Separation Of Sources Using Higher-Order Cumulants

The problem is to recover stochastic processes from an unknown stationary linear transform. Our contribution is two-fold. First we focus on instantaneous mixtures: observation e(t) is assumed to write as a regular linear transform of the sources, x(t), as e(t)=Box(t). The only assumption requested is that the sources xi(t) are mutually independent, and no additional knowledge upon their statistics is necessary provided they are not normal. Extensions to convolutional mixing are then pointed out, namely cases where e(t)=A(t)*x(t) where A(t) has a rational transfer function. Sensitive improvements to the algorithm of Giannakis et al for MA identification are included. Multivariate ARMA identification can be split into three successive estimation problems: AR identification, monic MA identification, and estimation of Bo in last position.