Back to search

Boston University

Foundations of uncertainty quantification with subsampling Markov chain Monte Carlo algorithms

Abstract

dc:description.abstract

Bayesian inference offers a principled framework for uncertainty quantification, but exact posterior computation is rarely tractable at modern scales. Stochastic gradient Markov chain Monte Carlo (SGMCMC) methods address this challenge by replacing full-data gradients with minibatch estimates. The central questions then become how to quantify the uncertainty in the samples SGMCMC produces and how to tune its hyperparameters. A natural framework for both is scaling-limit theory, which relates the discrete-time iterations to continuous-time diffusion processes whose stationary covariance and mixing directly translate into uncertainty estimates and tuning prescriptions. Existing scaling-limit results, however, are largely asymptotic; they leave open how accurately such approximations describe finite-step behavior, and how they extend to models with latent variables and to settings in which the subsampled gradient is biased, such as Gaussian process regression. This disserta- tion develops a theoretical foundation for uncertainty quantification and tuning of subsampling MCMC, organized around three contributions. The first contribution establishes the first quantitative, non-asymptotic error bounds between SGMCMC iterates and their Ornstein–Uhlenbeck scaling limits. Using a functional version of Stein’s method of exchangeable pairs, the bounds are explicit in the step size, batch size, inverse temperature, and number of iterations, and yield convergence rates in several process-level metrics. The second contribution develops a joint scaling-limit theory for SGMCMC that combines stochastic gradient updates of global parameters with Gibbs updates of latent variables. The limit is a jump–diffusion: the global parameters converge to a diffusion while the latent variables converge to Poisson jump processes. The analysis isolates how latent-variable randomness enters the effective diffusion matrix and yields concrete preconditioning and tuning prescriptions. The third contribution analyzes minibatch SGD for Gaussian process regression, where the kernel-coupled likelihood makes the subsampled gradient biased and stan- dard SGD theory does not apply. By exploiting the kernel spectrum, we obtain sharp coordinate-wise rates for both hyperparameters under polynomial and exponential Mercer decay, improving on prior worst-case bounds. Together, the three contribu- tions deliver a non-asymptotic foundation for SGMCMC at scale.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Mathematics & Statistics
Grantor
Boston University
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Wang, Xiaoyu
Advisor dc:contributor.advisor
  • Huggins, Jonathan

Subjects

dc:subject × 1

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/53663

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Wang, Xiaoyu. Foundations of uncertainty quantification with subsampling Markov chain Monte Carlo algorithms. doctoral thesis, Boston University, 2026. https://hdl.handle.net/2144/53663