Skip to Main content Skip to Navigation
Theses

High dimensional Markov chain Monte Carlo methods : theory, methods and applications

Abstract : The subject of this thesis is the analysis of Markov Chain Monte Carlo (MCMC) methods and the development of new methodologies to sample from a high dimensional distribution. Our work is divided into three main topics. The first problem addressed in this manuscript is the convergence of Markov chains in Wasserstein distance. Geometric and sub-geometric convergence with explicit constants, are derived under appropriate conditions. These results are then applied to thestudy of MCMC algorithms. The first analyzed algorithm is an alternative scheme to the Metropolis Adjusted Langevin algorithm for which explicit geometric convergence bounds are established. The second method is the pre-Conditioned Crank-Nicolson algorithm. It is shown that under mild assumption, the Markov chain associated with thisalgorithm is sub-geometrically ergodic in an appropriated Wasserstein distance. The second topic of this thesis is the study of the Unadjusted Langevin algorithm (ULA). We are first interested in explicit convergence bounds in total variation under different kinds of assumption on the potential associated with the target distribution. In particular, we pay attention to the dependence of the algorithm on the dimension of the state space. The case of fixed step sizes as well as the case of nonincreasing sequences of step sizes are dealt with. When the target density is strongly log-concave, explicit bounds in Wasserstein distance are established. These results are then used to derived new bounds in the total variation distance which improve the one previously derived under weaker conditions on the target density.The last part tackles new optimal scaling results for Metropolis-Hastings type algorithms. First, we extend the pioneer result on the optimal scaling of the random walk Metropolis algorithm to target densities which are differentiable in Lp mean for p ≥ 2. Then, we derive new Metropolis-Hastings type algorithms which have a better optimal scaling compared the MALA algorithm. Finally, the stability and the convergence in total variation of these new algorithms are studied.
Complete list of metadatas

Cited literature [142 references]  Display  Hide  Download

https://pastel.archives-ouvertes.fr/tel-01529750
Contributor : Abes Star :  Contact
Submitted on : Wednesday, May 31, 2017 - 12:42:07 PM
Last modification on : Friday, July 31, 2020 - 10:44:09 AM
Long-term archiving on: : Wednesday, September 6, 2017 - 3:52:38 PM

File

73643DURMUS2016archivage.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01529750, version 1

Citation

Alain Durmus. High dimensional Markov chain Monte Carlo methods : theory, methods and applications. Probability [math.PR]. Université Paris-Saclay, 2016. English. ⟨NNT : 2016SACLT001⟩. ⟨tel-01529750⟩

Share

Metrics

Record views

816

Files downloads

343