Higher Order Langevin Monte Carlo Algorithm

A new (unadjusted) Langevin Monte Carlo (LMC) algorithm with improved rates in total variation and Wasserstein distance is presented in the context of sampling from a target distribution $\pi$ under the assumption that its density on $\mathbb{R}^d$ is known up to a normalizing constant. Crucially, the Langevin SDE associated with the target distribution $\pi$ is assumed to have a locally Lipschitz drift coefficient such that its second derivative is locally H\"{o}lder continuous with exponent $\beta \in (0,1]$. In addition, non-asymptotic bounds are obtained for the convergence to stationarity of the new sampling method with convergence rate $1+ \beta/2$ in Wasserstein distance, while it is shown that the rate is 1 in total variation. The bounds between the invariant measure of the LMC algorithm and $\pi$ are also given.