ICA of correlated sources mismodeled as uncorrelated: performance analysis

In this contribution we present closed-form performance analysis for the problem of blind separation of correlated sources when the separation algorithm is ignorant of the source model, and assumes only scalar sources. Our data model is of one scalar source and one multidimensional source. We obtain a closed-form analytical expression for applying joint diagonalization on correlated sources. The obtained expression is different from the optimal expression (in the minimal mean square error (MMSE) sense), obtained by joint block diagonalization. Numerical evaluations quantify the loss of performance due to mismodeling.