News
Updates to the results, comparisons, and supporting material in samplewiki.
- Added sharp Wasserstein bounds for midpoint discretizations of Langevin dynamics, showing that deterministic midpoint methods can beat randomized ones in high dimension.
- Added Wasserstein bias bounds for stochastic gradient LMC, which improve under a fourth moment assumption on the gradient noise.
- Added the Gaussian cloud sampler: a new high-accuracy bound for strongly log-concave sampling with improved dimension dependence.
- Added the optimal fixed-dimensional bounds for sampling with bounded variance stochastic gradients.
- Added Proximal BPS and smoothed Picard HMC: new high-accuracy and KL bounds for strongly log-concave sampling.
- Added a lower bound showing that the FI rate for averaged LMC in the non-log-concave setting is tight.