All-at-once formulation meets the Bayesian approach: A study of two prototypical linear inverse problems

TL;DR

This paper introduces a novel all-at-once Bayesian formulation for linear inverse problems, allowing for joint prior modeling of parameters and states, and analyzing its effectiveness through two classical examples.

Contribution

It combines all-at-once and Bayesian approaches, enabling prior for both parameters and states, and accounts for model errors, improving inverse problem analysis.

Findings

Effective reconstruction of unknown parameters and states.

Analysis of ill-posedness via singular values.

Numerical validation on Poisson and heat equations.

Abstract

In this work, the Bayesian approach to inverse problems is formulated in an all-at-once setting. The advantages of the all-at-once formulation are known to include the avoidance of a parameter-to-state map as well as numerical improvements, especially when considering nonlinear problems. In the Bayesian approach, prior knowledge is taken into account with the help of a prior distribution. In addition, the error in the observation equation is formulated by means of a distribution. This method naturally results in a whole posterior distribution for the unknown target, not just point estimates. This allows for further statistical analysis including the computation of credible intervals. We combine the Bayesian setting with the all-at-once formulation, resulting in a novel approach for investigating inverse problems. With this combination we are able to chose a prior not only for the parameter, but also for the state variable, which directly influences the parameter. Furthermore, errors not only in the observation equation, but additionally, in the model can be taken into account. %The aim of this approach is not only to accomplish reasonable reconstructions of the unknown parameter but also to maximize the information gained from measurements through combining it with prior knowledge, obtained either from certain expertise or former investigation in the model. We analyze this approach with the help of two linear standard examples, namely the inverse source problem for the Poisson equation and the backwards heat equation, i.e. a stationary and a time dependent problem. Appropriate function spaces and derivation of adjoint operators are investigated. To assess the degree of ill-posedness, we analyze the singular values of the corresponding all-at-once forward operators. %as well as the convergence of the method. Finally, joint priors are designed and numerically tested.