Performances en séparation de sources

Nous donnons trois r esultats analytiques sur les performances asymptotiques de m ethodes de s eparation de sources dites`` a blanchiment'. i) Le crit ere de Comon-Lacoume et celui dèdiagonalisation conjointe' donnent les m^ emes performances. ii) Celles-ci sont donn ees explicitement a bruit nul. Elles ap-para^ ssent alors ne d ependre que de la distribution des sources et pas de la matrice de m elange. iii) Il existe a tout niveau de bruit une borne simple aux performances des m ethodes a blanchiment. Abstract We give three results on the asymptotic performance of source separation methods which are based on data prewhitening. i) The Comon-Lacoume criterion performs asymptotically as well as thèjoint diagonalization' criterion. ii) Explicit expressions for the performance in the noiseless case are derived, showing the dependence on the distribution of the sources. In the noiseless case, it is also shown that the performance is independent of the mixing matrix. iii) A simple universal noise-dependent bound exists on the performance of any separation technique relying on a prewhitening.