Searcharxiv⌕ Search

arXiv subjects

Luis Fredes

Publications and source records attributed to Luis Fredes.

16 recordsLinked to original sources
arXiv2026

Kantorovich Distance via Spanning Trees: Properties and Algorithms

We study optimal transport between probability measures supported on the same finite metric space, where the ground cost is a distance induced by a weighted connected graph. Building on recent work showing that the resulting Kantorovich distance can be expressed as a minimization problem over the set of spanning trees of this underlying graph, we investigate the implications of this reformulation on the construction of an optimal transport plan and a dual potential based on the solution of such an optimization problem. In this setting, we derive an explicit formula for the Kantorovich potential in terms of the imbalanced cumulative mass (a generalization of the cumulative distribution in R) along an optimal spanning tree solving such a minimization problem, under a weak non-degeneracy condition on the pair of measures that guarantees the uniqueness of a dual potential. Our second contribution establishes the existence of an optimal transport plan that can be computed efficiently by a dynamic programming procedure once an optimal spanning tree is known. Finally, we propose a stochastic algorithm based on simulated annealing on the space of spanning trees to compute such an optimal spanning tree. Numerical experiments illustrate the theoretical results and demonstrate the practical relevance of the proposed approach for optimal transport on finite metric spaces.

math.OC↗
arXiv2025

An hybrid stochastic Newton algorithm for logistic regression

In this paper, we investigate a second-order stochastic algorithm for solving large-scale binary classification problems. We propose to make use of a new hybrid stochastic Newton algorithm that includes two weighted components in the Hessian matrix estimation: the first one coming from the natural Hessian estimate and the second associated with the stochastic gradient information. Our motivation comes from the fact that both parts evaluated at the true parameter of logistic regression, are equal to the Hessian matrix. This new formulation has several advantages and it enables us to prove the almost sure convergence of our stochastic algorithm to the true parameter. Moreover, we significantly improve the almost sure rate of convergence to the Hessian matrix. Furthermore, we establish the central limit theorem for our hybrid stochastic Newton algorithm. Finally, we show a surprising result on the almost sure convergence of the cumulative excess risk.

stat.CO↗
arXiv2025

A combinatorial perspective on the Kemeny constant and more

Let $M$ be an irreducible transition matrix on a finite state space $V$. For a Markov chain $C=(C_k,k\geq 0)$ with transition matrix $M$, let $\tau^{\geq 1}_u$ denote the first positive hitting time of $u$ by $C$, and $\rho$ the unique invariant measure of $M$. Kemeny proved that if $X$ is sampled according to $\rho$ independently of $C$, the expected value of the first positive hitting time of $X$ by $C$ does not depend on the starting state of the chain: all the values $(E(\tau^{\geq 1}_X~|~C_0=u), u \in V)$ are equal. \par In this paper, we show that this property follows from a more general result: the generating function $\sum_{v\in V}E(x^{\tau_v^{\geq 1}}~|~C_0=u)\det(Id-xM^{(v)})$ is independent of the starting state $u$, where $M^{(v)}$ is obtained from $M$ by deleting the row and column corresponding to the state $v$. The factors appearing in this generating function are: first, the probability generating function of $\tau^{\geq 1}_v$, and second, the sequence of determinants $(det(Id-xM^{(v)}),v\in V),$ which, for $x=1$, is known to be proportional to the invariant measure $(\rho_u,u\in V)$. From this property, we deduce several further results, including relations involving higher moments of $\tau_X^{\geq 1}$, which are of independent interest.

math.PR↗
arXiv2025

Characterisation of Markov property on planar maps

We revisit, in a self contained way, the Markov property on planar maps and decorated planar maps from three perspectives. First, we characterize the laws on these planar maps that satisfy both the Markov property and rerooting invariance, showing that they are Boltzmann-type maps. Second, we provide a comprehensive characterization of random submaps, that we call stopping maps, satisfying the Markov property, demonstrating that they are not restricted to those obtained through a peeling procedure. Third, we introduce decorated metric planar maps in which edges are replaced by copies of random length intervals $[0,w_e]$, and the decorations are given by continuous functions on the edges. We define a probability measure on them that is the analogue of the Boltzmann map and show that it satisfies the Markov property even for sets that halt exploration mid-edge.

math.PR↗
arXiv2024

Markov chains on trees: almost lower and upper directed cases

The transition matrix of a Markov chain $(X_k,k\geq 0)$ on a finite or infinite rooted tree is said to be almost upper-directed if, given $X_k$, the node $X_{k+1}$ is either a descendant of $X_k$ or the parent of $X_k$. It is said to be almost lower-directed if given $X_k$, $X_{k+1}$ is either an ancestor of $X_k$ or a child of $X_k$. These models include nearest neighbor Markov chains on trees. Under an irreducibility assumption, we show that every almost upper-directed transition matrix on infinite (locally finite) trees has some invariant measures. An invariant measure $\pi$ is expressed thanks to a determinantal formula. We give general explicit criteria for recurrence and positive recurrence. An efficient algorithm (the leaf addition algorithm) of independent interest allows $\pi$ to be computed on many trees, without resorting to linear algebra considerations. Flajolet, in a series of papers, provided some relations between continuous fractions, generating functions of weighted M\"otzkin paths, and used them in connection with the analysis of birth and death processes. These fruitful representations made it possible to establish many formulae for continuous fractions. Analogous considerations appear here: this type of study can be extended to weighted paths on trees, whose generating functions can also be expressed, this time in terms of multicontinuous fractions.

math.PR↗
arXiv2024

On the SAGA algorithm with decreasing step

Stochastic optimization naturally appear in many application areas, including machine learning. Our goal is to go further in the analysis of the Stochastic Average Gradient Accelerated (SAGA) algorithm. To achieve this, we introduce a new $\lambda$-SAGA algorithm which interpolates between the Stochastic Gradient Descent ($\lambda=0$) and the SAGA algorithm ($\lambda=1$). Firstly, we investigate the almost sure convergence of this new algorithm with decreasing step which allows us to avoid the restrictive strong convexity and Lipschitz gradient hypotheses associated to the objective function. Secondly, we establish a central limit theorem for the $\lambda$-SAGA algorithm. Finally, we provide the non-asymptotic $\mathbb{L}^p$ rates of convergence.

math.OC↗
arXiv2024

Sharp analysis on the joint distribution of the number of descents and inverse descents in a random permutation

Chatteerjee and Diaconis have recently shown the asymptotic normality for the joint distribution of the number of descents and inverse descents in a random permutation. A noteworthy point of their results is that the asymptotic variance of the normal distribution is diagonal, which means that the number of descents and inverse descents are asymptotically uncorrelated.The goal of this paper is to go further in this analysis by proving a large deviation principlefor the joint distribution. We shall show that the rate function of the joint distributionis the sum of the rate functions of the marginal distributions, which also means that the number of descents and inverse descents are asymptotically independent at the large deviation level. However,we are going to prove that they are finely dependent at the sharp large deviation level.

math.CO↗
arXiv2023

On the triviality of the shocked map

The (non-spanning) tree-decorated quadrangulation is a random pair formed by a quadrangulation and a subtree chosen uniformly over the set of pairs with prescribed size. In this paper we study the tree-decorated quadrangulation in the critical regime: when the number of faces of the map, $f$, is proportional to the square of the size of the tree. We show that with high probability in this regime, the diameter of the tree is between $o(f^{1/4})$ and $f^{1/4}/\log^\alpha(f)$, for $\alpha >1$. Thus after scaling the distances by $f^{-1/4}$, the critical tree-decorated quadrangulation converges to a Brownian disk where the boundary has been identified to a point. These results imply the triviality of the shocked map: the metric space generated by gluing a Brownian disk with a continuous random tree.

math.PR↗
arXiv2021

Almost triangular Markov chains on $\mathbb{N}$

A transition matrix $[U_{i,j}]_{i,j\geq 0}$ on $\mathbb{N}$ is said to be almost upper triangular if $U_{i,j}\geq 0\Rightarrow j\geq i-1$, so that the increments of the corresponding Markov chains are at least $-1$; a transition matrix $[L_{i,j}]_{i,j\geq 0}$ is said to be almost lower triangular if $L_{i,j}\geq 0\Rightarrow j\leq i+1$, and then, the increments of the corresponding Markov chains are at most $+1$. In the present paper, we characterize the recurrence, positive recurrence and invariant distribution for the class of almost triangular transition matrices. The upper case appears to be the simplest in many ways, with existence and uniqueness of invariant measures, when in the lower case, existence as well as uniqueness are not guaranteed. We present the time-reversal connection between upper and lower almost triangular transition matrices, which provides classes of integrable lower triangular transition matrices. These results encompass the case of birth and death processes (BDP) that are famous Markov chains (or processes) taking their values in $\mathbb{N}$, which are simultaneously almost upper and almost lower triangular, and whose study has been initiated by Karlin & McGregor in the 1950's. They found invariant measures, criteria for recurrence, null recurrence, among others; their approach relies on some profound connections they discovered between the theory of BDP, the spectral properties of their transition matrices, the moment problem, and the theory of orthogonal polynomials. Our approach is mainly combinatorial and uses elementary algebraic methods; it is somehow more direct and does not use the same tools.

math.PR↗
arXiv2021

Scaling limit of random plane quadrangulations with a simple boundary, via restriction

We prove that quadrangulations with a simple boundary converge to the Brownian disk. More precisely, we fix a sequence $(p_n)$ of even positive integers with $p_n\sim 2\alpha \sqrt{2n}$ for some $\alpha\in(0,\infty)$. Then, for the Gromov--Hausdorff topology, a quadrangulation with a simple boundary uniformly sampled among those with $n$ inner faces and boundary length $p_n$ weakly converges, in the usual scaling $n^{-1/4}$, toward the Brownian disk of perimeter $3\alpha$. Our method consists in seeing a uniform quadrangulation with a simple boundary as a conditioned version of a model of maps for which the Gromov--Hausdorff scaling limit is known. We then explain how classical techniques of unconditionning can be used in this setting of random maps.

math.PR↗
arXiv2021

Models of random subtrees of a graph

Consider a connected graph $G=(E,V)$ with $N=|V|$ vertices. The main purpose of this paper is to explore the question of uniform sampling of a subtree of $G$ with $n$ nodes, for some $n\leq N$ (the spanning tree case correspond to $n=N$, and is already deeply studied in the literature). We provide new asymptotically exact simulation methods using Markov chains for general connected graphs $G$, and any $n\leq N$. We highlight the case of the uniform subtree of $\mathbb{Z}^2$ with $n$ nodes, containing the origin $(0,0)$ for which Schramm asked several questions. We produce pictures, statistics, and some conjectures. A second aim of the paper is devoted to surveying other models of random subtrees of a graph, among them, DLA models, the first passage percolation, the uniform spanning tree and the minimum spanning tree. We also provide new models, some statistics, and some conjectures.

math.PR↗
arXiv2021

A combinatorial proof of Aldous-Broder theorem for general Markov chains

Aldous-Broder algorithm is a famous algorithm used to sample a uniform spanning tree of any finite connected graph $G$, but it is more general: given an irreducible and reversible Markov chain $M$ on $G$ started at $r$, the tree rooted at $r$ formed by the first entrance steps in each node (different from the root) has a probability proportional to $\prod_{e=(e^-,e^+)\in {\sf Edges}(t,r)} M_{e^{-},e^+}$, where the edges are directed toward $r$. In this paper we give proofs of Aldous-Broder theorem in the general case, where the kernel $M$ is irreducible but not assumed to be reversible (this generalized version appeared recently in Hu, Lyons and Tang )

math.CO↗
arXiv2019

Tree-decorated planar maps

We introduce the set of (non-spanning) tree-decorated planar maps, and show that they are in bijection with the Cartesian product between the set of trees and the set of maps with a simple boundary. As a consequence, we count the number of tree decorated triangulations and quadrangulations with a given amount of faces and for a given size of the tree. Finally, we generalise the bijection to study other types of decorated planar maps and obtain explicit counting formulas for them.

math.CO↗
arXiv2018

Coexistence for a population model with forest fire epidemics

We investigate the effect on survival and coexistence of introducing forest fire epidemics to a certain two-species competition model. The model is an extension of the one introduced by Durrett and Remenik [DR09], who studied a discrete time particle system running on a random 3-regular graph where occupied sites grow until they become sufficiently dense so that an epidemic wipes out large clusters. In our extension we let two species affected by independent epidemics compete for space, and we allow the epidemic to attack not only giant clusters, but also clusters of smaller order. Our main results show that, for the two-type model, there are explicit parameter regions where either one species dominates or there is coexistence; this contrasts with the behavior of the model without epidemics, where the fitter species always dominates. We also discuss the survival and extinction regimes for the model with a single species. In both cases we prove convergence to explicit dynamical systems; simulations suggest that their orbits present chaotic behavior.

math.PR↗
arXiv2017

Invariant measures of discrete interacting particle systems: Algebraic aspects

Consider a continuous time particle system $\eta^t=(\eta^t(k),k\in \mathbb{L})$, indexed by a lattice $\mathbb{L}$ which will be either $\mathbb{Z}$, $\mathbb{Z}/n\mathbb{Z}$, a segment $\{1,\cdots, n\}$, or $\mathbb{Z}^d$, and taking its values in the set $E_{\kappa}^{\mathbb{L}}$ where $E_{\kappa}=\{0,\cdots,\kappa-1\}$ for some fixed $\kappa\in\{\infty, 2,3,\cdots\}$. Assume that the Markovian evolution of the particle system (PS) is driven by some translation invariant local dynamics with bounded range, encoded by a jump rate matrix $T$. These are standard settings, satisfied by the TASEP, the voter models, the contact processes... The aim of this paper is to provide some sufficient and/or necessary conditions on the matrix $T$ so that this Markov process admits some simple invariant distribution, as a product measure (if $\mathbb{L}$ is any of the spaces mentioned above), as the law of a Markov process indexed by $\mathbb{Z}$ or $[0,n]\cap \mathbb{Z}$ (if $\mathbb{L}=\mathbb{Z}$ or $\{1,\cdots,n\}$), or a Gibbs measure if $\mathbb{L}=\mathbb{Z}/n\mathbb{Z}$. Multiple applications follow: efficient ways to find invariant Markov laws for a given jump rate matrix or to prove that none exists. The voter models and the contact processes are shown not to possess any Markov laws as invariant distribution (for any memory $m$) Footnote: As usual, a random process $X$ indexed by $\mathbb{Z}$ or $\mathbb{N}$ is said to be a Markov chain with memory $m\in\{0,1,2,\cdots\}$ if for $\mathbb{P}(X_k \in A ~| X_{k-i}, i \geq 1)= \mathbb{P}(X_k \in A ~| X_{k-i}, 1\leq i\leq m)$, for any $k$.}. We also prove that some models close to these models do. We exhibit PS admitting hidden Markov chains as invariant distribution and design many PS on $\mathbb{Z}^2$, with jump rates indexed by $2\times 2$ squares, admitting product invariant measures.

math.PR↗