Adaptive blind source separation by second order statistics and natural gradient

Separation of sources that are mixed by an unknown (hence, "blind") mixing matrix is an important task for a wide range of applications. This paper presents an adaptive blind source separation method using second order statistics (SOS) and natural gradient. The SOS of observed data is shown to be sufficient for separating mutually uncorrelated sources provided that the temporal coherences of all sources are linearly independent of each other. By applying the natural gradient, new adaptive algorithms are derived that have a number of attractive properties such as invariance of asymptotical performance (with respect to the mixing matrix) and guaranteed local stability. Simulations suggest that the new algorithms are highly efficient and outperform some of the best existing ones.