Robust adaptive algorithms for fast principal component analysis

Abstract

In this paper, we investigate the robust estimation of the principal eigenvectors of multivariate observations in an adaptive scheme. Indeed, this problem known as PCA (Principal Component Analysis) has already many efficient solutions in batch or non-impulsive noise environment. Our objective is to develop fast adaptive methods for the PCA when the observed data is corrupted by impulsive noise or outliers or in the case of missing data. Hence, different low complexity algorithms are introduced for the estimation and tracking of the desired principal components in the three previous scenarios. Their effectiveness is illustrated via simulation experiments and comparative study.