Selection of Correlation Matrices for Second-Order-Statistics-Based Blind Source Separation

The aim of blind source separation is to recover mutually independent unknown source signals from observations obtained through an unknown linear mixture system. A simultaneous diagonalization of correlation matrices (second-order statistics) of the observations is a possible resolution for the case when the unknown source signals are non-stationary. In general, unknown source signals are not strictly uncorrelated; this may cause a degradation in the separation performance. In this study, we propose a method for selecting a combination of correlation matrices that yields a better separation performance, and verify the efficacy of the proposed method by computer simulations.