The ballistic limit of the log-Sobolev constant equals the Polyak-Łojasiewicz constant

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Chewi, Sinho, Stromme, Austin J.
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866915024541319168
author Chewi, Sinho
Stromme, Austin J.
author_facet Chewi, Sinho
Stromme, Austin J.
contents The Polyak-Lojasiewicz (PL) constant of a function $f \colon \mathbb{R}^d \to \mathbb{R}$ characterizes the best exponential rate of convergence of gradient flow for $f$, uniformly over initializations. Meanwhile, in the theory of Markov diffusions, the log-Sobolev (LS) constant plays an analogous role, governing the exponential rate of convergence for the Langevin dynamics from arbitrary initialization in the Kullback-Leibler divergence. We establish a new connection between optimization and sampling by showing that the low temperature limit $\lim_{t\to 0^+} t^{-1} C_{\mathsf{LS}}(μ_t)$ of the LS constant of $μ_t \propto \exp(-f/t)$ is exactly the PL constant of $f$, under mild assumptions. In contrast, we show that the corresponding limit for the Poincaré constant is the inverse of the smallest eigenvalue of $\nabla^2 f$ at the minimizer.
format Preprint
id arxiv_https___arxiv_org_abs_2411_11415
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle The ballistic limit of the log-Sobolev constant equals the Polyak-Łojasiewicz constant
Chewi, Sinho
Stromme, Austin J.
Probability
Functional Analysis
Optimization and Control
The Polyak-Lojasiewicz (PL) constant of a function $f \colon \mathbb{R}^d \to \mathbb{R}$ characterizes the best exponential rate of convergence of gradient flow for $f$, uniformly over initializations. Meanwhile, in the theory of Markov diffusions, the log-Sobolev (LS) constant plays an analogous role, governing the exponential rate of convergence for the Langevin dynamics from arbitrary initialization in the Kullback-Leibler divergence. We establish a new connection between optimization and sampling by showing that the low temperature limit $\lim_{t\to 0^+} t^{-1} C_{\mathsf{LS}}(μ_t)$ of the LS constant of $μ_t \propto \exp(-f/t)$ is exactly the PL constant of $f$, under mild assumptions. In contrast, we show that the corresponding limit for the Poincaré constant is the inverse of the smallest eigenvalue of $\nabla^2 f$ at the minimizer.
title The ballistic limit of the log-Sobolev constant equals the Polyak-Łojasiewicz constant
topic Probability
Functional Analysis
Optimization and Control
url https://arxiv.org/abs/2411.11415