Performance surfaces of blind source separation algorithms

Abstract—Many source separation methods use a two-stage procedure of first equalising the second order statistics of the data by a whitening process and then seeking a separation matrix for this whitened data. The second part of this procedure is in fact a question of finding the rotation matrix that maximises (or minimises, as appropriate) a higher order statistics performance criterion. We illustrate the shape of the performance surface for some typical examples and draw conclusions about the possible efficacy of adaptive search procedures on these surfaces.