Multilevel-Langevin pathwise average for Gibbs approximation

TL;DR

This paper introduces a multilevel Langevin-based method for efficiently approximating Gibbs distributions, achieving near-optimal complexity bounds and demonstrating effectiveness through numerical experiments in Bayesian learning.

Contribution

It develops a novel multilevel occupation measure approach for Langevin diffusions, providing theoretical complexity bounds and practical algorithms for Gibbs distribution approximation.

Findings

Achieves b5$-approximation with cost a a e ext{or}e| ext{log}b5|b3

Provides complexity bounds of 3a a db5$^{-2} ext{log}^3(db5$^{-2})$ for strongly convex potentials

Numerical illustrations show competitive performance in Bayesian learning tasks

Abstract

We propose and study a new multilevel method for the numerical approximation of a Gibbs distribution $\pi$ on $\mathbb{R}^d$, based on (overdamped) Langevin diffusions. This method inspired by \cite{mainPPlangevin} and \cite{giles_szpruch_invariant} relies on a multilevel occupation measure, $i.e.$ on an appropriate combination of $R$ occupation measures of (constant-step) Euler schemes with respective steps $\gamma_r = \gamma_0 2^{-r}$, $r=0,\ldots,R$. We first state a quantitative result under general assumptions which guarantees an \textit{$\varepsilon$-approximation} (in a $L^2$-sense) with a cost of the order $\varepsilon^{-2}$ or $\varepsilon^{-2}|\log \varepsilon|^3$ under less contractive assumptions. We then apply it to overdamped Langevin diffusions with strongly convex potential $U:\mathbb{R}^d\rightarrow\mathbb{R}$ and obtain an \textit{$\varepsilon$-complexity} of the order ${\cal O}(d\varepsilon^{-2}\log^3(d\varepsilon^{-2}))$ or ${\cal O}(d\varepsilon^{-2})$ under additional assumptions on $U$. More precisely, up to universal constants, an appropriate choice of the parameters leads to a cost controlled by ${(\bar{\lambda}_U\vee 1)^2}{\underline{\lambda}_U^{-3}} d\varepsilon^{-2}$ (where $\bar{\lambda}_U$ and $\underline{\lambda}_U$ respectively denote the supremum and the infimum of the largest and lowest eigenvalue of $D^2U$). We finally complete these theoretical results with some numerical illustrations including comparisons to other algorithms in Bayesian learning and opening to non strongly convex setting.