71
167
v1v2v3v4 (latest)

Sharp convergence rates for Langevin dynamics in the nonconvex setting

Abstract

We study the problem of sampling from a distribution p(x)exp(U(x))p^*(x) \propto \exp\left(-U(x)\right), where the function UU is LL-smooth everywhere and mm-strongly convex outside a ball of radius RR, but potentially nonconvex inside this ball. We study both overdamped and underdamped Langevin MCMC and establish upper bounds on the number of steps required to obtain a sample from a distribution that is within ϵ\epsilon of pp^* in 11-Wasserstein distance. For the first-order method (overdamped Langevin MCMC), the iteration complexity is O~(ecLR2d/ϵ2)\tilde{\mathcal{O}}\left(e^{cLR^2}d/\epsilon^2\right), where dd is the dimension of the underlying space. For the second-order method (underdamped Langevin MCMC), the iteration complexity is O~(ecLR2d/ϵ)\tilde{\mathcal{O}}\left(e^{cLR^2}\sqrt{d}/\epsilon\right) for an explicit positive constant cc. Surprisingly, the iteration complexity for both these algorithms is only polynomial in the dimension dd and the target accuracy ϵ\epsilon. It is exponential, however, in the problem parameter LR2LR^2, which is a measure of non-log-concavity of the target distribution.

View on arXiv
Comments on this paper