Blind identification of mixtures of quasi-stationary sources using a khatri-rao subspace approach

This paper addresses the problem of blind identification of a linear instantaneous overdetermined mixture of quasi-stationary sources, using a new formulation based on Khatri-Rao (KR) subspace. A salient feature of this formulation is that it decomposes the blind identification problem into a number of per-source, structurally less complex, blind identification problems. We tackle the per-source problems by developing a specialized alternating projections (AP) algorithm. Remarkably, we prove that AP almost surely converges to a true mixing matrix column in its first iteration, assuming an ideal model condition. Simulation results show that the proposed algorithm yields competitive complexity and performance.