Joint blind source separation of multidimensional components: Model and algorithm

This paper deals with joint blind source separation (JBSS) of multidimensional components. JBSS extends classical BSS to simultaneously resolve several BSS problems by assuming statistical dependence between latent sources across mixtures. JBSS offers some significant advantages over BSS, such as identifying more than one Gaussian white stationary source within a mixture. Multidimensional BSS extends classical BSS to deal with a more general and more flexible model within each mixture: the sources can be partitioned into groups exhibiting dependence within a given group but independence between two different groups. Motivated by various applications, we present a model that is inspired by both extensions. We derive an algorithm that achieves asymptotically the minimal mean square error (MMSE) in the estimation of Gaussian multidimensional components. We demonstrate the superior performance of this model over a two-step approach, in which JBSS, which ignores the multidimensional structure, is followed by a clustering step.