Blind Source Separation

What is Blind Source Separation ?

Let S be a n rows, p columns matrix that describes a set of n functions defined for p values of the same variable. A set of m linear combinations of these n functions (m greater or equal to n) are described by a m rows, p columns matrix X so that X = AS where A contains the mixing coefficients.

Blind Source Separation (BSS) consists in finding S, the sources, and A, the mixing matrix, with only X as data !!

Obviously, if S is a solution, the matrix that is obtained by scaling its rows by any factor and/or by permuting them is also a solution. Solving a BSS problem requires some prior knowledge about the sources. The existing BSS algorithms mainly differ by the nature of this prior. I have participated to the elaboration of two such algorithms, one named f-SOBI, derived from the Second-Order Blind Identification method, and a really original one, named LP-BSS

Jean-Marc Nuzillard

March 25th 2005