Bernstein - von Mises theorems for statistical inverse problems I: Schrödinger equation
For statisticians and applied mathematicians, this provides theoretical justification for Bayesian methods in PDE-constrained inverse problems, though it is an incremental extension of known theory to a specific PDE model.
The paper proves a Bernstein-von Mises theorem for a nonparametric Bayesian inverse problem of recovering a potential in a Schrödinger equation from noisy data, showing that the posterior distribution is asymptotically Gaussian with minimal covariance, enabling optimal frequentist inference.
The inverse problem of determining the unknown potential $f>0$ in the partial differential equation $$\fracΔ{2} u - fu =0 \text{ on } \mathcal O ~~\text{s.t. } u = g \text { on } \partial \mathcal O,$$ where $\mathcal O$ is a bounded $C^\infty$-domain in $\mathbb R^d$ and $g>0$ is a given function prescribing boundary values, is considered. The data consist of the solution $u$ corrupted by additive Gaussian noise. A nonparametric Bayesian prior for the function $f$ is devised and a Bernstein - von Mises theorem is proved which entails that the posterior distribution given the observations is approximated in a suitable function space by an infinite-dimensional Gaussian measure that has a `minimal' covariance structure in an information-theoretic sense. As a consequence the posterior distribution performs valid and optimal frequentist statistical inference on $f$ in the small noise limit.