APPROXIMATE JOINT DIAGONALIZATION USING NON­ORTHOGONAL MATRICES

Arie Yeredor

Several Blind Source Separation algorithms require the approximate joint diagonalization of a set of matrices. Most existing algorithms for approximate joint diag­ onalization are restricted to finding an orthogonal di­ agonalizing matrix. It has therefore become common practice to first find the non­orthogonal factor of the diagonalizing matrix by pre­whitening the data, and then proceed to find the orthogonal factor using ex­ isting algorithms. However, such an approach favors exact diagonalization of one matrix (usually the data correlation matrix), possibly at the expense of poor di­ agonalization of the other matrices. In this paper we propose a joint diagonalization algorithm which uses general (not necessarily orthogonal) matrices. The use of this algorithm enables to eliminate the traditional ''hard whitening'' phase, which is known to limit the performance, especially under noisy conditions. We demonstrate the improved performance via simulations results.