On The Use of Non-orthogonal Approximate Joint Diagonalization Algorithms for Blind Source Separation in Presence of Additive Noise

Abstract—We present in this paper a non-orthogonal algo-rithm for the approximate joint diagonalization of a set of matrices. It is an iterative algorithm, using relaxation technique applied on the rows of the diagonalizer. The performances of our algorithm are compared with usual standard algorithms using blind sources separation simulations results. We show that the improvement in estimating the separating matrix can be wreaked when the level noise in the mixture is significant, the length of observed sequences is sufficiently large and when the mixing matrix is not an orthogonal matrix or just about. I