A1222
Title: Langevin-type Monte Carlo algorithms for weakly differentiable non-convex potentials
Authors: Shogo Nakakita - The University of Tokyo (Japan) [presenting]
Abstract: Langevin-type Monte Carlo algorithms for distributions with non-convex non-smooth potential functions are considered. If a potential has a weak gradient whose fluctuation within all balls of radius 1 is uniformly bounded, then spherical smoothing can be used to approximate the potential with smoother functions. The spherically smoothed Langevin Monte Carlo algorithm and spherically smoothed stochastic gradient Langevin Monte Carlo one are proposed, and their sampling complexities are discussed.
