Exponentially embedded families-new approaches to model order estimation

Abstract

The use of exponential embedding of two or more probability density functions (pdfs) is introduced. Termed the exponentially embedded family (EEF) of pdfs, its properties are first examined and then it is applied to the problem of model order estimation. The proposed estimator is compared with the minimum description length (MDL) and is found to be superior for cases of practical interest. Also, we point out there is a relationship between the embedded family model order estimator and the generalized likelihood ratio test (GLRT). The embedded family estimator appears to extend the GLRT to the case of multiple alternative hypotheses that have differing numbers of unknown parameters.