Blind source separation using closed-form estimators with optimal finite sample performance

We address the blind separation of two source signals from two real-valued instantaneous linear mixtures. Appropriate weighting of previous estimation expressions yields a family of closed-form estimators of the separation parameter which are consistent under rather general conditions. The weight coefficient for the asymptotically (large-sample) most efficient estimator of the family, i.e., the estimator with optimal finite-sample performance, is obtained as a function of the source statistics. Experimental results demonstrate the benefits of the optimal weighted estimator.