Robust Principal Components and Dispersion Matrices via Projection Pursuit

摘要

This paper discusses a new kind of robust procedure for estimating covariance/correlation matrices and their principal components. Robust eigenvectors and eigenvalues of a covariance matrix are obtained by the projection pursuit method (PP) with robust variance as a projection index. Monte Carlo simulation results show that the best of the three projection pursuit type procedures introduced in this study compares favorably with approaches based on M-estimators of covariance: the estimate obtained by the new procedure has about the same bias and variance as the best M-estimators, and a somewhat better breakdown point. (Author)