An Online Sparsification Algorithm from the Book
In their seminal paper [Cohen et al., 2016], Cohen, Musco, and Pachocki proposed a natural and simple online spectral sparsification algorithm: rows $a_1, a_2, \ldots \in \mathbb{R}^d$ of a matrix $A$ arrive one-by-one, and when row $a_i$ arrives, it is appended to sparsifier $\tilde{A}$ (after appropriately reweighting it) with probability proportional to its current leverage score $$ \tau^{\mathrm{OL}}(a_i)=a_i^\top(A_i^\top A_i)^\dagger a_i, \text{ where }A_i = [a_1, a_2, \ldots, a_i]^\top $$ or otherwise discarded forever. For oblivious streams, they showed that this maintains a $(1\pm\epsilon)$-spectral approximation $\tilde{A}$ of every $A$ with $O(d\epsilon^{-2}\log^2 d)$ many rows. A natural question is whether the same algorithm works for adaptive streams, where each row may depend on the algorithm's previous random choices. The original proof does not extend directly: it analyzes the process in isotropic position with respect to the final matrix $A$, which is not fixed in advance under adaptivity. As an extension of this proof framework remained elusive, various algorithmic variants have since been suggested. In this paper, we show that the original online leverage-score sampling algorithm is indeed robust to adaptive adversaries. Our main technical contribution is a Freedman-type matrix martingale inequality with an evolving isotropic map, allowing the isotropic map used in the concentration argument to change with the stream. As a consequence, this gives the first online sparsification algorithm for adaptive streams that yields a sparsifier of near-optimal size $O(d \varepsilon^{-2}\log^2 d)$ whose working memory is proportional to the size of the sparsifier. For the special case of spectral graph sparsification, we provide an implementation that additionally runs in time near-linear in the stream size.