A New Evolutionary Bayesian Approach Incorporating Additive Path Correction for Nonlinear Inverse Problems

TL;DR

This paper introduces a derivative-free, evolutionary Bayesian method with additive updates for nonlinear inverse problems, demonstrating improved stability and convergence over traditional approaches, especially in optical imaging applications.

Contribution

It proposes a novel, measure-change-based additive update scheme for Bayesian inverse problems, avoiding weight degeneracy and enhancing stability and convergence.

Findings

Improved reconstruction accuracy in optical tomography.

Enhanced stability against regularization noise.

Better convergence properties compared to Gauss-Newton methods.

Abstract

An evolutionary form of a generalized Bayesian update method, which is strictly derivative- free yet directed through an additive update term based purely on the statistical moments of the design variables, is proposed for nonlinear inverse problems in general and applied in particular to an optical imaging problem, the ultrasound modulated optical tomography (UMOT). The additive update term, which bypasses most pitfalls of a conventional weight- based Bayesian update, results from a change of measures aimed at driving appropriately derived observation-prediction error terms or increments of cost functionals to zero-mean Brownian martingales. This constitutes a novel characterization corresponding to the extremization of the cost functional(s), where the design unknowns are represented as diffusion processes evolving with respect to a continuously parameterized iteration variable. This leads to a recursive prediction-update algorithm to implement the search. The scheme offers freedom from sample degeneracy and the accompanying divergence of the conventional weight-based Bayesian update schemes. We obtain the order of convergence of the conditioned process and also establish that the solutions are stable against tolerable variations in the regularizing noise terms, even as the original inverse problem remains severely ill-posed. Numerical evidence on solutions to the UMOT problem also confirms substantive improvements in the reconstruction efficacy through the proposed method vis-\`a- vis a Gauss-Newton approach, especially where the regularized quasi-Newton direction has low sensitivity to variations in the design unknowns.