A Homotopy Bayesian Approach for Inverse Problems

Abstract

In solving Bayesian inverse problems, it is often desirable to use a common density parameterization to characterize the prior and posterior. Typically, we seek an optimal approximation of the true posterior from the same distribution family as the prior. As one of the most important classes of distributions in statistics, the exponential family is considered as the parameterization. The optimal parameter values for representing the approximated posterior are achieved by minimizing the deviation between the parameterized density and a homotopy that deforms the prior density into the posterior density. Instead of the usual moment parameters, we introduce a new parameter system in the process, by which we reduce the dimension of the parameter and therefore the computational cost. The parameters are ruled by a system of explicit first-order differential equations. Solving this system over finite “time” interval yields the desired optimal density parameters. This method is proven to be effective using some numerical examples.