Title: Nearly Optimal Robust Subspace Tracking and Dynamic Robust PCA

Abstract: We study the robust subspace tracking (RST) problem and obtain one of the first provable guarantees for it. The goal of RST is to track data that lies in a slowly changing low-dimensional subspace, and the subspaces themselves, while being robust to corruption by (often large magnitude) sparse outliers. It can be simply interpreted as a dynamic (time-varying) extension of robust PCA, with the minor difference that RST also requires an online algorithm (short tracking delay). We propose an algorithm called NORST (nearly optimal RST) and prove that it solves both RST and dynamic robust PCA under weakened versions of standard RPCA assumptions, slow subspace change, and two simple extra assumptions (a lower bound on outlier magnitudes, and independence of the columns of the low-rank matrix).
Our guarantee shows that NORST enjoys a near optimal tracking delay of $O(r \log n \log(1/\varepsilon))$. Its required delay between subspace change times is the same, and its memory complexity is $n$ times this value. Here $n$ is the ambient space dimension and $r$ is the dimension of the changing subspaces in which the true data lies. Thus both these are also nearly optimal. Finally, our guarantee also shows that NORST has the best outlier tolerance compared with all previous RPCA or RST methods, both theoretically and empirically (including for real videos), without requiring any model on outlier support sets.