Proposals which speed-up function-space MCMC

TL;DR

This paper proposes new methods to accelerate function-space MCMC algorithms by leveraging spectral properties and Hessian information to improve mixing times in Bayesian inverse problems.

Contribution

It introduces two novel rescaling techniques, characteristic function truncation and Hessian-based interpolation, to enhance the efficiency of MCMC in high-dimensional Bayesian inverse problems.

Findings

Improved mixing times demonstrated through theoretical analysis.

Spectral rescaling methods outperform standard MCMC schemes.

Enhanced convergence in high-dimensional inverse problems.

Abstract

Inverse problems lend themselves naturally to a Bayesian formulation, in which the quantity of interest is a posterior distribution of state and/or parameters given some uncertain observations. For the common case in which the forward operator is smoothing, then the inverse problem is ill-posed. Well-posedness is imposed via regularisation in the form of a prior, which is often Gaussian. Under quite general conditions, it can be shown that the posterior is absolutely continuous with respect to the prior and it may be well-defined on function space in terms of its density with respect to the prior. In this case, by constructing a proposal for which the prior is invariant, one can define Metropolis-Hastings schemes for MCMC which are well-defined on function space, and hence do not degenerate as the dimension of the underlying quantity of interest increases to infinity, e.g. under mesh refinement when approximating PDE in finite dimensions. However, in practice, despite the attractive theoretical properties of the currently available schemes, they may still suffer from long correlation times, particularly if the data is very informative about some of the unknown parameters. In fact, in this case it may be the directions of the posterior which coincide with the (already known) prior which decorrelate the slowest. The information incorporated into the posterior through the data is often contained within some finite-dimensional subspace, in an appropriate basis, perhaps even one defined by eigenfunctions of the prior. We aim to exploit this fact and improve the mixing time of function-space MCMC by careful rescaling of the proposal. To this end, we introduce two new basic methods of increasing complexity, involving (i) characteristic function truncation of high frequencies and (ii) hessian information to interpolate between low and high frequencies.