Use this URL to cite or link to this record in EThOS:
Title: Parallel Markov chain quasi-Monte Carlo methods
Author: Schwedes, Tobias
ISNI:       0000 0004 8499 6537
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Quasi-Monte Carlo (QMC) methods for estimating integrals are attractive since the resulting estimators typically converge at a faster rate than pseudo-random Monte Carlo. However, they can be difficult to set up on arbitrary posterior densities within the Bayesian framework, in particular for inverse problems. We propose a principled and efficient way of applying QMC to drive algorithms based on a general parallel Markov chain Monte Carlo (MCMC) framework, in which multiple proposals are generated per iteration. We provide numerous methodological extensions of the original algorithm, including the use of non-reversible transition kernels, adaptive proposal kernels and Rao-Blackwellisation. Further, we prove a law of large numbers and a central limit theorem, ergodicity of the proposed adaptive methods and asymptotic unbiasedness for estimates based on the Rao-Blackwellisation scheme. We consider the use of completely uniformly distributed (CUD) numbers within the previously stated algorithms, which leads to a general parallel Markov chain quasi-Monte Carlo (MCQMC) methodology. A scheme that efficiently produces CUD seeds for arbitrary problem dimensions and proposal numbers based on an already existing CUD sequence is developed. We prove ergodicity of the resulting adaptive methods and asymptotic unbiasedness for the Rao-Blackwellised estimates. For the latter, we demonstrate numerically in a number of statistical models that this approach scales close to n^{−2} as we increase parallelisation, instead of the usual n^{−1} that is typical of pseudo-random MCMC algorithms. The improved rate is proven theoretically in a special case. Simulations are performed for Bayesian linear and logistic regression, simple non-linear ODE models and a complex model for cardiac excitation. The CUD driven Rao-Blackwellisation algorithms yield multiple orders of magnitude reduction in the variance and MSE of the resulting estimates compared to their pseudo-driven counterparts and to reference algorithms such as Metropolis-Hastings.
Supervisor: Calderhead, Ben Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral