Accelerating Non-Smooth and Heavy-Tailed Sampling

Name: Accelerating Non-Smooth and Heavy-Tailed Sampling
Uploaded: 2026-10-14T16:41:18.021821+00:00
Duration: 1022 s
Description: The work introduces non-reversible anchored Langevin dynamics (NALD) for sampling heavy-tailed distributions in Euclidean space and non-reversible reflected anchored Langevin dynamics (NRALD) for constrained domains. These methods accelerate convergence compared to reversible versions by introducing a circulation drift, leading to faster mixing and reduced asymptotic variance.

The work introduces non-reversible anchored Langevin dynamics (NALD) for sampling heavy-tailed distributions in Euclidean space and non-reversible reflected anchored Langevin dynamics (NRALD) for constrained domains. These methods accelerate convergence compared to reversible versions by introducing a circulation drift, leading to faster mixing and reduced asymptotic variance.

Open the full episode