Advanced MCMC methods for sampling on diffusion pathspace (Q1939346)

The authors discuss advanced Markov chain Monte Carlo (MCMC) methods for the sampling of distributions defined as changes of measures from Gaussian laws on Hilbert spaces. The main motivation and application of this work is sampling from the laws of solutions of stochastic differential equations (as defined on the Hilbert space \(L^2([0,l],\mathbb{R})\)). In general, the attribute ``advanced'' of the method follows from the fact that its mixing times do not deteriorate when the dimension of the state space increases, i.e., when the finite-dimensional projections are refined to better approximate an infinite-dimensional Hilbert space. The main focus of the paper lies on the advanced hybrid Monte Carlo (HMC) method which is thoroughly defined and discussed. However, not all propositions concerning the algorithm are proven. Besides HMC also the random walk Metropolis-Hastings (RWM) and Metropolis-adjusted Langevin algorithm (MALA) are briefly discussed. The authors perform an analytical study of the HMC method applied to the linear case, i.e., sampling an Ornstein-Uhlenbeck bridge. Therein, they find that HMC performs better than the other two methods, i.e., it gains in orders of complexity. In the following sections, the framework is extended to accommodate a large class of applications and more general cases. Discussed examples are diffusions observed at discrete time points, diffusions observed with errors, stochastic volatility models and latent diffusion survival models. The paper is concluded by numerical applications which aim to assess and compare efficiency aspects of the different algorithms. As for the theoretical analysis in the linear case, the superiority of the HMC method is found.