Sampling from a log-concave distribution with projected Langevin Monte Carlo (Q1650786)

The authors establish a bound on the number of steps required for the Projected Langevin Monte Carlo to attain a small total variation distance to the target measure. The properties of reflected Brownian is used to obtain a total variation bound from the Wasserstein bound. The arguments are generalized to an arbitrary smooth potential. Some preliminary experimental comparison between Langevin Monte Carlo and hit-and-run is presented.