On the Performance of Orthogonal Source Separation Algorithms

Source separation consists in recovering a set of n independent signals from m n observed instantaneous mixtures of these signals, possibly corrupted by additive noise. Many source separation algorithms use second order information in a whitening operation which reduces the non trivial part of the separation to determining a unitary matrix. Most of them further show a kind of invariance property which can be exploited to predict some general results about their performance. Our rst contribution is to exhibit a lower bound to the performance in terms of accuracy of the separation. This bound is independent of the algorithm and, in the i.i.d. case, of the distribution of the source signals. Second, we show that the performance of invariant algorithms depends on the mixing matrix and on the noise level in a speciic way. A consequence is that at low noise levels, the performance does not depend on the mixture but only on the distribution of the sources, via a function which is characteristic of the given source separation algorithm.