Blind source separation of short duration finite alphabet signals

This paper presents a blind source separation technique for short duration, finite alphabet signals. The approach is derived by posing the BSS problem on the Grassman manifold and then successively fixing each column of the estimated source matrix. Two approaches have been described: one optimizes over the unfixed columns of S and the other optimizes over the mixing matrix, A. This approach is a compromise between the computationally intensive exhaustive search and a simple rounding approach applied to an unconstrained method. Simulation results show that this method is significantly superior to the modified SOBI algorithm, which uses a rounding approach in conjunction with second order statistical assumptions.