SOURCE SEPARATION USING A CRITERION BASED ON 2ND-ORDER STATISTICS

It is often assumed that blind separation of dynamically mixed sources cannot be done with second-order statistics. In this paper, it is sho wn that separation of dynamically mixed sources indeed can be performe d using second-order statistics only. A criterion based on second-orde r statistics for the purpose of separating cross,vise mixtures is stat ed. The criterion is used in order to derive a gradient-based separati on algorithm, as well as a Newton-type separation algorithm. The uniqu eness of the solution representing separation is also investigated. Th is reveals that 1) the channel system is parameter identifiable under weak conditions, and 2) if the sources have the same color, there exis ts at most two solutions. The local convergence behavior of the propos ed algorithm is studied and reveals a sufficient condition for local c onvergence. Futhermore, the estimates of the channel system are shown to be consistent or to locally minimize the criterion.