Spectral Guarantees for Adversarial Streaming PCA

In streaming PCA, we see a stream of vectors and want to estimate the top eigenvector of their covariance matrix. This is easier if the spectral ratio is large. We ask: how large does need to be to solve streaming PCA in space? Existing algorithms require . We show: (1) For all mergeable summaries, is necessary. (2) In the insertion-only model, a variant of Oja's algorithm gets error for . (3) No algorithm with space gets error for . Our analysis is the first application of Oja's algorithm to adversarial streams. It is also the first algorithm for adversarial streaming PCA that is designed for a spectral, rather than Frobenius, bound on the tail; and the bound it needs is exponentially better than is possible by adapting a Frobenius guarantee.
View on arXiv