An iterative-recursive SOS-based method for separation of Post-Nonlinear Mixtures

Independent Component Analysis (ICA) methods are widely used in the linear Blind Source Separation (BSS) problem. Nevertheless, for some practical cases, the linear assumption is not valid, requiring nonlinear mixing models. The Post-Nonlinear is one of the few nonlinear models in which ICA is able to perform source separation. In previous works, an iterative SOS-based algorithm was proposed, combining elements of AMUSE and SOBI. However, instantaneous estimations of the correlation matrix could lead to a loss of performance. In that sense, in this work, we modify the previous algorithm to use an iterative-recursive estimation of the correlation matrices. Due to the SOS-based approach, sources should present a temporal dependency and certain constraints are required on the separation structure. Results show a good performance of the proposed algorithm.