Convergence of Langevin-Simulated Annealing algorithms with multiplicative noise

TL;DR

This paper analyzes the convergence of Langevin-Simulated Annealing algorithms with multiplicative noise, providing theoretical guarantees and convergence rates in Wasserstein distance for optimization problems relevant to Machine Learning.

Contribution

It extends convergence results to the general case with multiplicative noise, including adaptive noise, and establishes rates for the associated Euler scheme.

Findings

Proves convergence of the process and Euler scheme to a measure supported on the minimizers.

Establishes convergence rates to the instantaneous Gibbs measure.

Identifies the classical cooling schedule for the annealing process.

Abstract

We study the convergence of Langevin-Simulated Annealing type algorithms with multiplicative noise, i.e. for $V : \mathbb{R}^d \to \mathbb{R}$ a potential function to minimize, we consider the stochastic equation $dY_t = - \sigma \sigma^\top \nabla V(Y_t) dt + a(t)\sigma(Y_t)dW_t + a(t)^2\Upsilon(Y_t)dt$, where $(W_t)$ is a Brownian motion, where $\sigma : \mathbb{R}^d \to \mathcal{M}_d(\mathbb{R})$ is an adaptive (multiplicative) noise, where $a : \mathbb{R}^+ \to \mathbb{R}^+$ is a function decreasing to $0$ and where $\Upsilon$ is a correction term. This setting can be applied to optimization problems arising in Machine Learning. The case where $\sigma$ is a constant matrix has been extensively studied however little attention has been paid to the general case. We prove the convergence for the $L^1$-Wasserstein distance of $Y_t$ and of the associated Euler-scheme $\bar{Y}_t$ to some measure $\nu^\star$ which is supported by $\text{argmin}(V)$ and give rates of convergence to the instantaneous Gibbs measure $\nu_{a(t)}$ of density $\propto \exp(-2V(x)/a(t)^2)$. To do so, we first consider the case where $a$ is a piecewise constant function. We find again the classical schedule $a(t) = A\log^{-1/2}(t)$. We then prove the convergence for the general case by giving bounds for the Wasserstein distance to the stepwise constant case using ergodicity properties.