Dimension-Free Multimodal Sampling via Preconditioned Annealed Langevin Dynamics
Researchers have developed a new sampling algorithm called Preconditioned Annealed Langevin Dynamics (PALD) designed to improve exploration across modes in multimodal targets. The algorithm's stability across dimensions is analyzed, providing conditions under which it can achieve a prescribed accuracy within a dimension-uniform time horizon. The study also demonstrates that PALD can maintain dimension-uniform control even with imperfect initialization and approximate scores, and can prevent error accumulation across coordinates when using a misspecified mixture score model. AI
IMPACT Introduces a novel sampling technique with theoretical guarantees for multimodal targets, potentially improving generative model training and data analysis.