Hutch++: Optimal Stochastic Trace Estimation

Code

• github.com/RaphaelArkadyMeyerNYU/HutchPlusPlus
  OfficialIn papernone★ 13

Abstract

We study the problem of estimating the trace of a matrix A that can only be accessed through matrix-vector multiplication. We introduce a new randomized algorithm, Hutch++, which computes a (1 ) approximation to tr(A) for any positive semidefinite (PSD) A using just O(1/) matrix-vector products. This improves on the ubiquitous Hutchinson's estimator, which requires O(1/^2) matrix-vector products. Our approach is based on a simple technique for reducing the variance of Hutchinson's estimator using a low-rank approximation step, and is easy to implement and analyze. Moreover, we prove that, up to a logarithmic factor, the complexity of Hutch++ is optimal amongst all matrix-vector query algorithms, even when queries can be chosen adaptively. We show that it significantly outperforms Hutchinson's method in experiments. While our theory mainly requires A to be positive semidefinite, we provide generalized guarantees for general square matrices, and show empirical gains in such applications.

Reproductions