Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2506.20185 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- This work considers the framework of Markov chain importance sampling~(MCIS), in which one employs a Markov chain Monte Carlo~(MCMC) scheme to sample particles approaching the optimal distribution for importance sampling, prior to estimating the quantity of interest through importance sampling. In rare event estimation, the optimal distribution admits a non-differentiable log-density, thus gradient-based MCMC can only target a smooth approximation of the optimal density. We propose a new gradient-based MCIS scheme for rare event estimation, called affine invariant interacting Langevin dynamics for importance sampling~(ALDI-IS), in which the affine invariant interacting Langevin dynamics~(ALDI) is used to sample particles according to the smoothed zero-variance density. We establish a non-asymptotic error bound when importance sampling is used in conjunction with samples independently and identically distributed according to the smoothed optiaml density to estimate a rare event probability, and an error bound on the sampling bias when a simplified version of ALDI, the unadjusted Langevin algorithm, is used to sample from the smoothed optimal density. We show that the smoothing parameter of the optimal density has a strong influence and exhibits a trade-off between a low importance sampling error and the ease of sampling using ALDI. We perform a numerical study of ALDI-IS and illustrate this trade-off phenomenon on standard rare event estimation test cases.