Nonasymptotic upper bounds for the reconstruction error of PCA (Q2196210)

This paper is a study on the principal component analysis (PCA). PCA variants are standard tools in high-dimensional statistics and unsupervised learning. The motivation for this work is that the understanding of the error incurred by PCA in high dimensions is so far limited. Its possible exhibit upper bounds for the excess risk of the reconstruction error which give different rates in sample size and dimensionality depending on spectral properties of the covariance operator, and thus exhibit complex facets of this classical statistical method. From the introduction: ``By combining spectral projector calculus with concentration inequalities, we are able to give tight bounds for the excess risk which clarify the underlying error structure. This gives rise to oracle risk bounds which in wide generality prove that the error due to projecting on empirical principal components is negligible compared to the error due to optimal dimension reduction via the population version of PCA.'' The authors include functional PCA and kernel PCA in the standard multivariate PCA setting by allowing for general Hilbert spaces concentration inequalities. The main results shown by the authors are about the PCA reconstruction error, where they present new bounds for the excess risk and applications.