Novel blind source separation algorithms using cumulants

This paper investigates new algorithms for blind source separation that use cumulants instead of nonlinearities matched to the probability distribution of the sources. It is demonstrated that separation is a saddle point of a cumulant-based entropy cost function. To determine this point we propose two quasi-Newton algorithms whose convergence is isotropic and does not depend on the sources distribution. Moreover, convergence properties remain the same when there is Gaussian noise in the mixture