Markovian source separation in post-nonlinear mixtures
Résumé
In linear mixtures, priors, like temporal coloration of the sources, can be used for designing simpler and better algorithms. Especially, modeling sources by Markov models is very efficient, and Markov source separation can be achieved by minimizing the conditional mutual information. This model allows to separate temporally colored Gaussian sources. In this paper, we extend this result for post-nonlinear mixtures ({PNL}) , and show that algorithms based on a Markov model of colored sources leads to better separation results than without prior, {\em i.e.} assuming iid sources. The paper contains theoretical developments, and experiments with auto-regressive (AR) source mixtures. PNL algorithms for Markov sources point out a performance improvement of about $7$dB with respect to PNL algorithms for iid sources.