Scaling limits in computational Bayesian inversion

Abstract

Computational Bayesian inversion of operator equations with distributed uncertain input parameters is based on an infinite-dimensional version of Bayes’ formula established in M. Dashti and A.M. Stuart [Handbook of Uncertainty Quantification, edited by R. Ghanem, D. Higdon and H. Owhadi. Springer (2015).] and its numerical realization in C. Schillings and Ch. Schwab [Inverse Problems 29 (2013) 065011; Inverse Problems 30 (2014) 065007.] Based on the sparsity of the posterior density shown in C. Schillings and Ch. Schwab [Inverse Problems 29 (2013) 065011; Inverse Problems 30 (2014) 065007.]; C. Schwab and A.M. Stuart [Inverse Problems 28 (2012) 045003.], dimension-adaptive Smolyak quadratures can afford higher convergence rates than MCMC in terms of the number M of solutions of the forward (parametric operator) equation in C. Schillings and Ch. Schwab [Inverse Problems 29 (2013) 065011; Inverse Problems 30 (2014) 065007.]. The error bounds and convergence rates obtained in C. Schillings and Ch. Schwab [Inverse Problems 29 (2013) 065011; Inverse Problems 30 (2014) 065007.] are independent of the parameter dimension (in particular free from the curse of dimensionality) but depend on the (co)variance Γ> 0 of the additive, Gaussian observation noise as exp(bΓ-1) for some constant b> 0. It is proved that the Bayesian estimates admit asymptotic expansions as Γ ↓ 0. Sufficient (nondegeneracy) conditions for the existence of finite limits as Γ ↓ 0 are presented. For Gaussian priors, these limits are shown to be related to MAP estimators obtained from Tikhonov regularized least-squares functionals. Quasi-Newton (QN) methods with symmetric rank-1 updates are shown to identify the concentration points in a non-intrusive way, and to obtain second order information of the posterior density at these points. Based on the theory, two novel computational Bayesian estimation algorithms for Bayesian estimation at small observation noise covariance Γ> 0 with performance independent of Γ ↓ 0 are proposed: first, dimension-adaptive Smolyak quadrature from C. Schillings and Ch. Schwab [Inverse Problems 29 (2013) 065011; Inverse Problems 30 (2014) 065007.] combined with a reparametrization of the parametric Bayesian posterior density near the MAP point (assumed unique) and, second, generalized Richardson extrapolation to the limit of vanishing observation noise variance. Numerical experiments are presented which confirm Γ-independent convergence of the curvature-rescaled, adaptive Smolyak algorithm. Dimension truncation of the posterior density is justified by a general compactness result for the posterior’s Hessian at the MAP point.