Sampling high dimensional measures with computational guarantees beyond convexity and smoothness

Ziel

This research proposal aims to address key computational challenges in sampling high dimensional measures, with a focus on problems from statistical inference that lack the usual convexity and smoothness assumptions. Many statistical models present complex computational problems that require numerical integration of high dimensional measures which can be as challenging as those encountered in computer science and statistical physics, yet they remain poorly understood. This project develops MCMC algorithms and a quantitative mixing time theory in two main thrusts. First, we study sampling for polytope-constrained log-concave measures, typically arising in probit or censored regression, where the non-smoothness at boundary causes issue for Langevin-type algorithms. We design interior-point-method-inspired barrier functions tailored for sampling, refine Riemannian HMC analysis for implementable integrators, and develop instance-adaptive weighted Dikin/Vaidya/John-type walks to surpass worst-case analysis. Second, we take sparse PCA and Bayesian variable selection as canonical problems to study MCMC without log-concavity and with mixed-discrete-continuous parameters. Tuning signal-to-noise ratio (SNR) in these problems is known to make the computational problem from easy to hard. We aim to determine the optimal SNR threshold for MCMC to succeed efficiently and to resolve the open problem on whether MCMC are among the best algorithms in statistical problems with statistical-to-computational gaps, by leveraging recent progress on expander graphs, spectral independence, local-to-global approaches and stochastic localization. We further pursue a ``discrete KLS'' result on the Boolean hypercube, aiming for a new functional inequality framework with fewer checks than that in the state-of-the-art spectral independence framework.

Wissenschaftliches Gebiet (EuroSciVoc)

Dieses Projekt wurde noch nicht bei EuroSciVoc klassifiziert.
Schlagen Sie die Wissenschaftsbereiche vor, die Ihrer Einschätzung nach besonders relevant sind, und helfen Sie uns, unseren Klassifizierungsdienst zu verbessern.