A likelihood framework for nonlinear signal processing with finite normal mixtures

We introduce a likelihood framework for nonlinear signal processing using partial likelihood and use the result to derive the information geometric em algorithm for distribution learning through information-theoretic projections. We demonstrate the superior convergence of the em algorithm as compared to least relative entropy (LRE) algorithm by simulations. The performance of finite normal mixtures (FNM) based equalizers with different number of mixtures and different dimension observation vectors is also discussed.