Bayesian marginal model selection for low-rank sources

Abstract

We consider a Bayesian evidence approach to model parameter estimation and order selection. We specifically consider multiple-source parameter estimation and model order selection using data from an array of general configuration. A source observation is assumed to be of time-varying orientation in a low-rank subspace of the observation space. The conditional maximum likelihood (CML) framework is assumed, where we eliminate the large number of unknown nuisance parameters (i.e., signal amplitudes/orientations and noise power) by marginalization using noninformative priors which are proper as required for model order selection. We compare this Bayesian evidence-based parameter estimator to the CML estimator and a previously proposed Bayesian estimator.