PAC-Bayesian bounds for Principal Component Analysis in Hilbert spaces

Abstract: Based on some new robust estimators of the covariance matrix, we propose stable versions of Principal Component Analysis (PCA) and we qualify it independently of the dimension of the ambient space. We first provide a robust estimator of the orthogonal projector on the largest eigenvectors of the covariance matrix. The behavior of such an estimator is related to the size of the gap in the spectrum of the covariance matrix and in particular a large gap is needed in order to get a good approximation. To avoid the assumption of a large eigengap in the spectrum of the covariance matrix we propose a robust version of PCA that consists in performing a smooth cut-off of the spectrum via a Lipschitz function. We provide bounds on the approximation error in terms of the operator norm and of the Frobenius norm.