Generalized Proximal Langevin Algorithms via Backward Differentiation Formula

Read full text Summarise with VIRAT
Authors: Fan Jia, Yuhao Huang, Bao Wang
Cited by: 0
References: 38

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Sampling from complex probability distributions is a cornerstone of statistical inference and machine learning. Classical algorithms such as the unadjusted Langevin algorithm (ULA) and the Metropolis-adjusted Langevin algorithm (MALA) provide theoretical mixing guarantees under certain conditions. However, their practical performance often deteriorates in high-dimensional or ill-conditioned settings due to instability and slow convergence. Recent advances—including higher-order discretization (e.g., Runge–Kutta (RK) methods) and the proximal Langevin algorithm (PLA)—have demonstrated potential in mitigating these challenges, offering improved stability and efficiency. Proximal algorithms, in particular, reformulate computationally challenging implicit updates as optimization problems, thereby enabling the leveraging of advances in gradient-based optimization algorithms, e.g., momentum methods.In this paper, we propose a class of higher-order proximal algorithms, referred to as Prox-BDF, which are derived from the proximal formulation of the backward differentiation formula (BDF). These algorithms synergistically combine the accuracy benefits of higher-order temporal discretization with the stability advantages of implicit proximal schemes. Notably, Prox-BDF algorithms exhibit superior stability when applied to stiff or ill-conditioned potentials and achieve higher temporal accuracy than explicit methods such as ULA, existing proximal sampling approaches, and RK-based schemes. Through numerical experiments, we show that Prox-BDF algorithms substantially reduce discretization bias, enable using larger step sizes, and outperform the baseline Langevin methods in terms of both accuracy and robustness.

Fields of Study & Subjects

Markov Chains and Monte Carlo Methods Numerical Solvers for Ordinary Differential Equations Stochastic Gradient Optimization Techniques

Publication details

Journal / source
Communications on Pure &amp Applied Analysis
Volume
33
Issue
0
Pages
116–137
Publication date
1 January 2026
Language
EN