Rapid convergence of Langevin AIS for multimodal distributions on $\mathbb{R}^d$

Abstract

We study annealed importance sampling (AIS) driven by the overdamped Langevin diffusion, targeting a multimodal Gibbs measure $π_\varepsilon \propto e^{-U/\varepsilon}$ on $\mathbb{R}^d$ at low temperature $\varepsilon$. For a potential $U$ with nondegenerate critical points, finitely many wells of comparable mass, and at least quadratic growth at infinity, we show that AIS with a harmonically spaced temperature schedule produces a weighted empirical measure that approximates $π_\varepsilon$ to accuracy $δ$ at a total cost of order $\varepsilon^{-2}δ^{-2}$, up to logarithmic factors, and the exponents do not depend on the dimension. This is an exponential improvement over plain Langevin Monte Carlo, whose relaxation time obeys the Arrhenius law. The assumptions cover, in particular, the potentials of Gaussian mixtures, as long as their critical points are nondegenerate and their wells carry comparable mass. The analysis extends that of arXiv:2604.17526 [math.PR] from the torus $\mathbb{T}^d$ to the noncompact state space $\mathbb{R}^d$. The two new ingredients are a weighted $L^\infty$ bound on the eigenfunctions of the Langevin generator, obtained by combining Agmon estimates at infinity with a local maximum principle near the critical points, and a Weyl-type eigenvalue count derived from the Cwikel-Lieb-Rozenblum and Lieb-Thirring inequalities. For potentials with more than two wells the eigenfunction bound requires an additional argument, which we give by constructing a perturbation of $U$ that renders every saddle an effective local maximum.