An ensemble Kalman approach to randomized maximum likelihood estimation

TL;DR

This paper introduces an ensemble Kalman randomized maximum likelihood method for Bayesian inverse problems, demonstrating exponential convergence, posterior sampling capabilities, and computational efficiency via model reduction strategies.

Contribution

It presents a novel derivative-free ensemble Kalman approach for randomized maximum likelihood estimation, with theoretical convergence and practical model reduction techniques.

Findings

Ensemble members converge exponentially fast to estimators.

The method produces samples from the Bayesian posterior.

Model reduction accelerates computations significantly.

Abstract

This work proposes ensemble Kalman randomized maximum likelihood estimation, a new derivative-free method for performing randomized maximum likelihood estimation, which is a method that can be used to generate approximate samples from posterior distributions in Bayesian inverse problems. The new method has connections to ensemble Kalman inversion and works by evolving an ensemble so that each ensemble member solves an instance of a randomly perturbed optimization problem. Linear analysis demonstrates that ensemble members converge exponentially fast to randomized maximum likelihood estimators and, furthermore, that the new method produces samples from the Bayesian posterior when applied to a suitably regularized optimization problem. The method requires that the forward operator, relating the unknown parameter to the data, be evaluated once per iteration per ensemble member, which can be prohibitively expensive when the forward model requires the evolution of a high-dimensional dynamical system. We propose a strategy for making the proposed method tractable in this setting based on a balanced truncation model reduction method tailored to the Bayesian smoothing problem. Theoretical results show near-optimality of this model reduction approach via convergence to an optimal approximation of the posterior covariance as a low-rank update to the prior covariance. Numerical experiments verify theoretical results and illustrate computational acceleration through model reduction.