Proper prior marginalization of the conditional ML model for combined model selection/source localization

We present a Bayesian evidence technique for the parameter estimation/model selection problem within the conditional maximum likelihood (CML) framework. The CML is chosen because of its flexibility: it allows for a wide range of source amplitude models (e.g., no unreasonable or restrictive assumptions, such as Gaussian signals are necessary). In contrast to other CML studies, we eliminate the large number of unknown amplitude parameters by marginalization with a proper (normalizable), yet every broad prior. The resulting marginal is used to derive a new model selection/parameter estimation procedure, based on the Bayesian evidence of each considered model, given the observed data. Monte Carlo simulations for a scenario consisting of two narrowband, far-field sources demonstrate the effectiveness of the proposed method in low SNR, small temporal/spatial sample situations