Accelerating the diffusion-based ensemble sampling by non-reversible dynamics

This work proposed a novel ensemble Monte Carlo sampling scheme by using the non-reversible technique.

Download paper here

Outline

• We propose a new ensemble sampling method based on the non-reversible Markov chain technique. Then, we theoretically analyze the proposed sampling scheme in terms of the 2-Wasserstein distance.
• We find that the interaction causes a trade-off between a larger discretization error and faster convergence to the stationary distribution.
• We conduct numerical experiments to confirm that we can control the trade-off by tuning the interaction appropriately.