]. Quench-condensed Be films are chosen for this study because such films were found to be nearly amorphous [@Bielejec]. These Be films undergo a transition from insulating to superconducting [@Bielejec] when the normal state sheet resistance, $R_{N}$, is reduced below $\sim$ 10 k$\Omega/\Box$ with increasing film thickness. We have now studied the field-tuned SI transition in several films of $R_{N}$ between 5.6 and 12 k$\Omega$/$\Box$. The superconducting transition temperature, $T_{c}$, of these films varied between 0.5 to 4 K in zero-field.
The film in Fig. 1, with $R_{N} =$ 10.7 k$\Omega$/$\Box$ at 15 K, is considered marginally superconducting, as we will discuss later. With increasing field, corresponding to curves from the bottom to the top in Fig. 1, this film was driven from superconducting to insulating, with a rather flat $R_{\Box}$ vs. $T$ curve at a critical field of $B_{c} =$ 0.66 T. The main part of Fig. 2(a) shows $R_{\Box}$ vs. $B$ at various temperatures for the same film as shown in Fig. 1. In the vicinity of a quantum critical point, the resistance of a two-dimensional system is predicted to obey the scaling law in Eq. (1). Determining the critical exponents involves plotting the $R_{\Box}$ vs. $B$ data at various temperatures according to the scaling law. The good crossing point, over one decade in temperature, in the $R_{\Box}$ vs. $B$ plot in the main part of Fig. 2(a) identifies $B_{c} =$ 0.66 T and $R_{c} =$ 4.4 k$\Omega$/$\Box$ for this film. We have used two methods to determine the scaling exponent $\nu z$. First, we can find $\nu z$ by evaluating the derivative of the $R_{\Box}$ vs. $B$ curve at the critical field. In this case, we have the following scaling equation: $$\frac{ \partial R}{ \partial B} \vert_{B_{c}} \propto
R_{c}T^{-1/\nu z} f'(0).$$ A plot of $\frac{ \partial R} {\partial B}\vert_{B_{c}}$ vs. $T^{-1}$ on a log-log scale, shown in the inset to Fig. 2(a), yields a straight line with a slope equal to 1/$\nu z$, from which we determine $\nu z$ $=$ 1.36 $\pm$ 0.10. Alternatively, we can plot $R$/$R_{c}$ vs. $\vert B - B_{c} \vert t$ and treat $t(T)$ as an unknown variable. The values of $t(T)$ at various temperatures are determined by obtaining the best collapse of the data. Following this procedure presented in recent papers [@Markovic; @Markovic2] by N. Markovic $\em{et}$ $\em{al.}$, we obtain the temperature dependence of $t(T)$ which we plot on a log-log scale in the inset to Fig. 3(a). The collapse of the data is shown in the main part of Fig. 3(a) in a $R$/$R_{c}$ vs. $\vert
B - B_{c}\vert t$ plot. Figure 3(a) shows good collapse of the data over three orders of magnitude in $\vert B - B_{c}\vert t$ and two orders of magnitude in $R$/$R_{c}$. The straight line in the inset to Fig. 3(a) shows a power-law fit, as expected by the scaling function, Eq. (1). The slope of the line in the inset gives -1/$\nu z$, from which we find $\nu z$ $=$ 1.34 $\pm$ 0.10. The exponents obtained from the above two methods agree with each other, showing the consistency of the scaling analysis. The scaling exponents obtained in our Be films in the zero-bias limit appear to be in very good agreement with the predictions of theories based on the “dirty boson” model [@Fisher2; @Fisher3]. Our result is also in very good agreement with renormalization group theories [@Zhang; @Singh] and Monte Carlo [@Cha] simulations.
Another important prediction of the bosonic model is the universal critical resistance $R_{c}$ $=$ $R_{Q} = h/4e^{2}$ $\approx$ 6.5 k$\Omega$/$\Box$ at the quantum critical point. This remains a controversial issue since only in the Bi/Ge system [@Markovic; @Markovic2] has a critical sheet resistance close to the predicted value of $R_{Q}$ been observed. In fact, careful investigations in the Bi/Ge system have revealed that $R_{c}$ is 1.1$R_{Q}$ to 1.2$R_{Q}$ and decreases in a narrow range as $R_{N}$ is reduced with increasing film thickness. It was suggested [@Markovic] that the slightly larger $R_{c}$ than $R_{Q}$ was consistent with calculations in two-dimensions for a bosonic model including Coulomb interactions [@Herbut], which predicted $R_{c}$ $\sim$ 1.4$R_{Q}$, as well as Monte Carlo simulations in the (2+1)-dimensional $XY$ model without disorder [@Cha2], which found $R_{c}$ $=$ 1.2$R_{Q}$. The small variation of $R_{c}$ with $R_{N}$ could be explained [@Markovic2] as a finite-temperature effect, since strictly speaking the critical resistance is predicted to be universal only at $T =$ 0. In other systems, such as MoGe [@Yazdani] films, $R_{c}$ varies in a much wider range among films of varying $R_{N}$. In order to explain the low $R_{c}$ of their MoGe films, Yazdani and Kapitulnik proposed [@Yazdani] a two-channel model in which the conductance due to normal electrons add to the conductance due to the Cooper pairs. If the unpaired electrons are localized, this model collapses to a single channel bosonic model with $R_{c}$ $=$ $R_{Q}$. Otherwise, unpaired electrons add to the conduction at the transition. It has been noted [@Markovic2] that the resulting film resistance can be either larger or smaller than $R_{Q}$ since the magnetoresistance contribution [@Bergmann] from the unpaired electrons can be either positive, if the spin-orbit coupling is strong, or negative, if the spin-orbit coupling is very weak.
Our Be films showed a zero-field SI transition as $R_{N}$ was reduced below $\sim$ 10 k$\Omega$/$\Box$ with increasing film thickness [@Bielejec]. Films of $R_{N}$ near 11 k$\Omega$/$\Box$ are considered marginally superconducting since they have near-zero critical field values. Further increasing film thickness, the critical field, $B_{c}$, increases with decreasing $R_{N}$, as shown in Fig. 4(a). In Fig. 4(b), we plot how $R_{c}$ in the field-tuned transition varies with $R_{N}$. For marginally superconducting films of $R_{N}$ between 9 to 12 k$\Omega$/$\Box$, $R_{c}$ was significantly smaller than $R_{Q}$. Nevertheless, we see clearly by comparing Fig. 4(a) with Fig. 4(b) that, as $R_{N}$ was reduced by increasing film thickness, films with robust critical field values have a $R_{c}$ of 7.1 $\sim$ 8.0 k$\Omega$/$\Box$ in the field-tuned SI transition, which is about 1.2$R_{Q}$. This result is in excellent agreement with results from the Bi/Ge system [@Markovic2]. However, our results disagree with the suggestion [@Markovic2] that the magnetoresistance of the unpaired electrons caused the discrepancy between $R_{c}$ and $R_{Q}$. Since Be has the weakest spin-orbit coupling among metals, the magnetoresistance should be negative, leading to $R_{c}$ $<$ $R_{Q}$ following the two-channel model. We note that we have shown in Fig. 1, Fig. 2(a), and Fig. 3(a) data measured on a marginally superconducting film of $R_{N} =$ 10.7 k$\Omega$/$\Box$ at 15 K. Although the critical resistance of this film, $R_{c}$ $\sim$ 4.4 k$\Omega$/$\Box$, is significantly smaller than $R_{Q}$, we have found that, for all the marginally and robust superconducting films of $R_{N}$ ranging from 5.6 to 12 k$\Omega$/$\Box$, the critical exponents are the same with $\nu z$ $=$ 1.35 $\pm$ 0.10 in the zero-current limit.
Below, we describe results of the field-tuned SI transition in the presence of a dc bias current, $I_{bias}$. For such studies, $I_{bias}$ was varied between 125 nA and 2.5 $\mu$A and kept below the zero-field critical current ($\sim$ 15 $\mu$A). At each fixed $I_{bias}$, we used a magnetic field to tune a film from superconducting to insulating. We believe that joule heating was insignificant in our experiments based on the following arguments. First, we have performed extensive [*I-V*]{} measurements over the entire temperature and magnetic field range of our experiments, with the time span for the [*I-V*]{} sweeps ranging from ten minutes to one hour. The [*I-V*]{} curves were completely reproducible without any observable hysteresis and independent of the sweep rate. Secondly, we can estimate the temperature increase, $\Delta T$, on the Be films due to joule heating. The thin Be films were deposited on glass microslide substrates of thickness 0.23 mm, which was attached to a copper sample holder by a very thin layer of grease. The glass substrate was the dominant source of heat resistance, with a thermal conductivity of $\sim$ 0.0003 W/Km at 100 mK [@White]. For a typical film square of size 3$\times$3 mm$^{2}$ and resistance 10 k$\Omega$, joule heating for $I_{bias}=$ 125 nA is about 0.15 nW, resulting in $\Delta T \sim$ 0.012 mK at 100 mK. For $I_{bias}=$ 2.5 $\mu$A, $\Delta T \sim$ 5.0 mK at 100 mK. We note that the heating power at 2.5 $\mu$A was 400 times larger than the heating power at 125 nA. If joule heating were significant, the data obtained with $I_{bias}=$ 2.5 $\mu$A should show a flattening of the data in the low temperature region when compared to the data obtained with $I_{bias}=$ 125 nA. The fact that the scaling results at 125 nA and 2.5 $\mu$A agree well suggests that heating was insignificant. This also argues against the existence of significant electron heating decoupled from the lattice.
We plot in Fig. 2(b) and Fig. 3(b), for $I_{bias} =$ 2.5 $\mu$A, the results of scaling analyses based on the two previously described methods. In Table I, we list the parameters from scaling analyses of the field-tuned transition at various $I_{bias}$. Results from the data collapsing method are presented for $I_{bias}$ values of 250 nA and 2.5 $\mu$A, for which the amount of data taken was adequate for such analyses. It appears that $\nu
z$ $\sim$ 0.75 $\pm$ 0.10, showing no systemic change with $I_{bias}$ in the range we have studied. Nevertheless, it is significantly smaller than the $\nu z$ found in the zero-current limit. We can only speculate that the bias current could lead to a symmetry-breaking, resulting in different critical behavior at the transition. We note that experiments need to be carried out in other systems in order to find out whether this result is universal. In addition, experiments at low $I_{bias}$ values should be carried out to determine whether $\nu z$ changes abruptly or gradually as $I_{bias}$ is increased from zero. Such experiments can probe the threshold $I_{bias}$ for the change in the scaling exponents, and have the potential of revealing whether the new scaling exponents are produced by certain nonlinear effects in the vortices under a bias current. On the other hand, we need to discuss other possible origins that might have led to the new scaling exponents. For example, transport in marginally superconducting films, or, in the presence of a dc bias, could be dominated by narrow superconducting filaments, leading to a change in the dimensionality of the system. While it is difficult to determine to what extent film inhomogeneity affects our experiments, we would like to comment on the critical currents measured in our films. For the film shown in Fig. 2(b) and Fig. 3(b), the critical current in zero field was about 1.5$\times$10$^{-5}$ A. This film was about 10 [Å]{} in thickness and 3 mm in width. Thus, the critical current density would be about 5$\times$10$^{6}$ A/m$^{2}$ if the current were uniformly distributed in the bulk of the film. On the other hand, if we assume that the current runs through a few filaments of 100 [Å]{} in total width, the critical current density would be 1.5$\times$10$^{12} $A/m$^{2}$. The critical currents of amorphous Be films have been measured by other groups, for example by Okamoto [*et al.*]{} [@Okamoto]. Their Be films had a resistivity of 3.6$\times$10$^{-6}$ $\Omega$m. This value gives rise to $R_{\Box} \sim$ 3.6 k$\Omega$/${\Box}$ for a film of thickness same as ours ($\approx$ 10 [Å]{}), meaning that their films were roughly 3 times less resistive than ours. They found that their films had critical current density of about 1$\times$10$^{8}$ A/m$^{2}$. Compared to their critical current density, the assumption in our films of conduction dominated by a few narrow filaments leads to an unreasonably large critical current density. In addition, we can compare our marginally superconducting films with robust superconducting films which are unlikely to be dominated by narrow filaments. As we have discussed, with $I_{bias} =$ 0, these two types of films showed identical scaling exponents $\nu z$. It is possible that film inhomogeneity has lead to a depressed $R_{c}$ in marginally superconducting films, but has not changed the effective dimensionality of the system. Therefore, the change in the scaling exponents with a dc bias is unlikely due to a dimensional cross-over as a result of filamentary superconductivity.
In conclusion, the scaling exponents found in our Be films in the zero-biased, field-tuned SI transition agree very well with results from InO$_{x}$ [@Hebard] and MoGe [@Yazdani] films, but disagrees with recent results [@Markovic; @Markovic2] from quench-condensed Bi/Ge films. For Be films having a robust superconducting phase, the critical sheet resistance in the field-tuned SI transition was about 1.2$R_{Q}$. Our field-tuned transition in the presence of a dc current reveals a set of new scaling exponents, suggesting that the field-tuned transitions with and without dc bias belong to different universality classes. We gratefully acknowledge S. Teitel and Y. Shapir for numerous useful discussions.
---------------------------------------------------------------------------------------------------------------------
$I_{bias}$ (nA) $B_{c}$ (T) $R_{c}$ (k$\Omega$/$\Box$) $R_{N}$ (k$\Omega$/$\Box$) $\nu z$ [^1] $\nu
z$ [^2]
----------------- -- ------------- ---------------------------- ---------------------------- -------------- ---------
125 0.33 15.5 9.36 0.81 –
250 0.35 16.1 9.36 0.76 0.73
1000 0.36 15.5 9.36 0.73 –
2500 0.38 15.0 9.36 0.77 0.75
---------------------------------------------------------------------------------------------------------------------
: Bias currents and derived parameters[]{data-label="Table 2"}
S. L. Sondhi [*et al.*]{}, Rev. Mod. Phys. [**69**]{}, 315 (1997).
M. P. A. Fisher [*et al.*]{}, Phys. Rev. B [**40**]{}, 546 (1989).
M. P. A. Fisher, Phys. Rev. Lett. [**65**]{}, 923 (1990).
M. P. A. Fisher, G. Grinstein, and S. M. Girvin, Phys. Rev. Lett. [**64**]{}, 587 (1990).
M.-C. Cha [*et al.*]{}, Phys. Rev. B [**44**]{}, 6883 (1991)
J. M. Valles, Jr., R. C. Dynes, and J. P. Garno, Phys. Rev. Lett. [**69**]{}, 3567 (1992); S-Y. Hsu, J. A. Chervenak, and J. M. Valles, Jr., [*ibid.*]{} [**75**]{}, 132 (1995).
A. F. Hebard and M. A. Paalanen, Phys. Rev. Lett. [**65**]{}, 927 (1990); M. A. Paalanen, A. F. Hebard, and R.R. Ruel, [*ibid*]{} [**69**]{}, 1604 (1992).
Ali Yazdani and Aharon Kapitulnik, Phys. Rev. Lett. [**74**]{}, 3037 (1995); D. Ephron, A. Yazdani, A. Kapitulnik, and M. R. Beasley, [*ibid.*]{} [**76**]{}, 1529 (1996).
N. Markovic, C. Christiansen, and A. M. Goldman, Phys. Rev. Lett. [**81**]{}, 5217 (1998).
N. Markovic [*et al.*]{}, Phys. rev. B [**60**]{}, 4320 (1999).
E. Bielejec, J. Ruan, and Wenhao Wu, Phys. Rev. B [**63**]{}, 100502 (2001); E. Bielejec, J. Ruan, and Wenhao Wu, Phys. Rev. Lett. [**87**]{}, 036801 (2001).
L. Zhang and M. Ma, Phys. Rev. B [**45**]{}, 4855 (1992).
K. G. Singh and D. S. Rokhsar, Phys. Rev. B [**46**]{}, 3002 (1992).
Igor F. Herbut, Phys. Rev. Lett. [**81**]{}, 3916 (1998).
M. -C. Cha and S. M. Girvin, Phys. Rev. B [**49**]{}, 9794 (1994).
G. Bergmann, Phys. Rep, [**107**]{}, 1 (1984).
G. K. White, [*Experimental Techniques in Low Temperature Physics*]{}, 3rd ed., (Clarendon Press, Oxford, 1979), p. 320.
M. Okamoto, K. Takei, and S. Kubo, J. Appl. Phys. [**62**]{}, 212 (1987)
[^1]: $\nu z$ obtained by the $\frac{\partial R}{\partial B}\vert_{B_{c}}$ method
[^2]: $\nu z$ obtained by the data collapse method
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Entanglement is not only important for understanding the fundamental properties of many-body systems, but also the crucial resource enabling quantum advantages in practical information processing tasks. While previous works on entanglement formation and networking focus on discrete-variable systems, light—as the only travelling carrier of quantum information in a network—is bosonic and thus requires a continuous-variable description in general. In this work, we extend the study to continuous-variable quantum networks. By mapping the ensemble-averaged entanglement dynamics on an arbitrary network to a random-walk process on a graph, we are able to exactly solve the entanglement dynamics and reveal unique phenomena. We identify squeezing as the source of entanglement generation, which triggers a diffusive spread of entanglement with a parabolic light cone. The entanglement distribution is directly connected to the probability distribution of the random walk, while the scrambling time is determined by the mixing time of the random walk. The dynamics of bipartite entanglement is determined by the boundary of the bipartition; An operational witness of multipartite entanglement, based on advantages in sensing tasks, is introduced to characterize the multipartite entanglement growth. A surprising linear superposition law in the entanglement growth is predicted by the theory and numerically verified, when the squeezers are sparse in space-time, despite the nonlinear nature of the entanglement dynamics. We also give exact solution to the equilibrium entanglement distribution (Page curves), including its fluctuations, and found various shapes dependent on the average squeezing density and strength.'
author:
- 'Bingzhi Zhang$^{1,2}$'
- 'Quntao Zhuang$^{2,3}$'
title: 'Entanglement formation in continuous-variable random quantum networks'
---
Introduction
============
Quantum information science has brought to us capabilities to enhance the performance of computing [@Shor_1997], sensing [@Giovannetti2004] and communication [@Bennett2002; @gisin2007quantum], through entangling local or distant processing nodes. Therefore, a quantum network [@kimble2008quantum] that enables entanglement establishment is important for achieving the promised quantum advantages. The study of entanglement formation and quantum information scrambling has been fruitful in complex systems such as random quantum networks [@biamonte2019complex; @brito2019statistical; @acin2007entanglement], and circuits [@nahum2017quantum; @nahum2018operator; @Keyserlingk:2018aa; @Khemani:2018aa; @Rakovszky:2018aa], many-body systems [@kim2013ballistic; @luitz2017information; @huang2017out; @chen2017out; @fan2017out; @gopalakrishnan2018hydrodynamics; @you2018entanglement; @banerjee_2017; @patel2017quantum; @patel2017quantum_prx], models of holography [@sachdev1993; @kitaev2015simple; @gu2017local; @kitaev2018soft] and quantum gravity [@hayden2007black; @hosur2016chaos; @Yoshida:2017aa; @gao2017traversable; @maldacena2017diving; @sekino2008fast; @maldacena2016bound; @shenker2014black; @Lashkari13; @Roberts:2014isa; @piroli2020random; @agarwal2020toy; @liu2020dynamical]. Universal scaling laws and dynamical models of entanglement formation has been established, based on nonlinear surface growth models [@nahum2017quantum; @nahum2018operator]. Recently, experimental probing [@landsman2019verified; @fan2017out; @garttner2017measuring] of scrambling is also made possible; from the quantum network perspective, protocol designs [@chakraborty2019distributed; @vardoyan2019performance; @pant2019routing] for entanglement establishment has also been a recent focus.
The above works, whether on the basic understanding of scrambling or practical design of networking, mainly focus on entanglement in discrete-variable (DV) systems, which is natural for computing. However, as quantum networks inevitably utilize light as the carrier of quantum information in transmission, the bosonic nature of light makes it necessary to consider entanglement in a continuous-variable (CV) description. Moreover, various applications in the photonic or microwave domain, including universal quantum computing based on cluster states [@menicucci2006universal], quantum illumination [@tan2008quantum; @zhuang2017optimum; @Zheshen_15], quantum reading [@pirandola2011quantum], distributed sensing [@zhuang2018distributed; @zhuang2019physical; @guo2020distributed; @xia2019entangled] and entanglement-assisted communication [@shi2019practical; @guha2020infinite], require CV entanglement in the form of Gaussian states [@Weedbrook_2012]. In this regard, noiseless linear amplifiers [@seshadreesan2018continuous] and novel error correction codes [@noh2019encoding; @zhuang2019distributed] provide initial tools for CV networking, and an out-of-time-order correlator (OTOC) has revealed a unique squeezing-dependent butterfly-velocity of operator spreading [@zhuang2019scrambling].
In this paper, we study quantum information scrambling in CV quantum networks (see Fig. \[fig:general\_graph\]) focusing on the entanglement formation dynamics. Inspired by the classical statistical theory of complex networks [@barabasi1999emergence; @watts1998collective; @barabasi1999mean], we consider random quantum networking protocols to enable analytical solutions, through a mapping to a random-walk process on graph; at the same time, random protocols are expected to reveal typical and universal characteristics. Our results apply to quantum networks on general graphs, therefore provide a foundation for the statistical theory of complex quantum networks.
We provide an analytical formula connecting the entanglement entropy to weights in the passive linear optical transforms, therefore establishing a mapping between ensemble-averaged entanglement dynamics to the probability evolution of a random-walk process on a general graph (see Fig. \[fig:general\_graph\]). The change of the entanglement entropy $S({{\cal L}},t)$ of a subsystem ${{\cal L}}$, similar to the random walker’s probability in subsystem ${{\cal L}}$, is determined by the boundary $\partial{{\cal L}}$. Moreover, we also solve the fluctuations of the entanglement entropy.
In discrete space-time, the ensemble-averaged weights dynamics can be described by a Markov chain, with the transition matrix determined by the graph connectivity. Going into continuous space-time, we can derive simple diffusive partial differential equations (PDEs) to describe the entanglement evolution. In both cases, the model is completed with the analytical formula connecting weights to entanglement entropy. Alternatively, a phenomenological model coupling an epidemiology equation with a nonlinear diffusion can directly capture the entanglement dynamics.
To go beyond the bipartite characterization of entanglement through entanglement entropy, we also give an operational witness of multipartite entanglement. This witness is directly connected to entanglement’s advantage over classical correlations in distributed sensing protocols [@zhuang2019distributed]. Maximum values of the entanglement witness are achieved towards the late time, therefore verifying the full scrambling of the entire network.
Through the mapping between quantum dynamics and random walk, we also connect the scrambling time—the time it takes for the entire system to be maximally entangled—directly to the mixing time of the random walk. Moreover, at infinite time, the equilibrium entanglement distribution—analog to the Page curve in DV systems [@Page93; @nakagawa2018universality; @fujita2018Page]—can be solved analytically from the stationary state of the random walk. Surprisingly, the Page curve is independent of the topology of the network, as long as the graph is connected. And in general it depends on two statistical properties of the quantum network—the squeezer’s density and the average squeezing strength. Interestingly, a small subsystem can almost get close to the maximum entanglement entropy, while in DV systems, half of the system size is necessary.
While our theory works for general graphs, we also apply to networks respecting ‘locality’ of interactions—$D$-dimensional Cartesian graphs where links only exist between nearest neighbors (see Figs. \[fig:schematic\_1D\] and \[fig:schematic\_2D\]). In this regard, we identify a diffusive entanglement light cone at the early time, which divides the regions with almost no entanglement and regions with substantial entanglement. After the entanglement light cone reaches each node, there is a period of entanglement sudden growth, where the entanglement entropy quickly gets close to its equilibrium value. In the end, there is a long period of saturation, determined by the mixing time $\sim M^2$ quadratic in the length $M$ on each dimension.
Our theory framework provides a unique complement to the DV counterparts [@nahum2017quantum; @nahum2018operator; @Keyserlingk:2018aa; @Khemani:2018aa; @Rakovszky:2018aa; @you2018entanglement]. And our results provide insights into not only CV quantum networks being engineered, but also quantum information scrambling in various physical systems, as any form of bosonic radiation is intrinsically CV.
The paper is organized as the following. In Sec. \[sec:intro\_Gaussian\], we specify the model in details and give a more specific overview of results; in Sec. \[sec:statistical\_theory\], we present the statistical theory for the single-squeezer case based on a mapping to random-walk dynamics; in Sec. \[sec:multiple\_squeezers\], we generalize the single-squeezer results to the general case through providing a linear superposition law.
Quantum networks: modelling and main results {#sec:intro_Gaussian}
============================================
Our overall goal is to characterize generic entanglement formation dynamics towards equilibrium in a CV quantum network (see Fig. \[fig:general\_graph\] for a schematic). In general, a quantum network can have complicated topology, which makes the problem difficult. Moreover, each node can possess multiple optical modes, and perform local operations coordinated by classical communication to entangle them.
Considering the optical modes, we can reduce a general entanglement generation protocol to a quantum circuit on a graph, as we illustrate in an one-dimensional (1-D) hopping quantum network in Fig. \[fig:schematic\_1D\] (a). To establish entanglement, each node performs the following protocol repetitively: it receives a light mode from a neighbor, which gets entangled with a stored mode through a local unitary; then, it sends out a mode to another neighbor and stores one mode locally. For simplicity, the nodes send light to the left and right neighbors alternatively in even and odd steps. If we focus on the dynamics of the optical modes, the above protocol reduces to a 1-D local circuit in Fig. \[fig:schematic\_1D\] (b), where local gates apply alternatively on the light modes [@zhuang2019scrambling].
The transmission links in a quantum network are in general lossy. To cope of loss, error correction [@noh2019encoding] can be applied in each link transmission. On the physical layer, this means including additional components that seemingly complicate the analyses. However, on the logical layer, up to some small residual errors from imperfect error correction, the state being protected is identical to the state being generated in a lossless quantum network, as demonstrated in Ref. [@zhuang2019distributed] for sensing purposes. Therefore, we start with the lossless case.
With the mapping between quantum networks and quantum circuits in mind, we specify the set-up of the circuits on an arbitrary (undirected) graph (see Fig. \[fig:general\_graph\](c)). In general, the topology can be described by an un-directed graph $({{\cal G}}, {{\cal E}})$, where ${{\cal G}}$ denotes the set of all vertices, each described by a coordinate system $\bm x$ [@Note3]. The set of edges ${{\cal E}}$ can be described by a generalized connection matrix $E_{\bm x, \bm x^\prime}$. When $E_{\bm x, \bm x^\prime}=1$, the vertices $\bm x, \bm x^\prime$ are connected by an edge $\overline{\bm x \bm x^\prime}$, zero when not connected. For simplicity, we write the set of vertices that are directly connected to $\bm x$ (neighbors) as ${{\cal N}}(\bm x)$. We are interested in the entanglement between a set of vertices ${{\cal L}}$ and the rest ${{\cal R}}={{\cal G}}\backslash {{\cal L}}$. In Fig. \[fig:schematic\_1D\](b), we give a $1$-D example of the notations.
Unitaries are applied on the edges ${{\cal E}}$. We separate the edges into disjoint sets $\{{{\cal E}}_{k}\}_{k=1}^K$, such that the edges in each set ${{\cal E}}_k$ do not have common vertices. The dynamics repeat in a period of $K$ steps; in the $k$-th step of each period, one applies unitaries $U_{t,\bm x,\bm x^\prime}$ on each edge $\overline{\bm x\bm x^\prime}\in {{\cal E}}_k$. The particular separation of the unitaries is not essential to the dynamics and equilibrium. As an example, in a 1-D local circuit, $K=2$ and we alternative between gates $\{U_{t,k,k+1}\}$ on $k$ odd and even; in a 2-D local circuit, we have $K=4$, as shown in Fig. \[fig:schematic\_2D\].
To produce the Gaussian states that enables various applications in communication, sensing and computing, we consider Gaussian unitaries [@Weedbrook_2012], which are unitaries generated by Hamiltonians that are second order in the quadrature operators (see Appendix \[app:random\_mat\]). Gaussian unitaries include squeezing, which creates asymmetry in quadrature noises; and passive linear optics, which includes beamsplitters and phase-shifters.
Squeezing is essential for entanglement generation. However, as an ‘active’ component, squeezing is relatively difficult to implement. Thus, we consider the gates $\{U_{t,\bm x,\bm x^\prime}\}$ to be passive linear-optics gates. And squeezing operations are added in between in a sparse way. As an example, in Fig. \[fig:schematic\_1D\](b), to establish entanglement, in this case a single vertex performs a squeezing operation (the cyan box), and then entanglement is generated by passing it around through passive components (the orange boxes), with vacuum on the other input modes.
We expect random protocols to reveal universal characteristics, therefore we choose the passive linear optics gates $\{U_{t,\bm x,\bm x^\prime}\}$ to be Haar random (see Appendix \[app:random\_mat\]). This is also justified by the following reasons: (1) In classical complex network theory [@barabasi1999emergence; @watts1998collective; @barabasi1999mean], various networks can be modeled as random networks with a proper degree distribution. (2) In condensed matter theory, random quantum circuits and Hamiltonian systems [@nahum2017quantum; @nahum2018operator; @Keyserlingk:2018aa; @Khemani:2018aa; @Rakovszky:2018aa; @you2018entanglement] are able to capture the essential quantum information spreading features in generic many-body interacting systems. (3) In real quantum networks, the form of entanglement required can be complicated, depending on the purpose, e.g. the weights of the global parameter of interest in a distributed sensing protocol [@zhuang2018distributed].
We aim to characterize the entanglement dynamics in the above random circuits. The entanglement entropy, measured by von Neumann entropy or Renyi entropy, can be numerically evaluated efficiently (see Appendix \[app:entropy\]). Two examples of time evolution of von Neumann entropy in 1-D are given in Fig. \[fig:lightcone\]. In Fig. \[fig:lightcone\](a), we have a single squeezer at the center in the first step, which is identical to the case depicted in Fig. \[fig:schematic\_1D\]. The entanglement entropy grows diffusively from the source of squeezing (see Sec. \[sec:dynamic\]). Different from the DV case, we can identify an entanglement light cone (green lines), which is the boundary between regions with substantial entanglement and regions with almost-zero entanglement. These phenomena can also be found in higher dimensional random local circuits, as shown in Fig. \[fig:schematic\_2D\] for the 2-D case. When choosing a subsystem ${{\cal L}}$ as an individual mode at $(x_1,x_2)$, we can see similar entanglement light cone (green dashed).
As shown in Fig. \[fig:general\_graph\](b), the above entanglement dynamics can be solved by mapping to a random walk on a graph, which gives the exact ensemble-averaged entanglement entropy (see Sec. \[sec:mapping\]) $\braket{S({{\cal L}},t)}$ as a function of the total probability $\eta_{{{\cal L}},t}$ of having the walker in region ${{\cal L}}$ in the corresponding random walk, S(\_[[[L]{}]{},t]{}) = g, \[SL\_preliminary\] where $
g(x)
$ is the (von Neumann or Renyi) entropy of a thermal state with mean photon number $x$ (see Appendix \[app:entropy\]) and $r$ is the original squeezing strength. Note that the mapping holds for arbitrary graphs beyond the local Cartesian graphs shown above. The scrambling time—the time that the entire system becomes maximally entangled—can be calculated by the mixing time of the random walk (see Sec. \[sec:mixing\_time\]). The mapping also gives the Page curves—the late-time equilibrium entanglement entropy = S([|[[L]{}]{}|]{}/[|[[G]{}]{}|]{}) \[SL\_page\_preliminary\] as the CV analog to Page curve (see Sec. \[sec:equi\_solu\_general\]), while the fluctuation can be solved as $\propto |{{\cal L}}||{{\cal R}}|$. Moreover, when there are multiple squeezers, we can regard the entanglement dynamics as the superposition of all single-squeezer dynamics, as depicted in Fig. \[fig:lightcone\](b) and will be detailed in Sec. \[sec:superposition\]. When there are multiple squeezers, we find that the Page curve is determined by the average squeezing strength and density of the squeezers (see Sec. \[sec:Page\]). Therefore, combining the results, we have a complete understanding of the entanglement dynamics in a CV quantum network.
Statistical theory of random CV quantum networks {#sec:statistical_theory}
================================================
In this section, we present a statistical theory of the entanglement growth. We will focus on the single squeezer case in Fig. \[fig:lightcone\](a), while the extension to multiple squeezers is presented in Sec. \[sec:multiple\_squeezers\]. We introduce the mapping between random unitary circuits and random walk on a graph in Sec. \[sec:mapping\], which allows us to solve the Page curves in Sec. \[sec:equi\_solu\_general\] and scrambling time in Sec. \[sec:mixing\_time\] for general graphs. Explicit closed-form solutions can be obtained for local Cartesian graphs of an arbitrary dimension in Sec. \[sec:dynamic\]. Finally, we present the entanglement witness for multipartite entanglement in Sec. \[sec:multipartite\_E\].
Mapping to random walk on graphs {#sec:mapping}
--------------------------------
Consider the entire unitary evolution $U(t)$ of the random circuit. In the single-squeezer case, the mode annihilation operator $a_{\bm x,t}$ at vertex $\bm x\in {{\cal G}}$ experiences a passive transform, which in general can be expresses as a\_[x,t]{}=e\^[i\_[x,t]{}]{} a\_[SV]{}+[vac]{}, \[axt\] where mode $a_{\rm SV}$ is in a squeezed-vacuum (SV) state with strength $r$ and ‘vac’ denotes all vacuum terms that complete the commutation relation. Here the phase $\theta_{\bm x,t}$ is entirely random, and the positive weights $\bm w_t\equiv \{w_{\bm x,t}\}_{\bm x\in {{\cal G}}}$ describe the overall energy splitting of the single SV among all modes.
For any subsystem ${{\cal L}}$ with a density operator $\rho_{{\cal L}}(t)$, we can design a passive linear optics circuit $U_{{{\cal L}},t}$ such that $U_{{{\cal L}},t}\rho_{{\cal L}}(t)U_{{{\cal L}},t}^\dagger$ concentrates all the squeezing parts to a single mode a\_[[[L]{}]{},t]{}=a\_[SV]{}+[vac]{}, with the total transmissivity \_[[[L]{}]{},t]{}=\_[x[[L]{}]{}]{}w\_[x,t]{}, \[etat\_def\] and all other modes are in vacuum required by energy conservation. Because unitary operations preserve entropy, the entanglement entropy of ${{\cal L}}$ can be calculated from the entropy of mode $a_{{{\cal L}},t}$ as $$\begin{aligned}
&S\left({{\cal L}},t\right)=S\left(\eta_{{{\cal L}},t}\right)
\label{SL}
\\
&\simeq \frac{1}{2}\log_2\left[\eta_{{{\cal L}},t}\left(1-\eta_{{{\cal L}},t}\right)\right]+\frac{1}{\ln2}\left(r+1\right)-1,
\label{SL_sim}\end{aligned}$$ where $S(\eta_{{{\cal L}},t})$ is defined in Eq. (\[SL\_preliminary\]). We will focus on von Neumann entropy, but all of our results can be adapted to Renyi entropy easily. At the large squeezing limit of $\sqrt{\eta_{{{\cal L}},t}(1-\eta_{{{\cal L}},t})}e^r\gg1$, for von Neumann entropy we have Eq. (\[SL\_sim\]) to the leading order. When $\eta_{{{\cal L}},t}=0,1$, subsystem ${{\cal L}}$ has zero or entire portion of the SV, indeed from Eq. (\[SL\]) we have $S\left({{\cal L}},t\right)=0$. When $\eta_{{{\cal L}},t}=1/2$, we have the maximum entropy $S_0(r)=g(\sinh^2(r/2))\simeq \log_2(e^r/4)$. When one has large number of modes, this should agree with the result of $\max_{{{\cal L}}} S\left({{\cal L}},t\right)$ at any time.
So far we have the [*exact*]{} result $S\left(\eta_{{{\cal L}},t}\right)$ of the entanglement entropy of an arbitrary subsystem ${{\cal L}}$, given the weights $\bm w_t$ (which determines $\eta_{{{\cal L}},t}$) obtained in each random circuit realization. Due to the self-averaging in the random circuit, we expect =S() up to corrections that decay with the system size. Thus, we have reduced the problem of solving the ensemble-averaged entanglement dynamics to solving the ensemble-averaged dynamics of the weights $\braket{\bm w_t}$.
We start by focusing on a single gate $U_{t,\bm x,\bm x^\prime}$ on the modes at $\bm x$ and $\bm x^\prime$. By considering the Haar random ensemble averaging, we can derive the exact equation of motion of the weights as (see Appendix \[app:random\_derivation\]) ==(+). \[ensemble\_dynmaics\] The overall dynamics alternatives in $K$ steps, in the $k$-th step the transition of Eq. (\[ensemble\_dynmaics\]) on all edges in ${{\cal E}}_k$ is applied.
An immediate observation from Eq. \[ensemble\_dynmaics\] is that the change of the entanglement entropy of ${{\cal L}}$, determined by the total weights $\braket{\eta_{{{\cal L}},t}}$, is related only to boundary $\partial {{\cal L}}$ (schematic in Fig. \[fig:general\_graph\](c)), in the sense that $$\begin{aligned}
&\braket{\eta_{{{\cal L}},t+1}}-\braket{\eta_{{{\cal L}},t}}=
\nonumber
\\
&\frac{1}{2}\left[\sum_{\bm x \in \partial {{\cal L}}^+} \braket{w_{\bm x,t}}- \sum_{\bm x \in \partial {{\cal L}}^-} \braket{w_{\bm x,t}}\right],
\label{eta_boundary}\end{aligned}$$ which equals the net flow of the weights from the vertices on the outer boundary $\partial {{\cal L}}^+$ towards ${{\cal L}}$ and the weights from the inner boundary $\partial {{\cal L}}^-$ out from ${{\cal L}}$.
Another observation is that the weights update rule in Eq. (\[ensemble\_dynmaics\]) also describes the probability evolution of a lazy symmetric random-walk step, where the walker have half probability of staying and half probability of taking a step along $\overline{\bm x\bm x^\prime}$ (see Fig. \[fig:random\_walk\]). Combining the $K$ steps, the underlying transition matrix for the weights \_[x, x\^]{} = \_[k = 1]{}\^[K]{} (I+E\_[k,x,x\^]{}), \[total\_transition\_matrix\] where $I$ is the identity matrix and $E_{k,\bm x,\bm x^\prime}$ describes the adjacency matrix for the corresponding graph $({{\cal G}},{{\cal E}}_k)$ (an isolated mode is regarded as a vertex with a loop). Eq. \[total\_transition\_matrix\] describes a modified symmetric random walker on the graph (see Fig. \[fig:random\_walk\]), with $K$ steps combined to implement a single random-walk step from the current position $\bm x$ to all neighbors ${{\cal N}}(\bm x)$ (including $\bm x$) with equal probability.
Utilizing the $K$-step transition matrix, the ensemble-averaged weights can be solved at any time $t$ as = \_[x, x\^]{}\^[\[t/K\]]{}, \[ensemble\_dynamics\_matrix\] with the initial condition $\braket{\bm w_0}=\delta_{\bm x_0}$ as the Kronecker delta at the squeezer position $\bm x_0$. Thus, one can obtain $\braket{\eta_{{{\cal L}},t}}$ and the exact result of $S(\braket{\eta_{{{\cal L}},t}})$ from Eq. (\[SL\]) on any graph.
We give examples of the random walk in Fig. \[fig:random\_walk\] in 1-D and 2-D Cartesian graphs, whose entanglement evolution can be found in Figs. \[fig:schematic\_1D\] and \[fig:schematic\_2D\]. For the later use, we also introduce a general $D$ dimensional Cartesian lattice ${{\cal G}}_D$, with the coordinates $\bm x=(x_1,
\cdots, x_D)$ on a grid ($x_d\in[-N,N]$). The total number of modes is $|{{\cal G}}_D|=M^D$, with $M=2N+1$ modes on each dimension.
In the following, we will consider the equilibrium and the dynamics. Some results hold for general graphs, while some analytical results are made possible by considering the special case of ${{\cal G}}_D$.
Equilibrium of CV random networks: Page curves and fluctuations {#sec:equi_solu_general}
---------------------------------------------------------------
In this section, we focus on the Page curves—the equilibrium entanglement distribution at infinite time. In order to share entanglement, squeezers are applied, which are then followed up by the random beamsplitters and phase shifters. As the layers of gates increases, the overall passive linear transform will approach the Haar measure (see Appendix \[app:random\_mat\]). Therefore, we can regard the equilibrium entanglement distribution as the CV analog to Page curves.
Considering the mapping from the circuit to the random walk, the equilibration of the entanglement also corresponds to the full mixing of the random walk on the graph. Assuming the full connectivity of the graph, due to the special transform matrix in Eq. (\[total\_transition\_matrix\]), the equilibrium (stationary) state of weights is uniform among all vertices, i.e., =1/|[[G]{}]{}|, \[w\_stationary\] where $|{{\cal G}}|$ is the total number of vertices, despite how one arranges the set of edges ${{\cal E}}_k$. Note that this is different from normal random walks on a graph, where the stationary state has weights proportional to the degree of the vertex $|{{\cal N}}(\bm x)|$ [@lovasz1993random]. Therefore, the total transmissivity (i.e., total weights) =[|[[L]{}]{}|]{}/[|[[G]{}]{}|]{}. In fact, assuming fully random weights from a Haar random unitary, one can obtain the probability density of total weights as (see Appendix \[app:eta\_var\]) P(\_[[[L]{}]{},]{}=)\^[|[[L]{}]{}|-1]{}(1-)\^[|[[R]{}]{}|-1]{}. \[P\_eta\_dis\] From Eq. (\[SL\]), the Page curve is therefore given by $$\begin{aligned}
\braket{S({{\cal L}},\infty)}&= S(\frac{|{{\cal L}}|}{|{{\cal G}}|})
\label{SL_page}
\\
&\simeq \frac{1}{2}\log_2\left[\frac{|{{\cal L}}|}{|{{\cal G}}|}\left(1-\frac{|{{\cal L}}|}{|{{\cal G}}|}\right)\right]+\frac{1}{\ln2}\left(r+1\right)-1.
\label{SL_page_sim}\end{aligned}$$ The maximum is achieved at ${|{{\cal L}}|}/{|{{\cal G}}|}=1/2$, which equals $\braket{S_0(r)}$ introduced following Eq. (\[SL\]). The second equality is the leading order result similar to Eq. (\[SL\_sim\]).
Note that in terms of Page curves, the graph topology is irrelevant as the entire dynamics is equivalent to a single passive global Haar unitary; therefore, we can simply stretch the coordinates $\bm x$ of a general graph ${{\cal G}}$ to a single coordinate $x$ in 1-D. It then suffices to verify our theory of Page curves in ${{\cal G}}_1$, which allows simple visualization. In the 1-D system, when subsystem ${{\cal L}}$ contains the left side of mode $x$, we have $
\braket{\eta_{{{\cal L}},\infty}}={|{{\cal L}}|}/{|{{\cal G}}|}=({x+N})/({2N+1})\simeq\tilde{x}+1/2.
$ Therefore, it is convenient to choose the parameterization $\tilde{x}$. Fig. \[fig:pages\_1D\](a) plots $\braket{S(\tilde{x},\infty)}$ for various squeezing values of $r$ and system sizes of $M$, where we see perfect overlapping among curves with identical $r$ for different system sizes. And they all agree with Eq. (\[SL\_page\]) very well, as shown in Fig. \[fig:pages\_1D\] (b). As a by-product, the maximum entanglement—the maximum height of the Page curve $\max_x \braket{S(x,\infty)}$—agrees with the theory prediction $\braket{S_0(r)}$ following Eq. (\[SL\]), as shown in Fig. \[fig:pages\_1D\](b) inset.
Furthermore, we can also consider the fluctuations around the Page curves, as exemplified in Fig. \[fig:auto\_correlation\] (a) and (c). Since the entanglement entropy is determined by $\eta_{{{\cal L}},t}=\sum_{\bm x\in {{\cal L}}}w_{\bm x,t}$, it takes some time for $\{w_{\bm x,t}\}_{\bm x\in {{\cal L}}}$ to entirely change its values; thus, there will be correlations in the entanglement entropy at different times. We consider the equilibrium auto-correlation $$\begin{aligned}
&G_\infty\left(\Delta t\right)=\lim_{t\to\infty}
\nonumber
\\
&\braket{(S({{\cal L}},t)-\braket{S({{\cal L}},\infty)})(S({{\cal L}},t+\Delta t)-\braket{S({{\cal L}},\infty)})}.\end{aligned}$$ We expect the decay of this auto-correlation should have a similar time-scale with the mixing time of the entire system (which will be detailed in Sec. \[sec:mixing\_time\]), as demonstrated in Fig. \[fig:auto\_correlation\]. When $\Delta t=0$, the auto-correlation goes to the variance $$\begin{aligned}
&G_\infty\left(0\right)= \braket{{\rm var}\left(S({{\cal L}},\infty)\right)},
\\
&=
\left(\left.\frac{\partial S\left(\eta_{{{\cal L}},t}\right)}{\partial \eta_{{{\cal L}},t}}\right\vert_{\eta_{{{\cal L}},t}=\frac{|{{\cal L}}|}{|{{\cal G}}|}}\right)^2 \frac{|{{\cal L}}||{{\cal R}}|}{|{{\cal G}}|^2(|{{\cal G}}|+1)},
\label{eq:fluctuation_static}\end{aligned}$$ where we have used the chain rule of variance and the variance of $\eta_{{{\cal L}},\infty}$ can be obtained in Appendix \[app:eta\_var\]. Numerical results in 1-D and 2-D Cartesian graphs ${{\cal G}}_D$ agree well with Eq. (\[eq:fluctuation\_static\]), as shown in Fig. \[fig:static\_fluctuation\].
On the other hand, if we look at the change of subsystem entropy within a short period of time, e.g., a single-step, similar to Eq. (\[eta\_boundary\]) the short-time fluctuating $\braket{\left(S\left({{\cal L}},t+1\right)-S\left({{\cal L}},t\right)\right)^2}$ will mainly come from the boundary. This interplay of short time fluctuation related to the boundary, while long time fluctuation related to the bulk manifests the rich entanglement dynamics in random quantum networks.
Entanglement scrambling time {#sec:mixing_time}
----------------------------
The entanglement scrambling time—the time for the system to be maximally entangled and reach the equilibrium is an important quantity for many physical problems, especially those related to the black hole [@hayden2007black; @sekino2008fast]. In the CV quantum network, this scrambling time can be directly obtained from the mixing time $t_\epsilon^\star$ of random walks on a graph, where the probability measure (weights) get $\epsilon$-close to the stationary state in Eq. (\[w\_stationary\]). Formally, we define the mixing time $t_\epsilon^\star$ to be the time when the deviation $|\braket{w_{\bm x,t}}-{1}/{|{{\cal G}}|}|\le \epsilon$ for all $\bm x\in {{\cal G}}$. Multiple estimates can be obtained, as we explain bellow [@lovasz1993random].
The first estimate relies on the eigenvalues of the transition matrix $\mathbb{E}_{\bm x, \bm x^\prime}$ in Eq. (\[total\_transition\_matrix\]). The stationary state $\braket{w_{\bm x,\infty}}={1}/{|{{\cal G}}|}$ corresponds to the largest eigenvalue of unity, and the second largest eigenvalue $\lambda^\star<1$ gives the decay of deviations |-|\~[\^]{}\^[(t/K)]{}, \[graph\_bound\_1\] From the above, one can obtain $t_\epsilon^\star\sim K \ln \left(1/\epsilon\right)/\ln (1/\lambda^\star)$. Here we have taken into account that, one needs $K$ steps to implement the transition in Eq. (\[total\_transition\_matrix\]).
For the Cartesian graphs ${{\cal G}}_D$, we can obtain the convergence towards $\braket{\bm w_{\bm x,\infty}}=1/|{{\cal G}}|$ from the expected hitting time in the large $t$ limit as [@lovasz1993random] |-|\~6\^[-t/DM\^2]{}, \[graph\_bound\_2\] where $M=|{{\cal G}}|^{1/D}$ is the length of the Cartesian graph in each dimension. This gives another estimate $t_\epsilon^\star\sim\ln\left(1/\epsilon\right)DM^2/\ln(6)$.
A third bound can be obtained by calculating the conductance of a graph as ([[G]{}]{})=\_[[[L]{}]{}]{}, \[conductance\] where $|\nabla {{\cal L}}|$ denotes the set of boundary edges connecting ${{\cal L}}$ to ${{\cal R}}$ and $|{{\cal E}}|$ is the total number of edges. The minimization will be able to capture the slowest part of the mixing. We can obtain from Ref. [@lovasz1993random] as |-|\~(1-)\^t, \[graph\_bound\_3\] therefore the another estimate on the mixing time can be obtained as $t_\epsilon^\star\sim \ln \left(\epsilon\right)/\ln \left(1-\Phi^2/8\right)$.
For the Cartesian graph ${{\cal G}}_D$ of length $M$ on each direction, the total number of edges of ${{\cal G}}_D$ is $
|{{\cal E}}| = DM^{D-1}(M-1),
$ and total number of vertices in subsystem $|{{\cal G}}_D| = M^D$. And we need $K=2D$ steps to implement a single symmetric walker step on each direction. For $D=1$, one has $|\nabla {{\cal L}}|=1$, $|{{\cal G}}_D|=M$ and $|{{\cal E}}|=M-1$, therefore it is straightforward to obtain $\Phi({{\cal G}}_1)=2/(M-1)$. For $D=2$, we give an estimation from simple heuristic numerical minimization.
We compare the three estimates of mixing time on ${{\cal G}}_D$ for $D=1,2$ in Fig. \[fig:mixing\_time\], and find a good agreement in the scaling of the scrambling time as $t_\epsilon^\star\propto D M^2$.
Closed form solutions to Cartesian graphs {#sec:dynamic}
-----------------------------------------
With the understanding of the equilibrium Page curves, we now proceed to characterize the dynamical evolution towards the equilibrium. We will focus on the Cartesian graphs ${{\cal G}}_D$, which allows closed-from solutions by analog to random walkers. In the continuum limit, the dynamics can be well-described by a diffusive PDE (Sec. \[sec:continuum\]). In particular, we identify a unique parabolic entanglement light cone, followed by an entanglement sudden growth phenomenon, in CV networks (Sec. \[sec:light\_cone\]), as has already been shown in Figs. \[fig:lightcone\] and \[fig:schematic\_2D\].
In Cartesian graphs, the random walk analog to Eq. (\[ensemble\_dynmaics\]) can be understood as independent along each dimension. Thus, we modify the Pascal’s triangle from a usual random walk to obtain the solution \^[(Bi)]{}={\_[d=1]{}\^D [[n\_t]{}]{} }\_[x\_d=-N]{}\^N, \[D\_binomial\] with $n_t=[t/D], n_{x_d,t}=[\frac{x_d}{2}]+[\frac{t}{2D}]$ and ${a\choose b}$ as the binomial factor of $a$-choose-$b$.
With the weights in hand, we can calculate the entanglement entropy of an arbitrary subsystem ${{\cal L}}$. For example, we can consider ${{\cal L}}=\{\bm x^\prime| x_d^\prime<x_d, 1\le d \le D\}$, i.e., the system is cut into two parts by a high-dimensional plane. Then we can obtain the ensemble averaged total transmissivity $$\begin{aligned}
&\braket{\eta_{{{\cal L}},t}}^{(\rm Bi)}=\sum_{\bm x\in {{\cal L}}}\braket{ w_{\bm x^\prime,t}}^{(\rm Bi)}=\frac{1}{2^{t+D}}\prod_{d=1}^D2\sum_{n^\prime=0}^{n_{x_d,t}} {n_t \choose n^\prime}=\prod_{d=1}^D
\nonumber
\\
& \left[ 1-\frac{1}{2^t}{{n_t}\choose{1+n_{x_d,t}}}F(1,1-n_t+n_{x_d,t},2+n_{x_d,t}, -1)\right],
\label{etat}\end{aligned}$$ where $F$ is the hypergeometric function. We expect this to hold up to rounding errors from the integers. In the following, we compare the exact solution of weights $\braket{\bm w_t}$ from numerically solving Eq. (\[ensemble\_dynamics\_matrix\]) and the binomial solution of weights $\braket{\bm w_t}^{(\rm Bi)}$ in Eq. (\[D\_binomial\]).
As shown in Fig. \[fig:1D\_solution\_compare\], we see a very good agreement in the 1-D case, up to rounding errors before boundary effects comes in. When there is a finite boundary, standard techniques like image source methods can give more precise solutions. In fact, the continuum limit of Gaussian solutions, as we will present in Sec. \[sec:continuum\], also agrees well with the above results. The perfect agreement of the weights directly indicates the validity of the solution for the entanglement entropy, as demonstrated in Fig. \[fig:1D\_entropy\_compare\]. Note that due to the symmetry among different dimensions, it suffices to consider the 1-D case; however, results in higher dimension reveals more interesting dynamics.
The 2-D random circuit results are already shown in Fig. \[fig:schematic\_2D\], where the region is chosen as each single mode at $(x_1,x_2)$. There, we can see a clear light cone similar to the 1-D case. Here we consider two alternative choices. First, as an analog to the 1-D choice, we can choose the region on the corner ${{\cal L}}=[-N,x_1]\times [-N,x_2]$, as shown in Fig. \[fig:2D\_squares\](a1)-(a4). We see a gradual saturation to the equilibrium, where entanglement entropy along $(x_1+N)(x_2+N)={\rm constant}$ are about equal, as the total weights $\eta_{{{\cal L}},t}$ are equal along this line. Second, we can choose ${{\cal L}}=[-|x_1|,|x_1|]\times [-|x_2|,|x_2|]$ as squares centered at the origin. In Figs. \[fig:2D\_squares\](b1)-(b4) We can also see gradual saturation to equilibrium, where the entanglement entropy along $|x_1||x_2|={\rm constant}$ are equal, as the total weights $\eta_{{{\cal L}},t}$ are equal along this line. To enable comparison in 2-D, we consider cross sections with $x_1=x_2$. As shown in Fig. \[fig:2D\_cross\_section\], in both choices of the region ${{\cal L}}$, good agreement between the circuit results and the random-walk results can be seen.
In the above, we see the entanglement entropy from exact ensemble-averaged evolution of Eq. (\[ensemble\_dynamics\_matrix\]) (combined with Eq. (\[SL\])) and the actual results from numerical solving the entropy agree well, therefore verifying the underlying random-walk model.
Below, we further address the continuum limit and the entanglement light cone.
### Continuum limit {#sec:continuum}
We can take the continuum limit of Eq. (\[D\_binomial\]) and Eq. (\[etat\]), which give a $D$ dimensional Gaussian function \^[(Ga)]{} = . \[D\_gaussian\] and the corresponds Gaussian error function \^[(Ga)]{} = \_[d=1]{}\^D \[etaxt\_Gaussian\] In Fig. \[fig:1D\_solution\_compare\], we see a good agreement of the above Gaussian approximation with the other solutions.
The above solutions can be written as a function of $\tilde{\bm x}=\bm x/M$ and $\tilde{t}=t/M^2$ up to normalization, which is well-defined in the continuum limit of $M\to\infty$. We verify the continuum limit in 1-D by calculating the deviation measured by the relative 1-norm between the entanglement entropy $\braket{S(\tilde{x},t)}$ and the static value $\braket{S(\tilde{x},\infty)}$ as $$\delta_{\rm dy}(t) = \frac{\|\braket{S(\tilde{x},t)}-\braket{S(\tilde{x},\infty)}\|_1}{\|\braket{S(\tilde{x},\infty)}\|_1} \label{eq:delta_rn}$$ in the dynamic (short as ‘dy’) process (see Fig. \[fig:dynamic\_scale\_invariance\] (a)), where $\|f(\tilde{x})\|_1=\sum_{\tilde{x}} |f(\tilde{x})|$ sums over the spatial coordinates. We see the re-scaled curves overlap well for systems with different sizes. We can also directly verify in Fig. \[fig:dynamic\_scale\_invariance\] (b)(c) that the entanglement entropy agrees well after re-scaling. It is also worthy to point out that the continuum limit identified for the above single-squeezer case also holds for the multiple-squeezer cases (see Appendix \[app:conti\_multi\]).
The continuum limit in Eq. (\[D\_gaussian\]) naturally brings a PDE that descrbes the dynamics \_t w\_[x,t]{}=\^2 w\_[x,t]{}. \[PDE\_weights\] The above holds for Cartesian graphs ${{\cal G}}_D$, in general one needs to adopt the ‘$\nabla^2$’ operator to a graph. Nevertheless, we can obtain observations from ${{\cal G}}_D$. First, for any region ${{\cal L}}$, as the continuum limit of Eq. (\[etat\_def\]), we have $\eta_{{{\cal L}},t}= \int_{{\cal L}}d\bm x w_{\bm x,t} $, therefore the time derivative \_[t]{} \_[[[L]{}]{},t]{}=\_[[L]{}]{}dx \^2 w\_[x,t]{}=\_[[[L]{}]{}]{} d\_[x]{}w\_[x,t]{} \[PDE\_eta\] becomes a loop integral on the boundary $\partial {{\cal L}}$ of the ‘flow’ $\nabla w_{\bm x,t}$ along the normal direction $\hat{\bm n}_{\bm x}$. As the continuum limit of Eq. (\[eta\_boundary\]), it shows that the entanglement entropy’s dynamics is governed by the boundary.
The above PDEs (\[PDE\_weights\]) and (\[PDE\_eta\]) describe the evolution of weights, one relies on Eq. (\[SL\]) to connect to the entanglement evolution. Alternatively, one can directly focus on the entanglement entropy and design a coupled nonlinear diffusive epidemiology model to describe the entanglement dynamics, as we present in Appendix \[app:epidemiology\]. We note that both nonlinear diffusion equations [@nahum2017quantum] and epidemiology models [@qi2018quantum] have been separately used in modeling quantum information scrambling. This phenomenological model shows an interesting combination of both to describe a unique CV entanglement growth process.
### Entanglement light cone and sudden growth {#sec:light_cone}
We now proceed to identify unique phenomena for the entanglement evolution. One of such an evolution is depicted in Fig. \[fig:lightcone\](a), where we see entanglement diffusively spreads from the source at the origin. We can introduce an entanglement light cone (green lines) and a wave front of the entanglement sudden growth (black lines), as will be explained in the following paragraphs.
To further understand the dynamics, we focus on particular modes and consider $\braket{S(x,t)}$ as a function of $t$ (see Fig. \[fig:spot\_evo\_M801\](a) for examples). We observe a three-stage evolution: (1) In the first period $0\le t < T_1$, the entanglement $\braket{S(x,t)}$ is almost zero. This is the time period before the entanglement light cone reaches location $x$. (2) In the second period $T_1 \le t < T_1+T_2$, the entanglement light cone reaches the spot and causes a rapid increase in $\braket{S(x,t)}$, after which $\braket{S(x,t)}$ gets close to $\braket{S(x,\infty)}$. (3) In the last period $T_1+T_2 \le t $, $\braket{S(x,t)}$ gradually saturates towards $\braket{S(x,\infty)}$. We numerically investigate the scaling of $T_1, T_2$ with respect to the distance $\Delta \tilde{x}$ to the initial squeezer, which reveals the physics of entanglement growth.
In the first period $t<T_{1}$, the entanglement entropy at each spot is negligible. As we see in Fig. \[fig:lightcone\](a), the green curves show the threshold $T_1$ for spots at different distances $\Delta \tilde{x}$—a parabolic entanglement light cone much slower than the usual linear light cone of operator spreading. The second period $T_1\le t <T_1+T_2$ describes the wave-front of entanglement rapid growth. As we see in Fig. \[fig:lightcone\](a), the black curve depicts the threshold $T_1+T_2$ for spots at different distances $\Delta \tilde{x}$. The parabolic shape again indicates a diffusion behavior.
This parabolic light cone can be explained by our statistical theory. We want a constant fraction of the maximum entanglement $\epsilon S_0(r)=S(\eta_{{{\cal L}},t})$ in Eq. (\[SL\]), combining with Eq. (\[etat\]) or Eq. (\[etaxt\_Gaussian\]) we can solve $T_1, T_1+T_2$ precisely, despite the analytical formula being lengthy, one immediately recognizes the scaling T\_1, T\_2 \~()\^2 f(r), \[eq:T1T2\] with some function $f$ of the squeezing strength $r$. Indeed, as shown by Fig. \[fig:spot\_evo\_M801\](b), the green curve (entanglement light cone) and the black curve (entanglement sudden growth) both agrees well with the quadratic fitting. This is indeed consistent with the OTOC diffusion identified in [@zhuang2019scrambling], revealing an unique universal behavior intrinsic to CV quantum networks and absent in DV circuits [@nahum2017quantum; @nahum2018operator].
Growth of multi-partite entanglement measured by distributed sensing {#sec:multipartite_E}
--------------------------------------------------------------------
So far we have focused on bipartite entanglement between a subsystem ${{\cal L}}$ and its complement ${{\cal R}}$. In quantum networks, many applications often require multipartite entanglement, which is in general difficult to characterize [@adesso2014]. Here we take an operational approach from a quantum sensing perspective. An important application of the entanglement generated in such a random quantum network is distributed sensing [@zhuang2018distributed; @zhuang2019physical; @guo2020distributed; @xia2019entangled], where multi-partite entanglement enables an improvement in the measurement sensitivity. In the case of measuring uniform real displacements of amplitude $\alpha$ on all modes, one can prove that considering a total mean photon number $|{{\cal G}}|N_S$, the optimum $|{{\cal G}}|$-mode separable state can only offer a variance V\_C=\~ in estimating the displacement $\alpha$ (the standard quantum limit). Therefore beating the above precision limit is an evidence of entanglement. In fact, one can show that the optimum precision attainable by [*all*]{} entangled state is V\_E=\~, which possesses the Heisenberg scaling of $V_E\sim 1/|{{\cal G}}|^2$ that is only possible with multipartite entanglement. Therefore, we can define an entanglement witness for any $|{{\cal G}}|$-mode state $\rho$ with total energy $|{{\cal G}}|N_S$ as [[E]{}]{}()=\_[[LOCC]{}\^]{}\_2(V\_C/V()), where $V(\rho)$ is variance achievable by performing a local-operations (LO) and classical communications (CC) on the input $\rho$ (hence ${\rm LOCC}^\prime$), while keeping the total energy conserved (see Fig. \[fig:witness\]). We maximize over all such LOCC schemes. We have ${{\cal E}}(\rho)= 0$ for all separable states [@Note4], and ${{\cal E}}(\rho)\le \log_2(V_C/V_E)\sim \log_2|{{\cal G}}|$ for all states.
For the state $\rho(t)$ generated in the single-squeezer random network at time step $t$, we can design an estimator to obtain a lower bound of ${{\cal E}}(\rho(t))$. Given the weight $\{w_{\bm x,t}\}$ on distributing the SV state to $|{{\cal G}}|$ modes, we can design the following measurement protocol. First, we perform a phase rotation on each mode such that the displacements act on the corresponding squeezed quadrature; then we perform homodyne measurements on the corresponding quadratures to obtain the results $\{\tilde{\alpha}_{\bm x,t}\}_{\bm x\in {{\cal G}}}$. The estimator $\tilde{\alpha}=\sum_{\bm x\in {{\cal G}}} \sqrt{w_{\bm x,t}} \tilde{\alpha}_{\bm x,t}/\sum_{\bm x\in {{\cal G}}} \sqrt{w_{\bm x,t}}$, which gives the variance V(t)=, where the effective number of modes $\overline{|{{\cal G}}|}=(\sum_{\bm x \in {{\cal G}}} \sqrt{w_{\bm x,t}})^2\in[1,|{{\cal G}}|]$ that are entangled provides the advantage.
Combing the weights in Eqs. (\[D\_binomial\]) and (\[D\_gaussian\]), we can obtain the effective entangled mode number for the $D$ dimensional Cartesian graph ${{\cal G}}_D$ d\^[D]{}x\^[(Ga)]{} = (2(2t/D)\^[1/2]{})\^D which leads to the entanglement witness [[E]{}]{}((t))\_2(V\_C/V(t))\_2(8t/D), before the boundary effect comes in, when the effective modes become comparable to $|{{\cal G}}|$ at $t\sim |{{\cal G}}|^2$.
Multiple squeezers: sparse limit {#sec:multiple_squeezers}
================================
In Sec. \[sec:statistical\_theory\], we focus on random networks with a single squeezer and present a thorough theory for the entanglement dynamics and equilibrium, via an exact mapping to random walk on a graph. Quantum networks are likely to have multiple squeezers; therefore, we extend our analyses to random networks with multiple squeezers in this section. A surprising linear superposition law is numerically observed and theoretically explained.
We begin with an intuitive example of three squeezers in 1-D Cartesian graph in Fig. \[fig:lightcone\](b). The overall evolution of the entanglement entropy looks like a linear superposition of three independent squeezers, despite the nonlinearity of the entanglement dynamics. Following this observation, we consider a random circuit $\mathbb{C}$ with $N_q$ squeezers at different space-time coordinates $\{\bm \xi_k^\star=(\bm x_k^\star,t_k^\star)\}_{k=1}^{N_q}$, with squeezing strengths $\{r_k\}_{k=1}^{N_q}$. Linear superposition (spp) means the entanglement entropy of subsystem ${{\cal L}}$, $\braket{S({{\cal L}},t)}\simeq \braket{S_{\rm spp}({{\cal L}},t)}$, where $$\braket{S_{\rm spp}({{\cal L}},t)} = \sum_{k=1}^{N_q} \braket{S_k({{\cal L}},t)}
\label{eq:superposition}$$ is a simple sum of the ensemble averages $\braket{S_k({{\cal L}},t)}$. Here $S_k({{\cal L}},t)$ is generated from a random circuit $\mathbb{C}_k$ with a single squeezer of strength $r_k$ at $\bm \xi_k^\star$, therefore can be calculated by the random-walk mapping in Eqs. (\[ensemble\_dynamics\_matrix\]) and (\[SL\]). Note that the random beamsplitters in all $N_q$ single-squeezer circuits $\{\mathbb{C}_i\}_{k=1}^{N_q}$ and the original circuit $\mathbb{C}$ are independent. To test linear superposition, we numerically calculate the deviation per mode S\_[spp]{}(t) = -\_1. \[eq:ab\_deviation\_spp\] To evaluate the relative deviation, we can also rescale the deviation relative to the steady state value, as $\delta_{\rm spp}(t)=\Delta S_{\rm spp}(t)/(\|\braket{S({{\cal L}},\infty)}\|_1/|{{\cal G}}|)$. Both deviations are system-size independent in the continuum limit.
In Fig. \[fig:dynamic\_spacetime\_spp\_N3\](a), we evaluate the deviation for the 1-D three-squeezer case considered in Fig. \[fig:lightcone\](b). We see that the relative deviation $\delta_{\rm spp}(t)$ is small ($<2\%$) through the entire dynamical evolution. To be more explicit, in Figs. \[fig:dynamic\_spacetime\_spp\_N3\](b) and (c), we directly plot $\braket{S(\tilde{x},t)}$ (blue) at various times, which agrees well with the superposition result $\braket{S_{\rm spp}(\tilde{x},t)}$ (red).
Following the above observation, Sec. \[sec:spp\_theory\] provides a theory at the sparse squeezers limit, which predicts linear superposition for both the equilibrium Page curves and the dynamical evolution; These two aspects are then investigated in Sec. \[sec:Page\] and Sec. \[sec:superposition\].
Theory of the sparse squeezers limit {#sec:spp_theory}
------------------------------------
Inspired by the above numerical findings in 1-D, we present the following theory to explain the linear superposition law. For the circuit $\mathbb{C}$, similar to Eq. (\[axt\]), each mode at $\bm x\in {{\cal G}}$ and time $t$ can be written as a\_=\_[k=1]{}\^[N\_q]{}e\^[i\_[;k]{}]{} a\_[SV;k]{}+[vac]{}, \[axt\_multi\] where we use the simplified notation $\bm \xi=(\bm x,t)$. Since the squeezers’ locations $\bm \xi_k^\star$ are different, all fully random phases $\{\theta_{\bm \xi;k}\}$ are independent. Each set of positive weights $\{w_{\bm \xi;\bm \xi_k^\star}\}_{\bm x\in {{\cal G}}}$ describes the energy-splitting among the modes of each single-mode SV $a_{\rm SV;k}$. From these weights, it immediately follows that a total portion $\eta_{{{\cal L}},t;k}=\sum_{\bm x\in {{\cal L}}} w_{\bm \xi;\bm \xi_k}$ of each single mode squeezer end up in subsystem ${{\cal L}}$.
To evaluate the entropy of ${{\cal L}}$, similar to Sec \[sec:mapping\], we consider a set of passive linear optics unitaries $\{U_{{{\cal L}}_{s},t}\}_{s=1}^{N_q}$ to manipulate the power distribution of the SV within ${{\cal L}}$ (see Fig. \[fig:schematic\_transform\_multi\]). The first transform $U_{{{\cal L}}_1,t}$ acts on the entire system ${{\cal L}}_1={{\cal L}}$ to concentrate the first SV part to the mode at an arbitrary mode $\bm y_1\in {{\cal L}}_1$ as a\_[[[L]{}]{}\_1;1]{}=a\_[SV;1]{}+\_[k=2]{}\^[N\_q]{}e\^[i\_[;k]{}\^[(1)]{}]{} a\_[SV;k]{}+[vac]{}, where $\eta_{{{\cal L}}_1,t;1}=\sum_{\bm x\in {{\cal L}}_1} w_{\bm \xi; \bm \xi_k}$ is the total portion of $a_{\rm SV;1}$ in the subsystem ${{\cal L}}_1$. For the first transform, we have $\eta_{{{\cal L}}_1,t;1}= \eta_{{{\cal L}},t;1}$. We can denote the remaining $(M-1)$ modes after the transform as ${{\cal L}}_2$, with the weights $\{w_{\bm \xi;\bm \xi_k}^{(1)}\}_{\bm x\in {{\cal L}}}$ for squeezers $\{a_{\rm SV;k}\}_{k= 2}^{N_q}$. Note that a $w_{(\bm y_1,t);\bm \xi_k}^{(1)}$ portion of each squeezer $a_{\rm SV;k}$ ($k\ge2$) is also mixed in the mode $a_{{{\cal L}}_1;1}$, and not in the new subsystem ${{\cal L}}_2$.
After the first $(s-1)$ transforms ($N_q\ge s\ge2$), we have subsystem ${{\cal L}}_{s}$ with $|{{\cal G}}|-s+1$ modes . The $s$-th transform $U_{{{\cal L}}_s,t}$ acts on the subsystem ${{\cal L}}_{s}$ to concentrate the power of the SV $a_{\rm SV; s}$ to the mode at $\bm y_s\in {{\cal L}}_s$ as a\_[[[L]{}]{}\_s;s]{}=a\_[SV; s]{}+\_[k=s+1]{}\^[N\_q]{} e\^[i\_[;k]{}\^[(s)]{}]{} a\_[SV;k]{}+[vac]{}, where $\eta_{{{\cal L}}_{s},t;s}=\sum_{\bm x\in {{\cal L}}_s} w_{\bm \xi; \bm \xi_k}^{(s-1)} \le \eta_{{{\cal L}},t;s}$ is total power portion of $a_{\rm SV; s}$ in the subsystem ${{\cal L}}_s$.
After all transforms $\{U_{{{\cal L}}_{s},t}\}_{s=1}^{N_q}$, we obtain modes $\{a_{{{\cal L}}_s;s}\}_{s=1}^{N_q}$, which contains all the SVs in ${{\cal L}}$ [@Note1]; while the rest of the modes are all in vacuum. Due to the unitarity of all transforms, the entropy of the original system $S({{\cal L}},t)=S\left(\{a_{{{\cal L}}_s;s}\}_{s=1}^{N_q}\right)$ equals the entropy of these modes. To further evaluate the entropy, we impose two constraints—independence of weights and $N_q\ll |{{\cal G}}|$—on the random circuit to enable an approximate solution.
First, we assume the weights $\{w_{\bm \xi;\bm \xi_k^\star}\}_{\bm x\in {{\cal G}}}$ are independent between different squeezers (indexed by $k$). This assumption is true in two scenarios: (1) when any two squeezers are far away from each other in terms of the shortest path on the graph and/or the time difference. In this case, the independence of weights hold at any time $t$; (2) as long as the number of squeezers $N_q\ll |{{\cal L}}|$ and the squeezers do not act on top of each other, the correlations between the weights will be small at late time. This is especially true for the Page curve $\braket{S({{\cal L}},\infty)}$. The independence of the weights means that after the $s$-th passive linear unitary, the new weights $\{w_{\bm \xi;\bm \xi_k}^{(s)}\}_{\bm x\in {{\cal L}}}$ for the remaining squeezers are still randomly distributed across the entire system (not concentrated on any mode).
Second, we assume that the total number of squeezers is small, i.e., $N_q\ll |{{\cal G}}|$. This allows us to approximate the weights concentrated from subsystem ${{\cal L}}_{s}$ as the global total weights, i.e., $
\eta_{{{\cal L}}_{s},t;s}\simeq \eta_{{{\cal L}},t;s}, 1\le s \le N_q.
$ Therefore, we conclude that after the set of passive unitaries $\{U_{{{\cal L}}_{s},t}\}_{s=1}^{N_q}$, we can approximately obtain a set of modes {a\_[[[L]{}]{};s]{}a\_[SV;s]{}+[vac]{}}\_[s=1]{}\^[N\_q]{}, which is a product of lossy single-mode SVs. From additivity of entropy, we arrive at linear superposition $\braket{S({{\cal L}},t)}\simeq \braket{S_{\rm spp}({{\cal L}},t)}$, note that each single-squeezer term can also be written out explicitly as $S(\eta_{{{\cal L}},t;s})$ from Eq. (\[SL\_preliminary\]) by replacing $r$ with $r_k$.
Note that linear superposition law holds not only for the dynamical evolution at any $t$ (as systematically explored in Section \[sec:superposition\]), but also for the equilibrium Page curves at $t=\infty$ (see Section \[sec:Page\]).
Page curves with multiple squeezers {#sec:Page}
-----------------------------------
In terms of Page curves, as we explained in Sec. \[sec:equi\_solu\_general\], the graph topology is irrelevant as the entire dynamics is equivalent to a passive, Haar-random, global unitary; therefore, we can simply consider the 1-D case. Furthermore, a Gaussian unitary can be decomposed into a layer of squeezers concatenated by passive unitaries (Euler decomposition, see Appendix \[app:random\_mat\]), we can effectively push all squeezers to the first step and then apply a global Haar random passive linear transform (see Fig. \[fig:schematic\_Page\]). Therefore, each Page curve is characterized by a list of squeezing strength $\{r_k\}_{k=1}^{N_q}$. In this case, as long as $N_q\ll M$, we can regard the squeezers as sparse. We denote the average squeezing strength as $\overline{r}=\sum_{\ell=1}^{N_q} r_\ell/N_q$ and the density of squeezers $\overline{n_q}=N_q/M$. When $N_q\ll |{{\cal L}}|$, superposition $\braket{S({{\cal L}},\infty)}\simeq \braket{S_{\rm spp}({{\cal L}},\infty)}$ holds, with each single-squeezer result given in Eq. (\[SL\_page\]). In the large squeezing limit, we can further utilize Eq. (\[SL\_page\_sim\]) to obtain $$\begin{aligned}
&S\left({{\cal L}},\infty\right)
\nonumber
\\
&\simeq M \overline{n_q}\left\{\frac{1}{2}\log_2\left[\frac{|{{\cal L}}|}{|{{\cal G}}|}\left(1-\frac{|{{\cal L}}|}{|{{\cal G}}|}\right)\right]+\frac{1}{\ln2}\left(\overline{r}+1\right)-1\right\},
\label{SL_sim_page}\end{aligned}$$ We see a dependence on statistical quantities $\overline{r}, \overline{n_q}$, while the shape of the curve is invariant in the bulk at large squeezing limit, as can be seen in Fig. \[fig:static\_spp\] (b) and (c).
We numerically examine the validity of the linear superposition in Page curves via the relative deviations $\delta_{\rm spp}(t)$ in Fig. \[fig:static\_spp\]. Indeed when the squeezer density is low, we see a good agreement, as shown in Fig. \[fig:static\_spp\] (b); while when the squeezers are dense, substantial deviation can be found, as shown in Fig. \[fig:static\_spp\] (c). The transition is captured by the relative deviation $\delta_{\rm spp}(\infty)$ in Fig. \[fig:static\_spp\] (a), where $\delta_{\rm spp}(\infty)$ increases linearly with the squeezer density $\overline{n_q}$.
Although when $\overline{n_q}$ is not small, superposition does not hold, we can further numerically explore the Page curves’ dependence on the parameters. To consider the typical case, we randomly generate $\{r_k\}_{k=1}^{N_q}$ and plot the normalized Page curves in Fig. \[fig:pages\](b) for different system sizes. We find that all Page curves coincide as long as $\overline{r}, \overline{n_q}$ are the same; and independent of $M$ when $M$ is large Similar to the single-squeezer case. Therefore, we can plot the case with identical squeezer $\overline{r}$ as a benchmark. To demonstrate the coincidence, we consider random cases with $r_i \in [0.8\overline{r}, 1.2\overline{r}], 1\le i \le N_q$, while guaranteeing the average equaling $\overline{r}$ [@Note2]. Indeed, all the random points lie right on top of the benchmark of uniform squeezing with the same $\overline{r}, \overline{n_q}$. Since $\overline{r}$ and $\overline{n_q}$ are both scale-free statistical quantities, this indicates a well-defined continuum limit of CV Page curves.
Dynamics with linear superposition {#sec:superposition}
----------------------------------
Fig. \[fig:dynamic\_spacetime\_spp\_N3\] already confirms superposition through the entire evolution for a simple case, to verify it in a more general setting, we consider circuits with squeezers randomly distributed in space-time. Guided by the theory in Sec. \[sec:spp\_theory\], we expect the minimum distance between the squeezers to be the dominating factor of the deviation $\Delta S_{\rm spp}(t)$; However, simple random sampling methods inevitably lead to the appearance of clusters, where squeezers can be close to each other. In order to tune the minimum distance $d$, while preserving the random nature of the circuit set-up so that the results apply in general, we adopt the random Poisson sampling method [@bridson2007] to control the space-time distances between the squeezers.
With the squeezers randomly chosen, we evaluate the entanglement dynamics for circuits with random squeezers of different minimum distances. Two examples are given in Fig. \[fig:sparse\_dense\]: When the squeezers are sparse ($d=60$ is large) as indicated by the cyan dots in Fig. \[fig:sparse\_dense\](c), the superposition results $\braket{S_{\rm spp}(\tilde{x},t)}$ (dashed lines) agree well with the true values $\braket{S(\tilde{x},t)}$ (solid lines) at various time steps in Fig. \[fig:sparse\_dense\](d); when the squeezers are dense ($d$ is small), substantial deviations from the linear superposition can be observed, as demonstrated in Fig. \[fig:sparse\_dense\](a)(b) for the case of $d=20$.
To systematically examine the transition from ‘dense’ to ‘sparse’ squeezers, we calculate the deviation $\Delta_{\rm spp}(t)$ for distances $0\le d\le 500$ at various time steps in Fig. \[fig:sqs\_spp\_rn\], where the deviation decreases monotonically with $d$ up to numerical precision. Although for a smaller $d$, the deviation is larger, the relative deviation is still below $10\%$ of the static value $\|\braket{S(\tilde{x},\infty)}\|_1/M$ (orange). Therefore, we conclude that in generic CV quantum networks with sparse squeezers, linear superposition of entanglement growth holds.
Discussion
==========
In this paper, we reveal a mapping between entanglement formation dynamics in random CV networks to random walk on general graphs. This mapping allows analytical solutions of the entanglement entropy dynamics, Page curves and scrambling time for an arbitrary network topology. On networks respecting locality, the solution enables the understanding of three unique features of entanglement formation dynamics—an entanglement light cone, an entanglement sudden-growth period and parameter-dependent Page curves. Our results have implications in quantum network protocol design, e.g., the entanglement light cone will place bounds on the latency in entanglement distribution, and also on the fundamental understanding of many-body systems. Lastly, let us point out some future directions: it will be interesting to extend the entanglement light cone to long-range interacting systems; extending the multipartite entanglement witness to more general input states will bring further insights of quantum entanglement; exploration of the connection between entanglement dynamics with statistical properties of random networks such as connectivity distribution will lead to a full statistical theory of CV quantum networks.
This work is supported by Army Research Office under Grant No. W911NF-19-1-0418 and the University of Arizona. Q.Z. thanks Norman Yao and Beni Yoshida, Saikat Guha, Bihui Zhu for discussions.
Details of Gaussian unitaries and random matrices {#app:random_mat}
=================================================
A zero-mean general Gaussian unitary $U_{\bm S}$ is specified by a symplectic matrix $\bm S$, which can be decomposed into a product of ‘passive linear optics’ operations and ‘squeezing’ operations (Euler decomposition), U\_[S]{}=U\_[K]{} U\_[S({r\_k})]{} U\_[L]{}, \[Euler\_decomp\] where $\bm K, \bm L$ are symplectic orthogonal, and correspond to beamsplitters and phase-shifters. These two components are often named ‘passive linear optics’ since they preserve the mean photon number. Single-mode squeezing operations, which changes photon number, are characterized by their strengths $r_k$ and represented by the diagonal symplectic matrix $\bm S(\{r_k\}) = \bigoplus_k \text{Diag}\left( e^{-r_k}, e^{r_k} \right)$, where ‘$\bigoplus$’ denotes the direct sum. An additional property of zero-mean Gaussian unitary is that concatenating multiple Gaussian gates still produces a Gaussian gate, i.e., $U_{\bm S_1} U_{\bm S_2}=U_{\bm S_1 \bm S_2}$.
First, we specify a Haar random two-mode passive linear transform parameterized by angles $\theta$, $\phi_1, \phi_2$ and $\phi_3,\phi_4$, as shown below.
It consists of a beamsplitter described by the unitary $\exp\left[\theta \left(\hat{a}\hat{b}^{\dagger} - \hat{a}^{\dagger}\hat{b}\right)\right]$ on input modes $a,b$, where $\cos^2(\theta)$ is the transmissivity of the beamsplitter and four additional phase shifters, $R(\phi_1)$, $R(\phi_2)$, $R(\phi_3)$ and $R(\phi_4)$, two of which are ahead of beamsplitters and the others after it. The unitary of a phase shifter with amount $\phi$ on mode $\hat{a}$ is described as $R(\phi) = \exp\left(i\phi\hat{a}\hat{a}^{\dagger}\right)$. The corresponding unitary matrix and symplectic matrix are respectively $$\begin{aligned}
&\bm U_{B} = \begin{pmatrix}
e^{i(\phi_1 + \phi_3)}\cos\theta & -e^{i(\phi_2 + \phi_3)}\sin\theta \\
e^{i(\phi_1+\phi_4)}\sin\theta & e^{i(\phi_2+\phi_4)}\cos\theta
\end{pmatrix},
\label{eq:bs_unitary}
\\
&\bm S_{B} = \bm P^{T} \begin{pmatrix}
\operatorname{Re}(\bm U_{B}) & -\operatorname{Im}(\bm U_B)\\
\operatorname{Im}(\bm U_B) & \operatorname{Re}(\bm U_{B})
\end{pmatrix} \bm P.
\label{eq:bs_symplectic}\end{aligned}$$ Here the $4\times 4$ permutation matrix $\bm P$ exchanges the second and third item, i.e., $\bm P=\bm I_{14}\bigoplus \bm X_{23}$, with Pauli matrices $\bm I, \bm X$. The notations $\operatorname{Re}(.), \operatorname{Im}(.)$ denote the real and imaginary part of a complex matrix. To ensure Haar random of Eq. (\[eq:bs\_unitary\]), all phase angles $\phi_1,\phi_2,\phi_3, \phi_4$ are chosen uniform in $[0,2\pi)$ and $\cos^2\theta$ uniform in $[0, 1]$. Here we also reveal a one-to-one correspondence between orthogonal symplectic matrix and unitary matrix through Eq. (\[eq:bs\_symplectic\]) (and its generalization to more modes), which allows us to sample multi-mode Haar random passive linear transform from standard algorithms in Ref. [@mezzadri2006] directly.
Entanglement entropy {#app:entropy}
====================
The entanglement entropy $S({{\cal L}},t)$ can be efficiently calculated through keeping track of the covariance matrix of ${{\cal L}}$ (Equivalently, one can consider the covariance matrix of subsystem ${\protect {{\cal R}}}$.) V\_[ij]{}=, i,j [[L]{}]{}, where $\{,\}$ is the anti-commutator and $\bm X_i,\bm X_j$ are components of quadrature operators $\bm X$ that corresponds to subsystem ${{\cal L}}$. Under the dynamics of each local unitary $U_{\bm S}$, the covariance matrix evolves as $\bm V\to \bm S \bm V \bm S^T$. From the symplectic eigenvalues [@Weedbrook_2012] $\{\nu_i, 1\le i \le |{{\cal L}}|\}$ of $\bm V$, we can obtain the von Neumann entropy of ${{\cal L}}$ (which equals that of ${{\cal R}}$ due to purity of the global system) S([[L]{}]{},t)=S([[R]{}]{},t)=\_[i=1]{}\^[|[[L]{}]{}|]{} g((\_i-1)/2), where each term $
g(x) = (x+1)\log_{2}(x+1) - x\log_{2}\left(x\right)
$ is the entropy of a thermal state with mean photon number $x$. Alternatively, we can also choose Renyi entropy of the order $\alpha$, which can be calculated as [@Camilo2019] S\_([[L]{}]{},t) = \_[i=1]{}\^[|[[L]{}]{}|]{} \_2[g\_((\_i-1)/2)]{}, \[eq:Renyi entropy\] where the $g_{\alpha}\left(\cdot\right)$ is defined as $g_{\alpha}(x) = \left(x+1\right)^{\alpha} - \left(x\right)^\alpha$
Combining the above, when there is a single non-zero symplectic eigenvalue, we can write a single function $g(x)$, as we have done in Eq. (\[SL\_preliminary\]): for von Neumann entropy $g(x)= (x+1)\log_{2}(x+1) - x\log_{2}\left(x\right)$ and for Renyi entropy $g(x)=\log_2\left[(x+1)^\alpha-x^\alpha\right]/(\alpha-1)$.
In the large squeezing limit, to the first order, we have $S\left({{\cal L}},t\right)= S({{\cal R}},t)\simeq \log_2(\prod_{i=1}^{|{{\cal L}}|}\nu_i)=\log_2 (\sqrt{\det \bm V})$, which equals a projected phase space volume similar to the one identified through average OTOCs in Ref. [@zhuang2019scrambling]. The purity of the entire system also guarantees a conservation of the entire phase space volume, analog to the “Quantum Liouville’s theorem” in Ref. [@zhuang2019scrambling]. This interpretation also holds for Renyi entropy: e.g. for the Renyi-2 entropy $S_{2}({{\cal L}})=S_{2}({{\cal R}})=\log_2(\sqrt{\det \bm V})$, which can be interpreted as the logarithmic volume; For other choices of $\alpha\neq 1$, when $\nu_i$’s are large, we still have $S_\alpha(\rho)\simeq \log_2(\sqrt{\det \bm V})$. In this sense, the entanglement growth dynamics can be connected to projected phase-space volume growth (also see Ref. [@lerose2020]), and it is clear that any choices of entanglement entropy will have similar dynamics.
Note that here Renyi entropy is not directly connected with t-designs [@liu2018generalized], since the symplectic eigenvalues are related to eigenvalues of $i\bm\Omega\bm V$, which are not simple polynomials in general. Here $\bm \Omega=\bigoplus_{k=1}^N -i\bm Y$ is the symplectic metric.
It is worthy to comment that the diffusive growth of entanglement identified in Sec. \[sec:light\_cone\] does not contradict the known result that good approximations ($t$-design) of the global Haar random unitary only requires a random circuit with the number of layers linear in system size [@harrow2018approximate]. Mathematically, in a CV system it is the covariance matrix $\bm V$ that is transformed as $\bm V\to \bm S \bm V \bm S^T $ under the Haar random matrix $\bm S$. Both the von Neumann and Renyi entanglement entropies are given by the symplectic eigenvalues of $\bm V$, which is not a simple polynomial of $\bm S$. This contrasts with the case in DV systems, where the random matrices act on the density matrix directly and Renyi entropies are simple polynomials of them.
Details of the random-walk mapping {#app:random_derivation}
==================================
In terms of the weights, consider two modes $\bm x, \bm x^\prime$ on a connected edge
\[bs1\] $$\begin{aligned}
a_{\bm x,t}&=e^{i\theta_{\bm x,t}}\sqrt{w_{\bm x,t}} a_{\rm SV}+{\rm vac},
\\
a_{\bm x^\prime,t}&=e^{i\theta_{\bm x^\prime,t}}\sqrt{w_{\bm x^\prime,t}} a_{\rm SV}+{\rm vac}.\end{aligned}$$
A general passive gate $U(t,\bm x,\bm x^\prime)$, with transmissivity $\tau$, on modes $\bm x,\bm x^\prime$ would lead to the evolution
\[bs2\] $$\begin{aligned}
&a_{\bm x,t+1}=
e^{i\theta_{\bm x,t+1}^\prime}\left(\sqrt{\tau}\sqrt{w_{\bm x,t}}+\sqrt{1-\tau} e^{i\theta}\sqrt{w_{\bm x^\prime,t}}\right) a_{\rm SV},
\\
&a_{\bm x^\prime,t+1}=
e^{i\theta_{\bm x^\prime,t+1}^\prime}\left(\sqrt{1-\tau} \sqrt{w_{\bm x,t}}-\sqrt{\tau} e^{i\theta}\sqrt{w_{\bm x^\prime,t}}\right) a_{\rm SV},\end{aligned}$$
where the angles $\theta=(\theta_{\bm x^\prime,t}-\theta_{\bm x,t}), \theta_{\bm x,t+1}^\prime,\theta_{\bm x^\prime,t+1}^\prime$ are uniform in $[0,2\pi)$ and we left out the vacuum terms. Therefore we can write the new modes in the same form as Eq. (\[axt\]) with
\[bs3\] w\_[x,t+1]{}&=&w\_[x,t]{}+(1-)w\_[x\^,t]{}\
&& +2,\
w\_[x\^,t+1]{}&=&(1-) w\_[x,t]{}+w\_[x\^,t]{}\
&& -2,
and the angles $\theta_{\bm x,t+1},\theta_{\bm x^\prime,t+1}$ uniformly random in $[0,2\pi)$. We can obtain the ensemble averaged (over the gates $U(t,\bm x,\bm x^\prime)$, with transmissivities $\tau$ uniform in $[0,1)$) evolution as ==(+). This describes a symmetric random walk along the edge $\overline{\bm x\bm x^\prime}$.
Calculation of the variance of $\eta_{{{\cal L}},\infty}$ at equilibrium {#app:eta_var}
========================================================================
The weights $\{w_{\bm x,t}\}$ determine the entanglement, essentially they can be considered as the amplitude squared of a random complex number $\alpha_{\bm x,t}$, i.e. $w_{\bm x,t}=|\alpha_{\bm x,t}|^2$. Normalization requires $\sum_{\bm x\in {{\cal G}}} |\alpha_{\bm x,t}|^2=1$. Due to the Haar randomness at $t=\infty$, we assume that this is the only constraint, therefore $\alpha_{\bm x,t}$’s are random complex numbers on a high dimension sphere. We can calculate the distribution of weights $\eta_{{{\cal L}},\infty}$ through $$\begin{aligned}
&P(\eta_{{{\cal L}},\infty}=\eta)=
\nonumber
\\
&\frac{\int \prod_{\bm x\in {{\cal G}}}d^2\alpha_{\bm x,t}\ \delta\left(\eta-\sum_{\bm x\in {{\cal L}}} |\alpha_{\bm x,t}|^2\right) \delta\left(1-\sum_{\bm x\in {{\cal G}}} |\alpha_{\bm x,t}|^2\right)}{\int \prod_{\bm x\in {{\cal G}}}d^2\alpha_{\bm x,t} \ \delta\left(1-\sum_{\bm x\in {{\cal G}}} |\alpha_{\bm x,t}|^2\right)}
\\
&=\frac{\int \prod_{\bm x\in {{\cal R}}}d^2\alpha_{\bm x,t}\ S_{2|{{\cal L}}|-1}(\sqrt{\eta}) \delta\left(1-\eta-\sum_{\bm x\in {{\cal R}}} |\alpha_{\bm x,t}|^2\right)}{\sqrt{\eta} S_{2|{{\cal G}}|-1}(1)}
\\
&=\frac{ S_{2|{{\cal L}}|-1}(\sqrt{\eta}) S_{2|{{\cal R}}|-1}(\sqrt{1-\eta})}{2\sqrt{\eta}\sqrt{1-\eta}S_{2|{{\cal G}}|-1}(1)}
\\
&=\frac{\Gamma(|{{\cal G}}|)}{\Gamma(|{{\cal L}}|)\Gamma(|{{\cal R}}|)}\eta^{|{{\cal L}}|-1}(1-\eta)^{|{{\cal R}}|-1}.
\label{P_eta}\end{aligned}$$ Here we utilized the $N$-dimensional sphere area formula $$\begin{aligned}
&S_N(R)\equiv \int \prod_{\ell=1}^N dx_\ell \ \delta\left(R-\sqrt{\sum_{\ell=1}^{N+1} x_\ell^2}\right)
\\
&=2R\int \prod_{\ell=1}^N dx_\ell \ \delta\left(R^2-\sum_{\ell=1}^{N+1} x_\ell^2\right)=\frac{2\pi^{\frac{N+1}{2}}}{\Gamma(\frac{N+1}{2})}R^N,\end{aligned}$$ where we have used $\delta(x^2-a^2)=(\delta(x-a)+\delta(x+a))/(2a)$. One can easily verify that Eq. (\[P\_eta\]) is normalized and $\braket{\eta_{{{\cal L}},\infty}}=|{{\cal L}}|/|{{\cal G}}|$. Moreover, the fluctuation can be obtained as (\_[[[L]{}]{},]{}) = .
Continuum limit of the multiple-squeezer case {#app:conti_multi}
=============================================
In this section, we show that the same continuum limit, identified in Sec. \[sec:continuum\] (see Fig. \[fig:dynamic\_scale\_invariance\]) for a single-squeezer, still holds when there are multiple squeezers. In Fig. \[fig:sqs\_log(delta)2\], one layer of $N_q=M$ squeezers are applied in the first time step, and we still see a good agreement between the relative 1-norm (Fig. \[fig:sqs\_log(delta)2\](a)) at different times. Snapshots of entanglement entropy curve at various times also overlap entirely (Fig. \[fig:sqs\_log(delta)2\](b), (c)), after a re-scaling of space-time. Note that in this case, even with a dense squeezer distribution where the superposition principle does not hold anymore, this continuum limit still holds.
Phenomenological model: epidemiology with diffusion {#app:epidemiology}
===================================================
Although the dynamics of $\braket{S(x,t)}$ can be captured by the diffusion of weights and Eq. (\[SL\]) that connects weights to entanglement entropy. Alternative models are possible to capture major phenomena in Sec. \[sec:light\_cone\] for the 1-D Cartesian graph. We notice that in our CV circuits, squeezing behaves as the source of entanglement generation [@tserkis2020]; and while it diffusively spreads out, its strength at each mode also decays due to the effective loss during interactions with other modes; therefore we introduce a field $G(x,t)$ to model this diffusive source that triggers entanglement growth and device the following set of coupled diffusion-growth equations to give a theory prediction $S_T\left(x,t\right)$ for the ensemble averaged entanglement entropy $\braket{S\left(x,t\right)}$
\[couple\_KPZ\] $$\begin{aligned}
&\partial_{t}S_T(x,t) = AG(x,t)\left[c - S_T\left(x,t\right)\right],
\label{infection}
\\
&\partial_{t}G(x,t) = D\partial^2_{x} G(x,t) + fG(x,t)^{2},
\label{KPZ}\end{aligned}$$
where $A, c, D, f$ are four constants. Eq. (\[KPZ\]) describes a nonlinear diffusion process of the ‘source’, where the nonlinear term $fG{\left(x,t\right)}^2$ corrects the early stage dynamics. Eq. (\[infection\]) describes an ‘infectious’ saturation process with a growth rate proportional to ‘uninfected population’, $c - S_T\left(x,t\right)$, triggered by the source $G(x,t)$. We note that both nonlinear diffusion equations [@nahum2017quantum] and epidemiology models [@qi2018quantum] have been separately used in modeling quantum information scrambling. This model shows an interesting combination of both to describe a unique CV entanglement growth process.
The initial condition for the source field $G\left(x,t\right)$ is a delta-function at the squeezer’s position, and uniform zero for the entanglement $S_T\left(x,t\right)$. We choose $G\left(x,t\right)$ obeying the von Neumann boundary condition (reflection boundary conditions) such that $\partial_{x} G(\pm N,t) = 0$. It is straightforward to check that $S_T\left(x,t\right)=c$ is a steady state solution of Eqs. (\[couple\_KPZ\]). Considering the steady state of the real entanglement entropy in Fig. \[fig:pages\_1D\], we expect Eqs. (\[couple\_KPZ\]) to only describe the evolution of entanglement in the bulk. Indeed, when we numerically solve Eqs. (\[couple\_KPZ\]) in a system with $M=501$ modes and fix the steady state $c$ to be the maximal height $S_0(r)$, we find good agreement before boundary effects come in at late time (see Fig. \[fig:model\](a)). In particular, this model does not guarantee zero entanglement entropy on the boundary.
Therefore, Eqs. (\[couple\_KPZ\]) will be able to capture the entanglement growth of the bulk, before boundary effects become important. More precise models can be constructed by further fine-tuning Eqs. (\[couple\_KPZ\]) to different boundaries. As an example, we can implement a position-dependent $c=\braket{S\left(x,\infty\right)}$, such that steady state is consistent with the Page curve. Then, we numerically solve Eqs. (\[couple\_KPZ\]) and find the theory prediction in a nice agreement with the real data from random circuit simulations in Fig. \[fig:model\](b).
[74]{}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 [****, ()](https://doi.org/10.1137/S0097539795293172) [****, ()](https://doi.org/10.1103/PhysRevLett.92.027902) [****, ()](https://doi.org/10.1109/TIT.2002.802612) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevX.8.021013) [****, ()](https://doi.org/10.1103/PhysRevX.8.031057) [****, ()](https://doi.org/10.1103/PhysRevX.8.031058) @noop [****, ()]{} @noop [****, ()]{} @noop [**** ()]{} @noop [**** ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevB.95.134302) @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevLett.70.3339) in @noop [**]{}, Vol. () @noop [****, ]{} @noop [****, ]{} @noop [****, ]{} @noop [****, ]{} [ ](https://arxiv.org/abs/1710.03363) @noop [****, ]{} @noop [****, ()]{} @noop [****, ]{} @noop [****, ]{} @noop [****, ]{} [****, ](https://doi.org/10.1007/JHEP04(2013)022) [****, ](https://doi.org/10.1007/JHEP03(2015)051) @noop [ ()]{} @noop [ ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevLett.114.110506) @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevA.97.032329) @noop [****, ()]{} @noop [****, ()]{} [****, ()](https://doi.org/10.1103/PhysRevLett.124.150502) [****, ()](https://doi.org/10.1103/PhysRevApplied.13.034029) @noop [ ()]{} [****, ()](https://doi.org/10.1103/RevModPhys.84.621) @noop [ ()]{} @noop [ ()]{} [****, ()](https://doi.org/10.1088/1367-2630/ab7257) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} A. Barab[á]{}si, R. Albert, and H. Jeong, Mean-field theory for scale-free random networks, Physica A, [**272**]{}, 173 (1999) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ]{} @noop [ ()]{} Lov[á]{}sz, L[á]{}szl[ó]{}, Random walks on graphs: A survey, Combinatorics, Paul erdos is eighty, [**2**]{}, 1-46, Keszthely (Hungary) (1993) Alessio Lerose and Silvia Pappalardi, Bridging entanglement dynamics and chaos in semiclassical systems, arXiv:2005.03670 (2020) Gerardo Adesso and Samanta Piano, Theory of Genuine Tripartite Nonlocality of Gaussian States, Phys. Rev. Lett. [**112**]{}, 010401 (2014) @noop [ ()]{} S. Tserkis, J. Thompson, A. P. Lund, T. C. Ralph, P. K. Lam, M. Gu, S. M. Assad, Maximum entanglement of formation for a two-mode Gaussian state over passive operations, arXiv:2004.13948 (2020) @noop [ ]{} in @noop [**]{} (, ) p. [****, ()](https://doi.org/10.1103/PhysRevB.99.045155) @noop [****, ()]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We provide new results on the existence of nonzero positive weak solutions for a class of second order elliptic systems. Our approach relies on a combined use of iterative techniques and classical fixed point index. Some examples are presented to illustrate the theoretical results.'
address:
- 'Jos[é]{} [Á]{}ngel Cid, Departamento de Matemáticas, Universidade de Vigo, 32004, Pabellón 3, Campus de Ourense, Spain'
- 'Gennaro Infante, Dipartimento di Matematica e Informatica, Università della Calabria, 87036 Arcavacata di Rende, Cosenza, Italy'
author:
- José Ángel Cid
- Gennaro Infante
title: 'A non-variational approach to the existence of nonzero positive solutions for elliptic systems'
---
Introduction
============
In the recent papers [@lan1; @lan2] Lan, by means of classical fixed point index theory, studied under *sublinear* conditions the existence of nonzero positive solutions of systems of second order elliptic boundary value problems of the kind $$\label{sys-intro}
\left\{
\begin{array}{ll}
Lu_i(x)= f_i(x,u(x)), & x\in \Omega,\quad i=1,2,\ldots,n, \\
Bu_i(x)=0, & x\in \partial \Omega,\quad i=1,2,\ldots,n,
\end{array}
\right.$$ where $\Omega\subset {\mathbb{R}}^m$ is a suitable bounded domain, $L$ is a strongly uniformly elliptic operator, $u(x)=(u_1(x),u_2(x),\ldots,u_n(x))$, $f_i:\bar{\Omega}\times {\mathbb{R}}^m_+\to {\mathbb{R}}_+$ are continuous functions and $B$ is a first order boundary operator. A key assumption utilized by Lan is an asymptotic comparison between the nonlinearities and the principal eigenvalue of the corresponding linear system.
The formulation is rather general and covers systems of quasilinear elliptic equations of the type $$\label{lap-sys}
\left\{
\begin{array}{ll}
-\Delta u_1(x)=\lambda_1 f_1(x, u_1(x), u_2(x)), & x\in \Omega, \\
-\Delta u_2(x)=\lambda_2 f_2(x, u_1(x), u_2(x)), & x\in \Omega, \\
u_1(x)=u_2(x)=0, & x\in \partial \Omega,\end{array}\right.$$where $\lambda_i>0$. Elliptic systems of the kind have been widely investigated, by variational and topological methods, under a variety of growth conditions; we refer the reader to the reviews by de Figuereido [@defig] and Ruf [@bruf], the papers [@chzh; @chzh07; @cui3; @defigetal; @gimmrp; @ma2; @zhch] and references therein.
Within the topological framework, fixed point index plays a key role and, typically, the methodology is to compute the index on a small ball and on a large ball, which, in turn, involves controlling the behaviour of an associated operator on two boundaries, see for example [@alves-defig; @defigetal; @defig; @lan1; @lan-zhang; @zou]. In the context of ODEs Franco and co-authors [@df-gi-jp-prse] motivated by earlier works of Persson [@persson] and continuing the work by Cabada and Cid [@cabcid] and Cid et al. [@jc-df-fm], managed to replace the comparison between two boundaries with a comparison between a boundary and a point, at the cost of having some extra monotonicity requirement on the nonlinearities. These ideas were further developed, also in the case of ODEs, by Cabada and co-authors [@acjcgi1; @CCI2].
Here we apply a refinement of the results in [@CCI2] to the elliptic system $$\label{ellbvp-intro}
\left\{
\begin{array}{ll}
Lu_i(x)=\lambda_i f_i(x,u(x)), & x\in \Omega,\quad i=1,2,\ldots,n, \\
Bu_i(x)=0, & x\in \partial \Omega,\quad i=1,2,\ldots,n,
\end{array}
\right.$$ where $f_i\in C\bigl(\bar{\Omega}\times \prod_{j=1}^n[0,\rho_j]\bigr)$ and $\rho_i>0$ for $i=1,2,\ldots,n$. Our approach relies on the use of the monotone iterative method jointly with the fixed point index. We stress that Cheng and Zhang [@chzh] combined the method of sub-supersolutions with the Leray-Schauder degree, in the case of elliptic systems with Dirichlet BCs. Here we prove, under natural assumptions on the nonlinearities $f_i$, that nonzero positive weak solutions of the BVP exist for certain values of the parameters $\lambda_i$. We also present some numerical examples to illustrate our theory. As far as we know, our results are new and complement the results of [@chzh; @lan1].
Preliminaries
=============
A subset $P$ of a real Banach space $X$ is a [*cone*]{} if it is closed, $P+P\subset P$, $\lambda P\subset P$ for all $\lambda\ge 0$ and $P\cap(-P)=\{\theta\}$. A cone $P$ defines the partial ordering in $X$ given by $$x\le y \quad \mbox{if and only if $y-x\in P$}.$$ For $x,\ y \in X$, with $x\le y$, we define the ordered interval $$[x,y]=\{z\in X : x\le z\le y\}.$$
The cone $P$ is [*normal*]{} if there exists $d>0$ such that for all $x, y\in X$ with $\theta \le x\le y$ then $\|x\|\le d \|y\| $. The cone $P$ is [*total*]{} if $X=\overline{P-P}$. We remark that every cone $P$ in a normed space has the [*Archimedean property*]{}, that is, $n x\le y$ for all $n\in {\mathbb{N}}$ and some $y\in X$ implies $x\le \theta$. In the sequel, with abuse of notation, we will use the same symbol “$\ge$" for the different cones appearing in the paper since the context will avoid confusion.
We denote the closed ball of center $x_0\in X$ and radius $r>0$ as $$B[x_0,r]=\{x\in X : \|x-x_0\|\le r \},$$ and the intersection of the cone with the open ball centered at the origin and radius $r>0$ as $$P_r=P \cap \{x\in X : \|x\|<r\}.$$ In the sequel the closure and the boundary of subsets of $P$ are understood to be relative to $P$. An operator $T:D\subseteq X\to X$ is said to be monotone non-decreasing if for every $x,y\in D$ with $x\leq y$ we have $Tx\leq Ty$. We recall the following theorem, known as the monotone iterative method (see, for example, [@Amann-rev; @zeidler]).
\[mim\] Let $X$ be a real Banach space with normal order cone $P$. Suppose that there exist $\alpha\le \beta$ such that $T\colon [\alpha,\beta]\subset P\to P$ is a completely continuous monotone non-decreasing operator with $\alpha \le T \alpha $ and $T\beta \le \beta$. Then $T$ has a fixed point and the iterative sequence $\alpha_{n+1}= T \alpha_n$, with $\alpha_0=\alpha$, converges to the smallest fixed point of $T$ in $[\alpha,\beta]$, and the sequence $\beta_{n+1} = T \beta_n$, with $\beta_0=\beta$, converges to the greatest fixed point of $T$ in $[\alpha,\beta]$.
In the following Proposition we recall the main properties of the fixed point index of a completely continuous operator relative to a cone, for more details see [@Amann-rev; @guolak].
\[propindex\] Let $D$ be an open bounded set of $X$ with $0\in D_{P}$ and $\overline{D}_{P}\ne P$, where $D_{P}=D\cap P$. Assume that $T:\overline{D}_{P}\to P$ is a completely continuous operator such that $x\neq Tx$ for $x\in \partial D_{P}$. Then the fixed point index $i_{P}(T, D_{P})$ has the following properties:
- If there exists $e\in P\setminus \{0\}$ such that $x\neq Tx+\lambda e$ for all $x\in \partial D_P$ and all $\lambda>0$, then $i_{P}(T, D_{P})=0$.
- For example $(i)$ holds if $Tx\not\leq x$ for $x\in
\partial D_P$.
- If $\|Tx\|\ge \|x\|$ for $x\in
\partial D_P$, then $i_{P}(T, D_{P})=0$.
- If $Tx \neq \lambda x$ for all $x\in
\partial D_P$ and all $\lambda > 1$, then $i_{P}(T, D_{P})=1$.
- For example $(iii)$ holds if either $Tx\not\geq x$ for $x\in
\partial D_P$ or $\|Tx\|\le \|x\|$ for $x\in
\partial D_P$.
- Let $D^{1}$ be open bounded in $X$ such that $\overline{D^{1}_{P}}\subset D_P$. If $i_{P}(T, D_{P})=1$ and $i_{P}(T,
D_{P}^{1})=0$, then $T$ has a fixed point in $D_{P}\setminus
\overline{D_{P}^{1}}$. The same holds if $i_{P}(T, D_{P})=0$ and $i_{P}(T, D_{P}^{1})=1$.
The following result, ensuring the existence of a non-trivial fixed point, is a modification of [@CCI2 Theorem 2.3]. The short proof is presented here for completeness.
\[thCCI\] Let $X$ be a real Banach space, $P$ a normal cone with normal constant $d\ge 1$ and nonempty interior (i.e. solid) and for $\rho>0$ let $T\colon
\overline{P_{\rho}}\to P$ be a completely continuous operator. Assume that
- there exist $\beta \in \overline{P_{\rho/d}}$, with $T\beta\le \beta$, and $0<{R}<\rho$ such that $B[\beta,{R}]\subset P$,
- the map $T$ is non-decreasing in the set $$\tilde{\mathcal{P}}=\Bigl\{x\in P : x\le \beta \quad \mbox{and} \quad
\frac{{R}}{d}\le \|x\|
\Bigr\},$$
- there exists a (relatively) open bounded set $V\subset P_{\rho}$ such that $\theta\in V$, $i_P(T,V)=0$ and either $\overline{P_{R}}\subset V$ or $\overline{V}\subset P_{R}$.
Then the map $T$ has at least one non-zero fixed point $x_1$ in $P$, that $$\mbox{either belongs to $\tilde{\mathcal{P}}$ or belongs to}
\left\{ \begin{array}{ll} V\setminus \overline{P_{R}}, & \mbox{in case $\overline{P_{R}}\subset V$,} \\ P_{R}\setminus \overline{V}, & \mbox{in case $\overline{V}\subset P_{R}$.}\end{array}\right.$$
Since $B[\beta,{R}]\subset P$ we have that if $x\in
P$ with $\|x\|={R}$ then $x\leq \beta$.
Suppose first that there exists $\alpha\in P_{\rho}$ with $\|\alpha
\|={R}$ and $T\alpha \geq \alpha$. Then $\alpha\le \beta$ and due to the normality of the cone $P$ we have that $$\alpha \le x \implies R=\| \alpha \| \le d \|x\|.$$ So $[\alpha,
\beta]\subset \mathcal{\tilde P}$ and from (2) it follows that $T$ is non-decreasing on $[\alpha, \beta]$ . Then we can apply Theorem \[mim\] to ensure the existence of a fixed point of $T$ on $[\alpha, \beta]$, which, in particular, is a non-trivial fixed point.
Now suppose that such $\alpha$ does not exist and that $T$ is fixed point free on $\partial P_R$ (if not, we would have the desired non-trivial fixed point). Thus $Tx \ngeq x$ for all $x\in P_{\rho}$ with $\|x\|={R}$, which by Proposition \[propindex\], $(iii)$ implies that $i_P(T,P_{R})=1$. Since, by assumption, $i_P(T,V)=0$ we get the existence of a fixed point $x_1$ belonging to the set $V\setminus \overline{P_{R}}$ (when $\overline{P_{R}}\subset V$) or to the $P_{R}\setminus \overline{V}$ (when $\overline{V}\subset P_{R}$). In both cases $x_1$ would be a non-trivial fixed point of $T$.
In Theorem \[thCCI\], (3) with the condition $i_P(T,V)=0$, due to the construction of the fixed point index, we implicitly assume that $T$ is fixed point free on $\partial V$. On the other hand, notice that Condition (i) in Proposition \[propindex\] provides a rather general condition implying index zero and, typically in applications, one would try $e\equiv 1$ or $e$ a positive eigenfunction related to a suitable compact linear operator, see [@Amann-rev]. We develop further this idea in the following section.
Sufficient conditions for index zero
====================================
In this Section we deal with an abstract Hammerstein fixed point equation in a product space. For $i=1,2,\ldots,n$ let $X_i$ a Banach space ordered by a normal cone $P_i$ and consider the product Banach space $X=\prod_{i=1}^n X_i$ with the maximum norm and ordered by the cone $P=\prod_{i=1}^n P_i$. Let us consider the operators $F_i:D\subset X\to X_i$ and $K_i:X_i\to X_i$ and the fixed point equation $$\label{eqfixpoint}
u=Tu:=(K_1 F_1 u,K_2 F_2 u,\ldots, K_n F_n u) \quad \mbox{for $u\in D\subset X$.}$$
By exploiting the structure of the operator $T$ in the equation we are able to prove the following result which generalizes [@lan2 Lemma 2.2] in two ways: firstly, the inequality in our condition (H1) is asked only for one component $F_i$ and, secondly, we do not need to add a positive $\varepsilon$ to its right-hand side. Thus, the applicability of our result is wider and it includes for instance [@jw-tmna Theorem 4.3] which is not covered by [@lan2 Lemma 2.2].
\[thindex0-2\] Assume the following conditions:
- For each $i=1,2,\ldots,n$ the operator $K_i$ is linear, compact, $K_i(P_i)\subset P_i$ and there exist $ \alpha_i>0$ and $e_i\in P_i\setminus \{0\}$ such that $K_i e_i\ge \alpha_i e_i$.
- There exists $\rho_0>0$ such that $\bar{P}_{\rho_0}\subset D$, for each $i=1,2,\ldots,n$ the operator $F_i$ is continuous and bounded on $\bar{P}_{\rho_0}$, $F_i(\bar{P}_{\rho_0})\subset P_i$, and there exist $i_0\in \{1,2,\ldots,n\}$ such that $$F_{i_0}(u)\ge \frac{1}{\alpha_{i_0}} u_{i_0} \quad \mbox{for $u\in \partial P_{\rho_0}$.}$$
If $u \not=Tu$ for $u\in \partial P_{\rho_0}$ then $i_P(T, P_{\rho_0})=0$.
Defining $e=(e_1,e_2,\ldots,e_n)$ by Proposition \[propindex\] it is enough to prove that $$\label{eqindex0-1} u\neq Tu+\mu e \quad \mbox{for all $u\in \partial P_{\rho_0}$ and all
$\mu >0$.}$$ If not, there exists $u\in \partial P_{\rho_0}$ and $\mu >0$ such that $$\label{eqindex0-2} u= Tu+\mu e.$$ Then $u\ge \mu e$ and in particular $u_{i_0}\ge \mu e_{i_0}$. Now by (H0) and (H1) we have $$\begin{gathered}
u_{i_0}=K_{i_0} F_{i_0} u+\mu e_{i_0}\ge K_{i_0} \left(\frac{1}{\alpha_{i_0}} u_{i_0}\right) +\mu e_{i_0}\ge K_{i_0}\left(\frac{1}{\alpha_{i_0}} (\mu e_{i_0})\right) +\mu e_{i_0}\\
=\frac{1}{\alpha_{i_0}} \mu K_{i_0}(e_{i_0}) +\mu e_{i_0}\ge \frac{1}{\alpha_{i_0}} \mu \alpha_{i_0} e_{i_0} +\mu e_{i_0}=2(\mu e_{i_0}).\end{gathered}$$ Then $u_{i_0}\ge 2 (\mu e_{i_0})$ and by induction it can be proven that actually $u_{i_0}\ge n (\mu e_{i_0})$ for all $n\in {\mathbb{N}}$. By the Archimedian property we obtain $\mu e_{i_0} \le 0$, a contradiction.
\[remlbsr\] Notice that if (H0) holds then $\alpha_{i_0}$ is a lower bound for the spectral radius $r(K_{i_0})$, that is $r(K_{i_0})\ge \alpha_{i_0}>0$ (see [@kras Theorem 5.4] or [@jw-tmna Theorem 2.7]). So the best value for $\alpha_{i_0}$ in order to make (H1) as weak as possible would be $r(K_{i_0})$, and in fact we can choose $\alpha_{i_0}=r(K_{i_0})$ in (H1) if the assumptions of the celebrated Krein-Rutman theorem hold.
\[corindex0-3\] Suppose that (H0) holds and also
- There exists $\rho_0>0$ such that $\bar{P}_{\rho_0}\subset D$, for each $i=1,2,\ldots,n$ the operator $F_i$ is continuous, bounded on $\bar{P}_{\rho_0}$, $F_i(\bar{P}_{\rho_0})\subset P_i$, and there exist $i_0\in \{1,2,\ldots,n\}$ such that $P_{i_0}$ is a total cone and $$F_{i_0}(u)\ge \frac{1}{r(K_{i_0})} u_{i_0} \quad \mbox{for $u\in \partial P_{\rho_0}$.}$$
If $u \not=Tu$ for $u\in \partial P_{\rho_0}$ then $i_P(T, P_{\rho_0})=0$.
As explained in Remark \[remlbsr\] we have $r(K_{i_0})\ge \alpha_{i_0}>0$ so the Krein-Rutman theorem applies (see [@zeidler Theorem 7.26]) and then $r(K_{i_0})$ is an eigenvalue with an eigenvector $\varphi_{i_0} \in P_{i_0}\setminus \{0\}$, that is $K_{i_0}\varphi_{i_0}=r(K_{i_0})\varphi_{i_0}$. So (H0) also holds taking $\alpha_{i_0}=r(K_{i_0})$ which together with (H1’) give us the desired conclusion applying Lemma \[thindex0-2\].
A fixed point formulation for elliptic systems
==============================================
The assumptions
---------------
We shall deal with the following system of second order elliptic BVPs $$\label{eqellipticsystem}
\left\{
\begin{array}{ll}
Lu_i(x)=\lambda_i f_i(x,u(x)), & x\in \Omega,\quad i=1,2,\ldots,n, \\
Bu_i(x)=0, & x\in \partial \Omega,\quad i=1,2,\ldots,n,
\end{array}
\right.$$ where $\Omega\subset {\mathbb{R}}^m$, $m\ge 2$, is a suitable bounded domain, $L$ is a strongly uniformly elliptic operator, $u(x)=(u_1(x),u_2(x),\ldots,u_n(x))$, $\lambda_i>0$, $f_i\in C\bigl(\bar{\Omega}\times \prod_{j=1}^n[0,\rho_j]\bigr)$ and $\rho_i>0$ for $i=1,2,\ldots,n$ and $B$ is a first order boundary operator. More precisely, we shall assume the conditions in [@Amann-rev Section 4 of Chapter 1] (see also [@lan1; @lan2]), namely:
1. We suppose that $\Omega\subset {\mathbb{R}}^m$, $m\ge 2$, is a bounded domain such that its boundary $\partial \Omega$ is an $(m-1)-$dimensional $C^{2,\mu}-$manifold for some $\mu\in (0,1)$, such that $\Omega$ lies locally on one side of $\partial \Omega$ (see [@zeidler Section 6.2] for more details).
2. $L$ is the second order elliptic operator given by $$\label{eqL} Lz(x)=-\sum_{i,j=1}^m a_{ij}(x)\frac{\partial^2 z}{\partial x_i \partial x_j}(x)+\sum_{i=1}^m a_{i}(x) \frac{\partial z}{\partial x_i} (x)+a(x)z(x), \quad \mbox{for $x\in \Omega$,}$$ where $a_{ij},a_{i},a\in C^{\mu}(\overline{\Omega})$ for $i,j=1,2,\ldots,m$, $a(x)\ge 0$ on $\bar{\Omega}$, $a_{ij}(x)=a_{ji}(x)$ on $\bar{\Omega}$ for $i,j=1,2,\ldots,m$. Moreover $L$ is strongly uniformly elliptic, that is, there exists $\mu_0>0$ such that $$\sum_{i,j=1}^m a_{ij}(x)\xi_i \xi_j\ge \mu_0 \|\xi\|^2 \quad \mbox{for $x\in \Omega$ and $\xi=(\xi_1,\xi_2,\ldots,\xi_m)\in{\mathbb{R}}^m$.}$$
3. $B$ is the boundary operator given by $$Bz(x)=b(x)z(x)+\delta \frac{\partial z}{\partial v}(x) \quad \mbox{for $x\in\partial \Omega$},$$ where $v$ is an outward pointing and nowhere tangent vector field on $\partial \Omega$ of class $C^{1,\mu}$ (not necessarily a unit vector field), $\frac{\partial z}{\partial v}$ is the directional derivative of $z$ with respect to $v$, $b:\partial \Omega \to {\mathbb{R}}$ is of class $C^{1,\mu}$ and moreover one of the following conditions holds:
1. $\delta =0$ and $b(x)\equiv 1$ (Dirichlet boundary operator).
2. $\delta =1$, $b(x)\equiv 0$ and $a(x)\not\equiv 0$ (Neumann boundary operator).
3. $\delta =1$, $b(x)\ge 0$ and $b(x)\not\equiv 0$ (Regular oblique derivative boundary operator).
The solution operator
---------------------
It is well-known, since the pioneering work by Schauder, that under the previous conditions for each $g\in C^{\mu}(\bar{\Omega})$, the boundary value problem $$\label{eqelliptic}
\left\{
\begin{array}{ll}
Lz(x)=g(x), & x\in \Omega, \\
Bz(x)=0, & x\in \partial \Omega,
\end{array}
\right.$$ has a unique classical solution $z\in C^{2,\mu}(\bar{\Omega})$ and the solution operator $K:C^{\mu}(\bar{\Omega})\to C^{2,\mu}(\bar{\Omega})$ defined as $Kg=z$ is linear, continuous and positive (due to the maximum principle). It is well known that if $g\in C(\bar{\Omega})$ and $m\ge 2$ then problem need not have a classical solution $z\in C^{2}(\bar{\Omega})$ (see a counterexample in [@zeidler Problem 6.3]). For our purposes it will be fundamental to consider a generalized solution operator extending $K$ to $C(\bar{\Omega})$.
\[proK1\] The operator $K$ can be extended uniquely to a continuous, linear and compact operator $K:C(\bar{\Omega})\to C(\bar{\Omega})$ (that we will denote again by the same name).
Consider the cone of positive functions $P=C(\bar{\Omega},{\mathbb{R}}_+)$. The generalized solution operator $K$ is not only positive (i.e. $K(P)\subset P$), but also $e$-positive, as we stress in the following result, see [@amann-JFA Lemma 5.3].
\[proK2\] Let $e=K1\in C(\bar{\Omega},{\mathbb{R}}_+)\setminus \{0\}$. Then $K:C(\bar{\Omega})\to C(\bar{\Omega})$ is $e$-positive (and in particular positive), that is for each $g\in C(\bar{\Omega},{\mathbb{R}}_+)\setminus \{0\}$ there exist $\alpha_g>0$ and $\beta_g>0$ such that $\alpha_g e \le K g \le \beta_g e$. In particular, there exists $\alpha_e>0$ such that $Ke\ge \alpha_e$e.
The Nemytskii operator
----------------------
If $I=\prod_{j=1}^n I_j\subset {\mathbb{R}}^n$, where $I_j\subset {\mathbb{R}}$ is a closed interval with nonempty interior for $j=1,2,\ldots,n$, and $D\subset C(\bar{\Omega},{\mathbb{R}}^n)$ is a nonempty set we define $$D_I=\{u\in D: u(x)\in I \quad \mbox{for all $x\in \bar{\Omega}$}\}.$$
We utilize the space $C(\bar{\Omega},{\mathbb{R}}^n)$, endowed with the norm $\|u\|:=\displaystyle\max_{i=1,2,\ldots,n} \{\|u_i\|_{\infty}\}$, where $\|z\|_{\infty}=\displaystyle\max_{x\in \bar{\Omega}}|z(x)|$, and consider (with abuse of notation) the cone $P=C(\bar{\Omega},{\mathbb{R}}^n_+)$.
For a function $f:\bar{\Omega}\times I\to{\mathbb{R}}$ the Nemytskii (or superposition) operator $F$ is defined as $Fu(x):=f(x,u(x))$ for $u\in C(\bar{\Omega},{\mathbb{R}}^n)_I$ and $x\in \bar{\Omega}$. We recall some useful properties of the Nemytskii operator, see for example [@Amann-rev]; here by *non-decreasing* we mean non-decreasing with respect to the components.
\[proF\] If $f\in C(\bar{\Omega}\times I)$ then the following claims hold:
1. The Nemytskii operator $F:C(\bar{\Omega},{\mathbb{R}}^n)_I \to C(\bar{\Omega})$ is continuous and bounded.
2. If $f(x,u)\ge 0$ for all $(x,u)\in \bar{\Omega}\times I$ with $u\in{\mathbb{R}}^n_+$ then $F$ is positive, that is, $F(P_I)\subset P$.
3. If for every $x\in \bar{\Omega}$ the function $f(x,\cdot):I\to {\mathbb{R}}$ is non-decreasing then $F$ is non-decreasing.
4. If $f\in C^{\mu}(\bar{\Omega}\times I)$ and $u\in C^1(\bar{\Omega},{\mathbb{R}}^n)_I$ then $Fu\in C^{\mu}(\bar{\Omega})$.
The fixed point formulation and weak solutions
----------------------------------------------
Now we can use the previous results in order to formulate the elliptic system as a fixed point problem in the space of continuous functions.
Let us define $I=\prod_{i=1}^n [0,\rho_i]$ and the operator $T:C(\bar{\Omega},{\mathbb{R}}^n)_I \to C(\bar{\Omega},{\mathbb{R}}^n)$ as $$\label{opT}
T(u):=(\lambda_1 K F_1(u),\lambda_2 K F_2(u),\ldots,\lambda_n K F_n(u)),$$ where $F_i$ is the Nemytskii operator given by $F_i (u)(x)=f_i(x,u(x))$ for all $x\in \bar{\Omega}$.
We shall say that $u\in C(\bar{\Omega},{\mathbb{R}}^n)_I $ is a [*weak solution*]{} of the problem if and only if $u$ is a fixed point of operator $T$, that is, $u=Tu$.
Note that if $f$ is sufficiently smooth then, by regularity theory, a weak solution of is also a classical one.
\[proT\] Suppose that $f_i\in C(\bar{\Omega}\times I)$ for all $i=1,2,\ldots,n$. Then the following claims hold:
1. $T:C(\bar{\Omega},{\mathbb{R}}^n)_I \to C(\bar{\Omega},{\mathbb{R}}^n)$ is a completely continuous operator.
2. If moreover $f_i \ge 0$ for all $i=1,2,\ldots,n$, then $T(P_I)\subset P$.
3. If $f_i\in C^{\mu}(\bar{\Omega}\times I)$ for all $i=1,2,\ldots,n$, then $u\in C^{2,\mu}(\bar{\Omega},{\mathbb{R}}^n)_I$ is a solution of if and only if $u\in C(\bar{\Omega},{\mathbb{R}}^n)_I$ and $u=Tu$.
Existence of positive solutions for elliptic systems
====================================================
The following is our main result. Remember that by Proposition \[proK2\] and Remark \[remlbsr\] we have that $r(K)>0$, where $r(K)$ is the spectral radius of $K$. We denote by $\mu_1=1/r(K)$.
\[thmain\] Assume that $f_i\in C(\bar{\Omega}\times \prod_{j=1}^n [0,\rho] )$ for all $i=1,2,\ldots,n$ and moreover:
- For each $x\in \bar{\Omega}$ and for every $i=1,2,\ldots,n$, $f_i(x,\cdot):\prod_{j=1}^n [0,\rho] \to [0, +\infty)$ is non-decreasing.
- There exist $\delta \in (0,+\infty)$, $i_0\in \{1,2,\ldots,n\}$ and $\rho_0 \in (0,\rho)$ such that $f_{i_0}(x,u)\ge \delta u_{i_0}$ for all $x\in \bar{\Omega}$ and $u\in\prod_{j=1}^n [0,\rho_0]$.
- There exist $(\beta_1, \beta_2,\ldots ,\beta_n)\in \prod_{j=1}^n (0,\rho]$ such that $m_j(\beta_1, \beta_2,\ldots ,\beta_n)>0$ for all $j=1,2,\ldots,n$, $j\not=i_0$, where $
m_j(\beta_1, \beta_2,\ldots ,\beta_n)= \displaystyle\max_{x\in \bar{\Omega}}f_j(x,\beta_1, \beta_2,\ldots ,\beta_n)$ for all $j=1,2,\ldots,n$.
Then the system has a nonzero weak positive solution $u$ such that $0<\|u\|\leq \rho$ provided that $$\label{eqlambda-1}
0< \lambda_j \leq
\frac{\beta_j}{m_j(\beta_1, \beta_2,\ldots ,\beta_n) \| K(1)\|_{\infty}} \quad \mbox{for all $j=1,2,\ldots,n$, $j\not=i_0$,}$$ and $$\label{eqlambda-2}
\frac{\mu_1}{\delta}< \lambda_{i_0} \leq
\frac{\beta_{i_0}}{m_{i_0}(\beta_1, \beta_2,\ldots ,\beta_n) \| K(1)\|_{\infty}}.$$
We consider the cone $P=C(\bar{\Omega},{\mathbb{R}}^n_+)$, which is normal with constant $d=1$, in the space $C(\bar{\Omega},{\mathbb{R}}^n)$ with the maximum norm. $T$ maps $\overline{P_\rho} \rightarrow P$ and is completely continuous and nondecreasing by (a).
Note that by (c) we have $\beta=(\beta_1, \beta_2,\ldots ,\beta_n)\in \overline{P_\rho}$, $\beta$ lies in the interior of $P$ and furthermore, for every $i=1,2,\ldots,n$, $$\begin{gathered}
\lambda_i KF_i(\beta_1,\beta_2,\ldots ,\beta_n)(x) \leq \| \lambda_i KF_i(\beta_1,\beta_2,\ldots ,\beta_n)\|_{\infty}\\
\leq \|\lambda_i K(m_i(\beta_1, \beta_2,\ldots ,\beta_n))\|_{\infty} \leq \lambda_i m_i(\beta_1, \beta_2,\ldots ,\beta_n) \| K(1)\|_{\infty}\leq \beta_i.\end{gathered}$$ Finally, from Corollary \[corindex0-3\] and (b) we obtain that either $T$ has a fixed point in $\partial P_{\rho_0}$ or either $i_P(T, P_{\rho_0})=0$ for $\lambda_1>\mu_1/\delta$. Thus, in the second case we achieve also a nonzero fixed point of $T$ by Theorem \[thCCI\].
\[rem-reg\] Notice that by Proposition \[proT\], part (3), if $f_i\in C^{\mu}(\bar{\Omega}\times I)$ for all $i=1,2,\ldots,n$, then any weak solution belongs to $C^{2,\mu}(\bar{\Omega},{\mathbb{R}}^n)_I$ and it is also a classical solution of .
As an application, we present an existence result for a system with Dirichlet boundary conditions and the Laplacian operator.
Take $\Omega=\{(x_1,x_2)\in \mathbb{R}^2 : x_1^2+x_2^2<1\}$ and consider the system $$\label{example}
\left\{
\begin{array}{ll}
-\Delta u_1=\lambda_1 (|(u_1,u_2)|^\frac{1}{2} + \tan |(u_1,u_2)| ),& \text{in }\Omega , \\
-\Delta u_2=\lambda_2 |(u_1,u_2)|^2, & \text{in }\Omega , \\
u_1,u_2=0, & \text{on }\partial \Omega ,\end{array}\right.$$where $|(u_1,u_2)|:=\max \{ |u_1|,|u_2|\}$.
If we define $\rho=\frac{15 }{64}\pi$ then the system has a nontrivial positive solution $(u_1,u_2)\in C^{2,\frac1 2}(\bar{\Omega},{\mathbb{R}}^2_+)$ such that $0<\|(u_1,u_2)\|\le \frac{15 }{64}\pi$ for every $$0< \lambda_1 \leq
\frac{4\rho}{m_1(\rho, \rho) } \approx 1.669,\ 0< \lambda_2 \leq
\frac{4\rho}{m_2(\rho, \rho) } \approx 5.432.$$
Let $f_1,f_2:\bar{\Omega}\times[0,\rho]\times [0,\rho] \to {\mathbb{R}}$ be given by $$f_1(x_1,x_2,u_1,u_2)=|(u_1,u_2)|^\frac{1}{2} + \tan |(u_1,u_2)| \quad \mbox{and} \quad f_2(x_1,x_2,u_1,u_2)= |(u_1,u_2)|^2.$$ Note that $f_1,f_2\in C^{1/2}(\bar{\Omega}\times [0,\rho]\times [0,\rho])$, are non-decreasing and, given $\delta>0$, $f_1$ satisfies condition (b) in Theorem \[thmain\] for $\rho_0$ sufficiently small, due to the behaviour near the origin. By direct calculation, $K(1)=\frac{1}{4}(1-x_1^2-x_2^2)$ so $\|K(1)\|=\frac{1}{4}$ and fixing $\beta_1=\beta_2=\rho=\frac{15 }{64}\pi$ the result follows from Theorem \[thmain\] and Remark \[rem-reg\].
In the case of a single equation $$\label{eqlaplacian}
\left\{
\begin{array}{ll}
L z=\lambda f(x,z), & x\in \Omega, \\
Bz=0, & x\in \partial \Omega,
\end{array}
\right.$$ where $\lambda>0$ and $f:\bar{\Omega}\times [0,\rho]\to {\mathbb{R}}$, a more precise result than Theorem \[thmain\] can be obtained.
\[thm-single\] Assume that $f\in C(\bar{\Omega}\times [0,\rho])$ and moreover:
- For each $x\in \bar{\Omega}$ the function $f(x,\cdot):[0,\rho] \to [0, +\infty)$ is non-decreasing.
- There exist $\delta \in (0,+\infty)$ and $\rho_0 \in (0,\rho)$ such that $f(x,z)\ge \delta z$ for all $x\in \bar{\Omega}$ and $z\in [0,\rho_0]$.
Then the BVP admits a non-zero weak positive solution $z$ such that $0<\|z\|_{\infty}<\rho$ provided that $$\label{eqlambda}
\frac{\mu_1}{\delta}\leq \lambda < \sup_{0<s\le \rho} \frac{s}{M(s)\|K(1)\|_{\infty}},$$ where $M(s)= \displaystyle\max_{x\in \bar{\Omega}}
f(x,s)>0$ for all $s\in (0,\rho]$.
By our assumptions, $T:P_{\rho}\to P$ is completely continuous and non-decreasing. Now, let $0<\beta\le \rho$ such that $\lambda < \displaystyle\frac{\beta}{M(\beta) \|K(1)\|_{\infty}}$. Then we have $$T\beta = \|T\beta \|_{\infty}=\| K(\lambda F \beta)\|_{\infty} \leq\| \lambda M(\beta) K(1) \|_{\infty}= \lambda M(\beta) \|K(1)\|_{\infty}<\beta.$$
Finally, if we define $V=P_{\rho_0}$ then by Corollary \[corindex0-3\] we have that either $T$ has a fixed point in $\partial V$ (which will be a non-zero positive solution) or either $i_P(T,V)=0$. In this last case Theorem \[thCCI\] applies and we obtain the existence of a non-zero weak positive solution.
\[exlaplacian\] The problem $$\label{eqexample}
\left\{
\begin{array}{ll}
-\Delta z=\lambda (\sqrt{z}+\tan(z)), & x\in \Omega=\{(x_1,x_2)\in {\mathbb{R}}^2 : x_1^2+x_2^2<1\}, \\
z(x)=0, & x\in \partial \Omega,
\end{array}
\right.$$ has a non-zero positive solution $z$ such that $0<\|z\|_{\infty}<\frac{\pi}{2}$ provided that $$\label{ eqlambdaexample}
0 <\lambda < \sup_{ 0<s<\frac \pi 2} \frac{4 s}{\sqrt{s}+\tan(s)} \approx 1.66924.$$
Firstly, take into account that $K(1)(x_1,x_2)=\displaystyle \frac{1}{4}(1-x_1^2-x_2^2)$ for $\Omega=\{(x_1,x_2)\in {\mathbb{R}}^2 : x_1^2+x_2^2<1\}$ so $\|K(1)\|=\frac{1}{4}$. Therefore, if $\lambda>0$ satisfies we can choose $\delta>0$ and $0<\rho<\frac{\pi}{2}$ such that is also satisfied. Since $f(x_1,x_2,z)=\sqrt{z}+\tan(z)$ is non-decreasing in $[0,\rho]$ and the growth condition at $z=0$, we have that conditions i) and ii), for every $\delta>0$, in Theorem \[thm-single\] are satisfied, and then the existence of a non-zero weak positive solution follows taking $\beta=\rho$. Since $f\in C^{\frac 1 2}(\bar{\Omega}\times [0,\rho])$ the solution is also a classical one (see Remark \[rem-reg\]).
Since $f(0)=0$ then clearly $u\equiv 0$ is a solution of . Example \[exlaplacian\] shows a [*non-trivial*]{} positive solution that cannot be obtained from [@Amann-rev Theorem 9.4] and neither seems to be covered by the results in [@lions Section 1].
Acknowledgements {#acknowledgements .unnumbered}
================
The authors wish to thank Dr. Mateusz Maciejewski for the useful comments and the anonymous Referee for the constructive remarks. J. A. Cid was supported by Xunta de Galicia (Spain), project EM2014/032. G. Infante was partially supported by G.N.A.M.P.A. - INdAM (Italy).
[xxx]{}
C. O. Alves and D. G. de Figueiredo, Nonvariational elliptic systems, *Discrete Contin. Dyn. Syst.*, **8** (2002), 289–302.
H. Amann, On the number of solutions of nonlinear equations in ordered Banach spaces, [*J. Functional Analysis*]{}, [**11**]{} (1972), 346–384.
H. Amann, Fixed point equations and nonlinear eigenvalue problems in ordered Banach spaces, *SIAM. Rev.*, **18** (1976), 620–709.
A. Cabada and J. A. Cid, Existence of a non-zero fixed point for nondecreasing operators via Krasnoselskii’s fixed point theorem, *Nonlinear Anal.*, **71** (2009), 2114–2118.
A. Cabada, J. A. Cid and G. Infante, New criteria for the existence of non-trivial fixed points in cones, *Fixed Point Theory Appl.*, [**2013:125**]{} (2013).
J. A. Cid, D. Franco and F. Minhós, Positive fixed points and fourth-order equations, *Bull. Lond. Math. Soc.*, **41** (2009), 72–78.
A. Cabada, J. A. Cid and G. Infante, A positive fixed point theorem with applications to systems of Hammerstein integral equations, *Bound. Val. Probl.*, **2014:254**, (2014), 10 pp.
X. Cheng and Z. Zhang, Positive solutions for a class of multi-parameter elliptic systems, *Nonlinear Anal. Real World Appl.*, **14** (2013), 1551–1562.
X. Cheng and C. Zhong, Existence of three nontrivial solutions for an elliptic system, *J. Math. Anal. Appl.*, **327** (2007), 1420–1430.
R. Cui, P. Li, J. Shi and Y. Wang, Existence, uniqueness and stability of positive solutions for a class of semilinear elliptic systems, *Topol. Methods Nonlinear Anal.*, **42** (2013), 91–104.
D. G. de Figueiredo, J. M. do [Ó]{} and B. Ruf, Non-variational elliptic systems in dimension two: a priori bounds and existence of positive solutions, *J. Fixed Point Theory Appl.*, **4** (2008), 77–96.
D. G. de Figueiredo, *Semilinear elliptic systems: existence, multiplicity, symmetry of solutions*. In: Handbook of Differential Equations, Vol. 5: Stationary Partial Differential Equations, M. Chipot (ed.), Elsevier, 2008, 1–48.
D. Franco, G. Infante and J. Perán, A new criterion for the existence of multiple solutions in cones, *Proc. Roy. Soc. Edinburgh Sect. A*, [**142**]{} (2012), 1043–1050.
D. Guo and V. Lakshmikantham, *Nonlinear problems in abstract cones*, Academic Press, Boston, 1988.
G. Infante, M. Maciejewski and R. Precup, A topological approach to the existence and multiplicity of positive solutions of $(p,q)$-Laplacian systems, *Dyn. Partial Differ. Equ.*, **12** (2015), 193–215.
M. A. Krasnosel’skii, G. M. Vainikko, P. P. Zabreiko, Ya. B. Rutitskii and V. Ya. Stetsenko, *Approximate solution of operator equations*, Wolters-Noordhoff Publishing, Groningen, 1972.
K. Q. Lan, Nonzero positive solutions of systems of elliptic boundary value problems, [*Proc. Amer. Math. Soc.*]{}, **139** (2011), 4343–4349.
K. Q. Lan, Existence of nonzero positive solutions of systems of second order elliptic boundary value problems *J. Appl. Anal. Comput.*, **1** (2011), 21–31.
K. Q. Lan and Z. Zhang, Nonzero positive weak solutions of systems of $p$-Laplace equations, *J. Math. Anal. Appl.*, **394** (2012), 581–591.
P. L. Lions, On the existence of positive solutions of semilinear elliptic equations, *SIAM. Rev.*, **24** (1982), 441–467.
R. Ma, R. Chen and Y. Lu, Positive solutions for a class of sublinear elliptic systems, *Bound. Value Probl.*, **2014:28**, (2014), 15 pp.
H. Persson, A fixed point theorem for monotone functions, *Appl. Math. Lett.*, **19** (2006), 1207–1209.
B. Ruf, *Superlinear Elliptic Equations and Systems*. In: Handbook of Differential Equations, Vol. 5: Stationary Partial Differential Equations, M. Chipot (ed.), Elsevier, 2008, 211–276.
J. R. L. Webb, A class of positive linear operators and applications to nonlinear boundary value problems, *Topol. Methods Nonlinear Anal.*, **39** (2012), 221–242.
E. Zeidler, [*Nonlinear functional analysis and its applications. I. Fixed-point theorems*]{}, Springer-Verlag, New York (1986).
Z. Zhang and X. Cheng, Existence of positive solutions for a semilinear elliptic system, *Topol. Methods Nonlinear Anal.*, **37** (2011), 103–116.
H. Zou, A priori estimates for a semilinear elliptic system without variational structure and their applications, *Math. Ann.*, **323** (2002) 713–735.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Sebastian Höppner, Bernhard Vogginger, Yexin Yan, Andreas Dixius, Stefan Scholze, Johannes Partzsch, Felix Neumärker, Stephan Hartmann, Stefan Schiefer, Georg Ellguth, Love Cederstroem, Luis Plana, Jim Garside, Steve Furber, Christian Mayr [^1] [^2] [^3]'
title: 'Dynamic Power Management for Neuromorphic Many-Core Systems'
---
[^1]: Manuscript received . The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7) under grant agreement no 604102 and the EUs Horizon 2020 research and innovation programme under grant agreements No 720270 and 785907 (Human Brain Project, HBP) and the Center for Advancing Electronics Dresden (cfaed). The authors thank ARM and Synopsis for IP and the Vodafone Chair at Technische Universität Dresden for contributions to RTL design.
[^2]: S. Höppner, B. Vogginger, Y. Yan, A. Dixius, S. Scholze, J. Partzsch, F. Neumärker, S. Hartmann, S. Schiefer, G. Ellguth, L. Cederstroem and C. Mayr are with the Faculty of Electrical and Computer Engineering, Technische Universität Dresden, Germany (e-mail: [email protected])
[^3]: L. Plana, J. Garside and S. Furber are with the Advanced Processor Technologies Research Group, University of Manchester
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We report on our analysis of the 1 Ms Chandra observation of the supernova remnant Cas A in order to localize, characterize and quantify its non-thermal X-ray emission. More specifically, we investigated whether the X-ray synchrotron emission from the inside of the remnant is from the outward shock, but projected toward the inner ring, or from the inner shell. We tackle this problem by employing a Lucy-Richardson deconvolution technique and measuring spectral indices in the 4.2-6 keV band.
We show that most of the continuum emission is coming from an inner ring that coincides with the location of the reverse shock. This inner ring includes filaments, whose X-ray emission has been found to be dominated by X-ray synchrotron emission. The X-ray emission from these filaments, both at the forward shock and from the inner ring, have relatively hard spectra with spectral index $> -3.1$. The regions emitting hard X-ray continuum contribute about 54% of the total X-ray emission in the 4.2-6 keV. This is lower than suggested by extrapolating the hard X-ray spectrum as measured by BeppoSAX-PDS and INTEGRAL. This can be reconciled by assuming a gradual steepening of the spectrum toward higher energies. We argue that the X-ray synchrotron emission is mainly coming from the Western part of the reverse shock. The reverse shock in the West is almost at rest in our observation frame, corresponding to a relatively high reverse shock velocity of $\sim 6000$ [[kms$^{-1}~$]{}]{} in the frame of the freely expanding ejecta.
author:
- 'E. A. Helder and J. Vink'
title: 'Characterizing the non-thermal emission of Cas A '
---
Introduction
============
Supernova remnants (SNRs) are the main candidates for producing Galactic cosmic rays, with energies at least up to the so-called knee of the cosmic ray spectrum at $\sim 3 \times 10^{15}$ eV. The first direct evidence for this is the detection of X-ray synchrotron emission caused by $\sim 10^{14}$ eV electrons [first established for SN1006, @koyama]. Since the energy of electrons suffers from radiation losses, this might indicate even higher energies for ions. Moreover, hard X-ray tails up to 80 keV have been discovered for several Galactic SNRs [@Allen1999]. This has been contributed to either non-thermal bremsstrahlung [@Laming2001a] or to synchrotron radiation [@Allen1997]. In recent years, additional direct evidence for efficient cosmic ray acceleration has come from detection of TeV $\gamma$-rays for several SNRs by the High Energy Gamma-Ray Astronomy (HEGRA) experiment, the High Energy Spectroscopic System [H.E.S.S, e.g. @Aharonian2004] and MAGIC [@Albert]. The $\gamma$-ray emission is either caused by inverse Compton scattering by the same electrons that cause X-ray synchrotron emission, or by pion production caused by collisions of accelerated ions with the background plasma.
Cassiopeia A (Cas A) is one of the supernova remnants with a hard X-ray tail [@The] and has recently also been detected in $\gamma$-rays [@Aharonian2001; @Albert]. This remnant was until recently[^1] the youngest known supernova remnant in the Galaxy; its supernova was probably around 1671 [@Thorstensen]. In 2001 Chandra detected thin, X-ray synchrotron emitting, filaments at the forward shock of the remnant . This implies the presence of electrons with energies of $\sim 10^{13}$ eV for the magnetic fields in Cas A, estimated to be 0.1 mG to 0.6 mG [@Vink2003; @Berezhko]. These synchrotron rims can be understood in the context of diffusive shock acceleration and synchrotron cooling downstream of the shock.
In addition, the Chandra image shows thin filaments at the inside of the remnant. Some of these inner filaments show a featureless spectrum (Fig. \[Example\]). [@Hughes] identified one of the inner filaments at the West side of the remnant (‘region D’) as being the projected forward shock, based on its featureless spectrum. [@DeLaney] found that the kinematics of the inner filaments, which they interpreted as projected forward shock filaments, are different from the forward shock; they have a lower velocity.
Diffusive shock acceleration is a process which accelerates cosmic rays at a shock front [for a review, see @Malkov]. This mechanism accelerates charged particles of sufficient energy, which scatter on turbulent magnetic fields/plasma waves on both sides of the shock front. Each time the shock front is crossed, the particle gains energy, due to the difference in plasma velocity between both sides of the shock front. The higher the difference in the velocities is, the more energy is gained in one iteration and the higher the magnetic field and magnetic field turbulence, the more often particles cross the shock front. Since at the location where efficient particle acceleration takes place recently accelerated electrons are present, these locations show X-ray synchrotron radiation. However, further downstream from the shock front, synchrotron losses result in lower maximum energies of the synchrotron radiation.
The synchrotron spectrum can be approximated over a large range in frequencies with a power-law in flux density: $F_\nu \propto \nu^{-\alpha}$ with an index ($\alpha$) related to the power-law index of the electron distribution ($p$) as: $\alpha = (p-1)/2$. In what follows, we use index $\Gamma$, which refers to the number density index $\Gamma = -(\alpha + 1)$ and $n(E) \propto E^\Gamma$. Near the maximum electron energies, the electron spectrum has an exponential cut-off, but the resulting synchrotron spectrum cuts off less abruptly, roughly as $\exp(-{\sqrt{\nu/\nu_{max}}})$ [@Zirakashvili]. In contrast, the other important continuum emission process, thermal bremsstrahlung, has an exponential cut-off ($\propto \exp({-h\nu/kT})$). In Cas A, the plasma temperature ranges between 0.6 and 3.6 keV [@Yang]. One should take into account that these temperatures originate from a thermal model. If a partly non-thermal spectrum is fitted with a thermal model, the fitted temperature tends to increase with respect to the real temperature of the plasma. For a thermal bremsstrahlung spectrum with a temperature of 3.5 keV, the power-law index between 4.2 and 6.0 keV is -2.8. For a synchrotron spectrum at the forward shock of Cas A, we typically measure a power-law index of -2.1. We therefore expect the bremssstrahlung continuum to be steeper than the synchrotron continuum in the 4 to 6 keV continuum band.
In this paper we investigate the shape of the continuum spectrum and its spatial distribution in order to address several questions pertaining to the shock acceleration in Cas A: What is the location of the X-ray synchrotron filaments? What fraction of the overall X-ray continuum is thermal and what fraction is non-thermal? And what are the implications for the hard X-ray emission, above 10 keV. We do this by analyzing the Chandra megasecond observation of Cas A.
The used data {#data}
=============
Chandra observed Cas A for one million second in 2004 from Februari to May [@megasecond] with the ACIS S3. For extracting the images, we use the CIAO package, version 3.4. To align the separate pointings, the central compact object is taken as a reference point. We mostly concentrate on the 4.2 to 6 keV energy band, but in addition, we also extracted an image in the line of Si XIII He$\alpha$ from 1.80 to 1.90 keV. We corrected this for continuum emission by subtracting the average of two images next to the Si line: in the 1.63 - 1.64 and 2.13 - 2.14 keV bands.
We compared the results of our analysis of the Chandra data with a VLA radio map, made in the 6 cm band in 2000-2001. This map was kindly provided to us by Tracey DeLaney [@DeLaney2005]. The VLA image has a resolution of $0.38\arcsec \times 0.33 \arcsec$, which is comparable to the Chandra telescope resolution of $0.42\arcsec$. However, the Chandra pixels size of $0.49\arcsec$ slightly undersamples the Chandra resolution.
In section \[spectra\], we make use of BeppoSAX observations made in June 2001, with an exposure time of 501 ks. The hard X-ray spectrum obtained with the PDS instrument is described in [@Vink2001]. The data obtained with the MECS instruments were never published before, but the details of the analysis are similar to the analysis described in [@Vink1999] for a shorter exposure.
Finally, the INTEGRAL-ISGRI spectrum used in section \[spectra\] is described in [@Renaud].
Separating the forward and reverse shock {#Lucy}
========================================
The surface brightness ($\Sigma(r)$) of an optically thin object consists of the emissivity function of this object ($\epsilon(r)$), integrated along the line-of-sight. For a spherically symmetric object, this integral is as follows: $$\label{sigma}\Sigma(r) = 2\int ^R _r \epsilon(r')\frac{r'}{\sqrt{r'^2 - r^2}}dr'$$ In which $R$ denotes the outer radius of this object. We do the deconvolution in cylindrical coordinates, with $\theta$ perpendicular to $r$, and $\theta = 0$ is defined in the West, increasing counterclockwise. The number of counts in a sector at inner radius $r$, angular width $d\theta$ (in radians) and thickness $dr$ is: $$\label{counts}C_r = \Sigma(r)d\theta\left (rdr+\frac{(dr)^2}{2}\right )$$ By dividing Cas A in 18 sectors of $20^{\circ}$ each, we first make surface brightness profiles of the X-ray continuum between 4.2 and 6 keV. For the surface brightness profiles, we adopt the center of expansion [@Thorstensen]: $\alpha = 23^{\rm{h}}23^{\rm{m}}27^{\rm{s}}.77$ and $\delta = 58^{\circ}48\arcmin49\farcs4$ (Equinox J2000) and take step-sizes of 05. Furthermore, we assume sperical symmetry for each sector individually. For an example of a surface brightness profile, see the radial surface brightness profile between an angle of $10^{\circ}$ to $30^{\circ}$ in Fig. \[profile\]. In this figure, we can clearly see the outer shock coming up at $~$160, as already found by [@Gotthelf]. Using equations \[sigma\] and \[counts\], we now make an emission profile in such a way that it fits the surface brightness of the outer shock (smooth solid line). As the line indicates, this thin, outer shell, can not account for all the surface brightness in the center of the remnant (the line which fits the outer peak is for a small $R$ at least 15 times lower than the surface brightness profile). We also see a peak in the surface brightness at 126. We can identify this surface brightness peak with the featureless filament described by [@Hughes]. If we now include a second emission peak at 126(dashed line), we see that a large part of the surface brightness in the center is covered.
To go from surface brightness to emissivity we use a general de-convolution method described by [@Lucy] and previously used on SN1006 by [@Willingale]. To test this algorithm, we simulated emission functions, convolved them into a surface brightness profile and add Poisonian noise using the IDL routine ‘poidev’ from the NASA IDL Astronomy User’s Library [@IDL], we made them in such a way that the peak of the number of counts in one bin of $20^{\circ}\times 0\farcs5$ is 11000, which corresponds to the peak of Fig. \[profile\] in terms of counts per $0\farcs5 \times 20^{\circ}$ bin. After constructing this surface brightness profile we used the Lucy algorithm to recover the emissivity function. We stopped the de-convolution after 30 iterations. The resulting emissivity function is similar to the original one, within 10% for $R>40$and even within 5% for $R>75$.
\[results\]Results of the deconvolution
---------------------------------------
The results of the de-convolution of the Chandra data of Cas A are given in Fig. \[diagram\]. We see the forward shock coming up in the X-ray continuum at 160and with a width of 10. What we also see, is another ring inside the forward shock with a radius of 115with a width of 30, shifted to the right by 15. The forward shock and this inner ring are two almost complete circles, except for the NE region, where the jet is present. If we do the same for the radio and Si images of Cas A, we see a ring at the same location as the inner ring of the X-ray continuum. These inner rings in radio and Si have in previous researches been identified as the reverse shock [@Gotthelf].\
We note that the results of the deconvolution are far from perfect; the $\chi^2_{red}$ of the individual fits range from 20 to 253 for 4.2 to 6.0 keV and from 18 up to 1150 for the Si band. This is to be expected, since the filamentary structure is clearly inconsistent with the assumed spherical symmetry. Moreover, the X-ray Doppler maps of Cas A [@Willingale2002] also show deviations from spherical symmetry. However, the fact that we can trace the inner and outer shells consistently from 18 independent deprojected sectors, argues for the veracity of the two rings.
In order to estimate the total contribution of each shell to the total emissivity of Cas A in each band, we multiply the deprojected shells with the volume of the shells ($4\pi R^2 \Delta R$), taking into account the relative contributions of each sector. For the continuum band in the East to South ($\theta \in [180^{\circ}, 260^{\circ}]$), for which the overall X-ray radiation is dominated by silicon line emission, 16 % comes from the outer ring and 84 % comes from the inner ring. For the Western part of the remnant ($\theta \in [-100^{\circ}, 20^{\circ}]$), in which most of the filaments dominated by continuum are situated, we find that 18 % of the continuum emission is due to the outer ring and 82% due to the inner ring.
Thermal Bremsstrahlung versus Synchrotron radiation {#spectra}
===================================================
The hard X-ray emission from Cas A is clearly non-thermal in nature [@The; @Allen1997; @Favata]. In order to see how this relates to the Chandra spectrum of Cas A, we show in Fig. \[powerlaw\] both the broadband (1-8 keV) Chandra spectrum and the hard X-ray spectrum (above 20 keV) as obtained with BeppoSAX/PDS [@Vink2001] and INTEGRAL/ISGRI [@Renaud]. The model shown is fitted to the hard X-ray data only and is similar to the one used by Renaud et al., which includes the contributions of $^{44}$Ti decay at 67.8 and 78.4 keV. We included absorption with N$_{\rm H} = 1.3\times10^{22}\rm{cm}^{-2}$. Of interest here are the best fit parameters of the non-thermal component, fitted with a power-law. Our best fit parameters are $\Gamma = -3.4 \pm 0.2$ and the normalisation is 3.22 $\pm 1.9$ counts keV$^{-1}$ cm$^{-2}$ s$^{-1}$ at 1 keV. We checked these results with previous results on the hard X-ray tail of the spectrum of Cas A. [@Rothschild] observed Cas A for 226 ks with RXTE and fitted a power-law to the HEXTE data from 20 to 200 keV. They find a power-law index of -3.125 $\pm$ 0.050 with a normalization of $2.0\pm 0.6$ counts cm$^{-2}$s$^{-1}$ at 1 keV.
We see that the BeppoSAX power-law, extrapolated back to Chandra energies, fits the continuum between 2.0 - 7.0 keV well. Within a confidence level of 90%, the extrapolated powerlaw contains a minimum of 93% of the counts in the Chandra continuum bands. The results of [@Rothschild] amounts to a continuum flux from 4 to 6 keV which is 28% of the continuum flux measured by Chandra. Note that if we trust the power-law continuum model, there is little room for an additional thermal component. This is surprising, since we know that the continuum in the 1-10 keV range is at least partially due to thermal bremsstrahlung, which must accompany the copious X-ray line emission. In fact, fitting the spectrum in the 0.5-10 keV band can be done with a pure thermal model [@Vink1996; @Willingale2002].
To pinpoint the different mechanisms which contribute to the continuum in the Chandra band, we fit a power-law to regions of $4.9\arcsec \times 4.9 \arcsec$ ($10 \times 10$ pixels) for the band between 4.2 and 6.0 keV. We use the eventfiles of the megasecond observation, which we merge, using the central compact object as a reference point. We use one ARF[^2] file, made for the spectrum of Fig. \[powerlaw\]. We estimated the background contribution, using an annulus around Cas A, with $R_{min}=205\farcs 8$ and $R_{max} = 235\farcs 2$ and a center as defined in section \[Lucy\]. We neglected the effects of differential absorption over this small band; in the most extreme cases, the absorption varies from $1\times 10^{22}$ to $1.7 \times 10^{22} \rm{cm}^{-2}$ which has an effect of at maximum 0.09 in the power-law index. Fitting the continuum of the total remnant between 4.2 to 6.0 keV, using N$_{\rm H} = 1.3\times 10^{22}\rm{cm}^{-2}$, we found that the average index is -3.14.
Fig. \[specsel\] shows that the power-law index varies considerably inside Cas A. We see that some of the regions with hard spectra overlap with known regions of non-thermal X-ray emission. For example the forward shock region appears to have a harder power-law. And also the non-thermal filaments in the West show up in the spectral index map. However, there are also is also hard continuum emission in regions where there is also line emission. This is illustrated by Figures \[specsel\] and \[sum\].
The spectral map of Fig \[specsel\] implies that if we observe Cas A at higher energies, the Western part of the remnant will become more prominent. We validated this by comparing an extrapolation of our best fit power-law spectra with the 501 ks BeppoSAX/MECS observation in the 9-11 keV band.[^3] We deconvolved the BeppoSAX image using the Lucy deconvolution method, following the procedure described in [@Vink1999]. The results are shown in Fig \[911extra\]. for comparison, we also show the 4 to 6 keV BeppoSAX image next to the Chandra image smoothed to roughly the same resolution (the extrapolation without smoothing is shown in Fig \[911extraChandra\]). Qualitatively the 9-11 keV image of BeppoSAX agrees with the extrapolated Chandra image: the South-Eastern part of the remnant is relatively less bright, whereas the Western part and the Southern part of the center are becoming more prominent.
Since for a bremsstrahlung continuum, we expect an exponential cut-off and thus a soft spectrum and for synchrotron radiation, we expect a hard spectrum, we tentatively identify the hard spectra with synchrotron emission. Although there is a likely overlap in spectral index between thermal and non-thermal emission we can nevertheless estimate the total contribution of non-thermal emission by noting that in Fig. \[specsel\] left, those regions that have abundant line emission (the Eastern part of the shell), have power-law spectra steeper than -3.1. The total flux associated with those power-law indices amounts to 46 % of the total flux in the 4.2 to 6 keV band (Fig. \[integrated\]). Suggesting that the other 54% of the flux, is due to non-thermal radiation. This corresponds roughly to a total non-thermal flux above 4 keV of 2.7 $\times 10^{-10}{\rm erg~s^{-1}~cm^{-2}}$. If we take a power-law index of -2.8 as an upper limit, this corresponds to a thermal bremsstrahlung model with kT = 3.5 keV, the non-thermal contribution comes down to 33 %. This is indeed a large fraction of the total contintuum emission, but not as large as the 93% suggested by extrapolation of the hard X-ray power-law. The reason for this is probably that the true spectral shapes are not exactly power-laws, but steepen at higher energies. So apparently, in the Chandra band the addition of soft thermal emission and hard non-thermal gives by coincidence almost the same power law index, as the steepening non-thermal spectra at high energies. The lack of an obvious spectral break is probably due to the variation in power law indices for the non-thermal spectra, and the variation in cut-off energies across the remnant.
\[West\]West side of Cas A
--------------------------
So far, we have shown that a large part of the non-thermal emission is associated with the reverse shock region and that most of this emission is coming the Western part of the remnant. This prompted us to investigate the spectrum of the Western part more closely. Specifically we were interested in whether the reverse shock in the Western part shocks the less dense material. The reason is that the Western region not only shows more non-thermal emission, but there also seems to be less thermal emission. A low density of a shocked material results, apart from less flux from shock-heated plasma, also in a low ionization time scale, $n_{\rm e}t$. With this in mind, we extracted spectra from one of the featureless filaments, located at the reverse shock. This is the same filament as used by [@Hughes] (filament ‘D’). We extracted the spectrum of this region in the individual eventfiles of the megasecond image and added them with the `addspec` tool of ftools in heasoft. We fit the resulting spectrum with a model consisting of a power-law and the non-equilibrium ionization (NEI) model of SPEX [@spex]. The best fit parameters are listed in Table \[nei\].
tab1.tex
Discussion
==========
Using the a combination of deprojection and characterization of the X-ray continuum power-law slope of the one megasecond Chandra observation, we have determined the location of the non-thermal X-ray emission of Cas A. We have found that a significant fraction of the X-ray continuum emission is coming from two separate rings; one can be identified with the forward shock, the other with the shell dominated by bright ejecta. About 54% of all continuum emission in the 4.2 to 6 keV band is likely to be non-thermal.
The non-thermal emission from the forward shock is likely due to synchrotron radiation from electrons accelerated at the forward shock, confined to the forward shock region by synchrotron cooling [@Vink2003; @Berezhko]. However, the non-thermal emission of Cas A as a whole is dominated by the contribution of the inner ring. This inner region of emission is shifted to the West of the remnant by $\sim15\arcsec$. A similar shift was found for the location of the reverse shock by [@Reed; @Gotthelf], on the basis of optical data and a deprojection of radio data and a 50 ks Chandra observation, in the Si band. [@Gotthelf] also found marginal evidence for a higher emissivity at the reverse shock in the 4 to 6 keV band using the 50 ks Chandra observation.
Nature and location of the inner non-thermal emission
-----------------------------------------------------
Since some of the spectra in the West hardly show any line emission, we think it is likely that this is synchrotron emission, since non-thermal bremsstrahlung without line emission is hard to establish, unless one has peculiar abundances. Such a bremsstrahlung model was once invoked for SN1006 [@Hamilton1986], but has now been abandoned in favor of synchrotron emission. Furthermore, a recent paper by [@Vink2008] shows that electrons with energies close to the thermal electron energy distribution, lose their energy relatively fast due to Coulomb losses. This process is already important in the Chandra energy band for $n_{\rm e}t < 8.6 \times 10^{10} {\rm cm}^{-3}{\rm s}$; the value reported here for a filament, dominated by continuum emission in the West (section \[West\], Table \[nei\]). For $n_{\rm e}t \sim 10^{11} {\rm cm}^{-3}{\rm s}$, typical for Cas A [@Willingale2002], one only expects to see non-thermal bremsstrahlung for photon energies $\gtrsim 100~\rm{keV}$.
When it comes to the location of the synchrotron emission, there are two possibilities: the reverse shock and the contact discontinuity. The contact discontinuity marks the border between shocked ejecta and shocked circumstellar medium. Hydrodynamical solutions show that the density and the magnetic field peak at this radius [@Chevalier1982; @Lyutikov]. For a supernova remnant evolving in a stellar wind, the contact discontinuity is close to the reverse shock. So from our deprojections it difficult to judge whether the X-ray synchrotron emission is coming from the reverse shock or the contact discontinuity. It is unlikely that electrons are accelerated to TeV energies at the contact discontinuity, since no viable acceleration mechanism is known [however, see @Lyutikov], but there are two ways of generating X-ray synchrotron emission at the contact discontinuity: 1) Due to an increase of the magnetic field, electrons with relatively low energies suddenly light up in X-rays. 2) High energy electrons and positrons are created through the decay of charged pions ($\pi^\pm$), caused by hadronic cosmic ray collisions [@Gaisser1990]. Charged pions decay into muons and muon neutrinos. The muons decay into electrons/positrons and electron and muon neutrinos. The electrons and positrons thus created are often called secondary electrons and positrons.
For option 1: For synchrotron radiation the relation between electron energy and photon energy is $$E_{{\rm ph}} = 19 E_{{\rm TeV}}^2B_{\rm G}~{\rm keV} \label{Eph}$$ [@Ginzburg1965]. For electrons to be invisible in X-ray synchrotron radiation, the peak energy should be an order of magnitude lower than the X-ray continuum at 4-6 keV. Therefore, the magnetic field should increase an order of magnitude at the contact discontinuity. As an example, an electron, accelerated at the forward shock, with an energy of 3 TeV in the typical magnetic field of Cas A of 0.5 mG, typically emits photons of 0.09 keV. When this electron suddenly enters a region of 5 mG, it will emit photons of typically 1 keV; detectable in X-rays. However, the synchrotron loss time for such an electron in a 0.5 mG magnetic field, 18 years, is rather short compared to the age of Cas A. In this time the particle has to diffuse from the forward shock region to the contact discontinuity, which is about 0.5 pc, for a distance to Cas A of 3.4 kpc [@Reed]. For 3 TeV and B=0.5 mG the diffusion constant is $D=\eta\cdot 2\times 10^{23}~{\rm cm^2s^{-1}}$, with $\eta=1$ corresponding to Bohm diffusion [@Malkov]. In 18 yr a particle can diffuse by $R \approx \sqrt{2Dt} = 5\times 10^{-3} \sqrt{\eta}$ pc. In other words, this model only works if the magnetic field turbulence is very low, corresponding to $\eta > 10000$. This unlikely, since the ample presence of cosmic rays results in magnetic field turbulence. Moreover, at the shock front $\eta \sim 1$ [@Vinkreview2006; @Stage2006]. For low values of $\eta$, the diffusion length ($l_{\rm diff}$ the length for which advection dominates over diffusion), is short: $l_{\rm diff} = \eta 4.5\times10^{-4} $ parsec. So, for low $\eta$, the advection velocity, ($u = v_{\rm{fs}}/4$, with $v_{\rm{fs}}$ the forward shock velocity) is the relevant velocity. For a $v_{\rm{fs}}$ of 5800 [[kms$^{-1}~$]{}]{}, $u=1450$ [[kms$^{-1}~$]{}]{}. Using this velocity, it takes 674 years to go from the forward shock to the contact discontinuity, longer than the synchrotron loss time of 18 years.
It has been argued that the magnetic field at the shock front is high, but rapidly decays towards the inside [@Pohl], increasing the synchrotron loss times, and increasing the diffusion constant. However, the decay in magnetic field should be reflected in the radio emissivity from the forward shock to the inside, which is contrary to observations, that show gradual increase in emissivity toward the center starting at the shock front [@Gotthelf].
For option 2 (secondary electrons): if the X-ray synchrotron emission from the inside is due to secondary electrons, this would be an important discovery. It would be evidence for the presence of TeV ion cosmic rays. Neutral pions ($\pi^0$) are made in comparable quantities to charged pions. The power in secondary electrons should therefore be comparable to the luminosity in pion decay. In order to compare the fluxes from X-ray synchrotron and $\gamma$-ray radiation one should take into account the conversion from pion energy ($\pi^0$ & $\pi^\pm$) to photon energies, both in the TeV band (through $\pi^0$ decay) and in the X-ray band, due to synchrotron radiation from secondary electrons. Taking into account these various decay channels, one finds that $E_{{\rm ph}} \sim 20 E_{{\gamma \rm TeV}}^2B_{\rm G}~{\rm keV}$ [^4], so the X-ray synchrotron flux above 4 keV should correspond to roughly the $\gamma$-ray flux above 20 TeV. We found for the X-ray synchrotron flux above 4 keV 2.7$\times 10^{-10}$ erg s$^{-1}$cm$^{-2}$. This is a factor 130 higher than the $\gamma$-ray flux above 1 TeV, which we calculated to be $2.1\times10^{-12}~\rm{erg}~\rm{cm}^{-2}~\rm{s}^{-1}$ using the photon-flux and photon index above 1 TeV reported by the MAGIC collaboration [@Albert]. Since we should actually have evaluated the flux above 20 TeV, we can rule out that the X-ray synchrotron emission is caused by secondary electrons.
In our view it is therefore most likely that the X-ray synchrotron emission is caused by electrons accelerated at the reverse shock. Similarly to the forward shock region the electrons are likely to be confined to a region near the shock itself. For a long time the reverse shock as location for acceleration has been neglected, because of its putative low magnetic field. Moreover, the abundance pattern of cosmic rays is consistent with acceleration from plasmas with solar abundances [@Hoerandel]. As far as the magnetic field is concerned, however, @ellison have argued that if magnetic field amplification [@bell] works at the forward shock it is likely to operate at the reverse shock as well. Also for the SNR RCW 86 it has been suggested that the X-ray synchrotron emission in the Southwest of the remnant is coming from electrons accelerated at the reverse shock [@rho].
Reverse shock velocity in the West and the presence of synchrotron emission
---------------------------------------------------------------------------
The presence of X-ray synchrotron radiation from shock accelerated electrons is only expected if the shock velocity is high enough: shock acceleration theory predicts for the maximum photon energies [@Aharonian1999]: $$E_{\rm ph} = 0.5 \eta^{-1} \bigl( \frac{v_{\rm s}}{2000{\ \rm km s^{-1}}}\bigr)^2 \ {\rm keV},\label{eq-emax}$$ with $v_s$ the shock velocity. Note that the photon energy is independent of the magnetic field. For the reverse shock the velocity in Eq. \[eq-emax\] refers to the shock speed in the frame of the ejecta. The ejecta velocity is equal to the free expansion velocity $v_{\rm f, ej} =r_{\rm ej}/t$. For the reverse shock the shock velocity as seen by the ejecta is therefore $v_{\rm s, ej} = v_{\rm f, ej} - v_{\rm s, obs}$, with $v_{\rm s, obs}$ the shock velocity in the frame of the observer. The presence of X-ray synchrotron radiation from the reverse shock in the Western half of Cas A suggests a higher reverse shock than in the rest of the remnant. Proper motions of knots at the inside of Cas A in X-rays were most recently measured by [@DeLaney], but some details of the measurements, including measurements of the proper motions as a function of azimuth, only appeared in appendix 4.4 of @DeLaneyPhD [Fig. 4.6]. In this Figure, we see that in the West the expansion rate is between -0.1%yr$^{-1}$ and 0.1%$^{-1}$ this corresponds to $v_{\rm{r,ob}}$ at the reverse shock of approximately -2000 [[kms$^{-1}~$]{}]{}to 2000 [[kms$^{-1}~$]{}]{}, implying a shock velocity [^5] $v_{\rm{s,ej}} \approx 3900 - 7900~$[[kms$^{-1}~$]{}]{}. In the Northern and Eastern part of Cas A, @DeLaneyPhD finds expansion rates of 0.2%yr$^{-1}$, corresponding to $v_{\rm{r,ob}} \approx 4000$ [[kms$^{-1}~$]{}]{}and a shock velocity of $v_{\rm{s,ej}} \approx 1900~$[[kms$^{-1}~$]{}]{}at the reverse shock.
In the optical @Morse reports a shock velocity of $v_{\rm{s,ej}} \approx 3000~$[[kms$^{-1}~$]{}]{}for a filament in the Northwest. In our maps (Fig. \[specsel\] left) we find that this region has a hard power law slope, suggesting the presence of synchrotron radiation, but the overall X-ray emission is dominated by thermal X-ray line emission (Fig. \[specsel\] right).
It indeed looks like the presence of X-ray synchrotron emission from the inner ring corresponds with reverse shock velocities in excess of 2000 [[kms$^{-1}~$]{}]{}. To some extend this is surprising, since the X-ray synchrotron emissivity function is rather broad and can result in some synchrotron emission beyond the maximum photon energy as defined in Eq. \[eq-emax\]. On the other hand, $\eta = 1$ corresponds to Bohm diffusion, and represents the case for maximum acceleration efficiency. The fact that 2000 [[kms$^{-1}~$]{}]{}seems close to the velocity dividing the presence or absence of X-ray synchrotron radiation again suggests that acceleration takes place close to the Bohm limit. Equation \[eq-emax\] suggests that the spectra in the West of the remnant will be harder than in the rest, where $v_{\rm{s,ej}}$ is lower.
Note that the high reverse shock velocities in the West (and low observed velocities) are not in agreement with analytic hydrodynamic solutions for an SNR evolving in a stellar wind [@Laming]. Adapting the parameters of this model such that the forward shock radius and velocity and the reverse shock radius, match those of Cas A, we find $v_{\rm s, ej}$ ranging between 1600-2300 [[kms$^{-1}~$]{}]{}, significantly lower than in the West, but in agreement with those in the Eastern part of Cas A. This suggests that the circumstellar density structure is more complex in the West.
Conclusions {#Conclusions}
===========
We have presented an anaysis of the spatial and spectral variation of the X-ray continuum emission of Cas A, based on the 1 Ms Chandra observation. We find that harder continuum spectra are associated with the filaments, dominated by continuum emission, suggesting that the harder spectra are caused by non-thermal radiation. A dominant fraction of the non-thermal emission appears to come from the reverse shock region. We have discussed various options for the nature of the non-thermal emission and its origins. Some of our conclusions were independently also obtained by [@Uchiyama], but based on X-ray variability. Based on our analysis and discussion we come to the following conclusions:
- The power-law index of the spectrum between 4.2-6.0 keV is an indicator for X-ray synchrotron emission: there is a correlation between filaments, dominated by continuum emission and hard spectra,
- hard X-ray spectra are not exclusively associated with filaments, dominated by continuum emission, suggesting that non-thermal emission comes also from other regions,
- the non-thermal X-ray emission is likely to be synchrotron radiation,
- the non-thermal accounts for about 54% of the overall continuum emission in the 4-6 keV band,
- in the Western part of Cas A, most X-ray synchrotron comes from the reverse shock,
- the dominance of X-ray synchrotron emission from the West is probably the result of a locally higher reverse shock velocity of $v_s \sim 6000$ [[kms$^{-1}~$]{}]{}(corresponding to a lower proper motion) than in the Eastern region ($v_s \sim1900~$[[kms$^{-1}~$]{}]{}).
We want to thank Tracey DeLaney for providing us with a recent radio map of Cas A. We would like to thank Frank Verbunt for carefully reading the manuscript and Michele Cappellari for the software we used to display some of our results. This work is supported by the NWO-VIDI grant of J.V.
[*Facilities:*]{} .
[54]{} natexlab\#1[\#1]{}
, F. [et al.]{} 2001, , 370, 112
, F. A. & [Atoyan]{}, A. M. 1999, , 351, 330
, F. A. [et al.]{} 2004, , 432, 75
, J. [et al.]{} 2007, , 474, 937
, G., [Gotthelf]{}, E. V., & [Petre, R.]{} 1999, in International Cosmic Ray Conference, Vol. 3, International Cosmic Ray Conference, 480–+
, G. E. [et al.]{} 1997, , 487, L97+
, A. R. & [Lucek]{}, S. G. 2001, , 321, 433
, E. G. & [V[ö]{}lk]{}, H. J. 2004, , 419, L27
, J. A. M. [et al.]{} 2001, , 365, L225
, R. A. 1982, , 258, 790
, T. [et al.]{} 2004, , 613, 343
, T. [et al.]{} 2005, in X-Ray and Radio Connections (eds. L.O. Sjouwerman and K.K Dyer) Published electronically by NRAO, http://www.aoc.nrao.edu/events/xraydio Held 3-6 February 2004 in Santa Fe, New Mexico, USA, (E4.05) 7 pages, ed. L. O. [Sjouwerman]{} & K. K. [Dyer]{}
, T. A. 2004, PhD thesis, AA(UNIVERSITY OF MINNESOTA)
, D. C., [Decourchelle]{}, A., & [Ballet]{}, J. 2005, , 429, 569
, F. [et al.]{} 1997, , 324, L49
, T. K. 1990, [Cosmic rays and particle physics]{} (Cambridge and New York, Cambridge University Press, 1990, 292 p.)
, V. L. & [Syrovatskii]{}, S. I. 1965, , 3, 297
, E. V. [et al.]{} 2001, , 552, L39
, D. A., [Reynolds]{}, S. P., [Borkowski]{}, K. J., [Hwang]{}, U., [Harrus]{}, I., & [Petre]{}, R. 2008, , 387, L54
, A. J. S., [Sarazin]{}, C. L., & [Szymkowiak]{}, A. E. 1986, , 300, 698
, J. R. 2008, Advances in Space Research, 41, 442
, J. P., [Rakowski]{}, C. E., [Burrows]{}, D. N., & [Slane]{}, P. O. 2000, , 528, L109
, U. [et al.]{} 2004, , 615, L117
, J. S., [Mewe]{}, R., & [Nieuwenhuijzen]{}, H. 1996, in UV and X-ray Spectroscopy of Astrophysical and Laboratory Plasmas : Proceedings of the Eleventh Colloquium on UV and X-ray ... held on May 29-June 2, 1995, Nagoya, Japan. Edited by K. Yamashita and T. Watanabe. Tokyo : Universal Academy Press, 1996. (Frontiers science series ; no. 15)., p.411, ed. K. [Yamashita]{} & T. [Watanabe]{}, 411–+
, J. W. 1996, in Astronomical Society of the Pacific Conference Series, Vol. 99, Cosmic Abundances, ed. S. S. [Holt]{} & G. [Sonneborn]{}, 362–365
, K. [et al.]{} 1995, , 378, 255
, J. M. 2001, , 546, 1149
, J. M. & [Hwang]{}, U. 2003, , 597, 347
, W. B. 1993, in Astronomical Society of the Pacific Conference Series, Vol. 52, Astronomical Data Analysis Software and Systems II, ed. R. J. [Hanisch]{}, R. J. V. [Brissenden]{}, & J. [Barnes]{}, 246–+
, L. B. 1974, , 79, 745
, M. & [Pohl]{}, M. 2004, , 609, 785
, M. A. & [Drury]{}, L. 2001, , 64, 429
, J. A. [et al.]{} 2004, , 614, 727
, M., [Yan]{}, H., & [Lazarian]{}, A. 2005, , 626, L101
, J. E., [Hester]{}, J. J., [Fabian]{}, A. C., & [Winkler]{}, P. F. 1995, , 440, 706
, M. [et al.]{} 2006, , 647, L41
, S. P., [Borkowski]{}, K. J., [Green]{}, D. A., [Hwang]{}, U., [Harrus]{}, I., & [Petre]{}, R. 2008, , 680, L41
, J., [Dyer]{}, K. K., [Borkowski]{}, K. J., & [Reynolds]{}, S. P. 2002, , 581, 1116
, R. E. & [Lingenfelter]{}, R. E. 2003, , 582, 257
, M. D., [Allen]{}, G. E., [Houck]{}, J. C., & [Davis]{}, J. E. 2006, Nature Physics, 2, 614
, L.-S. [et al.]{} 1996, , 120, C357+
, J. R., [Fesen]{}, R. A., & [van den Bergh]{}, S. 2001, , 122, 297
, Y. & [Aharonian]{}, F. A. 2008, , 677, L105
, J. 2006, ArXiv Astrophysics e-prints
—. 2008, , Accepted
, J., [Bloemen]{}, H., [Kaastra]{}, J. S., & [Bleeker]{}, J. A. M. 1998, , 339, 201
, J., [Kaastra]{}, J. S., & [Bleeker]{}, J. A. M. 1996, , 307, L41
, J. & [Laming]{}, J. M. 2003, , 584, 758
, J. [et al.]{} 1999, , 344, 289
—. 2001, , 560, L79
, R., [Bleeker]{}, J. A. M., [van der Heyden]{}, K. J., [Kaastra]{}, J. S., & [Vink]{}, J. 2002, , 381, 1039
, R., [West]{}, R. G., [Pye]{}, J. P., & [Stewart]{}, G. C. 1996, , 278, 749
, X.-j., [Lu]{}, F.-j., & [Chen]{}, L. 2007, ArXiv e-prints, 712
, V. N. & [Aharonian]{}, F. 2007, , 465, 695
[^1]: Recent expansion measurements of G1.9+0.3 show an age for this remnant of around 100 years [@Green; @Reynolds2008].
[^2]: Ancillary Response File, which contains a description of the effective area of the instrument.
[^3]: There is also an observation made by XMM-Newton of Cas A up to 15 keV [@Bleeker], because of errors in the exposure map of this observation, we choose to use the BeppoSAX data.
[^4]: The decay product of $\pi^0$ decay is two photons, with each, on average, an energy $E_{\gamma \rm TeV} = 0.5E_{\pi^0}$. For the charged pions, $\pi^{\pm}$, the final decay product consists of electrons (positrons) and neutrinos ( $\pi^\pm \rightarrow \mu^\pm + \nu_\mu(\bar{\nu}_\mu)$, $ \mu^{\pm}\rightarrow e^\pm + \nu_e(\bar{\nu}_e) + \nu_\mu(\bar{\nu}_\mu)$). In this case the final electron or positron also take up, on average, half the initial pion energy $E_{e^\pm} \approx 0.5 E_{\pi^\pm}$; hence, for a given pion energy, $E_{e^\pm}\approx E_{\gamma \rm TeV}$. This electron or positron emits synchrotron radiation at a peak frequency given by Eq. \[Eph\]. Hence the close resemblance of this equation with Eq. \[Eph\].
[^5]: We have checked this result using the Chandra 1 Ms observation and an observation of Chandra made in 2000 using the procedure described in @Vink1998. We concentrated on filament ’D’. The result confirms the lower expansion or even backward velocities: we found $-970\pm 140$ [[kms$^{-1}~$]{}]{}, i.e. the filament seems to move towards the center. This implies a $v_{\rm{s,ej}} \approx 6900~$[[kms$^{-1}~$]{}]{}. @Vink1998 already reported a lower expansion for the whole Western part of Cas A.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
For a normed linear space $(X,\|\cdot\|)$ and $p>0$ we characterize all $n$-tuples $(\mu_1,\cdots,\mu_n)\in\mathbb{R}^{n}$ for which the generalized triangle inequality of the second type $$\|x_1+\cdots+x_n\|^p\leq\frac{\|x_1\|^p}{\mu_1}+\cdots+\frac{\|x_n\|^p}{\mu_n}$$ holds for any $x_1,\cdots,x_n\in
X$. We also characterize $(\mu_1,\cdots,\mu_n)\in\mathbb{R}^{n}$ for which the reverse of the inequality above holds.
title: Characterization of a generalized triangle inequality in normed spaces
---
Introduction
============
The triangle inequality is one of the most significant inequalities in mathematics. It has many interesting generalizations, refinements and reverses, which have been obtained over the years, see [@AM; @Drag; @KMM; @MAL; @MSKT; @Pec-Raj] and references therein. The generalized triangle inequalities are useful to study the geometrical structure of normed spaces, see e.g. [@Kato; @Hsu; @HL]. In this direction some results have been based on the triangle inequality of the second type $$\begin{aligned}
\|x+y\|^2\leq2(\|x\|^2+\|y\|^2)\end{aligned}$$ in a normed linear space, see [@Belb; @Saitoh; @Takahasi-Ras-Sai-Takahashi] for more information about this inequality.\
In framework of Hilbert spaces the Euler-Lagrange type identity (see [@Ras]) $$\frac{\|x\|^2}{\mu}+\frac{\|y\|^2}{\nu}-\frac{\|ax+by\|^2}{\lambda}=\frac{\|\nu
bx-\mu ay\|^2}{\lambda\mu\nu}\ \ \ (\lambda=\mu a^2+\nu b^2)$$ follows the more general triangle inequality of the second type (see [@Takahasi-Ras-Sai-Takahashi]) $$\begin{aligned}
\frac{\|ax+by\|^2}{\lambda}\leq\frac{\|x\|^2}{\mu}+\frac{\|y\|^2}{\nu}\,,\end{aligned}$$ where $\lambda=\mu a^2+\nu b^2$ and $\lambda\mu\nu>0$. Also Takahasi et al. [@Takahasi-Ras-Sai-Takahashi] investigate the inequality $$\begin{aligned}
\label{mos}
\frac{\|ax+by\|^p}{\lambda}\leq\frac{\|x\|^p}{\mu}+\frac{\|y\|^p}{\nu}\end{aligned}$$ for $p\geq1$. We should notice that one can assume that $\lambda=\pm 1$ by dividing the both sides of by $|\lambda|$ as well as $a\neq 0, b \neq 0$ since if, e.g. $a=0$, then $\mu$ can be suitably chosen arbitrary. By replacing $x$ and $y$ by $\frac{x}{a}$ and $\frac{y}{b}$ and changing $\mu$ and $\nu$ accordingly, inequality turns into $\frac{\|x+y\|^p}{\pm 1}\leq\frac{\|x\|^p}{\mu}+\frac{\|y\|^p}{\nu}$.
In this paper, for a normed linear space $(X,\|\cdot\|)$ and $p>0$ we characterize all $n$-tuples $(\mu_1,\cdots,\mu_n)\in\mathbb{R}^{n}$ for which the generalized triangle inequality of the second type $$\|x_1+\cdots+x_n\|^p\leq\frac{\|x_1\|^p}{\mu_1}+\cdots+\frac{\|x_n\|^p}{\mu_n}$$ holds for any $x_1,\cdots,x_n\in
X$. We also characterize $(\mu_1,\cdots,\mu_n)\in\mathbb{R}^{n}$ for which the reverse of the inequality above holds.
Main results
============
We need the following three lemmas which generalize some results due to Takagi et al. [@Takagi-Miu-Hay-Takahasi Theorem 2.1, Lemma 2.2 and Theorem 2.3]. Let us first recall the concept of an envelope.
An envelope of a family of surfaces is a surface that is tangent to each member of the family at some point. Let $\mathscr{L}$ be an $(n-1)$-parameters family of surfaces in $\mathbb{R}^n$ given by $F(a_1,\cdots,a_n;s_1,\cdots,s_{n-1})=0$ depending on real parameters $s_1,\cdots,s_{n-1}$ and variables $a_1,\cdots,a_n$. The envelope of $\mathscr{L}$ is the set of points $(a_1,\cdots,a_n)\in\mathbb{R}^n$ for which the following equations hold: $$\begin{aligned}
&F(a_1,\cdots,a_n;s_1,\cdots,s_{n-1})=0\,,&\\
&\frac{\partial F}{\partial
s_i}(a_1,\cdots,a_n;s_1,\cdots,s_{n-1})=0&\ \ \ (1\leq i\leq n-1)\,.\end{aligned}$$
\[lemma1\] Suppose that $p>1$, $S=\{(s_1,\cdots,s_n)\colon s_1,\cdots,s_n\geq0,
\sum_{i=1}^{n}s_i=1\}$ and $h_p(a_1,\cdots,a_{n-1})=\Big(1-\sum_{i=1}^{n-1}a_i^{\frac{1}{1-p}}\Big)^{1-p}$ for $a_1,\cdots,a_{n-1}>0$ with $\sum_{i=1}^{n-1}a_i^{\frac{1}{1-p}}<1$. For each $(s_1,\cdots,s_n)\in S$, let\
$L_p(s_1,\cdots,s_n):=\{(a_1,\cdots,a_n)\colon
a_1s_1^p+\cdots+a_ns_n^p=1\}\ ,$
$\Delta_p(s_1,\cdots,s_n):=\{(a_1,\cdots,a_n)\colon
a_1s_1^p+\cdots+a_ns_n^p\geq1\}\,.$
Then the following assertions hold:\
[(i)]{} If $\mathscr{L}=\{L_p(s_1,\cdots,s_n)\colon(s_1,\cdots,s_n)\in S\}$, then the envelope of $\mathscr{L}$ is given by $a_n=h_p(a_1,\cdots,a_{n-1})$;\
[(ii)]{} $\displaystyle{\bigcap_{(s_1,\cdots,s_n)\in
S}}\Delta_p(s_1,\cdots,s_n)=\{(a_1,\cdots,a_n)\colon a_n\geq
h_p(a_1,\cdots,a_{n-1})\}\,.$
\(i) Putting\
$F(a_1,\cdots,a_n;s_1,\cdots,s_{n-1})=a_1s_1^p+\cdots+a_{n-1}s_{n-1}^p+a_n(1-(s_1+\cdots+s_{n-1}))^p-1$,
we can consider $\mathscr{L}$ as a family of $(n-1)$-parameters of surfaces as follows:\
$$\begin{aligned}
\mathscr{L}&=&\{L_p(s_1,\cdots,s_n)\colon(s_1,\cdots,s_n)\in S\}\\
&=&\Big\{\{(a_1,\cdots,a_n)\colon a_1s_1^p+\cdots+a_ns_n^p=1\}\colon s_1,\cdots,s_n\geq0, \sum_{i=1}^{n}s_i=1\Big\}\\
&=&\Big\{\{(a_1,\cdots,a_n)\colon a_1s_1^p+\cdots+a_{n-1}s_{n-1}^p+a_n(1-(s_1+\cdots+s_{n-1}))^p=1\}\colon\\
&&s_1,\cdots,s_{n-1}\geq 0, \sum_{i=1}^{n-1}s_i\leq1\Big\}\\
&=&\Big\{\{(a_1,\cdots,a_n)\colon
F(a_1,\cdots,a_n;s_1,\cdots,s_{n-1})=0\}\colon
s_1,\cdots,s_{n-1}\geq0, \sum_{i=1}^{n-1}s_i\leq1\Big\}\,.\end{aligned}$$ The envelope of $\mathscr{L}$ is given by the solutions of the following simultaneous equations: $$\begin{aligned}
\label{L1}
&F(a_1,\cdots,a_n;s_1,\cdots,s_{n-1})=0\,,&\nonumber\\
&\frac{\partial F}{\partial
s_i}(a_1,\cdots,a_n;s_1,\cdots,s_{n-1})=0&\ \ \ (1\leq i\leq n-1)\,.
\label{L2}\end{aligned}$$ Equations yield $$\begin{aligned}
\label{L2.5}
pa_is_i^{p-1}-pa_n(1-(s_1+\cdots+s_{n-1}))^{p-1}=0\ \ \ (1\leq i\leq n-1)\end{aligned}$$ It must be $s_i\neq 0$ for all $i=1,\cdots,n-1$. Indeed, if $s_j=0$ for some $1 \leq j\leq n-1$, then $a_1s_1=\cdots=a_{n-1}s_{n-1}=a_n\big(1-\sum_{i=1}^{n-1}s_i\big)^{p-1}=0$ by . So that $F(a_1,\cdots,a_n;s_1,\cdots,s_{n-1})=-1$, which contradicts . Moreover, by the same way, it must be $\sum_{i=1}^{n-1}s_i< 1$.\
From we get $$\begin{aligned}
\label{L3}
a_i=\frac{a_n(1-(s_1+\cdots+s_{n-1}))^{p-1}}{s_i^{p-1}}\qquad (1\leq
i\leq n-1)\,.\end{aligned}$$ Using , equation turns into\
$$\begin{aligned}
a_ns_1(1-(s_1+\cdots+s_{n-1}))^{p-1}+\cdots+a_ns_{n-1}(1-(s_1+\cdots+s_{n-1}))^{p-1}\\
+a_n(1-(s_1+\cdots+s_{n-1}))^{p}-1=0\,,\end{aligned}$$ or equivalently we get $$\begin{aligned}
\label{L4}
a_n=\frac{1}{(1-(s_1+\cdots+s_{n-1}))^{p-1}}\,.\end{aligned}$$ From and we get $$\begin{aligned}
\label{L5}
a_i=\frac{1}{s_i^{p-1}}\ \ \ (1\leq i\leq n-1)\,.\end{aligned}$$ We note that $s_1,\cdots,s_{n-1}>0$ and $\sum_{i=1}^{n-1}s_i<1$ if and only if $a_1,\cdots,a_{n-1}>0$ and $\sum_{i=1}^{n-1}a_i^{\frac{1}{1-p}}<1$. Now we remove the parameters $s_1,\cdots,s_{n-1}$ in equations and to get $$\begin{aligned}
a_n&=&\frac{1}{\left(1-\left(\left(\frac{1}{a_1}\right)^{\frac{1}{p-1}}+\cdots+\left(\frac{1}{a_{n-1}}\right)^{\frac{1}{p-1}}\right)\right)^{p-1}}
=\left(1-\left(a_1^{\frac{1}{1-p}}+\cdots+a_{n-1}^{\frac{1}{1-p}}\right)\right)^{1-p}\\
&=&\left(1-\sum_{i=1}^{n-1}a_i^{\frac{1}{1-p}}\right)^{1-p}=h_p(a_1,\cdots,a_{n-1})\,.\end{aligned}$$ (ii) It is not difficult to check that the function $h_p$ is strictly convex and thus the domain $\{(a_1,\cdots,a_n)\colon a_n\geq
h_p(a_1,\cdots,a_{n-1})\}$ is a strictly convex set in the Euclidean space $\mathbb{R}^n$. Now the result follows from part (i).
\[lemma2\] Let $p>1$ and $\Omega\subseteq \big\{(s_1,\cdots,s_n)\colon
s_1,\cdots,s_n\geq0,\ \sum_{i=1}^{n}s_i\geq1\big\}$. Let $D_p(\Omega):=\{(a_1,\cdots,a_n)\colon a_1,\cdots,a_n\geq 0,
a_1s_1^p+\cdots+a_ns_n^p\geq 1\ for\ all\ (s_1,\cdots,s_n)\in\Omega\}$. Then the following assertions hold:\
[(i)]{} $\{(a_1,\cdots,a_n)\colon a_n\geq
h_p(a_1,\cdots,a_{n-1})\}\subseteq D_p(\Omega)\,;$\
[(ii)]{} If $S\subseteq\overline{\Omega}$ where $\overline{\Omega}$ is the usual closure of $\Omega$, then $D_p(\Omega)=\{(a_1,\cdots,a_n)\colon a_n\geq
h_p(a_1,\cdots,a_{n-1})\}\,.$
\(i) By Lemma \[lemma1\] (ii) it is sufficient to show that\
$\displaystyle{\bigcap_{(s_1,\cdots,s_n)\in
S}}\Delta_p(s_1,\cdots,s_n)\subseteq D_p(\Omega)$.
Let $(a_1,\cdots,a_n)\in \displaystyle{\bigcap_{(s_1,\cdots,s_n)\in
S}}\Delta_p(s_1,\cdots,s_n)$ and $(t_1,\cdots,t_n)\in\Omega$ be arbitrary. We can easily find $(r_1,\cdots,r_n)\in S$ such that $r_i\leq t_i \ (1\leq i\leq n)$ whence $a_1r_1^p+\cdots+a_nr_n^p\leq
a_1t_1^p+\cdots+a_nt_n^p$.\
Also from $(a_1,\cdots,a_n)\in \displaystyle{\bigcap_{(s_1,\cdots,s_n)\in
S}}\Delta_p(s_1,\cdots,s_n)$ we get $a_1r_1^p+\cdots+a_nr_n^p\geq1$.\
Thus $a_1t_1^p+\cdots+a_nt_n^p\geq1$ whence $(a_1,\cdots,a_n)\in
D_p(\Omega)$.\
(ii) By (i) and Lemma \[lemma1\] (ii) it is sufficient to show that $$D_p(\Omega)\subseteq \displaystyle{\bigcap_{(s_1,\cdots,s_n)\in
S}}\Delta_p(s_1,\cdots,s_n)\,.$$ Let $(a_1,\cdots,a_n)\in
D_p(\Omega)$ and $(s_1,\cdots,s_n)\in S$. There exists a sequence $\{(t_{1,m},\cdots,t_{n,m})\}_{m=1}^{\infty}$ in $\Omega$ satisfying $(t_{1,m},\cdots,t_{n,m})\longrightarrow(s_1,\cdots,s_n)$ as $m\longrightarrow\infty$, or equivalently $t_{i,m}\longrightarrow
s_i$ as $m\longrightarrow\infty\,\,(1\leq i\leq n)$, since $S\subseteq\overline{\Omega}$.\
Now $(a_1,\cdots,a_n)\in D_p(\Omega)$ implies that $a_1t_{1,m}^p+\cdots+a_nt_{n,m}^p\geq1\ (m\in\mathbb{N})$. Getting limit as $m\longrightarrow\infty$, it follows that $a_1s_1^p+\cdots+a_ns_n^p\geq1$.\
Hence $(a_1,\cdots,a_n)\in\displaystyle{\bigcap_{(s_1,\cdots,s_n)\in
S}}\Delta_p(s_1,\cdots,s_n)$.
Next, we identify $D_p(\Omega)$ when $0<p\leq1$.
\[lemma3\] Let $0<p\leq1,\ \Omega\subseteq\{(s_1,\cdots,s_n)\colon
s_1,\cdots,s_n\geq0,\ \sum_{i=1}^{n}s_i\geq1\}$ and $D_p(\Omega)$ be as in Lemma \[lemma2\]. Then the following assertions hold:\
[(i)]{} $\{(a_1,\cdots,a_n)\colon a_1\geq1,\cdots,a_n\geq1\}\subseteq D_p(\Omega)$;\
[(ii)]{} If $\{e_1,\cdots,e_n\}\subseteq\overline{\Omega}$ where $\{e_1,\cdots,e_n\}$ is the standard basis of $\mathbb{R}^n$, then\
$D_p(\Omega)=\{(a_1,\cdots,a_n)\colon a_1\geq1,\cdots,a_n\geq1\}.$
\(i) Let $a_1\geq1,\cdots,a_n\geq1$ and $(s_1,\cdots,s_n)\in\Omega$ be arbitrary. First we show that $s_1^p+\cdots+s_n^p\geq1$.\
Case 1. Let $s_i\leq1$ for all $i=1,\cdots,n$.\
It follows that $s_i\leq s_i^p$ for all $i=1,\cdots,n$, whence $s_1+\cdots+s_n\leq s_1^p+\cdots+s_n^p$. Also from $(s_1,\cdots,s_n)\in\Omega$ we get $s_1+\cdots+s_n\geq1$. Thus $s_1^p+\cdots+s_n^p\geq1$.\
Case 2. Let $s_j>1$ for some $1 \leq j\leq n$.\
We can easily obtain that $s_1^p+\cdots+s_n^p\geq s_j^p>1$.
Now we observe that $a_1s_1^p+\cdots+a_ns_n^p\geq
s_1^p+\cdots+s_n^p\geq1$. Hence $(a_1,\cdots,a_n)\in D_p(\Omega)$.\
(ii) By (i) it suffices to show that $D_p(\Omega)\subseteq\{(a_1,\cdots,a_n)\colon
a_1\geq1,\cdots,a_n\geq1\}\,.$ Let $(a_1,\cdots,a_n)\in D_p(\Omega)$ and $k\in\{1,\cdots,n\}$. There exists a sequence $\{(s_{1,m},\cdots,s_{n,m})\}_{m=1}^{\infty}$ in $\Omega$ satisfying $(s_{1,m},\cdots,s_{n,m})\longrightarrow e_k$ as $m\longrightarrow\infty$, or equivalently $s_{k,m}\longrightarrow1$ and $s_{l,m}\longrightarrow0$ as $m\longrightarrow\infty$ $(l\in\{1,\cdots,n\}\setminus\{k\})$, since $\{e_1,\cdots,e_n\}\subseteq\overline{\Omega}$. Also from $(a_1,\cdots,a_n)\in D_p(\Omega)$ we get $a_1s_{1,m}^p+\cdots+a_ns_{n,m}^p\geq1\ \ (m\in\mathbb{N})$. Taking limit as $m\longrightarrow\infty$, it follows that $a_k\geq1$.
Let $(X,\|\cdot\|)$ be a normed space and $p>0$. Our main aim is to a characterize all $n$-tuples $(\mu_1,\cdots,\mu_n)\in\mathbb{R}^{n}$ satisfying $$\begin{aligned}
\label{L6}
\|x_1+\cdots+x_n\|^p\leq\frac{\|x_1\|^p}{\mu_1}+\cdots+\frac{\|x_n\|^p}{\mu_n}\
\ (x_1,\cdots,x_n\in X)\end{aligned}$$ or its reverse $$\begin{aligned}
\label{L6.5}
\|x_1+\cdots+x_n\|^p\geq\frac{\|x_1\|^p}{\mu_1}+\cdots+\frac{\|x_n\|^p}{\mu_n}\
\ (x_1,\cdots,x_n\in X)\,.\end{aligned}$$ We put $$F(p)=\left\{(\mu_1,\cdots,\mu_n)\in\mathbb{R}^{n}\colon\ \ \left\|\sum_{i=1}^nx_i\right\|^p
\leq\sum_{i=1}^{n}\frac{\|x_i\|^p}{\mu_i}\ {\rm for\ all}\
x_1,\cdots,x_n\in
X\right\}$$and $$G(p)=\left\{(\mu_1,\cdots,\mu_n)\in\mathbb{R}^{n}\colon\
\ \left\|\sum_{i=1}^nx_i\right\|^p \geq\sum_{i=1}^{n}\frac{\|x_i\|^p}{\mu_i}\
{\rm for\ all}\ x_1,\cdots,x_n\in X\right\}\,.$$ Also for each $k=0,1,\cdots,n$ we correspond $F(p\ ;k)$ ($G(p\ ;k)$, resp.) as the subset of $F(p)$ ($G(p)$, resp.) consisting of all $n$-tuples $(\mu_1,\cdots,\mu_n)\in\mathbb{R}^n$ for which inequality (, resp.) holds and exactly $k$ numbers of $\mu_1,\cdots,\mu_n$ are negative. We note that $$\begin{aligned}
\label{L14}
F(p)=\bigcup_{k=0}^nF(p\ ;k)\end{aligned}$$ and $$\begin{aligned}
\label{L14.5}
G(p)=\bigcup_{k=0}^nG(p\ ;k)\,.\end{aligned}$$
In the next two theorems we characterize $F(p)$. First we consider the case where $p>1$.
\[T1\] Let $(X,\|\cdot\|)$ be a normed space and $p>1$. Then the following assertions hold:\
[(i)]{} $F(p\
;0)=\left\{(\mu_1,\cdots,\mu_n)\colon\mu_1,\cdots,\mu_n>0\
and\ \sum_{i=1}^n\mu_i^{\frac{1}{p-1}}\leq 1\right\}\,;$\
[(ii)]{} $F(p\ ;k)=\varnothing$, for all $k=1,\cdots,n$;\
[(iii)]{} $F(p)=F(p\ ;0)\,.$
Let $\mu_1,\cdots,\mu_n$ be arbitrary positive numbers for which $$\begin{aligned}
\|x_1+\cdots+x_n\|^p\leq\frac{\|x_1\|^p}{\mu_1}+\cdots+\frac{\|x_n\|^p}{\mu_n}\
\ (x_1,\cdots,x_n\in X)\,.\end{aligned}$$ Putting $x_i=\mu_i^\frac{1}{p-1}x$ (for some $x\neq0$ and for all $i=1,\cdots,n$) we get $$\begin{aligned}
\left\|\mu_1^\frac{1}{p-1}x+\cdots+\mu_n^\frac{1}{p-1}x\right\|^p\leq\frac{\left\|\mu_1^\frac{1}{p-1}x\right\|^p}{\mu_1}+\cdots+\frac{\left\|\mu_n^\frac{1}{p-1}x\right\|^p}{\mu_n}\,,\end{aligned}$$ or equivalently we obtain $\mu_1^\frac{1}{p-1}+\cdots+\mu_n^\frac{1}{p-1}\leq1$ because of $p>1$.\
Conversely, if $\mu_1,\cdots,\mu_n>0$ and $\mu_1^\frac{1}{p-1}+\cdots+\mu_n^\frac{1}{p-1}\leq1$, then the desired inequality is deduced from following inequalities: $$\begin{aligned}
\|x_1+\cdots+x_n\|^p&\leq&(\|x_1\|+\cdots+\|x_n\|)^p\\
&=&\left(\mu_1^\frac{1}{p}\left\|\frac{x_1}{\mu_1^\frac{1}{p}}\right\|+\cdots+\mu_n^\frac{1}{p}\left\|\frac{x_n}{\mu_n^\frac{1}{p}}\right\|\right)^p\\
&\leq&\left(\mu_1^\frac{1}{p-1}+\cdots+\mu_n^\frac{1}{p-1}\right)^{p-1}\left(\frac{\|x_1\|^p}{\mu_1}+\cdots+\frac{\|x_n\|^p}{\mu_n}\right)\\
&\leq&\frac{\|x_1\|^p}{\mu_1}+\cdots+\frac{\|x_n\|^p}{\mu_n}\,.\end{aligned}$$ We note that the second inequality follows from the well-known Hölder inequality.\
(ii) Let $(\mu_1,\cdots,\mu_n)\in F(p\ ;k)$ for some $k=1,\cdots,n$. There exists $1\leq j\leq n$ such that $\mu_j<0$ and inequality holds.\
Putting $x_i=0\ \left(i\in\{1,\cdots,n\}\setminus\{j\}\right)$ and $x_j\neq0$ in inequality we obtain $\|x_j\|^p\leq\frac{\|x_j\|^p}{\mu_j}$. This is a contradiction since $\mu_j<0$.\
(iii) It follows from (i), (ii) and .
\[T3\] Let $(X,\|\cdot\|)$ be a normed space and $0<p\leq1$. Then the following assertions are valid:\
[(i)]{} $F(p\ ;0)=(0,1]\times\cdots\times(0,1]$;\
[(ii)]{} $F(p\ ;k)=\varnothing$, for all $k=1,\cdots,n$;\
[(iii)]{} $F(p)=F(p\ ;0)\,.$
\(i) Let $\mu_1,\cdots,\mu_n$ be arbitrary positive numbers. We observe that inequality holds if and only if $$\begin{aligned}
\label{I1}
\sum_{i=1}^n\frac{\|x_i\|^p}{\mu_i\|x_1+\cdots+x_n\|^p}\geq1\end{aligned}$$ for all $x_1,\cdots,x_n\in X$ for which $\sum_{i=1}^nx_i\neq0$. Putting $$\Omega=\left\{\left(\frac{\|x_1\|}{\|\sum_{i=1}^nx_i\|},\cdots,\frac{\|x_n\|}{\|\sum_{i=1}^nx_i\|}\right)\colon
x_1,\cdots,x_n\in X,\ \sum_{i=1}^nx_i\neq0 \right\}$$ we get $\Omega\subseteq\big\{(s_1,\cdots,s_n)\colon s_1,\cdots,s_n\geq0,\
\sum_{i=1}^ns_i\geq1\big\}$, $\{e_1,\cdots,e_n\}\subseteq\Omega\subseteq\overline{\Omega}$ and inequality turns into $\sum_{i=1}^n\frac{s_i^p}{\mu_i}\geq1$ for all $(s_1,\cdots,s_n)\in\Omega$, or equivalently $(\frac{1}{\mu_1},\cdots,\frac{1}{\mu_n})\in D_p(\Omega)$. From Lemma \[lemma3\] we deduce that $\mu_i\leq1$ for all $i=1,\cdots,n$.\
(ii) It is similar to the proof of Theorem \[T1\] (ii).\
(iii) It follows from (i), (ii) and .
Now we want to characterize $G(p)$ for any $p>0$. The next theorem deals with the case where $p>1$.
\[T2\] Let $(X,\|\cdot\|)$ be a normed space and $p>1$. Then the following hold:\
[(i)]{} $G(p\ ;k)=\varnothing$, for all $k=0,\cdots,n-2$;\
[(ii)]{} $G(p\ ;n-1)\\=\left\{(\mu_1,\cdots,\mu_n)\colon\exists\
j=1,\cdots,n\ ; \mu_j>0,\ \mu_i<0\ (i\neq j)\ and\ \mu_j^\frac{1}{p-1}\geq1+\sum_{i=1,i\neq j}^n|\mu_i|^\frac{1}{p-1}\right\};$\
[(iii)]{} $G(p\ ;n)=\mathbb{R}^-\times\cdots\times\mathbb{R}^-$, where $\mathbb{R}^-=\{t\in\mathbb{R}: t <0\};$\
[(iv)]{} $G(p)=G(p\ ;n-1)\cup G(p\ ;n)\,.$
\(i) Let $(\mu_1,\cdots,\mu_n)\in G(p\ ;k)$ for some $k=0,\cdots,n-2$. Thus there exist $j_1,j_2\in\{1,\cdots,n\}$ such that $\mu_{j_1},\mu_{j_2}>0\ (j_1\neq j_2)$ and also inequality holds. Setting $x_i=0\
(i\in\{1,\cdots,n\}\setminus\{j_1,j_2\})$ in inequality we have $$\begin{aligned}
\label{L15}
\|x_{j_1}+x_{j_2}\|^p\geq\frac{\|x_{j_1}\|^p}{\mu_{j_1}}+\frac{\|x_{j_2}\|^p}{\mu_{j_2}}\,.\end{aligned}$$ Also we can consider $x_{j_1},x_{j_2}\in X$ for which $x_{j_1}+x_{j_2}=0$ and $(x_{j_1},x_{j_2})\neq(0,0)$. Thus inequality implies that $\frac{\|x_{j_1}\|^p}{\mu_{j_1}}+\frac{\|x_{j_2}\|^p}{\mu_{j_2}}\leq0$, which is a contradiction.\
(ii) Let $(\mu_1,\cdots,\mu_n)\in\mathbb{R}^n$ satisfy $\mu_j>0$ for some $j=1,\cdots,n$ and $\mu_i<0$ for all $i\in\{1,\cdots,n\}\setminus\{j\}$.\
We observe that inequality holds if and only if $$\begin{aligned}
\label{L9}
\sum_{i=1,i\neq
j}^n\frac{\mu_j\|x_i\|^p}{|\mu_i|\|x_j\|^p}+\frac{\mu_j\|x_1+\cdots+x_n\|^p}{\|x_j\|^p}\geq1\end{aligned}$$ for all $x_1,\cdots,x_n\in X, x_j\neq0$.\
Putting $\Omega=\left\{\left(\frac{\|x_1\|}{\|x_j\|},\cdots,\frac{\|x_{j-1}\|}{\|x_j\|},\frac{\|x_{j+1}\|}{\|x_j\|},\cdots,\frac{\|x_n\|}{\|x_j\|},\frac{\|x_1+\cdots+x_n\|}{\|x_j\|}\right)\in\mathbb{R}^n\colon
x_1,\cdots,x_n\in X, x_j\neq0\right\}$\
we have $\Omega\subseteq\big\{(s_1,\cdots,s_n)\colon s_1,\cdots,s_n\geq0,\
\sum_{i=1}^ns_i\geq1\big\}$, $S\subseteq\Omega\subseteq\overline{\Omega}$ and inequality turns into $$\frac{\mu_j}{|\mu_1|}s_1^p+\cdots+\frac{\mu_j}{|\mu_{j-1}|}s_{j-1}^p+\frac{\mu_j}{|\mu_{j+1}|}s_j^p+\cdots+\frac{\mu_j}{|\mu_n|}s_{n-1}^p+\mu_js_n^p\geq1$$ for all $(s_1,\cdots,s_n)\in\Omega$, or equivalently $$\begin{aligned}
\label{L10}
\left(\frac{\mu_j}{|\mu_{1}|},\cdots,\frac{\mu_j}{|\mu_{j-1}|},\frac{\mu_j}{|\mu_{j+1}|},\cdots,\frac{\mu_j}{|\mu_{n}|},\mu_j\right)\in
D_p(\Omega)\,.\end{aligned}$$ From Lemma \[lemma2\] and we deduce that $$\mu_j\geq
h_p\left(\frac{\mu_j}{|\mu_{1}|},\cdots,\frac{\mu_j}{|\mu_{j-1}|},\frac{\mu_j}{|\mu_{j+1}|},\cdots,\frac{\mu_j}{|\mu_{n}|}\right)\,,$$ that is $$\begin{aligned}
\label{L11}
\mu_j\geq\left(1-\sum_{i=1,i\neq
j}^n\frac{|\mu_i|^\frac{1}{p-1}}{\mu_j^\frac{1}{p-1}}\right)^{1-p}\,.\end{aligned}$$ By a straightforward calculation the following inequality follows from . $$\mu_j^\frac{1}{p-1}\geq1+\sum_{i=1,i\neq
j}^n|\mu_i|^\frac{1}{p-1}\,.$$ (iii) It is trivial.\
(iv) It easily follows from (i), (ii), (iii) and .
\[T4\] Let $(X,\|\cdot\|)$ be a normed space and $0<p\leq1$. Then the following hold.\
[(i)]{} $G(p\ ;k)=\varnothing$, for all $k=0,\cdots,n-2$;\
[(ii)]{} $G(p\ ;n-1)=\\
\left\{(\mu_1,\cdots,\mu_n)\colon\exists\ j=1,\cdots,n;\ \mu_j>0,\ \mu_i<0\ (i\neq j)\ and\ \mu_j\geq \displaystyle{\max_{i\in\{1,\cdots,n\}\setminus\{j\}}}\{1,|\mu_i|\}\right\};$\
[(iii)]{} $G(p\ ;n)=\mathbb{R}^-\times\cdots\times\mathbb{R}^-;$\
[(iv)]{} $G(p)=G(p\ ;n-1)\cup G(p\ ;n)\,.$
\(i) It is similar to the proof of Theorem \[T2\] (i).\
(ii) Let $(\mu_1,\cdots,\mu_n)\in\mathbb{R}^n$ satisfying $\mu_j>0$ for some $j=1,\cdots,n$ and $\mu_i<0$ for all $i\in\{1,\cdots,n\}\setminus\{j\}$.\
Putting $\Omega=\left\{\left(\frac{\|x_1\|}{\|x_j\|},\cdots,\frac{\|x_{j-1}\|}{\|x_j\|},\frac{\|x_{j+1}\|}{\|x_j\|},\cdots,\frac{\|x_n\|}{\|x_j\|},\frac{\|x_1+\cdots+x_n\|}{\|x_j\|}\right)\in\mathbb{R}^n\colon
x_1,\cdots,x_n\in X, x_j\neq0\right\}$\
we have $\Omega\subseteq\left\{(s_1,\cdots,s_n)\colon s_1,\cdots,s_n\geq0,\
\sum_{i=1}^ns_i\geq1\right\}$ and $\{e_1,\cdots,e_n\}\subseteq\Omega\subseteq\overline{\Omega}$. Passing the proof of Theorem \[T2\] (ii) we observe that $\left\|\sum_{i=1}^nx_i\right\|^p\geq\sum_{i=1}^{n}\frac{\|x_i\|^p}{\mu_i}$ for all $x_1,\cdots,x_n\in X$ if and only if $$\left(\frac{\mu_j}{|\mu_{1}|},\cdots,\frac{\mu_j}{|\mu_{j-1}|},\frac{\mu_j}{|\mu_{j+1}|},\cdots,\frac{\mu_j}{|\mu_{n}|},\mu_j\right)\in
D_p(\Omega)\,.$$ From Lemma \[lemma3\] we deduce that $$\mu_j\geq1\ \ \ {\rm and}\ \ \mu_j\geq|\mu_i|\ \ (i\in\{1,\cdots,n\}\setminus\{j\}).$$ or equivalently we get $\mu_j\geq
\displaystyle{\max_{i\in\{1,\cdots,n\}\setminus\{j\}}}\{1,|\mu_i|\}$.\
(iii) It is trivial.\
(iv) It follows from (i), (ii), (iii) and .
Suppose that $(X,\|\cdot\|)$ is a normed space and $p>0$. Let $$H(p):=\left\{(\mu_1,\cdots,\mu_n)\in\mathbb{R}^{n}\colon\ \left\|\sum_{i=1}^nx_i\right\|^p
\leq\left|\sum_{i=1}^{n}\frac{\|x_i\|^p}{\mu_i}\right|\ {\rm for\
all}\ x_1,\cdots,x_n\in X\right\}\,.$$ Then the following assertions hold:\
[(i)]{} If $p>1$, then $$\begin{aligned}
H(p)&=&F(p)\cap-F(P)\\&=&\left\{(\mu_1,\cdots,\mu_n)\colon \ \mbox{either}\ \mu_1,\cdots,\mu_n>0 \ \mbox{or}\
\mu_1,\cdots,\mu_n<0\ \mbox{as well as}\
\sum_{i=1}^n|\mu_i|^{\frac{1}{p-1}}\leq 1\right\}\,;\end{aligned}$$ [(ii)]{} If $0<p\leq1$, then $H(p)=(0,1]^n\cup[-1,0)^n$.
One can easily observe that $$H(p)=(F(p)\cap-G(p))\cup(-F(p)\cap G(p)).$$ Now (i) follows from Theorems \[T1\] and \[T2\] and (ii) follows from Theorems \[T3\] and \[T4\].
The special cases of our results in the case where $n=2$ and $p \geq 1$ give rise to some of main results of Takahasi et al [@Takahasi-Ras-Sai-Takahashi Theorems 1.1 and 4.1].
[99]{}
A.H. Ansari, M.S. Moslehian, More on reverse triangle inequality in inner product spaces, Intern. J. Math. Math. Sci. 18 (2005) 2883–2893.
S.S. Dragomir, Reverses of the triangle inequality in Banach spaces, J. Inequal. Pure Appl. Math. 6(5) (2005) Art. 129.
M. Khosravi, H. Mahyar, M.S. Moslehian, Reverse triangle inequality in Hilbert $C^*$-modules, J. Inequal. Pure Appl. Math. 10(4) (2009) Art. 110.
L. Maligranda, Some remarks on the triangle inequality for norms, Banach J. Math. Anal. 2(2) (2008) 31-–41.
K.-I. Mitani, K.-S. Saito, M. Kato, T. Tamura, On sharp triangle inequalities in Banach spaces, J. Math. Anal. Appl. 336 (2007) 1178-–1186.
J.E. Pečarić, R. Rajić, The Dunkl–Williams inequality with $n$-elements in normed linear spaces, Math. Inequal. Appl. 10 (2007) 461–470.
M. Kato, K.-S. Saito, T. Tamura, Sharp triangle inequality and its reverse in Banach spaces, Math. Inequal. Appl. 10 (2007) 451–460.
C.-Y. Hsu, Sen-Yen Shaw, H.-J. Wong Refinements of generalized triangle inequalities, J. Math. Anal. Appl. 344 (2008) 17–31.
H. Hudzik, T. R. Landes, Characteristic of convexity of K¨othe function spaces, Math. Ann. 294 (1992) 117-–124.
H. Belbachir, M. Mirzavaziri, M.S. Moslehian, q-norms are really norms, Austral. J. Math. Anal. Appl. 3 (2006) Article 2, 1–3.
S. Saitoh, Generalization of the triangle inequality, J. Inequal. Pure Appl. Math. 4(3) (2003) Art. 62.
S.-E. Takahasi, J.M. Rassias, S. Saitoh, Y. Takahashi, Refined generalizations of the triangle inequality on Banach space, Math. Inequal. Appl. 13 (2010) 733–741.
J.M. Rassias, Solutions of the Ulam stability problem for Euler–Lagrange quadratic mappings, J. Math. Anal. Appl. 220 (1998) 613–639.
H. Takagi, T. Miura, T. Hayata, S.-E. Takahasi, A reconsideration of Hua’s inequality II, J. Inequal. Appl. (2006) Art. ID 21540.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A central question for active learning (AL) is: “what is the optimal selection?” Defining optimality by classifier loss produces a new characterisation of optimal AL behaviour, by treating expected loss reduction as a statistical target for estimation.
This target forms the basis of *model retraining improvement* (MRI), a novel approach providing a statistical estimation framework for AL. This framework is constructed to address the central question of AL optimality, and to motivate the design of estimation algorithms.
MRI allows the exploration of optimal AL behaviour, and the examination of AL heuristics, showing precisely how they make sub-optimal selections. The abstract formulation of MRI is used to provide a new guarantee for AL, that an unbiased MRI estimator should outperform random selection. This MRI framework reveals intricate estimation issues that in turn motivate the construction of new statistical AL algorithms. One new algorithm in particular performs strongly in a large-scale experimental study, compared to standard AL methods. This competitive performance suggests that practical efforts to minimise estimation bias may be important for AL applications.
author:
- |
Lewis P. G. Evans [email protected]\
Department of Mathematics\
Imperial College London\
London, SW7 2AZ, United Kingdom Niall M. Adams [email protected]\
Department of Mathematics\
Imperial College London\
London, SW7 2AZ, United Kingdom\
Heilbronn Institute for Mathematical Research\
University of Bristol\
PO Box 2495, Bristol, BS8 9AG, United Kingdom Christoforos Anagnostopoulos [email protected]\
Department of Mathematics\
Imperial College London\
London, SW7 2AZ, United Kingdom
bibliography:
- 'Shared.bib'
title: Estimating Optimal Active Learning via Model Retraining Improvement
---
active learning, model retraining improvement, estimation framework, expected loss reduction, classification
Introduction {#section:Introduction}
============
Classification is a central task in statistical inference and machine learning. In certain cases unlabelled data is plentiful, and a subset can be queried for labelling. Active learning (AL) seeks to intelligently select this subset of unlabelled examples, to improve a base classifier. Examples include medical image diagnosis and document categorisation [@Dasgupta2008; @Hoi2006]. Many AL methods are heuristic, alongside a few theoretical approaches reviewed by @Settles2009 [@Olsson2009]. AL method performance is often assessed by large-scale experimental studies such as @Guyon2011 [@Kumar2010; @Evans2013].
A prototypical AL scenario consists of a classification problem and a classifier trained on a small labelled dataset. The classifier may be improved by retraining with further examples, systematically selected from a large unlabelled pool. This formulation of AL raises the central question for AL, “what is the optimal selection?” Performance in classification is judged by loss functions such as those described in @Hand1997, suggesting that optimality in AL selection should be characterised in terms of classifier loss. This suggests that the optimal selection should be defined as the example that maximises the expected loss reduction. This statistical quantity forms the basis of model retraining improvement (MRI), a novel statistical framework for AL. Compared to heuristic methods, a statistical approach provides strong advantages, both theoretical and practical, described below. This MRI estimation framework addresses the central question by formally defining optimal AL behaviour. Creating a mathematical abstraction of optimal AL behaviour allows reasoning about heuristics, e.g. showing precisely how they make sub-optimal choices in particular contexts. Within this framework, an ideal unbiased MRI estimator is shown to have the property of outperforming random selection, which is a new guarantee for AL. Crucially, MRI motivates the development of novel algorithms that perform strongly compared to standard AL methods. MRI estimation requires a series of steps, which are subject to different types of estimation problem. Algorithms are constructed to approximate MRI, taking different estimation approaches. A large-scale experimental study evaluates the performance of the two new MRI estimation algorithms, alongside standard AL methods. The study explores many sources of variation: classifiers, AL algorithms, with real and abstract classification problems (both binary and multi-class). The results show that the MRI-motivated algorithms perform competitively in comparison to standard AL methods.
This work is structured as follows: first the background of classification and AL are described in Section \[section:Background\]. Section \[section:Example Quality\] defines MRI, illustrated by an abstract classification problem in Section \[subsection:Theoretical Example to Illustrate Example Quality\]. MRI estimation algorithms are described in Section \[section:Algorithms to Estimate Example Quality\] and evaluated in a large-scale experimental study of Section \[section:Experiments and Results\], followed by concluding remarks.
Background {#section:Background}
==========
The background contexts of classification and AL are described, followed by a brief review of relevant literature, with particular focus on methods that are used later in the paper.
Classification {#subsection:Classification}
--------------
The categorical response variable $Y$ is modelled as a function of the covariates ${\bf X}$. For the response $Y$ there are $k$ classes with class labels $\{c_1, c_2, ..., c_k\}$. Each classification example is denoted $ ( {\bf x}, y) $, where ${\bf x}$ is a $d$-dimensional covariate vector and $y$ is a class label. The class prior is denoted $\boldsymbol\pi$.
The Bayes classifier is an idealisation based on the true distributions of the classes, thereby producing optimal probability estimates, and class allocations given a loss function. Given a covariate vector ${\bf x}$, the Bayes classifier outputs the class probability vector of $Y|{\bf x}$ denoted ${\bf p} = (p_j)_1^k$. A *probabilistic* classifier estimates the class probability vector as ${\bf \hat{p}} = (\hat{p}_j)_1^k$, and allocates ${\bf x}$ to class $\hat{y}$ using decision theoretic arguments, often using a threshold. This allocation function is denoted $h$: $\hat{y} = h({\bf \hat{p}})$. For example, to minimise misclassification error, the most probable class is allocated: $\hat{y} = h ({\bf \hat{p}}) = \operatorname*{arg\,max}_{j} (\hat{p}_j)$. The objective of classification is to learn an allocation rule with good generalisation properties.
A somewhat non-standard notation is required to support this work, which stresses the dependence of the classifier on the training data. A dataset is a set of examples, denoted $D = \{ {\bf x}_i,y_i \}_{i=1}^n$, where $i$ indexes the example. This indexing notation will be useful later. A dataset $D$ may be subdivided into training data $D_T$ and test data $D_E$. This dataset division may be represented by index sets, for example, $T \cup E = \{1, ..., n\}$, showing the data division into training and test subsets. First consider a parametric classifier, for example linear discriminant analysis or logistic regression [@Bishop2007 Chapter 4]. A parametric classifier has estimated parameters $\boldsymbol{\hat{\theta}}$, which can be regarded as a fixed length vector (fixed given $d$ and $k$). These parameters are estimated by model fitting, using the training data: $\boldsymbol{\hat{\theta}} = \theta (D_T)$, where $\theta()$ is the model fitting function. This notation is intended to emphasize the dependence of the estimated parameters $\boldsymbol{\hat{\theta}}$ on the training data $D_T$.
Second, this notation is slightly abused to extend to non-parametric classifiers. The complexity of non-parametric classifiers may increase with sample size, hence they cannot be represented by a fixed length object. In this case $\boldsymbol{\hat{\theta}}$ becomes a variable-length object containing the classifier’s internal data (for example the nodes of a decision tree, or the stored examples of $K$-nearest-neighbours).
While the contents and meaning of $\boldsymbol{\hat{\theta}}$ would be very different, the classifier’s functional roles are identical: model training produces $\boldsymbol{\hat{\theta}}$, which is used to predict class probabilities. This probability prediction is denoted ${\bf \hat{p}} = \phi(\boldsymbol{\hat{\theta}}, \mathbf{x})$. These predictions are in turn used to assess classifier performance.
To consider classifier performance, first assume a fixed training dataset $D_T$. Classifier performance is assessed by a loss function, for example error rate, which quantifies the disagreement between the classifier’s predictions and the truth. The empirical loss for a single example is defined via a loss function $g(y, {\bf \hat{p}})$. Many loss functions focus on the allocated class, for example error rate, $g_e(y, {\bf \hat{p}}) = \mathbbm{1} (y \ne h({\bf \hat{p}}))$. Other loss functions focus on the predicted probability, for example log loss, $g_o({\bf \hat{p}}) = \sum_{j=1}^k ( p_j \textrm{ log } \hat{p}_j )$.
The estimated probabilities ${\bf \hat{p}}$ are highly dependent on the estimated classifier $\boldsymbol{\hat{\theta}}$. To emphasize that dependence, the empirical loss for a single example is denoted $M(\boldsymbol{\hat{\theta}}, {\bf x}, y) = g(y, {\bf \hat{p}})$. For example, error rate empirical loss is denoted $M_e(\boldsymbol{\hat{\theta}}, {\bf x}, y)$.
In classification, generalisation performance is a critical quantity. For this reason, empirical loss is generalised to expected loss, denoted $L(\boldsymbol{\hat{\theta}})$: $$L (\boldsymbol{\hat{\theta}}) = E_{ {\bf X},Y} [M (\boldsymbol{\hat{\theta}}, {\bf x}, y)] = E_{Y|{\bf X}} E_{\bf X} [M (\boldsymbol{\hat{\theta}}, {\bf x}, y)].$$ This expected loss $L$ is defined as an expectation over all possible test data, given a specific training set. The expected error rate and log loss are denoted $L_e$ and $L_o$. Hereafter loss will always refer to the expected loss $L$. The loss $L(\boldsymbol{\hat{\theta}})$ is dependent on the data $D$ used to train the classifier, emphasized by rewriting $L(\boldsymbol{\hat{\theta}})$ as $L(\theta(D))$ since $\boldsymbol{\hat{\theta}} = \theta(D)$.
The change in the loss as the number of labelled examples increases is of great methodological interest. This function is known as the *learning curve*, typically defined as the change of expected loss with the number of examples. Learning curves are illustrated in Figure \[figure:Performance Comparison of AL and RS\], and discussed in @Provost2003 [@Gu2001; @Kadie1995].
Active Learning {#subsection:Active Learning}
---------------
The context for AL is an abundance of unlabelled examples, with labelled data either expensive or scarce. Good introductions to AL are provided by @Dasgupta2011, @Settles2009 and @Olsson2009.
An algorithm can select a few unlabelled examples to obtain their labels from an oracle (for example a human expert). This provides more labelled data which can be included in the training data, potentially improving a classifier. Intuitively some examples may be more informative than others, so systematic example selection should maximise classifier improvement.
In *pool-based* AL, there is an unlabelled pool of data $X_P$ from which examples may be selected for labelling. This pool provides a set of examples for label querying, and also gives further information on the distribution of the covariates. Usually there is also a (relatively small) initial dataset of labelled examples, denoted $D_I$, typically assumed to be iid in AL. This work considers the scenario of pool-based AL.
In AL it is common to examine the learning curve, by repeating the AL selection step many times (*iterated AL*). At each selection step, the loss is recorded, and this generates a set of losses, which define the learning curve for the AL method. Iterated AL allows the exploration of performance over the learning curve, as the amount of labelled data grows. This repeated application of AL selection is common in both applications and experimental studies [@Guyon2011; @Evans2013]. In contrast to iterated AL, the AL selection step may occur just once (*single-step AL*). The question of iterated or single-step AL is critical, because iterated AL inevitably produces covariate bias in the labelled data. The covariate bias from iterated AL creates a selection bias problem, which is intrinsic to AL. At each selection step, an AL method may select a single example from the pool (*individual AL*) or several examples at once (*batch AL*). AL applications are often constrained to use batch AL for pragmatic reasons [@Settles2009].
Turning to AL performance, consider random selection (RS) where examples are chosen randomly (with equal probability) from the pool. By contrast, AL methods select some examples in preference to others. Under RS and AL, the classifier receives exactly the same number of labelled examples; thus RS provides a reasonable benchmark for AL [@Guyon2011; @Evans2013]. The comparison of methods to benchmarks is available in experiments but not in real AL applications [@Provost2010].
Classifier performance should improve, at least on average, even under the benchmark RS, since the classifier receives more training data (an issue explored below). AL performance assessment should consider how much AL outperforms RS. Hence AL performance addresses the relative improvement of AL over RS, and the relative ranks of AL methods, rather than the absolute level of classifier performance. Figure \[figure:Performance Comparison of AL and RS\] shows the losses of AL and RS as the number of labelled examples increases.
![Performance comparison of active learning and random selection, showing that a classifier often improves faster under AL than under RS. In both cases the loss decreases as the number of labelled examples increases; however, AL improves faster than RS. These curves are smoothed averages from multiple experiments. The black vertical line illustrates the *fixed-label comparison*, whereas the blue horizontal line shows the *fixed-loss comparison* (see Section \[subsection:Active Learning\]). The classification problem is “Abalone” from UCI, a three-class problem, using classifier $5$-nn, and Shannon entropy as the AL method. []{data-label="figure:Performance Comparison of AL and RS"}](AL_RS_PerfComp1.pdf)
Figure \[figure:Performance Comparison of AL and RS\] shows two different senses in which AL outperforms RS: first AL achieves better loss reduction for the same number of labels (*fixed-label comparison*), and second AL needs fewer labels to reach the same classifier performance (*fixed-loss comparison*). Together the fixed-label comparison and fixed-loss comparison form the two fundamental aspects of AL performance. The fixed-label comparison first fixes the number of labels, then seeks to minimise loss. Several established performance metrics focus on the fixed-label comparison: AUA, ALC and WI [@Guyon2011; @Evans2013]. The fixed-label comparison is more common in applications where the costs of labelling are significant [@Settles2009].
Under the fixed-loss comparison, the desired level of classifier loss is fixed, the goal being to minimise the number of labels needed to reach that level. Label complexity is the classic example, where the desired loss level is a fixed ratio of asymptotic classifier performance [@Dasgupta2011]. Label complexity is often used as a performance metric in contexts where certain assumptions permit analytically tractable results, for example @Dasgupta2011.
Overview of Active Learning Methods {#subsection:Literature Review}
-----------------------------------
A popular AL approach is the uncertainty sampling heuristic, where examples are chosen with the greatest class uncertainty [@Thrun1992; @Settles2009]. This approach selects examples of the greatest classifier uncertainty in terms of class membership probability. The idea is that these uncertain examples will be the most useful for tuning the classifier’s decision boundary. Example methods include Shannon entropy (SE), least confidence and maximum uncertainty. For a single unlabelled example [**x**]{}, least confidence is defined as $ U_L ({\bf x}, \theta(D)) = 1 - \hat{p}(\hat{y} | {\bf x})$, where $\hat{p}(\hat{y} | {\bf x})$ is the classifier’s estimated probability of the allocated class $\hat{y}$. Shannon entropy is defined as $ U_E ({\bf x}, \theta(D)) = \sum_{j=1}^k \hat{p}_j \, \textrm{log} (\hat{p}_j)$. The uncertainty sampling approach is popular and efficient, but lacks theoretical justification.
Version space search is a theoretical approach to AL, where the version space is the set of hypotheses (classifiers) that are consistent with the data [@Mitchell1997; @Dasgupta2011]. Learning is then interpreted as a search through version space for the optimal hypothesis. The central idea is that AL can search this version space more efficiently than RS.
Query by committee (QBC) is a heuristic approximation to version space search [@Seung1992]. Here a committee of classifiers is trained on the labelled data, which then selects the unlabelled examples where the committee’s predictions disagree the most. This prediction disagreement may focus on predicted classes (for example vote entropy) or predicted class probabilities (for example average Kullback-Leibler divergence); see @Olsson2009. These widely used versions of QBC are denoted QbcV and QbcA. A critical choice for QBC is the classifier committee, which lacks theoretical guidance. In this sense version space search leaves the optimal AL selection unspecified.
Another approach to AL is exploitation of cluster structure in the pool. Elucidating the cluster structure of the pool could provide valuable insights for example selection. @Dasgupta2011 gives a motivating example: if the pool clusters neatly into $b$ class-pure clusters where $b=k$, then $b$ labels could suffice to build an optimal classifier. This very optimistic example does illustrate the potential gain.
A third theoretical approach, notionally close to our contribution, is error reduction, introduced in @Roy2001. This approach minimises the loss of the retrained classifier, which is the loss of the classifier which has been retrained on the selected example. Roy and McCallum consider two loss functions, error rate and log loss, to construct two quantities, which are referred to here as expected future error (EFE) and expected future log loss (EFLL). Those authors focus on methods to estimate EFE and EFLL, before examining the experimental performance of their estimators.
Given a classifier fitting function $\theta$, labelled data $D$ and a single unlabelled example ${\bf x}$, EFE is defined as $$EFE({\bf x}, \theta, D) = - E_{Y | {\bf x}} [ L_e(\theta(D \cup ({\bf x},Y)) ] = - \sum_{j=1}^k \{ p_j \, L_e(\theta(D \cup ({\bf x},c_j)) \},$$ where $L_e$ is error rate (see Section \[subsection:Classification\]). EFLL is defined similarly to EFE, with log loss $L_o$ replacing error rate $L_e$. Both of these quantities average over the unobserved label $Y|{\bf x}$.
Roy and McCallum define an algorithm to calculate EFE, denoted EfeLc, which approximates the loss using the unlabelled pool for efficiency. Specifically it approximates error rate $L_e$ by the total least confidence over the entire pool: $$L_e(\theta(D)) \approx \sum_{ {\bf x}_i \in X_P} U_L({\bf x}_i, \theta(D)),$$ where $X_P$ are the unlabelled examples in the pool. The uncertainty function $U_L$ is intended to capture the class uncertainty of an unlabelled example.
Roy and McCallum propose the following approximation for the value of EFE by calculating $$\label{eq:efelc}
\begin{split}
f_1({\bf x}, \theta, D) = - \sum_{j=1}^k \left\{ \hat{p}_j \sum_{ {\bf x}_i \in X_P} U_L({\bf x}_i, \theta(D \cup ({\bf x}_i,c_j))) \right\}
= - \sum_{j=1}^k \left\{ \hat{p}_j \sum_{ {\bf x}_i \in X_P} \left( 1 - \hat{p}(\hat{y_i} | {\bf x}_i) \right) \right\}.
\end{split}$$ Here $\hat{p}_j$ is the current classifier’s estimate of the class probability for class $j$, while $\hat{y}_i$ is the predicted label for $\mathbf{x}_i$ after a training update with the example $(\mathbf{x}, c_j)$. Note that EfeLc uses the the classifier’s posterior estimates *after* an update (to estimate the loss), whereas the uncertainty sampling approaches use the *current* classifier’s posterior estimates (to assess uncertainty).
This approximation of $L_e$ by the total least confidence over the pool is potentially problematic. It is easy to construct cases (for example an extreme outlier) where a labelled example would reduce a classifier’s uncertainty, but also increase the overall error; such examples call into question the approximation of error by uncertainty. In the absence of further assumptions or motivation, it is hard to anticipate the statistical properties of $f_1$ in Equation \[eq:efelc\] as an estimator. Further, EfeLc uses the same data to train the classifier and to estimate the class probabilities, thereby risking bias in the estimator (an issue explored further in Section \[section:Algorithms to Estimate Example Quality\]).
The error reduction approach is similar in spirit to MRI, since the optimal example selection is first considered, and then specified in terms of classifier loss. In that sense, the quantity EFE is a valuable precursor to model retraining improvement, which is defined later in Equation \[eq:eqc\]. However EFE omits the loss of the current classifier, which proves important when examining improvement (see Section \[subsection:Theoretical Example to Illustrate Example Quality\]). Further, EFE is only defined for individual AL, while MRI defines targets for both batch and individual AL.
The estimation of a statistical quantity, consisting of multiple components, raises several statistical choices, in terms of component estimators and how to use the data. These choices are described and explored in Section \[section:Algorithms to Estimate Example Quality\], whereas Roy and McCallum omit these choices, providing just a single algorithmic approach. In that sense, Roy and McCallum do not use EFE to construct an estimation framework for algorithms. Nor do Roy and McCallum use EFE to examine optimal AL behaviour, or compare it to the behaviour of known AL methods; Section \[subsection:Theoretical Example to Illustrate Example Quality\] provides such an examination and comparison using MRI. Finally, the EFE algorithms do not show strong performance in the experimental results of Section \[section:Experiments and Results\]. The current literature does not provide a statistical estimation framework for AL; MRI addresses this directly in Section \[section:Example Quality\].
Model Retraining Improvement {#section:Example Quality}
============================
Here the statistical target, model retraining improvement, is defined and motivated as an estimation target, both theoretically and for applications. This further lays the groundwork for MRI as a statistical estimation framework for AL. This Section defines the statistical target as an expectation, while Section \[section:Algorithms to Estimate Example Quality\] describes estimation problems, and algorithms for applications.
The Definition of Model Retraining Improvement {#subsection:The Definition of Example Quality}
----------------------------------------------
[|c|c|]{}\
Symbol & Description\
& Underlying distribution of the classification problem\
& Bayes class probability vector, for covariate $\mathbf{x}$: $\mathbf{p} = p(Y|\mathbf{x}) = \{ p(c_j|\mathbf{x}) \}_{j=1}^k$\
& Classifier training function\
& Classifier estimated parameters, where $\boldsymbol{\hat{\theta}} = \theta(D_T)$\
& Classifier prediction function; class probability vector $\mathbf{\hat{p}} = \phi(\boldsymbol{\hat{\theta}}, \mathbf{x})$\
& The labelled data: $D_S = (X_S, Y_S) = \{ \mathbf{x}_i, y_i \}^{i \in S}$\
& The unlabelled pool\
& Statistical target, optimal for individual AL\
& Statistical target, optimal for batch AL\
& Classifier loss\
& Classifier future loss, after retraining on $(\mathbf{x}, c_j)$: ${L '}_j = L(\theta(D_S \cup (\mathbf{x}, c_j))$\
& Classifier future loss vector, for covariate $\mathbf{x}$: $\mathbf{L '} = \{ L(\theta(D_S \cup (\mathbf{x}, c_j)) \}_{j=1}^k$\
The notation is summarised in Table \[table:Notation 1\]. To define the statistical target, expectations are formed with respect to the underlying distribution $({\bf X}, Y)$. Assume a fixed dataset $D_S$ sampled i.i.d. from the joint distribution $({\bf X}, Y)$. The dependence of the classifier $\boldsymbol{\hat{\theta}}$ on the data $D_S$ is critical, with the notation $\boldsymbol{\hat{\theta}} = \theta(D_S)$ intended to emphasize this dependence. First assume a base classifier already trained on a dataset $D_S$. Consider how much a single labelled example improves performance. The single labelled example $({\bf x},y)$ will be chosen from a *labelled* dataset $D_W$. The loss from retraining on that single labelled example is examined in order to later define the loss for the expected label of an unlabelled example.
Examine the selection of a single labelled example $({\bf x},y)$ from $D_W$, given the labelled data $D_S$, the classifier training function $\theta$ and a loss function $L$. The reduction of the loss for retraining on that example is defined as actual-MRI, denoted $Q^a$:
$$Q^a({\bf x}, y, \theta, D_S) = L(\theta(D_S)) - L(\theta(D_S \cup ({\bf x},y)). $$
$Q^a$ is the actual classifier improvement from retraining on the labelled example $({\bf x},y)$. The goal here is to maximise the reduction of loss. The greatest loss reduction is achieved by selecting the example $({\bf x_*},y_*)$ from $D_W$ that maximises $Q^a$, given by $${(\bf x_*}, y_*) = \operatorname*{arg\,max}_{ ({\bf x}, y) \in D_W} Q^a({\bf x}, y, \theta, D_S).$$
Turning to AL, the single example ${\bf x}$ is unlabelled, and will be chosen from the unlabelled pool $X_P$. Here the unknown label of ${\bf x}$ is a random variable, $Y|{\bf x}$, and taking its expectation allows the expected loss to defined, this being the classifier loss after retraining with the unlabelled example and its unknown label. Thus the expected loss is defined using the expectation over the label $Y|{\bf x}$ to form conditional-MRI, denoted $Q^c$:
$$\label{eq:eqc}
\begin{split}
Q^c({\bf x}, \theta, D_S) = E_{Y | {\bf x}} [Q^a({\bf x}, Y, \theta, D_S)] = L(\theta(D_S)) - E_{Y | {\bf x}} [ L(\theta(D_S \cup ({\bf x},Y)) ] \\
= L(\theta(D_S)) - \sum_{j=1}^k \{ p_j \, L(\theta(D_S \cup ({\bf x},c_j)) \}
= \underbrace{ L(\theta(D_S)) }_{\text{Term } T_c} - \underbrace{ \sum_{j=1}^k p_j \, {L '}_j }_{\text{Term } T_e} = L(\theta(D_S)) - \mathbf{p} \cdot \mathbf{L '},
\end{split}$$
where $\mathbf{p}$ denotes the Bayes class probability vector $p(Y|{\bf x})$. ${L '}_j$ denotes a single future loss, from retraining on $D_S$ together with one example ${\bf x}$ given class $c_j$. $\mathbf{L '}$ denotes the future loss vector, i.e. the vector of losses from retraining on $D_S$ together with one example, that example being ${\bf x}$ combined with each possible label $c_j$: $\mathbf{L '} = \{ {L '}_j \}_{j=1}^k = \{ L(\theta(D_S \cup ({\bf x}, c_j)))\}_{j=1}^k$.
Term $T_c$ is the loss of the current classifier, given the training data $D_S$. Term $T_e$ is the expected future loss of the classifier, after retraining on the enhanced dataset $(D_S \cup (\mathbf{x}, Y|\mathbf{x}))$. $Q^c$ is defined as the difference between Terms $T_c$ and $T_e$, i.e. the difference between the current loss and the expected future loss. Thus $Q^c$ defines the expected loss reduction, from retraining on the example ${\bf x}$ with its unknown label. In this sense $Q^c$ is an improvement function, since it defines exactly how much this example will improve the classifier.
The unlabelled example ${\bf x_*}$ from the pool $X_P$ that maximises $Q^c$ is the optimal example selection: $$\label{eq:maximise_eqc}
{\bf x_*} = \operatorname*{arg\,max}_{ {\bf x} \in X_P} Q^c({\bf x}, \theta, D_S).$$
Novel algorithms are constructed to estimate the target $Q^c$, given in Section \[section:Algorithms to Estimate Example Quality\]. For an abstract classification problem, the target $Q^c$ can be evaluated exactly, to reveal the best and worst possible loss reduction, by maximising and minimising $Q^c$. Figure \[figure:Maximum and minimum AL performance with simulated data, exact EQ\] shows that the best and worst AL performance curves are indeed obtained by maximising and minimising $Q^c$.
The statistical quantity $Q^c$ defines optimal AL behaviour for any dataset $D_S$, whether iid or not, including the case of iterated AL, which generates a covariate bias in $D_S$ (see Section \[subsection:Active Learning\]). Given $Q^c$ for the selection of a single example, i.e. for individual AL, the optimal behaviour is now extended to batch AL, the selection of multiple examples, via the target $B^c$, given below.
![The best and worst AL performance curves are obtained by maximising and minimising the target $Q^c$, which demonstrate the extremes of AL performance. With simulated data, $Q^c$ can be calculated exactly; here the classification problem is the Four-Gaussian problem (illustrated in Figure \[fig:Ripley Four-Gaussian Problem\]). These curves are smoothed from multiple experiments, using the classifier $5$-nn. []{data-label="figure:Maximum and minimum AL performance with simulated data, exact EQ"}](Ripley_EQc_minimax1_knn5.pdf)
### Model Retraining Improvement for Batch Active Learning {#subsubsection:Example Quality for Batch Active Learning}
In batch AL, multiple examples are selected from the pool in one selection step. Each chosen batch consists of $r$ examples. Here MRI provides the statistical target $B^c$, the batch improvement function, defined as the expected classifier improvement over an unknown set of labels. First examine a *fully labelled* dataset $({\bf x}_R, {\bf y}_R)$, where $R$ denotes the index set $\{ {1, ..., r} \}$. For that fully labelled dataset, the actual loss reduction is denoted $B^a$: $$B^a({\bf x}_R, {\bf y}_R, \theta, D_S) = L(\theta(D_S)) - L(\theta(D_S \cup ({\bf x}_R,{\bf y}_R)).$$
Second consider the AL context, with a set of *unlabelled* examples ${\bf x_R}$, which is a single batch of examples selected from the pool. The expected loss reduction for this set of examples is denoted $B^c$: $$\begin{split}
B^c({\bf x}_R, \theta, D_S) = E_{{\bf Y}_R | {\bf x}_R} [B^a({\bf x}_R, {\bf Y}_R, \theta, D_S)] = L(\theta(D_S)) - E_{{\bf Y}_R | {\bf x}_R} [ L(\theta(D_S \cup ({\bf x}_R,{\bf Y}_R)) ] \\
= L(\theta(D_S)) - \sum_{j_1=1}^k \sum_{j_2=1}^k ... \sum_{j_r=1}^k \{ p_{j_1} p_{j_2} ... p_{j_r} \times L(\theta(D_S \cup ({\bf x_1},c_{j_1}) \cup ({\bf x_2},c_{j_2}) ... \cup ({\bf x_r},c_{j_r})) \}.
\end{split}$$ This expected loss reduction $B^c$ is an expectation taken over the unknown set of labels $({\bf Y}_R | {\bf x}_R)$. $B^c$ is the statistical target for batch AL, and the direct analog of $Q^c$ defined in Equation \[eq:eqc\].
Estimating $B^c$ incurs two major computational costs, in comparison to $Q^c$ estimation. First there is the huge increase in the number of selection candidates. For individual AL, each selection candidate is a single example, and there are only $n_p$ candidates to consider (where $n_p$ is the pool size). Under batch AL, each selection candidate is a set of examples, each set having size $r$; the number of candidates jumps to $n_p \choose r$. Thus batch AL generates a drastic increase in the number of selection candidates, from $n_p$ to $n_p \choose r$, which presents a major computational cost.
The second cost of $B^c$ estimation lies in the number of calculations per selection candidate. In individual AL, each candidate requires one classifier retraining and one loss evaluation per class, for all $k$ classes. However in batch AL, each candidate requires multiple classifier retraining and loss evaluations, each candidate now requiring $k^r$ calculations. Hence the number of calculations increases greatly, from $k$ to $k^r$, which is a severe computational cost.
These major computational costs make direct estimation of the target $B^c$ extremely challenging. Thus for batch AL the more practical option is to recommend algorithms that estimate $Q^c$, such as the algorithms given in Section \[section:Algorithms to Estimate Example Quality\]. $Q^c$ and $B^c$ together define the optimal AL behaviour for individual AL and batch AL. These targets provide optimal AL behaviour for both single-step and iterated AL. The rest of this work focusses on the target $Q^c$ as the foundation of MRI’s estimation framework for AL.
Abstract Example {#subsection:Theoretical Example to Illustrate Example Quality}
----------------
An example using an abstract classification problem is presented, to illustrate MRI in detail. The stochastic character of this problem is fully specified, allowing exact calculations of the loss $L$, and the statistical target $Q^c$ as functions of the univariate covariate $x$. To reason about $Q^c$ as a function of $x$, an infinite pool is assumed, allowing any $x \in \mathbb{R}$ to be selected. These targets are then explored as functions of $x$, and the optimal AL selection $x_*$ is examined (see Equation \[eq:maximise\_eqc\]). The full stochastic description allows examination of the AL method’s selection, denoted $x_r$, and comparison to the optimal selection $x_*$. This comparison is made below for the popular AL heuristic Shannon entropy, and for random selection.
Imagine a binary univariate problem, defined by a balanced mixture of two Gaussians: $ \{ {\boldsymbol\pi} = (\frac{1}{2}, \frac{1}{2}), (X|Y=c_1) \sim \textrm{N}(-1,1), (X|Y=c_2) \sim \textrm{N}(1,1) \}$. The true means are denoted ${\mu}_1 = -1, {\mu}_2 = 1$. The loss function is error rate $L_e$ (defined in Section \[subsection:Classification\]), while the true decision boundary to minimise error rate is denoted $t = \frac{1}{2} ({\mu}_1 + {\mu}_2)$. Every dataset $D$ of size $n$ sampled from this problem is assumed to split equally into two class-pure subsets $D_j = \{ y_i = c_j, (x_i, y_i) \in D \}$ each of size $n_j = \frac{n}{2}$; this is sampling while holding the prior fixed. Consider a classifier that estimates only the class-conditional means, given the true prior $\boldsymbol\pi$ and the true common variance of 1. The classifier parameter vector is $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$, where $\hat{\mu}_j$ is the sample mean for class $c_j$. This implies that the classifier’s estimated decision boundary to minimise error rate is denoted $\hat{t} = \frac{1}{2} (\hat{\mu}_1 + \hat{\mu}_2)$.
### Calculation and Exploration of $Q^c$ {#subsubsection:Calculation of Qc}
Here $Q^c$ is calculated, then explored as a function of $x$. The classifier’s decision rule $r_1(x)$ minimises the loss $L_e(\boldsymbol{\hat{\theta}})$, and is given in terms of a threshold on the estimated class probabilities by $$r_1(x) = \left\{
\begin{array}{lr}
\hat{y} = c_1 & : \hat{p}_1(x) > \frac{1}{2}, \\
\hat{y} = c_2 & : \hat{p}_1(x) < \frac{1}{2}, \\
\end{array}
\right.$$ or equivalently, in terms of a decision boundary on $x$, by $$r_2(x) = \left\{
\begin{array}{lr}
\hat{\mu}_1 < \hat{\mu}_2 & : \hat{y} = c_1 \textrm{ if } x < \hat{t}, c_2 \textrm{ otherwise}, \\
\hat{\mu}_1 > \hat{\mu}_2 & : \hat{y} = c_1 \textrm{ if } x > \hat{t}, c_2 \textrm{ otherwise}. \\
\end{array}
\right.$$
The classifier may get the estimated class means the wrong way around, in the unlikely case that $\hat{\mu}_1 > \hat{\mu}_2$. As a result the classifier’s behaviour is very sensitive to the condition $(\hat{\mu}_1 > \hat{\mu}_2)$, as shown by the second form of the decision rule $r_2(x)$, and by the loss function in Equation \[eq:le\_equation\_1\].
It is straightforward to show that the loss $L_e(\boldsymbol{\hat{\theta}})$ is given by $$\label{eq:le_equation_1}
\begin{split}
L_e(\boldsymbol{\hat{\theta}}) = \frac{1}{2} \{ 1 - F_1(\hat{t}) + F_2(\hat{t}) + \mathbbm{1}(\hat{\mu}_1 > \hat{\mu}_2) [2 F_1(\hat{t}) - 2 F_2(\hat{t})] \},
\end{split}$$ where $F_j(x)$ denotes the cdf for class-conditional distribution $(X|Y=c_j)$. In individual AL an unlabelled point $x$ is chosen for the oracle to label, before retraining the classifier. Retraining the classifier with a single new example $(x, c_j)$ yields a new parameter estimate denoted $\boldsymbol{\hat{\theta}}^{\prime}_j$, where the mean estimate for class $c_j$ has a new value denoted $\hat{\mu}^{\prime}_j$, with a new estimated boundary denoted $\hat{t}^{\prime}_j$.
Here $\hat{\mu}^{\prime}_j = (1-z)\hat{\mu}_j + z x$ where $z = \frac{2}{n+2}$, $z$ being an updating constant which reflects the impact of the new example on the mean estimate $\hat{\mu}_j$. To calculate $Q^c$ under error loss $L_e$, observe that the Term $T_e$ from Equation \[eq:eqc\] is $[ p_1 L_e(\boldsymbol{\hat{\theta}}^{\prime}_1) + p_2 L_e(\boldsymbol{\hat{\theta}}^{\prime}_2) ]$. Term $T_c$ in Equation \[eq:eqc\] is directly given by Equation \[eq:le\_equation\_1\]. From Equations \[eq:eqc\] and \[eq:le\_equation\_1\], $Q^c (x, \theta, D) = L_e(\boldsymbol{\hat{\theta}}) - [ p_1 L_e(\boldsymbol{\hat{\theta}}^{\prime}_1) + p_2 L_e(\boldsymbol{\hat{\theta}}^{\prime}_2) ]$, hence $$\begin{split}
Q^c (x, \theta, D) = \frac{1}{2} \{ 1 - F_1(\hat{t}) + F_2(\hat{t}) + \mathbbm{1}(\hat{\mu}_1 > \hat{\mu}_2) [2 F_1(\hat{t}) - 2 F_2(\hat{t})] \} \\
- \frac{p_1}{2} \{ 1 - F_1(\hat{t}^{\prime}_1) + F_2(\hat{t}^{\prime}_1) + \mathbbm{1}(\hat{\mu}^{\prime}_1 > \hat{\mu}_2) [2 F_1(\hat{t}^{\prime}_1) - 2 F_2(\hat{t}^{\prime}_1)] \} \\
- \frac{p_2}{2} \{ 1 - F_1(\hat{t}^{\prime}_2) + F_2(\hat{t}^{\prime}_2) + \mathbbm{1}(\hat{\mu}_1 > \hat{\mu}^{\prime}_2) [2 F_1(\hat{t}^{\prime}_2) - 2 F_2(\hat{t}^{\prime}_2)] \},
\end{split}$$ where $p_j$, $\hat{\mu}^{\prime}_j$, and $\hat{t}^{\prime}_j$ are functions of $x$.
Even for this simple univariate problem, $Q^c(x, \theta, D)$ is a complicated non-linear function of $x$. Given this complication, $Q^c$ is explored by examining specific cases of the estimated parameter $\boldsymbol{\hat{\theta}}$, shown in Figure \[fig:Qc for specific thetaHat cases, Example 1\]. In each specific case of $\boldsymbol{\hat{\theta}}$, $x_*$ yields greatest correction to $\boldsymbol{\hat{\theta}}$ in terms of moving the estimated boundary $\hat{t}$ closer to the true boundary $t$. This is intuitively reasonable since error rate is a function of $\hat{t}$ and minimised for $\hat{t} = t$. In the first two cases (Figures \[fig:Qc for specific thetaHat cases, Example 1\_case\_mu1h=\[-0\_5\]\_\_mu2h=\[1\_5\]\] and \[fig:Qc for specific thetaHat cases, Example 1\_case\_mu1h=\[-0\_9\]\_\_mu2h=\[1\_1\]\]), the estimated threshold is greater than the true threshold, $\hat{t} > t$. In these two cases, $x_*$ is negative, hence retraining on $x_*$ will reduce the estimated threshold $\hat{t}$, bringing it closer to the true threshold $t$, thereby improving the classifier. In the third case (Figure \[fig:Qc for specific thetaHat cases, Example 1\_case\_mu1h=\[-1\_1\]\_\_mu2h=\[1\_1\]\]), $\hat{t} = t$ and here the classifier’s loss $L_e$ cannot be reduced, shown by $Q^c(x) < 0$ for all $x$. The fourth case (Figure \[fig:Qc for specific thetaHat cases, Example 1\_case\_mu1h=\[1\]\_\_mu2h=\[-1\]\]) is interesting because the signs of the estimated means are reversed compared to the true means, and here the most non-central $x$ offer greatest classifier improvement. Together these cases show that even for this toy example, the improvement function $Q^c$ is complicated and highly dependent on the estimated parameters.
[0.45]{} ![Illustration of the target $Q^c$ as a function of $x$, for specific cases of the estimated classifier parameters $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$. The class mean parameters are shown in solid blue and red, with the estimated means shown in dotted blue and red. The green line indicates $Q^c(x) = 0$ (zero improvement); in all cases, $n_s = 18$. In each specific case, the optimal selection $x_*$ yields greatest correction to $\boldsymbol{\hat{\theta}}$ in terms of moving the estimated boundary $\hat{t}$ closer to the true boundary $t$. []{data-label="fig:Qc for specific thetaHat cases, Example 1"}](Case_Qc_Lerror__n=18__mu1h=-0_5__mu2h=1_5.pdf "fig:"){width="\textwidth"}
[0.45]{} ![Illustration of the target $Q^c$ as a function of $x$, for specific cases of the estimated classifier parameters $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$. The class mean parameters are shown in solid blue and red, with the estimated means shown in dotted blue and red. The green line indicates $Q^c(x) = 0$ (zero improvement); in all cases, $n_s = 18$. In each specific case, the optimal selection $x_*$ yields greatest correction to $\boldsymbol{\hat{\theta}}$ in terms of moving the estimated boundary $\hat{t}$ closer to the true boundary $t$. []{data-label="fig:Qc for specific thetaHat cases, Example 1"}](Case_Qc_Lerror__n=18__mu1h=-0_9__mu2h=1_1.pdf "fig:"){width="\textwidth"}
[0.45]{} ![Illustration of the target $Q^c$ as a function of $x$, for specific cases of the estimated classifier parameters $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$. The class mean parameters are shown in solid blue and red, with the estimated means shown in dotted blue and red. The green line indicates $Q^c(x) = 0$ (zero improvement); in all cases, $n_s = 18$. In each specific case, the optimal selection $x_*$ yields greatest correction to $\boldsymbol{\hat{\theta}}$ in terms of moving the estimated boundary $\hat{t}$ closer to the true boundary $t$. []{data-label="fig:Qc for specific thetaHat cases, Example 1"}](Case_Qc_Lerror__n=18__mu1h=-1_1__mu2h=1_1.pdf "fig:"){width="\textwidth"}
[0.45]{} ![Illustration of the target $Q^c$ as a function of $x$, for specific cases of the estimated classifier parameters $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$. The class mean parameters are shown in solid blue and red, with the estimated means shown in dotted blue and red. The green line indicates $Q^c(x) = 0$ (zero improvement); in all cases, $n_s = 18$. In each specific case, the optimal selection $x_*$ yields greatest correction to $\boldsymbol{\hat{\theta}}$ in terms of moving the estimated boundary $\hat{t}$ closer to the true boundary $t$. []{data-label="fig:Qc for specific thetaHat cases, Example 1"}](Case_Qc_Lerror__n=18__mu1h=1__mu2h=-1.pdf "fig:"){width="\textwidth"}
### Exploration of Shannon Entropy and Random Selection {#subsubsection:Exploration of Shannon Entropy and Random Selection}
The abstract example is used to compare is two selection methods, SE and RS, against optimal AL behaviour. SE always selects $x_r$ at the estimated boundary $\hat{t}$. RS selects uniformly from the pool, assumed to be i.i.d. in AL, hence the RS selection probability is given by the marginal density $p(x)$. In contrast to $Q^c$ and SE, RS is a stochastic selection method, with expected selection $x_r = 0$ in this problem. Figure \[fig:Se-and-RS-and-Qc for specific thetaHat cases, Example 1\] illustrates $Q^c$, SE and $p(x)$ as contrasting functions of $x$, with very different maxima.
[0.45]{} ![Comparison of $Q^c$ against SE and RS as functions of $x$, for specific cases of the estimated classifier parameters $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$. $Q^c$ is shown in black, SE in purple and RS in orange (for RS, the density $p(x)$ is shown). The class mean parameters are shown in solid blue and red, with the estimated means shown in dotted blue and red. The green line indicates $Q^c(x) = 0$ (zero improvement); in all cases, $n_s = 18$. The three functions are scaled to permit this comparison. []{data-label="fig:Se-and-RS-and-Qc for specific thetaHat cases, Example 1"}](Qc_Lerror_mu1hat=-0_5__mu2hat=1_5_Overlay_n=18.pdf "fig:"){width="\textwidth"}
[0.45]{} ![Comparison of $Q^c$ against SE and RS as functions of $x$, for specific cases of the estimated classifier parameters $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$. $Q^c$ is shown in black, SE in purple and RS in orange (for RS, the density $p(x)$ is shown). The class mean parameters are shown in solid blue and red, with the estimated means shown in dotted blue and red. The green line indicates $Q^c(x) = 0$ (zero improvement); in all cases, $n_s = 18$. The three functions are scaled to permit this comparison. []{data-label="fig:Se-and-RS-and-Qc for specific thetaHat cases, Example 1"}](Qc_Lerror_mu1hat=-0_9__mu2hat=1_1_Overlay_n=18.pdf "fig:"){width="\textwidth"}
[0.45]{} ![Comparison of $Q^c$ against SE and RS as functions of $x$, for specific cases of the estimated classifier parameters $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$. $Q^c$ is shown in black, SE in purple and RS in orange (for RS, the density $p(x)$ is shown). The class mean parameters are shown in solid blue and red, with the estimated means shown in dotted blue and red. The green line indicates $Q^c(x) = 0$ (zero improvement); in all cases, $n_s = 18$. The three functions are scaled to permit this comparison. []{data-label="fig:Se-and-RS-and-Qc for specific thetaHat cases, Example 1"}](Qc_Lerror_mu1hat=-1_1__mu2hat=1_1_Overlay_n=18.pdf "fig:"){width="\textwidth"}
[0.45]{} ![Comparison of $Q^c$ against SE and RS as functions of $x$, for specific cases of the estimated classifier parameters $\boldsymbol{\hat{\theta}} = ( \hat{\mu}_1, \hat{\mu}_2 )$. $Q^c$ is shown in black, SE in purple and RS in orange (for RS, the density $p(x)$ is shown). The class mean parameters are shown in solid blue and red, with the estimated means shown in dotted blue and red. The green line indicates $Q^c(x) = 0$ (zero improvement); in all cases, $n_s = 18$. The three functions are scaled to permit this comparison. []{data-label="fig:Se-and-RS-and-Qc for specific thetaHat cases, Example 1"}](Qc_Lerror_mu1hat=1__mu2hat=-1_Overlay_n=18.pdf "fig:"){width="\textwidth"}
$Q^c$ is asymmetric in the first two cases, and symmetric for the final two. By contrast, SE and RS are always symmetric (for all possible values of $\boldsymbol{\hat{\theta}}$). In the first two cases (Figures \[fig:Se-and-Qc for specific thetaHat cases, Example 1\_case\_mu1h=\[-0\_5\]\_\_mu2h=\[1\_5\]\] and \[fig:Qc for specific thetaHat cases, Example 1\_case\_mu1h=\[-0\_9\]\_\_mu2h=\[1\_1\]\]), SE selects a central $x_r$, thereby missing the optimal selection $x_*$. In the second case (Figure \[fig:Qc for specific thetaHat cases, Example 1\_case\_mu1h=\[-0\_9\]\_\_mu2h=\[1\_1\]\]), SE selects $x_r$ with $Q^c(x_r) < 0$, failing to improve the classifier, whereas the optimal selection $x_*$ does improve the classifier since $Q^c(x_*) > 0$. The third case is unusual, since $\hat{t} = t$ and this classifier’s loss $L_e$ cannot be improved, hence $Q^c(x) < 0$ for all $x$. In the fourth case (Figure \[fig:Se-and-Qc for specific thetaHat cases, Example 1\_case\_mu1h=\[1\]\_\_mu2h=\[-1\]\]) SE makes the worst possible choice of $x$. In all four cases, SE never chooses the optimal point; SE may improve the classifier, but never yields the greatest improvement. These specific cases of $\boldsymbol{\hat{\theta}}$ show that SE often makes a suboptimal choice for $x_r$, for this abstract example. Turning to consider RS, the stochastic nature of RS suggests that the expected RS selection is the quantity of interest. For these four cases of $\boldsymbol{\hat{\theta}}$, the expected RS selection is a suboptimal choice of $x_*$ for this problem. It is notable that the expected RS selection is usually close to the SE selection. The stochastic nature of RS implies that it often selects far more non-central $x$ values than SE.
Unbiased $Q^c$ Estimation Outperforms Random Selection {#subsection:Unbiased $Q^c$ Estimation Yields AL Performance Matching or Exceeding RS}
------------------------------------------------------
We present an argument that an unbiased estimator of $Q^c$ will always exceed RS in AL performance. This formal approach opens the door to a new guarantee for AL, which motivates the estimation framework that MRI provides. This guarantee is not tautological since RS generally improves the classifier, making RS a reasonable benchmark to outperform. By contrast, heuristic AL methods such as SE lack any estimation target, making arguments of this kind difficult to construct. This argument also motivates the algorithm bootstrapMRI (see Section \[subsection:Algorithm PartitionEQ\]). The context is an AL scenario with a specific classification problem, classifier and loss function. We examine the selection of a single example from a pool $X_P$ consisting of just two examples $X_P = \{ x_1, x_2 \}$.
From Equation 2, the target function $Q^c$ depends on both the labelled data $D_S$ and the population distribution $(X,Y)$. This dependency on both data and population is somewhat unusual for an estimation target, but other statistical targets share this property, for example classifier loss. Here the labelled dataset $D_S$ is considered a random variable, hence the values of $Q^c$ over the pool are also random. Consider a hypothetical $\hat{Q^c}$ estimator, unbiased in this sense: $ (\forall x_i \in \mathbb{R}) E [\hat{Q^c}(x_i, \theta, D_S)] = [Q^c(x_i, \theta, D_S)]$. For a single example $x_i$, the true and estimated values of $Q^c$ are denoted by $Q_i = Q^c(x_i, \theta, D_S)$ and $\hat{Q}_i = \hat{Q^c}(x_i, \theta, D_S)$ respectively. Since the estimator is unbiased, the relationship between these quantities can be conceptualised as $\hat{Q}_i = Q_i + M_i$, where $M_i$ is defined as a noise term with zero mean and variance $\sigma^2$, with $E_{D_S} [M_i] = 0$. We assume that $M_i \condindep Q_i$, and make the moderate assumption that $M_i \sim N(0, \sigma^2)$. The difference between the true $Q^c$ values is defined as $R = Q_1 - Q_2$. We begin by addressing the case where $(R > 0)$ i.e. $(Q_1 > Q_2)$. The probability that the optimal example is chosen, denoted $\lambda$, will illustrate the estimator’s behaviour under different noise variances, $\sigma^2$.
We now quantify the selection probability $\lambda$ explicitly in terms of estimator variance. This selection probability $\lambda$ is given by $$\begin{aligned}
\lambda &= p(\hat{Q_1} > \hat{Q_2}) \\
&= p(Q_1 + M_1 > Q_2 + M_2) \\
&= p(Q_1 - Q_2 > M_2 - M_1) \\
&= p(M_2 - M_1 < Q_1 - Q_2),
$$ which can be rewritten as $\lambda = p(N < \Delta)$ where $N= M_2 - M_1$ is defined as a mean zero RV, and $\Delta = Q_1 - Q_2$ is strictly positive (since $R > 0$). $N$ is Gaussian, since $M_1$ and $M_2$ are both Gaussian. This variable $\Delta$ provides a ranking signal for example selection: its sign shows that $x_1$ is a better choice than $x_2$, and its magnitude shows how much better.
Further defining $\alpha = p(N < 0)$ and $\beta = p(0 \leq N < \Delta)$ and combining with $p(N < \Delta) = p(N < 0) + p(N \leq 0 < \Delta)$ gives $\lambda = \alpha + \beta$. Here $\alpha \condindep \Delta$ whereas $\beta \notcondindep \Delta$, showing that $\alpha$ is a pure noise term devoid of any $Q^c$ ranking information, while $\beta$ contains ranking information by its dependency on $\Delta$.
We now establish that $\alpha = \frac{1}{2}$ by examining the special case of $\hat{Q^c}$ estimator variance tending towards infinity. This value of $\alpha = \frac{1}{2}$ proves important in relating the selection behaviour of the infinite-variance estimator to random selection.
As $\sigma^2 \uparrow \infty$, $\beta \downarrow 0$, this result being shown in Appendix A. Hence as $\sigma^2 \uparrow \infty$, $\lambda \downarrow \alpha$. Thus as $\sigma^2 \uparrow \infty$, $\lambda \condindep (Q_1, Q_2)$ since $\alpha \condindep (Q_1, Q_2)$, hence $\lambda$ becomes independent of true $Q^c$ values, depending only on noise. Hence the limiting case, as the estimator variance approaches infinity, corresponds to uniform selection over the pool. A closely related argument for $\alpha = \frac{1}{2}$ is the impossibility of selection by signal-free noise $\alpha$ outperforming RS. Again considering $\sigma^2 \uparrow \infty$, if $\alpha > \frac{1}{2}$ then $\lambda > \frac{1}{2}$, which will consistently prefer the better example $x_1$, and therefore consistently outperform RS. Whereas $\alpha < \frac{1}{2}$ gives $\lambda < \frac{1}{2}$, which will consistently prefer the worse example $x_2$, and therefore consistently underperform RS. However, outperforming RS when selecting examples by noise alone is impossible, which implies $\alpha = \frac{1}{2}$. Further, $N$ is Gaussian with mean-zero which directly gives $\alpha = \frac{1}{2}$.
From $\alpha = \frac{1}{2}$, $\lambda$ can be expressed purely in terms of $\beta$ as $$\label{eq:lambda}
\lambda = \frac{1}{2} + \beta.$$ As $\sigma^2 \uparrow \infty$, $\beta \downarrow 0$ hence $\lambda \downarrow \frac{1}{2}$. When $\sigma^2 \downarrow 0$, $\beta \uparrow \frac{1}{2}$, as shown in Appendix B. Hence as $\sigma^2 \downarrow 0$, $\lambda \uparrow 1$. Since N is Gaussian, $\beta \in (0, \frac{1}{2}]$, hence $\lambda \in (\frac{1}{2}, 1]$. Having examined the case where $(R>0)$, we now consider all of the possibilities for $R$. The zero probability case ($R = 0$) is discarded, leaving only the second case defined by $(R < 0)$. In this second case $(R < 0)$ i.e. $(Q_1 < Q_2)$, the optimal selection is $x_2$, with $$\begin{aligned}
\lambda &= p(\hat{Q}_2 > \hat{Q}_1) \\
&= p(Q_2 + M_2 > Q_1 + M_1) \\
&= p(M_2 - M_1 > Q_1 - Q_2),
$$ rewritten as $\lambda = p(N > -\Delta_2)$ where $\Delta_2 = Q_2 - Q_1$ is strictly positive (since $R < 0)$. Hence $$\begin{aligned}
\lambda &= p(N > -\Delta_2) \\
&= p(N > 0) + p(-\Delta_2 < N \leq 0) \\
&= (1 - \alpha) + \beta_2 \\
&= \frac{1}{2} + \beta_2,
$$ where $\beta_2 = p(-\Delta_2 < N \leq 0)$. Since $N$ is Gaussian, it is symmetric, giving $\beta_2 = p(-\Delta_2 < N \leq 0) = p(0 \leq N < \Delta_2)$. Here $\Delta$ and $\Delta_2$ differ only in magnitude, and their magnitudes do not feature in the proofs in Appendices F and G. As a result, $\beta_2$ takes the very same values as $\beta$ when $\sigma \downarrow 0$ or $\sigma \uparrow \infty$, namely $\{ \frac{1}{2}, 0 \}$ (see the proofs in Appendices F and G). Thus the selection behaviour of the unbiased estimator is the same for both cases of $(R>0)$ and $(R<0)$, both cases being described by Equation \[eq:lambda\]. The RHS of Equation \[eq:lambda\] quantifies the combination of signal and noise in AL selection, with the estimator variance $\sigma^2$ determining $\beta$ and $\lambda$. Now the AL performance under the $\hat{Q^c}$ estimator can be elucidated in terms of the estimator variance.
The extreme case of infinite variance where $\lambda = \frac{1}{2}$ implies that the selection of examples is entirely random, and here the estimator’s behaviour is identical to random selection (RS), an established AL benchmark. By contrast, if $\lambda$ exceeds $\frac{1}{2}$, examples with better $Q^c$ values are more likely to be selected, leading to better expected AL performance than RS.
This argument applies directly to a pool of two elements. The ranking of a larger pool can be decomposed into pairwise comparisons, which may extend this argument to any pool. This argument serves to illustrate that an unbiased $\hat{Q^c}$ estimator outperforms RS, which is a new guarantee for AL. This argument receives experimental support from the results described in Section \[subsection:Primary Results\].
We make no attempt to prove the existence of such an unbiased $\hat{Q^c}$ estimator. The bootstrapMRI algorithm given in Section \[subsection:Algorithm PartitionEQ\] is constructed, as far as is practical, to capture the key characteristics of an ideal unbiased $\hat{Q^c}$ estimator.
Algorithms to Estimate Model Retraining Improvement {#section:Algorithms to Estimate Example Quality}
===================================================
For practical estimation of $\hat{Q^c}$, Term $T_c$ in Equation \[eq:eqc\] can be ignored since it is independent of ${\bf x}$. Thus the central task of practical $\hat{Q^c}$ estimation is the calculation of Term $T_e$ in Equation \[eq:eqc\], this Term $T_e$ being the expected classifier loss after retraining on the new example ${\bf x}$ with its unknown label $Y|{\bf x}$. The definition of Term $T_e$ in Equation \[eq:eqc\] includes two components: ${\bf p}$ and $\boldsymbol{L '}$. Consequently, $Q^c$ estimation requires estimating these two components from one labelled dataset $D_S$.
Estimating multiple quantities from a single dataset raises interesting statistical choices. One major choice must be made between using the same data to estimate both components (termed *naïve reuse*), or to use bootstrapping to generate independent resampled datasets, producing independent component estimates. This choice between naïve reuse and boostrapping has implications for the bias of $\hat{Q}^c$ estimates, discussed below. Here we assume that loss estimation itself requires two datasets, for training and testing, denoted $D_T$ and $D_E$ respectively. Since ${\bf p}$ estimation requires one dataset, then three datasets are needed in total, denoted $D_P$, $D_T$ and $D_E$, to estimate the two components ${\bf p}$ and $\mathbf{L '}$:
- The class probability vector, ${\bf p} = p(Y|{\bf x})$, estimated by $\hat{\bf p}$ using dataset $D_P$,
- The future loss vector, $\mathbf{L '}$, estimated by $\mathbf{\hat{L '}}$ using datasets $D_T$ and $D_E$.
Each of these three datasets ($D_P$, $D_T$ and $D_E$) must be derived from $D_S$. In the case of naïve reuse, all three datasets equal $D_S$, yielding the algorithm simpleMRI described below. For bootstrapping, the three datasets are all resampled from $D_S$ with replacement, giving the algorithm bootstrapMRI described below. These two algorithms are extreme cases, chosen for clarity and performance; numerous variations are possible here. A statistical estimate is considered *precise* when it has low estimation error. Literature on empirical learning curves suggests that classifier loss $L$ is larger for smaller training data samples [@Provost2003; @Gu2001; @Kadie1995]. This implies that ${\bf p}$ is difficult to estimate precisely, since precise estimates of ${\bf p}$ would directly produce a near-optimal classifier (one close to the optimum Bayes classifier, in terms of loss). The increased loss for smaller samples further implies that loss $L$ itself is hard to estimate precisely for a small training dataset; for if loss could be precisely estimated, a near-optimal classifier could be found by direct optimisation.
This line of reasoning suggests that the two main components of $Q^c$, ${\bf p}$ and $\boldsymbol{L '}$, are both very difficult to estimate precisely from small data samples. In practical applications where all quantities must be estimated from data, the estimates will inevitably suffer from imprecision.
Algorithm SimpleMRI {#subsection:Algorithm SimpleEQ}
-------------------
We present the simpleMRI algorithm to estimate $Q^c$, to illustrate the statistical framework. The pseudocode for simpleMRI is provided in Algorithm \[algorithm:SEQ\]. This first algorithm takes a simple approach where all of $D_S$ is used to estimate all three components. The algorithm uses the maximum amount of data for each component estimate, broadly intending to reduce the variance of these component estimates. The class probability vector $\mathbf{\hat{p}}$ is estimated by training a second classifier $\theta_2$ on $D_P$, then using its predicted probability vector $\mathbf{\hat{p}}$ for the example $\mathbf{x}$. This second classifier is $5$-nn, or random forest when the base classifier is $k$-nn [@Breiman2001]. For the future loss vector $\mathbf{\hat{L '}}$, each element ${L '}_j$ is estimated by training the base classifier $\theta()$ on $D_T \cup (\mathbf{x}, c_j)$, then computing a loss estimate using $D_E$. The simpleMRI algorithm immediately encounters a problem in estimating Term $T_e$: the same data $D_S$ is used both to train the classifier and also to estimate the loss. This in-sample loss estimation is known to produce optimistic, biased estimates of the loss [@Tibshirani2009 Chapter 7]. The simpleMRI algorithm suffers another potential problem with bias: the same data $D_S$ is used to estimate the class probability and estimate the loss, leading to dependence between the estimates of $\hat{\bf p}$ and $\mathbf{L '}$. This dependence of component estimates may produce bias in the estimate $\hat{Q^c}$ from simpleMRI, since the argument of Equation \[eq:Tehat-unbiased\] for unbiased $Q^c$ estimation requires independent component estimates. These two problems of biased and dependent component estimates under naïve reuse motivates the development of a second algorithm, termed bootstrapMRI, described below. For computational efficiency, $\hat{Q^c}$ values are only evaluated on a randomly (uniformly) chosen subset of the pool. This popular AL optimisation is commonly termed random sub-sampling.
$D_P \gets D_S$ $D_T \gets D_S$ $D_E \gets D_S$ *estimate class probability vector $\mathbf{\hat{p}}$ at $\mathbf{x}$* $\boldsymbol{\hat{\theta}}_2 \gets \theta_2(D_P)$ $\mathbf{\hat{p}} \gets \phi_2(\boldsymbol{\hat{\theta}}_2, \mathbf{x})$ *estimate future loss vector $\mathbf{\hat{L '}}$* $\boldsymbol{\hat{\theta}}_j \gets \theta(D_T \cup (\mathbf{x}, c_j))$ $\hat{L '}_j \gets \frac{1}{|D_E|} \sum_{(\mathbf{x}_e, y_e) \in D_E} M_e(\boldsymbol{\hat{\theta}}_j, \mathbf{x_e}, y_e)$ $\hat{T_e} \gets \mathbf{\hat{p}} \cdot \mathbf{\hat{L '}}$ $\hat{Q^c} \gets \hat{T_e}$
Algorithm BootstrapMRI {#subsection:Algorithm PartitionEQ}
----------------------
BootstrapMRI seeks to minimise $Q^c$ estimator bias in two ways: by generating independent component estimates, and by providing component estimators of reasonably low bias. If the two component estimators $\mathbf{\hat{p}}$ and $\mathbf{\hat{L '}}$ are independent, and both unbiased, then the $\hat{Q^c}$ estimator will be unbiased, as shown below in Section \[subsection:Algorithm PartitionEQ Provides Approximately Unbiased $Qc$ Estimation\]. The pseudocode for bootstrapMRI is provided in Algorithm \[algorithm:BEQ\]. The labelled dataset $D_S$ is resampled by bootstrapping to form three datasets $D_P, D_T$ and $D_E$. These three datasets are independent draws from the ecdf of $D_S$, yielding independent estimates [@Efron1983 Chapter 6].
The first dataset $D_P$ provides an estimate for the class probability $\mathbf{\hat{p}}$, by classifier training on $D_P$ and class probability prediction on $\mathbf{x}$. As before, for $\mathbf{\hat{p}}$ estimation, a second classifier $\theta_2$ is used, chosen in the very same way as simpleMRI above. The second and third datasets $D_T$ and $D_E$ together provide an estimate of the future losses vector $\mathbf{\hat{L '}}$. Each element ${L '}_j$ is estimated by training the base classifier $\theta()$ on $D_T \cup (\mathbf{x}, c_j)$, then computing a loss estimate using $D_E$. In the experimental study of Section \[section:Experiments and Results\], the stochastic resampling is repeated, $n_b = 25$ times, and the resulting estimates are averaged. Random sub-sampling of the pool is used for efficiency.
${\bf q} \gets \textit{zero vector of length } n_b$ $I_P \gets \textit{Sample With Replacement} (1:|D_S|)$ $I_T \gets \textit{Sample With Replacement} (1:|D_S|)$ $I_E \gets \textit{Sample With Replacement} (1:|D_S|)$ $D_P \gets D_S[I_P]$ $D_T \gets D_S[I_T]$ $D_E \gets D_S[I_E]$ *estimate class probability vector $\mathbf{\hat{p}}$ at $\mathbf{x}$* $\boldsymbol{\hat{\theta}}_2 \gets \theta_2(D_P)$ $\mathbf{\hat{p}} \gets \phi_2(\boldsymbol{\hat{\theta}}_2, \mathbf{x})$ *estimate future loss vector $\mathbf{\hat{L '}}$* $\boldsymbol{\hat{\theta}}_j \gets \theta(D_T \cup (\mathbf{x}, c_j))$ $\hat{L '}_j \gets \frac{1}{|D_E|} \sum_{(\mathbf{x}_e, y_e) \in D_E} M_e(\boldsymbol{\hat{\theta}}_j, \mathbf{x}_e, y_e)$ $\hat{T_e} \gets \mathbf{\hat{p}} \cdot \mathbf{\hat{L '}}$ $q[b] \gets \hat{T_e}$ *the final estimate is the average of the estimate vector $\mathbf{q}$* $\hat{Q^c} \gets median(\mathbf{q})$
BootstrapMRI Algorithm Properties {#subsection:Algorithm PartitionEQ Provides Approximately Unbiased $Qc$ Estimation}
---------------------------------
BootstrapMRI seeks to minimise $\hat{Q^c}$ estimation bias by generating independent component estimates, as shown below in Equation \[eq:Tehat-unbiased\]. Practical $Q^c$ estimation requires calculating only Term $T_e$ in Equation 2 (Term $T_c$ can be ignored for practical estimation, since $\text{Term } T_c \condindep \mathbf{x}$). Term $T_e$ is a product of $\mathbf{p}$ and $\mathbf{L '}$, the two components of $Q^c$ to be estimated.
The definitions of unbiased estimation are given below:
- Unbiasedness for $\mathbf{\hat{p}}$ is defined as $(\forall \mathbf{x}_i \in \mathbb{R}^d) E [\mathbf{\hat{p}}(\mathbf{x}_i)] = \mathbf{p}(\mathbf{x}_i)$.
- Unbiasedness for $\mathbf{\hat{L '}}$ is defined as $(\forall \mathbf{x}_i \in \mathbb{R}^d) E [\mathbf{\hat{L '}}(\mathbf{x}_i)] = \mathbf{L '}(\mathbf{x}_i)$.
- Unbiasedness for $\hat{Q^c}$ is defined as $(\forall \mathbf{x}_i \in \mathbb{R}^d) E [\hat{Q^c}(\mathbf{x}_i, \theta, D_S)] = Q^c(\mathbf{x}_i, \theta, D_S)$,
where the expectations are taken over the variability of the estimators.
The independence of $\mathbf{\hat{p}}$ and $\mathbf{\hat{L '}}$ is classical statistical independence: $(\mathbf{\hat{p}} \condindep \mathbf{\hat{L '}})
\Leftrightarrow [ p(\mathbf{\hat{p}} = \mathbf{a}, \mathbf{\hat{L '}} = \mathbf{b}) = p(\mathbf{\hat{p}} = \mathbf{a}) \, p(\mathbf{\hat{L '}} = \mathbf{b}) ]$, for constant vectors $\mathbf{a}$ and $\mathbf{b}$. By generating independent component estimates, bootstrapMRI provides a guarantee: that if the two component estimates $\mathbf{\hat{p}}$ and $\mathbf{\hat{L '}}$ are both unbiased, then the $\hat{Q^c}$ estimate will be unbiased. This is shown by $E [\hat{Q^c}(x)] = Q^c(x)$, since $$\begin{aligned}
\label{eq:Tehat-unbiased}
E[\hat{T_e}] &= E[\hat{\mathbf{p}} \cdot \hat{\mathbf{L '}}] \\
&= E[\hat{\mathbf{p}}] \cdot E[\hat{\mathbf{L '}}] \nonumber \\ &= \mathbf{p} \cdot \mathbf{L '} \nonumber \\ &= T_e \nonumber.\end{aligned}$$
An ideal scenario would include the Bayes classifier and a large test dataset, providing the exact probabilities $\mathbf{p}$ and precise, unbiased estimates of $\mathbf{L '}$. In that scenario, the $\hat{Q^c}$ estimate will be completely unbiased. In the real application context, neither the Bayes classifier nor a large test dataset are available, and it is hard to estimate either component $\mathbf{\hat{p}}$ or $\mathbf{\hat{L '}}$ precisely or unbiasedly from a small data sample, these being open research problems [@Acharya2013; @Rodriguez2013]. Small finite samples do not permit guarantees of unbiased estimation. In practice, the estimates of $\mathbf{\hat{p}}$ and $\mathbf{\hat{L '}}$ will suffer from both imprecision and bias. The development of bootstrapMRI algorithm intends to approach the ideal of unbiased $\hat{Q^c}$ estimation, given the component estimators available. For practical approximations to unbiased component estimators, we estimate $\mathbf{\hat{p}}$ and $\mathbf{\hat{L '}}$ by the $5$-nn classifier and by cross-validation respectively. The classifier $k$-nn has well-known low asymptotic bounds on its error rate, for continuous covariates and a reasonable distance metric [@Ripley1996 Chapter 6]. These results suggest that this classifier’s probability estimates should have good statistical properties, such as reasonably low bias in the finite sample case. The estimation of $\mathbf{\hat{L '}}$ is nearly unbiased for cross-validation [@Efron1983].
The class probability vector $\mathbf{p}$ is a component of $Q^c$, which raises a question for $\hat{Q^c}$ estimation, of whether $\mathbf{\hat{p}}$ estimates need to be precise for reasonable $\hat{Q^c}$ estimation. The argument of Section \[subsection:Unbiased $Q^c$ Estimation Yields AL Performance Matching or Exceeding RS\], and the experimental results of bootstrapMRI in Section \[section:Experiments and Results\], both suggest that the $\mathbf{\hat{p}}$ estimates do not need to be very precise, but should merely have reasonably low bias.
The computational cost of EfeLc at each selection step is given by $t_a = (t_r + t_p) + (n_p \, k \, (t_r + t_l))$, where $n_p = |X_P|$ is the size of the pool, $k$ is the number of classes, $t_r$ is the cost of classifier retraining, $t_p$ is the cost of classifier prediction and $t_l$ is the cost of classifier loss estimation. The cost for simpleMRI is the same cost as EfeLc, except that the $\hat{L}$-estimation method differs and hence $t_l$ is different. The cost for bootstrapMRI is $n_b$ times that of simpleMRI, where $n_b$ is the number of bootstrap resamples.
Experiments and Results {#section:Experiments and Results}
=======================
A large-scale experimental study explores the performance of the new $Q^c$-estimation AL methods. The intention is to compare those methods with each other, and to standard AL methods from the literature (described in Section \[subsection:Literature Review\]). The focus is on the relative classifier improvements of each AL method, rather than absolute classifier performance.
The base classifier is varied, since AL performance is known to depend substantially on the classifier [@Guyon2011; @Evans2013]. To provide model diversity, the study uses several classifiers with different capabilities: LDA, $5$-nn, naïve Bayes, SVM, QDA and Logistic Regression. The classifiers and their implementations are described in Appendix C. Many different classification problems are explored, including real and simulated data, described in Appendix D. These problems are divided into three problem groups to clarify the results, see Section \[subsection:Assembly of Aggregate Results\]. The experimental study uses error rate for the loss function $L$ (see Section \[subsection:Classification\]). Further results are available for another loss function, the H measure, but are omitted for space[^1]; the choice of loss function does not affect the primary conclusion of Section \[subsection:Primary Results\]. The experimental study explores several sources of variation: the AL algorithms, the classifier $\theta$, and the classification problem $({\bf X}, Y)$.
Active Learning Methods in the Experiment {#subsection:Active Learning Methods}
-----------------------------------------
The experimental study evaluates many AL methods, to compare their performance across a range of classification problems. These methods fall into three groups: RS as the natural benchmark of AL, standard AL methods from the literature, and algorithms estimating $Q^c$. The second group consists of four standard AL methods: SE, QbcV, QbcA, and EfeLc (all described in Section \[subsection:Literature Review\]). The third group contains the two $Q^c$-estimation algorithms, simpleMRI and bootstrapMRI, defined in Section \[section:Algorithms to Estimate Example Quality\] and abbreviated as SMRI and BMRI. For the two Qbc methods, a committee of four classifiers is chosen for model diversity: logistic regression, $5$-nn, $21$-nn, and random forest. Random forest is a non-parametric classifier described in @Breiman2001; the other classifiers are described in Appendix C. This committee is arbitrary, but diverse; the choices of committee size and constitution are open research problems.
Density weighting is sometimes recommended in the AL literature, see @Olsson2009. However, the effects of density weighting are not theoretically understood. The experimental study also generated results from density weighting, omitted due to space[^2], which left unaltered the primary conclusion that the $Q^c$-estimation algorithm bootstrapMRI is very competitive with standard methods from the literature. The issue of density weighting is deferred to future work.
Experimental AL Sandbox
-----------------------
Iterated AL provides for the exploration of AL performance across the whole learning curve, see Section \[subsection:Active Learning\] and @Guyon2011 [@Evans2013]. In this experimental study, the AL iteration continues until the entire pool has been labelled. The pool size is chosen such that when all of the pool has been labelled, the final classifier loss is close to its asymptotic loss (that asymptotic loss being the loss from training on a much larger dataset). The AL performance metrics described below examine the entire learning curve. Each single realisation of the experiment has a specific context: a classification problem, and a base classifier. The classification data is randomly reshuffled. To examine variation, multiple Monte Carlo replicates are realised; ten replicates are used for each specific context. Given this experimental context, the experimental AL sandbox then evaluates the performance of all AL methods over a single dataset, using iterated AL. Each AL method produces a learning curve that shows the overall profile of loss as the number of labelled examples increases. The amount of initial labelled data is chosen to be close to the number of classes $k$. To illustrate, Figure \[figure:Single result for simulated data 1\] shows the learning curve for several AL methods, for a single realisation of the experiment.
![Result for a single experiment of iterated AL. Each AL method performs multiple selection steps, generating a set of losses that define the learning curve. For clarity, a smoothed representation of the data is presented. The early part of the learning curve is shown. The classification problem is the Four-Gaussian problem (see Figure \[fig:Ripley Four-Gaussian Problem\] and Appendix D), with the base classifier being $5$-nn. []{data-label="figure:Single result for simulated data 1"}](Ripley_Single_Sim_Result1___Copy2.pdf)
Assessing Performance {#subsection:Assessing Performance}
---------------------
As discussed in Section \[subsection:Active Learning\], AL performance metrics assess the relative improvements in classifier performance, when comparing one AL method against another (or when comparing AL against RS). Thus the real quantity of interest is the ranking of the AL methods.
The AL literature provides a selection of metrics to assess AL performance, such as AUA [@Guyon2011], WI [@Evans2013] and label complexity [@Dasgupta2011]. The experimental study evaluates four metrics: AUA, WI with two weighting functions (exponential with $\alpha=0.02$, and linear), and label complexity (with $\epsilon=5$). Each of these four metrics is a function of the learning curve, creating a single numeric summary from the learning curve, this curve being generated by iterated AL. The *overall rank* is also calculated as the ranking of the mean ranks, as employed, for example, by @Brazdil2000. This yields five AL metrics in total: four primary metrics (label complexity, AUA, WI-linear, WI-exponential) and one aggregate metric (overall rank). The overall rank avoids any arbitrary choice of one single metric. In this experimental study, AL performance is assessed by overall rank, as used in @Brazdil2000.
For a single experiment, there is a single classification problem and base classifier. In such an experiment, all five metrics are evaluated for every AL method, so that each metric produces its own ranking of the AL methods. Since there are seven AL methods (see Section \[subsection:Active Learning Methods\]), the ranks fall between one and seven, with some ties. For brevity, the tables show the best six methods, chosen by overall rank.
The experimental results show that the AL metrics substantially agree on AL method ranking (see Tables \[table:Single Pairing Result-NEW 1\] and \[table:Problem Set Result-NEW 1\]). This agreement suggests that the results are reasonably insensitive to the choice of AL metric.
Aggregate Results {#subsection:Assembly of Aggregate Results}
-----------------
To address the variability of AL, multiple Monte Carlo draws are conducted for each classification problem. Thus for each experiment, the labelled, pool and test data are drawn from the population, as different independent subsamples. This random sampling addresses two primary sources of variation, namely the initially labelled data and the unlabelled pool.
[|c|c|c|c|c|c|c|]{}\
& **BMRI** & QbcV & QbcA & EfeLc & **SMRI** & RS\
& 1 & 2 & 3 & 4 & 5 & 6\
& 1 & 2 & 4 & 5 & 6 & 3\
& 1 & 2 & 4 & 3 & 5 & 6\
& 2 & 1 & 3 & 5 & 4 & 6\
& 1 & 2 & 3 & 5 & 4 & 5\
The experimental study examines many Monte Carlo draws, classification problems in groups, and classifiers. The goal here is to determine the relative performance of the AL methods, namely to discover which methods perform better than others, on average across the whole experimental study. To that end, the aggregate results are calculated by averaging. First the losses are averaged, over Monte Carlo replicates. From those losses, AL metrics are calculated, which imply overall rankings. Finally those overall rankings are averaged, over classification problems, and then over problem groups, and finally over classifiers.
[|c|c|c|c|c|c|c|]{}\
& **BMRI** & **SMRI** & EfeLc & SE & RS & QbcV\
& 1 & 2 & 3 & 4 & 5 & 6\
& 5 & 7 & 4 & 6 & 3 & 1\
& 1 & 2 & 3 & 4 & 5 & 6\
& 2 & 1 & 3 & 4 & 5 & 6\
& 2 & 1 & 3 & 4 & 5 & 6\
For a single pairing of classifier and problem, there are ten Monte Carlo replicates. Consider the true distribution of AL metric scores for each AL method, where the source of the variation is the random sampling. The performance of each AL method is encapsulated in the score distribution, which is summarised here by the mean. The set of mean scores implies a performance ranking of the AL methods. These rankings are then averaged to produce a final overall ranking. Integer rankings of the AL methods are shown for clarity. The frequency with which each AL method outperforms random selection is also of great interest, and calculated from the group-classifier rankings.
[|c|c|c|c|c|c|c|]{}\
& QbcV & QbcA & **BMRI** & SE & **SMRI** & RS\
& SE & **BMRI** & **SMRI** & QbcA & QbcV & RS\
& **BMRI** & QbcV & **SMRI** & SE & RS & QbcA\
& **BMRI** & QbcV & SE & **SMRI** & QbcA & RS\
To summarise the aggregate result calculations:
- For a single problem-classifier pairing, the average losses are calculated, over the ten Monte Carlo replicates. This averaging of the losses reduces the variability in the learning curve. From these average losses, four AL metric numbers are calculated, leading to five rankings of the AL methods, see Table \[table:Single Pairing Result-NEW 1\].
- For a single group-classifier pairing, the overall rankings of all problem-classifier pairings in the group are averaged, see Table \[table:Problem Set Result-NEW 1\].
- For a single classifier, the overall rankings for all three group-classifier pairings are averaged, see Table \[table:LDA Result-NEW Original 1\] (and Tables \[table:KNN Result-NEW Original 1\]-\[table:LogReg Result-NEW Original 1\] in Appendix E).
- Finally, the overall rankings for all six classifiers are averaged, see Table \[table:AggOverSixClassifiers Result-NEW Original 1\].
- The frequency counts show how often each AL method outperforms RS. These are calculated from the group-classifier rankings (18 in total), see Table \[table:AggOverSixClassifiers\_CompVsRs Result-NEW Original 1\]. For example, Table \[table:LDA Result-NEW Original 1\] shows BMRI and SE outperforming RS three times out of three, whereas QbcA only outperforms RS twice.
Thus the aggregate results are calculated by averaging over successive levels, one level at a time: over Monte Carlo replicates, over problems within a group, over groups, and finally over classifiers. This successive averaging is shown by the progression from specific realisations to the whole experiment, which starts at Figure \[figure:Single result for simulated data 1\], then moves through Tables \[table:Single Pairing Result-NEW 1\] to \[table:AggOverSixClassifiers Result-NEW Original 1\] inclusive[^3].
Results {#subsection:Primary Results}
-------
The overall performance of the AL methods is summarised by the final ranking, shown in Table \[table:AggOverSixClassifiers Result-NEW Original 1\], and the frequency of outperforming RS, given in Table \[table:AggOverSixClassifiers\_CompVsRs Result-NEW Original 1\]. These two tables provide the central results for the experimental study.
[|c|c|c|c|c|c|c|c|]{}\
& Rank 1 & Rank 2 & Rank 3 & Rank 4 & Rank 5 & Rank 6 & Rank 7\
& **BMRI** & SE & QbcV & QbcA & RS & **SMRI** & EfeLc\
[|c|c|c|c|c|c|c|]{}\
& Rank 1 & Rank 2 & Rank 3 & Rank 4 & Rank 5 & Rank 6\
& **BMRI** & SE & QbcV & **SMRI** & QbcA & EfeLc\
& 15 & 14 & 13 & 9 & 8 & 2\
The primary conclusion is that the $Q^c$-motivated bootstrapMRI algorithm performs well in comparison to the standard AL methods from the literature. This conclusion holds true over different classifiers and different classification problems. Table \[table:AggOverSixClassifiers\_CompVsRs Result-NEW Original 1\] suggests that just three methods consistently outperform RS: bootstrapMRI, SE and QbcV. BootstrapMRI outperforms RS fifteen times out of eighteen. This provides experimental confirmation for the argument that unbiased $Q^c$ estimation algorithms consistently outperform RS, given in Section \[subsection:Unbiased $Q^c$ Estimation Yields AL Performance Matching or Exceeding RS\].
Comparing the $Q^c$-estimation algorithms against each other, the algorithm bootstrapMRI outperforms the algorithm simpleMRI, in all cases except two. This suggests that minimising bias in $Q^c$ estimation may be important for AL performance.
Examining the AL methods from the literature, QBC and SE consistently perform well. For QBC, vote entropy (QbcV) mostly outperforms average Kullback-Leibler divergence (QbcA). EfeLc performs somewhat less well, perhaps because of the way it approximates loss using the unlabelled pool (see Section \[subsection:Literature Review\]). For most classifiers, RS performs poorly, with many AL methods providing a clear benefit; SVM is the exception here, where RS performs best overall. The detailed results for each individual classifier are given in Appendix E.
Section \[section:Algorithms to Estimate Example Quality\] describes the difficulty of estimating the $Q^c$ components, ${\bf p}$ and $\boldsymbol{L '}$, from small data samples. For the practical algorithms, the estimates of $\boldsymbol{\hat{p}}$ and $\boldsymbol{\hat{L '}}$ will suffer from imprecision and bias. The experimental results show that despite these estimation difficulties, strong AL performance can still be achieved.
Conclusion
==========
Model retraining improvement is a novel statistical framework for AL, which characterises optimal behaviour via classifier loss. This approach is both theoretical and practical, giving new insights into AL, and competitive AL algorithms for applications. The MRI statistical estimation framework begins with the targets $Q^c$ and $B^c$. These quantities define optimal AL behaviour for the contexts of pool-based AL: individual and batch, single-step and iterated. Exploring the abstract definition of optimal AL behaviour generates new insights into AL. For a particular abstract problem, the optimal selection is examined and compared to known AL methods, revealing exactly how heuristics can make suboptimal choices. The framework is used to show that an unbiased $Q^c$ estimator will outperform random selection, bringing a new guarantee for AL.
The MRI framework motivates the construction of new algorithms to estimate $Q^c$. A comprehensive experimental study compares the performance of $Q^c$-estimation algorithms to several standard AL methods. The results demonstrate that bootstrapMRI is strongly competitive across a range of classifiers and problems, and is recommended for practical use. There are many more statistical choices for $Q^c$-estimation algorithms. These choices include various methods to estimate the class probability $\mathbf{\hat{p}}$ (e.g. via the base classifier, a different classifier, or a classifier committee); different methods to estimate the loss $\mathbf{L'}$ (e.g. in-sample, cross-validation, bootstrap, or via the unlabelled pool); and many further ways to use the data (e.g. full reuse, resampling, or partitioning). More sophisticated estimators are the subject of future research, and hopefully MRI will motivate others to develop superior estimators.
The estimation framework enables reasoning about AL consistency behaviour and stopping rules, which are the subject of future work. The MRI framework opens the door to potential statistical explanations of AL heuristics such as SE and Qbc, whose experimental effectiveness may otherwise remain mysterious.
Appendix A. {#appendix-a. .unnumbered}
===========
This Appendix shows that given a zero-mean univariate Gaussian RV denoted $N$ with variance $\sigma$, a positive constant $\delta$, a fixed-sized interval $[0, \delta)$, and the probability $\beta = p(0 \leq N < \delta)$, then as $\sigma^2 \uparrow \infty$, $\beta \downarrow 0$. $N$ is Gaussian with mean zero, hence it has cdf $F_N(x) = \frac{1}{2} [1 + \text{erf}(\frac{x}{\sigma \sqrt{2}})]$ where $\text{erf}(y) = \frac{1}{\sqrt{\pi}} \int^y_{-y} e^{-t^2} dt$. By definition $$\begin{aligned}
\beta &= p(0 \leq N < \delta) \\
&= F_N(\delta) - F_N(0) \\
&= F_N(\delta) - \frac{1}{2} \\
&= \frac{1}{2} [1 + \text{erf}(\frac{\delta}{\sigma \sqrt{2}})] - \frac{1}{2} \\
&= \frac{1}{2} \, \text{erf}(\frac{\delta}{\sigma \sqrt{2}}).\end{aligned}$$ As $\sigma^2 \uparrow \infty$, $$\begin{aligned}
\text{erf}(\frac{\delta}{\sigma \sqrt{2}}) =& \frac{1}{\sqrt{\pi}} \int^{\frac{\delta}{\sigma \sqrt{2}}}_{-{\frac{\delta}{\sigma \sqrt{2}}}} e^{-t^2} dt \\
\downarrow& \frac{1}{\sqrt{\pi}} \int^0_0 e^{-t^2} dt \\
=& 0.\end{aligned}$$ Hence as $\sigma^2 \uparrow \infty$, $\beta \downarrow 0$.
The above argument applies to a RV $N$ and a fixed interval $[0, \delta)$, but also applies to a random interval $[0, \Delta)$ with $\Delta$ being a strictly positive RV, since the argument relies to every realisation of $\Delta$.
Appendix B. {#appendix-b. .unnumbered}
===========
This Appendix shows that given a zero-mean univariate Gaussian RV denoted $N$ with variance $\sigma$, a positive constant $\delta$, a fixed-sized interval $[0, \delta)$, and the probability $\beta = p(0 \leq N < \delta)$, then as $\sigma^2 \downarrow 0$, $\beta \uparrow \frac{1}{2}$.
By definition, $$\begin{aligned}
p(N \geq 0) = p(0 \leq N < \delta) + p(N \geq \delta), $$ i.e. $$\begin{aligned}
\frac{1}{2} = \beta + p(N \geq \delta),\end{aligned}$$ giving $$\begin{aligned}
\beta = \frac{1}{2} - p(N \geq \delta).\end{aligned}$$
By definition $$\begin{aligned}
p(N \geq \delta) \leq p(|N| \geq \delta),\end{aligned}$$ and Chebyshev’s Inequality gives $$\begin{aligned}
p(|N| \geq \delta) \leq \frac{\sigma^2}{\delta^2},\end{aligned}$$ hence $$\begin{aligned}
p(N \geq \delta) \leq \frac{\sigma^2}{\delta^2}.\end{aligned}$$ As $\sigma^2 \downarrow 0$, $\frac{\sigma^2}{\delta^2} \downarrow 0$ hence $p(N \geq \delta) \downarrow 0$. Thus as $\sigma^2 \downarrow 0$, $p(N \geq \delta) \downarrow 0$, combining with $\beta = \frac{1}{2} - p(N \geq \delta)$ yields $\beta \uparrow \frac{1}{2}$ as $\sigma^2 \downarrow 0$.
The above argument applies to a RV $N$ and a fixed interval $[0, \delta)$, but also applies to a random interval $[0, \Delta)$ with $\Delta$ being a strictly positive RV, since the argument relies to every realisation of $\Delta$.
Appendix C. {#appendix-c. .unnumbered}
===========
This Appendix describes the six classifiers used in the experimental study of Section \[section:Experiments and Results\], and their implementation details.
Section \[section:Experiments and Results\] describes experiments with six classifiers: linear discriminant analysis, quadratic discriminant analysis, $K$-nearest-neighbours, naïve Bayes, logistic regression and support vector machine. Linear discriminant analysis (LDA) is a linear generative classifier described in @Tibshirani2009 [Chapter 4]. Quadratic discriminant analysis (QDA) is a non-linear generative classifier described in @Tibshirani2009 [Chapter 4]. $K$-Nearest-Neighbours ($K$-nn) is a well-known non-parametric classifier discussed in @Duda2001 [Chapter 4]. Naïve Bayes is a probabilistic classifier which assumes independence of the covariates, given the class; see @Hand2001. Logistic Regression is a linear parametric discriminative classifier described in @Tibshirani2009 [Chapter 4]. The support vector machine (SVM) is a popular non-parametric classifier described in @Vapnik1995. Standard R implementations are used for these classifiers, see below. The classifier implementation details are now described. For LDA, the standard R implementation is used. For QDA, the standard R implementation is used. For $5$-nn, the R implementation from package kknn is used.[^4] This implementation applies covariate scaling: each covariate is scaled to have equal standard deviation (using the same scaling for both training and testing data). For naïve Bayes, the R implementation from package e1071 is used.[^5] For continuous predictors, a Gaussian distribution (given the target class) is assumed. This approach is less than ideal, but tangential to the statistical estimation framework and experimental study. For Logistic Regression, the Weka implementation from package RWeka is used.[^6] For SVM, the R implementation from package e1071 is used.The SVM kernel used is radial basis kernel. The probability calibration of the scores is performed for binary problems by MLE fitting of a logistic distribution to the decision values, or for multi-class problems, by computing the a-posteriori class probabilities using quadratic optimisation.
Appendix D. {#appendix-d. .unnumbered}
===========
A diverse set of classification problems is chosen to explore AL performance. The classification problems fall into two sets: real problems and abstract problems.
First the real data classification problems are shown in Tables \[table:Small Real Data 1\] and \[table:Large Real Data 1\]. The real data problems are split into two groups, one for smaller problems of fewer examples, and another of larger problems. The class prior is shown, since the experimental study uses error rate as loss. The sources for this data include UCI [@Bache2013], @Guyon2011, @Anagnostopoulos2012 and @Adams2010.
The intention here is to provide a wide variety in terms of problem properties: covariate dimension $d$, number of classes $k$, the class prior $\boldsymbol{\pi}$, and the underlying distribution. The number and variety of problems suggests that the results in Section \[section:Experiments and Results\] have low sensitivity to the presence or absence of one or two specific problems.
Name Dim. $d$ Classes $k$ Cases $n$ Class Prior $\boldsymbol{\pi}$
------ ---------- ------------- ----------- ---------------------------------
14 2 690 (0.44, 0.56)
4 3 625 (0.08, 0.46, 0.46)
10 6 214 (0.33,0.36,0.08,0.06,0.04,0.14)
13 2 270 (0.65, 0.44)
6 2 432 (0.5, 0.5)
6 2 432 (0.5, 0.5)
6 2 432 (0.5, 0.5)
8 2 768 (0.35, 0.65)
60 2 208 (0.47, 0.53)
13 3 178 (0.33, 0.4, 0.27)
: Real Data Classification Problems, Smaller[]{data-label="table:Small Real Data 1"}
Name Dim. $d$ Classes $k$ Cases $n$ Class Prior $\boldsymbol{\pi}$
------ ---------- ------------- ----------- --------------------------------
20 2 5999 (0.167, 0.833)
6 2 27552 (0.585, 0.415)
16 2 17076 (0.406, 0.594)
29 2 4406 (0.007, 0.993)
29 2 8493 (0.091, 0.909)
: Real Data Classification Problems, Larger[]{data-label="table:Large Real Data 1"}
Second the abstract classification problems are illustrated in Figure \[fig:Contour graphs to show the tasks\]. These abstract problems are generated by sampling from known probability distributions. The class-conditional distributions $({\bf X} | y = c_j)_1^k$ are either Gaussians or mixtures of Gaussians. This set of problems presents a variety of decision boundaries to the classifier. All have balanced uniform priors, and the Bayes Error Rates are approximately 0.1.
\_Problem\_Textwidth\_Multiplier [0.19]{}
[\_Problem\_Textwidth\_Multiplier ]{} ![Density contour plots showing the abstract classification problems. The class-conditional distributions $({\bf X} | y = c_j)_1^k$ are shown in red for class 1 and blue for class 2. These class-conditional distributions $({\bf X} | y = c_j)_1^k$ are either Gaussians or mixtures of Gaussians. The decision boundary is shown in green. []{data-label="fig:Contour graphs to show the tasks"}](Ripley1.pdf "fig:"){width="\textwidth"}
[\_Problem\_Textwidth\_Multiplier ]{} ![Density contour plots showing the abstract classification problems. The class-conditional distributions $({\bf X} | y = c_j)_1^k$ are shown in red for class 1 and blue for class 2. These class-conditional distributions $({\bf X} | y = c_j)_1^k$ are either Gaussians or mixtures of Gaussians. The decision boundary is shown in green. []{data-label="fig:Contour graphs to show the tasks"}](Tgm.pdf "fig:"){width="\textwidth"}
[\_Problem\_Textwidth\_Multiplier ]{} ![Density contour plots showing the abstract classification problems. The class-conditional distributions $({\bf X} | y = c_j)_1^k$ are shown in red for class 1 and blue for class 2. These class-conditional distributions $({\bf X} | y = c_j)_1^k$ are either Gaussians or mixtures of Gaussians. The decision boundary is shown in green. []{data-label="fig:Contour graphs to show the tasks"}](sd2.pdf "fig:"){width="\textwidth"}
[\_Problem\_Textwidth\_Multiplier ]{} ![Density contour plots showing the abstract classification problems. The class-conditional distributions $({\bf X} | y = c_j)_1^k$ are shown in red for class 1 and blue for class 2. These class-conditional distributions $({\bf X} | y = c_j)_1^k$ are either Gaussians or mixtures of Gaussians. The decision boundary is shown in green. []{data-label="fig:Contour graphs to show the tasks"}](sd3.pdf "fig:"){width="\textwidth"}
[\_Problem\_Textwidth\_Multiplier ]{} ![Density contour plots showing the abstract classification problems. The class-conditional distributions $({\bf X} | y = c_j)_1^k$ are shown in red for class 1 and blue for class 2. These class-conditional distributions $({\bf X} | y = c_j)_1^k$ are either Gaussians or mixtures of Gaussians. The decision boundary is shown in green. []{data-label="fig:Contour graphs to show the tasks"}](sd7.pdf "fig:"){width="\textwidth"}
Appendix E. {#appendix-e. .unnumbered}
===========
This Appendix shows the results for each individual classifier in the experimental study described in Section \[section:Experiments and Results\]. The results for LDA, $K$-nn, naïve Bayes, SVM, QDA and Logistic Regression are shown in Tables \[table:LDA Result-NEW Original 1\], \[table:KNN Result-NEW Original 1\], \[table:Naive Bayes Result-NEW Original 1\], \[table:SVM Result-NEW Original 1\], \[table:QDA Result-NEW Original 1\] and \[table:LogReg Result-NEW Original 1\] respectively. These results are the detailed results of the experimental study, covering the six classifiers, all the problems in three groups, and multiple Monte Carlo replicates. In each table, the average rank is calculated as the numerical mean, with ties resolved by preferring lower variance of the rank vector.
[|c|c|c|c|c|c|c|]{}\
& SE & **BMRI** & QbcV & **SMRI** & RS & QbcA\
& QbcA & **BMRI** & SE & QbcV & RS & **SMRI**\
& **BMRI** & SE & RS & QbcV & **SMRI** & QbcA\
& **BMRI** & SE & QbcV & RS & QbcA & **SMRI**\
[|c|c|c|c|c|c|c|]{}\
& SE & **BMRI** & QbcV & QbcA & RS & **SMRI**\
& QbcV & SE & EfeLc & **BMRI** & **SMRI** & QbcA\
& SE & QbcV & **BMRI** & **SMRI** & RS & QbcA\
& SE & QbcV & **BMRI** & **SMRI** & QbcA & RS\
[|c|c|c|c|c|c|c|]{}\
& QbcV & RS & QbcA & SE & **BMRI** & **SMRI**\
& RS & QbcA & QbcV & EfeLc & **SMRI** & **BMRI**\
& QbcV & RS & QbcA & **BMRI** & **SMRI** & SE\
& RS & QbcV & QbcA & **BMRI** & **SMRI** & SE\
[|c|c|c|c|c|c|c|]{}\
& SE & **BMRI** & QbcV & QbcA & **SMRI** & RS\
& **BMRI** & **SMRI** & EfeLc & SE & RS & QbcV\
& SE & **BMRI** & RS & QbcV & QbcA & **SMRI**\
& **BMRI** & SE & QbcV & **SMRI** & RS & QbcA\
[|c|c|c|c|c|c|c|]{}\
& QbcV & QbcA & **BMRI** & SE & RS & **SMRI**\
& SE & QbcV & QbcA & **SMRI** & **BMRI** & RS\
& **BMRI** & RS & SE & **SMRI** & QbcV & QbcA\
& QbcV & SE & **BMRI** & QbcA & RS & **SMRI**\
The results of Section \[subsection:Primary Results\] quantify the benefit of AL over RS: the rankings of Tables \[table:LDA Result-NEW Original 1\] and \[table:AggOverSixClassifiers Result-NEW Original 1\] show how much AL methods outperform RS. Another way to quantify AL benefit is provided by the regret difference between an AL method and RS. Here AL regret is naturally defined as the loss difference, between the optimal performance given by maximising $Q^c$, and the actual performance of any given AL method. Another aspect of AL benefit is the question of where AL outperforms RS, and this aspect is quantified by the frequency results in Table \[table:AggOverSixClassifiers\_CompVsRs Result-NEW Original 1\].
0.2in
[^1]: For these results see http://www.lewisevans.com/JMLR-Extra-Experimental-Results-Feb-2015.pdf.
[^2]: For these results see http://www.lewisevans.com/JMLR-Extra-Experimental-Results-Feb-2015.pdf.
[^3]: For further details see http://www.lewisevans.com/JMLR-Extra-Experimental-Results-Feb-2015.pdf.
[^4]: For details see http://cran.r-project.org/web/packages/kknn/kknn.pdf.
[^5]: For details see http://cran.r-project.org/web/packages/e1071/e1071.pdf.
[^6]: For details see http://cran.r-project.org/web/packages/RWeka/RWeka.pdf.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a new general, complete closed-form solution of the three-dimensional Stark problem in terms of Weierstrass elliptic and related functions. With respect to previous treatments of the problem, our analysis is exact and valid for all values of the external force field, and it is expressed via unique formulæ valid for all initial conditions and parameters of the system. The simple form of the solution allows us to perform a thorough investigation of the properties of the dynamical system, including the identification of quasi-periodic and periodic orbits, the formulation of a simple analytical criterion to determine the boundness of the trajectory, and the characterisation of the equilibrium points.'
author:
- |
Francesco Biscani$^1$[^1] and Dario Izzo$^2$\
$^{1}$Castle Mews 27, St. Thomas Street,\
$\hphantom{^{2}}$Oxford, Oxfordshire, OX1 1JR, United Kingdom\
$^{2}$ESA – Advanced Concepts Team, European Space Research Technology Centre (ESTEC),\
$\hphantom{^{2}}$Keplerlaan 1, Postbus 299, 2200 AG Noordwijk The Netherlands
bibliography:
- 'biblio.bib'
title: The Stark problem in the Weierstrassian formalism
---
\[firstpage\]
Celestial mechanics - Gravitation - Stark problem
Introduction
============
The dynamical system consisting of a test particle subject simultaneously to an inverse-square central field and to a force field constant both in magnitude and direction is known under multiple denominations. Historically, this system was first studied in detail in the context of particle physics (where it is known as *Stark problem* [@stark_beobachtungen_1914]) in connection with the shifting and splitting of spectral lines of atoms and molecules in the presence of an external static electric field.
In astrophysics and dynamical astronomy, the Stark problem is sometimes known as the *accelerated Kepler problem*, and it is studied in several contexts. Models based on the accelerated Kepler problem have been used to study the excitation of planetary orbits by stellar jets in protoplanetary disks and to explain the origin of the eccentricities of extrasolar planets [@namouni_origin_2005; @namouni_accelerated_2007; @namouni_excitation_2013]. The Stark problem has also been used in the study of the dynamics of dust grains in the Solar System [@belyaev_dynamics_2010; @pastor_influence_2012].
In astrodynamics, the Stark problem is relevant in connection to the *continuous-thrust problem*, describing the dynamics of spacecrafts equipped with ion thrusters. In such a context, the trajectory of the spacecraft is often considered as a series of non-Keplerian arcs resulting from the simultaneous action of the gravitational acceleration and the constant thrust provided by the engine [@sims_preliminary_1999].
From a purely mathematical perspective, the importance of the Stark problem lies mainly in fact that it belongs to the very restrictive class of Liouville-integrable dynamical systems of classical mechanics [@arnold_mathematical_1989]. Action-angle variables for the Stark problem can be introduced in a perturbative fashion, as explained in @born_mechanics_1927 and @berglund_averaged_2001.
Different types of solutions to the Stark problem are available in the literature. If the constant acceleration field is much smaller than the Keplerian attraction along the orbit of the test particle, the problem can be treated in a perturbative fashion, and the (approximate) solution is expressed as the variation in time of the Keplerian (or Delaunay) orbital elements of the osculating orbit [@vinti_effects_1966; @berglund_averaged_2001; @namouni_accelerated_2007; @belyaev_dynamics_2010; @pastor_influence_2012]. A different approach is based on regularisation procedures such as the Levi-Civita and Kustaanheimo-Stiefel transformations [@kustaanheimo_perturbation_1965; @saha_interpreting_2009], which yield exact solutions in a set of variables related to the cartesian ones through a rather complex nonlinear transformation [@kirchgraber_problem_1971; @rufer_trajectory_1976; @poleshchikov_one_2004]. A third way exploits the formulation of the Stark problem in parabolic coordinates to yield an exact solution in terms of Jacobi elliptic functions and integrals [@lantoine_complete_2011].
The aim of this paper is to introduce and examine a new solution to the Stark problem that employs the Weierstrassian elliptic and related functions. The main features of our solution can be summarised as follows:
- it is an exact (i.e., non-perturbative), closed-form and explicit solution;
- it is expressed as a set of unique formulæ independent of the values of the initial conditions and of the parameters of the system;
- it is a solution to the full three-dimensional Stark problem (whereas many previous solutions deal only with the restricted case in which the motion is confined to a plane).
The simple form of our solution allows us to examine thoroughly the dynamical features of the Stark problem, and to derive several new results (e.g., regarding questions of (quasi) periodicity and boundness of motion). Our method of solution is in some sense close to the one employed in @lantoine_complete_2011. However, we believe that our solution offers several distinct advantages:
- by adopting the Weierstrassian formalism (instead of the Jacobian one), we sidestep the issue of categorising the solutions based on the nature of the roots of the polynomials generating the differential equations, and thus our formulæ do not depend on the initial conditions or on the parameters of the system;
- we provide explicit formulæ for the three-dimensional case;
- we avoid introducing a second time transformation in the solution.
These advantages are critical in providing new insights in the dynamics of the Stark problem. On the other hand, the use of the Weierstrassian formalism presents a few additional difficulties with respect to the approach described in @lantoine_complete_2011, the most notable of which is probably the necessity of operating in the complex domain. Throughout the paper, we will highlight these difficulties and address them from the point of view of the actual implementation of the formulæ describing our solution to the Stark problem.
In this paper, we will focus our attention specifically on the full three-dimensional Stark problem, where the motion of the test particle is not confined to a plane, and we will only hint occasionally at the bidimensional case (where instead the motion is constrained to a plane).
Problem formulation
===================
From a dynamical point of view, the Stark problem is equivalent to a one-body gravitational problem with an additional force field which is constant both in magnitude and direction. The corresponding Lagrangian in cartesian coordinates $\boldsymbol{r} = \left(x,y,z\right)$ and velocities $\boldsymbol{v} = \left(\dot{x},\dot{y},\dot{z}\right)$ is then $$L\left(\boldsymbol{v};\boldsymbol{r}\right)=\frac{1}{2}v^{2}+\frac{\mu}{r}+\varepsilon z,
\label{eq:orig_lagr}$$ where the inertial coordinate system has been centred on the central body, $v = \left| \boldsymbol{v}\right|$, $r = \left| \boldsymbol{r}\right|$, $\mu$ is the gravitational parameter of the system and $\varepsilon > 0$ is the constant acceleration imparted to the test particle by the force field. Without loss of generality, the coordinate system has been oriented so that the force field is directed towards the positive $z$ axis.
Following the lead of @epstein_zur_1916 and @born_mechanics_1927, we proceed by expressing the Lagrangian in parabolic coordinates $\left( \xi, \eta, \phi \right)$ via the coordinate transformation $$\begin{aligned}
x & =\xi\eta\cos\phi, & \dot{x} & =\left(\dot{\xi}\eta+\xi\dot{\eta}\right)\cos\phi-\xi\eta\dot{\phi}\sin\phi,\label{eq:x_para}\\
y & =\xi\eta\sin\phi, & \dot{y} & =\left(\dot{\xi}\eta+\xi\dot{\eta}\right)\sin\phi+\xi\eta\dot{\phi}\cos\phi,\\
z & =\frac{\xi^{2}-\eta^{2}}{2}, & \dot{z} & =\xi\dot{\xi}-\eta\dot{\eta},\label{eq:z_para}\end{aligned}$$ where $\xi \geq 0$, $\eta \geq 0$ and $\phi \in \left( -\pi,\pi\right]$ is the azimuthal angle. The inverse transformation from cartesian to parabolic coordinates is $$\begin{aligned}
\xi & =\sqrt{r+z}, & \dot{\xi} & =\frac{\dot{r}+\dot{z}}{2\sqrt{r+z}},\label{eq:xi_inv}\\
\eta & =\sqrt{r-z}, & \dot{\eta} & =\frac{\dot{r}-\dot{z}}{2\sqrt{r-z}},\label{eq:eta_inv}\\
\phi & =\arctan\left(y,x\right), & \dot{\phi} & =\frac{\dot{y}x-\dot{x}y}{x^{2}+y^{2}}\label{eq:phi_inv},\end{aligned}$$ where $\dot{r}=\left(\boldsymbol{v}\cdot\boldsymbol{r}\right)/r$ and $\arctan$ is the two-argument inverse tangent function. In the new coordinate system, $$\begin{aligned}
v^{2} & =\left(\xi^{2}+\eta^{2}\right)\left(\dot{\xi}^{2}+\dot{\eta}^{2}\right)+\xi^{2}\eta^{2}\dot{\phi}^{2},\\
r & =\frac{\xi^{2}+\eta^{2}}{2},\end{aligned}$$ and the Lagrangian becomes $$\begin{gathered}
L=\frac{1}{2}\left[\left(\xi^{2}+\eta^{2}\right)\left(\dot{\xi}^{2}+\dot{\eta}^{2}\right)+\xi^{2}\eta^{2}\dot{\phi}^{2}\right]\\
+\frac{2\mu}{\xi^{2}+\eta^{2}}+\varepsilon\frac{\xi^{2}-\eta^{2}}{2}.\end{gathered}$$ Switching now to the Hamiltonian formulation through a Legendre transformation, the momenta are defined as $$\begin{aligned}
p_{\xi} & =\frac{\partial L}{\partial\dot{\xi}}=\left(\xi^{2}+\eta^{2}\right)\dot{\xi},\label{eq:p_xi}\\
p_{\eta} & =\frac{\partial L}{\partial\dot{\eta}}=\left(\xi^{2}+\eta^{2}\right)\dot{\eta},\label{eq:p_eta}\\
p_{\phi} & =\frac{\partial L}{\partial\dot{\phi}}=\xi^{2}\eta^{2}\dot{\phi},\label{eq:p_phi}\end{aligned}$$ and the Hamiltonian is written as $$\begin{aligned}
\mathcal{H} & =\dot{\xi}p_{\xi}+\dot{\eta}p_{\eta}+\dot{\phi}p_{\phi}-L\\
& =\frac{1}{2}\frac{p_{\xi}^{2}+p_{\eta}^{2}}{\xi^{2}+\eta^{2}}+\frac{1}{2}\frac{p_{\phi}^{2}}{\xi^{2}\eta^{2}}-\frac{2\mu}{\xi^{2}+\eta^{2}}-\varepsilon\frac{\xi^{2}-\eta^{2}}{2}.
\label{eq:Ham_def}\end{aligned}$$ Since the coordinate $\phi$ is absent from the Hamiltonian, the momentum $p_{\phi}$ is a constant of motion. It can be checked by substitution that $p_{\phi}$ is the $z$ component of the total angular momentum of the system. Thus, when $p_{\phi}$ vanishes, the motion is confined to a plane perpendicular to the $xy$ plane and intersecting the origin, and we can refer to this subcase as the *bidimensional* problem (as opposed to the *three-dimensional* problem when $p_{\phi}$ is not null).
We now employ a Sundman regularisation [@sundman_memoire_1912], introducing the fictitious time $\tau$ via the differential relation $$dt=\left(\xi^{2}+\eta^{2}\right)d\tau,\label{eq:fic_time}$$ and the new, identically null, function $$\mathcal{H}_{\tau}\left(p_\xi,p_\eta,p_\phi;\xi,\eta,\phi\right)=\left(\mathcal{H}-h\right)\left(\xi^{2}+\eta^{2}\right),$$ where $h$ is the energy constant of the system (obtained by substituting the initial conditions into the expression of $\mathcal{H}$). We have then for $p_{\xi}$ and $\xi$ $$\begin{aligned}
\frac{dp_{\xi}}{d\tau} & =\frac{dp_{\xi}}{dt}\frac{dt}{d\tau}=-\frac{\partial\mathcal{H}}{\partial\xi}\left(\xi^{2}+\eta^{2}\right)=-\frac{\partial\mathcal{H}_{\tau}}{\partial\xi},\\
\frac{d\xi}{d\tau} & =\frac{d\xi}{dt}\frac{dt}{d\tau}=\frac{\partial\mathcal{H}}{\partial p_{\xi}}\left(\xi^{2}+\eta^{2}\right)=\frac{\partial\mathcal{H}_{\tau}}{\partial p_{\xi}},\end{aligned}$$ and, similarly for $p_{\eta}$, $\eta$, $p_{\phi}$ and $\phi$, $$\begin{aligned}
\frac{dp_{\eta}}{d\tau} & = -\frac{\partial\mathcal{H}_{\tau}}{\partial\eta},\\
\frac{d\eta}{d\tau} & = \frac{\partial\mathcal{H}_{\tau}}{\partial p_{\eta}},\\
\frac{dp_{\phi}}{d\tau} & = -\frac{\partial\mathcal{H}_{\tau}}{\partial\phi},\\
\frac{d\phi}{d\tau} & = \frac{\partial\mathcal{H}_{\tau}}{\partial p_{\phi}}.\end{aligned}$$ $\mathcal{H}_\tau$ can thus be considered as an Hamiltonian function describing the evolution of the system in fictitious time[^2]. Explicitly, $$\begin{gathered}
\mathcal{H}_{\tau}=-\varepsilon\frac{\xi^{4}}{2}-h\xi^{2}+\frac{1}{2}p_{\xi}^{2}+\frac{1}{2}\frac{p_{\phi}^{2}}{\xi^{2}}\\
+\varepsilon\frac{\eta^{4}}{2}-h\eta^{2}+\frac{1}{2}p_{\eta}^{2}+\frac{1}{2}\frac{p_{\phi}^{2}}{\eta^{2}}-2\mu,\end{gathered}$$ and the Hamiltonian $\mathcal{H}_\tau$ has thus been separated into the two independent constants of motion $$\begin{aligned}
\alpha_{1} & =-\varepsilon\frac{\xi^{4}}{2}-h\xi^{2}+\frac{1}{2}p_{\xi}^{2}+\frac{1}{2}\frac{p_{\phi}^{2}}{\xi^{2}},\label{eq:alpha1}\\
\alpha_{2} & =\varepsilon\frac{\eta^{4}}{2}-h\eta^{2}+\frac{1}{2}p_{\eta}^{2}+\frac{1}{2}\frac{p_{\phi}^{2}}{\eta^{2}}.\label{eq:alpha2}\end{aligned}$$ These constants represent the conservation of a component of the generalised Runge-Lenz vector [@redmond_generalization_1964]. By inversion of $\alpha_{1}$ and $\alpha_{2}$ for $p_{\xi}$ and $p_{\eta}$, Hamilton’s equations finally yield $$\begin{aligned}
p_\xi=\frac{d\xi}{d\tau}&=\pm\frac{1}{\xi}\sqrt{\varepsilon\xi^{6}+2h\xi^{4}+2\alpha_{1}\xi^{2}-p_{\phi}^{2}},\label{eq:xi_diffeq}\\
p_\eta=\frac{d\eta}{d\tau}&=\pm\frac{1}{\eta}\sqrt{-\varepsilon\eta^{6}+2h\eta^{4}+2\alpha_{2}\eta^{2}-p_{\phi}^{2}},\label{eq:eta_diffeq}\\
\frac{d\phi}{d\tau}&=p_\phi\left(\frac{1}{\xi^2}+\frac{1}{\eta^2}\right).\label{eq:phi_diffeq}\end{aligned}$$ The solution of the Stark problem has thus been reduced to the integration by quadrature of eqs. –. Before proceeding, it is useful to outline the general features of the functions on the right-hand side of eqs. and .
Study of $p_{\xi}\left(\xi\right)$ and $p_{\eta}\left(\eta\right)$ {#subsec:poly_study}
------------------------------------------------------------------
Both $p_{\xi}\left(\xi\right)$ and $p_{\eta}\left(\eta\right)$ are functions of $\xi$ and $\eta$ symmetric with respect to both the horizontal and vertical axes. The zeroes of both functions are given by the roots of the bicubic polynomial radicands on the right-hand side of eqs. and . Hence, the number of real roots of $p_{\xi}\left(\xi\right)$ and $p_{\eta}\left(\eta\right)$ will depend on the initial conditions and on the physical parameters of the system (namely, the gravitational parameter and the value of the constant force field).
For any given set of initial conditions, it is clear that the polynomial radicand on the right-hand side of eq. will tend to $+\infty$ for $\xi\to\pm\infty$, since $\varepsilon>0$ by definition. Thus, $p_{\xi}\left(\xi\right)$ will always tend to $\pm\infty$ in the limit $\xi\to\pm\infty$. Conversely, for $\eta\to\pm\infty$, the radicand in $p_{\eta}\left(\eta\right)$ will eventually start assuming negative values, thus implying the existence of a real root. For both $p_{\xi}\left(\xi\right)$ and $p_{\eta}\left(\eta\right)$, moving along the horizontal axis towards the origin from the initial conditions means encountering another root, as for $\xi=\eta=0$ both functions result in the computation of the square root of the negative value $-p_{\phi}^{2}$. This also implies that, in the three-dimensional problem, the trajectories in the phase planes $\left(\xi,p_\xi \right)$ and $\left(\eta,p_\eta\right)$ will not cross the vertical axes, and $p_{\xi}\left(\xi\right)$ and $p_{\eta}\left(\eta\right)$ always have at least two real roots. Figure \[fig:tridimensional\_case\] shows a selection of representative trajectories in the phase space for $p_{\xi}\left(\xi\right)$ and $p_{\eta}\left(\eta\right)$ in the three-dimensional case.
The bidimensional case requires a separate analysis. When $p_{\phi}$ is null, the bicubic polynomials collapse to biquadratic polynomials (via the inclusion of the external factors $1/\xi$ and $1/\eta$). As in the three-dimensional case, the evolution of $p_{\xi}$ can be either bound or unbound, while the evolution of $p_{\eta}$ is always bound. The first difference is that, when $\alpha_{1}>0$, $p_{\xi}$ might have no real roots. Secondly, when the signs of the constants $\alpha_{1}$ and $\alpha_{2}$ are positive, $p_{\xi}$ and $p_{\eta}$ assume real values for $\xi=0$ and $\eta=0$, and the trajectories in the phase plane thus seemingly cross the vertical axes. Physically, the conditions $\xi=0$ and $\eta=0$ correspond (via eqs. and ) to polar transits (i.e., the test particle is passing through the negative or positive $z$ axis). But, according to the definition of parabolic coordinates, $\xi$ and $\eta$ are strictly non-negative quantities and thus the trajectories in the phase planes cannot enter the regions $\xi < 0$ and $\eta < 0$. In order to solve this apparent contradiction it can be shown how, in correspondence of a transit through $\xi = 0$ or $\eta = 0$, the corresponding momentum ($p_\xi$ or $p_\eta$) switches discontinuously its sign (and, concurrently, the azimuthal angle $\phi$ discontinuously changes by $\pm\pi$). In the phase plane, upon reaching the vertical axis from a positive $\xi$ or $\eta$, the trajectory will be discontinuously reflected with respect to the horizontal axis, and its evolution will proceed again towards positive $\xi$ or $\eta$. Figure \[fig:bidimensional\_case\] shows a selection of representative trajectories in the phase space for $p_{\xi}\left(\xi\right)$ and $p_{\eta}\left(\eta\right)$ in the bidimensional case.
{width="100.00000%"}
{width="100.00000%"}
We proceed now to determine the explicit solutions for $\xi\left(\tau\right)$ and $\eta\left(\tau\right)$ in the three-dimensional case. We will focus on the study of the solution for $\xi$, as the solution for $\eta$ differs only by notation. We will then use $\xi\left(\tau\right)$ and $\eta\left(\tau\right)$ to determine the solution for $\phi\left(\tau\right)$.
Solution by quadrature
======================
The integration of eq. yields $$\int_{0}^{\tau}du=\pm\int_{\xi_{0}}^{\xi}\frac{udu}{\sqrt{\varepsilon u^{6}+2hu^{4}+2\alpha_{1}u^{2}-p_{\phi}^{2}}},
\label{eq:tau_inverse}$$ where the initial fictitious time has been set to zero[^3], $u$ is a dummy integration variable and $\xi_{0}$ is the initial value of $\xi$. Before proceeding, we need to discuss briefly the nature of the sign ambiguity in this formula.
The left-hand side of eq. represents the fictitious time needed by the dynamical system to evolve from the initial coordinate $\xi_0$ to an arbitrary coordinate $\xi$. As pointed out in the previous section, all phase plots are symmetric with respect to the horizontal axis, and thus, along a trajectory in phase space, each coordinate $\xi$ will be visited twice: once with a positive $p_\xi$ coordinate, and once with a negative $p_\xi$ coordinate[^4]. It follows that we can choose either sign in , and the left-hand side will then represent the evolution time along a portion of trajectory in which $p_\xi$ remains positive ($+$) or negative ($-$).
Changing now integration variable in the right-hand side of eq. to $$s=\frac{1}{2}u^{2},$$ we can rewrite the equation as $$\begin{aligned}
\tau&=\pm\int_{\frac{1}{2}\xi_{0}^{2}}^{\frac{1}{2}\xi^{2}}\frac{ds}{\sqrt{8\varepsilon s^{3}+8hs^{2}+4\alpha_{1}s-p_{\phi}^{2}}} \label{eq:tau_t}\\
&=\pm\int_{\frac{1}{2}\xi_{0}^{2}}^{\frac{1}{2}\xi^{2}}\frac{ds}{\sqrt{f_{\xi}\left(s\right)}},\end{aligned}$$ where $f_{\xi}\left(s\right)$ is a cubic polynomial in $s$. The integral in this expression is an elliptic integral, which can be computed and inverted to yield $\xi^{2}$ as function of $\tau$ using a formula by Weierstrass [see @whittaker_course_1927 §20.6]. After electing $$\begin{aligned}
f_{\xi}\left(s\right) & =a_{4}+4a_{3}s+6a_{2}s^{2}+4a_{1}s^{3},\end{aligned}$$ and defining $$\begin{aligned}
g_{2} & =-4a_{1}a_{3}+3a_{2}^{2},\label{eq:g2_def}\\
g_{3} & =2a_{1}a_{2}a_{3}-a_{2}^{3}-a_{1}^{2}a_{4},\label{eq:g3_def}\\
\wp_{\xi}\left(\tau\right) & \equiv\wp\left(\tau;g_{2},g_{3}\right),\end{aligned}$$ where $\wp\left(\tau;g_{2},g_{3}\right)$ is a Weierstrass elliptic function defined in terms of the invariants $g_{2}$ and $g_{3}$ (see @whittaker_course_1927, Chapter XX, and @abramowitz_handbook_1964, Chapter 18), the evolution of $\xi^{2}$ in fictitious time is given by $$\begin{aligned}
\xi^{2} & =\xi_{0}^{2}+\frac{1}{\left[\wp_{\xi}\left(\tau\right)-\frac{1}{24}f_{\xi}^{\prime\prime}\left(\frac{\xi_{0}^{2}}{2}\right)\right]^{2}}\nonumber \\
& \quad\cdot\left\{ \frac{1}{2}f_{\xi}^{\prime}\left(\frac{\xi_{0}^{2}}{2}\right)\left[\wp_{\xi}\left(\tau\right)-\frac{1}{24}f_{\xi}^{\prime\prime}\left(\frac{\xi_{0}^{2}}{2}\right)\right]\vphantom{\sqrt{f_{\xi}\left(\frac{\xi_{0}^{2}}{2}\right)}}\right.\nonumber \\
& \quad\left.+\frac{1}{24}f_{\xi}\left(\frac{\xi_{0}^{2}}{2}\right)f_{\xi}^{\prime\prime\prime}\left(\frac{\xi_{0}^{2}}{2}\right)\pm\sqrt{f_{\xi}\left(\frac{\xi_{0}^{2}}{2}\right)}\wp_{\xi}^{\prime}\left(\tau\right)\right\}.\label{eq:xi_tau}\end{aligned}$$ Here the $\pm$ sign represents the sign ambiguity discussed earlier, and the derivatives of $f_\xi$ are calculated with respect to the polynomial variable, while the derivative $\wp_\xi^\prime$ is calculated with respect to $\tau$. $\wp_\xi^\prime$ is related to $\wp$ via the relation $$\left[\wp_\xi^{\prime}\left( z \right)\right] ^2 = 4 \wp_\xi^3\left( z \right) - g_2 \wp_\xi\left( z \right) - g_3 \label{eq:Pprime_rel}$$ [@abramowitz_handbook_1964 eq. 18.1.6]. If $\xi_{0}^{2}/2$ is chosen as a root $\xi_{r}^{2}/2$ of $f_{\xi}$, then $f_\xi\left( \xi_r^2/2 \right) = 0$ and eq. simplifies to $$\xi^{2}=\xi_{r}^{2}+\frac{1}{2}\frac{f_{\xi}^{\prime}\left(\frac{\xi_{r}^{2}}{2}\right)}{\wp_{\xi}\left(\tau-\tau_{\xi}\right)-\frac{1}{24}f_{\xi}^{\prime\prime}\left(\frac{\xi_{r}^{2}}{2}\right)},
\label{eq:xi_simpl}$$ where $\tau_{\xi}$ is the fictitious time for which $\xi$ assumes the value $\xi_{r}$. The analogous expressions for $\eta$ are $$\begin{aligned}
\eta^{2} & =\eta_{0}^{2}+\frac{1}{\left[\wp_{\eta}\left(\tau\right)-\frac{1}{24}f_{\eta}^{\prime\prime}\left(\frac{\eta_{0}^{2}}{2}\right)\right]^{2}}\nonumber \\
& \quad\cdot\left\{ \frac{1}{2}f_{\eta}^{\prime}\left(\frac{\eta_{0}^{2}}{2}\right)\left[\wp_{\eta}\left(\tau\right)-\frac{1}{24}f_{\eta}^{\prime\prime}\left(\frac{\eta_{0}^{2}}{2}\right)\right]\vphantom{\sqrt{f_{\eta}\left(\frac{\eta_{0}^{2}}{2}\right)}}\right.\nonumber \\
& \quad\left.+\frac{1}{24}f_{\eta}\left(\frac{\eta_{0}^{2}}{2}\right)f_{\eta}^{\prime\prime\prime}\left(\frac{\eta_{0}^{2}}{2}\right)\pm\sqrt{f_{\eta}\left(\frac{\eta_{0}^{2}}{2}\right)}\wp_{\eta}^{\prime}\left(\tau\right)\right\} \label{eq:eta_tau}\end{aligned}$$ and $$\eta^{2}=\eta_{r}^{2}+\frac{1}{2}\frac{f_{\eta}^{\prime}\left(\frac{\eta_{r}^{2}}{2}\right)}{\wp_{\eta}\left(\tau-\tau_{\eta}\right)-\frac{1}{24}f_{\eta}^{\prime\prime}\left(\frac{\eta_{r}^{2}}{2}\right)}.
\label{eq:eta_simpl}$$ The formulæ and represent a general and complete closed-form solution for the squares $\xi^2$ and $\eta^2$ of the parabolic coordinates $\xi$ and $\eta$. Since $\xi$ and $\eta$ are non-negative by definition, in order to recover the solution for $\xi$ and $\eta$ it will be enough to take the principal square root of $\xi^2$ and $\eta^2$. The cartesian positions and velocities can be reconstructed using eqs. -, where the derivatives of the parabolic coordinates with respect to the real time can be computed by inverting eqs. - (and by keeping in mind that $p_\xi$ and $p_\eta$ can be calculated by differentiating eqs. and with respect to $\tau$ – see eqs. and ).
For simplicity’s sake, notational convenience and further analysis, however, it is desirable to be able to use the simplified formulæ and whenever possible. To this end, we first note how, from the considerations presented in the previous section, the polynomials $f_\xi$ and $f_\eta$ will always have at least one positive real root, with the exception of the bidimensional case for $f_\xi$ with $\alpha_1 > 0$ (displayed in Figure \[fig:bidimensional\_case\] (b)). The roots $\xi_r$ and $\eta_r$ of the cubic polynomials $f_\xi$ and $f_\eta$ can be computed exactly using the general formulæ for the roots of a cubic function. Secondly, in order to determine $\tau_\xi$ (and, analogously, $\tau_\eta$) we can use eq. to write: $$\tau_\xi=\pm\int_{\frac{1}{2}\xi_{0}^{2}}^{\frac{1}{2}\xi_r^{2}}\frac{ds}{\sqrt{8\varepsilon s^{3}+8hs^{2}+4\alpha_{1}s-p_{\phi}^{2}}}.
\label{eq:depressed_transform}$$ Following then @byrd_handbook_1971 [eqs. (A7)–(A13)], we introduce the Tschirnhaus transformation [@cayley_tschirnhausens_1861] $$s = \sqrt[3]{\frac{1}{2\varepsilon}}s_1 - \frac{1}{3}\frac{h}{\varepsilon}$$ in order to reduce the polynomial $f_\xi$ to a depressed cubic: $$\tau_\xi=\pm\sqrt[3]{\frac{1}{2\varepsilon}}\int_{\sqrt[3]{2\varepsilon}\left(\frac{1}{2}\xi_0^2+\frac{1}{3}\frac{h}{\varepsilon}\right)}^
{\sqrt[3]{2\varepsilon}\left(\frac{1}{2}\xi_r^2+\frac{1}{3}\frac{h}{\varepsilon}\right)}
\frac{ds_1}{
\sqrt{4s_{1}^{3}-h_2s_{1}-h_3}},$$ where $$\begin{aligned}
h_2 & = \sqrt[3]{\frac{1}{2\varepsilon}}\left(\frac{8}{3}\frac{h^{2}}{\varepsilon}-4\alpha_{1}\right), \\
h_3 & = \frac{4}{3}\frac{\alpha_{1}h}{\varepsilon}-\frac{16}{27}\frac{h^{3}}{\varepsilon^{2}}+p_{\phi}^{2}.\end{aligned}$$ The integral can now be split into two separate Weierstrass normal elliptic integrals of the first kind, $$\begin{gathered}
\tau_\xi=\pm\sqrt[3]{\frac{1}{2\varepsilon}}\left[
\int_{\sqrt[3]{2\varepsilon}\left(\frac{1}{2}\xi_0^2+\frac{1}{3}\frac{h}{\varepsilon}\right)}^{\infty}
\frac{ds_1}{\sqrt{4s_{1}^{3}-h_2s_{1}-h_3}}
\right.\\
\left.
-\int_{\sqrt[3]{2\varepsilon}\left(\frac{1}{2}\xi_r^2+\frac{1}{3}\frac{h}{\varepsilon}\right)}^{\infty}
\frac{ds_1}{\sqrt{4s_{1}^{3}-h_2s_{1}-h_3}}
\right],\end{gathered}$$ and solved in terms of the inverse Weierstrass elliptic function $\wp^{-1}$ as $$\begin{gathered}
\tau_\xi=\pm\sqrt[3]{\frac{1}{2\varepsilon}}\left\{
\wp^{-1}\left[\sqrt[3]{2\varepsilon}\left(\frac{1}{2}\xi_0^2+\frac{1}{3}\frac{h}{\varepsilon}\right);h_2,h_3\right]\right.\\
\left.-\wp^{-1}\left[\sqrt[3]{2\varepsilon}\left(\frac{1}{2}\xi_r^2+\frac{1}{3}\frac{h}{\varepsilon}\right);h_2,h_3\right]
\right\}.\label{eq:tau_xi}\end{gathered}$$ The corresponding formula for $\tau_\eta$ can be obtained by switching $\varepsilon$ to $-\varepsilon$ and $\alpha_1$ to $\alpha_2$. It must be noted though that in this formula there are ambiguities regarding the computation of the inverse Weierstrass elliptic function, as $\wp^{-1}\left(z\right)$ is a multivalued function[^5]. The values of $\wp^{-1}$ in eq. have then to be chosen appropriately in order to yield the correct result [as explained, e.g., in @hoggatt_inverse_1955]. As an alternative, it is possible to compute directly the integral in eq. in terms of Legendre elliptic integrals using known formulæ [e.g., @gradshtein_table_2007 §3.131 and §3.138].
The solution for the third coordinate $\phi$ can now be computed directly by integrating eq. with respect to $\tau$: $$\int_{\phi_0}^\phi du = p_\phi\left[\int_0^\tau\frac{du}{\xi^2\left( u \right)} + \int_0^\tau\frac{du}{\eta^2\left( u \right)}\right].
\label{eq:phi_inteq}$$ It is easier to tackle the calculation via the simplified formulæ and . The integrals on the right-hand side of eq. are then in a form which can be solved through a formula involving $\wp^\prime$, $\wp^{-1}$ and the Weierstrass $\sigma$ and $\zeta$ functions (see @jules_tannery_elements_1893, chapter CXII, and @gradshtein_table_2007, §5.141.5): $$\int \frac{\wp\left( u \right) + \beta}{\gamma\wp\left( u \right) + \delta}du =
\frac{u}{\gamma}+\frac{\delta-\beta\gamma}{\gamma^2\wp^\prime\left(v\right)}\left[\ln\frac{\sigma
\left( u+v \right)}{\sigma\left( u-v \right)} - 2u\zeta\left( v \right)\right],
\label{eq:phi_integral}$$ where $v=\wp^{-1}\left(-\delta / \gamma\right)$. In this case, the multivalued character of $\wp^{-1}$ does not matter: it can be verified via the reduction formulæ of the Weierstrassian functions [@abramowitz_handbook_1964 §18.2] that *any* value of $v$ such that $\wp\left( v \right) = -\delta/\gamma$ will yield the same result in the right-hand side of eq. . The final result for $\phi$ is then[^6]: $$\begin{gathered}
\phi = \phi_0 + 2p_\phi\left\{\tau\left(\frac{1}{\gamma_\xi}+\frac{1}{\gamma_\eta}\right)+\frac{\delta_\xi-\beta_\xi\gamma_\xi}{\gamma_\xi^2\wp_\xi^\prime\left(u_\xi\right)}\right.\\
\cdot\left[\ln\frac{\sigma_\xi\left(\tau-\tau_\xi+u_\xi\right)}{\sigma_\xi\left(\tau-\tau_\xi-u_\xi\right)}
-\ln\frac{\sigma_\xi\left(-\tau_\xi+u_\xi\right)}{\sigma_\xi\left(-\tau_\xi-u_\xi\right)}-2\tau\zeta_\xi\left( u_\xi \right)\right]\\
+\frac{\delta_\eta-\beta_\eta\gamma_\eta}{\gamma_\eta^2\wp_\eta^\prime\left(u_\eta\right)}\\
\left.\cdot\left[\ln\frac{\sigma_\eta\left(\tau-\tau_\eta+u_\eta\right)}{\sigma_\eta\left(\tau-\tau_\eta-u_\eta\right)}
-\ln\frac{\sigma_\eta\left(-\tau_\eta+u_\eta\right)}{\sigma_\eta\left(-\tau_\eta-u_\eta\right)}
-2\tau\zeta_\eta\left( u_\eta \right)\right]
\vphantom{\frac{\tau}{\gamma_\xi}+\frac{\delta_\xi-\beta_\xi\gamma_\xi}{\gamma_\xi^2\wp_\xi^\prime\left(u_\xi\right)}}
\right\},\label{eq:phi_tau}\end{gathered}$$ where the following constants have been defined for notational convenience: $$\begin{aligned}
\beta_\xi & = -\frac{1}{24}f_{\xi}^{\prime\prime}\left(\frac{\xi_{r}^{2}}{2}\right),&\beta_\eta & = -\frac{1}{24}f_{\eta}^{\prime\prime}\left(\frac{\eta_{r}^{2}}{2}\right),\\
\gamma_\xi & = 2\xi_r^2, & \gamma_\eta & = 2\eta_r^2,\\
\delta_\xi & = f_\xi^\prime\left(\frac{\xi_r^2}{2}\right)+2\xi_r^2\beta_\xi,&\delta_\eta & = f_\eta^\prime\left(\frac{\eta_r^2}{2}\right)+2\eta_r^2\beta_\eta,\\
u_\xi&=\wp_\xi^{-1}\left(-\frac{\delta_\xi}{\gamma_\xi}\right),&u_\eta&=\wp_\eta^{-1}\left(-\frac{\delta_\eta}{\gamma_\eta}\right).\end{aligned}$$ Starting from eq. , we adopt the subscript notation $\sigma_\xi$ and $\zeta_\xi$ to indicate Weierstrass $\sigma$ and $\zeta$ functions defined in terms of the same invariants as $\wp_\xi$.
The time equation {#sec:time_equation}
=================
The final step in the solution of the Stark problem is to establish an explicit connection between real and fictitious time. To this end, we need to integrate eq. : $$dt=\left[\xi^{2}\left(\tau\right)+\eta^{2}\left(\tau\right)\right]d\tau.$$ In the general case, according to eqs. and , the exact solutions for $\xi^{2}\left(\tau\right)$ and $\eta^{2}\left(\tau\right)$ are of the form $$A+B\wp^{\prime}\left(\tau\right),\label{eq:wp_rational}$$ where A and B are rational functions of $\wp\left(\tau\right)$. Then, according to the theory of elliptic functions, the antiderivative of can be calculated in terms of $\wp$, $\wp^\prime$, $\wp^{-1}$ and the Weierstrass $\sigma$ and $\zeta$ functions. The integration method, due to @halphen_traite_1886 [chapter VII] (and explained in detail in @greenhill_applications_1959 [chapter VII]), involves the decomposition of $A$ and $B$ into separate fractions, resulting in the split of the integral into fundamental forms that can be integrated using the Weierstrassian functions.
It is again easier to use the simplified solutions and , and thus obtain the time equation $$\begin{aligned}
t &= \int_0^\tau\left[\xi^{2}\left(u\right)+\eta^{2}\left(u\right)\right]du \\
& = \left(\xi_r^2+\eta_r^2\right)\tau + \frac{1}{2}\int_0^\tau
\frac{f_{\xi}^{\prime}\left(\frac{\xi_{r}^{2}}{2}\right)}{\wp_{\xi}\left(u-\tau_{\xi}\right)
-\frac{1}{24}f_{\xi}^{\prime\prime}\left(\frac{\xi_{r}^{2}}{2}\right)}du\notag\\
&\quad +\frac{1}{2}\int_0^\tau
\frac{f_{\eta}^{\prime}\left(\frac{\eta_{r}^{2}}{2}\right)}{\wp_{\eta}\left(u-\tau_{\eta}\right)
-\frac{1}{24}f_{\eta}^{\prime\prime}\left(\frac{\eta_{r}^{2}}{2}\right)}du.\label{eq:tau_01}\end{aligned}$$ The integrals appearing in eq. are known and they can be computed directly. To this end, it would be tempting to apply the formulæ in @gradshtein_table_2007 [§5.141]. However, as it can be verified by direct substitution using the exact solution of the cubic equation[^7], $\frac{1}{24}f_{\xi}^{\prime\prime}\left(\frac{\xi_{r}^{2}}{2}\right)$ and $\frac{1}{24}f_{\eta}^{\prime\prime}\left(\frac{\eta_{r}^{2}}{2}\right)$ are always roots of the characteristic cubic equations $$4t^3 - g_2t - g_3 = 0,\label{eq:char_cubic}$$ associated to $\wp_\xi$ and $\wp_\eta$. Consequently, the formulæ in @gradshtein_table_2007 [§5.141] will be singular, and we have to use instead the results in @jules_tannery_elements_1893 [§CXII], which yield the formula $$\int\frac{du}{\wp\left(u\right)-e_i} = \frac{1}{g_2/4-3e_i^2}\left[ue_i + \zeta\left(u-\omega_i\right)\right].$$ In this formula, the $e_i$ represents the three roots of the characteristic cubic equation, while the $\omega_i$ are defined by the relation $e_i=\wp\left(\omega_i\right)$ (so that, following @abramowitz_handbook_1964 [eq. 18.3.1], two of the $\omega_i$ are the fundamental half-periods of $\wp$ and the third one is the sum of the fundamental half-periods). The solution of eq. is thus: $$\begin{gathered}
t = \left(\xi_r^2+\eta_r^2\right)\tau+\frac{1}{2}\frac{f_{\xi}^{\prime}\left(\frac{\xi_{r}^{2}}{2}\right)}{g_{2,\xi}/4-3e_{i,\xi}^2}\cdot
\\\left[\tau e_{i,\xi} + \zeta_\xi\left(\tau-\tau_\xi-\omega_{i,\xi}\right)-\zeta_\xi\left(-\tau_\xi-\omega_{i,\xi}\right)\right]+\\
\frac{1}{2}\frac{f_{\eta}^{\prime}\left(\frac{\eta_{r}^{2}}{2}\right)}{g_{2,\eta}/4-3e_{i,\eta}^2}\cdot\\
\left[\tau e_{i,\eta} + \zeta_\eta\left(\tau-\tau_\eta-\omega_{i,\eta}\right)-\zeta_\eta\left(-\tau_\eta-\omega_{i,\eta}\right)\right].
\label{eq:time_equation}\end{gathered}$$ This equation can be considered as the equivalent of Kepler’s equation for the Stark problem. Similarly to the two-body problem, it is constituted of a linear part modulated by two quasi-periodic transcendental parts (with the Weierstrass $\zeta$ function replacing the sine function appearing in Kepler’s equation). In this sense, the fictitious time $\tau$ can be regarded as a kind of eccentric anomaly for the Stark problem. According to eq. , the time equation is a monotonic function and its inversion can thus be achieved numerically using standard techniques (Newton-Raphson, bisection, etc.).
Analysis of the results
=======================
After having determined the full formal solution of the Stark problem in the previous sections, we now turn our attention to the interpretation of the results.
Before proceeding, we first need to point out how our solution to the Stark problem, as developed in the previous sections, is directly applicable to the three-dimensional case, but not in general to all bidimensional cases. As explained in §\[subsec:poly\_study\], in certain bidimensional cases (specifically, when the constant of motion $\alpha_1$ is positive) the polynomial $p_\xi\left(\xi\right)$ might have no real zeroes, and thus the simplified formula cannot be used. While in this case it is still possible to proceed to a complete solution via the full formula in conjunction with the general theory for the integration of rational functions of elliptic functions (see @halphen_traite_1886 [chapter VII] and @greenhill_applications_1959 [chapter VII]), the resulting expressions for $\phi\left(\tau\right)$ and $t\left(\tau\right)$ will be more complicated than the formulæ obtained for the three-dimensional case.
An additional complication in the bidimensional case is the presence of the discontinuity discussed in §\[subsec:poly\_study\]. In correspondence of a polar transit, either $p_\xi$ or $p_\eta$ will switch sign. This discontinuity must be taken into account in the computation and inversion of the integral , and ultimately it has the effect of introducing a branching in the solutions for $\xi\left(\tau\right)$ and $\eta\left(\tau\right)$.
Quasi-periodicity and periodicity
---------------------------------
Our solution to the Stark problem is based on the Weierstrass elliptic and related functions. Without giving a full account of the theory of the Weierstrassian functions (for which we refer to standard textbooks such as @whittaker_course_1927), we will recall here briefly a few fundamental notions[^8]. To this end, we will employ the notation of @abramowitz_handbook_1964 [chapter 18].
The elliptic function $\wp\left( z;g_2,g_3\right)$ is a doubly-periodic complex-valued function of a complex variable $z$ defined in terms of two complex parameters $g_2$ and $g_3$, called *invariants*. The complex primitive half-periods $\omega$ and $\omega^\prime$ of $\wp$ can be related to the invariants via formulæ involving elliptic integrals and the roots $e_1$, $e_2$ and $e_3$ of the characteristic cubic equation $$4t^3-g_2t-g_3 = 0 \label{eq:cubic_wp}$$ [e.g., see @abramowitz_handbook_1964 §18.9]. The sign of the *modular discriminant* $$\Delta = g_2^3-27g_3^2$$ determines the nature of the roots $e_1$, $e_2$ and $e_3$. In the case of the Stark problem, the invariants are by definition real (see eqs. –), and thus the $(\omega,\omega^\prime)$ pairs can be chosen as (real, imaginary) or complex conjugate (depending on the sign of $\Delta$). It is known from the theory of elliptic functions that there actually exist infinite pairs of fundamental half-periods for $\wp$, related to each other via integral linear combinations with unitary determinant [@hancock_lectures_1910 §79]. We can then always introduce two new half-periods $\omega_R$ (the *real* period) and $\omega_C$ (the *complex* period) such that $\omega_R$ is real and positive, and $\omega_C$ complex with positive imaginary part. The relation with the fundamental half-periods $\omega$ and $\omega^\prime$ from @abramowitz_handbook_1964 is $$\begin{aligned}
\omega_R & = \omega + \delta \omega^\prime,\\
\omega_C & = \omega^\prime,\end{aligned}$$ where $\delta = 0$ if $\Delta > 0$ and $\delta = 1$ if $\Delta < 0$. Since we are interested in the behaviour of $\wp$ on the real axis (as $\tau$ is a real-valued variable), we can then regard $\wp\left( \tau;g_2,g_3\right)$ as a singly-periodic real-valued function of period $2\omega_R$.
It follows then straightforwardly from eqs. – that $\xi\left( \tau \right)$ and $\eta\left( \tau \right)$ are both periodic in $\tau$ with periods that, in general, will be different. Conversely, from eq. , it follows immediately that $\phi\left( \tau \right)$ is not periodic. Indeed, $\phi\left( \tau \right)$ is a function of the form $$f\left(\tau\right)=A+B\tau+C_\xi\ln\frac{\sigma_\xi\left(\tau + a_\xi\right)}{\sigma_\xi\left(\tau + b_\xi\right)}
+C_\eta\ln\frac{\sigma_\eta\left(\tau + a_\eta\right)}{\sigma_\eta\left(\tau + b_\eta\right)},\label{eq:generic_f_sigma}$$ where $A$, $B$, $C$, $a$ and $b$ are constants. It is now interesting to note that, according to eqs. and , if $\xi$ and $\eta$ have real half-periods $\omega_{R,\xi}$ and $\omega_{R,\eta}$ such that $$\frac{\omega_{R,\xi}}{\omega_{R,\eta}}=\frac{n}{m},$$ with $n$ and $m$ coprime natural numbers (or, in other words, $\omega_{R,\xi}$ and $\omega_{R,\eta}$ are commensurable), then $d\phi/d\tau$ becomes a periodic function with period $T = 2m\omega_{R,\xi} = 2n\omega_{R,\eta}$. Recalling the quasi-periodicity of $\sigma$ via the relation [@abramowitz_handbook_1964 eq. 18.2.20] $$\begin{gathered}
\sigma\left(z+2M\omega+2N\omega^\prime\right) = \left(-1\right)^{M+N+MN}\sigma\left(z\right)\cdot \\
\mathrm{e}^{\left(z+M\omega+N\omega^\prime\right)\left[2M\zeta\left(\omega\right)+2N\zeta\left(\omega^\prime\right)\right]},\end{gathered}$$ with $M,N\in\mathbb{Z}$, we can then write for eq. $$\begin{gathered}
f\left(\tau + T\right) = A+B\tau+C_\xi\ln\frac{\sigma_\xi\left(\tau + a_\xi\right)}{\sigma_\xi\left(\tau + b_\xi\right)}
+C_\eta\ln\frac{\sigma_\eta\left(\tau + a_\eta\right)}{\sigma_\eta\left(\tau + b_\eta\right)}\\
+ BT + 2 m C_\xi\left(a_\xi - b_\xi \right)\zeta_\xi\left(\omega_{R,\xi}\right)\\
+ 2 n C_\eta\left(a_\eta - b_\eta \right)\zeta_\eta\left(\omega_{R,\eta}\right),\end{gathered}$$ or, more succinctly, $$f\left(\tau + T\right) = f\left(\tau\right) + D,$$ where $D$ is the constant $$\begin{gathered}
D = BT + 2 m C_\xi\left(a_\xi - b_\xi \right)\zeta_\xi\left(\omega_{R,\xi}\right)\\
+ 2 n C_\eta\left(a_\eta - b_\eta \right)\zeta_\eta\left(\omega_{R,\eta}\right).\end{gathered}$$ Thus, if $\xi\left( \tau \right)$ and $\eta\left( \tau \right)$ have commensurable periods, $\phi\left( \tau \right)$ is an arithmetic quasi-periodic function of $\tau$. The geometric meaning of this quasi-periodicity is that, after a quasi-period $T$, the test particle will be in a position that results from a rotation around the $z$ axis of the original position. The particle’s trajectory will thus draw a rotationally-symmetric figure in space.
Quasi-periodic orbits can be found via a numerical search for a set of initial conditions and constant acceleration field $\varepsilon$ that satisfies the commensurability relation on the periods of $\xi$ and $\eta$. The numerical search can be setup as the minimisation of the function $\left(m\omega_{R,\xi}-n\omega_{R,\eta}\right)^2$ for two chosen coprime integers $n$ and $m$. A representative quasi-periodic orbit found this way using the PaGMO optimiser [@biscani_global_2010] is displayed in Figure \[fig:quasi\_periodic\].
Periodic orbits can also be found in a similar way by imposing the additional condition $p \phi(T) = 2\pi$, where $p \in \mathbb Z$. For any triplet of $\left( n,m,p \right)$ integers, one has then to solve numerically an optimisation problem that yields periodic orbits such as the one displayed in Figure \[fig:periodic\] for a case $n=1$, $m = 2$, and $p = 7$.
{width="100.00000%"}
{width="100.00000%"}
Bound and unbound orbits
------------------------
The solution of the Stark problem in terms of the Weierstrassian functions allows to determine the conditions under which the motion is bound. As we have seen in the previous sections, the parabolic coordinate $\eta$ is always bound, whereas $\xi$ can be either bound or unbound. From the general solution , it is easily deduced that the formula for $\xi\left(\tau\right)$ has a pole (and thus $\xi$ is unbound) when the denominator is zero, i.e., under the condition $$\wp_{\xi}\left(\tau\right)-\frac{1}{24}f_{\xi}^{\prime\prime}\left(\frac{\xi_{0}^{2}}{2}\right) = 0.$$ Recalling now that $\wp_\xi\left(\tau\right)$ is analytical everywhere except at the poles (where it behaves like $1/\tau^2$ around $\tau = 0$), it can be deduced from the properties of parity and periodicity that $\wp_\xi\left(\tau\right)$ must have a global minimum within the real period $2\omega_R$. Moreover, since $\wp_\xi$ satisfies the differential equation , the condition for the existence of a stationary point is $$\wp_\xi\left(\tau\right) = e_i,$$ where $e_i$ represents the roots of the cubic equation . It is known [@abramowitz_handbook_1964 eq. 18.3.1] that $\wp_\xi\left(\omega_i\right) = e_i$, where $$\begin{aligned}
\omega_1 &= \omega,\\
\omega_2 &= \omega + \omega^\prime,\\
\omega_3 &= \omega^\prime,\end{aligned}$$ which implies that the global minimum of $\wp_\xi\left(\tau\right)$ is in correspondence of $\tau = \omega_R$. We can then conclude that the condition for bound motion is $$e_R > \frac{1}{24}f_{\xi}^{\prime\prime}\left(\frac{\xi_{0}^{2}}{2}\right),$$ where we have denoted with $e_R$ the root of the cubic equation for which $\wp_\xi\left(\omega_R\right) = e_R$. Figure \[fig:plot\_3d\] displays the evolution of two representative bound orbits in the three-dimensional space.
{width="100.00000%"}
Figure \[fig:bound\_vs\_unbound\] displays the evolution in $\tau$ of the parabolic coordinates and of the real time $t$ in a bound and an unbound case. It is interesting to note that in the unbound case only $\xi$ and $t$ present vertical asymptotes, whereas $\eta$ and $\phi$ assume finite values when $\xi$ and $t$ go to infinity. With respect to the evolution in real time $t$, this means that $\eta$ and $\phi$ tend asymptotically to finite values for $t\to \infty$. At infinity, the trajectory of the test particle is determined solely by the constant acceleration field and will thus be a parabola. The plane in which such asymptotic parabola lies is perpendicular to the $xy$ plane and its orientation is determined by the value to which the azimuthal angle $\phi$ tends asymptotically (which can be determined exactly by calculating the value of $\phi$ at the end of one period in fictitious time). This result could prove to be particularly useful in the design of powered planetary kicks (or flybys), a technique vastly used in modern interplanetary trajectory design [@danby_fundamentals_1988]. Planetary kicks are traditionally designed assuming an unperturbed hyperbolic motion around a certain planet. The outgoing conditions are then simply determined by the analytical expression governing Keplerian motion (i.e., a rotation of the hyperbolic access velocity). A different type of powered flyby can be considered, in which the spacecraft thrusts continuously in a fixed inertial direction. In such a case, and ignoring the fuel mass loss, the spacecraft conditions at infinity (i.e., when leaving the planet’s sphere of influence) can be determined exactly by a fully-analytical solution such as the one presented here.
{width="100.00000%"}
Equilibrium points and displaced circular orbits
------------------------------------------------
We turn now our attention to the analysis of the equilibrium points of the Stark problem. It is useful to consider initially the Hamiltonian in cartesian coordinates and real time $t$ resulting from the Lagrangian . The equations of motion are, trivially, $$\begin{aligned}
\frac{dx}{dt}&=v_x,&\frac{dv_x}{dt}&=-\frac{\mu x}{r^3},\\
\frac{dy}{dt}&=v_y,&\frac{dv_y}{dt}&=-\frac{\mu y}{r^3},\\
\frac{dz}{dt}&=v_z,&\frac{dv_z}{dt}&=-\frac{\mu z}{r^3}+\varepsilon.\end{aligned}$$ The only equilibrium point for this system is for $v_x=v_y=v_z=x=y=0$ and $z=\sqrt{\mu / \varepsilon}$. That is, the test particle is stationary on the positive $z$ axis at a distance from the origin such that the Newtonian attraction and the external acceleration field counterbalance each other. We refer to this unstable critical point as the *cartesian stationary equilibrium*.
Back in parabolic coordinates and fictitious time $\tau$, a first straightforward observation is that the cartesian stationary equilibrium cannot be handled in this coordinate system, as it corresponds to a position in which the azimuthal angle $\phi$ is undefined. Secondly, since $d\phi/d\tau$ is a monotonic function according to , it follows that there cannot be a parabolic stationary equilibrium point, and that only the coordinates $\xi$ and $\eta$ can be in a stationary point. From the definition we can deduce how a trajectory in which $\xi$ is constant is constrained to a circular paraboloid symmetric with respect to the $z$ axis and defined by the equation $$z = \frac{\xi_0^4-x^2-y^2}{2\xi_0^2},$$ resulting from the inversion of eq. . Similarly, a trajectory with constant $\eta$ will be constrained to the paraboloid defined by $$z = \frac{x^2+y^2-\eta_0^4}{2\eta_0^2}$$ (via inversion of eq. ).
It is then interesting to note how a trajectory in which both $\xi$ and $\eta$ are constant will be constrained to the intersection of two coaxial circular paraboloids with opposite orientation. That is, the trajectory will follow a circle centred on the $z$ axis and parallel to the $xy$ plane. Additionally, according to eqs. and , such a circular trajectory will have constant angular velocity both in fictitious and real time. Such orbits are known in the literature as *static orbits* [@forward_statite_1991], *displaced circular orbits* [@dankowicz_special_1994; @lantoine_complete_2011], *displaced non-Keplerian orbits* [@mcinnes_dynamics_1998], or *sombrero orbits* [@namouni_accelerated_2007].
From a physical point of view, displaced circular orbits are possible when the initial conditions satisfy the following requirements:
- the distance from the $xy$ plane is such that the net force acting on the test particle is perpendicular to the $z$ axis (i.e., the total force has zero $z$ component),
- the initial velocity vector is lying on the plane $\Pi$ of the displaced circular orbit, it is perpendicular to the projection of the position vector on $\Pi$ and its magnitude has the same value it would assume in a circular Keplerian orbit with a fictitious central body lying in correspondence of the $z$ axis on the $\Pi$ plane (where the mass of the fictitious body is generating the total force experienced by the test particle).
In other words, with these initial conditions the test particle evolves along a Keplerian planar circular orbit under the influence of a fictitious body lying on the positive $z$ axis. These requirements are satisfied by the following cartesian initial conditions: $$\begin{aligned}
\boldsymbol{r}_0 & = \left(\sqrt{\left(\frac{z\mu}{\varepsilon} \right)^\frac{2}{3}-z^2},0,z \right),\label{eq:dco_init_cond}\\
\boldsymbol{v}_0 & = \left(0,\sqrt{\frac{\varepsilon}{z}\left[\left(\frac{z\mu}{\varepsilon} \right)^\frac{2}{3}-z^2\right]},0\right),\label{eq:dco_init_vel}\end{aligned}$$ where $z>0$ and where we have taken advantage of the cylindrical symmetry of the problem by choosing, without loss of generality, a set of initial conditions on the $xz$ plane. It is clear from eqs. and that there exist a limit on the value of $z$ after which displaced circular orbits are not possible because the radicand in the expression for the $x$ coordinate becomes negative. Physically, this means that the gravitational force cannot counterbalance the constant acceleration field in the $z$ direction. This limit value is clearly in correspondence of the cartesian stationary equilibrium.
From a mathematical point of view, a displaced circular orbit must turn the solutions $\xi\left(\tau\right)$ and $\eta\left(\tau\right)$ into constants. From eqs. and it is clear that these expressions can become constants only when $f_{\xi}^{\prime}\left(\frac{\xi_{r}^{2}}{2}\right)$ and $f_{\eta}^{\prime}\left(\frac{\eta_{r}^{2}}{2}\right)$ are zero. This condition is equivalent to the requirement that the two polynomials $f_\xi$ and $f_\eta$ have roots of multiplicity greater than one. From the point of view of the theory of dynamical systems, the two polynomials need to have roots of multiplicity greater than one because otherwise the zeroes of the differential equations and are in correspondence of a point in which the equations lose their properties of differentiability and Lipschitz continuity, and the resulting equilibria are thus spurious.
It can be verified by direct substitution that the initial conditions and , after the transformation into parabolic coordinates, are roots of both the characteristic polynomials $f_\xi$ and $f_\eta$ and of their derivatives. Our solution in terms of Weierstrassian functions is thus consistent with known results [e.g., see @namouni_accelerated_2007] regarding the existence and characterisation of the equilibrium points in the Stark problem.
Conclusions
===========
In this paper we introduced a new solution to the Stark problem based on Weierstrass elliptic and related functions. Our treatment yields an exact (i.e., non-perturbative) and explicit solution of the full three-dimensional problem in terms of a set of unique formulæ valid for all initial conditions and physical parameters of the system. Formally, the result is remarkably similar to the solution of the two-body problem: the evolution of the coordinates is given as a function of an anomaly (or, a fictitious time) connected to the real time by a transcendental equation.
The simplicity of our formulation allows us to derive several new results. In particular, we were able to formulate conditions for the existence of quasi-periodic and periodic orbits, and to successfully identify instances of (quasi) periodic orbits using numerical techniques. We were also able to formulate a new simple analytical criterion to study the boundness of the motion, a result that can be particularly interesting for astrodynamical applications (e.g., in the study of the ejection of dust grains in the outer Solar System – see @belyaev_dynamics_2010 and [@pastor_influence_2012]). Another result of astrodynamical interest (in connection to the design of powered flyby manoeuvres) is the identification of an analytical formula for the determination of the orientation of the asymptotic planes of motion at infinity in case of unbound orbits.
Our analysis shows how the Weierstrassian formalism can be fruitfully applied to yield a new insight in the dynamics of the Stark problem. We hope that our results will contribute to revive the interest in this beautiful and powerful mathematical tool.
Acknowledgements {#acknowledgements .unnumbered}
================
F. Biscani would like to thank Dr. Santiago Nicolas Lopez Carranza for helpful discussion, and E. S. for providing the motivation to complete the manuscript.
The authors would also like to thank the reviewers, Prof. Ryan Russell, Noble Hatten and Nick Bradley, for their insightful input and suggestions during the review process.
Implementation details
======================
Implementation of the Weierstrassian functions
----------------------------------------------
The Weierstrassian functions are not as readily available in scientific computation packages as other special functions. Following @abramowitz_handbook_1964 [§18.9 and §18.10], it is possible to express them in terms of Jacobi elliptic and theta functions. The recipes in @abramowitz_handbook_1964 do not present particular difficulties in terms of implementation details. A minor complication is that the cases in which the Weierstrass invariant $g_3$ is negative are transformed in non-negative $g_3$ via the homogeneity relation $$\wp\left(z;g_2,g_3\right) = -\wp\left(\imath z;g_2,-g_3\right)\label{eq:homo_g3}$$ (and similar relations hold for $\zeta$ and $\sigma$). This transformation is not problematic for the computation of the values of the functions, but it needs to be properly taken into account when computing auxiliary quantities such as the half-periods and the roots of the characteristic cubic equations.
Regarding the half-periods, it is seen from eq. how the effect of the homogeneity relation is that of a rotation of the half-periods by $-\pi/2$ in the complex plane (via the $\imath$ factor applied to the argument $z$ on the right-hand side). The half-periods can then be first calculated in the transformed non-negative $g_3$ case, and afterwards they can be rotated back to obtain the original half-periods.
Regarding the roots of the characteristic polynomial $$y = 4x^3 - g_2x - g_3,$$ one can see how a change in sign in $g_3$ corresponds to a reflection with respect to both the $x$ and $y$ axes. The net effect will thus be equivalent to a simple change of the sign of all roots.
For the actual implementation of the Weierstrassian functions, we used the elliptic functions module of the multiprecision Python library *mpmath* [@mpmath].
On the computation of the complex logarithms in equation {#sec:complex_log}
---------------------------------------------------------
The solution for the evolution of the $\phi$ coordinate in fictitious time, eq. , involves, in the general case, the computation of complex logarithms. Since the complex logarithm is a multivalued function, care must be taken in order to select values that yield physically meaningful solutions.
The standard way of proceeding when dealing with complex logarithms is to restrict the computation to the principal value ${ \operatorname{Log}}$ of the logarithm, i.e., the unique value whose imaginary part lies in the interval $\left(-\pi,\pi\right]$. In doing so, if one takes the logarithm of a complex function whose values cross the negative real axis (i.e., the branch cut of ${ \operatorname{Log}}$), a discontinuity will arise – the imaginary part of the logarithm of the function will jump from $\pi$ to $-\pi$ (or vice versa). In the case of the Stark problem, this means that $\phi\left(\tau\right)$ will be discontinuous. These discontinuities are merely an artefact of the way of choosing a particular logarithm value among all the possible ones, and they need to be dealt with in order to produce a physically correct (i.e., continuous) solution.
We start by recalling the following series expansion for the logarithm of $\sigma$ [@jules_tannery_elements_1893 §CVI]: $$\begin{gathered}
{ \operatorname{Log}}\sigma\left(u\right)={ \operatorname{Log}}\frac{2\omega_{R}}{\pi}+\frac{\eta_Ru^{2}}{2\omega_{R}}+{ \operatorname{Log}}\sin\frac{\pi u}{2\omega_{R}}\\
+\sum_{r=1}^{\infty}\frac{q^{2r}}{r\left(1-q^{2r}\right)}\left(2\sin\frac{r\pi u}{2\omega_{R}}\right)^{2},\label{eq:log_exp}\end{gathered}$$ where $\eta_{R}=\zeta\left(\omega_{R}\right)$ and $q=\exp\left(\imath\pi\frac{\omega_{C}}{\omega_{R}}\right)$, and $u$ is decomposed into its components along the fundamental periods as $$u=2\alpha\omega_{R}+2\beta\omega_{C},$$ with $\alpha,\beta\in\mathbb{R}$. This series expansion is convergent for $\left|\beta\right|<1$, or, in other words, as long as $u$ is confined to the strip in the complex plane defined by $\left|\Im\left(u\right)\right|<2\Im\left(\omega_{C}\right)$.
We turn now to the study of the behaviour of the series expansion within the real period $2\omega_{R}$ and in the positive half of the strip of convergence. That is, we study the behaviour of the series expansion of ${ \operatorname{Log}}\sigma\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)\right]$, with $x_{\ast}$ as a real variable in the interval $\left[0,2\omega_{R}\right)$ and $0<\beta<1$. We first note that, from eq. , there exists a potential discontinuity in the computation of the complex logarithm $${ \operatorname{Log}}\sin\frac{\pi\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)\right]}{2\omega_{R}},\label{eq:log_sin}$$ when its argument crosses the negative real axis. However, by applying elementary trigonometric identities, we can write $$\begin{aligned}
\Re\left\{ \sin\frac{\pi\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)\right]}{2\omega_{R}}\right\} & =\sin\frac{\pi x_{\ast}}{2\omega_{R}}\cosh\frac{\pi\beta\Im\left(\omega_{C}\right)}{\omega_{R}},\label{eq:R_log}\\
\Im\left\{ \sin\frac{\pi\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)\right]}{2\omega_{R}}\right\} & =\cos\frac{\pi x_{\ast}}{2\omega_{R}}\sinh\frac{\pi\beta\Im\left(\omega_{C}\right)}{\omega_{R}}.\end{aligned}$$ That is, the argument of the logarithm in crosses the real axis when $x_{\ast}=\omega_{R}$. But then, for $x_{\ast}=\omega_{R}$, the real part of the argument of the logarithm is strictly positive (as the hyperbolic cosine is a strictly positive function), and hence the crossing of the real axis does not happen in correspondence of the branch cut of the principal value of the logarithm. This means that, for $x_{\ast}\in\left[0,2\omega_{R}\right)$, the series expansion of ${ \operatorname{Log}}\sigma\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)\right]$ is a continuous function.
Outside the interval $\left[0,2\omega_{R}\right)$, we can represent a variable $x\in\mathbb{R}$ as $x=x_{\ast}+2N\omega_{R}$, where $N\in\mathbb{Z}$. Recalling now the definition of the Weierstrass sigma function [@greenhill_applications_1959 §195], we can write $$\begin{gathered}
\sigma\left[x+\imath 2\beta\Im\left(\omega_{C}\right)\right]=\sigma\left[x_{\ast}+2N\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)\right]\\
=\exp\left\{ { \operatorname{Log}}\left[x_{\ast}+2N\omega_{R}+\imath 2\beta\Im\left(\omega_{C}\right)\right]+
\vphantom{\int_{0}^{x_{\ast}+2N\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)}\left[\zeta\left(z\right)-\frac{1}{z}\right]dz}
\right.\\
\left.\int_{0}^{x_{\ast}+2N\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)}\left[\zeta\left(z\right)-\frac{1}{z}\right]dz\right\}.\label{eq:sigma_def}\end{gathered}$$ We can split the integral in eq. as $$\begin{gathered}
\int_{0}^{x_{\ast}+2N\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)}\left[\zeta\left(z\right)-\frac{1}{z}\right]dz\\
=\int_{0}^{x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)}\left[\zeta\left(z\right)-\frac{1}{z}\right]dz\\
+\int_{x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)}^{x_{\ast}+2N\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)}
\left[\zeta\left(z\right)-\frac{1}{z}\right]dz,
\label{eq:three_integrals}\end{gathered}$$ and, following [@jules_tannery_elements_1893 §CXVII], the third integral in eq. can be computed as $$\begin{gathered}
\int_{x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)}^{x_{\ast}+2N\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)}\left[\zeta\left(z\right)-\frac{1}{z}\right]dz\\
={ \operatorname{Log}}\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)\right]-{ \operatorname{Log}}\left[x_{\ast}+2N\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)\right]\\
+2N\eta_{R}\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)+N\omega_{R}\right]-\imath N\pi.\end{gathered}$$ In other words, $$\begin{gathered}
{ \operatorname{Log}}\sigma\left[x+\imath2\beta\Im\left(\omega_{C}\right)\right]={ \operatorname{Log}}\sigma\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)\right]\\
+2N\eta_{R}\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)+N\omega_{R}\right]-\imath N\pi,\label{eq:final_log_sigma}\end{gathered}$$ which corresponds to the homogeneity relation in @abramowitz_handbook_1964 [§18.2]. Since, as we have seen, ${ \operatorname{Log}}\sigma\left[x_{\ast}+\imath2\beta\Im\left(\omega_{C}\right)\right]$ is a continuous function, the only possible discontinuities in eq. are in the neighbourhood of $x=2N\omega_{R}$, where $x_{\ast}$ changes discontinuously by $\pm2\omega_{R}$ and $N$ by $\pm1$. For $x=2N\omega_{R}$, $x_{\ast}$ is zero and the limit from the right is $$\begin{gathered}
L^{+}=\lim_{x\to\left(2N\omega_{R}\right)^{+}}{ \operatorname{Log}}\sigma\left[x+\imath2\beta\Im\left(\omega_{C}\right)\right]\\
={ \operatorname{Log}}\sigma\left[\imath2\beta\Im\left(\omega_{C}\right)\right]+2N\eta_{R}\left[\imath2\beta\Im\left(\omega_{C}\right)+N\omega_{R}\right]-\imath N\pi.\end{gathered}$$ The limit from the left instead is $$\begin{gathered}
L^{-}=\lim_{x\to\left(2N\omega_{R}\right)^{-}}{ \operatorname{Log}}\sigma\left[x+\imath2\beta\Im\left(\omega_{C}\right)\right]\\
={ \operatorname{Log}}\sigma\left[2\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)\right]\\
+2\left(N-1\right)\eta_{R}\left[\imath2\beta\Im\left(\omega_{C}\right)+\left(N+1\right)\omega_{R}\right]-\imath\left(N-1\right)\pi.\end{gathered}$$ By using the series expansion , we can write $$\begin{gathered}
L^{+}={ \operatorname{Log}}\frac{2\omega_{R}}{\pi}+\frac{\eta_{R}\left[\imath2\beta\Im\left(\omega_{C}\right)\right]^{2}}{2\omega_{R}}+
{ \operatorname{Log}}\sin\frac{\imath\pi\beta\Im\left(\omega_{C}\right)}{\omega_{R}}\\
+\sum_{r=1}^{\infty}\frac{q^{2r}}{r\left(1-q^{2r}\right)}\left\{ 2\sin\frac{r\pi\left[\imath2\beta\Im\left(\omega_{C}\right)\right]}{2\omega_{R}}\right\} ^{2}\\
+2N\eta_{R}\left[\imath2\beta\Im\left(\omega_{C}\right)+N\omega_{R}\right]-\imath N\pi\end{gathered}$$ and $$\begin{gathered}
L^{-}={ \operatorname{Log}}\frac{2\omega_{R}}{\pi}+\frac{\eta_{R}\left[2\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)\right]^{2}}{2\omega_{R}}\\
+{ \operatorname{Log}}\sin\frac{\pi\left[2\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)\right]}{2\omega_{R}}\\
+\sum_{r=1}^{\infty}\frac{q^{2r}}{r\left(1-q^{2r}\right)}\left\{ 2\sin\frac{r\pi\left[2\omega_{R}+\imath2\beta\Im\left(\omega_{C}\right)\right]}
{2\omega_{R}}\right\} ^{2}\\
+2\left(N-1\right)\eta_{R}\left[\imath2\beta\Im\left(\omega_{C}\right)+\left(N+1\right)\omega_{R}\right]-\imath\left(N-1\right)\pi.\end{gathered}$$ By noting that $${ \operatorname{Log}}\left[\pm\sin\frac{\imath\pi\beta\Im\left(\omega_{C}\right)}{\omega_{R}}\right]=
{ \operatorname{Log}}\sinh\frac{\pi\beta\Im\left(\omega_{C}\right)}{\omega_{R}}\pm\imath\frac{\pi}{2}$$ (as $\beta$, $\Im\left(\omega_{C}\right)$ and $\omega_{R}$ are all real positive quantities by definition), it can be verified, after a few algebraic passages, that $L^{+}=L^{-}$, and thus the right-hand side of eq. is a continuous function.
Going back to the Stark problem, we can immediately see how the logarithmic forms in eq. are in the same form as in eq. . For instance, in $$\ln\sigma_\xi\left(\tau-\tau_\xi+u_\xi\right)$$ the real variable is $\tau$, while $u_\xi$ is defined as $$u_\xi=\wp_\xi^{-1}\left(-\frac{\delta_\xi}{\gamma_\xi}\right).$$ Since $u_\xi$ is the result of an inverse $\wp$, it can always be chosen inside the fundamental period parallelogram, where the condition of convergence of the series expansion ($\left| \beta \right| < 1$) is always satisfied[^9]. Eq. can thus be substituted into eq. to provide a formula for $\phi\left(\tau\right)$ free of discontinuities.
Solution algorithm {#subsec:sol_algorithm}
------------------
In this section, we are going to detail the steps of a possible implementation of our solution to the Stark problem, starting from initial conditions in cartesian coordinates. The algorithm outlined below requires the availability of an implementation of the Weierstrassian functions $\wp$, $\wp^\prime$, $\wp^{-1}$, $\zeta$ and $\sigma$, and of a few related ancillary functions (e.g., for the conversion of the invariants $g_2$ and $g_3$ to the half-periods $\omega$ and $\omega^\prime$). Chapter 18 in @abramowitz_handbook_1964 details how to implement these requirements in terms of Jacobi theta and elliptic functions and integrals.
The algorithm is given as follows:
1. transform the initial cartesian coordinates into parabolic coordinates via eqs. -, and compute the initial Hamiltonian momenta $p_\xi$, $p_\eta$ and $p_\phi$ via eqs. -;
2. compute the constants of motion $h$, $\alpha_1$ and $\alpha_2$, through the substitution of the initial Hamiltonian coordinates and momenta into eqs. , and ;
3. calculate the roots of the bicubic polynomials on the right-hand sides of eqs. and . Among the positive roots, choose one for each of the two polynomials as the $\xi_r$ and $\eta_r$ values. In the case of the $\xi$ coordinate, $\xi_r$ must be a *reachable* root, i.e., a value that will actually be assumed by $\xi$ at some point in time[^10];
4. compute the fictitious times of “pericentre passage” $\tau_\xi$ and $\tau_\eta$ via eq. . The integral can be solved either via the inverse Weierstrass function [@hoggatt_inverse_1955] or via elliptic integrals [e.g., @gradshtein_table_2007 §3.131 and §3.138]. The signs of $\tau_\xi$ and $\tau_\eta$ must be chosen in accordance with the choice of $\xi_r$ and $\eta_r$ and with the initial signs of $p_\xi$ and $p_\eta$. For instance, in our implementation of this algorithm we always pick as $\xi_r$ the smallest reachable root, so that the sign of $\tau_\xi$ is the opposite of the sign of the initial value of $p_\xi$ (i.e., if initially $p_\xi < 0$ then $\xi_r$ will be reached in the future and thus $\tau_\xi > 0$);
5. at this point, it will be possible to compute the evolution in fictitious time of $\xi$, $\eta$ and $\phi$ via eqs. , and . The complex logarithm appearing in the equation for $\phi$, eq. , should be computed using the methodology described in Appendix \[sec:complex\_log\] in order to avoid discontinuities;
6. in order to compute the time equation, eq. , determine the roots $e_i$ of the characteristic cubic equations and the fundamental half-periods $\omega_i$ they correspond to, as explained in §\[sec:time\_equation\]. It will now be possible to compute $t\left( \tau \right)$, and to invert it via numerical techniques to yield $\tau\left( t \right)$.
Code availability
=================
The Weierstrassian functions, the analytical formulæ presented in this paper, and the algorithm outlined in Appendix \[subsec:sol\_algorithm\] have been implemented in the Python programming language. The implementation is available under an open-source license from the code repository <https://github.com/bluescarni/stark_weierstrass>
\[lastpage\]
[^1]: Previously at the Advanced Concepts Team, ESA. E-mail: [email protected]
[^2]: This regularisation procedure is sometimes referred to as *Poincaré trick* or *Poincaré time transform* [@siegel_lectures_1971; @carinena_time_1988; @saha_interpreting_2009].
[^3]: Note that one can always set the initial fictitious time to zero, as the relation between real and fictitious time is differential – see eq. .
[^4]: In the particular case in which $p_\xi\left( \xi \right)$ has no real roots, there will be no sign ambiguity: $p_\xi$ will always be positive or negative, and the sign can be chosen once and for all in accordance with the initial sign of $p_\xi$.
[^5]: Not only $\wp\left(z\right)$ is doubly periodic in $z$, but even within its fundamental periods it assumes all complex values twice [@whittaker_course_1927].
[^6]: There is an insidious technical difficulty in the direct use of formula , related to the multivalued character of the complex logarithm. The issue is presented and addressed in Appendix \[sec:complex\_log\].
[^7]: Such a check is best performed using a computer algebra tool. In this specific case, we used the Python library SymPy [@sympy].
[^8]: It is interesting to note that the study of the Weierstrassian formalism for the theory of elliptic functions is today no longer part of the typical background of physicists and engineers. Recently, the Weierstrassian formalism has been successfully applied to dynamical studies in General Relativity [e.g., @hackmann_analytical_2010; @scharf_schwarzschild_2011; @gibbons_application_2012; @biscani_first-order_2013].
[^9]: From the point of view of practical implementation, one can choose among two possible values for $u_\xi$ in the fundamental period parallelogram. In order to improve the convergence properties of the series expansion, it is convenient to select the value with the smaller imaginary part.
[^10]: Consider, for instance, a phase space portrait like the one depicted in Figure \[fig:tridimensional\_case\](b). Depending on the initial conditions, the test particle will be confined either to a circulation lobe (in which case there are two reachable roots, where the lobe intersects the horizontal axis) or to the parabolic arm (in which case there is only one reachable root, where the parabolic arm intersects the horizontal axis).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '**Arising from emergent electromagnetic field of magnetic skyrmions due to their nontrivial topology, the skyrmion Hall effect might be a roadblock for practical applications since any longitudinal motions of skyrmions in nanotrack is accompanied by a transverse motion. A direct consequence of such an effect is easy destruction of skyrmions at the nanotrack edges during their fast motions along the nanotrack, despite their topological protection. Here we propose an entirely novel solution of completely inhibiting such skyrmion Hall effect without affecting its topological properties based on a antiferromagnetic-coupling bilayer system. We show that a pair of magnetic skyrmions can be nucleated in such a bilayer system through vertical current injection or converted from a current-driven domain-wall pair. Once nucleated, the skyrmion pair can be displaced through current-induced spin torque either from a vertical injected current or in-plane current. The skyrmion Hall effect is completely suppressed due to the cancellation of back-action forces acting on each individual skyrmion, resulting in a straight and fast motion of skyrmions along the current direction. This proposal will be of fundamental interests by introducing the bilayer degree of freedom into the system. Moreover, it provides an easy way to engineer the transport properties of the skyrmionic devices to achieve desired performance, making it highly promising for practical applications such as ultradense memory and information-processing devices based on skyrmions.**'
address: '[Department of Applied Physics, University of Tokyo, Hongo 7-3-1, 113-8656, Japan]{}'
author:
- 'Xichao Zhang$^{1}$'
- 'Yan Zhou$^{1,2}$'
- 'Motohiko Ezawa$^{3}$'
title: '[Magnetic bilayer-skyrmions without skyrmion Hall effect]{}'
---
Since the first experimental observations of magnetic skyrmion lattices in bulk non-centrosymmetric magnets[@Mol; @Yu] and films[@Heinze], there has been tremendous interest in these topologically protected spin configurations with a quantized topological number[@SkRev; @Fert]. However, the creation and transmission of isolated magnetic skyrmion in magnetic thin films are required for any practical applications such as encoding information in individual skyrmion to allow entirely novel devices and circuitry[@Sampaio; @Fert; @Tchoe; @Yan; @Marco; @XichaoSR2015; @XichaoSR2014]. Significant efforts and progress have been made towards realizing such ultrathin film based on perpendicularly magnetized magnetic layer$|$heavy metal structure to host a rich variety of chiral spin textures including skyrmions[@vincent; @klaui]. The strong spin orbit coupling at the interfaces between the magnetic layer with perpendicular magnetic anisotropy (PMA) and the underlying heavy metal layer provides a sizeable Dzyaloshinsky-Moriya interaction (DMI) to stabilize skyrmions[@Fert; @parkin; @Bogdanov_1; @Bogdanov_2; @Bogdanov_3].
To move a skyrmion in nanotrack for information-processing applications, a convenient and efficient way is by means of spin current which can transfer the angular momentum from itinerant conduction electrons to the magnetic moments of the skyrmion[@IwasakiNL; @Iwasaki]. However, one major roadblock to the manipulation and transmission of skyrmions in nanotrack is the skyrmion Hall effect, *i.e.*, skyrmions exhibit the Hall effect driven by spin currents due to the presence of the Magnus force which in turn originates from its nontrivial topology[@SkRev]. Thus a skyrmion will not move parallel with the direction of the current. Instead it will gain a transverse velocity with a magnitude proportional to the spin current density which displaces the skyrmion towards the edge of the nanotrack. Therefore a skyrmion will be easily destroyed for a distance of much less than 1$\mu $m for a nanotrack made of typical magnetic layer$|$heavy metal system[@IwasakiNL; @Iwasaki; @IwasakiNC].
Since both the skyrmion Hall effect and its topological protection arise from the same mechanism, *i.e*., its nontrivial topological number of $\pm 1$, it seems to be impossible to inhibit the skyrmion Hall effect without breaking its topological protection. In this work, we propose a novel solution of two perpendicularly magnetized sublayers strongly coupled via the antiferromagnetic (AFM) exchange interaction with a heavy metal layer beneath the bottom magnetic layer[@ParkinNNano2015] (Fig.\[Fig1\]a, b, c). When a skyrmion is created in one of the sublayers, another skyrmion is simultaneously created in the other sublayer under certain conditions. We refer to such a pair of magnetic skyrmions as a magnetic bilayer-skyrmion (Fig.\[Fig1\]d, e). Moreover, we show a bilayer-skyrmion can be displaced over arbitrarily long distances driven by spin currents without touching the edges due to the absence of the skyrmion Hall effect, which is distinct from a skyrmion in the monolayer thin-film structure. In addition to address the above-mentioned dilemma, bilayer-skyrmions have many distinct characteristics from conventional skyrmions in monolayer thin film, which allow for versatile and multifunctional ultradense and ultrafast information processing and logic applications.
{width="96.00000%"}
These features are true for both the current-in-plane (CIP) and current-perpendicular-to-plane (CPP) cases. The CIP implementation has been extensively studied in the past ten years for displacing domain walls or switching magnetization. Recently, the spin Hall effect has been demonstrated to be a more efficient means of manipulating magnetization. The current for the CIP case is along the nanowire axial direction. By contrast, the spin current is perpendicular to the heavy metal/ferromagnetic interface for the CPP case. In our nomenclature for CPP, the spin current is perpendicular-to-film plane whereas the charge current can be either in the film plane (such as the spin Hall effect scenario) or perpendicular to the film plane (such as the perpendicular MRAM where a perpendicularly magnetized polarizer is incorporated). The CPP scheme has been proven to be much more efficient to move the domain wall or skyrmion than the CIP method[@parkin; @Sampaio; @Iwasaki; @IwasakiNL].
In this work, we first describe the nucleation process of an isolated skyrmion within the top layer by utilizing the CPP injection into a disk composed of two AFM-coupled magnetic sublayers. Owing to the interlayer AFM exchange coupling, another skyrmion will automatically emerge in the bottom layer. In so doing we explore various properties of a bilayer-skyrmion. Furthermore, we move a bilayer-skyrmion in the nanotracks either by the CPP or CIP. We also study the generation of a domain-wall (DW) pair and the conversion process from a DW pair to a skyrmion in the bilayer nanotracks.
{width="90.00000%"}
Bilayer System coupled with AFM interaction
===========================================
**Hamiltonian.** We investigate the bilayer system where the top and bottom ferromagnetic (FM) layers are coupled antiferromagnetically by the exchange interaction, as illustrated in Fig.\[Fig1\]a. The Hamiltonian for each layer reads $$\begin{aligned}
H_{\tau }=&-A_{\text{intra}}\sum_{\langle i,j\rangle }\boldsymbol{m}_{i}^{\tau }
\cdot \boldsymbol{m}_{j}^{\tau }+\sum_{\langle i,j\rangle }
\boldsymbol{D}\cdot (\boldsymbol{m}_{i}^{\tau }\times \boldsymbol{m}_{j}^{\tau })\notag\\
&+K\sum_{i}[1-(m_{i}^{\tau ,z})^{2}]+H_{\text{DDI}}, \label{HamilAF}\end{aligned}$$ where $\tau $ is the layer index ($\tau =$T, B), $\boldsymbol{m}_{i}^{\tau }$ represents the local magnetic moment orientation normalized as $|\boldsymbol{m}_{i}^{\tau }|=1$ at the site $i$, and $\left\langle i,j\right\rangle $ runs over all the nearest neighbor sites in each layer. The first term represents the FM exchange interaction with the FM exchange stiffness $A_{\text{intra}}$. The second term represents the DMI with the DMI vector $\boldsymbol{D}$. The third term represents the PMA with the anisotropic constant $K$. $H_{\text{DDI}}$ represents the dipole-dipole interaction. There exists an AFM coupling between the top and bottom layers, $$H_{\text{inter}}=-A_{\text{inter}}\sum_{i}\boldsymbol{m}_{i}^{\text{T}}\cdot
\boldsymbol{m}_{i}^{\text{B}}.$$ The sign of $A_{\text{inter}}$ is negative reflecting that the interlayer interaction is antiferromagnetic. We assume that the spins in the top layer are pointing upward. Then the spins in the bottom layer are pointing downward due to the interlayer AFM couplings.
{width="96.00000%"}
**Topological Number.** The classical field $\boldsymbol{m}^{\tau }(\boldsymbol{x})$ is introduced for the spin texture in the FM system by considering the zero limit of the the lattice constant, $a\rightarrow 0$. The ground-state spin textures are $\boldsymbol{m}^{\text{T}}=(0,0,1)$ and $\boldsymbol{m}^{\text{B}}=(0,0,-1)$. A magnetic skyrmion is a spin texture which has a quantized topological number. Spins swirl continuously around the core and approach the ground-state value asymptotically. The skyrmion is characterized by the topological number $Q_{\tau }$ in each layer, $$Q_{\tau }=-{\frac{1}{4\pi }}\int d^{2}x\boldsymbol{m}^{\tau }(\boldsymbol{x})
\cdot \left( \partial _{x}\boldsymbol{m}^{\tau }(\boldsymbol{x})\times
\partial _{y}\boldsymbol{m}^{\tau }(\boldsymbol{x})\right) .
\label{PontrNumbe}$$ We obtain $Q_{\tau }=\pm 1$ for a skyrmion in a sufficiently large area. We also call $Q_{\tau }$ the skyrmion number. Even if the skyrmion spin texture is deformed, its skyrmion number does not change, as far as the boundary condition is not modified. It can be neither destroyed nor separated into pieces, *i.e.*, a skyrmion is topologically protected.
The spins in the top and bottom layers are tightly bounded due to the AFM coupling. Accordingly, if one skyrmion is created in the top layer, a second skyrmion is also created in the bottom layer (Fig.\[Fig2\]a) simultaneously. The topological number of the bottom layer is opposite to that of the top layer since all the spins are inverted, $Q_{\text{B}}=-Q_{\text{T}}$. See Fig.\[Fig2\]b how these topological numbers evolve after the creation of a skyrmion in the top layer.
**Excitation Energy.** We may determine numerically the spin profile of each skyrmion and estimate the excitation energy based on the Hamiltonian $H_{\text{total}}=H_{\text{T}}+H_{\text{B}}+H_{\text{inter}}$. We compared the energy of one bilayer-skyrmion state and the energy of the homogeneous state in Fig.\[Fig2\]c as a function of the DMI. Similar result is obtained for any value of the interlayer AFM coupling strength. This is because the spin directions between the top and bottom layers are always opposite.
The energy of the bilayer-skyrmion becomes lower than that of the uniform ground state for $D\gtrsim 4$ mJ m$^{-2}$, which is the threshold value. Consequently, the bilayer-skyrmion excitation is energetically favorable when the DMI strength is larger than the threshold value. However, spontaneous generation of skyrmions does not occur due to the topological protection. Nevertheless, the topological protection can be violated in condensed matter physics because of the existence of the lattice structure and the boundary of the sample. We may create skyrmions by leveraging these properties.
**Skyrmion Hall effect.** It is well understood[@Stone] that the center-of-mass motion of a skyrmion is determined by the Lagrangian, $$L=L_B-U,$$ where $L_B$ is the Berry phase term of the spin texture, $$L_B=\frac{G}{2}(\dot{X}Y-X\dot{Y})=\frac{1}{2}\mathbf{G}\cdot (\dot{\mathbf{R}}\times\mathbf{R}),$$ and $U$ is the potential. Here, $\mathbf{G}=(0,0,G)$ is the gyromagnetic coupling constant representing the Magnus force with $G=4\pi Q$, and $\mathbf{R}=(X,Y)$ is the center-of-mass coordinate of the skyrmion. The Euler-Lagrange equation yields $$\dot{\mathbf{R}}=\frac{1}{G}\mathbf{e}_z\times \mathbf{F},$$ where $\mathbf{e}_z=(0,0,1)$, and $\mathbf{F}=-\nabla U$ is the force acting on the skyrmion. Consequently, a moving skyrmion feels the Magnus force and bends toward the direction perpendicular to the force. It is called the skyrmion Hall effect.
The direction of the Magnus force is opposite when the sign of the skyrmion number $Q$ is opposite. For instance, if a skyrmion is driven by the current along the $+ x$ direction it will be bended toward the $+ y$ ($-y$) direction in the top (bottom) layer.
Landau-Lifshitz-Gilbert-Slonczewski equation for CPP
====================================================
We may apply a CPP spin-polarized current injection from magnetic tunnel junction (MTJ) or the spin Hall effect in heavy metal layer[@Khv; @Sampaio; @Fino]. We numerically solve the Landau-Lifshitz-Gilbert-Slonczewski (LLGS) equation, which governs the dynamics of the magnetization $\boldsymbol{m}_{i}$ at the lattice site $i$. By suppressing the layer index, it reads $$\begin{aligned}
\frac{d\boldsymbol{m}_{i}}{dt} =&-|\gamma |\boldsymbol{m}_{i}\times
\boldsymbol{H}_{i}^{\text{eff}}+\alpha \boldsymbol{m}_{i}\times
\frac{d\boldsymbol{m}_{i}}{dt} \notag \\
& +\left\vert \gamma \right\vert u(\boldsymbol{m}_{i}\times \boldsymbol{p}\times \boldsymbol{m}_{i})
-\left\vert \gamma \right\vert u^{\prime }(\boldsymbol{m}_{i}\times \boldsymbol{p}), \label{LLGS}\end{aligned}$$where $\boldsymbol{H}_{i}^{\text{eff}}\mathbf{=-}
\partial H_{\text{total}}/\partial \mathbf{m}_{i}$ is the effective magnetic field induced by the Hamiltonian $H_{\text{total}}=H_{\text{T}}+H_{\text{B}}+H_{\text{inter}}$, $\gamma $ is the Gilbert gyromagnetic ratio, $\alpha $ is the Gilbert-damping coefficient originating from spin relaxation, $u$ is the Slonczewski STT coefficient, $u^{\prime }$ is the out-of-plane STT coefficient, and $\boldsymbol{p}$ represents the electron polarization direction. Here, $u=|\frac{\hbar }{\mu
_{0}e}|\frac{j|\boldsymbol{p}|}{2dM_{s}}$ with $\mu _{0}$ the vacuum magnetic permittivity, $d$ the film thickness, $M_{s}$ the saturation magnetization, and $j$ the current density. We take $-z$ direction for creating the skyrmion, while $+y$ direction for moving the skyrmion. The STT is induced either by injection from a magnetic tunnel junction polarizer or by the spin Hall effect[@Sampaio; @Fino; @Khv].
{width="94.00000%"}
LLG equation for CIP
====================
Alternatively we may apply a CIP injection[@Khv; @Sampaio; @Fino] to move skyrmions. We numerically solve the LLG equation, $$\begin{aligned}
\frac{d\boldsymbol{m}_{i}}{dt} =&-|\gamma |\boldsymbol{m}_{i}\times
\boldsymbol{H}_{i}^{\text{eff}}+\alpha \boldsymbol{m}_{i}\times \frac{d\boldsymbol{m}_{i}}{dt} \notag \\
&+\frac{\boldsymbol{p}|a^{3}}{2eM_{s}}(\boldsymbol{j(r)}\cdot \nabla )\boldsymbol{m}_{i} \notag \\
&-\frac{\boldsymbol{p}|a^{3}\beta }{2eM_{s}}[\boldsymbol{m}_{i}\times (\boldsymbol{j(r)}\cdot \nabla )\boldsymbol{m}_{i}],\end{aligned}$$ where $\beta$ is the strength of the non-adiabatic torque and $a$ is the lattice constant.
**Creation of a bilayer-skyrmion by vertical spin current.** We employ a CPP injection with a circular geometry in a nanodisk. The spin-polarized current (polarized along $-z)$ is injected into the top layer in the central circle region, as illustrated in Fig.\[Fig1\]a.
We demonstrate how the spin textures develop in Fig.\[Fig2\]a. The spins start to flip in both layers following the spin current injection only in the top layer. When there is no interlayer AFM coupling, a skyrmion is formed only in the top layer (). By contrast, a skyrmion is formed also in the bottom layer upon the current injection in the presence of the interlayer AFM coupling ().
We show the evolution of the skyrmion number in Fig.\[Fig2\]b. It oscillates at the initial stage for $t<0.2$ ns, and rapidly increases to $1$. The skyrmion remains stable even when the current is switched off, demonstrating that it is topologically protected. During this process, the spins in the top and bottom layers are always anti-parallel. A comment is in order. The saturated skyrmion number is not exactly $Q=1$ but $Q=0.89$. This is due to the fact that there is a background skyrmion number which originates from the tilting the edge spins. It is $Q=-0.09$. Accordingly, the calibrated skyrmion number is $Q=0.98$, which is almost unity.
We present a nucleation phase diagram of a bilayer-skyrmion pair as a function of the current density and the interlayer AFM coupling in Fig.\[Fig2\]d and Fig.\[Fig2\]e. When the magnitude of the injected current is strong enough, the bilayer-skyrmion is created. This is due to the fact that spin flip costs a certain energy. On the other hand, if the interlayer AFM coupling is too strong, the bilayer-skyrmion is suppressed due to the fact that the nucleation field and the coercivity increases with the interlayer AFM exchange, leading to a larger current density for nucleation.
**Creation of a bilayer-skyrmion from a bilayer DW pair.** A magnetic skyrmion can be created from a DW pair by using a junction geometry[@NC]. In this scenario we first make a DW pair into a nanotrack of the top layer through the local CPP injection with $-z$ direction. We show how the spins start to flip in the top layer and subsequently in the bottom layer driven by the AFM exchange force in Fig.\[Fig3\]a. Then, the bilayer DW pair is shifted by applying CPP current, as shown in the process from $t=50$ ps to $t=120$ ps in Fig.\[Fig3\]b. Here we consider the vertical injection of a spin current towards $+z$ and polarized along $+y$ in the bottom layer. The CPP injection moves the bilayer DW in the rightward direction. When the bilayer DW arrives at the junction interface ($t=170$ ps), both the end spins of the DW are pinned at the junction, whereas the central part of the DW continues to move due to STT in the wide part of the nanotrack. Therefore, the structure is deformed into a curved shape and a bilayer-skyrmion texture forms at $t=190$ ps ().
**Current-driven motion of a bilayer-skyrmion in a nanotrack.** The magnetic bilayer-skyrmion can be displaced by the vertical spin-polarized current as in the case of the magnetic skyrmion. We may employ the CPP injection or the CIP injection to drive a bilayer-skyrmion. In general, a moving skyrmion is easily destroyed by the sample edges due to the skyrmion Hall effect. Therefore, the maximum velocity of skyrmion in FM nanotrack is typically much less than 10$^{3}$ m/s, limited by the edge confining force[@IwasakiNL] of $\sim D^{2}/J$.
The skyrmion in the top layer follows the motion of the skyrmion in the bottom layer even when the current is not injected into the top layer. This is because that two skyrmions are bounded by the interlayer AFM coupling. There is no skyrmion Hall effect for a magnetic bilayer-skyrmion. This can be explained as follows. If there is no interlayer AFM coupling, a skyrmion in the top layer moves left-handed and the skyrmion in the bottom layer moves right-handed. However, when the AFM coupling is strong enough, two skyrmions are tightly bounded and the Magnus forces acting on the skyrmions between the top and bottom layers are exactly cancelled. Accordingly the bilayer-skyrmion will move straightly. This mechanism works both for the CPP and CIP cases ().
We show the relation between the magnitude of the injected current and the velocity in Fig.\[Fig4\]. The velocity is proportional to the injected current density. For strong enough current, the bilayer-skyrmion is destroyed and split into two independent skyrmions. This is because that the skyrmion Hall effect increases as the current increases, which acts as the repulsive force between two skyrmions ().
With strong interlayer AFM exchange coupling, the coupled skyrmions move along the central line of the nanotrack at a high speed of a few hundred meters per second, without any transverse motion. However, with small interlayer AFM exchange coupling, the skyrmions in the top and bottom layers will be decoupled due to the fast motion of the skyrmion in the bottom layer driven by large current. Once the skyrmions in the top and bottom layers are decoupled, the skyrmion Hall effect becomes active, leading to the destruction of skyrmions in the top and/or bottom layer by touching edge. At the same time, in the CPP case, the skyrmion in the top layer will stop motion. It can be seen that the coupled skyrmions driven by CIP current doesn’t require a fine tuning of damping and non-adiabatic torque coefficients.
It is also worth noting that the interlayer AFM coupling does not produce a mass of the bilayer-skyrmion. When the driving current is suddenly turned off, the bilayer-skyrmion stops high-speed motion immediately (see Supplementary Movie 8). On the other hand, we also investigated the case where the DMI constant $D$ is different between the top and bottom layers. It is found that the results of current-driven motion of the bilayer-skyrmion do not change much since the DMI only changes the radius of the skyrmion (see Supplementary Movies 9-10). The massless property and its robustness make the bilayer-skyrmion an ideal candidate for practical applications.
Perspectives
============
We have presented a novel solution of inhibiting the Hall effect of skyrmions without affecting their topological properties, by exploring a new device made of antiferromagnetically exchange-coupled bilayer nanodisks and nanotracks. Compared with the mostly investigated skyrmion in the FM monolayer system, the bilayer-skyrmion exhibits entirely distinct characteristics with regard to the current-transport behaviour and robustness. First, it can move strictly along the direction of the spin current flow, which makes it more appealing for motions in nanowires for ultradense memory applications. This is in high contrast with the case of monolayer skyrmion, where the skyrmion information carrier can be easily destroyed by the edges of the nanotracks. Second, it will be immune to magnetic field perturbations which might be generated externally or internally within the device circuitry since the net magnetic moment is zero. Third, by introducing the bilayer degree of freedom, the transport properties of the device can be engineered to achieve desired performance. For example, the skyrmion Hall effect can be intentionally suppressed or enhanced by tuning the magnetic properties of individual layers. This newly proposed solution of transporting skyrmion information carrier for arbitrarily long distances at much enhanced velocity may be very appealing for versatile applications such as ultradense memory and information processing. Similar ideas can be extended to multilayer or superlattice where the skyrmions are strongly coupled to realize a better manipulation of skyrmions in nanotrack or extended thin films.
Methods
=======
**Modeling and simulation.** The micromagnetic simulations are performed using the Object Oriented MicroMagnetic Framework (OOMMF) including the Dzyaloshinskii-Moriya interaction (DMI) extended module [@Fert; @Sampaio; @Dona; @Rohart]. The time-dependent magnetization dynamics is governed by the Landau-Lifshitz-Gilbert (LLG) equation including spin torque[@Brown; @Gilbert; @LL; @Thia; @Thia2]. The average energy density $E$ is a function of $\textbf{M}$, which contains the intralayer exchange, the interlayer exchange, the anisotropy, the applied field (Zeeman), the demagnetization and the DMI energy terms. For micromagnetic simulations, the intrinsic magnetic parameters are adopted from Refs.[@Fert; @Sampaio]: Gilbert damping coefficient $\alpha
= 0.3$ and the value for Gilbert gyromagnetic ratio is -2.211$\times$ 10$^{5}$ m A$^{-1}$ s$^{-1}$. Saturation magnetization $M_{S}$ = 580 kA m$^{-1}$, intralayer exchange stiffness $A$ = 15 pJ m$^{-1}$, DMI constant $D$ = 0 $\sim$ 6 mJ m$^{-2}$ and perpendicular magnetic anisotropy (PMA) $K$ = 0.8 MJ m$^{-3}$ unless otherwise specified. The interlayer exchange coefficient $A_\text{inter}$ is set from 0 to -10 pJ m$^{-1}$, whereas the corresponding interface exchange coefficient $\sigma$ equals from 0 to -10 mJ m$^{-2}$ ($\protect\sigma$ = $A_{\text{inter}}$ / 1 nm), where “-” denotes that the interface is antiferromagnetically coupled. The field-like out-of-plane STT coefficient $u^{\prime }$ is set to zero. All samples are discretized into cells of 2 nm $\times$ 2 nm $\times$ 1 nm in the simulation, which is sufficiently smaller than the typical exchange length ($\sim$ 4.3 nm) and the skyrmion size to ensure the numerical accuracy. For all simulation of current-driven skyrmions reported throughout this paper, the skyrmions are firstly created at the designed spot of the nanotrack ($x$ = 100 nm) by a local spin current perpendicular to the plane of the top layer. Then the system is relaxed to an energy minimum state without applying any current. Next, we start the timer and the spin current ($P$ = 0.4) is injected into the nanotrack with the geometry of current-in-plane (CIP) or current-out-of-plane (CPP) as shown in Fig.\[Fig1\]. In the configuration of CIP, the electrons flow toward the right in both the top and bottom layers, $i.e.$, the currents flow toward the left, while in the configuration of CPP, the electrons flow toward the top only in the bottom layer.
[99]{} S. Mhlbauer et. al. Science **323**, 915 (2009).
X. Z. Yu et. al. Nature **465**, 901 (2010).
S. Heinze et. al. Nature Phys. **7**, 713 (2011).
N. Nagaosa and Y. Tokura, Nat. Nanotech. **8**, 899 (2013).
A. Fert, V. Cros, and J. Sampaio, Nat. Nanotech. **8**, 152 (2013).
J. Sampaio, V. Cros, S. Rohart, A. Thiaville, and A. Fert, Nat. Nanotech. **8**, 839 (2013).
Y. Tchoe and J. H. Han, Phys. Rev. B **85**, 174416 (2012).
Y. Zhou, E. Iacocca, A. Awad, R. K. Dumas, F. C. Zhang, H. B. Braun, and J. Akerman, cond-mat/arXiv:1404.3281.
M. Finazzi, M. Savoini, A. R. Khorsand, A. Tsukamoto, A. Itoh, L. Duo, A. Kirilyuk, Th. Rasing, and M. Ezawa, Phys. Rev. Lett. **110**, 177205 (2013).
X. C. Zhang, M. Ezawa, and Y. Zhou, Sci. Rep. **5**, 9400 (2015).
X. C. Zhang, G. P. Zhao, H. Fangohr, J. P. Liu, W. X. Xia, J. Xia, and F. J. Morvan, Sci. Rep. **5**, 7643 (2015).
C. Moreau-Luchaire, C. Moutafis, N. Reyren, J. Sampaio, N. Van Horne, C.A.F. Vaz, K. Bouzehouane, K. Garcia, C. Deranlot, P. Warnicke, P. Wohlhüter, J.M. George, J. Raabe, V. Cros, and A. Fert, arXiv:1502.07853 (2015).
S. Woo, K. Litzius, B. Krüger, M. Y. Im, L. Caretta, K. Richter, M. Mann, A. Krone, R. Reeve, M. Weigand, P. Agrawal, P. Fischer, M. Kläui, G. S. D. Beach, arXiv:1502.07376 (2015).
S. Parkin, and S. Yang, Nat. Nanotech. **10**, 195 (2015).
N. Bogdanov and D. A. Yablonskii, Sov. Phys. JETP **68**, 101 (1989).
N. Bogdanov and A. J. Hubert, Magn. Magn. Mater. **138**, 255 (1994).
U. K. Roessler, N. Bogdanov, and C. Pfleiderer, Nature **442**, 797 (2006).
J. Iwasaki, W. Koshibae, and N. Nagaosa, Nano Lett., **14**, 4432 (2014).
J. Iwasaki, M. Mochizuki, and N. Nagaosa, Nat. Nanotech. **8**, 742 (2013).
J. Iwasaki, M. Mochizuki, and N. Nagaosa, Nat. Commun. **4**, 1463 (2012).
S. H. Yang, K. S. Ryu, and S. Parkin, Nat. Nanotech. **10**, 221 (2015).
M. Stone, Phys. Rev. B **53**, 16573 (1996).
A.V. Khvalkovskiy, V. Cros, D. Apalkov, V. Nikitin, M. Krounbi, K.A. Zvezdin, A. Anane, J. Grollier, and A. Fert, Phys. Rev. B **87**, 020402(R) (2013).
R. Tomasello, E. Martinez, R. Zivieri, L. Torres, M. Carpentieri, and G. Finocchio, Sci. Rep. **4**, 6784 (2014).
Y. Zhou and M. Ezawa, Nat. Com. **5**, 4652 (2014).
M. J. Donahue and D. G. Porter, National Institute of Standards and Technology, Interagency Report NISTIR, **6376** (1999 OOMMF user’s guide, version 1.0.).
S. Rohart and A. Thiaville, Phys. Rev. B **88**, 184422 (2013).
W. F. Brown J., Micromagnetics (Krieger, New York, 1978).
T. L. Gilbert, Phys. Rev. **100**, 1243 (1955).
L. Landau and E. Lifshitz, Physik. Z. Sowjetunion **8**, 153 (1935).
A. Thiaville, S. Rohart, E. Jue, V. Cros, and A. Fert, Europhys. Lett. **100**, 57002 (2012).
A. Thiaville, Y. Nakatani, J. Miltat, and Y. Suzuki, Europhys. Lett. **69**, 990 (2005).
Acknowledgements
================
Y.Z. thanks the support by the Seed Funding Program for Basic Research and Seed Funding Program for Applied Research from the HKU, ITF Tier 3 funding (ITS/171/13), the RGC-GRF under Grant HKU 17210014, and University Grants Committee of Hong Kong (Contract No. AoE/P-04/08). M.E. thanks the support by the Grants-in-Aid for Scientific Research from the Ministry of Education, Science, Sports and Culture, No. 25400317. M.E. is very much grateful to N. Nagaosa for many helpful discussions on the subject.
Author contributions
====================
M.E. conceived the idea and designed the project. M.E. and Y.Z. coordinated the project. X.Z. performed the numerical simulations supervised by Y.Z. All authors discussed the results and wrote the manuscript.
Additional information
======================
Supplementary information is available.
Competing financial interests
=============================
The authors declare no competing financial interests.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Sub-Doppler, saturation dip, spectra of lines in the $v_1 + v_3$, $v_1 + 2v_4$ and $v_3 + 2v_4$ bands of $^{14}$NH$_3$ have been measured by frequency comb-referenced diode laser absorption spectroscopy. The observed spectral line widths are dominated by transit time broadening, and show resolved or partially-resolved hyperfine splittings that are primarily determined by the $^{14}$N quadrupole coupling. Modeling of the observed line shapes based on the known hyperfine level structure of the ground state of the molecule shows that, in nearly all cases, the excited state level has hyperfine splittings similar to the same rotational level in the ground state. The data provide accurate frequencies for the line positions and easily separate lines overlapped in Doppler-limited spectra. The observed hyperfine splittings can be used to make and confirm rotational assignments and ground state combination differences obtained from the measured frequencies are comparable in accuracy to those obtained from conventional microwave spectroscopy. Several of the measured transitions do not show the quadrupole hyperfine splittings expected based on their existing rotational assignments. Either the assignments are incorrect or the upper levels involved are perturbed in a way that affects the nuclear hyperfine structure.'
author:
- Sylvestre Twagirayezu
- 'Gregory E. Hall'
- 'Trevor J. Sears'
title: 'Quadrupole splittings in the near-infrared spectrum of $^{14}$NH$_3$'
---
[^1]
Introduction
============
Ammonia, NH$_3$, is a symmetric top rotor possessing a large amplitude inversion vibrational mode. The inversion motion is associated with a double-well potential[@Kroto] with an effective barrier height of 2020 [cm$^{-1}$]{}. Tunneling though the barrier leads to the well-known inversion splittings[@Kroto; @Townes] in the zero point, and higher, levels of the molecule. Further, the spins of the three equivalent protons result in two nuclear spin states of the molecule, ortho- ($\bm{I_H} = 3/2$) and para- ($\bm{I_H} = 1/2$) each associated with a distinct set of rotational levels.[@Bunkerbook] The two sets of levels do not interact under most circumstances. These characteristics, together with its practical importance, have made ammonia a prototype molecule for spectroscopic studies.\
There has been considerable published work on the near-infrared spectrum of ammonia during the past few years.[@Sung2012; @Cacciani2012; @Foldesi2014; @AlDerzi2015; @Lehmann; @Douglas; @Czajkowski2009; @Berden1999] The spectrum in this region consists of a mixture of combination and overtone bands, of which the perpendicular bands $v_1 + v_3$ , $v_1 + 2v_4$, $v_3 + 2v_4$, and $2v_3$ are the most prominent. However, despite many years of work, spectroscopic analysis remains incomplete because of the many overlapping rotational features even at Doppler-limited resolution, multiple uncharacterized perturbations caused by anharmonic and Coriolis mixing, and the presence of hot band lines in the low-frequency inversion mode. The great majority of the observed features in recent high resolution measurements in the near-infrared remain rotationally unassigned.[@Sung2012; @Cacciani2012] Isotopically labeled samples[@Lees2008] permitted assignments of some of the strongest features of $^{14}$NH$_3$ and $^{15}$NH$_3$ spectra, but again leave many features unassigned. Variable (low) temperature studies[@Foldesi2014] of the spectrum have led to estimates of the lower state energies for many of the spectroscopic features and have permitted further assignments and corrections to earlier ones. Very recently, the Tennyson group[@AlDerzi2015] critically reviewed and validated all the available high resolution spectroscopic data for ammonia, to create a database of known levels and band origins. Combined with previously calculated absorption intensities, this work provides a detailed and accurate map of the assigned spectrum of ammonia from the microwave though the near-infrared. The work resulted in a Measured Active Rotational-Vibrational Energy Levels (MARVEL) database that can be dynamically updated as new data become available.\
In the present work, we report sub-Doppler measurements of a number of transitions in the $v_1+v_3$, $v_1+2v_4$ and $v_3+2v_4$ bands. These measurements fully resolve many overlapped vibration-rotation features in the Doppler-limited spectra whose presence have previously hindered assignments.[@Lees2008] The new measurements also exhibit partially resolved nuclear hyperfine structure due to the $^{14}$N quadrupole and proton hyperfine couplings. The observed hyperfine patterns are spectroscopic signatures of the rotational levels involved in the transitions[@Kukolich67; @Kukolich68; @Kukolich70; @Hougen72] and may therefore be used to confirm or deny spectroscopic assignments. Previous saturation spectroscopy measurements of ammonia at these wavelengths, carried out on a sample in a hollow-core fiber, have been reported.[@Cubillas2008; @Petersen2010] The resolution reported was sufficient to resolve lines blended in the Doppler-limited spectrum, but not sufficient to resolve any hyperfine structure. Czajkowski et al.[@Czajkowski2009] have also previously reported the measurement of a few sub-Doppler lines in the ammonia spectrum in this region, but again did not report the observation of quadrupole structure, presumably because the measurements were conducted at higher pressures than the current work, resulting in some collisional broadening.\
Two of the transitions observed, the $v_1+v_3$ band, $^pP(J,K=5,4)_a$ line at 6537.6806 [cm$^{-1}$]{}, frequency measured here at 195 994.734 57(2) GHz, and the $(v_1+2v_4)$, $^RP(7,5)_a$ line at 6488.200 [cm$^{-1}$]{}, 194 511.328 21(2) GHz here, exhibit hyperfine structure that does not conform to that expected, based on the assumption that the known lower state hyperfine splittings[@Hougen72] do not change on vibrational excitation. Sub-Doppler measurements of other transitions to confirm the relevant ground state combination differences are needed. For now, these assignments, based on Doppler-resolved spectra,[@Sung2012; @AlDerzi2015] must be regarded with suspicion. If the rotational assignments are correct, the observed change in the quadrupole splitting must be due to a perturbation in the upper level involved in the transitions. Examination of the upper energy levels listed in the MARVEL database,[@AlDerzi2015] shows that the $v_1+v_3$, $(4,3)_a$ level is within approximately 0.04 [cm$^{-1}$]{}in energy of $(5,4)_s$ of $(v_1+2v_4)$. The level $(4,3)_a$ of $(v_1+v_3)$ belongs to the para- set of proton hyperfine levels,[@Bunkerbook] while $(5,4)_s$ of $(v_1+2v_4)$ is an ortho- level. However, the estimated magnitude of the ortho-para coupling terms in ammonia[@Cacciani2009] is too small to account for such a (relatively) large observed perturbation between these levels. In future studies of the spectrum of ammonia in this region, the possibility of ortho-para coupling between nearly degenerate levels of the appropriate symmetry should be kept in mind, but accidental degeneracies close enough to allow significant ortho-para mixing are likely to be rare.
Experimental Methods
====================
The spectrometer used in this work has been described in detail previously.[@Twagirayezu1] Samples of anhydrous ammonia gas (Matheson Gas, Inc.) were introduced in the cavity-type absorption cell at pressures of between 1 and 20 mTorr (0.133-2.66 Pa) depending on the strength of the absorption. Sub-Doppler, saturation dip, spectra of rotational lines in the ammonia spectrum near 1.5$\mu$m were recorded. Collisional broadening of the saturation features could be observed at higher pressures, but no attempt was made to quantify this effect in this work. As in previous work,[@Twagirayezu1] the stronger saturation dip features could also be power-broadened and distorted at higher laser powers. Therefore, spectra were recorded at the lowest pressure and laser power consistent with satisfactory signal-to-noise ratios. Examples of observed derivative signals of the saturation dip profiles are shown in Figures \[fig1\] and \[fig2\]. These were obtained by scanning the comb repetition rate by 0.25 Hz/step, corresponding to approximately 60 kHz/step in the optical frequency across the saturation dip, typically collecting an averaged signal for 3 seconds at each frequency step. A few weak lines were recorded with more averaging to obtain satisfactory signal-to-noise ratios. Note that the modulation depth used to record Figure \[fig2\] was reduced compared to that shown in Figure \[fig1\] to reduce modulation broadening and highlight the splittings. The figures also show the results of the line analysis described in detail below.
Results and Analysis
====================
Hyperfine Level Structure
-------------------------
The observed saturation line shapes arise from partially resolved hyperfine splittings. The dominant hyperfine splitting is due to the nuclear quadrupole of the $^{14}$N nucleus, $\bm{I_N}=\bm{1}$, splitting each rotational level ($J>0$) into three sub-levels, labeled here by the quantum number $F_1$ with $F_1$ = $J$, $J\pm1$. Superimposed on the quadrupole structure, each level is further split by the proton nuclear hyperfine structure due to the $^1$H nuclear spin-rotation coupling and nuclear spin-spin dipolar coupling. The three equivalent protons lead to ortho- ($\bm{I_H}=\bm{3/2}$) and para- ($\bm{I_H}=\bm{1/2}$) proton nuclear spin functions,[@Bunkerbook] and the total angular momentum is $\bm{F}\, =\, \bm{F_1}+\bm{I_H}$, quantum number $F$. Measurements by Kukolich[@Kukolich67; @Kukolich68; @Kukolich70] and additional analysis by Hougen[@Hougen72] fully characterized the hyperfine structure in the lowest inversion doublet ($v_2 = 0$) of the $^{14}$NH$_3$ molecule. However, the proton hyperfine splittings are less than a few tens of kHz, too small to be resolved in the current experiments. The quadrupole coupling energy contributions are determined by the matrix elements:[@Kukolich67; @Hougen72] $$\label{WQ1}
W_Q(F_1,J^{\prime}, J, K) = (-1)^{J+I_N+F_1} { \begin{Bmatrix} F_1 & I_N & J' \\ 2 & J& I_N \end{Bmatrix} }
<J'~K\mid V \mid J~ K> < I_N\mid\mid Q \mid\mid I_N >$$ where V and Q are matrix elements which depend on $J^{\prime}$, $J$, $K$ and $I_N$. The $V$ contribution is: $$\label{V1}
<J'~K\mid V \mid J~ K>= \frac{1}{2} q\left[(2J'+1)(2J+1) \right]^{1/2}
(-1)^{J'+ K} \begin{pmatrix} J' & 2 & J \\ -K & 0 & K \end{pmatrix}$$ and $Q$: $$\label{Q1}
< I_N\mid\mid Q \mid\mid I_N > = \frac{1}{2} eQ (2I_N+1)\left[ \frac{(2I_N+3)(I_N+1) }{I_N(2I_N-1)(2I_N+1)} \right]^{1/2}$$\
The energy contributions from matrix elements off-diagonal in $J$ in eq. (\[WQ1\]) are negligible compared to the experimental resolution here because they connect levels separated by a rotational term value. Therefore, we computed just the diagonal contributions to the quadrupolar splittings in simulating the observed spectral features *i.e.* $J'$=$ J$ in eq. (\[WQ1\]) and (\[V1\]) above. These expressions were used to calculate the quadrupolar hyperfine splittings for each rotational level of interest. Also, although the inversion doublet levels have slightly different quadrupolar splittings, determined by the parameter $\Delta Q^*$ in Hougen’s notation, [@Hougen72] these differences are small compared to the current experimental measurement precision, so they too were neglected. Hence, all the experimental splittings will depend on the one quadrupolar parameter $\frac{eQq}{4}$, assumed to be the same in both the ground and excited vibrational levels. The calculated $^{14}$N quadrupole splittings based on this model for the rotational levels in NH$_3$ of interest here are given in Table \[table1\]. We note that the splittings for the $(J,K) = (3,2)$ levels are zero in this approximation because of the factor $J(J+1)-3K^2$ arising from the expansion of the 3-$j$ symbol in equation \[V1\].\
Saturation line profiles
------------------------
Spectroscopic transitions are expected to obey the selection rules, $\Delta F_1= 0,\pm1$ and the observed line shapes were modeled as a convolution of Lorentzian derivatives with an adjustable modulation broadening[@Axner2001] for each possible quadrupolar hyperfine transition as detailed below. Relative intensities of the quadrupole split transitions are given by: $$\label{IntQ}
Intensity_Q(rel.) =\ (2F'_1+1)(2 F_1+1) \begin{Bmatrix} I_N & J'& F'_1 \\ 1 & F_1& J \end{Bmatrix} ^{2}$$
Here, $J^{\prime}$ and $F^{\prime}$ are the quantum numbers in the upper state of the transition. With the aid of eq.(\[WQ1\]) and eq.(\[IntQ\]), the observed line shapes were modeled in steps: (i) the quadrupole energy level patterns for the rotational levels involved in the transition were estimated as described above (ii) the component quadrupole transition frequencies were then obtained by subtracting the lower energy levels ($W_Q(F_1,J,K)$) from the upper split levels ($W_Q(F'_1,J',K')$) with the selection rules, $\Delta F_1 = 0,\pm1$, and the relative intensities computed from eq. (\[IntQ\]); (iii) crossover resonances, due to pairs of two-level transitions sharing a common upper or lower hyperfine component, were also computed with a resonance halfway between the contributing two-level saturation frequencies and an intensity given by the geometrical mean of their intensities.[@Demtroder] Finally, (iv) the calculated saturation resonance frequencies and relative intensities were used to create an overall simulated line shape, from an intensity weighted sum of Lorentzian derivative lines using the Axner *et al.* model,[@Axner2001] each with a width determined by the estimated transit-time broadening (HWHM = 290 kHz) and an empirical modulation broadening to account for other broadening as detailed below, to compare to the experimental trace.\
Figure \[fig3\] illustrates the variety of quadrupolar split hyperfine patterns that can be expected for $P$, $Q$, and $R$ type transitions with $\Delta K=0, \pm 1$, given the experimental resolution of this work. For purposes of illustration, the line shapes for transitions originating in the state $(J,K)=(5,3)$ were computed, with only the three dominant $\Delta F_1 = \Delta J$ hyperfine transitions included for each rotational line. Ammonia only has perpendicular bands in this spectral region of interest, so the examples in the center column of the figure are not expected to be observed. Experimentally, we have found transitions involving rotational levels with $J > 3$ are reasonably well modeled by the line shapes in this figure. For transitions involving lower rotational levels, additional hyperfine components, *i.e.* $\Delta F_1 \ne \Delta J$, and crossover resonances need to be included to model the observed shapes more reliably. Examples are shown in Figs. \[fig1\] and \[fig2\].\
All the measured saturation dip transitions showed either resolved splittings or distortions compared to a wavelength-modulated derivative of a single Lorentzian transmission line shape.[@Axner2001] Individual line widths are dominated by transit-time broadening,[@Twagirayezu1] but poorly defined contributions from power, collisional (at pressures $>$5 mT) and modulation broadening also contribute, and we found that the chosen function provided a better approximation to the observations compared to the expected Gaussian resulting from purely transit-time broadening effects.[@Demtroder]\
The calculated profile was then matched to the observed feature with an adjustable central frequency ($\nu_0$) and the fixed set of quadrupolar offsets, $\Delta$W$_Q$, derived from the line assignment and the quadrupole splittings in table \[table1\]. Except in a few cases, discussed below, the predicted line shapes were in good accord with the observed features. The predictions might be refined by slightly adjusting the estimated quadrupolar shifts ($\Delta$W$_Q$) for the component transitions by varying the upper, quadrupole split, energy levels while keeping ground state splittings fixed. However, adjusting the upper state quadrupole coupling parameter to fit one or two poorly modeled line shapes would result in worse agreement with the great majority of the measurements. Figure \[fig1\] illustrates the simulation results for the $^rP(3,0)_s$ transition at 196 193.824 GHz[@Sung2012; @AlDerzi2015] while Figure \[fig2\] illustrates the more complex line shape observed for a $^pP(2,1)_s$ transition.\
In Fig. \[fig1\], the calculated hyperfine pattern is close to that observed, but the match would be improved if the $F_1^{\prime}=2$ component is slightly lower in energy. For the $^pP(2,1)_s$ transition in Fig. \[fig2\], the the positions of the hyperfine components lie close to the observed positions, but the relative intensities of the weaker components are underestimated, or the strongest component overestimated, compared to the observed. We attribute this is varying degrees of saturation for the different components, but trials using a model explicitly varying the relative intensities based on their linear line strengths were not successful in improving the qualitative agreement. Figure \[fig4a\] shows examples of a series of transitions from levels with $J=5$. The difference between the $^pP(5,4)$ $a-$ and $s-$ components is noteworthy. In the model, the two should have identical shapes, and the observations for the $K=5$ and $K=3$ transitions match this expectation.\
Hyperfine-free rest frequencies
-------------------------------
Values of the hyperfine-free rest frequencies $\nu _{0}$ were extracted for each frequency measured line and summarized in Table \[table2\]. The offset between $\nu _{0}$ and the dominant zero-crossing frequency of the partially resolved hyperfine pattern comes from the line shape modeling, so that even though the observed central zero crossing frequency is determined with high precision, the reported $\nu _{0}$ value will include an additional model-dependent error. These errors are of the same order of magnitude as errors expected from the neglect of the proton hyperfine and other neglected contributions to the line shape and are close to the experimental resolution. For this reason, we report a conservative estimate of the line center measurement errors as $\pm 20$ kHz for most of the data. Exceptions are noted in table \[table2\] when the lines were weak, for transitions where the observed quadrupole splitting patterns were complicated, typically low-$J$ transitions, and for those transitions where the predicted line shape did not match the observed. The accuracy of the measurement of the position of the zero-crossing of an observed line shape is actually better than this, and of the order of 3-10 kHz.[@Twagirayezu1]\
The observed $^pP(5,4)_a$ line shape shown in figure \[fig4a\] is an outlier compared to the other $^pP(5,K)$ lines. Comparison of the shape with simulations in figure \[fig3\] suggests that the pattern more closely resembles that for a $^pQ$ or $^pR$ transition. Alternatively, if the transition assignment is correct, the upper level quadrupole splitting has to be perturbed. One other measured transition, $(v_1+2v_4)$, $^rP(7,5)_a$ was expected to show an extended splitting pattern such as the one in the lower right corner of figure \[fig3\] but actually exhibits only a slightly asymmetrical shape, with a broader lower frequency lobe. Finally, the feature at 6635.4971 [cm$^{-1}$]{}has been assigned[@Sung2012] to three different rotational transitions based on combination differences in the Doppler-limited spectra: $^rR(5,3)_a\, \mathrm{in\, the} (v_1+2v_4)$ band, $^rQ(6,4)_s\, \mathrm{in\, the} (v_1+v_3)$ band, and $^rQ(6,4)_s\, \mathrm{in\, the} (v_3+2v_4)$ band. Searching near this frequency, we found only one sub-Doppler feature with a line shape that supports the first of these alternatives. We attempted to record sub-Doppler spectra of other transitions connected to the upper state in all these transitions, but were unsuccessful, possibly due to insufficient sensitivity.\
Absolute frequencies given in Table \[table2\] are generally within the estimated errors of those reported by Sung *et al.*[@Sung2012] and Földes *et al.*[@Foldesi2014] when allowing for the fact that many of the new measurements are of components of overlapped features in the Doppler-limited spectra. The accuracy and precision of the present measurements places much tighter constraints on ground state combination differences in the ammonia spectrum in this region.\
The present data permit the determination of six ground state combination differences. These are given in Table \[table3\] where they are also compared to numbers derived from the published energy levels[@Urban1984] for the lowest inversion doublet levels. Comparison shows that the present energy differences are systematically slightly larger than those derived from the published data, with the differences increasing with rotational energy. Even so, the largest deviations are less than 1 MHz, and unresolved hyperfine splittings in the earlier data could be the major contributor to the differences. Future improvements to the current spectrometer sensitivity will permit many more combination differences to be determined in these vibrational bands, and improvement in the accuracy of the rotational energy levels of ammonia could result from such measurements.
Conclusions
===========
The new data illustrate the precision that can be obtained from sub-Doppler measurements in the near-infrared. The data provide accurate rest frequencies for transitions that are overlapped in the best Doppler-limited spectra, even at reduced temperatures, and we have illustrated how they can resolve ambiguities in spectral assignments. The precision of lower state energy level combination differences derived from the measurements compares well with previous determinations derived from microwave and Doppler-limited far-infrared data. The observed quadrupole splittings in the sub-Doppler data provide a signature to help confirm rotational assignments because the quadrupole patterns are distinctive to a given rotational quantum number change. All but a few of the 55+ sub-Doppler measurements exhibit observed quadrupole hyperfine patterns that match those expected based on predictions from the known splittings in the lowest inversion doublet levels. While the upper, [$E^{\prime}$]{}symmetry, level $(4,3)_a$ in the 1010(10) vibrational excited state accessed in the $^pP(5,4)_a$ transition appears perturbed, and lies close to the $(5,4)_s$ level of 1002(02) with A$_2^{\prime}$ overall symmetry, these levels belong to different proton nuclear spin symmetries and, in ammonia, will be mixed by nuclear spin-rotation terms in the Hamiltonian. These are too small[@Cacciani2009] to account for the observed perturbation. We therefore conclude that the rotational assignment is probably not correct in this case.\
In a similar vein, the assignment of the $^rP(7,5)_a$ transition in the 0012(12) band at 6488.200 [cm$^{-1}$]{}is also probably incorrect because the observed hyperfine pattern does not match the expected. Also, the upper level energy is calculated to be 6488.199520 + 463.70701 [cm$^{-1}$]{}from Table \[table1\]. This is 0.03 [cm$^{-1}$]{}from the value for the energy in the MARVEL database,[@AlDerzi2015] which is greater than expected based on the literature uncertainties. There are no other, potentially perturbing, levels close to this energy in the database, so we conclude this assignment is also questionable. Finally, many of the data in Table \[table1\] represent measurement of transitions that are overlapped in the Doppler-limited spectra. For the most part, the sub-Doppler line measurements confirm the assignments in the literature. But, for the line at 6635.4971 [cm$^{-1}$]{}, assigned to three separate rotation-vibration transtions[@Sung2012], we could only find one sub-Doppler feature, that matched only one of the postulated assignments, raising questions about the other assignments.\
Efforts are underway to modify the spectrometer to increase its sensitivity and resolution still further. As it stands, sub-Doppler spectra can reliably be measured for unblended lines in ammonia with linestrengths, $S_{HITRAN}$, as low as 3$\times 10^{-22}$ cm.molecule$^{-1}$. The major source of noise is due to vibrational and acoustic perturbations of the cavity and efforts are underway to counteract this. The line widths could also be improved by increasing the beam waist diameter as described by Abe et al.[@Abe2014] With these improvements, several of the assignment questions raised above could be definitively resolved. In any case, the spectrum of ammonia in this region remains fertile ground for future sub-Doppler measurements.
Acknowledgements
================
Work at Brookhaven National Laboratory was carried out under Contract No. DE-SC0012704 with the U.S. Department of Energy, Office of Science, and supported by its Division of Chemical Sciences, Geosciences and Biosciences within the Office of Basic Energy Sciences.
Figures and Tables
==================
\[h!\] {width="40.00000%"} {width="60.00000%"}
\[h\] ![[]{data-label="fig2"}](Fig2-revised.pdf "fig:"){width="50.00000%"}
\[h\] ![[]{data-label="fig3"}](Figure3JCP.pdf "fig:"){width="50.00000%"}
\[h\] ![[]{data-label="fig4a"}](Figure4JCPrev.pdf "fig:"){width="50.00000%"}
[ccddd]{} & & & &\
1 & 0 & 2043.8 & -1021.9 & 204.4\
1 & 1 & -1022.3 & 511.1 & -102.2\
2 & 0 & 1020.8 & -1020.8 & 291.7\
2 & 1 & 510.6 & -510.6 & 145.9\
2 & 2 & -1022.4 & 1022.4 & -292.1\
3 & 0 & 815.4 & -1019.3 & 339.8\
3 & 1 & 611.8 & -764.7 & 254.9\
3 & 2 & -0.0 & 0.0 & -0.0\
3 & 3 & -1022.7 & 1278.4 & -426.1\
4 & 0 & 726.6 & -1017.2 & 369.9\
4 & 1 & 617.8 & -864.9 & 314.5\
4 & 2 & 291.1 & -407.5 & 148.2\
4 & 3 & -255.1 & 357.2 & -129.9\
4 & 4 & -1023.2 & 1432.5 & -520.9\
5 & 0 & 676.4 & -1014.6 & 390.2\
5 & 1 & 609.0 & -913.4 & 351.3\
5 & 2 & 406.4 & -609.6 & 234.5\
5 & 3 & 67.9 & -101.8 & 39.2\
5 & 4 & -408.3 & 612.4 & -235.5\
5 & 5 & -1024.0 & 1536.1 & -590.8\
6 & 0 & 643.6 & -1011.4 & 404.6\
6 & 1 & 597.9 & -939.5 & 375.8\
6 & 2 & 460.4 & -723.5 & 289.4\
6 & 3 & 230.6 & -362.4 & 145.0\
6 & 4 & -92.5 & 145.4 & -58.1\
6 & 5 & -510.5 & 802.1 & -320.9\
6 & 6 & -1025.1 & 1610.8 & -644.3\
7 & 0 & 620.1 & -1007.7 & 415.0\
7 & 1 & 587.1 & -954.1 & 392.9\
7 & 2 & 488.0 & -793.0 & 326.5\
7 & 3 & 322.2 & -523.6 & 215.6\
7 & 4 & 89.1 & -144.8 & 59.6\
7 & 5 & -212.4 & 345.1 & -142.1\
7 & 6 & -583.7 & 948.5 & -390.5\
7 & 7 & -1026.3 & 1667.8 & -686.7\
[llllldddd]{}
\
& & & & & & & &\
[[** ...continued from previous page**]{}]{}\
& & & & & & & &\
\
2 & 2 & 3 & 3 & 0012(12)s & 85.861 590 & 198 955.689 131 & 6 636.447 443 0 & 6636.448\
2 & 2 & 3 & 3 & 0012(12)a & 85.657 810 & 198 963.757 520 & 6 636.716 575 4 & 6636.716\
3 & 2 & 4 & 3 & 0012(12)a & 166.087 889 & 198 361.659 685 & 6 616.632 753 5 & 6616.633\
3 & 2 & 4 & 3 & 0012(12)s & 165.331 083 & 198 369.730 830 & 6 616.901 977 9 & 6616.902\
4 & 4 & 5 & 5 & 0012(12)a & 206.087 431 & 198 230.227 211 & 6 612.248 638 1 & 6612.249\
4 & 4 & 5 & 5 & 0012(12)s & 205.269 098 & 198 221.617 673 & 6 611.961 454 8 & 6611.962\
6 & 6 & 7 & 5 & 0012(12)a & 463.707 007 & 194 511.328 210$$^e$$ & 6 488.199 520 0 & 6488.200\*\
3 & 2 & 4 & 3 & 1002(02)a & 166.087 888 & 195 073.628 560 & 6 506.955 840 8 & 6506.956\
3 & 2 & 4 & 3 & 1002(02)s & 165.331 083 & 195 113.933 570 & 6 508.300 271 2 & 6508.301\
4 & 2 & 5 & 3 & 1002(02)a & 265.226 620 & 194 549.591 940$$^e$$ & 6 489.475 860 7 & 6489.476\
4 & 2 & 5 & 3 & 1002(02)s & 264.516 615 & 194 558.668 340$$^e$$ & 6 489.778 616 8 & 6489.779\
4 & 4 & 5 & 5 & 1002(02)a & 206.087 431 & 195 165.624 892 & 6 510.024 508 1 & 6510.025\
5 & 4 & 6 & 5 & 1002(02)a & 325.127 182 & 194 607.591 710$$^e$$ & 6 491.410 524 7 & 6491.411\
6 & 4 & 5 & 3 & 1002(02)a & 265.226 620 & 198 927.198 079 & 6 635.497 083 8 & 6635.497\
1 & 0 & 1 & 1 & 1010(10)a & 16.963 349 & 198 243.691 814$$^e$$ & 6 612.697 768 9 & 6612.704\*\
1 & 0 & 1 & 1 & 1010(10)s & 16.172 993 & 198 241.337 160$$^e$$ & 6 612.619 226 1 & 6612.619\
1 & 0 & 2 & 1 & 1010(10)s & 55.938 722 & 197 049.190 313$$^e$$ & 6 572.853 487 6 & 6572.854\
2 & 0 & 2 & 1 & 1010(10)a & 56.709 214 & 198 244.484 501$$^e$$ & 6 612.724 210 1 & 6612.726\*\
2 & 0 & 2 & 1 & 1010(10)s & 55.938 722 & 198 244.855 520$$^e$$ & 6 612.736 585 9 & 6612.726\*\
2 & 0 & 3 & 1 & 1010(10)a & 116.278 269 & 196 458.648 780 & 6 553.155 142 4 & 6553.155\
2 & 1 & 3 & 0 & 1010(10)s & 119.237 839 & 196 193.824 176$$^e$$ & 6 544.321 544 5 & 6544.322\
2 & 2 & 1 & 1 & 1010(10)s & 16.172 933 & 198 945.712 286$$^e$$ & 6 636.114 651 2 & 6636.115\
2 & 2 & 1 & 1 & 1010(10)a & 16.172 933 & 198 948.000 247$$^e$$ & 6 636.190 969 4 & 6636.191\
3 & 0 & 3 & 1 & 1010(10)a & 116.278 269 & 198 244.327 123 & 6 612.718 960 5 & 6612.726\*\
3 & 0 & 3 & 1 & 1010(10)s & 115.536 605 & 198 247.689 375 & 6 612.831 113 1 & 6612.833\*\
3 & 0 & 4 & 1 & 1010(10)a & 195.611 277 & 195 865.982 837 & 6 533.385 934 5 & 6533.386\
3 & 0 & 4 & 1 & 1010(10)s & 194.906 311 & 195 868.244 853 & 6 533.461 387 2 & 6533.461\
3 & 1 & 4 & 0 & 1010(10)a & 199.293 900 & 195 618.241 805 & 6 525.122 183 2 & 6525.122\
3 & 1 & 4 & 2 & 1010(10)a & 184.553 024 & 196 106.363 145 & 6 541.404 158 5 & 6541.405\
3 & 1 & 4 & 2 & 1010(10)s & 183.829 075 & 196 105.846 984 & 6 541.386 941 2 & 6541.387\
3 & 2 & 3 & 1 & 1010(10)a & 116.278 269 & 197 752.153 450 & 6 596.301 813 9 & 6596.302\*\
3 & 2 & 4 & 1 & 1010(10)a & 195.611 277 & 195 373.809 131 & 6 516.968 786 8 & 6516.970\
3 & 2 & 4 & 1 & 1010(10)s & 194.906 311 & 195 375.305 456 & 6 517.018 698 9 & 6517.020\
3 & 2 & 4 & 3 & 1010(10)a & 166.087 888 & 196 328.038 380 & 6 548.798 448 4 & 6548.798\
3 & 2 & 4 & 3 & 1010(10)s & 165.331 083 & 196 322.408 681 & 6 548.610 661 9 & 6548.611\
4 & 0 & 4 & 1 & 1010(10)a & 195.611 277 & 198 242.993 623 & 6 612.674 479 7 & 6612.673\*\
4 & 0 & 4 & 1 & 1010(10)s & 194.906 311 & 198 247.789 736 & 6 612.834 460 8 & 6612.824\*\
4 & 0 & 5 & 1 & 1010(10)a & 294.629 992 & 195 274.486 618 & 6 513.655 744 4 & 6513.656\
4 & 0 & 5 & 1 & 1010(10)s & 293.968 256 & 195 277.986 831 & 6 513.772 498 9 & 6513.773\
4 & 1 & 5 & 2 & 1010(10)a & 283.616 663 & 195 531.648 422 & 6 522.233 738 8 & 6522.234\
4 & 1 & 5 & 2 & 1010(10)s & 282.937 143 & 195 523.772 061 & 6 521.971 011 7 & 6521.971\
4 & 2 & 5 & 3 & 1010(10)a & 265.226 620 & 195 727.557 554 & 6 528.768 564 1 & 6528.769\
4 & 2 & 5 & 3 & 1010(10)s & 264.516 615 & 195 740.182 541 & 6 529.189 688 3 & 6529.190\
4 & 3 & 5 & 4 & 1010(10)a & 239.408 225 & 195 994.734 505 & 6 537.680 627 9 & 6537.681\
4 & 3 & 5 & 4 & 1010(10)s & 238.652 596 & 195 962.122 314 & 6 536.592 802 3 & 6536.593\
4 & 4 & 5 & 5 & 1010(10)a & 206.087 431 & 196 136.937 270 & 6 542.424 001 5 & 6542.424\
4 & 4 & 5 & 5 & 1010(10)s & 205.269 098 & 196 142.933 714 & 6 542.624 021 4 & 6542.624\
5 & 1 & 5 & 2 & 1010)10)a & 283.616 663 & 198 372.131 726 & 6 616.982 063 2 & 6616.982\
5 & 2 & 6 & 3 & 1010(10)a & 383.977 459 & 195 171.619 988 & 6 510.224 483 0 & 6510.225\
5 & 4 & 6 & 5 & 1010(10)a & 325.127 182 & 195 560.864 922 & 6 523.208 296 4 & 6523.208\
5 & 4 & 6 & 5 & 1010(10)s & 324.368 904 & 195 583.055 630 & 6 523.948 498 7 & 6523.949\*\
5 & 5 & 6 & 6 & 1010(10)a & 284.410 125 & 195 727.661 582 & 6 528.772 034 1 & 6528.772\
5 & 5 & 6 & 6 & 1010(10)s & 283.574 345 & 195 731.505 651 & 6 528.900 258 4 & 6528.901\
6 & 0 & 6 & 1 & 1010(10)s & 412.624 301 & 198 242.568 537 & 6 612.660 300 4 & 6612.653\*\
6 & 2 & 7 & 3 & 1010(10)a & 522.222 914 & 194 500.699 730 & 6 487.844 992 1 & 6487.845\
6 & 2 & 7 & 3 & 1010(10)s & 521.621 923 & 194 570.952 480 & 6 490.188 371 6 & 6490.189\
6 & 5 & 7 & 6 & 1010(10)s & 422.458 103 & 195 194.649 722 & 6 510.992 672 2 & 6510.993\
7 & 0 & 7 & 1 & 1010(10)s & 550.758 585 & 198 257.822 506 & 6 613.169 118 0 & 6613.169\
*e*. Line center measurement less accurate due to poor signal-to-noise or complex hyperfine pattern. Estimated errors up to $\pm$30kHz.
[cdd]{} & &\
(2,1) - (1,1)s & 1 192.146 847 & 1 192.146 564\
(3,1) - (2,1)a & 1 785.835 721 & 1 785.835 342\
(4,1) - (3,1)s & 2 379.444 522 & 2 379.443 925\
(4,1) - (3,1)a & 2 378.344 319 & 2 378.343 747\
(5,1) - (4,1)a & 2 968.507 005 & 2 968.506 396\
(5,1) - (4,1)s & 2 969.802 905 & 2 969.802 399\
[24]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [**]{} (, ) @noop [**]{} (, ) @noop [**]{} (, ) [****, ()](\doibase
http://dx.doi.org/10.1016/j.jqsrt.2012.02.037) [****, ()](\doibase
http://dx.doi.org/10.1016/j.jqsrt.2012.02.026) @noop [****, ()]{} [****, ()](\doibase
http://dx.doi.org/10.1016/j.jqsrt.2015.03.034) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, ) @noop [****, ()]{} @noop [****, ()]{}
[^1]: Present Address: Department of Chemistry & Biochemistry, Lamar University, Beaumont, TX 77710
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The robustness of neural networks to adversarial examples has received great attention due to security implications. Despite various attack approaches to crafting visually imperceptible adversarial examples, little has been developed towards a comprehensive measure of robustness. In this paper, we provide a theoretical justification for converting robustness analysis into a local Lipschitz constant estimation problem, and propose to use the Extreme Value Theory for efficient evaluation. Our analysis yields a novel robustness metric called CLEVER, which is short for **C**ross **L**ipschitz **E**xtreme **V**alue for n**E**twork **R**obustness. The proposed CLEVER score is attack-agnostic and computationally feasible for large neural networks. Experimental results on various networks, including ResNet, Inception-v3 and MobileNet, show that (i) CLEVER is aligned with the robustness indication measured by the $\ell_2$ and $\ell_\infty$ norms of adversarial examples from powerful attacks, and (ii) defended networks using defensive distillation or bounded ReLU indeed achieve better CLEVER scores. To the best of our knowledge, CLEVER is the first attack-independent robustness metric that can be applied to any neural network classifier.'
author:
- |
Tsui-Wei Weng^1^[^1],Huan Zhang^2^$^{*}$,Pin-Yu Chen^3^,Jinfeng Yi^4^,Dong Su^3^,Yupeng Gao^3^,\
**Cho-Jui Hsieh^2^, Luca Daniel^1^**\
^1^Massachusetts Institute of Technology, Cambridge, MA 02139\
^2^University of California, Davis, CA 95616\
^3^IBM Research AI, Yorktown Heights, NY 10598\
^4^Tencent AI Lab, Bellevue, WA 98004\
`[email protected], [email protected],`\
`[email protected], [email protected],`\
`{dong.su,yupeng.gao}@ibm.com, [email protected], [email protected]`
bibliography:
- 'iclr2018\_conference.bib'
title: 'Evaluating the Robustness of Neural Networks: An Extreme Value Theory Approach'
---
Conclusion {#sec:6}
==========
In this paper, we propose the CLEVER score, a novel and generic metric to evaluate the robustness of a target neural network classifier to adversarial examples. Compared to the existing robustness evaluation approaches, our metric has the following advantages: (i) attack-agnostic; (ii) applicable to any neural network classifier; (iii) comes with strong theoretical guarantees; and (iv) is computationally feasible for large neural networks. Our extensive experiments show that the CLEVER score well matches the practical robustness indication of a wide range of natural and defended networks.
**Acknowledgment.** Luca Daniel and Tsui-Wei Weng are partially supported by MIT-Skoltech program and MIT-IBM Watson AI Lab. Cho-Jui Hsieh and Huan Zhang acknowledge the support of NSF via IIS-1719097.
Appendix {#appendix .unnumbered}
========
Proof of Theorem \[thm:delta\_bnd\]
-----------------------------------
According to Lemma \[prop:Lips\], the assumption that $g(\bm{x}):=f_c(\bm{x})-f_j(\bm{x})$ is Lipschitz continuous with Lipschitz constant $L_q^j$ gives $$\label{eq:g_Lq}
| g(\bm{x}) - g(\bm{y}) | \leq L_q^j \| \bm{x} - \bm{y} \|_p .$$ Let $\bm{x} = \bm{x_0} + \bm{\delta}$ and $\bm{y} = \bm{x_0}$ in (\[eq:g\_Lq\]), we get $$| g(\bm{x_0} + \bm{\delta}) - g(\bm{x_0}) | \leq L_q^j \| \bm{\delta} \|_p,$$ which can be rearranged into the following form $$\label{eq:g_Lq_delta_rearrange}
g(\bm{x_0}) - L_q^j \| \bm{\delta} \|_p \leq g(\bm{x_0} + \bm{\delta}) \leq g(\bm{x_0}) + L_q^j \| \bm{\delta} \|_p.$$
When $g(\bm{x_0} + \bm{\delta}) = 0$, an adversarial example is found. As indicated by , $g(\bm{x_0} + \bm{\delta})$ is lower bounded by $g(\bm{x_0}) - L_q^j \| \bm{\delta} \|_p$. If $\| \bm{\delta} \|_p$ is small enough such that $g(\bm{x_0}) - L_q^j \| \bm{\delta} \|_p \geq 0$, no adversarial examples can be found:
$$\label{eq:bnd_for_all_delta}
g(\bm{x_0}) - L_q^j \| \bm{\delta} \|_p \geq 0 \Rightarrow \| \bm{\delta} \|_p \leq \frac{g(\bm{x_0})}{L_q^j} \Rightarrow \| \bm{\delta} \|_p \leq \frac{f_c(\bm{x_0})-f_j(\bm{x_0})}{L_q^j},$$
Finally, to achieve $\operatorname*{argmax}_{1\leq i \leq K} f_i(\bm{x_0}+\bm{\delta}) = c$, we take the minimum of the bound on $\| \bm{\delta} \|_p$ in (\[eq:bnd\_for\_all\_delta\]) over $j \neq c$. I.e. if $$\| \bm{\delta} \|_p \leq \min_{j \neq c} \frac{f_c(\bm{x_0})-f_j(\bm{x_0})}{L_q^j},$$ the classifier decision can *never* be changed and the attack will *never* succeed.
Proof of Corollary \[cor:aaa\]
------------------------------
By Lemma \[prop:Lips\] and let $g = f_c - f_j$, we get $L_{q,x_0}^j = \max_{y \in B_p(x_0,R)} \| \nabla g(y)\|_q = \max_{y \in B_p(x_0,R)} \| \nabla f_j(y) - \nabla f_c(y) \|_q$, which then gives the bound in Theorem 2.1 of [@hein2017formal].
Proof of Lemma \[lemma:3.3\]
----------------------------
For any $\bm{x}, \bm{y}$, let $\bm{d}=\frac{\bm{y}-\bm{x}}{\|\bm{y}-\bm{x}\|_p}$ be the unit vector pointing from $\bm{x}$ to $\bm{y}$ and $r=\|\bm{y}-\bm{x}\|_p$. Define uni-variate function $u(z)=h(\bm{x}+z\bm{d})$, then $u(0) = h(\bm{x})$ and $u(r) = h(\bm{y})$ and observe that $D^+ h(\bm{x}+z\bm{d}; \bm{d})$ and $D^+ h(\bm{x}+z\bm{d}; \bm{-d})$ are the right-hand and left-hand derivatives of $u(z)$, we have $$u'(z) = \begin{cases}
D^+ h(\bm{x}+z\bm{d}; \bm{d}) \leq L_q &\text{ if }
D^+ h(\bm{x}+z\bm{d}; \bm{d})= D^+ h(\bm{x}+z\bm{d}; -\bm{d})
\\
\text{undefined} &\text{ if }
D^+ h(\bm{x}+z\bm{d}; \bm{d}) \neq D^+ h(\bm{x}+z\bm{d}; -\bm{d})
\end{cases}$$ For ReLU network, there can be at most finite number of points in $z\in (0,r)$ such that $g'(z)$ does not exist. This can be shown because each discontinuous $z$ is caused by some ReLU activation, and there are only finite combinations. Let $0=z_0<z_1 < \dots < z_{k-1} < z_{k}=1$ be those points. Then, using the fundamental theorem of calculus on each interval separately, there exists $\bar{z}_i\in (z_i, z_{i-1})$ for each $i$ such that $$\begin{aligned}
u(r) - u(0) &\leq \sum_{i=1}^k |u(z_i)-u(z_{i-1}) | \\
& \leq \sum_{i=1}^k |
u'(\bar{z}_i) (z_i-z_{i-1}) | && \text{(Mean value theorem)}\\
& \leq \sum_{i=1}^k L_q |z_i - z_{i-1}|_p \\& = L_q \|x-y\|_p. && \text{($z_i$ are in line $(x,y)$)}\end{aligned}$$ Theorem \[thm:delta\_bnd\] and its corollaries remain valid after replacing Lemma \[prop:Lips\] with Lemma \[lemma:3.3\].
Theorem \[thm:Fx\_one\_hidden\] and its proof
---------------------------------------------
[D.1]{}\[$F_Y(y)$ of one-hidden-layer neural network\] \[thm:Fx\_one\_hidden\] Consider a neural network $f: {\mathbb{R}}^d \rightarrow {\mathbb{R}}^K$ with input $\bm{x_0} \in {\mathbb{R}}^d$, a hidden layer with $U$ hidden neurons, and rectified linear unit (ReLU) activation function. If we sample uniformly in a ball $B_p(\bm{x_0},R)$, then the cumulative distribution function of $ \| \nabla g(\bm{x}) \|_q$, denoted as $F_Y(y)$, is piece-wise linear with at most $M = \sum_{i=0}^d {U \choose i}$ pieces, where $ g(\bm{x}) = f_c(\bm{x}) - f_j(\bm{x})$ for some given $c$ and $j$, and $\frac{1}{p}+\frac{1}{q} = 1, 1 \leq p,q \leq \infty$.
The $j_{\text{th}}$ output of a one-hidden-layer neural network can be written as $$f_j(\bm{x}) = \sum_{r=1}^{U} \bm{V}_{jr} \cdot \sigma \left ( \sum_{i=1}^{d} \bm{W}_{ri} \cdot x_i + b_r \right ) = \sum_{r=1}^{U} \bm{V}_{jr} \cdot \sigma \left ( \bm{w}_r \bm{x} + b_r \right ),$$ where $\sigma (z) = \max (z, 0)$ is ReLU activation function, $\bm{W}$ and $\bm{V}$ are the weight matrices of the first and second layer respectively, and $\bm{w}_r$ is the $r_{\text{th}}$ row of $\bm{W}$. Thus, we can compute $g(\bm{x})$ and $\| \nabla g(\bm{x}) \|_q$ below: $$\begin{aligned}
g(\bm{x}) = f_c(\bm{x}) - f_j(\bm{x})
&= \sum_{r=1}^{U} \bm{V}_{cr} \cdot \sigma \left ( \bm{w}_r \bm{x} + b_r \right ) - \sum_{r=1}^{U} \bm{V}_{jr} \cdot \sigma \left ( \bm{w}_r \bm{x} + b_r \right ) \\
&= \sum_{r=1}^{U} ( \bm{V}_{cr} - \bm{V}_{jr} ) \cdot \sigma \left ( \bm{w}_r \bm{x} + b_r \right )\end{aligned}$$ and $$\begin{aligned}
\| \nabla g(\bm{x}) \|_q &= \left \| \sum_{r=1}^{U} \mathbb{I}(\bm{w}_r \bm{x} + b_r) ( \bm{V}_{cr} - \bm{V}_{jr} ) \bm{w}_r^\top \right \|_q,\end{aligned}$$ where $\mathbb{I}(z)$ is an univariate indicator function: $$\mathbb{I}(z) = \Big \{
\begin{tabular}{cc}
1, & \text{if $z>0$,} \\
0, & \text{if $z \leq 0$.}
\end{tabular}$$
As illustrated in Figure \[fig:fig\_bound\], the hyperplanes $\bm{w}_r \bm{x} + b_r = 0, r \in \{ 1, \ldots, U \}$ divide the $d$ dimensional spaces ${\mathbb{R}}^d$ into different regions, with the interior of each region satisfying a different set of inequality constraints, e.g. $\bm{w}_{r_+} \bm{x} + b_{r_+} > 0$ and $\bm{w}_{r_-} \bm{x} + b_{r_-} < 0$. Given $\bm{x}$, we can identify which region it belongs to by checking the sign of $\bm{w}_r \bm{x} + b_r$ for each $r$. Notice that the gradient norm is the same for all the points in the same region, i.e. for any $\bm{x}_1$, $\bm{x}_2$ satisfying $\mathbb{I}(\bm{w}_r \bm{x}_1 + b_r) = \mathbb{I}(\bm{w}_r \bm{x}_2 + b_r)$ $\forall r$, we have $\| \nabla g(\bm{x}_1) \|_q = \| \nabla g(\bm{x}_2) \|_q$. Since there can be at most $M = \sum_{i=0}^d {U \choose i}$ different regions for a $d$-dimensional space with $U$ hyperplanes, $\| \nabla g(\bm{x}) \|_q$ can take at most $M$ different values.
Therefore, if we perform uniform sampling in a ball $B_p(\bm{x_0},R)$ centered at $\bm{x_0}$ with radius $R$ and denote $\| \nabla g(\bm{x}) \|_q$ as a random variable $Y$, the probability distribution of $Y$ is discrete and its CDF is piece-wise constant with at most $M$ pieces. Without loss of generality, assume there are $M_0 \leq M$ distinct values for $Y$ and denote them as $m_{(1)}, m_{(2)}, \ldots, m_{(M_0)}$ in an increasing order, the CDF of $Y$, denoted as $F_Y(y)$, is the following: $$F_Y(m_{(i)}) = F_Y(m_{(i-1)}) + \frac{\mathbb{V}_d(\{\bm{x} \mid \| \nabla g(\bm{x}) \|_q = m_{(i)} \}) \cap \mathbb{V}_d(B_p(\bm{x_0},R)))}{\mathbb{V}_d (B_p(\bm{x_0},R))}, i = 1, \ldots, M_0,$$ where $F_Y(m_{(0)}) = 0$ with $m_{(0)} < m_{(1)}$, $\mathbb{V}_d(E)$ is the volume of $E$ in a $d$ dimensional space.
Additional experimental results
-------------------------------
### Percentage of examples having p value $> 0.05$
Table \[tab:p-values\] shows the percentage of examples where the null hypothesis cannot be rejected by K-S test, indicating that the maximum gradient norm samples fit reverse Weibull distribution well.
### CLEVER v.s. number of samples
Figure \[fig:bnd\_samp\_app\] shows the $\ell_2$ CLEVER score with different number of samples ($N_b=50,100,250,500$) for MNIST and CIFAR models. For most models except MNIST-BReLU, reducing the number of samples only change CLEVER scores very slightly. For MNIST-BReLU, increasing the number of samples improves the estimated lower bound, suggesting that a larger number of samples is preferred. In practice, we can start with a relatively small $N_b = a$, and also try $2a, 4a, \cdots$ samples to see if CLEVER scores change significantly. If CLEVER scores stay roughly the same despite increasing $N_b$, we can conclude that using $N_b = a$ is sufficient.
[^1]: Tsui-Wei Weng and Huan Zhang contributed equally
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'With the rapid advancement of Big Data platforms such as Hadoop, Spark, and Dataflow, many tools are being developed that are intended to provide end users with an interactive environment for large-scale data analysis (e.g., IQmulus). However, there are challenges using these platforms. For example, developers find it difficult to use these platforms when developing interactive and reusable data analytic tools. One approach to better support interactivity and reusability is the use of micro-level modularisation for computation-intensive tasks, which splits data operations into independent, composable modules. However, modularizing data and computation-intensive tasks into independent components differs from traditional programming, e.g., when accessing large scale data, controlling data-flow among components, and structuring computation logic. In this paper, we present a case study on modularizing real world computation-intensive tasks that investigates the impact of modularization on processing large scale image data. To that end, we synthesize image data-processing patterns and propose a unified modular model for the effective implementation of computation-intensive tasks on data-parallel frameworks considering reproducibility, reusability, and customization. We present various insights of using the modularity model based on our experimental results from running image processing tasks on Spark and Hadoop clusters.'
author:
-
bibliography:
- 'IEEEbig.bib'
title: 'Micro-level Modularity of Computaion-intensive Programs in Big Data Platforms: A Case Study with Image Data'
---
Modularisation; computation-intensive; image processing; map-reduce;
Introduction
============
With the rapid advancement of Big Data platforms, software systems [@genap; @pairs; @tomogram] are being developed that provide end-users’ an interactive environment for large-scale data analysis in the area of scientific research, business, governments, and journalism. An interactive environment provides a drag-and-drop facility for composing reproducible computation (workflows/pipelines) from a collection of sub-tasks without much technical knowledge. Big Data platforms such as Hadoop [@hadop], Spark [@Spark], Google Dataflow [@data_model], and so on provide high-level abstract interfaces for implementing distributed-cluster processing of data using commodity hardware. Recently, a number of researchers [@cloud_archi; @graywulf; @module_soft; @iabdt; @hbaselarge] focused on developing architectures and frameworks for large-scale data analysis tools utilizing these platforms. Most of these architectural frameworks are adopting workflows and pipelines [@traverna; @kepler; @illumina] for reproducible data analysis tasks according to the user requirements. However, data-storage models, data-structures, data-operations, accessing and visualization of large data are complex to handle. Existing literature [@model_deployment] suggest that significant effort is spent developing data processing pipelines. Besides, a recent empirical study [@notall] reports that data engineers are facing great difficulties to work with Big Data platforms. In order to reduce the development efforts and providing better programming flexibility, a few studies attempted to develop more abstract and unified programming interfaces (especially in Bioinformatics and GIS research) [@sparkseq; @biosspark; @spatialhadoop; @stormcv; @kira; @hipi] as a layer on top of these platforms (e.g., Hadoop and Spark). However, most of them are still in the development phase (e.g., SparkSeq [@sparkseq]), and some of them only implemented and tested a few specific tasks within a certain domain. Although, a few works implemented large scale image processing tasks [@stormcv; @kira; @hipi; @iabdt], they did not provide an analysis study about the underlying challenges and solutions of using these platforms for real world image processing pipelines. Moreover, common unified frameworks are not readily available to implement reproducible image processing pipelines covering a wide area with Big Data platforms. A few researchers [@tomogram; @highregistration] have attempted to tune cluster resources for performance optimization of the tasks. Nevertheless, resource enhancement may not provide a feasible solution even with the availability of enough computing power. Therefore, interactive large-scale data-analysis with a Big-Data platform is still a challenging task for the programmers and developers.
Modularization is an important paradigm in software design which provides special program constructs, such as shared data structures or abstract and unified frameworks. Modularization is the action of *“de-composing a system into modules”* [@leveragedr]. Moreover, modularization is essential for scalable and interactive application development with Big Data platforms [@mrrunner; @model_deployment; @module_soft]. Our focus is on modularising interactive, data-intensive programs so that they operate effectively on map-reduce frameworks in order to support reusability, reproducibility, and customization. Splitting tasks considering large scale data processing and computation logic reusability may have adverse effects when run on Big Data platforms. However, micro-level modularity has been shown to work successfully on map-reduce frameworks for a number of applications, including machine learning [@knnis; @bigkmean] and graph data processing [@biggraph; @graphx]. We are also motivated to support developers of Big Data analytic tools. By separating tasks into further independent micro-components based on data-processing patterns, we hope to develop a unified programming interface that will provide the flexibility for accelerating the development of interactive, re-usable Big Data analytics tools.
Although Big Data platforms hide the complexity of distributed computing, they provide a limited number of methods (e.g., *map, filter, reduce*) for data parallel operations. Adding an extra data processing step with those methods could increase the computation and memory overhead in a significant way. For example, Smith and Albarghouthi [@mapreduce] discuss the challenge of partitioning computation with data-parallel operators (*map, filter, reduce*). Due to the complexity of partitioning, they avoid optimization of their technique. In order to reduce working efforts, a few authors [@mrrunner; @model_deployment] focused on developing frameworks for running and re-deploying modular jobs and a statistical model provided by the users. Unfortunately, none of them conducted extensive study on any effective techniques for modularity to examine the impact of modularity on computation-intensive tasks. Moreover, the mechanism of controlling data-flow among intermediate steps is really important for reproducible computation of large scale data. All things considered, we propose a modularity model and observe the behaviors of different applications in terms of modularization. Overall, in our work, we mainly focus on two research questions:
**RQ1.** How to modularize data and computation-intensive programs to provide a unified abstract framework for developing interactive tools?
**RQ2.** How does splitting up of run time job-data and processing logic affect the performance of computation-intensive tasks in map-reduce platforms?
In order to answer **RQ1**, we analyzed various open source image processing tools and state-of-the-art image processing techniques that cover a wide range of tasks. We look into the programming models and data-types that are produced during a full image processing task (some of them are presented in Table \[tab:steps\]). Then, we categorized the image operations and defined a data processing pattern that is fruitful for modularizing the tasks with Big Data platform. After that, we proposed a micro-level modularity model consisting of four major data-parallel modules each having three core layers (the second layer controls parallel data-flow). For answering **RQ2**, we implemented six image processing applications following the proposed modular model. Then we experimented with both the minimal and modularised version with various datasets in a Spark cluster. From the experiment, we found that the task modularization affects system’s performance and flexibility of pipeline development. Performance varies case by case with some tasks improving, some decreasing, and others unaffected. For all the cases, it opens up the facility of flexible implementation with data-parallel components. Notably, we also identified the challenges of image processing with data-parallel frameworks from our experimentation. In summary, our case study provides a modularization technique and helpful knowledge-base for interactive tools developers for large scale image processing. Our defined data-pattern and modularity model can be used as a design pattern and design rule in this domain.
The rest of the paper is organized as follows. Section II describes the process of extracting data-processing patterns. Section III presents our proposed modular model. Section IV provides our experimental results. Section V provides discussion and some useful insights from the lesson learned. Section VI describes related work. Finally, section VII presents the conclusion and future work of the paper.
Modularising Data-intensive Tasks
=================================
In our study we focused on image data since a framework that supports various image processing pipelines is not readily available. On the contrary, few abstract frameworks [@kira; @hipi; @icp] are being developed for a few specific image processing applications and most of the Big Data frameworks support workflows for text data processing [@spatialhadoop; @biosspark]. We conduct our case study following three major strategies: (i) Background and Contextual Analysis, (ii) Data-processing Pattern Extraction, and (iii) Transformation to Data-parallel Components.
Background and Contextual Analysis
----------------------------------
To develop a unified framework, understanding the context is essential. To that end, literature review and analysis of various architectures [@cloud_archi; @graywulf; @module_soft; @iabdt], frameworks, tools, techniques, and open-source APIs in the scientific data analysis are essential to determine the exact support needed for the data scientist. Analyzing the recent development strategy of analytic tools for large scale data, we notice that some of the developed applications follow a workflow based modularity architecture [@module_soft; @graywulf], whereas others follow a layered architecture [@tomogram; @iabdt; @pairs]. In the workflow based modularity architecture, applications are designed using a special data model which is much different than the traditional model view controller model. For example, the architecture of IQmulus [@module_soft], a GIS data processing system, is heavily dependent on data-analysis workflows. High-level components, job manager, processing services etc. are designed focusing on the on-the-fly workflow compositions. Still users need to learn a considerable amount of script for composing workflows for GIS. Similarly, GrayWulf [@graywulf] handles two types of workflows: (i) one is for data manager, and (ii) another is for end-users. The architectural model is based on these workflows composition. However, using GrayWulf, a smaller amount of processed result can be shared and retrieved in the cloud. Another application for image analysis, IABDT [@iabdt] followed multi-layer architecture and primarily used HadoopImageBundle (HIB) for performing basic operations on image data. In a recent study, Roy et al. [@cloud_archi] focus on data-centric component development for an application that supports large scale data analysis. Besides, most of the unified frameworks to support applications development as mentioned above followed specialized data-models for large scale data processing with distributed clusters. Among existing popular frameworks in the scientific analysis, SparkSeq is based on Hadoop-BAM [@hadoopbam] data frameworks. Hadoop-BAM is created to solve the issue of map-reduce implementation and attempted to include all data formats in bioinformatics. A unified framework for large scale Geospatial data analysis, SpatialHadoop [@spatialhadoop] added three more layers on top of Hadoop to drive efficient map-reduce based processing of GIS data. KIRA [@kira] is written using SEP library and FITS data model for analyzing the astronomical object. All of the evidence prompt that tool development in Big Data platforms requires different design rule and modularity models. Yet, the common obvious advantages of modularization [@leveragedr] in software development are: (i) Easier to Debug and Problem Detection, (ii) Reusable Code, (iii) Readability, and (iv) Reliability. Debugging time is lengthy during the development of Big Data analytic tools. Most of the time, ultimate problems cannot be detected until the application is run on the live cluster with the full set of data. In summary, for large-scale data analysis the following trends are emerging:
\(i) Suitable architectural model [@cloud_archi; @module_soft; @iabdt],
\(ii) Work-flow processing and management [@traverna; @kepler; @module_soft],
\(iii) Data-pipelines [@illumina],
\(iv) Data-flow management [@data_model; @pig; @udf_dataflow],
\(v) Data-centric decoupling of programs [@graphx; @biggraph],
\(vi) Efficient data-storage model [@spatialhadoop; @pairs], and
\(vii) Intelligent modularization [@mrrunner; @model_deployment].
The central objective of all of these paradigms is to make scientific computation reproducible [@reproducable_computation] with minimal technical knowledge. Figure \[workflow\_comp\] demonstrates how reproducible workflows/pipelines are constructed. However, large scale image processing domain requires more focus on all of the above-mentioned directions. In order to understand and develop a knowledge-base, we look into the properties of various open source image processing tools [@imagej; @htpheno; @plantcv; @imageharvest; @bisque; @kira; @hipi; @stormcv].
{height="1.10in" width="3.2in"}
\[workflow\_comp\]
Analyzing source code of the open-source tools facilitate us more intuitive insight about the implementation of real world image analysis applications, programming models, I/O operations, and data entities that analysts, researchers, or end-users might re-use later. We also observe that in image processing tools following attributes are influential (some of them are identified by Heit et al. [@model_deployment] in data mining as well): (i) Image pipeline composition, reuse and management, (ii) Image processing workflow modeling language, (iii) Image storage service, (iv) Collaboration between data scientists, (v) Deployment and third party service communication, (vi) Scalability, and (vii) Plugin development and integration. Another key thing to remember is that data structure and computation model of images are complex and diverse [@imagej]. Images ($I$) consist of different data units (8-bit, 16-bit and 32-bit), formats (TIFF, GIF, JPEG, BMP, DICOM, FITS), or dimensions/channels (2D, 3D, 3-channels and so on). Additionally, we found that many algorithms are operated on an individual image except for machine learning/statistical model and template generations. For developing a desired image processing task, programmers and researchers need to experiment with various combinations of techniques and algorithms (hundreds of algorithms are available) along with parameters tuning for each of the canonical operations. Moreover, these operations are tested on large collections of images multiple times and the experimental setup needs to be stored for the future run. Many core image operations can be found within a popular open-source image processing API called OpenCV [@opencv]. Therefore, if a framework can be devised that also facilitates an automatic transformation of iterative operations for a single image into parallel (processing with multiple computers) one for multiple images will be valuable for the data scientists. All of this knowledge-base is useful for unified framework development and reshaping the modularization for Big Data frameworks. In the following sections, we will present analysis study of various image processing tasks to extract data-processing patterns and transforming the concept into the data-parallel framework.
Program Synthesis and Extracting Data Processing Patterns {#data_pattern}
---------------------------------------------------------
From the previous discussion it is persuasive that in large-scale data processing, most of the techniques, algorithms, frameworks, and software models are extensively data-centric [@cloud_archi; @bigdataopportunity; @udf_dataflow]. In data-centric development, at first a core-feature model is developed, then data processing patterns are extracted from feature model, targeted technologies and real-world experience, and finally, components are designed and modularized based on the pattern. Therefore, understanding data-processing patterns is an important part of implementing modularised, split and decoupled data processing applications. Our selected algorithms and techniques cover various image processing tasks in plant science, agriculture, biomedical, astronomy, and general computer vision. A total number of 30 applications we analyzed are presented in Table \[tab:itasks\]. Some of the selected image processing tasks with the major steps and their corresponding produced entities are presented in Table \[tab:steps\]. The high-level steps are shown in Table \[tab:steps\] of extracting texts [@banglaocr] from a video are: gray-scale conversion and noise removal, feature calculation, detecting text areas, then extract texts from the segmented areas of the video images. We notice that produced output of various steps has different data structures in most of the cases. However, it is necessary to figure out an optimal and unified model that might be fitted for a wide area. Most of the image analysis tasks can be divided into four re-usable tasks (please note that here we consider single image analysis task, while two or more image analysis tasks are used to compose a complex pipeline like HtPheno [@htpheno]). Many tasks have more than four steps. However, in terms of data and program reusability, for image registration (presented in Table \[tab:steps\]), matching points calculation ($S_{3}$) and Homography generation ($S_{4}$) can be considered as one logical and independent step. Similarly, in Tomograms generation, refined class objects are reused later, thus $S_{3}$ and $S_{4}$ can be combined into a single step logically.
\[tab:itasks\]
**Area of Application** **Number of Image processing Applications**
------------------------- ---------------------------------------------
Plant and Agriculture 14
Medical and Biology 5
Astronomy 4
General computer vision 7
\[tab:steps\]
[p[.20in]{}|p[0.60in]{}|p[0.60in]{} |p[0.55in]{}|p[0.60in]{} ]{}
**** & **Img Classification [@img_class]** & **Img Registration [@img_regi]** & **Text Extraction[@banglaocr]** & **Pattern in Tomograms[@tomogram]**\
& & & &\
& & & &\
& & & &\
& & & &\
& – & & &\
\
\
Therefore, analyzing the above-mentioned tools and techniques, we categorize the canonical operations of image analysis tasks into major four steps:
**Preprocess/conversion ($S_{1}$):** This step is the first and very common for every image analysis pipeline. This step may produce different kinds of output (such as grayscale image and Canny edge image) based on applied techniques or algorithms, such as Gaussian blurring, wavelet transform, image contrasting, enhanced, noise reduction and so on [@opencv].
**Estimate/Extraction ($S_{2}$):** In this step, different kinds of algorithms such as SURF, SIFT, ORB, HOG [@opencv] are applied for calculating features, metrics, and key points. However, other texture generation techniques are also employed after the feature and keypoint extraction step. This step produces an array, vector or list type data-structures.
**Model/Fitting ($S_{3}$):** This step uses extracted features, metrics or composed data for fitting, training or developing models for generating templates based on which final analysis and processing are done.
**Analysis/Postprocess ($S_{4}$):** This final step mainly produces processed images and analysis result based on the generated template or the model in the model-fitting step. The produced results of this step include matched images, extracted objects, clusters of images, and registered images along with statistical results.
Such a categorization of the operations based on produced data and computation logic would help developers and programmers to wrap image processing tasks into a common data model and abstract frameworks. Furthermore, program synthesis of the above-mentioned image processing tools, their operations, and I/O operations allowed us to come to a conclusion that produced data in various steps as discussed in Table \[tab:steps\] can also be saved for later reuse. Consequently, these tasks should be modularised not only for program reuse but also for data entities re-use. In summary, the data processing pattern for image processing tasks can be described as follows (as shown in Table \[tab:steps\]):
- Input of first step ($S_{1}$) is generally $\{I, R_{S1}\}$, where I is raw images and $R_{S1}$ is parameters, produced data are processed images $I_{P}$. Parameters may be numeric values, meta-data, vectors, or even raw images.
- Other two steps ($S_{2}$,$S_{3}$) input are $\{I_{P}, R_{S(2/3)}\}$, where $I_{P}$ is the produced entity of the previous step and $R_{S(2/3)}$ represents parameters of these steps.
- However, in some cases the input of the last step ($S_{4}$) is $\{I, I_{P}, D_{M}, R_{S4}\}$; where $D_{M}$ is the model or template generated in the third steps, and $I_{P}$ is the outcome from the first step.
- We observe that many cases, $I$ and $I_{P}$ are required to flow and retain up to the last step which is handled with disk storage in localized processing.
- For a few cases, images are required to group or bundle during $S_{3}$ and $S_{4}$ (e.g., image registration and panoramic view generation). Likewise, produced results in image processing tasks have various types. Those are a single image, a collection of images or image objects ($I_{R}$), list of string or numeric values ($L_{s}$), collection of matrix or vectors ($L_{M}$), dictionary ($D_{S}$), tuple of lists ($T_{L}$), and so on.
This common pattern is the basis for interactive image analytic tools development for both usual and large scale data. In the next subsequent sections, we will discuss in details how to implement this data-processing patterns into map-reduce frameworks considering a unified programming interface.
Transformation to Data-parallel Modules
=======================================
In this section, we discuss how to implement the image processing tasks into modularised and abstract steps in Big Data frameworks following the extracted data-processing pattern. We focus on Apache Spark (with HDFS) implementation which is optimized and the mostly used [@mapreduce; @knnis] framework. Here, the data-processing pattern serves as modularity properties. We will use many terms and symbols to avoid frequent use of the phrases in our description (many of them are introduced by Smith and Albarghouthi [@mapreduce]).
### Challenges
Recent works [@mapreduce; @knnis; @graphx; @bigkmean; @biggraph] with map-reduce frameworks provide firm evidence that map-reduce based implementation is non-trivial for flexible and scalable data processing. Moreover, many applications are yet to a good fit for Big Data platforms using traditional map-reduce techniques due to network induced non-determinism, data shuffling [@mapreduce], and run-time data increment [@knnis]. For example, researchers are still working to make KNN more feasible for large data with Big Data platforms [@knnis]. Storage files of text and Genome data could be partitioned into further smaller blocks for efficient distributed processing. But data file of each image and associated meta-data is required to treat as a single unit for image processing. Few images among thousands of collection might be corrupted and disrupt the whole processing task. This scenario is also required to handle during large scale processing. However, all the operations in map-reduce based platform (i.e., Spark) should be done with the data parallel components ($\sum _{DP}$): *map(), reduce(), filter(), join(), repartition(), subtractbykey(), count(), collect()* along with *$\lambda-$expressions* (PABS) [@mapreduce]. All the image operations cannot be easily paralleled with this platform. When data size is big enough, a single additional operation with $\sum _{DP}$ takes a significant amount of time. Moreover, broadcasting data entity frequently to the worker processes might add further overhead. Consequently, programmers are required to be more careful and thoroughly test with a full dataset. For reusable computation, each step should be independent in terms of execution, data sharing, and data storing. Having said that, steps should not be divided arbitrarily like usual programming. As we discussed in Section \[data\_pattern\], raw-data, processed data, and external parameters need to flow from one step to another, and this might increase both memory and time overhead (with the number of steps). Apart from these, handling of various types of produced results (as described in the data processing pattern) requires a well-defined rule to store in a distributed environment.
### Proposed Modularity Model
Image processing tasks can be implemented in a various number of modularized steps (one or more) with data-parallel frameworks as shown in Figure \[modular\]. Here we introduce data-parallel module, $M_{DP}=\bigcup\limits_{i=1}^{n} DP_{i}$ as a combination of one or more data-parallel components in $\sum _{DP}$. The split into $M_{DP}$ is followed by the corresponding data-processing patterns presented in Section \[data\_pattern\]. From the analysis of data processing patterns of image processing applications, we identified four canonical steps: $S_{1}, S_{2}, S_{3},S_{4}$. A step is a combination of many operations (some of them are canonical also), and it is essential to detect which operations require parallelism and which parts do not. We can represent $S_{i}=\{\sum_{OS}, NP_{S}\}$, where $\sum_{OS}$ represents operations that require parallelism, and $NP_{S}$ represents not parallel. All $\sum_{OS}$ within a $M_{DP}(S_{i})$ should be combined in such a way that the number of $\sum _{DP}$ are minimal (i.e., this rule restricts the modularity of usual computation). However, for a few steps in some cases, run-time data should be partitioned (based on heuristics [@knnis]) for further optimization (as shown in Listings \[goodflowercount\]). A module, $M_{DP}$ must produce a meaningful outcome that can be reused in future either by one of the independent operations in $S_{1}$ to $S_{4}$ or another task (or pipelines). However, a complete Image analysis task could be implemented with one or two minimal steps in map-reduce frameworks (Figure \[modular\]). As we observe, most of the cases first two steps– $S_{1}$ and $S_{2}$ can be executed with one component in $\sum _{DP}$. These two steps can be combined into one $M_{DP}$. Other two steps– $S_{3}$ and $S_{4}$ require more than one components in $\sum _{DP}$. Another key thing to remember is that in data-parallel components, input entities and parameters are a different thing (Smith and Albarghouthi [@mapreduce] define them as *arity* and *free variable* respectively). Sometimes, step $S_{4}$ requires the input parameters which value is calculated from either $S_{1}$ or $S_{2}$. Consequently, $S_{3}$ should be in a separate $M_{DP}$. Similarly, $S_{4}$ requires the input parameter calculated from all collective elements from $S_{3}$. Therefore, $S_{4}$ is separate from $S_{3}$. For many image processing tasks, $S_{3}$ and $S_{4}$ combined into a single data-parallel step. However, for reusable and customization perspective, we propose to wrap up the independent meaningful four steps into four $M_{DP}$– $M_{DP}(S_{1}), M_{DP}(S_{2}), M_{DP}(S_{3}), M_{DP}(S_{4})$. As we noticed in data-processing patterns, in many cases outcomes of the steps are required to flow and retain among intermediate steps (even up to the last step). That poses a challenge to data-parallel implementation as this data flow may increase both run-time memory and execution overhead. We present a solution considering a common list of data-entities with defined order to link-up data-flows among the $M_{DP}$. We recommend a three-layers vertical implementation of $M_{DP}$ for image pipelines which are presented in Figure \[mdp\]. Layer-1 consists of abstract interfaces and $\sum _{DP}$, layer-2 handles parallel data-flow (DPF) and order of data entities (pseudo code is shown in Listing \[api\_code\]), and layer-3 contains $S_{i}$ on images. Data-parallel operations could be optimized using layer-1 without considering others. Layer-3 also works as a bridge to include image processing libraries (Skimage, OpenCV). Processing logic in this layer can be improved without the knowledge of layer-1. Components of Layer-1 call components in Layer-2, and Layer-2 call components in the lower layer. Therefore, three layers version of data-parallel module, $M_{DP}(S_{i}) = \sum _{DP} .>DPF.>\sum _{OS}[N,R]\{NP_{S}\}$. Here *N* is the input entities (similar to RDD elements in Spark) populated by PABS, and *R* is the list of parameters as described in data-processing patterns, $I$ and $I_{P}$ can be common in *N*. Only $\sum _{DP}$ (via PABS) will call $\sum _{OS}$ through $DPF$. This modularity model provides a multidimensional (3x4, three layers and four modules) separation of concerns and dependency inversion principle (which is valuable for parallel development as distributed programming experts and image processing experts are not the same people usually). This will give the tool developer a common programming model to rapidly implementing the sequential tasks into Big Data platform. Finally, we recommend to save $I_{R}$ and $L_{M}$ into distributed storage, other types of result should be stored either in flat storage or databases. A block diagram is shown in Figure \[interface\] on how common programming interface could be utilized using the $M_{DP}$ and processing patterns for interactive workflow/pipeline development. However, in the experimentation phase, we will discuss on what will be the impact of modularization and maintain a common list of data entities for each $M_{DP}$.
{height="1.70in" width="3.60in"}
\[modular\]
{height="2.50in" width="2.80in"}
\[mdp\]
{height="1.75in" width="3.5in"}
\[interface\]
Impact Analysis of Modularization
=================================
In this section, we will discuss our experiments to observe the modularising effect of the image processing applications based on the modularity model presented in the previous section. We have three different Big Data infrastructure: (i) built with stand alone frameworks (Hadoop and Spark), (ii) built with Cloudera platform, and (iii) another is built with OpenStack. Big Data infrastructure built with Cloudera provides much flexibility of cluster setup and configuration, but our experimentation reveal it is a bit slower than stand-alone system. On the other hand, OpenStack facilitates dynamic node creation without further installing the frameworks and libraries (Cassandra, OpenCV, and so on) for each machine. In this paper, we present our experimental result with the second and third infrastructure cluster. The first infrastructure, Cloudera cluster consists of seven worker nodes (physical machines) with total 58 cores and 56GB RAM. The model of the processors is Intel Xeon L5420 and the speed is 2.50GHz. The master node is configured to 36GB main memory for the cluster driver. In ComputeCanada resource, our infrastructure is built by OpenStack with 5 instances each has 8 cores and 30GB RAM (Intel Xeon CPU E5-2650, 2.60GHz). This infrastructure is slightly better in terms of processor speed and main memory. Therefore, experiments with different configurations provide us the intuitive understanding of how the DCM behaves in different contexts. We conduct our experiment on four datasets: a set of CANOLA field images (most of them contain flowers, each image size is 1280x720), two sets of crop field images (each image size is 1280x960), and LSVRC2015 [^1]. We implemented the programs with Spark-2.0.1 and Python 2.7. We compared the execution time between modularised and compact versions of six image processing tasks: (i) Image matching [@imgmatching], (ii) Image classification [@imgclass], (iii) CANOLA flower count (modified version of the base algorithm [@flowercount] for B-Channel), (iv) Object extraction [@segment_water], (v) Image registration [@img_regi], and (vi) Mosaic image generation [@panoramic]. We utilize the OpenCV [@opencv] image processing library to implement the operations in Layer-3 for the DCM. We found interesting behavior of differences between modularised and non-modularised versions. In the case of time-intensive operations, there is a significant performance issue when modularization is used for some tasks (i.e., Image Registration). However, in many cases, there is no significant differences such as object extraction, matching, and mosaic image generation.
From the experimentation with the Compute Canada infrastructure, shown in Table \[performance\_computecanada\], we notice that the difference in execution time $\Delta(t)$ for counting flowers from 2K images, $T_{min}-T_{op}$ = 2.0 $-$1.9 = +0.1 minutes (+0.6 for slower cluster Table\[performance\_little\]), while for 8.6K (3.9GB) images, the modular version overcomes the run-time memory exceeding issue of the master node (with 30GB RAM). That means modularising the tasks facilitates performance optimization for individual step (we improved the performance by splitting runtime data at the third step according to the number of map partitions). For image registration, $\Delta(t)$ is about $-4$ minutes slower for both data size for the modular version, that means arbitrary modularization affects performance significantly. The performance is affected because of extensive data flow (image bundles) from the first data-parallel module to the last one. Whereas, the minimal step version has virtually no parallel data-flow. Likewise, for the larger dataset, $\Delta(t)$ for modular image classification is $-7$ minutes (for first infrastructure in Table \[performance\_little\] this difference is $-13$). All other cases, $\Delta(t)$ is almost 0, meaning no impact on execution time between modular and non-modular versions. We also experimented with the LSVRC2015 dataset for matching, clustering, and object extraction (as other pipelines require crops field images) and found that modularity does not impact execution time (presented in Table \[iccv\_performance\]) negatively for the large collection of images. On the other hand, for lower configured infrastructure, the time difference is higher despite enough resources.
However, we found some challenging tasks during our experiment. For the 2K images, the performance of the flower counting task is feasible, but we found that for the 8.6K dataset either the execution time is unusual or causing memory exceeding issues during runtime. Modularization and further splitting up the third data-parallel module of the flower counting task solved the memory issue, and the execution time decreases. However, in the mosaic image generation task, processing 300 images takes more than 400 minutes (while the well-configured machine takes 210 minutes for the usual program) for both the minimal and modular version; with the increment of images, the complexity arises (603 images take more than 600 minutes). Above all, although for a few tasks, modularity increases execution time, we can say that data-centric modularity increases the opportunity to further optimization for many cases.
\[performance\_computecanada\]
[p[.60 in]{}|p[0.35in]{} |p[.50in]{}|p[0.50in]{} || p[0.35in]{} |p[.50in]{}|p[0.50in]{} |p[0.20in]{}]{} **Tasks** & **\#Img** & **$T_{min}$** & **$T_{mod}$**& **\#Img** & **$T_{min}$** & **$T_{mod}$**& **$N_{S}$**\
IMatch & 2K & 2.2& 2.2& 8.6K &9& 9 & 1\
Clustering & 2K &6.6& 6.4& 4K & 15& 18 & 2\
FCount& 2K & **2**& **1.9** & 8.6K & **MemI**& **7.2**& 3\
OBE & 2K & 0.30& 0.30& 8.6K & 0.83& 0.83 & 1\
IMReg& 0.5K & **8** & **12**& 1.5K & **25** & **32** & 1\
Mosaic & 0.2K & $>$300& $>$300& 0.3K & $>$400& $>$400 & 2\
\
\
\
\[performance\_little\]
**Tasks** **\#Img** **$T_{min}$** **$T_{mod}$** **\#Img** **$T_{min}$** **$T_{mod}$** **$N_{S}$**
------------ ----------- --------------- --------------- ----------- --------------- --------------- -------------
IMatch 2K 3.3 3.3 8.6K 13 13 1
Clustering 2K 11 11 4K 18 27 2
FCount 2K **6.1** **5.5** 8.6K **MemI** **19** 3
OBE 2K 0.8 0.8 8.6K 2.1 2.1 1
IMReg 0.5K **9.3** **13** 1.5K **27** **40** 1
Mosaic 0.2K $>$300 $>$300 0.3K $>$400 $>$400 2
\[iccv\_computecanada\]
**Tasks** **\#Img** **$T_{min}$** **$T_{mod}$** **\#Img** **$T_{min}$** **$T_{mod}$** **$N_{S}$**
----------- ----------- --------------- --------------- ----------- --------------- --------------- -------------
IMatch 30K 4.6 4.6 60K 8 8 1
OBE 30K 1.9 1.9 60K 3.2 3.2 1
\[iccv\_performance\]
**Tasks** **\#Img** **$T_{min}$** **$T_{mod}$** **\#Img** **$T_{min}$** **$T_{mod}$** **$N_{S}$**
----------- ----------- --------------- --------------- ----------- --------------- --------------- -------------
IMatch 30K 11 11 60K 22 22 1
OBE 30K 3.3 3.2 60K 6.5 6.5 1
Apart from these, if we follow common data processing patterns as described in Section \[data\_pattern\], it is possible to write common data-parallel modules at the micro-level for the high-level components of large data analytic tools. This not only facilities re-usable module development but also we can write a common abstract interface to work with image processing without much knowledge of data-parallel operations and tuning. If separate operations are implemented as common method signatures within a class and its object instances are passed through corresponding data-parallel module, then processing logic can be reused or customized willingly without the knowledge of data-parallel components. With our model, the image processing tasks which contains only parallelisable operations in a processing step $S_{i}$ can be easily transformed into data-parallel programs without knowing the details of Layer-1 and Layer-2. Furthermore, if the input parameters for the DCM modules do not depend on the outcome of any of the steps in $S_{i}$ (and no non-parallel operations) then those DCMs for image processing tasks can be automatically converted into a data-parallel module. Wrapping all of these concepts, we also developed a library called SHIPPI which is being used by other programmers for developing and deploying image processing pipelines. Coupled with the data-processing pattern (in Section \[data\_pattern\]), our model represents a strong design rule [@leveragedr] in this domain. For instance, consider a project where one team is working on the web part, one team is working on the efficient large-scale data processing support, and another team is working on the image processing part; here dependency inversion principle (depend upon abstractions; do not depend upon concretions [@dependency_inversion]) is essential. We are working for more sophisticated techniques and algorithms to design a module to auto-transform the image pipelines into data-parallel modules, but this is out of the scope of our study. One critical observation is that for a map-reduce framework, developers should avoid arbitrary modularisation, unlike usual programming. Along with the large-scale data-handling and modules re-usability, our data-centric model explicitly supports the scenarios S3, S5, S6, S11, S12 described in [@cloud_archi]. We believe, this modularity model decreases the architectural changes of the core modules in a greater extent since a DCM senses and pre-defines most of the possibilities of a sub-domain as we have shown in the case study.
Challenges of Unified Data Interface for Large Scale Images
-----------------------------------------------------------
IO and storage model is crucial for the performance of massive data processing. In order to solve the problems, research on unified data interface has been gaining traction [@diana]. However, we have not found extensive experimentation on image data for unified data interface. Here, we provide (only for large data among our many other IOs) our experience with various IO and storage model with the state-of-the-art technologies for our designed unified data interface for the data-centric modules. Primary options for the IO and storage model for large-scale images are: (i) Distributed storage (HDFS), (ii) Flat storage (both local and remote machine), and (iii) Database. IO with the distributed storage HDFS is faster but one problem is that processed images are saved as text data and another local program needs to convert it to images for making the result usable by the researchers. However, we experimented four types of model; one of the models we implemented is parallel loading and saving images via worker nodes utilizing SSH protocol which can handle up to 2K images only. The experimental outcome of other three models is presented in Table \[io\_performance\]. We found that IO for images are still challenging for feasible image analysis tasks, for instance, reading HDFS and writing to flat storage (master node) takes $\sim16$ minutes for 60K (5.7GB) images while HDFS-HDFS takes 84 seconds. However, despite more IO time, the benefit of HDFS-flat model is that the analysts get the ready-to-use processed images. On the other hand, for data and result management, Cassandra reveals the most effective data storage model (among other NoSQL storage, Cassandra is one of the most efficient models [@casperform; @hadoop_cas] for Big Data). In summary, an effective I/O model is essential for the unified interface for handling large-scale images. Our data-centric module supports configurable storage-models at run-time for the storage of intermediate results as shown in Table \[modular\].
\[io\_performance\]
**IO Model** **1.5k(Crp)** **4k(Flwr)** **8.6K(Flwr)** **30K(VRC)** **60K(VRC)**
-------------- --------------- -------------- ---------------- -------------- --------------
Cndra-Cndra 12s 16s 26s 24s 30s
Hdfs-Hdfs 11s 16s 25s 46s 84s
Hdfs-flat 1.4mins 4.8mins 9.2mins 8.1mins 16.1 mins
Apart from these, if we follow common data processing patterns, it is possible to write common data-parallel modules ($M_{DP}$) at the micro-level for the high-level components of large data analytic tools. This not only facilities re-usable module development but also we can write a common abstract interface to work with image processing without much knowledge of data-parallel operations ($\sum _{DP}$) and tuning. If separate operations are implemented as common method signatures (as shown in Figure \[interface\] and Listing \[api\_code\]) within a class and its object instances are passed through corresponding $M_{DP}$, then processing logic can be reused or customized (shown in Figure \[reusable\]) willingly without the knowledge of data-parallel components. With our model, the image processing tasks which contains only $\sum_{OS}$ in $S_{i}$ can be easily transformed into data-parallel programs without knowing the details of Layer-1 and Layer-2. Furthermore, if $R_{S1}$ to $R_{S4}$ do not depend on the outcome of any of the steps in $S_{i}$ (and no $NP_{S}$) then those image processing tasks can be automatically converted into $M_{DP}$. Coupled with the data-processing pattern, it is logical to treat the model as a strong design pattern and design rule [@leveragedr] in this domain (for instance, consider a project where one team is working for the web part, one team is working for the efficient large scale data processing support, and another team is working for the image processing part; here dependency inversion principle is essential). More sophisticated techniques and algorithms might provide a framework to auto-transform the image pipelines into $M_{DP}$ in future which are out of the scope of our study.
DPFEstimate(N, obj, Rs2)
unpack(N)-->im_id, I, Ip
metrics = obj.estimate(Ip, Rs2)
...
return pack(im_id, I, Ip, metrics)
DPFModel(N, obj, Rs3)
unpack(N)-->im_id,I, Ip, metrics
Dm = obj.model(Ip, metrics, Rs3)
...
return pack(im_id, I, Ip, Dm)
{height="1.25in" width="3.6in"}
\[reusable\]
Discussion
==========
Pr_RDD <-- RDDraw.map(preprocess)
Fs_RDD <-- Pr_RDD.map(faetureExtract)
Mosaic <-- Fs_RDD.first()
Tmp_RDD <-- Fs_RDD.zipWithIndex().cache()
#split = #elements / split_size
traversed.add(Mosaic)
do until all_images traversed
for i=0 to #split
start = i * split_size
end = start + split_size
Filter_RDD <-- Tmp_RDD.filter(not(traversed) and in(start, end))
MFeature <-- broadcast(faetureExtract(Mosaic))
Matched_RDD <-- Filter_RDD.map(matchpoints, MFeature)
MaxImg <-- Matched_RDD.reduce(max(matchratio))
if(MaxImg.ratio > previous_ratio)
SeletedImg <-- MaxImg
traversed.add(SeletedImg)
Mosaic <-- mergeHomography(Mosaic,SeletedImg)
result <-- Mosaic
**Lesson learned**: In summary, from our case study we extracted the following important insights:
- Still, I/O operations create a bottleneck for optimal image processing with data-parallel frameworks
- Considering modular data-processing patterns will reduce the implementation efforts and increase the reusability of both the program and the processed entities in data-parallel frameworks.
- Programmer should avoid arbitrary modularisation.
- Programmer should not rely on usual map-reduce concepts and tune hardware resources only for computation intensive tasks.
- Intelligently splitting up the map-reduce operations and run-time data further might solve the limited resource problems as well as increase performance.
- All image processing tasks may not be a good fit for traditional map-reduce techniques.
Related Work
============
A number of studies [@knnis; @bigkmean; @hadoop_mosaic; @biggraph] have pointed out the challenges and problems of implementing the computation intensive tasks for scientific data with the abstract data-parallel frameworks in spite of having enough computing resources. To reduce the efforts of the data-scientists for large scale data analysis, some applications and frameworks are being developed [@spatialhadoop; @hadoopbam; @sparkseq; @module_soft] for GeoSpatial and Bioinformatic data processing by adding more abstract layers on top of map-reduce frameworks. Despite enough progress, they do not support image processing operations. However, a few studies attempted to develop software and tools [@icp; @iabdt; @tomogram; @imageharvest; @dedip; @kira; @hipi] for large scale image processing for few specific cases. Nonetheless, they do not provide a common framework for diverse image processing pipelines. Our objective is to develop a scalable, unified and abstract framework for developing interactive image processing pipelines for large scale data.
Nowadays, large-scale images are used for analysis in various scientific works and general computer vision. Although a few studies provide techniques [@hipi; @kira; @stormcv; @hadoop_mosaic; @highregistration] for specific image processing tasks with data-parallel frameworks, they do not describe the challenges and optimization techniques to overcome the challenges. In this study, we highlight the challenges of real world large scale image processing tasks as well as recommend optimization technique with data-parallel components.
It is proven that program modularization is a key concept for developing unified frameworks on top of distributed and map-reduce programming environment. However, modularising map-reduce job (computation and data-intensive) is still challenging as data-parallel frameworks only provide a limit of few strict API methods. Yang et al. [@mrrunner] attempted to develop a framework for running modular map-reduce jobs, but users need to provide modularized jobs and dependency information. Recently, Heit et al. [@model_deployment] proposed a modular architecture for working with statistical models for data mining. However, these studies do not provide any technique of micro-level modularity and impact of modularizing the tasks. In our work, we propose a strategy for modularising large-scale image processing tasks at the micro-level and illustrate the pros and cons of modularising tasks with data-parallel frameworks.
In summary, our study on computation-intensive task analysis strategy, modularity model, and experimental insights will provide the researchers to focus on such challenges extensively for devising better techniques, and developers to consider the insights during large scale image processing tools development.
Conclusion
==========
In this paper, we presented a case study on modularising data and computation intensive tasks into micro-level components. Our focus is on large image data as there is a lack of studies on the implications of running a wide variety of image processing tasks on Big Data platforms. We synthesize image data-processing patterns and propose a unified modular model for the effective implementation of computation-intensive tasks on data-parallel frameworks considering reproducibility, reusability, and customization. Our experimental results with six real world image processing tasks show that splitting and modularising the computation tasks is crucial to utilize the power of Big Data platforms. However, not all tasks show similar performance in execution time after modularising. A few of them need more sophisticated techniques for optimization with data-parallel frameworks. Therefore, our study provides a valuable knowledge-base for abstract and unified frameworks development for large scale data analysis. In future, we will work on techniques for automatic transformation of sequential tasks to data-parallel modules.
[^1]: http://image-net.org/challenges/LSVRC/2015/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we show that to a unital associative algebra object (resp. co-unital co-associative co-algebra object) of any abelian monoidal category ${\mathcal{C}}$ endowed with a symmetric $2$-trace, one can attach a cyclic (resp. cocyclic) module, and therefore speak of the cyclic (co)homology of the (co)algebra “with coefficients in $F$". We observe that if ${\mathcal{M}}$ is a ${\mathcal{C}}$-bimodule category equipped with a stable central pair then ${\mathcal{C}}$ acquires a symmetric 2-trace. The dual notions of symmetric $2$-contratraces and stable central contrapairs are derived as well. As an application we can recover all Hopf cyclic type (co)homology theories, obtain a conceptual understanding of anti-Yetter-Drinfeld modules, and give a formula-free definition of cyclic cohomology. The machinery can also be applied in settings more general than Hopf algebra modules and comodules.'
author:
- 'Mohammad Hassanzadeh, Masoud Khalkhali, Ilya Shapiro'
title: 'Monoidal Categories, 2-Traces, and Cyclic Cohomology'
---
monoidal category (18D10), abelian and additive category (18E05), cyclic homology (19D55), Hopf algebras (16T05).
Introduction
=============
One of the major advances in cyclic cohomology theory in recent years was the introduction of a new cohomology theory for Hopf algebras by Connes and Moscovici and its extension to a cohomology theory for (co)algebras endowed with an action of a Hopf algebra and with coefficients in a local system [@CM98; @hkrs1]. The local systems are closely related, and in a sense they are a mirror image of, Yetter-Drinfeld modules over a Hopf algebra. Beyond Hopf algebras, one often encounters interesting algebraic objects: for example quasi Hopf algebras and weak Hopf algebras, that only possess some of the axioms of Hopf algebras. Developing a Hopf cyclic-type theory for these Hopf-like objects is one of the motivations for this paper.
We find that the language of monoidal categories is a suitable framework to discuss this question and to discover an answer. The key point, for us, of the categorical machinery is the notion of a trace. Traces in (monoidal) categories are well-known and have been used in different settings. In [@DP] the categorical definition of ordinary traces of square matrices are used for endomorphisms of a dualizable object in a symmetric monoidal category. In the derived category of a ring, traces are called Lefschetz numbers. Categorical traces are also used to study fixed-point theory [@P]. The authors in [@JSV] showed that traces can be defined for dualizable objects in a monoidal category. On the other hand, if we think of a monoidal category as a bicategory with one object, then the notion of trace that we want is similar to the generalized traces of [@PS]. Finally the authors in [@FSS] found a relationship between the category-valued traces and the twisted center of a monoidal category.
The technical aspects of this paper would have been much more involved and would have required much delicacy, had the notions of bimodule categories and their centers not been already extensively studied. In particular, as we realized, the center of a certain bimodule category of the monoidal category of (left) modules over a Hopf algebra (sometimes called the twisted center of the monoidal category) provides the suitable coefficients for Hopf cyclic cohomology. These coefficients were already known, called (stable) anti Yetter-Drinfeld modules [@hkrs1], but were defined much less conceptually. We use these (or dual) central elements to form a suitable categorical trace to define the desired homological objects.
More precisely, in this paper we use the notion of a [*symmetric $2$-trace*]{} (compare with the shadow structure in [@PS]) for an abelian monoidal category to show that for a monoidal category endowed with such a trace one can attach a cyclic module to any unital associative algebra object. Later we introduce the notion of stable central pair in a monoidal category which is a practical way of obtaining a symmetric $2$-trace.
If ${\mathcal{M}}$ is a ${\mathcal{C}}$-bimodule category then so is $Fun({\mathcal{M}}, {\text{Vec}})$; let ${\mathcal{Z}}_{{\mathcal{C}}}({\mathcal{M}})$ and ${\mathcal{Z}}_{{\mathcal{C}}}Fun({\mathcal{M}}, {\text{Vec}})$ denote their respective centers. If $F\in{\mathcal{Z}}_{{\mathcal{C}}}Fun({\mathcal{M}}, {\text{Vec}})$ and $m\in {\mathcal{Z}}_{{\mathcal{C}}}({\mathcal{M}})$ then the pair $(F, m)$ is called a stable central pair if it satisfies one additional mutual compatibility condition given in Definition \[stablepair\]. We show that any such pair gives us a symmetric $2$-trace.
As an example we see that the monoidal category ${{_H\mathcal{M}}}$ of left modules over a Hopf algebra $H$ can be endowed with a stable central pair and therefore a symmetric $2$-trace. To construct this pair, we consider the ${\mathcal{C}}$ bimodule category ${{^{\#^{-1}}_H\mathcal{M}}}$ where the right ${\mathcal{C}}$-action is given by the monoidal tensor product and the left action is twisted by $S^{-2}$, where $S$ is the antipode of $H$, as we will explain in Section \[ydsection\]. Then if we start with a central element $M\in {\mathcal{Z}}_{{\mathcal{C}}}({{^{\#^{-1}}_H\mathcal{M}}})$, and take $$F:=\Hom_H(1,-)\in {\mathcal{Z}}_{{\mathcal{C}}}Fun({{^{\#^{-1}}_H\mathcal{M}}}, {\text{Vec}}),$$ we observe that $(F, M)$ is a stable central pair (provided that an extra stability condition is satisfied). More interestingly, we prove that the elements of the center of the ${\mathcal{C}}$-bimodule category ${{^{\#^{-1}}_H\mathcal{M}}}$ are nothing but the “duals" of the well-known anti Yetter-Drinfeld modules over $H$. On the other hand, if we pursue a contravariant theory, then the bimodule category of interest is ${{^{\#}_H\mathcal{M}}}$ and we show that $(F,M)$ is a stable central contrapair if $$F:=\Hom_H(-,1),$$ $M$ is the usual anti Yetter Drinfeld module and the stability condition is the usual one [@hkrs1].
We recall that the center of a monoidal category has been studied for different reasons. It is known that the elements of the center of the monoidal category of modules over a Hopf algebra, a weak Hopf algebra, and a quasi Hopf algebra are in fact the Yetter-Drinfeld modules which are the solutions of the quantum Yang-Baxter equations. If we consider this monoidal category as a bimodule category over itself by the left and right actions given by the monoidal tensor product, then the Yetter-Drinfeld modules do indeed form the center of this bimodule category.
One notes that the language of monoidal categories is fundamental to the study of Hopf-like objects. More precisely, for such an object, the category of left modules over it is a monoidal category. Often the axioms that specify the type of the Hopf-like object are themselves dictated by exactly this requirement. There are a great many results about recovering the original Hopf-like object from its associated monoidal category provided that some extra structure (a variation on the fiber functor theme) is provided. This explains the important relation between monoidal categories and Hopf algebras. There are also other categorical approaches to cyclic homology [@BS], [@KP]. We observe that our categorical machinery can be applied to the monoidal categories associated to interesting Hopf-like objects such as weak Hopf algebras, Hopf algebroids, quasi-Hopf algebras and Hopfish algebras to obtain homological constructions such as cyclic homology.\
**Acknowledgments**: The authors would like to thank the organizers of the “Noncommutative Geometry Workshop" at the University of Western Ontario, June 2015, where this paper began.
Motivation
----------
The authors in [@hkrs1] introduced Hopf cyclic cohomology with coefficients for four types of symmetries. In the case of a $H$-module coalgebra $C$, for a right-left stable anti Yetter-Drinfeld module (SAYD) over $H$, they assign a cocyclic module structure to $C^n=\Hom_H(k, M\ot C^{\ot n+1})$. This theory generalizes Connes-Moscovici’s Hopf cyclic cohomology theory [@CM98]. In the case of a $H$-module algebra $A$, and also for a right-left stable anti Yetter-Drinfeld module (SAYD) over $H$, they assign a cocyclic module structure to $C^n=\Hom_H(M\ot A^{\ot n+1}, k)$. The comodule part of the anti-Yetter-Drinfeld (AYD) module structure appears in the cyclic map $\tau$. The mysterious AYD structure has not been conceptually well-understood in the literature, although it is known that this structure is obtained by replacing the antipode $S$ by $S^{-1}$ in the definition of a Yetter-Drinfeld (YD) module. On the other hand, the YD modules are well-understood as they form the center of the monoidal category of $H$-modules, ${{_H\mathcal{M}}}$. Not only was the categorical meaning of AYD modules not understood, but also it was not clear why such a mysterious structure is needed to obtain a cocyclic module and therefore cyclic cohomology.
### Contravariant cohomology theory.
To answer this question, one can start from scratch and try to define the cyclic map $\tau$ on $C^n=\Hom_H(M\ot A^{\ot n+1}, k)$ directly. This is the only significant addition to the already apparent cosimplicial structure (in the case that $A$ is an algebra, unital and associative of course). More precisely, we need to *slide the first copy of $A$ past $M$ and then to the back.* To understand the idea better let us consider a special case when the monoidal category is rigid, such is the category of *finite* dimensional left modules over a Hopf algebra $H$. Later we will see that the finiteness assumption can be removed. Using the standard adjunction properties of rigidity, for any $V, W\in {{_H\mathcal{M}}}$, we have:
$$\Hom_{H}(V\ot W, 1)\simeq \Hom_{H}(V, 1\ot {W^*})\simeq \Hom_{H}(V, {W^*}\ot 1)\simeq \Hom_H({W^{**}}\ot V, 1).$$
Thus for an $M\in {{_H\mathcal{M}}}$ and an algebra object $A$ (though the algebra structure plays no role at this stage) in ${{_H\mathcal{M}}}$ we obtain
$$\label{adj22} \Hom_{H}(M\ot A^{\ot n+1}, 1)\simeq \Hom_{H}({A^{**}}\ot M\ot A^{\ot n}, 1).$$
If we have an extra condition on $M$, namely that $$\label{centre22}{A^{**}}\ot M \simeq M\ot A ,$$ then we obtain the desired $\tau$, i.e., $$\tau_n:\Hom_{H}(M\ot A^{\ot n+1}, 1)\simeq \Hom_{H}({A^{**}}\ot M\ot A^{\ot n}, 1) \simeq \Hom_{H}(M\ot A^{\ot n}\ot A,1),$$ where we first use , followed by the inverse of . This suggests that for a rigid monoidal category ${\mathcal{C}}$ $$AYD({\mathcal{C}})=\{M \in {\mathcal{C}}: {A^{**}}\ot M \xrightarrow{\sim} M\ot A , \, \forall A\in {\mathcal{C}}\},$$ with some compatibility conditions.
Knowing that $YD$ modules form the center of the monoidal category, i.e., $YD=\mathcal{Z}({\mathcal{C}})$, we see that $AYD$ is to $YD$ as ${A^{**}}\ot M \xrightarrow{\sim} M\ot A$ is to $A\ot M\xrightarrow{\sim} M\ot A$. The stability condition in the case of the usual stable anti Yetter-Drinfeld modules ensures that $\tau^{n+1}_n=\Id.$ To obtain the same conclusion in our general case leads us to the requirement that the single cyclic map $\tau_0$ be the identity. In summary, from the above considerations, we guess that whereas $YD= \mathcal{Z}( {{_H\mathcal{M}}}_{fd})$ (center of a monoidal category), $AYD=\mathcal{Z}({{{^{**}_H\mathcal{M}}}}_{fd})$ (center of a bimodule category) where ${{{^{**}_H\mathcal{M}}}}_{fd}$ is simply ${{_H\mathcal{M}}}_{fd}$ with the left action modified by ${(-)^{**}}$. This guess turns out to be correct. Note that both $\tau_n$ and $$AYD({\mathcal{C}})=\mathcal{Z}_{{\mathcal{C}}}({^{**}{\mathcal{C}}})$$ make sense for any rigid category.
### Covariant cohomology theory.
If instead we consider $C^n=\Hom_H(k, M\ot C^{\ot n+1})$ and try to define $\tau$ directly, we need to *slide the first copy of $C$ past $M$ and then to the back.* Again let us consider the *finite* dimensional left modules over a Hopf algebra $H$. Using the standard adjunction properties of rigidity, for any $V, W\in {{_H\mathcal{M}}}$, we have:
$$\Hom_{H}(1, V\ot W)\simeq \Hom_{H}( 1\ot {^* W}, V)\simeq \Hom_{H}( {^* W}\ot 1, V)\simeq \Hom_H(1, {^{**}W}\ot V).$$
Thus for an $M\in {{_H\mathcal{M}}}$ and a coalgebra object $C$ (though the coalgebra structure plays no role at this stage) in ${{_H\mathcal{M}}}$ we obtain
$$\label{adj2} \Hom_{H}(1, M\ot C^{\ot n+1})\simeq \Hom_{H}(1, {^{**}C}\ot M\ot C^{\ot n}).$$
If we have an extra condition on $M$, namely that $$\label{centre2}{^{**}C}\ot M \simeq M\ot C ,$$ then we obtain the desired $\tau$, i.e., $$\tau_n:\Hom_{H}(1, M\ot C^{\ot n+1})\simeq \Hom_{H}(1, {^{**}C}\ot M\ot C^{\ot n}) \simeq \Hom_{H}(1, M\ot C^{\ot n}\ot C),$$ where we first use the inverse of followed by the inverse of . This suggests that for a rigid monoidal category ${\mathcal{C}}$ we also need $$YD_1({\mathcal{C}})=\{M \in {\mathcal{C}}: {^{**}C}\ot M \xrightarrow{\sim} M\ot C , \, \forall C\in {\mathcal{C}}\},$$ with some compatibility conditions, to serve as coefficients. Note that $YD_1$ is not the same as $AYD$, it is “dual" to it.
Preliminaries {#prelim}
=============
Here we collect some background material that should facilitate the reading of this paper. The content of Sections \[ydsection\] and \[stabilitysection\] is new. The discussion involving Yetter-Drinfeld modules, anti-Yetter-Drinfeld modules, and their generalizations contained in Section \[ydsection\] is especially important. The conceptual reinterpretation of these objects and their associated complicated formulas was one of the motivations for this paper.
(Co)cyclic modules.
-------------------
The main goal of this paper is to introduce a suitable categorical language to unify different notions of cyclic homology under a single theory. Therefore we need to recall the definitions of cyclic and cocyclic modules from [@Connes-cyclic] and [@loday].
Recall that the simplicial category $\Delta$ has as its objects non-negative integers considered as totally ordered sets $[n]=\{0,1,\cdots,n\}$ and its morphisms are non-decreasing functions $[n]\to[m]$. A simplicial module is a contravariant functor from $\Delta$ to ${\text{Vec}}$. Similarly, a cosimplicial module is a covariant functor. By keeping the same objects and adding cyclic permutations we obtain Connes cyclic category $C$. A cyclic module is again a contravariant functor from $C$ to ${\text{Vec}}$, while a cocyclic module is a covariant one.
More explicitly, a [*cosimplicial module*]{} is given by the data $(C_{n},\delta_{i}, \sigma_{i})$ where $\{C_{n}\}$, $n\geq 0$ is a sequence of vector spaces over the field $k$. The maps $\delta_{i}:C^{n}\rightarrow C^{n+1}$ are called cofaces, and $\sigma_{i}:C^{n}\rightarrow C^{n-1}$ are called codegeneracies. These are $k$-linear maps satisfying the following cosimplicial relations: $$\begin{aligned}
\label{rel1}
\begin{split}
\delta_{j} \delta_{i} &= \delta_{i} \delta_{j-1}, \quad i <j,\\
\sigma_{j} \sigma_{i} &= \sigma_{i} \sigma_{j+1}, \quad i \leq j,\\
\sigma_{j} \delta_{i} &=
\begin{cases}
\delta_{i} \sigma_{j-1}, &
i<j\\
\text{Id}, & i=j \,\,\text{or}\,\, i=j+1,\\
\delta_{i-1} \sigma_{j}, & i>j+1.
\end{cases}
\end{split}\end{aligned}$$
A [*cocyclic module* ]{} is a cosimplicial module equipped with the extra morphisms $\tau_n :C^n\rightarrow C^n$, called cocyclic maps such that the following relations hold: $$\begin{aligned}
\label{rel2}
\begin{split}
\tau_{n}\delta_{i}&=\delta_{i-1} \tau_{n-1}, \quad 1\le i\le n,\\
\tau_{n} \delta_{0}&=\delta_{n},\\
\tau_n \sigma_{i}&=\sigma_{i-1} \tau_{n+1}, \quad 1\le i\le n,\\
\tau_n^{n+1}&=\Id.\\
\end{split}\end{aligned}$$
In a dual manner, one can define a cyclic module as a simplicial module with extra cyclic maps. More precisely, a cyclic module is given by the data $(C_{n},\delta_i, \sigma_{i}, \tau_{n})$, where $C_{n}$, $n\geq 0$ is a $k$-vector space and $\delta_{i}: C_{n}\rightarrow C_{n-1}, \quad \sigma_{i}:C_n \ra C_{n+1}, \quad 0 \leq i \leq n$, and $\tau_n:C_n \ra C_n$, are called faces, degeneracies and cyclic maps respectively. These are $k$-linear maps, satisfying the following relations: $$\begin{aligned}
\label{rel11}
\begin{split}
\delta_{i} \delta_{j}&= \delta_{j-1} \delta_{i}, \quad i <j,\\
\sigma_{i} \sigma_{j}&= \sigma_{j+1} \sigma_{i}, \quad i \leq j,\\
\delta_{i} \sigma_{j} &=
\begin{cases}
\sigma_{j-1} \delta_{i}, & i<j,\\
\Id, & i=j \,\,\text{or}\,\, i=j+1,\\
\sigma_{j} \delta_{i-1}, & i>j+1.
\end{cases}
\end{split}\end{aligned}$$ and $$\begin{aligned}
\label{rel22}
\begin{split}
\delta_{i}\tau_n&= \tau_{n-1}\delta_{i-1}, \quad 1\le i\le n,\\
\delta_{0} \tau_n&= \delta_{n}\\
\sigma_{i} \tau_n&= \tau_{n+1} \sigma _{i-1},\quad 1\le i\le n,\\
\tau_n^{n+1}&= \Id. \\
\end{split}\end{aligned}$$ One notes that the relation $\sigma_{0} \tau_n = \tau_{n+1}^2 \sigma_{n}$ that is usually listed along with the above is an extra relation [@loday section 5.2] which can be obtained from $\tau^{n+1}_n=\Id$ and $\sigma_{i} \tau_n= \tau_{n+1} \sigma _{i-1}$. Similarly for a cocyclic module $\tau_n \sigma_{0} = \sigma_{n} \tau_{n+1}^2$ can be obtained from the other relations. From a (co)cyclic module, one can define Hochschild, cyclic and periodic cyclic (co)homology [@loday].
$H$-modules, $H$-comodules, and compatibility conditions. {#ydsection}
---------------------------------------------------------
Recall that the center of the monoidal category of left $H$-modules, ${{_H\mathcal{M}}}$, is equivalent to the category of left-right Yetter-Drinfeld modules. For finite dimensional Hopf algebras the center $\mathcal{Z}({{_H\mathcal{M}}})$ is also equivalent to the representations of the quantum double $_{D(H)}\mathcal{M}$. For details, we refer the reader to [@kassel]. We recall from [@majid; @sch] that for a Hopf algebra (or a bialgebra) $H$, a left $H$-module, right $H$-comodule $M$ is called a Yetter-Drinfeld module if $$\label{YD1}
h{~\hspace{-3pt}^{^{\left(1\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(2\right)}}} m{~\hspace{-3pt}_{\left<1\right>}}=(h{~\hspace{-3pt}^{^{\left(2\right)}}} m){~\hspace{-3pt}_{\left<0\right>}}\ot (h{~\hspace{-3pt}^{^{\left(2\right)}}} m){~\hspace{-3pt}_{\left<1\right>}}h{~\hspace{-3pt}^{^{\left(1\right)}}}.$$ For Hopf algebras with an invertible antipode this is equivalent to $$\label{YD2}
(hm){~\hspace{-3pt}_{\left<0\right>}}\ot (hm){~\hspace{-3pt}_{\left<1\right>}}= h{~\hspace{-3pt}^{^{\left(2\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(3\right)}}}m{~\hspace{-3pt}_{\left<1\right>}}S^{-1}(h{~\hspace{-3pt}^{^{\left(1\right)}}}).$$ The isomorphism giving the central structure of a left right YD module $M$ is $$\begin{aligned}
V\ot M&\xrightarrow\sim M\ot V\\v\ot m&\mapsto m{~\hspace{-3pt}_{\left<0\right>}}\ot m{~\hspace{-3pt}_{\left<1\right>}}v\end{aligned}$$ and the YD condition simply ensures that the map above is that of $H$-modules.
We recall from [@hkrs2] that a left-right anti Yetter-Drinfeld module $M$ over a Hopf algebra $H$ is a left $H$-module and a right $H$-comodule satisfying $$\label{AYD1}
(hm){~\hspace{-3pt}_{\left<0\right>}}\ot (hm){~\hspace{-3pt}_{\left<1\right>}}= h{~\hspace{-3pt}^{^{\left(2\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(3\right)}}}m{~\hspace{-3pt}_{\left<1\right>}}S(h{~\hspace{-3pt}^{^{\left(1\right)}}}).$$ We denote the category of left-right AYD modules over a Hopf algebra $H$ by $_H\mathcal{AYD}^H$. Note that there are three additional flavors of AYD modules: left-left, right-left, right-right. All of them are equivalent and thus we focus only on the left-right variety. We will need to generalize.
Let $M$ be a left module and a right comodule over $H$, and let $i\in\mathbb{Z}$. We say that $M\in {_H\mathcal{YD}}^H_i$ if $$\label{genyd1}(h{~\hspace{-3pt}^{^{\left(2\right)}}}m){~\hspace{-3pt}_{\left<0\right>}}\ot (h{~\hspace{-3pt}^{^{\left(2\right)}}}m){~\hspace{-3pt}_{\left<1\right>}} S^{-2i}(h{~\hspace{-3pt}^{^{\left(1\right)}}})=h{~\hspace{-3pt}^{^{\left(1\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(2\right)}}}m{~\hspace{-3pt}_{\left<1\right>}}.$$ We say that $M$ is a generalized Yetter-Drinfeld module.
The following lemma provides a characterization of the generalized Yetter-Drinfeld modules akin to the one above for the Yetter-Drinfeld modules.
For a Hopf algebra $H$ the generalized $i$th YD condition is equivalent to $$\label{genyd2}
\rho(hm)=(hm){~\hspace{-3pt}_{\left<0\right>}}\ot(hm){~\hspace{-3pt}_{\left<1\right>}}=h{~\hspace{-3pt}^{^{\left(2\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(3\right)}}}m{~\hspace{-3pt}_{\left<1\right>}} S^{-1-2i}(h{~\hspace{-3pt}^{^{\left(1\right)}}}).$$
Thus ${_H\mathcal{YD}}^H_0={_H\mathcal{YD}}^H$, while ${_H\mathcal{YD}}^H_{-1}={_H\mathcal{AYD}}^H$, and we will also need ${_H\mathcal{YD}}^H_1$ to serve as coefficients of the covariant theory. Note that if instead of ${{_H\mathcal{M}}}$ we consider ${{_H\mathcal{M}}}_{fd}$ then $\mathcal{YD}_1=\mathcal{AYD}^*=\mathcal{AYD}\, contramodules$.
First, we show that $\eqref{genyd2}\Rightarrow\eqref{genyd1}$: $$\begin{aligned}
(h{~\hspace{-3pt}^{^{\left(2\right)}}}m){~\hspace{-3pt}_{\left<0\right>}}\ot (h{~\hspace{-3pt}^{^{\left(2\right)}}}m){~\hspace{-3pt}_{\left<1\right>}} S^{-2i}(h{~\hspace{-3pt}^{^{\left(1\right)}}}) &=h{~\hspace{-3pt}^{^{\left(3\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(4\right)}}} m{~\hspace{-3pt}_{\left<1\right>}}S^{-1-2i}(h{~\hspace{-3pt}^{^{\left(2\right)}}})S^{-2i}(h{~\hspace{-3pt}^{^{\left(1\right)}}})\\
&=h{~\hspace{-3pt}^{^{\left(2\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(3\right)}}}m{~\hspace{-3pt}_{\left<1\right>}}\varepsilon(h{~\hspace{-3pt}^{^{\left(1\right)}}})\\
&=h{~\hspace{-3pt}^{^{\left(1\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(2\right)}}}m{~\hspace{-3pt}_{\left<1\right>}}.
\end{aligned}$$ Now we show that $\eqref{genyd1}\Rightarrow\eqref{genyd2}$: $$\begin{aligned}
h{~\hspace{-3pt}^{^{\left(2\right)}}}m{~\hspace{-3pt}_{\left<0\right>}}\ot h{~\hspace{-3pt}^{^{\left(3\right)}}}m{~\hspace{-3pt}_{\left<1\right>}}S^{-1-2i}(h{~\hspace{-3pt}^{^{\left(1\right)}}}) &=(h{~\hspace{-3pt}^{^{\left(3\right)}}}m){~\hspace{-3pt}_{\left<0\right>}}\ot (h{~\hspace{-3pt}^{^{\left(3\right)}}}m){~\hspace{-3pt}_{\left<1\right>}}S^{-2i}(h{~\hspace{-3pt}^{^{\left(2\right)}}})S^{-1-2i}(h{~\hspace{-3pt}^{^{\left(1\right)}}})\\
&=(h{~\hspace{-3pt}^{^{\left(2\right)}}}m){~\hspace{-3pt}_{\left<0\right>}}\ot (h{~\hspace{-3pt}^{^{\left(2\right)}}}m){~\hspace{-3pt}_{\left<1\right>}}\varepsilon(h{~\hspace{-3pt}^{^{\left(1\right)}}})\\
&=\varepsilon(h{~\hspace{-3pt}^{^{\left(1\right)}}})\rho(h{~\hspace{-3pt}^{^{\left(2\right)}}}m)\\
&=\rho(\varepsilon(h{~\hspace{-3pt}^{^{\left(1\right)}}})h{~\hspace{-3pt}^{^{\left(2\right)}}}m)\\
&=\rho(hm).
\end{aligned}$$
If ${\mathcal{C}}$ is a monoidal category, ${\mathcal{M}}$ a ${\mathcal{C}}$-bimodule category, and $F: {\mathcal{M}}\rightarrow {\mathcal{M}}$ a monoidal endofunctor, then we use ${\mathcal{M}}^F$ and ${^F{\mathcal{M}}}$ to denote the bimodule categories with the right and respectively left actions twisted by $F$. More precisely, for $V,W\in{\mathcal{C}}$ and $M\in{^F{\mathcal{M}}}$ we have $$V\cdot_{new} M\cdot_{new} W=F(V)\cdot_{old} M\cdot_{old} W,$$ with ${\mathcal{M}}^F$ defined analogously. Note that if $F$ is an equivalence then ${^F{\mathcal{M}}}\simeq{\mathcal{M}}^{F^{-1}}$.
Let $\#: {{_H\mathcal{M}}}\rightarrow {{_H\mathcal{M}}}$ be the functor which takes a left $H$-module $M$ to $M^\#\in {{_H\mathcal{M}}}$ where $M^\#$ is the same as $M$ as a vector space but the left action is modified by $S^{2}$, i.e., is now given by $h\cdot m= S^{2}(h)m$. If, as we always assume, $S$ is invertible, then $\#$ is an autoequivalence of ${{_H\mathcal{M}}}$. Thus for $i\in\mathbb{Z}$ we can consider ${{^{\#^i}_H\mathcal{M}}}$, i.e., $$V\cdot M\cdot W=V^{\#^i}\ot M\ot W.$$
We can now repeat verbatim the same arguments as in the usual, YD modules Vs center, case. Roughly speaking, let $M$ be in ${\mathcal{Z}}_{{{_H\mathcal{M}}}}{{^{\#^i}_H\mathcal{M}}}$, then $M$ is already a left $H$ module and for every $V\in{{_H\mathcal{M}}}$ we have an isomorphism $\Phi:V^{\#^i}\ot M\xrightarrow\sim M\ot V$. Take $V=H$ and define the right comodule structure on $M$ via $$\rho(m)=\Phi(1\ot m).$$ Conversely, suppose that $M$ is in ${_H\mathcal{YD}}^H_{-i}$. Then for every $V\in{{_H\mathcal{M}}}$ define $\Phi$ by $$\label{mcent}\Phi(v\ot m)=m{~\hspace{-3pt}_{\left<0\right>}}\ot m{~\hspace{-3pt}_{\left<1\right>}}v.$$ Note that the requirement that $\Phi$ be an $H$ module map is exactly the equation . Furthermore, $$\Phi^{-1}(m\ot v)=S(m{~\hspace{-3pt}_{\left<1\right>}})v\ot m{~\hspace{-3pt}_{\left<0\right>}}.$$ We have arrived at the following theorem:
\[ydcenter\] For a Hopf algebra $H$ with an invertible antipode and $i\in\mathbb{Z}$,
$${\mathcal{Z}}_{{{_H\mathcal{M}}}}{{^{\#^i}_H\mathcal{M}}}\simeq {_H\mathcal{YD}}^H_{-i}.$$
If we consider the action of $\mathbb{Z}$ on ${{_H\mathcal{M}}}$ via $\#$, then ${{_H\mathcal{M}}}\rtimes\mathbb{Z}$ is a $\mathbb{Z}$ graded monoidal category. If we write ${{_H\mathcal{M}}}\rtimes\mathbb{Z}=\bigoplus_i{\mathcal{M}}_i$ then ${\mathcal{M}}_i={{_H\mathcal{M}}}^{\#^i}$ as an ${{_H\mathcal{M}}}$ bimodule category and $\bigoplus_i{_H\mathcal{YD}}^H_{i}$ is a $\mathbb{Z}$-equivariant $\mathbb{Z}$-braided monoidal category.
Stability. {#stabilitysection}
----------
Recall that a left $H$ module and right $H$ comodule $M$ is called stable if $m{~\hspace{-3pt}_{\left<1\right>}}m{~\hspace{-3pt}_{\left<0\right>}}=m$. We will need a slightly more general notion for the covariant theory. The classical stability will be precisely correct for the contravariant version.
Let $i\in\mathbb{Z}$. A left $H$ module and right $H$ comodule $M$ is called $i$-stable if $S^{2i}(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}=m$.
Thus the usual stability is now $0$-stability. The following lemma shows what happens to the odd powers of the antipode.
\[stability\] For a left $H$ module and right $H$ comodule $M$ we have $$S^{2i}(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}=m\quad\Leftrightarrow\quad S^{2i-1}(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}=m.$$
This is a direct computation, however in the instances where we see it, a more conceptual explanation can be found in terms of the $\tau_0$ map and its inverse that play a key role in our more conceptual understanding of stability.
“$\Rightarrow$" $$\begin{aligned}
S^{2i-1}(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}&=S^{2i-1}(m{~\hspace{-3pt}_{\left<2\right>}})S^{2i}(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}\\
&=S^{2i-1}(S(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<2\right>}})m{~\hspace{-3pt}_{\left<0\right>}}\\
&=\varepsilon(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}\\
&=m.\end{aligned}$$
“$\Leftarrow$" $$\begin{aligned}
S^{2i}(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}&=S^{2i}(m{~\hspace{-3pt}_{\left<2\right>}})S^{2i-1}(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}\\
&=S^{2i-1}(m{~\hspace{-3pt}_{\left<1\right>}}S(m{~\hspace{-3pt}_{\left<2\right>}}))m{~\hspace{-3pt}_{\left<0\right>}}\\
&=\varepsilon(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}\\
&=m.\end{aligned}$$
Monoidal categories and 2-traces {#moncat2tr}
================================
This section develops the core of the conceptual machinery that we need in order to understand the Hopf-type cyclic homology theories. For convenience we start with the covariant case and derive the contravariant case from it. We note that ignoring the non-strictness of the monoidal category by suppressing the explicit formulas for associators would have cleaned up the exposition. Our choice to include them was motivated by future applications of this machinery to monoidal categories where the associator appears as an explicit formula and so would have to appear in the definitions of the cyclic structure once it is unpacked from the conceptual definitions. The ease with which such laborious formulas are safely hidden from view demonstrates the power of the categorical machinery.
Let $(\mathcal{C}, \ot)$ be a monoidal category. We will need the following conventions. Let $A$ be an object in ${\mathcal{C}}$, by $A^{\ot n}$ we mean an object defined inductively as $$A^{\ot n}=A^{\ot n-1}\ot A.$$ For $\vec{n}=(n_1,\cdots, n_k)$ with $n_i$ non-negative integers, by $A^{\ot \vec{n}}$ we denote an object defined inductively as $$A^{\ot \vec{n}}=A^{\ot (n_1,\cdots, n_{k-1})}\ot A^{\ot n_k}.$$ We interpret $A^{\ot 0}$ as the unit object ${1}$. We apply a similar convention to morphisms. Let $|\vec{n}|=n_1+\cdots+n_k.$ Then for $\vec{n}$ and $\vec{m}$ with $|\vec{n}|=|\vec{m}|$ denote by $${\mathfrak{a}}_{\vec{n}}^{\vec{m}}: A^{\ot\vec{n}}\rightarrow A^{\ot\vec{m}}$$ the unique isomorphism ensured by the monoidal structure. Omitting the brackets enclosing the vector components to reduce clutter, we thus have ${\mathfrak{a}}_{n+1}^{n,1}=Id$, while ${\mathfrak{a}}_{n,1}^{1,n}$ is in general highly non-trivial and will play a central role below. Though ${\mathfrak{a}}_{n,1}^{1,n}$ is invisible for Hopf algebras, it will be needed for quasi-Hopf algebras and similar objects which lack “on the nose" coassociativity.
Symmetric 2-traces
------------------
Let ${\text{Vec}}$ be the category of vector spaces, and $\mathcal{M}$ be a ${\mathcal{C}}$-bimodule category. Then the functor category $Fun(\mathcal{M}, {\text{Vec}})$ is a $\mathcal{C}$-bimodule category with the left and right actions defined by
$$c\cdot F(-):= F(-\cdot c), \quad \text{and} \quad F\cdot c(-):= F(c\cdot -)$$
for all $c\in {\mathcal{C}}$. The center of a ${\mathcal{C}}$-bimodule category $\mathcal{M}$ is denoted by ${\mathcal{Z}}_{{\mathcal{C}}}\mathcal{M}$. Since ${\mathcal{C}}$ is a ${\mathcal{C}}$-bimodule category using its tensor product, we can set $\mathcal{M}={\mathcal{C}}$. To simplify the notation the center of a monoidal category ${\mathcal{C}}$ will be denoted by ${\mathcal{Z}}({\mathcal{C}})$.
Let $(\mathcal{C}, \ot)$ be a monoidal category.
- A functor $$F\in \mathcal{Z}_{\mathcal{{\mathcal{C}}}}Fun(\mathcal{C}, {\text{Vec}}),$$ is called a 2-trace. In particular we have natural isomorphisms $$\iota_c(-): F(-\ot c) \rightarrow F(c\ot -).$$
- A 2-trace $F$ is called a symmetric 2-trace (compare with the shadow structure in [@PS]) if $$\iota_{c}(1)=F({\mathfrak{a}}_{0,1}^{1,0}).$$
Note that the symmetry condition is indeed worthy of its name as it ensures that $$\iota_c(c')\iota_{c'}(c)=Id_{F(c\ot c')}.$$
[ Let $A$ be an associative algebra and let ${\mathcal{C}}=\text{Bimod}(A)$ denote the tensor category of $A$-bimodules. Then an example of a symmetric $2$-trace is provided by the functor $HH_0(A,-),$ the $0$th Hochschild homology of an $A$-bimodule [@FSS]. We note that for our purposes as outlined below, this example is not very interesting, its only advantage is that it is easy to explain. ]{}
Recall that we denote the subcategory of unital associative algebras in ${\mathcal{C}}$ by $Alg({\mathcal{C}})$. We denote the multiplication morphism of an algebra object $A\in {\mathcal{C}}$ by $m: A\ot A\rightarrow A$ and its unit morphism by $u: 1\rightarrow A$. Given an algebra $A\in Alg({\mathcal{C}})$ and a symmetric 2-trace $F: {\mathcal{C}}\to {\text{Vec}}$, we define a cyclic object in ${\text{Vec}}$ as follows.
Let $$C_n(A)=F(A^{\ot n+1}), n \geq 0.$$ We define the cyclic structure on $C_n (A)$ by
- $\tau_n = F({\mathfrak{a}}_{1,n}^{n,1})\circ\iota_A(A^{\ot n})$,
- $\delta_i = F({\mathfrak{a}}^{n-1,1}_{i,1,n-i-1})\circ F(Id^{\ot i}\ot m\ot \Id^{\ot n-i-1})\circ F({\mathfrak{a}}_{n,1}^{i,2,n-i-1})$, for $0 \leq i \leq n-1$,
- $\delta_{n} = \delta_0\circ\tau_n$,
- $\sigma_i = F({\mathfrak{a}}^{n+1,1}_{i+1,1,n-i})\circ F(\Id^{\ot i+1}\ot u\ot \Id^{\ot n-i})\circ F({\mathfrak{a}}_{n,1}^{i+1,0,n-i})$, for $0 \leq i \leq n$.
Note that for $0 \leq i \leq n-1$ we have $$\delta_i=F(\delta_i^{(n)})$$ and for $0 \leq i \leq n-2$ we have $$\delta_i=F(\delta_i^{(n-1)}\ot Id).$$ Similarly, for $0 \leq i \leq n$ we have $$\sigma_i=F(\sigma_i^{(n)})=F({\mathfrak{a}}_{1,n+1}^{n+1,1})F(Id\ot \sigma_{i-1}^{(n-1)})F({\mathfrak{a}}_{n,1}^{1,n}),$$ and observe that $\sigma_{-1}$ makes sense and is useful. These observations become relevant in the following Proposition.
\[trace1\] For any $A\in Alg({\mathcal{C}})$ and any symmetric 2-trace $F: {\mathcal{C}}\to {\text{Vec}}$ we have a cyclic object $C_{\bullet}(A)=F(A^{\ot \bullet +1})$ in ${\text{Vec}}$.
To see the simplicial relations we apply the functor $F$ to the simplicial relations that are classically satisfied by $\delta_i^{(n)}$’s and $\sigma_i^{(n)}$’s, with the exception of those involving the special $\delta_n$. One can check that the latter all follow formally from the former simplicity relations and the cyclicity relations below.
Here we check the cyclicity relations. First for any $1\leq i \leq n$, we show that $\delta_{i}\tau_n= \tau_{n-1} \delta_{i-1}$. We begin with the case $1\leq i \leq n-1$: $$\begin{aligned}
\tau_{n-1} \delta_{i-1} &= F({\mathfrak{a}}_{1,n-1}^{n-1,1})\iota_A(A^{\ot n-1})F(\delta_{i-1}^{(n-1)}\ot Id)\\
&= F({\mathfrak{a}}_{1,n-1}^{n-1,1}) F(Id\ot \delta_{i-1}^{(n-1)})\iota_A(A^{\ot n})\\
&= F(\delta_i^{(n)})F({\mathfrak{a}}_{1,n}^{n,1})\iota_A(A^{\ot n})\\
&= \delta_i \tau_n.\end{aligned}$$ For $i=n$ we observe that $\delta_n\tau_n=\tau_{n-1}\delta_{n-1}$ iff $\delta_0\tau^2_n=\tau_{n-1}\delta_{n-1}$, since $\delta_n=\delta_0\tau_n$ by definition, and so: $$\begin{aligned}
\tau_{n-1} \delta_{n-1} &= F({\mathfrak{a}}_{1,n-1}^{n-1,1})\iota_A(A^{\ot n-1})F(Id^{\ot n-1}\ot m)F({\mathfrak{a}}_{n,1}^{n-1,2})\\
&= F({\mathfrak{a}}_{1,n-1}^{n-1,1}) F(m\ot Id^{\ot n-1})\iota_{A^{\ot 2}}(A^{\ot n-1})F({\mathfrak{a}}_{n,1}^{n-1,2})\\
&= F({\mathfrak{a}}_{1,n-1}^{n-1,1}) F(m\ot Id^{\ot n-1})F({\mathfrak{a}}_{n,1}^{2,n-1})F({\mathfrak{a}}^{n,1}_{2,n-1})\iota_{A^{\ot 2}}(A^{\ot n-1})F({\mathfrak{a}}_{n,1}^{n-1,2})\\
&= \delta_0 \tau_n^2.\end{aligned}$$
Here we show that $\sigma_i \tau_n=\tau_{n+1} \sigma_{i-1}$ for all $0\leq i\leq n$: $$\begin{aligned}
\sigma_i \tau_n &= F({\mathfrak{a}}_{1,n+1}^{n+1,1})F(Id\ot \sigma_{i-1}^{(n-1)})F({\mathfrak{a}}_{n,1}^{1,n})F({\mathfrak{a}}^{n,1}_{1,n})\iota_A(A^{\ot n})\\
&= F({\mathfrak{a}}_{1,n+1}^{n+1,1})F(Id\ot \sigma_{i-1}^{(n-1)})\iota_A(A^{\ot n})\\
&= F({\mathfrak{a}}_{1,n+1}^{n+1,1})\iota_A(A^{\ot n+1})F(\sigma_{i-1}^{(n-1)}\ot Id)\\
&= \tau_{n+1} \sigma_{i-1}.
\end{aligned}$$
Finally we demonstrate that $\tau_n^{n+1}=Id$: $$\begin{aligned}
\tau^{n+1}_n &= F({\mathfrak{a}}_{n+1,0}^{n,1})\iota_{A^{\ot n+1}}(1)F({\mathfrak{a}}_{n,1}^{0,n+1})\\
&= F({\mathfrak{a}}_{n+1,0}^{n,1})F({\mathfrak{a}}_{0,n+1}^{n+1,0})F({\mathfrak{a}}_{n,1}^{0,n+1})\\
&= F({\mathfrak{a}}_{n,1}^{n,1})\\
&= Id.
\end{aligned}$$
Note that there was nothing special about ${\text{Vec}}$ in the above considerations, namely the results would still hold if ${\text{Vec}}$ was replaced by any target category ${\mathcal{T}}$, namely a symmetric ${\mathcal{T}}$-valued $2$-trace would still produce cyclic objects in ${\mathcal{T}}$ from elements of $Alg({\mathcal{C}})$.
Let ${\mathcal{C}}_{op}$ denote the opposite monoidal category *with only the arrows reversed*. Thus the associator is replaced by its inverse. Let ${\mathcal{M}}$ be a ${\mathcal{C}}$ bimodule category, then ${\mathcal{M}}_{op}$ is a ${\mathcal{C}}_{op}$ bimodule category via $$c'\cdot m'\cdot d'=(c\cdot m\cdot d)',$$ where we use $m'$ to denote the element $m\in{\mathcal{M}}$ when we consider it as an element of ${\mathcal{M}}_{op}$. Recall that for $m\in{\mathcal{Z}}_{\mathcal{C}}({\mathcal{M}})$ we have isomorphisms $\iota^m_c:c\cdot m\xrightarrow\sim m\cdot c$. We note that $$\begin{aligned}
{\mathcal{Z}}_{\mathcal{C}}{\mathcal{M}}&\simeq{\mathcal{Z}}_{{\mathcal{C}}_{op}}{\mathcal{M}}_{op}\\m&\mapsto m'\\\iota^m_c&\mapsto \iota'^{m'}_{c'}=((\iota^m_c)^{-1})'.\end{aligned}$$
Consider ${\mathcal{T}}={\text{Vec}}_{op}$ and replace ${\mathcal{C}}$ by ${\mathcal{C}}_{op}$. More precisely, let $Coalg({\mathcal{C}})$ denote the subcategory of coassociative counital coalgebra objects of $({\mathcal{C}}, \ot)$. Then $$\begin{aligned}
Fun({\mathcal{C}},{\text{Vec}})&=Fun({\mathcal{C}}_{op},{\text{Vec}}_{op})_{op},\\
\mathcal{Z}_{{\mathcal{C}}}Fun({\mathcal{C}},{\text{Vec}})&\simeq\mathcal{Z}_{{\mathcal{C}}_{op}}Fun({\mathcal{C}},{\text{Vec}})_{op}\\&=\mathcal{Z}_{{\mathcal{C}}_{op}}Fun({\mathcal{C}}_{op},{\text{Vec}}_{op}),\\
Coalg({\mathcal{C}})&=Alg({\mathcal{C}}_{op})_{op}.\end{aligned}$$ Furthermore, a cyclic object in ${\text{Vec}}_{op}$ is the same as a cocyclic object in ${\text{Vec}}$ and we have arrived at the following:
\[trace2\] If $C\in Coalg({\mathcal{C}})$ and $F$ a symmetric 2-trace then $C^\bullet= F(C^{\ot \bullet+1})$ is a cocyclic object in ${\text{Vec}}$.
Recall that for an algebra $A$, we had $$\tau_n = F({\mathfrak{a}}_{1,n}^{n,1})\circ\iota_A(A^{\ot n}).$$ However, after unraveling the above identifications, we have for a coalgebra $C$: $$\tau_n = \iota^{-1}_C(C^{\ot n})\circ F({\mathfrak{a}}_{n,1}^{1,n}).$$
### The contravariant functor $F$.
While the covariant theory discussed above is suitable for explaining the cocyclic structure for the case $C^n=\Hom_H(k, M\ot C^{\ot n+1})$. If we want to deal with the case of $C^n=\Hom_H(M\ot A^{\ot n+1}, k)$ and obtain a cocyclic structure on it, then we need a contravariant $F$. This is not hard to do in light of the above.
We say that a contravariant functor $F$ from ${\mathcal{C}}$ to ${\text{Vec}}$ is a symmetric $2$-contratrace if $F$ is a symmetric $2$-trace on ${\mathcal{C}}_{op}$.
By recalling that $Coalg({\mathcal{C}})=Alg({\mathcal{C}}_{op})_{op}$ we immediately obtain the following:
\[trace3\] If $C\in Coalg({\mathcal{C}})$ and $F$ a symmetric 2-contratrace then $C^\bullet= F(C^{\ot \bullet+1})$ is a cyclic object in ${\text{Vec}}$.
While $Alg({\mathcal{C}})=Coalg({\mathcal{C}}_{op})_{op}$ implies that:
\[trace4\] If $A\in Alg({\mathcal{C}})$ and $F$ a symmetric 2-contratrace then $C^\bullet= F(A^{\ot \bullet+1})$ is a cocyclic object in ${\text{Vec}}$.
Now for an algebra $A$, we have $$\tau_n = \iota^{-1}_A(A^{\ot n})\circ F({\mathfrak{a}}^{n,1}_{1,n}).$$ Furthermore, for a coalgebra $C$ we get: $$\tau_n = F({\mathfrak{a}}^{1,n}_{n,1})\circ\iota_C(C^{\ot n}).$$
Stable central pairs
--------------------
The concept of a stable central pair introduced in the following definition arises naturally in settings generalizing the Hopf-cyclic theory. The Hopf-cyclic theory itself is implicitly based on it. As the Lemma below demonstrates the reason for its usefulness is that it is a natural way of constructing symmetric $2$-traces, which lead, as we saw above, to cyclic objects.
\[stablepair\] Let $(\mathcal{C}, \ot)$ be a monoidal category, and $\mathcal{M}$ a ${\mathcal{C}}$-bimodule category. Let $F\in Fun (\mathcal{M}, {\text{Vec}})$ and $m\in \mathcal{M}$. The pair $(F, m)$ is called a central pair if
- $F\in \mathcal{Z}_{{\mathcal{C}}} Fun(\mathcal{M}, {\text{Vec}})$, in particular $\iota^F_c(-): F(-\cdot c)\simeq F(c\cdot -)$.
- $m\in \mathcal{Z}_{{\mathcal{C}}}(\mathcal{M})$, in particular $\iota^m_c: c\cdot m \simeq m\cdot c$.
The central pair $(F, m)$ is called a stable central pair if
- $F(\iota^m_c)\iota^F_c(m)=Id_{F(m\cdot c)}$.
\[pairtotrace\] If $(F, m)$ is a (stable) central pair then $F(m \cdot -)$ is a (symmetric) 2-trace.
Define the structure of a 2-trace on $F(m \cdot -)$, i.e., an isomorphism $$\iota_c: F(m\cdot (-\ot c))\simeq F(m\cdot(c\ot -))$$ via the chain of isomorphisms: $$\begin{aligned}
F(m\cdot (-\ot c))&\rightarrow F((m\cdot -)\cdot c)\\
&\rightarrow F(c\cdot(m\cdot -))\\
&\rightarrow F((c\cdot m)\cdot -)\\
&\rightarrow F((m\cdot c)\cdot -)\\
&\rightarrow F(m\cdot(c\ot -))\end{aligned}$$
For the symmetry condition consult the following commutative diagram with all arrows being the obvious isomorphisms:
$$\xymatrix{
F(m\cdot c)\ar[r]\ar[dr]\ar[dd] & F(m\cdot (1\ot c))\ar[d]\\
& F((m\cdot 1)\cdot c)\ar[d]\\
F(c\cdot m)\ar[r]\ar[dr]\ar[dd] & F(c\cdot(m\cdot 1))\ar[d]\\
& F((c\cdot m)\cdot 1)\ar[d]\\
F(m\cdot c)\ar[r]\ar[dr] & F((m\cdot c)\cdot 1)\ar[d]\\
& F(m\cdot (c\otimes 1))\\
}$$ then the first column composes to $Id$ by stability, and the second column composes to $\iota_c(1)$ by definition, the claim follows.
This shows that a stable central pair gives us a symmetric 2-trace and therefore by Propositions \[trace1\] and \[trace2\] produces cyclic and cocyclic objects from algebras and coalgebras. More precisely, $F(m\cdot A^{\ot \bullet+1})$ and $F(m\cdot C^{\ot \bullet+1})$ are cyclic and cocyclic objects, for $A$ an algebra and $C$ a coalgebra respectively. Let us write out the cyclic map for these cases. Roughly speaking, i.e., ignoring the associativity isomorphisms we have $$\tau: F(m\cdot A^{\ot n+1})\xrightarrow{\iota^F_A(m\cdot A^{\ot n})} F(A\cdot m\cdot A^{\ot n})\xrightarrow{F(\iota^m_A\cdot Id^{\ot n})} F(m\cdot A^{\ot n+1})$$ while $$\tau: F(m\cdot C^{\ot n+1})\xrightarrow{F((\iota^m_C)^{-1}\cdot Id^{\ot n})} F(C\cdot m\cdot C^{\ot n})\xrightarrow{(\iota^F_C)^{-1}(m\cdot C^{\ot n})} F(m\cdot C^{\ot n+1}),$$ so that in the algebra case $\tau$ moves “last to first" while in the coalgebra case it does the opposite.
### Stable central contrapairs.
Let us now mirror the above discussion for the contravariant case.
\[stablecontrapair\] Let $(\mathcal{C}, \ot)$ be a monoidal category, and ${\mathcal{M}}$ a ${\mathcal{C}}$-bimodule category. Let $F\in Fun ({\mathcal{M}}_{op}, {\text{Vec}})$ and $m\in {\mathcal{M}}$. The pair $(F, m)$ is called a stable central contrapair if $(F,m')$ is a stable central pair for ${\mathcal{C}}_{op}$.
We immediately obtain:
\[contrapairtocontratrace\] If $(F, m)$ is a stable central contrapair then $F(m \cdot -)$ is a symmetric $2$-contratrace.
Thus as above, $F(m\cdot A^{\ot \bullet+1})$ and $F(m\cdot C^{\ot \bullet+1})$ are cocyclic and cyclic objects, for $A$ an algebra and $C$ a coalgebra respectively. Ignoring the associativity isomorphisms we have $$\tau: F(m\cdot A\ot A^{\ot n})\rightarrow F(A\cdot m\cdot A^{\ot n})\rightarrow F(m\cdot A^{\ot n}\ot A)$$ while $$\tau: F(m\cdot C^{\ot n}\ot C)\rightarrow F(C\cdot m\cdot C^{\ot n})\rightarrow F(m\cdot C\ot C^{\ot n}),$$ so that in the coalgebra case $\tau$ moves “last to first" while in the algebra case it does the opposite.
The monoidal category of left modules over a Hopf algebra
=========================================================
In this section we apply our results from Section \[moncat2tr\] to the monoidal category of left modules over a Hopf algebra $H$. Our aim is to construct a symmetric 2-trace via a stable central pair. The idea was sketched in Section \[motivation\] for module coalgebras, here we briefly recap for module algebras. The only difference is that in the definition of $\tau$ the order of what gets used first: the centrality of the functor or the centrality of the element gets reversed.
Let us consider a simpler version of what we want, namely ${{_H\mathcal{M}}}_{fd}$ which is a rigid monoidal category. Using the rigid structure we have the following isomorphism
$$\label{adj1}
\Hom_H(1, -\ot V)\xrightarrow\sim \Hom_H(1, {^{**}V}\ot -),$$
and furthermore
$$\Hom_H(1, -)\in {\mathcal{Z}}_{{{_H\mathcal{M}}}_{fd}} Fun({{^{**}_H\mathcal{M}}}_{fd}, {\text{Vec}}).$$ If in addition $$M\in{\mathcal{Z}}_{{{_H\mathcal{M}}}_{fd}}({{^{**}_H\mathcal{M}}}_{fd}),$$ in particular we have $$\label{centre}{^{**}-}\ot M \xrightarrow\sim M\ot -,$$ then we can make a cyclic map $\tau$ as follows.
Consider an (algebra) object $A$ in ${{_H\mathcal{M}}}_{fd}$ then we obtain
$$\tau: \Hom_{H}(1, M\ot A^{\ot n+1})\xrightarrow\sim \Hom_{H}(1, {^{**}A}\ot M\ot A^{\ot n}) \xrightarrow\sim \Hom_{H}(1, M\ot A\ot A^{\ot n}),$$ where we first used and then thus *sliding the last copy of $A$ to the front and past $M$*. Of course we only need the algebra structure to define the simplicial structure, the map $\tau$ above does not need it. The resulting structure on $\Hom_{H}(1, M\ot A^{\ot n+1})$ is that of a cyclic module, provided that $$\tau_0:\Hom_{H}(1, M\ot A)\xrightarrow\sim \Hom_{H}(1, M\ot A)$$ is the identity map. If the latter condition is dropped then the result is a paracyclic module.
We note that the conditions on $M$ as outlined above are equivalent to the $1$-stable $YD_1$ condition, see below for details.
The covariant theory for ${{_H\mathcal{M}}}$ {#2traceforhmod}
--------------------------------------------
In this subsection we consider the monoidal category ${{_H\mathcal{M}}}$ and show that if $M$ is a left-right $\mathcal{YD}_1$ module, then the functor $\Hom_H(1,-)$ paired with $M$ forms a central pair $\left(\Hom_H(1, -), M\right)$ for a suitable bimodule category, namely ${\mathcal{M}}={{^{\#^{-1}}_H\mathcal{M}}}$. Furthermore if $M$ is $1$-stable, then $\left(\Hom_H(1, -), M\right)$ is a stable central pair.
Recall that to prove that $\left(\Hom_H(1, -), M\right)$ is a stable central pair for ${{_H\mathcal{M}}}$ and its bimodule category ${{^{\#^{-1}}_H\mathcal{M}}}$, we need to show that
- $\Hom_H(1, -)\in {\mathcal{Z}}_{{_H\mathcal{M}}}Fun({{^{\#^{-1}}_H\mathcal{M}}},{\text{Vec}})$.
- If $M\in\mathcal{YD}_1$ then $M\in {\mathcal{Z}}_{{_H\mathcal{M}}}({{^{\#^{-1}}_H\mathcal{M}}})$.
- $1$-stability of $M$ ensures that $\tau_0=\Id$ and therefore the stability of the central pair.
The second point is the content of Theorem \[ydcenter\]. Now to address the first point.
If $A, B\in {{_H\mathcal{M}}}$, then the left $H$-module map $f\in \Hom_H(1,A\ot B)$ is equivalent to the data of an $H$-invariant element $a\ot b\in A\ot B$. More precisely, $$h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}} b=\varepsilon(h)a\ot b,$$ for all $h\in H$. Note that we write $a\ot b$ when we actually mean a sum of such elements in $(A\ot B)^H$.
\[cov-H-mod\] Let $A, B\in {{_H\mathcal{M}}}$ and $a\ot b\in A\ot B$. Then $$h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}} b=\varepsilon(h)a\ot b \quad\Leftrightarrow\quad S(h)a\ot b= a\ot hb.$$
If $S(h)a\ot b=a\ot hb$ then $\varepsilon(h)a\ot b=h{~\hspace{-3pt}^{^{\left(1\right)}}}S(h{~\hspace{-3pt}^{^{\left(2\right)}}})a\ot b=h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}}b.
$ Conversely if $h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}}b=\varepsilon(h) a\ot b$ then $$\begin{aligned}
S(h)a\ot b&= S(h{~\hspace{-3pt}^{^{\left(1\right)}}}\varepsilon(h{~\hspace{-3pt}^{^{\left(2\right)}}}))a\ot b\\
&= S(h{~\hspace{-3pt}^{^{\left(1\right)}}})\varepsilon(h{~\hspace{-3pt}^{^{\left(2\right)}}})a\ot b\\
&= S(h{~\hspace{-3pt}^{^{\left(1\right)}}})h{~\hspace{-3pt}^{^{\left(2\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(3\right)}}}b\\
&= \varepsilon(h{~\hspace{-3pt}^{^{\left(1\right)}}})a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}}b\\
&= a\ot \varepsilon(h{~\hspace{-3pt}^{^{\left(1\right)}}})h{~\hspace{-3pt}^{^{\left(2\right)}}}b\\
&= a\ot hb.
\end{aligned}$$
For the sake of reducing notational clutter, let us, for an element $B\in{{_H\mathcal{M}}}$, denote by ${^\# B}$ what was until now called $B^{\#^{-1}}$.
\[homcentral\] $$\Hom_H(1, -)\in {\mathcal{Z}}_{{_H\mathcal{M}}}Fun({{^{\#^{-1}}_H\mathcal{M}}}, {\text{Vec}}).$$
Since $$S(h)a\ot b= a\ot hb\quad\forall h\in H\quad\Leftrightarrow\quad S^{-1}(h)b\ot a=b\ot ha\quad\forall h\in H,$$
so in view of the Lemma \[cov-H-mod\], we see that $$a\ot b\in(A\ot B)^H\quad\Leftrightarrow\quad b\ot a\in({^\#B}\ot A)^H,$$ i.e, $$\label{homcent} a\ot b\longmapsto b\ot a$$ obviously gives rise to the following natural isomorphisms
$$\Hom_H(1, -\ot B)\xrightarrow\sim \Hom_H(1,{^\#B}\ot -);$$
it is not hard to check the rest given that the map itself is very simple.
On to the third point: stability. We need to check that $\tau_0=Id$ if $M$ is $1$-stable. Note that strictly speaking $\tau_0$ depends on wether we want to use the theory on algebras or coalgebras, otherwise we might need its inverse. Yet $\tau_0=Id$ if and only if $\tau_0^{-1}=Id$; this is not surprising as the notion of a stable central pair doesn’t depend on what you intend to use it for.
If $M$ is $1$-stable then $\tau_0=Id$.
Using followed by we see that for $m\ot v\in(M\ot V)^H$ we have $\tau_0(m\ot v)=m{~\hspace{-3pt}_{\left<0\right>}}\ot m{~\hspace{-3pt}_{\left<1\right>}}v$ which is $S(m{~\hspace{-3pt}_{\left<1\right>}})m{~\hspace{-3pt}_{\left<0\right>}}\ot v$ by Lemma \[cov-H-mod\], and the latter is $m\ot v$ by the $1$-stability of $M$ in view of Lemma \[stability\].
Thus $(\Hom_H(1,-), M)$ is a stable central pair. As usual let $A$ be an algebra and $C$ a coalgebra. We write out the formulas for the cyclic and cocyclic module structures.
For $C_n=\Hom_H(1, M\ot A^{\ot n+1})$: $$\begin{aligned}
\delta_i(m\ot a_0\ot \cdots \ot a_n)&=m\ot a_0\ot \cdots \ot a_i a_{i+1}\ot \cdots \ot a_n,\\
\delta_n(m\ot a_0\ot \cdots \ot a_n)&=m{~\hspace{-3pt}_{\left<0\right>}}\ot (m{~\hspace{-3pt}_{\left<1\right>}}a_n) a_0\ot a_1\ot \cdots \ot a_{n-1}, \\
\sigma_i(m\ot a_0\ot \cdots \ot a_n)&=m\ot a_0\ot \cdots \ot a_i \ot 1\ot \cdots \ot a_n,\\
\tau_n(m\ot a_0\ot \cdots \ot a_n)&=m{~\hspace{-3pt}_{\left<0\right>}}\ot m{~\hspace{-3pt}_{\left<1\right>}}a_n\ot a_0\ot \cdots \ot a_{n-1}.\end{aligned}$$
For $C^n=\Hom_H(1, M\ot C^{\ot n+1})$: $$\begin{aligned}
\delta_i(m\ot c_0\ot \cdots \ot c_{n-1})&=m\ot c_0\ot \cdots \ot c_i{~\hspace{-3pt}^{^{\left(1\right)}}}\ot c_i{~\hspace{-3pt}^{^{\left(2\right)}}}\ot \cdots \ot c_{n-1},\\
\delta_n(m\ot c_0\ot \cdots \ot c_{n-1})&=m{~\hspace{-3pt}_{\left<0\right>}}\ot c_0{~\hspace{-3pt}^{^{\left(2\right)}}}\ot c_1\ot \cdots \ot c_{n-1}\ot S(m{~\hspace{-3pt}_{\left<1\right>}})c_0{~\hspace{-3pt}^{^{\left(1\right)}}}, \\
\sigma_i(m\ot c_0\ot \cdots \ot c_{n+1})&=m\ot c_0\ot \cdots \ot \varepsilon(c_{i+1})\ot \cdots \ot c_{n+1},\\
\tau_n(m\ot c_0\ot \cdots \ot c_n)&=m{~\hspace{-3pt}_{\left<0\right>}}\ot c_1\ot \cdots \ot c_{n}\ot S(m{~\hspace{-3pt}_{\left<1\right>}})c_0.\end{aligned}$$
The contravariant theory for ${{_H\mathcal{M}}}$
------------------------------------------------
Here we redo the previous section for the case of $\Hom_H(-,1)$. More precisely, we show that if $M$ is a left-right $\mathcal{YD}_{-1}$ module ($\mathcal{AYD}$ module), then the functor $\Hom_H(-,1)$ paired with $M$ forms a central contrapair $\left(\Hom_H(-, 1), M\right)$ for ${\mathcal{M}}={{^{\#}_H\mathcal{M}}}$. Furthermore if $M$ is $0$-stable (classically stable), then $\left(\Hom_H(-, 1), M\right)$ is a stable central contrapair.
We will need a characterization of $H$ module maps from $A\ot B$ to the monoidal unit $k$.
\[H-maps\] Let $H$ be a Hopf algebra over a field $k$, $A, B\in {{_H\mathcal{M}}}$ and $f: A\ot B\rightarrow k$ a $k$-linear map. Then $f(h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}}b)=\varepsilon(h) f(a\ot b)$ if and only if $f(ha\ot b)=f(a\ot S(h)b)$.
The computation is similar to that of Lemma \[cov-H-mod\]. If $f(ha\ot b)=f(a\ot S(h)b)$ then $f(h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}}b)=f(a\ot S(h{~\hspace{-3pt}^{^{\left(1\right)}}})h{~\hspace{-3pt}^{^{\left(2\right)}}}b)=\varepsilon(h) f(a\ot b).
$ Conversely if $f(h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}}b)=\varepsilon(h) f(a\ot b)$ then $$\begin{aligned}
f(a\ot S(h)b)&= f(a\ot S(\varepsilon(h{~\hspace{-3pt}^{^{\left(1\right)}}})h{~\hspace{-3pt}^{^{\left(2\right)}}})b)\\
&= \varepsilon(h{~\hspace{-3pt}^{^{\left(1\right)}}})f(a\ot S(h{~\hspace{-3pt}^{^{\left(2\right)}}})b)\\
&= f(h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot h{~\hspace{-3pt}^{^{\left(2\right)}}} S(h{~\hspace{-3pt}^{^{\left(3\right)}}})b)\\
&= f(h{~\hspace{-3pt}^{^{\left(1\right)}}}a\ot \varepsilon(h{~\hspace{-3pt}^{^{\left(2\right)}}})b)\\
&= f(ha\ot b).
\end{aligned}$$
\[homcentral1\] $$\Hom_H(-, 1)\in {\mathcal{Z}}_{{{_H\mathcal{M}}}_{op}} Fun({{^{\#}_H\mathcal{M}}}_{op}, {\text{Vec}}).$$
This boils down to observing that for $f\in\Hom_H(A^\#\ot B,1)$ we have $\gamma f\in\Hom_H(B\ot A,1)$ where $\gamma f(b\ot a)=f(a\ot b)$. This is an immediate consequence of Lemma \[H-maps\].
We recall that by Theorem \[ydcenter\], if $M\in\mathcal{AYD}$, i.e., $M\in\mathcal{YD}_{-1}$ then $M\in{\mathcal{Z}}_{{{_H\mathcal{M}}}}({{^{\#}_H\mathcal{M}}})$ and the latter is equivalent to ${\mathcal{Z}}_{{{_H\mathcal{M}}}_{op}}({{^{\#}_H\mathcal{M}}}_{op})$. So all that remains is to investigate the stability condition needed for $\tau_0=Id$.
If $M$ is $0$-stable (classically stable) then $\tau_0=Id$.
Let $f\in\Hom_H(M\ot V,1)$, then $\tau_0 f(m\ot v)=f(m{~\hspace{-3pt}_{\left<0\right>}}\ot S(m{~\hspace{-3pt}_{\left<1\right>}})v)=f(m{~\hspace{-3pt}_{\left<1\right>}}m{~\hspace{-3pt}_{\left<0\right>}}\ot v)=f(m\ot v)$.
Thus $(\Hom_H(-,1), M)$ is a stable central contrapair. As usual let $A$ be an algebra and $C$ a coalgebra. We write out the formulas for the cocyclic and cyclic module structures.
For $C^n=\Hom_H(M\ot A^{\ot n+1},1)$: $$\begin{aligned}
\delta_if(m\ot a_0\ot \cdots \ot a_n)&=f(m\ot a_0\ot \cdots \ot a_i a_{i+1}\ot \cdots \ot a_n),\\
\delta_nf(m\ot a_0\ot \cdots \ot a_n)&=f(m{~\hspace{-3pt}_{\left<0\right>}}\ot (m{~\hspace{-3pt}_{\left<1\right>}}a_n) a_0\ot a_1\ot \cdots \ot a_{n-1}), \\
\sigma_if(m\ot a_0\ot \cdots \ot a_n)&=f(m\ot a_0\ot \cdots \ot a_i \ot 1\ot \cdots \ot a_n),\\
\tau_nf(m\ot a_0\ot \cdots \ot a_n)&=f(m{~\hspace{-3pt}_{\left<0\right>}}\ot m{~\hspace{-3pt}_{\left<1\right>}}a_n\ot a_0\ot \cdots \ot a_{n-1}).\end{aligned}$$
For $C_n=\Hom_H(M\ot C^{\ot n+1},1)$: $$\begin{aligned}
\delta_if(m\ot c_0\ot \cdots \ot c_{n-1})&=f(m\ot c_0\ot \cdots \ot c_i{~\hspace{-3pt}^{^{\left(1\right)}}}\ot c_i{~\hspace{-3pt}^{^{\left(2\right)}}}\ot \cdots \ot c_{n-1}),\\
\delta_nf(m\ot c_0\ot \cdots \ot c_{n-1})&=f(m{~\hspace{-3pt}_{\left<0\right>}}\ot c_0{~\hspace{-3pt}^{^{\left(2\right)}}}\ot c_1\ot \cdots \ot c_{n-1}\ot S(m{~\hspace{-3pt}_{\left<1\right>}})c_0{~\hspace{-3pt}^{^{\left(1\right)}}}), \\
\sigma_if(m\ot c_0\ot \cdots \ot c_{n+1})&=f(m\ot c_0\ot \cdots \ot \varepsilon(c_{i+1})\ot \cdots \ot c_{n+1}),\\
\tau_nf(m\ot c_0\ot \cdots \ot c_n)&=f(m{~\hspace{-3pt}_{\left<0\right>}}\ot c_1\ot \cdots \ot c_{n}\ot S(m{~\hspace{-3pt}_{\left<1\right>}})c_0).\end{aligned}$$
In this paper we have investigated the four (co)homology theories that arise naturally in the consideration of the monoidal category of left $H$-modules. These come from the considerations of the covariant and the contravariant theories in the sense of their behavior with respect to maps of (co)algebras. If we consider the contravariant theory of the algebra case, we recover the type $A$ cohomology theory of [@hkrs2] on the nose. By considering the covariant theory of the coalgebra case we obtain a different cohomology theory than that of [@hkrs2]; this explains the need for *new* coefficients of opposite “charge" than AYD. The type $C$ theory which generalizes Connes-Moscovici Hopf cyclic cohomology [@CM98] is actually obtained from the $2$-contratrace giving the type $A$ theory by considering its predual which is a $2$-trace. The other two possibilities considered are both homology theories, one requiring AYDs and the other “anti" AYDs.
Our explicit calculations do not extend to the type $B$ theory of [@hkrs2] which is a contravariant cohomology theory for $H$ *comodule* algebras. However we point out that our machinery can be applied to the monoidal category ${^H{\mathcal{M}}}$ of left comodules over $H$. In that case the type $B$ theory is a straightforward consequence, though with modifications.
Let us summarize. If we are given a rigid monoidal category ${\mathcal{C}}$ then there is a covariant cyclic theory with coefficients in ${\mathcal{Z}}_{\mathcal{C}}{\mathcal{C}}^{**}$ that turns algebras into cyclic modules and coalgebras into cocyclic modules. There is also a contravariant cyclic theory with coefficients in ${\mathcal{Z}}_{\mathcal{C}}{^{**}{\mathcal{C}}}$ that turns algebras into cocyclic modules and coalgebras into cyclic modules. In the above $^{**}$ is the functor that sends $c\in{\mathcal{C}}$ to $c^{**}$. Note that the coefficients need to be more than just central in a correct bimodule category, they have to be stable as well. If the category is not rigid *anymore*, such as was the case of the general ${{_H\mathcal{M}}}$ with infinite dimensional representations allowed, we can still proceed: we would need a replacement for $(-)^{**}$, such as $\#$ was in the case of ${{_H\mathcal{M}}}$.
[9]{}
G. Bohm, D. Stefan, *A categorical approach to cyclic duality* J. Noncommutative Geometry 6 (2012), no. 3, 481-538.
A. Connes, *Cohomologie cyclique et foncteurs $Ext^n$*, (French) (Cyclic cohomology and functors $Ext^n$) C. R. Acad. Sci. Paris Sr.I Math.296(1983), no. 23, 953-958.
A. Connes, H. Moscovici, *Hopf algebras, cyclic cohomology and the transverse index theorem*, Commun. Math. Phys. 198 (1998), 199–246.
A. Connes, H. Moscovici, *Cyclic cohomology and Hopf algebras*, Lett. Math. Phys. 48 (1999), 97–108.
A. Dold, D. Puppe, *Duality, trace, and transfer*, Proceedings of the Steklov Institute of Mathematics, 154(1984), 85–103.
J. Fuchs, G. Schaumann, and C. Schweigert, *A trace for bimodule categories*, http://arxiv.org/abs/14126968.
J. Greenough, *Monoidal 2-structure of bimodule categories*, Volume 324, Issue 8, (2010), 1818-1859.
J. L. Loday, *Cyclic homology*, Springer Verlag, (1992).
P. M. Hajac, M. Khalkhali, B. Rangipour, and Y. Sommerh�auser, *Hopf-cyclic homology and cohomology with coefficients*, C. R. Math. Acad. Sci.Paris 338 (2004), no. 9, 667-672. P. M. Hajac, M. Khalkhali, B. Rangipour, and Y. Sommerh�auser, *Stable anti-Yetter-Drinfeld modules*, C. R. Acad. Sci. Paris, Ser. I, 338 (2004) 587-590.
M. Khalkhali, A. Pourkia, *Hopf cyclic cohomology in braided monoidal categories*, Homology, Homotopy, Appl, Volume 12, Number 1 (2010), 111-155.
A. Joyal, R. Street, and D. Verity, *Traced monoidal categories*, Math. Proc. Cambridge Philos. Soc, 119(3)(1996), 447-468.
C. Kassel,*Quantum Groups*, Graduate Texts Math.155, Springer Verlag, Berlin, (1995).
S. Majid, *Foundations of Quantum Group Theory*, Cambridge University Press, Cambridge, (1995).
K. Ponto, *Relative fixed point theory*, Algebr. Geom. Top., 11(2), (2011),839-886.
K. Ponto, M. Shulman, *Shadows and traces in bicategories*, Journal of Homotopy and Related Structures 8 (2013), 151-200.
P. Schauenburg, *Hopf Modules and Yetter-Drinfel’d Modules*, Journal of Algebra 169(3), (1994), 874-890.
Department of Mathematics and Statistics, University of Windsor, 401 Sunset Avenue, Windsor, Ontario N9B 3P4, Canada
*E-mail address*: **[email protected]**
Department of Mathematics, Middlesex College, The University of Western Ontario,
London, Ontario, N6A 5B7, Canada.
*E-mail address*: **[email protected]**
Department of Mathematics and Statistics, University of Windsor, 401 Sunset Avenue, Windsor, Ontario N9B 3P4, Canada
*E-mail address*: **[email protected]**
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a combinatorial method of proving non-specialty of a linear system of curves with multiple points in general positions. As an application we obtain a classification of special linear systems on ${\mathbb{P}}^1\times{\mathbb{P}}^1$ for which the multiplicities do not exceed $3$.'
author:
- Tomasz Lenarcik
title: 'Linear systems over ${\mathbb{P}}^1\times{\mathbb{P}}^1$ with base points of multiplicity bounded by three'
---
Introduction {#introduction .unnumbered}
============
Let $p_1,\ldots,p_r$ denote points in ${\mathbb{P}}^1\times{\mathbb{P}}^1$ in general position and let $m_1,\ldots,m_r$ be positive integers. Consider a blowing up $\pi:X\longrightarrow{\mathbb{P}}^1\times{\mathbb{P}}^1$ at $p_1,\ldots,p_r$, and denote the exeptional divisors respectively by $E_1,\ldots,E_r$. For given $d,e\geq 0$ we define ${\mathcal{L}}_{(d,e)}(p_1 m_1,\ldots,p_r m_r)$ to be a complete linear system of the following divisor: $$dH_1+eH_2-m_1E_1-\ldots-m_rE_r$$ where $H_1$ and $H_2$ are pullbacks of classes of ${\mathbb{P}}^1\times\{a_1\}$ and $\{a_2\}\times{\mathbb{P}}^1$ respectively, and $a_1,a_2\in{\mathbb{P}}^1$ are arbitrary. It can be understood as a linear space of curves of bidegree $(d,e)$ that vanish at $p_i$ with a multiplicity of at least $m_i$ for $i=1,\ldots,r$. For a sufficiently general choice of affine coordinates, each curve from ${\mathcal{L}}_{(d,e)}(p_1 m_1,\ldots,p_r m_r)$ can be uniquely represented (up to a constant factor) by a polynomial in two variables, namely $X$ and $Y$, that contains monomials of the form $X^{\alpha}Y^{\beta}$, where $0\leq\alpha\leq d$ and $0\leq\beta\leq e$. Therefore, from linear algebra it follows that the projective dimension of ${\mathcal{L}}_{(d,e)}(p_1 m_1,\ldots,p_r m_r)$ is not less than: $$\label{eq_edim}
\max\{-1,(d+1)(e+1)-\sum_{i}\binom{m_i+1}{2}-1\}.$$ The actual dimension, however, does not have to equal the *expected dimension* , as the equations may happen to be linearly dependent even for a general choice of $p_1,\ldots,p_r$. In such an instance, we say that the linear system is *special*. A similar definition can be formulated for special linear systems over ${\mathbb{P}}^2$ (see for example [@Dum]). Further information about linear systems shall be presented in Section 1.
In Section 2 we propose a combinatorial technique of proving non-specialty of a linear system. Several approaches of this kind have been recently developed including degenerations techniques by Ciliberto, Dumitrescu and Miranda [@CDM], application of tropic geomentry by Baur and Draisma [@Dra; @Bau-Dra], and reduction method by Dumnicki and Jarnicki [@Dum; @Dum-Jar].
Comprehensive research has been done on linear systems over ${\mathbb{P}}^2$ due to the Gimigliano-Harbourne-Hirschowitz Conjecture (see [@Hir] for orginal statement or [@Dum] for further references). In [@Dum-Jar], Dumnicki and Jarnicki gave a classification of all special systems over ${\mathbb{P}}^2$ with multiplicities bounded by $11$, which made it possible to verify that Gimigliano-Harbourne-Hirschowitz Conjecture holds for all systems of this type. The case of ${\mathbb{P}}^1\times{\mathbb{P}}^1$ seems to be less developed in terms of such classification. As long as all multiplicities of the base points equal $2$ the problem of specialty of a linear system has been widely studied by many authors for varieties of type ${\mathbb{P}}^{n_1}\times\ldots\times{\mathbb{P}}^{n_k}$. This is due to the fact that special linear systems of this kind are closely related to defective Segre-Veronese embeddings (see [@CGG1; @CGG2] by Catalisano, Geramita, Gimigliano and [@Bau-Dra; @CDM]).
As an application of the method introduced in Section 2, we state Theorem \[main1\], which gives a characterization of special linear systems over ${\mathbb{P}}^1\times{\mathbb{P}}^1$ with base points of multiplicity bounded by $3$. While writing this paper we found that such characterization had already been known for homogenous systems, i.e. systems with base points of all the multiplicities equal to 3 (Laface [@Laf]). The proof of Theorem \[main1\], which is the main result of this paper, shall be presented in Section 3.
Linear systems
==============
Let ${\mathbb{K}}$ be an arbitrary field, ${\mathbb{N}}=\{0,1,2,\ldots\}$, ${\mathbb{N}}^*=\{1,2,3,\ldots\}$. For any $\delta\in{\mathbb{N}}^2$ we write $\delta=(\delta_1,\delta_2)$.
\[def01\] Any finite and non-empty set $D{\subset}{\mathbb{N}}^2$ shall be called a *diagram*. Let $r\in{\mathbb{N}}^*$ and $m_1,\ldots,m_r\in{\mathbb{N}}$ (there is a technical reason to consider zero, see for example Definition \[def\_small\_matrix\]). Let $L$ be a field of rational functions over $K$ in variables $x_1,y_1,\ldots,x_r,y_r$. Define a *linear system* spanned over a diagram $D$ with base points of multiplicities $m_1,\ldots,m_r$ to be an $L$-vector space ${\mathcal{L}}_D(m_1,\ldots,m_r){\subset}L[X,Y]$ of polynomials $f=\sum_{\delta\in D} A_{\delta}X^{\delta_1}Y^{\delta_2}$, such that: $$\label{eq01}
\frac{\partial^{\alpha+\beta}f}
{\partial X^{\alpha}\partial Y^{\beta}}(x_i,y_i)=0,\text{ for }
i=1,\ldots,r\text{ and } \alpha+\beta<m_i$$ Let $M=M_D(m_1,\ldots,m_r)$ be the matrix of the system of equations , which are linear with respect to unknown coefficients $\{A_{\delta}\}_{\delta\in D}$. We say that the system ${\mathcal{L}}_D(m_1,\ldots,m_r)$ is special when $M$ is not of the maximal rank. Observe that the enteries of $M$ belong to the polynomial ring ${\mathbb{K}}[x_1,y_1,\ldots,x_r,y_r]$.
Given $d,e\in{\mathbb{N}}$ we denote by ${\mathcal{L}}_{(d,e)}(m_1,\ldots,m_r)$ a linear system spanned over the diagram $\{0,1,\ldots,d\}\times\{0,1,\ldots,e\}$. Let $D$ be a diagram and $m_1,\ldots,m_t\in{\mathbb{N}}$, $q_1,\ldots,q_t\in{\mathbb{N}}^*$. We shall use the following notation: $${\mathcal{L}}_D(m_1^{\times q_1},\ldots,m_t^{\times q_t}):=
{\mathcal{L}}_D(\overbrace{m_1,\ldots,m_1}^{q_1},\ldots,
\overbrace{m_t,\ldots,m_t}^{q_t})
.$$
\[main1\] Let us assume that $0\leq d\leq e$. A linear system of the form ${\mathcal{L}}_{(d,e)}(1^{\times p},2^{\times q},3^{\times r})$ is special if and only if one of the following conditions hold:
- $d=0$, $p+2q+3r\leq e$ and $q\geq 1$ or $r\geq 1$
- $d=1$, $p+3q+5r\leq 2e+1$ and $r\geq 1$
- $d=2$, $p=0$, $e=q+2r-1$ and $2\nmid q+r$
- $d=3$, and for some $n\geq 1$:
- $e=3n$, $p=q=0$ and $r=2n+1$
- $e=3n$, $p\leq 1$, $q=1$ and $r=2n$
- $e=3n+1$, $p\leq 2$, $q=0$ and $r=2n+1$
- $e=3n+2$, $p=0$, $q=2$ and $r=2n+1$
- $d=4$, $e=5$, $p=q=0$ and $r=5$
The proof of Theorem \[main1\] shall be presented in Section 3. Throughout the proof we will take advantage of a relation between geometrical properties of diagram $D$ and the rank of matrix $M_D(m_1,\ldots,m_r)$ (see Theorem \[prop08\]). For linear systems which contain only one base point, this relation can be expressed as follows (the proof can be found in [@Dum]):
\[prop02\] Let $D=\{\delta_1,\ldots,\delta_s\}$ be a diagram, and $\#D=s=\binom{m_1+1}{2}$. Then $\det M_D(m_1)=0$ (i.e. sytem ${\mathcal{L}}_D(m_1)$ is special) if and only if there extists a curve of degree $m_1-1$ that contains all of the points of $D$. Moreover, if $\det M_D(m_1)\neq 0$ then the following equation holds: $$\det M_D(m)=A\cdot x_1^{\delta_{1,1}+\ldots+\delta_{s,1}-s}
y_1^{\delta_{1,2}+\ldots+\delta_{s,2}-s},
\quad A\in{\mathbb{K}}\setminus0$$
\[def03\] Let $D$ be a diagram and $\#D=\binom{m+1}{2}$. We say that $D$ is *non-special (special)* of degree $m$ if $\det M_D(m)\neq 0(=0)$.
Unique tilings
==============
We introduce the notion of a *unique tiling* and state Theorem \[prop08\] and Theorem \[tw13\]. Thanks to these theorems we will be able to prove non-specialty of linear systems in terms of finding a solution for some specific problem of exact covering.
\[def04\] Given a diagram $D$ we define its *center of mass* as follows: $${\mathbb{Q}}^2\ni(c_1(D),c_2(D))=c(D):=\frac{1}{\#D}\sum_{d\in D}d$$
We say that a finite and non-empty set of diagrams $T$ is *a tiling*, if any two elements of $T$ are disjoint. Let $T$ and $T'$ be tilings, and consider a mapping $f:T\longrightarrow T'$. We say that $T$ and $T'$ are congruent through $f$, and denote it by $f:T\simeq T'$, if the following conditions hold:
- $f$ is one to one, and $\#f(D)=\#D$, $c(f(D))=c(D)$ for any $D\in T$
- $\bigcup T=\bigcup T'$
A tiling $T$ that contains only non-special diagrams is said to be *unique*, if $f:T\simeq T'$ implies that either $T=T'$ and $f=\operatorname{id}_U$, or $T'$ contains a special diagram.
\[def\_small\_matrix\] Given a diagram $D$ and numbers $m,r,i>0$, where $i\leq r$, we define: $$M_D^{(i)}(m):=M_D(0^{\times i-1},m,0^{\times {r-i}})$$ Observe, that the entries of the matrix $M_D^{(i)}(m)$ depend only on variables $x_i,y_i$ (see Definition \[def01\]).
The following theorem states a relation between uniqueness of a particular tiling and non-specialty of a linear system.
\[prop08\] Let $T=\{D_1,\ldots,D_r\}$ be a unique tiling such that $\#D_i=\binom{m_i+1}{2}$. If $D$ is a diagram for which one of the following conditions holds:
- $D_1\cup\ldots\cup D_r{\subset}D$
- $D{\subset}D_1\cup\ldots\cup D_r$
then ${\mathcal{L}}_D(m_1,\ldots,m_r)$ is non-special.
Let us denote $M=M_D(m_1,\ldots,m_r)$. We shall group the rows of $M$ into submatricies $M_1,\ldots,M_r$ such that $M_i$ coresponds to $\binom{m_i+1}{2}$ equations which depend on variables $x_i,y_i$ (see Definition \[def01\]). The columns of $M$ are indexed by the elements of $D$ in a natural way (each element of $D$ coresponds to a monomial).
It is sufficient to prove the theorem under the assumption of $(i)$. Let $D'=D_1\cup\ldots\cup D_r$. From $(i)$ it follows that the minor $M(D')$, i.e. the submatrix of $M$ consisting of columns indexed by the elements of $D'$, is of the maximal rank. By the Laplace decomposition we get: $$\begin{gathered}
\label{eq02}
\det M(D')
=\sum \varepsilon(D_1',\ldots,D_r')
\det M_1(D_1')\cdots\det M_r(D_r') \\
=\sum \varepsilon(D_1',\ldots,D_r')
\det M^{(1)}_{D_1'}(m_1)\cdots\det M^{(r)}_{D_r'}(m_r)\end{gathered}$$ where $\varepsilon(D_1',\ldots,D_r')\in\{-1,1\}$ and the summation runs over all partitions $\{D_1',\ldots,D_r'\}$ of $D'$ for witch $\#D_i'=\binom{m_i+1}{2}$.
Let $s_i=\#D_i'$. From Proposition \[prop02\] it follows that each non-zero component of the sum is non-zero if and only if, $D_1',\ldots,D_r'$ are non-special. In which case this component is a monomial of the form: $$\begin{gathered}
\label{eq03}
\det M^{(1)}_{D_1'}(m_1)\cdots\det M^{(r)}_{D_r'}(m_r)\\
=A\cdot
x_1^{c_1(D_1')s_1-s_1}
y_1^{c_2(D_1')s_1-s_1}
\cdots
x_r^{c_1(D_r')s_r-s_r}
y_r^{c_2(D_r')s_r-s_r}\end{gathered}$$ where $A\in{\mathbb{K}}$. As the tiling $T$ is unique, the non-zero monomial: $$\det M^{(1)}_{D_1}(m_1)\cdots\det M^{(r)}_{D_r}(m_r)$$ turnes up as a component of only once. Since it can not be reduced we get $\det M(D')\neq 0$.
A weak point of Theorem \[prop08\] is its assumption about the uniqueness of the tiling. Verifying whether a given tiling is unique or not may seem to be even more challanging than evaluating ”by hand” the rank of the matrix related to a linear system. This problem is partialy addressed by Theorem \[tw13\], which aims at giving some conditions that are sufficient for the uniqueness of a tiling. Before we state this result, we need to introduce some necessary definitions.
\[def09\] Given a diagram $D$ we define its *boundary distributions* and *innertial momentum* as follows: $$\begin{gathered}
\varphi_1(D):{\mathbb{N}}\ni\alpha\mapsto\#(D\cap\{\alpha\}\times{\mathbb{N}})\in{\mathbb{N}}\\
\varphi_2(D):{\mathbb{N}}\ni\beta\mapsto\#(D\cap{\mathbb{N}}\times\{\beta\})\in{\mathbb{N}}\\
i(D)=\sum_{\delta\in D}\Vert \delta-c(D)\Vert^2\end{gathered}$$ We say that a diagram $D$ is *a stable diagram* if it is non-special, its vertical and horizontal sections are segments, and for any diagram $D'$ the equations $\#D=\#D'$ and $c(D)=c(D')$ imply that at least one of the following conditions holds:
- $D'$ is special
- $i(D')>i(D)$
- $i(D')=i(D)$, $\varphi_1(D')=\varphi_1(D)$ and $\varphi_2(D')=\varphi_2(D)$
\[ex\_sing\] Any $1-$diagram, i.e. a diagram which is a singleton, is a stable diagram.
We define a relation on the set of all diagrams: $$\begin{gathered}
D\preceq D'\iff
\text{there exists }\delta=(\delta_1,\delta_2)\in D\text{ and }
\delta'=(\delta_1',\delta_2')\in D',\\
\text{such that }\delta_1=\delta_1'\text{ and }\delta_2\preceq \delta_2'\end{gathered}$$ When restricted to a particular tiling, this relation may be extended to a partial ordering. The following observation contains further details. We omit the simple proof.
\[obs12\] Suppose that the projection on the first coordinate of any diagram from a tiling $T$ is a segment. Then the following conditions are equivalent:
- relation $\preceq$ can be extended to a partial ordering on $T$
- for any $D,D'\in T$ the relations $D\preceq D'$ and $D'\preceq D$ imply $D=D'$.
\[tw13\] Suppose that a tiling $T$ consists of stable diagrams. If the relation $\preceq$ can be extended to a partial ordering on $T$ then $T$ is unique.
Whether a given diagram is special or not, can be usually verified with the help of Bézout Theorem due to Proposition \[prop02\]. Thanks to Observation \[obs12\]$(ii)$ it is very easy to state, for a given tiling, that $\preceq$ can be extended to a partial ordering. Meanwhile, determining if a diagram is stable or not seems to be a more complex task. As the condition of being a stable diagram is an indeterminant of an isometry, the problem of finding all stable diagrams of a bounded degree leads to a finite number of cases, and so it can be solved through effective, but harmfull, computation.
The figure below represents all stable diagrams, up to an isometry, consisting of $3$ or $6$ elements. Every diagram that is isometric to one of the following shall be called either *3-diagram* or *6-diagram*.
We follow by induction on the number of elements of $T$. It is clear that a tiling consisting of one diagram is always unique.
Let us denote $T=\{D_1,\ldots,D_s\}$ and $D=\bigcup T$. Suppose that $T'=\{D_1',\ldots,D_s'\}$ consists of non-special diagrams and $f:T\simeq T'$, where $f:D_j\mapsto D_j'$ for $i=1,\ldots,s$. Our goal is to prove that $D_j=D_j'$ for any $j$. From the elementary properties of innertial momentum we get: $$\label{eq04}
\sum_{j=1}^s\#D_j\Vert c(D_j)\Vert^2+i(D_j)
=\sum_{d\in D}\Vert d\Vert^2
=\sum_{j=1}^s\#D_j'\Vert c(D_j')\Vert^2+i(D_j')$$ As $D_j$ are stable diagrams and $D_j'$ are non-special, it follows that $i(D_j)\leq i(D_j')$. Due to we get $i(D_j)=i(D_j')$ for any $j=1,\ldots,s$. From the stability of $D_j$ we get $\varphi_1(D_j)=\varphi_1(D_j')$ for any $j=1,\ldots,s$.
By assumption we can extend $\preceq$ to a partial ordering on $T$. Without loss of generality we can assume that $D_1$ is minimal. We define $m(\alpha):=\min\{\beta\in{\mathbb{N}}:(\alpha,\beta)\in D_1\}$ for $\alpha\in{\mathbb{N}}$ such that $\varphi_1(D_1)(\alpha)>0$, and $m(\alpha):=0$ otherwise. Since the vertical sections of $D_1$ are segments, $\varphi_1(D_1)(\alpha)=\varphi_1(D_1')(\alpha)$, and $D_1$ is minimal we get: $$\begin{gathered}
\label{eq05}
m(\alpha)+(m(\alpha)+1)+\ldots+(m(\alpha)+\varphi_1(D_1)(\alpha)-1)\\
=\frac{(\varphi_1(D_1)(\alpha)-1)\varphi_1(D_1)(\alpha)}{2}
+\varphi_1(D_1)(\alpha)\cdot m(\alpha)
\leq
\sum_{(\alpha,\beta)\in D_1'}\beta\end{gathered}$$ Summing up for all possible $\alpha$, we get: $$\begin{gathered}
\label{eq06}
\#D_1\cdot c_2(D_1)
=\sum_{\alpha\in{\mathbb{N}}}\frac{(\varphi_1(D_1)(\alpha)-1)\varphi_1(D_1)(\alpha)}{2}
+\varphi_1(D_1)(\alpha)\cdot m(\alpha)\\
\leq\sum_{\alpha\in{\mathbb{N}}}\sum_{(\alpha,\beta)\in D_1'}\beta
=\#D_1'\cdot c_2(D_1')\end{gathered}$$ Due to $\#D_1\cdot c_2(D_1)=\#D_1'\cdot c_2(D_1')$ we get equations in both and . This implies $D_1=D_1'$. According to the induction hypothesis we already know that the tiling $T\setminus\{D_1\}$ is unique. Hence, from $f:T\setminus\{D_1\}\simeq T'\setminus\{D_1\}$, we get $D_2=D_2',\ldots,D_r=D_r'$.
The proof of Theorem \[main1\]
==============================
The proof shall be divided into several lemmas. The first of these gives an explanation of why the linear system which fullfils one of the conditions from Theorem \[main1\] is a special linear system.
One can consider using the Cremona transformation as a method of verifing the specialty of these linear systems (see for example [@Dum-Jar]). This is due to the fact that every complete linear system of the form ${\mathcal{L}}_{(d,e)}(m_1,\ldots,m_r)$ (over ${\mathbb{P}}^1\times{\mathbb{P}}^1$) is isomorphic to a linear system over ${\mathbb{P}}^2$ (see [@CGG1] for more details).
The linear systems listed in the hypothesis of Theorem \[main1\] are special.
Throughout the proof we will refer to the following two observations. We omit their proofs, as they are very simple.
\[obs\_aux1\] Consider a diagram $D$ and numbers $m_1,\ldots,m_r\geq 1$. The following properties hold:
- if $\#D>\sum_i\binom{m_i+1}{2}$ then system ${\mathcal{L}}_D(m_1,\ldots,m_r)$ is non-empty (i.e. it contains a non-zero polynomial)
- if we have equation in $(i)$ then the system ${\mathcal{L}}_D(m_1,\ldots,m_r)$ is non-empty if and only if it is special
\[obs\_aux2\] Let $D$, $D'$ be diagrams, and $m_1,m_1',\ldots,m_r,m_r'\in{\mathbb{N}}$ (some of them can be zero). Then, the following is true: $${\mathcal{L}}_D(m_1,\ldots,m_r)\cdot{\mathcal{L}}_{D'}(m_1',\ldots,m_r'){\subset}{\mathcal{L}}_{D+D'}(m_1+m_1',\ldots,m_r+m_r')$$ where $D+D'=\{d+d'\mid d\in D, d'\in D'\}$. Furthermore, if the systems on the left are non-empty, then: $$\begin{gathered}
\dim_L{\mathcal{L}}_{D+D'}(m_1+m_1',\ldots,m_r+m_r')\\
\geq\max\{\dim_L{\mathcal{L}}_D(m_1,\ldots,m_r),\dim_L{\mathcal{L}}_D'(m_1',\ldots,m_r')\}\end{gathered}$$
Let us begin with a system of form $(2)$. W assume that $q+r=2k+1$ for some $k\geq 0$ and claim that the following system is special: $${\mathcal{L}}_{(2,q+2r-1)}(2^{\times q},3^{\times r})
.$$ Thanks to Observation \[obs\_aux1\]$(ii)$ it is sufficient to state that the system is non-empty. From Observation \[obs\_aux2\] it follows that: $$({\mathcal{L}}_{(1,k)}(1^{\times q},1^{\times r}))^2
\cdot {\mathcal{L}}_{(0,r)}(0^{\times q},1^{\times r}){\subset}{\mathcal{L}}_{(2,q+2r-1)}(2^{\times q},3^{\times r})$$ The factors are non-empty due to Observation \[obs\_aux1\]$(i)$ as the coresponding diagrams have respectively $2\cdot(k+1)$ and $r+1$ elements. The thesis is now a consequence of Observation \[obs\_aux2\].
The same type of reasoning can be applied for $(3.1)$, $(3.4)$ and $(4)$. The following inclusions should be considered: $$\begin{aligned}
{\mathcal{L}}_{(2,2n)}(0,2,2^{\times 2n})\cdot
{\mathcal{L}}_{(1,n)}(0,1,1^{\times 2n})\;{\subset}\; &
{\mathcal{L}}_{(3,3n)}(0,3,3^{\times 2n})
\\
{\mathcal{L}}_{(2,2n+2)}(2^{\times 2},2^{\times 2n+1})\cdot
{\mathcal{L}}_{(1,n)}(0^{\times 2},1^{\times 2n+1})\;{\subset}\; &
{\mathcal{L}}_{(3,3n+2)}(2^{\times 2},3^{\times 2n+1})
\\
{\mathcal{L}}_{(2,4)}(2^{\times 5})\cdot
{\mathcal{L}}_{(2,1)}(1^{\times 5})\;{\subset}\; &
{\mathcal{L}}_{(4,5)}(3^{\times 5})\end{aligned}$$ The factors that contain base points of multiplicity $2$ are non-empty due to the case $(2)$.
We will need a more datailed estimation for $(3.2)$ and $(3.3)$. Let us consider the following inclusions: $$\begin{aligned}
{\mathcal{L}}_{(2,2n)}(0,2,2^{\times 2n})\cdot
{\mathcal{L}}_{(1,n)}(0,0,1^{\times 2n})\:{\subset}\; &
{\mathcal{L}}_{(3,3n)}(0,2,3^{\times 2n})
\\
{\mathcal{L}}_{(2,2n)}(0,2,2^{\times 2n})\cdot
{\mathcal{L}}_{(1,n)}(1,0,1^{\times 2n})\;{\subset}\; &
{\mathcal{L}}_{(3,3n)}(1,2,3^{\times 2n})
\\
{\mathcal{L}}_{(2,2n)}(0,0,2^{\times 2n+1})\cdot
{\mathcal{L}}_{(1,n+1)}(0,0,1^{\times 2n+1})\;{\subset}\; &
{\mathcal{L}}_{(3,3n+1)}(0,0,3^{\times 2n+1})
\\
{\mathcal{L}}_{(2,2n)}(0,1,2^{\times 2n+1})\cdot
{\mathcal{L}}_{(1,n+1)}(0,1,1^{\times 2n+1})\;{\subset}\; &
{\mathcal{L}}_{(3,3n+1)}(0,1,3^{\times 2n+1})
\\
{\mathcal{L}}_{(2,2n)}(1,1,2^{\times 2n+1})\cdot
{\mathcal{L}}_{(1,n+1)}(1,1,1^{\times 2n+1})\;{\subset}\; &
{\mathcal{L}}_{(3,3n+1)}(1,1,3^{\times 2n+1})\end{aligned}$$ For each case we can easily compute the dimension of the second factor, which equals respectively: $2,1,3,2,1$ (the systems containing points of multiplicity $1$ are alway non-special). From Observation \[obs\_aux2\] we know that the dimension of the system on the right side of the inclusion is at least: $2,1,3,2,1$, which is more than the expected dimension. Therefore, all systems on the right are special.
Finally, we move to the cases of $(0)$ and $(1)$. Consider the system ${\mathcal{L}}_{(1,e)}(1^{\times p},2^{\times q},3^{\times r})$. Suppose that $p+3q+5r\leq 2e+1$ and $r\geq 1$. Let us denote $M=M_{(1,e)}(1^{\times p},2^{\times q},3^{\times r})$ Our goal is to show that each minor of $M$ of the maximal rank is zero.
Evry $6$ points arbitrarily choosen from the diagram $\{0,1\}\times\{0,\ldots,e\}$ are contained in two lines. Therefore, it follows from Proposition \[prop02\] that the rows of $M$ corresponding to a point of multiplicity $3$ (we assumed that there is at least one such point) are linearly dependent. Hence, if only $p+3q+6r\leq 2e+2$ (i.e. the number of rows does not exceed the number of columns), the rank of $M$ can not be maximal.
Suppose that $p+3q+6r>2e+2$ (i.e. there are more rows than columns). From $p+3q+5r\leq 2e+1$ it follows that among any $2e+2$ rows there are at least $6$ rows coresponding to the same point of multiplicity $3$. But, as before observed, these rows are linearly dependent.
Let us consider the system ${\mathcal{L}}_{(0,e)}(1^{\times p},2^{\times q},3^{\times r})$. We assume that $p+2q+3r\leq e$ and $q\geq 1$ or $r\geq 1$. Let $M=M_{(0,e)}(1^{\times p},2^{\times q},3^{\times r})$. If the number of rows does not exceed the number of columns we proceed as before. Otherwise, from $p+2q+3r\leq e$ it follows that among any $e+1$ rows there are at least $3$ rows corresponding to a point of multiplicity $2$ or at least $4$ rows corresponding to a point of multiplicity $3$. We can apply the previous arguments to the former case. To deal with the latter case, observe that given any $4$ rows coresponding to a point of multiplicity $3$, at least one of them is zero. Indeed, each monomial of the form $X^{\alpha}Y^{\beta}$, where $\alpha\leq 1$, becomes zero after calculating the second derivative with respect to $X$.
In the following lemmas we prove non-specialty of a large class of linear systems. What we actually show, is the following:
*Given a system of the form ${\mathcal{L}}_{(d,e)}(1^{\times p},2^{\times q},3^{\times})$, for which none of the conditions $(0)-(4)$ from Theorem \[main1\] hold, there exists a tiling that fulfills both the hypothesis of Theorem \[prop08\] and Theorem \[tw13\].*
For $k,l\geq 1$ we shall use the notion $k\times l:=\{0,\ldots,k-1\}\times\{0,\ldots,l-1\}$. In several cases we will write ${\mathcal{L}}_{k\times l}(\ldots)$ instead of ${\mathcal{L}}_{(k-1,l-1)}(\ldots)$. It is obvious, that if a proper tiling is already constructed for a system of the form ${\mathcal{L}}_D(2^{\times q},3^{\times r})$ we can add some $1-$diagrams (see Example \[ex\_sing\]) so as to achive the desired value of $p$. Therefore, for simplicity, we always assume that $p=0$. The assumption from Theorem \[tw13\] concerning the possibility of extending relation $\preceq$ to a partial ordering, shall be alway fulfield. This will follow immediately from Observation \[obs12\]$(ii)$.
\[lem\_aux1\] For any $d,e\geq 5$ and $p,q,r\geq 0$, a system of the form ${\mathcal{L}}_{(d,e)}(1^{\times p},2^{\times q},3^{\times r})$ is non-special.
Let $k=d+1$, $l=e+1$. Without losing generality we can assume that $kl-3<3q+6r<kl+6$ and $p=0$. We shall prove that it is sufficient to consider $k,l<12$. Let us observe that any rectangular diagram of height $6$ can be tiled with either $6-$diagrams or $3-$diagrams by means of the schemes presented in Figure \[fig:6x\].
Now suppose, that $6\leq k\leq l$ and $l\geq 12$. From $kl-3\leq 3p+6q$ it follows that $q\geq k$ or $p\geq 2k$. If the former holds then we can tile a rectangle $k\times 6$ and reduce the problem to a smaller rectangle $k\times (l-6)$, while still having $l-6\geq 6$ and $k(l-6)-3<3p+6(q-k)$. We proceed in the same way in case of $p\geq 2k$.
Therefore, the problem of finding a proper tiling can be reduced to a fininte number of cases, namely $k,l<12$ and $kl-3<3q+6r<kl+6$. Some possible schemes that covers all neccessery constructions can be found in Figure \[fig:6\_schemes\] at the end of this paper. Because each $6-$diagram, apart from the ”triangular” one, can be divided into $3-$diagrams, even the use of only $6-$diagrams satisfies all cases.
For $k=6$ and $k=9$ the ”triangular” diagram has to be used (see Figure \[fig:6\_schemes\]). To address this problem one can find a way of covering a ”triangular” diagram together with one of its ”neighbours” using four $3-$diagrams.
\[lem\_aux2\] For any $e\geq 6$ and $p,q,r\geq 0$, a linear system of the form ${\mathcal{L}}_{(4,e)}(1^{\times p},2^{\times q},3^{\times r})$ is non-special.
Let us denote $k=e+1$. We proceed as in the proof of Lemma \[lem\_aux1\]. We will reduce the problem of finding appropriate tiling to a finite number of cases, namely $k<19$.
We can assume that $p=0$ nad $5k-3<3q+6r$. Therefore, we get $q\geq 10$ or $r\geq 10$. If $q\geq 10$ then the problem can be reduced to the rectangle $5\times (k-12)$ as one can tile the rectangle $5\times 12$ using the example presented in Figure \[fig:5x\].
If $r<10$ and $q\geq 10$, we reduce to the rectangle $5\times (k-6)$. The schemes that cover the case of $k<19$ can be found in Figure \[fig:5\_schemes\].
\[lem\_aux4\] The system ${\mathcal{L}}_{(4, 5)}(1^{\times p},2^{\times q},3^{\times r})$ is special if and only if $p=q=0$ and $r=5$.
The schemes presented in Figure \[fig:5x6\] cover all possible cases.
\[lem\_aux5\] Consider numbers $3\leq e$, $p,q,r\geq 0$ and let $n$ be such that $0\leq e-3n<3$. Supposing that none of the following conditions hold (see Theorem \[main1\]):
- $e=3n$, $p=q=0$ and $r=2n+1$
- $e=3n$, $p\leq 1$, $q=1$ and $r=2n$
- $e=3n+1$, $p\leq 2$, $q=0$ and $r=2n+1$
- $e=3n+2$, $p=0$, $q=2$ and $r=2n+1$
then the system ${\mathcal{L}}_{(3,e)}(1^{\times p},2^{\times q},3^{\times r})$ is non-special.
Observe that if $r\geq 2n+2$ then the rectangle $4\times (e+1)$ can be covered with $n+1$ rectangle $4\times 3$, and each such rectangle can be tiled witch two $6-$diagrams.
When $r\leq 2n-1$ then one of the algorithms presented in Figure \[fig:alg4x\] can be used to consturct a tiling that fulfills the hypothesis of Theorem \[prop08\]. The block on the right should be used if the number of $6-$diagrams is even.
We still need to construct a tiling for $r=2n$ and $r=2n+1$. Observe that using $n-1$ rectangles $4\times 3$, we reduce the problem to $r=2$ or $r=3$ and $e+1\in\{4,5,6\}$. Since we assumed that none of the conditions (3.1-4) hold, one of the schemes in Figure \[fig:4x456\] can be used to produce the desired tiling.
[10]{} K. Baur, J. Draisma, [***Secant dimensions of low-dimensional homogeneous varieties***]{}, arXiv:math.AG/0707.1605v1 J. Draisma, [***A tropical approach to secant dimensions***]{}, J. Pure Appl. Algebra 212(2), 2008, 349–363 M. V. Catalisano, A. V. Geramita, A. Gimigliano, [***Higher Secant Varieties of Segre-Veronese varieties, Projective varieties with unexpected properties***]{}, Walter de Gruyter GmbH and Co. KG (2005), 81–107 M. V. Catalisano, A. V. Geramita, A. Gimigliano, [***Segre-Veronese embeddings of ${\mathbb{P}}^1\times{\mathbb{P}}^1\times{\mathbb{P}}^1$ and their secant varieties***]{}, Collect. Math. 58, 1 (2007), 1–24 M. Dumnicki, [***Reduction method for linear systems of plane curves with base points***]{}, Ann. Polon. Math. 90.2 (2007), 131–143 M. Dumnicki, W. Jarnicki, [***New effective bounds on the dimension of a linear system in ${\mathbb{P}}^2$***]{}, J. Symbolic Comput. 42 (2007), 621–635 C. Ciliberto, R. Miranda, [***Degeneration of planar linear systems***]{}, J. Reine Angew. Math. 501 (1998), 191–220 A. Hirschowitz, [***Une conjecture pour la cohomologie des diviseurs sur les surfaces rationnelles génériques***]{}, J. reine angew. Math. 397 (1989), 208–213 A. Laface, [***On linear systems of curves on rational scrolls***]{}, Geom. Dedicata 90 (2002), 127–144
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The spatial averaging used for the splitting of the local scale factor on the homogeneous background and small inhomogeneous perturbation leads to a non-local relationship between locally and globally defined comoving curvature perturbations. We study this relationship within a quasi-homogeneous, nearly spatially flat domain of the Universe. It is shown that, on scales larger than the size of the observed patch, the Fourier components of the locally defined comoving curvature perturbation are suppressed. We have also shown that the statistical properties of local and global comoving curvature perturbations are coincide on a small scale. Several examples are discussed in detail.'
author:
- 'N.A. Koshelev'
date: 'Received: date / Accepted: date'
title: 'Spatial averaging and a non-Gaussianity. '
---
Introduction. {#intro}
=============
The consideration of cosmological inhomogeneities within non-linear perturbation theory has attracted much attention in recent years. Investigation of non-Gaussian features of cosmological perturbations gives new opportunities to test inflationary models, theoretical models of thermal reheating of the Universe. It also provides a possibility to get new cosmological constraints on a number of high energy physics theories.
Currently, the non-linear perturbation theory is rapidly developing. The gradient expansion approach [@SB],[@9411040],[@9905064] led to the generalization of the $\delta N$ formalism [@Starobinsky],[@Sasaki_Stewart],[@Sasaki_Tanaka] to the non-linear case [@0411220],[@0504045] that gives a handy tool for an estimation and calculation of non-Gaussianities. The quantum field theory methods are increasingly penetrating into the cosmological perturbation theory. Non-perturbative approaches are under development. For example, there are such techniques as renormalized cosmological perturbation theory [@0509418], [@0509419] and the time renormalization group method [@0806.0971]. However, there are still many unresolved theoretical issues.
The non-linear generalization of the curvature perturbation implies the need to take into account the finiteness of the observable patch of the Universe in determining the local comoving curvature perturbation. This makes itself evident in the fact that the observer identifies the perturbation as deviation from the spatial average over the observable patch. The consequences of such definition probably related to the problem of infrared divergences in the calculation of $n$-point correlation functions by perturbation theory [@0707.0361] are still not well understood. The potential consequences have been considered recently [@1301.3128],[@1303.3549] in the context of a landscape picture of the Universe. In this picture, observed statistical properties of perturbations depend on the position of the observable region. A landscape Universe arises in eternal inflation [@Linde],[@Steinhardt], and some of its aspects were discussed in connection with curvaton scenario [@0511736], [@1012.0549]. However, it is worth noting that several questionable approximations have been applied in the papers [@1301.3128], [@1303.3549].
This paper is organized as follow. In Section \[ch2\], we consider a quasi-homogeneous domain of the Universe and its observable patch. We introduce globally and locally defined comoving curvature perturbations ($\zeta (x^\mu)$ and $\zeta_S(x^\mu)$, correspondingly) and briefly describe the relationship between them. The exact relations between the Fourier components of $\zeta $ and $\zeta_S$ are considered in Section \[ch3\]. We show that, on a scale much smaller than the size of an observable patch, the values of globally and locally defined comoving curvature perturbations are coincided with good accuracy. Section \[ch4\] serves to clarify some issues of application obtained in Section \[ch3\] equations. In Section \[ch5\], we discuss the obtained discrepancy with some claims of papers [@1301.3128], [@1303.3549]. We conclude the paper in Section \[ch6\].
The comoving curvature perturbation. {#ch2}
====================================
Let’s consider a domain of the Universe, which can be described by nearly homogeneous and spatially flat metric. The comoving curvature perturbation $\zeta (\mathbf{x})$ is defined by the local scale factor [@0411220] $$\label{locala}\tilde{a}(\mathbf{x},t) = a(t) e^{\zeta (\mathbf{x})}.$$ Equivalently, one can write $$\label{localzeta}\zeta (\mathbf{x})=\ln \left(\frac{\tilde{a}}{a}\right).$$ The ambiguity of this definition is eliminated by condition $$\label{constr}\langle\zeta (\mathbf{x})\rangle_L
\equiv\frac{1}{V_L}\int_{V_L} \zeta (\mathbf{x}) d^3\mathbf{x}=0,$$ where $ V_L $ is the comoving three-dimensional volume of the domain.
From equations (\[locala\]) and (\[constr\]), it follows that $$\label{backgra}a(t) = e^{\frac{1}{V_L}\int_{V_L} \ln\tilde{a} d^3\mathbf{x}}$$ and $$\label{zetagl}\zeta (\mathbf{x}) =\ln \tilde{a} - \frac{1}{V_L}
\int_{V_L} \ln\tilde{a} d^3\mathbf{x} .$$
If the considered domain is infinite, one can use the Fourier integral transformation $$\begin{aligned}
\label{Fourie1}\zeta (\mathbf{x}) &=& \int \frac{d^3k}{(2\pi)^3}
\zeta_\mathbf{k}e^{i\mathbf{k}\mathbf{x}}, \\
\label{Fourie2}\zeta _\mathbf{k}&=&\int \zeta(\mathbf{x})
e^{-i\mathbf{k} \mathbf{x}}d^3x.\end{aligned}$$
Assuming statistical homogeneity, the $n$-point correlators of $\zeta_\mathbf{k}$ are of the form $$\begin{aligned}
\label{2point}\langle \zeta_ {\mathbf{k}_1}\zeta_ {\mathbf{k}_2}\rangle
&=&(2\pi)^3\delta^{(3)}(\mathbf{k}_1+\mathbf{k}_2)P_\zeta(\mathbf{k}_1), \\
\label{3point}\langle\zeta _ {\mathbf{k}_1}\zeta_{\mathbf{k}_2}
\zeta_ {\mathbf{k}_3} \rangle &=&(2\pi)^3\delta^{(3)}
(\mathbf{k}_1+\mathbf{k}_2
+\mathbf{k}_3) B_\zeta(\mathbf{k}_1,\mathbf{k}_2, \mathbf{k}_3), \\
\label{4point}\langle\zeta_ {\mathbf{k}_1}\zeta _ {\mathbf{k}_2}
\zeta _ {\mathbf{k}_3} \zeta _ {\mathbf{k}_4} \rangle
&=&(2\pi)^3\delta^{(3)} (\mathbf{k} _1 +\mathbf{k}_2
+\mathbf{k}_3+\mathbf{k}_4) T_\zeta(\mathbf{k}_1), \mathbf{k}_2,
\mathbf{k}_3, \mathbf{k}_4).\end{aligned}$$ Bispectrum $B_\zeta(\mathbf{k}_1,\mathbf{k}_2, \mathbf{k}_3)$ and trispektrum $T_\zeta(\mathbf{k}_1, \mathbf{k}_2, \mathbf{k}_3,
\mathbf{k}_4)$ are usually parameterized in terms of the spectrum $P_\zeta(\mathbf{k})$ and the non-linear parameteres $f_{NL}$, $\tau_{NL}$, $g_{NL}$ by $$\begin{aligned}
\label{bs} B_\zeta(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3)
&=&\frac{6}{5}f_{NL}(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3)
\left(P_\zeta(\mathbf{k}_1\!)P_\zeta(\mathbf{k}_2\!) +
P_\zeta(\mathbf{k}_2\!)P_\zeta(\mathbf{k}_3\!)
+P_\zeta(\mathbf{k}_3\!)P_\zeta(\mathbf{k}_1\!) \right), \\
T_\zeta(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3,\mathbf{k}_4\!)
&=&\frac{1}{2}\tau_{NL}
(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3,\mathbf{k}_4\!)
\left(P_\zeta(\mathbf{k}_1\!)P_\zeta(\mathbf{k}_2\!)
P_\zeta(\mathbf{k}_1+\mathbf{k}_4\!) + cyclic~ permutations
\right) \nonumber \\
\label{trs}&+& \frac{54}{25}g_{NL}
(\!\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3,\mathbf{k}_4\!)
\left(P_\zeta(\!\mathbf{k}_1\!)P_\zeta(\!\mathbf{k}_2\!)
P_\zeta(\!\mathbf{k}_3\!) + cyclic~ permutations \right)\! .\end{aligned}$$ In the case of a finite domain, it is necessary to introduce modifications of the equations (\[2point\])-(\[4point\]).
On a sufficiently large scale, it is very convenient to use the ansatz $$\label{zetaanzatz}\zeta (\mathbf{x}) = \zeta_G (\mathbf{x})
+\frac{3}{5}f_{NL}^0\left(\zeta_G^2 (\mathbf{x})-\langle\zeta
_G^2\rangle\right)+\left(\frac{3}{5}\right)^2g_{NL}^0\left(\zeta_G^3
(\mathbf{x})-\langle\zeta _G^3\rangle \right) +... ,$$ where the auxiliary quantity $ \zeta_G (\mathbf{x}) $ is a Gaussian random variable and the non-linearity parameters $f_{NL}^0$, $g_{NL}^0$ are some dimensionless constants. This expression can be treated as $ \delta N $-formula for a simple inflationary model with one single scalar field $ \varphi $ $$\zeta (\mathbf{x}) = \sum_{n=1}^\infty \frac{N^{(n)}
(\varphi_0(t_*))}{n!}\delta\varphi^n(t_*,\mathbf{x}),$$ where $\varphi(t)$ is the background scalar field, $ t_ * $ is some moment immediately after horizon crossing and the scalar field perturbation $\delta\varphi^n(t_*,\mathbf{x}) = \zeta_G
(\mathbf{x})/ N '$ is assumed Gaussian.
Let us consider the patch $\Omega_S$ with comoving volume $ V_S
$, which is located within the treated above domain. In this patch, one can define the local comoving curvature perturbation, denoted as $ \zeta_S $. Analogically to equation (\[zetagl\]), this quantity is given by $$\label{zetaloc}\zeta_S (\mathbf{x}) =\ln \tilde{a} -
\frac{1}{V_S}\int_{V_S} \ln\tilde{a} d^3\mathbf{x} .$$
Since the volume averaging does not change the background scale factor ($\langle a(t)\rangle_S=a(t)$), one can obtain $$\zeta_S (\mathbf{x}) =\ln \left(\frac{\tilde{a}}{a}\right) -
\frac{1}{V_S}\int_{V_S} \ln \left(\frac{\tilde{a}}{a}\right)
d^3\mathbf{x} =\zeta(\mathbf{x}) - \frac{1}{V_S}\int_{V_S}
\zeta(\mathbf{x}) d^3\mathbf{x}.$$ This equation can be rewritten as [@1303.3549] $$\label{zetaSzeta}\zeta_S (\mathbf{x}) =\zeta(\mathbf{x}) -
\langle \zeta\rangle_S.$$
The equation (\[zetaSzeta\]) shows that the relationship between $\zeta_S$ and $\zeta$ is non-local. To calculate the value of the quantity $ \zeta_S $ at one point, it is necessary to know $\zeta(\mathbf{x})$ in the whole patch at the same moment of time. This non-locality is essential if the region $\Omega_S$ is not fixed, i.e., if $\Omega_S=\Omega_S(\mathbf{x})$. In this case, to find $\zeta_S (\mathbf{x})$ at all points of $\Omega_S(\mathbf{x}_0)$, it is necessary to know the comoving curvature perturbation $\zeta(\mathbf{x})$ at all points of some larger patch which includes $\Omega_S(\mathbf{x}_0)$.
Mapping between $\zeta$ and $\zeta_S$ in the momentum space. {#ch3}
============================================================
The procedure of spatial averaging plays an important role in the decomposition of the quantity $\tilde{a}(\mathbf{x},t) $ on the homogeneous background and small inhomogeneous perturbation. Here, we consider the relations in Fourier space, which follows from the equation (\[zetaSzeta\]).
We use the fact that the operation of spatial averaging is linear and consider only one Fourier mode $$\label{pw}\zeta(\mathbf{x})=\zeta (\mathbf{k})e^{i\mathbf{k}\mathbf{x}}.$$ We denote the comoving coordinates of the patch as $ \mathbf{x} _0
$. For example, if the region has a spherical shape, then $\mathbf{x}_0$ is the center of this sphere. We also introduce the integration variable $\mathbf{x}'$, so that $\mathbf{x}=\mathbf{x}_0+\mathbf{x}'$. The spatial averaging gives $$\langle \zeta\rangle_S=\frac{1}{V_S}\int \zeta (\mathbf{k})
e^{i\mathbf{k}(\mathbf{x}_0+\mathbf{x}')}d(\mathbf{x}_0+\mathbf{x}')
= W(\mathbf{k})\zeta (\mathbf{k})e^{i\mathbf{k}\mathbf{x}_0},$$ where $$\label{Wk}W(\mathbf{k})=\frac{1}{V_S}\int e^{i\mathbf{k}
\mathbf{x}'} d\mathbf{x}'.$$
Thus, the spatial averaging of the function (\[pw\]) gives a plane wave field with the same wave vector, but with a different amplitude. We obtain the known results [@1303.3549] $$\left[\langle \zeta\rangle_S \right]_\mathbf{k} = W(\mathbf{k})
\zeta_\mathbf{k}$$ and $$\label{zetak}\left[\zeta_S \right]_\mathbf{k} =(1-W(\mathbf{k}))
\zeta_\mathbf{k}.$$ The equation (\[zetak\]) is exact regardless of the form of the quantity $\zeta (\mathbf{x})$. Although this equation is known, its consequences for cosmological perturbation theory have not been investigated.
In what follows, we assume that the considered patch is Hubble sized at the present time. This choice allows us to get some physical consequences of equation (\[zetak\]).
Let’s denote the characteristic comoving size of the patch as $x_S$. The observer can confidently separate the perturbation from the background if the perturbation experienced several oscillations in the patch. This is carried out for the Fourier modes with wave vector $\mathbf{k}$ satisfying the inequality $|\mathbf{k}|x_S\gg 2\pi$. Such perturbations we call short wavelength ones. The long wavelength Fourier modes satisfy the opposite condition $|\mathbf{k}|x_S\ll 2\pi$.
For short wavelength perturbations, the equations (\[Wk\]), (\[zetak\]) give $$\label{largek}\left[\zeta_S \right]_{\mathbf{k}} \Big|_{k\gg k_S}
= \left[\zeta \right]_{\mathbf{k}}\Big|_{k\gg k_S} ,$$ where $ k = |\mathbf{k}| $ and $ k_S = 2\pi/x_S $. This equation show that the difference between $\left[\zeta_S
\right]_{\mathbf{k}}$ and $\left[\zeta \right]_{\mathbf{k}}$ can be neglected on a small scale.
Expanding the exponential function in a Taylor series, we get for long wavelength perturbation $$\label{smallk}\left[\zeta_S \right]_{\mathbf{k}}\Big|_{k\ll k_S}
\approx \frac{1}{2}\langle(\mathbf{k}\mathbf{x}')^2
\rangle_S\left[\zeta \right]_{\mathbf{k}}\Big|_{k\ll k_S}\propto
k^2\left[\zeta \right]_{\mathbf{k}}\Big|_{k\ll k_S}.$$ One can see that quantity $\left[\zeta_S \right]_{\mathbf{k}}$ is suppressed on a large scale. This is the consequence of the fact that the observer can not separate surely the long-wavelength perturbation from the background. Perhaps, the equation (\[smallk\]) will allow to solve still existing problem of infrared divergences in the calculation of correlation functions at loop level (see [@1306.4461] for a recent review of this topic), although the reformulation of the perturbation theory in terms of locally defined variables is difficult due the non-locality of the equation (\[zetaSzeta\]).
The form of the function $W(\mathbf{k})$ is of interest over a wide range of scales. Assuming that the observable patch can be approximated as a sphere of radius $ x_S $, a simple calculation gives $$\label{WkSphere}W(\mathbf{k})=3\left(\frac{\sin (kx_S)}{(kx_S)^3}
- \frac{\cos (kx_S)}{(kx_S)^2} \right).$$ Equations (\[largek\]) and (\[smallk\]) are main results of this work.
Since the quantity $\zeta_S(\mathbf{x})$ is statistically homogeneous, there are an expressions which are analogous to the equations (\[2point\])- (\[trs\]). In particular, we have $$\label{bsS} B_{\zeta_S}\!(\mathbf{k}_1,\mathbf{k}_2,\mathbf{k}_3)
\!=\!\frac{6}{5}f_{NL}^{(S)} \!(\mathbf{k}_1, \mathbf{k}_2,\mathbf{k}_3)
\left(P_{\zeta_S}\!(\mathbf{k}_1\!)\!P_{\zeta_S}\!(\mathbf{k}_2\!) +
P_{\zeta_S}\!(\mathbf{k}_2\!)\!P_{\zeta_S}\!(\mathbf{k}_3\!)
+P_{\zeta_S}\!(\mathbf{k}_3\!)\!P_{\zeta_S}\!(\mathbf{k}_1\!) \right).$$
For equal-$k$ ($k_1=k_2=k_3\equiv k$), the equations (\[bs\]),(\[bsS\]) gives $$f_{NL}^{(S)} (\mathbf{k}_1, \mathbf{k}_2, \mathbf{k}_3) =
\frac{f_{NL}}{1-W(k)}.$$ In the squeezed limit ($k_1 \approx k_2\equiv k\gg k_3$), the mutual relations of $f_{NL}^{(S)}$ and $f_{NL}$ depends on the scale. If $k_1$, $k_2$ are infrared and the primordial spectrum is scale-invariant, then $$f_{NL}^{(S)} (\mathbf{k}_1, \mathbf{k}_2, \mathbf{k}_3) \approx
f_{NL } (\mathbf{k}_1, \mathbf{k}_2, \mathbf{k}_3)
\frac{2}{1-W(k)} \frac{k}{k_3}.$$ In both cases, we obtain $$f_{NL}^{(S)} (\mathbf{k}_1, \mathbf{k}_2, \mathbf{k}_3) \gg f_{NL}
(\mathbf{k}_1, \mathbf{k}_2, \mathbf{k}_3).$$ Formally, the non-linearity parameter $f_{NL}^{(S)}$ is enhanced on a large scale due the suppression of $\left[\zeta_S \right]
_\mathbf{k}$.
Examples. {#ch4}
=========
The equation (\[largek\]) indicates that, on a small scale, the local comoving curvature perturbation $ \zeta_S $ has the same statistical properties as the quantity $ \zeta $. The only question is whether it is possible to recover the global statistical properties of $ \zeta_S $ using the observed data sample of $\left[ \zeta_S \right]_\mathbf{k}$. In other words, are there any reasons for which the local observer can fail to notice the coincidence of statistical properties of $\left[ \zeta_S
\right]_\mathbf{k}$ and $\left[ \zeta \right]_\mathbf{k}$ on a small scale?
At first, let us consider how to implement the equation (\[largek\]). For simplicity, we use the ansatz (\[zetaanzatz\]), in which we assume that only $f_{NL}^0\neq 0$. The ansatz (\[zetaanzatz\]) is reduced to $$\label{zetafnl}\zeta (\mathbf{x}) = \zeta_G (\mathbf{x})
+\frac{3}{5}f_{NL}^0\left(\zeta_G^2 (\mathbf{x})-\langle\zeta
_G^2\rangle\right)$$ and leads to the equation $$\label{zetaSfnl}\zeta_S (\mathbf{x}) = \zeta_G (\mathbf{x}) -
\langle \zeta_G\rangle_S +\frac{3}{5}f_{NL}^0\left(\zeta_G^2
(\mathbf{x}) -\langle\zeta_G^2\rangle_S \right).$$
In momentum space it gives $$\label{zetaSfnlk}\left[ \zeta_S \right]_\mathbf{k} =\left[ \zeta_G
- \langle \zeta_G\rangle_S \right]_\mathbf{k} +\frac{3}{5}
f_{NL}^0 \left[\zeta_G^2 -\langle\zeta_G^2 \rangle_S
\right]_\mathbf{k}.$$ The volume averaging over the Hubble patch cuts off short wavelength perturbations, so that on a small scale $$\label{zetaSk1}\left[ \zeta_S \right]_\mathbf{k} =\left[ \zeta_G
\right]_\mathbf{k} +\frac{3}{5}f_{NL}^0\left[\zeta_G^2
\right]_\mathbf{k} .$$ It coincides with the result (\[largek\]).
Observable patch is differs from the large domain by size, and the discrepancy between observed statistical properties of $\left[
\zeta_S \right]_\mathbf{k}$ and theoretical statistical properties of $\left[ \zeta \right]_\mathbf{k}$ may be expected to be due to the influence of infrared perturbations. Let us examine this issue in some detail.
Let’s the realization of the auxiliary quantity $ \zeta_G $ has the form $$\label{twomode}\zeta_G=\zeta_1+\zeta_2,$$ where $\zeta_1=\zeta_{\mathbf{k}_1}e^{\mathbf{k}_1 \mathbf{x}}$, $\zeta_2=\zeta_{\mathbf{k}_2}e^{\mathbf{k}_2\mathbf{x}}$ and $k_1
\ll k_S$ (long wave), $k_2 \gg k_S$ (short wave). Then the equation (\[zetaSfnlk\]) yelds $$\label{zetatm}\left[\zeta_S \right]_{\mathbf{k}_2} = \left[
\zeta_G \right]_{\mathbf{k}_2} +\frac{3}{5}f_{NL}^0\left[\zeta_G^2
-\langle\zeta_G^2\rangle_S\right]_{\mathbf{k}_2} .$$
In the coordinate space, we obtain from equations (\[twomode\]), (\[zetatm\]) $$\zeta_G^2 -\langle\zeta_G^2\rangle_S = \left(\zeta_{2}^2 -
\langle\zeta_{2}^2 \rangle_S \right)+ 2\zeta_{1} \left( \zeta_{2}
- \langle\zeta_{2}\rangle_S\right).$$
In the momentum space it gives $$\left[\zeta_G^2 -\langle\zeta_G^2\rangle_S \right]
_{\mathbf{k}_2} = \left[ \left(\zeta_{2}^2 - \langle\zeta_{2}^2
\rangle_S \right)+ 2\zeta_{1} \left( \zeta_{2} -
\langle\zeta_{2}\rangle_S\right)\right]_{\mathbf{k}_1}=0.$$
Using this result one can obtain $$\left[\zeta_S \right]_{\mathbf{k}_2} = \left[ \zeta_G
\right]_{\mathbf{k}_2},$$ i. e., the long wavelength Fourier mode with wave vector $\mathbf{k}_1$ does not affect the mode with wave vector $\mathbf{k}_2$. It is clear also from consideration of the convolution theorem $$\label{conv}\left[\zeta_G ^2 \right]_{\mathbf{k}}=\int
\left[\zeta_G \right]_{\mathbf{q}}\left[\zeta_G \right]
_{\mathbf{k}-\mathbf{q}}\frac{d^3\mathbf{q}}{(2\pi)^3}.$$
If the quantity $\zeta_G$ has two modes with wave vectors $\mathbf{k}_1$ and $\mathbf{k}_2$, then the quantity $\zeta_G^2$ contain modes with wave vectors $2\mathbf{k}_1$, $2\mathbf{k}_2$, $\mathbf{k}_1 + \mathbf{k}_2$ only.
To gain influence on mode with the wave vector $\mathbf{k}_2$, the expansion of $\zeta_G$ should include à mode with $\mathbf{k}_3 =
\mathbf{k}_2-\mathbf{k}_1$. However, just like the Fourier mode with $\mathbf{k}_1$ is indistinguishable from the background, the modes with $\mathbf{k}_2$ and $\mathbf{k}_3$ are indistinguishable between themselves. In other words, infrared perturbations do not contribute to the correlation functions of physically distinguishable modes.
A landscape Universe. {#ch5}
=====================
Recently, in the papers [@1301.3128],[@1303.3549], it was made several claims, which seem to contradict the equation (\[largek\]). However, the authors of [@1301.3128],[@1303.3549] have used several approximations which require very careful handling when working in Fourier space. On the other hand, the parameters of ansatz (\[zetaanzatz\]) are sensitive to the choice of the auxiliary quantity $\zeta_G$. This can be verified by direct calculation in the simplest case of anzatz (\[zetaSfnl\]).
We introduce the notation $$\label{GS}\zeta_{G,S}=\zeta_G-\bar{\zeta},$$ where $\bar{\zeta}=\langle \zeta_G\rangle_S$. It is important that the quantity $\tilde{\zeta}$ can be considered as a constant on a small scale, i.e., one can assume $$\label{kmean}\bar{\zeta} _{\mathbf{k}}\Big|_{k\gg k_S} =0.$$
The equation (\[zetaSfnl\]) and the equality $$\label{eqconstr}\zeta_{G,S}^2-\langle\zeta_{G,S}^2 \rangle_S =
\zeta_G^2 - \langle\zeta_G^2\rangle-2\bar{\zeta}\zeta_{G,S}$$ gives $$\label{lanscape1}\zeta_S (\mathbf{x}) =\left(1 + \frac{6}{5}f_{NL}
\bar{\zeta}\right) \zeta_{G,S} +\frac{3}{5}f_{NL}^0
\left(\zeta_{G,S}^2 -\langle\zeta_{G,S}^2\rangle_S \right).$$ This result is in full agreement with the result of the paper [@1303.3549]. On a small scale, the equations (\[kmean\]), (\[lanscape1\]) yields now $$\left[ \zeta_S \right]_\mathbf{k} =\left(1 + \frac{6}{5}f_{NL}^0
\bar{\zeta}\right) \left[\zeta_{G,S}\right]_\mathbf{k}
+\frac{3}{5}f_{NL}\left[\zeta_{G,S}^2\right]_\mathbf{k}.$$
One can make the variable transformation $$\label{chix}\chi_G=\left(1 + \frac{6}{5}f_{NL}^0 \bar{\zeta}
\right) \zeta_{G,S}.$$ It gives [@1303.3549] $$\label{zetaSk2}\left[ \zeta_S \right]_\mathbf{k} = \left[\chi_{G}
\right] _\mathbf{k} +\frac{3}{5}f_{NL}^{0(S)} \left[\chi_{G}^2
\right] _\mathbf{k},$$ where $$\label{fnl0s}f_{NL}^{0(S)}=\frac{f_{NL}^0}{\left(1 +
\frac{6}{5}f_{NL}^0 \bar{\zeta}\right)^2}.$$ Consequently, both parameterizations (\[zetaSk2\]) and (\[zetaSk1\]) are possible at once. This can easily be checked on a small scale using the equalities $$\begin{aligned}
\left[\zeta_{G,S}\right]_\mathbf{k} &=& \left[\zeta_{G}
\right]_\mathbf{k},\\
\label{GS2} \left[\zeta_{G,S}^2\right]_\mathbf{k} &=&
\left[\zeta_{G}^2\right]_\mathbf{k} -2\bar{\zeta} \left[\zeta_{G}
\right]_\mathbf{k}.\end{aligned}$$ which follows from the equations (\[GS\]) and (\[eqconstr\]) correspondingly. The parameterizations (\[zetaSk1\]) and (\[zetaSk2\]) yields to the same numerical values for the Fourier components of the local comoving curvature perturbation, as it should be.
It is significant that the convolution theorem (\[conv\]) allows us to rewrite the equation (\[GS2\]) as $$\left[\zeta_{G,S}^2\right]_\mathbf{k} = \int \left[\zeta_G
\right]_{\mathbf{q}}\left[\zeta_G \right]_{\mathbf{k}
-\mathbf{q}}\frac{d^3\mathbf{q}}{(2\pi)^3} -2\bar{\zeta}
\left[\zeta_{G} \right]_\mathbf{k}.$$ This expression differs from the approximate one used in the papers [@1301.3128], [@1303.3549].
Equations (\[chix\]), (\[GS2\]) and (\[fnl0s\]) indicate that the quantities $\chi_G$ and $f_{NL}^{0(S)}$ are influenced by the same random parameter $\bar{\zeta}$. Though, the associated uncertainty of parametrization did not affect the numerical value of the local comoving curvature perturbation. One can see that some arbitrariness in definition of the auxiliary function $\chi_G$ is compensated by a corresponding change of the non-linearity parameter $f_{NL}^{0(S)}$ without affecting $\left[
\zeta_S \right]_\mathbf{k}$ and its correlation functions. The stochastic quantity $\bar{\zeta}$ also does not affect the calculation of the parameter $f_{NL}^{(S)}$ defined by the equation (\[bsS\]).
Conclusions. {#ch6}
============
We have studied the relationship between locally and globally defined comoving curvature perturbations within a quasi-homogeneous domain of the Universe. It is shown that, on scales larger than the size of the observed patch, the Fourier components of the locally defined comoving curvature perturbation are suppressed. It is shown also that the statistical properties of local and global comoving curvature perturbations are coincide on a small scale. Several examples are discussed in detail. We have shown that, in the simplest cases, the long wavelength perturbations do not contribute to the bispectrum and trispektrum on a sub-Hubble scale.
D. S. Salopek and J. R. Bond, “Nonlinear evolution of long-wavelength metric fluctuations in inflationary models,” Phys. Rev. **D42** 3936 (1990).
N. Deruelle and D. Langlois, “Long wavelength iteration of Einstein’s equations near a spacetime singularity,” Phys. Rev. **D52**, 2007 (1995).
M. Shibata and M. Sasaki, “Black hole formation in the Friedmann universe: Formulation and computation in numerical relativity,” Phys. Rev. **D60**, 084002 (1999).
A. A. Starobinsky, “Multicomponent de Sitter (inflationary) stages and the generation of perturbations,” JETP Lett. **42**, 152 (1985).
M. Sasaki and E. D. Stewart,“A General Analytic Formula for the Spectral Index of the Density Perturbations produced during Inflation,” Prog. Theor. Phys. **95**, 71 (1996).
M. Sasaki and T. Tanaka, “Super-Horizon Scale Dynamics of Multi-Scalar Inflation, ” Prog. Theor. Phys. **99**, 763 (1998).
D. H. Lyth, K. A. Malik and M. Sasaki, “A general proof of the conservation of the curvature perturbation,” JCAP **0505**, 004 (2005).
D. H. Lyth and Y. Rodriguez, “The Inflationary prediction for primordial non-Gaussianity,” Phys. Rev. Lett. **95**, 121302 (2005).
M. Crocce, R. Scoccimarro, “Renormalized Cosmological Perturbation Theory,” Phys.Rev.**D73** 063519 (2006).
M. Crocce, R. Scoccimarro, “Memory of Initial Conditions in Gravitational Clustering,” Phys.Rev.**D73** 063520 (2006).
M. Pietroni, “Flowing with Time: a New Approach to Nonlinear Cosmological Perturbations,” JCAP **0810** 036 (2008).
D. H. Lyth, “The curvature perturbation in a box,” JCAP **0712** 016 (2007).
S. Nurmi, C. T. Byrnes and G. Tasinato, “A non-Gaussian landscape,” JCAP **06** 004 (2013).
M. LoVerde, E. Nelson, S. Shandera, “Non-Gaussian Mode Coupling and the Statistical Cosmological Principle,” JCAP **06**, 024 (2013).
A. D. Linde, “Eternally existing selfreproducing chaotic inflationary universe,” Phys.Lett. **B 175**, 395–400 (1986).
P. J. Steinhardt, “Natural inflation,” in: *The Very Early Universe*, ed. G.W. Gibbons, S.W. Hawking and S. Siklos, Cambridge University Press, (1983).
A. D. Linde and V. Mukhanov, “The Curvaton Web,” JCAP **0604** 009 (2006).
V. Demozzi, A. Linde and V. Mukhanov, “Supercurvaton,” JCAP **1104** 013 (2011).
T. Tanaka, Y. Urakawa, “Loops in inflationary correlation functions,” arXiv:1306.4461 \[hep-th\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present , a novel hybrid (mixed-protocol) framework for secure function evaluation (SFE) which enables two parties to jointly compute a function without disclosing their private inputs. combines the best aspects of generic SFE protocols with the ones that are based upon additive secret sharing. In particular, the framework performs linear operations in the ring $\mathbb{Z}_{2^l}$ using additively secret shared values and nonlinear operations using Yao’s Garbled Circuits or the Goldreich-Micali-Wigderson protocol. departs from the common assumption of additive or linear secret sharing models where three or more parties need to communicate in the online phase: the framework allows two parties with private inputs to communicate in the online phase under the assumption of a third node generating correlated randomness in an offline phase. Almost all of the heavy cryptographic operations are precomputed in an offline phase which substantially reduces the communication overhead. is both scalable and significantly more efficient than the ABY framework (NDSS’15) it is based on. Our framework supports signed fixed-point numbers. In particular, ’s vector dot product of signed fixed-point numbers improves the efficiency of mining and classification of encrypted data for algorithms based upon heavy matrix multiplications. Our evaluation of on a 5 layer convolutional deep neural network shows 133x and 4.2x faster executions than Microsoft CryptoNets (ICML’16) and MiniONN (CCS’17), respectively.'
author:
- 'M. Sadegh Riazi'
- Christian Weinert
- Oleksandr Tkachenko
- 'Ebrahim M. Songhori'
- Thomas Schneider
- Farinaz Koushanfar
bibliography:
- 'main\_bib.bib'
title: |
: A Hybrid Secure Computation Framework\
for Machine Learning Applications
---
<ccs2012> <concept> <concept\_id>10002978.10002991.10002995</concept\_id> <concept\_desc>Security and privacy Privacy-preserving protocols</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
#### Acknowledgements
This work has been co-funded by the DFG as part of project E4 within the CRC 1119 CROSSING and by the German Federal Ministry of Education and Research (BMBF) as well as by the Hessen State Ministry for Higher Education, Research and the Arts (HMWK) within CRISP.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'By presenting an approximated performance-complexity tradeoff (PCT) algorithm, a low-complexity non-binary low density parity check (LDPC) code over $q$-ary-input symmetric-output channel is designed in this manuscript which converges faster than the threshold-optimized non-binary LDPC codes in the low error rate regime. We examine our algorithm by both hard and soft decision decoders. Moreover, simulation shows that the approximated PCT algorithm has accelerated the convergence process by $30\%$ regarding the number of the decoding iterations.'
author:
- '[Yang Yu, and Wen Chen, ]{}[^1] [^2] [^3]'
bibliography:
- 'IEEEabrv.bib'
- 'QLDPC.bib'
title: 'Design of Low Complexity Non-binary LDPC Codes with an Approximated Performance-Complexity Tradeoff'
---
Nonbinary LDPC, EXIT chart, performance-complexity tradeoff, Gallager decoding algorithm b.
Introduction
============
Investigation over Galois field $GF(q)$, $q=2^p$, shows that $q$-ary LDPC codes have potentially better performance than binary LDPC codes for not very long block length at the cost of higher decoding complexity, and irregular LDPC codes can outperform the regular LDPC codes[@Davey98montecarlo]. The design of high-performance nonbinary LDPC codes has been studied in the literature[@Byers05ExitNonbinary; @Rathi05DeNonbinary; @ge09NLDPC]. A major concern of $q$-ary LDPC is the decoding complexity. PCT analysis in [@Wei05PctB; @smith10LdpcPCT] utilizes the nature of binary iterative decoder, in which messages passing through each iteration, can be profiled by a single parameter. The code design problem is then reduced to the shaping of the decoding trajectory of extrinsic information transfer (EXIT) chart for an optimal PCT[@smith10LdpcPCT], where they show that the (decoding) complexity optimized binary LDPC codes outperforms the threshold optimized binary LDPC codes. However, messages, passing through the nonbinary LDPC decoder, are vectors[@ge09NLDPC]. The main challenge in cooperating PCT in nonbinary LDPC codes design is how to characterize the decoding complexity as a uni-parametric transfer function. To solve this problem, we present an irregular EXIT chart by using an upper bound of the message error probability, based on which, an approximated performance-complexity tradeoff (PCT) algorithm is put forward to design irregular nonbinary LDPC codes with optimized decoding complexity.
However, advantages of the proposed approximated PCT algorithm are obvious: firstly, [@Byers05ExitNonbinary; @Rathi05DeNonbinary] give methods to predict the performance threshold for nonbinary LDPC codes, but the complexity can not be optimized based on these procedures. The presented EXIT chart can be also used to reduce the decoding complexity. Secondly, the complexity optimization algorithm in [@Wei05PctB; @smith10LdpcPCT] is applicable for binary LDPC codes with rates greater than $1/4$. But the optimization algorithm in this manuscript is a universal method in the sense that, when $q=2$, the algorithm coincides with binary case.
Preliminaries
=============
LDPC Codes
----------
An LDPC code is called regular if the column and row weight of the parity check matrix is constant, respectively. The irregular LDPC codes can be characterized by variable degree distribution $$\lambda (x) = \sum \limits_{i \geqslant 2}\lambda_i x^{i-1},$$ and check degree distribution $$\rho (x) = \sum \limits_{i \geqslant 2}\rho_i x^{i-1},$$ from an edge-perspective, where $\lambda_i$ and $\rho_i$ are the fraction of edges belonging to degree-$i$ variable and check node, respectively. Using this characterization, code rate $R$ is given by $R = 1 - \frac{\int_0^1 \rho (x)dx}{\int_0^1 \lambda (x)dx}$, and $\lambda(1)=\rho(1)=1$. Due to this characterization, Fig. \[fig:one\_tree\] gives the depth-one decoding tree for a degree-$i$ variable node. During one iteration, messages (beliefs) are passed from the input to the output of the tree. The EXIT chart based on message error probability of LDPC codes can be given by $$\label{p_out}
p_{out} = \sum \limits_{i \geqslant 2} \lambda_i f_i(p_{in}),$$ where $p_{in}$ is the input error probability and $f_i$ is the elementary EXIT chart associated with degree-$i$ depth-one tree[@ardakani04EXIT] as in Fig. \[fig:one\_tree\]. The initial probability is calculated by $p_0 = {\bf{P_e}}(D_{ch})$, where $\bf{P_e}$ denotes the error probability and $D_{ch}$ is the conditional probability distribution function (pdf) of the message from channel. Then the number of decoding iterations is given[@Wei05PctB; @smith10LdpcPCT] by $$N = \int_{p_t}^{p_0}{\left( p\ln \left( \frac{p}{\sum_{i \geqslant 2} \lambda_i f_i(p)} \right)\right)^{-1}dp},$$ where $p_t$ is the target error probability.
![Depth-one decoding tree.[]{data-label="fig:one_tree"}](chen_CL2011-2467.R1_fig1.eps){width="1.6in"}
Symmetric Conditions for Nonbinary LDPC Codes
---------------------------------------------
A log-domain FFT-QSPA (The fast Fourier transform q-ary sum-product algorithm) decoder is used in[@ge09NLDPC]. The log likelihood ratio vectors (LLRV) are fed into the decoder. The $i$th element of LLRV can be calculated as $l_i = \ln \frac{p_0}{p_i}$, where $p_i = {\bf Pr}({\bf y}|x=i)$, and ${\bf y}$ is the channel observation of variable node $x$. In[@ge09NLDPC], a generalized symmetric condition for $q$-ary-input symmetric-output channel ($q$-ary PSK modulated channels for prime $q$ and binary-modulated channels for $q=2^p$) is given by $$\nonumber
{\bf Pr}({\bf y}|x=a) = {\bf Pr}({{\mathcal{I}}}[a]{\bf y}|x=0), \forall a \in GF(q),$$ where ${{\mathcal{I}}}[a]$ is a $(q-1) \times (q-1)$ diagonal matrix with the $i$-th diagonal entry $r^{i\otimes a}$, $i = 1,2,...,q-1$, $r$ is the primitive root of the corresponding field, and $\otimes$ is the mod-$q$ multiplication. Further, it’s proven that, under this symmetric condition, the error performance of an LDPC code is independent of the transmitted codeword. So, analysis (EXIT chart) for $q$-ary LDPC codes based on all-zero codeword will suffice for the decoder.
Complexity-Optimized Nonbinary LDPC Codes
=========================================
This section proposes a irregular nonbinary EXIT chart based on an upper bound of message error probability. Further, a complexity optimization algorithm based on the EXIT chart is put forward to design low decoding complexity $q$-ary LDPC codes which are examined by both hard and soft decision decoders.
Irregular EXIT Chart for Nonbinary LDPC Codes {#section_EXIT}
---------------------------------------------
Assuming all zero codewords are sent, a well designed EXIT chart can be adopted to construct $q$-ary LDPC codes with optimized PCT over $q$-ary-input symmetric channel. Based on symmetric conditions, EXIT chart is first developed for Turbo codes as pictorial demonstration of iterative decoding process[@Brink99EXIT]. Later, a more accurate approximation is applied to binary LDPC to design good performance code ensemble according to their degree distributions[@ardakani04EXIT]. When it is applied to $q$-ary LDPC, [@ge09NLDPC] generalizes the symmetric condition, gives a Gaussian approximation to non-binary density evolution, and shows that, by using a channel adapter, static channel can be forced to be symmetric. A more systematic approach to design $q$-ary LDPC codes is given in[@Bennatan06NonbinaryMem], where they use coset codes to symmetrize the memoryless channels, and design good coset $GF(q)$ LDPC codes too. An EXIT chart based on new mutual information metric is given in[@Byers05ExitNonbinary] using a Gaussian mixture distribution which is less computationally intensive. The EXIT chart for $q$-ary LDPC is also studied in[@Rathi05DeNonbinary].
These methods can well predict the performance thresholds of LDPC codes with infinite block length. But the decoding complexity can not be optimized based on these design procedure. So, instead of giving method for predicting the precise performance of $q$-ary LDPC codes, we present a complexity optimization algorithm by using Gallager’s formula which is an upper bound of message error probability for FFT-QSPA decoder and can be also used as an extended analysis for Gallager decoding algorithm b (Gal-b) [@Gallager63low-densityparity-check].
The reasons why we adopt the Gallager’s formula to extend the PCT analysis to non-binary LDPC codes are as follows. *(i)*. This formula has been shown of great potential in designing excellent irregular LDPC codes for soft decision decoders in [@Luby01IrrLdpc], where they show that given the degree distributions, one can construct decoding graphs for any number of nodes with the correct edge fractions, under *belief propagation* algorithm, by using Gallager’s formula. The designed results can be directly applied to soft decision decoders. *(ii)* For practical considerations, this formula simplifies the analysis of convergence behavior of $q$-ary LDPC codes and makes the design of complexity-optimized $q$-ary LDPC codes possible. From this formula[@Gallager63low-densityparity-check], it is known that for a degree-$k$ check node, the probability of either no errors or of the summation of errors is equal to 0 (mod-$q$) in one of the $k-1$ parity check sets is $$\label{Q_out}
Q_{{out},k} = \frac{1+(q-1)(1-\frac{qp_{{in}}}{q-1})^{k-1}}{q},$$ where $p_{in}$ is the input error probability of messages from a variable node to a check node. For an irregular-check-degree depth-one tree, define $Q_{out}$ as $$\begin{aligned}
\label{Q_out_all}
Q_{{out}} =\sum \limits_{k \geqslant 2}\rho_k Q_{{out},k}.
$$
For a variable with degree $d_v = i$, the output message error probability $p_{i,out} = f_i(p_{in})$, where $f_i$ is the uni-parametric element EXIT chart given by $$\begin{gathered}
\label{f_out}
f_i(p_{in}) = p_0 - p_0\sum \limits_{l=l_0} \limits^{i-1} {{i-1} \choose {l}} Q_{out}^{l}(1-Q_{out})^{i-1-l}+\\
(1-p_0)(q-1)\sum \limits_{l=l_0} \limits^{i-1} {{i-1} \choose {l}}\left ( \frac{1-Q_{out}}{q-1} \right )^{l} \left ( 1- \frac{1-Q_{out}}{q-1} \right )^{i-1-l},\end{gathered}$$ where $p_0$ is the initial error probability from the channel. The second additive term in Eq. is the probability of message received in error in the variable and then corrected, while the third additive term is the probability that $l_0$ check nodes agree on the same error message. $l_0$ is the smallest integer chosen to minimize $p_{out}$, subject to $l_0>(i-1)/2$, for which $$\label{choose_B}
\frac{1-p_0}{p_0}\leqslant \frac{Q_{out}^{l_0}(q-1)^{i-2}}{(1-Q_{out})^{(2l_0 +1-i)}(q-2-Q_{out})^{(i-1-l_0)}}.$$
From[@ge09NLDPC; @smith10LdpcPCT], it is known that the overall decoding complexity is proportional to $NE$, where $N$ is the number of decoding iterations and $E$ is the number of edges in Tanner graph. Since each codeword encodes $Rn\log q$ information bits, the decoding complexity per information bit is $O(\frac{NE}{Rn\log q})$. Then the decoding complexity can be formulated as $$\label{decod_compl}
K = \frac{NE}{Rn\log q} = \frac{N(1-R)}{R\log q \sum_{i \geqslant 2} \frac{\rho_i}{i}}.$$ So, complexity optimization is equivalent to finding the unique local minimum of $K$ in general, because the convexity can not be always guaranteed[@Wei05PctB; @smith10LdpcPCT].
A General Method for Constructing Irregular $q$-ary LDPC Codes with Optimized PCT {#section_PCT}
---------------------------------------------------------------------------------
The fact that $q$-ary LDPC codes with small mean column weight $\bar{d}_v$ can outperform other LDPC codes, has been known for years[@ge09NLDPC; @Davey98montecarlo]. For large field order, average columns weight $\bar{d}_v$ of the best $q$-ary LDPC[@ge09NLDPC; @Davey98montecarlo] will tend to $2$, which is also called $q$-ary cycle LDPC codes[@Huang10CycleNLDPC]. Irregular $q$-ary LDPC codes with small $\bar{d}_v$, [*i.e.* ]{}$2<\bar{d}_v<3$, can outperform other LDPC codes[@ge09NLDPC; @Davey98montecarlo]. In this manuscript, we do not restrict the variable degree to only two small numbers as in [@jie08Underwater], hoping to find better codes.
Considering irregular $q$-ary LDPC codes with degree distribution $\lambda(x)$ and $\rho(x)$, we set a target rate $R_0$, $R\geqslant R_0$. Then the optimization algorithm in[@smith10LdpcPCT] is modified as $$\begin{aligned}
\label{PCT}
\nonumber
\text{minimize} && \frac{1-R_0}{R_0\log(q) \sum \frac{\rho_i}{i}} \int_{p_t}^{p_0}{\left( p\ln \left( \frac{p}{\sum \lambda_i f_i(p)} \right)\right)^{-1}dp}.\\
\nonumber
\text{subject to}
&& p < \sum \lambda_i f_i(p);\\ \nonumber
&& \sum_i(\lambda_i/i) \geqslant \frac{\sum_i(\rho_i/i)}{1-R_0};\\
\nonumber
&& \lambda_i\geqslant 0,\rho_i \geqslant 0;\\
\nonumber
&& \sum_{i} \lambda_i = \sum_{i} \rho_i = 1;\\
&& \|{\lambda}-\bar{ \lambda}\|_\infty<\zeta_1, \|\rho-\bar{\rho}\|_\infty<\zeta_2.
$$ where $\bar{\lambda}$ and $\bar{\rho}$ can be initialized as the threshold-optimized LDPC codes suggest [@ge09NLDPC; @smith10LdpcPCT]. $R_0$ is fixed which is lower than the rate of the code $(\bar{\lambda},\bar{\rho})$. $\zeta_1$ and $\zeta_2$ are carefully set to be small values to guarantee finding the unique local maximum [@Wei05PctB; @smith10LdpcPCT]. The constraint $p<\sum \lambda_i f_i(p)$ is substantial for which this optimization algorithm is valid.
![Performance comparison according to BER.[]{data-label="fig_pct_ber"}](chen_CL2011-2467.R1_fig2.eps){width="3in"}
Note that, this *irregular* algorithm is different to the *quasi-regular* optimization in[@smith10LdpcPCT] in the sense that the proposed algorithm updates $\bar{\lambda}$ and $\bar{\rho}$ by the recent optimal values in each iteration through which we obtain the convergence-optimized $q$-ary LDPC codes.. More importantly, a mild condition, [*i.e.* ]{}$\{\lambda_i|f(p_{{in}})\geqslant e^2 p_{{in}} \}$, is given in[@Wei05PctB; @smith10LdpcPCT], under which $f(p_{{in}})$ is a convex function of $\lambda_i$. The complexity-optimized $q$-ary LDPC codes, resulting from our irregular algorithm, has a little lower threshold than the original one, but converges faster at higher SNR regime. We take the $q$-ary LDPC codes with variable degrees restricted to $2$ and $3$ [@Huang10CycleNLDPC; @jie08Underwater] for example. If the message error probability is sufficiently small, then $Q_{out,k}\approx 1-(k-1)p_{in}$. From Eq. , calculate $Q_{out}\approx 1-(\tau_1 + \rho_{\tau_2}-1)p_{in}$, and $Q^2_{out}\approx 1-2(\tau_1+\rho_{\tau_2}-1)p_{in}$, where $\tau_1$ and $\tau_2$ is the check degrees. In addition, the element EXIT charts of the designed $q$-ary LDPC codes are $$\begin{aligned}
\nonumber
f_2(p_{in}) & = & 1-(2-p_0)Q_{out},\\
\nonumber
f_3(p_{in}) & = & p_0 + \frac{1+p_0}{q-1}(1-2Q_{out}+Q^2_{out})-Q^2_{out}.\end{aligned}$$ Then, we have $$\label{p_out_approx}
f(p_{in})\approx (p_0 - 1) + (2-\lambda_2 p_0)(\tau_1+\rho_{\tau_2}-1)p_{in}.$$
It is easy to verify that Eq. does not always satisfy the convex condition. Numerical simulations nevertheless suggest that, there exists a unique local optimum. In table \[tb\_dv\], we give the minimum average column weight of the parity check matrix, [*i.e.* ]{}$T_{\bar{d}_v}$, in terms of the code rate, such that the optimization algorithm is valid.
Simulation Results {#section_sim}
==================
------------------------- ------------------------------------------------------------------------------------------------ ----------- --------
$(\bar{d}_v,\bar{d}_c)$ $f(p_{in})$ estimated actual
$(2.7,3.75)$ $0.62p_{in}+4.97p^2_{in}-18.24p^3_{in}+27.53p^4_{in}-23.28p^5_{in}+10.75p^6_{in}-2.09p^7_{in}$ $21.1$ $22$
$(2.7,3.6)$ $0.59p_{in}+5.3p^2_{in}-16.25p^3_{in}+23.20p^4_{in}-18.20p^5_{in}+8.01p^6_{in}-1.45p^7_{in}$ $19.04$ $18$
$(2.65,3.53)$ $0.69p_{in}+4.71p^2_{in}-14.46p^3_{in}+20.11p^4_{in}-15.53p^5_{in}+6.48p^6_{in}-1.13p^7_{in}$ $26.67$ $26$
$(2.68,3.94)$ $0.70p_{in}+5.79p^2_{in}-20.19p^3_{in}+32.23p^4_{in}-28.81p^5_{in}+14.11p^6_{in}-2.93p^7_{in}$ $28.81$ $28$
$(2.65,3.68)$ $0.72p_{in}+5.00p^2_{in}-16.32p^3_{in}+24.15p^4_{in}-19.92p^5_{in}+8.95p^6_{in}-1.69p^7_{in}$ $30.97$ $$31$$
------------------------- ------------------------------------------------------------------------------------------------ ----------- --------
\[tb\_iter\]
The $q$-ary LDPC codes in the manuscript are construct by the modified progressive edge-growth (PEG) algorithm. If the variable degrees are restricted to $2$ and $3$, We estimate the number of iterations when the message error probability is reduced to $10^{-6}$ from $10^{-2}$. Table \[tb\_iter\] gives the estimated and actual number of iterations according to different $\bar{d}_v$ and $\bar{d}_c$ for Gal-b. Table \[tb\_dv\] gives the required smallest $\bar{d}_v$, [*i.e.* ]{}$T_{\bar{d}_v}$, for different code rate $R$, such that the proposed optimization algorithm is valid for the soft decision decoder.
Then, we show how to reduce the decoding complexity of a given code. Considering the threshold optimized $4$-ary LDPC codes with block length 30000 bits reported in [@ge09NLDPC; @smith10LdpcPCT], characterized by $\lambda(x)=0.249009x+0.200042x^2+0.02177703x^5+0.161403x^6+0.0489424x^8+
0.0381342x^{16}+0.0874772x^{18}+0.0154621x^{19}+0.177761x^{49}$ and $\rho(x)=0.439929x^7+0.560007x^8$, the complexity optimized $4$-ary LDPC code characterized by $\lambda(x)=0.5503x+0.0297x^3+0.1304x^4+0.2003x^{15}+0.0893x^{20}$ and $\rho(x)=0.2998x^3+0.7002x^4$. We give the bit error rate (BER) and word error rate (WER) in Fig. \[fig\_pct\_ber\] and Fig. \[fig\_pct\_wer\] by calculating the average error rate from 100 times experiments. We expect that the complexity optimized code will reach a BER of $10^{-7}$ faster at a smaller number of iterations, while maintaining the excellent performance as the original one. Let $C_1(N)$ and $C_2(N)$ be the original and optimized codes, respectively, where $N$ is the number of iterations. Fig. \[fig\_pct\_ber\] shows that the optimized code outperforms the original one with faster convergence rate at a small $N$. $C_2(19)$ even converges faster than $C_1(27)$. The convergence process has been accelerated by $30\%$ regarding the number of decoding iterations.
----------------- ------ ------ ------ ------ ------ ------
$R$ 1/6 1/5 1/4 1/3 1/2 2/3
$T_{\bar{d}_v}$ 2.37 2.40 2.48 2.56 2.70 2.81
----------------- ------ ------ ------ ------ ------ ------
: The smallest $\bar{d}_v$ required for different rates
\[tb\_dv\]
![Performance comparison according to WER.[]{data-label="fig_pct_wer"}](chen_CL2011-2467.R1_fig3.eps){width="3in"}
Conclusion and Discussions
==========================
The proposed PCT algorithm is used to design irregular nonbinary LDPC codes with optimized decoding complexity. However, the encoding complexity is not optimized during the design procedure. A future work of this manuscript is to construct structured nonbinary LDPC codes that can achieve optimized decoding complexity and optimized encoding complexity at the same time. In addition, upper bounds of message error probability are used to analyze the performance of nonbinary LDPC codes, which results in an approximated PCT analysis for the soft decision decoder. In order to achieve faster convergence performance, we need to construct more accurate PCT algorithms. Another future work of this manuscript is to find more accurate uni-parametric representation of the decoding trajectory for the nonbinary soft decision decoders.
[^1]: Manuscript received December 6th, 2011; revised February 1st, 2012, accepted February 8th, 2012. The associate editor coordinating the review of this paper and approving it for publication was M. Lentmaier.
[^2]: The authors are with Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, and SKL for ISN, Xidian University, China. e-mail: {yuyang83;wenchen}@sjtu.edu.cn.
[^3]: This work is supported by NSFC \#60972031, by national 973 project \#2012CB316106 and \#2009CB824900, by NSFC \#61161130529, by national key laboratory project \#ISN11-01.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $C$ be a smooth curve of genus $g$. For each positive integer $r$ the $r$-gonality $d_r(C)$ of $C$ is the minimal integer $t$ such that there is $L\in \mbox{Pic}^t(C)$ with $h^0(C,L) =r+1$. In this paper for all $g\ge 40805$ we construct several examples of smooth curves $C$ of genus $g$ with $d_3(C)/3< d_4(C)/4$, i.e. for which a slope inequality fails.'
address: |
Dept. of Mathematics\
University of Trento\
38123 Povo (TN), Italy
author:
- 'E. Ballico'
title: 'On the gonality sequence of smooth curves: normalizations of singular curves in a quadric surface'
---
[^1]
Introduction {#S1}
============
Let $C$ be a smooth and connected projective curve of genus $g\ge 3$. For each integer $r\ge 1$ the $r$-gonality $d_r(C)$ of $C$ is the minimal integer $d$ such that there is a degree $d$ line bundle $L$ on $C$ with $h^0(C,L) \ge r+1$ ([@lm]). The sequence $\{d_r(C)\}_{r\ge 1}$ is called the [*gonality sequence*]{} of $C$. This sequence is important to understand the Brill-Noether theory of vector bundles on $C$ ([@ln1], [@ln2], [@ms]). See [@lm], §3, for general properties of this sequence for an arbitrary curve $C$. For most curves we have $$\label{eqa1}
\frac{d_r(C)}{r} \ge \frac{d_{r+1}(C)}{r+1}$$ for all $r\ge 2$ ([@lm], Proposition 4.1). In [@lm] H. Lange and G. Martens introduced the following notion. The curve $C$ is said to satisfy the [*slope inequality*]{} if (\[eqa1\]) is satisfied for all $r\ge 2$. Since $d_2(C) \le 2d_1(C)$ for all $C$, the slope inequality is always satisfied for $r=1$. Hence $C$ does not satisfy the slope inequality if and only if there is at least one integer $r \ge 2$ for which (\[eqa1\]) fails. Many different examples of such curves are constructed in [@lm]. In this paper we look at the case $r=3$ of (\[eqa1\]) and prove the following result.
\[i1\] Fix an integer $g\ge 40805$. Then there exists a smooth curve $C$ of genus $g$ such that $d_3(C)/3< d_4(C)/4$.
The curves $C$ used to prove Theorem \[i1\] are the normalization of nodal curves $Y$ contained in a smooth quadric surface $Q \subset \mathbb {P}^3$. These families of examples are an extension of [@lm], Example 4.12. We prove that for the normalization of many of them the rational number $d_4(C)/4 -d_3(C)/3$ is rather large (Propositions \[g5\], \[h1\] and Corollary \[h2\]). As an obvious consequence we get the following statement.
\[i2\] There is a sequence $\{C_g\}_{g\ge 3}$ of smooth curves such that $C_g$ has genus $g$, $$\lim _{g\to \infty} \frac{d_4(C_g)}{d_3(C_g)}
= 3/2 \ \mbox{and} \ \lim _{g\to \infty} \frac{d_4(C_g)/4 -d_3(C_g)/3}{\sqrt{g}} = 1/12.$$
To prove Theorems \[i1\] and \[i2\] we need to study the cohomology of certain finite subsets $S\cup B$ of the smooth quadric surface $Q$. These preliminary lemmas are proved in section \[S2\]. In section \[S3\] we use these lemmas in the following way. Fix an integral nodal curve $Y\in \vert \mathcal {O}_Q(a,a+m)\vert$ and set $S:= \mbox{Sing}(Y)$. Let $C$ be the normalization of $Y$. Fix any $L\in \mbox{Pic}^z(C)$ evincing $d_4(C)$. To a general divisor $A\in \vert L\vert$ we associate a set $B\subset Q\setminus S$ such that $\sharp (B)=z$ and $h^1(Q,\mathcal {I}_{S\cup B}(a-2,a+m-2)) >0$. The lemmas proved in section \[S2\] show that $z\ge 3a-15$ for a general $S$ and $m$ not too large, while obviously $d_3(C) \le 2a$. Taking only smooth curves inside $Q$ we only get a sequence of genera, enough to prove the weaker form of Theorem \[i2\] with “ $\limsup$ ” instead of “ $\lim$ ” (as implicit in [@lm], Example 4.12).
For all integers $r\ge 2$ and $g\ge 2$ let $\alpha (r,g)$ be the supremum of all rational numbers $d_{r+1}(C)/d_r(C)$ with $C$ a smooth curve of genus $g$.
\[iq1\] Compute $\alpha ' (r):= \liminf _{g\to \infty } \alpha (r,g)$ and $\alpha ''(r):= \limsup _{g\to \infty} \alpha (r,g)$. Is $\alpha ''(3) = 3/2$ ?
We work over an algebraically closed base field with characteristic zero.
I want to thank the referee for several extremely useful remarks (in this version Step ($\diamond$) of the proof of Proposition \[g5\] is due to the referee).
Preliminaries {#S2}
=============
Let $Q\subset \mathbb {P}^3$ be a smooth quadric surface. For any coherent sheaf $\mathcal {F}$ on $Q$ and any $i\in \mathbb {N}$ set $H^i(\mathcal {F}):= H^i(Q,\mathcal {F})$ and $h^i(\mathcal {F}):= \dim (H^i(\mathcal {F}))$. For all $(a,b)\in \mathbb {Z}^{2}$ let $\mathcal {O}_Q(a,b)$ denote the line bundle on $Q$ with bidegree $(a,b)$. We have $h^0(\mathcal {O}_Q(a,b))=(a+1)(b+1)$ and $h^1(\mathcal {O}_Q(a,b))=0$ if $(a,b)\in \mathbb {N}^2$, while $h^0(\mathcal {O}_Q(a,b)) =0$ if either $a<0$ or $b<0$. If $(a,b)\in \mathbb {N}^2$ and $T\in \vert \mathcal {O}_Q(a,b)\vert$, then we say that $T$ has type $(a,b)$. The lines contained in $Q$ are the curves $D\subset Q$ with either type $(1,0)$ or type $(0,1)$.
\[g5\] Fix a integer $x>0$ and a general $S\subset Q$ such that $\sharp (S)=x$. Since $h^0(\mathcal {O}_Q(1,1)) =4$, $h^0(\mathcal {O}_Q(2,1)) =h^0(\mathcal {O}_Q(1,2))=6$ and $S$ is general, we have $\sharp (S\cap T_1)\le 3$ for every $T_1\in \vert \mathcal {O}_Q(1,1)\vert$, $\sharp (S\cap T_2)\le 5$ for every $T_2\in \vert \mathcal {O}_Q(2,1)\vert$ and $\sharp (S\cap T_3)\le 5$ for every $T_3\in \vert \mathcal {O}_Q(1,2)\vert$.
\[e2.0\] Fix integers $u>0$ and $v >0$. Fix a reduced $D\in \vert \mathcal {O}_Q(2,1)\vert$ and a set $S\subset D$ such that $\sharp (S) \le 2v+u+1$, $\sharp (S\cap T) \le 1$ for every line $T\subset D$ (if any) and $\sharp (S\cap T) \le u+v+1$ for every component $T$ of type $(1,1)$ of $D$ (if any). Then $h^1(D,\mathcal {I}_{S,D}(u,v))=0$.
First assume that $D$ is irreducible. Since $D \cong \mathbb {P}^1$ and $\deg (\mathcal {O}_D(u,v)) = 2v+u \ge \sharp (S)-1$, we have $h^1(D,\mathcal {I}_{S,D}(u,v))=0$. Now assume that $D$ has an irreducible component $A$ of type $(1,1)$ and write $D = A\cup T$ with $T$ a line of type $(1,0)$. Since $\sharp (A\cap S) \le u+v+1$, $A\cong \mathbb {P}^1$ and $\deg (\mathcal {O}_A(u,v)) = u+v$, we have $h^1(A,\mathcal {I}_{S\cap A,A}(u,v))=0$. Since $\sharp (T\cap S)\le 1$ and $S\subset D$, we have $\sharp (S\setminus S\cap A) \le 1$. Since $\deg (T\cap A)=1$, we also have $h^1(T,\mathcal {I}_{(S\setminus S\cap A)\cup (A\cap T),T}(u,v))=0$. Hence a Mayer-Vietoris exact sequence gives $h^1(D,\mathcal {I}_{S,D}(u,v))=0$. Now assume that $D$ is the union of $3$ lines $T_1,T_2,T_3$ with $T_2$ of type $(0,1)$. Since $\sharp (T_i\cap S)\le 1$ for all $i$, we have $h^1(T_1,\mathcal {I}_{S\cap T_1}(u,v))=0$, $h^1(T_2,\mathcal {I}_{S\cap T_2\setminus
S\cap T_1\cap T_2,T_2}(u-1,v))=0$ and $h^1(T_3,\mathcal {I}_{S\setminus (T_1\cup T_2)\cap S,T_3}(u-1,v-1))=0$. We use two Mayer-Vietoris exact sequences and get first $h^1(T_1\cup T_2,\mathcal {I}_{S\cap (T_1\cup T_2),T_1\cup T_2}(u,v)) =0$ and then $h^1(D,\mathcal {I}_{S,D}(u,v))=0$.
\[e4\] Fix integers $v \ge u\ge 9$ and set $\alpha := \lfloor u/3\rfloor$. Fix a finite set $E\subset Q$ such that $\sharp (E) \le v-u +10\alpha $, no $2$ of the points of $E$ are contained in a line of $Q$, at most $2u+1$ of the points of $E$ are contained in a curve of type $(1,1)$, at most $3u+1$ of the points of $E$ are contained in a curve of type $(2,1)$ and at most $3u-4$ of the points of $E$ are contained in a curve of type $(1,2)$. Then $h^1(\mathcal {I}_E(u,v)) =0$.
Notice that $\alpha \ge3$. Set $\beta := u-3\alpha$
\(i) In this step we assume $v=u$. Set $E_0:=E$. Take any $A_1\in
\vert
\mathcal {O}_Q(2,1)\vert$ such that $a_1:= \sharp (E_0\cap A_1)$ is maximal. Set $F_1:= A_1\cap E_0$ and $E_{1,0}:= E_0\setminus F_1$. Let $D_1\in \vert \mathcal \mathcal {O}_Q(1,2)\vert$ be a curve such that $b_1:= \sharp (E_{1,0}\cap D_1)$ is maximal. Set $G_1:= E_{1,0}\cap D_1$ and $E_1:= E_{1,0}\setminus G_1$. For each $i\in \{2,\dots ,\alpha \}$ we define recursively the integers $a_i$ and $ b_i$, the curves $A_i\in \vert \mathcal {O}_Q(2,1)\vert$, $D_i\in \vert \mathcal {O}_Q(1,2)\vert$ and the sets $F_i$, $E_{i,0}$, $G_i$, $E_i$ in the following way. Take $A_i\in \vert \mathcal {O}_Q(2,1)\vert$ such that $a_i:= \sharp (A_i\cap E_{i-1})$ is maximal. Set $F_i:= E_{i-1}\cap A_i$ and $E_{i,0}:= E_{i-1}\setminus F_i$. Take $D_i\in \vert \mathcal {O}_Q(1,2)\vert$ such that $b_i:= \sharp (D_i\cap E_{i,0})$ is maximal and set $G_i:= D_i\cap E_{i,0}$ and $E_i:= E_{i,0}\setminus G_i$. For each $i\in \{1,\dots ,\alpha \}$ we have the exact sequences $$\begin{aligned}
\label{eqe6}
&0 \to \mathcal {I}_{E_{i,0}}(u-3i+1,u-3i+2) \to \mathcal {I}_{E_{i-1}}(u-3i+3,u-3i+3) \notag \\
&\to \mathcal {I}_{F_i,A_i}(u-3i+3,u-3i+3) \to 0\end{aligned}$$ $$\begin{aligned}
\label{eqe7}
&0 \to \mathcal {I}_{E_i}(u-3i,u-3i) \to \mathcal {I}_{E_{i,0}}(u-3i+1,u-3i+2) \notag \\
& \to \mathcal {I}_{G_i,D_i}(u-3i+1,u-3i+2) \to 0\end{aligned}$$ Notice that the sequences $\{a_i\}_{1 \le i \le \alpha}$ and $\{b_i\}_{1\le i \le \alpha }$ are non-increasing. If $a_i \le 4$, then $E_{i,0} = \emptyset$, because $h^0(Q,\mathcal {O}_Q(2,1)) =6$. If $b_i \le 4$, then $E_i=\emptyset$, because $h^0(Q,\mathcal {O}_Q(1,2)) =6$. Since $\sharp (E) \le 10\alpha$, we get $E_\alpha =\emptyset$.
For every $i\in \{1,\dots ,\alpha \}$ we have $a_i \le 3u-9i+10$.
Assume $a_i \ge 3u-9i+11$. Since at most $3u+1$ of the points of $E$ are contained in a curve of type $(2,1)$, we have $i\ge 2$. Since the sequences $\{a_n\}$, $\{b_n\}$ are non-increasing and $b_j \ge 5$ if $a_{j+1}>0$, we get $\sharp (E) \ge i(3u-9i+11) + 5(i-1) = i(3u +16 -9i) -5$. If $i=2$, then we get $\sharp (E) \ge 6u-9 > 10\alpha $, a contradiction. For any $t\in \mathbb {R}$ set $\phi (t) = t(3u+16-9t)-5$. The function $\phi$ is increasing in the interval $[0,(3u+16)/18]$ and decreasing if $t\ge (3t+16)/18$. Since $\sharp (E) < \phi (2)$ and $\phi (u/3) =16u/3-5 > \sharp (E)$, we get a contradiction.
For each $i \in \{1,\dots ,\alpha \}$ we have $h^1(A_i,\mathcal {I}_{F_i,A_i}(u-3i+3,u-3i+3))=0$.
By Claim 1 we have $\sharp (F_i) \le 3u-9i+10$. If $A_i$ is irreducible, then Claim 2 is true (e.g. by Lemma \[e2.0\]). If $A_i$ is the union of $3$ lines, then $a_i\le 3$ and Claim 2 is true (Lemma \[e2.0\]). Now assume $A_i = T\cup D$ with $T$ a smooth conic and $D$ of type $(1,0)$. By Lemma \[e2.0\] Claim 2 is true if $\sharp (F_i\cap T) \le 2u-6i+7$. Assume $\sharp (F_i\cap T) \ge 2u-6i+8$. Our assumptions on $E$ imply $i\ge 2$. We have $a_i \ge \sharp (A_i\cap T) \ge 2u-6i+8$. Since $b_j \ge 5$ if $E_{j+1} \ne \emptyset$, we get $\sharp (E) \ge i(2u-6i+8) +5(i-1) = i(2u+13-6i)-5$. If $i=2$, then $\sharp (E) \ge 4u-3 > 10\alpha$, a contradiction. For every $t\in \mathbb {R}$ set $\psi (t):=
t(2u+13-6t) -5$. Since the function $\psi$ is increasing in the interval $[0,(2u+13)/12]$ and decreasing for $t>(2u+13)/12$, $\sharp (E) < \psi (2)$ and $\psi (\alpha ) \ge 13\alpha -5 > 10\alpha$, we get a contradiction.
For each $i\in \{1,\dots ,\alpha \}$ we have $b_i\le 3u-9i+5$.
Assume $b_i \ge 3u-9i+6$. Since $G_i\subseteq E$, our assumptions on $E$ imply $i \ge 2$. Since $b_i>0$, we have $a_j\ge 5$ for all $j\le i$. Hence $\sharp (E) \ge 5i + i(3u-9i+6) = i(3u+11 -9i)$. Set $\tau (t) = t(3u+11-9t)$. The function $\tau (t)$ is increasing in the interval $[0,(3u+11)/18]$ and decreasing if $t>(3u+11)/18$. Since $\tau (2) = 6u-14 > \sharp (E)$ and $\tau (\alpha ) \ge 11\alpha > \sharp (E)$, we get a contradiction.
For each $i \in \{1,\dots ,\alpha \}$ we have $h^1(D_i,\mathcal {I}_{G_i,D_i}(u-3i+1,u-3i+2))=0$.
We apply Lemma \[e2.0\] taking $D_i\in \vert
\mathcal {O}_Q(1,2)\vert$ instead of an element of $\vert \mathcal {O}_Q(2,1)\vert$. If $D_i$ is irreducible, then Claim 4 follows from Claim 3, because $D_i\cong \mathbb {P}^1$ and $\deg (\mathcal {O}_{D_i}(u-3i+1,u-3i+2))= 2(u-3i+1)+(u-3i+2)$. If $D_i$ is a union of $3$ lines, then $b_i\le 3$; in this case we just use that $u-3i+1>0$ and $u-3i+2 > 0$. Now assume $D_i = T\cup D$ with $T$ a smooth conic and $D$ a line. It is sufficient to have $\sharp (T\cap G_i) \le 2u-6i+4$ (Lemma \[e2.0\] for curves of type $(1,2)$). Assume $\sharp (T\cap G_i) \ge 2u-6i+5$. Since $T\cup I \in \vert \mathcal {O}_Q(2,1)\vert$ for all $I\in \vert \mathcal {O}_Q(1,0)\vert$ and $\sharp (T\cap E_{i-1}) \ge 2u-6i+5$, we get $a_i \ge 2u-6i+6$. Hence $\sharp (E) \ge i(4u-12i+11)$. For any $t\in \mathbb {R}$ set $\eta (t) := t(4u-12t+11)$. The function $\eta (t)$ is increasing in the interval $0 \le t \le (4u+11)/24$ and decreasing if $t> (4u+11)/24$. Since $\eta (1) = 4u-1 > 10\alpha$ and $\eta (\alpha) = \alpha (4u-12\alpha +11)\ge 11\alpha$, we get a contradiction.
By Claims 2 and 4 and the exact sequences (\[eqe6\]) and (\[eqe7\]) we get $h^1(\mathcal {I}_E(u,u)) \le h^1(\mathcal {I}_{E_{\alpha }}(\beta ,\beta))$. Since $E_{\alpha } =\emptyset$, we have $h^1(\mathcal {I}_{E_{\alpha }}(\beta ,\beta ))=0$.
\(ii) Now assume $v>u$. Write $E = F\sqcup F'$ with $\sharp (F') = \min \{\sharp (E), v-u\}$. Since $\sharp (F')\le v-u$ and no two points of $E$ are contained in an element of $\vert \mathcal {O}_Q(0,1)\vert$, there is a union $T\subset Q$ of $v-u$ disjoint elements of $\vert \mathcal {O}_Q(0,1)\vert$ such that $F'\subset T$ and $T\cap F=\emptyset$. We have an exact sequence $$\label{eqe2}
0\to \mathcal {I}_F(u,u)\to \mathcal {I}_E(u,v)\to \mathcal {I}_{F',T}(u,v)\to 0$$ Since $T$ is a disjoint union of $v-u$ lines, each of them containing at most one point of $F'$, we have $h^1(T,\mathcal {I}_{F',T}(u,v))=0$. Step (i) gives $h^1(Q,\mathcal
{I}_F(u,u))=0$. Hence (\[eqe2\]) gives $h^1(Q,\mathcal {I}_E(u,v))=0$.
\[g4\] Fix integers $x, \alpha, \beta ,z$ such that $\alpha \ge 3$, $\beta \ge 2$, $0 \le x\le (\beta +1)^2$ and $0 \le z\le
10\alpha$. Set $u:= 3\alpha +\beta$. Fix a general $S\subset Q$ such that $\sharp (S)=x$. Fix $B\subset Q\setminus S$ such that $\sharp (B)
=z$, no line of $Q$ contains $2$ points of $S\cup B$, $\sharp (B\cap T_1)\le 2u-2$ for every $T_1\in \vert \mathcal {O}_Q(1,1)\vert$, $\sharp (B\cap T_2)\le 3u-4$ for every $T_2\in \vert \mathcal {O}_Q(2,1)\vert$ and $\sharp (B\cap T_3)\le 3u-9$ for every $T_3\in \vert \mathcal {O}_Q(1,2)\vert$. Then $h^1(\mathcal {I}_{S\cup B}(u,u))
= 0$.
Set $E_0:= S\cup B$ and $B_0:= B$. Since $S$ is general, we have $\sharp (S\cap T_1)\le 3$ for every $T_1\in \vert \mathcal {O}_Q(1,1)\vert$, $\sharp (S\cap T_2)\le 5$ for every $T_2\in \vert \mathcal {O}_Q(2,1)\vert$ and $\sharp (S\cap T_3)\le 5$ for every $T_3\in \vert \mathcal {O}_Q(1,2)\vert$ (Remark \[g5\]). Hence $\sharp
(E_0\cap T_1)\le 2u+1$ for every $T_1\in \vert \mathcal {O}_Q(1,1)\vert$, $\sharp (E_0\cap T_2)\le 3u+1$ for every element of $\vert \mathcal {O}_Q(2,1)\vert$ and $\sharp (E_0\cap T_3)\le 3u-4$ for every element of $\vert \mathcal {O}_Q(1,2)\vert$. Take any $A_1\in
\vert
\mathcal {O}_Q(2,1)\vert$ such that $a_1:= \sharp (B_0\cap A_1)$ is maximal. Set $F_1:= A_1\cap E_0$, $B'_1:= A_1\cap B_0$, $E_{1,0}:= E_0\setminus F_1$ and $B_{1,0}:= B_0\setminus B'_1$. Let $D_1\in \vert \mathcal {O}_Q(1,2)\vert$ be a curve such that $b_1:= \sharp (B_{1,0}\cap D_1)$ is maximal. Set $G_1:= E_{1,0}\cap D_1$, $B''_1:= B_{1,0}\cap D_1$, $B_1:= B_{1,0}\setminus B''_1$ and $E_1:= E_{1,0}\setminus G_1$. For each $i\in \{2,\dots ,\alpha \}$ we define recursively the integers $a_i$, $ b_i$, the curves $A_i\in \vert \mathcal
{O}_Q(2,1)\vert$, $D_i\in \vert \mathcal {O}_Q(1,2)\vert$ and the sets $F_i$, $E_{i,0}$, $G_i$, $B'_i$, $B_{i,0}$, $B_i$, $B''_i$, $E_i$ in the following way. Take $A_i\in \vert \mathcal {O}_Q(2,1)\vert$ such that $a_i:= \sharp (A_i\cap B_{i-1})$ is maximal. Set $F_i:= E_{i-1}\cap A_i$, $B'_i:= A_i\cap B_{i-1}$, $B_{i,0}:= B_{i-1} \setminus B'_i$ and $E_{i,0}:= E_{i-1}\setminus F_i$. Take $D_i\in \vert \mathcal {O}_Q(1,2)\vert$ such that $b_i:= \sharp (D_i\cap B_{i,0})$ is maximal and set $G_i:= E_{i,0}\cap D_i$, $E_i:= E_{i,0}\setminus G_i$, $B''_i:= B_{i,0}\cap D_i$ and $B_i:= B_{i,0}\setminus B''_i$. For each $i\in \{1,\dots ,\alpha \}$ we have the exact sequences (\[eqe6\]) and (\[eqe7\]). Notice that the sequences $\{a_i\}_{1 \le i \le \alpha}$ and $\{b_i\}_{1\le i \le \alpha }$ are non-increasing. If $a_i \le 4$, then $E_{i,0} = \emptyset$, because $h^0(Q,\mathcal {O}_Q(2,1)) =6$. If $b_i \le 4$, then $E_i=\emptyset$, because $h^0(Q,\mathcal {O}_Q(1,2)) =6$.
For every $i\in \{1,\dots ,\alpha \}$ we have $a_i \le 3u-9i+5$ and $\sharp (A_i\cap E_{i-1})\le
3u-9i+10$.
Since $\sharp (A_i\cap S)\le 5$, it is sufficient to prove the inequality $a_i\le 3u-9i+5$. Assume $a_i \ge 3u-9i+6$. Since at most $3u-4$ of the points of $B$ are contained in a curve of type $(2,1)$, we have $i\ge 2$. Since the sequences $\{a_j\}$ and $\{b_j\}$ are non-increasing and $b_j \ge 5$ if $a_{j+1}>0$, we get $\sharp (B) \ge i(3u-9i+6) + 5(i-1) = i(3u +11 -9i) -5$. If $i=2$, then we get $\sharp (B) \ge 6u-19 \ge 18\alpha +6\beta -19$, contradicting the assumptions $\sharp (B) \le 10\alpha $, $\alpha \ge 3$ and $\beta >0$. For any $t\in \mathbb {R}$ set $\phi (t) = t(3u+11-9t)-5$. The function $\phi$ is increasing in the interval $[0,(3u+11)/18]$ and decreasing if $t\ge (3t+11)/18$. Since $\sharp (B) < \phi (2)$ and $\phi (\alpha ) =\alpha (3u+11-9\alpha) -5 = \alpha (3\beta +11) -5 > 10\alpha \ge \sharp (B)$, we get a contradiction.
For each $i \in \{1,\dots ,\alpha \}$ we have $h^1(A_i,\mathcal {I}_{F_i,A_i}(u-3i+3,u-3i+3))=0$.
By Claim 1 we have $\sharp (F_i) \le 3u-9i+10$. If $F_i$ is irreducible, then Claim 2 is true (e.g. by Lemma \[e2.0\]). If $A_i$ is the union of $3$ lines, then $\sharp (F_i)\le 3$ and Claim 2 is true (Lemma \[e2.0\]). Now assume $A_i = T\cup D$ with $T$ a smooth conic and $D$ of type $(1,0)$. By Lemma \[e2.0\] Claim 2 is true if $\sharp (E_{i-1}\cap T) \le 2u-6i+7$. Assume $\sharp (E_{i-1}\cap T) \ge 2u-6i+8$. Since $\sharp (S\cap T)\le 3$, we get $\sharp (B_{i-1}\cap T) \ge 2u-6i+5$. Since $\sharp (T_1\cap B)\le 2u-2$ for every $T_1\in \vert \mathcal {O}_Q(1,1)\vert$, we have $i\ge 2$. We have $a_i \ge \sharp (B_{i-1}\cap T) \ge 2u-6i+5$. Since $b_j \ge 5$ if $B_{j+1} \ne \emptyset$, we get $\sharp (B) \ge i(2u-6i+5) +5(i-1) = i(2u+10-6i)-5$. If $i=2$, then $\sharp (B) \ge 4u-9$, a contradiction. For every $t\in \mathbb {R}$ set $\psi (t):=
t(2u+10-6t) -5$. The function $\psi$ is increasing in the interval $[0,(u+5)/6]$ and decreasing for $t>(u+5)/6$. Since $\sharp (B) < \psi (2)$ and $\psi (\alpha ) = \alpha (2u+10-6\alpha ) =\alpha (10+2\beta )> 10\alpha \ge \sharp (B)$, we get a contradiction.
For each $i\in \{1,\dots ,\alpha \}$ we have $b_i\le 3u-9i$ and $\sharp (G_i) \le 3u-9i+5$.
Since $\sharp (S\cap D_i)\le 5$, it is sufficient to prove $b_i\le 3u-9i$. Assume $b_i \ge 3u-9i+1$. Since $G_i\subseteq D_i$, our assumptions on $B$ gives $i \ge 2$. Since $b_i>0$, we have $a_j\ge 5$ for all $j\le i$. Hence $\sharp (B) \ge 5i + i(3u-9i+1) = i(3u+6 -9i)$. Set $\tau (t) = t(3u+6-9t)$. The function $\tau (t)$ in increasing in the interval $[0,(3u+6)/18]$ and decreasing if $t>(3u+6)/18$. Since $\tau (2) = 6u-24
\ge 18\alpha +3\beta -24 > 10\alpha \ge \sharp (B)$ and $\tau (\alpha ) = \alpha (3u+6-9\alpha )
\ge \alpha (6+3\beta ) \ge 12\alpha$, we get a contradiction.
For each $i \in \{1,\dots ,\alpha \}$ we have $h^1(D_i,\mathcal {I}_{G_i,D_i}(u-3i+1,u-3i+2))=0$.
We apply Lemma \[e2.0\] taking $D_i\in \vert
\mathcal {O}_Q(1,2)\vert$ instead of an element of $\vert \mathcal {O}_Q(2,1)\vert$. If $D_i$ is irreducible, then Claim 4 follows from Claim 3, because $D_i\cong \mathbb {P}^1$ and $\deg (\mathcal {O}_{D_i}(u-3i+1,u-3i+2))= 2(u-3i+1)+(u-3i+2)$. If $D_i$ is a union of $3$ lines, then $b_i\le 3$; in this case we just use that $u-3i+1>0$ and $u-3i+2 > 0$. Now assume $D_i = T\cup D$ with $T$ a smooth conic and $D$ a line. It is sufficient to have $\sharp (G_i\cap T) \le 2u-6i+4$ (Lemma \[e2.0\] for curves of type $(1,2)$). Assume $\sharp (T\cap G_i) \ge 2u-6i+5$. Since $S$ is general and $h^0(\mathcal {O}_Q(1,1))=4$, we have $b_i = \sharp (T\cap B''_i) \ge 2u-6i+2$. Since $T\cup I\in \vert \mathcal {O}_Q(2,1)\vert$ for each $I\in \vert \mathcal {O}_Q(1,0)\vert$ and $\sharp (T\cap B_{i-1}) \ge 2u-6i+2$, we have $a_i \ge 2u-6i+3$. Hence $a_j\ge 2u-6i+3$ for all $j\le i$. Hence $\sharp (B) \ge i(4u-12i+5)$. Set $\eta _1(t):= t(4u+5-12t)$. The function $\eta _1(t)$ is increasing if $0 \le t \le (4u+5)/24$ and decreasing if $t>(4u+5)/24$. We have $\eta _1(1) = 4u-7 = 12\alpha +4\beta -7 > 10\alpha$. We have $\eta _1(\alpha ) = \alpha (4\beta +5) \ge 13\alpha > 10\alpha$. Since $\sharp (B) \le 10 \alpha$, we get a contradiction.
By Claims 2 and 4 and the exact sequences (\[eqe6\]) and (\[eqe7\]) we get $h^1(\mathcal {I}_E(u,u)) \le h^1(\mathcal {I}_{E_{\alpha }}(\beta ,\beta ))$. We have $E_{\alpha } \subseteq S $. Since $\sharp (S) = x \le (\beta +1)^2$ and $S$ is general, we have $h^1(\mathcal {I}_S(\beta ,\beta )) =0$. Hence $ h^1(\mathcal {I}_{E_{\alpha }}(\beta ,\beta ))=0$.
$d_4(C)$ for the normalization $C$ of a nodal $Y\subset Q$ {#S3}
==========================================================
For any finite set $S\subset Q$ let $2S$ denote the first infinitesimal neighborhood of $S$ in $Q$, i.e. the closed subscheme of $Q$ with $(\mathcal {I}_S)^2$ as its ideal sheaf. The scheme $2S$ is zero-dimensional, $(2S)_{red}= S$ and $\deg (2S)=3\cdot \sharp (S)$.
\[g1\] Fix integers $a, b, x$ such that $b \ge a \ge 4$ and $0 \le 3x \le ab$. Fix a general $S\subset Q$ such that $\sharp (S)=x$. We have $h^0(Q,\mathcal {I}_{2S}(a,b))
= (a+1)(b+1) -3x$. Fix a general $Y\in \vert \mathcal {I}_{2S}(a,b)\vert$. Then $Y$ is integral, nodal and $\mbox{Sing}(Y)=S$.
We have $h^1(Q,\mathcal {I}_{2S}(a-1,b-1)) =0$ ([@bd], Theorem 1.1). Hence $h^1(Q,\mathcal {I}_{2S}(a,b)) =0$, i.e. $h^0(Q,\mathcal {I}_{2S}(a,b))
= (a+1)(b+1) -3x$. Since $h^1(Q,\mathcal {I}_{2S}(a-1,b-1)) =0$ and the line bundle $\mathcal {O}_Q(1,1)$ is very ample, Castelnuovo-Mumford’s lemma implies that the sheaf $\mathcal {I}_{2S}(a,b)$ is spanned. Hence $\vert \mathcal {I}_{2S}(a,b)\vert$ has no base points outside $S$. Bertini’s theorem implies $S = \mbox{Sing}(Y)$. Fix $P\in S$ and set $S':= S\setminus \{P\}$. Take a general $D\in \vert \mathcal {I}_{2S'}(a-1,b-1)\vert$. Since $h^1(Q,\mathcal {I}_{2S}(a-1,b-1)) =0$, we have $h^1(Q,\mathcal {I}_{2S'\cup \{P\}}(a-1,b-1)) =0$. Hence $h^0(\mathcal {I}_{2S'\cup \{P\}}(a-1,b-1))
= h^0(\mathcal {I}_{2S'}(a-1,b-1))-1$. Since $D$ is general, we get $P\notin D$. Let $D'\cup D''
\subset Q$ be the reducible conic with $P$ as its singular locus. Since $P\notin D$, $D\cup D'\cup
D''$ is an element of $\vert \mathcal {I}_{2S}(a,b)\vert$ with an ordinary node at $P$. Since $Y$ is general, it has an ordinary node at $P$. Since this is true for all $P\in S$, $Y$ is nodal. For every irreducible component $T$ of $Y$ we have $\omega _Q\cdot T = \mathcal {O}_Q(-2,-2) \cdot T<0$. Since $b\ge a \ge 4$ and $3x \le ab$, we have $p_a(Y) = ab-a-b +1 \ge x$. Since $Y$ is nodal, no component of $Y$ appears with multiplicity $\ge 2$. Since $S$ is general and $S = \mbox{Sing}(Y)$, the curve $Y$ is irreducible ([@ac], Proposition 4.1).
\[g2\] Fix integers $a, b, x$ such that $b \ge a \ge 4$ and $0 \le 3x \le (a-1)(b-1)$. Fix a general $S\subset Q$ such that $\sharp (S)=x$. Fix zero-dimensional schemes $Z, Z'\subset Q$ such that $\deg (Z)=\deg (Z')=2$, $Z_{red}$ and $(Z')_{red}$ are distinct points, $Z$ is contained in a line $D_1\in \vert \mathcal {O}_Q(1,0)\vert$, $Z'$ is contained in a line $D_2\in \vert \mathcal {O}_Q(0,1)\vert$ and $Z\cap D_2=Z'\cap D_1=S\cap (D_1\cup D_2) =\emptyset$. Take a general $Y\in \vert \mathcal {I}_{Z\cup Z'\cup 2S}(a,b)\vert$. Then $h^0(Q,\mathcal {I}_{Z\cup Z'\cup 2S}(a,b)) = (a+1)(b+1)-3x-4$, $Y$ is nodal, integral, $\mbox{Sing}(Y)=S$, $\sharp ((Y\cap D_1)_{red}) =b-1$ and $\sharp ((Y\cap D_2)_{red}) =a-1$.
We have $h^1(Q,\mathcal {I}_{2S}(a-2,b-2)) =0$ ([@bd], Theorem 1.1). We immediately get $h^1(Q,\mathcal {I}_{Z\cup Z'\cup 2S}(a,b)) = 0$, i.e. $h^0(Q,\mathcal {I}_{Z\cup Z' \cup 2S}(a,b)) = (a+1)(b+1)-3x-4$. We also see that $h^1(Q,\mathcal {I}_{Z\cup Z'\cup 2S}(a-1,b-1)) = 0$. Hence $\mathcal {I}_{Z\cup Z'\cup 2S}(a,b)$ is spanned by Castenuovo-Mumford’s lemma. Hence $Y$ is smooth outside $S\cup \{Z_{red},(Z')_{red}\}$. Lemma \[g1\] applied to the integers $a-1$ and $b-1$ gives the existence of an integral and nodal curve $T\in \vert \mathcal {I}_{2S}(a-1,b-1)\vert$ such that $S= \mbox{Sing}(T)$. Since $\mathcal {I}_{2S}(a-1,b-1)$ is spanned, we may find $T$ as above and with $Z_{red}\notin T$ and $(Z')_{red}\notin T$. Since $T\cup D_1\cup D_2
\in \vert \mathcal {I}_{Z\cup Z' \cup 2S}(a,b)\vert$ and $S\cap (D_1\cup D_2) =\emptyset$, we get that $Y$ is smooth at $(Z)_{red}$ and at $(Z')_{red}$ and nodal at each point of $S$. Since $T$ is irreducible and $Y$ is general, either $Y$ is irreducible or $Y = T_1\cup A_1$ with $T_1\in \vert \mathcal {O}_Q(a,b-1)\vert$ and $A_1\in \vert \mathcal {O}_Q(0,1)\vert$ or $Y = T_2\cup A_2$ with $T_2\in \vert \mathcal {O}_Q(a-1,b)\vert$ and $A_2\in \vert \mathcal {O}_Q(1,0)\vert$ or $Y = T_3\cup A_3$ with $T_3\in \vert \mathcal {O}_Q(a-1,b-1)\vert$ and $A_3\in \vert \mathcal {O}_Q(1,1)\vert$. The last three cases are impossible, because $Y$ is nodal, $\mbox{Sing}(Y) =S$, no $a-1$ of the points of $S$ are contained in a line and no conic contains $a+b-2$ points of $S$.
Since $Z\subseteq D_1\cap Y$, we have $\sharp (Y\cap D_1) \le b-1$. Since $\mathcal {I}_{2S\cup Z\cup Z'}(a,b)$ is spanned and $Y$ is general, $Y$ does not contain the degree $3$ divisor of $D_1$ with $Z_{red}$ as its support. Hence $Z_{red}$ appears with multiplicity two in the scheme $Y\cap D_1$. We need to prove that the other points of $(Y\cap D_1)_{red}$ appear with multiplicity one in the scheme $Y\cap
D_1$. Fix $P\in D_1\setminus Z_{red}$ and let $W\subset D_1$ be the divisor of degree two with $P$ as its support. Since $h^1(\mathcal {I}_{2S}(a-1,b-1))=0$ and $b\ge a \ge 3$, we have $h^1(\mathcal {I}_{2S\cup Z\cup Z'\cup W}(a,b))=0$. Hence $\vert \mathcal {I}_{2S\cup Z\cup Z'\cup W}(a,b)\vert$ has codimension two in $\vert \mathcal {I}_{2S\cup Z\cup
Z'}(a,b)\vert$. Since $\dim (D_1)=1$ and $Y$ is general, $Y$ contains no such scheme $W$. Hence $\sharp ((Y\cap D_1)_{red})
=b-1$. In the same way we prove that $\sharp ((Y\cap D_2)_{red}) =a-1$.
\[f2\] Fix integers $a, m, x$ such that $a\ge 4$, $0 \le m < a$ and $0 \le 3x \le (a-1)(a+m-1)$. Fix a general $S\subset
Q$ such that $\sharp (S)=x$. Let $C$ be the normalization of a general $Y \in \vert \mathcal {I}_{2S}(a,a+m)\vert$. Let $u_1: C \to \mathbb {P}^1$ (resp. $u_2: C\to \mathbb {P}^1$) be the $g^1_a$ (resp. $g^1_{a+m}$) induced by the projection of $Q$ onto its second (resp. first) factor. Then neither $u_1$ nor $u_2$ is composed with an involution, i.e. there are no triple $(C_i,v_i,w_i)$ with $C_i$ a smooth curve, $w_i: C \to C_i$, $v_i: C_i\to \mathbb {P}^1$, $u_i = v_i\circ w_i$, $\deg (v_i)>1$ and $\deg (w_i)>1$.
It is sufficient to find $O\in \mathbb {P}^1$ and $O'\in \mathbb {P}^1$ such that $u_2^{-1}(O)$ is formed by $a+m-1$ points, one of them being an ordinary ramification point of $u_2$, and $u_1^{-1}(O')$ is formed by $a-1$ points, one of them being an ordinary ramification point of $u_2$. Fix zero-dimensional schemes $Z, Z'\subset Q$ such that $\deg (Z)=\deg (Z')=2$, $Z_{red}$ and $(Z')_{red}$ are distinct points, $Z$ is contained in a line $D\in \vert \mathcal {O}_Q(1,0)\vert$, $Z'$ is contained in a line $D'\in \vert \mathcal {O}_Q(0,1)\vert$ and $S\cap (D\cup D') =\emptyset$. Take a general $M\in \vert \mathcal {I}_{Z\cup Z'\cup 2S}(a,a+m)\vert$. Lemma \[g2\] gives that $M$ is a nodal and irreducible curve, $\mbox{Sing}(M) =S$, $ \sharp ((D\cap M)_{red}) = a+m-1$ and $\sharp ((D'\cap M)_{red}) =a-1$. Let $u': C' \to M$ be the normalization of $M$. Call $u'_1: C'\to \mathbb {P}^1$ and $u'_2: C'\to \mathbb {P}^1$ the pencils induced by the projections of $Q$. The scheme $M\cap D$ is the disjoint union of $Z$ and $a+m-2$ distinct points and the scheme $M\cap D'$ is the disjoint union of $Z'$ and $a-2$ distinct points. Since $(Z')_{red}\cap S = \emptyset$ and $Z_{red}\cap S =\emptyset$, there are unique points $O_1, O'_1\in C$ such that $u'(O_1) =Z_{red}$ and $u'(O'_1) = (Z')_{red}$. Set $O':= u_2(O'_1)$ and $O:= u_1(O_1)$. The $a-2$ (resp. $a+m-2$) points of $u_1^{-1}(O')\setminus \{O'_1\}$ (resp. $u_2^{-1}(O)\setminus \{O_1\}$) appear with multiplicity one in the fiber $u_1^{-1}(O')$, because $u$ is a local isomorphism at each of these points and $D'$ (resp. $D$) is transversal to $M$ outside $(Z')_{red}$ (resp. $Z_{red}$). Hence $O_1$ (resp. $O'_1$) is an ordinary ramification point of $u'_2$ (resp. $u'_1$). Since $Y$ is a general equitrivial deformation of $M$ inside $\vert \mathcal {I}_{2S}(a,a+m)\vert$, each $u_i$ has a fiber with a unique ramification point and this ramification point is an ordinary one.
\[g5\] Fix integers $a, m, x$ such that $a \ge 18$, $x\ge 0$, $0\le m<a$ and $$\label{eqg00}
x\le a/3 +m$$ Fix a general $S\subset Q$ such that $\sharp (S)=x$. A general $Y\in \vert \mathcal {I}_{2S}(a,a+m)\vert$ is an integral nodal curve with $S$ as its singular locus. Let $u: C \to Y$ denote the normalization map. We have $g(C)=
a^2+am-2a-m+1 -x$ and $d_4(C) \ge 3a-14$. If $a \ge 4m+43$, then $d_4(C)/4 > d_3(C)/3$.
Lemma \[g2\] gives that $Y$ is integral and nodal and that $S =\mbox{Sing}(Y)$. Hence $C$ has genus $a^2+am-2a-m+1-x$.
Since the line bundle $u^\ast (\mathcal {O}_Y(1,1))$ has degree $2a+m$, we have $d_3(C)\le 2a+m$. Notice that $4(2a+m) < 3(3a-14)$ if $a \ge 4m +43$. Set $z:= d_4(C)$ and assume $z \le 3a-15$.
Fix $L\in \mbox{Pic}^z(C)$ evincing $d_4(C)$. The line bundle $L$ is spanned ([@lm], Lemma 3.1 (b)). Fix a general $A\in \vert L\vert$. Set $B:= u(A)$. Since $L$ has no base points and $A$ is general, $S\cap B =\emptyset$. Since $Q$ has only finitely many lines intersecting $S$ and $A$ is general, we may assume $B$ disjoint from these finitely many lines. Hence no line of $Q$ contains a point of $S$ and at least one point of $B$.
$h^1(\mathcal {I}_{S\cup B}(a-2,a+m-2))>0$.
Since $L$ has no base points, we have $h^0(C,\mathcal {O}_C(A\setminus \{O\})) =
h^0(C,\mathcal {O}_C(A)) -1$ for every $O\in A$. Hence $h^0(C,\omega _C(-A+\{O\})) = h^0(C,\omega _C(-A))$ for every $O\in A$ (Riemann-Roch and Serre duality). We have $\omega _Q\cong \mathcal {O}_Q(-2,-2)$. Hence the adjunction formula gives $\omega _Y \cong \mathcal {O}_Y(a-2,a+m-2)$. Since $h^i(\mathcal {O}_Q(-2,-2)) =0$, $i=0, 1$, the restriction map $H^0(\mathcal {O}_Q(a-2,a+m-2)) \to H^0(Y,\omega _Y)$ is bijective. Since $Y$ has only ordinary nodes as singularities, we have $H^0(C,\omega _C) \cong H^0(\mathcal {I}_S(a-2,a+m-2))$. Hence for any $O\in A$ we have $h^0(\mathcal {I}_{S\cup (B\setminus \{u(O)\})}(a-2,a+m-2))
= h^0(C,\omega _C(-(A\setminus \{O\}))) = h^0(C,\omega _C(-A)) = h^0(\mathcal {I}_{S\cup B}(a-2,a+m-2))$. Hence $h^1(\mathcal {I}_{S\cup B}(a-2,a+m-2))>0$, concluding the proof of Claim 1.
Let $v_2: C \to \mathbb {P}^1$ (resp. $v_1: C \to \mathbb {P}^1$) denote the degree $a$ (resp. degree $a+m$) morphism obtained composing $u$ with the pencil associated to $\vert \mathcal {O}_Q(0,1)\vert$ (resp. $\vert \mathcal {O}_Q(1,0)\vert$). Lemma \[f2\] shows that none of these two pencils factors non-trivially.
For a general $B$ no line of $Q$ contains two or more points of $B$.
Assume for instance that for a general $B$ there is a line $D_B\in \vert \mathcal {O}_Q(0,1)\vert$ such that $\sharp (D_B\cap B)\ge 2$. Set $\Psi := \{(P,Q)\in C\times C: P\ne Q$ and $v_2(P) =v_2(Q)\}$. For any $D\in \vert \mathcal {O}_Q(0,1)\vert$ the scheme $D\cap Y$ is zero-dimensional. Since $\dim (\vert L\vert )=3 + \dim (\vert \mathcal {O}_Q(0,1)\vert )$, there is a one-dimensional irreducible set $\Phi \subseteq \Psi$ such that for all $(P,Q)\in \Phi$ the set $\{P,Q\}$ is contained in a $3$-dimensional family $F_{\{P,Q\}}$ of elements of $\vert L\vert$. Fix $(P,Q)\in \Phi$. Since $L$ has no base points, we have $h^0(L(-P)) =4$. Hence the existence of the family $F_{\{P,Q\}}$ implies $h^0(L(-P-Q)) = h^0(L(-P))$. Hence $Q$ is a base point of $\vert L(-P)\vert$. The two projections $C\times C\to C$ induce dominant maps $\Phi \to C$. Hence $P$ may be seen as a general point of $C$. Since $Q\ne P$, the morphism $\varphi : C \to \mathbb {P}^4$ associated to $\vert L\vert$ is not birational onto its image, i.e. $\varphi = u _2\circ u_1$ with $\deg (u_1)\ge 2$, $u_1: C \to C'$ a morphism of degree $\ge 2$ with $C'$ a smooth curve and $u_2: C' \to \varphi (C)\hookrightarrow \mathbb {P}^4$ birational onto its image. We have $z = \deg (u_1)\cdot \deg (\varphi (C)) \ge 4 \deg (u_1)$. Since $u_1(P)=u_1(Q)$ for a general $(P,Q)\in \Phi $, a general fiber of $u_1$ intersects in at least two points a fiber of $v_2$. Since $v_2$ is not composed with a pencil, $u_1$ factors through $v_2$. Hence $z \ge 4a$, a contradiction. Hence Claim 2 is true.
Since $S$ is finite, there are only finitely many lines of $Q$ containing at least one point of $S$. Call $\Gamma$ their union. Since $S$ is general, no such a line contains at least two points of $S$. Since $\vert L\vert$ has no base points, and $\Gamma \cap Y$ is finite, for general $B$ we may assume $B\cap \Gamma =\emptyset$. Hence Claim 2 implies that no line of $Q$ contains at least two points of $S\cup B$. Claim 1 gives $h^1(Q,\mathcal {I}_{S\cup B}(a-2,a+m-2)) >0$. By (\[eqg00\]) we have $x+z \le 3a-15 +a/3 +m \le
10 \lfloor a/3 \rfloor +20/3 +m -15 \le 10\lfloor a/3\rfloor +m-8$. To apply Lemma \[e4\] with $E=S\cup B$, $a=u$ and $v =a+m$ and get a contradiction it is sufficient to prove that $\sharp ((S\cup B)\cap T_1) \le 2a-3$ for all $T_1\in \vert \mathcal {O}_Q(1,1)\vert$, $\sharp ((S\cup B)\cap T_2) \le 3(a-2)+1$ for every $T_2\in \vert \mathcal {O}_Q(2,1)\vert$ and $\sharp ((S\cup B)\cap T_2) \le 3(a-2)-4$ for every $T_2\in \vert \mathcal {O}_Q(2,1)\vert$. Fix $T_1\in \vert \mathcal {O}_Q(1,1)\vert$, $T_2\in \vert \mathcal {O}_Q(2,1)\vert$ and $T_3\in \vert \mathcal {O}_Q(1,2)\vert$. Set $y_i:= \sharp (S\cap T_i)$ and $a_i:= \sharp (B\cap T_i)$. Assume either $y_1+a_1 \ge 2a-2$ or $y_2+a_2 \ge 3a-4$ or $y_3+a_3\ge 3a-9$. Since $S$ is general, we have $y_1\le 3$, $y_2 \le 5$ and $y_3\le 5$. Hence either $a_1\ge 2a-5$ or $a_2\ge 3a-9$ or $a_3 \ge 3a-14$. Since $z\ge a_i$ for every $i$ such that $T_i$ exists and $z\le 3a-15$, we get the existence of $T_1\in \vert \mathcal {O}_Q(1,1)\vert$ such that $\sharp (B\cap T_1) \ge 2a-5$. Since any line of $Q$ contains at most one element of $B$, $T_1$ is irreducible.
Step ($\diamond$) (Proof due to the referee; a simpler form would also prove Claim 2) Let $\varphi : C \to \mathbb {P}^4$ be the morphism defined by $\vert L\vert$. Set $\Gamma := \varphi (C)$. Let $C'$ be the normalization of $\Gamma$ and let $f: C \to C'$ be the covering induced by $\varphi$. Let $g^4_{z'}$, $z' = z/\deg (f)$, be the linear series on $C'$ induced by the inclusion $\Gamma \hookrightarrow \mathbb {P}^4$. Take $A'\in g^4_{z'}$ with $A = f^{-1}(A')$. Since $A\in \vert L\vert$ is general, $A'$ is general. The monodromy group of the general hyperplane section of $\Gamma$ is the full symmetric group. Hence any $4$ points of a general hyperplane section of $\Gamma$ span a 3-dimensional projective space. Hence for any $E\subset A'$ with $\sharp (E) \ge 4$, $A'$ is the only element of $g^4_{z'}$ containing $E$. For any $T\in \vert \mathcal {O}_Q(1,1)\vert$ the set $T\cap Y$ is finite. Since $\dim (\vert \mathcal {O}_Q(1,1)\vert )<4$, there is an infinite family $\mathcal {F} \subset \vert L\vert$ such that $u(D)$ contains $B\cap T_1$ for all $D\in \mathcal {F}$. Fix $D'\in g^4_{z'}$ such that $D:= f^{-1}(D') \in \mathcal {F}$ and $D' \ne A'$. Since $u(f^{-1}(D'\cap A'))$ contains $B\cap T_1$, there is $P\in A'$ such that $\sharp (u(f^{-1}(P)) \cap (B\cap T_1)) \ge (2a-5)/3$. Since $(2a-5)/3 \ge 3$, $T_1$ is the only element of $\vert \mathcal {O}_Q(1,1)\vert$ containing $u(f^{-1}(P))\cap (B\cap T_1)$. Moving $A'$ generally in $g^4_{z'}(-P)$ we get that $D'$ and $B\cap T_1$ moves into a subscheme $B'$ of $u(f^{-1}(D'))$ contained in some $T'\in \vert \mathcal {O}_Q(1,1)\vert$. Since $P\in D'$ and $\sharp (u(f^{-1}(P)) \cap (B\cap T_1)) \ge 3$, we have $T'=T_1$. Hence $B\cap T_1$ does not move moving $A'$ in $g^4_{z'}(-P)$. Hence $\sharp (f^{-1}(P)) \ge 2a-5$, i.e. $\deg (f) \ge 2a-5$. Since $z'\ge 4$, we get $z\ge 4(2a-5) $, a contradiction.
\[h1\] Fix integers $x, \alpha, \gamma $ such that $\alpha \ge 3$, $\gamma \ge 4$, $0 \le x \le (\gamma -1)^2$. Set $a:= 3\alpha +\gamma$. Fix a general $S\subset Q$ such that $\sharp (S)=x$ and a general $Y\in \vert \mathcal {I}_{2S}(a,a)\vert$. $Y$ is integral, nodal and $\mbox{Sing}(Y)=S$. Let $C$ be the normalization of $Y$. Then $d_4(C) \ge \min \{10\alpha +1,3a-14\}$.
Set $z:= d_4(C)$ and assume $z\le 3a-15$ and $z\le 10\alpha$. Take the set-up of the proof of Proposition \[g5\] with $m =0$. In particular we get a finite set $B\subset Q\setminus S$ such that $\sharp (B) =z$, $h^1(Q,\mathcal {I}_{S\cup B}(a-2,a-2)) >0$ and no line of $Q$ contains two points of $S\cup B$. To get a contradiction we cannot apply Lemma \[e4\] with $u=v = a-2$ and $E:= S\cup B$, because $x$ may be large. We need to check that we may apply Lemma \[g4\] with $u=a-2$ and $\beta =\gamma -2$, i.e. we need to check that no line of $Q$ contains two points of $S\cup B$, $\sharp (B\cap T_1)\le 2a-6$ for every $T_1\in \vert \mathcal {O}_Q(1,1)\vert$, $\sharp (B\cap T_2)\le 3a-10$ for every $T_2\in \vert \mathcal {O}_Q(2,1)\vert$ and $\sharp (B\cap T_3)\le 3a-14$ for every $T_3\in \vert \mathcal {O}_Q(1,2)\vert$. Since $z\le 3a-15$, we only need to test the conditions for the lines of $Q$ and that $\sharp (B\cap T_1)\le 2a-6$ for each $T_1\in \vert \mathcal {O}_Q(1,1)\vert$. Step ($\diamond$) of the proof of Proposition \[g5\] proves the condition for $T_1\in \vert \mathcal {O}_Q(1,1)\vert$. We may also copy the proof of Claim 2 of the proof of Proposition \[g5\], because the assumptions of Lemma \[g2\] are satisfied.
\[h4\] Fix integers $a, x$ such that $a \ge 24$ and $0 \le x \le 2a-4$. Fix a general $S\subset Q$ such that $\sharp (S)=x$ and a general $Y\in \vert \mathcal {I}_{2S}(a,a)\vert$. $Y$ is integral, nodal and $\mbox{Sing}(Y)=S$. Let $C$ be the normalization of $Y$. Then $2a -5 \le d_3(C) \le 2a$.
Lemma \[g1\] gives that $Y$ is integral, nodal and smooth outside $S$. The pull-back of the line bundle $\mathcal {O}_Q(1,1)$ gives $d_3(C) \le 2a$. Assume $z:= d_3(C) \le 2a-6$ and fix $L\in \mbox{Pic}^z(C)$ evincing $d_3(C)$. As in Claims 1 and 2 of the proof of Proposition \[g5\] we get a set $B\subset Q\setminus S$ such that $\sharp (B) =z$, no line of $Q$ contains $2$ of the points of $S\cup B$ and $h^1(\mathcal {I}_{S\cup B}(a-2,a-2))>0$. Since $z \le 2a-6$ and $a\ge 16$, we have $z+5 \le 3a-15$. Hence $\sharp (T\cap (S\cup B)) \le 3a-15$ for every $T\in \vert \mathcal {O}_Q(2,1)\vert$ and every $T\in \vert \mathcal {O}_Q(1,2)\vert$. Since $z\le 2a-6$, we have $\sharp (B\cap T) \le 2(a-2)-2$ for every $T\in \vert \mathcal {O}_Q(1,1)\vert$. Set $\gamma := \max \{2,-1+\sqrt{x}\}$, $\alpha := \lfloor (a-2-\gamma )/3\rfloor$ and $\beta := a-2 -3\alpha$. We have $\beta \ge \gamma \ge 2$ and $x \le (\gamma +1)^2 \le (\beta +1)^2$. To apply Lemma \[g4\] (and hence to get $h^1(\mathcal {I}_{S\cup B}(a-2,a-2))=0$, i.e. a contradiction) it is sufficient to prove that $z\le 10\alpha$. This is true if $x \le 8$, because in this case $\gamma =2$. Hence we may assume $\gamma = -1+\sqrt{x}$. Hence $\alpha \ge (a-4 -\sqrt{2a})/3$. We have $10(a-4-\sqrt{2a})/3 \ge 2a-6$ if and only if $4a-22 \ge 10\sqrt{2a}$. Hence it is sufficient to assume $a\ge 24$.
\[h3\] Take $C$ as in Proposition \[h1\]. Set $M:= u^\ast (\mathcal {O}_Y(2,1))\in \mbox{Pic}^{3a}(C)$. Since $h^0(Q,\mathcal {O}_Q(2,1)) =6$ and $Y$ is contained in no element of $\vert \mathcal {O}_Q(2,1)\vert$, we have $h^0(M)\ge 6$. Fix $P\in Y$ such that $P\in \mbox{Sing}(Y)$ if $x>0$. Let $F\subset C$ be the scheme-theoretic pull-back of the scheme $P$. We have $\deg (F) \ge 1+\min \{1,x\}$. Since $h^0(Q,\mathcal {I}_E(2,1)) =4$, we have $h^0(M(-F)) \ge 5$. Hence $d_4(C) \le 3a - 1 -\min \{1,x\}$.
\[h2\] Fix integers $a, x$ such that $a\ge 204$ and $0 \le x \le 2a-4$. Fix a general $S\subset Q$ such that $\sharp (S)=x$ and a general $Y\in \vert \mathcal {I}_{2S}(a,a)\vert$. $Y$ is integral, nodal and $\mbox{Sing}(Y)=S$. Let $C$ be the normalization of $Y$. Then $2a-5 \le d_3(C) \le 2a$ and $3a-15
\le d_4(C) \le 3a -1-\min \{1,x\}$.
Since $3x \le 6a-12 \le a^2$, $Y$ is integral, nodal and $\mbox{Sing}(Y) = S$ (Lemma \[g1\]). Lemma \[h4\] gives $2a-5 \le d_3(C) \le 2a$. Remark \[h3\] gives $d_4(C) \le 3a-1-\min \{x,1\}$. Assume $z:= d_4(C) \le 3a-16$. Take $B$ as in the proofs of Proposition \[g5\] and \[h1\]. We have $h^1(\mathcal {I}_{S\cup B}(a-2,a-2)) >0$. Set $\delta := -1
+\lceil \sqrt{2a-4}\rceil$ and $\alpha := \lfloor (a-2 -\delta )/3\rfloor$. Notice that $x \le 2a-4 \le (\delta +1)^2$ and that $\alpha \ge (a-5-\sqrt{2a})/3$. Since $a\ge 204$, we have $a-2 \ge 10\sqrt{2a}$, i.e. $10(a-5-\sqrt{2a})/3 \ge 3a-16$. Hence $10\alpha \ge 3a-16$. By assumption we have $d_3(C) \le 3(a-2)-9$. Claim 2 and Step ($\diamond$) of the proof of Proposition \[g5\] show that we may apply Lemma \[g4\] with the integers $u:=a-2$ and $\beta:= a-2-3\alpha$ (notice that $\beta \ge \delta$) and get $h^1(\mathcal {I}_{S\cup B}(a-2,a-2)) =0$, a contradiction.
For all integers $a, x$ set $g_a:= a^2-2a+1$ and $g_{a,x} = g_a-x$. If $a>0$, then $g_a = p_a(Y)$ for any $Y\in\vert \mathcal {O}_Q(a,a)\vert$. Hence if $Y$ is a nodal curve of type $(a,a)$ with exactly $x$ nodes, then $g_{a,x}$ is the genus of the normalization of $Y$. Now assume $a\ge 2$. We have $g_a-g_{a-1} = a^2-2a+1-a^2+2a-1+2a-2-1 =2a-3$. Hence the set $\{g_{a,x}\}_{0 \le x \le 2a-4}$ contains every integer between $g_{a-1}+1$ and $g_a$. We take the set-up of the proof of Corollary \[h2\]. Fix an integer $g \ge 40805$. Let $a$ be the only integer such that $g_{a-1} < g \le g_a$. Since $g_{203} =40804$, we have $a\ge 204$. We have $g = g_a -x$ with $0\le x\le 2a-4$. Apply Corollary \[h2\].
Of course, the lower bound $g\ge 40805$ is not sharp.
For any $g< 40805$ we take as $C_g$ an arbitrary smooth curve of genus $g$. Fix an integer $g \ge 40805$ and call $a$ the minimal positive integer such that $g \le a^2-2a+1$. Set $x:= a^2-2a+1-g$. Since $g> (a-1)^2-2(a-1)+1$, we have $x \le 2a-4$. Fix a general $S\subset Q$ such that $\sharp (S) =x$. Take as $C_g$ the normalization of a general $Y\in \vert \mathcal {I}_{2S}(a,a)\vert$. Corollary \[h2\] gives $2a-6 \le d_3(C) \le 2a$ and $3a -15\le d_4(C) \le 3a -1 -\min \{1,x\}$. We have $ g_{a-1} < g \le g_a$. Hence the limits are as in the statement of Theorem \[i2\].
[99]{}
E. Arbarello and M. Cornalba, Footnotes to a paper of Beniamino Segre, *Math. Ann.* 256 (1981), 341–362.
K. Baur and J. Draisma, Secant dimensions of low-dimensional homogeneous varieties, *Adv. Geom.* 10 (2010), 1–29.
H. Lange and G. Martens, On the gonality sequence of an algebraic curve, *Manuscripta Math.* 137 (2012), 457–473.
H. Lange and P. E. Newstead, Clifford indices for vector bundles on curves, in: Schmitt, A. (ed.) Affine Flag Manifolds and Principal Bundles, Trends in Mathematics, Birkhäuser, Basel (2010), pp. 165–202.
H. Lange and P. E. Newstead, Lower bounds for Clifford indices in rank three, *Math. Proc. Cambridge Philos. Soc.* 150 (2011), no. 1, 23–33.
E. C. Mistretta and L. Stoppino, Linear series on curves: stability and Clifford index, arXiv:111.0304v1.
[^1]: The author was partially supported by MIUR and GNSAGA of INdAM (Italy).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The dynamics of a binary mixture of large and small discs are studied at temperatures approaching the glass transition using an analysis based on the topology of the Voronoi polygon surrounding each atom. At higher temperatures we find that dynamics is dominated by fluid-like motion that involves particles entering and exiting the nearest-neighbour shells of nearby particles. As the temperature is lowered, the rate of topological moves decreases and motion becomes localised to regions of mixed pentagons and heptagons. In addition we find that in the low temperature state particles may translate significant distances without undergoing changes in their nearest neighbour shell. These results have implications for dynamical heterogeneities in glass forming liquids.'
address:
- '1 Physics Department, Simon Fraser University, 8888 University Drive, Burnaby, British Columbia, V5A 1S6, Canada'
- '2 Institut für Theoretische Physik, Technicshe Universität Berlin, 10623 Berlin, Germany'
author:
- 'Matthew T. Downton$^{1,2}$ and Malcolm P. Kennett$^{1}$'
title: 'Following microscopic motion in a two dimensional glass-forming binary fluid'
---
Introduction {#sec:introduction}
============
The relationship between the dynamics and structure of a glass-forming liquid is a non-trivial problem. Structurally, a simple liquid is isotropic and has smooth pair distribution functions that decay to the background value within a few nearest neighbour spacings. The positions of the maxima and minima of these functions can be explained in terms of coordination shells of particles packed around a central reference particle. This gives an intuitive understanding of the microscopic structure of a liquid. The corresponding picture of dynamics is based around atoms that are trapped by their surrounding neighbours, and make rare “cage-breaking” jumps to nearby positions. At high temperatures, the separation in time between collisions with neighbours and jumps is small, but at lower temperatures this time starts to diverge. How different arrangements of atoms promote or inhibit this motion is an open question.
Molecular dynamics simulations provide a powerful tool to address this problem by allowing one to follow structure and dynamics simultaneously. It has been observed in simulations that systems that exhibit glass-like behavior at lower temperatures can have characteristics of an interplay between crystal and liquid structure [@shintani06]. Structural changes have also been observed in fluids approaching the fluid-solid transition [@truskett98]. Additionally, it has been observed that as the glass transition is approached, dynamics become increasingly spatially heterogeneous [@Sillescu; @Ediger; @Glotzer; @Richert; @Castillo; @Israeloff; @VandenBout; @Reinsberg; @Weeks; @Kob; @Donati1; @Donati2]. The character of these dynamical heterogeneities has been well studied, both computationally [@Donati1; @HarrowellDH; @Yamamoto; @Andersen; @Perera2; @Gebremichael1; @Stevenson] and in experiments on colloidal glasses [@Kegel; @Weeks2], in which “caging” of particles, and the escape from cages via string-like motion has been seen at temperatures slightly above the glass transition temperature. Despite all of this effort, the origin of dynamical heterogeneities is unclear, as is their connection to the dramatic slowing down of dynamics seen as a liquid becomes glassy.
A simulation approach that has been recently introduced, the iso-configurational ensemble, sheds light on the relationship between structure and dynamics [@widmer-cooper04]. In this technique, repeated short simulations are performed using the same starting structure, but randomized velocities. Averages of single particle displacements over the ensemble of simulations give a map of the propensity to move. It has been observed that the local Debye-Waller factor appears to correlate very closely with the dynamic propensity at long times [@widmer-cooper06], however it is harder to answer the more subtle question of how the propensity correlates with the dynamics of a single trajectory. It appears that on the single-particle level, such a connection cannot be made [@berthier07], but that the connection between structure and dynamics will exist at longer lengthscales.
An alternative theoretical approach to understanding the behaviour of glass-forming liquids, and in particular their kinetic arrest in the vicinity of the glass transition temperature, has been through studying phenomenological models. Such models, in particular, kinetically constrained models (KCMs) have received considerable attention in recent years [@ritort03; @FA84; @FA85; @Graham97; @Garrahan02; @Berthier03; @PNAS03; @Leonard07] as they have had success in reproducing diverging timescales observed in glass-formers. These models are generally understood as an effective model for some coarse-grained degree of freedom (often referred to as spins), and at low temperatures, the dynamics can be understood as a few up-spins or “defects” in a background sea of down spins [@BerthierPRE03]. We have recently attempted to provide a coarse-graining procedure to map the dynamics of a glass-former onto a KCM, using local mean squared displacement as a means to define spins [@Downton07]. This approach found kinetically constrained behaviour became increasingly important at low temperatures, but only on timescales longer than the alpha relaxation time, whereas one would hope that the alpha relaxation time was in fact a consequence of the KCM, rather than an input. For shorter coarse-graining times, there was no evidence of kinetically constrained dynamics. This suggests that alternative defect variables should be identified if such a coarse-graining procedure is to be successful.
In this paper we examine the interplay between structure and dynamics by investigating the dynamics of a two-dimensional liquid in terms of the coordination number given by a Voronoi polygonisation. Previously, the population densities of different coordinations have been used to make predictions regarding the glass transition temperature [@aharonov07]. Here we use this choice to give an unambiguous definition of ‘fluid-like’ motion within a dense particle system in terms of a simple topological process. The approach we take can also be related to previous work on lattice based models that obey a similar dynamical rule [@davison00; @davison01; @sherrington02]. We find that there are a few dominant particle moves, of which, only one has significant temperature dependence, becoming relatively more important with decreasing temperature.
The structure of the paper is as follows: in Sec. \[sec:model\] we give details of the model and simulation procedures, along with some results that show the development of glassy phenomenology at lower temperatures. In Sec. \[sec:results-simulations\] we define the “defects” we consider and discuss their spatial distribution, and in Sec. \[sec:topol-based-dynam\] we present the microscopic dynamics of the fluid in terms of local topology changing transitions. These results are followed by concluding observations and discussion in Sec. \[sec:discussion\].
Model and simulation details {#sec:model}
============================
We used a two-dimensional model that exhibits all of the key ingredients of a true glass-former. The model consists of a binary mixture of large and small soft discs with size ratio 1:1.4 interacting with a potential of the form $V_{\alpha\beta}=\epsilon \left(\frac{\sigma_{\alpha\beta}}{r}\right)^{12}$. The mass, $m$, of all particles was set to unity; time was measured in units $\tau = \sqrt{m\sigma_{11}^{2}/\epsilon}$; distance was measured in terms of the small particle diameter, $\sigma_{11}$; temperature was measured in units of $\epsilon$.
Molecular dynamics simulations of 1600 particles in the ratio 75:25 small to large were performed at constant volume and temperature with the number density fixed to $0.85\sigma_{11}^{-2}$. Configurations were equilibrated at high temperatures then slowly cooled to create a series of configurations at different temperatures. These were then extensively equilibrated at fixed temperature using a Gaussian isokinetic thermostat. At temperatures below $T=0.330$ the system was found to age as indicated by a continuous decrease in the pressure and total potential energy. In a separate series of calculations it was found that when the initial configuration consisted of segregated small and large particles, there was complete mixing for temperatures $T\ge 0.450$; below this temperature there was no mixing for initially segregated configurations other than relaxation at the phase boundary.
The phenomenology of this mixture is almost identical to the mixture studied by Perera and Harrowell: multiple relaxational timescales develop at lower temperatures; a change in the scaling between structural relaxation and diffusion occurs at a temperature $T^{*}$ [@perera98; @perera99; @perera99a]. These results are summarised in figure \[fig:background\] where we take the structural relaxation time, $\tau_{\alpha}$, to be the time it takes the function $$F(q,t) = \frac{\left<\rho(\mathbf{q},t)\rho(\mathbf{-q},0)\right>}{\left<\rho(\mathbf{q},0)\rho(\mathbf{-q},0)\right>},$$ to decay to the value $1/e$, where $\rho(\mathbf{q},t) = \sum_{j}\exp\left( i\mathbf{q}\cdot\mathbf{r}_{j}\right)$ and $q=|\mathbf{q}|$ is the position of the first peak in the static structure factor. Similar results were found if the self-intermediate scattering function for a single particle species was used to measure structural relaxation. It can be seen that $T^{*}\approx 0.450$ for this particular mixture: at temperatures below $T^*$, $F(q,t)$ begins to exhibit two stage decay.
![Development of glassy behavior at lower temperatures. (a) The dynamic structure factor, $F(q,t)$, calculated at $q=6.23\sigma^{-1}$ ($T=0.360 - 1.440$) (b) Change in scaling of the structural relaxation time, $\tau_{\alpha}$, and the diffusion coefficient for small particles.[]{data-label="fig:background"}](figure1a.eps "fig:"){width="0.35\columnwidth"} ![Development of glassy behavior at lower temperatures. (a) The dynamic structure factor, $F(q,t)$, calculated at $q=6.23\sigma^{-1}$ ($T=0.360 - 1.440$) (b) Change in scaling of the structural relaxation time, $\tau_{\alpha}$, and the diffusion coefficient for small particles.[]{data-label="fig:background"}](figure1b.eps "fig:"){width="0.35\columnwidth"}
Static results {#sec:results-simulations}
==============
Distribution and concentration of defects {#sec:distribution-defects}
-----------------------------------------
The Voronoi procedure partitions the system into polygonal cells surrounding each atom. Nearest neighbours are identified as atoms which share an edge, whereas next-nearest neighbours have common nearest neighbours but are not themselves neighbours. For convenience we refer to atoms with coordination number, $c$, differing from six as defects and use the “topological charge”, $q=c-6$, to characterise the coordination of individual atoms. Within simulations such as this where periodic boundary conditions are used, the total coordination is conserved with $\sum_{i}q_{i}=0$.
The variation with temperature of the concentration of atoms with different $q$ is shown in figure \[fig:defect\]. Although occasional configurations contained four-fold or nine-fold coordinated atoms these were found to be relatively rare and the majority of atoms have $q$ in the range -1 to +2. At higher temperatures there is an excess of $-1$ atoms to $+1$ atoms that is compensated for by a small number of $+2$ atoms. At all temperatures within the range of study, the majority of large atoms have charge $+1$ and these constitute almost all of the +1 atoms at the lower end of this range (inset to figure \[fig:defect\]). The number of $q=0$ atoms also increases as a function of $T$, before levelling out at $T^{*}$.
![Population of defect species as a function of temperature in the non-aging temperature range. Inset: Topological charge of large atoms as a proportion of the total number of atoms.[]{data-label="fig:defect"}](figure2.eps){width="0.5\columnwidth"}
Two configurations representative of high and low temperatures are shown in figure \[fig:configurations\]. In both configurations, the distribution of large and small particles is similar and there is no apparent demixing of the system at lower temperatures. However, a greater degree of hexagonal coordination is seen within the lower temperature configuration; this is especially noticeable in larger domains of small particles. Accompanying these snapshots are pictures of the system showing the positions of atoms with non-zero $q$. For $T=0.870$ there is a slightly higher concentration of lone defects. In both systems there is a network of defects which almost spans the system. Inspection of figure \[fig:defect\] shows that in fact at the temperature at which the Stokes-Einstein relation breaks down, $T^{*}$, the population of $q=0$ sites reaches $\sim 0.65$, and the threshold for bond percolation on a triangular lattice is 0.347 [@Sykes]. Relationships between percolation and breakdown of the Stokes-Einstein relation have been noticed by earlier authors in the context of a lattice gas [@Nicodemi] and a three dimensional glass former [@Dzugutov]. The defect network is more tightly connected at lower temperatures, and leaves larger regions that are defect free and consist of small particles. Defects are generally arranged in pairs, so that a +1 defect is the neighbor to a -1 defect; no net separation of charge can be observed at either high or low temperatures.
![Particle positions and coordination at high and low temperatures. From left to right: $T=0.36$ and $T=0.84$. The top configuration shows the spatial distribution of large and small particles as filled and open circles respectively. In the lower configuration, atoms with $q=+1$ are filled circles, atoms with $q=-1$ are open circles, atoms with $q=+2$ are marked with a (+).[]{data-label="fig:configurations"}](figure3.eps){width="0.5\columnwidth"}
In figure \[fig:spin\] we plot the temperature dependence of the autocorrelation function of $q$. This is defined as $$\label{eq:autocorrelation}
C_{i}(t) = \frac{\left<q_{i}(t)q_{i}(0)\right>-\left<q_{i}\right>^{2}}{\left<q_{i}^2\right>-\left<q_{i}\right>^{2}}.$$ At high temperatures $C(t)$ decays rapidly before plateauing and then decaying more slowly. Reducing the temperature gives rise to a more complicated decay pattern and a second long-time decay rate appears that grows at a similar rate as the dynamic structure factor, $F(q,t)$. Hence the dynamics of structural “defects” clearly mirrors that of density correlations, suggesting that there may be some connection between the two.
![Autocorrelation function of topological charge averaged over small atoms. From top: temperatures in the range $T=0.360-1.470$ in steps of $0.03$.[]{data-label="fig:spin"}](figure4.eps){width="0.5\columnwidth"}
Topology based dynamics {#sec:topol-based-dynam}
=======================
One output of a Voronoi tesselation is a graph of nearest neighbours within the system. If fluid-like motion in real space can be thought of as the random motion of particles between the nearest and next-nearest neighbour shells then this motion has a simple topological rule that governs the creation and destruction of edges on the graph or equivalently the changes in the number of edges on each Voronoi cell. These dual views of the dynamics are shown in figure \[fig:T1\]. It is clear from the figure that pairs of particles that become nearest neighbours each gain an edge from the displaced particles. Provided that the number of particles is conserved, this single rule can be used to describe all of the changes in the coordination that occur during a simulation or experiment [@larson]. This gives a suitable coarse-graining of configurational changes and allows us to examine the microscopic origin of the decay of the topological charge found in the previous section.
In the following section we proceed as follows: A standard molecular dynamics simulation is performed and the Voronoi procedure is performed frequently enough that the majority of nearest neighbour changes can be directly identified[^1]. We then examine the frequency with which moves occur, the type and spatial distribution of nearest neighbour changes and finally examine the relationship with particle displacements at low temperatures. In the present study we ignore the identity of the atoms.
![The T1 move in both representations of the particle positions (a) from the Voronoi cells (b) from the Delaunay triangulation of nearest neighbours. In these dual views, $i$ and $l$ move together to become nearest neighbours, $j$ and $k$ are pushed apart, becoming next-nearest neighbours. All other neighbours of the four atoms remain the same.[]{data-label="fig:T1"}](figure5.eps){width="0.5\columnwidth"}
Identity and spatial distribution of nearest neighbour exchanges {#sec:rate-near-neighb}
----------------------------------------------------------------
The majority of nearest neighbour exchanges consist of the six basic moves listed in table \[tab:processes\]. In this table, we have listed the topological charge of each atom *before* the nearest neighbour exchange occurs. The atoms in each case are labelled according to figure \[fig:T1\]: $i$ and $l$ are acceptor atoms that each gain an extra nearest neighbour after the T1 move; $j$ and $k$ both lose nearest neighbours after the exchange. Many other exchange types are observed in the simulations, but are relatively infrequent in comparison to those listed.
The temperature dependence of the relative frequency of these move types is plotted in figure \[fig:temp-dependence\](a). Two aspects of the dynamics are immediately apparent. First, there is no sudden change in the relative frequency of any particular T1 move at the temperature where the change in scaling between diffusion and structural relaxation occurs. Second, moves which are the inverse of one another are found to occur with similar frequency. This can be seen as an indication that detailed balance holds for the liquid and that the simulations are in equilibrium on the time scales studied. In figure \[fig:temp-dependence\](b), the rate at which T1 moves are observed is plotted as a function of temperature and can be seen to be surprisingly linear.
At all temperatures, the most significant move is the dislocation glide (move $\mathbf{1}$) where a pair of $+1$ and $-1$ defects translate by a single particle spacing in a direction that is roughly perpendicular to the line segment that connects the two atoms. The frequency of this move has been noted in previous simulation studies of liquid dynamics [@deng89b]. The move $\mathbf{2}_\mathrm{a}$ involves the combination of a $-1$ defect with three atoms that have coordination of six to produce two $-1$ defects, a single $+1$ defect and a six-fold coordinated atom. $\mathbf{2}_\mathrm{b}$ is the reverse of this move. This pair of moves is therefore one route by which a lone $\pm 1$ defect pair may be created or annihilated and is the only move of the most frequent that shows any strong temperature dependence. The final move that we consider in detail is the pair $\mathbf{3}_\mathrm{a}$ and $\mathbf{3}_\mathrm{b}$. This involves the destruction of two $+1$ and two $-1$ defects in one direction or the creation of these defects in the other direction and will play an important role in the melting and annealing of crystal lattices.
$i$ $j$ $k$ $l$ $c$ $i$ $j$ $k$ $l$ $c$
------------------------- ----- ----- ----- ----- ----- ------------------------- ----- ----- ----- ----- -----
$\mathbf{1}$ -1 1 0 0 0 $\mathbf{4}_\mathrm{a}$ -1 1 1 0 1
$\mathbf{2}_\mathrm{a}$ -1 0 0 0 -1 $\mathbf{4}_\mathrm{b}$ 0 0 1 0 1
$\mathbf{2}_\mathrm{b}$ -1 0 1 -1 -1 $\mathbf{5}$ -1 0 0 -1 -2
$\mathbf{3}_\mathrm{a}$ -1 1 1 -1 0 $\mathbf{6}_\mathrm{a}$ -1 0 0 1 0
$\mathbf{3}_\mathrm{b}$ 0 0 0 0 0 $\mathbf{6}_\mathrm{b}$ -1 0 2 -1 0
: Processes that comprise most of the nearest neighbour exchanges observed in the simulations. The columns $i, j, k$ and $l$ use the same atom labels as figure \[fig:T1\] and give the topological charge on each atom *before* the exchange occurs. The number in bold refers to the label used in figure \[fig:temp-dependence\]. Exchanges which are the mutual reverse of one other are labelled with the subscripts $a$ and $b$.[]{data-label="tab:processes"}
![(a) Temperature dependence of the probability of the six most frequent T1 moves. The labelled curves refer to the moves in listed table \[tab:processes\]; the thick solid line indicates the probability of all other moves. For moves $\mathbf{2}, \mathbf{3}, \mathbf{4}$ and $\mathbf{6}$ both sets of curves are plotted, but are almost indistinguishable from one another. (b) Number of T1 moves per particle per unit time.[]{data-label="fig:temp-dependence"}](figure6a.eps "fig:"){width="0.34\columnwidth"} ![(a) Temperature dependence of the probability of the six most frequent T1 moves. The labelled curves refer to the moves in listed table \[tab:processes\]; the thick solid line indicates the probability of all other moves. For moves $\mathbf{2}, \mathbf{3}, \mathbf{4}$ and $\mathbf{6}$ both sets of curves are plotted, but are almost indistinguishable from one another. (b) Number of T1 moves per particle per unit time.[]{data-label="fig:temp-dependence"}](figure6b.eps "fig:"){width="0.35\columnwidth"}
Having identified moves as they occur, the next stage in understanding the dynamics of the liquid is to examine the spatial distribution of nearest neighbour exchanges. In figure \[fig:moves\](a), we show the distribution of the first 1000 T1 moves observed after the $T=0.360$ snapshot shown in figure \[fig:configurations\] was taken. These are strongly localised around the mixed regions of +1 and -1 defects that were previously observed in figure \[fig:configurations\]. Figure \[fig:moves\](b) shows that there is not a strong correlation between the number of T1 moves observed for a given particle and the distance that it translates. In fact there seem to be a number of particles that displace relatively large distances during the short simulations without any T1 moves. However, the maximum displacement for particles that experience large numbers of T1 moves does tend to be smaller than those with fewer T1 moves.
One problem with the method that we present here is that we have no simple way of accounting for quartets of atoms are in almost square conformations. Small fluctuations around particle positions could then lead to the mis-counting of a large number of T1 moves. To address this we have repeated the simulations and changed the frequency with which the Voronoi procedure is performed and do not find any changes in the overall rate of T1 moves. Such counting problems are therefore unlikely to be significant.
![(a) Distribution of 1000 T1 moves (white circles) at $T=0.360$ superimposed on the positions of the large atoms in the starting configuration (black circles) (b) Correlation between the number of T1 moves that an atom undertakes and the mean squared displacement, $\Delta r^{2}$ (inset) Probability distribution function of $\Delta r^{2}$.[]{data-label="fig:moves"}](figure7a.eps "fig:"){width="0.35\columnwidth"} ![(a) Distribution of 1000 T1 moves (white circles) at $T=0.360$ superimposed on the positions of the large atoms in the starting configuration (black circles) (b) Correlation between the number of T1 moves that an atom undertakes and the mean squared displacement, $\Delta r^{2}$ (inset) Probability distribution function of $\Delta r^{2}$.[]{data-label="fig:moves"}](figure7b.eps "fig:"){width="0.35\columnwidth"}
Conclusions {#sec:discussion}
===========
We have presented results of simulations of a two dimensional glass-forming binary fluid that shows identifiable changes in its structure when the control temperature is lowered. These changes are related to the spatial distribution of atoms with coordination that differs from the mean value. Using the topological properties of the two-dimensional nearest neighbour network we have shown that it is possible to identify a strong relationship between one aspect of the spatial distribution of dynamics and the local level of disorder in structure. There is apparently still motion within the simulation cell that is not accounted for by the topological process that we study, but our work may contribute to understanding of the puzzle presented by heterogeneous dynamics in slow and glassy systems. The observation that the rate of T1 moves decreases linearly with temperature, with the relative proportions of different types of T1 moves fairly insensitive to temperature, suggests that the slowing down associated with the glass transition is likely not to be able to be ascribed to any one set of T1 processes. Recent work has suggested that a particular class of “liquid-like” defects (which presumably have some set of associated topological moves) may play a role in the glass transition [@aharonov07], but we have not explored this possibility here. However, the temperature dependence of the time decay of the correlations in topological charge, and its similarity to the decay of the dynamic structure factor suggest that the dynamics of T1 moves may be relevant to dynamic facilitation as is required for kinetically constrained models of glasses.
Michael Plischke is thanked for encouragement and helpful discussions. NSERC and Westgrid are thanked for research funding and computing facilities.
References {#references .unnumbered}
==========
[99]{}
Shintani H and Tanaka H 2006 [*Nature Phys.*]{} [**2**]{} 200
Truskett TM, Torquato S, Sastry S, Debenedetti PG and Stillinger FH 1998 [*Phys. Rev. E*]{} [**58**]{} 3083
Sillescu H 1999 [*J. Non-Cryst. Solids* ]{} [**243**]{} 81
Ediger MD 2000 [*Annu. Rev. Phys. Chem.*]{} [**51**]{} 99
Glotzer SC 2000 [*J. Non-Cryst. Solids*]{} [**274**]{} 342
Richert R 2002 [*J. Phys.: Condens. Matter*]{} [**14**]{} R703
Castillo HE, Chamon C, Cugliandolo LF and Kennett MP 2002 [*Phys. Rev. Lett.*]{} [**88**]{} 237201
Vidal-Russell E and Israeloff NE 2000 [*Nature*]{} [**408**]{} 695
Deschenes LA and Vanden Bout DA 2001 [*Science*]{} [**292**]{} 255
Reinsberg SA, Qiu XH, Wilhelm M, Spiess HW and Ediger MD 2001 [*J. Chem. Phys.*]{} [**114**]{} 7299
Weeks ER and Weitz DA 2002 [*Phys. Rev. Lett.*]{} [**89**]{} 095704
Kob W, Donati C, Plimpton SJ, Poole PH and Glotzer SC 1997 [*Phys. Rev. Lett.*]{} [**79**]{} 2827
Donati C, Douglas JF, Kob W, Plimpton SJ, Poole PH and Glotzer SC 1998 [*Phys. Rev. Lett.*]{} [**80**]{} 2338
Donati C, Glotzer SC, Poole PH, Kob W and Plimpton SJ 1999 [*Phys. Rev. E*]{} [**60**]{} 3107
Perera DN and Harrowell P 1996 [*Phys. Rev. E*]{} [**54**]{} 1652
Yamamoto R and Onuki A 1998 [*Phys. Rev. Lett.*]{} [**81**]{} 4915
Andersen HC 2005 [*Proc. Nat. Acad. Sci.*]{} [**102**]{} 6686
Perera DN and Harrowell P 1998 [*J. Non-Cryst. Solids*]{} [**235**]{} 314
Gebremichael Y, and Vogel M and Glotzer SC 2004 [*J. Chem. Phys.*]{} [**120**]{} 4415
Stevenson JD, Schmailian J and Wolynes PG 2006 [*Nature Phys.*]{} [**2**]{} 268
Kegel WK and van Blaarderen A 2000 [*Science*]{} [**287**]{} 290
Weeks ER, Crocker JC, Levitt AC, Schofield A and Weitz DA 2000 [*Science*]{} [**287**]{} 627
Widmer-Cooper A, Harrowell P and Fynewever H 2004 [*Phys. Rev. Lett.*]{} [**93**]{} 135701
Widmer-Cooper A and Harrowell P 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 185701
Berthier L and Jack RL 2007 [*Phys. Rev. E*]{} [**76**]{} 041509
Ritort F and Sollich P 2003 [*Adv. Phys.*]{} [**52**]{} 219
Fredrickson GH and Andersen HC 1984 [*Phys. Rev. Lett.*]{} [**53**]{} 1244
Fredrickson GH and Andersen HC 1985 [*J. Chem. Phys.*]{} [*83*]{} 5822
Graham IS, Piché L and Grant M 1997 [*Phys. Rev. E*]{} [**55**]{} 2132
Garrahan JP and Chandler D 2002 [*Phys. Rev. Lett.*]{} [**89**]{} 035704
Berthier L and Garrahan JP 2003 [*J. Chem. Phys.*]{} [*119*]{} 4367
Garrahan JP and Chandler D 2003 [*Proc. Nat. Acad. Sci.*]{} [**100**]{} 9710
Léonard S, Mayer P, Sollich P, Berthier L and Garrahan JP 2007 [*J. Stat. Mech.: Theory Exp.*]{} 07017
Berthier L and Garrahan JP 2003 [*Phys. Rev. E*]{} [**68**]{} 041201
Downton MT and Kennett MP 2007 [*Phys. Rev. E*]{} [**76**]{} 031502
Aharonov E, Bouchbinder E, Hentschel HGE, Ilyin V, Makedonska N, Procaccia I and Schupper N 2007 [*Europhys. Lett.*]{} [**77**]{} 56002
Davison L and Sherrington D 2000 [*J. Phys. A: Math. Gen.*]{} [**33**]{} 8615
Davison L, Sherrington D, Garrahan JP and Buhot A 2001 [*J. Phys. A: Math. Gen.*]{} [**34**]{} 5147
Sherrington D, Davison L, Buhot A and Garrahan JP 2002 [*J. Phys.: Condens. Matter*]{} [**14**]{} 1673
Perera DN and Harrowell P 1998 [*Phys. Rev. Lett.*]{} [**98**]{} 120
Perera DN and Harrowell P 1999 [*Phys. Rev. E*]{} [**59**]{} 5721
Perera DN and Harrowell P 1999 [*J. Chem. Phys.*]{} [**111**]{} 5441
Sykes MF and Essam JW 1964 [*J. Math. Phys.*]{} [**5**]{} 1117
Nicodemi N and Coniglio A 1998 [*Phys. Rev. E*]{} [**57**]{} R39
Dzugutov M, Simdyankin SI and Zetterling FHM 2002 [*Phys. Rev. Lett.*]{} [**89**]{} 195701
Larson RG 1999 [*The Structure and Rheology of Complex Fluids*]{} (Oxford: Oxford University Press)
Deng D, Argon AS and Yip S 1989 [*Phil. Trans. R. Soc. Lond. A*]{} [**329**]{} 575
[^1]: In practice, with a timestep of $\Delta t = 0.0005\tau$, 94% of T1 moves can be identified at $T=0.840$ and 98% at $T=0.360$. The remaining moves are groups of 5-8 atoms that change coordination simultaneously.
| {
"pile_set_name": "ArXiv"
} |
---
address: 'Department of Mathematics, California Institute of Technology, Pasadena, CA 91125'
author:
- Mayank Pandey
title: On the mean value of the magnitude of an exponential sum involving the divisor function
---
Introduction
============
For a sequence $(a_n)_{n\ge 1}$ of arithmetical interest, it is often desirable in additive number theory to have an understanding of the $L^p$ norms of the exponential sum $M(\alpha) = \sum_{n\le X} a_ne(n\alpha)$ for large $X$. A sufficiently good understanding of these can also lead to estimates for the measure of $\{\alpha\in [0, 1] : |M(\alpha)| > \lambda\}$ for $\lambda$ of an appropriate size.
The case of $L^1$ norms has recieved particular attention, and there are few tools to study them. For these problems, obtaining good lower bounds is particularly difficult in many cases. Littlewood conjectured, and McGehee, Pigno, and Smith [@MPS] proved that if $S$ is some set of $n$ integers, then $\int_0^1 |\sum_n \mathbf 1_S(n)e(n\alpha)|d\alpha\gg\log n$, and this is optimal. For example, one may take $S$ to be some arithemtic progression.
For sequences sensitive to the multiplicative structure of $n$, one expects the true value of the $L^1$ norm to be closer to the $L^2$ norm. This is what happens in the case of sequences with elements chosed uniformly at random from $\Set{-1, 1}$ by Khintchine’s inequality. It is reasonable to expect that, if the coefficients are sensitive to multiplicative structure, then they should behave randomly with regards to additive considerations.
In the case of primes, following Vaughan [@Va] and Goldston [@G], we have that $$\sqrt{X}\ll\int_0^1 |\sum_{n\le X} \Lambda(n)e(n\alpha)| d\alpha\le \left(\frac{\sqrt{2}}{2} + o(1)\right)\sqrt{X\log X}.$$
Other sequences have also recieved attention, such as the indicator function of $r$-free numbers [@BGP], [@BR], and the Möbius function [@BR], [@BR2], [@BP]. In the case of “$\GL(1)$ sequences" (i.e. characters), it is easy to show that if $\chi$ is some Dirichlet character, then $$\int_0^1 \bigg|\sum_{n\le X}\chi(n)e(n\alpha)\bigg|d\alpha\sim C_{\chi}\log X$$ for some $C_{\chi} > 0$. Similar estimates for sequences such as $\chi(n)n^{it}$ are also quite simple to show.
In this paper, we make the first such contribution towards the $\GL(2)$ case by finding an asymptotic formula for the exponential sum with $\tau(n) = \sum_{d|n} 1$ as its coefficients. In particular, with $$M(\alpha) = \sum_{n\le X} \tau(n)e(n\alpha),$$ we prove the following result.
We have that $$\int_0^1 |M(\alpha)| = C\sqrt{X}\log X + O(\sqrt{X})$$ where
$$C = \frac{18}{\pi^3} - \frac{12\log 2}{\pi^3} - \frac{1}{2\pi}\approx 0.153\dots.$$
This improves on the bound $$\sqrt{X}\ll \int_0^1 |M(\alpha)|d\alpha\ll \sqrt{X}\log X.
\label{eq:prevResult}$$ obtained by Goldston and the author in [@GP]. This problem was previously attempted by Brüdern [@Br], who claimed to have shown that the lower bound from (\[eq:prevResult\]) reflected the true order of magnitude. It turned out that there was an error in the proof of a key lemma, which turned out to be false.
Let us now briefly describe the main ideas of the proof. We use Kloosterman’s refinement of the circle method. This leads us to require good estimates on $M\left(\frac{a}{q} + \beta\right)$ for $q\le\sqrt{X}, |\beta|\ll (q\sqrt X)^{-1}$.
To obtain such estimates, we first replace the cutoff $\mathbf 1_{[1, X]}(n)$ with $w\left(\frac nX\right)$ for $w$ some smooth function supported on $[0, 1]$ equal to $1$ on $(\eta, 1 - \eta)$ for some sufficiently small $\eta$. Due to the fact that the size of $\tau(n)$ may be large for various points, we require that $\eta$ depends on $X$ and is at most $o\left(\frac{1}{\log X}\right)$. The fact that we cannot suppose that the derivatives $w^{(j)}$ are $O_j(1)$ for all $j$ requires us to be quite careful in some parts of the argument.
At this point, we use Voronoi summation to obtain that $\sum_{n} \tau(n)e\left(\frac{an}{q}\right)e(n\beta)w(n/X)$ is equal to a relatively easily understood main term plus an error term roughly of the form $\frac{X}{q}\sum_{n} \tau(|n|)e\left(\frac{\conj an}{q}\right)f_{q, \beta}(n)$ where $f_{q, \beta}$ is some somewhat well-understood function that can be thought of as (with the dependence on $\beta$ largely coming from bounds on derivatives of $w$) being concentrated near the points $|n|\ll\frac{q^2}{X}\cdot X^{\eps}$.
We may bound this error term quite straightforwardly, using integration by parts to control the complications that arise if $\beta$ is large and $q$ is small. However even assuming that $w$ does not depend on $X$ (so that $w^{(j)}\ll_j 1$), the bound one obtains is off from what we desire by a power of $\log X$, due to the potentially large size of $\tau(n)$. Furthermore, poor bounds for these derivatives lead to poor bounds for $f_{q, \beta}$ when one integrates by parts to control the oscillation of $e(x\beta)$ and the Bessel functions in the definition of $f_{q, \beta}$.
For the rest of the outline, we shall focus on the most difficult case when $|\beta|\asymp 1/X, q\asymp\sqrt{X}$. Note that $f_{q, \beta}$ does not depend on $a$, so on average, as $a$ ranges over coprime residue classes $\mod q$, there must be cancellation between summands in the error term for most $a$. This is exploited by Cauchy-Schwarz and orthogonality, giving a saving of $X^{1/4}$ for large $n$. The problem then reduces to bounding the contribution from the first $O(1)$ values of $n$, so the potentially large size of $\tau(n)$ ceases to be a issue. In addition, the reduced effective range of $n$ allows us to treat the remaining terms differently in a way which bypasses the issues that come from the dependence of $w$ on $X$ that arise when trying to control the size of $f_{q, \beta}$. In the end, we obtain a bound of $O(X^{1/2})$ on average for the error term, which suffices.
Let us close by making a few comments on possible extensions of this work. Obtaining an asymptotic formula when $a_n = \lambda_f(n)$, the $n$th Fourier coefficient of some holomorphic cusp form $f$, seems significantly harder. The exponential sum is small everywhere, since there is no “main term" around rationals with small denominators like we have in the case of the divisor function. However, determining the $L^1$ norm up to a constant factor is straightforward. In particular, the estimate $\int_0^1 |\sum_{n\le X} \lambda_f(n)e(n\alpha)| d\alpha\asymp \sqrt{X}$ follows from Hölder’s inequality, the asymptotic $\sum_{n\le X} |\lambda_f(n)|^2\sim c_fX$ for some $c_f > 0$, and the pointwise bound $\sum_{n\le X} \lambda_f(n)e(n\alpha)\ll \sqrt{X}$ of Jutila [@J].
Another natural extension is to consider $\tau_3(n) = \sum_{d_1d_2d_3 = n} 1$ instead of $\tau(n)$. The problem becomes significantly harder in this case however if one follows the same methods, since the effective length of the error term from Voronoi summation is significantly longer.
In the proof, some of our estimates, particularly in our treatment of $\beta$, are more precise than necessary for the benefit of those who may try to improve our result.
Setup
=====
Let $\eta = \eta(X)$ be a parameter to be decided on later, and let $w$ be a smooth function taking values in $[0, 1]$ supported on $(1/(2X), 1)$ such that $$\begin{aligned}
&w(u) = 1\text{ for }1/X\le u \le (1 - \eta)\\
&w^{(j)}(u)\ll_j \eta^{-j}\text{ for }j\in\{0, 1, 2\}, u > 1/X\\
&w^{(j)}(u)\ll_j X^j\text{ for }j\in\{0, 1, 2\}, 1/(2X)\le u\le 1/X.\end{aligned}$$ Instead of working with the $L^1$-norm of $M(\alpha)$, it suffices to work with the $L^1$-norm of $$M^*(\alpha) = \sum_{n} \tau(n)w\left(\frac nX\right)e(n\alpha)$$ if we take $\eta = X^{-\delta}$ with $\delta = \frac{1}{100}$ since we have that by Parseval and Cauchy-Schwarz $$\begin{aligned}
&\bigg|\int_0^1 |M(\alpha)|d\alpha - \int_0^1 |M^*(\alpha)|d\alpha\bigg| \\
&\le \int_0^1 \bigg|\sum_{(1 - \eta)X < n\le X}|1 - w\left(\frac{n}{X}\right)|\tau(n)e(n\alpha)\bigg|d\alpha\\
&\le \bigg(\int_0^1 \bigg|\sum_{(1 - \eta)X < n\le X}|1 - w\left(\frac nX\right)|\tau(n)e(n\alpha)\bigg|^2d\alpha\bigg)^{1/2}\\
&\le \bigg(\sum_{(1 - \eta)X < n \le X} \tau(n)^2\bigg)^{1/2}\ll \eta^{1/2}X^{1/2 + \eps}.\end{aligned}$$ Let $Q = \sqrt{X}$. Then, we have by (20.9) and the proof of 20.7 in [@IK] $$\int_0^1 |M^*(\alpha)|d\alpha =
\sum_{q\le Q}\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}
\left[\int_{-1/(aq)}^0 \bigg|M^*\left(\frac{\conj a}{q} + \beta\right)\bigg|d\beta
+ \int_{0}^{1/(aq)}\bigg|M^*\left(-\frac{\conj a}{q} + \beta\right)\bigg|d\beta\right],
\label{eq:majorArcDecomp}$$ where $\conj a$ is so that $a\conj a\equiv 1\pmod{q}$. From (4.49) in [@IK], we have that for all $\beta$ and $q\le Q, (a, q) = 1$ $$\begin{aligned}
M^*\left(\frac aq + \beta\right) = \sum_{n\ge 1}\tau(n)e\left(\frac{an}q\right)w\left(\frac nX\right)e(n\beta)
&= \frac{1}{q}\int_0^\infty (\log x + 2\gamma - 2\log q)w\left(\frac xX\right)e(x\beta)dx \\
&+ \sum_{n\in\ZZ}e\left(-\frac{\conj an}{q}\right)\Delta(n, q, \beta)\end{aligned}$$ where here $$\Delta(n, q, \beta) =
\begin{cases}
-\frac{2\pi\tau(n)}{q}\int w\left(\frac xX\right)e(x\beta)Y_0\left(\frac{4\pi\sqrt{xn}}{q}\right)dx & n\ge 1\\
0 & n = 0 \\
-\frac{4\tau(-n)}{q}\int w\left(\frac xX\right)e(x\beta)K_0\left(\frac{4\pi\sqrt{x|n|}}{q}\right)dx & n < 0.
\end{cases}$$ Here, $Y_\nu, K_\nu$ are the standard Bessel functions. It follows that $$\int_0^1 |M^*(\alpha)| d\alpha = E + R$$ where $$E = \sum_{q\le Q}\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}\int_{-1/(aq)}^{1/(aq)}\frac{1}{q}|I_q(\beta)|d\beta$$ with $$I_q(\beta) = \int_0^\infty (\log x + 2\gamma - 2\log q)w\left(\frac xX\right)e(x\beta)dx,$$ and $$R\le\sum_{q\le Q}\sumCp_{a(q)} \int_{-1/(qQ)}^{1/(qQ)} |R_{a,q}(\beta)| d\beta$$ where $$R_{a,q}(\beta) = \sum_{n}e\left(-\frac{\conj an}{q}\right)\Delta(n, q, \beta).$$ Here, we write $\sumCp_{a(q)}$ to denote a sum over the coprime residue classes $a$ modulo $q$. Then, the main theorem follows from the following two lemmas, which we prove in the next two sections.
We have $$R\ll\sqrt{X}.$$ \[lem:errBound\]
We have $$E = C\sqrt{X}\log X + O(\sqrt{X}).$$ \[lem:mainTermBound\]
The error term
==============
In order to prove Lemma \[lem:errBound\], we shall in fact show the following bound for a very large range of choices for $\eta$.
We have that for all $q\le Q, |\beta|\le 1/(qQ)$, $$\begin{aligned}
{\sum_{a(q)}}^* |R_{a, q}(\beta)|\ll q^2 + &\min(q^{3/2}X^{1/4}, \eta^{-1}q^{5/2}X^{-1/4}) + |\beta|^2q^{5/2}X^{7/4} + |\beta|^2q^{3/2}X^{7/4}\log^{3/2} (q + 2) \\
&+ |\beta|q^{3/2}X^{3/4}\log^{3/2}(q + 2) + \eta^{-1}q^{3/2}X^{-1/4}\log^{3/2}(q + 2).$$ \[lem:majorArcCancel\]
Specializing to our choice of $\eta = X^{-1/100}$ we have that by Lemma \[lem:majorArcCancel\] $$\begin{aligned}
R &= \sum_{q\le Q}{\sum_{a(q)}}^* \int_{-1/(qQ)}^{1/(qQ)} |R_{a,q}(\beta)| d\beta\\
&\ll \sum_{q\le Q}\int_{-1/(qQ)}^{1/(qQ)} q^{3/2}X^{1/4} + |\beta|^2X^{7/4}(q^{5/2} + q^{3/2}\log^{3/2}(q + 2)) + |\beta|q^{3/2}X^{3/4}\log^{3/2}(q + 2) d\beta \\
&\ll \sum_{q\le Q} q^{1/2}X^{-1/4} + X^{1/4}q^{-1/2} + X^{1/4}q^{-3/2}\log^{3/2}(q + 2) + q^{-1/2}X^{-1/4}\log^{3/2}(q + 2)\\
&\ll X^{1/2}\end{aligned}$$ as desired.
Note that by Cauchy-Schwarz, we have $$\begin{aligned}
R_q(\beta) &= {\sum_{a(q)}}^*\bigg|\sum_{n} e\left(-\frac{\conj an}{q}\right)\Delta(n, q, \beta)\bigg|\\
&\le\sum_{m\in\ZZ}{\sum_{a(q)}}^*
\bigg|\sum_{qm\le n < q(m + 1)} e\left(-\frac{\conj an}{q}\right)\Delta(n, q, \beta)\bigg|\\
&\le \sum_{m\in\ZZ}q^{1/2}\left({\sum_{a(q)}}^*\bigg|\sum_{qm\le n < q(m + 1)} e\left(-\frac{an}{q}\right)\Delta(n, q, \beta)\bigg|^2\right)^{1/2}\end{aligned}$$
Let $$B_{q, m}(\beta) = {\sum_{a(q)}}^*\bigg|\sum_{qm\le n < q(m + 1)} e\left(-\frac{an}{q}\right)\Delta(n, q, \beta)\bigg|^2.$$ Note that then we have $$\begin{aligned}
B_{q, m}(\beta) &= {\sum_{a(q)}}^*\bigg|\sum_{qm\le n < q(m + 1)} e\left(-\frac{an}{q}\right)\Delta(n, q, \beta)\bigg|^2\le
{\sum_{a(q)}}\bigg|\sum_{qm\le n < q(m + 1)} e\left(-\frac{an}{q}\right)\Delta(n, q, \beta)\bigg|^2\\
&= {\sum_{a(q)}}\sum_{qm\le n_1, n_2 < q(m + 1)} e\left(-\frac{a(n_1 - n_2)}{q}\right)\Delta(n_1, q, \beta)\conj{\Delta(n_2, q, \beta)}\\
&= \sum_{qm\le n_1, n_2 < q(m + 1)} \Delta(n_1, q, \beta)\conj{\Delta(n_2, q, \beta)}{\sum_{a(q)}}e\left(-\frac{a(n_1 - n_2)}{q}\right)\\
&= q\sum_{qm\le n < q(m + 1)} |\Delta(n, q, \beta)|^2\end{aligned}$$ so it follows that $$R_q(\beta)\le\sum_{m\in\ZZ}q\left(\sum_{qm\le n < q(m + 1)}|\Delta(n, q, \beta)|^2\right)^{1/2}.
\label{eq:q_err_bound}$$
Integrating by parts twice, and noting that $w$ is supported on $[1/(2X), 1]$ we obtain (with $B_\nu$ denoting either $Y_\nu$ or $K_\nu$ depending on the sign of $n$) that for all $n\ne 0$ $$\begin{aligned}
\int_0^{\infty} & e(x\beta)w\left(\frac xX\right)B_0\left(\frac{4\pi\sqrt{x|n|}}{q}\right)dx\\
&= \frac{q^2}{4\pi^2|n|}\int_{1/2}^{\infty} e(x\beta)(X^{-2}w'' + 4\pi iX^{-1}\beta w' - 4\pi^2\beta^2 w)\left(\frac xX\right)\left(xB_2\left(\frac{4\pi\sqrt{x|n}|}{q}\right)\right)dx. {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:intpart1}\end{aligned}$$ We have that from the bounds on $w$ that $$\begin{aligned}
(X^{-2}w'' + 4\pi i\beta X^{-1}w' - & 4\pi^2\beta^2 w)\left(\frac xX \right)\\
&\ll |\beta|^2 + \mathbf 1_{[1/2, 1]}(x) + (\eta^{-1}|\beta|X^{-1} + \eta^{-2}X^{-2})\mathbf 1_{[(1 - \eta)X, X]}(x).{\addtocounter{equation}{1}\tag{\theequation}}\label{eq:deriv_bounds}\end{aligned}$$ so from (\[eq:intpart1\]) we obtain that for $|n|\gg q^2$, by (4.9) in [@Zh] $$\begin{aligned}
&\int_0^{\infty} e(x\beta)w\left(\frac{x}{X}\right)B_0\left(\frac{4\pi\sqrt{x|n|}}{q}\right)dx\\
&\ll\frac{q^2}{|n|}\int_{1/2}^{X}
\big(|\beta|^2 + \mathbf 1_{[1/2, 1]}(x) + (\eta^{-1}|\beta|X^{-1} + \eta^{-2}X^{-2})\mathbf 1_{[(1 - \eta)X, X]}(x)\big)
\bigg|xB_2\left(\frac{4\pi\sqrt{x|n}|}{q}\right)\bigg| dx\\
&\ll\frac{q^2}{|n|}\bigg(\int_{1/2}^{X} |\beta|^2q^{1/2}x^{3/4}|n|^{-1/4}dx + \int_{1/2}^{1}q^{1/2}|n|^{-1/4} dx + \frac{1}{\eta X}\int_{(1 - \eta)X}^X (|\beta| + \eta^{-1}X^{-1})q^{1/2}x^{3/4}|n|^{-1/4}dx\bigg)\\
&\ll |\beta|^2q^{5/2}|n|^{-5/4}X^{7/4} + q^{5/2}|n|^{-5/4} + |\beta|q^{5/2}X^{3/4}|n|^{-5/4} + \eta^{-1}q^{5/2}X^{-1/4}$$ so we have that $$\Delta(n, q, \beta)\ll \tau(|n|)(|\beta|^2q^{3/2}|n|^{-5/4}X^{7/4} + q^{3/2}|n|^{-5/4} + |\beta|q^{3/2}X^{3/4}|n|^{-5/4} + \eta^{-1}q^{3/2}X^{-1/4}|n|^{-5/4}).$$ It follows that for $q^{100} > |m| > q$, by a result of Shiu [@Sh] $$\begin{aligned}
\sum_{qm\le n < q(m + 1)}|\Delta(n, q,\beta)|^2&\ll \log^3 m(|\beta|^4X^{7/2} + 1 + |\beta|^2X^{3/2} + \eta^{-2}X^{-1/2})q^4(q|m|)^{-5/2}
$$ and for $|m|\ge q^{100}$, the same bound holds but with $m^\eps$ in place of $\log^3 m$ by the divisor bound $\tau(d)\ll_\eps d^\eps$, so $$\begin{aligned}
\sum_{|m| > q}&\bigg(\sum_{qm\le n < q(m + 1)}|\Delta(n, q,\beta)|^2\bigg)^{1/2}\\ &\ll
q^{1/2}\log^{3/2} (q + 2)(1 + |\beta|^2X^{7/4} + |\beta|X^{3/4} + \eta^{-1}X^{-1/4}).{\addtocounter{equation}{1}\tag{\theequation}}\label{eq:tailTerms}
$$
From now on, until specified otherwise, we shall restrict ourselves to $|n|\le q^{2}$, $|m|\ll q$.
For $n\ne 0$, by integration by parts $$\Delta(n, q, \beta)\ll\frac{\tau(n)}{q}(|\Delta_1(n)| + |\Delta_2(n)| + |\Delta_3(n)|)$$ where $$\Delta_1
= \Delta_1(n, q, \beta) = \frac{q\beta}{\sqrt{|n|}}\int_0^\infty e(x\beta)w\left(\frac{x}{X}\right)x^{1/2}B_1\left(\frac{4\pi\sqrt{x|n|}}{q}\right)dx$$ $$\Delta_2
= \Delta_2(n, q, \beta) = \frac{q}{X\sqrt{|n|}}\int_{(1 - \eta)X}^X e(x\beta)w'\left(\frac{x}{X}\right)x^{1/2}B_1\left(\frac{4\pi\sqrt{x|n|}}{q}\right)dx$$ $$\Delta_3
= \Delta_3(n, q, \beta) = \frac{q}{X\sqrt{|n|}}\int_{1/2}^1 e(x\beta)w'\left(\frac{x}{X}\right)x^{1/2}B_1\left(\frac{4\pi\sqrt{x|n|}}{q}\right)dx.$$ We have, from (4.9) in [@Zh] and integration by parts that $$\begin{aligned}
\Delta_1 &= \frac{q^2\beta}{|n|}\int \left(2\pi i\beta e(x\beta)w\left(\frac xX\right) + e(x\beta)X^{-1}w'\left(\frac xX\right) \right)\left(xB_2\left(\frac{4\pi\sqrt{x|n|}}{q}\right)\right)dx\\
&\ll q^2|\beta||n|^{-5/4}(I_1 + |n|^{-3/4}I_2) $$ where $$I_1 = \int_{q^2/|n|}^{X} \left(|\beta| + X^{-1}\abS{w'\left(\frac xX\right)}\right)q^{1/2}x^{3/4}dx,$$ and $$I_2 = \int_{1/2}^{q^2/|n|} \left(|\beta| + X^{-1}\abS{w'\left(\frac{x}{X}\right)}\right)q^{2}dx.$$ We have that $$I_1\ll q^{1/2}X^{3/4}\int_{(1 - \eta)X}^X X^{-1}\abS{w'\left(\frac{x}{X}\right)}dx + q^{1/2}|\beta|X^{3/4}\int_{1}^X dx\ll q^{1/2}X^{3/4} + |\beta|q^{1/2}X^{7/4}$$ and $$I_2\ll |\beta|q^2|n|^{-1} + q^2\ll q^2.$$ Combining these, we obtain that $$\Delta_1\ll |\beta|q^{5/2}X^{3/4}|n|^{-5/4} + |\beta|^2q^{5/2}X^{7/4}|n|^{-5/4}.
\label{eq:delta1Bound}$$ Similarly, integrating by parts, we obtain that $$\begin{aligned}
\Delta_2 &\ll \frac{q^2}{|n|}\int_{(1 - \eta)X}^X \left(|\beta|X^{-1}\abS{w'\left(\frac xX\right)} + X^{-2}\abS{w''\left(\frac xX\right)}\right)\bigg|xB_2\left(\frac{4\pi\sqrt{x|n|}}{q}\right)\bigg| dx\\
&\ll \frac{q^2}{|n|}\int_{(1 - \eta)X}^X (|\beta|\eta^{-1}X^{-1} + \eta^{-2}X^{-2})\bigg|xB_2\left(\frac{4\pi\sqrt{x|n|}}{q}\right)\bigg| dx\\
&\ll|\beta|q^{5/2}X^{3/4}|n|^{-5/4} + \eta^{-1}q^{5/2}X^{-1/4}|n|^{-5/4} {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:delta2Bound_n_aspect}.\end{aligned}$$ In addition, we have from (2.5) in [@Zh] the bound $B_1(x)\ll x^{-1/2}$ for $x\gg 1$, so we obtain that $$\begin{aligned}
\Delta_2 &\ll \frac{q}{\sqrt{|n|}}\int_{(1 - \eta)X}^X \eta^{-1}X^{-1}x^{1/2}(q^{1/2}x^{-1/4}|n|^{-1/4})dx\ll q^{3/2}X^{1/4}|n|^{-3/4} {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:delta2Bound_lastTerms}. $$ This bound is superior to (\[eq:delta2Bound\_n\_aspect\]) for $|n|\ll \eta^{-1}$ in the critical case that $q\asymp Q, |\beta|\asymp 1/X$. Also, from the bound $B_1(y)\ll y^{-1}$, we obtain that $$\Delta_3\ll \frac{q^2}{|n|}.
\label{eq:delta3Bound}$$ Therefore, combining (\[eq:delta1Bound\]), (\[eq:delta2Bound\_n\_aspect\]), (\[eq:delta2Bound\_lastTerms\]), and (\[eq:delta3Bound\]) we have that $$\begin{aligned}
|\Delta(n, q, \beta)|\ll \tau(|n|)\bigg(\frac{q}{|n|} &+ |\beta|q^{3/2}X^{3/4}|n|^{-5/4} + |\beta|^2q^{3/2}X^{7/4}|n|^{-5/4} \\
&+ \min\left(\eta^{-1}q^{3/2}X^{-1/4}|n|^{-5/4}, q^{1/2}X^{1/4}|n|^{-3/4}\right)\bigg),\end{aligned}$$ so we have that $$\begin{aligned}
\sum_{|n|\le 2q} |\Delta(n, q, \beta)|^2
&\ll \sum_{0 < n\le 2q} q^2\frac{\tau(n)^2}{n^2} + |\beta|^2q^3X^{3/2}\frac{\tau(n)^2}{n^{5/2}} + |\beta|^4q^3X^{7/2}\frac{\tau(n)^2}{n^{5/2}} + \min(qX^{1/2}, \eta^{-2}q^{3}X^{-1/2})\frac{\tau(n)^2}{n^{3/2}}\\
&\ll q^2 + |\beta|^2q^3X^{3/2} + |\beta|^4q^3X^{7/2} + \min(qX^{1/2}, \eta^{-2}q^3X^{-1/2}), {\addtocounter{equation}{1}\tag{\theequation}}\label{eq:firstTerms}\end{aligned}$$ and for $1 < |m|\le q$ $$\begin{aligned}
\sum_{qm\le n < q(m + 1)}&|\Delta(n, q, \beta)|^2\\
&\ll q\log^3 (q + 2)\bigg(q^2\cdot\frac{1}{|m|^2q^2} + |\beta|^4q^3X^{3/2}\cdot\frac{1}{|m|^{5/2}q^{5/2}} + \eta^{-2}q^3X^{-1/2}\cdot\frac{1}{q^{5/2}|m|^{5/2}}\bigg)\\
&\ll \frac{q\log^3(q + 2)}{|m|^2} + \frac{|\beta|^4q^{3/2}X^{3/2}}{|m|^{5/2}} + \frac{\eta^{-2}q^{3/2}X^{-1/2}}{|m|^{5/2}}.\end{aligned}$$ Therefore, it follows that $$\begin{aligned}
&\sum_{1 < |m|\le q}q\bigg(\sum_{qm\le n < q(m + 1)}|\Delta(n, q, \beta)|^2\bigg)^{1/2}\\
&\ll q\log^{3/2} (q + 2)\sum_{1 < |m|\le q}q^{1/2}\frac{1}{|m|} + |\beta|^2q^{3/4}\log^{3/2}(q + 2)X^{3/4}\frac{1}{|m|^{5/4}} + \eta^{-1}q^{3/4}\log^{3/2}(q + 2)X^{-1/4}\frac{1}{|m|^{5/4}}\\
&\ll q^{3/2}\log^{5/2} (q + 2) + |\beta|^2q^{7/4}X^{3/4}\log^{3/2}(q + 2) + \eta^{-1}q^{7/4}X^{-1/4}\log^{3/2}(q + 2).\end{aligned}$$ Putting this bound together with (\[eq:firstTerms\]), (\[eq:tailTerms\]), and (\[eq:q\_err\_bound\]) yields that $$\begin{aligned}
R_q(\beta)\ll q^2 + &\min(q^{3/2}X^{1/4}, \eta^{-1}q^{5/2}X^{-1/4}) + |\beta|^2q^{5/2}X^{7/4} + |\beta|^2q^{3/2}X^{7/4}\log^{3/2} (q + 2) \\
&+ |\beta|q^{3/2}X^{3/4}\log^{3/2}(q + 2) + \eta^{-1}q^{3/2}X^{-1/4}\log^{3/2}(q + 2).{\addtocounter{equation}{1}\tag{\theequation}}\label{eq:final_bound}\end{aligned}$$ as desired.
The main term
=============
In this section, we prove Lemma \[lem:mainTermBound\]. Note that we have that $$$$ $$E = \sum_{q\le Q}\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}\int_{-1/(aq)}^{1/(aq)}\frac{1}{q}|I_q(\beta)|d\beta
= 2\sum_{q\le Q}\frac{1}{q}\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}\int_0^{1/(aq)}|I_q(\beta)|d\beta$$ For $q\le Q$, we have that by the triangle inequality $$\begin{aligned}
|I_q(\beta)|&\le \int_{1/2}^X |\log x + 2\gamma - 2\log q| dx\\ &\ll \int_{1/2}^{q^2} \log(q^2/x)dx + \int_{q^2}^X \log(x/q^2)dx + O(X)\\
&= q^2\log(q^2) - (q^2\log(q^2) - q^2) + X\log X - X - q^2\log q^2 + q^2 - (X - q^2)\log q^2 + O(X)\\
&= X\log(X/q^2) + O(X).\end{aligned}$$ It can then be checked that the contribution due to $\beta < 1/X$ can be disregarded. Indeed, we have that $$\sum_{q\le Q} \frac{\varphi(q)}{q}\int_{0}^{1/X} |I_q(\beta)|d\beta
\le \frac{1}{X}\sum_{q\le Q} X(\log(X/q^2) + 1)\ll X^{1/2}.$$ We shall therefore now restrict our attention to $\beta\ge 1/X$. Note that we have that by integration by parts $$I_q(\beta) = \int (\log x + 2\gamma - 2\log q)w\left(\frac xX\right)e(x\beta)dx = I_q^1(\beta) + I_q^2(\beta)$$ where $$I_q^1(\beta) = \frac{1}{2\pi i\beta}\int \frac{w\left(\frac xX\right)}{x}e(x\beta)dx$$ $$I_q^2(\beta) = \frac{1}{2\pi i\beta}\int_{(1 - \eta)X}^X (\log x + 2\gamma - 2\log q)X^{-1}w'\left(\frac xX\right)e(x\beta)dx.$$ We have that $$I_q^2(\beta)\ll \frac{1}{\beta \eta X}\int_{(1 - \eta)X}^X \log x + 2\gamma - 2\log q dx\ll \frac{1}{\beta}(\log(X/q^2) + 2\gamma).$$ Also, we have that $$I_q^1(\beta) =
\frac{1}{2\pi i\beta}\int_{1/2}^{\beta^{-1}} \frac{w\left(\frac xX\right)}{x} dx + O\left(\beta^{-1}\int_{1/2}^{\beta^{-1}}\frac{|e(x\beta) - 1|}{x} dx
+ \beta^{-1}\bigg|\int_{\beta^{-1}}^X \frac{w\left(\frac xX\right)}{x}e(x\beta)dx\bigg|\right).$$ From the inequality $|e(\alpha) - 1|\le |\alpha|$, it follows that $$\int_{1/2}^{\beta^{-1}}\frac{|e(x\beta) - 1|}{x} dx\le 1$$ and by integration by parts, we have that for $\beta > 1/X$ $$\begin{aligned}
\int_{\beta^{-1}}^X \frac{w\left(\frac xX\right)}{x}e(x\beta)dx
&= \frac{1}{2\pi i}w\left(\frac{\beta^{-1}}{X}\right) + \frac{1}{2\pi i\beta}\int_{\beta^{-1}}^X \frac{w\left(\frac xX\right)}{x^2}e(x\beta)dx - \frac{1}{2\pi i \beta}\int_{(1 - \eta)X}^X
\frac{X^{-1}w'\left(\frac xX\right)}{x}e(x\beta)dx\\
&\ll 1 + \beta^{-1}\int_{\beta^{-1}}^X \frac{1}{x^2}dx + \beta^{-1}\int_{(1 - \eta)X}^X \eta^{-1}X^{-1}x^{-1}dx\ll 1.\end{aligned}$$ Therefore, we have that $$I_q^1(\beta) = \frac{\log(\beta^{-1})}{2\pi i\beta} + O(\beta^{-1})$$ so $$\begin{aligned}
&E = 2\sum_{q\le Q}\frac 1q\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}\int_{1/X}^{1/(aq)} |I_q(\beta)|d\beta + O(X^{1/2}) \\
&= \frac{1}{\pi}\sum_{q\le Q}\frac{1}{q}\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}
\int_{1/X}^{1/(aq)} \frac{\log(\beta^{-1})}{\beta} d\beta
+ O\bigg(\sum_{q\le Q}\frac{\varphi(q)}q\int_{1/X}^{1/(qQ)} \frac{1}{\beta}(\log(X/q^2) + 1)d\beta\bigg).\end{aligned}$$ The term inside the $O(-)$ is $$\ll \sum_{q\le Q} \log(X/q^2)^2 + \log(X/q^2)\ll\sqrt{X}.$$ At this point, the lower bound of $\gg \sqrt{X}\log X$ follows. Indeed, we have $$\sum_{q\le Q}\frac{1}{q}\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}
\int_{1/X}^{1/(aq)} \frac{\log(\beta^{-1})}{2\pi\beta} d\beta\gg \sum_{q\le Q}\frac{\varphi(q)}q((\log X)^2 - \log(2qQ)^2)\gg \sqrt{X}\log X$$ by partial summation.
Now we shall show that with $$\begin{aligned}
S &= \frac{1}{\pi}\sum_{q\le Q}\frac{1}{q}\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}\int_{1/X}^{1/(aq)} \frac{\log(\beta^{-1})}{\beta} d\beta \\
&= \frac{1}{2\pi}\sum_{q\le Q}\frac{1}{q}\sum_{\substack{Q < a\le q + Q\\ (a, q) = 1}}\log^2X - \log^2(aq),\end{aligned}$$ $$S = C\sqrt{X}\log X + O(\sqrt{X})$$ with $C$ as in the main theorem. From the identity $$\mathbf 1_{(a, q) = 1} = \sum_{\substack{d | a\\ d | q}} \mu(d)$$ we obtain that $$S = \frac{1}{2\pi}\sum_{q\le Q}\frac{1}{q}\sum_{Q < a\le q + Q}(\log^2X - \log(aq))\sum_{\substack{d | a\\ d | q}}\mu(d)
= S_2 + S_3$$ where $$S_2 = \frac{1}{2\pi}\sum_{Q\ge d > X^{1/4}}\mu(d)\sum_{\substack{q\le Q\\ d | q}}\frac{1}{q} \sum_{\substack{Q < a\le q + Q \\ d | a}}\log^2X - \log^2(aq),$$ and $$S_3 = \frac{1}{2\pi}\sum_{d\le X^{1/4}}\mu(d)\sum_{\substack{q\le Q\\ d | q}}\frac{1}{q} \sum_{\substack{Q < a\le q + Q \\ d | a}}\log^2X - \log^2(aq).$$ Note that $$\begin{aligned}
S_2&\ll \log^2 X\sum_{Q\ge d > X^{1/4}}\sum_{\substack{q\le Q\\ d | q}}\frac{1}{q} \sum_{\substack{Q < a\le q + Q \\ d | a}} 1\\
&\ll\log^2 X\sum_{Q\ge d > X^{1/4}}\sum_{\substack{q\le Q\\ d | q}}\frac{1}{q}\cdot \frac{Q}{d}\\
&\ll Q\log^3 X\sum_{Q\ge d > X^{1/4}}\frac{1}{d^2}\\
&\ll X^{1/4}\log^3 X.{\addtocounter{equation}{1}\tag{\theequation}}\label{eq:bigdBound}
\end{aligned}$$
In addition, we have that $$\begin{aligned}
S_3 &= \frac{1}{2\pi}\sum_{d\le X^{1/4}}\mu(d)\sum_{\substack{q\le Q\\ d | q}}\frac{1}{q} \sum_{\substack{Q < a\le q + Q \\ d | a}}\log^2X - \log^2(aq)\\
&= \frac{1}{2\pi}\sum_{d\le X^{1/4}}\mu(d)\sum_{q_0\le Q / d}\frac{1}{q_0d} \sum_{Q / d < a_0\le q_0 + Q / d}\log^2X - \log^2(a_0q_0d^2)\\
&= \frac{1}{2\pi}\sum_{d\le X^{1/4}}\mu(d)\sum_{q_0\le Q / d}\frac{1}{q_0d}\bigg(\int_{Q / d}^{q_0 + Q / d}\log^2X - \log^2(yq_0d^2) dy + O(\log^2 X)\bigg)\\
&= \frac{1}{2\pi}\sum_{d\le X^{1/4}}\mu(d)\sum_{q_0\le Q / d}\frac{1}{q_0d}\int_{Q / d}^{q_0 + Q / d}\log^2X - \log^2(yq_0d^2) dy + O(-)\end{aligned}$$ where the error in the $O(-)$ is $$\ll\log^2X\sum_{d\le X^{1/4}}\sum_{q_0\le Q/d}\frac{1}{q_0d}\ll\log^3X\sum_{d\le X^{1/4}}\frac{1}{d}\ll \log^4 X.$$ The main term is equal to $S_4 - S_5$ with $$S_4 = \frac{1}{2\pi}\sum_{d\le X^{1/4}}\frac{\mu(d)}{d}\sum_{q_0\le Q/d} \log^2X,$$ $$S_5 = \frac{1}{2\pi}\sum_{d\le X^{1/4}}\frac{\mu(d)}{d}\sum_{q_0\le Q/d} \frac{1}{q_0}\int_{Q / d}^{q_0 + Q / d}\log^2(yq_0d^2) dy.$$ Note that $$S_4 = \frac{1}{2\pi}Q\log^2X\sum_{d\le X^{1/4}} \frac{\mu(d)}{d^2}(1 + O(X^{-1/4})) = \frac{1}{2\pi\zeta(2)}\sqrt{X}\log^2X + O(X^{1/4}\log^2X).
\label{eq:log2Sum}$$ In addition, we have that $$S_5 = \frac{Q}{2\pi}\sum_{d\le X^{1/4}} \frac{\mu(d)}{d^2}\sum_{q_0\le Q/d} \frac{1}{q_0}\int_{1}^{1 + dq_0/Q}\log^2(ydq_0Q)dy.$$ With $A = dq_0Q$, the inner integral is equal to $$\begin{aligned}
\bigg[&y(\log^2(Ay) - 2\log(Ay) + 2)\bigg]_1^{1 + dq_0/Q} \\
&= \bigg[y(\log^2y + 2(\log A - 1)\log y + 2 - 2\log A + \log^2 A)\bigg]_1^{1 + dq_0/Q}\\
&= \frac{dq_0}{Q}(2 - 2\log A + \log^2 A) + \left(1 + \frac{dq_ 0}Q\right)\left(\log^2\left(1 + \frac{dq_0}Q\right) + 2(\log A - 1)\log\left(1 + \frac{dq_0}Q\right)\right)\end{aligned}$$ so by various manipulations including the identities $$\int_1^{Q/d} \log(1 + td/Q)\log(tdQ)\frac{dt}{t} = -Li_2(-1)\log X+ O(1)$$ and $$\frac{d}{Q}\int_1^{Q/d} \log(1 + td/Q)\log(tdQ)dt = (2\log 2 - 1)\log X + O(1)$$ where $\Li_s$ is the polylogarithm function, we obtain that $$\begin{aligned}
S_5 &= \frac{Q}{2\pi}\sum_{d\le X^{1/4}}\frac{\mu(d)}{d^2}
\int_1^{Q/d}\frac{d}{Q}(2 - 2\log(tdQ) + \log^2 (tdQ)) \\
&\hspace{1cm} + \left(\frac 1t + \frac{d}Q\right)\left(\log^2\left(1 + \frac{dt}Q\right) + 2(\log (tdQ) - 1)\log\left(1 + \frac{dt}Q\right)\right) dt + O(X^{1/4 + \eps})\\
&= \frac{Q}{2\pi}\sum_{d\le X^{1/4}}\frac{\mu(d)}{d^2}\int_1^{Q/d} \frac{d}{Q}(- 2\log(tdQ) + \log^2 (tdQ))\\
&\hspace{1cm} + \left(\frac 1t + \frac{d}Q\right)\left(2\log (tdQ)\log\left(1 + \frac{dt}Q\right)\right) dt + O(X^{1/2})\\
&= \frac{Q}{2\pi\zeta(2)}(\log^2 X - 4\log X + (4\log 2 - 2)\log X - 2\Li_2(-1)\log X) + O(X^{1/2}) \\
&= \frac{1}{2\pi\zeta(2)}\sqrt{X}\log^2X - C\sqrt{X}\log X + O(X^{1/2})\end{aligned}$$ where $$C = \frac{3}{\pi^3}(4 + 2 - 4\log 2 + 2\Li_2(-1)) = \frac{18}{\pi^3} - \frac{12\log 2}{\pi^3} - \frac{1}{2\pi}\approx 0.153\dots.$$ The desired result then follows.
Acknowledgements
================
The author is thankful to Maksym Radziwiłł and Liyang Yang for discussions of this problem and Vinayak Kumar and Erich Liang for reading earlier drafts of this paper.
[9]{} A. Balog and A. Perelli, On the L1 mean of the exponential sum formed with the Möbius function, [*J. Lond. Math. Soc.*]{}, 57 (1998), 275-288. A. Balog, I. Z. Ruzsa, On the exponential sum over r-free integers. [*Acta Math. Hungar.*]{} 90 (2001), 219-230. A. Balog and I. Z. Ruzsa, A new lower bound for the L1 mean of the exponential sum with the Möbius function, [*Bull. Lond. Math. Soc.*]{}, 31 (1999), 415-418. J., Brüdern, Exponential sums over products and their $L_1$-norm. [*Arch. Math.*]{} 76 (2001), 196-201. J. Brüdern, A. Granville, A. Perelli, R.C. Vaughan and T. D. Wooley, On the exponential sum over k-free numbers. [*R. Soc. Lond. Philos. Trans. Ser. A*]{} 356 (1998), 739-761. D. A. Goldston, The major arcs approximation of an exponential sum over primes. [*Acta Arith.*]{} 92.2 (2000), 169-179. D. A. Goldston and M. Pandey, On the $L^1$ norm of an exponential sum involving the divisor function. [*Arch. Math.*]{} 112.3 (2019), 261-268. H. Iwaniec and E. Kowalski, Analytic number theory, Amer. Math. Soc. Colloquium Publ. 53, Amer. Math. Soc., Providence RI, 2004 M. Jutila. On exponential sums involving the Ramanujan function, [*Proc. Indian Acad. Sci.*]{} 97 (1987), 157-166. E. Keil, Moment estimates for exponential sums over k-free numbers. [*Int. J. Number Theory*]{} 9 (2013), 607-619. O. McGehee, L. Pigno, B. Smith, Hardy’s Inequality and the $L^1$ norm of Exponential Sums. [*Ann. Math.*]{} 113 (1981), 613-618. P. Shiu, A Brun-Titchmarsh theorem for multiplicative functions, [*J. Reine Angew. Math.*]{} 313 (1980), 161-170. R.C. Vaughan, The $L^1$ mean of exponential sums over primes, [*Bull. London Math. Soc.*]{} 20(1988), 121-123. L. L. Zhao, The sum of divisors of a quadratic form, [*Acta Arith.*]{} 163 (2014), 161-177.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Nonlinear contraction theory is a comparatively recent dynamic control system design tool based on an exact differential analysis of convergence, in essence converting a nonlinear stability problem into a linear time-varying stability problem. Contraction analysis relies on finding a suitable [*metric*]{} to study a generally nonlinear and time-varying system. This paper shows that the computation of the metric may be largely simplified or indeed avoided altogether by extending the exact differential analysis to the higher-order dynamics of the nonlinear system. Simple applications in economics, classical mechanics, and process control are described.'
---
**Higher-Order Nonlinear Contraction Analysis**
Introduction
============
Nonlinear contraction theory is a comparatively recent dynamic control system design tool based on an exact differential analysis of convergence [@Lohm1]. In essence, it allows one to convert a nonlinear stability problem into a linear time-varying stability problem. Contraction analysis relies on finding a suitable [*metric*]{} to study a generally nonlinear and time-varying system. Depending on the application, the metric may be trivial (identity or rescaling of states), or obtained from physics, combination of contracting subsystems [@Lohm1], semi-definite programming [@Lohm4], or recently sums-of-squares programming [@Parrilo].
The goal of this paper is to show that the computation of the metric may be largely simplified or avoided altogether by extending the exact differential analysis to the higher-order dynamics of the nonlinear system. Intuitively this is not surprising, since, as an elementary instance, a scalar linear time-invariant system would require in the original approach a non-identity metric (obtained from a Lyapunov Matrix Equation).
After a brief review of contraction theory in Section 2, the main results are presented in Section 3, first in the discrete-time case (with a simple application to price dynamics in economics) and then in the continuous-time case. Simple examples and applications are discussed in Section 4, in the contexts of classical mechanics, process control, and observer design (see [@Rouchon; @Nguyen; @Jouff; @inertial] for other recent applications of contraction theory to observer design). Hamiltonian systems are studied in section 5. Concluding remarks are offered in section 6.
Contraction theory
==================
The basic theorem of contraction analysis [@Lohm1] can be stated as
Consider the deterministic system $ \ \dot{\bf x} =
{\bf f}({\bf x},t) \ $, where ${\bf f}$ is a smooth nonlinear function. If there exist a uniformly positive definite metric
${\bf M(\bf x}, t) \ = \ {\bf \Theta}({\bf x}, t)^T \ {\bf
\Theta}^{\ast}({\bf x}, t)$
such that the Hermitian part of the associated generalized Jacobian ${\bf F} \ = \ \left(\dot{\bf \Theta} + {\bf \Theta}
\frac{\partial {\bf
f}} {\partial {\bf \bf x}} \right){\bf \Theta}^{-\ast}$
is uniformly negative definite, then all system trajectories converge exponentially to a single trajectory, with convergence rate $| \lambda_{max} |$, where $\lambda_{max}$ is the largest eigenvalue of the Hermitian part of $\ {\bf F}$. The system is said to be contracting. \[th:theoremF\]
In the above, $\ ^{\ast}$ denotes complex conjugation, $\ ^{-\ast}$ for a matrix denotes the inverse of the conjugate matrix, and the state-space is $R^n$ (in this paper) or $C^n$. The system is said to be [*semi-contracting*]{} (for the metric ${\bf M(\bf x}, t)$) if ${\bf F}$ is always negative semi-definite, and [*indifferent*]{} if ${\bf F}$ is always zero.
It can be shown conversely that the existence of a uniformly positive definite metric with respect to which the system is contracting is also a necessary condition for global exponential convergence of trajectories. In the linear time-invariant case, a system is globally contracting if and only if it is strictly stable, with ${\bf F}$ simply being a normal Jordan form of the system and ${\bf \Theta}$ the coordinate transformation to that form. Conceptually, approaches closely related to contraction, although not based on differential analysis, can be traced back to [@Hart] and even to [@Lew].
Similarly, a discrete system $${\bf x}_{i+1} = {\bf f}_i({\bf x}_i, i)$$ will be contracting in a metric ${\bf \Theta}^T{\bf \Theta}^{\ast}$ if the largest singular value of the discrete Jacobian ${\bf \Theta}_{i+1}
\frac{\partial {\bf f}_i}{\partial {\bf x}_i} {\bf \Theta}_i^{-\ast}$ is strictly smaller than 1. In the particular case of real autonomous systems with identity metric, the basic contraction theorem corresponds in the continuous-time case to Krasovkii’s theorem [@Sl91], and in the discrete-time case to the contraction mapping theorem [@Bert].
Contraction theory proofs make extensive use of [*virtual displacements*]{}, which are differential displacements at fixed time borrowed from mathematical physics and optimization theory. Formally, if we view the position of the system at time $t$ as a smooth function of the initial condition ${\bf x}_o$ and of time, $\ {\bf x} = {\bf
x}({\bf x}_o ,t)\ $, then $\ \delta {\bf x} = \frac{\partial {\bf
x}}{\partial {\bf x}_o} \ d {{\bf x}_o}\ $. For instance [@Lohm1], for the system of Theorem 1, one easily computes $$\label{differential}
\frac{d}{dt} ({\bf \Theta} \delta {\bf x}) \ = \ {\bf F} ({\bf \Theta} \delta {\bf x})$$
An appropriate metric to show that the system is contracting may be obtained from physics, combination of contracting subsystems [@Lohm1], semi-definite programming [@Lohm4], or sums-of-squares programming [@Parrilo]. The goal of this paper is to show that the computation of the metric may be largely simplified or avoided altogether by considering the system’s [*higher-order*]{} virtual dynamics (rather than merely its first-order virtual dynamics, as in equation (\[differential\])).
Higher-order contraction
========================
The discrete-time case
----------------------
Technically, the extension to higher-order contraction is simplest in the discrete-time case, which we discuss first. The main idea is to construct an exponential bound on the virtual displacement $\delta
{\bf x}$ over $n$ time-steps, rather than over a single time-step as in [@Lohm1].
Consider for $i \ge 0$ the $n$-dimensional ($n \ge 1$) virtual dynamics $$\delta {\bf x}_{i+n} = {\bf A}^{n-1}_i \delta {\bf x}_{i+n-1} +
\ldots + {\bf A}^o_i \delta {\bf x}_i$$ Taking the norm (denoted by $|\ \ |$ ) on both sides, and bounding, yields $$|\delta {\bf x}_{i+n}| \ \le \ \ |A^{n-1}_i| \ |\delta {\bf x}_{i+n-1}|\ + \
\ldots \ + \ |A^o_i| \ |\delta {\bf x}_i| \nonumber$$ where the norm of a matrix is the largest singular value of that matrix. Let us bound for $i=0$ the initial conditions using real positive constants $\lambda$ and $K$ as $$|\delta {\bf x}_{j}| \le K \lambda^{j}, \ 0 \le j < n$$ Assume now that the following characteristic equation is verified, $$\lambda^n \ge |A^{n-1}_i| \lambda^{n-1} + \ldots + |A^o_i|, \ \forall
i \ge 0 \nonumber$$ We then get $$|\delta {\bf x}_{i+n}| \le |A^{n-1}_i| K \lambda^{i+n-1} + \ldots +
|A^o_i| K \lambda^i \le K \lambda^{i+n}$$
(200,200)(0,-100) (-2,0)[(1,0)[210]{}]{} (3,-100)[(0,1)[200]{}]{} (215,0)[$i$]{} (0,105)[$\delta x_i$]{} (-10,0)[$0$]{} (25,-2)[$:$]{} (50,-2)[$:$]{} (75,-2)[$:$]{} (100,-2)[$:$]{} (125,-2)[$:$]{} (150,-2)[$:$]{} (175,-2)[$:$]{} (0,30)[$\bullet$]{} (25,-30)[$\bullet$]{} (50,40)[$\bullet$]{} (75,-35)[$\bullet$]{} (100,-15)[$\bullet$]{} (125,-28)[$\bullet$]{} (150,17)[$\bullet$]{} (175,-15)[$\bullet$]{} (0,90)(100,27)(200,13) (0,-90)(100,-27)(200,-13)
Repeating the above recursively for $ i \ge 0$ we get by complete induction $$|\delta {\bf x}_i| \le K \lambda^i$$ and hence exponential convergence of $|\delta {\bf x}_i|$, as illustrated in Figure \[fig:discretecontraction\].
Consider for $i \ge 0$ the $n$-dimensional ($n \ge 1$) virtual dynamics $$\delta {\bf x}_{i+n} = {\bf A}^{n-1}_i \delta {\bf x}_{i+n-1} +
\ldots + {\bf A}^o_i \delta {\bf x}_i$$ Let us define $\forall i \ge 0$ a constant $\lambda$ with the characteristic equation $$\begin{aligned}
\lambda^n &\ge& |A^{n-1}_i| \lambda^{n-1} + \ldots + |A^o_i|, \
\forall i \ge 0 \ \label{eq:discretecharacteristic}\end{aligned}$$ We can then conclude $$|\delta {\bf x}_{i+n}| \le K \lambda^{i+n}$$ where $K$ is defined by $$|\delta {\bf x}_{j}| \le K \lambda^{j}, \ 0 \le j < n \
\label{eq:discretedefinitionofK}$$ Thus, the system is contracting if $\ \lambda < 1$. \[th:higherorderdiscrete\]
[ ]{}
[**:** ]{} Consider first a second-order linear time invariant (LTI) dynamics $$x_{i+2} \ + \ 2\ \gamma \ x_{i+1} \ + \ \alpha \gamma^2 x_i \ = \ u_i$$ where $u_i$ is an input, and $\gamma$ and $\alpha$ are constants. The virtual dynamics is $$\delta x_{i+2} = - \ 2\ \gamma \ \delta x_{i+1}\ - \ \alpha \gamma^2
\ \delta x_i$$ The characteristic equation (\[eq:discretecharacteristic\]) for $\lambda \ge 0$ is then given by $$\lambda^2 \ \ge \ 2\ |\gamma| \ \lambda + |\alpha| \gamma^2 \ \ \ \ \
\ \ \ \ i.e. \ \ \ \ \ \ \ \ \ \lambda \ \ge \ |\gamma| (1 + \sqrt{
1 + |\alpha |\ }) \nonumber$$ Thus, the contraction condition $\lambda < 1$, or $$|\gamma| ( 1 + \sqrt{ 1 + |\alpha | \ }) \ < \ 1$$ simply means that both eigenvalues of the system have to lie for the conjugate complex case ($\alpha > 1$) within the red half circles in (\[fig:discretecirecle\]) or on the green line for the real case ($\alpha \le 1$). Note that Theorem \[th:higherorderdiscrete\] simply bounds the possibly oscillating discrete system with a non-oscillating system of the same convergence rate for the real case.
Consider now the virtual dynamics of an arbitrary second-order nonlinear time-varying system, $$\delta x_{i+2} + \ 2\ \gamma(i)\ \delta x_{i+1} + \alpha(i) \
\gamma^2(i) \delta x_i \ = \ 0$$ The characteristic equation and the contraction condition are the same as above, except that $\ \gamma\ $ and $\ \alpha\ $ are now time-dependent.
\[exLTVdiscrete\]
[ ]{}
[**:** ]{} In economics, consider the price dynamics $$\begin{aligned}
{\bf n}_{i+1} &=& {\bf f}_i({\bf p}_i, i) \nonumber \\
{\bf p}_{i+1} &=& {\bf g}_i({\bf n}_i, i) \nonumber\end{aligned}$$ with ${\bf n}_i$ the number of sold products at time $i$ and corresponding price ${\bf p}_i$.
The first line above defines the customer demand as a reaction to a given price. The second line defines the price, given by the production cost under competition, as a reaction to the number of sold items. The dynamics above corresponds to the second-order economic growth cycle dynamics $${\bf n}_{i+2} = {\bf f}_{i+1} \left( {\bf g}_i ({\bf n}_i, i) \right)$$
Contraction behavior of this economic behavior with contraction rate $\lambda$ can then be concluded with Theorem \[th:higherorderdiscrete\] for $$\lambda^2 \ge | \frac{\partial {\bf f}_{i+1}}{\partial {\bf p}_{i+1}}
\frac{\partial {\bf g}_i}{\partial {\bf n}_i} | \label{eq:game}$$ That means we get stable (contraction) behavior if the product of customer demand sensitivity to price and production cost sensitivity to number of sold items has singular values less than 1. We can get unstable (diverging) behavior for the opposite case.
Note that this result even holds when no precise model of the sensitivity is known, which is usually the case in economic or game situations.
Whereas the above is well known for LTI economic models we can see that the economic behavior is unchanged for a non-linear, time-varying economic environment.
Let as assume now that the above corresponds to a game situation (see e.g. [@Shamma] or [@Bryson]) between two players with strategic action ${\bf p}_i$ and ${\bf n}_i$. Both players optimize their reaction ${\bf g}$ and ${\bf f}$ with respect to the opponent’s action.
We can then again conclude for (\[eq:game\]) to global contraction behavior to a unique time-dependent trajectory (in the autonomous case, the Nash equilibrium). \[execonomics\]
Of course, and throughout this paper, in some cases the analysis may yet be further streamlined by [*first*]{} applying a simplifying metric transformation of the form $\ \delta {\bf z} = {\bf \Theta} \delta
{\bf x}\ $, and then applying the results to $\ \delta {\bf z}$.
The continuous-time case
------------------------
Let us now derive the continuous-time version of the previous results. Consider for $t\ge0$ the $n$-dimensional ($n \ge 1$) virtual dynamics $$\delta {\bf x}^{(n)} = - {\bf A}_{n-1} \delta {\bf x}^{(n-1)} - \
... \ - {\bf A}_o \delta {\bf x}$$ The proof is based on splitting up the dynamics into a stable part, described by a block diagonal matrix composed of identical negative definite blocks ${\bf F}$ which we select, and an unstable higher-order part. Let $\delta {\bf x}_o = \delta {\bf x}$, and define recursively $$\begin{aligned}
\delta \dot{\bf x}_o &=& {\bf F} \delta {\bf x}_o + \delta {\bf
x}_1 \nonumber \\
&...& \nonumber \\
\delta \dot{\bf x}_{n-2} &=& {\bf F} \delta {\bf x}_{n-2} +
\delta {\bf x}_{n-1} \nonumber \\
\delta \dot{\bf x}_{n-1} &=& - {\bf A}_{n-1} \delta {\bf x}_o^{(n-1)} -
{\bf A}_{n-2} \delta {\bf x}_o^{(n-2)} - \ ... \ - {\bf A}_o \delta
{\bf x}_o \nonumber \\
&&- \left( {\bf F} \delta{\bf x}_o \right)^{(n-1)} - ... - \left( {\bf
F} \delta {\bf x}_{n-2} \right)^{(1)} \nonumber \\
&=& -{\bf A}_{n-1} \left( {\bf F} \delta {\bf x}_o + \delta {\bf
x}_1 \right)^{(n-2)} - ... - {\bf A}_o \delta {\bf x}_o \nonumber \\
&&- \left( L^1 {\bf F} \delta {\bf x}_o + L^o {\bf F} \delta {\bf x}_1
\right)^{(n-2)} - ... - \left( L^1 {\bf F} \delta {\bf x}_{n-2} + L^o
{\bf F} \delta {\bf x}_{n-1} \right) \nonumber \\
&=& {\bf F} \delta {\bf x}_{n-1}
- {\bf A}_{n-1}^{\ast} \delta {\bf x}_{n-1}
- {\bf A}_{n-2}^{\ast} \delta {\bf x}_{n-2}
- {\bf A}_{n-3}^{\ast} \delta {\bf x}_{n-3} -
... \label{eq:superimposed}\end{aligned}$$ where $$\begin{aligned}
L^o {\bf F} &=& {\bf F} \nonumber \\
L^j {\bf F} &=& \frac{d}{dt} L^{j-1} {\bf F} + L^{j-1} {\bf F} \
{\bf F} \ \ \ \ \ \ \ \ \ j \ge 1 \nonumber\end{aligned}$$ and $$\begin{aligned}
{\bf A}_{n-1}^{\ast} &=& {\bf A}_{n-1} + \left( \begin{array}{c} n \\
1 \end{array} \right) L^o {\bf F} \label{eq:Aast} \\
{\bf A}_{n-2}^{\ast} &=& {\bf A}_{n-2} + \left( \begin{array}{c} n-1
\\ 1 \end{array} \right) {\bf A}_{n-1} L^o {\bf F} + \left(
\begin{array}{c} n \\ 2 \end{array} \right) L^1 {\bf F} \nonumber \\
{\bf A}_{n-3}^{\ast} &=& {\bf A}_{n-3} + \left( \begin{array}{c} n-2
\\ 1 \end{array} \right) {\bf A}^{n-2} L^o {\bf F} + \left(
\begin{array}{c} n-1 \\ 2 \end{array} \right) {\bf A}^{n-1} L^1 {\bf
F} + \left( \begin{array}{c} n \\ 3 \end{array} \right) L^2 {\bf F}
\nonumber \\ &...& \nonumber\end{aligned}$$ Equation (\[eq:superimposed\]) represents the superposition of a higher-order-system and a block diagonal dynamics in the chosen ${\bf
F}$. Let us assess the contraction behavior of the higher-order part by taking the norm $$|\delta {\bf x}^{(n)}| \le |{\bf A}_{n-1}^{\ast}| |\delta {\bf
x}^{(n-1)}| + |{\bf A}_{n-2}^{\ast}| |\delta {\bf x}^{(n-2)}| + \
\ldots $$ where the norm of a matrix is the largest singular value of that matrix. Let us bound for $t=0$ the initial conditions with real and constant $\lambda, K \ge 0$ and assume the following characteristic equation $$\begin{aligned}
|\delta {\bf x}^{(j)}| &\le& K \lambda^j e^{\lambda t}, \ \ \ \ \ \ \ \
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ 0 \le j < n \ \label{eq:definitionofK}
\\ \lambda^n &\ge& |{\bf A}_{n-1}^{\ast}| \lambda^{n-1} + \ \ldots \ +
|{\bf A}_o^{\ast}|, \ \forall t \ge 0 \label{eq:characteristic}\end{aligned}$$ Figure \[fig:discretecontraction\] shows how $K$ has to be selected for a given $\lambda$ for a second-order system ($n=2$).
(200,200)(0,-100) (0,0)[(1,0)[200]{}]{} (103,-100)[(0,1)[200]{}]{} (205,0)[t]{} (200,20)[$\delta x(t)$]{} (200,100)[$K e^{\lambda t}$]{} (0,-50)(100,107)(200,30) (0,55)(100,72)(200,115)
With (\[eq:characteristic\]) we can bound the $n$’th derivative of $\delta {\bf x}$ as $$|\delta {\bf x}^{(n)}| \le |{\bf A}_{n-1}^{\ast}| K \lambda^{n-1}
e^{\lambda t} + \ \ldots \ + |{\bf A}_o^{\ast}| K e^{\lambda t} \le K
\lambda^n e^{\lambda t}$$ Integrating the above for $t \ge 0$ we can exponentially bound the higher-order dynamics $\delta {\bf x}$ as $$|\delta {\bf x}| \le K e^{\int_o^t \lambda d \tau}$$ Using the above this allow to conclude:
Consider for $t\ge0$ the $n$-dimensional ($n \ge 1$) virtual dynamics $$\delta {\bf x}^{(n)} = - {\bf A}^{n-1} \delta {\bf x}^{(n-1)} - \
\ldots \ - {\bf A}^o \delta {\bf x}$$ Let us define a constant $\lambda \ge 0$ such that $\forall t \ge 0$ we fulfill the characteristic equation $$\lambda^n \ge {\bf A}_{n-1}^{\ast} \lambda^{n-1} + \ \ldots \ + {\bf
A}_o^{\ast} \label{eq:theoremcharacteristic}$$ where ${\bf A}_j^{\ast}$ is defined in (\[eq:Aast\]) for a given choice of the matrix ${\bf F}$.
We can then conclude on contraction rate (i.e., the largest eigenvalue of the symmetric part of) ${\bf F} + \lambda {\bf I}$, where $|\delta {\bf x}|$ is initially bounded with $K$, defined in (\[eq:definitionofK\]). \[th:higherordercontinuous\]
One specific choice of ${\bf F}$ is $-\frac{{\bf A}_{n-1}}{n}$, which cancels the highest time-derivative on the right-hand side, and is known for LTV systems as the reduced or unstable form [@Kailath] of the original higher-order dynamics. We will use this definition of ${\bf F}$ in most of the following examples. Also note that more general forms could be chosen for the stable part.
Examples and Applications
=========================
In this section, we discuss simple examples (section 4.1), applications to nonlinear observer design (section 4.2), and adding an indifferent system (section 4.3).
Some simple examples
--------------------
[ ]{}
[**:** ]{} Consider the second-order LTI dynamics $$\ddot{x} = - 2 \zeta \omega \dot{x} - \omega^2 x$$ with constant $\zeta$ and $\omega \ge 0$. The virtual dynamics is $$\delta \ddot{x} = - 2 \zeta \omega \delta \dot{x} - \omega^2 \delta x$$ The characteristic equation (\[eq:theoremcharacteristic\]) is then given with $F = \zeta \omega$ for constant, positive $\lambda$ by $$\begin{aligned}
\lambda^2 \ge | - \omega^2 + \frac{(2 \zeta \omega)^2}{4} | =
| \zeta^2 - 1 | \omega^2 \ \ \ \ \ \ \ \ \ \rm{i.e.} \ \ \ \ \ \ \ \ \
\lambda \ge \omega \sqrt{| \zeta^2 - 1 |} \nonumber\end{aligned}$$ Using Theorem \[th:higherordercontinuous\] we can then conclude on contraction behavior with convergence rate $$\omega (-\zeta + \sqrt{| \zeta^2 - 1 |})$$ This means that we require the poles to lie within the $\pm45^o$ quadrant of the left-half complex plane. \[exLTI\]
While Theorem \[th:higherordercontinuous\] can thus be overly conservative for LTI systems, this is not the case for general nonlinear time-varying systems, as we now illustrate.
[ ]{}
[**:** ]{} Consider the second-order LTV dynamics $$\ddot{x} + a_1 \ \dot{x} + a_o(t)\ \ x = u(t)$$ with $a_1, a_o \ge 0$, which would be sufficient conditions for LTI stability. Let us assume a small damping gain $a_1$ and strong spring gains $a_o$ such that the system oscillates.
If now the time-varying gain $a_o(t)$ is chosen to be very large when the system oscillates back to $0$ and small otherwise then the energy is constantly increased, which makes the system unstable (Figure \[fig:example\_34\]). This is precisely what is excluded by Theorem \[th:higherordercontinuous\]. \[exLTV\]
[ ]{}
[**:** ]{} Consider the second-order LTV dynamics $$\ddot{x} + a_1(t)\ \dot{x} + a_o(t)\ x = u(t)$$ The virtual dynamics is $$\delta \ddot{x} = - a_1 \delta \dot{x} - a_o \delta x$$ The characteristic equation (\[eq:theoremcharacteristic\]) is then given with $F = - \frac{a_1}{2}$ by $$\lambda^2 \ge | - a_o + \frac{a_1^2}{4} + \frac{\dot{a_1}}{2}|$$ for constant positive $\lambda$. Using Theorem \[th:higherordercontinuous\] we can then conclude on contraction behavior with convergence rate $-\frac{a_1}{2} + \lambda\ $ (Figure \[fig:example\_35\]). \[exLTV\]
[ ]{}
[**:** ]{} Let us illustrate a case where choosing ${\bf F}$ other than $\ - \frac{{\bf A}_{n-1}}{n}\ $ can simplify the result. Consider the generalized Van der Pol or Lienard dynamics $$\ddot{x} = - a_1(x) \dot{x} - a_o(x, t)$$ with $a_1, \frac{\partial a_o}{\partial x} \ge 0$. The virtual dynamics is $$\delta \ddot{x} = - a_1 \delta \dot{x} - \left( \frac{\partial
a_o}{\partial x} + \dot{a_1} \right) \delta x$$ The characteristic equation (\[eq:theoremcharacteristic\]) is then given with $F = - a_1 + \frac{\min(a_1)}{2}$ by $$\lambda^2 = ( a_1 - \min(a_1) ) \lambda + |-\frac{ \min(a_1)^2}{4} +
\frac{a_1 \min( a_1)}{2} - \frac{\partial a_o}{\partial x}|$$ and hence $\lambda \ge 0.5 \left( a_1 - \min(a_1) + \sqrt{ ( a_1 -
\min(a_1) )^2 + |-\min(a_1)^2 + 2 a_1 \min(a_1) - 4 \frac{\partial
a_o}{\partial x}|} \right)$. Thus the contraction behavior of $\delta x$ is then given by $$\lambda + F \le - 0.5 a_1 + 0.5 \sqrt{ ( a_1 - \min(a_1) )^2 +
|-( a_1 - \min(a_1) )^2 + a_1^2 - 4 \frac{\partial a_o}{\partial x}|}$$ which is negative if the argument of the square root is less than $a_1^2$, i.e. for $\min(a_1)^2 - 2 a_1 \min(a_1) + 4 \frac{\partial
a_o}{\partial x} \le 0$.
Hence we can conclude on contraction behavior if the poles of the minimal damping $\min(a_1)$ with the actual spring gain $\frac{\partial a_o}{\partial x}$ lie within the $\pm 45^o$ quadrant of the left-half complex plane (Figure \[fig:example\_37\]). Note that this result can be extended to the vector case ${\bf x}$ if the corresponding matrix $A_1({\bf x})$ is integrable. \[exVanderPol\]
[ ]{}
[**:** ]{} Consider the second-order nonlinear vector system $$\ddot{\bf x} + {\bf D} \ \dot{\bf x} + \frac{\partial V}{\partial {\bf
x}} = u(t)$$ with potential energy $V = x_1^2+x_2^2+x_1x_2\sin t$ and constant damping gain ${\bf D} =diag(1,4)$.
The corresponding variational dynamics is $$\delta \ddot{\bf x} + {\bf D} \ \delta \dot{\bf x} + \frac{\partial^2
V}{\partial {\bf x}^2} \ \delta {\bf x} = 0$$ The characteristic equation (\[eq:theoremcharacteristic\]) for ${\bf
F} = - \frac{\bf D}{2}$ is then given by $$\lambda^2 \ge | \frac{\partial^2
V}{\partial {\bf x}^2} - \frac{\bf D D}{4}\ |$$ for constant positive $\lambda$. Using Theorem \[th:higherordercontinuous\] we can then conclude on contraction behavior with convergence rate $\ \lambda {\bf I}\ -\frac{\bf D}{2}$ in Figure \[fig:example\_36\]. \[exsecond-order-system\]
[ ]{}
[**:** ]{} Let us consider two coupled systems of the same dimensions, with the virtual dynamics $$\frac{d}{dt}
\left(
\begin{array}{c}
\delta {\bf x}_1 \\
\delta {\bf x}_2
\end{array}
\right) \ = \
\left(
\begin{array}{cc}
{\bf F}_{11} & {\bf F}_{12} \\
{\bf F}_{21} & {\bf F}_{22}
\end{array}
\right)
\left(
\begin{array}{c}
\delta {\bf x}_1 \\
\delta {\bf x}_2
\end{array}
\right)$$ Let us transform this dynamics in the following second-order dynamics $$\begin{aligned}
\frac{d^2}{dt^2} \ \delta {\bf x}_1 &=& \dot{\bf F}_{11} \delta
{\bf x}_1 + {\bf F}_{11} \ \frac{d}{dt}\ \delta {\bf x}_1 +
(\dot{\bf F}_{12} + {\bf F}_{12} {\bf F}_{22}) \delta {\bf x}_2 +
{\bf F}_{12} {\bf F}_{21} \delta {\bf x}_1 \nonumber \\ &=& \left(
{\bf F}_{11} + {\bf F}_{22}^{\ast} \right) \frac{d}{dt} \ \delta
{\bf x}_1 + \left( \dot{\bf F}_{11} + {\bf F}_{12} {\bf F}_{21} -
{\bf F}_{22}^{\ast} {\bf F}_{11} \right) \delta {\bf x}_1 \nonumber\end{aligned}$$ with the generalized Jacobian ${\bf F}_{22}^{\ast} = (\dot{\bf
F}_{12} + {\bf F}_{12} {\bf F}_{22}) {\bf F}_{12}^{-1}$. The shrinking rate of this system is now the average of ${\bf F}_{11}$ and ${\bf F}_{22}^{\ast}$. Using a direct contraction approach with e.g. ${\bf F}_{12}\ = -\ k \ {\bf F}_{21}^T$ the guaranteed contraction rate would be a more conservative value, namely the largest of the individual contraction rates of ${\bf F}_{11}$ and ${\bf F}_{22}^{\ast}$. \[exsecond-order-state-space\]
Higher-order observer design
----------------------------
While a controller for an $n^{\rm th}$ order system simply has to add stabilizing feedback in ${\bf x}^{(n-1)}, \ \ldots \ , {\bf x}, t$ according to Theorem \[th:higherordercontinuous\], the situation is not such straightforward for observers since here only a part of the state is measured. Motivated by the linear Luenberger observer and the linear reduced-order Luenberger observer, we derive such an observer design for higher-order nonlinear systems.
Consider the $n$-dimensional nonlinear system dynamics $${\bf x}^{(n)} = {\bf a}_o({\bf x}, \dot{\bf x}, t) + \ \ldots \ + {\bf
a}_{n-2}^{(n-2)} ({\bf x}, \dot{\bf x}, t) + {\bf
a}_{n-1}^{(n-1)}({\bf x}, t)$$ with the measurement ${\bf y}({\bf x}^{(n-1)}, \ \ldots \ , {\bf x},
t)$. Note that for a linear Luenberger observer ${\bf y}$ is equivalent to ${\bf x}$ and all ${\bf a}_i$ are linear functions of ${\bf x}$.
Consider now the corresponding nonlinear observer $$\begin{aligned}
\dot{\hat{\bf x}}_{n-1} &=& {\bf a}_o - {\bf e}_{o}(\hat{\bf y}) +
{\bf e}_{o}({\bf y}) \nonumber \\ \dot{\hat{\bf x}}_{n-2} &=& \hat{\bf
x}_{n-1} + {\bf a}_1 - {\bf e}_{1}(\hat{\bf y}) + {\bf e}_{1}({\bf y})
\nonumber \\ &\ \ldots \ & \nonumber \\ \dot{\hat{\bf x}}_o &=&
\hat{\bf x}_1 + {\bf a}_{n-1}- {\bf e}_{n-1}(\hat{\bf y}) + {\bf
e}_{n-1}({\bf y})\nonumber\end{aligned}$$ with ${\bf a}_i(\hat{\bf x}_o, \hat{\bf x}_1 + {\bf a}_{n-1}- {\bf
e}_{n-1}(\hat{\bf y}) + {\bf e}_{n-1}({\bf y}), t)$ and ${\bf x}_o =
{\bf x}$. Note that the coordinate transformation in the bracket is a nonlinear generalization of the reduced Luenberger observer. The above dynamics is equivalent to $$\begin{aligned}
\hat{\bf x}^{(n)} &=& {\bf a}_o(\hat{\bf x}, \dot{\hat{\bf x}}, t) + \
\ldots \ + {\bf a}_{n-2}^{(n-2)} (\hat{\bf x}, \dot{\hat{\bf x}}, t) +
{\bf a}_{n-1}^{(n-1)}(\hat{\bf x}, t) \nonumber \\ &-& {\bf
e}_o(\hat{\bf y}) + {\bf e}_o({\bf y}) -\ \ldots \ - {\bf
e}_{n-1}^{(n-1)}(\hat{\bf y}) + {\bf e}_{n-1}^{(n-1)}({\bf y})
\nonumber\end{aligned}$$ whose variational dynamics is $$\delta \hat{\bf x}^{(n)} = \delta {\bf a}_o + \ \ldots \ + \delta
{\bf a}_{n-1}^{(n-1)} - \delta {\bf e}_o -\ \ldots \ - \delta {\bf
e}_{n-1}^{(n-1)}$$ where the varation is performed on $\hat{\bf x}^{n-1}, \ \ldots \ ,
\hat{\bf x}$. Hence the feedback is not only performed in $\hat{\bf
y}$, but implicitly also up to the $(n-1)^{\rm th}$ time-derivative of $\hat{\bf y}$.
[ ]{}
[**:** ]{} Consider the second-order nonlinear system $$\ddot{x} + \frac{\partial a_1}{\partial x} (x)\ \dot{x} + a_o(x) = 0$$ where $x$ is measured. Consider now the corresponding nonlinear observer $$\begin{aligned}
\dot{\hat{x}}_1 &=& a_o(\hat{x}) - e_o (\hat{x} - x)
\nonumber \\
\dot{\hat{x}}_o &=& \hat{x}_1 + a_1(x) - e_1 (\hat{x} - x)
\nonumber\end{aligned}$$ with constant $e_o$ and $e_1$ and where we have replaced with the feedback $a_1$ as a function of $x$. The corresponding second-order variational dynamics is $$\delta \ddot{\hat{x}} + e_1 \ \delta \dot{\hat{x}} + (e_o -
\frac{\partial a_o}{\partial \hat{x}})\ \delta \hat{x} = 0$$ Contraction behavior can then be shown with Theorem \[th:higherordercontinuous\]. \[exsecond-order-system-observer\]
[ ]{}
[**:** ]{} Consider the temperature-dependent reaction $A \rightarrow B$ in a closed tank $$\frac{d}{dt}
\left(
\begin{array}{c}
c_A \\
T
\end{array}
\right) =
\left(
\begin{array}{c}
-1 \\
-10
\end{array}
\right)
e^{-\frac{E}{T}} c_A$$ with $c_A$ the concentration of A, $T$ the measured temperature, and $E$ the specific activation energy. This reaction dynamics is equivalent to the following second-order dynamics in temperature $$\ddot{T} + \frac{-E}{T^2} \dot{T}^2 = -e^{-\frac{E}{T}} \dot{T}$$ Letting $\tau = \int_o^T e^{\frac{-E}{T}} dT$ yields $$\ddot{\tau} = -e^{-\frac{E}{T}} \dot{\tau}$$ Contraction can then be shown as in [Example \[exVanderPol\]]{}. The observer dynamics $$\begin{aligned}
\dot{\hat{T}}_1 &=& -e^{-\frac{E}{\hat{T}_o}} \dot{\hat{T}}_o +
\frac{E}{\hat{T}_o^2} \dot{\hat{T}}_o^2 - e_o e^{\frac{E}{\hat{T}_o}}
\int^{\hat{T}}_T e^{\frac{-E}{T}} dT \nonumber \\
\dot{\hat{T}}_o &=& \hat{T}_1 - e_1 (\hat{T}_o - T_o) \nonumber\end{aligned}$$ with $\hat{T}_o = \hat{T}$ and constant $e_o$ and $e_1$ leads to $$\ddot{\hat{\tau}} = (-e^{-\frac{E}{T}} - e_1) \dot{\hat{\tau}} - e_o \
\hat{\tau} + e_1 \dot{\tau}(t) + e_o \tau(t)$$ whose contraction behavior can now be tuned as in [Example \[exVanderPol\]]{}. \[exchemical\]
[ ]{}
[**:** ]{} Consider the system $$\dddot{x}\ = \ \sin (\dot{x})\ +\ 0.1 \sin t\ - \ 0.015$$ with measurement $y=x$. Letting $e_0(y)=y$, $e_1(y)=4y$, and $e_2(y)=3y$, the variational equation is $$\delta \dddot{\hat{x}}+3\ \delta \ddot{\hat x}+(4-\cos(\dot{\hat x}))\
\delta \dot{\hat x}+\delta \hat{x}\ =\ 0$$ The corresponding observer is illustrated in Figure \[fig:example\_4\]. \[ex\]
Adding an indifferent system
----------------------------
The analysis may be further simplified by using superposition to compare the system to one whose contraction behavior is known analytically. We illustrate this idea on second-order systems using an indifferent added dynamics. In principle, the approach can be extended to higher-order systems as well as other types of added dynamics.
Consider the indifferent system [@Lohm1] $$\delta \dot{\bf x} = i \Omega \delta {\bf x}$$ with real and invertible $\Omega(\dot {\bf x}, {\bf x}, t)$. The above corresponds to the second-order dynamics $$\delta \ddot{\bf x} = \dot{\Omega} \Omega^{-1} \delta \dot{\bf x} -
\Omega \Omega \delta {\bf x}$$ which is thus itself indifferent. We can write the reduced form of Theorem \[th:higherordercontinuous\] as $$\begin{aligned}
\frac{d}{dt}
\left(
\begin{array}{c}
\delta \dot{\bf x} \\
\delta {\bf x}
\end{array}
\right) &=&
\left(
\begin{array}{cc}
- \Theta({\bf x}, \dot{\bf x}, t) & - \Omega \Omega(\dot {\bf x}, {\bf
x}, t) \\
{\bf I} & {\bf 0}
\end{array}
\right)
\left(
\begin{array}{c}
\delta \dot{\bf x} \\
\delta {\bf x}
\end{array}
\right) \nonumber \\ &=&
\left( \left(
\begin{array}{cc}
\dot{\Omega} \Omega^{-1} & - \Omega \Omega \\
{\bf I} & {\bf 0}
\end{array}
\right)
+
\left(
\begin{array}{cc}
- \dot{\Omega} \Omega^{-1} - \Theta & \ {\bf 0} \\
{\bf 0} & \ {\bf 0}
\end{array}
\right)
\right)
\left(
\begin{array}{c}
\delta \dot{\bf x} \\
\delta {\bf x}
\end{array}
\right) \nonumber\end{aligned}$$ which thus corresponds to the superposition of an indifferent system with a semi-contracting system of rate $-\Theta - \dot{\Omega} \Omega^{-1}$.
The reduced form $$\delta \ddot{\bf x} + \Theta({\bf x}, \dot{\bf x}, t) \delta
\dot{\bf x} + \Omega \Omega(\dot {\bf x}, {\bf x}, t)\ \delta{\bf x} =
0 \nonumber$$ is semi-contracting with rate $-\Theta - \dot{\Omega}
\Omega^{-1}$. The corresponding unreduced form (see Theorem \[th:higherordercontinuous\]) has an additional contraction rate $-\frac{\bf F}{2}$. \[th:variationalenergy\]
[ ]{}
[**:** ]{} Consider the restricted Three-Body Problem [@ThreeBody] $$\ddot{\bf x} =
2 \left(
\begin{array}{ccc}
0 & 1 & 0 \\
-1 & 0 & 0 \\
0 & 0 & 0
\end{array}
\right)
\dot{\bf x} +
\left(
\begin{array}{ccc}
0 & 1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 0
\end{array}
\right)
{\bf x}
-\nabla V({\bf x})$$ with state ${\bf x} = (x, y, z)^T$, potential energy $V({\bf x})
= -\nu \sqrt{y^2 + z^2 + (-1+x+y)^2}^{-1} - (1-\nu)
\sqrt{y^2+z^2+(x+y)^2}^{-1}$, $\nu$ the ratio of the smaller mass to the larger mass, and the third body of zero mass.
Using ${\bf F} = \left(
\begin{array}{ccc}
0 & 1 & 0 \\
-1 & 0 & 0 \\
0 & 0 & 0
\end{array}
\right)$ in Theorem \[th:higherordercontinuous\], the reduced variational dynamics is $$\delta \ddot{\bf x} = - \Delta V \delta {\bf x}$$ Using Theorem \[th:variationalenergy\], the contraction behavior of the three-body problem is the superposition of
$\bullet$ $\ i \sqrt{\Delta V}$, which is indifferent for $\Delta V \ge
0$ and unstable otherwise.
$\bullet$ $\ -\frac{d \sqrt{\Delta V}}{dt} \sqrt{\Delta V}^{-1}$, where a tightening (relaxing) potential force adds semi-contracting (diverging) behavior. \[exThreeBodyProblem\]
Hamiltonian system dynamics {#Hamilton}
===========================
Consider the general $n$-dimensional Hamiltonian dynamics [@Lovelock] with sum convention over the free index $$\ddot{x}^j + \gamma_{kh}^j \dot{x}^k \dot{x}^h = - H^{jh} f_h - D_h^j
\dot{x}^h$$ with external forces $f_h(x^l, t)$, damping gain $D_h^j(x^l)$ and the Christoffel term $\gamma^j_{kh} = \frac{1}{2} H^{mj} \left(
\frac{\partial H_{km}}{\partial x^h} + \frac{\partial H_{hm}}{\partial
x^k} - \frac{\partial H_{kh}}{\partial x^m} \right)$ (section 7.2. in [@Lovelock]) of the symmetric inertia tensor $H_{lh}(x^m)$ and where we use the convention $H^{mj} = H_{mj}^{-1}$. The variational dynamics of the above is $$\delta \ddot{x}^j + \frac{\partial \gamma_{kh}^j}{\partial x^l} \delta
x^l \dot{x}^k \dot{x}^h + 2 \gamma_{kh}^j \dot{x}^k \delta \dot{x}^h =
- \frac{\partial (H^{jh} f_h)}{\partial x^l} \delta x^l - \frac{\partial
D_h^j}{\partial x^l} \delta x^l \dot{x}^h - D_h^j \delta \dot{x}^h$$ Let us now compute with the above the covariant time-derivative (section 3.6. in [@Lovelock] and generalized contraction with metric in [@Lohm1]) of the covariant velocity variation $\delta \dot{x}^j + \gamma_{lh}^j \dot{x}^h \delta x^l$ with respect to the metric $H_{lh}$ as $$\begin{aligned}
\frac{d}{dt} \left( \delta \dot{x}^j + \gamma_{lh}^j
\dot{x}^h \delta x^l \right) + \gamma_{ih}^j \dot{x}^h \left( \delta
\dot{x}^i + \gamma_{kl}^i \dot{x}^k \delta x^l \right) &=&
\nonumber \\ \delta \ddot{x}^j + \frac{\partial
\gamma_{lh}^j}{\partial x^k} \dot{x}^k \dot{x}^h \delta x^l +
\gamma_{lh}^j \ddot{x}^h \delta x^l + \gamma^j_{lh} \dot{x}^h \delta
\dot{x}^l + \gamma_{ih}^j \dot{x}^h \left( \delta \dot{x}^i +
\gamma_{kl}^i \dot{x}^k \delta x^l \right) &=& \nonumber \\
K^j_{lkh} \dot{x}^k \dot{x}^h \delta x^l - H^{jh} \left(
\frac{\partial f_h}{\partial x^l} - \gamma^k_{hl} f_k \right) \delta
x^l - D_h^j \left( \delta \dot{x}^h + \gamma^h_{kl} \dot{x}^k \delta
x^l \right) \label{eq:varacceldynamics}\end{aligned}$$ with curavture tensor $K^j_{lkh} = \frac{\partial
\gamma_{lh}^j}{\partial x^k} - \frac{\partial \gamma_{kh}^j}{\partial
x^l} + \gamma_{ih}^j \gamma_{kl}^i - \gamma_{li}^j \gamma_{kh}^i$ (section 7.3. in [@Lovelock]), covariant derivative of the external forces $\frac{\partial f_h}{\partial x^l} - \gamma^k_{hl}
f_k$, and where we have assumed the covariant derivative of $D_h^j$, that is $\frac{\partial D_h^j}{\partial x^l} + \gamma_{lk}^j D_h^k -
\gamma_{lh}^k D_k^j$ (see section 3.6 in [@Lovelock]), to vanish which is usually the case for mechanical damping.
Note that according section 7.3 in [@Lovelock] corresponds $K^i_{lkh} H_{ji} Y^k Y^h X^j X^l$ to the Riemannian or Gaussian curvature of the 2-D subspace span by $X$ and $Y$. Hence the convexity of $H$ orthogonal to $\dot{x}$ acts as a spring, whose gain increases linearly with velocity.
Also note that (\[eq:varacceldynamics\]) can be used directly in combination with Theorem 2 in [@Lohm5] to show under which condition the covariant Hessian of the action $\phi$ becomes convex, which implies contraction behavior.
However we use here Theorem \[th:higherordercontinuous\] for ${\bf
F} = {\bf D}$, which is more general than the above (i.e. it allows to assess complex solutions of the Hessian dynamics in Theorem 2 in [@Lohm5]), to conclude:
Consider for $t\ge0$ the $n$-dimensional ($n \ge 1$) dynamics with sum convention over the free index $$\ddot{x}^j + \gamma_{kh}^j \dot{x}^k \dot{x}^h = - H^{jh} f_h - D_h^j
\dot{x}^h
\label{eq:Hamiltonian}$$ with external forces $f^j(x^l, t)$, damping gain $D_h^j(x^l)$ with covariant derivative $\frac{\partial D_h^j}{\partial x^l} +
\gamma_{lk}^j D_h^k - \gamma_{lh}^k D_k^j = 0$, and the Christoffel term $\gamma^j_{kh} = \frac{1}{2} H^{mj} \left( \frac{\partial
H_{km}}{\partial x^h} + \frac{\partial H_{hm}}{\partial x^k} -
\frac{\partial H_{kh}}{\partial x^m} \right)$ of the symetric, u.p.d. and invertible inertia tensor $H_{lh}(x^m)$.
Let us define a constant $\lambda \ge 0$ as the largest singular value $\forall t \ge 0$ as
$$\left( -K_{ipkh} \dot{x}^k \dot{x}^h + \frac{\partial f_i}{\partial
x^p} - \gamma^h_{ip} f_h + H_{ih} \frac{D_k^h D_p^k}{4} \right) H^{ij}
\left( -K_{jlkh} \dot{x}^k \dot{x}^h + \frac{\partial f_j}{\partial
x^l} - \gamma^h_{jh} f_h + H_{jh} \frac{D_k^h D_l^k}{4} \right) \le
\lambda^2 H_{pl} \label{eq:characteristicHamiltonian}$$
with curavture tensor $K_{jlkh} = H_{jo} \frac{\partial
\gamma_{lh}^o}{\partial x^k} - \frac{\partial \gamma_{kh}^o}{\partial
x^l} + \gamma_{ih}^o \gamma_{kl}^i - \gamma_{li}^o \gamma_{kh}^i$.
We can then conclude on contraction rate $- \frac{D_h^j}{2} +
\lambda$. \[th:Hamiltoniancontinuous\]
Also note that taking the double integral of the dynamics leads to an exponential Lyaponov energy stability proof for the autonomous case. In this sense Theorem 5 represents a variational extension of the classical energy based Lyaponov proofs for autonomous systems. Note that at a variational energy approach a tightening spring (positive $\dot{\Omega}$) leads to semi-contraction behavior.
[ ]{}
[**:** ]{} Let us now illustrate the simplicity of the stability results using the inertia tensor as metric. Consider the Euler dynamics of a rigid body, with Euler angles ${\bf x} = (\psi, \theta,
\phi)^T$ and measured rotation vector ${\bf \omega}$ in body coordinates [@Goldstein] $$\dot{\bf x} = \left(
\begin{array}{ccc}
1 & 0 & - \sin \theta \\
0 & \cos \psi & \cos \theta \sin \psi \\
0 & -\sin \psi & \cos \theta \cos \psi
\end{array}
\right)^{-1} {\bf \omega} \label{eq:rotationdynamics}$$ The underlying energy is $$\begin{aligned}
h &=& \frac{1}{2} {\bf \omega}^T {\bf \omega} \nonumber \\
&=& \frac{1}{2}\dot{\bf x}^T
\left(
\begin{array}{ccc}
1 & 0 & -\sin \theta \\
0 & 1 & 0 \\
- \sin \theta & 0 & 1
\end{array}
\right)
\dot{\bf x} \nonumber\end{aligned}$$ After a straightforward but tedious calculation we can compute $$\frac{d}{dt} \left( \delta {\bf x}^T {\bf H} \delta {\bf x} \right) = 0$$ Thus, the Euler dynamics (\[eq:rotationdynamics\]) is globally indifferent. Note that this can also be seen from the quaternion angular dynamics, whose Jacobian is skew-symmetric [@inertial]. \[exEulerdynamics\]
[ ]{}
[**:** ]{} The convexity of a the inertia tensor can act similarly to a stabilizing potential force in the variational Hamiltonian dynamics (\[eq:varacceldynamics\]). Consider a rotating point mass of mass $m$ on a ball with radius $R$. The Hamiltonian energy is $$h = \frac{m R^2}{2} \dot{\bf x}^T
\left(
\begin{array}{cc}
1 & 0 \\
0 & \sin^2 \phi
\end{array}
\right)
\dot{\bf x}$$ with latitude $\phi$ and longitude $\psi$ in ${\bf x} = (\phi,
\psi)^T$. The curvature tensor can be computed e.g. with MAPLE as $$K_{jlkh} \dot{x}^k \dot{x}^h =
\frac{m R^2 \sin^2 \phi}{2}
\left( \begin {array}{cc}
-\dot{\psi}^2 & \dot{\psi} \dot{\phi} \\
\dot{\psi} \dot{\phi} & -\dot{\phi}^2
\end {array}
\right)$$ which scales as the inertia tensor with $\frac{m R^2}{2}$ and is negative orthogonal to the velocity and indifferent along the velocity. Hence the convexitiy in Theorem \[th:Hamiltoniancontinuous\] acts under motion as a stabilizing spring, that lets two moving neighboring trajectories oscillate around each other. \[exball\]
[ ]{}
[**:** ]{} Theorem \[th:Hamiltoniancontinuous\] can be used to define observers or tracking controllers for time-varying Hamiltonian systems. Consider a two-link robot manipulator, with kinetic energy $$\frac{1}{2}
\left( \begin{array}{cc} \dot{q}_1 & \dot{q}_2 \end{array} \right)
\left( \begin{array}{cc}
a_1 + 2 a_2 \cos q_2 & a_2 \cos q_2 + a_3 \\
a_2 \cos q_2 + a_3 & a_3
\end{array} \right)
\left( \begin{array}{c} \dot{q}_1 \\ \dot{q}_2 \end{array} \right)$$ with $a_1 = m_1 l_{c1}^2 + I_1 + m_2(l_1^2+l_{c2}^2) + I_2$, $a_2 =
m_2 l_1 l_{c2}$ and $a_3 = m_2 L_{c2}^2 + I_2$ and $-\pi \le q_1, q_2
\le \pi$. Let us assume that $q^j$ is measured and define the observer $$\begin{aligned}
\dot{\hat{\omega}}^j + \gamma_{kh}^j(\hat{q}^j) \dot{\hat{q}}^k
\dot{\hat{q}}^h &=& - H^{hj} f_h \nonumber \\ \dot{\hat{q}}^j &=&
\hat{\omega}^j + d_h^j ( q^h - \hat{q}^h ) \nonumber\end{aligned}$$ with the external forces $$f_h =
\left(
\begin{array}{c}
g (m_1 l_{c1} + m_e l_1) \cos q_1 + g m_e l_{ce} \cos(q_1 + q_2) +
\tau_1 \\ g m_e l_{ce} \cos(q_1 + q_2) + \tau_2
\end{array}
\right)
+ k_h^i \left(\hat{q}^i -q^i \right)$$ and external torques $\tau_1$, $\tau_2$. The above is equivalent to (\[eq:Hamiltonian\]) $$\ddot{\hat{q}}^j + \gamma_{kh}^j(\hat{q}^j) \dot{\hat{q}}^k
\dot{\hat{q}}^h = - H^{jh} \left( f_h - d_{hk} \dot{q}^k \right)
- d_h^j \dot{\hat{q}}^h$$ where the covariant derivative of a constant scalar $d_h^j$ vanishes. The curvature tensor can be computed e.g. with MAPLE as $$K_{jlkh} \dot{x}^k \dot{x}^h =
\frac{ \left( a_1 a_3 - a_3^2 - a_2^2 \right) a_2 \cos q_2} {a_1 a_3
-a_2^2 \cos^2 q_2 -a_3^2}
\left( \begin {array}{cc}
-\dot{q}_2^2 & \dot{q}_1 \dot{q}_2 \\
\dot{q}_1 \dot{q}_2 & -\dot{q}_1^2
\end {array}
\right)$$ The curvature is for $a_1 a_3 \ge a_2^2 + a_3^2$ convex (concave) for $-\frac{\pi}{2} \le q_2 \le \frac{\pi}{2}$ ($-\frac{\pi}{2} \ge q_2 \
or \ q_2 \ge \frac{\pi}{2}$) and accordingly (de)-stabilizes the dynamics when the arm is retracted (extended). Let us now compute the covariant derivative of the external forces $$\frac{\partial f_i}{\partial x^p} + \gamma^h_{ip} \left( f_h - d_{hk}
\dot{q}^k \right) = -k_p^i + \gamma^h_{ip} \left( f_h - d_{hk}
\dot{q}^k \right)$$ with $-\gamma^1_{11} = \gamma^2_{12} = \gamma^2_{21} = \gamma^2_{22} =
\frac{ \left( a_2 \cos q_2 + a_3 \right) a_2 \sin q_2} {a_1 a_3 -a_2^2
\cos^2 q_2 -a_3^2}$, $\gamma^1_{12} = \gamma^1_{21} = \gamma^1_{22} =
- \frac{a_3 a_2 \sin q_2} {a_1 a_3 -a_2^2 \cos^2 q_2 -a_3^2}$ and $\gamma^2_{11} = \frac{ \left( a_1 + 2 a_2 \cos q_2 \right) a_2 \sin
q_2} {a_1 a_3 -a_2^2 \cos^2 q_2 -a_3^2}$.
The spring gain $k_p^i$ stabilizes the system, whereas the supporting force $f_h - d_{hk} \dot{q}^k$ can (de)-stabilize the system proportional to the magnitude of the supporting force. Note that $f_h
- d_{hk} \dot{q}^k$ can (de)-stabilize a curved system is unavoidable since here no constant or parallel (force) vectors exist, whose covariant derivative vanishes [@Lovelock].
Computing $\lambda$ from (\[eq:characteristicHamiltonian\]) then allows with Theoreom \[th:Hamiltoniancontinuous\] to compute bounds on the velocity $\dot{q}^j$ and external forces $f_h - d_{hk}
\dot{q}^k$ for which global contraction behavior can be concluded.
System responses to a control input $\tau_i = (\sin t, \cos 5 t)$, initial conditions $q^i(0) = (-\frac{\pi}{2}, \pi)$ rad, $\dot{q}^i(0)
= (3, -3)$ rad/s, $\hat{q}^i(0) = (- \frac{\pi}{2}, \pi)$ rad, $\hat{\dot{q}}^i(0) = (-5, 5)$ rad/s and parameters $m_1 = 1$ kg, $l_1
= 1$ m, $m_e = 2$ kg, $I_1 = 0.12$ kgm$^2$, $l_{c1} = 0.5$ m, $I_e =
0.25$ kgm$^2$, $l_{ce} = 0.6$ m, $d_h^j = 5$ Nms/rad, $k_h^i = 5$ Nm/rad are illustrated in figure \[fig:robotqhat1\] and \[fig:robotwhat1\]. The solid lines represent the real plant, and the dashed lines the observer estimate.
![Positions of two-link robot[]{data-label="fig:robotqhat1"}](robotqhat1.eps)
![Velocities of two-link robot[]{data-label="fig:robotwhat1"}](robotwhat1.eps)
The above observers provide a simple alternative to current design methods \[see e.g., Berghuis and Nijmeyer, 1993; Marino and Tomei, 1995\], and guarantees local (for bounded velocities and time-varing inputs) exponential convergence.
Note that Theoreom \[th:Hamiltoniancontinuous\] can also be used to bound the diverging behavior, caused by the concave inertia when the robot arm is pointing inwards, of the double inverted pendulum, when no damping or stabilizing potential force is applied. \[exrobot\]
[ ]{}
[**:** ]{} Biology found a solution to the problem that a time-varying supporting torque can destabilize a system.
Recently, there has been considerable interest in analyzing feedback controllers for biological motor control systems as combinations of simpler elements, or motion primitives. For instance [@Bizzi] and [@Mussa] stimulate a small number of areas (A, B, C, and D) in a frog’s spinal cord and measure the resulting torque angle relations.
Force fields seem to add when different areas are stimulated at the same time so that [@Bizzi] and [@Mussa] propose the following biological control inputs $$f_i = - \sum_{l=1}^n k_l(t) f_{il} (q^j)$$ where each single torque $k_l(t) f_{il} (q^j)$ results from the stimulation of area $l$ in the spinal cord. With force measurements the above authors did show that the covariant derivative $\frac{\partial f_{il}}{\partial x^p} + \gamma^h_{ip} f_{hl}$ of $f_{il}$ with respect to the inertia tensor of the frog’s leg or body is uniformly positive definite.
Likely candidates for $k_i(t)$ are positive upper and lower bounded sigmoids and pulses and periodic activation patterns.
Using Theorem \[th:Hamiltoniancontinuous\] or the discussion in [Example \[exrobot\]]{} with sufficient damping then allows to compute a maximal $\dot{q}^j$ for which exponential convergence to a single motion is guaranteed.
Note that the achievement of tracking control with a proportional gain $k_l(t)$ rather than an additional supporting force as in [Example \[exrobot\]]{} has the advantage that the supporting force has no impact on the contraction behavior anymore. \[exmotionprimitives.\]
Concluding remarks
==================
The research in this paper can be extended in several directions, as the development suggests.
Some of the extensions will likely require the combination of the above results with a simplifying metric pre-transformation, as mentioned in section 3.1. In particular, classical transformation ideas in nonlinear control such as feedback linearization and flatness [@flatness] typically use linear time-invariant target dynamics, while the framework provided in this paper should allow considerably more flexibility. This, combined with the fact that a metric transformation such as $\ \delta {\bf z} = {\bf \Theta} \delta
{\bf x}\ $ need not be integrable (i.e. does not require an explicit ${\bf z}$ to exist), could potentially lead to useful generalisations of these methods.
[**Acknowledgement**]{} The authors are grateful to Yong Zhao for performing the simulations and for stimulating discussions, and to Wei Wang for thoughtful comments and suggestions.
[XX]{}
Aghannan, N., Rouchon, P., An Intrinsic Observer for a Class of Lagrangian Systems, [*I.E.E.E. Trans. Aut. Control*]{}, [**48(6)**]{} (2003). Aylward E., Parrilo P., and J.J.E. Slotine, Algorithmic search for contraction metrics via SOS programming, [*submitted to the 2006 American Control Conference*]{}. Bertsekas, D., and Tsitsiklis, J., Parallel and distributed computation: numerical methods, [*Prentice-Hall*]{}, 1989. Bizzi E., Giszter S.F., Loeb E., Mussa-Ivaldi F.A., Saltiel P., [*Trends in Neurosciences. Review 18:442*]{}, 1995. Bryson A., Ho, Y., Applied Optimal Control, [*Taylor and Francis*]{}, 1975. Fliess M., Levine J., Martin Ph., and Rouchon P., Flatness and defect of nonlinear systems: introductory theory and examples. [*Int. J. Control, 61(6)*]{}, 1995. Goldstein H., Classical Mechanics, [*Addison Wesley*]{}, 1980. Hartmann, P. Ordinary differential equations, [*John Wiley $\&$ Sons, New York*]{}, 1964. Jouffroy J. and J. Opderbecke. Underwater vehicle trajectory estimation using contracting PDE-based observers, [*American Control Conference*]{}, Boston, Ma, 2004. Kailath, T., Linear Systems, [*Prentice Hall*]{}, 1980. Lewis, D.C., Metric properties of differential equations, [*American Journal of Mathematics,*]{} [**71**]{}, pp. 294-312, 1949. Lohmiller, W., and Slotine, J.J.E., On Contraction Analysis for Nonlinear Systems, [*Automatica, 34(6)*]{}, 1998. Lohmiller, W., and Slotine, J.J.E., Nonlinear Process Control Using Contraction Theory, [*A. I. Che. Journal*]{}, March 2000. Lohmiller, W., and Slotine, J.J.E., Contraction Analysis of Nonlinear Distributed Systems, [*International Journal Of Control, 78(9)*]{}, 2005. Lovelock D., and Rund, H., Tensors, Differential Forms, and Variational Principles, [*Dover*]{}, 1989. Mussa-Ivaldi, F.A., [*I.E.E.E. International Symposium on Computational Intelligence in Robotics and Automation*]{}, 1997. Nguyen, T.D., and Egeland, O. Observer Design for a Towed Seismic Cable, [*American Control Conference, Boston*]{} (2004) Shamma, J., and Gurdal, A., Dynamic Fictitious Play, Dynamic Gradient Play and Distributed Convergence to Nash Equilibra, [*IEEE Transactions on Automatic Control*]{}, March 2005 Slotine and Li, Applied Nonlinear Control, [*Prentice Hall*]{}, 1991. Zhao, Y., and Slotine, J.J.E., Discrete Nonlinear Observers for Inertial Navigation, [*Systems and Control Letters, 54(8)*]{}, 2005. The Restricted Three Body Problem, http://www.physics.cornell.edu/ sethna/teaching/sss/jupiter/Web/Rest3Bdy.htm
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The series solution to Laplace’s equation in a helical coordinate system is derived and refined using symmetry and chirality arguments. These functions and their more commonplace counterparts are used to model solenoidal magnetic fields via linear, multidimensional curve-fitting. A judicious choice of functional forms, a small number of free parameters and sparse input data can lead to highly accurate, fine-grained modeling of solenoidal magnetic fields, including helical features arising from the winding of the solenoid, with overall field accuracy at better than one part per million.'
author:
- Brian Pollack
- Ryan Pellico
- Cole Kampa
- Henry Glass
- Michael Schmitt
bibliography:
- 'main.bib'
title: 'Modeling Magnetic Fields with Helical Solutions to Laplace’s Equation'
---
The authors would like to thank Matt Bonakdarpour for helpful discussions. We gratefully acknowledge the support provided by the Department of Energy under award number DE-SC0015910. This document was prepared by members of the Mu2e Collaboration using the resources of the Fermi National Accelerator Laboratory (Fermilab), a U.S. Department of Energy, Office of Science, HEP User Facility. Fermilab is managed by Fermi Research Alliance, LLC (FRA), acting under Contract No. DE-AC02-07CH11359.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We provide a uniform framework to study the exceptional homogeneous compact geometries of type $\mathsf{C}_3$. This framework is then used to show that these are simply connected, answering a question by Kramer and Lytchak, and to calculate the full automorphism groups.'
author:
- 'Jeroen Schillewaert\'
- 'Koen Struyve\'
bibliography:
- 'literaturliste.bib'
date:
title: 'On exceptional homogeneous compact geometries of type $\mathsf{C}_3$'
---
Introduction
============
In [@Tit:81] Tits introduced geometries of Coxeter type, which are geometries which locally are buildings. In particular, buildings are examples of such geometries. The motivation for these objects comes from applications to (finite) group theory, see for example [@Kan:86 C.7].
Flag-transitive finite geometries of type $\mathsf{C}_3$ have been classified by Aschbacher and Yoshiara in [@Asc:84] and [@Yos:96]. Such a geometry is either a building, or it is isomorphic with the Neumaier geometry on seven points (see [@Neu:84]).
In the compact connected case a similar classification has been obtained by Kramer and Lytchak in [@Kra-Lyt:14] (albeit missing one of the two exceptional cases). The eventual conclusion here is that such a geometry is either covered by a building, or it is isomorphic to one of two exceptional geometries. It are precisely these geometries which are the subject of the current paper.
The two exceptional geometries were first encountered in the study and classification of polar actions on manifolds, see [@Pod-Tho:99], [@Fan-Gro-Tho1], [@Fan-Gro-Tho2] and [@Gor-Kol]. They are associated to actions of ${\mathsf{SU}}(3) \times {\mathsf{SU}}(3)$ and ${\mathsf{SO}}(3) \times \mathsf{G}_2$ on the Cayley plane $\mathbb{OP}^2$.
In this paper we provide a uniform description for both geometries, and use this to show that these geometries are simply connected. This answers Problem 5 of [@Kra-Lyt:14]. We also use this description to obtain their full automorphism groups.
In Sections \[section:comp\] and \[section:geom\] we introduce basic notions needed later on. In Section \[section:geomcomp\] both of the exceptional geometries are constructed in an uniform way, and we discuss briefly their structure. Section \[section:auto\] discusses the automorphism group, the last section \[section:simple\] is devoted to the proof of their simple connectedness.
[**Acknowledgement.**]{} The first author wants to thank Linus Kramer profoundly for the hospitality and the interesting mathematical discussions during his stay at the University of Münster as a Von Humboldt Fellow. Both authors would like to thank Alexander Lytchak for helpful comments and discussions on this problem.
Composition algebras {#section:comp}
====================
Let ${\mathbb{A}}$ be a composition algebra over the real numbers ${\mathbb{R}}$, so ${\mathbb{A}}$ is either ${\mathbb{R}}$ itself or ${\mathbb{C}}$, ${\mathbb{H}}$ or ${\mathbb{O}}$ (by Hurwitz’s theorem). The algebra ${\mathbb{A}}$ comes equipped with a norm $ | \cdot |: {\mathbb{A}}\to {\mathbb{R}}$ and standard involution $\bar{\cdot}$. Let $k$ be either ${\mathbb{R}}$ of ${\mathbb{C}}$ and assume that $k$ is a subfield of ${\mathbb{A}}$.
The norm on ${\mathbb{A}}$ induces a positive-definite ${\mathbb{R}}$-bilinear form $\< \cdot , \cdot \> : {\mathbb{A}}\times {\mathbb{A}}\to {\mathbb{R}}$ on ${\mathbb{A}}$ when interpreted as an ${\mathbb{R}}$-vector space. Therefore we can consider orthogonality in ${\mathbb{A}}$. The orthogonal complement of a subspace $K$ in ${\mathbb{A}}$ is denoted by $\perp_{\mathbb{A}}K$.
We define the *$k$-pure* elements in ${\mathbb{A}}$ as $\operatorname{Pu}_k({\mathbb{A}}) {\mathrel{\mathop:}=}\perp_{\mathbb{A}}k$. In particular we have a direct sum decomposition
$${\mathbb{A}}= k \oplus \operatorname{Pu}_k ({\mathbb{A}}).$$
Note that the standard involution acts on $\operatorname{Pu}_k({\mathbb{A}})$ as the map $x\mapsto -x$..
We now define an Hermitian inner product $( \cdot | \cdot ): {\mathbb{A}}\times {\mathbb{A}}\to k$ on ${\mathbb{A}}$ (as a right vector space) over $k$ by setting $(x | y)$ to be the $k$-part of the product $xy$. Note that if $k={\mathbb{R}}$, then $\< \cdot , \cdot \>$ and $( \cdot | \cdot )$ agree. This Hermitian inner product has the following properties.
$$\begin{aligned}
\operatorname{Re}(a | b) &= \< a, b \> \\
| (a | b)| &\leq |a | \cdot |b| \\
| (a | b)| &= |a | \cdot |b| \iff a,b \mbox{ are } k\mbox{-linearly dependent} \label{eq:CS} \\
(a |a) &= |a|^2\end{aligned}$$
where $a,b \in {\mathbb{A}}$. Proofs of these are either direct or are basic properties of inner products. For an element $a \in \operatorname{Pu}_k({\mathbb{A}})$ we denote the orthogonal complement of the vector line $ak$ in $\operatorname{Pu}_k({\mathbb{A}})$ by $a^\perp$.
The following result will be crucial for our construction.
\[prop:map\] Let ${\mathbb{A}}$ and ${\mathbb{B}}$ be two composition algebras over ${\mathbb{R}}$, both containing a common subfield $k$ which is either ${\mathbb{R}}$ or ${\mathbb{C}}$. Assume additionally that the dimension of ${\mathbb{A}}$ over $k$ is four, and where the $k$-dimension of ${\mathbb{B}}$ is at least the $k$-dimension of ${\mathbb{A}}$. Let $a,c \in \operatorname{Pu}_k({\mathbb{A}})$ and $b,d \in \operatorname{Pu}_k({\mathbb{B}})$ be of norm 1 such that $(a | c)=(b | d)$ and $ak \neq ck$. Then there exists a $k$-algebra morphism of ${\mathbb{A}}$ into ${\mathbb{B}}$ mapping the pair $(a,c)$ to $(b,d)$.
[*Proof. *]{}Note that $bk \neq dk$ by Equation \[eq:CS\]. We may assume without loss of generality that $(a|c) = (b|d)= 0$ by replacing $c$ with a suitable scalar multiple of $c - (a | c) a$ being of unit norm (and similarly for $d$). In particular this implies that $ac$ and $bd$ are again in $\operatorname{Pu}_k({\mathbb{A}})$, see e.g. Proposition 11.10 of [@CPP].
The elements $1,a,c$ and $ac$ then form a $k$-basis for ${\mathbb{A}}$, satisfying the following identities: $$\begin{aligned}
\overline{a} = a^{-1} &= -a \\
\overline{c} = c^{-1} &= -c \\
\overline{ac} = \overline{c} \overline{a} &= ca = -ac \\
le = - \overline{le} &= - \overline{e} \overline{l} = e \overline{l}\end{aligned}$$ where $l \in k$ and $e \in \operatorname{Pu}_k({\mathbb{A}})$. Note that these relations completely determine ${\mathbb{A}}$ as a $k$-algebra. As analogous properties hold for $b$ and $d$, we can extend the map $(1,a,c,ac) \mapsto (1,b,d,bd)$ to a $k$-algebra morphism from ${\mathbb{A}}$ into ${\mathbb{B}}$.
Geometries {#section:geom}
==========
A *geometry over a set $I$* is a system $\Gamma {\mathrel{\mathop:}=}(V,\tau, *)$, consisting of a set $V$, a surjective map $\tau: V \to I$, and a binary symmetric relation $*$ on $V$ such that for any two elements $x,y$ of $V$ whose images under $\tau$ are identical, the relation $x * y $ holds if and only if $x=y$. The relation ‘$*$’ is the *incidence relation*, the image by $\tau$ of an element or a subset of $V$ is its *type*.
A *flag* of the geometry $\Gamma$ is a set of pairwise incident elements of $V$. The *type* of a flag is its image under $\tau$. The *corank* of a flag is $| I |$ minus the size of its type. Let $X$ be a flag, and let $Y$ be the set of all elements of $V \setminus X$ incident to $X$. The system $\Gamma_X := (Y, \tau\vert_Y, * \cap (Y \times Y))$ forms a geometry over $I \setminus \tau(X)$ and is called the *residue of $X$* in $\Gamma$. For a $J \subseteq I$, the set of all flags of the type $J$ as subset of $V^J$ is called the *flag variety* $V_J$.
The geometry $\Gamma$ is *connected* if the graph with vertices $V$ and adjacency relation ‘$*$’ is connected. A geometry is *thick* if every flag of corank 1 is contained in at least three maximal flags.
A *generalized $n$-gon* (with $2 \leq n < \infty$) is a thick rank 2 geometry such that the associated graph has girth $2n$ (i.e. the smallest cycle has length $2n$) and diameter $n$. Generalized triangles correspond exactly with (axiomatic) projective planes.
We call the geometry $\Gamma$ a *compact geometry* if the set of vertices $V$ carries a compact Hausdorff topology such that for every $J \subseteq I$ the flag variety $V_J$ is closed in the compact product space $V^J$.
Geometries of type $\mathsf{C}_3$
---------------------------------
A connected geometry over the set $\{1,2,3\}$ is a *geometry of type $\mathsf{C}_3$* if
- the residue of a vertex of type 1 is a generalized quadrangle,
- the residue of a vertex of type 2 is a generalized digon,
- the residue of a vertex of type 3 is a generalized triangle.
One often calls the vertices of type 1 the *points*, those of type 2 *lines*, and of type 3 *planes*. In this sense we can speak of *collinearity* of points and *coplanarity* of lines.
By a result of Tits ([@Tit:81 6.2.3]), the buildings of type $\mathsf{C}_3$ are precisely those geometries of type $\mathsf{C}_3$ in which no pair of points is incident with more than one common line.
The next lemma captures part of the structure of these geometries, which is well known, see e.g. [@Pas:94 Exercise 7.7].
\[lem:c3\] Let $\Gamma$ be a geometry of type $\mathsf{C}_3$.
i. For a point $p$ and a line $L$ there exists at least one point $q$ incident with $L$ and collinear with $p$.
ii. For a point $p$ and plane $\pi$, there exists at least one line $L$ incident with $\pi$ and having a common incident plane with $p$.
[*Proof. *]{}Since the geometry is connected and since the residue of a vertex of type 3 is a generalized triangle (i.e. a projective plane) there exists a (finite) path from $p$ to $L$ in the incidence graph alternating between points and lines. Let $a*D*b*E*c*F$ be a path where $a,b,c$ are points and $D,E,F$ are lines. Consider a plane $\pi* F$. Since the residue of $c$ is a generalized quadrangle there exist a plane $\pi' * E$ with $\pi' \cap \pi=L_c$. Since the residue in $b$ is a generalized quadrangle there exists a plane $\pi'' * D$ with $\pi'\cap \pi"=L_b$. Then $d\in L_b\cap L_c$ is a point in $\pi$ collinear with $a$. Considering the residue in $d$ yields there exists a plane containing $a$ intersecting $\pi$ in a line $L_d$. Since $\pi$ is a projective plane the intersection $L_d\cap F$ contains a point $e$. Hence we have found a shorter path from $a$ to $F$, namely $a * \langle a,e\rangle * e *F$, which implies (i).
For (ii) consider any line $L$ in $\pi$. Then by (i) there is a point $q$ incident with $L$ and collinear with $p$. Considering the residue of $q$ proves the statement.
Coverings
---------
Let $\Gamma{\mathrel{\mathop:}=}(V,\tau, *)$ be a geometry of type $\mathsf{C}_3$. From $\Gamma$ one can construct a connected simplicial complex with vertices the elements of $V$ and simplices the flags of $\Gamma$. We denote the metric realization of this simplicial complex by $| \Gamma |$. Connected covers of $| \Gamma |$ then correspond with 2-covers of the geometry $\Gamma$ (see [@Pas:94 Ch. 12]), which are again geometries of type $\mathsf{C}_3$.
By this correspondence one can consider the *universal cover* $\widetilde{\Gamma}$ of $\Gamma$, and say that the geometry is simply connected if it is its own universal cover.
Geometries from composition algebras {#section:geomcomp}
====================================
In this section we construct a class of geometries of type $\mathsf{C}_3$ starting from two composition algebras with a common subfield. The construction itself is contained in Section \[sec:constr\], the verification of the claim in Section \[sec:verif\].
Lastly, in Section \[sec:pos\], we list the different possibilities for the composition algebras and show the existence of a compact flag-transitive automorphism group.
Construction of $\Gamma$ {#sec:constr}
------------------------
We define a geometry $\Gamma$ as follows. Let ${\mathbb{A}}$ and ${\mathbb{B}}$ be two composition algebras over ${\mathbb{R}}$, containing a common subfield $k$ which we assume to be either ${\mathbb{R}}$ or ${\mathbb{C}}$, such that ${\mathbb{A}}$ is four-dimensional over $k$ and where the ($k$-)dimension of ${\mathbb{B}}$ is at least the ($k$-)dimension of ${\mathbb{A}}$.
Scalar multiplication, as well as projectivization, is always understood to be over $k$.
- The points are the vector lines in $\operatorname{Pu}_k({\mathbb{A}})$ as vector space over $k$.
- Lines are of the form $[a,b]$ where $a \in \operatorname{Pu}_k({\mathbb{A}})$, $b \in \operatorname{Pu}_k({\mathbb{B}})$ and $|a|= |b| \neq 0$, up to multiplication by a common scalar multiple.
- The planes of the geometry are formed by the embeddings $\phi: {\mathbb{A}}\to {\mathbb{B}}$ as composition algebras over $k$.
Incidence is defined as follows. Every point is incident with every plane, a point $uk = \<u\>$ ($u \in \operatorname{Pu}_k({\mathbb{A}})$) is incident with a line $[a,b]$ if and only if $ a \in u^\perp$. A line $[a,b]$ and a plane $\phi$ are incident if and only if $\phi(a) = b$.
\[rem:flat\] Often one calls a geometry of type $\mathsf{C}_3$ where every point is incident with every plane *flat* (see for example [@Pas:94 4.3.2]). A second property that our geometry $\Gamma$ has is that if two lines are incident with at least two common points, then every point incident with one line is also incident with the other. (Another way to say this is that the point shadows of both lines agree in this case.) The set of points of $\Gamma$ and the set of point shadows of lines (with duplicity removed) form a projective plane over $k$.
$\Gamma$ is a geometry of type $\mathsf{C}_3$ {#sec:verif}
---------------------------------------------
Before we prove that $\Gamma$ is indeed a geometry of type $\mathsf{C}_3$, we obtain a criterion for coplanarity of lines. (Note that this criterion is not influenced by taking scalar multiples.)
\[lem:coplanar\] Two lines $[a,b]$ and $[c,d]$ are coplanar if and only if $( a | c ) = (b | d)$.
[*Proof. *]{}The ‘only if’ direction is clear as embeddings of composition algebras preserve the inner product. The other direction follows from Proposition \[prop:map\].
The geometry $\Gamma$ is of type $\mathsf{C}_3$.
[*Proof. *]{}Clearly $\Gamma$ is thick, connected, with the residue of any given line being a digon, and the residue of any plane being a projective plane over $k$. So it remains to verify that the residue of a given point $\< u \>$ is a generalized quadrangle.
All planes belong to this residue and a line $[a,b]$ belongs to it if and only if $a \in u^\perp$.
In order to verify that the residue is a generalized quadrangle consider a plane $\phi$ and a line $[a,b]$ with $\phi(a) \neq b$. It suffices to show existence and uniqueness of a line $[c,\phi(c)]$ (so incident with $\phi$) such that $( a | c )=(b |\phi(c) )$ (up to taking a scalar product of $c$) which states that $[c,\phi(c)]$ coplanar with $[a,b]$.
This is equivalent to $( \phi(a)- b | \phi(c) )=0$, which implies that $\phi(c)$ is contained in the hyperplane $\xi:=(\phi(a)- b)^\perp$ of $\operatorname{Pu}_k({\mathbb{A}})$.
In order to have a unique possibility for $c$ (up to scalar products), it suffices that the subspace $\xi' = \phi(u^\perp)$ intersects $\xi \cap \phi(\operatorname{Pu}_k({\mathbb{A}}))$ in a vector line. This is equivalent with $\xi' \neq \xi \cap \phi(\operatorname{Pu}_k({\mathbb{A}}))$.
Assume by way of contradiction that we do have equality. Then $\phi(a)$, which is an element of $\xi'$, is also contained in $\xi$. This would imply that $( \phi(a)- b | \phi(a) ) = 0 $, which is equivalent to $(b, \phi(a) ) = | \phi(a) |^2$. As $|b| = |a|$, the Cauchy-Schwarz inequality now implies that $b = l \phi(a)$, with $l \in k$ of unit norm. Plugging this back into $( \phi(a)- b | \phi(a) ) = 0 $ yields that $l=1$ or equivalently $b = \phi(a)$, which is a contradiction. This proves the proposition.
Another way to get the same conclusion is to interpret the condition $|a| = |b|$ for the lines through a fixed point as a Hermitian form on a $(\dim_k {\mathbb{A}}+ \dim_k {\mathbb{B}}- 3)$-dimensional vector space over $k$, and obtain a generalized quadrangle in this way.
Possibilities for $k$, $A$ and $B$ {#sec:pos}
----------------------------------
By Hurwitz’s theorem on composition algebras we only have the following three possibilities for $k$, ${\mathbb{A}}$ and ${\mathbb{B}}$ (up to isomorphism). $$\begin{aligned}
k& = {\mathbb{R}}, {\mathbb{A}}= {\mathbb{H}}, {\mathbb{B}}= {\mathbb{H}}, \\
k& = {\mathbb{R}}, {\mathbb{A}}= {\mathbb{H}}, {\mathbb{B}}= {\mathbb{O}}, \\
k& = {\mathbb{C}}, {\mathbb{A}}= {\mathbb{O}}, {\mathbb{B}}= {\mathbb{O}}.\end{aligned}$$
In each of these cases one can construct the following compact flag-transitive automorphism group of the associated geometry $\Gamma$. Set $G = {\mathrm{Aut}}_k {\mathbb{A}}\times {\mathrm{Aut}}_k {\mathbb{B}}$, and the action of an element $g {\mathrel{\mathop:}=}(\alpha, \beta) \in G$ on the geometry to be
- points: $ka \mapsto \alpha(ka)$
- lines: $[a,b] \mapsto [\alpha(a),\beta(b)]$
- planes: $\phi \mapsto \beta\phi\alpha^{-1}$.
It is straightforward to check this action is bijective and preserves incidence. As ${\mathrm{Aut}}_{\mathbb{R}}{\mathbb{H}}\simeq {\mathsf{SO}}(3)$, ${\mathrm{Aut}}_{\mathbb{R}}{\mathbb{O}}\simeq \mathsf{G}(2)$ and ${\mathrm{Aut}}_{\mathbb{C}}{\mathbb{O}}\simeq {\mathsf{SU}}(3)$ (see [@CPP]), we obtain the following possibilities for $G$ (using the same ordering as before).
$$\begin{aligned}
{\mathsf{SO}}(3) \times & {\mathsf{SO}}(3), \\
{\mathsf{SO}}(3) \times & \mathsf{G}_2, \\
{\mathsf{SU}}(3) \times & {\mathsf{SU}}(3). \end{aligned}$$
The action is faithful except in the last case, for which the action has a kernel of size three.
The group $G$ clearly admits a compact topology and acts continuously for the standard topology on $\Gamma$. Flag-transitivity is easily seen by the transitivity of ${\mathrm{Aut}}_k {\mathbb{A}}$ on $\operatorname{Pu}_k({\mathbb{A}})$, and the fact that, when assuming ${\mathbb{A}}$ to be a $k$-subalgebra of ${\mathbb{B}}$, each $k$-algebra injective morphism from ${\mathbb{A}}$ into ${\mathbb{B}}$ extends to an automorphism of ${\mathbb{B}}$.
We can hence apply the classification made in Theorem A of [@Kra-Lyt:14] and conclude that the geometry $\Gamma$ is either covered by a building, or is one of two unique exceptional geometries of type $\mathsf{C}_3$. Our last two possibilities correspond with the two exceptional geometries, the geometry for the first possibility is therefore covered by a building of type $\mathsf{C}_3$.
The case $G = {\mathsf{SO}}(3) \times SO(3)$
--------------------------------------------
For this case we provide an explicit description of the covering. The covering geometry is the projective quadric $\Delta$ defined by the following equation.
$$Q(X_0,X_1,X_2,X_3,X_4,X_5,X_6)=X_0^2 + X_1^2 + X_2^2 - X_3^2 - X_4^2 -X_5^2 - X_6^2=0$$
The group ${\mathsf{SO}}(4)$ acts on the last four coordinates, in particular on the unit sphere $\mathbb{S}^3$ in the three-dimensional space spanned by these. Let $H$ be the subgroup of left isoclinic rotations (which is isomorphic with ${\mathsf{SU}}(2)$ as well as the multiplicative group of unit quaternions, which themselves can be identified with the sphere $\mathbb{S}^3$). This group $H$ acts sharply transitive on $\mathbb{S}^3$.
Let $G$ be the group ${\mathsf{SO}}(3) \times {\mathsf{SO}}(4)$ (where ${\mathsf{SO}}(3)$ acts on the first three coordinates, and ${\mathsf{SO}}(4)$ is as before). Then $H\leq G$ and the quotient $G / H$ is isomorphic to ${\mathsf{SO}}(3) \times {\mathsf{SO}}(3)$.
The action of $G$ on $\Delta$ is well-understood in the context of Veronese and polar representations, in particular one has the following lemma.
\[lem:transitiveG\] The group $G$ acts chamber-transitively on the projective quadric $\Delta$.
If we can show that the group $H$ does not fix any point, line or plane, then the quotient geometry $\Delta'$ is a $\mathsf{C}_3$ geometry on which the group quotient $G / H \simeq {\mathsf{SO}}(3) \times {\mathsf{SO}}(3)$ acts chamber-transitively.
No non-identity element of $H$ maps a point of $\Delta$ to a collinear point on it.
[*Proof. *]{}Suppose that an element $g \in G$ maps some point $p$ of $\Delta$ to a collinear point $q$. By Lemma \[lem:transitiveG\] and the fact that $H$ is a normal subgroup of $G$ we may assume w.l.o.g. that the point $p$ is represented by $(1,0,0,1,0,0,0)$. The point $q$ can then be represented by $(1,0,0,a,b,c,d)$ where $a^2 +b^2 + c^2 + d^2 = 1$.
In order for these points to be collinear on $\Delta$ we need that for the associated bilinear form $B(x,y)=\frac{1}{2}(Q(x+y)-Q(x)-Q(y))$ we have $B(p,q)=0$. This implies that $\frac{1}{2}(4 - (1+a)^2-b^2-c^2-d^2) = 0$. Combined with the previous condition $a^2 +b^2 + c^2 + d^2 = 1$ it follows that $a=1$ and hence $b=c=d=0$, so $p = q$. As $H$ acts sharply transitive on the sphere $\mathbb{S}^3$, we have that $g$ is the identity element. This proves the lemma.
The group $H$ acts freely on the projective quadric $\Delta$.
[*Proof. *]{} This follows directly from the previous lemma.
One can verify that the geometry and group action constructed in this section is the same one as in Section \[sec:pos\] by comparing the stabilizer of a chamber and the subsimplices of it.
The full automorphism group of the geometry $\Gamma$ {#section:auto}
====================================================
This section is devoted to determining the full automorphism group ${\mathrm{Aut}}(\Gamma)$ of $\Gamma$, which will turn out to be very close to the compact group given in Section \[sec:pos\].
Statement of results {#section:fullclaim}
--------------------
We claim that the full automorphism group is given by the following table (listed case-by-case):
$$\begin{aligned}
k& = {\mathbb{R}}, {\mathbb{A}}= {\mathbb{H}}, {\mathbb{B}}= {\mathbb{H}}: {\mathsf{SO}}(3) \times {\mathsf{SO}}(3), \\
k& = {\mathbb{R}}, {\mathbb{A}}= {\mathbb{H}}, {\mathbb{B}}= {\mathbb{O}}: {\mathsf{SO}}(3) \times \mathsf{G}_2, \\
k& = {\mathbb{C}}, {\mathbb{A}}= {\mathbb{O}}, {\mathbb{B}}= {\mathbb{O}}: (( {\mathsf{SU}}(3) \times {\mathsf{SU}}(3)) / C_3) \rtimes C_2.\end{aligned}$$
The cyclic group $C_2$ consisting of two elements arises from complex conjugation. The cyclic group $C_3$ is the kernel of the map of the group ${\mathsf{SU}}(3)$ to its projectivization $\mathsf{PSU}(3)$, i.e. the center of ${\mathsf{SU}}(3)$, matrices $\zeta I$, where $\zeta$ is a third root of unity and $I$ is the identity matrix.
From the results of Section \[sec:pos\] it easily follows that these are indeed automorphism groups of the associated geometries. It only remains to check if these are the full groups.
Image into $\mathsf{P\Gamma L}(3, k)$ {#section:image}
-------------------------------------
The set of points and the set of point shadows of lines (removing duplicity) form a projective plane $\Sigma$ over $k$, see also Remark \[rem:flat\]. Hence the automorphism group of $\Gamma$ can be mapped into the full automorphism group ${\mathrm{Aut}}(\Sigma)$ of this plane, which is $\mathsf{P\Gamma L}(3, k)$. We aim to show that the image of ${\mathrm{Aut}}(\Gamma)$ into ${\mathrm{Aut}}(\Sigma)$ is contained in the centralizer of the polarity induced by the Hermitian inner product $(\cdot | \cdot )$, as expected.
As the compact automorphism group listed in Section \[sec:pos\] is flag-transitive, we may fix the point $x {\mathrel{\mathop:}=}\<a \>$, and line $L {\mathrel{\mathop:}=}[a',b]$, where $a, a', a''$ is an orthogonal $k$-basis for $\operatorname{Pu}_k({\mathbb{A}})$. Let $M$ be a second line with the same point shadow as $L$, which we may assume to be of the form $M {\mathrel{\mathop:}=}[a', d]$.
Consider the lines $N$ through $x$ which are coplanar with both $L$ and $M$. Such a line is of the form $[c,f]$.
The point shadows for varying $N$ are all identical if and only if $\< b \> = \<d \>$. In this case the common point shadow is formed exactly by the points corresponding to the vector lines in the $k$-span of $\{a,a'\}$.
[*Proof. *]{}The point shadows are identical if and only if there is a unique solution for $c$ up to scalar multiples.
The equations that have to be satisfied in order for a line $N {\mathrel{\mathop:}=}[c,f]$ go through $x$ and being coplanar with both $L$ and $M$ are the following. $$\begin{aligned}
&c \in \< a', a'' \> \\
(a' | c)& = (b | f) = (d | f)\end{aligned}$$ Note that $c \notin \<a' \>$. If $c = a'l + a'' $ is a valid solution, then $c = a'l - a'' $ is also. Hence in order to have unicity for $c$ up to scalar multiples we need that $l=0$, or equivalently that $(b |f) =0$ for every $f$ such that $(b |f) = (d|f)$. The latter condition can be rewritten as $(b-d | f) =0$. As the subspace $(b-d)^\perp$ is contained in $b^\perp$ if and only if the vector lines through $b$ and $d$ are identical, it follows that $\< b \> = \<d \>$.
The element $c$ is then a scalar multiple of $a''$, which determines the point shadow of $N$.
Note that the point shadow of such an $N$ forms a line through $x$ orthogonal to the point shadow of $L$. One can hence recognize the polarity $\rho$.
We conclude that the image of ${\mathrm{Aut}}(\Gamma)$ in ${\mathrm{Aut}}(\Sigma)$ is contained in $\mathsf{SO}(3)$ if $k = {\mathbb{R}}$, and $\mathsf{PSU}(3) \rtimes C_2$ if $k = {\mathbb{C}}$, where the cyclic group $C_2$ corresponds with complex conjugation. Note that this corresponds with the possibilities listed in Section \[section:fullclaim\].
Kernel of the map {#section:kernel}
-----------------
In this section we take a look at the kernel of the map into $\mathsf{P\Gamma L}(3, k)$ as considered in Section \[section:image\]. If we can show that this is exactly ${\mathrm{Aut}}_k {\mathbb{B}}$, then the claim from Section \[section:fullclaim\] readily follows. As the kernels for the groups of automorphisms listed in Section \[section:fullclaim\] still act transitive on the set of planes of the geometry $\Gamma$, we may restrict ourselves to considering the stabilizer of one such plane. This choice of a plane corresponds with an embedding of ${\mathbb{A}}$ into ${\mathbb{B}}$, hence we may assume that ${\mathbb{A}}$ is a $k$-subalgebra of ${\mathbb{B}}$.
Let $H$ be the stabilizer of the plane corresponding with the (natural) embedding of ${\mathbb{A}}$ into ${\mathbb{B}}$ in the kernel of the map into $\mathsf{P\Gamma L}(3, k)$. So $H$ fixes every line of the form $[a,a]$ ($a \in \operatorname{Pu}_k{\mathbb{A}}$), as well as every point of $\Gamma$.
Let $a$ in $\operatorname{Pu}_k({\mathbb{A}})$ be an element of unit norm, and consider the set of lines $K {\mathrel{\mathop:}=}\{[a,b] \vert b \in \operatorname{Pu}_k({\mathbb{B}}), |a| = |b| =1 \}$, which is naturally parametrized by the elements in $\operatorname{Pu}_k({\mathbb{B}})$ of unit norm and stabilized by $H$. In the next few lemmas we investigate how $H$ acts on $K$, which via the parametrization corresponds with an action of $H$ on the elements of unit norm in $\operatorname{Pu}_k({\mathbb{B}})$.
\[lemma:spaces\] The action of $H$ on the set $K$ preserves the intersection of vector subspaces of $\operatorname{Pu}_k({\mathbb{B}})$ (as a vector space over $k$) with the subset of elements of unit norm.
[*Proof. *]{}The lines $[a,b]$ in $K$ coplanar with some given line $[c,d]$ are determined by the linear equation $(a|c) = (b|d)$ over $k$ by Lemma \[lem:coplanar\]. We can therefore recognize the hyperplanes, and hence any vector subspace of $\operatorname{Pu}_k({\mathbb{B}}) \cup \{0\}$ intersected with the elements of unit norm.
\[lemma:ortho\] The action of $H$ on $K$ preserves orthogonality.
[*Proof. *]{}Consider a set $K' {\mathrel{\mathop:}=}\{[c,d] \vert b \in \operatorname{Pu}_k({\mathbb{B}}), |c| = |d| \}$ of lines with the same point shadow, where $a$ and $c$ are not scalar multiples of each other.
Then for a fixed line $[a,b] \in K$, the lines in $K$ which are coplanar with a line in $K'$ which is on its turn coplanar to $[a,b]$ are those lines $[a,f]$ where $f$ is such that there exists a $d \in \operatorname{Pu}_k({\mathbb{B}})$ with $|d| = |f| = |a|$ and $(a|c)=(b|d) = (f|d)$.
If this is not the entirety of lines in $K$, it will certainly not contain those lines $[a,f]$ with $(c|f) = 0$. This characterizes exactly those lines.
\[lemma:spread\] Let $a \in \operatorname{Pu}_k {\mathbb{A}}$ and $b \in \operatorname{Pu}_k {\mathbb{B}}$ such that $|a| = |b|$. It is then possible to reconstruct the lines of the form $[c,d]$ where $d$ is a scalar multiple of $b$ from the geometry.
[*Proof. *]{}We first consider the case where $(c|a) = 0$. The line $[a,b]$ is coplanar with a line $[c,f]$ ($f \in \operatorname{Pu}_k {\mathbb{B}}$ and $|f| = |c|$) if and only if $(f|d) =0$. As one can recognize orthogonality by Lemma \[lemma:ortho\], it follows that one can reconstruct the lines $[c,d]$ where $d$ is a scalar multiple from $b$ from this.
The case where $(c|a) \neq 0$ follows from applying the previous case twice to an element $e$ in $\operatorname{Pu}_k {\mathbb{A}}$ such that $(e|a) = (e|c) =0$, which is always possible to find (as $(e|a) = (e|c) =0$ define two linear equations in a three-dimensional space).
### The case $\dim_k {\mathbb{B}}= 4$ {#section:dim4}
We start by considering the case where ${\mathbb{B}}$ is four-dimensional over $k$, or, by the assumptions at the beginning of Section \[section:kernel\], that ${\mathbb{A}}$ equals ${\mathbb{B}}$. We will show that $H$ acts trivially on any line and plane of the geometry, which proves the claim.
Assume that $[a, b]$ is line of the geometry $\Gamma$ (so $|a| = |b|$) not fixed by a certain element $h\in H$. By Lemma \[lemma:spread\] and the fact that ${\mathbb{A}}= {\mathbb{B}}$ the element $h$ can map this line only to a line $[a,bl]$, with $l \in k$ and $|l| = 1$.
The lines of type $[c, c]$ ($c \in \operatorname{Pu}_k {\mathbb{A}}$) (which are fixed by $H$) coplanar with $[a,b]$ are those lines such that $(c|a) = (c|b)$. Note that there exists such $c$ up to scalar multiples, by considering the residue of a point incident with $[a,b]$, which is a generalized quadrangle. If $(c|a) \neq 0$, then we would have that $(c|b) = (c|b l) =(c|b) l$ implying that $l =1$ and that $[a,b]$ is fixed by $h$, which is a contradiction.
If there are only such coplanar lines $[c,c]$ where $(c|a) = 0$, then, by considering the residue of each point incident with $[a,b]$, the line $[a,b]$ is coplanar with each line of the form $[c,c]$ containing the point $\<a\>$. So $a$ and $b$ are orthogonal to the same elements and are hence scalar multiples from each other. Such a line is however also coplanar with fixed lines of the form $[c,d]$ where $c$ and $d$ are not scalar multiples of each other, and where $(c |a ) \neq 0$. As before one deduces that the line $[a,b]$ is fixed.
From this it follows that every line, and by extension every plane of $\Gamma$ will be fixed by $H$. This proves the claim.
### The case $\dim_k{\mathbb{B}}= 8$
We now assume that $\dim_k{\mathbb{B}}= 8$. In particular this implies that $k = {\mathbb{R}}$, ${\mathbb{A}}= {\mathbb{H}}$ and ${\mathbb{B}}= {\mathbb{O}}$. The following lemma allows us to recognize quaternion subalgebras of ${\mathbb{B}}$.
\[lemma:4sub\] The action of $H$ on $K$ preserves the intersection with four-dimensional $k$-subalgebras of ${\mathbb{B}}$.
[*Proof. *]{}Let $a'$ and $a''$ be elements of unit norm in $\operatorname{Pu}_k({\mathbb{A}})$ such that $(a|a') = 0$ and $a'' = a a'$. So $1, a, a'$ and $a''$ form an orthonormal $k$-basis for ${\mathbb{A}}$.
Let $1,b,b'$ and $b'' {\mathrel{\mathop:}=}b b'$ be an orthonormal $k$-basis for a four-dimensional $k$-subalgebra of ${\mathbb{B}}$. We are going to construct the line $[a,b'' l'']$ (so $b''$ up to a scalar product) given the lines $[a,b]$ and $[a,b']$ in $K$. By Lemma \[lemma:spread\] we can construct lines $[a',bl]$ and $[a'',b'l']$ from these lines. The unique plane containing both $[a',bl]$ and $[a'',b'l']$ is given by the embedding defined by $$\phi:
\left\{
\begin{array}{l}
a \mapsto b'' ll' \\
a'\mapsto bl \\
a'' \mapsto b'l' \\
\end{array}
\right.$$ The unique line of the form $[a,f]$ incident with the plane is the line $[a, b''ll']$, as desired. The statement of the lemma now follows from Lemma \[lemma:spaces\].
This lemma allows us to consider the subgeometry of $\Gamma$ corresponding with the choice ${\mathbb{B}}= {\mathbb{A}}= {\mathbb{H}}$, and conclude by the results of Section \[section:dim4\] that this subgeometry is completely fixed by $H$.
Recall the 3 cases we are considering and their proposed group of automorphism, see Section \[section:fullclaim\]. $$\begin{aligned}
k& = {\mathbb{R}}, {\mathbb{A}}= {\mathbb{H}}, {\mathbb{B}}= {\mathbb{H}}: {\mathsf{SO}}(3) \times {\mathsf{SO}}(3), \\
k& = {\mathbb{R}}, {\mathbb{A}}= {\mathbb{H}}, {\mathbb{B}}= {\mathbb{O}}: {\mathsf{SO}}(3) \times \mathsf{G}_2, \\
k& = {\mathbb{C}}, {\mathbb{A}}= {\mathbb{O}}, {\mathbb{B}}= {\mathbb{O}}: (( {\mathsf{SU}}(3) \times {\mathsf{SU}}(3)) / C_3) \rtimes C_2.\end{aligned}$$
The proposed group of automorphisms acts sharply transitively on the elements in ${\mathbb{B}}$ of unit norm and orthogonal to all elements in ${\mathbb{A}}$ (see for example [@Bae:02 4.1]), so fix such an element $b \in {\mathbb{B}}$.
Let $H'$ be the stabilizer of a line $[c,b]$ with $|c| = |b|$. (Note that the choice of $c$ does not matter.) Each element $a \in \operatorname{Pu}_k{\mathbb{A}}$ of unit norm generates a four-dimensional $k$-subalgebra of ${\mathbb{B}}$ together with $b$, which we can recover geometrically in the form of subgeometry $\Gamma'$ fixed by $H'$ by Lemma \[lemma:4sub\]. As we know that the group of automorphisms fixing all points of this subgeometry will be the group ${\mathsf{SO}}(3)$, and that $H$ additionally fixes the lines $[a,a]$ and $[a, b]$ of $\Gamma'$, we conclude that $H'$ fixes the subgeometry $\Gamma'$. Repeating this argument for different subgeometries we obtain that $H'$ acts trivially on $\Gamma$, whence the claim.
On simple connectedness of the geometry $\Gamma$ {#section:simple}
================================================
In this section we prove that the geometry $\Gamma {\mathrel{\mathop:}=}(V, \tau, *)$ of type $\mathsf{C}_3$ constructed in Section \[sec:constr\] is either simply connected, or covered by a building.
#### Sketch of proof. –
We start by considering the edge-path group of $\Gamma$, which we use to study the universal cover $\widetilde{\Gamma}$. From Proposition \[prop:control\] (proved in Sections \[section:red\] up to \[section:control\]) we show that this cover admits a compact topology with a flag-transitive continuous group acting on it. Such a geometry is subject to the classification of Kramer and Lytchak ([@Kra-Lyt:14]) and is hence known, from which one concludes that the cover is either a building or the exceptional geometry itself.
The edge-path group of $\Gamma$
-------------------------------
Instead of working with the fundamental group of $| \Gamma |$ it makes more sense to work with the equivalent edge-path group, which is more natural in the simplicial setting. We refer to [@Sei-Thr:80 Ch. 44] or to [@Spanier] for a detailed exposition.
For us, an *edge path* is a finite sequence of vertices in $V$ such that each two subsequent vertices are incident. (Note that such a sequence defines a unique sequence of edges on $| \Gamma |$.) The *length* of a path is the number of vertices in it minus one.
We now introduce two kinds of *elementary combinatorial deformations*. These are:
i. Replacing a subpath $(u)$ with $(u,v,u)$, where $u$ and $v$ are incident, or the converse operation.
<!-- -->
ii. Replacing a subpath $(u,v)$ with $(u,w,v)$, where $w$ is incident with both $u$ and $v$ (which is equivalent with $\{u,v,w\}$ being a flag), or the converse operation.
Two edge paths $\gamma$ and $\gamma'$ which can be transformed into each other by a finite number of elementary combinatorial deformations are said to be *combinatorially homotopic*. An edge path is *contractible* if it is combinatorially homotopic to a path of length zero. As combinatorial homotopies do not alter the begin or endpoint of a path, this implies that a contractible edge path begins and ends at the same point.
For a fixed vertex $v \in V$, the *edge path group* $E(\Gamma, v)$ consists of the equivalence classes of combinatorially homotopic edge paths starting and ending at the vertex $v$. The group multiplication is defined as concatenation.
The universal cover $\widetilde\Gamma$
--------------------------------------
It will be useful to have a concrete model of the universal cover. We will do this by constructing a geometry $\widetilde\Gamma {\mathrel{\mathop:}=}(\widetilde{V}, \widetilde{\tau}, \widetilde{*})$ using edge paths.
Fix a point $x \in V$. The set of vertices $\widetilde{V}$ consists of the equivalence classes of edge paths starting at $x$ under combinatorial homotopy. We set the type (i.e. its image under $\widetilde{\tau}$) of such an equivalence class of edge paths to be the type of the last vertex in any path in the class. Two equivalence classes of edge paths are incident if one can find an edge path of length $k$ in one class, and an edge path of length $k+1$ in the other class from which the first path can be obtained by removing the last vertex in the path.
Finally, the covering map $\rho: \widetilde{V} \to V$ is then defined by mapping an equivalence edge path to the last vertex of any path in it.
Lifting the compact topology to the universal cover
---------------------------------------------------
The goal of this section is to define a compact topology $\mathcal{T}$ on $\widetilde{V}$ starting from the compact topology on $V$. Our approach is based on Section 4 of [@Fan-Gro-Tho2] and Section 8 of [@Lyt:14].
Proving the following proposition will be the subject of Sections \[section:red\] up to \[section:control\].
\[prop:control\] For every natural number $k$ there exists a natural number $D(k)$ with the following property. Any two edge paths $\gamma$ and $\gamma'$ in the universal cover $\widetilde{\Gamma}$ of length at most $k$ with the same extremities are homotopic by a combinatorial homotopy consisting of at most $D(k)$ elementary combinatorial deformations.
We fix a point $\tilde{x}$ in the universal cover $\widetilde\Gamma$, such that $x {\mathrel{\mathop:}=}\rho(\tilde{x})$ is its image under the covering map $\rho$.
We say that a sequence of vertices $(\tilde{v}_n)_{n \in \mathbb{N}}$ of vertices in $\widetilde\Gamma$ converges to a vertex $\tilde{v} \in \widetilde{V}$, if one can choose representative edge paths $(x, u_1^n, u_2^n, \dots, u_k^n)$ for each $\tilde{v}_n$, all of the same length $k$, and such that each of the sequences $(u_i^n)_{n \in {\mathbb{N}}}$ converges to some vertex $u_i$ (in the compact topology on $V$) and $(x, u_1, \dots, u_k)$ is a valid edge path representing $\tilde{v}$.
This notion of convergence defines closed sets and hence a topology $\mathcal{T}$ on $\widetilde{V}$. This topology is sequentially compact, as the set of paths of fixed length $k$ starting at $x$ is closed in $V^{k+1}$. The topology admits a dense countable subset as this also holds for $V$, and by the fact that every vertex in $\widetilde{V}$ can be represented by an edge path whose length is globally bounded.
\[secondcountable\] One can define the topology $\mathcal{T}$ in a slightly different way as follows. Consider the set of edge paths starting at a single point $p$ (denote this set by $\overline{V}$). One can interpret this a subset of the countable product space $\Pi_{i=0}^\infty V$. On this set of edge paths we define equivalences as above, thus defining a quotient topology $\mathcal{T}'$. Since countable products of second countable spaces are again second countable, and since subspaces of second countable spaces are second countable and since the quotient map defined by this equivalence relation is open, $\mathcal{T}'$ is second countable. By Proposition 2.4 of [@Siwiec] a convergent sequence in the quotient topology lifts to a convergent sequence in the topology itself. As $\mathcal{T}$ was described by convergence and since first countability implies that a set is closed if and only if it is sequentially closed, we can conclude that $\mathcal{T}'$ is equal to $\mathcal{T}$. Hence $\mathcal{T}$ is second countable.
This topology will have nice properties, as exhibited by the following lemmas.
\[lem:onelimit\] Each sequence has at most one limit.
[*Proof. *]{}Let $(\tilde{v}_n)_{n \in \mathbb{N}}$ be a sequence of vertices, and let $(x, u_1^n, u_2^n, \dots, u_k^n)$ and $(x, {u'}_1^n, {u'}_2^n, \dots, {u'}_l^n)$ be two edge paths both representing $\tilde{v}_n$, one of length $k$, the other of length $l$, for every $n \in \mathbb{N}$. Suppose that the paths $(x, u_1^n, u_2^n, \dots, u_k^n)$ converge to an edge path $(x, u_1^n, u_2^n, \dots, u_k^n)$, and the paths $(x, {u'}_1^n, {u'}_2^n, \dots, {u'}_l^n)$ to an edge path $(x, u'_1, u'_2, \dots, u'_l)$. We have to proof that these two limit edge paths are combinatorially homotopic.
Each pair of paths $(x, u_1^n, u_2^n, \dots, u_k^n)$ and $(x, {u'}_1^n, {u'}_2^n, \dots, {u'}_l^n)$ are combinatorially homotopic by a bounded number of elementary combinatorial deformations (see Proposition \[prop:control\]). By sequential compactness these combinatorial homotopies (which can be thought of as sequences of edge paths) for each $n$ admit a convergent subsequence, which will be the desired combinatorial homotopy.
\[lem:independent\] The topology $\mathcal{T}$ is independent of the choice of points $x$ and $\tilde{x}$.
[*Proof. *]{}For another point $\tilde{y} \in \widetilde{V}$, where $y {\mathrel{\mathop:}=}\rho(\tilde{y})$, and any vertex $\tilde{v}$ in the cover, we can extend an edge path starting from $x$ and representing $\tilde{v}$ to an edge path starting from $y$, by adding a fixed path from $y$ to $x$ in front. This operation does not influence the notion of convergence in $\mathcal{T}$, hence this topology is independent of the choice of $x$ and $\tilde{x}$.
The universal cover $\widetilde{\Gamma}$ of $\Gamma$ admits a compact metrizable topology with connected panels, invariant under a flag-transitive automorphism group.
[*Proof. *]{}The topology $\mathcal{T}$ is second countable by Remark \[secondcountable\] and hence Hausdorff by Lemma \[lem:onelimit\], is separable and sequentially compact, hence this topology is compact and metrizable. As the panels of $\Gamma$ are connected, the panels of $\Gamma'$ which are homeomorphic are also connected.
By the results of Section \[sec:pos\] we know that a compact flag-transitive group $G$ of continuous automorphisms acts on $\Gamma$. This group lifts to a group $\widetilde{G}$ acting flag-transitively on the universal cover $\widetilde{\Gamma}$, where the kernel of the map $\widetilde{G} \to G$ is the group of deck transformations of the universal cover. As the topology $\mathcal{T}$ is independent of the choice of $x$ and $\tilde{x}$ by Lemma \[lem:independent\], the group $\widetilde{G}$ acts continuously.
We are now in the position to apply Theorem A of [@Kra-Lyt:14] yielding that the geometry $\widetilde{\Gamma}$ is either a building of type $\mathsf{C}_3$, or is one of two possible exceptional geometries.
So if $$k = {\mathbb{R}}, A= {\mathbb{H}}, B = {\mathbb{H}},$$ then $\Gamma$ cannot be covered by one of the two exceptional possibilities (as the rank two residues do not match), hence it is not simply connected and covered by a building, and if $$k = {\mathbb{R}}, A= {\mathbb{H}}, B = {\mathbb{O}}$$ or $$k = {\mathbb{C}}, A= {\mathbb{O}}, B = {\mathbb{O}},$$ so where $\Gamma$ is an exceptional geometry associated to the actions of ${\mathsf{SU}}(3) \times {\mathsf{SU}}(3)$ or ${\mathsf{SO}}(3) \times \mathsf{G}_2$ on the Cayley plane, then $\widetilde\Gamma$ would be homeomorphic to $\Gamma$ (as it is shown in [@Kra-Lyt:14] that these cannot be covered by a building, and the fact that the rank two residues of the two cases are different). As the universal cover is simply connected by definition, we conclude that the geometry $\Gamma$ is simply connected and its own universal cover.
Reducing planes from edge paths {#section:red}
-------------------------------
The goal from now on is to produce a proof for Proposition \[prop:control\]. A first step is to reduce the set of edge paths one needs to consider.
If $(u,\pi,v)$ is an edge path where $u$ and $v$ are two vertices incident with a common plane $\pi$ of the geometry $\Gamma$, then this edge path is combinatorially homotopic to any edge path of the form $(u, w_1, w_2, w_3, \dots, w_{k-1}, v)$ completely contained in the residue of $\pi$ (so assuming that each of the $w_i$ is incident with $\pi$) by applying $k$ elementary combinatorial deformations.
Hence, if an edge path $\gamma$ of length $k$ does not start or end at a plane, we can find a combinatorially homotopic edge path of length at most $\floor{\frac{3k}{2}}$ containing only points and lines using at most $3\floor{\frac{k}{2}}$ elementary combinatorial deformations (as $\floor{\frac{k}{2}}$ is the maximum amount of planes in $\gamma$, and 3 is the diameter of a projective plane).
Moreover if $(u_0, u_1, \dots, u_{k-1}, u_k)$ and $(v_0 {\mathrel{\mathop:}=}u_0, v_1, \dots, v_{l-1}, v_l {\mathrel{\mathop:}=}u_k)$ are two edge paths both completely contained in the residue of some plane $\pi,$, then they are combinatorially homotopic by applying at most $k+l$ elementary combinatorial deformations.
Primitive edge paths {#section:prim}
--------------------
We call an edge path of the form $(x,L,y,M,x)$, where $L$ and $M$ are two different lines through two different points $x$ and $y$, *primitive*.
In this section we want to show that if a primitive edge path in $\Gamma$ is contractible, then it can be reduced to the trivial edge path $(x)$ by at most $K$ elementary combinatorial deformations, where $K$ is a universal constant. (Which is a special case of Proposition \[prop:control\].)
Two points $x$ and $y$ of $\Gamma$ are said to be *orthogonal* if the corresponding vector lines in $\operatorname{Pu}_k({\mathbb{A}})$ are orthogonal w.r.t. the inner product $(\cdot | \cdot)$.
Every primitive edge path $(x, L, y, M, x)$ is homotopic (by using at most 12 elementary combinatorial deformations) to some primitive edge path $(x, L', y', M', x)$, where $x$ and $y'$ are orthogonal.
[*Proof. *]{}If $x$ and $y$ are orthogonal, there is nothing to prove, so assume that this is not the case. We start by noting that the polar line of $x$ (i.e. the points orthogonal to $x$), is a line of the projective plane obtained from projectivizing $\operatorname{Pu}_k({\mathbb{A}})$.
Fix a line $N$ through $y$ which is coplanar to both $L$ and $M$. This is always possible as the residue of $y$ is a generalized quadrangle. The line $N$ intersects the polar line of $x$ in some point $y'$. Let $L'$, respectively $M'$, be the unique lines through $x$ and $y'$, in the unique plane containing $L$ and $N$, respectively $M$ and $N$. By Section \[section:red\] we know that $(x, L, y, M, x)$ and $(x, L', y', M', x)$ are combinatorially homotopic by performing twelve elementary combinatorial deformations. This proves the lemma.
Note that this lemma implies that if $j$ elementary combinatorial deformations suffice for a contractible edge path $(x, L, y, M, x)$, then $j+ 12$ elementary combinatorial deformations suffice for the homotopic edge path $(x, L', y', M', x)$.
As the automorphism group $G$ of $\Gamma$ described in Section \[sec:pos\] acts transitively on pairs of orthogonal points of $X$, we may fix an orthogonal pair of points $x,y$, represented by vector lines $\<a \>$ and $\<a'\>$ with $a$ and $a'$ of unit norm in $\operatorname{Pu}_k({\mathbb{A}})$ with $(a|a') =0$, and reduce the question to the problem whether there is a global bound for the needed number of elementary combinatorial deformations for a contractible primitive edge path $(x, L, y, M, x)$.
Let $a'' {\mathrel{\mathop:}=}a a' \in \operatorname{Pu}_k({\mathbb{A}})$, then $a''$ is of norm one and orthogonal to both $a$ and $a'$, see Proposition 11.10 of [@CPP]. We now can represent $L$ and $M$ by respectively $[a'', b]$ and $[a'',c]$ (where $b$ and $c$ are of unit norm). By the action of the $k$-automorphisms of ${\mathbb{B}}$ the orbit of a primitive edge path $(x, L, y, M, x)$ under the corresponding automorphisms of $\Gamma$ (which fix $x$ and $y$) is completely determined by the inner product $( b \vert c )$. We call this inner product the *PL-invariant* of $(x, L, y, M, x)$.
If a primitive edge path $(x, L, y, M, x)$ with PL-invariant $l$ is contractible, where $l\in k \setminus \{ \pm 1\}$ and $| l |=1$, then there exists a contractible primitive edge path $(x, L', y, M', x)$ with PL-invariant $l'$ with $| l' | <1$. Note that in order for such an element $l$ to exist it is necessary that $k = {\mathbb{C}}$.
[*Proof. *]{}If we represent $L$ by $[a'',b]$ (as before), then the condition on the PL-invariant implies that $M$ is represented by $[a'',bl]$.
Let $b'$ in $\operatorname{Pu}_k({\mathbb{B}})$ be of unit norm such that $(b| b') =0$. The line $N {\mathrel{\mathop:}=}[a, b']$ is incident with the point $y$ and coplanar with both $L$ and $M$ (for different planes), as $$( a' | a ) = ( a''| a ) = ( b | b') = ( bl | b') = 0,$$ where we make use of Lemma \[lem:coplanar\].
If we put $b'' = b b'$, then $1, b, b'$ and $b''$ form an orthogonal $k$-basis of a subalgebra of ${\mathbb{B}}$. This leads to two embeddings
$$\phi:
\left\{
\begin{array}{l}
a \mapsto b' \\
a'\mapsto b'' \\
a'' \mapsto b \\
\end{array}
\right.$$ and $$\psi:
\left\{
\begin{array}{l}
a \mapsto b' \\
a' \mapsto b''\bar{l} \\
a'' \mapsto bl \\
\end{array}
\right.$$ of ${\mathbb{A}}$ into ${\mathbb{B}}$ (similar as in the proof of Proposition \[prop:map\]). The corresponding planes contain respectively the lines $L$, $N$ and $M, N$. Set $d {\mathrel{\mathop:}=}\frac{1}{\sqrt{2}} (a'+a'')$ (note that this is of norm one and orthogonal to $a$). The lines $L' {\mathrel{\mathop:}=}[d, \phi(d)] = [d, \frac{1}{\sqrt{2}} (b + b'')]$ and $M' {\mathrel{\mathop:}=}[d, \psi(d)] = [d, \frac{1}{\sqrt{2}} ( bl + b'' \bar{l})]$ lie in a common plane with respectively $L, N$ and $M, N$, contain $x$ and a common point $y'$ of $N$, which is orthogonal to $x$. By this reasoning $(x, L, y, M, x)$ will be combinatorially homotopic with $(x, L', y', M', x)$, in particular the latter will be contractible.
By using a suitable automorphism of our geometry we can map this last primitive pl-path to a pl-path $(x, L', y', M', x)$, with PL-invariant $(\frac{1}{\sqrt{2}} (b + b'') \vert \frac{1}{\sqrt{2}} ( bl + b'' \bar{l}) ) = \frac{1}{2} (l + \bar{l} ) = \operatorname{Re} (l) < 1$. This proves the lemma.
We can distinguish between two possibilities at the moment, either there exists some contractible primitive edge path $(x, L, y, M, x)$ (where $L \neq M$) with PL-invariant different from $-1$, or every such contractible edge path has PL-invariant $-1$. (If the PL-invariant would be $1$, then we would have that $L = M$.)
In the second case the group of automorphisms fixing $x$ and $y$ acts transitively on such paths, hence there exists a global bound on the number of elementary combinatorial deformations needed to reduce a contractible primitive edge path to the trivial path in this case.
In the remainder of the section we handle the first case.
\[lem:diam\] Fix an $l \in k$. Let $\Gamma$ be the directed graph where the vertices are the norm one elements in $\operatorname{Pu}_k ({\mathbb{B}})$, and where $(b,c)$ is a directed edge if $( b |c ) = l$. If $\vert l \vert < 1$, then $\Gamma$ is connected and of finite diameter.
[*Proof. *]{}The proof will go in different steps, each one asserting that if $(b \vert c )$ equals some $l'$, then $c$ can be reached from $b$ using a finite number of directed edges. We call such an $l'$ *valid*. By the action of the automorphism group of ${\mathbb{B}}$ over $k$, the number of edges needed depends only $l'$, not on $c$.
First we set $l' = l^2$. Then we may assume without loss of generality, that $$c {\mathrel{\mathop:}=}b l^2 + b' r$$ where $b$, $b'$ and $b''$ are mutually orthogonal and of norm 1, and such that $| r | = \sqrt{1 - | l | ^2}$. Set $d {\mathrel{\mathop:}=}bl + b '' r$, then $(b,d)$ and $(d,c)$ are directed edges. We conclude that $l'$ is valid.
By repeating this argument we obtain that any power $l^n$, where $n$ is a power of two, is valid. Note that $| l^n |$ approaches zero.
We now claim that $0$ is valid. To see this we can take elements $$\begin{aligned}
b& \\
d& {\mathrel{\mathop:}=}b l^n + b' {\sqrt{1- | l^n |^2} } b' \\
c& {\mathrel{\mathop:}=}b' \frac{l^n}{\sqrt{1- | l^n |^2} } b' + b'' s \end{aligned}$$ where $b$, $b'$ and $b''$ are as before, and $s$ is such that $|c| = 1$. The latter is only possible when $ | \frac{l^n}{\sqrt{1- | l^n |^2} } | \leq 1$, which is true for a large enough $n$ (which we fix for the remainder of the proof). This proves that $0$ is valid. In particular if $( b \vert c )$ = 0, then $c$ can be reached by at most $2n$ steps from $b$ (as $(b|d) = l^n$ and $(d|c) = l^n$).
This implies that the directed graph $\Gamma$ is connected and has diameter at most $4n$, as one can find for every pair of elements a mutually orthogonal element.
We now make use of the following observation. If the primitive edge paths $(x, L, y, M, x)$ and $(x, M, y, N, x)$, both containing $x$ and $y$, are both contractible, then $(x, L, y, N, x)$ will be contractible as well. Moreover if the needed combinatorial homotopies of $(x, L, y, M, x)$ and $(x, M, y, N, x)$ both consist of at most $j$ elementary combinatorial deformations, then there exists a combinatorial homotopy from $(x, L, y, N, x)$ to the constant path $(x)$ consisting of at most $2j+2$ elementary combinatorial deformations. (There are two of them needed to obtain $(x, L, y , M, x, M,y, N, x)$, and then $2j$ deformations to reduce the subpaths.)
Combining this observation with Lemma \[lem:diam\] implies that every primitive edge path is contractible using at most $K$ elementary combinatorial deformations (where $K$ is a constant).
Our claim holds in both cases, and hence in general.
In the first case one can, at this point, directly show that $\Gamma$ is simply connected. This is done by exploiting the fact that every primitive edge path is contractible.
Homotopy control {#section:control}
----------------
Throughout this section we work with edge paths in the universal cover $\widetilde{\Gamma}$ of $\Gamma$, in particular any two edge paths with same begin and endpoint are homotopic.
We begin with some lemmas on edge paths consisting only of points and lines.
\[lemma:pinching\] Let $\gamma {\mathrel{\mathop:}=}(x, L, y, M,z)$ be some edge path, and $L'$ a line (where $L \neq L'$) through $x$ coplanar with $L$. Then there exists an edge path $\gamma' {\mathrel{\mathop:}=}(x,L',y',M',z)$ which can be obtained from $\gamma$ by 12 elementary combinatorial deformations.
[*Proof. *]{}Let $\pi$ be the plane incident with both $L$ and $L'$. There exists some line $N$ incident with $\pi$ and a plane $\xi$ incident with both $N$ and $z$ (by Lemma \[lem:c3\]). Let $y'$ be the intersection point of $L'$ and $N$ and let $M'$ be the line incident with $\xi$ and both points $y'$ and $z$.
With six elementary combinatorial deformations we get from $(x, L, y, M,z)$ to $(x, L',y',N,y, M,z)$, and with another six to $(y, L', y', M',z)$.
\[lem:1and2\] Let $(x, L, y, M,z)$ and $(x, N, z)$ be two edge paths starting and ending at the same point, one of length four, the other of length two. Then we can obtain one path out of the other with at most $K+30$ elementary combinatorial deformations.
[*Proof. *]{}By applying Lemma \[lemma:pinching\] at most twice, we may assume that $L$ and $N$ are identical (at the cost of at most 24 elementary combinatorial transformations). As the path $(x, N, z)$ is combinatorially homotopic with $(x, N, y, N, z)$ (using six elementary combinatorial deformations), we are reduced with the question of transforming the subpath $(y,M,z)$ to the subpath $(y,N,z)$, which takes at most $K$ elementary combinatorial homotopies.
\[lem:2and2\] Let $ (x, L, y, M,z)$ and $(x, L', y', M', z)$ be two edge paths starting and ending at the same point, both of length four. Then we can obtain one path out of the other with at most $K+55$ elementary combinatorial deformations.
[*Proof. *]{}As before we may assume by Lemma \[lemma:pinching\] that $L$ and $L'$ are identical (at the cost of at most 24 elementary combinatorial deformations). One additional elementary combinatorial deformation transforms the path $(x,L,y,M,z)$ into $(x,L,y', L, y,M,z)$, which reduces the problem to the subpaths $ (y', L, y,M,z)$ and $(y', M', z)$, which is covered by Lemma \[lem:1and2\].
\[lem:3\] If $\gamma {\mathrel{\mathop:}=}(x_0, L_1, x_1, \dots , x_l)$ is an edge path (containing only of points and lines) of length $2l$, with $l \geq 3$ then there is an edge path $\gamma'$ of length $2l-2$ which one can be obtained from $\gamma$ by applying at most $K+56$ elementary combinatorial deformations.
[*Proof. *]{}Let $y$ be a point on $L_3$ collinear to $x_0$ (which exists by Lemma \[lem:c3\]), and $M$ a line incident with both $y$ and $x_0$. The subpath $(x_0, L_1, x_1, L_2, x_2)$ is now combinatorially homotopic to the path $(x_0, M, y, L_3, x_2)$, using at most $K+55$ elementary combinatorial deformations, by Lemma \[lem:2and2\].
So the edge path $(x_0, L_1, x_1, L_2, x_2, L_3, x_3, \dots ,x_l)$ is combinatorially homotopic to $(x_0, M, y, L_3, x_2, L_3, x_3, \dots ,x_l)$, which in turn is combinatorially homotopic to the edge path $(x_0, M, y, L_3, L_3, x_3, \dots ,x_l)$ of length $l-1$.
The following lemma yields a more general homotopy control.
\[lem:contr1\] For every natural number $k$ there is a number $C(k)$ with the following property. Any two edge paths $\gamma$ and $\gamma'$ starting at the same point and ending at the same point, and containing only points and lines, in the universal cover $\widetilde{X}$ of length at most $k$ are homotopic by a homotopy consisting of at most $C(k)$ elementary combinatorial deformations.
[*Proof. *]{}Note that the length of such paths is necessarily even.
For $k = 0$, $C(0) = 0$ trivially suffices. For $k=2$ we need $C(2) = K$, and for $k=4$ we can set $C(4) = K+55$ by Lemmas \[lem:1and2\] and \[lem:2and2\].
For $k \geq 6$, we can reduce edge paths of length at most $k$ to paths of length at most four by applying at most $\frac{(k-4)(K+56)}{2}$ elementary homotopies by repeated application of Lemma \[lem:3\], which leaves us in the previously handled case of $k=4$. It hence suffices to set $C(k)$ to be $(k-4)(K+56) + K+55$ in this case.
Finally, the next proposition removes the conditions concerning types, arriving at our claim in Proposition \[prop:control\].
For every natural number $k$ there is a number $D(k)$ with the following property. Any two edge paths $\gamma$ and $\gamma'$ with the same extremities are homotopic by a homotopy consisting of at most $D(k)$ elementary combinatorial deformations.
[*Proof. *]{}Let $u$ and $v$ be respectively the begin and end vertex of the edge paths $\gamma$ and $\gamma'$. If $u$ is a line or a plane, then we can apply an elementary combinatorial deformation to the beginning subpath $(u)$ of both $\gamma$ and $\gamma'$ to the subpath $(u,x,u)$ where $x$ is a point incident with $u$. If we apply a similar operation to the end vertex $v$, we can by omitting the extremities reduce the question to two edge paths $\gamma$ and $\gamma'$ of length at $k+2$ starting and ending at a point (at the cost of at most four elementary combinatorial deformations).
By applying Section \[section:red\] we may assume that $\gamma$ and $\gamma'$ are edge paths of length at most $\floor{\frac{3k+6}{2}}$ consist of only points and lines (using at most $3\floor{\frac{k+2}{2}}$ elementary combinatorial deformations).
At this point we can apply Lemma \[lem:contr1\] and conclude that $D(k) = C(\floor{\frac{3k}{2}}) + 4 + 6\floor{\frac{k+2}{2}}$ is sufficient.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
This paper is the first step in a systematic project to study examples of Kähler manifolds with positive holomorphic sectional curvature ($H > 0$). Previously Hitchin proved that any compact Kähler surface with $H>0$ must be rational and he constructed such examples on Hirzebruch surfaces $M_{2, k}=\mathbb{P}(H^{k}\oplus
1_{\mathbb{CP}^1})$. We generalize Hitchin’s construction and prove that any Hirzebruch manifold $M_{n, k}=\mathbb{P}(H^{k}\oplus
1_{\mathbb{CP}^{n-1}})$ admits a Kähler metric of $H>0$ in each of its Kähler classes. We demonstrate that the pinching behaviors of holomorphic sectional curvatures of new examples differ from those of Hitchin’s which were studied in the recent work of Alvarez-Chaturvedi-Heier. Some connections to recent works on the Kähler-Ricci flow on Hirzebruch manifolds are also discussed.
It seems interesting to study the space of all Kähler metrics of $H>0$ on a given Kähler manifold. We give higher dimensional examples such that some Kähler classes admit Kähler metrics with $H>0$ and some do not.
address:
- 'Bo Yang. Department of Mathematics, 310 Malott Hall, Cornell University, Ithaca, NY 14853-4201, USA.'
- 'Fangyang Zheng. Department of Mathematics, The Ohio State University, 231 West 18th Avenue, Columbus, OH 43210, USA and Department of Mathematics, Zhejiang Normal University, Jinhua, 321004, Zhejiang, China'
author:
- Bo Yang
- Fangyang Zheng
title: Hirzebruch manifolds and positive holomorphic sectional curvature
---
[^1]
[^2]
Introduction
============
Let $(M, J, g)$ be a Kähler manifold, then one can define the holomorphic sectional curvature of any $J$-invariant real $2$-plane $\pi=\operatorname{Span}\{X, JX\}$ by $$H(\pi)=\frac{R(X, JX, JX, X)}{||X||^4}.$$ It is the Riemannian sectional curvature restricted on any $J$-invariant real $2$-plane (p165 [@KN]). Compact Kähler manifolds with positive holomorphic sectional ($H>0$) form an interesting class of Kähler manifolds. For example, these manifolds are simply-connected, by the work of Tsukamoto [@Tsukamoto]. An averaging argument due to Berger [@Berger1966] showed that $H>0$ implies positive scalar curvature, which further leads to the vanishing of its pluri-canonical ring by a Bochner-Kodaira type identity, see [@KW]. In 1975 Hitchin [@Hitchin] proved that any Hirzebruch surface $M_{2, k}=\mathbb{P}(H^{k}\oplus 1_{\mathbb{CP}^1})$ admits a Kähler metrics with $H>0$. Moreover, he proved that any rational surface admits a Kähler metric with positive scalar curvature.
Recall that one can define another positive curvature condition on Kähler manifolds, the so-called (holomorphic) bisectional curvature, which is stronger than $H>0$. Any compact Kähler manifold with positive holomorphic bisectional curvature is biholormophic to $\mathbb{CP}^n$ by the work of Mori [@Mori] and Siu-Yau [@SiuYau]. Motivated by these results and Hitchin’s example, Yau [@Yau1] asked if the positivity of holomorphic sectional curvature can be used to characterize the rationality of algebraic manifolds. More precisely, the following question was asked.
\[rationality\] Consider a compact Kähler manifold with positive holomorphic sectional curvature, is it unirational? Is it projective? If a projective manifold is obtained by blowing up a compact manifold with positive holomorphic sectional curvature along a subvariety, does it still carry a metric with positive holomorphic sectional curvature? In general, can we find a geometric criterion to distinguish the concept of unirationality and rationality?
There are some recent progress on Question \[rationality\]. For example, an important citerion on non-uniruledness of projective manifolds in terms of pseudoeffective canonical line bundles was established by Boucksom-Demailly-P[ă]{}un-Peternell [@BDPP]. Heier-Wong [@HeierWong2012] gave a nice application of this result to show that any projective manifold with a Kähler metric with positive total scalar curvature is uniruled, i.e., having a rational curve passing through every point. Very recently, they [@HeierWong2015] proved that any projective manifold with a Kähler metric of $H>0$ is rationally connected, i.e., any two points can be connected by a rational curve. The latter paper also contains analogous results for $H \geq 0$ and nonnegative Ricci curvatures. There are also some recent studies on Hermitian manifolds with $H \geq 0$ ([@YangX]).
In this paper, we focus on examples of Kähler metrics with $H>0$. More specifically, we want to carry out a detailed study of such metrics on any Hirzebruch manifold $M_{n, k}=\mathbb{P}(H^{k}\oplus
1_{\mathbb{CP}^{n-1}})$.
One of our motivation is the work of Chen-Tian ([@ChenTian1] and [@ChenTian2]), where they studied Kähler-Ricci flow with positive bisectional curvature and proved that the space of all Kähler metrics with positive bisectional curvature on $\mathbb{CP}^n$ is path-connected. Similarly, we would like to understand the space of all Kähler metrics with $H>0$ on any Hirzebruch manifold $M_{n, k}=\mathbb{P}(H^{k}\oplus
1_{\mathbb{CP}^{n-1}})$.
\[space of H\] Given any Hirzebruch manifold $M_{n, k}=\mathbb{P}(H^{k}\oplus
1_{\mathbb{CP}^{n-1}})$, what can we say about the space of all Kähler metrics with $H>0$? Is it path-connected?
As a first step to answer Question \[space of H\], we prove the following result.
\[main in intro\] Given any Hirzebruch manifold $M_{n,k}$, there exists a Kähler metric of $H>0$ in each of its Kähler classes.
Let us explain the background of Theorem \[main in intro\] and its relation to Hitchin’s examples in [@Hitchin]. Recall any Hirzebruch surface $M_{2, k}=\mathbb{P}(H^{k}\oplus
1_{\mathbb{CP}^{1}})$ can be obtained from $\mathbb{CP}^2$ by blowing up and down, i.e. is a rational surface. It is the projective bundle associated to the rank-$2$ vector bundle $H^{k}\oplus 1_{\mathbb{CP}^{1}}$, $H$ being the hyperplane bundle over $\mathbb{CP}^{1}$ and $1_{\mathbb{CP}^{1}}$ the trivial bundle. We call a surface minimal if it has no rational curve of self-intersection $-1$, then the only minimal rational surface are $\mathbb{CP}^2$, $\mathbb{CP}^1 \times \mathbb{CP}^1$, and $M_{2,
k}$ with $k>1$. In other words, any rational surface is the blow up of $\mathbb{CP}^2$, $\mathbb{CP}^1 \times \mathbb{CP}^1$, and $M_{2,
k}$ with $k>1$. We refer the reader to Chapter 4 of Griffiths-Harris [@GH] for a detailed description of rational surfaces.
Now we focus on the Kähler metrics on Hirzebruch manifolds $M_{n,
k}$ where $\pi: M_{n,k} \rightarrow \mathbb{CP}^{n-1}$ for $n \geq
2$. Note that $M_{n,k}$ can also be described as $\mathbb{P}(H^{-k}\oplus 1_{\mathbb{CP}^{n-1}})$, the projective bundle associated to $H^{-k}\oplus 1_{\mathbb{CP}^{1}}$. Let $E_0$ denote the divisor in $M_{n,k}$ corresponding to the section $(0,1)$ of $H^{-k}\oplus 1_{\mathbb{CP}^{n-1}}$, $E_{\infty}$ the divisor in $M_{n,k}$ corresponding to the section $(0,1)$ of $H^{k}\oplus
1_{\mathbb{CP}^{n-1}}$, and $F$ the divisor corresponding to the pull-back line bundle $\pi^{\ast} H$ over $M_{n.k}$ [^3]. Then the Picard group of $M_{n,k}$ is generated by the divisors $E_0$ and $F$, while $E_{\infty}=E_0+kF$. The integral cohomology ring of $M_{n,k}$ is $${\mathbb Z}[F,E_0] / \langle F^n, E_0^2+kE_0F \rangle.$$ The anti-canonical class of $M_{n,k}$ can be expressed as $$K^{-1}_{M_{n,k}}=2 E_{\infty}-(k-n)F=\frac{n+k}{k}
E_{\infty}-\frac{n-k}{k} E_{0},$$ and every class $\alpha$ in the Kähler cone of $M_{n,k}$ can be expressed as $$\alpha=\frac{b}{k} [E_{\infty}]-\frac{a}{k} [E_0]. \label{kahler
cone}$$ for any $b>a>0$.
In [@Hitchin] Hitchin proved that each $M_{2,k}$ admits Kähler metrics with $H>0$. His example was motivated by a natural choice of Kähler metric on any projective vector bundles over Kähler manifold. Namely let $\pi: (E, h) \rightarrow (M,g)$ be any Hermitian vector bundle over a compact Kähler manifold. The Chern curvature form $\Theta (\mathcal{O}_{{\mathbb P}(E)} (1))$ of $\mathcal{O}_{{\mathbb P}(E) (1)}$ over ${\mathbb P}(E)$ has the fiber direction components given by the Fubini-Study form, hence is positive. Therefore $$\tilde{\omega}=\pi^{\ast} \omega_g+ s \sqrt{-1} \Theta
(\mathcal{O}_{{\mathbb P}(E)} (1)) \label{Hitchin intro}$$ is a well-defined Kähler metric on $P(E)$ when $s>0$ is sufficiently small. Fix a Hirzebruch surface $M_{2, k}$, one picks $(E, h)=(H^{k} \oplus 1_{\mathbb{CP}^{1}}, h)$ and $(M, g)$ as $(\mathbb{CP}^1, g_{FS})$ where $g_{FS}$ is the standard Fubini-Study metric and $h$ the induced metric. Under this situation, $\tilde{\omega}$ has a natural $U(2)$ isometric action. Hitchin showed that $\tilde{\omega}$ has $H>0$ if $0<s(1+ks)^2
<\frac{1}{k(2k-1)}$ by an explicit calculation. It was further observed by Alvarez-Chaturvedi-Heier [@ACH] that Hitchin’s examples satisfy $H>0$ if and only if $s<\frac{1}{k^2}$. We note that Hitchin’s examples can only exist on a proper open set of Kähler cone of $M_{2,k}$, for example, on $M_{2,1}$, by a scaling his examples lie in Kähler class $b[E_{\infty}]-aE_0$ where $a<b<2a$. Now Theorem \[main in intro\] implies the existence of Kähler metric of $H>0$ in each of Kähler class on any $M_{n, k}$ where $n \geq 2$.
To prove Theorem \[main in intro\], we follow Calabi’s ansatz ([@Calabi1] and [@Calabi2]). The crucial observation (pointed out in [@Calabi2]) that the group of holomorphic transformations of $M_{n,k}$ contains $U(n)/Z_k$ as its maximal compact group. Therefore it is natural to study Kähler metrics with $U(n)$-symmetry. Calabi’s method has been very fruitful to produce examples of special Kähler metrics in various settings, including Kähler-Einstein metrics, Kähler-Ricci solitons, Kähler metrics with constant scalar curvature, extremal Kähler metrics, etc., see for example [@KS1986], [@Lebrun], [@Simanca], [@Koiso], [@Apostolov], [@Cao], [@F-I-K], [@HwangSinger]. To be more precise, we follow the slightly general approach due to Koiso-Sakane [@KS1986]. It turns out that for $M_{n,k}$ this method is equivalent to Calabi’s ansatz.
Let us view $M_{n,k}$ as a compactification of the $\mathbb{C}^{\ast}$ bundle over $\mathbb{CP}^{n-1}$ obtained by $H^{-k}\setminus E_0$ where $E_0$ is its zero section. In general, given any $\mathbb{C}^{\ast}$ bundle $\pi: L^{\ast} \rightarrow M$ obtained by a Hermitian line bundle $(L, h)$, we may consider the following metric on the total space of $L^{\ast}$: $$\tilde{g}=\pi^{\ast}g_{t}+dt^{2}+(dt\circ \tilde{J})^{2}, \label{def
of g intro}$$ where $g_t$ is a continuous family of Kähler metrics on the base $(M, J)$, $t$ is a function which only depends on the norm of Hermitian metric $h$, and $\tilde{J}$ the complex structure on the total space of $L$. Koiso-Sakane [@KS1986] gave a sufficient condition to enure the resulting metric $\tilde{g}$ is Kähler and studied the compactification of such metrics. They were able to give new examples of non-homogeneous Kähler-Einstein metrics.
Now let us focus on $M_{n,k}$, where we will pick $L=H^{-k}$ in the approach of Koiso-Sakane. After a suitable reparametrization and some careful analysis, one can show that $\tilde{g}$ in (\[def of g intro\]) can be compactified to produce Kähler metrics on $M_{n,k}$ if a generating function $\phi(U)$ of a single variable $U$ satisfying suitable boundary conditions. For simplicity, let use still call such a metric $(M_{n,k}, \tilde{g})$. Making full use of the $U(n)$-isometric action, it can be shown that the curvature tensors of $(M_{n,k}, \tilde{g})$ are completely determined by three components, namely
\(1) $A$ which is the holomorphic sectional curvature along the fiber direction $F$,
\(2) $B$ which is the bisectional curvature along the fiber and any direction in the base $E_{0}$,
\(3) $C$ which is the holomorphic sectional curvature along $E_0$.
Thus we can reduce the problem of constructing Kähler metric with $H>0$ in the form of $\tilde{g}$ by looking for a suitable generating function $\phi(U)$ satisfying some differential inequalities related to $A, B, C$ defined above. In particular, we show that Hitchin’s example is canonical among all Kähler metrics with $H>0$ on $M_{n,k}$ in the following sense:
Hitchin’s examples can be uniquely characterized as $U(n)$-invariant Kähler metrics on $M_{n,k}$ which have the constant curvature component $A$.
In the level of the generating function $\phi(U)$ in the setting of Koiso-Sakane, each of Hitchin’s example corresponds to some quadratic even function defined on $[-c, c]$ with $0<c<\frac{n}{k(2k+1)}$. Indeed, the boundary conditions of the generating function $\phi(U)$ where $U \in [-c, c]$ reflects the Kähler class of the resulting metric $\tilde{g}$, the volume of the zero section $E_0$ is $(1-\frac{k}{n} c) V_{FS}$, where $V_{FS}$ denote the volume of $\mathbb{CP}^{n-1}$ endowed with $Ric(g_{FS})=g_{FS}$, and the volume of the infinity section $E_{\infty}$ is $(1+\frac{k}{n} c) V_{FS}$. To produce examples in each of the Kähler class of $M_{n,k}$ is equivalent to produce examples of $\phi(U)$ for any $c \in (0, \frac{n}{k})$ and yet satisfying inequalities related to $H>0$. We are able to construct such $\phi(U)$ by establishing some delicate estimates on even polynomials with large degrees.
Recently, Alvarez-Chaturvedi-Heier [@ACH] studied the pinching constants of Hitchin’s examples on $M_{2,k}$.
\[Alvarez-Chaturvedi-Heier [@ACH]\] \[ACH intro\] The local and global pinching constant of holomorphic sectional curvature are the same for any of the Hitchin’s examples on $M_{2,k}$. The maximum among them is $\frac{1}{(2k+1)^2}$ and the ray of the corresponding Kähler classes is $b[E_{\infty}]-aE_0$ of the slope $\frac{b}{a}=\frac{2k+2}{2k+1}$.
Recall that the local holomorphic pinching constant is the maximum of all $\lambda \in (0,1]$ such that $0<\lambda H(\pi^{,}) \leq
H(\pi) $ for any $J-$invariant real $2-$planes $\pi, \pi^{,} \subset
T_p(M)$ at any $p \in M$, while the global holomorphic pinching constant is the maximum of all $\lambda \in (0,1]$ such that there exists a positive constant $C$ so that $\lambda C \leq H(p,\pi) \leq
C$ holds for any $p \in M$ and any $J$-invariant real 2-plane $\pi
\subset T_p(M)$. Obviously the global holomorphic pinching constant is no larger than the local one. We show that the conclusion of Theorem \[ACH intro\] is not always true for other Kähler metrics with $U(n)$-symmetry and with $H>0$, which again reflects the specialness of Hitchin’s examples.
There exist Kähler metrics with $H>0$ on $M_{n,k}$ whose local and global pinching constants for holomorphic sectional curvature are not equal.
In general, the local holomorphic pinching constant of any $U(n)$-invariant Kähler metric on $M_{n,k}$ is bounded from above by $\frac{1}{k^2}$.
A direct calculation enables us to generalize the result of Alvarez-Chaturvedi-Heier [@ACH] to $M_{n, k}$. It is interesting to see that the optimal pinching constant is dimension free. It is the same constant $\frac{1}{(2k+1)^2}$ discovered in [@ACH], with the corresponding Kähler class on $M_{n,k}$ satisfies $$b[E_{\infty}]-a[E_0] \ \ \text{where}\ \ b=\frac{2k+2}{2k+1}a>0.$$ It is unclear to us how to determine the optimal holomorphic pinching constant among all $U(n)$-invariant Kähler metrics on $M_{n,k}$, though we believe it is achieved among Hitchin’s examples. Motivated by the result of Alvarez-Chaturvedi-Heier [@ACH], we would like to propose the following question:
\[pinching intro\] Is the following statement true? If a compact Kähler surface with $H>0$ has its local pinching constant $\lambda > \frac{1}{9}$, then it must be biholomorphic to $\mathbb{CP}^{2}$ or $\mathbb{CP}^{1}
\times \mathbb{CP}^{1}$.
As a partial evidence on Question \[pinching intro\], by using some previous results on positive orthogonal bisectional curvature, we give a complete classification of compact Kähler manifolds with local holomorphic pinching constant $\lambda \geq \frac{1}{2}$. In the case of Kähler surfaces, they are biholomorphic to $\mathbb{P}^{2}$ or $\mathbb{P}^{1} \times \mathbb{P}^{1}$.
It is of course desirable to make further studies on Question \[space of H\]. We are able to show that the proof of Theorem \[main in intro\] can be used to establish the path-connectedness of all $U(n)$-invariant Kähler metrics of $H>0$ on any Hirzebruch manifold $M_{n,k}$.
\[U(n) path connected\] The space of all $U(n)$-invariant Kähler metrics of $H>0$ on $M_{n,k}$ is path-connected.
At this moment, it is not clear to use how to prove the path-connectedness without the assumption of $U(n)$-symmetry. In the meantime it seems impossible to make use of the path constructed in Corollary \[U(n) path connected\] and improve the holomorphic pinching constants. In the other direction, as illustrated in the work of Chen-Tian [@ChenTian1] and [@ChenTian2], one may wonder if the Kähler-Ricci flow can be used to study the space of all Kähler metric of $H>0$ and Question \[pinching intro\]. To that end, we calculate the holomorphic sectional curvature for the Kähler-Ricci shrinking soliton on Fano Hirzebuch manifold $M_{n,k}$ with $n>k$ due to Koiso [@Koiso] and Cao [@Cao], and also for the noncompact ones on the total space of $H^{-k}
\rightarrow \mathbb{CP}^{n-1}$ with $n>k$ due to Feldman-Ilmanen-Knopf [@F-I-K].
The Cao-Koiso shrinking soliton on Hirzebruch manifold $M_{n,k}$ have $H>0$ as the ratio $\frac{n}{k}$ is sufficiently large. If we fix $k=1$, the first example with $H>0$ is on $M_{3,1}$; if we fix $k=2$, the first one is on $M_{7,2}$.
In the complete noncompact case, the Feldman-Ilmanen-Knopf shrinking solitons do not have $H>0$ if $k<n \leq k^2+2k$.
In fact we expect that none of Feldman-Ilmanen-Knopf shrinking solitons will have $H>0$. It would be interesting to know if any complete Kähler-Ricci soliton with $H>0$ must be compact, in view of the recent work Munteanu-Wang [@MW] and the previous work of Ni [@Ni].
As a corollary to the previous works of Zhu [@Zhu], Weinkove-Song [@SW2011], Fong [@Fong], Guo-Song [@GuoSong] on Kähler-Ricci flow on Hirzebruch manifolds with Calabi’s symmetry, we exhibit various pinching behaviors along the Kähler-Ricci flow when the initial metrics are chosen from examples constructed in Theorem \[main in intro\], in particular we have the following:
$H>0$ is not preserved under the Kähler-Ricci flow.
The above corollary entails the following question: Can we construct a suitable one-parameter family of deformation of Kähler metrics $M_{n,k}$ so that the holomorphic pinching constant is monotone along the deformation?
We would like to point out another generalization of Hitchin’s exmaples. In a very recent work of Alvarez, Heier, and the second-named author [@AHZ], it was proved that the projectivization ${\mathbb P}(E)$ of any Hermitian vector bundle $E$ over a compact Kähler manifold with $H>0$ also admits a Kähler metric of $H>0$. The resulting metric on ${\mathbb P}(E)$ is of the form (\[Hitchin intro\]) for $s$ sufficiently small. Instead of working with the line bundle $H^{-k}$ on $\mathbb{CP}^{n-1}$, it is possible to apply the method of Koiso-Sakane developed in the proof of Theorem \[main in intro\] to get more examples of Kähler metrics of $H>0$ on some $\mathbb{CP}^{k}$ bundles. For example, consider $M=\mathbb{CP}^{n_1-1} \times \mathbb{CP}^{n_2}$ and $L=\pi_1^{\ast} H_1^{-1} \otimes \pi_1^{\ast} H_2^{-k_2}$ where $H_1$ and $H_2$ are hyperplanes bundles on $\mathbb{CP}^{n_1-1}$ and $\mathbb{CP}^{n_2}$, $\pi_1$ and $\pi_2$ are projections to its factors. Then we can produce a $\mathbb{CP}^{n_1}$-bundle over $\mathbb{CP}^{n_2}$ as a suitable compactification of $L^{\ast}
\rightarrow M$. It is also interesting to study the space of all Kähler metrics of $H>0$ on it.
In complex dimensions higher than one, it is highly desirable to find examples of compact Kähler manifold with $H>0$. Among the known such examples are all compact Hermitian symmetric spaces and some Kähler $C$-spaces (rational homogeneous space). In the last section of the paper, we study the holomorphic pinching constant for the canonical Kähler-Einstein metric on the flag 3-fold, the only Kähler $C$-space which is not Hermitian symmetric in dimension 3. We also demonstrate a higher dimensional projective manifold such that some of its classes admit Kähler metric with $H>0$ while some do not. More precisely, we prove:
\[some classes not\] Let $M$ be the hypersurface in $\mathbb{CP}^n \times \mathbb{CP}^n$ defined by $ \sum_{i=1}^{n+1} z_i w_i = 0$ equipped with the restriction of the product of the Fubini-Study metric, where $z$, $w$ are the homogeneous coordinates. Then the holomorphic pinching constant of $M$ is $\frac{1}{4}$.
Consider $N$ which is a smooth bidegree $(p,1)$ hypersurface in $\mathbb{CP}^r \times \mathbb{CP}^s$ where $r, s \geq 2$, $p \geq
1$, and $p > r+1$, then some Kähler classes of $N$ admit Kähler metrics of $H>0$ and some do not.
Similarly as in Question \[pinching intro\], we may ask
\[pinching 3fold\] Is it true that if a compact Kähler $3$-fold with $H>0$ has its local holomorphic pinching constant $\lambda > \frac{1}{4}$, then it must be biholomorphic to a compact Hermitian symmetric space?
In a sequel of this paper, we will study examples of Kähler metrics with $H>0$ on other rational surfaces, other Kähler $C$-spaces, and higher dimensional projective manifolds.
This paper is organized as follows: In Section 2, we prove the classification theorem of compact Kähler manifolds with local holomorphic pinching constant $\lambda \geq \frac{1}{2}$. In Section 3, we prove the main Theorem \[main in intro\] and studies the relation between Kähler-Ricci flow and $H>0$ on $M_{n,k}$. In the last Section 4, we study the canonical Kähler-Einstein metric on the flag $3$-fold and prove Theorem \[some classes not\]. We end the paper with some discussions on $H>0$ in the higher dimension case from the submanifold point of view.
Holormophic sectional curvature: preliminary results
====================================================
Let us begin the definition of various curvatures on a Kähler manifold.
Let $(M, g, J)$ be a Kähler manifold of complex dimension $n
\geq 2$ with the Levi-Civita connection $\nabla$ and Riemannian curvature tensor $R$.
\(1) Sectional curvature for any real $2$-plane $\pi \subset T_p(M)$ is defined by $K(\pi)=\frac{R(X, Y, Y, X)}{|X|^2|Y|^2-g(X,Y)^2}$ where $\pi=\operatorname{span}\{X,Y\}$.
\(2) Holomorphic sectional curvature (H) for any $J$-invariant real $2$-plane $\pi \subset T_p(M)$ is defined by $H(\pi)=\frac{R(X, JX,
JX, X)}{|X|^4}$ where $\pi=\operatorname{span}\{X,JX\}$. For the sake of simplicity, we freely use $H(X)$, $H(X-\sqrt{-1}JX)$ or $H(\pi)$ for holomorphic sectional curvature.
\(3) (Holomorphic) bisectional curvature for any two $J$-invariant real $2$-planes $\pi, \pi^{,} \subset T_p(M)$ is defined by $B(\pi,
\pi^{,})=\frac{R(X, JX, JY, Y)}{|X|^2|Y|^2}$ where $\pi=\operatorname{span}\{X,JX\}$ and $\pi^{,}=\operatorname{span}\{Y,JY\}$.
In the study of Kähler manifolds with positive curvature, it is useful to consider some pinching condition in either a local or a global sense.
\[Local pinching and global pinching\]\[pinching def\]
Let $\lambda, \delta \in (0,1)$, we define the following pinching conditions on a Kähler manifold $(M, g)$.
\(1) $\lambda \leq H \leq 1$ in the local sense if for any $p \in M$, $0<\lambda H(\pi^{,}) \leq H(\pi) \leq \frac{1}{\lambda} H(\pi^{,})$ for any $J$-invariant real $2$-planes $\pi, \pi^{,} \subset T_p(M)$. In other words, there exists a function $\varphi(p)>0$ on $M^n$ such that $0<\lambda \varphi(p) \leq H(p, \pi) \leq \frac{1}{\lambda}
\varphi(p)$ for any $p$ and any holomorphic plane $\pi \subset
T_p(M)$.
\(2) $\delta \leq K \leq 1$ in the local sense if for any $p \in M$, $0<\delta K(\pi^{,}) \leq K(\pi) \leq \frac{1}{\delta} K(\pi^{,})$ for any two real $2$-planes $\pi, \pi^{,} \subset T_p(M)$.
\(3) $\lambda \leq H \leq 1$ in the global sense if $\lambda \leq
H(\pi) \leq 1$ for any $p \in M$ and any $J$-invariant real 2-plane $\pi \subset T_p(M)$. $\delta \leq K \leq 1$ in the global sense is defined similarly.
Kähler manifolds with $H>0$ are less understood and somewhat mysterious. For example, if one works with linear algebra aspects of curvature tensors, then $H>0$ alone does not give any helpful information on the Ricci curvature. In fact, most of the Hirzebruch surfaces in Hitchin’s examples are not Fano, thus do not admit any Kähler metric with positive Ricci curvature. Nonetheless one may study Kähler manifolds with $H>0$ pinched by a large constant. In this regard, the following results of Berger [@Berger1960] and Bishop-Goldberg [@BG1963] are very interesting.
\[Berger1960\] Let $(M^{n},g)$ be Kähler, then $0<\lambda \leq H \leq 1$ in the local sense implies $\frac{7\lambda-5}{8} \leq K \leq
\frac{4-\lambda}{3}$ in the local sense.
\[BG\] If $(M^{n},g)$ is Kähler, then $0<\lambda \leq H(p) \leq 1$ implies $$\frac{1}{4}[3(1+\cos^2\theta) \lambda-2] \leq K(X,Y) \leq
1-\frac{3}{4} \lambda \sin^2\theta$$ for any unit tangent vectors $X, Y$ at $p$ with $g(X,Y)=0$ and $g(X,JY)=\cos \theta$. In particular, $\lambda$-holomorphic pinching implies $\frac{1}{4}(3\lambda-2)$-pinching on sectional curvatures.
In the proof of the above Proposition \[Berger1960\], Berger discovered an interesting inequality.
\[Berger lemma\] Let $(M^n, g)$ be a Kähler manifold and $0<\lambda \leq H \leq 1$ in the local sense, then for any unit vector $X, Y$ with $g(X,Y)=0$ and $g(X,JY)=\cos {\theta}$, we have $$\lambda-\frac{1}{2}+\frac{\lambda}{2} \cos^2 \theta \leq R(X, JX,
JY,Y) \leq 1-\frac{\lambda}{2}+\frac{1}{2}\cos^2 \theta.
\label{Berger pinching}$$
For the convenience of the readers, we sketch Berger’s proof of Lemma \[Berger lemma\], as it will be crucial in the proof of Proposition \[half pinching\] below.
Given any unit vector $X, Y$ with $g(X,Y)=0$ and $g(X,JY)=\cos
{\theta}$, consider $$\lambda \leq \frac{1}{2}\Big[H(aX+bY)+H(aX-bY)\Big] \leq 1
\label{Bergerlemma1}$$ By the left half of inequality (\[Bergerlemma1\]), we conclude that $$(H(X)-\lambda) a^4+(R(X, JX, JY, Y)+2R(X, JY, JY, X)-\lambda) 2a^2
b^2+(H(Y)-\lambda) b^4 \geq 0 \label{Bergerlemma2}$$ holds for any real numbers $a$, $b$. Apply $H(X), H(Y) \leq 1$, it follows from (\[Bergerlemma2\]) that $$R(X, JX, JY, Y)+2R(X, JY, JY, X) \geq 2\lambda-1.
\label{Bergerlemma3}$$ Next consider $$\lambda \leq \frac{1}{2}\Big[H(aX+bJY)+H(aX-bJY)\Big] \leq 1.
\label{Bergerlemma4}$$ Since $$\frac{1}{2} \{ (a^2+b^2+2ab \cos \theta)^2 + (a^2+b^2-2ab \cos \theta)^2 \} =
(a^2+b^2)^2+4a^2b^2 \cos^2 \theta,$$ a similar argument as in (\[Bergerlemma2\]) and (\[Bergerlemma3\]) leads to $$3R(X, JX, JY, Y)-2R(X, JY, JY, X) \geq 2\lambda+2\lambda \cos^2
\theta-1. \label{Bergerlemma5}$$ By adding (\[Bergerlemma3\]) and (\[Bergerlemma5\]) we have $$R(X, JX, JY, Y)\geq \lambda+\frac{\lambda}{2} \cos^2
\theta-\frac{1}{2}. \label{Bergerlemma6}$$ The right half of inequality (\[Berger pinching\]) can be proved similarly if we work on the right halves of inequalities in both (\[Bergerlemma1\]) and (\[Bergerlemma4\]).
It is possible to get some characterization of Kähler manifolds with a large holomorphic pinching constant $\lambda$. For example, Bishop-Goldberg [@BG1963] proved that if $\frac{4}{5}<\lambda
\leq H \leq 1$ holds in the local sense on a compact Kähler manifold $(M, g)$, then $M$ has the homotopy type of $\mathbb{CP}^n$. They also proved in [@BG1965] that $\lambda>
\frac{1}{2}$ implies $b_2(M)=1$. Note that a direct calculation shows that $\mathbb{CP}^k \times \mathbb{CP}^l$ with the product of Fubini-Study metric has exactly $\frac{1}{2} \leq H \leq 1$ (see [@ACH] for a general result on holomorphic pinching of product metrics). In light of these results, it is natural to ask if $\frac{1}{2}<\lambda \leq H \leq 1$ in the local sense implies that $M^n$ is biholomorphic to $\mathbb{CP}^n$. This is indeed the case and we have the following:
\[half pinching\] Let $(M^n,g)$ be a compact Kähler manifold with $0<\lambda \leq H
\leq 1$ in the local sense, then the following holds:
\(1) If $\lambda>\frac{1}{2}$, then $M^n$ is bibolomorphic to $\mathbb{CP}^{n}$.
\(2) If $\lambda=\frac{1}{2}$, then $M^n$ is one of the following
(2a) $M^n$ is biholomorphic to $\mathbb{CP}^{n}$.
(2b) $M^n$ is holomorphically isometric to $\mathbb{CP}^k \times
\mathbb{CP}^{n-k}$ with a product of Fubini-Study metrics.
(2c) $M^n$ is holomorphically isometric to an irreducible compact Hermitian symmeric space of rank $2$ with its canonical Kähler-Einstein metric.
Let us consider $n \geq 2$, the crucial observation is that $\frac{1}{2} \leq \lambda \leq 1$ in the local sense implies that $(M^n,g)$ has nonnegative orthogonal holomorphic bisectional curvature. Namely for any two $J$-invariant planes $\pi=\operatorname{span}\{X,JX\}$ and $\pi^{,}=\operatorname{span}\{Y,JY\}$ in $T_p (M)$ which are orthogonal in the sense that $g(X, Y)=g(X, JY)=0$, then $$R(X, JX, JY, Y) \geq 0.$$ This follows from Berger’s inequality (\[Berger pinching\]).
Nonnegative and positive orthogonal bisectional curvature is well studied in [@ChenX], [@GuZhang], and [@Wilking].
If $\lambda>\frac{1}{2}$ then $(M^n, g)$ has positive orthogonal bisectional curvature, it is proved in [@ChenX], [@GuZhang], and [@Wilking] that the Kähler-Ricci flow evolves such a metric to positive bisectional curvature, which is biholomorphic to $\mathbb{CP}^n$ by [@Mori] and Siu-Yau [@SiuYau].
If $\lambda>\frac{1}{2}$ then $(M^n, g)$ has nonnegative orthogonal bisectional curvature, according to a classification result due to Gu-Zhang (Theorem in 1.3 in [@GuZhang]), combining the fact $H>0$, then the universal covering manifold $(\tilde{M}, \tilde{g})$ is holomorphically isometric to $$(\mathbb{CP}^{k_1}, g_{k_1}) \times \cdots \times
(\mathbb{CP}^{k_r}, g_{k_r}) \times (N^{l_1}, h_{l_1}) \times \cdots
(N^{k_r}, h_{l_s}) \label{product list}$$ Where each of $(N^{l_i}, h_{l_i})$ is a compact irreducible Hermitian symmetric spaces of rank $\geq 2$ with its canonical Kähler-Einstein metric, The holomorphic pinching constant of such a metric was well-studied and it is exactly the reciprocal of its rank, see for example [@Chen1977]. On the other hand, the pinching of product Kähler metrics was studied by [@ACH], the proved that the pinching constant of a produce Kähler metric $(M_1
\times M_2, g_1 \times g_2)$ is $\frac{\lambda_1
\lambda_2}{\lambda_1+\lambda_2}$ where $\lambda_1$ and $\lambda_1$ are pinching constants of $M_1$ and $M_2$ respectively. It is clearly that $\frac{\lambda_1
\lambda_2}{\lambda_1+\lambda_2}=\frac{1}{2}$ is equivalent to $\lambda_1=\lambda_2=1$.
Therefore if $\lambda=\frac{1}{2}$, The decomposition (\[product list\]) reduces to either a single $\mathbb{CP}^{n}$ or a single irreducible Hermitian symmetric space of rank $2$, or a product of $\mathbb{CP}^k \times \mathbb{CP}^{n-k}$ with product Fubini-Study metrics. Obviously this decomposition descends to the original manifold $(M^n , g)$.
A natural question following Proposition \[half pinching\] is what is the next threshold, if any, for the holomorphic pinching constants for Kähler manifolds with $H>0$. In general, the situation might be complicated. Note that the canonical Kähler-Einstein metric on a compact Hermitian symmetric space has holomorphic pinching constant determined by its rank ([@Chen1977]). The Kähler-Einstein metrics on a lot of the Kähler $C$-spaces also have $H>0$, and in general one has to work with the corresponding Lie algebra carefully to determine its holomorphic pinching constant. Nonetheless, in this paper we focus on the case of dimension $2$ and $3$, we will see in Section 3 and 4 that Hirzebruch surfaces and the flag 3-space might be the right objects to provide the next interesting threshold for the holomorphic pinching constant.
Kähler metrics with $H>0$ on Hirzebruch manifolds
=================================================
In this section we first review Hitchin’s examples on Hirzebruch surfaces $M_{2,k}$, then we prove the main Theorem \[main in intro\] and study the relation between the Kähler-Ricci flow and $H>0$.
A review of Hitchin’s construction
----------------------------------
Hitchin [@Hitchin] proved that any compact Kähler surface with positive sectional curvature is rational. Any rational surface can be obtained by blowing up points on $\mathbb{P}^{2}$, $\mathbb{P}^{1} \times \mathbb{P}^{1}$, and Hirzebruch surfaces $M_{2, k}$. The natural question is which rational surface admits Kähler metric with $H>0$. In this regard, Hitchin proved that any Hirzebruch surface $M_{2, k}$ admits a Hodge metric of $H>0$. Moreover, he proved that the blow up of any compact Kähler manifold with positive scalar curvature admits a Kähler metric with positive scalar curvature when the complex dimension $n \geq
2$. As a corrolary, he showed that any rational surface admits a Kähler metric with positive scalar curvature.
In general, given any Hermitian vector bundle $(E, h) \rightarrow
(M,g)$ where $(M,g)$ is a compact Kähler manifold, the Chern curvature form $\Theta (\mathcal{O}_{{\mathbb P}(E)} (1))$ of $\mathcal{O}_{{\mathbb P}(E) (1)}$ over ${\mathbb P}(E)$ has the fiber direction component given by the Fubini-Study form, hence is positive. Therefore $$\tilde{\omega}=\pi^{\ast} \omega_g+ s \sqrt{-1} \Theta
(\mathcal{O}_{{\mathbb P}(E)} (1)) \label{Hitchin metric}$$ is a well-defined Kähler metric on $P(E)$ when $s>0$ is sufficiently small.
Hitchin [@Hitchin] studied Kähler metrics of the from (\[Hitchin metric\]) on Hirzebruch surfaces $M_{2, k}$, Here we pick $(E, h)=(H^{k} \oplus 1_{\mathbb{CP}^{1}}, h)$ and $(M, g)$ as $(\mathbb{CP}^1, g_{FS})$ where $g_{FS}$ is the standard Fubini-Study metric and $h$ the induced metric. If we use the local parametrization $(z_1, (dz_1)^{-\frac{k}{2}}, z_2)$ and write down the metric locally $$\tilde{\omega}=\sqrt{-1} \partial {\bar{\partial}}\log (1+|z_1|^2)+
s\sqrt{-1} \partial {\bar{\partial}}\log [(1+|z_1|^2)^k+|z_2|^2]$$ In this case, since the vector bundle $H^{k} \oplus
1_{\mathbb{CP}^{1}}$ has nonnegative curvature, the component of the Chern curvature form $\Theta (\mathcal{O}_{{\mathbb P}(E)} (1))$ along the base direction is nonnegative, so $\tilde{\omega}$ is in fact a Kähler metric for all $s>0$.
Hitchin [@Hitchin] proved that $\tilde{\omega}$ has $H>0$ if $0<s(1+ks)^2 <\frac{1}{k(2k-1)}$. Alvarez-Chaturvedi-Heier [@ACH] further proved that it suffices to assume $s<\frac{1}{k^2}$ to guarantee $H>0$. Let us define the optimal local (global) holomorphic pinching constant to be the maximum value among all the pinching constants of Hitchin’s examples according to Definition \[pinching def\]. It was calculated in [@ACH] that the optimal local and global holomorphic pinching constants are the same and equal to $\frac{1}{(2k+1)^2}$, and the corresponding $s=\frac{1}{2k^2+k}$. The corresponding Kähler class is $b[E_{\infty}]-aE_0$ where $b=\frac{2k+2}{2k+1}a>0$. In particular, if $k=1$, then $s=\frac{1}{3}$, the corresponding Kähler metric $\tilde{\omega}$ is not in the anti-canonical class of $2\pi c_1 (M_{2,1})$, note that $M_{2,1}$ is the only Fano Hirzebruch surface.
Let us rephrase the question we proposed in Section 1 of this paper.
Hitchin’s examples produce a family of Kähler metrics with $H>0$ whose Kähler classes only stay in a subset of the Kähler cone. The path does not approach both sides of the essential boundary of the Kähler cone of $M_{2,k}$. Here by essential we mean that here the vertex of the cone is not counted as the boundary.
Are there Kähler metric with $H>0$ from each of the Kähler classes of $M_{2,k}$? In particular, since $c_1 (M_{2,1})>0$, it would be interesting to know there is any metric with $H>0$ from the anti-canonical class of $M_{2,1}$.
What is the best holomorphic pinching constant $\lambda_k$ among all Kähler metrics of $H>0$ on the Hirzebruch surfaces $M_{2, k}$? Note that Hitchin’s examples are of $U(2)$-symmetry, it seems reasonable to expect the optimal holomorphic pinching constant $\lambda_k$ to be realized by some Kähler metric with a large symmetry.
Let $\lambda_k$ denote the optimal holomorphic pinching constant among all Kähler metrics of $H>0$ on $M_{2, k}$. Is it true that any compact Kähler surface with pinching constant strictly greater $\lambda_1$ must be biholomorphic to $\mathbb{CP}^{2}$ or $\mathbb{CP}^{1} \times \mathbb{CP}^{1}$?
Hirzebruch manifolds by Calabi’s ansatz
---------------------------------------
Let us recall a powerful method to construct canonical metrics pioneered by Calabi (Calabi’s ansatz). Our exposition follows more closely from Koiso-Sakane [@KS1986]. As we shall see later, for Hirzebruch manifolds $M_{n, k}$, Calabi’s ansatz can be applied to produce $U(n)-$invariant Kähler metrics which include Hitchin’s examples as a special case.
### Kähler metrics on $\mathbb{C}^{\ast}$-bundles reviewed
First we review some facts on the construction of a Kähler metric on a $\mathbb{C}^{\ast}$-bundle over a compact Kähler manifold where $\mathbb{C}^{\ast}=\mathbb{C}-\{0\}$. Given a holomorphic line bundle $L \rightarrow M$ on a complex manifold $M$, where $\pi$ is the natural projection, we consider the $\mathbb{C}^{\ast}$-action on $L^{\ast}=L\setminus {L_0}$, where $L_0$ is the zero section of $L$. Denote by $H$ and $S$ the two holomorphic vector fields generated by the $\mathbb {R}^{+}$ and $\mathbb{S}^{1}$ action, respectively.
Let $\pi: (L, h) \rightarrow (M, g)$ be a Hermitian line bundle over a compact Kähler manifold $(M, g)$. Denote by $\tilde{J}$ the complex structure on $L$. Assume $t$ is a smooth function on $L$ depending only on the norm and is increasing in the norm of Hermitian metric $h$. Consider a Hermitian metric on $L^{\ast}$ of the form $$\tilde{g}=\pi^{\ast}g_{t}+dt^{2}+(dt\circ \tilde{J})^{2}, \label{def
of g}$$ where $g_t$ is a family of Riemannian metrics on $M$. Denote by $u(t)^{2}=\tilde{g}(H,H)$. It can be checked that $u$ depends only on $t$.
The following results were proved in [@KS1986].
The Hermitian metric $\tilde{g}$ defined by (\[def of g\]) is Kähler on $L^{\ast}$ if and only if each $g_{t}$ is Kähler on $M$, and $g_{t}=g_{0}-U \Theta(L)$, where $U=\int_{0}^{t} u(\tau )d\tau $, and if the range of $t$ includes $0$, then the corresponding value of $U$ is $0$.
\[assump 1\] We further assume the eigenvalues of the curvature $\Theta(L)$ with respect to $g_{0}$ are constant on M.
Let $z_{1}\cdots z_{n-1}$ be local holomorphic coordinates on $M$ and $z_{0}\cdots z_{n-1}$ be local coordinates on $L^{\ast}$ such that $\frac{\partial}{\partial z_{0}}=H-\sqrt{-1}S$.
$\tilde{g}_{0\bar 0}=2u^2,\
\tilde{g}_{\alpha \bar 0}=2u\partial_{\alpha} t,\ \tilde{g}_{\alpha
\bar{\beta}}=g_{t \alpha \bar{\beta}}+2\partial_{\alpha}
t\partial_{\bar{\beta}} t.$ Define $p=det(g_{0}^{-1}\cdot g_{t})$, then $det(\tilde{g})=2u^2 \cdot p \cdot det(g_{0})$.
\[Koiso local\] If we assume that $\partial_{\alpha} t=\partial_{\bar{\alpha}} t=0\
\ (1 \leqslant \alpha \leqslant n-1)$ on a fiber, and if a function f on $L^{\ast}$ depends only on $t$, then $\partial_{0}\partial_{\bar{0}}f=u\frac{d}{dt}(u\frac{df}{dt}),\
\partial_{\alpha}\partial_{\bar{0}}f=0,\
\partial_{\alpha}\partial_{\bar{\beta}}f=-\frac{1}{2}u\frac{df}{dt} \Theta(L)_{\alpha
\bar{\beta}}$. Moreover, the Ricci curvature of $\tilde{g}$ becomes: $\tilde{R}_{0\bar 0}=-u\cdot \frac{d}{dt}(u
\cdot\frac{d}{dt}(log(u^{2}p))),\ \tilde{R}_{\alpha \bar 0}=0, \ \
\tilde{R}_{\alpha \bar{\beta}}=R_{0\alpha
\bar{\beta}}+\frac{1}{2}u\cdot \frac{d}{dt}(log(u^{2}p))\cdot
\Theta(L)_{\alpha \beta}$.
It is convenient to introduce the new functions $\phi(U)=u^2(t)$ and $ Q(U)=p$. Recall that $\sqrt{h}$ is the norm of Hermitian metric on $L$. Since $\frac{dU}{\sqrt{\phi(U)}}=dt$ and $
\frac{dU}{\phi(U)}=\frac{d\sqrt{h}}{\sqrt{h}}$, for any given $\phi(U)$ we can solve for $t$ with respect to $\sqrt{h}$, hence recover the metric $\tilde{g}$. The following lemma characterizes any $\phi(U)$ which corresponds to a well-defined (maybe incomplete) Kähler metric in the form of (\[def of g\]) on the total space of $L^{\ast}$.
\[well defined on star\]
Given any hermitian line bundle $(L, h)$ over a compact Kähler manifold $(M,g_0)$ and Assumption \[assump 1\] holds. Fix $-\infty
< U_{min}<U_{max} \leq +\infty$ such that $g_t \doteq g_0-U
\Theta(L)$ remains positive on $(U_{min},U_{max})$.
Let $\phi(U)$ be a smooth positive function on $(U_{min},U_{max})$ with $\phi(U_{min})=\phi(U_{max})=0$. We further assume that $\int_{U_{min}}^{U} \frac{dU}{\phi(U)}=+\infty$, $\int_{U}^{U_{max}}
\frac{dU}{\phi(U)}=+\infty$ and $\int_{U_{min}}^{U}
\frac{dU}{\sqrt{\phi(U)}}$ is finite for all $U \in
(U_{min},U_{max})$.
Then we can solve for $t$ as a function of $\sqrt{h}$ which is strictly increasing on $\sqrt{h}$, and $t$ has the range $(t_{min},t_{max})$ which contains $0$. Therefore we can get a well-defined smooth Kähler metric $\tilde{g}$ in the form of (\[def of g\]) on $L^{\ast}$.
### Metric completion by compactification, Kähler metrics on ${\mathbb P}(L \oplus 1)$.
Koiso-Sakane [@KS1986] had a general discussion on when the Kähler metric $\tilde{g}$ on $L^{\ast}$ admits a compactification so that it can be extended onto ${\mathbb P}(L
\oplus 1)$. We summarize their results below.
\[Koiso-Sakane [@KS1986]\] \[basic compactification\] Let $(t_{min}, t_{max})$ be the range of function $t$ on $L^{\ast}$, and assume that $t$ extends to ${\mathbb P}(L \oplus 1)$ with the range $[t_{min}, t_{max}]$, where the subset $M_{min}$ (or $M_{max}$) defined by $t=t_{min}$ (or $t=t_{min}$) is a complex submanifold with codimension $D_{min}$ or $D_{max}$. Moreover, assume the Kähler metric $\tilde{g}$ extends to ${\mathbb P}(L
\oplus 1)$, which is also denoted by $\tilde{g}$.
Then it implies that near $U=U_{min}$ the Taylor expansion of $\phi(U)$ has the first term $2(U-U_{min})$, and near $U=U_{max}$, it has the first term $2(U_{max}-U)$. In other words, $t-t_{min}$ gives the distance from $M_{min}$ to points in ${\mathbb P}(L \oplus
1)$ and from $M_{max}$ in ${\mathbb P}(L \oplus 1)$, and $t_{max}-t$ the distance from $M_{min}$.
A standard example which satisfies the assumptions of Lemma \[basic compactification\] is Hirzebruch manifold $M_{n,k}$. Indeed We may view any $M_{n,k}$ as the compactification of the total space of $\mathbb{C}^{\ast}$-bundle induced from $k$-th power of the tautological bundle $H^{-k} \rightarrow \mathbb{CP}^{n-1}$. Here we assume the base $\mathbb{CP}^{n-1}$ is endowed with the Fubini-Study metric $Ric(g_0)=g_0$, hence $R_{i \bar{j} k \bar{l}}
(g_0)=\frac{1}{n}(g_{i \bar{j}}g_{k \bar{l}}+g_{k \bar{j}}g_{i
\bar{l}})$. Then from the previous results due to Koiso-Sakane we get $$\tilde{g}_{i\overline{j}}=(1+\frac{k}{n}U)g_{i \overline{j}},
\ \ \Theta(H^{-k})=-\frac{k}{n}g_{i\overline{j}}, \ \ t_{i
\overline{j}}=\frac{k}{2n}ug_{i\overline{j}}.$$ Here we need to pick $[U_{min}, U_{max}]$ such that $u_{min}>-\frac{k}{n}$ and $U_{max}<\infty$.
From now on, we will focus our consideration on $M_{n,k}$. But before that, let us remark that there are other types of compactification covered in Lemma \[basic compactification\]. For example, consider the tautological bundle $H^{-1} \rightarrow
\mathbb{CP}^{n-1}$, if we pick $U_{min}=-n$ and $U_{max}<\infty$, the corresponding compactification will produce a Kähler metric on $\mathbb{CP}^n$. Also it is clear that a similar discussion can be carried out on $C^{\ast}$-bundles obtained from a vector bundle other than a line bundle.
The following proposition gives the formulas of curvature tensors of $M_{n,k}$. Note that $U(n)$ acts isometrically on the base $\mathbb{CP}^{n-1}$, and it can be lifted as isometric actions on the total space $M_{n,k}$, so the local calculation along a fiber in Lemma \[Koiso local\] works on the total space $M_{n,k}$.
\[curvature tensors\] Let us assume that $\partial_{\alpha}
t=\partial_{\overline{\alpha}} t=0\ \ (1 \leqslant \alpha \leqslant
n-1)$ on a fiber. Consider the unitary frame $\{e_0, e_1, \cdots ,
e_{n-1}\}$ on $M_{n,k}$: $$e_0=\frac{1}{2\phi} \frac{\partial}{\partial z_0}, \
e_i=\frac{1}{\sqrt{(1+\frac{k}{n}U)g_{i\bar{i}}}}
\frac{\partial}{\partial z_i}\,\, (1\leq i\leq n-1),$$ where $\{e_1, \cdots, e_{n-1}\}$ is a unitary frame on $(\mathbb{CP}^{n-1},g_0)$. The only nonzero curvature components of $\tilde{g}$ on $M_{n,k}$ are: $$\begin{aligned}
A&=\tilde{R}_{0 \bar{0}0 \bar{0}}=-\frac{1}{2} \frac{d^2
\phi}{dU^2},\\
B&=\tilde{R}_{0 \bar{0}i \bar{i}}=
\frac{k^2 \phi-k(n+kU)\frac{d\phi}{dU}}{2(n+kU)^2},\\
C&=\tilde{R}_{i \bar{i}i \bar{i}}=2\tilde{R}_{i \bar{i}j
\bar{j}}=\frac{[2(n+kU)-k^2\phi]}{(n+kU)^2}.\end{aligned}$$ where $1\leq i, j\leq n-1$ and $i\neq j$.
Proposition \[curvature tensors\] leads to the following characterization of $U(n)$-invariant Kähler metrics with $H>0$ on $M_{n,k}$.
Any $U(n)$-invariant Kähler metric on $M_{n,k}$ has positive holomorphic sectional curvature if and only if $$A>0, C>0, 2B>-\sqrt{AC}.$$ In other words, it is characterized by a smooth concave function $\phi(U)$ where $-\infty<U_{min} \leq U \leq U_{max}<+\infty$ with $1+\frac{k}{n}U_{min}>0$ such that the following conditions hold:
\(1) $\phi>0$ on $(U_{min}, U_{max})$, $\phi(U_{min})=\phi(U_{max})=0$, $\phi'(U_{min})=2$, and $\phi'(U_{max})=-2$.
\(2) $$\phi(U)<\frac{2}{k^2}(n+kU), \ \
\frac{k\phi}{n+kU}-\phi'>-\sqrt{\phi^{\prime\prime}
\big[\frac{\phi}{2}-\frac{1}{k^2}(n+kU)\Big]}$$ for any $U \in [U_{min},U_{max}]$.
As a corollary of Proposition \[curvature tensors\], we also have the following rough estimates on holomorphic pinching constants of $U(n)$-invariant Kähler metrics on $M_{n,k}$.
Any $U(n)$-invariant Kähler metric on $M_{n,k}$ with $H>0$ have its local holomorphic pinching constant bounded from above by $\frac{1}{k^2}$, Fix $0<c<\frac{n}{k+2}$, if we consider any $U(n)$-invariant Kähler metric with $H>0$ whose corresponding Kähler class lies in the following ray $S$ in the Kähler cone $$S=\{\ b[E_{\infty}]-a[E_0] \ |\ \text{where}\ \
a=b\,\frac{n-kc}{n+kc} \ \}.$$ Then its holomorphic pinching constant is bounded from above by $\frac{2c}{n-kc}$.
Another consequence of Proposition \[curvature tensors\] is the path-connectedness of Kähler metrics of $H>0$ in the same Kähler class on $M_{n,k}$.
\[convexity\] If $\phi_1$ and $\phi_2$ are two generating functions of two Kähler metrics of $H>0$ in the same Kähler class on $M_{n,k}$, so is any convex combination $t\phi_1+(1-t)\phi_2$ with $0<t<1$.
### Hitchin’s examples reformulated
Hitchin’s construction gives a family of Kähler metrics with $U(2)$ symmetry on $M_{2,k}$. We observe a similar construction works for $M_{n,k}$.
Given $s>0$, $U_{min}=0$, $U_{max}=ns$, define $\phi_{s}(U)=-\frac{2}{ns} U^2+2U$. Now $$A=\frac{2}{ns},\,\,
B=\frac{\frac{k^2}{ns}U^2+\frac{4k}{s}U-kn}{(n+kU)^2},\,\,
C=\frac{\frac{2k^2}{ns}U^2+(2k-2k^2)U+2n}{(n+kU)^2}.$$
For all $s>0$, $\phi_s$ gives a family of Kähler metric on $M_{n,k}$. Now assume $n=2$, Hitchin proved that these metrics have $H>0$ if $s$ suitably small. Indeed it is observed by Alvarez-Chaturvedi-Heier [@ACH] that it suffices to assume $0<s<\frac{1}{k^2}$ to get $H>0$. They also calculated the pinching constants of these metrics and concluded that the optimal value is $\frac{1}{(2k+1)^2}$ when $s=\frac{1}{2k^2+k}$. When $k=1$, the optimal metric lies in the Kähler class $\frac{4}{3}[D_{\infty}]-[D_0]$.
In fact, we observe that Hitchin’s example is canonical in the following sense.
Hitchin’s examples can be uniquely characterized as $U(n)$-invariant Kähler metrics on $M_{n,k}$ with the constant radial curvature $A$. In particular, the following example gives the unique form of $\phi(U)$ up to a scaling and a translation of $[U_{min}, U_{max}]$.
\[other class\]
Let $c>0$, $U_{min}=-c$, $U_{max}=c$, define $\phi_c(x)=c-\frac{x^2}{c}$ on $[-c,c]$. Since $\phi^{\prime}(-c)=2$, $\phi^{\prime}(c)=-2$, and $\phi(\pm c)=0$, we have a Kähler metric on $M_{2,1}$. Now $$A=\frac{1}{c},\,\,
B=\frac{\frac{1}{c}U^2+\frac{4}{c}U+c}{2(U+2)^2},\,\,
C=\frac{\frac{1}{c}U^2+2U+4-c}{(U+2)^2}.$$
If we assume $0<c<2$, then obviously $1-\frac{1}{2}U>0$, $A>0$, and $C>0$ on $[-c,c]$. Consider $$D \doteq
2B+\sqrt{AC}=\frac{\frac{1}{c}U^2+\frac{4}{c}U+c+\frac{1}{c}
\sqrt{U^2+2cU+c(4-c)}\, \cdot (U+2)}{(U+2)^2}.$$ Then $D(-c)>0$ is equivalent to $c<\frac{2}{3}$. Moreover, one can check that the numerator of $D(U)$ is increasing on $U \in (-c,c)$, hence $D(U)>0$ for any $-c<U<c$ and $0<c<\frac{2}{3}$. Therefore $\phi_c$ provides a family of Kähler metrics of $H>0$.
Next let us find the pinching constant for $\phi_{c}(x)=c-\frac{x^2}{c}$. For any given $\phi_{c}$, the expression of the holomorphic sectional curvature is $$H(X)=A|x_1|^4+4B|x_1 x_2|^2+C |x_2|^4,$$ where $X=x_1 e_1+x_2 e_2$ with $|x_1|^2+|x_2|^2=1$.
If we set $t=|x_1|^2$, then $H(X)=(A+C-4B)t^2+t(4B-2C)+C$ with $t\in
[0,1]$. it is elementary to discuss its extremal values. In particular, we will show that for any $c \in (0,\frac{2}{3})$, the pinching constant $\inf_{U \in (-c,c)} \frac{\min H}{\max H}(U)$ is always attained at $U=-c$, i.e. along the zero section of $M_{2,1}$.
Indeed $$\min_{||v||=1} H(U,v)=\frac{AC-4B^2}{(A+C-4B)}, \max_{||v||=1} H(U,
v)=A.$$ Therefore, the local pinching constant equals $$\frac{AC-4B^2}{A(A+C-4B)}=\frac{2U^3+(6-3c)U^2-12cU-c^3-2c^2-8c}
{(2U-2-3c)(U+2)^2}.$$
It is direct to check that the above expression is increasing on $U
\in [-c,c]$. If $U=-c$, it becomes $\frac{2c(c-3c)}{(2-c)(5c+2)^2}$, When $c=\frac{2}{7} \approx 0.2857$, it attains the maximum $\frac{1}{9}$. The optimal pinching constant among the family $\phi_{c}$ agrees with the result of Alvarez-Chaturvedi-Heier [@ACH], and the corresponding optimal Kähler metrics are just multiples of those in [@ACH].
It is also straightforward to solve the optimal holomorphic pinching constant of Hitchin’s examples on any Hirzebruch manifold $M_{n,k}$. Given any $1 \leq k<n$, pick $c>0$, $U_{min}=-c$, $U_{max}=c$, define $\phi_{c}(U)=c-\frac{x^2}{c}$ on $[-c,c]$. We claim that $\phi_c$ gives a Kähler metric on $M_{n,k}$ with $H>0$ as long as $c < \frac{n}{k(2k+1)}$. Note that $$A=\frac{1}{c},\,\,
B=\frac{\frac{k^2}{c}U^2+\frac{2nk}{c}U+c k^2}{2(n+kU)^2},\,\,
C=\frac{\frac{k^2}{c}U^2+2kU+2n-c k^2}{(n+kU)^2}.$$
Similarly $$D \doteq 2B+\sqrt{AC}=\frac{\frac{k^2}{c}U^2+\frac{2nk}{c}U+ck^2+
\frac{n+kU}{c}\sqrt{k^2U+2kcU+(2n-ck^2)c}}{(n+kU)^2}$$
First note that $0<c < \frac{n}{k(2k+1)}$ is equivalent to $D(-c)>0$, then similarly we could show under this condition on $c$ the numerator of $D$ is strictly increasing on $U \in (-c, c)$, therefore $D(U)>0$ holds on $[-c,c]$.
Note that when $n\geq 4$, for any positive integer $k$ satisfying $k(2k+1) < n$, we may pick $c=1$. In this case the Kähler class of the resulting metric is proportional to the anti-canonical class.
For the above metric on $M_{n,k}$ given by $\phi_c (x) = c - \frac{x^2}{c}$ on $[-c,c]$, where $0<c<\frac{n}{k(2k+1)}$ is a constant, one can carry out a similar calculation to conclude that maximal local pinching constant achieves its maximum at $U=-c$, which is $$\frac{2c(n-c(2k^2+k))}{(n-ck)((3k+2)c+n)}.$$ It can be shown that it obtains its maximum value at $c=\frac{n}{4k^2+3k}$, and the optimal pinching constant is $\frac{1}{(2k+1)^2}$. Note that the optimal pinching constant is dimension free.
New examples and the proof of Theorem \[main in intro\]
-------------------------------------------------------
Next let us consider $M_{2,1}$ which is the only Fano Hirzebruch surface. Note that Kähler metrics of Hitchin’s examples on $M_{2,1}$ can not be proportional to the anti-cananical class. A natural question is whether there exists a Kähler metric with $H>0$ in $2\pi c_1(M_{2,1})$. Note that the corresponding $\phi(U)$ of such a metric must satisfy $(2+U_{max})=3(2+U_{min})$ besides $\phi(U_{min})=\phi(U_{max})=0$, $\phi'(U_{min})=2$, and $\phi'(U_{max})=-2$ required by the smooth compactification.
In the following we exhibit such an example with different global and local holomorphic pinching constants.
\[A new family of Kähler metrics of $H>0$ on $M_{2,1}$\] \[canonical class\]
Given any real number $0<c<\frac{6}{5}$, pick a real number $\mu \in
(\frac{1}{2}c, c)$, define $\phi_{c, \mu}: [-c, c] \rightarrow
\mathbb{R}$ by $$\phi_{c,\mu}(U)=\mu-(\frac{1}{c^3}-\frac{\mu}{c^4})
U^4-(\frac{2\mu}{c^2}-\frac{1}{c})U^2$$ Thus $\phi_{c, \mu}$ determines a family of Kähler metrics on $M_{2,1}$, and in particular when $c=1$, the Kähler class of $\phi_{c, \mu}$ is proportional to the anti-canonical class of $M_{2,1}$.
There exists some $\delta \in (0,\frac{1}{2})$ which depends on $c$ such that for any $\frac{1}{2}c < \mu <(\frac {1}{2}+\delta)c$, $\phi_{c,\mu}(U)$ defines a Kähler metric on $M_{2,1}$ with $H>0$.
We begin with the curvature tensors of $\phi_{c, \mu}(U)$: $$\begin{aligned}
&A=6(\frac{1}{c^3}-\frac{\mu}{c^4})U^2+(\frac{2\mu}{c^2}-\frac{1}{c}), \\
&B=\frac{3(\frac{1}{c^3}-\frac{\mu}{c^4})U^4+
8(\frac{1}{c^3}-\frac{\mu}{c^4})U^3+
(\frac{2\mu}{c^2}-\frac{1}{c})U^2+
4(\frac{2\mu}{c^2}-\frac{1}{c})U+\mu}{2(U+2)^2},\\
&C=\frac{(\frac{1}{c^3}-\frac{\mu}{c^4})U^4+
(\frac{2\mu}{c^2}-\frac{1}{c})U^2+2(U+2)-\mu}{(U+2)^2}.\end{aligned}$$
First note that for any $c \in (0,2)$, $\mu \in (\frac{1}{2}c,c)$, we have $A>0$ for any $U \in [-c, c]$, then since $C(-c)=\frac{2}{2-c}$ and $C(U+2)^2$ is increasing on $[-c,c]$, we also have $C>0$.
Next one can check that $$2B+\sqrt{AC}|_{U=-c}=-\frac{2}{2-c}+\frac{1}{c}\sqrt{\frac{2(5c-4\mu)}{2-c}},$$
From here, it is direct to see that given any $c \in (0,
\frac{6}{5})$, there exists some $\delta>0$ such that for any $
\frac{1}{2}c < \mu< (\frac{1}{2}+\delta)c$, $2B+\sqrt{AC}>0$ at $U=-c$.
From now on let us consider $T_{\mu}(U)=(U+2)^2(2B+\sqrt{AC})$, it suffices to show that $T_{\mu} (U)>0$ for any $U \in [-c,c]$ if $\mu$ is sufficiently close to $\frac{1}{2}c$. Note that $$T_{\mu} (U)=P_{\mu}(U)+(U+2)\sqrt{Q_{\mu}(U)}$$ where $$P_{\mu}(U)=3(\frac{1}{c^3}-\frac{\mu}{c^4})U^4
+8(\frac{1}{c^3}-\frac{\mu}{c^4})U^3
+(\frac{2\mu}{c^2}-\frac{1}{c})U^2
+4(\frac{2\mu}{c^2}-\frac{1}{c})U+\mu$$ and $$\begin{aligned}
Q_{\mu}(U)= & \ \ 6(\frac{1}{c^3}-\frac{\mu}{c^4})^2 U^6
+7(\frac{1}{c^3}-\frac{\mu}{c^4})(\frac{2\mu}{c^2}-\frac{1}{c})U^4
+12(\frac{1}{c^3}-\frac{\mu}{c^4})U^3 \\
&
+\Big[(\frac{2\mu}{c^2}-\frac{1}{c})^2+6(4-\mu)(\frac{1}{c^3}-\frac{\mu}{c^4})\Big]U^2
+2(\frac{2\mu}{c^2}-\frac{1}{c})U
+(-\frac{2}{c^2}\mu^2+\frac{8+c}{c^2} \mu -\frac{4}{c}).\end{aligned}$$
\[claim1\] $T_{\frac{c}{2}}(U) > 0$ on $[-c,c]$ and in particular it has a positive lower bound at $0$.
*Proof of Claim \[claim1\].* To see it is true, note that: $$T_{\frac{c}{2}}(U)=(\frac{3}{2c^3}U^4+\frac{4}{c^3}U^3+\frac{c}{2})
+(U+2)\sqrt{\frac{3}{2c^6}U^6+\frac{6}{c^3}U^3+\frac{3(8-c)}{2c^3}U^2}.$$ It suffices to consider the interval $[-c, 0]$, we will show that $T_{\frac{c}{2}}(U)$ is strictly concave on $[-c, 0]$, thus it attains its minimum either at $U=-c$ or at $U=0$, which are both positive. $$\dfrac{d^2 T_{\frac{c}{2}} (U)}{ d
U^2}=\frac{6}{c^3}U(3U+4)+\frac{\frac{9}{2c^{12}}U^3 \cdot
R(U)}{\Big(\sqrt
{\frac{3}{2c^6}U^6+\frac{6}{c^3}U^3+\frac{3(8-c)}{2c^3}U^2}\Big)^3},$$ where $$\begin{aligned}
R(U)=& \ 6U^8+6U^7+36 c^3 U^5+(108c^3-9c^4)
U^4\\
&+(80c^3-10c^4)U^3+30c^6 U^2+(108c^6-12c^7) U +c^8-20c^7+96c^6.\end{aligned}$$
Note that $R(-c)=4c^{6}(2-c)^2>0$, we will prove that $\frac{d R(U)}{d
U}>0$ on $[-c, 0]$, which leads to $R(U)>0$ for any $0<c<\frac{6}{5}$. Indeed, $$\begin{aligned}
\frac{1}{6} \frac{d R(U)}{d U}=& \ 8 U^7+7 U^6+30 c^3 U^4+6(12
c^3-c^4)U^3\\
&+5(8c^3-c^4)U^2+10 c^6 U+(18 c^6 - 2c^7)\\
=&I_1+I_2+I_3.\end{aligned}$$ where we have $$\begin{aligned}
I_1 =& \ 10 c^6 U + \frac{72}{5} c^6-2c^7 \geq \frac{72}{5}
c^6-12 c^7 = 12 (\frac{6}{5}-c) \geq 0, \\
I_2 =& \ 8 U^7+ 7 U^6+ \frac{13}{6} c^3 U^4 \geq U^6
(8U+7+\frac{13}{6} c) \geq 7U^6(1-\frac{5}{6} c) \geq 0,\\
I_3 =& \ \frac{167}{6} c^4 U^4+\frac{18}{5}c^6+6(12
c^3-c^4)U^3+5(8c^3-c^4)U^2,\end{aligned}$$for any $U \in [-c, 0]$ where $0<c<\frac{6}{5}$.
Next we prove $I_3>0$ on $[-c, 0]$. $$\begin{aligned}
I_3 \geq & \ U^2 \Big[\frac{167}{6} c^3 U^2+ 6(12 c^3-c^4)
U+5(8c^3-c^4) +\frac{18}{5} c^4 \Big].\end{aligned}$$ Let $S(U)$ denote the quadratic function inside the bracket: $$S(U)=\frac{167}{6} c^3 U^2+ 6(12 c^3-c^4) U+5(8c^3-c^4)
+\frac{18}{5} c^4.$$ Now it is straightforward to see that under the assumption $0<c<\frac{6}{5}$, $S(U)$ attains its minimum at $U=-c$, and $$S(-c)=c^3 (\frac{203}{6} c^2-\frac{367}{5}c+40)>0$$
Putting these together, we have proved that $\dfrac{d^2
T_{\frac{c}{2}} (U)}{ d U^2}$ is strictly negative on $[-c, 0]$, and therefore $T_{\frac{c}{2}}(U) > 0$ on $[-c,c]$.
Now let us continue with the proof of Proposition \[canonical class\], note that as $\mu \rightarrow (\frac{c}{2})^+$, $T_{\mu}(U)$ converges to $T_{\frac{c}{2}} (U)$ uniformly on $[-1, -U_0] \cup [U_0,1]$ for some fixed small number $U_0>0$, $$\begin{aligned}
\frac{\partial T_{\mu}(U)}{\partial \mu} = & \ \ \frac{\partial
P}{\partial \mu}+
\frac{(U+2)}{2\sqrt{Q}} \frac{\partial Q}{\partial \mu} \\
=& \ (-\frac{3}{c^4}U^4-\frac{8}{c^4}U^3+\frac{2}{c^2}U^2+\frac{8}{c^2}U+1)\\
&+\frac{(U+2)}{2\sqrt{Q}}
\Big\{-\frac{12}{c^4}(\frac{1}{c^3}-\frac{\mu}{c^4})U^6+
7\Big[\frac{2}{c^2}(\frac{1}{c^3}-\frac{\mu}{c^4})-
\frac{1}{c^4}(\frac{2\mu}{c^2}-\frac{1}{c})\Big]U^4\\
&
-\frac{12}{c^4}U^3-\Big[\frac{4}{c^2}(\frac{2\mu}{c^2}-\frac{1}{c})
-6(\frac{1}{c^3}-\frac{\mu}{c^4})-\frac{6}{c^4}(4-\mu)\Big]U^2
+\frac{4}{c^2}U+ \frac{8+c-4\mu}{c^2} \Big\}\end{aligned}$$
Take $U_0=\min\{\frac{1}{100}, \frac{1}{100}c^2\}$, we see that $\frac{\partial T_{\mu}(U)}{\partial {\mu}}$ is strictly positive for any $|U|<U_0$ as long as $\mu-\frac{c}{2}$ is small enough. In other words, we can find some $\delta>0$ such that $T_{\mu }(U)>0$ on $[-U_0, U_0]$ for any $\frac{c}{2}<\mu<(\frac{1}{2}+\delta)c$. Moreover, $T_{\mu}(U)$ converges to $T_{\frac{c}{2}} (U)$ outside $[-U_0, U_0]$, hence we get $T_{\mu}(U)>0$ for $\frac{1}{2}c < \mu
<(\frac {1}{2}+\delta)c$.
Let us now focus on the anti-canonical examples constructed in the previous proposition, namely, with $c=1$. We expect that Proposition \[canonical class\] is still true for any $\frac{1}{2} < \mu < \frac{3}{4}$. Numerical tests suggest that $2B+\sqrt{AC}$ is indeed positive on $U
\in [-1,1]$ for any $\frac{1}{2} < \mu < \frac{3}{4}$. However, it seems rather tedious to prove it rigorously, as $T_{\mu}(U)$ is not always increasing on $[-1,1]$. We need some better estimates on the critical points of $T_{\mu}(U)$ which lie in $[-1,0)$.
The following table shows that if $\mu$ is close to $\frac{1}{2}$, then the local pinching constants of Kähler metric generated by $\phi_{\mu}$ differs from the global one.
Intervals of $U$ $[-1, U_1]$ $[U_1, U_2]$ $[U_2, U_3]$ $[U_3, U_4]$ $[U_4, 1]$
-------------------------- -------------------------- -------------------------- -------------- -------------------------- -------------------------- --
$\min_{||v||=1} H(U, v)$ $\frac{AC-4B^2}{A+C-4B}$ $\frac{AC-4B^2}{A+C-4B}$ $A$ $\frac{AC-4B^2}{A+C-4B}$ $\frac{AC-4B^2}{A+C-4B}$
$\max_{||v||=1} H(U, v)$ $A$ $C$ $C$ $C$ $A$
In the above, $U_1<U_4$ are values which corresponds to $A=C$, and $U_2<U_3$ are values which corresponds to $A=2B$.
For example along the zero section $U=-1$, $\min
H=\frac{AC-B^2}{A+C-4B}=\frac{6-8\mu}{11-4\mu}$ and $\max
H=A(-1)=5-4\mu$. Therefore the pinching constant along zero section is $\frac{6-8\mu}{(5-4\mu)(11-4\mu)}$, which is close to $\frac{2}{27}$ as $\mu$ goes close to $\frac{1}{2}$. It is clear that the global maximum of holomorphic sectional curvature is attained at $U=-1$ by $A=5-4\mu$ while the global minimum is attained at $U=0$ by $A=2\mu-1$. Therefore, we conclude that the local pinching constants of Kähler metric generated by $\phi_{\mu}$ differs from the global one. Indeed, we have $$\min_{U \in [-1,1]} \frac{\min_{||v||=1} H(U, v)}{\max_{||v||=1}
H(U, v)}=\frac{4(2\mu-1)}{4-\mu}, \, \, \frac{\min_{U, ||v||=1} H(U,
v)}{\max_{U, ||v||=1} H(U, v)}=\frac{2\mu-1}{5-4\mu}.$$
Now we are ready to prove our main theorem.
\[main any class\] Given any Hirzebruch manifold $M_{n,k}=\mathbb{P}(H^{k}\oplus
1_{\mathbb{CP}^{n-1}})$, there exists a Kähler metric of $H>0$ in each of its Kähler classes.
Theorem \[main any class\] is a corollary of the following proposition.
\[any class\] Let $n \geq 2$ and $k \geq 1$ be any two integers, there exists some $p_0(n,k) \in \mathbb{N}$ and a sequence of positive real numbers $\{\epsilon_p\}_{p \geq p_0}$ with $\lim_{p \rightarrow \infty}
\epsilon_p=0$, such that for any $c \in
(0,\frac{n}{k}-2\epsilon_p]$, there exists $p_1(n ,k, c)>p_0$ with the following property:
Given any $p \geq p_1$ there exists some $\delta_1>0$ and $\delta_2>0$ such that for any $\alpha_2 \in (0,\delta_1)$ and $\mu=\frac{c}{p}+\delta_2$, $\phi(x)$ defined on $[-c, c]$ by $$\begin{aligned}
\phi(x)=\mu-\alpha_2 x^2-\alpha_{2p-2} x^{2p-2}-\alpha_{2p} x^{2p}.
\label{polynomial formula}\end{aligned}$$ where $$\alpha_{2p-2}=\frac{p\mu-c-(p-1)\alpha_2 c^2}{c^{2p-2}}, \ \ \
\alpha_{2p}=\frac{c-(p-1)\mu+(p-2)\alpha_2 c^2}{c^{2p}},$$ generates a Kähler metric with $H>0$ on $M_{n,k}$.
Let $\epsilon_p=\frac{2n}{2p+2k-1}$ and pick any $c<\frac{n}{k}-2\epsilon_{p}$, we will determine constants $\delta_1$ and $\delta_2$ step by step. A quick observation is that in order to make sure both $\alpha_{2p-2}$ and $\alpha_{2p}$ positive, we need $$\delta_2<\frac{c}{p(p-1)}, \ \text{and} \ (p-1)\delta_1 c^2< p \,
\delta_2. \label{condition 1}$$ Note that $$\begin{aligned}
A=& \ p(2p-1) \alpha_{2p} U^{2p-2}+ (p-1)(2p-3)\alpha_{2p-2} U^{2p-4}
+\alpha_2, \label{general A}\\
B=& \ \frac{(2p-1)\alpha_{2p}U^{2p}+ 2p \, nk \, \alpha_{2p} \,
U^{2p-1}+ (2p-3)k^2
\alpha_{2p-2} U^{2p-2}}{2(n+kU)^2} \nonumber\\
&+\frac{(2p-2)nk \alpha_{2p-2} U^{2p-3} +k^2 \alpha_2 U^2
+2nk\alpha_2 U+ k^2\mu}{2(n+kU)^2}, \label{general B}\\
C=& \ \frac{k^2 \alpha_{2p} U^{2p}+k^2 \alpha_{2p-2} U^{2p-2}+k^2
\alpha_2 U^2 + 2kU + 2n-k^2 \mu}{(n+kU)^2}. \label{general C}\end{aligned}$$
Obviously $A>0$ on $[-c, c]$, note that $$C>\frac{2n-k^2
\mu-2kc}{(n+kU)^2} \ \ \text{for any} \ U \in [-c,c].$$ By plugging $c \leq \frac{n}{k}-2\epsilon_p$ and $\mu=\frac{c}{p}+\delta_2$ into the above one can conclude that is a sufficient condition for $C>0$ is $$\delta_2<\frac{n(2p+2k+1)}{kp(2p+2k-1)}. \label{condition 2}$$
Note that $2B+\sqrt{AC}|_{U=-c}>0$ is equivalent to $\phi^{\prime
\prime} (-c)<-\frac{4k^2}{n-kc}$. By a direct calculation we see that it is further equivalent to $$\frac{k(2p-1)+2k^2}{c(n-kc)} \big[ \frac{n}{k}-\epsilon_p -c \big]
>\frac{2p(p-1)}{c^2} \delta_2 -\alpha_2 (2p^2-6p+4).
\label{condition 3}$$ In order that (\[condition 3\]) holds, it suffices to pick $\delta_2$ so that the following holds: $$\frac{k(2p-1)+2k^2}{(n-kc)} \big[ \frac{n}{k}-\epsilon_p -c \big]
>\frac{2p(p-1)}{c} \delta_2.
\label{condition 4}$$ In other words, for any $c \in (0,\frac{n}{k}-2\epsilon_p]$, it is easy to pick $\delta_1$ and $\delta_2$ such that all the inequalities in (\[condition 1\]), (\[condition 2\]), and (\[condition 4\]) are satisfied.
It remains to show $2B+\sqrt{AC}$ is positive on $[-c,c]$. Motivated by the proof of Proposition \[canonical class\] let us introduce $$\begin{aligned}
T(\mu, \alpha_2, U)=P(\mu, \alpha_2, U)+ (n+kU) \sqrt{Q(\mu,
\alpha_2, U)},\end{aligned}$$ where $$\begin{aligned}
P(\mu, \alpha_2, U)=2(n+kU)^2 B(U), \ \text{and}\ Q(\mu, \alpha_2,
U)=(n+kU)^2 A(U) \cdot C(U).\end{aligned}$$ where $A, B, C$ are given in (\[general A\]), (\[general B\]), and (\[general C\]).
We need to show that $T(\mu, \alpha_2, U)>0$ for any $U \in [-c,c]$ under the assumption $c \in (0,\frac{n}{k}-2\epsilon_p]$, $\alpha_2
\in (0,\delta_1)$, and $\mu=\frac{c}{p}+\delta_2$, where $\delta_1$ and $\delta_2$ satisfy (\[condition 1\]), (\[condition 2\]), and (\[condition 3\]). By a similar argument as in the proof of Proposition \[canonical class\], one checks that for $\alpha_2$, $\delta_2$ small, there exists $U_0$ sufficiently small. $$\dfrac{\partial T(\mu, \alpha_2, U)}{\partial {\mu}}>0, \
\text{and}\ \dfrac{\partial T(\mu, \alpha_2, U)}{\partial
{\alpha_2}}>0$$ for any $|U|<U_0$.
For example, note that $$\frac{\partial \alpha_{2p-2}}{\partial \mu}=\frac{p}{c^{2p-2}},
\frac{\partial \alpha_{2p}}{\partial \mu}=-\frac{p-1}{c^{2p}}$$ Therefore, $$\begin{aligned}
\frac{\partial P}{\partial \mu} = & \ -\frac{(p-1)(2p-1)k^2}{c^{2p}}
U^{2p} +(2p-3)\frac{k^2p}{c^{2p-2}} U^{2p-2} +
(2p-2) \frac{nkp}{c^{2p-2}} U^{2p-3}+k^2,\\
\frac{\partial Q}{\partial \mu} = & \ \frac{\partial A}{\partial \mu}
C+A\frac{\partial C}{\partial \mu}\\
=& \ \Big( -\frac{p(p-1)(2p-1)}{c^{2p}}
U^{2p}+\frac{p(p-1)(2p-3)}{c^{2p-2}} U^{2p-4} \Big) \cdot C\\
& \ + \Big( -\frac{k^2(p-1)}{c^{2p}} U^{2p}+\frac{k^2 p}{c^{2p-2}}
U^{2p-2} -k^2\Big) \cdot A,\\
\geq & \ \frac{p(p-1)(2p-3)}{2c^{2p-2}} U^{2p-4} (2n-k^2 \mu)
-2k^2\Big(\alpha_2+(p-1)(2p-3)\alpha_{2p-2} U^{2p-4}\Big).\end{aligned}$$ It follows that when $|U|$, $\alpha_2$, $\delta_2$ are small enough, $$\begin{aligned}
\frac{\partial T(\mu, \alpha_2, U)}{\partial \mu} \ = & \ \
\frac{\partial P}{\partial \mu}+ \frac{(n+kU)}{2\sqrt{Q}}
\frac{\partial Q}{\partial \mu} \ \geq \ \frac{k^2}{2}-\frac{4nk^2
\alpha_2}{2\sqrt{\alpha_2 \cdot n}} \ > \ 0. \label{derivative
positive}\end{aligned}$$
Therefore to prove Theorem \[any class\] it suffices to show that $T(\frac{c}{p}, 0, U)>0$ for $U \in [-c,c]$. Now we have $$\alpha_{2p-2}=0, \ \text{and} \ \alpha_{2p}=\dfrac{1}{p \,
c^{2p-1}}.$$ Therefore, $$\begin{aligned}
P(\frac{c}{p}, 0, U)=& \ \frac{(2p-1)k^2}{p c^{2p-1}} U^{2p} +
\frac{2nk}{c^{2p-1}} U^{2p-1}+\frac{c}{p}k^2, \\
Q(\frac{c}{p}, 0, U)=& \ \frac{2p-1}{c^{2p-1}} U^{2p-2}
\Big(\frac{k^2}{p c^{2p-1}} U^{2p} + 2 k U + 2n-\frac{c}{p} k^2
\Big).\end{aligned}$$ Let us reparametrize $x=\frac{U}{c}$, then $T(x)$ is defined on $[-1,1]$. $$\begin{aligned}
T(\frac{c}{p}, 0, x)=& \ P(x)+(n+kcx)\sqrt{Q(x)}\\
=& \ \frac{(2p-1)k^2 c}{p} x^{2p} + 2nk
x^{2p-1}+ \frac{c}{p} k^2 \nonumber\\
& \ +(n+k c \,x)\sqrt{(2p-1) x^{2p-2} \Big(\frac{k^2}{p} x^{2p} + 2kx
+ \frac{2n} {c}-\frac{k^2}{p} \Big)}\end{aligned}$$
Note that $P(x)$ is increasing on $[-1,0]$ and has a unique zero $x_0 \in (-1,0)$. We already have $T(-1)>0$ and $T(x_0)>0$ from (\[condition 1\]), (\[condition 2\]), and (\[condition 3\]). It suffices to show $T(x)>0$ on $[-1, x_0]$. To that end, let us introduce $$W(x)=-(P(x))^2+(n+kxc)^2 Q(x).$$
Obviously we also have $W(-1)>0$ and $W(x_0)>0$. By Lemma \[T positive\] below we conclude that $W(x)>0$ on $[-1,x_0]$, which implies $T(x)>0$ for any $x \in [-1, 0]$, thus completing the proof of Proposition \[any class\].
\[T positive\] Given any $n \geq 2$ and $k$ positive integers and $c \in (0,
\frac{n}{k}-2\epsilon_p]$ for $p \geq p_0(n,k)$, there exists $p_1(n, k, c)$ such that for any $p>p_1$, either there exists $-1<x_1<x_0$ such that $W(x)$ is increasing on $[-1, x_1]$ and decreasing on $[x_1, x_0]$, or $W(x)$ is increasing on $[-1, x_0]$.
*Proof of Lemma \[T positive\].* A straightforward calculation shows that $$\begin{aligned}
\frac{d W}{d x}& = \ -2P(x)P^{\prime}(x)+2(n+kcx) kc Q(x)+(n+kcx)^2
Q^{\prime}(x)\\
& = \ (2p-1)(n+kcx)x^{2p-3} J(x),\end{aligned}$$ where $$\begin{aligned}
J(x)=& \ -4k^3 c \frac{p-1}{p} x^{2p+1} -2nk^2 \frac{2p+1}{p} x^{2p}
+2k^2 c (2p+1) x^2\\
& \ +(8npk-2kn-\frac{4k^3c}{p}-2k^3c)x+(\frac{4pn^2}{c}-\frac{4n^2}{c}
-2k^2n+\frac{2nk^2}{p}).\end{aligned}$$
We add a brief remark on $J(-1)$ when $c=\frac{n}{k}-2\epsilon_p$. It follows that $$\begin{aligned}
J(-1)=& \ \frac{4p}{c}(n-kc)^2 +6k^3c-6nk^2+2k^2c+2kn-\frac{4n^2}{c} \nonumber\\
=& \ \frac{4pk^2}{\frac{n}{k}-2\epsilon_p} 4\epsilon_p^2+\epsilon_p
(-12k^3-4k^2-\frac{8nk}{\frac{n}{k}-2\epsilon_p}) \nonumber\\
=& \ \epsilon_p \Big(-12k^3-4k^2-\frac{8nk}{\frac{n}{k}-2\epsilon_p}+
\frac{16 p k^2 \epsilon_p}{\frac{n}{k}-2\epsilon_p} \Big).\end{aligned}$$ Recall that $\epsilon_p=\frac{2n}{2p+2k-1}$. A tedious calculation will lead to the fact that for $k$ suitably large ($k \geq 5$), $J(-1)>0$ for $p=p(n,
k)$ large enough. On the other hand, if $k$ is small ($k \leq 2$ for example), then $J(-1)<0$.
The crucial observation leads to the proof of the lemma is that, for any $0< c \leq \frac{n}{k}-2\epsilon_p$, we can find $p_1=p(n, k,
c)$ so that $J^{\prime}(x)>0$ for any $x \in [-1, x_0]$.
First let us estimate $x_0$, the unique zero of $P(x)$ on $(-1,0)$. We have $$\begin{aligned}
P(x_0)& = \ \frac{(2p-1)k^2 c}{p} x_0^{2p} + 2nk x_0^{2p-1}+ \frac{c}{p}
k^2 \label{root 1}\\
& = \ x_0^{2p-1} \Big[(2-\frac{1}{p})k^2 c x + 2nk \Big] + \frac{c}{p}
k^2. \label{root 2}\end{aligned}$$
Note that (\[root 1\]) implies that $2nk (-x_0)^{2p-1}
>\frac{c}{p}k^2$, while (\[root 2\]) implies that $$(-x_0)^{2p-1} \Big[2nk -(2-\frac{1}{p})k^2 c \Big] <\frac{c}{p}k^2.$$
To sum up, we have the following estimates on $x_0$ $$(\frac{ck}{2pn})^{\frac{1}{2p-1}}<-x_0<(\frac{ck}{n})^{\frac{1}{2p-1}}.
\label{x0 estimate}$$
Next we compute $J^{\prime}(x)$: $$\begin{aligned}
\frac{1}{p} J^{\prime}(x) \ =& \ -4k^3 c(1-\frac{1}{p})(2+\frac{1}{p})
x^{2p}-4nk^2 (2+\frac{1}{p}) x^{2p-1} \nonumber\\
& \ +4k^2c (2+\frac{1}{p})
x+\Big(8nk-\frac{2kn}{p}-\frac{4k^3c}{p^2}-\frac{2k^3c}{p}\Big)
\label{J prime formula 1}.\end{aligned}$$
In order to do calculation in the $O(\frac{1}{p})$ order, we note that when $p=p(n, k)$ is large enough, we have $$\frac{2n}{3p}<\epsilon_p <\frac{n}{p} . \label{epsilon estimate}$$ Plugging into $c=\frac{n}{k}-s$ where $2\epsilon_p \leq
s<\frac{n}{k}$, it follows from (\[J prime formula 1\]) that $$\begin{aligned}
\frac{1}{p} J^{\prime}(x)=& \ -4k^3
(\frac{n}{k}-s)(2-\frac{1}{p}-\frac{1}{p^2})
x^{2p}-4nk^2 (2+\frac{1}{p}) x^{2p-1} \nonumber\\
&\ +4k^2 (\frac{n}{k}-s) (2+\frac{1}{p})
x+\Big(8nk-\frac{2kn}{p}-\frac{4 n k^2}{p^2}+ \frac{4 k^3
s}{p^2}-\frac{2n k^2}{p}+\frac{2k^3 s}{p}\Big)
\\
\geq & \ -8 n k^2 x^{2p-1} (x+1)+8nk(1+x) \nonumber \\
& \ +4k^3 \frac{n}{kp} x^{2p}-4 \frac{nk^2}{p} x^{2p-1}+4k^2
(-2s+\frac{n}{kp})x-\frac{2kn}{p}-\frac{2n k^2}{p}+O(\frac{1}{p^2})
\label{J prime formula 2}\end{aligned}$$
It follows from (\[x0 estimate\]), (\[epsilon estimate\]), and (\[J prime formula 2\]) that $$\begin{aligned}
\frac{1}{p} J^{\prime}(x) \ \geq \ 4k^2 \frac{5n}{3p}
\Big[\frac{ck}{2pn} \Big]^{\frac{1}{2p-1}}-\frac{2kn}{p}-\frac{2n
k^2}{p}+O(\frac{1}{p^2})>\frac{2kn(2k-1)}{p}>0 \label{J prime
formula 3}\end{aligned}$$ for any $p>p_1(n, k, c)$ large enough and any $-1 \leq x \leq x_0$. Note that in (\[J prime formula 3\]), we have used $$\begin{aligned}
\lim_{p \rightarrow +\infty} \Big[\frac{ck}{2pn}
\Big]^{\frac{1}{2p-1}}=1. \label{limit to 1}\end{aligned}$$
This completes the proof of Lemma \[T positive\].
Let use remark that if we consider $M_{2,1}$, the conclusion of Theorem \[main any class\] is not necessarily stronger than that of Proposition \[canonical class\]. The point is that the degree of the generating polynomial $\phi(U)$ in Proposition \[any class\] might depend on the $c$ where $[-c, c]$ is the domain of $\phi(U)$. In particular, the degree $p$ goes to infinity as $c$ approaches to $0$, while the generating function in Proposition \[canonical class\] is quartic polynomial. However, we am able to show that the proof of Proposition \[any class\] can be used to establish the path-connectedness of all $U(n)$-invariant Kähler metrics of $H>0$ on any Hirzebruch manifold $M_{n,k}$.
The space of all $U(n)$-invariant Kähler metrics of $H>0$ on $M_{n,k}$ is path-connected.
In view of Corollary \[convexity\], it suffices to show that given any $0<c_1<c_2<\frac{n}{k}$, we can construct a continuous family of generating functions $\phi_c(U)$ where $U \in [-c,c]$ for any $c_1
\leq c \leq c_2$.
Such a family can be constructed following the proof of Proposition \[any class\]. In particular, if we examine (\[condition 1\]), (\[condition 2\]), (\[condition 4\]), (\[derivative positive\]), and (\[limit to 1\]), we conclude by the continuous dependence of parameters that for any $c \in [c_1, c_2]$ there exists an sufficiently large integer $p=p(n,k)$ which is independent of the choice of $c$, $\delta_1=\delta_1(p, n, k, c)$, and $\delta_2=\delta_2(p, n, k, c)$ such that $\phi_c(U)$ defined by (\[polynomial formula\]) is a continuous path of Kähler metrics with $H>0$.
Complete Kähler-Ricci solitons revisited
----------------------------------------
In this subsection, we are interested in complete Kähler-Ricci solitons in the following two cases.
\(1) Compact shrinking Kähler-Ricci soliton on Hirzebuch manifolds $M_{n,k} (k<n)$, which is the compactification of the total space of $H^{-k} \rightarrow \mathbb{CP}^{n-1}$.
\(2) Complete noncompact shrinking Kähler-Ricci soliton on the total space of $H^{-k} \rightarrow \mathbb{CP}^{n-1}$ when $k<n$.
These were constructed by Cao [@Cao], Koiso [@Koiso], Feldman-Ilmanen-Knopf [@F-I-K].
\[shrinkers and H>0\]
\(1) What can we say about compact shrinking Kähler-Ricci soliton with $H>0$? (2) Are there any complete noncompact shrinking Kähler-Ricci soliton with $H>0$?
Before answering Question \[shrinkers and H>0\], let us review the construction of Cao, Koiso, Feldman-Ilmanen-Knopf. We follow Koiso’s approach in the compact case and it can be extended to the complete noncompact case. Recall the Kähler metric $\tilde{g}$ on the compactification of the $\mathbb{C}^{\ast}$-bundle obtained by $H^{-k} \rightarrow \mathbb{CP}^{n-1}$. In the compact case, We make the following assumption on $M_{n,k}$:
\[canonical compactification\] Kähler metric $\tilde{g}$ on $M_{n,k}$ is in $2\pi c_1(M_{n,k})$. i.e. there exists a function $f$ on $\hat{L}$ such that $$Ric(\tilde{g})-\tilde{\omega_g}=\sqrt{-1}\partial \bar{\partial} f.$$
\[Koiso soliton condition\]
Under Assumption \[basic compactification\] and Assumption \[canonical compactification\], if we further assume $U_{min}=-D_{min}$, then $f$ in Assumption \[canonical compactification\] is given by $$\frac{d \phi}{dU}+\frac{\phi}{Q}\frac{dQ}{dU}+2U-\phi \frac{df}{dU}
=0. \label{f equation}$$ Moreover, we have $U_{max}=D_{max}$ and $Ric(g_0)=g_0$ on $M$.
If we further assume that $f$ in Assumption \[canonical compactification\] is give by $f=-\alpha U$ for some constant $\alpha $, then it follows that $\nabla f=-\frac{\alpha}{2}H$ is a holomorphic vector field. Equation (\[f equation\]) becomes the shrinking soliton equation. $$\frac{d \phi}{dU}+\frac{\phi}{Q}\frac{dQ}{dU}+2U-\alpha \phi=0.
\label{shrinking equation}$$
Note that $Q=(1+\frac{k}{n}U)^{n-1}$, so Equation (\[shrinking equation\]) takes the form $$\frac{d \phi}{dU}+\frac{k(n-1)}{n+kU}\phi+2U-\alpha \phi=0.
\label{shrinking equation simple}$$
Equation (\[shrinking equation simple\]) can be solved explicitly: $$\phi(U)=\frac{2\eta(U,\alpha)}{Q(U)}-\frac{2
e^{\alpha(U-U_{min})}}{Q(U)} \eta(U_{min},\alpha), \label{formal
further}$$ where $\eta(U,\alpha)$ is a polynomial of degree $n$ defined by $$\int xe^{-\alpha x}Q(x) dx=-e^{-\alpha x} \eta(x,\alpha).
\label{integral}$$
In the compact case it follows from Proposition \[Koiso soliton condition\] that $U_{min}=-1$ and $U_{max}=1$, therefore $\phi>0$ is a smooth function on $(-1, 1)$ with $\frac{d \phi}{dU}|_{U=-1}=2$ and $\frac{d \phi}{dU}|_{U=1}=-2$. While in the noncompact case we set $U_{min}=-1$ and $U_{max}=+\infty$, therefore $\phi>0$ is a smooth function on $(-1, +\infty)$ with $\frac{d
\phi}{dU}|_{U=-1}=2$.
For any integer $1 \leq k<n$, consider $(\mathbb{CP}^{n-1}, g_0)$ where $g_0$ is the Fubini-Study metric with $Ric (g_0)=g_0$.
\(1) We assume that the shrinking soliton metric on $M_{n,k}$ is of the form (\[def of g\]), and satisfies Assumptions \[assump 1\], \[basic compactification\], and \[canonical compactification\]. Then there exists a unique shrinking Kähler-Ricci soliton on each $M_{n,k}$ when $1 \leq k<n$. It is unique in the sense that the value $\alpha>0$ in the associated holomorphic vector field $\nabla
f=-\frac{\alpha}{2}H$ is determined by the unique solution of $\phi(1)=0$. Cao proved that the Ricci curvature of the soliton metric is positive on $M_{n,k}$ if and only if $k=1$.
\(2) There exists a unique complete shrinking Kähler-Ricci soliton on the total space of $L^{k} \rightarrow \mathbb{CP}^{n-1}$ whose value $\alpha>0$ is determined by the unique solution to $\eta(-1,\alpha)=0$.
Indeed, by (\[shrinking equation simple\]) and Proposition \[curvature tensors\], it is easy to write down the curvature tensors for shrinking solitons. $$\begin{aligned}
&A=-\frac{1}{2}[{\alpha}^2-\frac{2 \alpha
k(n-1)}{n+kU}-\frac{k^2n(n-1)}{(n+kU)^2}]
\phi+(\alpha-\frac{k(n-1)}{n+kU})U+1\\
&B=\frac{[k^2 n-k(n+kU) \alpha]\phi+2k(n+kU)U}{2(n+kU)^2},\\
&C=\frac{2(n+kU)-k^2 \phi}{(n+kU)^2}.\end{aligned}$$
It is easy to see that along the zero section $U=-1$, $2B+\sqrt{AC}>0$ implies that $$\alpha<\alpha_0(n,k) \doteq \frac{(n-2k)(k+1)}{n-k} \label{critical
alpha}$$ For example, when $n=2, k=1$, the necessary condition for $H>0$ is $\alpha<\alpha_0(2,1)=0$, however the corresponding $\alpha$ on $M_{2,1}$ is the unique positive root of the equation $e^{2\alpha}(-{\alpha}^2+2)-3{\alpha}^2-4 \alpha-2=0$ ($\alpha
\simeq 0.5276195199$). Therefore the Cao-Koiso shrinking soliton on $M_{2,1}$ does not satisfy $H>0$.
Next we analyze the noncompact case in more details. Consider the following polynomial which is of degree $n$ in terms of $\alpha$. $${\alpha}^{n+1}\eta(x,\alpha) = -{\alpha}^{n+1} e^{\alpha x} \int
xe^{-\alpha x} (1+\frac{k}{n}x)^{n-1} dx$$ Making use of the integration formula $$\int z^n e^z dz=(\sum_{l=0}^{n} (-1)^{n-l} \frac{n !}{l !} z^l)
e^z+C,$$ Some routine calculation leads to $$\begin{aligned}
{\alpha}^{n+1}\eta(U,\alpha)=(\frac{k}{n})^{n-1}\Big(\sum_{l=1}^{n}
\frac{n!}{l!} \frac{(n+kU)^{l-1}(n+kU-l)}{k^{l}}
{\alpha}^{l}+n!\Big).\end{aligned}$$ Therefore the value of ${\alpha}_{\ast}$ which solves the shrinking soliton equation, namely, the root of the polynomial $\eta(-1,\alpha)$, is reduced to root of the following polynomial of degree $n$ $$\chi(\alpha)=\sum_{l=1}^{n} \frac{n!}{l!}
\frac{(n-k)^{l-1}(n-k-l)}{k^{l}} {\alpha}^{l}+n!.$$
Similarly as in Feldman-Ilmanen-Knopf (see the equation $f$ in P197 [@F-I-K]), since $\chi(\infty)<0$ and $\chi(0)>0$, Descartes’ rule of signs implies there exists a unique positive root for $\chi(\alpha)$. Call it ${\alpha}_{\ast}$. For this choice $\alpha={\alpha}_\ast$, $\phi(U)$ has the asymptotical behavior $\phi \sim \dfrac{1}{\alpha_{\ast}}U$ as $U \rightarrow \infty$, and it shows that the soliton metric is complete along infinity.
Now we are interested in a more precise estimate of ${\alpha}_{\ast}$. First note that ${\alpha}_{\ast} > k$. To see this, we check $$\begin{aligned}
\chi(\alpha)&= \ \sum_{l=1}^{n} \frac{n!}{l!}
\frac{(n-k)^{l-1}(n-k-l)}{k^{l}} {\alpha}^{l}+n! \nonumber \\
&= \ \Big(\frac{\alpha (n-k)}{k}\Big)^n+\sum_{l=0}^{n-1} \frac{n!}{l!}
(\frac{\alpha (n-k)}{k})^l (1-\frac{\alpha}{k}). \label{rough
estimate}\end{aligned}$$
We propose the following conjecture on a precise estimate on $\alpha_{\ast}$.
\[root range\] For any $1 \leq k<n$, ${\alpha}_{\ast}$, which is the unique positive root of of the polynomial $\chi(\alpha)$, satisfies $\alpha_0(n,k)<{\alpha}_{\ast}<k+1$. If so, then none of Feldman-Ilmanen-Knopf shrinking solitons have positive holomorphic sectional curvature.
Argue similarly as in (\[rough estimate\]), we can show that Conjecture \[root range\] is indeed true if $k<n \leq k^2+2k$. It is very likely that it is true in general, as numerical experiments suggest.
On the other hand, compact shrinking solitons on $M_{n,k}$ could have positive holomorphic sectional curvature as the ratio $\frac{n}{k}$ grows larger.
(Some compact shrinking solitons on $M_{n,k}$ have $H>0$). If we fix $k=1$, then the lowest dimension example of Cao-Koiso shrinking solitons which have $H>0$ is $n=3$. On $M_{3,1}$ its local pinching constant is $\frac{1-\alpha}{(2-\alpha)(5-\alpha)} \simeq
0.05587$ where $\alpha \simeq 0.6820161326$. If $k=2$, the first compact example with $H>0$ is on $M_{7,2}$, where $\alpha \simeq
1.742423694$.
Let us explain the calculation in the case of $M_{3,1}$. In this case the corresponding $\alpha$ is the unique positive solution of the following equation $$16{\alpha}^3+24{\alpha}^2+18{\alpha}+6-e^{2{\alpha}}
(-4{\alpha}^3+6{\alpha}+6)=0.$$In particular, $\alpha \simeq 0.6820161326 <\alpha_0 (3,1)=1.5$ where $\alpha_{0}(n,k)$ defined in (\[critical alpha\]). One can show that that $(2B+\sqrt{AC})(n+kU)^2$ is increasing on $U \in
[-1,1]$ by a direct calculation, therefore we have $H>0$. In general, for any $M_{n,1}$ with $n \geq 3$, one could expect to verify that $\alpha<\alpha_0 (n,k)$ and $(2B+\sqrt{AC})(n+kU)^2$ is increasing on $U \in [-1,1]$. Therefore the Cao-Koiso shrinking soliton on Hirzebruch manifold $M_{n,1}$ have $H>0$ for any $n \geq
3$.
Let us calculate the local holomorphic pinching constant of Cao-Koiso soliton on $M_{3,1}$. A similar argument as in Example \[canonical class\] shows that $$\frac{\min_{||v||=1} H(U, v)}{\max_{||v||=1} H(U, v)}=
\left\{ \begin{array}{lll}
\frac{AC-4B^2}{A(A+C-4B)}, & U \in [-1, U_{\ast}]\\
\mbox{} & \mbox{} \\
\frac{C}{A}, & U \in [U_{\ast}, 1]
\end{array} \right.$$ Here $U_{\ast}$ is the solution of $2B=C$ on $[-1,1]$, whose numerical value is about $-0.5.73003$. Therefore the pinching constant is obtained at $U=-1$, which is $\frac{1-E}{(2-E)(5-E)}
\simeq 0.05587$.
Let us remark that if we drop the assumption of the shrinking soliton, it is easy to write down examples of complete Kähler metrics of $H>0$ on the total space of $H^{-k} \rightarrow
\mathbb{CP}^1$. For instance, we have
Define a function $\omega(U)$ on $[-\frac{1}{2}, +\infty)$ as follows: $$\omega(U)= \left\{ \begin{array}{lll}
\frac{1}{2}-2U^2, & U \in [-\frac{1}{2}, -\frac{1}{4})\\
& \\
c(U+2)+\frac{1}{2}\ln(U+2), & U \in [-\frac{1}{4}, \infty)
\end{array} \right.$$ where $c=\frac{3}{14}-\frac{2}{7} \ln(\frac{7}{4}) \sim 0.05439$ so that $\omega$ is continuous at $U=-\frac{1}{4}$. Choose a small number $\delta >0$ so that $\omega$ admits a convex smoothing $\phi$ which equals to $\omega$ except inside $(-\frac{1}{4}-\delta,
-\frac{1}{4}+\delta)$, Note that for $\omega$, both $C>0$ and $2B+\sqrt{AC}$ have positive lower bounds in $(-\frac{1}{4}-\delta,
-\frac{1}{4})$ and $(-\frac{1}{4}, -\frac{1}{4}+\delta)$. It guarantees the existence of a convex smoothing $\phi$ which in turns gives a complete Kähler metric with $H>0$ on the total space of $L^{-1} \rightarrow \mathbb{CP}^1$. The metric actually has positive bisectional curvature outside a compact subset. Moreover, its bisectional curvatures decay quadratically along the infinity. Asymptotically such a metric has a conical end and there exist holomorphic functions with polynomial growth on it.
$H>0$ is not preserved along the Kähler-Ricci flow
--------------------------------------------------
Recently there have been much progress on Kähler-Ricci flow with Calabi-symmetry on Hirzebruch manifolds $M_{n,k}$. See for example Zhu [@Zhu], Weinkove-Song [@SW2011], Fong [@Fong], and Guo-Song [@GuoSong]. In this subsection, we apply Hitchin’s example and new examples constructed in Theorem \[main any class\] to show that in general $H>0$ is not preserved by the Kähler-Ricci flow.
These results imply that if the initial metric $g_0$ is of $U(n)$-symmetry and in the Kähler class $\frac{b_0}{k}[E_{\infty}]-\frac{a_0}{k}[E_0]$ where $b_0>a_0>0$, then the flow always develops a singularity in finite time. Let $T<\infty$ denote the maximal existence time. Their results can be summarized as:
\(1) Suppose that the initial metric $g_0$ satisfies $\frac{|E_{\infty}|}{|E_0|}=\frac{b_0}{a_0}=\frac{n+k}{n-k}$, where $|E_0|$ denotes the volume of the divisor $E_0$ with respect to $g_0$. In this case the Kähler class of $g_0$ is proportional to the anti-canonical class of $M_{n,k}$. Then the Kähler-Ricci flow shrinks the fiber and the base uniformly and collapses to a point. The rescaled flow converges to the Cao-Koiso soliton on $M_{n,k}$ ([@Zhu]).
\(2) If $g_0$ satisfies $\frac{b_0}{a_0}<\frac{n+k}{n-k}$ , then the Kähler-Ricci flow shrinks the fiber first, and the flow collapses to the base $\mathbb{CP}^{n-1}$ ([@SW2011]). The rescaled flow converges to $\mathbb{C}^{n-1} \times \mathbb{CP}^1$ ([@Fong]).
\(3) If $g_0$ satisfies $\frac{b_0}{a_0}>\frac{n+k}{n-k}$, then the Kähler-Ricci flow shrinks the zero section $E_0$ first and hence ‘contracts the exceptional divisor’ ([@SW2011]). The rescaled flow converges to the Feldman-Ilmanen-Knopf shrinking soliton on the total space of $H^{-k} \rightarrow \mathbb{CP}^{n-1}$ ([@GuoSong]).
\(4) If $n \leq k$, then the Kähler-Ricci flow shrinks the fiber first, and the flow converges to the base $\mathbb{CP}^{n-1}$ in the Gromov-Hausdorff sense as $t \rightarrow T$ ([@SW2011]).
Based on Hitchin’s examples reformulated in Example \[other class\] and Proposition \[any class\], we have the following examples of the Kähler-Ricci flow.
On $M_{2,1}$, let us take the metric in the anti-canonical class constructed in Example \[canonical class\] as the initial metric. Then the normalized flow converges to the Cao-Koiso soliton. Unfortunately, the positivity of $H$ breaks down. Therefore, in general, $H>0$ is not preserved along the Kähler-Ricci flow. If instead we start from initial metric corresponding to $\phi_{c}(U)=c-\frac{x^2}{c}$ on $[-c,c]$ s in Example \[other class\], where $0<c<\frac{2}{3}$, then the limiting metric of the unnormalized flow is $(\mathbb{CP}^1, cg_{FS})$. In this case, it is not clear whether $H>0$ is preserved, or how the holomorphic pinching constant of $g(t)$ evolves.
On $M_{3,1}$, if we pick initial metric as the Cao-Koiso shrinking soliton, it is a fixed point of the normalized flow, therefore the holomorphic pinching constant remains constant.
On $M_{4,1}$, if we pick initial metric by the examples $\phi_{c}(U)=c-\frac{x^2}{c}$ on $[-c,c]$ for any $0<c<\frac{4}{3}$. Then all three cases mentioned above could occur. For $c=1$, the normalized flow evolves $\phi_1(U)$ to the Cao-Koiso soliton. While the initial metric has the holomorphic pinching constant $2/27
\simeq 0.074$, the limit metric has the holomorphic pinching constant $\simeq 0.095$. It indeed improves after a long time, even though we do not know the short time effect of it. If $1<c<\frac{4}{3}$, then the normalized flow evolves $\phi_{c}(U)$ to the limiting Feldman-Ilmanen-Knopf shrinking soliton on the total space of $L^{-1} \rightarrow \mathbb{CP}^{3}$. However the limiting soliton no longer has $H>0$, and once again we see that $H>0$ is not preserved under the Kähler-Ricci flow.
Therefore a natural question arises, namely, is there an effective way to construct a one-parameter family of deformation of Kähler metrics with $H>0$? It would be ideal if the holomorphic pinching constant could enjoy some monotonicity properties along this deformation.
Kähler metrics of $H>0$ from the submanifold point of view
==========================================================
In this section, we discuss holomorphic pinching of the canonical Kähler-Einstein metrics on some Kähler $C$-spaces. In general, we would like to discuss the question of constructing $H>0$ metric from the submanifold point of view.
Consider the flag threefold, or more generally, let $M$ be the hypersurface in $\mathbb{CP}^n \times
\mathbb{CP}^n$ defined by $$\sum_{i=1}^{n+1} z_i w_i = 0, \label{defining flag}$$ where $n\geq 2$ and $([z],[w])$ are the homogeneous coordinates. Let $g$ be the restriction on $M$ of the product of the Fubini-Study metrics (each of which has $H=2$). Then the holomorphic sectional curvature of $g$ is between $2$ and $\frac{1}{2}$. So the holomorphic pinching constant is $\frac{1}{4}$, which is dimension free.
Let us work on the case $n=2$ and in the inhomogeneous coordinates $[1,
z_1, z_2]$ and $[w_1, 1, w_2]$. The hypersurface $M^3
\subset \mathbb{CP}^2 \times \mathbb{CP}^2$ is defined by $w_1+z_1+z_2w_2=0$ and can be parametrized by $$(t_1, t_2, t_3) \rightarrow [1, t_1, t_2] \times [-t_1 - t_2 t_3, 1,
t_3] .$$
It follows that $$\frac{\partial }{\partial t_1}= \frac{\partial }{\partial
z_1}-\frac{\partial }{\partial w_1}, \ \frac{\partial }{\partial
t_2}=\frac{\partial }{\partial z_2}-t_3 \frac{\partial }{\partial
w_1}, \ \frac{\partial }{\partial t_3}=-t_2 \frac{\partial
}{\partial w_1}+\frac{\partial }{\partial w_2}.$$
Therefore under $\{\frac{\partial }{\partial t_1}, \frac{\partial
}{\partial t_2}, \frac{\partial }{\partial t_3} \}$ the induced metric $\widetilde{g}$ has the form: $$\begin{pmatrix}
g_{1 \bar{1}}+h_{1 \bar{1}} & g_{1 \bar{2}}+\overline{t_3} h_{1 \bar{1}}
& \overline{t_{2}}h_{1 \bar{1}}+h_{1 \bar{2}} \\
g_{2 \bar{1}}+\overline{t_3} h_{1 \bar{1}} &
g_{2 \bar{2}}+|t_3|^2 h_{1 \bar{1}}
& t_3 \overline{t_2} h_{1 \bar{1}}-t_3 h_{1 \bar{2}} \\
t_2 h_{1 \bar{1}}-h_{2 \bar{1}} &
t_2 \overline{t_3} h_{1 \bar{1}}-\overline{t_3} h_{2 \bar{1}}
& |t_2|^2 h_{1 \bar{1}}+h_{2 \bar{2}}
\end{pmatrix}$$ where $g_{i \bar{j}}=g(\frac{\partial }{\partial z_i},
\frac{\partial }{\partial \overline{z_j}})$ and $h_{i
\bar{j}}=g(\frac{\partial }{\partial w_i}, \frac{\partial }{\partial
\overline{w_j}})$ where $1 \leq i, j \leq 2$ are Fubini-Study metrics on two factors $\mathbb{CP}^2$ respectively: $$g_{i \bar{j}}=\frac{\delta_{ij}}{1+|z|^2}-\frac{z_j
\overline{z_i}}{(1+|z|^2)^2}, \ \ h_{i
\bar{j}}=\frac{\delta_{ij}}{1+|w|^2}-\frac{w_j
\overline{w_i}}{(1+|w|^2)^2}.$$
Recall that the curvature tensor of $\widetilde{g}$ is given by the formula $$R_{i \bar{j} k \bar{l}}=-\frac{\partial^2 \widetilde{g}_{k
\bar{l}}}{ \partial {z_i} \partial \overline{z}_j}+\widetilde{g}^{p
\bar{q}} \frac{\partial \widetilde{g}_{k \bar{q}}}{\partial z_i}
\frac{\partial \widetilde{g}_{p \bar{l}}}{\partial \overline{z}_j}.$$ Note that there is a natural $U(3)$-action on $M$ because of the defining equation (\[defining flag\]), which acts transitively. Therefore it suffices to calculate the curvature tensor of the induced metric $\widetilde{g}$ at the point $(t_1, t_2,
t_3)=(0,0,0)$. Now pick an orthonormal frame $\{e_1, e_2,
e_3\}=\{\frac{1}{\sqrt{2}} \frac{\partial }{\partial t_1},
\frac{\partial }{\partial t_2}, \frac{\partial }{\partial t_3} \}$. Then a straight forward calculation shows that the only non vanishing curvature components under $\{e_1, e_2, e_3\}$ are the following: $$\begin{aligned}
R_{1 \bar{1} 1 \bar{1}}=1, \ \ R_{2 \bar{2} 2 \bar{2}}=2, \ \ R_{3
\bar{3} 3
\bar{3}}=2,\\
R_{1 \bar{1} 2 \bar{2}}=R_{1 \bar{1} 3 \bar{3}}=\frac{1}{2}, \ \ \
R_{2 \bar{2} 3 \bar{3}}=-\frac{1}{2}.\end{aligned}$$
From this, we get that $R_{i \bar{j}}=2\delta_{ij}$ for any $1 \leq i\neq j \leq 3$, so $\widetilde{g}$ is Kähler-Einstein. Once we have all the curvature components, it is direct to see that $\min_{||X||=1}
H(X)=H(\frac{e_2+e_3}{\sqrt{2}})=\frac{1}{2}$ and $\max_{||X||=1}
H(X)=2$.
The same argument works for the hypersurface defined by $\sum_{i=1}^{n+1} z_i w_i = 0$ in $\mathbb{P}^n \times \mathbb{P}^n$. It gives the same pinching constant $\frac{1}{4}$ for any $n \geq 2$.
However, if we try a similar calculation on other types of bidegree $(p, q)$ hypersurfaces in $\mathbb{CP}^n \times \mathbb{CP}^n$, the argument breaks down.
As a simple example, consider the bidegree $(2,1)$ hypersurface defined by $\sum_{i=0}^{2} z_i^2 w_i = 0$ in $\mathbb{CP}^2 \times
\mathbb{CP}^2$ given by homogeneous coordinates $([z],[w])$. Consider the following parametrization of the hypersurface: $$(t_1, t_2, t_3) \rightarrow [1, t_1, t_2,] \times [-t_1^2-t_2^2 t_3,
1, t_3].$$ and $$\frac{\partial }{\partial t_1}= \frac{\partial }{\partial z_1}-2t_1
\frac{\partial }{\partial w_1}, \ \frac{\partial }{\partial
t_2}=\frac{\partial }{\partial z_2}-2 t_2 t_3 \frac{\partial
}{\partial w_1}, \ \frac{\partial }{\partial t_3}=-t_2^2
\frac{\partial }{\partial w_1}+\frac{\partial }{\partial w_2}.$$
The corresponding $\widehat{g}$ induced from the product of Fubini-Study metric on $\mathbb{CP}^2 \times \mathbb{CP}^2$ is $$\begin{pmatrix}
g_{1 \bar{1}}+4|t_1|^2h_{1 \bar{1}} &
g_{1 \bar{2}}+4 t_1 \overline{t_2 t_3} h_{1 \bar{1}}
& 2t_1\overline{t_{2}}^2 h_{1 \bar{1}}-2t_1 h_{1 \bar{2}} \\
g_{2 \bar{1}}+4\overline{t_1}\overline{t_2 t_3} h_{1 \bar{1}} &
g_{2 \bar{2}}+4 |t_2|^2 |t_3|^2 h_{1 \bar{1}}
& 2 |t_2|^2 \overline{t_2} t_3 \overline{t_2} h_{1 \bar{1}}-2t_2t_3 h_{1 \bar{2}} \\
t_2^2 \overline{t_1} h_{1 \bar{1}}-2 \overline{t_1} h_{2 \bar{1}} &
2 |t_2|^2 t_2 \overline{t_2} \overline{t_3} h_{1 \bar{1}}-2\overline{t_2 t_3}\overline{t_3} h_{2 \bar{1}}
& |t_2|^4 h_{1 \bar{1}}+h_{2 \bar{2}}
\end{pmatrix}$$
If we only calculate the curvature at $(t_1, t_2, t_3)=(0,0,0)$, we already encounter some negativity of $H$. First note that $(\frac{\partial }{\partial t_1}, \frac{\partial }{\partial t_2},
\frac{\partial }{\partial t_3})$ is orthonormal at $(0,0,0)$, then it follows that $$R_{1 \bar{1} 1 \bar{1}}=-\frac{\partial^2 \widehat{g}_{1 \bar{1}}}{
\partial {t_1} \overline{t_1}}=-4 h_{1 \bar{1}}-
\frac{\partial^2 g_{1 \bar{1}}}{ \partial {z_1} \overline{z_1}}=-2.$$
The same problem occurs for a general bidegree $(2,1)$ hypersurface in $\mathbb{CP}^n \times \mathbb{CP}^n$. Of course, this just means that for a bidegree $(2,1)$ hypersurface in $\mathbb{CP}^n \times
\mathbb{CP}^n$, the restriction of the product of the Fubini-Study metric on the hypersurface does not have $H>0$. But presumably there could be other metrics on it with $H>0$. This is indeed the case, and we have the following result:
\[proper kahler cone\] Let $M^n$ be any smooth bidegree $(p,1)$ hypersurface in $\mathbb
{CP}^r \times \mathbb {CP}^s$, where $n=r+s-1$, $p\geq 1$, and $r$, $s\geq 2$. Then $M^n$ admits a Kähler metric with $H>0$. Morever, when $p>r+1$, the Kähler classes of all the Kähler metrics with $H>0$ form a proper subset of the Kähler cone of $M^n$.
\[Proof of Proposition \[proper kahler cone\]\] Let $[z]$ and $[w]$ be the homogeneous coordinates of $\mathbb{CP}^r$ and $\mathbb {CP}^s$, respectively. Let $\pi:
\mathbb{CP}^r \times \mathbb{ CP}^s \rightarrow \mathbb{CP}^r$ be the projection map. Suppose that $M^n$ is defined by $$\sum_{i=1}^{s+1} f_i(z_1, \cdots, z_{r+1})w_i =0,$$ where each $f_i$ is a homogeneous polynomial of degree $p$. Consider the sheaf map $h: {\mathcal O}^{\oplus (s+1)} \rightarrow
{\mathcal O}(p)$ on $\mathbb{CP}^r$ defined by $$h(e_i)=f_i(z), \ \
\ 1\leq i\leq s+1,$$ where $e_i = (0, \ldots , 0, 1, 0, \ldots , 0)$ has $1$ at the $i$-th position. Clearly, $h$ is surjective, and its kernel sheaf $E$ is locally free. Since $M^n ={\mathbb P}(E)$ over $\mathbb {CP}^r$, by the result of [@AHZ], we know that $M^n$ admits Kähler metrics with $H>0$.
To see the second part of the statement, let us denote by $H_1$, $H_2$ the hyperplane section from the two factors restricted on $M$, then we have $c_1(M)=(r+1-p)H_1+sH_2$. Clearly, $H_1^{r+1}=0$, $H_2^{s+1}=0$, and since $M^n \sim pH_1+H_2$, we have $H_1^r
H_2^{s-1}=1$ and $H_1^{r-1} H_2^{s}=p$ on $M^n$. For any Kähler class $[\omega ] = aH_1+bH_2$ where $a>0$ and $b>0$, we have $$c_1(M)\cdot [\omega ]^{n-1} = a^{r-2}b^{s-2} \Big[ \begin{pmatrix}
n-1 \\
r
\end{pmatrix} sa^2 +
\begin{pmatrix}
n-1 \\
r-1
\end{pmatrix} (r+1+sp-p)ab + \begin{pmatrix}
n-1 \\
r-2
\end{pmatrix} (r+1-p)b^2 \Big].$$ So when $p>r+1$ and $b>>a$, we know that the total scalar curvature of $(M^n,\omega )$ is negative. Thus the Kähler classes of metrics with $H>0$ can not fill in the entire Kähler cone.
For a smooth bidegree $(p,2)$ hypersurface $M^n$ in $\mathbb {CP}^r
\times \mathbb {CP}^s$, where $n=r+s-1$ and $p\geq 2$, one may raise the question of whether $M^n$ admits Kähler metrics with $H>0$? The answer might be yes in view of Proposition \[proper kahler cone\]. Note that if we project to $\mathbb {CP}^r$, then $M$ becomes a holomorphic fibration over $\mathbb {CP}^r$ whose generic fiber are smooth quadrics.
If $M^3$ is a compact Kähler manifold with local holomorphic pinching constant strictly greater $\frac{1}{4}$, then is it biholomorphic to a compact Hermitian symmetric space, i.e. $\mathbb{CP}^3$, $\mathbb{CP}^{2} \times \mathbb{CP}^{1}$, $\mathbb{CP}^{1} \times \mathbb{CP}^{1}\times \mathbb{CP}^{1}$, or $Q^{3}$ which is the smooth quadric in $\mathbb{CP}^4$?
Let us conclude the discussion here by giving a couple of general remarks. First, if we we want to construct metrics with $H>0$ from the submanifold point of view, in particular, as complete intersections. Then it seems difficult to find examples other than those already known (such as Hermitian symmetric spaces or Kähler $C$-spaces, or projectivized vector bundles covered in [@AHZ]). For instance, if we consider a cubic hypersurface $M^n \subset
\mathbb {CP}^{n+1}$ and $g$ be the restriction on $M$ of the Fubini-Study metric. Then it is unclear if $(M^n,g)$ can have $H>0$, though we expect the answer is no. As another example, if we consider the restriction of the ambient Fubini-Study metric onto a complete intersection, where typically we need to restrict to degree $1$ or $2$. Let us consider $M^n$ as the intersection of two quadrics, in the case of $n=2$, it is the del Pezzo surface of degree $4$, or $\mathbb{CP}^2$ blowing up $5$ points. It is unlikely that the induced metric on $M^n$ could have $H>0$. We plan to discuss these questions elsewhere.
Secondly, it is a general belief that the existence of a Kähler metric of $H>0$ is a ‘large open’ condition, as illustrated in this paper on any Hirzebruch manifold. Therefore, it is reasonable to expect if a projective manifold $M$ admits a Kähler metric $g_0$ of $H>0$, then there exsits a small deformation of such a metric $g_1$ which lies in a Hodge class and still has $H>0$. hence one can conclude from a theorem of Tian ([@Tian]) that $g_1$ can be approximated by pull backs of Fubini-Study metrics by a sequence of projective embeddings $\phi_k: M \rightarrow \mathbb{CP}^{N_k}$. However, it seems difficult to construct examples of Kähler metrics of $H>0$ in this way because of the implicit nature of $\phi_k$ and $N_k$.
0.2cm
**Acknowledgments.** We would like to thank Gordon Heier for helpful suggestions and a few corrections on some inaccuracies and typos on a preliminary version of this paper, and Bennett Chow for his interests. Bo Yang is grateful for Xiaodong Cao for the encouragement, the Math Department at Cornell University for the excellent working condition, as well as Bin Guo, Zhan Li, and Shijin Zhang for helpful discussions.
[99]{}
Apostolov, V.; Calderbank, D. M. J.; Gauduchon, P.; T[ø]{}nnesen-Friedman, C. W.. *Extremal Kähler metrics on ruled manifolds and stability.* G[é]{}om[é]{}trie diff[é]{}rentielle, physique math[é]{}matique, math[é]{}matiques et soci[é]{}t[é]{}. II Astérisque No. 322 (2008), 93-150.
Alvarez, A.; Chaturvedi, A.; Heier, G.. *Optimal pinching for the holomorphic sectional curvature of Hitchin’s metrics on Hirzebruch surfaces.* Contemp. Math., 654, 2015, 133-142.
Alvarez, A.; Heier, G.; Zheng, F.. *On projectivized vector bundles and positive holomorphic sectional curvature.* arXiv:1606.08347.
Berger, M.. *Sur quelques variétés riemanniennes suffisamment pincées*. Bull. Soc. Math. France **88**, 1960, 57-71.
Berger, M.. *Pincement riemannien et pincement holomorphe.* Ann. Scuola Norm. Sup. Pisa (3), **14**, 1960, 151-159. and *Correction d’un article antérieur*. ibid. (3), **16**, (1962), 297.
Berger, M. *Sur les variétés d’[E]{}instein compactes.* 1966 Comptes [R]{}endus de la [III]{}e [R]{}éunion du [G]{}roupement des [M]{}athématiciens d’[E]{}xpression [L]{}atine ([N]{}amur, 1965) pp. 35-55 Librairie Universitaire, Louvain.
Bishop, R. L.; Goldberg, S. I.. *On the topology of positively curved Kähler manifolds*. Tôhoku Math. J. (2), **15**, 1963, 359-364.
Bishop, R. L.; Goldberg, S. I.. *On the topology of positively curved Kähler manifolds.* II. Tôhoku Math. J. (2), **17**, 1965, 310-318.
Boucksom, S.; Demailly, J.-P.; P[ă]{}un, M.; Peternell, T.. *The pseudo-effective cone of a compact Kähler manifold and varieties of negative Kodaira dimension.* J. Algebraic Geom. 22 (2013), no. 2, 201-248.
Cao,H.-D., *Existence of gradient Kähler-Ricci solitons,* Elliptic and parabolic methods in geometry (Minneapolis, MN, 1994), 1-16.
Calabi, E.. *Métriques kählériennes et fibrés holomorphes.* Ann. Sci. École Norm. Sup. (4) 12 (1979), no. 2, 269-294.
Calabi, E.. *Extremal Kähler metrics.* Seminar on Differential Geometry, pp. 259-290, Ann. of Math. Stud., 102, Princeton Univ. Press, Princeton, N.J., 1982.
Chen, B.-y.. *Extrinsic spheres in Kähler manifolds, II.* Michigan Math. J. **24** (1977), no. 1, 97-102.
Chen, X. X.. *On Kähler manifolds with positive orthogonal bisectional curvature.* Adv. Math. 215 (2007), no. 2, 427-445.
Chen, X. X.; Tian, G.. *Ricci flow on Kähler-Einstein surfaces.* Invent. Math. 147 (2002), no. 3, 487-544.
Chen, X. X.; Tian, G.. *Ricci flow on Kähler-Einstein manifolds.* Duke Math. J. 131 (2006), no. 1, 17-73.
Feldman, M., IImanen, T. and Knopf, D., *Rotationally symmetric shrinking and expanding gradient Kähler-Ricci solitons,* J. Diff. Geom., 65, 2003, 169-209.
Fong, F. T.-H.. *Kähler-Ricci flow on projective bundles over Kähler-Einstein manifolds.* Trans. Amer. Math. Soc. 366 (2014), no. 2, 563-589.
Griffiths, P.; Harris, J.. *Principles of algebraic geometry.* Reprint of the 1978 original. Wiley Classics Library. John Wiley & Sons, Inc., New York, 1994.
Gu, H.; Zhang, Z.. *An extension of Mok’s theorem on the generalized Frankel conjecture.* Sci. China Math. 53 (2010), no. 5, 1253-1264.
Guo, B.; Song, J.. *On Feldman-Ilmanen-Knopf conjecture for the blow-up behavior of the Kähler Ricci flow.* ArXiv:1505.04869.
Heier, G.; Wong, B.. *Scalar curvature and uniruledness on projective manifolds.* Comm. Anal. Geom., 20(4):751-764, 2012.
Heier, G,; Wong, B.. *On projective Kähler manifolds of partially positive curvature and rational connectedness.* arXiv:1509.02149.
Hitchin, N.. *On the curvature of rational surfaces.* Differential geometry Proc. Sympos. Pure Math., Vol. XXVII, Part 2, 65-80. Amer. Math. Soc., Providence, R. I., 1975.
Hwang, A. D.; Singer, M. A.. *A momentum construction for circle-invariant Kähler metrics.* Trans. Amer. Math. Soc. 354 (2002), no. 6, 2285-2325.
Kobayashi, S.; Wu, H.-H.. *On holomorphic sections of certain hermitian vector bundles.* Math. Ann. 189, 1970, 1-4.
Koiso, N., *On rotionally symmetric Hamilton’s equation for Kähler-Einstein metrics,* Recent Topics in Diff. Anal. Geom., Adv. Studies in Pure Math., 18-I, Academic Press, Boston, MA, 1990, 327-337.
Koiso, N.; Sakane, Y.. *Nonhomogeneous Kähler-Einstein metrics on compact complex manifolds.* Curvature and topology of Riemannian manifolds, 165-179, Lecture Notes in Math., 1201, Springer, Berlin, 1986.
Kobayashi, S.; Nomizu, K.. Foundations of differential geometry. Vol. II. Reprint of the 1969 original. Wiley Classics Library. John Wiley & Sons, Inc., New York, 1996.
LeBrun, C.. *Counter-examples to the generalized positive action conjecture.* Comm. Math. Phys. 118 (1988), no. 4, 591-596.
Mori, S.. *Projective manifolds with ample tangent bundles.* Ann. of Math. (2) 110 (1979), no. 3, 593-606.
Munteanu, O.; Wang, J.. *Positively curved shrinking Ricci solitons are compact.* ArXiv:1504.07898.
Ni, L.. *Ancient solutions to Kähler-Ricci flow.* Math. Res. Lett. 12 (2005), no. 5-6, 633-653.
Simanca, S. R.. *Kähler metrics of constant scalar curvature on bundles over [${\bf C}{\rm P}_{n-1}$]{}.* Math. Ann. 291 (1991), no. 2, 239-246.
Siu, Y. T.; Yau, S. T.. *Compact Kähler manifolds of positive bisectional curvature.* Invent. Math. 59 (1980), no. 2, 189-204.
Song, J.; Weinkove, B.. *The Käler-Ricci flow on Hirzebruch surfaces.* J. Reine Angew. Math. 659 (2011), 141-168.
Tian, G.. *On a set of polarized Kähler metrics on algebraic manifolds.* J. Differential Geom. 32 (1990), no. 1, 99-130.
Tsukamoto, Y.. *On Kälerian manifolds with positive holomorphic sectional curvature.* Proc. Japan Acad. 33 (1957), 333-335.
Wilking, B.. *A Lie algebraic approach to Ricci flow invariant curvature conditions and Harnack inequalities.* J. Reine Angew. Math. 679 (2013), 223-247.
Yang, X.. *Hermitian manifolds with semi-positive holomorphic sectional curvature.* Math. Res. Lett. 23 (2016), no. 3, 939-952.
Yau, S.-T.. *A review of complex differential geometry.* Several complex variables and complex geometry, Part 2 (Santa Cruz, CA, 1989), 619-625, Proc. Sympos. Pure Math., 52, Part 2, Amer. Math. Soc., Providence, RI, 1991.
Yau, S.-T.. *Open problems in geometry.* 1-28, Proc. Sympos. Pure Math., 54, Part 1, Amer. Math. Soc., Providence, RI, 1993.
Zhu, X.. *Kähler-Ricci flow on a toric manifold with positive first Chern class.* ArXiv:math/0703486.
[^1]: Research partially supported by an AMS-Simons Travel Grant
[^2]: Research partially supported by a Simons Collaboration Grant
[^3]: Note that we follow the notations of $E_0$ and $E_{\infty}$ as in Calabi [@Calabi2], which is different with those in [@GH].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Jessica R. Kitamura'
- 'Lucimara P. Martins'
- Paula Coelho
bibliography:
- 'references.bib'
date: 'Received 03 October 2016 / Accepted 31 January 2017'
title: A study on missing lines in the synthetic solar spectrum near the Ca triplet
---
Introduction
============
Synthetic stellar spectra are extensively used in many different applications in astronomy, from the determination of atmospheric parameters of observed stellar spectra to the study of galaxy integrated spectra (as one of the main ingredients of stellar population models). Generating an accurate synthetic stellar spectrum requires the thermodynamic description of the atmosphere as a function of depth [[e.g.]{} @K93; @CK04; @Px11], the so-called model atmosphere. These model atmospheres may include simplifications for complex physical process, such as convection, or computational limitations, for example, geometry and/or boundary conditions; often they depend on incomplete or imprecise atomic and molecular opacities [@Cz15].
To generate the synthetic stellar spectra, a stellar synthesis code [[e.g. SYNTHE;]{} @KA81] solves the radiative transfer of a given model atmosphere using an atomic and molecular line list. Atomic and molecular opacities also greatly affect the radiative transfer in stars and consequently their physical structure. Despite this, we are far away from having complete and accurate line lists and, in fact, half of the known lines in the stellar spectra are not present in the line lists with wavelengths considered to have good accuracy [@K11]. Besides, detailed spectral models calibrated to a single star (like Sun or Vega) are very important but, at the same time, rare and poorly tested [e.g. @Cz15].
In recent decades, there were great efforts to complete and improve the line lists and nowadays there are large atomic databases ([*e.g.*]{} National Institute of Standards and Technology-Atomic Spectra Database[^1] and Vienna Atomic Line Database[^2]).
Although the current line lists include millions of absorption lines, only a small fraction of these lines were actually accurately measured in laboratory or have accurate parameters derived, and the uncertainty in the transition parameters for atomic lines can reach up to 200%. Even for the Sun, our closest and most studied star, the synthetic spectrum still does not completely reproduce all the features and many of the lines are missing in the synthetic spectrum [@K11]. For the purpose of spectrophotometric studies, the missing lines have been included via the so-called “predicted lines", in which, rather than measured, one or both energy levels of the transition was only predicted through quantum mechanics calculations [@K92]. These lines are essential to better describe the structure of atmospheric models and for spectrophotometric forecasts [e.g. @SL96; @C14]. However, these quantum mechanical predictions are accurate only at a few percent level and the wavelength of these lines can be largely incorrect. As such, the inclusion of the predicted lines is inappropriate to generate theoretical stellar spectra with high resolution [e.g. @B94; @CK04; @M05]. As a consequence of this complex scenario, synthetic libraries are only partially able to reproduce observed spectra and the quality of the model varies with spectral type and wavelength range modelled [e.g. @bertone05; @bertone+08; @MC07; @C14].
In an attempt to fill this gap, many groups have been working through the years to improve the quality of atomic data used to generate synthetic spectra, focusing on reducing the uncertainty of the transition probabilities [e.g. @T90; @K02; @FW06; @SS10; @B11; @P11; @Wi11; @C12; @R13] and the broadening parameters [e.g. @AO95; @BO97; @B98; @L99; @B00; @K02; @D03; @Dm03]. Besides, many limitations of these atomic line lists have been approached by different methods and authors, mainly aimed at chemical analyses of stellar photospheres using high resolution spectra [e.g. @S97; @BW08; @B03; @J04; @H11; @SV13; @W13]. However, not only known lines have to be improved, but the identification of missing lines is necessary to fill significant gaps that are poorly modelled in stellar spectra.
In this work, we identify and catalogue missing lines in the atomic and molecular line lists used to generate synthetic spectra of the Sun based on a recent line list available in literature, and, when possible, calibrate the atomic lines. For this task we used an observed solar spectrum [@W11], which is an obvious choice since it is the highest quality observed stellar spectrum available in the literature. Additionally, uncertainties involved in the determination of its effective temperature, superficial gravity, chemical abundance, etc. are smaller than for any other star. The identification of missing lines was performed in the wavelength range from $\mathrm {8\,470\,\AA ~\mbox{to}~ 8\,740\,\AA}$ since it is the spectral region covered by the Radial-Velocity Spectrometer on the GAIA telescope [@Lindegren+96; @Mignard05] launched in 2013. Even though the spectral resolution of the mentioned instrument is relatively low ($\lambda/\Delta\lambda \sim$ 11500 ), we can expect this wavelength region to be highly attractive to stellar and galactic studies for the years to come. We used the observed solar spectrum available from @W11 and compared with the synthetic spectrum generated using the atomic and molecular line list from [@S04], updated by @C14. We use the code ALiCCE [Atomic Lines Calibration using the Cross-Entropy Algorithm; @M14] to calibrate the atomic parameters.
This paper is structured as follows: in Section 2, we give details about the observed solar spectrum used in this work; in Section 3 we present the identification and characterisation of the missing lines in the synthetic solar spectrum; in Section 4 we show the calibration of the atomic parameters of some of these lines; and, in Section 5, we present our discussion and conclusions.
The observed solar spectrum
===========================
High resolution spectral atlases of the Sun have been produced since the middle of the 20th century [e.g. @M40]. However, most of the solar spectral atlases are of disk-centre regions. The solar flux spectra, taken over the integrated disk, are much less common. A flux atlas shows the mean effects of rotation, convection, and centre-to-limb variation. Thus, flux spectra are fundamental for comparisons with other stellar spectra and with synthetic spectra [@W11].
The observed solar spectrum used in this work was published by @W11. These authors claim that the quality of their spectrum is higher than @K05 because of the more efficient subtraction of the telluric lines. The observation was taken at McMath-Pierce Solar Telescope, located on Kitt Peak, using the Fourier Transform Spectrograph (hereafter FTS). The solar integrated light FTS spectra were obtained mainly on two occasions: in 1980–1981 for @K84 atlas and in 1989 for monitoring the irradiance spectrum over the solar cycle [@ML91]. Both sets were observed near the peaks of sunspot activity. These were the data used by @W11 to produce a new flux atlas.
The wavelength coverage of this observed spectrum ranges from $\mathrm {2\,958\,\AA ~\mbox{to}~ 9\,257\,\AA}$ and the spectral resolution varies from 350000 to 700000 ($R=\lambda/\Delta\lambda$). The signal to noise in the continuum goes beyond several hundreds. In principle, the FTS could have observed the entire spectral region in a single integration, however, to reduce the photon noise, six separate observations were carried out with the spectral coverage in each limited by optical bandpass filters. The Doppler shift correction was determined empirically by measuring the solar line positions and correcting them to the frequencies in @N94. Thus, the solar gravitational redshift was also removed from the wave number scales.
@W11 made the telluric emission correction, knowing that from $\mathrm {2\,958\,\AA ~\mbox{to}~ 5\,400\,\AA}$ the solar spectrum is free from any terrestrial lines. However, weak lines of $\mathrm {H_2O}$ begin to appear from $5\,400\,\AA$ and $\mathrm {O_2}$ lines also appear from $5\,790\,\AA$. The scheme developed by @W11 to correct for the telluric spectrum was to use the solar spectra obtained from many sets with different air masses, in the morning or evening, applied to the disk-centre spectra. This was carried out because the flux spectra were not taken in suitable airmass sets to allow the transmission spectra to be extracted. They found the best signal to noise and correction with the flux spectrum from October 1989, using a disk-centre observation from July 1983.
More details about the observations and the reduction processes are available in @W11. Figure \[sunspec\] shows the @W11 spectrum for the spectral region used in this work.
{width="19cm"}
Line identifications
====================
We used the solar model atmosphere from @CK04, which is based in ATLAS9 [@K70; @S04], to generate the synthetic spectrum of the Sun. There are two different downloadable atmosphere models: one using the @AGS05 chemical abundances and another using the @GS98 [^3]. Both are tested in this work. The effective temperature of the Sun considered is ${\rm T_{eff}} = $5777 K, the surface gravity is $\log g = $4.44 [@K70] and convection was taken into account. We chose ATLAS9 because it is a static and local thermodynamic equilibrium atmosphere model and is still one of the most commonly used models for chemical abundances studies to generate synthetic stellar libraries, and it reproduces well the colours of observed stars [@MC07].
To generate the synthetic spectrum, we used the SYNTHE code [@KA81] in its Linux version published by @S04. The atomic and molecular line lists used are publicly available with ATLAS9 and were updated according to @C14. The spectral synthesis code uses the atomic and molecular line lists to solve the equation of radiative transfer. The parameters necessary for the calculation of each absorption line are the central wavelength, the energy of the upper and lower levels, the oscillator strength, and the broadening parameters (natural, Stark and Van der Waals). The broadening parameters dominate the line wings, while the oscillator strengths dominate the line depths. The spectral resolution adopted for this work, given by $R=\lambda/\Delta\lambda$, was 676000, which is the same resolution of the observed spectrum in this region. To reproduce the line broadening from the solar rotational velocity, we applied a rotational velocity ($V_{\sin i}$) of 2.4 $km~s^{-1}$ [@S78] and microturbulence of 1.0 $km~s^{-1}$.
We produced synthetic spectra for three different chemical abundances: @AGS05 and @GS98, whose model atmosphere were downloaded from the website of F. Castelli as indicated above, and @A09. To produce a synthetic solar spectrum with the latest abundances we used the same model atmosphere as with @AGS05 abundances, but changing abundances in SYNTHE code accordingly. This approach is valid for small variations in the chemical abundances, which should not significantly change the atmosphere structure. The three different abundances were tested just to ensure that weak missing lines were not just an artefact of small abundance changes. The comparison between the spectra was carried out visually over the entire spectral range studied to identify the missing lines in the theoretical spectrum. The telluric spectrum used for the correction of the solar spectrum was also inspected to avoid the classification of possible residual subtractions as missing lines.
All lines in this wavelength range that were present in the observed spectrum and absent in the theoretical spectrum have been identified and catalogued. We found 39 missing lines in the wavelength range analysed. For the characterisation of these lines, each one was fitted by a Gaussian profile. @W11 also identified some lines in the spectrum of the Sun and therefore we compared our catalogued lines with the lines identified by these authors. Table \[lineident\] shows the central wavelength values of the catalogued lines, their identification by @W11 (when present), the equivalent width (EW), the full width at half maximum (FWHM), and the asymmetry of each line. Figures \[lines1\] and \[lines2\] show each of the missing lines identified.
$\#$ @W11 $\mathrm{\lambda_{central}}$ (Å) EW (Å) FWHM $\mathrm{(km~s^{-1})}$ asymmetry
------ -------- ---------------------------------- -------- ----------------------------- -----------
1 8470.36 0.0036 2.747 -0.259
2 8482.88 0.0046 2.309 -0.130
3 - 8483.44 0.0078 2.379 -0.274
4 CN red 8499.30 0.0179 3.311 0.003
5 - 8502.73 0.0192 5.581 0.851
6 CN red 8503.22 0.0042 6.556 1.034
7 - 8508.12 0.0008 1.564 -0.138
8 - 8509.59 0.0025 2.431 0.173
9 - 8513.45 0.0029 3.122 0.099
10 - 8515.65 0.0016 2.156 -0.182
11 - 8517.29 0.0130 3.040 0.651
12 8525.01 0.0096 2.180 -0.193
13 - 8526.96 0.0073 2.714 0.317
14 - 8535.50 0.0011 1.977 -0.047
15 - 8554.27 0.0707 8.576 0.679
16 8559.05 0.0051 2.502 -0.129
17 8559.74 0.0052 2.271 0.008
18 - 8560.64 0.0079 4.084 0.579
19 - 8570.17 0.0016 2.406 0.317
20 CN red 8575.79 0.0129 5.960 0.657
21 8585.58 0.0106 2.642 -0.228
22 - 8586.21 0.0195 4.114 -0.007
23 8592.12 0.0053 2.476 0.053
24 - 8601.69 0.0093 6.157 0.530
25 - 8602.19 0.0043 2.442 0.082
26 - 8608.33 0.0111 2.362 -0.466
27 - 8615.32 0.0080 2.335 -0.468
28 CN red 8619.08 0.0075 3.992 0.254
29 CN red 8622.75 0.0075 2.242 -0.129
30 - 8623.73 0.0043 2.218 -0.427
31 - 8624.45 0.0021 2.162 -0.141
32 8700.32 0.0039 2.254 -0.860
33 - 8705.17 0.0035 2.631 -0.154
34 - 8706.06 0.0035 2.594 -0.143
35 8707.15 1.3182 26.072 0.827
36 8707.34 1.7846 35.624 1.73
37 - 8725.21 0.0046 2.383 -0.376
38 CN red 8730.25 0.0037 2.855 -0.099
39 - 8732.72 0.0013 2.281 0.099
: Characterisation of the catalogued lines: Identification by @W11, central wavelength, EW, FWHM, and asymmetry for each missing line.
\[lineident\]
{width="17cm"}
{width="17cm"}
We can see in Table \[lineident\] that about half of the lines (21 lines) have widths of about 2.4 $km~s^{-1}$, which correspond to the net effect of rotational velocity, microturbulence, and macroturbulence in the Sun[^4]. We assume that the broadening of these lines is dominated by these listed combined effects, rather than an effect intrinsic of the lines. We can also speculate that these lines are produced or dominated by one atomic transition as opposed to molecular transitions or a blend of lines. Another reason that these lines are produced or dominated by only one electronic transition is their symmetry. Asymmetric lines are more likely produced by a blend of lines, whether atomic or molecular. The asymmetry was measured through the skewness of the line, where values close to zero mean very symmetrical distributions. Thirteen of these lines have asymmetry values close to zero (lines 2, 8, 10, 12, 16, 17, 23, 25, 29, 31, 33, 34, and 39). For these lines, |asymmetry| <0.20. However, line 29 was identified by @W11 as due to a CN transition, so we cannot rule out that some of these lines come from molecular transitions. Another caveat that might be taken into account is that asymmetries and wavelength shifts in stars might be correlated with convection in the sense that warm, rising convective elements are blueshifted and cool and falling convective elements are redshifted. However, according to @CONFIRMARdravins98 this effect in the Sun is around 300 m/s, which is much smaller than visible asymmetries seen here. For the value of the central wavelength, this implies in a $\Delta \lambda$ of 0.0085 Å at 8500 Å, which is also minor for the values observed here. Signatures with much larger widths are more likely to be produced by a blend of lines, although single atomic lines can be broadened by hyperfine and isotopic splitting. [@Wahlgren05]. Two signatures have widths that are smaller than 2.0 $km~s^{-1}$ (lines 7 and 14). This is possibly due to measurement errors caused by the difficulty of adjusting these signatures, which are in regions with a high density of lines.
After characterising each signature, we looked for the atomic data of these missing lines in the NIST database to verify whether their absence was not only a matter of outdated line lists. The atomic data collected from NIST were measured in the laboratory. Table \[linenist\] shows the relation between catalogued missing lines and the data found in NIST.
------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- --
$\#$ **$\mathrm{\lambda_{ident}}$ (Å) & **$\mathrm{\lambda_{NIST}}$ (Å) & **Element & **log [*gf*]{} & **$\mathrm{E_i (cm^{-1})}$ & **$\mathrm{E_k (cm^{-1})}$ & **Element Code & **Quality\
1 & 8470.36 & 8470.37 & & -2.5 & 54275.649 & 66078.272 & 26.01 & D\
2 & 8482.88 & 8482.67 & & - & - & - & 67.00 & -\
2 & 8482.88 & 8482.684 & & - & 66048.39 & 77833.88 & 21.01 & -\
3 & 8483.44 & 8483.39 & & - & - & - & 42.00 & -\
3 & 8483.44 & 8483.56 & & - & 32391.95 & 44176.23 & 44.00 & -\
4 & 8499.30 &8499.34 & & - & - & - & 26.00 & -\
5 & 8502.73 & 8502.7 & & - & - & - & 65.00 & -\
5 & 8502.73 & 8502.714 & & - & 92116.529 & 103874.261 & 26.01 & -\
5 & 8502.73 & 8502.714 & & - & 93328.553 & 105086.265 & 26.01 & -\
6 & 8503.22 & 8503.212 & & -0.14& 97062.844 & 108819.860 & 27.01 & C+\
7 & 8508.12 & 8508.08 & & - & - & - & 71.00 & -\
8 & 8509.59 & 8509.60 & & - & - & - & 26.00 & -\
9 & 8513.45 & 8513.38 & & - & - & - & 35.00 & -\
9 & 8513.45 & 8513.5 & & - & 71336.005 & 83078.8 & 80.00 & -\
9 & 8513.45 & 8513.57 & & - & - & - & 57.00 & -\
10 & 8515.65 & 8515.475 & & -1.341 & 121528.72 & 133268.68 & 16.01 & C\
11 & 8517.29 & 8517.37 & & - & - & - & 3.01 & -\
12 & 8525.01 & 8525.029 & & - & 36975.588 & 48702.535 & 26.00 & -\
13 & 8526.96 & 8526.99 & & - & - & - & 41.00 & D\
13 & 8526.96 & 8527.03 & & 0.08 & 169615.12 & 181339.27 & 22.02 & -\
14 & 8535.50 &\
15 & 8554.27 &\
16 & 8559.05 &\
17 & 8559.74 & 8559.741 & & - & 41178.412 & 52857.804 & 26.00 & -\
18 & 8560.64 & 8560.54 & & - & - & - & 41.00 & -\
19 & 8570.17 & 8570.099 & & - & 92358.625 & 104023.921 & 26.01 & -\
19 & 8570.17 & 8570.099 & & - & 45509.152 & 57174.430 & 26.00 & -\
20 & 8575.79 & 8575.78 & & - & - & - & 52.01 & -\
20 & 8575.79 & 8575.87 & & - & - & - & 41.00 & -\
20 & 8575.79 & 8575.92 & & - & - & - & 73.00 & -\
21 & 8585.58 & 8585.4403 & & -0.47 & 90386.533 & 102030.965 & 26.01 & C+\
21 & 8585.58 & 8585.52 & & - & 269388.34 & 281032.70 & 38.02 & -\
22 & 8586.21 &\
23 & 8592.12 & 8592.22 & & -1.59 & 135999.37 & 147634.61 & 49.01 & D+\
24 & 8601.69 &\
25 & 8602.19 &\
26 & 8608.33 & 8608.3067 & & - & 157572.1025 & 169185.5968 & 55.01 & -\
26 & 8608.33 & 8608.3882 & & - & 157572.1025 & 169185.4867 & 55.01 & -\
27 & 8615.32 &\
28 & 8619.08 &\
29 & 8622.75 &\
30 & 8623.73 & 8623.805 & & - & 189437.7396 & 201030.3698 & 18.01 & -\
31 & 8624.45 &\
32 & 8700.32 & 8700.25 & & - & 32915.539 & 44406.31 & 49.00 & -\
33 & 8705.17 &\
34 & 8706.06 &\
35 & 8707.15 & 8707.048 & & - & 246483.317 & 257965.11 & 8.01 & -\
35 & 8707.15 & 8707.14 & & - & 49346.729 & 60828.41 & 12.00 & -\
35 & 8707.15 & 8707.215 & & - & 32620.38 & 44101.99 & 43.00 & -\
36 & 8707.34 &\
37 & 8725.21 &\
38 & 8730.24 &\
39 & 8732.75 & 8732.679 & & - & 270011.52 & 281459.61 & 38.02 & -\
39 & 8732.72 & 8732.746 & & 0.02 & 98568.907 & 110016.915 & 26.01 & E\
****************
------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- --
\[linenist\]
In Table \[linenist\], the column $\lambda_{Ident}$ concerns all the lines identified as missing from the comparison of the observed solar spectrum with the synthetic spectrum listed in Table \[lineident\]. The data in columns $\lambda_{NIST}$, Element, log [*gf*]{}, and $E_i$ and $E_k$ energies are the values from the NIST database that correspond to the identified wavelengths. The data in column *Element Code* is the code used by Kurucz to characterise each ion. From all the identified missing lines, NIST suggests that about a quarter are from Fe transitions. For some of the identified lines, we have two or more matching lines found in NIST. This is likely because we allowed a search inside a range of $0.2\,\AA$ from the measured central wavelength to account for possible errors in the central wavelength measurement.
All lines found in NIST with measured log [*gf*]{} were included in the atomic line list used in the spectral synthesis code. Unfortunately, there are very few cases where this happens (seven lines) and we did not observe any significant improvement in the synthetic spectra generated after this inclusion. This means that despite the absence of these lines in the atomic list used, the missing lines are due to transitions other than those catalogued in NIST.
In an attempt to find candidates for the missing lines in the line list, we searched for them in the VALD database. The values from VALD are not necessarily measured in laboratory and many of them are theoretically calculated or empirically calibrated.
From the 39 lines missing in the solar spectrum characterised in this work, we found counterparts for 22 on NIST, although only 7 have a measured log [*gf*]{}, as previously mentioned. On VALD we found counterparts for 16 of these 22 lines, although 9 of them were already included in the Kurucz line list (lines 1, 7, 8, 11, 13, 18, 20, 30, and 35). The candidates found in VALD are listed in Table \[valdlines\]. Lines indicated in grey in this table are already present on the line list, either with the same exact parameters or with values very close to those found. For some of these lines we found more than one counterpart in VALD, which were not found in NIST.
All lines from Table \[valdlines\] that are missing in the line list are lines from Fe ions, which is not surprising given that Fe has a complex electron configuration and is very abundant. Weak lines from Fe are important because they are frequently used for the measurement of chemical abundances of stars. Since the atomic parameters from VALD are not obtained from laboratory, we decided to calibrate these iron lines before including them in the line list.
------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- -- --
$\#$ **$\mathrm{\lambda_{ident}}$ (Å) & **$\mathrm{\lambda_{VALD}}$ (Å) & **log [*gf*]{} & **Elem. & **$\mathrm{E_i~(cm^{-1})}$ & **$\mathrm{E_k~(cm^{-1})}$ & **log $\Gamma_{R}$ & **log $\Gamma_{S/N_{e}}$ & **log $\Gamma_{W/N_{H}}$\
1& 8470.36 & 8470.359& -2.519 & & 54275.6400 & 66078.2700 & 0.000 & 0.000 & 0.000\
1& 8470.36 & 8470.365& -2.734 & & 54275.6490 & 66078.2720 & 8.560 & -6.530 & -7.900\
1& 8470.36 & 8470.390& -1.443 & & 47177.2340 & 58979.8230 & 8.47 & -4.41 & -7.45\
4 & 8499.34 & 8499.330& -0.693 & & 47017.1880 & 58779.5900 & 8.470 & -4.560 & -7.430\
5 & 8502.70 & 8502.670& -6.818 & & 8154.7140 & 19912.4950 & 3.520 & -6.280 & -7.850\
5 & 8502.70 & 8502.705& -3.059 & & 92116.5290 & 103874.2610 & 8.920 & -5.840 & -7.750\
5 & 8502.70 & 8502.720& -5.427 & & 93328.5530 & 105086.2650 & 8.770 & -5.700 & -7.500\
7 & 8508.12 & 8508.112& -0.040 & & 23524.2400 & 35274.5000 & 0.00 & 0.00 & 0.00\
8 & 8509.60 & 8509.617& -3.436 & & 35257.3240 & 47005.5060 & 8.240 & -5.330 & -7.550\
11 & 8517.29 & 8517.369& -0.672 & & 490071.1000 & 501808.5900 & 10.410 & -5.570 & 0.000\
12 & 8525.01 & 8525.011& -1.523 & & 46889.1420 & 58616.1100 & 8.210 & -4.140 & -7.380\
12 & 8525.01 & 8525.026& -3.599 & & 36975.5880 & 48702.5350 & 7.730 & -5.990 & -7.800\
12 & 8525.01 & 8524.999& -7.819 & & 86124.3480 & 97851.3320 & 8.960 & -5.680 & -7.710\
12 & 8525.01 & 8525.060& -8.023 & & 113056.8470 & 124783.7480 & 8.610 & -4.370 & -7.420\
13 & 8526.96 & 8526.958& -1.560 & & 10922.7400 & 22647.0300 & 0.00 & 0.00 & 0.00\
13 & 8526.96 & 8527.060& 0.138 & & 169615.1200 & 181339.2700 & 9.090 & -4.950 & -7.520\
17 & 8559.74 & 8559.738& -1.510 & & 41178.4120 & 52857.8040 & 8.190 & -5.710 & -7.710\
18 & 8560.64 & 8560.538& -1.690 & & 8705.3200 & 20383.6200 & 0.00 & 0.00 & 0.00\
19 & 8570.17 & 8570.094& -3.231 & & 45509.1520 & 57174.4300 & 8.010 & -3.990 & -7.260\
19 & 8570.17 & 8570.081& -2.358 & & 92358.6250 & 104023.9210 & 9.010 & -5.900 & -7.740\
20 & 8575.79 & 8575.879& -1.710 & & 12357.7000 & 24015.1100 & 0.00 & 0.00 & 0.00\
20 & 8575.79 & 8575.901& -2.060 & & 13351.4500 & 25008.8300 & 0.00 & 0.00 & 0.00\
21 & 8585.58 & 8585.437& -1.011 & & 90386.5330 & 102030.9650 & 8.520 & -5.180 & -7.480\
21 & 8585.58 & 8585.540& -5.930 & & 102952.1700 & 114596.4620 & 8.790 & -5.020 & -7.360\
30 & 8623.73 & 8623.850& -0.780 & & 189437.7360 & 201030.3000 & 0.000 & 0.000 & 0.000\
35 & 8707.15 & 8707.135& -2.760 & & 49346.7290 & 60828.4100 & 0.000 & 0.000 & 0.000\
39 & 8732.73 & 8732.634& -3.392 & & 93840.4050 & 105228.5590 & 8.750 & -5.650 & -7.500\
39 & 8732.73 & 8732.746& -0.080 & & 98568.9070 & 110016.9150 & 8.920 & -5.730 & -7.690\
39 & 8732.73 & 8732.823& -1.424 & & 100492.0250 & 111939.9320 & 9.040 & -5.740 & -7.680\
39 & 8732.73 & 8732.852& -6.238 & & 45913.4970 & 57361.3660 & 8.030 & -4.890 & -7.290\
******************
------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- -- --
\[valdlines\]
Line calibration
================
We used the code ALiCCE [Atomic Lines Calibration using the Cross-Entropy Algorithm; @M14] for the calibration of the atomic parameters. The ALiCCE code was developed to automatically calibrate atomic lines using the cross-entropy method [e.g. @R97; @R99; @M04; @K06; @B05]. The cross-entropy method is a general Monte Carlo approach to combinatorial and continuous multi-extremal optimisation and importance sampling, which is a general technique for estimating properties of a particular distribution using samples generated randomly from a different statistical distribution rather than the distribution of interest [@R97; @R99; @kroese+06].
For each iteration, ALiCCE generates N different atomic line lists with each atomic parameter to be calibrated varying inside a given interval. It then makes external calls to the spectral synthesis code SYNTHE for each of the N lists. The output spectra generated is then compared with the observed spectrum of the Sun (the comparison star used for this work) and a performance function is calculated for each of the N lists, which is a measure of how well the synthetic spectrum represents the observed spectrum. The lists are then ranked and the interval for each atomic parameter is recalculated based on the mean and standard deviation of the top 5% solutions. The process starts again until the stopping criteria is fulfilled. More details about the code can be found in @M14.
In this work we calibrated only the oscillator strength but not the broadening values since the width of the lines are dominated by the rotational velocity (plus micro and macro turbulence) of the Sun. All possible atomic transitions for a given missing line, as shown in Table \[valdlines\], were calibrated together. However, since the wavelength range between each missing line was large enough, we calibrated them separately, running ALiCCE for each missing line identified. The log [*gf*]{} values obtained from VALD were used as initial guesses. Table \[resultallice\] shows the values obtained for each line calibrated by ALiCCE.
------ --------------- -------------- --------------------
VALD ALiCCE
$\#$ $\lambda$ (Å) log [*gf*]{} log [*gf*]{}
4 8499.330 -0.693 -1.066 $\pm$ 0.001
5 8502.705 -3.059 no convergence
12 8524.999 -7.819 no convergence
12 8525.011 -1.523 -1.223$\pm$ 0.001
12 8525.027 -2.208 -2.208 $\pm$ 0.001
12 8525.060 -8.023 no convergence
17 8559.738 -1.509 -1.941 $\pm$ 0.004
19 8570.081 -2.385 0.389 $\pm$ 0.002
19 8570.095 -3.281 -1.751 $\pm$ 0.014
21 8585.540 -5.93 no convergence
39 8732.852 -6.238 no convergence
------ --------------- -------------- --------------------
: Comparison between VALD and ALiCCE atomic parameters.
\[resultallice\]
A significant improvement in the reproduction of the identified lines was obtained after the calibration of the log [*gf*]{} with ALiCCE. These results can be seen in Figure \[calib\_lines\], which show the comparison of synthetic spectra generated without the presence of the lines, values of the atomic parameters from VALD, and log [*gf*]{} values obtained by ALiCCE.
{width="17cm"}
The oscillator strengths could not be derived for five lines given that they did not converge in less than 500 iterations and had errors greater than 0.010 (see details of error evaluation in @M14).
Discussion and conclusion
=========================
Improving the quality of the synthetic stellar spectra is of paramount importance to our understanding of stars and galaxies. Synthetic spectra libraries has an advantage over empirical spectra ones, which is the possibility of generating spectra with any atmospheric parameters desired for any chemical abundance pattern desired. However, synthetic spectra also have limitations, since they can only be as good as the ingredients used to generate them.
One of the major problems faced by theoretical stellar spectra is the uncertainty and incompleteness of the atomic and molecular line lists used to generate them. There are still many lines present in stellar spectra that are unknown. Moreover, among the millions of known lines present in the line lists, few of them have accurate and precise values, whether they are measured or computed.
We have identified and catalogued missing lines in the atomic and molecular line lists used for producing synthetic stellar spectra in the spectral region of $\mathrm {8\,470\,\AA ~\mbox{to}~ 8\,740\,\AA}$. The line lists that we used are based on previous work by @K70 [@KA81; @CK04] and updated according to @C14. For this, we used the observed spectrum of the Sun published by @W11 and we generated the synthetic spectrum using the spectral synthesis code SYNTHE [@KA81] with the model atmosphere of the Sun generated by ATLAS9 [@K70; @S04].
We found 39 lines missing from the atomic and molecular line lists within the analysed wavelength region. We performed a characterisation of each line by measuring their equivalent widths and their widths at half maximum (FWHM). From these values and from the analysis of the symmetry of the lines, we conclude that about one-third ($\sim$ 36%) of the identified lines can be produced or are dominated by a single atomic transition. These are the lines with smaller widths and most symmetric profiles. Wider and asymmetric lines are likely generated by line blends of multiple atomic species or molecules. We searched the identified lines in the NIST atomic database and we found counterparts for 22 of the 39 missing lines, but only 7 had all atomic parameters measured. We added these lines to the atomic list used by the spectral synthesis code, but obtained no improvement in the spectrum produced. This means that the missing lines likely have major contributions of other species.
Because some lines found in NIST lacked the atomic parameters needed to include them in the line list, we also looked for them on VALD. We found counterparts for 14 of these 22 lines, although 8 of them were already included in the line list adopted here. For some of these lines we found more than one counterpart for a given line. All the lines from VALD not present in the atomic line list were from and . Since the atomic parameters from VALD are mostly not produced by measurements in laboratory, but are instead determined empirically or theoretically, we attempted to calibrate the atomic parameters of these lines before including them in the line list.
We used the ALiCCE code [@M14] to calibrate the oscillator strength of the Fe lines missing in the line list. We did not try to calibrate the broadening parameters since the width of the lines is dominated by the rotational velocity of the Sun. The ALiCCE code found better log [*gf*]{} values for 5 of the 10 lines. The new values significantly improved the reproduction of the solar spectrum. For the remaining lines the code has not been able to find results that could improve the reproduction of the solar spectrum.
We would like to thank Robert Kurucz, the referee of this paper, for his valuable comments that definitely improved the paper. We would like to thank Gillian Nave for helping with the search of some of the unidentified lines. J.R.K. acknowledges FAPESP (2014/00502-9) and CAPES for financial support. L.M. thanks CNPQ for financial support through grant 303697/2015-6 and FAPESP through grant 2015/14575-0. P.C. acknowledges CNPQ for financial support through grant 305066/2015-3.
[^1]: <http://www.nist.gov/>
[^2]: <http://vald.astro.uu.se>
[^3]: <http://wwwuser.oats.inaf.it/castelli/sun.html>
[^4]: A significant contribution to the spectral line broadening might come from velocity fields in the stellar photosphere. In 1D model atmospheres these velocity fields are represented by microturbulence and macroturbulence, although physically they have little to do with turbulence [@Doyle+14].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '> Team formation is a core problem in AI. Remarkably, little prior work has addressed the problem of mechanism design for team formation, accounting for the need to elicit agents’ preferences over potential teammates. Coalition formation in the related hedonic games has received much attention, but only from the perspective of coalition stability, with little emphasis on the mechanism design objectives of true preference elicitation, social welfare, and equity. We present the first formal mechanism design framework for team formation, building on recent combinatorial matching market design literature. We exhibit four mechanisms for this problem, two novel, two simple extensions of known mechanisms from other domains. Two of these (one new, one known) have desirable theoretical properties. However, we use extensive experiments to show our second novel mechanism, despite having no theoretical guarantees, empirically achieves good incentive compatibility, welfare, and fairness.'
author:
- |
Mason Wright\
Computer Science & Engineering\
University of Michigan\
Ann Arbor, MI\
[email protected] Yevgeniy Vorobeychik\
Electrical Engineering and Computer Science\
Vanderbilt University\
Nashville, TN\
[email protected]
bibliography:
- 'mdtfPaper.bib'
title: Mechanism Design for Team Formation
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss the concept and the performance of a powerful future ground-based astronomical instrument, 5@5 - a [**5**]{} GeV energy threshold stereoscopic array of several large imaging atmospheric Cherenkov telescopes installed [**at**]{} a very high mountain elevation of about [**5**]{} km a.s.l. - for the study of the $\gamma$-ray sky at energies from approximately 5 GeV to 100 GeV, where the capabilities of both the current space-based and ground-based $\gamma$-ray projects are quite limited. With its potential to detect the “standard” EGRET $\gamma$-ray sources with spectra extending beyond several GeV in exposure times from 1 to $10^3$ seconds, such a detector may serve as an ideal “Gamma-Ray Timing Explorer” for the study of transient non-thermal phenomena like $\gamma$-radiation from AGN jets, synchrotron flares of microquasars, the high energy (GeV) counterparts of Gamma Ray Bursts, [*etc.*]{} 5@5 also would allow detailed $\gamma$-ray spectroscopy of persistent nonthermal sources like pulsars, supernova remnants, plerions, radiogalaxies, and others, with unprecedented for $\gamma$-ray astronomy photon statistics. The existing technological achievements in the design and construction of multi (1000) pixel, high resolution imagers, as well as of large, 20 m diameter class multi-mirror dishes with rather modest optical requirements, would allow the construction of such a detector in the foreseeable future, although in the longer terms from the point of view of ongoing projects of 100 GeV threshold IACT arrays like H.E.S.S. which is in the build-up phase. An ideal site for such an instrument could be a high-altitude, 5 km a.s.l or more, flat area with a linear scale of about 100 m in a very arid mountain region in the Atacama desert of Northern Chile.'
address:
- 'MPI für Kernphysik, D-69029 Heidelberg, Germany'
- 'Astronomy and Astrophysics Department, Pontificia Universidad Catolica de Chile, Casilla 306, Santiago 22, Chile'
author:
- 'F.A. Aharonian, A.K. Konopelko, H.J. Völk'
- 'H. Quintana'
title: '5@5 - a 5 GeV energy threshold array of imaging atmospheric Cherenkov telescopes at 5 km altitude'
---
[*Astroparticle Physics, accepted for publication*]{}
Atmospheric imaging Cherenkov technique, GeV detector
Introduction
============
The high detection rate, the ability of effective separation of electromagnetic and hadronic showers, and the good accuracy of reconstruction of the direction of primary $\gamma$-rays are three remarkable features of the imaging atmospheric Cherenkov telescope (IACT) technique (see e.g. [@Cawley; @Hillas; @Fegan]). The recent detections of TeV $\gamma$-rays from several galactic and extragalactic objects (see e.g. [@Weekes99]) provide the basis for the further development of [*ground-based*]{} gamma-ray astronomy.
Qualitative improvements of the IACT technique in the next few years will most probably be linked to [*stereoscopic*]{} observations of $\geq 100 \, \rm GeV$ $\gamma$-rays [@AhAk]. This approach not only allows an unambiguous determination of the energy and of the arrival direction of $\gamma$-ray primaries on an [*event-by-event*]{} basis, but also significantly improves the efficiency of rejection of hadronic showers produced by cosmic rays [@array], as it was recently demonstrated by the HEGRA system of 5 imaging telescopes operating in the energy region from 500 GeV to 20 TeV [@Daum; @hegra_MC].
One of the important issues for future detectors is the choice of the energy region based on two principal arguments: (a) astrophysical significance (goals) and (b) the experimental feasibility/reliability (cost). If one limits the energy region from above to a relatively modest energy threshold around 100 GeV, then the performance of IACT arrays and their practical implementation can be predicted with confidence. In practice, an energy threshold of $\sim 100$ GeV can be achieved by a stereoscopic system of IACTs consisting of 10 m diameter class optical mirrors and equipped with high resolution cameras, based on conventional photo-multipliers (PMTs). Currently three such arrays are under development/construction in Australia (CANGAROO-3 [@CANG]), in Namibia (H.E.S.S. [@HESS]), and in Arizona (VERITAS [@VERITAS]). With their superior angular resolution of several arcminutes and energy flux sensitivity close to $10^{-13} \, \rm erg/cm^2 s$, these projects perfectly suit to the energy range from 50 GeV to 10 TeV, which, from the point of view of scientific motivations and the potential astronomical targets, can be considered as a spectral domain in its own right.
On the other hand, it is expected that the next generation major satellite $\gamma$-ray mission GLAST (see e.g. [@Bloom; @GehMich]), the successor of the EGRET instrument aboard the Compton Gamma Ray Observatory, will extend the exploration of the $\gamma$-ray sky up to 100 GeV. Thus the gap between space-based and ground-based $\gamma$-ray instruments will eventually disappear. It should be noticed, however, that in many cases this statement has a rather conditional or even symbolic character. Although at GeV energies GLAST will improve the EGRET sensitivity by almost two orders of magnitude, the capability of GLAST (and likely that of any post-GLAST space-based project) at energies well beyond 10 GeV will be quite limited because of the limited detection area.
This circumstance justifies recent activities to reduce the energy threshold of atmospheric Cherenkov detectors below 100 GeV. Currently two low threshold projects, CELESTE in France [@celest] and STACEE in USA [@stace], based on the concept of conversion of existing solar power plants into atmospheric Cherenkov telescopes, are in their final stage of realization. Although this technique allows in principle a reduction of the energy threshold down to 20-30 GeV (see e.g. Ref. [@Pare]), the challenge remains to reach an adequate detection efficiency of $\gamma$-rays at such low energies. Another approach has been proposed by the MAGIC collaboration with a single imaging Cherenkov telescope having a large, 17 m diameter reflector. Equipped with a standard PMT-based high resolution camera this telescope is expected to allow effective detection of $\gamma$-rays at energies above 30 GeV [@magic].
With some exceptions, the GeV sources ($E \geq 0.1\, \rm GeV$) are expected to be quite different from TeV sources ($E \geq 0.1 \, \rm TeV$). The proximity of the intermediate domain below 100 GeV to the energy range covered by EGRET suggests that many objects established as GeV emitters have a good chance to be detected also by the above mentioned ground-based instruments. This argument, however, cannot yet guarantee definite success. Indeed, although the two largest $\gamma$-ray source populations identified by EGRET, radiopulsars and distant AGN, do not show a significant steepening or cutoff up to 10 GeV, the theoretical studies of production and absorption conditions in these objects, as well as rather general phenomenological considerations predict cutoffs in the energy spectra around 10 GeV or less. In addition, for any reasonable model of the diffuse extragalactic cosmic background radiation, we should expect sharp cutoffs in the spectra of distant extragalactic objects with redshift $z \sim 1$ at energies as low as 50 GeV (see e.g. [@primack]). This implies that for the study of cosmologically distant sources, like the GeV blazars discovered by EGRET, or Gamma Ray Bursts (GRBs), the energy threshold of the detectors should be less than 10 GeV at which energy the Universe is most likely transparent up to at least $z \simeq 3$. An instrument like GLAST, operating effectively in the 0.1 to 10 GeV energy region, nicely suits this task. In particular, it is expected that the number of AGN that GLAST will detect could exceed several thousands [@GehMich]. At the same time, the relatively small detection area of GLAST, $A_{\rm eff} \simeq 0.8 \, \rm m^2$, limits the potential of this instrument for detailed studies of the temporal and spectral characteristics of highly variable $\gamma$-ray sources like blazars, which have variability timescales less than a few hours, or of solitary events like GRBs with a duration of $10^{-2}$ to $10^3$ seconds. In this regard, GLAST can hardly match the performance of current X-ray detectors that have similar detection areas but operate in a regime of photon fluxes that exceed the fluxes of MeV/GeV $\gamma$-rays by many orders of magnitude.
The idea of a “Gamma-ray timing explorer” to study transient $\gamma$-ray phenomena with an adequate photon detection rate motivated, to a large extent, the present investigation. It concerns the possible extension of the domain of ground-based Cherenkov technique with its huge detection area of $10^4$ to $10^5 \, \rm m^2$ down to energies of several GeV. We shall argue that such a goal could be best achieved by stereoscopic systems of several large, 20 m class imaging atmospheric Cherenkov telescopes located at very high, $H \sim 5 \, \rm km$, mountain altitude.
Concept of an IACT array with 5 GeV threshold
=============================================
The concept of stereo imaging is based on the simultaneous detection of a single air shower in different projections by at least two telescopes, separated at a distance comparable with the “effective radius” $R_{\rm C} \sim 100 \, \rm m$ of the Cherenkov light pool. The stereoscopic approach allows [*(i)*]{} unambiguous and precise reconstruction of shower parameters on an [*event-by-event*]{} basis, [*(ii)*]{} superior rejection of hadronic showers, and [*(iii)*]{} effective suppression of the background light from different sources - the [*nigh sky background*]{} (N.S.B.), local muons, etc. [@array]. All these three advantages over single IACTs have been convincingly demonstrated at TeV energies by the HEGRA IACT system [@Daum; @hegra_MC].
Compared with single (“stand alone”) telescopes, which can adequately measure the shower inclination in the direction perpendicular to the plane containing the telescope axis, but poorly in the in-plane direction, the stereoscopic approach allows full reconstruction of the arrival direction of individual $\gamma$-ray showers. Apart from the good directional information, stereoscopic IACT systems make use of the fact that the Cherenkov images of a shower detected by several different, spatially separated telescopes, are only [*partially*]{} correlated. Therefore the stereoscopic measurements significantly improve the efficiency of rejection of hadronic (background) showers at both the hardware (trigger) and the software levels. The only disadvantage of the stereoscopic approach is a non-negligible loss in the detection rate because of the overlap of the shower detection areas of individual telescopes located from each other at distances $\leq 2R_{\rm C}$. However, this loss of statistics is compensated, especially for steep spectra of primary $\gamma$-rays, by a significant reduction of the energy threshold of the telescopes operating in coincidence mode.
At $\gamma$-ray energies above 100 GeV stereoscopic IACT arrays do provide an excellent angular resolution of about $0.1^{\circ}$ or less, and a “gamma/hadron” separation efficiency (including the hadron rejection at the trigger level) of 1000:1. This improves the flux sensitivity dramatically compared with the sensitivity of single telescopes. The efficiency of the imaging technique is somewhat lower at energies below 100 GeV. In particular, the Cherenkov images become less elongated and less regular. In practice this introduces significant uncertainties in the reconstruction of image parameters. Even so, below it will be shown that the performance of the stereoscopic imaging remains adequately high even below 10 GeV.
The [*effective energy threshold*]{} of IACTs is basically determined by two conditions: [*(i)*]{} the number of photoelectrons in the image should be sufficient for an appropriate image analysis; typically, “good imaging” requires $n_{\rm ph.e.}^{\min} \sim 50-100$ electrons[^1], [*(ii)*]{} the accidental trigger rate introduced by the N.S.B. should not exceed the detection rate of $\gamma$-rays. For an ideal IACT the first condition in principle should dominate over the second, technical condition.
The number of photoelectrons detected by the imager - a multi-pixel camera placed in the focal plane of the mirror - depends on (i) the telescope photo-electron response (or aperture), $S_{\rm ph.e.}=S_{\rm mir} \cdot \xi_{\rm ph \rightarrow e}$ (where $S_{\rm mir}$ is the geometrical area of the mirror, and $\xi_{\rm ph \rightarrow e}$ is the photon-to-photoelectron conversion factor), and (ii) the density of optical Cherenkov photons $\rho(R,E)$ produced at the typical distance $R \sim 100 \, \rm m$ from the axis of the shower initiated by a primary $\gamma$-ray of energy $E$.
The density of the Cherenkov light at an elevation of $H=5$ km a.s.l. within 100 m of the shower core produced by a primary $\gamma$-ray photon of energy $E=5 \, \rm GeV$ is close to $1 \, \rm photon/m^2$. Correspondingly, the number of photoelectrons detected by a telescope with aperture $S_{\rm ph.e.}$ is equal to $n_{\rm min} \approx 1 \ S_{\rm ph.e.}$. Thus, for reduction of the energy threshold down to 5 GeV, the telescope aperture $S_{\rm ph.e.}$ should be as large as $50 \, \rm m^2$, assuming that the minimum number of electrons required for an image analysis $n_{\rm ph.e.}^{\rm min} \sim 50$. Detailed Monte-Carlo calculations presented below confirm that the energy threshold of such an instrument, determined as the energy at which the differential $\gamma$-ray detection rate reaches to its maximum, could indeed be as low as 5 GeV. For conventional aluminized optical mirrors and PMT-based cameras with typical conversion factor $\xi_{\rm ph \rightarrow e} \sim 0.15-0.2$, $S_{\rm ph.e}=50 \, \rm m^2$ would require a large optical reflector of approximately 20 m diameter. With the successful development of novel, fast (nsec) detectors of optical radiation with a quantum efficiency exceeding $50 \, \%$, the energy threshold of the telescopes could be pushed further down to 2 or 3 GeV, which is an absolute limit determined by the minimum energy of secondary electrons capable of producing Cherenkov light in the upper atmosphere.
The Cherenkov light density increases monotonically with elevation. Thus, for the given telescope configuration, the installation of an IACT array at very high altitude would allow a straightforward and unconditional reduction of the energy threshold. At the same time the increase of the telescope aperture leads not only to a proportional increase of number of photoelectrons registered per shower, but also to an increase of the accidental rate caused by the N.S.B. Thus, a significant gain from large aperture telescopes, with the ultimate goal to operate the telescopes in the linear regime, $E_{\rm th} \sim 1/S_{\rm ph.e}$, can be achieved only through effective suppression of the N.S.B. Otherwise the reduction of the energy threshold would be rather slow, $E_{\rm th} \sim 1/S_{\rm ph.e}^{1/2}$, and therefore would be difficult to justify economically.
In the imaging technique, the accidental events introduced by the N.S.B are suppressed by a trigger condition that requires signals above a threshold $q_0$ (photoelectrons) in $m \geq 2$ adjacent pixels. A smaller pixel size not only reduces significantly the noise level due to the N.S.B., but it allows higher trigger multiplicity as well. At the same time the optimal size of the pixel is determined by the condition of “good imaging”, which implies that the area which is covered by the minimum number of pixels (typically 10) used in the image analysis, should not exceed the characteristic size of the image. At $E \sim 10 \, \rm GeV$, the image area of electromagnetic showers is less than $0.2$ square degree. Therefore at such low energies high resolution cameras with a pixel size close to $0.1^{\circ}$ provide an adequate imaging quality. Although such small pixels allow also significant reduction of the N.S.B., it appears that for a single IACT with aperture $50 \, \rm m^2$, even a pixel size of $0.1^{\circ}$ is still not sufficient to operate the telescope at the minimum possible energy threshold, i.e. in the regime when the detection threshold is determined by the Cherenkov light amplitudes rather than by the N.S.B. noise. Further reduction of the pixel size makes the design of the imager with many thousands of channels technically very difficult. A solution of this dilemma is the stereoscopic mode of observations which offers a more feasible and economic approach. It requires simultaneous detection of a shower by at least two telescopes. Because of the flat lateral distribution of the Cherenkov radiation from electromagnetic showers, this requirement does not effect the $\gamma$-ray detection efficiency (if the distance between telescopes does not significantly exceed 100 m), but significantly reduces the accidental rate caused by the N.S.B. Below we will indeed show that the suppression of the N.S.B. by a stereoscopic system of IACTs with camera pixel size $\sim 0.1^\circ$ is sufficient to operate the $50 \, \rm m^2$ aperture telescopes in the energy regime below 10 GeV.
The arrangement of an IACT array, in particular the number of telescopes, and the spacing between them, can be understood from the following simple considerations. The Cherenkov light pool on the ground produced by primary $\gamma$-rays of energies $E \geq 100 \, \rm GeV$ has a flat radial distribution with a pool radius of approximately 100 m. Hence the optimal spacing between the telescopes should be of the order of 100 m. A significantly smaller spacing reduces not only the detection area, but also the quality of images. A spacing of significantly more than 100 m reduces the coincidence rate dramatically, especially at low energies, and thus increases the energy threshold. The special investigation carried out for 100 GeV class telescopes [@array; @optim], generally confirms this simple conclusion. We expect that this should be the case also for the sub-10 GeV array, although detailed Monte-Carlo simulations are needed for optimization of the arrangement of reflectors in this array. Here we consider a reasonable arrangement with a spacing of 100 m. The question of optimization of the spacing will be discussed in a separate paper. The minimum number of telescopes is determined from the consideration that at least 3 stereoscopic views are needed to reconstruct the shower parameters reliably [@array]. For clarity, we assume here a square baseline with telescopes at the four corners, placed at an altitude $H$ of 5 km a.s.l. In order to have more homogeneous coverage of distances of impact parameters, especially for showers with cores outside the square, we assume in addition a further (5th) telescope in the center of the square. The main objective of this paper is to study the basic performance of a 5 GeV IACT array placed at very high altitude. The optimization of the arrangement of the array is outside of the scope of this paper, but we believe that the suggested layout is not far from the optimum design.
Basic characteristics of Cherenkov radiation induced by sub-10 GeV gamma-rays
=============================================================================
The results discussed in this section have been obtained by using the ALTAI code [@altai] that simulates electromagnetic and hadronic showers and their Cherenkov radiation in the Earth’s atmosphere. This code has been used before for calculations of the performance of the HEGRA system of IACTs [@hegra_MC], as well as for studies of the expected characteristics of the new generation “100 GeV - threshold” IACT arrays [@array] that are the basis of the H.E.S.S. (High Energy Stereoscopic System) project in Namibia [@HESS]. The predictions for the HEGRA IACT system have been thoroughly checked using the detected hadronic (background) showers produced by cosmic rays with a relatively well known energy spectrum and mass composition [@hegra_MC; @proton_paper]. More importantly, the calculations of the characteristics of the instrument for primary $\gamma$-rays, in particular the lateral distribution of the Cherenkov light, which determines the $\gamma$-ray detection area, the angular resolution, and the gamma/hadron separation efficiency, have been experimentally confirmed by the HEGRA collaboration using $\gamma$-ray data obtained during the active state of Mkn 501 in 1997 [@hegra_lat; @mkn501]. This extraordinary high state of the source, with a duration of several months, resulted in approximately 40,000 $\gamma$-rays in the energy interval from 500 GeV to 20 TeV, detected by the HEGRA IACT system under almost background-free conditions [@mkn501].
In this paper we are interested in the possibility of detecting primary $\gamma$-rays over an interval of very low energies from several GeV to 100 GeV, by a system of IACTs located at $H \sim 5 \, \rm km$ a.s.l. Both conditions are rather extreme, and exhibit features which differ significantly from traditional studies of air showers. In particular, at energies of primary $\gamma$-rays below 10 GeV, we deal with Cherenkov radiation from only a handful of first generation electrons, while at TeV energies the Cherenkov radiation is contributed by a large number of electrons produced during the full cascade development. At such low energies we have therefore to expect more fluctuations of the parameters characterizing the showers and their Cherenkov radiation. In particular, the Cherenkov images of sub-10 GeV showers are expected to have a less regular shape, as well as to be more strongly affected by the geomagnetic field, as compared with the showers in the $\geq 100 \, \rm GeV$ energy region. Also, both the very low energy domain and the very high elevation of the location of the suggested IACT array imply not only better transparency conditions for the Cherenkov radiation, but also a non-negligible reduction of the light produced above the telescopes from hadronic showers. All these effects add new features to the characteristics of the Cherenkov radiation of air showers.
Lateral and longitudinal development of showers
-----------------------------------------------
The lateral distribution of Cherenkov radiation from air showers, induced by $\gamma$-rays of energy 10 GeV and by protons of energy 100 GeV, are shown in Fig. 1 at 3 different observational levels - $H=$2.2 km, 3.5 km and 5 km. It is seen that within the radius of 100 m from the shower core which typically determines the detection area of IACTs, the density of Cherenkov radiation from $\gamma$-ray induced showers has [*(i)*]{} a rather flat lateral distribution, and [*(ii)*]{} increases significantly with elevation above sea level. In particular, the rise from 2.2 km to 5 km results in an increase of the density of the Cherenkov light by a factor of 2 to 3 (Fig. 1a). This elevation effect is less pronounced for proton-induced showers ($\leq 50$ per cent at $r=100 \, \rm m$; see Fig. 1b). This implies that the choice of very high elevation for the operation of telescopes would allow a significant reduction of the energy threshold, by a factor of $\sim 2.5$, as well as a noticeable improvement of the background rejection of hadronic showers.
![Lateral distribution of the Cherenkov light density from [**(a)**]{} a primary $\gamma$-ray photon of energy 10 GeV (left) and [**(b)**]{} a cosmic ray proton of energy 100 GeV (right) at three different observation levels above sea level: 2.2 km, 3 km, and 5 km.](FIG1a.eps "fig:"){width="0.45\linewidth"} ![Lateral distribution of the Cherenkov light density from [**(a)**]{} a primary $\gamma$-ray photon of energy 10 GeV (left) and [**(b)**]{} a cosmic ray proton of energy 100 GeV (right) at three different observation levels above sea level: 2.2 km, 3 km, and 5 km.](FIG1b.eps "fig:"){width="0.45\linewidth"}
The slow increase of the Cherenkov light density from hadronic showers with the rise of elevation is explained by the effect of deep penetration of the “$\pi^{\pm}$-branches” (sub-showers) of the cascade into the atmosphere (Fig. 2), which results in the production of a non-negligible amount of Cherenkov light below the observation level at $H \sim 5 \, \rm km$. This effect is quantitatively demonstrated in Fig. 3 where the longitudinal distributions of the Cherenkov light, produced by a 10 GeV $\gamma$-ray photon and a 100 GeV proton, are shown.
{width="0.75\linewidth"}
The low energy threshold itself already leads, in fact more strongly than the elevation effect, to a significant improvement of the gamma/hadron ratio. This effect, caused by the reduction of the yield of Cherenkov light by sub-100 GeV hadronic showers (see e.g. [@Pare]) is demonstrated in Fig. 4 for $H=$5 km a.s.l. The density of Cherenkov light in the electromagnetic showers is approximately proportional to the $\gamma$-ray energy down to $E \sim 10 \, \rm GeV$, the dependence becoming a bit stronger below 10 GeV (Fig. 4a). At the same time, the density of
{width="0.6\linewidth"}
Cherenkov light from hadronic showers drops significantly faster with reduction of the primary energy (Fig. 4b). In Fig. 5 we show the ratio of densities of the Cherenkov light $\rho^{(\gamma)}_{100}(E_\gamma)$ and $\rho^{(\rm p)}_{100}(E_{\rm p})$ produced by $\gamma$-rays and protons at 100 m distance from the shower core.
![Lateral distribution of the Cherenkov light density in [**(a)**]{} $\gamma$-ray (left) and [**(b)**]{} proton (right) induced showers at an altitude of 5 km a.s.l. The energies of primary particles are shown at the curves.](FIG4a.eps "fig:"){width="0.45\linewidth"} ![Lateral distribution of the Cherenkov light density in [**(a)**]{} $\gamma$-ray (left) and [**(b)**]{} proton (right) induced showers at an altitude of 5 km a.s.l. The energies of primary particles are shown at the curves.](FIG4b.eps "fig:"){width="0.45\linewidth"}
Dramatic rise of this ratio below 100 GeV implies a monotonic increase of the gap between the energies of electromagnetic and hadronic showers that produce the same amount of Cherenkov light. This effect is demonstrated in Fig. 6. The solid line corresponds to the relation between energies of $\gamma$-rays $E_\gamma$ and protons $E_{\rm p}$ defined from equation $\rho^{(\gamma)}_{100}(E_\gamma)=\rho^{(\rm p)}_{100}(E_{\rm p})$. For comparison, the line “$E_\gamma=E_{\rm p}$” is also shown. It is seen, in particular, that the gap between $E_\gamma$ and $E_{\rm p}$, which is close to $\approx 2$ at high energies (e.g. 1 TeV electromagnetic showers versus 2 TeV hadronic showers), becomes significantly larger at very low energies below 10 GeV (e.g. 2 GeV $\gamma$-ray events versus 30 GeV hadronic showers).
{width="0.48\linewidth"}
{width="0.48\linewidth"}
Field of view and pixel size
----------------------------
One of the principal parameters of the imaging camera is its field of view (FoV). At TeV energies the image centroids of showers detected at distances $R \sim 100 \, \rm m$ or beyond are shifted in the focal plane from the center of the camera by $\approx 1^{\circ}$. Therefore an inner region with diameter $ \sim 3^{\circ}$ provides high detection efficiency for $\gamma$-rays, and can be treated as optimal zone for the hardware trigger. On the other hand, the IACT technique requires that the camera should be larger than the trigger zone by about one degree, in order to avoid a distortion of the Cherenkov images because of a limited FoV; for lower $\gamma$-ray energies the shift of the image centroid decreases [@array]. This effect becomes significant especially at very low energies.
{width="0.7\linewidth"}
In particular at $E=10 \, \rm GeV$ the shift does not exceed $0.5^{\circ}$ (Fig. 7). Thus in the energy range $E \ll 100 \, \rm GeV$, the camera can be more compact compared with the conventional (FoV$\geq 4^{\circ}$) cameras designed for high energies. In fact, a FoV$\approx 3^{\circ}$ seems a reasonable for point-like or even moderately extended ($\leq 0.5^{\circ}$) $\gamma$-ray sources. This is an important circumstance which keeps the number of the channels of the imager within reasonable ($\leq 10^3$) limits, especially taking into account that at such low energies the pixel size must be small, $\approx 0.1^{\circ}$ or less, dictated by two the equally important conditions: effective suppression of the N.S.B., and high image quality (see Sec. 2).
Trigger integration gate
------------------------
Truncation of the Cherenkov pulse integration gate is another effective way to suppress the N.S.B. Ideally, the integration gate should be comparable with the duration of the time impulse of the Cherenkov light which for 10 GeV $\gamma$-rays, detected at 5 km a.s.l. altitude, is less than 5 ns (Fig. 8). The compression of the integration gate from conventional 20 ns to 5 ns reduces the average noise level - the number of photoelectrons produced by the N.S.B. - by a factor of 4, and correspondingly lowers the energy threshold significantly.
{width="0.6\linewidth"}
Note that both techniques, [*fast timing*]{} and [*small pixel size*]{}, have already been successfully employed by the CAT collaboration [@cat]. A further suppression of the N.S.B. can be effectively achieved in the stereoscopic approach by requiring simultaneous detection of showers by two or more telescopes. This technique has been convincingly demonstrated by the HEGRA collaboration [@Daum; @hegra_MC]. Below we will show that, even for telescopes with an aperture as large as $50 \, \rm m^2$, the combination of all three techniques - [*small pixel size*]{} ($\sim 0.1^{\circ}$), [*fast timing*]{} ($\leq 5 \rm \ ns$), and [*stereoscopy*]{} - allows effective operation of the IACT system in the energy regime as low as several GeV.
Effect of geomagnetic field
---------------------------
An important issue concerning the quality of Cherenkov images at very low energies
{width="0.6\linewidth"}
{width="0.4\linewidth"}
is connected with the deflection of the secondary (cascade) electrons in the geomagnetic field (see e.g. Ref. [@durham]). Generally, our results agree well with the conclusion of
{width="0.6\linewidth"}
Patterson and Hillas [@B_hillas] that down to $\gamma$-ray energies of 100 GeV the geomagnetic field has a rather small effect on the lateral distribution of the Cherenkov radiation as long as the perpendicular component of the magnetic field $B_{\perp}$ is less than 0.3 G. One cannot [*a priory*]{} exclude, however, that at very low energies the effect of the geomagnetic field is stronger. The main parameter characterizing the effect of the geomagnetic field is the average energy $\overline{E_{\rm e}}$ of the electrons radiating Cherenkov light at the shower maximum, rather than the energy of primary particles. In Fig. 9 we compare the parameter $\overline{E_{\rm e}}$ for showers produced by primary $\gamma$-rays at three energies - 10 GeV, 100 GeV, and 1 TeV. It is seen that, at any fixed depth, the average energy of electrons producing Cherenkov light is larger for cascades induced by higher energy $\gamma$-rays. However, the average electron energies become quite similar when we compare $\overline{E_{\rm e}}$ at the depths corresponding to the shower maxima for the given energy of the primary $\gamma$-ray photon. This interesting effect is explained by the fact that at low primary energies the maximum of an electromagnetic shower occurs at high altitudes where the density is low, and therefore the electrons need to be more energetic to produce Cherenkov light. In practice this implies that lowering the energy of $\gamma$-rays should not result in significant amplification of the geomagnetic effect. In Fig. 10 the density of the Cherenkov light from 10 GeV showers at the observation level are shown for three different values of the perpendicular component of the geomagnetic field - $B_{\perp}$=0, 0.5 G, 1 G. The split of the patterns (azimuthal asymmetry) produced by electrons and positrons due to their deflections in opposite directions becomes noticeable only for a very large perpendicular component of the field, $B_{\perp}$=1 G. More quantitatively, the effect of the geomagnetic field on the lateral distribution of Cherenkov photons from 10 GeV $\gamma$-rays is shown in Fig. 11. It is seen that for a reasonable field $B_{\perp} \leq 0.5 \, \rm G$, the effect is less than 15 per cent.
Detection areas
===============
For a given telescope configuration and arrangement of the IACT array, the energy threshold and the effective detection area of primary $\gamma$-rays are determined by the lateral and angular distributions of the Cherenkov light and the hardware trigger conditions.
{width="0.6\linewidth"}
The configuration of the IACT array assumed in this paper is the following: the system consisting of 5 IACTs is installed at the altitude $H=5 \, \rm km$ a.s.l., four telescopes are located at the corners, and one in the center of a square with a linear size $d=100 \, \rm m$; each telescope has an aperture $S=50 \, \rm m^2$, and is equipped with a $n=721$-channel camera of individual pixel size $0.12^{\circ}$. This comprises an effective field of view FoV$=3.2^{\circ}$. The hardware trigger is organized in a way that requires (1) signals above some critical threshold $q_0$ in $m$ adjacent pixels (“$m/n \geq q_0$”) in each individual telescope (the “local trigger”); (2) detection of a shower by at least 2 telescopes (the “system trigger”). For calculations of the threshold $q_0$ we assume a 5 ns trigger gate, and require that the accidental rate caused by the N.S.B. is less than 1 Hz, i.e. less than 1 per cent of the detection rate of cosmic ray electrons (see below).
{width="0.6\linewidth"}
In Fig. 12 we show the detection areas for $\gamma$-rays from 1.5 GeV to 100 GeV calculated for 3 local trigger conditions: $2/721 \geq$ 6 ph.-e., $3/721 \geq$ 5 ph.-e. $4/721 \geq$ 5 ph.-e., assuming a standard N.S.B. flux $F_{N.S.B.} \simeq 1.5 \times 10^{12} \, \rm ph/m^2 sr \ s$ (see e.g. [@Cawley]). It is seen that for a chosen pixel size of $0.12^{\circ}$, a multiplicity greater than two ($m > 2$) actually reduces the detection area, even though the requirement of higher multiplicity allows a lower trigger threshold $q_0$. Therefore all calculations presented below were performed for $m=2$.
In Fig. 13 the detection area of the IACT array for a somewhat higher threshold, $q_0=8$ ph.-e., is shown in order to demonstrate how sensitive the detection area is to the choice of $q_0$, especially at low, sub-10 GeV energies (compare curves 1 in Fig. 12 and Fig. 13). For comparison, we show also the overall detection area of 5 telescopes if they would be located at large, $\gg 100 \, \rm m$ distances from each other and operating independently in a stand alone mode, i.e. being not integrated in the system trigger. It is seen that at energies $E \geq 3 \, \rm GeV$ the array of independent telescopes has a larger detection area, but the difference compared with the array operating in the stereoscopic mode at energies $E \geq 5 \, \rm GeV$ does not exceed a factor of two. On the other hand, the analysis of images obtained in the stereoscopic mode provides much better suppression of the background caused by the cosmic ray protons and electrons which effectively results in a significant improvement of not only the quality of the data, but also of the flux sensitivity of the instrument.
At energies below 100 GeV, the cosmic ray background detected by the system of IACTs operating in the stereoscopic mode is essentially dominated by electromagnetic showers produced by cosmic ray electrons. This is seen in Fig. 14 where the detection rates of both cosmic ray protons and electrons are also shown. The strong dominance of the electronic component is explained by the combination of several effects, in particular [*(i)*]{} the large, up to a factor of 10, difference between the energies of electrons and protons producing the same amount of Cherenkov light, [*(ii)*]{} the high altitude of observations, [*(iii)*]{} the compact Cherenkov images of electromagnetic showers compared with hadronic showers, [*(ii)*]{} the noticeable increase (approximately $\propto E^{-0.5}$) of the electron-to-proton ratio of cosmic rays down to $E \sim 10 \, \rm GeV$.
In Fig.14 we show also the detection rate of $\gamma$-rays from a point source with a power-law spectrum ${\rm d} J_\gamma/{\rm d} E \sim E^{-2.5}$, and integral flux $J_\gamma(\geq 1 \, \rm GeV)=3 \times 10^{-7} \, \rm ph/cm^2 s$. The latter is somewhat larger than the fluxes of most of the EGRET sources [@Vela]. This implies that the $\gamma$-ray detection rate shown in Fig. 14 should be considered as an upper limit for “standard” EGRET sources. This curve still lies significantly below the rate of detection of cosmic ray electrons. However, for a point-like source the electron background can be reduced significantly if we select showers arriving from the direction of the $\gamma$-ray source.
{width="0.6\linewidth"}
The detection rates of electromagnetic showers within the detector’s energy-dependent Point Spread Function (PSF), given by the angle $\phi$, which is the half-angle of the cone around the source direction containing $67 \%$ of events at energy $E$, are determined as follows: $$R_\gamma=\frac{{\rm d}J}{{\rm d}E} \ A_{\rm eff}(E) \ \kappa_\gamma \ ,$$ and $$R_{\rm e}=\frac{{\rm d}J_{\rm e}}{{\rm d}E {\rm d}\Omega}
\ A_{\rm eff}(E) \Omega \ ,$$ for $\gamma$-rays and cosmic ray electrons, respectively, where ${\rm d}J/{\rm d}E$ is the differential flux of $\gamma$-rays from a point source, ${\rm d}J_{\rm e}/{\rm d}E {\rm d}\Omega$ is the differential flux of cosmic ray electrons per solid angle, $A_{\rm eff}$ is the detection area for electromagnetic showers, $\Omega=2 \pi \ (1- \cos \phi)$, and $\kappa_\gamma=0.67$ is (by definition) the $\gamma$-ray acceptance.
{width="0.6\linewidth"}
{width="0.6\linewidth"}
For the chosen configuration of the IACT array, the Monte-Carlo calculations of the effective detection area $A_{\rm eff}$ (Fig. 15) in the interval from 1.5 GeV to 100 GeV can be presented in the following form $$A_{\rm eff}=8.5 E^{5.2} [1+(E/5 \, \rm GeV)^{4.7}]^{-1} \, \rm m^2 \, ,$$ where E is the energy of a $\gamma$-ray or electron in units of GeV.
![The [alpha]{} distributions for a single telescope.](FIG15.eps){width="0.6\linewidth"}
Cut, deg $\kappa_\gamma$ $\kappa_{CR}$ Q-factor
---------- ----------------- --------------- ----------
5 0.25 0.0504 1.11
10 0.42 0.0972 1.35
15 0.54 0.1439 1.42
20 0.62 0.1907 1.42
25 0.69 0.2348 1.42
30 0.73 0.2868 1.36
: Efficiency of the [alpha]{}-cut for a single “low energy” telescope.
This presentation shows a strong energy-dependence of the detection area at energies below 10 GeV, $A_{\rm eff}(E) \propto E^{5.2}$, but at higher energies it gradually turns into a slow increase with energy, $A_{\rm eff}(E) \propto E^{0.5}$ Close to 1 TeV the detection area actually becomes constant, in essence because of the limited field of the view ($\approx 3^{\circ}$) of the camera.
The stereoscopic approach allows the determination of the arrival direction of primary $\gamma$-rays on an event-by-event basis. The determination of the arrival direction of primary $\gamma$-rays is described in ref.[@array], and its practical implementation in the case of the HEGRA IACT system can be found in ref. [@hegra_MC; @wh_utah]. In Fig. 16 we show the Monte-Carlo calculations of the PSF which can be approximated in the simple form $$\phi=0.8 (E/1 \, \rm GeV)^{-0.4} \, \rm degree \, .$$
Despite the small pixel size $\sim 0.1^{\circ}$, the angular resolution of the 5 GeV IACT array at energies $E \ll 100 \, \rm GeV$ is significantly poorer than the resolution of “100 GeV” threshold instruments, which could be as good as $0.1^{\circ}$, even for a larger pixel-size of about $0.25^{\circ}$ [@array]. Since for both categories of instruments the number of photoelectrons, or the so-called size of the image, from a detected $\gamma$-ray photon are comparable ($\sim 100 (E/E_{\rm th})$ photoelectrons), the lower performance of the IACT technique is rather an intrinsic feature of Cherenkov images at very low energies (the images are less elongated and less regular). This effect can be seen also in the so-called [alpha]{}-distribution of images in a single telescope, where the [alpha]{} parameter is indicative of the orientation of the image in the camera (see e.g. ref.[@Fegan]): while at TeV energies most of $\gamma$-rays from a point source have an angle [alpha]{} less than 5-8 deg [@Fegan], the [alpha]{} distribution is significantly broader in the energy region around 10 GeV (see Fig. 17). The suppression of the cosmic ray background at such low energies becomes correspondingly less effective. The characteristic values of the acceptance of the isotropic cosmic ray showers $\kappa_{\rm CR}$, and the point-source $\gamma$-ray showers $\kappa_\gamma$, as well as the so-called Q-factor, $Q=\kappa_\gamma/\kappa_{\rm CR}^{1/2}$, which characterizes the improvement of the signal-to-noise ratio after application of the image cuts, are shown in Table 1. It is seen that the best improvement of the signal-to-noise ratio is achieved for an [alpha]{} cut at $\sim 20-25^{\circ}$ which gives a rather modest Q-factor, $Q \simeq 1.4$. In contrast, at 10 GeV the stereoscopic measurements allow the determination of the shower direction with an accuracy of $0.3^{\circ}$ (Fig. 16). Therefore the $\gamma$-ray signal could be improved by a factor of $Q=\kappa_\gamma/\kappa_{\rm CR}^{1/2}=
\kappa_\gamma (\Psi/2 \phi) \approx 3.35$, where the efficiency of the rejection showers from cosmic ray electrons $\kappa_{\rm CR} \approx (2 \ \phi/\Psi)^2$ is determined by the FoV of the imager, $\Psi \simeq 3^{\circ}$, and the angular resolution $\phi \simeq 0.3^{\circ}$. Thus, despite the smaller (by a factor of two) detection area of the array operating in the stereoscopic mode compared with the overall area of 5 independent IACTs (see Fig. 13), the stereoscopic array would have at least by a factor of 2 better sensitivity, even disregarding other advantages of the stereoscopic approach, in particular, the complete removal of the hadronic background.
Detection rates and the energy threshold
========================================
The differential detection rates of $\gamma$-rays from a point source, calculated for the collection area given by Eq. (3) and assuming pure power-law spectrum in the form $$\frac{{\rm d}J}{{\rm d}E}=10^{-7} \ (E/1 \, \rm GeV)^{-\alpha} \
\rm cm^{-2} s^{-1} GeV^{-1} \ ,$$ are shown in Fig. 18a. Eq. (5) implies that, independent of the spectral index $\alpha$, the differential $\gamma$-ray flux is normalized at 1 GeV to $10^{-7} \ \rm cm^{-2} s^{-1} GeV^{-1}$. The latter corresponds to the typical flux of “standard” EGRET sources.
![Differential detection rates of $\gamma$-rays and cosmic-ray electrons within the cone determined by the PSF of the IACT array, calculated for two types of $\gamma$-ray spectra represented by [**(a)**]{} Eq.(5) (left) and [**(b)**]{} Eq.(6) (right).](FIG17a.eps "fig:"){width="0.45\linewidth"} ![Differential detection rates of $\gamma$-rays and cosmic-ray electrons within the cone determined by the PSF of the IACT array, calculated for two types of $\gamma$-ray spectra represented by [**(a)**]{} Eq.(5) (left) and [**(b)**]{} Eq.(6) (right).](FIG17b.eps "fig:"){width="0.45\linewidth"}
In Fig. 18b we show the differential detection rates of $\gamma$-rays with a hard power-law spectrum with $\alpha=2$ and an exponential cutoff at $E_0$: $$\frac{{\rm d}J}{{\rm d}E} \propto E^{-2} \ \exp(-E/E_0),$$ for 4 different values of $E_0$, and assuming the same absolute flux normalization at 1 GeV as in Fig. 18a.
![The flux of cosmic ray electrons. The compilation of the experimental fluxes is taken from Ref. [@taira; @heat]. The solid line corresponds to the fit represented by Eq.(7). The dashed line correspond to the function ${\rm d} J_{\rm e}/{\rm d}E \propto E^{-0.4} [1+(E/3.9 \, \rm GeV)]^{-3.1}$ which provides a better fit to the data at energies above 100 GeV.](FIG16.eps){width="0.6\linewidth"}
It is seen from Figs.18a and 18b that for a large variety of $\gamma$-ray spectra the peak of the detection rate appears in a rather narrow band between 4 and 6 GeV. Now, defining the energy threshold as [*the energy at which the differential $\gamma$-ray detection rate reaches to its maximum*]{}, we may conclude that the suggested IACT array has an effective energy threshold of about 5 GeV.
In Fig. 18a,b we show also the differential detection rate of cosmic ray electrons within the cone limited by the PSF of the instrument given by Eq. (4). The energy spectrum of cosmic ray electrons is shown in Fig. 19. At energies above 10 GeV the differential spectrum is very steep with a power-law index $\alpha_{\rm e} \sim 3.2$. Below 10 GeV it becomes flatter. Within the uncertainties of the measured fluxes, the electron spectrum can be approximated in the entire region from several GeV to 1 TeV by the following function shown by the solid line in Fig. 19 : $$\frac{{\rm d}J_{\rm e}}{{\rm d}E {\rm d}\Omega}=1.36 \times 10^{-7} \
E^{-1} [1+(E/5 \, \rm GeV)^{2.2}]^{-1} \, m^{-2} s^{-1} sr^{-1} GeV^{-1} \ .$$
In Fig. 20 we show integral detection rates $R_\gamma(\geq \rm E)$ for power-law $\gamma$-ray spectra represented by Eq.(5). It is seen that for a relatively flat $\gamma$-ray spectrum with $\alpha_\gamma \sim 2$ and for an integral flux above 1 GeV of $10^{-7} \, \rm ph/cm^2 s$ (this approximately corresponds to the total, i.e. pulsed plus unpulsed, flux from the Crab), the detection rate of $\gamma$-rays from the EGRET sources can be as high as 6 events per second, against the cosmic-ray background rate of about 25 events per sec caused by cosmic-ray electrons. This implies that an observation time of approximately 20-30 sec would be sufficient to detect a statistically significant signal from such a source. Remarkably, for the brightest persistent $\gamma$-ray source, the Vela pulsar with photon index $\alpha_\gamma \sim 1.7$ from 100 MeV to 10 GeV, and the integral flux $J_\gamma(\geq \rm 1 \ GeV) \approx 1.5 \times 10^{-6} \, \rm ph/cm^2 s$, the detection rates would exceed 100 events per 1 sec. Thus a statistically significant signal from the source could be obtained during an observation time less than 1 sec ! For the given normalization of the differential flux at 1 GeV, $10^{-7} \, \rm ph/cm^2 s \ GeV$, the detection of sources with steep $\gamma$-ray spectra would require significantly longer exposure. Even so, the time needed for detection of sources with very steep power-law spectra with an index $\alpha_\gamma \sim 3$ (like curve 4 in Fig. 18a), or with a sharp, e.g. exponential cutoff at a few GeV (like the curve 4 in Fig. 18b), does not significantly exceed 1 h.
{width="0.5\linewidth"}
Flux Sensitivity
================
The curves in Fig. 14 correspond to the detection rates before the image analysis. Remarkably, even after such effective rejection of hadronic showers at the [*trigger*]{} level, there still remains room for further suppression of the background from cosmic ray protons and nuclei by analyzing the shapes of the Cherenkov images of the detected showers. In Fig. 21 we show the so-called mean-scaled [width]{} parameter distribution of showers which have already passed the hardware trigger condition. This parameter represents the mean value of the [width]{} parameter measured by all telescopes and normalized to the impact distances and the image amplitudes [@hegra_MC]. It is seen that the distributions of the electromagnetic and hadronic showers are rather well separated. The efficiencies of the acceptance of both type of showers for different mean scaled [width]{} cuts are presented in Table 2. We see that even very loose cuts at the level of $<w>=1.3$ provide suppression of the hadronic showers by a factor of 5, while the $\gamma$-ray acceptance can be as high as 90 per cent. This implies that after such a loose cut, which practically does not reduce the $\gamma$-ray (or CR electron) statistics, we may push the detection rates of cosmic ray protons and nuclei further down, and thus make $\gamma$-ray detection in the entire energy region below 100 GeV essentially free from hadronic background.
![Distribution of the mean scaled [width]{} parameter for electromagnetic and hadronic showers.](FIG19.eps){width="0.5\linewidth"}
$< \tilde w>$ 0.8 0.9 1.0 1.1 1.2 1.3
--------------- ------- ------- ------ ------ ------ ------
$k_\gamma$ 0.28 0.45 0.60 0.74 0.83 0.90
$k_{cr}$ 0.025 0.044 0.07 0.11 0.15 0.20
Q-factor 1.8 2.1 2.3 2.2 2.1 2.0
: Efficiency of the cosmic ray rejection using the mean scaled [width]{} cut.
This makes the calculations for the [*differential*]{} flux sensitivity of the instrument straightforward and simple and, more importantly, there is no need to specify the spectrum of primary $\gamma$-rays. Indeed, because the showers produced by electrons are very similar to $\gamma$-ray showers[^2], the condition of detection of a $\gamma$-ray signal with statistical significance $m$-sigma in the energy interval $[E-\Delta E, E+\Delta E]$, $N(E) \Delta E=m \sqrt{N_{\rm e} \Delta E}$, provided that the number of detected $\gamma$-rays $N_{\rm min}=J(E) \ 2 \Delta E \ A_{\rm eff}(E) \ T \geq 10$, gives the minimum detectable differential flux for the observation time $T$: $$J_{\rm min}(E) \approx \frac{m}{\kappa_\gamma}
\frac{\phi}{(A_{\rm eff}(E) \ T)^{1/2}}
\sqrt{\frac{2 \pi J_{\rm e}(E)}{E}}$$ where it is assumed that $\Delta E=E/4$, which corresponds to a rather conservative 25 per cent accuracy of reconstruction of the energy of the primary electron or $\gamma$-ray photon. The results of calculations for the so-called spectral energy distribution (SED), $E^2 \ J(E)$, based on Eqs.(3),(4) and (7) requiring 3-sigma detection ($m=3$) at each energy $E$, are presented in Fig. 22 for 2 different observation times, $T=$1 h and 25 h.
{width="0.5\linewidth"}
In the same figure we present also the power-low fluxes of $\gamma$-rays represented by Eq.(6). It is seen that 1 hour observations by the IACT array would be sufficient to detect a statistically significant signal from a “standard” EGRET source even at the presence of an exponential cutoff in the $\gamma$-ray spectrum as low as 3 GeV. In the case of a cutoff at 10 GeV or higher energies, the detection time ($t \propto 1/J^2$) could be reduced to $\leq 1$ min.
{width="0.5\linewidth"}
For comparison in Fig. 22 we show the expected sensitivity of GLAST for an observation time $T=30 \, \rm days$. Since at energies above several GeV GLAST will operate at almost background free conditions (for point-like sources), the flux sensitivity is determined by the photon statistics, $N_{\rm min}=10$ (see e.g. [@Bloom; @GehMich]). Note that at energies above 5-10 GeV, 1 hour observation time by the IACT array could provide better sensitivity that the minimum detectable fluxes achievable by GLAST during 1 month of continuous observations. Moreover, even very short observations by the IACT array can give unusually rich (for $\gamma$-ray astronomical standards) photon statistics over the whole energy region from few GeV to 100 GeV; the number of detected $\gamma$-rays exceeds 100 at each energy interval $E \pm E/4$ (see Fig. 23). This would guarantee an appropriate $\gamma$-ray spectroscopy with energy resolution of about 20-25 per cent below 10 GeV, and better than 15 per cent at higher energies.
The flux sensitivity shown in Fig. 22 is obtained under a very robust and to a large extent non-standard condition which requires detection of a signal with at least 3-sigma significance in [*each*]{} energy band with width $E/2$ centered on $E$, provided that the number of detected $\gamma$-rays in this band exceeds 10. Note that this definition of sensitivity does not require any knowledge about the shape of the spectrum of the primary $\gamma$-rays.
{width="0.5\linewidth"}
In Fig.24 we present the flux sensitivities determined in a more traditional way, namely requiring 5-sigma detection of $\gamma$-rays above the given energy, $J_{\rm min}(\geq \rm E)$. This definition of the integral flux sensitivity obviously requires an assumption about the shape of the energy spectrum. The curves shown in Fig. 24 are calculated for $T=25 \, \rm h$ observation time, assuming power-law spectra of $\gamma$-rays with photon indices $\alpha_\gamma=$1.5 (dotted curve), 2 (dot-dashed curve), and 2.5 (dashed curve). The expected GLAST sensitivity shown by the solid curve corresponds to 1 year of continuous observations of the source.
Discussion
==========
Results presented in this paper show that a stereoscopic array of large, 20 m diameter class imaging Cherenkov telescopes installed at very high mountain altitudes could effectively enter into the domain of satellite-borne $\gamma$-ray astronomy. A [**5**]{} GeV energy threshold array of IACTs [**at**]{} [**5**]{} km a.s.l. - hereafter 5@5 - could provide a deeper probe of $\gamma$-ray sources compared with GLAST - the most powerful current satellite-borne $\gamma$-ray project. However, the scientific goals of these instruments are essentially different. While GLAST with its almost $2 \pi$ steradian field of view can provide very effective [*simultaneous*]{} monitoring of a very large number (hundreds or even thousands) quasi-stable $\gamma$-ray sources, as well as a study the galactic and extragalactic components of the diffuse $\gamma$-ray background radiation, 5@5 has an obvious advantage for the search and study of highly variable or transient $\gamma$-ray sources. The flux sensitivity of this instrument at 5 GeV of about $2 \times 10^{-11} \, \rm erg/cm^2 s$ (see Fig. 22) would allow the detection of any $\gamma$-ray flare with apparent luminosity $2 \times 10^{39} \, \rm (d/1 \ Mpc)^2$, lasting only 1 h, where $d$ is the distance to the source. Of special interest are the [*gamma-ray blazars*]{} detected by EGRET (see e.g. [@Vela]). A detailed study of the time structure of $\gamma$-radiation for these highly variable objects on timescales of several minutes by 5@5 would provide unique information about the relativistic non-thermal processes in astrophysical jets. An effective operation of this instrument in the sub-10 GeV regime guarantees detection of $\gamma$-rays arriving from cosmological distances up to $z \sim 3$ or so, for which the intergalactic medium becomes almost transparent. The dynamic range from several GeV to 100 GeV would allow important [*cosmological measurements*]{}, in particular a study of the diffuse ultraviolet extragalactic background by detecting intergalactic $\gamma-\gamma$ absorption features in the spectra of $\gamma$-rays below 100 GeV. The confusion problem (spectral cutoff due to the internal or extragalactic absorption ?) at such redshifts can probably be overcome by simultaneous observations at optical and X-ray wavelengths.
{width="0.5\linewidth"}
5@5 can also be effectively used for the study of [*galactic transient sources*]{}, in particular for the detection of short time ($\leq 1 \, \rm day$) $\gamma$-ray activity expected during synchrotron radio flares of microquasars.
And finally, 5@5 can serve as a very powerful instrument for a study of the phenomenon of [*Gamma Ray Bursts*]{} (GRBs). If the spectra of GRBs extend to high energies without abrupt cutoffs up to several GeV, which is the case at least for some of GRBs [@Hurley], then the sensitivity of 5@5 would allow very detailed studies of the spectral and temporal features of GRBs in this extremely important energy region. In Fig. 25 we show the minimum time $t_{\rm min}$ required for detection of GeV $\gamma$-ray flares with a given energy flux $f(E)$ at 4 different $\gamma$-ray energies: 2 GeV, 5 GeV, 10 GeV, and 50 GeV. The calculations correspond to the 3-sigma signal at each energy $E$ within the interval $[E-0.25 E, E+0.25 E]$, provided that the number of detected $\gamma$-rays exceeds 10. In the regime of low fluxes, typically $f \leq 10^{-9} \, \rm erg/cm^2 s$, the $\gamma$-rays are detected in the presence of the heavy background induced by cosmic ray electrons. Therefore $t_{\rm min} \propto f^{-2}$. For fluxes larger than $10^{-9} \, \rm erg/cm^2 s$, the detection occurs under almost background-free conditions, and therefore $t_{\rm min} \propto f^{-1}$.
The results shown in Fig. 25 demonstrate the capability of 5@5 for detection of GeV counterparts of GRBs. The detection of $\geq 5$ GeV episodic events with typical GRB fluxes between $10^{-8}$ and $10^{-6} \, \rm erg/cm^2 s$ would require only $0.1 \, \rm s$ observation time. Thus it would be possible to monitor the spectral evolution of the source with a typical GRB duration from several second to 100 seconds. Remarkably, even for fluxes as low as $10^{-10} \, \rm erg/cm^2 s$, the required exposure time does not exceed 100 sec. This implies that 5@5 could serve as a unique tool to study GRBs in the late stages of evolution, i.e. during the afterglows.
5@5 is a detector with a small field of view. Therefore it requires special strategies for the search and study of multi-GeV $\gamma$-ray emitters. The proximity of this energy region to the energy range covered by EGRET suggests that almost all (more than 300) EGRET sources should also be detected by 5@5. The typical observation time for detection of a “standard” EGRET source would not exceed 1 hour, even if the spectrum of $\gamma$-rays cuts off at energies of several GeV. The full overlap of the energy range of this instrument with the energy domain of GLAST would make the latter a “the best guide” for developing a strategy for the study of persistent galactic and extragalactic objects. Generally, all sources seen by GLAST can be potential targets for observations with the 5@5. These observations with very large photon statistics - not achievable by GLAST - could provide detailed studies of the spectral and temporal features of $\gamma$-ray sources in the multi-GeV region. For highly variable objects like blazars or galactic sources with relativistic jets, a multi-wavelength approach including observations with radio, optical and X-ray detectors would be very important. These observations would not only inform about the pre-flaring or flaring states of the sources, but also would provide complementary information for understanding and comprehensive modeling of the physical processes in these objects.
A special strategy should be developed for the search for GeV radiation from GRBs during and after the main event. Apparently, prompt information (within 10 sec or so) from the new generation GRB detectors like SWIFT (and possibly also from GLAST) would be needed, containing the angular coordinates of an event with an accuracy better than $1^{\circ}$. It would be very worthwhile to have also a nearby ground-based prompt optical telescope like ROTSE [@rotse]. In their turn, the telescopes of 5@5 should be rather fast in order to be directed to the source not later than 1 minute after receiving the alarm from these detectors.
Because of effective rejection of hadronic showers by 5@5, the cosmic-ray background below 100 GeV is dominated by the showers from cosmic-ray electrons. This component of electromagnetic showers remains a part of the background which can be hardly removed, and thus it is the most serious limiting factor of flux sensitivities, especially for extended sources. On the other hand, these electromagnetic showers with a known flux and spectrum of cosmic ray electrons, measured up to energies 1 TeV, can be used for absolute [*energy calibration*]{} of the instrument.
Energy calibration with the aid of the cosmic-ray electrons provides a unique tool for the [*continuous*]{} (on-line) control of the characteristics of the detector (e.g. the energy threshold, the detection area, [*etc.*]{}) during the observations. For example, a 10 min exposure will be enough for the detection of hundreds of electrons at any energy $E$ within $\pm E/4$ in the entire dynamical region from few GeV to 100 GeV (see Fig. 23). It is difficult to overestimate the significance of such calibration and control, especially for the study of the spectral characteristics of highly variable $\gamma$-ray sources on sub-hour timescales.
The basic elements of the suggested 5@5 detector are the large optical reflectors and the multichannel high resolution cameras. Presently, 20 m diameter alt-azimuth mounts with the required precision of about 1 arcminute could be designed and built by many companies specialized in the construction of large radio dishes. The area of the optical reflector could be composed of several hundreds to thousand $\leq 1 \, \rm m$ diameter glass mirrors with protective quartz coating, quite similar to the mirrors used in the current or planned imaging Cherenkov telescope projects. The general requirements on the imagers are a relatively large ($3^{\circ}$ or so) field of view with a pixel size of about $0.1^{\circ}$. This implies less than 1 thousand fast channels. Such a camera with similar parameters has been already built and successfully operated as part of the CAT imaging Cherenkov telescope [@cat].
The decrease of the energy threshold of the imaging atmospheric Cherenkov technique to several GeV would depend to a large extent on the availability of exceptional sites with a dry and transparent atmosphere at an altitude as high as 5 km. Nature does provide us with such an extraordinary site - The Llano de Chajnantor in the Atacama desert in Northern Chile. This site with its very arid atmosphere was recently chosen for the installation of one of the most powerful future astronomical instruments - the Atacama Large Millimeter Array (ALMA), a funded US-European project which will consist of 64 12-meter radio antennas with spacing from approximately 150 meters to 10 km (see http://www.alma.nrao.edu/). The large flat area on that site could certainly accommodate an additional Cherenkov telescope array as well which requires a relatively compact area with a radius of about 100 m. Another attractive feature of this site seems to be an adequate infrastructure which will be built up during the next several years for the ALMA project. The foreseen technological developments of ALMA, concerning in particular the construction of very large antennas operating in robotic or semi-robotic mode, could help very much in the design of the telescopes of 5@5. Moreover, the neighboring Cerro Toco site [@tocco] offers suitable areas at even higher altitudes, $H \simeq5.2$ and $5.6$ km a .s.l., with the same infrastructure advantages.
[**Acknowledgments**]{}
We thank the anonymous referee for her/his critical comments and remarks which helped us to improve the paper significantly. HQ is grateful for the award of a Presidential Chair in Science (Chile). His research is partially funded by FONDECYT Grant No. 8970009.
[999]{} M.F. Cawley and T.C. Weekes, [*Exp. Astron.*]{} [**6**]{} (1996) 7. A.M. Hillas, [*Space Sci. Rev.*]{} [**75**]{} (1996) 17. D.J. Fegan, [*J. Phys. G: Nucl. Part. Phys.*]{} [**23**]{} (1998) 1013. T.C. Weekes, in: Proc. [*GeV/TeV Gamma Ray Astrophysics Workshop*]{}, Eds. B. Dingus et al. (Snowbird, Utah), AIP Conf. Proc Ser. 515 (2000) 3. F.A. Aharonian and C.W. Akerlof, [*Annu Rev. Nucl. Part Sci.*]{} [**47**]{} (1997) 324. F.A. Aharonian, W. Hofmann, A.K. Konopelko, and H.J. Völk, [*Astropart. Phys.*]{} [**6**]{} (1997) 343. A. Daum et al. (HEGRA collaboration), [*Astropart. Phys.*]{} [**8**]{} (1997) 1. A.K. Konopelko et al. (HEGRA collaboration), [*Astropart. Phys.*]{} [**10**]{} (1997) 275. M. Mori et al. in: Proc. [*GeV/TeV Gamma Ray Astrophysics Workshop*]{}, Eds. B. Dingus et al. (Snowbird, Utah), AIP Conf. Proc Ser. 515 (2000) 485. W. Hofmann, in: Proc. [*GeV/TeV Gamma Ray Astrophysics Workshop*]{}, Eds. B. Dingus et al. (Snowbird, Utah), AIP Conf. Proc Ser. 515 (2000) 500. F. Krennrich et al., in: Proc. [*GeV/TeV Gamma Ray Astrophysics Workshop*]{}, Eds. B. Dingus et al. (Snowbird, Utah), AIP Conf. Proc Ser. 515 (2000) 515. E.D. Bloom, [*Space Sci. Rev.*]{} [**75**]{} (1996) 109. N. Gehrels, P. Michelson, [*Astropart. Phys.*]{} [**11**]{} (1997) 277. D. Smith, in: Proc. [*GeV/TeV Gamma Ray Astrophysics Workshop*]{}, Eds. B. Dingus et al. (Snowbird, Utah), AIP Conf. Proc Ser. 515 (2000) 416. D. Ong, in: Proc. [*GeV/TeV Gamma Ray Astrophysics Workshop*]{}, Eds. B. Dingus et al. (Snowbird, Utah), AIP Conf. Proc Ser. 515 (2000) 401. E. Pare, [*Space Sci. Rev.*]{} [**75**]{} (1996) 127. E. Lorenz, [*GeV/TeV Gamma Ray Astrophysics Workshop*]{}, Eds. B. Dingus et al. (Snowbird, Utah), AIP Conf. Proc Ser. 515 (2000) 510. J.R. Primack et al., [*Astropart. Phys.*]{} [**11**]{} (1999) 93. A.K. Konopelko, A.V. Plyasheshnikov, [*Nucl. Inst. and Methods*]{} [**450**]{} (2000) 419. F.A. Aharonian (HEGRA collaboration) [*Phys. Rev D*]{} [**59**]{} (1999) 092003. F.A. Aharonian et al. (HEGRA collaboration), [*Astropart. Phys.*]{} [**10**]{} (1998) 21. F.A. Aharonian et al. (HEGRA collaboration) [*Astron. Astrophys.*]{} [**349**]{} (1999) 11. A.K. Konopelko [*Astropart. Phys.*]{} [**11**]{} (1999) 263. A. Djannati-Atai et al. (CAT collaboration) [*Astron. Astrophys.*]{} [**350**]{} (1999) 17. P. Chadwick et al. , [*J. Phys. G: Nucl. Part. Phys.*]{} [**25**]{} (1999) 1223. J.R. Patterson, A.M. Hillas, [*J. Phys. G: Nucl. Phys.*]{} [**9**]{} (1983) 1433. C.E. Fichtel et al., [*ApJ Suppl.*]{} [**90**]{} (1994) 917. W. Hofmann, in: Proc. [*GeV/TeV Gamma Ray Astrophysics Workshop*]{}, Eds. B. Dingus et al. (Snowbird, Utah), AIP Conf. Proc Ser. 515 (2000) 318. T. Taira et al., in: [*Proc 23rd ICRC*]{} (Calgary) [**2**]{} (1993) 128. S.W. Barwick et al., [*ApJ*]{} [**498**]{} (1998) 779. K. Hurley, [*Space Sci. Rev.*]{} [**75**]{} (1996) 43. C. Akerlof et al., [*Nature*]{} [**398**]{} (1999) 400. A.D. Miller et al. [*ApJ*]{} [**524**]{} (1999) L1.
[^1]: For example, in the case of the HEGRA stereoscopic IACT system the minimum number of photoelectrons corresponding to the showers classified as “high quality events”, i.e. the showers which are accepted for the further image analysis, is close to 40 electrons per telescope [@hegra_MC].
[^2]: Actually there are some differences. In particular the primary electrons start to produce Cherenkov light earlier, but in this paper we will ignore these effects.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Junichi Shigezumi
title: |
On the zeros of certain modular functions for\
the normalizers of congruence subgroups of low levels\
II
---
Graduate School of Mathematics Kyushu University\
Hakozaki 6-10-1 Higashi-ku, Fukuoka, 812-8581 Japan\
[*E-mail address*]{} : [email protected]
> [**Abstract.**]{} We research the location of the zeros of the Eisenstein series and the modular functions from the Hecke type Faber polynomials associated with the normalizers of congruence subgroups which are of genus zero and of level at most twelve.
>
> In Part II, we will observe the location of the zeros of the above functions by numerical calculation.\
>
> [**Key Words and Phrases.**]{} Eisenstein series, locating zeros, modular forms.\
>
> 2000 [*Mathematics Subject Classification*]{}. Primary 11F11; Secondary 11F12.
\
In ‘Part I’, we will consider the general theory of modular functions for the normalizers of the congruence subgroups $\Gamma_0(N)$ of level $N \leqslant 12$. And in ‘Part II’, we will observe the location of the zeros of the Eisenstein series and the the modular functions from Hecke type Faber polynomials for the normalizers in Part I by numerical calculation.
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'library.bib'
---
[**** ]{}\
Vina Ayumi^1^, L.M. Rasdi Rere^1,2^, Mohamad Ivan Fanany^1^, Aniati Murni Arymurthy^1^,\
**[1]{} Machine Learning and Computer Vision Laboratory,\
Faculty of Computer Science, Universitas Indonesia\
**[2]{} Computer System Laboratory, STMIK Jakarta STI&K\
[email protected]****
Abstract {#abstract .unnumbered}
========
Convolutional neural network (CNN) is one of the most prominent architectures and algorithm in Deep Learning. It shows a remarkable improvement in the recognition and classification of objects. This method has also been proven to be very effective in a variety of computer vision and machine learning problems. As in other deep learning, however, training the CNN is interesting yet challenging. Recently, some metaheuristic algorithms have been used to optimize CNN using Genetic Algorithm, Particle Swarm Optimization, Simulated Annealing and Harmony Search. In this paper, another type of metaheuristic algorithms with different strategy has been proposed, i.e. Microcanonical Annealing to optimize Convolutional Neural Network. The performance of the proposed method is tested using the MNIST and CIFAR-10 datasets. Although experiment results of MNIST dataset indicate the increase in computation time (1.02x - 1.38x), nevertheless this proposed method can considerably enhance the performance of the original CNN (up to 4.60%). On the CIFAR10 dataset, currently, state of the art is 96.53% using fractional pooling, while this proposed method achieves 99.14%.
[***Keywords—*** Metaheuristic, Microcanonical Annealing, Convolutional Neural Network, MNIST, CIFAR10]{}
Introduction {#sec1}
============
Essentially, Deep learning (DL) is motivated by the artificial intelligent (AI) research, where the objective is to replicate the human brain capability, i.e. to observe, learn, analyze and make a decision, particularly for complex problems [@Naja]. DL is about learning the representation of a hierarchical feature, and it contains a variety of methods, such as neural network, hierarchical of probabilistic models, and supervised as well as unsupervised learning algorithms.[@Liangpei]. The current good reputation of DL is due to the decrease in the price of computer hardware, improvement in the computational processing capabilities, and advanced research in the Machine Learning and Signal Processing [@Deng].
In general, DL models can be classified into discriminative, generative, and hybrid models[@Deng]. Recurrent neural network (RNN), deep neural networks (DNN), and convolutional neural networks (CNN) are some examples of Discriminative models. Examples of generative models are deep Boltzmann machine (DBM), regularized autoencoder, and deep belief network (DBN). In the case of the hybrid model, it refers to a combination of generative and discriminative models. An example of such hybrid model is a pre-trained deep CNN using DBN, where it can improve the performance of deep CNN better than if it uses only random initialization. Among all of these DL techniques, this paper focuses on CNN.
Although the good reputation of DL for solving any learning problem is known, how to train it is challenging. The successful proposal to optimize this technique using layered-wise pre-training was proposed by Hinton and Salakhutdinov [@Hinton]. Some other methods are Hessian-free optimization suggested by Marten [@Martens], and Krylov Subspace Descent by Vinyal et al. [@Vinyal]
Recently, some of the metaheuristic algorithms have been used to optimize DL, especially CNN. Some papers[@You][@Oul][@Rasdi][@Gustavo] [@Laode] report that these methods can improve the accuracy of CNN. Metaheuristic is a powerful method to solve difficult optimization problems, and it has been used in almost all research area of engineering, science, and even industrial application [@Yang]. In general, this method works with three main objectives, i.e. solving big problems, solving the problem faster, and finding robust algorithms [@Talbi]. Besides, they are not difficult to be designed, flexible, and relatively easy to be applied.
Almost all metaheuristics algorithms inspired by nature, which is based on several principles of phenomena in physics, biology, and ethology. Some examples of biology phenomena are Differential Evolution (DE), Evolution Strategy (ES), Genetic Algorithm (GA). Phenomena of physics are Threshold Accepting method (TA), Microcanonical Annealing (MA), Simulated Annealing (SA), and Ethology phenomena are Ant Colony Optimization (ACO), Firefly Algorithm (FA), Particle Swarm Optimization (PSO)[@Boussaid]. Another metaheuristic phenomenon is inspired by music, such as Harmony Search algorithm [@Lee].
Classifications of metaheuristic can also be based on single-solution based metaheuristic (S-metaheuristic) and population-based metaheuristic (P-metaheuristic). Examples of S-metaheuristic are SA, TA, MA, Guided Local Search, and Tabu Search. In the case of P-metaheuristic, it can be divided into Swarm Intelligent (SI) and Evolutionary Computation (EC). Examples of SI are FA, PSO, ACO, Bee Colony Optimization and examples of EC are GA, ES, DE [@Boussaid].
Of the various types of the metaheuristic algorithm, in this paper we use the MA, with the consideration that the S-Metaheuristic is simple to implement on DL, and to the best of our knowledge, has never been used for optimizing CNN.
The Macrocanonic algorithm is the variant of Simulated Annealing. Uses an adaptation of the Metropolis algorithm, the conventional SA algorithm aims to bring a system to equilibrium at decreasing temperatures [@Stephen]. On the other hand, MA based on Creutz’s microcanonical simulation technique, where the system’s evolution is controlled by its internal energy, not by its temperature. The advantages Creutz algorithm over the Metropolis algorithm is since it does not require the generation of quality random numbers or the evaluation of transcendental functions, thus allowing much faster implementation. Experiments on the Creutz method indicate that, it can be programmed to run an order of magnitude faster than the conventional Metropolis method for discrete systems [@Bhanot]. A further significant advantage is that microcanonical simulation does not require high-quality random numbers.
The organization of this paper is as follows: Section 1 is an introduction; Section 2 provides an overview of Microcanonical Annealing; Section 3 describes the method convolutional neural network; Section 4 presents the proposed method; Section 5 gives the results of the experiment, and lastly, Section 6 presents the conclusion of this paper.
Microcanonical Annealing
========================
Microcanonical Annealing (MA) corresponds to a variant of simulated annealing (SA). This technique is based on the Creutz algorithm, known as “demon” algorithm or microcanonical Monte Carlo simulation. In which the algorithm tolerates attainment of the equilibrium of thermodynamic in an isolated system, where in this condition, total energy of the system $E_p$ is constant [@Boussaid].
Total energy is the sum of kinetic energy $E_k$ and potential energy $E_p$ of the system, as the equation (2) follow:
$$\label{eq1}
E_{total} = E_k + E_p$$
In case of minimum optimization problem, potential energy $E_p$ is the objective function to be minimized, and the kinetic energy is used as temperature in SA, that is forced to remain positive [@Boussaid]. When the change of energy is negative $(- \Delta E)$, while it increases the kinetic energy $(E_k \leftarrow E_k - \Delta E )$, this new states is accepted. Otherwise, it is accepted when $- \Delta E < E_k$, and the energy obtained in the form of potential energy is cut off from the kinetic energy. So that the total energy remains constant. The standard algorithm for MA is shown in **Algorithm 1**.[@Boussaid].
Randomly, select an initial solution $x$ Initialize the kinetic energy $E_k$
Convolutional neural network
============================
One variant of the standard multilayer perceptron (MLP) is CNN. Its capability in reducing the dimension of data, extracting the feature sequentially, and classifying in one structure of network are distinguished advantages of this method, especially, for pattern recognition compared with the conventional approaches [@Bengio].
The classical CNN by LeCun et al [@LeCun] is an extension of traditional MLP based on three ideas: local receive fields, weights sharing, and spatial/temporal sub-sampling. There are two types of processing layers, which are convolution layers and sub-sampling layers. As demonstrated in Fig.1, the processing layers contain three convolution layers C1, C3, and C5, combined in between with two sub-sampling layers S2 and S4, and output layer F6. These convolution and sub-sampling layers are arranged into planes called features maps.
In convolution layer, each neuron is locally linked to a small input region (local receptive field) in the preceding layer. All neurons with similar feature maps obtain data from different input regions until the whole plane input is skimmed, but the similar weights are used together (weights sharing).
The feature maps are spatially down-sampled in sub-sampling layer, in which the map size is reduced by a factor 2. For instance, the feature map in layer C3 of size 10x10 is sub-sampled to a conforming feature map of size 5x5 in the subsequent layer S4. The last layer is F6 that is the process of classification [@LeCun].
Basically, a convolution layer is correlated with some feature maps, the size of the kernel, and connections to the previous layer. Each feature map is the result of a sum of convolution from the maps of the previous layer, by their corresponding kernel and a linear filter. Furthermore, a bias term is added to the map then and applying it to a non-linear function. The k-th feature map $M_{ij}^k$ with the weights $W^k$ and bias $b_k$ is obtained using the $\tanh$ function as follow:
$$M_{ij}^k=\tanh((W^k \times x)_{ij} + b_k)$$
The purpose of a sub-sampling layer is the spatially invariant reached by reducing the feature maps resolution, where each feature map is pooled relating to one of the feature map of the previous layer
where each map feature is collected relating to one of the maps of the features of the previous layer. Where $a_i^{n \times n} $ are the inputs, $ \beta $ is a scalar of trainable, and $ b $ is bias of trainable, the sub-sampling function,is given by the following equation:
$$a_j=\tanh\left(\beta\sum_{N\times N}{a_i^{n \times n} + b}\right)$$
After several convolutions and sub-samplings, the last structure is a classification layer. This layer works as an input for a series of fully connected layers that will execute the classification task. In this layer, each output neuron is assigned to one class label, and in the case of CIFAR10 or MNIST data set, this layer contains ten neurons corresponding to their classes.
Design of proposed methods
==========================
In this proposed method, the algorithm of MA is used to train CNN to find the condition of best accuracy, as well as to minimize estimated error and indicator of network complexity. This objective can be realized by computing the loss function of vector solution or the standard error on the training set. The following is the loss function used in this paper: $$f= \frac {1}{2} \left({\frac{\sum_{i=N}^{N}{(x - y)^2}}{N}}\right)^{0.5}$$ where the expected output is $x$, the real output is $y$, and some of the training samples are $N$. The two situations are used in this method for termination criterion. The first is when the maximum iteration has been reached and the second is when the loss function is less than a certain constant. Both conditions mean that the most optimal state has been achieved.
The architecture of this proposed method is i-6c-2s-12c-2s, where the number of C1 is 6, and C3 is 12. The size of kernel for all convolution layer is 5x5, and the scale of sub-sampling is 2. This architecture is a simple CNN structure (LeNet-5), not a complex structure like AlexNet[@Alexnet], SPP[@KHe], and GoogLeNet[@Szegedy]. In this paper, these architecture is designed for MNIST dataset.
Technically in these proposed methods, CNN will compute the values of bias and weight. These values ($x$) are used to calculate the loss function $f(x)$.
The values of $x$ are used as a vector of solution in MA, which will be optimized, by adding a value of $\Delta x$ randomly. Meanwhile, $f(x)$ is used as a potential energy $E_k$ in MA.
In this proposed method, $\Delta x$ is one of the important parameters. The value of accuracy will be improved significantly by providing an appropriate value of the $\Delta x$ parameter. As an example of one epoch, if $ \Delta x = 0.001 \times rand$, then the maximum accuracy is 87.60%, in which this value is 5.21% greater than the original CNN (82.39%). However, if $ \Delta x = 0.0001 \times rand$, its accuracy is 85.45% and its is only 3.06% greater than the original CNN.
Another important parameter of the proposed method is the size of neighborhood. For example in one epoch, if neighborhood is 5, 10 or 20, and then the accuracy values are respectively 85.74%, 87.52%, or 88.06%. While the computing time are respectively 98.06 seconds, 99.18 seconds and 111.80 seconds.
Furthermore, this solution vector is updated based on MA algorithm. In case of termination criterion has been reached, all of biases and weights for all layers on the system will be updated.
Experiment and results
======================
In this paper, there are two categories of experiments conducted, based on the dataset. The first experiment was using MNIST dataset, and the second experiment using CIFAR10 dataset. Some of the examples image for MNIST dataset are shown in Figure 2 and for CIFAR10 dataset are shown in Figure 3.
![Examples of some image from MNIST data-set[]{data-label="fig:my_label"}](MNIST5)
Experiment using MNIST data set
-------------------------------
The experiment for MNIST data set was implemented in MATLAB-R2011a, windows 10, on a PC with processor Intel Core i7-4500u, and 8 GB RAM running memory, with five experiments for each epoch. The original program of this experiment is DeepLearn Toolbox from Palm[@Palm]. In this research, the program of CNN is modified with the algorithm of MA.
In all experiment, the size of neighborhood was set to 10, maximum of iteration (maxit) = 10, as well as kinetic energy = 100. We also set the parameter of CNN i.e., the learning rate ($\alpha = 1$) and the batch size (100).
On the MNIST dataset, all of the experiment results of the proposed methods are compared with the experiment result from the original CNN. The results of CNN and CNN based on MA is summarized in Table 1, for accuracy (A1, A2) and computation time (T1, T2), as well as Figure 4 for Error and Figure 5 for computation time.
In case of 100 epochs, as is shown in Figure 6 and 7, the accuracy of original CNN is 98.65% and the accuracy of CNN by MA is 98.75%. The computation time of both methods are 10731 seconds and 17090s seconds respectively.
[c c c c c]{} & &\
& & & &\
1 & 82.39 & 91.75 & 86.99 $\pm$ 0.53 & 109.99 $\pm$ 10.47\
2 & 89.06 & 193.39 & 91.33 $\pm$ 0.43 & 203.04 $\pm$ 0.83\
3 & 91.13 & 297.31 & 93.14 $\pm$ 0.31 & 302.84 $\pm$ 1.74\
4 & 92.33 & 379.44 & 94.48 $\pm$ 0.15 & 402.42 $\pm$ 0.57\
5 & 93.11 & 479.04 & 95.11 $\pm$ 0.28 & 514.52 $\pm$ 7.54\
6 & 93.67 & 576.38 & 95.72 $\pm$ 0.12 & 612.07 $\pm$ 2.33\
7 & 94.25 & 676.57 & 95.99 $\pm$ 0.22 & 781.32 $\pm$ 23.57\
8 & 94.77 & 768.24 & 96.26 $\pm$ 0.26 & 1062.11 $\pm$ 6.79\
9 & 95.37 & 855.85 & 96.49 $\pm$ 0.27 & 1144.01 $\pm$ 79.84\
10 & 95.45 & 954.54 & 96.89 $\pm$ 0.15 & 1274.00 $\pm$ 47.69\
100 & 98.65 & 10731 & 98.75 & 17090\
In general, the experiments conducted for MNIST data set shown that the proposed methods are better than the original CNN, for any given epoch. As an example for the second epoch, the accuracy of original CNN is 89.06%, while for CNNMA is 91.33%. Accuracy improvement of the proposed method, compared to original CNN, varies of each epoch, with a range of values between 1.12% (CNNMA, 9 epoch) up to 4.60% (CNNMA, 1 epoch).
The computation time for the proposed method, compared to the original CNN, is in the range of $1.02\times$ (CNNMA, three epochs : 302.84/297.31) up to $1.38\times$ (CNNMA, eight epochs: 1062.11/768.24).
Experiment using CIFAR10 dataset
--------------------------------
The experiment of CIFAR10 dataset was conducted in MATLAB-R2014a, Ubuntu 14.04 LTS 64, on a PC with Processor Intel Core i7-5820K, Four GPU GTX Titan X, Memory DDR2 RAM 64.00 GB, Hard disk 240 GB. The original program is MatConvNet from [@vedaldi]. In this paper, the program was modified with MA algorithm. The results can be seen in Fig. 8 for top-1 error and top-5 error.
The proposed method has proven very effective on CIFAR10 dataset with an accuracy of 99.6%, for the last epoch in the top-1 error. In Table II different results from state of the art approaches are listed as a comparison. Another work proposed fine-tuning CNN using metaheuristic algorithm, harmony search (HS) [@Rosa] also compared in Table II.
Method Accuracy (%)
---------------------------- --------------
**CNN-MA \[ours\]** **99,14**
CNN-HS [@Rosa] 72.28
Fractional Pooling [@BG] 96.53
Large ALL-CNN [@Jost] 95.59
Spatially Sparse CNN [@BG] 95,53
LSUV [@Dmytro] 94.16
CNN [@matconvnet] 80.46
: State of the art CIFAR10 dataset[]{data-label="tab:example"}
Conclusion
==========
This paper proposed a type of metaheuristic called Microcanonical Annealing algorithm to optimize the Convolutional Neural Network. Experimental result using MNIST and CIFAR-10 dataset demonstrated that although MA requires more computational time, the accuracy is reasonably better than the standard CNN without metaheuristic. This paper shows that on MNIST dataset, Microcanonical Annealing can improve the accuracy of Convolutional Neural Network, for all variations of epoch up to 4.60%. The results obtained for CIFAR10 dataset, with an accuracy of 99.14% (top-1 error), indicates that the proposed method is able to compete on the current state of the art approaches (96.53%), in the field of image classification. For the future study, fining the proper MA parameters need to be investigated. Furthermore application of this proposed method using the other benchmark data set need to be explored, such as MMI and CKP facial expression data set, as well as ORI and ImageNet. For future research, we will investigate further on the computation time comparison between Microcanonical Annealing to the Simulated Annealing. We also need to examine further the accuracy on the CIFAR-10 dataset using other GPU-based deep learning frameworks such as Torch, Theano, Tensorflow, and Keras with more number of iterations.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A map to the quark-gluon cascade on the basis of nonlinearity in the quark and gluon distributions in hadrons is proposed. Calculations of the quarks trajectories have shown the presence of the chaotic dynamics as a consequence of bifurcations.'
author:
- 'A.T. Temiraliev'
title: 'Chaotic dynamics in quark-gluon cascade'
---
At high energy, there are two types of the QCD processes: perturbative and non-perturbative (short and long distances, respectively). In the non-perturbative region, direct expansion in the strong coupling constant is inapplicable. Then, to describe the strong interactions, it is necessary to use in addition some assumptions and phenomenological constructions. Modern research in hadron and nuclear physics have deal more and more with nonlinear aspects of the quark-gluon dynamics. Analysis of the role of nonlinear effects in formation of the initial quark-gluon configurations in the hadrons at early stage of the nuclear impacts was given for the first time in the articles [@Kovch]. Later on, nonlinear effects were considered within the quasi-classical approach [@Leon].
Ambiguity of the account of nonlinear effects has led to significant variety of the cascade models. Various models of the behavior of quarks and gluons have been confronted with a large amount of experimental data on the deep-inelastic scattering (DIS) structure functions (SF), which describe the quarks and gluons distribution [@Adl]. The dependence of quarks and gluons SF on Bjorken’s variables $Q^2$ and $x$ is usually obtained from the numerical solution of the linear Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP) [@G-L] or Balitsky-Fadin-Kuraev-Lipatov (BFKL) equations [@KLF]. Evolution of the transverse-momentum dependent parton densities had been studied in [@Cher].
At present, one of the greatest achievements can consider the discovery of chaotic dynamics [@Fej] in different areas of science. Chaotic behavior has been observed in a variety of systems including lasers, electrical circuits, fluid dynamics as well as computer models of chaotic processes. In frameworks of fractal analysis the hypothesis of universality at formation of hadrons is discussed in [@Bat]. After those works, it becomes clear that the dynamics at the chaos border often manifests scaling regularities. Scenario of transition to chaos means a sequence of bifurcations observed under slow variation of a control parameter on a way from regular to chaotic behavior, for example, via period-doubling cascade, quasi-periodicity, intermittency.
The evolution of dynamic system usually is described by the differential equation. However in the modern researches there is a map method allowing to describe many phenomena of nonlinear dynamics. There are a merges and splitting in the quark and gluon evolution. The collective interactions of partons (quarks and gluons) led to universality of the hadron SF. We use the renormalization-group approach to the evolution of hadron’s SF, allowing to recreate a physical picture of the critical behavior. For the quark-gluon cascade, we enter an iterative map in which a number of the quarks and gluons in $(n+1)$-th generation are proportional to the number of them in $n$-th generation. To find the shares of the momentum $x$, we use the one-dimensional map for quark and gluon distributions: $$x_{n+1} = R f(x_n) \ ,$$ where $n=0,1,2,\ldots $ and $f(x)=xq(x)$ is the momentum distribution of partons with the momentum fraction $x$ and density $q(x)$. $R$ is the parameter, determining the character of observing regimes.
The function $f(x)$ in the map $x_{n+1}=f(x_{n})$ determines evolution for one step of the discrete time. For two steps we have $x_{n+2}=f(f(x_n))$. Let us introduce instead of $x$ a new variable re-scaled with the factor $\alpha$. Changing $x$ to $x/\alpha$ in both sides of the equation, we can write $x_{n+2}=f_1(x_n)$, where $f_1(x)=\alpha f(f(x/\alpha))$. We can take $f_1(x)$ as a new initial function and perform the same operations. Then, we obtain a renormalized evolution operator for four steps: $x_{n+4}=f_2(x_n)$, where $f_2(x)=\alpha f_1(f_1(x/\alpha))$. Repetition of the procedure yields a recurrent functional equation $f_{k+1}(x)=\alpha f_k(f_k(x/\alpha))$. If the original map $f_0(x)$ depends on the parameter and demonstrates period-doubling cascade, then, at the accumulation point of the period-doubling. The limit function f(x) will satisfy to the equation $f(x)=\alpha f(f(x/\alpha))$ [@Tur]. It is a fixed point of the functional equation.
The momentum distributions of quarks in a hadron for the momentum fraction $x \rightarrow 1$ possess, according to the QCD sums rules, asymptotic behavior of the valence quarks which define the quantum numbers of hadron:
$$\lim f(x)\rightarrow (1-x)^{2n_v-1} \ ,$$ where $n_v$ is the number of the valence quarks between which the rest of the hadron momentum ($n_v=2$ for nucleons and $n_v=1$ for mesons) is distributed. It is yet not possible to prove asymptotic formulas in limits of QCD for partons with small fraction of hadron momentum $(x\ll 1)$ at Regge region. We use simple parameterization of the SF from experimental data at great values of 4-momentum transferred at interaction ($Q^2\geq 10 \ Gev^2$) for the proton: $xu_v(x)=1.8x(1-x)^{2.5}$, $xd_v(x)=3.6x(1-x)^{1.5}$, $xS(x)=0.1(1-x)^6$, $xG(x)=2.76(1-x)^5$, where $u_v$, $d_v$ are the valence quark, $S$ is the sea quark, and $G$—the gluon distributions. The momentum distribution of the quarks reads: $$f(x)
=
1.8x(1-x)^{2.5}+3.6x(1-x)^{1.5}+0.1(1-x)^6 \ .$$
Thus, the quark momentum fraction in the nucleon is 0.54 and the gluon one is almost one half 0.46. Positive terms in the equation (1) with $f(x)$ from (2) describe the increase of the partons number and negative their is the reduction, i.e. quark-antiquark, quark-gluon and gluon-gluon recombination.
Numerical solution of the equation (1) has shown that there is the evolution termination in the field of small values of evolution parameter ($R < 0.3$) at any initial value $x_0$ of parton’s momentum fractions. At increasing of the parameter $R$ the transition is in established mode that corresponds to long enough cascade process at great values of iterations. It is observed that the number of partons do not change. Trajectories ($x_n$) after transient reach some steady value (motionless point). At $R=0.8$ after transient the display orbit becomes strictly periodic with the period two (bifurcation of motionless point). At $R> 0.9$ the orbit represents more difficult movement. These processes are shown in Figure 1 at $R=0.2$, $R=0.8$ and $R=0.99$. At the fixed value $R$ and changing initial values of momentum fractions $x_0$ of the partons after transition period don’t depend on the initial value $x_0$.
The change of trajectory character most visually shows the dependence schedule $x_{i,j}=f(x_{i-1,j},R_j)$ at iterations of parameter of evolution $R_j=0.99j/500=0\div 0.99$, where $i$ is the number of points on variable $x$ (number of iterations), $j$ is the number of values of the variable $R_j$. At large values of $R$, the system transfers in a chaotic mode when two close points run up on different trajectories that is shown in Figure 2 is observed. Thus, casual small initial changes of the system of the cooperating partons can lead to as much as big changes of the dynamics of the system.
In order to distinguish between the chaotic and not chaotic modes, we compare orbits to close entry conditions in these modes. As a measure of this difference, we choose the difference module between values of corresponding orbits of the display, carried to the value of one of the orbits.
At initial approach for the first trajectory $x_0 =0.5$ and $y_0=0.50001$ for the second one. Results of calculations of trajectories for $R=0.95$ and $R=0.75$ dynamics are presented in Figure 3. **Conclusion:** In the quark-gluon cascade, the essential role is played by not only linear parts of the parton density, but also the higher order terms. Dynamic quark-gluon systems are highly sensitive to the initial conditions. The strange quark trajectories (attracts) display a new nonlinear phenomenon in the hadron physics — dynamic chaos in the quark-gluon evolution. Scenario of the transition to the chaos means a sequence of bifurcations observed under the slow variation of the control parameter.
**Acknowledgments:** I am grateful to I.O. Cherednikov for his attention to the article and help.
*Yu. Kovchegov, A. Mueller*, Nucl. Phys. B329 (1998) 451. *A. Leonidov, D. Ostrovsky*, Eur. Phys. J. Cll (1999) 495. *C. Adloff*, Eur.Phys.J. C13 (2000) 609; hep-ex/9908059. *V. N. Gribov, L. N Lipatov*, Sov. J. Nucl. Phys. 15 (1972) 438;\
*Y. L. Dokshitzer*, Sov. Phys. JETP. 46 (1977) 641;\
*G. Altarelli, G. Parisi*, Nucl. Phys. B126 (1977) 298. *E. A. Kuraev, L. N. Lipatov, V. S. Fadin*, Sov. JETP. 44 (1976) 443;\
*I. I. Balitsky, L. N. Lipatov*, Sov. J. Nucl. Phys. 28 (1978) 822;\
*L. N. Lipatov*, Sov. Phys. JETP. 63 (1986) 904. *I.O.Cherednikov, N.G.Stefanis*, arXiv: 1104.0168 (2011). *M. Fejgenbaum*, Phys. Usp. 141 (1983) 343. *A. V. Batunin*, Phys. Lett. B327 (1994) 293. *L.V. Turukina, A. Pikovsky* Phys. Lett. A375 (2011) 1407.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We provide the explicit combinatorial structure of the Kirillov–Reshetikhin crystals $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$. This is achieved by constructing the crystal analogue $\sigma$ of the automorphism of the $D_n^{(1)}$ (resp. $B_n^{(1)}$ or $A_{2n-1}^{(2)}$) Dynkin diagram that interchanges the 0 and 1 node. The involution $\sigma$ is defined in terms of new $\pm$ diagrams that govern the $D_n$ to $D_{n-1}$ (resp. $B_n$ to $B_{n-1}$, or $C_n$ to $C_{n-1}$) branching. It is also shown that the crystal $B^{r,s}$ is perfect. These crystals have been implemented in MuPAD-Combinat; the implementation is discussed in terms of many examples.'
address: 'Department of Mathematics, University of California, One Shields Avenue, Davis, CA 95616-8633, U.S.A.'
author:
- Anne Schilling
title: 'Combinatorial structure of Kirillov–Reshetikhin crystals of type $D_n^{(1)}$, $B_n^{(1)}$, $A_{2n-1}^{(2)}$'
---
[^1] [^2]
Introduction
============
The irreducible finite-dimensional modules over a quantized affine algebra $U'_q({\mathfrak{g}})$ were classified by Chari and Pressley [@CP:1995; @CP:1998] in terms of Drinfeld polynomials. We are interested in the subfamily of such modules which possess a global crystal basis. Kirillov–Reshetikhin (KR) modules are finite-dimensional $U'_q({\mathfrak{g}})$-modules $W^{r,s}$ that were introduced in [@HKOTT:2002; @HKOTY:1999]. It is expected that each KR module has a crystal basis $B^{r,s}$, and that every irreducible finite-dimensional $U'_q({\mathfrak{g}})$-module with crystal basis, is a tensor product of the crystal bases of KR modules. KR crystals play an important role in lattice models of statistical mechanics and in the Kyoto path construction of highest weight $U_q({\mathfrak{g}})$-modules.
The KR modules $W^{r,s}$ are indexed by a Dynkin node $r$ of the classical subalgebra (that is, the distinguished simple Lie subalgebra) ${{\mathfrak{g}}_0}$ of ${\mathfrak{g}}$ and a positive integer $s$. In general the existence of $B^{r,s}$ remains an open question. For type $A_n^{(1)}$ the crystal $B^{r,s}$ is known to exist [@KKMMNN:1992a] and its combinatorial structure has been studied [@S:2002]. In many cases, the crystals $B^{1,s}$ and $B^{r,1}$ for nonexceptional types are also known to exist and their combinatorics has been worked out in [@KKM:1994; @KKMMNN:1992a] and [@JMO:2000; @Ko:1999], respectively. For type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$, which are subject of the current paper, the existence of $B^{r,s}$ was recently shown by Okado [@O:2006; @OS:2007]. For the twisted case, this relies on work by Hernandez [@H:2007].
Viewed as a $U_q({{\mathfrak{g}}_0})$-module by restriction, $W^{r,s}$ is generally reducible. Its decomposition into $U_q({{\mathfrak{g}}_0})$-irreducibles was conjectured in [@HKOTT:2002; @HKOTY:1999] and verified by Chari [@C:2001] for the nontwisted cases and Hernandez [@H:2007] for the twisted cases. Kashiwara [@Ka:2002] conjectured that as classical crystals, many of the KR crystals (the ones conjectured to be perfect in [@HKOTT:2002; @HKOTY:1999]) are isomorphic to certain Demazure subcrystals of affine highest weight crystals. Kashiwara’s conjecture was confirmed by Fourier and Littelmann [@FL:2004] in the untwisted cases and Naito and Sagaki [@NS:2005] in the twisted cases. In [@FSS:2006] it was shown that $0$-arrows of these KR crystals are in fact fixed by the Demazure structure and that this implies that the KR crystals are unique if they exist and satisfy certain properties (see Property \[A:KR\]).
In this paper, we provide an explicit combinatorial construction of the KR crystals $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$. The construction is based on the analogue $\sigma$ of the Dynkin diagram automorphism which interchanges the 0 and the 1 nodes as shown in Figure \[fig:Dynkin\].
(9,3) (1,0)(1,2)(2,1)(3,1)(4,1)(5,1)(6,1)(7,0)(7,2) (1,0)(2,1) (1,2)(2,1) (2,1)(6,1) (6,1)(7,2) (6,1)(7,0) (0.5,2)[0]{} (0.5,0)[1]{} (2,1.5)[2]{} (3,1.5)[3]{} (4,1.5)[$\ldots$]{} (5.8,1.6)[$n-2$]{} (7.9,2)[$n-1$]{} (7.5,0)[$n$]{} (0,0)[A]{} (0,2)[B]{} (-1,1)[$\sigma$]{}
(9,3) (1,0)(1,2)(2,1)(3,1)(4,1)(5,1)(6,1)(7,1) (1,0)(2,1) (1,2)(2,1) (2,1)(6,1) (6,1)(7,1) (0.5,2)[0]{} (0.5,0)[1]{} (2,1.5)[2]{} (3,1.5)[3]{} (4,1.5)[$\ldots$]{} (5.8,1.5)[$n-1$]{} (7,1.5)[$n$]{} (0,0)[A]{} (0,2)[B]{} (-1,1)[$\sigma$]{}
(9,3) (1,0)(1,2)(2,1)(3,1)(4,1)(5,1)(6,1)(7,1) (1,0)(2,1) (1,2)(2,1) (2,1)(6,1) (6,1)(7,1) (0.5,2)[0]{} (0.5,0)[1]{} (2,1.5)[2]{} (3,1.5)[3]{} (4,1.5)[$\ldots$]{} (5.8,1.5)[$n-1$]{} (7,1.5)[$n$]{} (0,0)[A]{} (0,2)[B]{} (-1,1)[$\sigma$]{}
A similar construction was given by Shimozono [@S:2002] for type $A_{n-1}^{(1)}$, for which the Dynkin diagram automorphism maps node $i$ to node $i+1 \pmod n$. On type $A_{n-1}^{(1)}$ crystals $\sigma$ is the promotion operator [@S:2002] and the affine crystal operators are expressed as $e_0=\sigma^{-1} \circ e_1 \circ \sigma$ and $f_0=\sigma^{-1} \circ f_1 \circ \sigma$. For type $D_n^{(1)}$, the case $B^{2,s}$ was treated in the author’s paper with Philip Sternberg [@StS:2006], by providing an explicit description of the Dynkin diagram automorphism $\sigma$ that interchanges nodes 0 and 1 on the level of crystals. In his thesis [@St:2006], Sternberg gave a conjecture for the crystal structure for general $B^{r,s}$ of type $D_n^{(1)}$. In this paper and [@OS:2007], we prove Sternberg’s conjecture by making it more explicit using $\pm$ diagrams and also extend the construction to type $B_n^{(1)}$ and $A_{2n-1}^{(2)}$. The $\pm$ diagrams govern the branching of the underlying classical algebra.
The main result of this paper is the definition of combinatorial crystals $B^{r,s}$ for type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ given in Definition \[def:comb B\]. The classical crystal structure is fixed by below and the affine crystal operators are defined as $e_0=\sigma \circ e_1 \circ \sigma$ and $f_0=\sigma \circ f_1 \circ \sigma$ with $\sigma$ as defined in Definition \[def:sigma\]. As shown in collaboration with Okado [@OS:2007 Theorem 1.2], the combinatorial crystals $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ constructed in this paper are in fact the Kirillov–Reshetikhin crystals associated with the KR $U'_q({\mathfrak{g}})$-module $W^{r,s}$. In addition we prove the following theorem. Here $\tilde{B}^{r,s}$ denotes the unique affine crystal structure of reference [@FSS:2006] satisfying Property \[A:KR\] below.
\[thm:affine\] For type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ and $r$ not a spin node, the crystals $\tilde{B}^{r,s}$ and $B^{r,s}$ are isomorphic.
The paper is outlined as follows. In section \[sec:KR\] we review the definition of crystals and the unique characterization of KR crystals coming from Demazure crystal theory as provided in [@FSS:2006]. In section \[sec:mupad\] we briefly describe the implementation of KR crystals in MuPAD-Combinat; throughout the paper it is demonstrated how to reproduce examples via MuPAD-Combinat, though the computer implementation is not used for the proofs. In the section \[sec:construction\] the explicit construction of $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ is given in terms of $\pm$ diagrams. The proof of Theorem \[thm:affine\] is provided in section \[sec:proof\]. It is shown in section \[sec:perfect\] that $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ is perfect.
Acknowledgements {#acknowledgements .unnumbered}
----------------
I would like to thank Ghislain Fourier, Peter Littelmann, Masato Okado, Mark Shimozono, and Philip Sternberg for many helpful discussions. Without the collaboration on [@StS:2006], [@FSS:2006] and [@OS:2007] the current paper would not have been possible. In particular, the notion of $\pm$ diagrams was developed in discussions with Mark Shimozono. Many thanks are also due to Christopher Creutzig, François Descouens, Teresa Gomez-Diaz, Florent Hivert, and Nicolas Thiéry for their support with MuPAD-Combinat [@HT:2003]. The implementation of the affine crystal $B^{r,s}$ in MuPAD-Combinat was essential for the progress of this project! Thanks to Adrien Boussicault for his help with ps-tricks and Robert Gutierrez for implementing some algorithms.
Kirillov–Reshetikhin crystals $B^{r,s}$ {#sec:KR}
=======================================
Some general definitions regarding crystals are reviewed in sections \[ss:axiom\] and \[ss:tensor\]. The classical crystal structure of type $D_n$, $B_n$, and $C_n$ crystals is given in section \[ss:classical\]. A unique characterization of Kirillov–Reshetikhin crystals of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ is reviewed in section \[ss:KR\].
Axiomatic definition of crystals {#ss:axiom}
--------------------------------
Let ${\mathfrak{g}}$ be a symmetrizable Kac-Moody algebra, $P$ the weight lattice, $I$ the index set for the vertices of the Dynkin diagram of ${\mathfrak{g}}$, $\{\alpha_i\in P \mid i\in I \}$ the simple roots, and $\{h_i\in P^* \mid i\in I \}$ the simple coroots. Let $U_q({\mathfrak{g}})$ be the quantized universal enveloping algebra of ${\mathfrak{g}}$. A $U_q({\mathfrak{g}})$-crystal [@K:1995] is a nonempty set $B$ equipped with maps ${\mathrm{wt}}:B\rightarrow P$ and $e_i,f_i:B\rightarrow B\cup\{\emptyset\}$ for all $i\in I$, satisfying $$\begin{aligned}
\label{eq:e-f}
f_i(b)=b' &\Leftrightarrow e_i(b')=b
\text{ if $b,b'\in B$} \\
{\mathrm{wt}}(f_i(b))&={\mathrm{wt}}(b)-\alpha_i \text{ if $f_i(b)\in B$} \\
\label{eq:string length}
{\langle h_i\,,\,{\mathrm{wt}}(b)\rangle}&=\varphi_i(b)-\varepsilon_i(b).\end{aligned}$$ Here for $b \in B$ $$\begin{split}
\varepsilon_i(b)&= \max\{n\ge0\mid e_i^n(b)\not=\emptyset \} \\
\varphi_i(b) &= \max\{n\ge0\mid f_i^n(b)\not=\emptyset \}.
\end{split}$$ (It is assumed that $\varphi_i(b),\varepsilon_i(b)<\infty$ for all $i\in I$ and $b\in B$.) A $U_q({\mathfrak{g}})$-crystal $B$ can be viewed as a directed edge-colored graph (the crystal graph) whose vertices are the elements of $B$, with a directed edge from $b$ to $b'$ labeled $i\in I$, if and only if $f_i(b)=b'$.
Let $\{\Lambda_i \mid i\in I\}$ be the fundamental weights of ${\mathfrak{g}}$. For every $b\in B$ define $\varphi(b)=\sum_{i\in I} \varphi_i(b) {\Lambda}_i$ and $\varepsilon(b)=\sum_{i\in I} \varepsilon_i(b) {\Lambda}_i$. An element $b\in B$ is called highest weight if $e_i(b) = \emptyset$ for all $i\in I$.
The center of an affine Kac–Moody algebra $\mathfrak{g}$ is one-dimensional and is generated by the canonical central element $c=\sum_{i\in I}a^\vee_i h_i$, where the $a^\vee_i$ are the numbers on the nodes of the Dynkin diagram of the algebra dual to $\mathfrak{g}$ given in [@Kac:1990 Table Aff of section 4.8]. Moreover, the imaginary roots of $\mathfrak{g}$ are nonzero integral multiples of the null root $\delta=\sum_{i\in I} a_i \alpha_i$, where the $a_i$ are the numbers on the nodes of the Dynkin diagram of $\mathfrak{g}$ given in [@Kac:1990 Table Aff]. Define $P_{{\mathrm{cl}}}=P/\mathbb{Z}\delta$ and $P_{{\mathrm{cl}}}^{+}=\{{\lambda}\in P_{{\mathrm{cl}}} \mid \langle h_{i},{\lambda}\rangle \geq 0 \textrm{ for all }i\in I\}$.
Tensor products of crystals {#ss:tensor}
---------------------------
Let $B_1,B_2,\dotsc,B_L$ be $U_q({\mathfrak{g}})$-crystals. The Cartesian product $B_L\times \dotsm \times B_2 \times B_1$ has the structure of a $U_q({\mathfrak{g}})$-crystal using the so-called signature rule. The resulting crystal is denoted $B=B_L\otimes\dots\otimes B_2\otimes B_1$ and its elements $(b_L,\dotsc,b_1)$ are written $b_L\otimes \dotsm \otimes b_1$ where $b_j\in B_j$. The reader is warned that our convention is opposite to that of Kashiwara [@K:1995]. Fix $i\in I$ and $b=b_L\otimes\dotsm\otimes b_1\in B$. The $i$-signature of $b$ is the word consisting of the symbols $+$ and $-$ given by $$\underset{\text{$\varphi_i(b_L)$ times}}{\underbrace{-\dotsm-}}
\quad \underset{\text{$\varepsilon_i(b_L)$
times}}{\underbrace{+\dotsm+}} \,\dotsm\,
\underset{\text{$\varphi_i(b_1)$ times}}{\underbrace{-\dotsm-}}
\quad \underset{\text{$\varepsilon_i(b_1)$
times}}{\underbrace{+\dotsm+}} .$$ The reduced $i$-signature of $b$ is the subword of the $i$-signature of $b$, given by the repeated removal of adjacent symbols $+-$ (in that order); it has the form $$\underset{\text{$\varphi$ times}}{\underbrace{-\dotsm-}} \quad
\underset{\text{$\varepsilon$ times}}{\underbrace{+\dotsm+}}.$$ If $\varphi=0$ then $f_i(b)=\emptyset$; otherwise $$f_i(b_L\otimes\dotsm\otimes b_1)= b_L\otimes \dotsm \otimes
b_{j+1} \otimes f_i(b_j)\otimes \dots \otimes b_1$$ where the rightmost symbol $-$ in the reduced $i$-signature of $b$ comes from $b_j$. Similarly, if $\varepsilon=0$ then $e_i(b)=\emptyset$; otherwise $$e_i(b_L\otimes\dotsm\otimes b_1)= b_L\otimes \dotsm \otimes
b_{j+1} \otimes e_i(b_j)\otimes \dots \otimes b_1$$ where the leftmost symbol $+$ in the reduced $i$-signature of $b$ comes from $b_j$. It is not hard to verify that this well-defines the structure of a $U_q({\mathfrak{g}})$-crystal with $\varphi_i(b)=\varphi$ and $\varepsilon_i(b)=\varepsilon$ in the above notation, with weight function $$\label{eq:tensor wt}
{\mathrm{wt}}(b_L\otimes\dotsm\otimes b_1)=\sum_{j=1}^L {\mathrm{wt}}(b_j).$$ This tensor construction is easily seen to be associative. The case of two tensor factors is given explicitly by $$\label{eq:f on two factors}
f_i(b_2\otimes b_1) = \begin{cases} f_i(b_2)\otimes b_1
& \text{if $\varepsilon_i(b_2)\ge \varphi_i(b_1)$} \\
b_2\otimes f_i(b_1) & \text{if $\varepsilon_i(b_2)<\varphi_i(b_1)$}
\end{cases}$$ and $$\label{eq:e on two factors}
e_i(b_2\otimes b_1) = \begin{cases} e_i(b_2) \otimes b_1 &
\text{if $\varepsilon_i(b_2)>\varphi_i(b_1)$} \\
b_2\otimes e_i(b_1) & \text{if $\varepsilon_i(b_2)\le \varphi_i(b_1)$.}
\end{cases}$$
Type $D_n$, $B_n$, and $C_n$ crystals {#ss:classical}
-------------------------------------
Crystals of most interest are those associated with a $U_q({\mathfrak{g}})$-module. In the case when ${\mathfrak{g}}$ is a simple Lie algebra of nonexceptional type, the crystals associated to the $U_q({\mathfrak{g}})$-modules were studied by Kashiwara and Nakashima [@KN:1994]. Here we review the combinatorial structure in terms of tableaux of the crystals of type $D_n$, $B_n$, and $C_n$ since these are the finite subalgebras relevant to the KR crystals of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$.
The Dynkin data for type $D_n$, $B_n$, and $C_n$ is given as follows. The simple roots are $$\begin{split}
\alpha_i&=\epsilon_i-\epsilon_{i+1} \qquad \text{for $1\le i<n$}\\
\alpha_n&= \begin{cases}
\epsilon_{n-1}+\epsilon_n & \text{for type $D_n$}\\
\epsilon_n &\text{for type $B_n$}\\
2 \epsilon_n & \text{for type $C_n$}
\end{cases}
\end{split}$$ and the fundamental weights are $$\begin{aligned}[3]
\text{Type $D_n$:} \qquad
&{\omega}_i = \epsilon_1+\cdots+\epsilon_i &&\text{for $1\le i\le n-2$}\\
&{\omega}_{n-1} = (\epsilon_1+\cdots+\epsilon_{n-1}-\epsilon_n)/2 &&\\
&{\omega}_n = (\epsilon_1+\cdots+\epsilon_{n-1}+\epsilon_n)/2&&\\[2mm]
\text{Type $B_n$:} \qquad
&{\omega}_i = \epsilon_1+\cdots+\epsilon_i &&\text{for $1\le i\le n-1$}\\
&{\omega}_n = (\epsilon_1+\cdots+\epsilon_{n-1}+\epsilon_n)/2&&\\[2mm]
\text{Type $C_n$:} \qquad
&{\omega}_i = \epsilon_1+\cdots+\epsilon_i &&\text{for $1\le i\le n$}
\end{aligned}$$ where $\epsilon_i\in \mathbb{Z}^n$ is the $i$-th unit standard vector. For type $D_n$ the nodes $n-1$ and $n$ are spin nodes and for type $B_n$ node $n$ is a spin node. Type $C_n$ does not have any spin node. In fact, ${\omega}_i={\Lambda}_i-\langle c,{\Lambda}_i\rangle {\Lambda}_0$ are the level 0 fundamental weights.
Let $X_n=D_n,B_n$, or $C_n$. Any $X_n$ dominant weight ${\omega}$ without a spin component can be expressed as ${\omega}=\sum_i c_i {\omega}_i$ for nonnegative integers $c_i$ and the sum runs over all $i=1,2,\ldots,n$ not a spin node. In the standard way we represent ${\omega}$ by the partition that has exactly $c_i$ columns of height $i$. Conversely, if ${\omega}$ is a partition with at most $n-2$ (resp. $n-1$ or $n$) parts we write $c_i({\omega})$ for the number of columns of ${\omega}$ of height $i$ for $X_n=D_n$ (resp. $X_n=B_n$ or $C_n$). From now on we identify partitions and dominant weights in this way.
The crystal graph $B({\omega}_1)$ of the vector representation for type $D_n$, $B_n$, and $C_n$ is given in Table \[tab:vr\] by removing the 0 arrows in the crystal $B^{1,1}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$, respectively.
$D_n^{(1)}$
------------------ --
$B_n^{(1)}$
$A_{2n-1}^{(2)}$
: \[tab:vr\]KR crystal $B^{1,1}$
The crystal $B({\omega}_\ell)$ for $\ell$ not a spin node can be realized as the connected component of $B({\omega}_1)^{\otimes \ell}$ containing the element $\ell\otimes \ell-1\otimes \cdots \otimes 1$. Similarly, the crystal $B({\omega})$ labeled by a dominant weight ${\omega}={\omega}_{\ell_1}+\cdots+{\omega}_{\ell_k}$ with $\ell_1\ge \ell_2 \ge \cdots \ge \ell_k$ not containing spin nodes can be realized as the connected component in $B({\omega}_{\ell_1}) \otimes \cdots \otimes B({\omega}_{\ell_k})$ containing the element $u_{{\omega}_{\ell_1}} \otimes \cdots \otimes u_{{\omega}_{\ell_k}}$, where $u_{{\omega}_i}$ is the highest weight element in $B({\omega}_i)$. As shown in [@KN:1994], the elements of $B({\omega})$ can be labeled by tableaux of shape ${\omega}$ in the alphabet $\{1,2,\ldots, n, \overline{n}, \ldots, \overline{1}\}$ for types $D_n$ and $C_n$ and the alphabet $\{1,2,\ldots, n, 0, \overline{n}, \ldots, \overline{1}\}$ for type $B_n$. For the explicit rules of type $D_n$, $B_n$, and $C_n$ tableaux we refer the reader to [@KN:1994]; see also [@HongKang:2002].
Kirillov–Reshetikhin crystal of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ {#ss:KR}
-----------------------------------------------------------------------------------
Let ${\mathfrak{g}}$ be an affine Kac–Moody algebra of rank $n$. Kirillov–Reshetikhin (KR) modules $W^{r,s}$, labeled by a positive integer $s$ and $r \in \{1,2,\ldots,n\}$, are finite-dimensional $U'_q({\mathfrak{g}})$-modules. See [@HKOTT:2002] for the precise definition. It was shown in collaboration with Okado [@O:2006; @OS:2007] that $W^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ has a global crystal basis $B^{r,s}$.
In [@HKOTT:2002] a conjecture is given for the decomposition of each KR module $W^{r,s}$ into its ${\mathfrak{g}}_0$-components, which was proven by Chari [@C:2001] for the nonexceptional untwisted algebras and Hernandez [@H:2007] for the twisted cases. For ${\mathfrak{g}}$ of type $D_n^{(1)}$, $B_n^{(1)}$, or $A_{2n-1}^{(2)}$ the underlying finite Lie algebra ${\mathfrak{g}}_0$ is of type $X_n=D_n,B_n$, or $C_n$, respectively. Explicitly, as a $X_n$-crystal, the KR crystal $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$ or $A_{2n-1}^{(2)}$ decomposes into the following irreducible components $$\label{eq:classical decomp}
B^{r,s} \cong \bigoplus_{{\omega}} B({\omega}),$$ for $1\le r\le n$ not a spin node. Here $B({\omega})$ is the $X_n$-crystal of highest weight ${\omega}$ and the sum runs over all dominant weights ${\omega}$ that can be obtained from $s{\omega}_r$ by the removal of vertical dominoes, where ${\omega}_i$ are the fundamental weights of $X_n$ as defined in section \[ss:classical\].
In [@FSS:2006 Corollary 4.6] we showed that Property \[A:KR\] below defines a unique affine crystal structure $\tilde{B}^{r,s}$. In this paper we construct the explicit combinatorial affine crystal structure $B^{r,s}$ of $\tilde{B}^{r,s}$ for type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ by showing that it satisfies these properties. In a subsequent paper with Okado [@OS:2007 Theorem 1.2] it is shown that the combinatorial crystals $B^{r,s}$ of this paper are indeed the crystals associated to the KR modules $W^{r,s}$. Combining these two results implies in particular that the KR crystals $B^{r,s}$ satisfy the conditions of Property \[A:KR\].
Before stating the Property \[A:KR\], we need to make a few definitions. An $I$-crystal $B$ is *regular* if, for each subset $K\subset
I$ with $|K|=2$, each $K$-component of $B$ is isomorphic to the crystal basis of an irreducible integrable highest weight $U'_q({\mathfrak{g}}_K)$-module where ${\mathfrak{g}}_K$ is the subalgebra of ${\mathfrak{g}}$ with simple roots $\alpha_i$ for $i\in K$.
The Dynkin diagram of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ all have an automorphism $\sigma$ interchanging nodes 0 and 1. See Figure \[fig:Dynkin\].
\[A:KR\] Let $\tilde{B}^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ be the crystal with the following properties:
1. \[A:classical\] As an $X_n$ crystal, $\tilde{B}^{r,s}$ decomposes according to , where $X_n=D_n,B_n$, and $C_n$, respectively.
2. \[A:regular\] $\tilde{B}^{r,s}$ is regular.
3. \[A:u\] There is a unique element $u\in \tilde{B}^{r,s}$ such that $$\begin{aligned}
\label{eq:u}
{\varepsilon}(u)=s{\Lambda}_0 \quad \text{and} \quad
{\varphi}(u) = \begin{cases} s {\Lambda}_0 & \text{for $r$ even},\\ s {\Lambda}_1 & \text{for $r$ odd.} \end{cases}\end{aligned}$$
4. \[A:auto\] $\tilde{B}^{r,s}$ admits the automorphism corresponding to $\sigma$ (also denoted $\sigma$) such that ${\varepsilon}\circ \sigma = \sigma\circ {\varepsilon}$ and ${\varphi}\circ\sigma=\sigma\circ{\varphi}$.
[@FSS:2006 Corollary 4.6] \[thm:specification\] The crystal $\tilde{B}^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ is uniquely determined by the conditions of Property \[A:KR\].
MuPAD-Combinat implementation {#sec:mupad}
=============================
The Kirillov–Reshetikhin crystals $B^{r,s}$ of type $A_n^{(1)}$, $B_n^{(1)}$, $D_n^{(1)}$, and $A_{2n-1}^{(2)}$ are implemented in MuPAD-Combinat [@HT:2003], an open source algebraic combinatorics package for the computer algebra system MuPAD [@Fuchssteiner:1996]. A KR crystal is declared via the command
combinat::crystals::kirillovReshetikhin($r$,$s$,type)
For example
>> KR:=combinat::crystals::kirillovReshetikhin(2,2,["D",4,1]):
defines the Kirillov–Reshetikhin crystal $B^{2,2}$ of type $D_4^{(1)}$. An element $t$ in this crystal is specified by its corresponding tableau via
>> t:=KR([[3],[1]])
+---+
| 3 |
+---+
| 1 |
+---+
Then $e_0(t)$ and $\sigma(t)$ are obtained via
>> t::e(0)
+----+
| -2 |
+----+
| 3 |
+----+
and
>> t::sigma()
+----+----+
| -2 | -1 |
+----+----+
| 2 | 3 |
+----+----+
The full crystal graph can be obtained by
>> KR:=combinat::crystals::kirillovReshetikhin(1,1,["D",4,1]):
>> KR::crystalGraph("filename.dot"):
followed by [dot -Tpdf filename.dot -o filename.pdf]{} in the command line. The result is presented in Figure \[fig:crystal graph\] [@graphviz] .
Throughout the paper more functionalities of the implementation are presented by examples. It should be noted that the computer implementation is not used for the proofs, but rather for illustration and intuition.
Explicit construction of $B^{r,s}$ {#sec:construction}
==================================
In this section we define a combinatorial crystal $B^{r,s}$ which satisfies all conditions of Property \[A:KR\]. By , the $X_n$-crystal structure of $B^{r,s}$ is fixed, that is, all Kashiwara operators $e_i$ and $f_i$ for $1\le i\le n$ are determined. Hence the complete crystal structure of $B^{r,s}$ is determined by specifying the affine Kashiwara crystal operators $e_0$ and $f_0$. We define $e_0$ and $f_0$ by constructing the crystal analogue $\sigma$ of the automorphism of the Dynkin diagram that interchanges the 0 and 1 node (see Definition \[def:sigma\] below). Then $$\label{eq:e0}
\begin{split}
f_0 &= \sigma \circ f_1 \circ \sigma,\\
e_0 &= \sigma \circ e_1 \circ \sigma.
\end{split}$$
\[def:comb B\] The combinatorial crystal $B^{r,s}$ is given by the classical decomposition and the affine crystal operators $e_0=\sigma\circ e_1\circ \sigma$ and $f_0=\sigma\circ f_1\circ \sigma$ with $\sigma$ as defined in Definition \[def:sigma\].
The involution $\sigma$ is first defined on $X_{n-1}$ highest weight elements and then extended to any element in $B^{r,s}$ (see section \[ss:sigma\]). In section \[ss:branching\] we discuss how the branching from $X_n$ to $X_{n-1}$ can be formulated in terms of $\pm$ diagrams. In section \[ss:bij\] a bijection between $X_{n-1}$ highest weight elements and $\pm$ diagrams is given.
$X_n$ to $X_{n-1}$ branching {#ss:branching}
----------------------------
We introduce combinatorial objects called $\pm$ diagrams to describe the branching from $X_n$ to the subalgebra of type $X_{n-1}$ obtained by removing the Dynkin node $1$. A $\pm$ diagram $P$ of shape ${\Lambda}/{\lambda}$ is a sequence of partitions ${\lambda}\subset \mu \subset {\Lambda}$ such that ${\Lambda}/\mu$ and $\mu/{\lambda}$ are horizontal strips. We depict this $\pm$ diagram by the skew tableau of shape ${\Lambda}/{\lambda}$ in which the cells of $\mu/{\lambda}$ are filled with the symbol $+$ and those of ${\Lambda}/\mu$ are filled with the symbol $-$. Write ${\Lambda}={\mathrm{outer}}(P)$ and ${\lambda}={\mathrm{inner}}(P)$ for the outer and inner shapes of the $\pm$ diagram $P$.
If ${\lambda}$ is a dominant weight for the simple Lie algebra ${\mathfrak{g}}$ write $B_{\mathfrak{g}}({\lambda})$ for the crystal graph of the highest weight $U_q({\mathfrak{g}})$-module of highest weight ${\lambda}$.
\[P:branch\] For an $X_n$ dominant weight ${\Lambda}$ with no spin weights, there is an isomorphism of $X_{n-1}$-crystals $$\begin{aligned}
B_{X_n}({\Lambda}) \cong \bigoplus_{\substack{\text{$\pm$ diagrams $P$} \\ {\mathrm{outer}}(P)={\Lambda}}}
B_{X_{n-1}}({\mathrm{inner}}(P)).\end{aligned}$$ That is, the multiplicity of $B_{X_{n-1}}({\lambda})$ in $B_{X_n}({\Lambda})$, is the number of $\pm$ diagrams of shape ${\Lambda}/{\lambda}$.
This follows directly from the branching rules for $X_n$ to $X_{n-1}$ (see for example [@FH:1991 pg. 426]).
Let ${\Lambda}={\omega}_4+{\omega}_2$. The corresponding $\pm$ diagrams are $$\begin{split}
&\yng(1,1,2,2) \quad \young(+,{\text{}},{\text{}}{\text{}},{\text{}}{\text{}}) \quad
\young({\text{}},{\text{}},{\text{}}+,{\text{}}{\text{}}) \quad \young(+,{\text{}},{\text{}}+,{\text{}}{\text{}}) \quad
\young(-,{\text{}},{\text{}}{\text{}},{\text{}}{\text{}}) \quad
\young({\text{}},{\text{}},{\text{}}-,{\text{}}{\text{}}) \quad \young(-,{\text{}},{\text{}}-,{\text{}}{\text{}}) \quad
\young(-,+,{\text{}}{\text{}},{\text{}}{\text{}}) \quad \young(+,{\text{}},{\text{}}-,{\text{}}{\text{}}) \quad\\
& \young(-,+,{\text{}}-,{\text{}}{\text{}}) \quad \young(-,{\text{}},{\text{}}+,{\text{}}{\text{}}) \quad
\young({\text{}},{\text{}},{\text{}}-,{\text{}}+) \quad
\young(-,{\text{}},{\text{}}-,{\text{}}+) \quad \young(-,+,{\text{}}+,{\text{}}{\text{}}) \quad
\young(+,{\text{}},{\text{}}-,{\text{}}+) \quad \young(-,+,{\text{}}-,{\text{}}+)
\end{split}$$ Therefore (suppressing the subscript $X_{n-1}$) we have $$\begin{aligned}
B_{X_n}({\omega}_4+{\omega}_2) &\cong
B({\omega}_4+{\omega}_2) \oplus B({\omega}_4+{\omega}_1)^{\oplus 2} \oplus B({\omega}_4)
\oplus B({\omega}_3+{\omega}_2)^{\oplus 2} \oplus \\
&\quad \,\,B({\omega}_3+{\omega}_1)^{\oplus 4} \oplus B({\omega}_3)^{\oplus 2}
\oplus B(2{\omega}_2) \oplus B({\omega}_2+{\omega}_1)^{\oplus 2} \oplus B({\omega}_2).\end{aligned}$$
$X_{n-1}$ highest weight elements {#ss:bij}
---------------------------------
We now suppose that $B^{r,s}$ is a KR crystal of type $D_n^{(1)}$, $B_n^{(1)}$, or $A_{2n-1}^{(2)}$ with $1\le r \le n$ not a spin node and ${\Lambda}$ is such that $B_{X_n}({\Lambda})$ is a $X_n$-summand in , where $X_n=D_n,B_n$, or $C_n$, respectively. That is, ${\Lambda}$ has exactly $s$ columns, and for each column, its height is at most $r$ and of the same parity as $r$. If $r$ is even then columns of height zero are allowed.
Let ${\lambda}$ be a partition associated with a “spinless" $X_{n-1}$-highest weight. We now give an explicit bijection ${\Phi}:P\mapsto b$ from the set of $\pm$ diagrams $P$ of shape ${\Lambda}/{\lambda}$ to the set of $X_{n-1}$-highest weight vectors $b$ of $X_{n-1}$-weight ${\lambda}$ in $B_{X_n}({\Lambda})$ (that is, the elements $b\in B_{X_n}({\Lambda})$ such that ${\varepsilon}_i(b)=0$ and ${\varphi}_i(b)=c_i({\lambda})$ for all $2\le i\le n$). This bijection has the property of preserving the $1$-weight, that is, $m_+(P)-m_-(P)=m_1(b)-m_{{\overline{1}}}(b)$ where $m_+(P)$ and $m_-(P)$ are respectively the number of symbols $+$ and $-$ in $P$. Here, for $b\in B_{X_n}({\Lambda})$ let $m_i(b)$ (resp. $m_{{\overline{i}}}(b)$) be the number of symbols $i$ (resp. ${\overline{i}}$) that occur in $b$, for $1\le
i\le n$. The bijection also satisfies the property that the barred letters in $b$ occur precisely in the positions in $P$ containing the symbol $-$.
The precise correspondence is given by the following algorithm. Start with a $\pm$ diagram $P$ of shape ${\Lambda}/{\lambda}$. Let $d=m_+(T)-m_-(T)$. We build the tableau $b\in B_{X_n}({\Lambda})$ as follows. Place the maximum number of symbols $1$ and $\bar{1}$ into the diagram of ${\Lambda}$, placing the $1$s into the first row and the $\bar{1}$s into the rightmost positions of $-$ in $P$, such that $m_1(b)-m_{{\overline{1}}}(b)=d$, and the placement is legal in a $X_n$-tableau. Assume that all letters $1,{\overline{1}},2,{\overline{2}},\dotsc,i-1,{\overline{i-1}}$ have already been placed. Fill the remainder of row $i-1$ with $i$’s. Then place the maximum number of symbols $i$ into row $i$ and symbols $\bar{i}$ into the available $-$ positions starting from the right, such that the $i$-weight is correct ($m_i(b)-m_{{\overline{i}}}(b)={\lambda}_{i-1}$) and the partial tableau is legal in a $X_n$-tableau.
During the placement of the letters $i$ and ${\overline{i}}$ in the computation of the map $P\mapsto b$ it suffices to check the following conditions:
1. The unbarred subtableau is semistandard with respect to the total order $1<2<\dotsm<i$;
2. A column containing $\bar{i}$ cannot contain all of the letters $1,2,\dotsc,i$;
3. there are no configurations of the form $$\begin{array}{c|c} & \bar{i}\\ &\\ i&i\end{array} \qquad
\text{or} \qquad
\begin{array}{c|c} \bar{i}&\bar{i}\\ & \\ i& \end{array}$$ in adjacent columns.
Alternatively, the bijection ${\Phi}:P\mapsto b$ from $\pm$ diagrams $P$ of shape ${\Lambda}/{\lambda}$ to the set of $X_{n-1}$-highest weight vectors $b$ of $X_{n-1}$-weight ${\lambda}$ is as follows. Place ${\overline{1}}$ in all positions in $P$ that contain a $-$ and fill the remainder of all columns by strings of the form $23\ldots k$. We move through the columns of $b$ from top to bottom, left to right. Each $+$ in $P$ (starting with the leftmost moving to the right) will alter $b$ as we move through the columns. Suppose the $+$ is at height $h$ in $P$. If one encounters a ${\overline{1}}$, replace ${\overline{1}}$ by ${\overline{h+1}}$. If one encounters a $2$, replace the string $23\ldots k$ by $12\ldots h h+2\ldots k$.
The $X_{n-1}$-highest weight elements corresponding to $$\young(+,{\text{}},{\text{}}+,{\text{}}{\text{}}) \qquad
\young(+,{\text{}},{\text{}}-,{\text{}}{\text{}}) \qquad
\young(-,{\text{}},{\text{}}+,{\text{}}{\text{}}) \qquad
\young(-,{\text{}},{\text{}}-,{\text{}}{\text{}})$$ are $$\young(4,3,22,11) \qquad
\young(4,3,2{\bar{1}},12) \qquad
\young({\bar{3}},4,33,22) \qquad
\young({\bar{1}},4,3{\bar{1}},22)$$ respectively.
\[ex:hw\] The $X_{n-1}$-highest weight element corresponding to $$\young(+-,{\text{}}+,{\text{}}{\text{}}--,{\text{}}{\text{}}{\text{}}+) \quad \text{under ${\Phi}$ is} \quad
\young(4{\bar{4}},34,23{\bar{1}}{\bar{1}},1122)\,.$$ In MuPAD-Combinat this example can be reproduced by
>> KR:=crystals::kirillovReshetikhin(4,5,["D",6,1]):
>> P:=[["+","-"],["","+"],["","","-","-"],["","","","+"]]:
>> KR::diagramsToTableaux(P)
inside the KR crystals $B^{4,5}$ of type $D_6^{(1)}$.
For a given $\pm$ diagram $P$ one can easily construct the string of operators $f_{{\mathbf{\stackrel{\rightarrow}{a}}}} := f_{a_1} f_{a_2} \cdots f_{a_\ell}$ such that ${\Phi}(P) = f_{{\mathbf{\stackrel{\rightarrow}{a}}}} u$, where $u$ is the highest weight vector in $B_{X_n}(\Lambda)$. Start with the empty string ${\mathbf{\stackrel{\rightarrow}{a}}}=()$. Scan the columns of $P$ from right to left. For each column of $P$ for which a $+$ can be added, append $(1,2, \ldots, h)$ to ${\mathbf{\stackrel{\rightarrow}{a}}}$, where $h$ is the height of the added $+$. Next scan $P$ from left to right and for each column that contains a $-$ in $P$, append to ${\mathbf{\stackrel{\rightarrow}{a}}}$ the string $(1,2,\ldots,n,n-2,n-3,\ldots, h)$ for type $D_n$, $(1,2,\ldots,n-1,n,n,n-1,\ldots,h)$ for type $B_n$, and $(1,2,\ldots,n-1,n,n-1,\ldots,h)$ for type $C_n$, where $h$ is the height of the $-$ in $P$.
\[prop:to hw\] Let $P$ be a $\pm$ diagram of shape $\Lambda/\lambda$. Let ${\mathbf{\stackrel{\rightarrow}{a}}}$ be as constructed above. Then ${\Phi}(P) = f_{{\mathbf{\stackrel{\rightarrow}{a}}}} u$, where $u$ is the $X_n$ highest weight vector in $B_{X_n}({\Lambda})$.
This follows immediately from the previous discussion.
\[ex:string\] Let $P$ be the $\pm$ diagram of Example \[ex:hw\]. Then for type $D_6$ we have ${\mathbf{\stackrel{\rightarrow}{a}}}=(1,\hspace{0.2cm}1,2,3,4,5,6,4,\hspace{0.2cm}1,2,3,4,5,6,4,3,2,\hspace{.2cm}
1,2,3,4,5,6,4,3,2)$. The command to obtain the string ${\mathbf{\stackrel{\rightarrow}{a}}}$ in MuPAD-Combinat is
>> KR::diagramsString(P,6)
Definition of $\sigma$ {#ss:sigma}
----------------------
Let $\sigma$ be the automorphism of the Dynkin diagram that interchanges 0 and 1 as shown in Figure \[fig:Dynkin\]. Here we construct the analogous automorphism on $B^{r,s}$ which, by abuse of notation, is also called $\sigma$. By construction $\sigma$ is an involution which commutes with $f_i$ and $e_i$ for $i=2,3,\ldots,n$. Hence it suffices to define $\sigma$ on $X_{n-1}$ highest weight elements where $X_{n-1}$ is the subalgebra whose Dynkin diagram is obtained from that of $X_n$ by removing node $1$. Because of the bijection ${\Phi}$ between $\pm$ diagrams and $X_{n-1}$-highest weight elements defined in section \[ss:bij\], it suffices to define the map on $\pm$ diagrams.
Let $P$ be a $\pm$ diagram of shape ${\Lambda}/{\lambda}$. Let $c_i=c_i({\lambda})$ be the number of columns of height $i$ in ${\lambda}$ for all $1\le i<r$ with $c_0=s-{\lambda}_1$. If $i\equiv r-1 \pmod{2}$, then in $P$, above each column of ${\lambda}$ of height $i$, there must be a $+$ or a $-$. Suppose there are $p_i$ of the $+$ symbols above the columns of height $i$, so that there are $(c_i-p_i)$ of the $-$ symbols. Change this to $(c_i-p_i)$ of the $+$ symbols and $p_i$ of the $-$ symbols. If $i\equiv r \pmod{2}$, then in $P$, above each column of ${\lambda}$ of height $i$, either there are no signs or a $\mp$ pair. Suppose there are $p_i$ $\mp$ pairs above the columns of height $i$. Change this to $(c_i-p_i)$ $\mp$ pairs. The result is ${\mathfrak{S}}(P)$, which has the same inner shape ${\lambda}$ as $P$ but a possibly different outer shape.
\[ex:pm\] Let $s=5$. Then we have $$P=\young(+-,{\text{}}+,{\text{}}{\text{}}--,{\text{}}{\text{}}{\text{}}+)\qquad
{\mathfrak{S}}(P)=\young(-,{\text{}},{\text{}}{\text{}}+-,{\text{}}{\text{}}{\text{}}+)\,.$$ In MuPAD-Combinat this is achieved via
>> KR::sigmaOnDiagrams(P)
To define $\sigma$ on any element $b\in B^{r,s}$, first go to the $X_{n-1}$-highest weight element in the component of $b$ using the crystal raising operators $e_i$, then use ${\mathfrak{S}}$ on the corresponding $\pm$-diagram, and go back to the same element within the new $X_{n-1}$ component using the string of $f_i$ that reverses the previously applied string of $e_i$.
\[def:sigma\] Let $b\in B^{r,s}$ and $e_{{\mathbf{\stackrel{\rightarrow}{a}}}} := e_{a_1} e_{a_2} \cdots e_{a_\ell}$ be such that $e_{{\mathbf{\stackrel{\rightarrow}{a}}}}(b)$ is a $X_{n-1}$ highest weight crystal element. Define $f_{{\mathbf{\stackrel{\leftarrow}{a}}}}:= f_{a_\ell} f_{a_{\ell-1}} \cdots f_{a_1}$. Then $$\label{eq:def sigma}
\sigma(b) := f_{{\mathbf{\stackrel{\leftarrow}{a}}}} \circ {\Phi}\circ {\mathfrak{S}}\circ {\Phi}^{-1} \circ e_{{\mathbf{\stackrel{\rightarrow}{a}}}}(b),$$ with ${\mathfrak{S}}$ as defined above and ${\Phi}$ as defined in section \[ss:bij\].
Take $$b=\young({\bar{4}}{\bar{2}},34,23{\bar{1}}{\bar{1}},1123) \; \in \; B^{4,5}$$ in type $D_6^{(1)}$. Then the corresponding $D_5$-highest weight vector $e_{{\mathbf{\stackrel{\rightarrow}{a}}}}(b)$ is the tableau of Example \[ex:hw\] where ${\mathbf{\stackrel{\rightarrow}{a}}}=(4,6,5,4,3,2,2)$. Using the results of Examples \[ex:hw\] and \[ex:pm\] we obtain $$\sigma(b) = \young({\bar{2}},{\bar{4}},334{\bar{1}},1223)\; .$$ In MuPAD-Combinat this example can be checked by
>> b:=KR([[-4,-2],[3,4],[2,3,-1,-1],[1,1,2,3]]):
>> b::sigma()
Note that the shape of $b\in B^{r,s}$, or equivalently, the classical component $B({\omega})$ in will in general change under the application of $\sigma$.
Proof of Theorem \[thm:affine\] {#sec:proof}
===============================
By Theorem \[thm:specification\] we know that $B^{r,s}$ is uniquely determined by the four conditions of Property \[A:KR\]. Hence to prove Theorem \[thm:affine\] it suffices to show that the crystal given by together with the affine crystal operators $e_0$ and $f_0$ of with $\sigma$ as in Definition \[def:sigma\] satisfies the conditions (\[A:classical\]), (\[A:regular\]), (\[A:u\]) and (\[A:auto\]) of Property \[A:KR\].
By construction Property \[A:KR\] (\[A:classical\]) holds and $\sigma$ satisfies Property \[A:KR\] (\[A:auto\]).
If $r$ is even, let $u$ be the unique element in $B(\emptyset)$. If $r$ is odd, let $u$ be the $X_n$-highest weight element in the component $B(s{\omega}_1)$. It is not hard to check that $u$ is the element of Property \[A:KR\] (\[A:u\]). Because of the condition ${\varepsilon}(u)=s\Lambda_0$, $u$ has to be a classical highest weight vector. Each classical highest weight vector $v\in B({\Lambda})$ corresponds to the $\pm$ diagram of outer shape ${\Lambda}$ with a $+$ in every column. It can be shown explicitly, that $u$ is the unique highest weight vector $v$ satisfying ${\varphi}(u)=s{\Lambda}_0$ for $r$ even and $s{\Lambda}_1$ for $r$ odd.
It remains to consider the regularity Property \[A:KR\] (\[A:regular\]). (In principle this follows from [@OS:2007 Theorem 1.2] which shows that $B^{r,s}$ is the crystal corresponding to the KR module $W^{r,s}$; the proof in [@OS:2007] uses only a special case of Lemma \[lem:e1 action\] below. We give an independent proof here). Since $B^{r,s}$ decomposes as as an $X_n$-crystal, each $K$-component of $B^{r,s}$ is an irreducible integrable highest weight $U_q({\mathfrak{g}}_K)$-crystal for $K=\{i,j\}$ with $i,j\neq 0$. For $K=\{0,i\}$ with $i\neq 1$ this is also true since $$e_0 e_i = \sigma e_1\sigma e_i = \sigma(e_1 \sigma e_i \sigma ) \sigma
= \sigma (e_1 e_i) \sigma,$$ so that by conjugation with $\sigma$ the crystal behaves just like a $\{1,i\}$-crystal. Hence it suffices to consider the case $K=\{0,1\}$ in which case we need to show that $e_0$ and $e_1$ commute since the 0 and 1 node are not adjacent in the $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ Dynkin diagrams.
For us $I=\{0,1,2,3,\ldots,n\}$ are the labels of the affine Dynkin diagram. In the following we will also consider the subalgebras $X_n$ with Dynkin set $I_{X_n} = \{1,2,3,\ldots,n\}$, $X_{n-1}$ with Dynkin set $I_{X_{n-1}} = \{2,3,\ldots,n\}$, and $X_{n-2}$ with Dynkin set $I_{X_{n-2}} = \{3,4,\ldots,n\}$. Since $e_i$ with $i\in I_{X_{n-2}}$ commutes with both $e_0$ and $e_1$, it suffices to prove the commutativity of $e_0$ and $e_1$ for $X_{n-2}$ highest weight vectors. As we have seen in section \[ss:bij\], the $X_{n-1}$-highest weight elements in the branching $X_n\to X_{n-1}$ can be described by $\pm$ diagrams. Similarly the $X_{n-2}$-highest weight elements in the branching $X_{n-1}\to X_{n-2}$ can be described by $\pm$ diagrams. Hence each $X_{n-2}$-highest weight vector is uniquely determined by a pair of $\pm$ diagrams $(P,p)$ such that ${\mathrm{inner}}(P)={\mathrm{outer}}(p)$. The diagram $P$ specifies the $X_{n-1}$-component $B_{X_{n-1}}({\mathrm{inner}}(P))$ in $B_{X_n}({\mathrm{outer}}(P))$, and $p$ specifies the $X_{n-2}$ component inside $B_{X_{n-1}}({\mathrm{inner}}(P))$. Let $\Psi$ denote the map $(P,p) \mapsto b$ from a pair of $\pm$ diagrams to a $X_{n-2}$ highest weight vector.
\[ex:Psi\] The tableau $b\in B^{4,3}$ of type $D_6^{(1)}$ corresponding to $$P=\young(-,+,{\text{}}+-,{\text{}}{\text{}}{\text{}}) \quad \text{and} \quad
p=\young(-,{\text{}}{\text{}}+) \quad \text{is} \quad
\Psi(P,p)=\young({\bar{3}},{\bar{4}},34{\bar{1}},133) \;.$$ In MuPAD-Combinat this is achieved via
>> KR:=crystals::kirillovReshetikhin(4,3,["D",6,1]):
>> P:=[["-"],["+"],["","+","-"],["","",""]]:
>> p:=[["-"],["","","+"]]:
>> KR::twoDiagramsToTableaux(P,p)
By the commutation of $e_0$ and $e_1$ is equivalent to the commutativity of the following diagram $$\label{eq:e1 sigma}
\xymatrix{
B^{r,s} \ar[d]_{\sigma} \ar[r]^{e_1} & B^{r,s} \cup \{\emptyset\} \ar[r]^{\sigma} &
B^{r,s} \cup \{\emptyset\} \ar[r]^{e_1} & B^{r,s} \cup \{\emptyset\} \ar[d]^{\sigma}\\
B^{r,s} \ar[r]_{e_1} & B^{r,s} \cup \{\emptyset\} \ar[r]_{\sigma} & B^{r,s} \cup \{\emptyset\}
\ar[r]_{e_1} & B^{r,s} \cup \{\emptyset\}
}$$
By , the action of $\sigma$ on a pair of $\pm$ diagrams $(P,p)$ is given by $({\mathfrak{S}}(P),p)$.
The operator $e_1$ will either change a 2 into a 1 or a ${\bar{1}}$ into a ${\bar{2}}$ in $b=\Psi(P,p)$. On the level of $(P,p)$ this means that either a $+$ from $p$ transfers to $P$, or a $-$ moves from $P$ to $p$. To describe the precise action of $e_1$ on $(P,p)$ perform the following algorithm:
1. Successively run through all $+$ in $p$ from left to right and, if possible, pair it with the leftmost yet unpaired $+$ in $P$ weakly to the left of it.
2. Successively run through all $-$ in $p$ from left to right and, if possible, pair it with the rightmost yet unpaired $-$ in $P$ weakly to the left.
3. Successively run through all yet unpaired $+$ in $p$ from left to right and, if possible, pair it with the leftmost yet unpaired $-$ in $p$.
\[lem:e1 action\] If there is an unpaired $+$ in $p$, $e_1$ moves the rightmost unpaired $+$ in $p$ to $P$. Else, if there is an unpaired $-$ in $P$, $e_1$ moves the leftmost unpaired $-$ in $P$ to $p$. Else $e_1$ annihilates $(P,p)$.
Let ${\mathbf{\stackrel{\rightarrow}{a}}}$ be the string of indices in $I_{X_n}=\{1,2,\ldots, n\}$ of Proposition \[prop:to hw\] such that ${\Phi}(P)=f_{{\mathbf{\stackrel{\rightarrow}{a}}}} u$, where $u$ is the $X_n$-highest weight vector of $B_{X_n}({\mathrm{outer}}(P))$. Similarly, let ${\mathbf{\stackrel{\rightarrow}{b}}}$ be the string of indices in $I_{X_{n-1}}=\{2,3,\ldots,n\}$ such that $\Psi(P,p) = f_{{\mathbf{\stackrel{\rightarrow}{b}}}} f_{{\mathbf{\stackrel{\rightarrow}{a}}}} u$. Again ${\mathbf{\stackrel{\rightarrow}{b}}}$ is determined by Proposition \[prop:to hw\], where this time $f_{{\mathbf{\stackrel{\rightarrow}{a}}}}u$ takes the role of the $X_{n-1}$-highest weight vector of $B_{X_{n-1}}({\mathrm{inner}}(P))$.
In the following we give the arguments for $X_n=D_n$. The cases $X_n=B_n,C_n$ can be treated in the same fashion. Recall that the action of $f_i$ on a tableau can be determined by reading the tableau columnwise. After determining the reduced $i$-signature, $f_i$ acts on the rightmost possible letter; see Sections \[ss:tensor\] and \[ss:classical\]. Match each string $2,3,\ldots,n,n-2,\ldots,h'$ in ${\mathbf{\stackrel{\rightarrow}{b}}}$ corresponding to a $-$ in $p$ with a string $1,2,\ldots,n,n-2,\ldots,h$ in ${\mathbf{\stackrel{\rightarrow}{a}}}$ which corresponds to a $-$ in a column in $P$ weakly to the left, meaning that $h'<h$. The raising operator $e_1$ cannot act on such a combination. This corresponds to the pairing of step (2). Similarly, match each string $2,3,\ldots,h'$ in ${\mathbf{\stackrel{\rightarrow}{b}}}$ with a string $1,2,\ldots,h$ in ${\mathbf{\stackrel{\rightarrow}{a}}}$, where $h'> h$. The operator $e_1$ cannot act on such a combination. This is equivalent to the pairing of step (1) (recall that the strings $(1,)2,3,\ldots,h$ correspond to columns in $p$ (resp. $P$) without a $+$). Finally, match every yet unmatched string $2,3,\ldots,n,n-2,\ldots,h'$ in ${\mathbf{\stackrel{\rightarrow}{b}}}$ corresponding to a $-$ in $p$ with a yet unmatched $1,2,\ldots,h$ corresponding to a column without $+$ in $P$. Again $e_1$ cannot act on such a combination. This is equivalent to step (3).
Note that $e_1$ acts on the leftmost possible unmatched letter in a tableau and all $f_i$ act rightmost. Hence $e_1$ will want to transform the smallest unmatched string $1,2,\ldots,h$ corresponding to a column in $P$ that does not contain a $+$ to a column in $p$. This corresponds to the case when an unpaired $+$ exists in $p$ which $e_1$ will move to $P$. Else $e_1$ will move the leftmost unpaired $-$ in $P$ to $p$.
Take $n>7$ and $$P=\young({\textcolor{blue}{\ominus}},{\text{}},{\text{}}{\textcolor{red}{\oplus}},{\text{}}{\text{}},{\text{}}{\text{}}{\textcolor{red}{\oplus}}-,{\text{}}{\text{}}{\text{}}{\text{}})
\quad \text{and} \quad
p=\young(+,{\text{}},{\text{}}{\textcolor{blue}{\ominus}},{\text{}}{\textcolor{red}{\oplus}},{\text{}}{\text{}}{\text{}}{\textcolor{red}{\oplus}}) \; .$$ The paired + of step (1) are circled, as are the paired - of step (2). There are no pairings for step (3). Hence $e_1(P,p)=(\tilde{P},\tilde{p})$ yields the pair $$\tilde{P}=\young({\textcolor{blue}{\ominus}},+,{\text{}}{\textcolor{red}{\oplus}},{\text{}}{\text{}},{\text{}}{\text{}}{\textcolor{red}{\oplus}}-,{\text{}}{\text{}}{\text{}}{\text{}})
\quad \text{and} \quad
\tilde{p}=\young({\text{}},{\text{}}{\textcolor{blue}{\ominus}},{\text{}}{\textcolor{red}{\oplus}},{\text{}}{\text{}}{\text{}}{\textcolor{red}{\oplus}}) \; .$$
Take $n>7$ and $$P=\young({\textcolor{blue}{\ominus}},{\text{}},{\text{}}{\textcolor{red}{\oplus}},{\text{}}{\text{}},{\text{}}{\text{}}{\text{}}+{\textcolor{blue}{\ominus}},{\text{}}{\text{}}{\text{}}{\text{}}{\text{}})
\quad \text{and} \quad
p=\young({\textcolor{green}{\raisebox{.07cm}{\scalebox{.5}{\framebox{+}}}}},{\text{}},{\text{}}{\textcolor{blue}{\ominus}},{\text{}}{\textcolor{red}{\oplus}}+,{\text{}}{\text{}}{\text{}}{\textcolor{green}{\raisebox{.07cm}{\scalebox{.5}{\framebox{-}}}}}{\textcolor{blue}{\ominus}}) \; .$$ The paired + of step (1) are circled, as are the paired - of step (2). The pairings for step (3) are denoted by a square frame. Hence $e_1(P,p)=(\tilde{P},\tilde{p})$ yields the pair $$P=\young({\textcolor{blue}{\ominus}},{\text{}},{\text{}}{\textcolor{red}{\oplus}},{\text{}}{\text{}},{\text{}}{\text{}}++{\textcolor{blue}{\ominus}},{\text{}}{\text{}}{\text{}}{\text{}}{\text{}})
\quad \text{and} \quad
p=\young({\textcolor{green}{\raisebox{.07cm}{\scalebox{.5}{\framebox{+}}}}},{\text{}},{\text{}}{\textcolor{blue}{\ominus}},{\text{}}{\textcolor{red}{\oplus}},{\text{}}{\text{}}{\text{}}{\textcolor{green}{\raisebox{.07cm}{\scalebox{.5}{\framebox{-}}}}}{\textcolor{blue}{\ominus}}) \; .$$
The commutative diagram holds for $X_{n-2}$ highest weight elements.
Since $X_{n-2}$ highest weight elements are in bijection with tuples $(P,p)$ of $\pm$ diagrams, it suffices to prove for tuples $(P,p)$. We need to distinguish the two cases when a $+$ moves from $p$ to $P$ as in and or when a $-$ moves from $P$ to $p$ as in and . The following diagrams schematically indicate the region of the $\pm$ diagrams on which $e_1$ and subsequently $\sigma$ acts. The usual $\pm$ symbols belong to $P$, whereas the circled $\pm$ symbols belong to $p$. $$\label{eq:+1}
\xymatrix{
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,6)
\psframe(0,4)(1,5) \rput(0.5,4.5){$-$}
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(4,2)(5,3) \rput(4.5,2.5){$-$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\gamma$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,5.5){$\overbrace{}^a$}
\rput(2,3.5){$\overbrace{\hspace{1.3cm}}^b$}
\rput(3.5,3.5){$\overbrace{}^c$}
\rput(4.5,3.5){$\overbrace{}^d$}
\end{pspicture}}
\qquad
\ar[d]_{\text{$\sigma \circ e_1$ if $\gamma>0$}}
\ar[dr]^{\text{$\sigma \circ e_1$ if $\gamma=0$}} \ar[r]^{\sigma} &
\qquad
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,6)
\psframe(0,4)(1,5) \rput(0.5,4.5){$-$}
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(4,2)(5,3) \rput(4.5,2.5){$-$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\gamma$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,5.5){$\overbrace{}^b$}
\rput(2,3.5){$\overbrace{\hspace{1.3cm}}^a$}
\rput(3.5,3.5){$\overbrace{}^d$}
\rput(4.5,3.5){$\overbrace{}^c$}
\end{pspicture}} \\
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,6)
\psframe(0,4)(1,5) \rput(0.5,4.5){$-$}
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(4,2)(5,3) \rput(4.5,2.5){$-$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_{\gamma-1}$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_{\epsilon+1}$}}
\rput(0.5,5.5){$\overbrace{}^{b-1}$}
\rput(2,3.5){$\overbrace{\hspace{1.3cm}}^a$}
\rput(3.5,3.5){$\overbrace{}^d$}
\rput(4.5,3.5){$\overbrace{}^{c+1}$}
\end{pspicture}}
&
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(4,6)
\psframe(0,4)(1,5) \rput(0.5,4.5){$-$}
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){$+$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$-$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_{\alpha-1}$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\delta$}
\rput(3.5,0.5){$\underbrace{}_{\epsilon}$}}
\rput(0.5,5.5){$\overbrace{}^{b-1}$}
\rput(1.5,3.5){$\overbrace{}^a$}
\rput(2.5,3.5){$\overbrace{}^d$}
\rput(3.5,3.5){$\overbrace{}^{c+1}$}
\end{pspicture}}
}$$
$$\label{eq:+2}
\xymatrix{
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,5)
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(1,3)(2,4) \rput(1.5,3.5){$-$}
\psframe(2,3)(3,4) \rput(2.5,3.5){$-$}
\psframe(3,3)(4,4) \rput(3.5,3.5){$-$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\gamma$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,4.5){$\overbrace{}^a$}
\rput(2,4.5){$\overbrace{\hspace{1.3cm}}^b$}
\rput(3.5,4.5){$\overbrace{}^c$}
\rput(4.5,2.5){$\overbrace{}^d$}
\end{pspicture}}
\qquad
\ar[d]_{\text{$\sigma \circ e_1$ if $\gamma>0$}}
\ar[dr]^{\text{$\sigma \circ e_1$ if $\gamma=0$}} \ar[r]^{\sigma} &
\qquad
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,5)
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(1,3)(2,4) \rput(1.5,3.5){$-$}
\psframe(2,3)(3,4) \rput(2.5,3.5){$-$}
\psframe(3,3)(4,4) \rput(3.5,3.5){$-$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\gamma$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,4.5){$\overbrace{}^b$}
\rput(2,4.5){$\overbrace{\hspace{1.3cm}}^a$}
\rput(3.5,4.5){$\overbrace{}^d$}
\rput(4.5,2.5){$\overbrace{}^c$}
\end{pspicture}} \\
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,5)
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(1,3)(2,4) \rput(1.5,3.5){$-$}
\psframe(2,3)(3,4) \rput(2.5,3.5){$-$}
\psframe(3,3)(4,4) \rput(3.5,3.5){$-$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_{\gamma-1}$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_{\epsilon+1}$}}
\rput(0.5,4.5){$\overbrace{}^{b-1}$}
\rput(2,4.5){$\overbrace{\hspace{1.3cm}}^a$}
\rput(3.5,4.5){$\overbrace{}^d$}
\rput(4.5,2.5){$\overbrace{}^{c+1}$}
\end{pspicture}}
&
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(4,5)
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(1,3)(2,4) \rput(1.5,3.5){$-$}
\psframe(2,3)(3,4) \rput(2.5,3.5){$-$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){$+$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_{\alpha-1}$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\delta$}
\rput(3.5,0.5){$\underbrace{}_{\epsilon}$}}
\rput(0.5,4.5){$\overbrace{}^{b-1}$}
\rput(1.5,4.5){$\overbrace{}^a$}
\rput(2.5,4.5){$\overbrace{}^d$}
\rput(3.5,2.5){$\overbrace{}^{c+1}$}
\end{pspicture}}
}$$
$$\label{eq:-1}
\xymatrix{
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,5)
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(1,3)(2,4) \rput(1.5,3.5){$-$}
\psframe(2,3)(3,4) \rput(2.5,3.5){$-$}
\psframe(3,3)(4,4) \rput(3.5,3.5){$-$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\gamma$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,4.5){$\overbrace{}^a$}
\rput(2,4.5){$\overbrace{\hspace{1.3cm}}^b$}
\rput(3.5,4.5){$\overbrace{}^c$}
\rput(4.5,2.5){$\overbrace{}^d$}
\end{pspicture}}
\qquad
\ar[d]_{\text{$\sigma \circ e_1$ if $\delta+\epsilon=c+d$}}
\ar[dr]^{\text{$\sigma \circ e_1$ if $\delta+\epsilon<c+d$}} \ar[r]^{\sigma} &
\qquad
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,5)
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(1,3)(2,4) \rput(1.5,3.5){$-$}
\psframe(2,3)(3,4) \rput(2.5,3.5){$-$}
\psframe(3,3)(4,4) \rput(3.5,3.5){$-$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\gamma$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,4.5){$\overbrace{}^b$}
\rput(2,4.5){$\overbrace{\hspace{1.3cm}}^a$}
\rput(3.5,4.5){$\overbrace{}^d$}
\rput(4.5,2.5){$\overbrace{}^c$}
\end{pspicture}} \\
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,5)
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(1,3)(2,4) \rput(1.5,3.5){$-$}
\psframe(2,3)(3,4) \rput(2.5,3.5){$-$}
\psframe(3,3)(4,4) \rput(3.5,3.5){$-$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_{\gamma+1}$}
\rput(3.5,0.5){$\underbrace{}_{\delta-1}$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,4.5){$\overbrace{}^{b}$}
\rput(2,4.5){$\overbrace{\hspace{1.3cm}}^{a+1}$}
\rput(3.5,4.5){$\overbrace{}^d$}
\rput(4.5,2.5){$\overbrace{}^{c-1}$}
\end{pspicture}}
&
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(4,5)
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(1,3)(2,4) \rput(1.5,3.5){$-$}
\psframe(2,3)(3,4) \rput(2.5,3.5){$-$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){$+$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_{\beta+1}$}
\rput(2.5,0.5){$\underbrace{}_\delta$}
\rput(3.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,4.5){$\overbrace{}^{b}$}
\rput(1.5,4.5){$\overbrace{}^{a+1}$}
\rput(2.5,4.5){$\overbrace{}^d$}
\rput(3.5,2.5){$\overbrace{}^{c-1}$}
\end{pspicture}}
}$$
$$\label{eq:-2}
\xymatrix{
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,6)
\psframe(0,4)(1,5) \rput(0.5,4.5){$-$}
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(4,2)(5,3) \rput(4.5,2.5){$-$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\gamma$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,5.5){$\overbrace{}^a$}
\rput(2,3.5){$\overbrace{\hspace{1.3cm}}^b$}
\rput(3.5,3.5){$\overbrace{}^c$}
\rput(4.5,3.5){$\overbrace{}^d$}
\end{pspicture}}
\qquad
\ar[d]_{\text{$\sigma \circ e_1$ if $\delta+\epsilon=c+d$}}
\ar[dr]^{\text{$\sigma \circ e_1$ if $\delta+\epsilon<c+d$}} \ar[r]^{\sigma} &
\qquad
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,6)
\psframe(0,4)(1,5) \rput(0.5,4.5){$-$}
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(4,2)(5,3) \rput(4.5,2.5){$-$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_\gamma$}
\rput(3.5,0.5){$\underbrace{}_\delta$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,5.5){$\overbrace{}^b$}
\rput(2,3.5){$\overbrace{\hspace{1.3cm}}^a$}
\rput(3.5,3.5){$\overbrace{}^d$}
\rput(4.5,3.5){$\overbrace{}^c$}
\end{pspicture}} \\
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(5,6)
\psframe(0,4)(1,5) \rput(0.5,4.5){$-$}
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$+$}
\psframe(4,2)(5,3) \rput(4.5,2.5){$-$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(4,1)(5,2) \rput(4.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_\beta$}
\rput(2.5,0.5){$\underbrace{}_{\gamma+1}$}
\rput(3.5,0.5){$\underbrace{}_{\delta-1}$}
\rput(4.5,0.5){$\underbrace{}_\epsilon$}}
\rput(0.5,5.5){$\overbrace{}^{b+1}$}
\rput(2,3.5){$\overbrace{\hspace{1.3cm}}^a$}
\rput(3.5,3.5){$\overbrace{}^{d-1}$}
\rput(4.5,3.5){$\overbrace{}^c$}
\end{pspicture}}
&
\scalebox{.8}{
\psset{unit=0.7cm}
\begin{pspicture}(4,6)
\psframe(0,4)(1,5) \rput(0.5,4.5){$-$}
\psframe(0,3)(1,4) \rput(0.5,3.5){$+$}
\psframe(0,2)(1,3) \rput(0.5,2.5){${\textcolor{red}{\oplus}}$}
\psframe(1,2)(2,3) \rput(1.5,2.5){${\textcolor{red}{\ominus}}$}
\psframe(2,2)(3,3) \rput(2.5,2.5){$+$}
\psframe(3,2)(4,3) \rput(3.5,2.5){$-$}
\psframe(2,1)(3,2) \rput(2.5,1.5){${\textcolor{red}{\oplus}}$}
\psframe(3,1)(4,2) \rput(3.5,1.5){${\textcolor{red}{\ominus}}$}
{\red \rput(0.5,1.5){$\underbrace{}_\alpha$}
\rput(1.5,1.5){$\underbrace{}_{\beta+1}$}
\rput(2.5,0.5){$\underbrace{}_{\delta}$}
\rput(3.5,0.5){$\underbrace{}_{\epsilon}$}}
\rput(0.5,5.5){$\overbrace{}^{b+1}$}
\rput(1.5,3.5){$\overbrace{}^a$}
\rput(2.5,3.5){$\overbrace{}^{d-1}$}
\rput(3.5,3.5){$\overbrace{}^c$}
\end{pspicture}}
}$$
Let $(P,p)$ be the $\pm$ diagram on the left top in each case. We show that $e_1$ acts in the same way on both the top right diagram $\sigma(P,p)$ and the bottom diagram $\sigma\circ e_1(P,p)$, except for a few special cases which are treated separately below. Up to the special cases, this proves . For case with $\gamma>0$, the net pairing according to the algorithm before Lemma \[lem:e1 action\] is the same for the top right and bottom diagram, except in the case when $e_1$ acts on one of the $\gamma$ pluses in $\sigma(P,p)$ and $\gamma=1$. Let us call this case S: $\gamma=1$ which will be treated separately later. For with $\gamma=0$, the condition that a plus moves from $p$ to $P$ requires that $\alpha>a$. Since $\alpha+\beta\le a+b$ we have $\beta<b$. Hence the only special case that needs to be checked explicitly, occurs when $e_1$ acts on the extra minus in the set of $c+1$ minuses in $\sigma \circ e_1(P,p)$. Call this case S: $\gamma=0$. For case with $\gamma=0$ the pairing is obviously the same for $\sigma(P,p)$ and $\sigma\circ e_1(P,p)$. For $\gamma>0$, the only special case occurs when $e_1$ acts on one of the $\gamma$ pluses in $\sigma(P,p)$ and $\gamma=1$. Call this case S: $\gamma=1$.
In case with $\delta+\epsilon=c+d$, the fact that a minus moves from $P$ to $p$ requires that $c>\epsilon$ so that $\delta>d$. Hence the action of $e_1$ is the same on $\sigma(P,p)$ and $\sigma\circ e_1(P,p)$ except for the special case when $\delta=d+1$ and $e_1$ acts on one of the $\gamma+1$ pluses in $\sigma\circ e_1(P,p)$ and on one of the $\delta$ pluses in $\sigma(P,p)$. Call this case S: $\delta=d+1$. For with $\delta+\epsilon<c+d$, the action of $e_1$ on $\sigma\circ e_1(P,p)$ and $\sigma(P,p)$ is the same. Similarly, for , a minus in the group of $d$ minuses in $P$ must move under the action of $e_1$. This requires $\beta\ge a$ or equivalently, since $\alpha+\beta\le a+b$, it follows that $\alpha\le b$. Similarly, we need $\epsilon<d$ and $\delta>c$. Hence the net pairing remains unchanged and $e_1$ acts in the same way on $\sigma\circ e_1(P,p)$ and $e_1(P,p)$, except possibly when $e_1$ acts on one of the $\delta$ pluses in $\sigma(P,p)$, but on one of the $\gamma+1$ pluses in $\sigma\circ e_1(P,p)$. But for $\delta+\epsilon=c+d$ this is impossible.
We are now going to treat the special cases. For S: $\gamma=1$, case $\gamma=1$ applies to $\sigma(P,p)$, and case $\gamma=0$ applies to both $\sigma\circ e_1(P,p)$ and $\sigma\circ e_1\circ\sigma(P,p)$. It can be shown explicitly that commutes. The same happens for S: $\gamma=1$ with replaced by everywhere. For S: $\gamma=0$, $e_1$ acts on one of the $c+1$ minuses in $\sigma\circ e_1(P,p)$. Then case with $\delta>0$ applies for $\sigma\circ e_1(P,p)$, case with $\delta>0$ applies for $\sigma(P,p)$ and case with $\gamma>0$ applies for $\sigma\circ e_1 \circ \sigma(P,p)$. Again it can be shown explicitly that commutes.
For S: $\delta=d+1$, $e_1$ acts on the group of $\gamma+1$ pluses in $\sigma\circ e_1(P,p)$, but the group of $\delta$ pluses in $\sigma(P,p)$. Then case with $\gamma>0$ applies for $\sigma\circ e_1(P,p)$, case with $\gamma=0$ applies for $\sigma(P,p)$, and case with $\delta+\epsilon<c+d$ applies for $\sigma\circ e_1\circ \sigma(P,p)$. Again it follows that commutes.
Perfectness {#sec:perfect}
===========
In this section we show that the KR crystals $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ are perfect. The notion of perfect crystal was introduced in [@KKMMNN:1992] in order to construct a path realization of highest weight modules. In a subsequent paper [@KKMMNN:1992a] a perfect crystal of arbitrary level was given for every nonexceptional affine algebra. Further perfect crystals were found and studied in [@BFKL:2006; @HN:2006; @JMO:2000; @Ka:2002; @Ko:1999; @KMOY:2006; @NS:2005; @Y:1998] for example. In [@HKOTT:2002; @HKOTY:1999] it is conjectured that $B^{r,s}$ is perfect of level $s/t_r$ if $s/t_r$ is an integer, where $t_r=\max(1,2/(\alpha_r,\alpha_r))$.
To define perfect crystals, we need a few preliminary definitions. Define the set of level $\ell$ weights to be $(P_{{\mathrm{cl}}}^{+})_{\ell}=\{{\lambda}\in P_{{\mathrm{cl}}}^{+} \mid \langle c,{\lambda}\rangle=\ell\}$. For a crystal basis $B$, we define $B_{\min}$ to be the set of crystal basis elements $b$ such that $\langle c,{\varepsilon}(b)\rangle$ is minimal over $b\in B$. We say that $B\otimes B$ is connected if it contains only one irreducible component.
\[def:perfect\] A crystal $B$ is a perfect crystal of level $\ell$ if:
1. $B\otimes B$ is connected;
2. there exists ${\lambda}\in P_{{\mathrm{cl}}}$ such that ${\mathrm{wt}}(B)\subset
{\lambda}+\sum_{i\neq0}\mathbb{Z}_{\leq0}\alpha_{i}$ and $\#(B_{{\lambda}})=1$;
3. there is a finite-dimensional irreducible $U_q({\mathfrak{g}})$-module $V$ with a crystal base whose crystal graph is isomorphic to $B$;
4. for any $b\in B$, we have $\langle c,{\varepsilon}(b)\rangle \geq
\ell$;
5. the maps ${\varepsilon}$ and $\varphi$ from $B_{\min}$ to $(P_{{\mathrm{cl}}}^{+})_{\ell}$ are bijective.
We use the notation ${\mathrm{lev}}(B)$ to indicate the level of the perfect crystal $B$.
We show here that $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ satisfies conditions (1), (2), (4), and (5) of Definition \[def:perfect\] with level $\ell=s$. Condition (3) is proved in [@OS:2007]. Let us first give an explicit construction of all elements in $B^{r,s}_{\min}$. Let $\Lambda\in (P_{\mathrm{cl}}^+)_s$ be a dominant weight of level $s$. That is $\Lambda=\ell_0\Lambda_0+\ell_1\Lambda_1+\cdots+\ell_n \Lambda_n$ with ${\mathrm{lev}}({\Lambda}):=\langle c, {\Lambda}\rangle = s$, where explicitly $$\begin{aligned}
{\mathrm{lev}}({\Lambda}) &= \ell_0+\ell_1+2\ell_2+2\ell_3+\cdots+2\ell_{n-2}+\ell_{n-1}+\ell_n && \text{for type $D_n^{(1)}$}\\
{\mathrm{lev}}({\Lambda}) &= \ell_0+\ell_1+2\ell_2+2\ell_3+\cdots+2\ell_{n-2}+2\ell_{n-1}+\ell_n && \text{for type $B_n^{(1)}$}\\
{\mathrm{lev}}({\Lambda}) &= \ell_0+\ell_1+2\ell_2+2\ell_3+\cdots+2\ell_{n-2}+2\ell_{n-1}+2\ell_n && \text{for type $A_{2n-1}^{(2)}$.}\end{aligned}$$
To a given fundamental weight $\Lambda_k$ we may associate the following $\pm$ diagram $$\label{eq:La pm}
{\mathrm{diagram}}: \Lambda_k \mapsto \begin{cases}
\phantom{k+1\left\{\right.} \emptyset & \text{if $r$ is even and $k=0$}\\[1mm]
\phantom{k+1\left\{\right.} \young(-,+) & \text{if $r$ is even and $k=1$}\\[3mm]
\phantom{k+1\left\{\right.} \young(+) & \text{if $r$ is odd and $k=0$}\\[2mm]
\phantom{k+1\left\{\right.} \young(-) & \text{if $r$ is odd and $k=1$}\\[2mm]
k+1\left\{\young(-,+,{\text{}}{\text{}},{\text{}}{\text{}})\right.
& \text{if $k\not \equiv r \bmod{2}$ and $2\le k\le r$}\\[8mm]
\qquad k\left\{\young(+-,{\text{}}{\text{}},{\text{}}{\text{}},{\text{}}{\text{}})\right.
& \text{if $k \equiv r \bmod{2}$ and $2\le k\le r$}\\[8mm]
\qquad r\left\{\young({\text{}}{\text{}},{\text{}}{\text{}},{\text{}}{\text{}},{\text{}}{\text{}})\right.
& \begin{array}{l} \text{if $r<k\le n-2$ for type $D_n^{(1)}$}\\
\text{if $r<k\le n-1$ for type $B_n^{(1)}$}\\
\text{if $r<k\le n$ for type $A_{2n-1}^{(2)}$} \end{array} \\[8mm]
\qquad r\left\{\young({\text{}},{\text{}},{\text{}},{\text{}})\right.
& \begin{array}{l} \text{if $k=n-1,n$ for type $D_n^{(1)}$}\\
\text{if $k=n$ for type $B_n^{(1)}$.} \end{array}
\end{cases}$$ This map can be extended to any dominant weight $\Lambda=\ell_0\Lambda_0+\cdots+
\ell_n\Lambda_n$ by concatenating the columns of the $\pm$ diagrams of each piece.
\[ex:minimal\] Let $s=9$ and $\Lambda=\Lambda_0+2\Lambda_1+\Lambda_2+\Lambda_3+\Lambda_5$ of type $D_8^{(1)}$. Then $$\begin{split}
{\mathrm{diagram}}(\Lambda) &= \young({\text{}}{\text{}}+--,{\text{}}{\text{}}{\text{}}{\text{}}+,{\text{}}{\text{}}{\text{}}{\text{}}{\text{}}{\text{}}+--)
\quad \text{for $r=3$ and}\\
{\mathrm{diagram}}(\Lambda) &= \young({\text{}}{\text{}}-,{\text{}}{\text{}}+,{\text{}}{\text{}}{\text{}}{\text{}}+---,{\text{}}{\text{}}{\text{}}{\text{}}{\text{}}{\text{}}++)
\quad \text{for $r=4$.}
\end{split}$$
To every fundamental weight $\Lambda_k$ we also associate a string of operators $f_i$ with $i\in\{2,3,\ldots,n\}$ as follows. Let $T(\Lambda_k)$ be the tableau assigned to $\Lambda_k$ as $$T(\Lambda_k) = \begin{cases}
\quad \emptyset & \text{if $r$ is even and $k=0$}\\[1mm]
\quad \young({\bar{2}},2) & \text{if $r$ is even and $k=1$}\\[4mm]
\quad \young(1) & \text{if $r$ is odd and $k=0$}\\[1mm]
\quad \young({\bar{1}}) & \text{if $r$ is odd and $k=1$}\\[2mm]
\quad \begin{array}{|c|c|} \cline{1-1} \overline{k+1} & \multicolumn{1}{c}{} \\ \cline{1-1}
k+1 & \multicolumn{1}{c}{}\\ \hline k&\bar{2}\\
\hline \vdots & \vdots\\ \hline 2& \bar{k}\\
\hline \end{array} & \text{if $2\le k\le r$ and $k\not \equiv r\bmod 2$} \\[1.4cm]
\quad \begin{array}{|c|c|} \hline k & {\bar{1}}\\ \hline \vdots & \vdots\\[2mm] \hline 1& \bar{k}\\
\hline \end{array} & \text{if $2\le k\le r$ and $k\equiv r\bmod 2$} \\[1cm]
\quad \begin{array}{|c|c|} \hline k&\overline{k-r+1}\\ \hline \vdots&\vdots\\ \hline
k-r+1 & \overline{k} \\ \hline \end{array}
& \begin{array}{l} \text{if $r<k\le n-2$ for type $D_n^{(1)}$}\\
\text{if $r<k\le n-1$ for type $B_n^{(1)}$}\\
\text{if $r<k\le n$ for type $A_{2n-1}^{(2)}$} \end{array} \\[0.9cm]
\quad \left.\begin{array}{|c|} \hline \vdots \\ \hline n\\ \hline
\overline{n}\\ \hline n \\ \hline\end{array} \right\} r
& \text{for $k=n-1$ for type $D_n^{(1)}$}\\[1.1cm]
\quad \text{previous case with $n\leftrightarrow \bar{n}$}
& \text{for $k=n$ for type $D_n^{(1)}$}\\[.2cm]
\quad \left.\begin{array}{|c|} \hline 0 \\ \hline \vdots \\ \hline 0 \\ \hline \end{array} \right\} r
& \text{for $k=n$ for type $B_n^{(1)}$}
\end{cases}$$ Then $f(\Lambda_k)$ for $0\le k\le n$ is defined such that $T(\Lambda_k) = f(\Lambda_k) Y(\Lambda_k)$, where $Y(\Lambda_k)$ is the $X_{n-1}$-highest weight tableau corresponding to ${\mathrm{diagram}}(\Lambda_k)$. Note that in fact $f(\Lambda_0)
=f(\Lambda_1)=1$.
The minimal element $b$ in $B^{r,s}$ that satisfies $\varepsilon(b)=\Lambda$ can now be constructed as follows $$b = f(\Lambda_n)^{\ell_n} \cdots f(\Lambda_2)^{\ell_2} {\Phi}({\mathrm{diagram}}(\Lambda)).$$
From the condition that ${\mathrm{wt}}(b)={\varphi}(b)-{\varepsilon}(b)$ it is not hard to see that ${\varphi}(b)={\varepsilon}(b)$ for $b\in B_{\min}$ and $r$ even. For $r$ odd, we have ${\varphi}(b)=\sigma \sigma' {\varepsilon}(b)$ for $b\in B_{\min}$, where $\sigma$ is the Dynkin diagram automorphism interchanging nodes 0 and 1 and $\sigma'$ is the Dynkin diagram automorphism interchanging nodes $n-1$ and $n$.
Continuing Example \[ex:minimal\], we find that the element $b\in B^{r,s}$ such that $\varepsilon(b)=\Lambda$ is $$\begin{split}
&\young(35{\bar{5}}{\bar{1}}{\bar{1}},225{\bar{5}}{\bar{2}},1115{\bar{5}}{\bar{3}}{\bar{2}}{\bar{1}}{\bar{1}}) \quad \text{for $r=3$ and}\\
&\young(5{\bar{5}}{\bar{2}},35{\bar{5}},225{\bar{5}}{\bar{2}}{\bar{1}}{\bar{1}}{\bar{1}},11125{\bar{5}}{\bar{3}}{\bar{2}}) \qquad \text{for $r=4$.}
\end{split}$$ In MuPAD-Combinat these examples can be reproduced via
>> KR:=combinat::crystals::kirillovReshetikhin(3,9,["D",8,1]):
>> KR::minimalElement([1,2,1,1,0,1,0,0,0])
and
>> KR:=combinat::crystals::kirillovReshetikhin(4,9,["D",8,1]):
>> KR::minimalElement([1,2,1,1,0,1,0,0,0])
where ${\Lambda}$ is entered via its coordinate vector in the basis of ${\Lambda}_i$.
The list of all minimal elements of $B^{2,2}$ of type $D_4^{(1)}$ is given by $$\emptyset \quad \young(4,{\bar{4}}) \quad \young({\bar{4}},4) \quad \young({\bar{2}},2) \quad
\young({\bar{2}}{\bar{1}},12) \quad \young(2{\bar{1}},1{\bar{2}}) \quad \young(4{\bar{1}},1{\bar{4}}) \quad
\young({\bar{4}}{\bar{1}},14) \quad \young(3{\bar{2}},2{\bar{3}}) \quad \young(4{\bar{3}},3{\bar{4}}) \quad
\young({\bar{4}}{\bar{3}},34) \quad .$$ In MuPAD-Combinat the minimal elements are listed via
>> KR := combinat::crystals::kirillovReshetikhin(2,2,["D",4,1]):
>> KR::minimal()
The crystals $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ are perfect of level $s$.
Conditions (1) and (2) of Definition \[def:perfect\] follow from [@FSS:2006]. Condition (3) is proven in [@OS:2007 Theorem 1.2]. We have shown that the crystals $B^{r,s}$ of type $D_n^{(1)}$, $B_n^{(1)}$, and $A_{2n-1}^{(2)}$ constructed in this paper satisfy [@FSS:2006 Assumption 1]. Hence by [@FSS:2006 Corollary 6.1] the tensor product $B^{r,s}\otimes B^{r,s}$ is connected. Furthermore, it follows from [@FSS:2006 Theorem 4.4] that $\lambda=s\omega_r$ is the weight required for property (2).
We prove conditions (4) and (5) by induction on $s$. Suppose the classical dominant weight ${\omega}$ is such that $B({\omega}) \subset B^{r,s-1}$ according to . Since also $B({\omega})\subset B^{r,s}$ this defines an embedding $\iota:B^{r,s-1} \hookrightarrow B^{r,s}$. The classical component remains unchanged under this embedding, so that ${\varepsilon}_i(b) = {\varepsilon}_i(\iota(b))$ for all $b\in B^{r,s-1}$ and $i=1,2,\ldots,n$. Furthermore, ${\varepsilon}_0(\iota(b)) = {\varepsilon}_0(b)+1$ for $b\in B^{r,s-1}$ due to the fact that $\sigma$ adds an extra $\young(-,+)$ to the corresponding $\pm$ diagram. Hence by induction $\langle c, {\varepsilon}(b) \rangle \ge s$ for $b\in B({\omega})\subset B^{r,s}$ with width of ${\omega}$ strictly less than $s$.
For (4) it remains to show that $\langle c, {\varepsilon}(b) \rangle \ge s$ for $b\in B({\omega})\subset B^{r,s}$ with ${\omega}$ of width $s$. Suppose that $b$ is in the $X_{n-2}$-highest weight component corresponding to the pair of $\pm$ diagrams $(P,p)$. If $p$ is not of width $s$, then $P$ has at least one column of the form $\young(-,+)$, $\young(-)$, or $\young(+)$. In the first case $\sigma(b)\in B({\omega})\subset B^{r,s}$ with ${\omega}$ of width strictly less than $s$. Hence by the same arguments as above $\langle c, {\varepsilon}(\sigma(b)) \rangle = \langle c, {\varepsilon}(b) \rangle \ge s$. If $P$ contains a column $\young(-)$, then the last column of $b$ is $\young({\bar{1}})$. In this case $b$ is in the image of the embedding $\tilde{\iota} : B^{r,s-1} \hookrightarrow B^{r,s}$ for $r$ odd, which adds a column $\young({\bar{1}})$. Under this embedding we have ${\varepsilon}_i(\tilde{b})
= {\varepsilon}_i(\tilde{\iota}(\tilde{b}))$ for $i =0,2,3,\ldots,n$ and ${\varepsilon}_1(\iota(\tilde{b})) = {\varepsilon}_1(\tilde{b})+1$ for all $\tilde{b}\in B^{r,s-1}$. Hence by induction on $s$, we have $\langle c, {\varepsilon}(b) \rangle \ge s$. Similarly, if $P$ contains a column $\young(+)$, then ${\mathfrak{S}}(P)$ contains a column $\young(-)$ and by the above arguments $\langle c, {\varepsilon}(b) \rangle = \langle c, {\varepsilon}(\sigma(b)) \rangle \ge s$. Hence from now on we may assume that $p$ has width $s$.
Let $k_+, k_-, k_\mp$, and $k_0$ be the number of columns in $p$ containing only a $+, -, \mp$, and no $+$ or $-$, respectively. Then we have $$\left.
\begin{array}{ll}
\sum_{i=2}^{n-2} 2{\varepsilon}_i(b) + {\varepsilon}_{n-1}(b) + {\varepsilon}_n(b) & \text{for type $D_n^{(1)}$}\\
\sum_{i=2}^{n-1} 2{\varepsilon}_i(b) + {\varepsilon}_n(b) & \text{for type $B_n^{(1)}$}\\
\sum_{i=2}^n 2{\varepsilon}_i(b) & \text{for type $A_{2n-1}^{(2)}$}
\end{array}
\right\}
\ge 2k_- + k_\mp + k_0 = s + s' - 2k_+$$ where $s'=s-k_0-k_\mp$. This inequality follows from the fact that for an $X_{n-2}$ highest weight vector corresponding to $p$ each $-$ and each column without a $+$ contributes 2 to the sum on the left-hand side, which is $(2k_- +2 k_{\mp}) + (2 k_0 + 2 k_-)=4k_- + 2 k_\mp + 2 k_0$. Using products of $f_i$ it is possible to either change a ${\varepsilon}_2$ to a ${\varepsilon}_{n-1}$ or ${\varepsilon}_n$ for type $D_n^{(1)}$ and $B_n^{(1)}$, or to match two contributions in two different columns (an $\bar{i}$ can support a $i$ in the tableau). In both cases the contribution can at most be halved.
By Lemma \[lem:e1 action\] we obtain $$\begin{split}
{\varepsilon}_1(b) &\ge \max(k_+-K_+ + K_- - k_-,0),\\
{\varepsilon}_0(b) &\ge \max(k_+-K_- + K_+ - k_-,0),
\end{split}$$ where $K_+$ and $K_-$ are the total number of $+$ and $-$ in $P$. Hence $$\label{eq:lev ineq}
\langle c, {\varepsilon}(b) \rangle \ge
s+s'-2k_++2\max(k_+-k_-,0).$$ If $k_+\le s'/2$, then certainly $\langle c, {\varepsilon}(b) \rangle \ge s$ by . If $k_+>s'/2$, then $k_-<s'/2$ since $k_+ + k_- = s'$, so that by we have $\langle c, {\varepsilon}(b) \rangle \ge s+s'-2k_->s$. This concludes the proof of property (4).
Finally let us prove property (5). For all weights ${\Lambda}\in (P_{{\mathrm{cl}}}^{+})_s$ which contain $\Lambda_0$ (resp. $\Lambda_1$), we know by induction on $s$ that ${\Lambda}-{\Lambda}_0$ (resp. ${\Lambda}-{\Lambda}_1$) is in bijection with a minimal element $b\in B^{r,s-1}$, and hence the weight ${\Lambda}$ corresponds to $\iota(b)$ or $\tilde{\iota}(\sigma(b))$ (resp. $\iota(\sigma(b))$ or $\tilde{\iota}(b)$) which is a minimal element in $B^{r,s}$. Hence we may restrict our attention to $b\in B^{r,s}$ in an $X_{n-2}$ highest weight component labeled by a tuple of $\pm$ diagrams $(P,p)$ such that $p$ has width $s$. By the arguments following it is clear that $\langle c, {\varepsilon}(b) \rangle = s$ is only possible if $k_+=s'/2=k_-$. In this case also $K_+=K_-$, and ${\varepsilon}_0(b)={\varepsilon}_1(b)=0$ from the previous inequalities. Lemma \[lem:e1 action\] then demands that for ${\varepsilon}_0(b)={\varepsilon}_1(b)=0$ to hold, the diagrams $P$ and $p$ are symmetrical, meaning that the number of columns of the same height containing only $+$ and only $-$ is equal, and the number of columns of height $h$ containing $\mp$ and the number of empty columns of height $h-2$ is equal. Therefore $P$ is indeed the concatenation of columns as in . Then it is not hard to see that one can only obtain a minimal element from $P$ by the application of a product of appropriate $f({\Lambda}_k)$ since otherwise there is either a gap in the string amounting to an additional contribution to $\langle c,{\varepsilon}(b) \rangle$ or one of the $-$ in $P$ gives an additional contribution to $\langle c,{\varepsilon}(b) \rangle$. The proof that ${\varphi}: B_{\min} \to (P_{\mathrm{cl}})_s$ is a bijection is similar.
[99]{}
G. Benkart, I. Frenkel, S.-J. Kang, H. Lee, *Level 1 perfect crystals and path realizations of basic representations at $q=0$*, Int. Math. Res. Not. 2006, Art. ID 10312, 28 pp.
V. Chari, *On the fermionic formula and the Kirillov–Reshetikhin conjecture*, Internat. Math. Res. Notices **12** (2001) 629–654.
V. Chari, A. Pressley, *Quantum affine algebras and their representations*, Representations of groups (Banff, AB, 1994), 59–78, CMS Conf. Proc., **16**, Amer. Math. Soc., Providence, RI, 1995.
V. Chari, A. Pressley, *Twisted quantum affine algebras*, Comm. Math. Phys. **196** (1998) 461–476.
G. Fourier, P. Littelmann, *Tensor product structure of affine Demazure modules and limit constructions*, Nagoya Math. J. **182** (2006), 171–198.
G. Fourier, A. Schilling, M. Shimozono, *Demazure structure inside Kirillov-Reshetikhin crystals*, J. Algebra **309** (2007), 386–404.
B. Fuchssteiner et al., MuPAD User’s Manual - MuPAD Version 1.2.2, John Wiley and sons, Chichester, New York, 1996. [http://www.mupad.de]{}
W. Fulton, J. Harris, Representation Theory. A First Couse, Graduate Texts in Mathematics, Springer Verlag, 1991. ISBN: 3-540-97495-4.
D. Hernandez, *Kirillov-Reshetikhin conjecture : the general case*, preprint arXiv:0704.2838v1 \[math.QA\].
G. Hatayama, A. Kuniba, M. Okado, T. Takagi, Z. Tsuboi, *Paths, crystals and fermionic formulae*, MathPhys odyssey, 2001, 205–272, Prog. Math. Phys., **23**, Birkhäuser Boston, Boston, MA, 2002.
G. Hatayama, A. Kuniba, M. Okado, T. Takagi, Y. Yamada, *Remarks on fermionic formula*, Recent developments in quantum affine algebras and related topics (Raleigh, NC, 1998), 243–291, Contemp. Math., **248**, Amer. Math. Soc., Providence, RI, 1999.
J. Hong, S.-J. Kang, Introduction to Quantum Groups and Crystal Bases, Graduate Studies in Mathematics, Volume **42**, American Mathematical Society, 2002.
D. Hernandez, H. Nakajima, *Level 0 monomial crystals*, Nagoya Math. J. **184** (2006), 85–153.
F. Hivert, N. M. Thiéry, *MuPAD-Combinat, an Open-Source Package for Research in Algebraic Combinatorics*, Séminaire Lotharingien de Combinatoire **51** (2003) \[B51z\] (70 pp).
N. Jing, K.C. Misra, M. Okado, *$q$-wedge modules for quantized enveloping algebras of classical type*, J. Algebra **230** (2000), no. 2, 518–539.
V.G. Kac, Infinite-dimensional Lie algebras. Third edition. Cambridge University Press, Cambridge, 1990. xxii+400 pp. ISBN: 0-521-37215-1.
M. Kashiwara, *On crystal bases*, in: Representations of groups (Banff, AB, 1994), 155–197, CMS Conf. Proc., 16, Amer. Math. Soc., Providence, RI, 1995.
M. Kashiwara, *On level-zero representation of quantized affine algebras*, Duke Math. J. **112** (2002), no. 1, 117–195.
M. Kashiwara and T. Nakashima, *Crystal graphs for representations of the $q$-analogue of classical Lie algebras*, J. Alg. **165** (1994) 295–345.
Y. Koga, *Level one perfect crystals for $B^{(1)}_n,C^ {(1)}_n$, and $D^ {(1)}_n$*, J. Algebra **217** (1999), no. 1, 312–334.
S.-J. Kang, M. Kashiwara, K. C. Misra, *Crystal bases of Verma modules for quantum affine Lie algebras*, Compositio Math. **92** (1994) 299–325.
S-J. Kang, M. Kashiwara, K. C. Misra, T. Miwa, T. Nakashima, A. Nakayashiki, *Affine crystals and vertex models*, Int. J. Mod. Phys. A [**7**]{} (suppl. 1A) (1992) 449–484.
S.-J. Kang, M. Kashiwara, K.C. Misra, T. Miwa, T. Nakashima, A. Nakayashiki, *Perfect crystals of quantum affine Lie algebras*, Duke Math. J. **68** (1992), no. 3, 499–607.
M. Kashiwara, K. .C. Misra, M. Okado, D. Yamada, *Perfect Crystals for $U_q(D_4^{(3)})$*, preprint math.QA/0610873.
S. Naito, D. Sagaki, *Construction of perfect crystals conjecturally corresponding to Kirillov–Reshetikhin modules over twisted quantum affine algebras*, Comm. Math. Phys. **263** (2006), no. 3, 749–787.
M. Okado, *Existence of Crystal Bases for Kirillov-Reshetikhin Modules of Type $D$*, preprint math.QA/0610874.
M. Okado, A. Schilling, *Existence of Kirillov-Reshetikhin crystals for nonexceptional types*, preprint arXiv:0706.2224.
M. Shimozono, *Affine type A crystal structure on tensor products of rectangles, Demazure characters, and nilpotent varieties*, J. Algebraic Combin. **15** (2002), no. 2, 151–187.
P. Sternberg, *Applications of crystal bases to current problems in representation theory*, PhD thesis, UC Davis 2006 (available at math.QA/0610704).
P. Sternberg, A. Schilling, *Finite-dimensional crystals $B^{2,s}$ for quantum affine algebras of type $D_n^{(1)}$*, J. Alg. Combin. **23** (2006) 317–354.
S. Yamane, *Perfect crystals of $U_q(G^{(1)}_2)$*, J. Algebra **210** (1998), no. 2, 440–486.
The crystal graph drawings rely on graphviz [http://www.graphviz.org/]{}
[^1]: *Date:* April 2007, revised October 2007
[^2]: Partially supported by NSF grant DMS-0501101.
| {
"pile_set_name": "ArXiv"
} |
=-02cm=23.cm=16.cm =0.25cm=0.25cm
\#1\#2\#3[255 -\#1255 height \#2255 depth \#3255 \#1255]{} \#1\#2\#3\#4[ [255]{} [255]{} [255]{} [255]{}]{} ø ¶[[P]{}]{}
ENSLAPP-L-422/93\
JINR E2-93-72
0.5cm
N=2 Superconformal Affine Liouville Theory
0.5cm
[E. Ivanov]{}
.5cm
Bogoliubov Theoretical Laboratory, JINR,
Dubna, Head Post Office, P.O. Box 79,
101000, Moscow, Russia
[F. Toppan]{}
.5cm
Laboratoire de Physique Théorique ENSLAPP,
Ecole Normale Supérieure de Lyon,
46 Allée d’Italie, 69007 Lyon, France
.5cm 1.5cm
**Abstract**
.5cm We present a new supersymmetric integrable model: the $N=2$ superconformal affine Liouville theory. It interpolates between the $N=2$ super Liouville and $N=2$ super sine-Gordon theories and possesses a Lax representation on the complex affine Kac-Moody superalgebra ${\hat {sl(2| 2)^{(1)}}}$. We show that the higher spin $W_{1+\infty}$-type symmetry algebra of ordinary conformal affine Liouville theory extends to a $N=2\; W_{1/2 + \infty}$-type superalgebra.
1
Introduction
============
It is a well-established fact \[1-4\] that an important class of conformally invariant integrable models can be produced starting from constrained WZNW theories based on affine Kac-Moody algebras (Conformally Affine Toda or CAT theories). The standard massive affine Toda theories can be recovered as a special limit of these models.
At the bosonic level any given simple Lie algebra admits an affine extension and the associate CAT model can be defined, with a Lax pair given on this affine extension.
In the supersymmetric case this is not always possible: to obtain a manifestly supersymmetric integrable model one needs to start from superalgebras admitting a set of purely fermionic simple roots, the reason for that being the fact that the Lax pair operators should be fermionic objects in this case. So, to learn which superalgebras admit an $N=1$ superextended integrable CAT model it is sufficient to look at the classification of superalgebras and their root systems as is given in [@FSS]. For instance, the $N=1$ superconformal affine Liouville (super CAL) theory [@TZ] which generalizes the massive super sinh-Gordon model [@OBL] is associated with the twisted affine superalgebra ${\hat { {osp(2|2)}^{(2)}}}$ having two fermionic simple roots.
As for the $N=2$ case, the demand that a superalgebra admits a set of purely fermionic simple roots is necessary but not sufficient; a detailed discussion of which superalgebras can give rise to integrable models possessing a second supersymmetry has been given in \[7-9\]. The simplest $N=2$ superconformal integrable model is the $N=2$ super Liouville theory [@IK] associated with the superalgebra $sl(2|1)$ and the simplest $N=2$ massive integrable model is the $N=2$ super sine-Gordon theory associated with the loop superalgebra $sl(2|2)^{(1)}$ .
In this letter we present an integrable $N=2$ superconformal affine Liouville model which reduces to the $N=2$ super sine-Gordon and the $N=2$ super Liouville models in two special limits. We will work in a manifestly supersymmetric $N=2$ superfield approach. At first we will introduce our model in the lagrangian formulation and then we will discuss it in a Lax pair context [@LS]. We will also construct supercurrents which generate an infinite-dimensional higher spin symmetry superalgebra generalizing the $W_{1+\infty}$ type symmetry algebra of the ordinary CAT models [@AFGZ2].
The action and equations of motion
==================================
We will work in a manifestly supersymmetric formalism.
Let $x^{\pm\pm}$ denote the bosonic $2D$ light-cone coordinates, then the $N=2\;2D$ superspace is parametrized by $x^{\pm\pm}$ and the complex fermionic coordinates $\theta^\pm$ (and their conjugate ${\overline{\theta}^\pm}$). The $N=1$ superspace is recovered by letting ${{\theta^\pm}}={\overline{\theta}^\pm}$. The $N=2$ spinor derivatives $D_{\pm},{\overline D}_{\pm}$ are defined as: D\_ &=&- i\^\_\
[D]{}\_ &=&-+ i[ ]{}\^\_ . The only non-vanishing bracket between them is given by {D\_, [D]{}\_} &=& 2i \_ .In particular we have \^2 = [[D]{}\_]{}\^2 =0&& .An $N=2$ chiral superfield $\Psi$, the simplest matter $N=2$ supermultiplet, can be defined by the constraint = 0, &&while its conjugate satisfies \_\^ =0 and so it is an anti-chiral $N=2$ superfield.
The bosonic conformal affine Liouville theory is formulated in terms of three real bosonic fields [@BB]. A natural way to promote it to a $N=2$ supersymmetric theory is to put these fields in proper minimal $N=2$ supermultiplets. So we will define the $N=2$ superconformal affine Liouville theory via three $N=2$ chiral superfields $\Phi,\;\Lambda,\;\Sigma$. By analogy with the $N=0$ and $N=1$ cases we choose the action ${\cal S}$ to be &=& d\^2 x d\^2[\_+]{} d\^2[\_-]{} { +++\
&& e\^\_+\_-+e\^[-]{}\_+\_- +\^ e\^[\^]{}\_-[\_+]{}+ \^ e\^[\^ -\^]{}\_-\_+ }. \[action1\] Without loss of generality, the constants $\alpha ,\;\beta $ can be chosen equal unity, $\alpha ,\;\beta =1$ (by means of proper constant shifts of the superfields $\Phi ,\;\Lambda$ and their conjugates).
From the above action the following superfield equations of motion can be derived: \_+[D]{}\_- &=& -e\^[\^]{} + e\^[\^ -\^]{}\
[D]{}\_+[D]{}\_- &=& - e\^[\^ - \^]{}\
[D]{}\_+ [D]{}\_- &=& 0 \[eqmo\] (and their conjugates). The $N=2$ super sine-Gordon theory is recovered as the special solution $\Lambda = \Lambda^\dagger = 0$. On the other hand, redefining $\Lambda$ as $\Lambda \rightarrow a\Lambda$ and letting $a \rightarrow \infty $, one recovers the $N=2$ super Liouville theory [@IK]. In the bosonic limit the theory constructed becomes a conformal affine generalization of both the sine- and sinh-Gordon theories, as it is evident from the equations of motion written in terms of the component fields (see below).
Let us introduce the component fields as follows: && |\
\_+ && [øD]{}\_+|\
F && [øD]{}\_+[øD\_-]{}|\
&& |\
\_+ && [øD]{}\_+|\
L && [øD]{}\_+[øD\_-]{}|\
&& |\
\_+ && [øD]{}\_+|\
S && [øD\_+]{}[øD\_-]{}|. All these fields are complex, $\varphi ,\;\lambda ,\; \sigma$ being bosonic, $\psi_+ ,\;\mu_+ , \;\rho_+ $ fermionic and $F,\; L,\; S $ auxiliary.
After eliminating the auxiliary fields by their equations of motion we are led to the following expression for the action $\cal S$ in terms of component fields: &=& d\^2x{\_[++]{}\_[–]{}\^+ \_[++]{} \_[–]{}\^+ \_[++]{}\_[–]{}\^\
&& - \_[–]{}\_+[ø]{}\_+ - \_[++]{}\_-[ø]{}\_- - \_[–]{}\_+[ø]{}\_+ - \_[++]{}\_-[ø]{}\_-\
&& - \_[–]{}\_+[ø]{}\_+ - \_[++]{}\_-[ø]{}\_- - \_-\_+e\^+ [[ø]{}\_-]{} [ø]{}\_+e\^[\^]{}\
&& - (\_–\_-)(\_+-\_+)e\^[-]{}+ ([ø]{}\_–[ø]{}\_-)([ø]{}\_+-[ø]{}\_+) e\^[\^-\^]{}\
&&- e\^[+ \^]{} - e\^[- ]{} e\^[\^ - \^]{} + e\^[-]{} e\^[\^]{} + e\^[\^ - \^]{} e\^ }. The equations of motion for the bosonic fields are: & \_[++]{}\_[–]{}- [14]{}[ø]{}\_-[ø]{}\_+e\^[\^]{} +[14]{} ([ø]{}\_–[ø]{}\_-)([ø]{}\_+-[ø]{}\_+) e\^[\^-\^]{} &\
&+[14]{} e\^[\^+]{} - [14]{}e\^[\^+]{} e\^[-(\^+) ]{} - [14]{} e\^[+\^-]{} + [14]{}e\^[\^-\^+]{} =0&\
&\_[–]{}\_[++]{}-[14]{}([ø]{}\_- -[ø]{}\_- ) ([ø]{}\_+ - [ø]{}\_+ ) e\^[\^-\^]{} + [14]{} e\^[\^++-\^]{} - [14]{} e\^[\^+-\^]{}=0&\
&\_[–]{}\_[++]{}=0.& In the bosonic limit, with all fermions discarded, they are reduced to the system \_[–]{}\_[++]{} (Re) +e\^[2Re]{} -e\^[2Re- 2 Re]{} &=&0\
\_-\_+ (Im) -e\^[Re]{} sin(Im-2Im) &=& 0\
\_[–]{}\_[++]{} (Re) +e\^[2Re]{} cos( 2Im) -e\^[Re]{} cos(Im- 2 Im )&=&0\
\_[–]{}\_[++]{} (Im) +e\^[2Re]{} sin(2Im) +e\^[Re]{}sin( Im-2Im) &=&0\
\_[–]{}\_[++]{} (Re) = \_[–]{}\_[++]{} (Im) &=&0. \[action2\] As was mentioned above, this set is a conformally invariant extension of both the sinh- and sine-Gordon equations (for the fields $Re \phi $ and $Im \phi $, respectively) which are restored in the limit $\lambda = 0$.
We end this section by giving the equations of motion for the fermionic fields \_[–]{} \_+ + [ø]{}\_- e\^[\^]{} - ([ø]{}\_–[ø]{}\_- ) e\^[\^-\^]{} &=&0\
\_[++]{}\_- -[ø]{}\_+ e\^[\^]{} + ([ø]{}\_+-[ø]{}\_+ ) e\^[\^-\^]{} &=&0\
\_[–]{} \_+ + ([ø]{}\_–[ø]{}\_- ) e\^[\^-\^]{} &=& 0\
\_[++]{} \_- + ([ø]{}\_+-[ø]{}\_+ ) e\^[\^-\^]{} &=& 0\
\_[–]{} \_+ = \_[++]{}\_- &=& 0.
Lax pair formulation
====================
The model defined in the previous section is the conformal affine extension of the $N=2$ super sine-Gordon theory considered in ref. . The latter theory is integrable: its equations can be cast in a Lax form with the Lax connections valued in the loop superalgebra $sl(2|2)^{(1)}$. In this section we will make explicit the integrability properties of our model by writing down its Lax pair formulation, which turns out to be based on an affine extension of $sl(2|2)^{(1)}$.
Before going on, let us first recall some basic facts about $sl(2|2)^{(1)}$ . This superalgebra contains four bosonic generators $h_i$ in the Cartan subalgebra and a set of four simple roots generators ${e^{\pm}}_i$ (these all are fermionic). The (anti)commutation relations between the Cartan and simple roots generators are given by &=& a\_[ij]{} [e\^]{}\_j\
{ [e\^+]{}\_i,[e\^-]{}\_j}&=& \_[ij]{}h\_j, with $ i,j =1,...,4$ and $ a_{ij}$ being the Cartan matrix a\_[ij]{} &=& (
[cccc]{} 0 & -1 & 0 & 1\
-1 & 0 & 1 & 0\
0&1&0&-1\
1&0&-1&0
). The above commutation relations are the only ones which are actually needed to completely define the massive integrable theory corresponding to $sl(2|2)^{(1)}$.
The key property of the above Cartan matrix is that it is degenerate, having rank $r=2$: the Cartan elements $h_1+h_3$ and $h_2+h_4$ are $c$-numbers which commute with all other generators.
This fact is an indication that the integrable theory associated with this superalgebra, i.e. the $N=2$ super sine-Gordon theory, is not conformally invariant: it is already known from the bosonic and $N=1$ supersymmetric [@TZ] cases that conformally invariant integrable models can be produced only starting from (super)algebras which have enough Cartan generators to remove the degeneracy among different roots. The best example is supplied by the sinh-Gordon theory related to the $sl(2)$-loop algebra. A conformally invariant generalization of such a model can be constructed by enlarging the $sl(2)$-loop algebra with additional new generators in its Cartan sector (a central extension and a derivative operator counting the powers of the loop spectral parameter): as a result one goes from the $sl(2)$-loop algebra to its affine extension.
A naive transfer of this procedure to the case of $N=2$ super CAL theory does not work basically for two reasons: first, because the introduction of the derivative operator cannot completely remove the degeneracy among roots (an additional operator is needed for this purpose) and, second, because such a derivative is a real operator while we are dealing with complex superfields in the $N=2$ case.
An essentially new feature of the superalgebra $sl(2|2)^{(1)}$ compared to the (super)algebras relevant to the previously known conformal affine theories consists in the existence of an involution between its generators. Just this feature allows for the second supersymmetry. One should therefore expect that the extension of $sl(2|2)^{(1)}$ appropriate to the present case is realized via the introduction of a pair of operators in involution. Indeed, our model involves chiral superfields and their conjugates.
Let us discuss in detail how such an extension can be constructed.
At first we notice that the involution just mentioned is realized, in the particular case of Cartan and simple roots generators, as &&
[cc]{} h\_1 h\_2 & h\_3h\_4\
[e\^+]{}\_1\_2 & [e\^+]{}\_3 \_4.
Since no confusion will arise, let us call for simplicity from now on the above involution as “conjugation" operation.
To enlarge the superalgebra we introduce a couple of conjugate bosonic operators $d,{\o d}$. These are defined as follows: the commutator between them and a generator $e_\alpha$ corresponding to the root $\alpha$, with the decomposition $\alpha =\sum_in_i\alpha_i$ for $i=1,...,4$, $n_i$ integer and $\alpha_i$ denoting the simple roots, is given by: &=& -n\_4e\_\
&=& n\_3e\_, \[rootop\] while their commutators with themselves and with the Cartan generators $h_i$ are assumed to vanish.
Due to the properties of the root decomposition the above relations are consistent with the Jacobi identities.
The operators ${\o d},\; { d}$ are grading operators with respect to the roots ${e^+}_3$, ${e^+}_4$, respectively. The choice of grading with respect to just these roots is to some extent arbitrary, the important point is that the root operators should be conjugate. For standard integrable theories associated with non-singular Cartan matrices such grading operators are not too useful notion since they can be re-expressed as combinations of the original Cartan generators. The importance of $d,{\o d}$ in the case at hand is that they allow to remove the degeneracy of the Cartan subalgebra.
We will denote the algebra enlarged by $d, {\o d}$ as ${\hat {sl(2|2)^{(1)}}}$: $${\hat{sl(2|2)^{(1)}}}= sl(2|2)^{(1)}\oplus
{\bf C} d\oplus {\bf C} {\o d}$$ Clearly, $d,\; {\o d}$ belong to the Cartan sector of ${\hat{sl(2|2)^{(1)}}}$.
It should be mentioned that, in the bosonic and $N=1$ supersymmetric cases \[1,4\], in order to obtain a conformally invariant theory one is forced to introduce not only the derivative operator, but also the central extension. In the present case we do not need to further enlarge the algebra with central extensions, because we have already at our disposal a couple of (conjugate) $c$-numbers given by $h_1+h_3$, $h_2+h_4$ which play such a role.
To express our formulas in a concise notation, it is convenient to split the Cartan subalgebra in the two conjugate parts &=& {h\_1, h\_3, d}\
[ø[H]{}]{}&=& { h\_2, h\_4, [ød]{} } and to denote the sum over the sets of two conjugate simple roots by && [E]{}\^ = [e\^]{}\_1 +[e\^]{}\_3\^ = [e\^]{}\_2+[e\^]{}\_4. Further, let us denote with $\Psi , \;{\o \Psi}$ two conjugate superfields taking values in the Cartan subsectors ${\cal H},\; {\o {\cal H}}$, respectively: &=& A\^h\_1 + B\^h\_3 + C\^d\
[ø]{} &=& A h\_2 + Bh\_4 + C [ød]{} . We possess now all the necessary ingredients to express the integrability properties of our system by introducing two conjugate linear systems of the Lax pair type: $$(D_\pm+{\cal L}_\pm){\cal T}=0$$ and $$({\overline D}_\pm+{\overline{\cal L}}_\pm){\overline {\cal T}}=0\;,$$ where ${\cal L}_\pm, {\o{\cal L}}\pm$ belong to the superalgebra ${\hat{sl(2|2)}^{(1)}}$ and ${\cal T}, {\o{\cal T}}$ to its associate affine KM supergroup.
The zero-curvature condition is the compatibility condition of the above linear systems and it is provided by the relations D\_+[L]{}\_-+D\_-[L]{}\_++{[L]{}\_+,[L]{}\_-}&=&0\
[D]{}\_+\_- + [D]{}\_-\_++ {[\_+]{},[\_-]{}} &=&0. To ensure integrability, the equations of motion of the lagrangian formulation (3) should be recovered from the above zero-curvature relations. This is indeed the case.
Let us define L\_+ &=& D\_++ e\^[ad [ø]{} ]{}[E]{}\^+\
L\_- &=& -[E]{}\^- for one copy of the Lax pair, and \_+ &=& - [ø[E]{}]{}\^+\
[øL]{}\_- &=& [øD]{}\_- [ø]{} + e\^[ad]{}[ø[E]{}]{}\^- for the conjugate copy. Explicitly we have, in the former case, L\_+ &=& (D\_+[A]{}\^) h\_1 +(D\_+[B]{}\^) h\_3 + (D\_+ [C]{}\^) d +\
&& e\^[(B-A)]{}[e\^+]{}\_1 + e\^[(C-(B-A))]{} [e\^+]{}\_3\
L\_- &=& -([e\^-]{}\_1+[e\^[-]{}]{}\_3) (and similar expressions in the latter case). It is a straightforward exercise to check that the zero-curvature condition is satisfied, once provided that the superfields $A,B,C$ and their conjugates satisfy the following relations D\_+D\_-A\^ &=& -e\^[(B-A)]{}\
D\_+D\_- B\^&=& -e\^[(C-(B-A))]{}\
D\_+D\_- C\^&=& 0 \[Laxmotion\] together with their conjugate counterparts.
It is worth mentioning that the zero appearing in the r.h.s. of the equation for the superfield $C^\dagger$ is due to the fact that the $d$ generator never appears in the commutators of generators of $sl(2|2)^{1}$. Moreover, in the r.h.s. of all equations the superfields $A,B$ come out only in the combination $B-A$ since $h_1+h_3$ is a $c$-number.
The Lax-pair version of the theory involves $3$ complex superfields just like the lagrangian formulation of the previous section. The zero-curvature equations (\[Laxmotion\]) reproduce just the equations of motion of the lagrangian theory under the following identification: B-A,C, -B (the resulting equations are conjugate of eqs. (3)).
Superconformal properties and higher spin primary fields
========================================================
A characteristic feature of the model we have constructed is its invariance under $N=2$ superconformal transformations which are infinitesimally given by &&
[cc]{} = -D\_+\^+ & =-2D\_+\^+\
= D\_+\^+ & D\_+ = -(D\_+\^+)D\_+\
[øD]{}\_+\^+ = 0
\[transf\] and by their conjugate counterparts. Here $\alpha$ is an arbitrary real parameter. It is straightforward to check that the action (\[action1\]) and the equations of motion (\[eqmo\]) are invariant with respect to these transformations. Similar relations hold for the superconformal transformations acting in the opposite light-cone sector of $N=2$ superspace (on the coordinates $x^{--}, \theta^-, {\o\theta}^-$).
The presence of extra superfields $\Lambda $, $ \Sigma $ allows to reestablish the superconformal invariance which is spoiled in the corresponding massive model ($N=2$ super sine-Gordon theory). Due to these superfields the constructed model turns out to have a richer algebraic structure than both its special limits, the $N=2$ super Liouville and super sine-Gordon theories. In particular, the action (\[action1\]) is invariant under the shifts $$\Sigma \rightarrow \Sigma + \delta \Sigma\;,\;\;\;
\Sigma^{\dagger} \rightarrow \Sigma^{\dagger} + \delta \Sigma^{\dagger}\;,$$ with $\delta \Sigma$ ($\delta \Sigma^{\dagger}$) being a sum of two arbitrary chiral (anti-chiral) superfunctions living in the left and right light-cone sectors of $N=2$ superspace. By the way, using this freedom, one may always redefine the superconformal transformation of $\Sigma$ so as to make it homogeneous (this amounts to choosing $\alpha = 0$ in eqs. (24)).
Recall that the bosonic analog of our model, the CAL theory (as well as its Toda generalizations), exhibits a $W_{1+\infty}$ type symmetry algebra generated by an infinite set of higher spin primary fields together with the spin 2 conformal stress-tensor and a quasi-primary spin 1 field [@AFGZ2]. We will repeat in our case the procedure developed in [@AFGZ2] and will find an infinite set of supercurrents generating a $N=2$ superextension of the algebra present in the CAL theory. Like in most (super)conformally invariant theories, this superalgebra splits into two commuting light-cone copies. So, without loosing generality, we will limit our study to one of them corresponding to the light-cone direction specified by $x^{++} , \theta^+ ,{\o\theta}^+$.
Let us first recall the definition of primary $N=2$ superfields which are a natural generalization of the spin $s$ primary fields used in the construction of ref. [@AFGZ2]. These are denoted by $T^{(s_1\;s_2)}$, with $s_1$, $s_2$ being arbitrary integers, and transform under the $N=2$ superconformal group (\[transf\]) according to the rule T\^[(s\_1s\_2)]{} & =& (s\_1[øD]{}\_+\^+ - s\_2D\_+\^+ )T\^[(s\_1s\_2)]{}. \[primary\] The external conformal spin of $T$ is defined as $s \equiv {1\over 2}
(s_1 + s_2)$ while $h \equiv s_1 - s_2$ has the meaning of the external $U(1)$ charge of $T$ (under the convention that the $U(1)$ charge of $\theta^{\pm}$ equals 1).
The basic ingredient of the construction of ref. [@AFGZ2] is the anomaly free spin 2 conserved current $T_{++++}$. Its analog in the $N=2$ case is the anomaly free conserved[^1] spin 1 supercurrent $\tilde{J}_{++}^{(1\;1)}$. In order to find it we proceed as follows.
As a first step we define the most general spin 1 real conserved supercurrent which turns out to be J\_[++]{} &=& [øD]{}\_+ D\_+\^+ 2i\_[++]{} (-\^) +\
&& [øD]{}\_+ D\_+\^+[øD]{}\_ +D\_+\^+ 4i \_[++]{} ( - \^)+\
&& a [øD]{}\_+ D\_+\^+ i b \_[++]{} (-\^), \[current\] with $a,\; b$ being arbitrary real constants.
The parameters $a,\; b$ in (\[current\]) are uniquely fixed via the parameter $\alpha$ defined in (\[transf\]) if we further demand $J_{++}$ to transform as a primary $(1,1)$ $N=2$ superfield, i.e. require the absence of an inhomogeneous piece in its superconformal transformation law. In this way we recover the expression for the anomaly-free supercurrent ${\tilde J}_{++}^{(1\;1)}$ which is given by eq. (\[current\]) with the following fixed values of the involved parameters && a= -[14]{} b = 2-1. Note that the canonical $N=2$ conformal supercurrent generating, via appropriate super Poisson brackets, superconformal transformations of the superfields $\Phi,\;\Lambda$ and $\Sigma$ in eqs. (26) corresponds to a different choice of the parameters, namely, && a=0 b = -2 . It is worth mentioning that it is anomaly-free at $\alpha = {1\over 4}$.
One more entity used in [@AFGZ2] to construct an infinite sequence of the primary higher spin currents generating a $W$ type symmetry algebra is a conserved spin 1 current. It is quasi-primary, i.e. inhomogeneously transforms under the conformal group. $N=2$ analogs of this current are two conjugate spin ${\t {1\over 2}}$ conserved supercurrents $$J_+ = D_+\Lambda^\dagger \;,\;\;\;\;{\o J}_+ ={\o D}_+\Lambda.$$ These are quasi-primary with respect to superconformal transformations, e.g., $$\delta J_+ = - D_+\delta\theta^+ J_+ + 4i \partial_{++}\delta\theta^+\;.$$
Now we are ready to generalize to our case the procedure employed in [@AFGZ2]. We introduce the operators $I_1$, $I_2$ which count, respectively, the weights $s_1$ and $s_2$ of the primary superfields $$I_{1,2} T^{(s_1\;s_2)} = s_{1,2} T^{(s_1\;s_2)}$$ and define two supercovariant derivatives \_+ &=& D\_+ - D\_+\^ I\_1\
[ø[D]{}]{}\_+ &=& [øD]{}\_+ - [øD]{}\_+ I\_2. They have the following properties: applied to a superfield of weights $(s_1, s_2)$, $T^{(s_1\;s_2)}$, they send it, respectively, into superfields of weights $(s_1, s_2+1)$ and $(s_1+1,s_2)$. Indeed, it is a simple exercise to check that the superfields $${\cal D}_+ T^{(s_1\;s_2)}\;,\;\;\; {\o{\cal D}}_+ T^{(s_1\;s_2)}$$ transform according to the generic transformation law (\[primary\]) with the aforementioned weights. Note that the supercovariant derivatives ${\cal D}_+$, ${\o{\cal D} }_+$ satisfy the following anticommutation relation: { [D]{}\_+, [ø[D]{}]{}\_+ } &=& 2i\_[++]{} +( [øD]{}\_+ D\_+\^ -i\_[++]{} \^)I\_1 + ( D\_+\^ [øD]{}\_+-i\_[++]{}\^)I\_2\
&& -D\_+ \^ [øD]{}\_+ (I\_1 +1) -[øD]{}\_+ D\_+(I\_2 +1)\
[D]{}\_+ [D]{}\_+ &=& [ø[D]{}]{}\_+[ø[D]{}]{}\_+ = 0.
Now, starting with the anomaly-free spin 1 supercurrent $\tilde{J}_{++}^{(1\;1)}$ and acting on it successively by ${\cal D}$ and ${\o {\cal D}}$, we may construct an infinite tower of conserved supercurrents with higher $s_1$ and $s_2$ which is a genuine $N=2$ generalization of the set of currents of the bosonic CAL model. All these are primary with respect to the canonical $N=2$ conformal supercurrent, i.e. have no anomalous terms in their superconformal transformations. The generic form of the basic supercurrents is as follows (taking account of the relations (30)) J\_[+2n]{}\^[(nn)]{} &=& [ø[D]{}]{}\_+ [D]{}\_+ ....\_+ [D]{}\_+ \_[++]{}\^[(11)]{}\
J\_[+(2n+1)]{}\^[(nn+1)]{} &=& [D]{}\_+ J\_[+2n]{}\^[(nn)]{}. The remaining primary supercurrents are obtained from these two basic sequences via complex conjugation and permutation $s_1 \leftrightarrow s_2$.
Together with the spin ${\textstyle{ 1\over 2}}$ conserved supercurrents and the canonical conformal supercurrent, the primary supercurrents defined above form a set which is nonlinearly closed under the super Poisson brackets and so is recognized as a kind of infinite-dimensional nonlinear $N=2$ $W_{1/2 + \infty}$ superalgebra. Based on analogy with the bosonic case [@AFGZ2], in the limit of “large” $\Lambda$ these supercurrents are expected to close on a $N=2$ supersymmetric extension of the linear area-preserving $w_{1+\infty}$ algebra. The detailed discussion will be reported elsewhere.
Concluding remarks
==================
To summarize, in this paper we have constructed the $N=2$ conformal affine super Liouville theory, shown its integrability by defining the appropriate superfield Lax pair (zero curvature) representation and found an infinite set of primary $N=2$ supercurrents which are counterparts of the analogous currents of the bosonic CAT theories and form a nonlinear $N=2\; W_{1/2 + \infty}$ type symmetry algebra of the model. It is straightforward to extend this construction to arbitrary $N=2$ super Toda theory [@EH] like this has been done for the bosonic case in ref. [@AFGZ2]. An interesting question is how to reproduce $N=2$ CAT theories from the appropriate supergroup $N=2$ supersymmetric WZNW sigma models via hamiltonian reduction (for the bosonic and $N=1$ supersymmetric CAT theories this has been done in [@AFGZ2] and [@TZ]). A trouble here is that up to now a manifestly $N=2$ supersymmetric superfield formulation of such sigma models for general target (super)groups is lacking. One more interesting problem we are planning to address in the nearest future is the construction of $N=4$ supersymmetric conformal affine Liouville theory which should be an extension of $N=4$ super Liouville theory [@IK2].
The authors are grateful to F. Delduc, L.A. Ferreira, J.F. Gomes, S. Krivonos for useful discussions. A special thank to P. Sorba who explained us ref. [@FSS] and who carefully read the manuscript. F.T. is grateful to Director of Bogoliubov Theoretical Laboratory of JINR, Prof. D.V. Shirkov, for his kind hospitality in Dubna, where part of this work has been accomplished.
.3in
[99]{} O. Babelon and L. Bonora, [*Phys. Lett.*]{} [**B244**]{} (1990) 220 L. Bonora, M. Martellini and Y.-Z. Zhang, [*Int. J. Mod. Phys.*]{} [**A6**]{} (1991) 1617; Y.-Z. Zhang, [*Mod. Phys. Lett.*]{} [**A6**]{} (1991) 2023 H. Aratyn, L.A. Ferreira, J.F. Gomes and A.H. Zimerman, [*Phys. Lett.*]{} [**B254**]{} (1991) 372 F. Toppan and Y.-Z. Zhang, [*Phys. Lett.*]{}[**B292**]{} (1992) 67 L. Frappat, A. Sciarrino and P. Sorba, [*Commun. Math. Phys.*]{} [**121**]{} (1989) 457 M.A. Olshanesky, [*Commun. Math. Phys.*]{} [**88**]{} (1983) 63; O. Babelon and F. Langouche, [*Nucl. Phys.*]{} [**B290**]{} (1987) 603 J. Evans and T. Hollowood, [*Nucl. Phys.*]{} [**B352**]{} (1991) 723 F. Delduc, E. Ragoucy and P. Sorba, [*Comm. Math. Phys.*]{} [**146**]{} (1992) 403 J. Evans and T. Hollowood, [*Phys. Lett.*]{} [**B293**]{} (1992) 100 E.A. Ivanov and S.O. Krivonos, [*Lett. Math. Phys.*]{} [**7**]{} (1983) 523; [**8**]{} (1984) 345 E T. Inami and H. Kanno, [*Comm. Math. Phys.*]{} [**136**]{} (1991) 519; [*Nucl. Phys.*]{} [**B359**]{} (1991) 201; K. Kobayashi and T. Uematsu, [*Phys. Lett.*]{} [**B275**]{} (1992) 361 A.N. Leznov and M.V. Saveliev, [*Lett. Math. Phys.*]{} [**3**]{}(1979) 49 H. Aratyn, L.A. Ferreira, J.F. Gomes and A.H. Zimerman, [*Phys. Lett.*]{} [**B281**]{} (1992) 245; [*preprint*]{} [**UICHEP-TH/92-1**]{} E.A. Ivanov and S.O. Krivonos, [*J. Phys. A: Math. Gen.*]{} [**17**]{} (1984) L671; E.A. Ivanov, S.O. Krivonos and V.M. Leviant, [*Nucl. Phys.*]{} [**B304**]{} (1988) 601; O. Gorovoy and E. Ivanov, [*Nucl. Phys.*]{} [**B381**]{} (1992) 394
[^1]: A supercurrent $J$ is called conserved if, by using the equations of motion for the involved superfields, it satisfies the relations $D_-J = {\o D}_-J =0$ which imply in particular the condition of light-cone chirality with respect to the ordinary space-time variables, $\partial_{--} J=0$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
-
-
-
title: 'Can network coding bridge the digital divide in the Pacific?'
---
Introduction
============
The present study has its genesis in a longitudinal project aimed at determining long-term trends in jitter, packet loss and other observables on long-distance Internet paths [@iibex]. The project also documents the challenges faced by users in remote locations with numerous nodes in various Pacific Island nations. This naturally poses the question how one might improve their connectivity.
Bandwidth upgrades can address the bandwidth bottleneck to an extent, but do not address latency issues unless an island simultaneously upgrades to a lower latency connection, i.e., from a geostationary (GEO) to a medium earth orbit (MEO) satellite [@O3b] or to a fibre-optic submarine cable.
Latency is particularly problematic when accompanied by packet loss. Satellite links to Pacific islands usually have much lower bandwidth than the international fibre networks supplying them, making them prime candidates for packet queue formation and hence for packet loss associated with tail drops.
Moreover, packet loss causes TCP sessions to lowering their data rate. However, long latency prevents TCP from doing so in a timely fashion, which can lead to queue oscillation at the satellite gateway [@queueoscillation]. This can result in frequent idleness of the link, i.e., unused satellite bandwidth being observed even over short time intervals in the order of seconds.
Many small Pacific island states are developing nations with small populations (with similar numbers living overseas, mostly in New Zealand, Australia and the U.S.). Only a few, such as Hawaii, Fiji, and Guam, are the lucky transit points of international fibre routes. Others such as Tonga or Samoa are located close enough to such a transit point and populous enough to be able to afford a spur connection from a transit point. Many others simply do not have the population or budget for a fibre connection.
MEO satellites represent a lower cost/bandwidth alternative to fibre, with latencies roughly comparable to a fibre link from New Zealand to California. However, small islands often cannot muster the resources to install and maintain the required tracking antennas. At the time of writing, a small number of O3b MEO connections operate in Rarotonga, Samoa and East Timor [@O3b].
This leaves in particular the smallest, poorest, and most remote island communities reliant on expensive low-bandwidth and very long latency GEO satellite links. In many cases, thousands of people share an international link bandwidth equivalent to an average residential ADSL connection.
Various fast TCP variants such as H-TCP [@htcp] and Hybla [@hybla] attempt to establish a stable flow across long-latency connections, but are generally aimed at large-bandwidth scenarios to transfer large amounts of data efficiently. Our results indicate that they do not perform significantly better than conventional TCP in low-bandwidth scenarios in the Pacific and do not routinely outperform the most widely deployed Linux TCP variant, Cubic [@cubic].
Network coded TCP [@nc-meets-tcp] (TCP/NC) is a forward error correction technique which may be used to hide packet losses between two hosts on a TCP network. This can mask tail drops at the satellite gateways, but an end-to-end TCP/NC solution requires both the sending and receiving host to “speak” TCP/NC. While this could conceivably be achieved for on-island hosts, it would be unrealistic to expect the rest of the Internet to follow suit.
An alternative solution is to deploy a network-coded “tunnel” between an off-island host with good connectivity to the Internet, and an on-island host acting as a gateway to an on-island network. This tunnel could mask satellite gateway losses and supply island end users with a low-loss connection to the Internet, avoiding the start-stop behaviour of their conventional TCP connections. But how well could this work in practice? This paper describes the initial results of a TCP/NC project in three Pacific Island locations with slightly different profiles.
We first give a brief birds-eye overview of TCP/NC and the basic network topology of our project, followed by an introduction to our deployment sites and initial observations from each.
Network-coded TCP
=================
Network coding has attracted considerable research interest in information theory and communication engineering over the last decade, with much of the focus on solutions for mobile and wireless, especially in broadcast and multicast settings and in cooperative networking. In recent years, a number of authors have also proposed its use in conjunction with conventional point-to-point TCP/IP networking [@nc-meets-tcp; @hansenetal]. Network coding has also been investigated on the link layer of satellite links [@bischl-et-al].
The fundamental idea of TCP/NC is to treat the original packets in a TCP connection as variables in a system of linear equations. As each packet is simply a binary string, it also represents a binary number, which can be multiplied and added. From a set of packets $p_1, p_2, \ldots, p_n$, [*random linear network coding*]{} (RLNC) generates a set of $n+\omega$ linear equations with random coefficients $c_{i,j}$ such that the $i$-th equation in the set is: $$\sum_{j=1}^{n}c_{i,j} p_j = r_i.$$ where $1\leq i \leq n+\omega$. We refer to $n$ as the [*generation size*]{} and to $\omega$ as the [*overhead*]{} and write $n+\omega$ to denote the size of the system. Instead of transmitting the original packets, TCP/NC transmits “equation” packets, i.e., the coded packet for the $i$-th equation encodes the $c_{i,j}$ for $1\leq j \leq n$ and $r_i$.
The receiving host can recover all of $p_1, p_2, \ldots, p_n$ once it has received $n$ linearly independent coded “equation” packets. Note that as the sender generates $n+\omega$ coded packets, the communication can in principle tolerate the loss of possibly up to $\omega$ arbitrary packets (readers without a background in linear algebra will probably still remember the “at least $n$ equations for $n$ variables” rule from school). Moreover, the sender can start generating redundancy as soon as two of the original packets are available, it does not have to wait for all $n$ packets like a block coder would have to.
The field from which one picks the $c_{i,j}$, the sparseness of the matrix $c_{i,j}$, the choice of $n$ and $\omega$ all influence robustness, coding rate, and decoding delay and complexity. A detailed discussion of these topics is beyond the scope of this paper. We merely note here that this is an active research area, some aspects of which we will return to in the rest of this paper.
IP networking considerations for TCP/NC tunnels
===============================================
The tunnel in our experiments transports coded packets in the payload of UDP packets. UDP packets heading to the island originate from an off-island gateway $G_w$ and terminate at a gateway $G_i$ on the island. In this case, $G_w$ acts as the NC encoder and $G_i$ as the decoder. In the opposite direction, $G_i$ generates coded packets and $G_w$ is the decoder. Note that $G_w$ can be anywhere on the Internet: Rather than being topologically close to the satellite gateway $G_s$ facing the island, it should be close to the source/destination of the data required on-island to minimise the latency of the connection links covered by conventional TCP only.
The UDP headers of the tunnel packets contain only the IP addresses of $G_w$ and $G_i$. The IP addresses of the actual endpoints of each connection travelling through the tunnel travel inside the UDP payload as part of the NC header, i.e., they are not visible between $G_w$ and $G_i$. We denote these endpoints as $E_i$ (on-island) and $E_w$ (outside world).
For an IP packet from $E_i$ to reach $G_i$ for encoding, it suffices to insert $G_i$ as a “gateway router” for the on-island subnet in which $E_i$ resides. The tunnelled and decoded packet then egresses from $G_w$ enroute to $E_w$. N.B.: The network within which $G_w$ resides must allow egress of packets from $E_i$’s subnet.
In the opposite direction, an IP packet from $E_w$ to $E_i$ must find its way to $G_w$ for encoding. This requires one of the following two configurations to be implemented:
1. $E_i$ must be part of the autonomous system (AS) of $G_w$, and the AS must be configured internally to route traffic to $E_i$ via $G_w$ once it enters the AS.
2. Alternatively, $G_w$ may be one of the (at least) two BGP gateway routers giving access to an on-island AS within which $E_i$ resides. In this case, all gateways for this AS also need to be TCP/NC encoders/decoders to ensure that only encoded traffic enters the AS.
Note that $E_i$ cannot simply use an address from the on-island network for which the satellite gateway $G_s$ acts as the IP gateway. If it did, BGP would route packets from $E_w$ to $E_i$ straight to $G_s$, bypassing the encoder $G_w$. In our experiments, we opted for the first configuration as it only needs a single off-island gateway $G_w$. The drawback of this solution is that the AS of $E_i$ starts off the island, potentially leaving an island ISP reliant on address blocks from the AS of a single offshore provider. Note also that the $G_i$’s IP address must be reachable via the satellite gateway $G_s$, so cannot be part of the same AS as $E_i$’s subnet.
Packet size also needs to be considered: As TCP/NC adds its own header to each encapsulated packet, the packet may exceed the MTU (maximum transmission unit) between $G_w$ and $G_i$, leading to IP fragmentation. While this does not affect principal functionality, it adds undesirable overhead and can affect timing and performance. For unfragmented UDP transit, the tunnel must limit the MTU for unencoded IP packets at $G_w$ and $G_i$ to the MTU of the path between $G_w$ and $G_i$ minus the size of the NC header.
In our experiment, we use a $G_w$ with two Ethernet interfaces, [*eth0*]{} and [*eth1*]{}, configured as follows:
- [**eth0**]{} is the tunnel interface. Its single exclusive route points at the “normal” island IP address of $G_i$’s tunnel interface only so it can reach it via $G_s$. [*eth0*]{} connects to its own router port in its own exclusive /30 subnet. The tunnel endpoint itself is a virtual network device supplied by Steinwurf ApS as a loadable Linux kernel module. This endpoint intercepts packets from $G_i$’s tunnel interface arriving at [*eth0*]{} and decodes them. Similarly, any unencoded packets sent to the virtual adapter are encoded and sent to $G_i$ via [*eth0*]{}.
- [**eth1**]{} is the default “world” interface. It also connects to its own router port in its own exclusive /30 subnet, separate from [*eth0*]{}’s subnet. Any outgoing traffic not destined for [*eth0*]{}’s subnet or $G_i$ egresses here. Similarly, the outside world routes traffic destined for $G_w$ and for $E_i$’s on-island subnet to [*eth1*]{}. Any IP packets destined for the latter subnet are sent to the virtual network device for encoding.
The $G_w$ in our study is a Dell PowerEdge R320 server with an Intel Xeon E5-2420 v2 Processor running at 2.20 GHz based at the University of Auckland. A further $G_w$ endpoint is planned for California.
At the island ends, the respective $G_i$ have three Ethernet interfaces (say [*eth0*]{}, [*eth1*]{} and [*eth2*]{}):
- As in $G_w$, [*eth0*]{} is the tunnel interface with a single exclusive route pointing at the IP of [*eth0*]{} on $G_w$. The tunnel endpoint is the same kernel module as in $G_w$, configured to peer with [*eth0*]{} on $G_w$ to intercept UDP from there for decoding. Any unencoded packets sent to the virtual adapter are encoded and sent to $G_w$ via [*eth0*]{}.
- [*eth1*]{} now serves as the interface to $E_i$’s on-island subnet. Decoded traffic for $E_i$ egresses here, and any traffic arriving at this interface from the outside is sent to the kernel module for encoding.
- [*eth2*]{} acts as a “maintenance interface” for emergency remote access from off-island machines via $G_s$ if [*eth0*]{} is unavailable. [*eth2*]{} and [*eth0*]{} must be in different networks. Policy routes let $G_i$ respond to incoming standard TCP traffic on both [*eth0*]{} and [*eth2*]{}.
Our $G_i$ at each island end is a Stealth LPC-630F Little PC industrial computer with an Intel i7-3520M Gen 3 processor.
Deployment sites
================
The study deployed $G_i$ in the following locations:
[**Rarotonga, Cook Islands**]{}. Rarotonga has a permanent population of over 13,000, with an estimated 1000 to 3000 visitors at any time [^1] Telecom Cook Islands (TCI), Rarotonga’s sole Internet provider, connects primarily via the O3b MEO satellite network. At time of deployment, TCI subscribed to 160 Mbps down- and 40 Mbps uplink, normally routed via O3b’s Hawaii teleport. Casual 802.11b/g/n Internet retailed for around US\$ 0.06 - 0.08 per MB on Rarotonga; US\$ 75 bought a month of ADSL connection limited to 4 Mbps with an 8 GB data cap.
[**Niue**]{}. This single-island country has a permanent population of around 1,600 with only a small number of visitors but a large overseas diaspora: Over 20,000 Niueans live in New Zealand alone [@nupopstats]. Niue’s only ISP is Internet Niue, with a GEO connection now providing 8 Mbps down- and 2 Mbps uplink after an upgrade. Niue has a free local public access 802.11 WiFi network covering most of the island. US\$ 70 buys a month of ADSL limited to 384 kbps / 2 GB.
[**Funafuti Atoll, Tuvalu**]{}. Funafuti is the most populous atoll in Tuvalu with $\approx$4,500 inhabitants. The sole ISP, Tuvalu Telecommunication Corporation (TTC), uses a GEO link. Its satellite provider also manages some of TTC’s on-island network near the local satellite gateway for TTC. The GEO downlink rate into Funafuti was not available to us, but our measurements indicate a rate of 16 Mbps. US\$100 buys 256 kbps ADSL with 3 GB data/month.
A somewhat unusual component of TTC’s network is a SilverPeak NX-3700 WAN Optimiser [@silverpeak], which carries all production IP traffic. Details on the configuration of the device were not available. However, its functions include forward error correction in tunnels via parity packets that can be added automatically or at pre-set ratios to protect against sporadic packet loss. We also observed that the device had network memory activated. Other features include packet reordering, coalescing of smaller packets into larger ones, IP header / payload compression, and TCP and other protocol acceleration.
Network memory in particular makes it challenging to compare TCP/NC and conventional TCP across a SilverPeak path, as some packets are served from local cache and never transit the link. We therefore positioned the TCP/NC encoder / decoder in Funafuti such that our traffic to it did not pass through the SilverPeak device. However, we measured conventional TCP goodput to a machine supplied through the NX-3700.
Both Niue and Tuvalu have signed up to the Kacific geostationary high-throughput broadband satellite system [@kacific], with a ready-for-service date in 2017. This will increase the bandwidth into both countries to levels broadly comparable to the current Rarotonga connection and will provide direct-to-building services.
Observations to date
====================
Data from our IIBEX database, collected using VoIP-like UDP traffic with small packets transmitted at 50 packets/s, showed packet loss on both up- and downlink to and from Rarotonga of between 0 and 0.25% before deployment. To and from Niue, packet loss was below 0.1%. No recent data was available for Tuvalu, but data from 2013 suggested packet losses of up to 2.5%. At deployment time, we encountered similar values with significant variation over time.
As part of the deployment, we undertook extensive throughput measurements in all locations using both conventional TCP and TCP/NC. We found the following in the respective locations:
[**Rarotonga**]{}: Conventional TCP seemed unable to access more than around 75% of the available satellite bandwidth, even over short time intervals in the order of seconds. Most of the time, the observed utilisation was below 60%. As satellite uplinks are not subject to medium contention, the utilisation below 100% implies that the input queue to the satellite modem in Hawaii [*empties*]{} completely during those time periods.
However, we also observed regular burst packet losses in inbound traffic. These do not occur between the various offshore locations involved, indicating an association with the satellite link. Burst losses are characteristic for tail drops on [*overflowing*]{} queues, in this case most likely the input queue to the satellite modem, where high bandwidth fibre or Ethernet traffic feeds into the lower bandwidth of the satellite link, but can also be caused by the radio link itself.
Link underutilisation with this behaviour is a well-known effect called [*queue oscillation*]{} [@queueoscillation], which often occurs at Internet bottlenecks with multiple parallel flows. All satellite links under consideration here qualify as bottlenecks with bandwidths below those of the networks connected at either end. Queue oscillation can occur when multiple TCP flows try to “fill the pipe” through the bottleneck. Traffic banks up until the queue overflows, resulting in packet loss. The TCP senders cannot detect this until an ACK becomes overdue or a selective repeat request arrives. Neither of these events reduces the queue arrival rate for a full end-to-end round-trip-time (RTT) period, long enough to turn most connections’ packet losses into bursts. This causes all senders to slow down and the arrival rate at the queue now drops substantially below the satellite bandwidth; the queue drains and the link sits idle. Without further packet losses, the sending rates increase again. This cycle can repeat within only a few RTTs, explaining why the effect is seen even at small timescales of just a few seconds.
Initially, we required $n+\omega=60+30$ (=50% redundancy) to mask the packet loss bursts for our TCP sender to sustain a higher packet rate. Between Ubuntu servers offshore and a Linux client on the island’s test network, goodput improved by up to a factor of 4 with the TCP/NC tunnel. Bandwidth utilisation increased to almost 90% for a single download via the tunnel. With a Windows 7 client, goodput improvement was only $\approx 25\%$, which was sustainable for several parallel connections, however. Possible reasons may include fragmentation or packet reordering across the tunnel.
When revisiting the Rarotonga link in late January 2015, we found that (a) the total downlink bandwidth had increased to 200 Mbps, (b) conventional TCP now slightly outperformed the $60+30$ tunnel, (c) the queue still seemed to be oscillating, with similar utilisation percentages, i.e., around 25% additional throughput, and (d) a $30+6$ tunnel (constant 20% overhead) was now sufficient to mask losses in most cases and gave better goodput than TCP, especially at times of high packet loss (see Fig. \[raro20pc\]). During periods of low or no packet loss, the $30+6$ tunnel mostly yielded goodput comparable to conventional TCP. This suggests that an adaptive overhead scheme could replace most of the overhead by goodput during these times.
[**Niue**]{}: The link into Niue sees sustained peak data rates of around 7.5 Mbps with $>$7 Mbps recorded for much of the day. Individual conventional TCP connections achieve around 0.3 Mbps. Packet loss into Niue is low but not zero. Closer inspection reveals that the link transports goodput without redundant retransmissions arriving at the Niue end. At the utilisation and data rates observed, the link can thus handle around 25 parallel connections.
Single TCP connections across a TCP/NC tunnel achieved around 2-2.4 Mbps goodput with very low overhead, i.e., the entire link capacity would be exhausted by 3-4 such TCP/NC connections. Given the high existing link utilisation, this additional performance of even a single connection comes at the expense of conventional TCP goodput. However, if these connections are downloads, the higher goodput rate also shortens the flow. This poses the question as to whether short wideband flows with TCP/NC are better from a user perspective than long thin ones, given that the bulk of bandwidth use consists of flows that download something.
In Niue, we also investigated the potential of H-TCP and Hybla compared to the standard Cubic TCP used by Ubuntu. While there were considerable differences between them and Cubic at certain times, neither of the two presented a convincingly strong alternative on this narrowband path.
[**Funafuti**]{}: No link utilisation data was available, but we were able to measure traffic directly between the sat gate and TTC’s local network with a line tap and nprobe/ntop [@ntop]. Without TCP/NC, utilisation for total IP data traffic into Funafuti seldom exceeded 2-3 Mbps – less than 20% of available bandwidth. Packet losses occurred as soon as a relatively modest load was offered. In combination, these observations once again suggest queue oscillation.
Conventional TCP to the Funafuti TCP/NC encoder/decoder (i.e., not through the NX-3700) did not exceed 1.6 Mbps at any time. During peak hours, rates were $<$0.4 Mbps and some connections timed out. Closer analysis revealed that peak time packet loss bursts during oscillation ran into the hundreds of packets – longer than the maximum overhead presently possible in our TCP/NC kernel module. Downloads into Funafuti via TCP/NC tunnel achieved a steady average goodput of around 4 Mbps for a single connection. Link utilisation with a $60+30$ generation tunnel observed regularly reached short-term peak rates of over 15 Mbps, i.e., longer downloads may well reach average rates $>$4 Mbps. Conventional TCP throughput on the local network side of the NX-3700 reaches up to 13 Mbps. Given the comparatively low data rate observed as going into Funafuti, it is likely that a significant part of the 13 Mbps is supplied by the NX-3700 network memory function.
Conclusions and future work
===========================
This paper reported on preliminary observations from our three deployment sites. These show that TCP/NC can provide significantly higher goodput rates for individual connections than conventional TCP and that it can exploit spare capacity on satellite links left idle due to queue oscillation. On such links, TCP/NC can provide additional goodput. On links already carrying mostly goodput (Niue), it can merely replace long and more or less parallel data transfers by short and mostly successive ones. A possible benefit of this is that individual downloads complete faster, reducing user temptation to abort downloads (and hence waste partially downloaded data).
There are many obvious questions that we have not yet been able to investigate and many challenges to address. These include finding optimal generation sizes and overheads, adapting them to conditions, investigating the potential use of a performance-enhancing proxy (PEP) that breaks up the TCP connection into multiple sections, and the performance of TCP/NC if used in conjunction with network memory and other functions of a WAN optimiser. Also, we still expend a lot of TCP/NC overhead on a problem primarily caused by the conventional TCP traffic our TCP/NC streams mingle with. Would an all-TCP/NC link work with less overhead? That said, the pace at which these investigations will be able to proceed is limited by the very problem we are trying to solve: the long-latency, low-bandwidth satellite links into Pacific Islands.
Acknowledgements {#acknowledgements .unnumbered}
================
This research was supported by the Information Society Innovation Fund Asia through the Pacific Island Chapter of the Internet Society (PICISOC) and by Internet New Zealand. We would also like to thank the many Internet users and staff of Telecom Cook Islands, Internet Niue, and the Tuvalu Telecommunication Corporation for their patience during this study and for sharing their precious bandwidth with us.
[99]{} IIBEX home page, <http://iibex.auckland.ac.nz/>. Visited 15 February 2015. O3b Networks, home page, <http://www.o3bnetworks.com/>. Visited 15 February 2015. J. Kim, I. Yeom, [*Reducing Queue Oscillation at a Congested Link*]{}, IEEE Transactions on Parallel and Distríbuted Systems, 19(3), 394–407, 2008 D. Leith, [*H-TCP: TCP Congestion Control for High Bandwidth-Delay Product Paths*]{}, Internet Draft, IETF, April 7, 2008, <https://tools.ietf.org/html/draft-leith-tcp-htcp-06> C. Caini, R. Firrincieli, [*TCP Hybla: a TCP enhancement for heterogeneous networks*]{}, Int. J. of Satellite Communications and Networks, 22, 2004 S. Ha, I. Rhee, L. Xu, [*CUBIC: A New TCP-Friendly High-Speed TCP Variant*]{}, ACM SIGOPS Operating System Review, 42(5), 64–74, 2008 J. K. Sundararajan, D. Shah, M. Médard, S. Jakubczak, M. Mitzenmacher, J. Barros, [*Network Coding Meets TCP: Theory and Implementation*]{}, Proc. IEEE, 99(3), 490–512, 2011 J. Hansen, J. Krigslund, D.E. Lucani, F.H.P. Fitzek, [*Sub-Transport Layer Coding: A Simple Network Coding Shim for IP Traffic*]{}, IEEE VTS Vehicular Technology Conference (VTC), 2014 H. Bischl, H. Brandt, F. Rossetto, [*An experimental demonstration of Network Coding for satellite networks*]{}, CEAS Space Journal 2.1-4 (2011): 75-83 Ministry of Finance and Economic Management, Cook Islands: Tourism and Migration Statistics <http://www.mfem.gov.ck/population-and-social-statistics/tourism-migration-stats>. Visited 15 February 2015. Statistics Niue: Niue at a Glance, <http://www.spc.int/prism/niue/>. Visited 1 January 2015. Kacific Broadband Satellite Home Page, <http://kacific.com>. Visited 15 February 2015. Silver Peak WAN Optimization Appliances, Appliance Manager Operator’s Guide, VXOA 6.2, December 2014, <http://www.silver-peak.com/sites/default/files/userdocs/appliancemgr_operators_guide_r6-2-5_revn_december2014_0.pdf>. Visited 15 February 2015. ntop home page, <http://www.ntop.org/>. Visited 15 February 2015.
[^1]: Authors’ estimate based on monthly government visitor arrival statistics [@ckislandsarrivalstats].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Elaine Angelino[^1]\
Harvard University\
`[email protected]`
- |
Varun Kanade[^2]\
UC Berkeley\
`[email protected]`
bibliography:
- 'all-refs.bib'
title: 'Attribute-Efficient Evolvability of Linear Functions'
---
Introduction {#sec:introduction}
============
Model and Preliminaries {#sec:notation}
=======================
Evolving Sparse Linear Functions {#sec:algorithms}
================================
Evolving Sparse Linear Functions Using $\optsel$ {#sec:greedy}
------------------------------------------------
Conclusion and Future Work {#sec:conclusion}
==========================
Acknowledgments {#acknowledgments .unnumbered}
---------------
We would like to thank Leslie Valiant for helpful discussions and comments on an earlier version of this paper. We are grateful to Frank Solomon for discussing biological aspects related to this work.
Omitted Proofs
==============
Proofs from Section \[sec:notation-class\] {#app:notation-class}
------------------------------------------
Proofs from Section \[sec:sparse\_linear\] {#app:sparse_linear}
------------------------------------------
Proofs from Section \[sec:greedy\] {#app:greedy}
----------------------------------
[^1]: This author is supported in part by a grant from the National Library of Medicine (4R01LM010213-04) and NSF grant CCF-09-64401
[^2]: This author is supported by a Simons Postdoctoral Fellowship.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss low-energy heterotic M-theory with five-branes in four and five dimensions and its application to moving brane cosmology.'
author:
- |
André Lukas\
[Centre for Theoretical Physics]{}\
[University of Sussex]{}\
[Falmer, Brighton BN1 9QJ, United Kingdom]{}\
[E-mail: [email protected]]{}
title: ' Moving Five-Branes and Cosmology [^1]'
---
Introduction
============
If string/M-theory is the correct fundamental theory of nature time-evolution of branes is likely to have played an important role in the history of the early universe. In this talk, I will review several aspects of this early universe brane dynamics in the context of strongly-coupled heterotic string-theory. This theory can be formulated as 11-dimensional supergravity on the orbifold $S^1/Z_2$ coupled to two 10-dimensional $E_8$ super-Yang-Mills multiplets localised on the two orbifold fixed planes (or boundaries) [@Horava:1996ma]. Upon compactification on a Calabi-Yau three-fold one arrives at a five-dimensional $N=1$ supergravity theory on $S^1/Z_2$ where the now four-dimensional boundaries carry $N=1$ gauge-theories which originate from the 10-dimensional Yang-Mills theories [@Lukas:1998yy; @Brandle:2001ts]. Further, M-theory five-branes can be included in this compactification. They wrap two-cycles within the internal Calabi-Yau space and, hence, appear as three-branes from the viewpoint of the effective five-dimensional theory. It is the dynamics of these M five-branes/three-branes which we are going to focus on in the following.
It should be stressed that our approach to early universe cosmology has the virtue of being embedded in a meaningful and successful environment for particle phenomenology from M-theory. Specifically, one can construct models within the class of compactifications described above which lead to phenomenologically promising theories close to the standard model of particle physics located on one of the boundaries [@Donagi:1999xe; @Donagi:2000zf].
The plan of this talk is as follows. To set the scene, I will present the effective five- and four-dimensional theories of heterotic M-theory including three-branes in their minimal versions. The status and interpretation of these effective theories as well as their general implications for three-brane dynamics will be reviewed. Subsequently, specific classes of solutions, both in four and five dimensions, describing simple three-brane evolution will be presented. Finally, we will discuss a new approach to brane-dynamics where three-branes are modelled as topological defects (kinks) of a bulk scalar field theory. Within this approach, more complicated brane evolution can be analysed by studying the dynamics of the modelling bulk scalar field. Particularly, what amounts to topology-changing transitions of the original M-theory action can be explicitly described by certain transitions within this bulk scalar field theory.
Effective actions in five and four dimensions
=============================================
The five-dimensional effective action for heterotic M-theory is given by a specific $D=5$, $N=1$ gauged supergravity in the bulk coupled to four-dimensional $N=1$ gauge theories on the boundaries and the three-branes. In its minimal version there are two bulk supermultiplets, namely the supergravity multiplet containing the metric, the gravi-photon and the gravitini and the universal hypermultiplet containing the dilaton $\F$ along with three pseudo-scalar partners and the associated fermions. On the boundaries $M_4^i$, where $i=1,2$, one obtains $N=1$ gauge theories with gauge groups $G_i\subset E_8$ and chiral matter fields transforming under $G_i$. The details of this boundary field content depend on the compactification from eleven to five dimensions which is the main subject of particle physics model building in heterotic M-theory. Each three-brane $M_4^3$ (and its $Z_2$ mirror $\tilde{M}_4^3$) carries on its world-volume a universal chiral multiplet containing the field $Y$, specifying the three-brane position in the transverse space, along with an axionic partner and $U(1)$ vector multiplets. The number of these vector multiplets is given by the genus $g$ of the curve wrapped by the M five-brane.
The effective action for these fields [@Lukas:1998yy; @Brandle:2001ts] can be consistently truncated to an even simpler version containing the metric $g_{\a\b}$ and the dilaton $\F$ as the only bulk fields and the embedding coordinates $X^\a = X^\a(\s^\m )$ as the only brane fields. For a single three-brane, the associated effective action is then given by $$\begin{aligned}
S_5 &=& -\frac{1}{2\k_5^2}\left\{\int_{M_5}\sqrt{-g}\left[
\frac{1}{2}R+\frac{1}{4}\partial_\a\F\partial^\a\F
+\frac{1}{3}\a^2 e^{-2\F}\right]\right. \nonumber \\
&&\qquad +\int_{M_4^1}\sqrt{-g}\; 2\a_1 e^{-\F}
+\int_{M_4^2}\sqrt{-g}\; 2\a_2 e^{-\F} \nonumber \\
&&\qquad \left. +\int_{M_4^3\cup\tilde{M}_4^3}\sqrt{-\g}\;
\a_3 e^{-\F}\right\}\; . \label{S5}\end{aligned}$$ where $\a_i=\s\b_i$, $\b_i\in {\bf Z}$ for $i=1,2,3$ are the charges on the boundaries and the three-brane quantised in units of $\s$. These charges satisfy the cohomology condition $$\sum_{i=1}^3\a_i = 0 \label{cohomology}\; .$$ The gauge-charge $\a$ is defined as a sum of step-functions $$\a = \a_1\theta (M_4^1) + \a_2\theta (M_4^2) + \a_3\left(\theta (M_4^3)+
\theta (\tilde{M}_4^3)\right) \; . \label{alpha}$$ Note that two actions of the above type with two different sets of charges $\a_i$ correspond to topologically different M-theory compactifications.
The four-dimensional effective action obtained from (\[S5\]) by compactifying on the orbifold is given by $$S_4 =
-\frac{1}{2\k_P^2}\int_{M_4}\sqrt{-g_4}\left[\frac{1}{2}R_4+
\frac{1}{4}\partial_\m\f\partial^\m\f +\frac{3}{4}\partial_\m\b\partial^\m\b
+\frac{q_3}{2}e^{\b -\f}\partial_\m z\partial^\m z\right]\; \label{S4}$$ where $q_3=\p\r\a_3$. Here, the field $\f$, as the zero mode of the five-dimensional dilaton $\F$ measures the Calabi-Yau size, while the field $\b$ originates from the $(55)$ component of the metric and measures the orbifold size. Finally, the field $z$ specifies the three-brane position and is normalised so that $z\in [0,1]$ with $z=0,1$ corresponding to the two boundaries. This effective action can be obtained from the Kähler potential [@Derendinger:2001gy] $$K = -\ln\left( S+\bar{S}-q_3\frac{(Z+\bar{Z})^2}{T+\bar{T}}\right)
-3\ln\left( T+\bar{T}\right)\; , \label{K}$$ by truncating off the axionic fields contained in the chiral superfields $S\leftrightarrow \f$, $T\leftrightarrow \b$ and $Z\leftrightarrow z$. Perturbatively, these moduli superfields are flat directions but one expects non-perturbative contributions to their superpotential from a number of sources [@Moore:2000fs; @Lima:2001nh]. Assuming the axions can still be integrated out, these contributions can be included by adding a suitable potential $V=V(\f ,\b ,z)$ to the above action. Note that this potential will be a function of all three fields rather than $z$ alone. It is quite clear even from the free action (\[S4\]) that three-brane evolution in a static geometry (that is, for constant $\f$ and $\b$) is not possible. For this reason, the above action (possibly with a potential) represents the minimal system in which heterotic moving-brane cosmology can be discussed in a meaningful way.
Cosmological solutions with a moving three-brane
================================================
The most general cosmological solutions of (\[S4\]) with flat spatial sections [@Copeland:2001zp] show a number of interesting generic features. Asymptotically at early and late time, both in the negative and positive time branch, the three-brane is practically at rest while the dilaton $\f$ and the T-modulus $\b$ evolve according to a standard rolling radii solution. However, the early and late rolling radii solutions are generally different. Interpolation between these rolling radii solutions is achieved at intermediate time due to non-trivial evolution of the three-brane. It turns out that not all rolling radii solutions can be approached asymptotically but only those for which the strong-coupling expansion parameter $$\e \sim e^{\b -\f}$$ diverges at early and late time. The brane-evolution, therefore, drives the system to strong coupling asymptotically. This is a direct consequence of the nontrivial kinetic term for $z$ in the action (\[S4\]) and it is a feature that one would have missed had one - inconsistently - studied the brane-evolution in a static geometry.
Further, for all solutions the three-brane moves for a finite coordinate distance only. As a consequence, the brane may or may not collide with one of the boundaries, depending on initial conditions. The negative-time branch of these solution constitutes the correct starting point for moving-brane pre-big-bang cosmology which may terminate in a brane collision. Whether or not this collision will turn pre-big-bang contraction of the universe into expansion, thereby generating a “graceful exit”, must currently be viewed as an open problem.
The five-dimensional origin of these four-dimensional moving brane solution is not known at present except for a number of special cases. There are two four-dimensional solutions with $\e =$ const which can be directly lifted to five dimensions [@Copeland:2001zp]. A less trivial example of a five-dimensional moving brane solution which reduces to one of the above $D=4$ solutions in a specific limit only has been found more recently [@Copeland:2002fv].
Modelling three-branes with kinks
=================================
Studying more complicated dynamics of three-branes, such as brane-brane or brane-boundary collision, requires a more microscopic understanding of M-theory which is currently out of reach. However, these processes can be studied in the context of a toy model where the three-branes are replaced by smooth kink solutions of a bulk scalar field theory. The following action [@Antunes:2002hn] for this toy model, replacing the M-theory action (\[S5\]), $$\begin{aligned}
\tilde{S}_5 &=& -\frac{1}{2\k_5^2}\left\{\int_{M_5}\sqrt{-g}\left[
\frac{1}{2}R+\frac{1}{4}\partial_\a\F\partial^\a\F
+\frac{1}{2}e^{-\F}\partial_\a\c\partial^\a\c
+V(\F ,\c )\right]\right.\nonumber \\
&&\qquad\left. +\int_{M_4^1}\sqrt{-g}\; 2W
-\int_{M_4^2}\sqrt{-g}\; 2W\right\}\; .
\label{S5t}\end{aligned}$$ has been proposed recently. Here, the potential $V$ is obtained from a “superpotential” $W=e^{-\F}w(\c )$ by $$V = \frac{1}{3}e^{-2\F}w^2+\frac{1}{2}e^{-\F}U\; ,\qquad
U = \left(\frac{dw}{d\c}\right)^2\; . \label{U}$$ The potential $U$ for the new bulk field $\c$ is required to be periodic with period $v$, that is $U(\c +v)=U(\c)$ and to have minima for all $\c =nv$, where $n\in{\bf Z}$, with $U(nv)=0$. These conditions can be easily translated into conditions on the function $w(\c )=\int_0^\c d\tilde{\c}\sqrt{U(\tilde{\c})}$ which specifies the superpotential. An example for $U$ is provided by the sine-Gordon potential $$U = m^2\left[1-\cos\left(\frac{2\p\c}{v}\right)\right]\; ,$$ although the specific form of the potential is largely irrelevant for the subsequent discussion as long as the above general properties are satisfied.
How does (\[S5t\]) relate to the original M-theory action (\[S5\])? Consider first a situation where the field $\c$ is constant throughout space-time and is located in the $n^{\rm th}$ minimum of $U$, that is, $\c = nv$. Eq. (\[S5t\]) then reduces to the M-theory action (\[S5\]) without a three-brane, corresponding to a charge configuration $$(\b_1,\b_2,\b_3)=(n,-n,0)\label{c1}\; ,$$ if the identification $\s = w(v)$ is being used.
The equations of motion derived from the action (\[S5t\]) have Bogomol’nyi-type first integrals. These first order equations can be used to obtain a BPS single-kink solution which interpolates between $\c=nv$ on the first boundary and $\c =(n+1)v$ on the second boundary. In a thin-kink approximation the $3+1$–dimensional surface defined by the core of the kink can be described by a Nambu-Goto action and can, therefore, be identified with the three-brane in Eq. (\[S5\]). Hence, in the background of a single kink, the action (\[S5t\]) reduces to the M-theory action with a single-charged three-brane, corresponding to charges $$(\b_1,\b_2,\b_3)=(n,-(n+1),1)\; .\label{c2}$$ Note that different configurations of the scalar field $\c$ in the defect model correspond to topologically distinct M-theory models. It has been shown [@Antunes:2002hn] that a collision of a kink with a boundary generically leads to a transition between a vacuum state (\[c2\]) with a single-charged three-brane and a vacuum state (\[c1\]) without a three-brane. Topological transitions on the M-theory side can, therefore, be modelled by simple scalar field dynamics using the action (\[S5t\]). This opens up the possibility of modelling the evolution of more complicated configurations such as networks of three-branes in the early universe.
Acknowledgements {#acknowledgements .unnumbered}
================
The author is supported by a PPARC Advanced Fellowship.
[0]{}
P. Horava and E. Witten, “Eleven-Dimensional Supergravity on a Manifold with Boundary,” Nucl. Phys. B [**475**]{} (1996) 94 \[hep-th/9603142\]. A. Lukas, B. A. Ovrut, K. S. Stelle and D. Waldram, “The universe as a domain wall,” Phys. Rev. D [**59**]{} (1999) 086001 \[hep-th/9803235\]. M. Brandle and A. Lukas, “Five-branes in heterotic brane-world theories,” Phys. Rev. D [**65**]{} (2002) 064024 \[hep-th/0109173\]. R. Donagi, A. Lukas, B. A. Ovrut and D. Waldram, “Non-perturbative vacua and particle physics in M-theory,” JHEP [**9905**]{} (1999) 018 \[hep-th/9811168\]. R. Donagi, B. A. Ovrut, T. Pantev and D. Waldram, JHEP [**0108**]{} (2001) 053 \[hep-th/0008008\]. J. Derendinger and R. Sauser, “A five-brane modulus in the effective N = 1 supergravity of M-theory,” Nucl. Phys. B [**598**]{} (2001) 87 \[hep-th/0009054\].
G. W. Moore, G. Peradze and N. Saulina, instantons,” Nucl. Phys. B [**607**]{} (2001) 117 \[hep-th/0012104\]. E. Lima, B. A. Ovrut and J. Park, Nucl. Phys. B [**626**]{} (2002) 113 \[hep-th/0102046\]. E. J. Copeland, J. Gray and A. Lukas, “Moving five-branes in low-energy heterotic M-theory,” Phys. Rev. D [**64**]{} (2001) 126003 \[hep-th/0106285\]. E. J. Copeland, J. Gray, A. Lukas and D. Skinner, \[hep-th/0207281\]. N. D. Antunes, E. J. Copeland, M. Hindmarsh and A. Lukas, \[hep-th/0208219\].
[^1]: Based on a talk given at the 1st International Conference on String Phenomenology, Oxford, 6-11 Jul 2002.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.