New natural gradient algorithm for cyclostationary sources

A new natural gradient type algorithm (NGA) for the separation of cyclostationary sources is introduced. Based on the interpretation of blind source separation (BSS) as a two-stage process, including prewhitening and rotation, the cyclostationary NGA (CSNGA) algorithm is constructed such that it also ensures that the recovered sources are decorrelated in the cyclostationary sense. The method is generalised to the case of complex valued source signals, and modified so that adequate algorithm performance is attained even when only one source cycle frequency is known. The properties of the new algorithm are investigated when additive white Gaussian noise is present, and it is found that, in general, the CSNGA approach improves the convergence properties of the natural gradient algorithm. Computer simulations support the validity of the approach.