19
2

Improved Langevin Monte Carlo for stochastic optimization via landscape modification

Abstract

Given a target function HH to minimize or a target Gibbs distribution πβ0eβH\pi_{\beta}^0 \propto e^{-\beta H} to sample from in the low temperature, in this paper we propose and analyze Langevin Monte Carlo (LMC) algorithms that run on an alternative landscape as specified by Hβ,c,1fH^f_{\beta,c,1} and target a modified Gibbs distribution πβ,c,1feβHβ,c,1f\pi^f_{\beta,c,1} \propto e^{-\beta H^f_{\beta,c,1}}, where the landscape of Hβ,c,1fH^f_{\beta,c,1} is a transformed version of that of HH which depends on the parameters f,βf,\beta and cc. While the original Log-Sobolev constant affiliated with πβ0\pi^0_{\beta} exhibits exponential dependence on both β\beta and the energy barrier MM in the low temperature regime, with appropriate tuning of these parameters and subject to assumptions on HH, we prove that the energy barrier of the transformed landscape is reduced which consequently leads to polynomial dependence on both β\beta and MM in the modified Log-Sobolev constant associated with πβ,c,1f\pi^f_{\beta,c,1}. This yield improved total variation mixing time bounds and improved convergence toward a global minimum of HH. We stress that the technique developed in this paper is not only limited to LMC and is broadly applicable to other gradient-based optimization or sampling algorithms.

View on arXiv
Comments on this paper