Row Sampling for Matrix Algorithms via a Non-Commutative Bernstein Bound

TL;DR

This paper introduces new row sampling algorithms for matrix computations that depend on leverage scores, using non-commutative Bernstein bounds, and provides the first efficient methods to approximate these probabilities without full SVD.

Contribution

It presents novel algorithms for row sampling based on leverage scores that avoid full SVD computations, improving efficiency for matrix approximation tasks.

Findings

Algorithms achieve near-linear dependence on stable rank

First algorithms to approximate leverage scores without SVD

Use of non-commutative Bernstein bounds enhances analysis

Abstract

We focus the use of \emph{row sampling} for approximating matrix algorithms. We give applications to matrix multipication; sparse matrix reconstruction; and, \math{\ell_2} regression. For a matrix \math{\matA\in\R^{m\times d}} which represents \math{m} points in \math{d\ll m} dimensions, all of these tasks can be achieved in \math{O(md^2)} via the singular value decomposition (SVD). For appropriate row-sampling probabilities (which typically depend on the norms of the rows of the \math{m\times d} left singular matrix of \math{\matA} (the \emph{leverage scores}), we give row-sampling algorithms with linear (up to polylog factors) dependence on the stable rank of \math{\matA}. This result is achieved through the application of non-commutative Bernstein bounds. We then give, to our knowledge, the first algorithms for computing approximations to the appropriate row-sampling probabilities without going through the SVD of \math{\matA}. Thus, these are the first \math{o(md^2)} algorithms for row-sampling based approximations to the matrix algorithms which use leverage scores as the sampling probabilities. The techniques we use to approximate sampling according to the leverage scores uses some powerful recent results in the theory of random projections for embedding, and may be of some independent interest. We confess that one may perform all these matrix tasks more efficiently using these same random projection methods, however the resulting algorithms are in terms of a small number of linear combinations of all the rows. In many applications, the actual rows of \math{\matA} have some physical meaning and so methods based on a small number of the actual rows are of interest.