A Provable Comparison Scheme via Scaling Limits for Emerging Monte Carlo Samplers

PDMP
Poster
News
Author

Hirofumi Shiba

Published

6/28/2026

概要

A class of Monte Carlo methods with a qualitatively different dynamical nature from traditional MCMC methods is emerging within the framework of piecewise deterministic Markov processes (PDMPs). Its design space is begin actively explored, and our work aims to provide theoretical grounding for comparing their performance. We propose to measure their performance by examing their log-likelihood processes, and prove a theoretically rigorous ordering of asymptotic efficiency between algorithms, through scaling analysis and weak convergence results. Specifically, within this framework, Forward Event-Chain Monte Carlo (FECMC) is asymptotically more efficient than Bouncy Particle Sampler (BPS), regardless of the choice of hyperparameters. Moreover, we show that FECMC attains its best efficiency when this hyperparameter is set to zero, showing its design choices successfully lead to a robust exploration of the target distribution.

ISBA World Meeting 2026

Date Location
Jun. 28-Jul. 2, 2026 WINC AICHI (Aichi Industry & Labor Center)