Performance of blind discrete source separation

The aim of this paper is two-fold. In a first part, we investigate the ultimate performances of source separation in the case of BPSK and 4-PSK sources. These ultimate bounds are computed by looking first for the most favourable mixing matrices, for various SNRs, number of sources and number of sensors, and then by calculating the associated error probabilities. In a second part, we compare the behaviour of two 4-QAM source separation algorithms. One is based on the Constant Modulus (CM) property of these sources, and the other is based on the Constance of their fourth power. The goal here is to see what kind of improvement brings the knowledge of the source distribution.