Concentration of the Langevin Algorithm's Stationary Distribution

A canonical algorithm for log-concave sampling is the Langevin Algorithm, aka the Langevin Diffusion run with some discretization stepsize . This discretization leads the Langevin Algorithm to have a stationary distribution which differs from the stationary distribution of the Langevin Diffusion, and it is an important challenge to understand whether the well-known properties of extend to . In particular, while concentration properties such as isoperimetry and rapidly decaying tails are classically known for , the analogous properties for are open questions with direct algorithmic implications. This note provides a first step in this direction by establishing concentration results for that mirror classical results for . Specifically, we show that for any nontrivial stepsize , is sub-exponential (respectively, sub-Gaussian) when the potential is convex (respectively, strongly convex). Moreover, the concentration bounds we show are essentially tight. Key to our analysis is the use of a rotation-invariant moment generating function (aka Bessel function) to study the stationary dynamics of the Langevin Algorithm. This technique may be of independent interest because it enables directly analyzing the discrete-time stationary distribution without going through the continuous-time stationary distribution as an intermediary.
View on arXiv