English

Faster Sampling from Log-Concave Densities over Polytopes via Efficient Linear Solvers

Data Structures and Algorithms 2024-09-09 v1 Machine Learning Machine Learning

Abstract

We consider the problem of sampling from a log-concave distribution π(θ)ef(θ)\pi(\theta) \propto e^{-f(\theta)} constrained to a polytope K:={θRd:Aθb}K:=\{\theta \in \mathbb{R}^d: A\theta \leq b\}, where ARm×dA\in \mathbb{R}^{m\times d} and bRmb \in \mathbb{R}^m.The fastest-known algorithm \cite{mangoubi2022faster} for the setting when ff is O(1)O(1)-Lipschitz or O(1)O(1)-smooth runs in roughly O(md×mdω1)O(md \times md^{\omega -1}) arithmetic operations, where the mdω1md^{\omega -1} term arises because each Markov chain step requires computing a matrix inversion and determinant (here ω2.37\omega \approx 2.37 is the matrix multiplication constant). We present a nearly-optimal implementation of this Markov chain with per-step complexity which is roughly the number of non-zero entries of AA while the number of Markov chain steps remains the same. The key technical ingredients are 1) to show that the matrices that arise in this Dikin walk change slowly, 2) to deploy efficient linear solvers that can leverage this slow change to speed up matrix inversion by using information computed in previous steps, and 3) to speed up the computation of the determinantal term in the Metropolis filter step via a randomized Taylor series-based estimator.

Keywords

Cite

@article{arxiv.2409.04320,
  title  = {Faster Sampling from Log-Concave Densities over Polytopes via Efficient Linear Solvers},
  author = {Oren Mangoubi and Nisheeth K. Vishnoi},
  journal= {arXiv preprint arXiv:2409.04320},
  year   = {2024}
}

Comments

The conference version of this paper appears in ICLR 2024

R2 v1 2026-06-28T18:36:33.731Z