Fast Regression for Structured Inputs

TL;DR

This paper introduces efficient subsampling algorithms for $\, ext{l}_p$ regression on structured matrices, achieving polynomial dependence on $p$, and extends to $ ext{l}_\, ext{infinity}$ regression, improving computational feasibility.

Contribution

The paper presents novel polynomial-time subsampling methods for $\, ext{l}_p$ regression on structured matrices, reducing dependence on $p$ compared to prior exponential-time algorithms.

Findings

Algorithms for Vandermonde matrices with runtime depending polynomially on $p$.

Extension of methods to $ ext{l}_\, ext{infinity}$ regression via approximation.

A new, simpler subsampling algorithm for arbitrary matrices for $p \,\geq 4$.

Abstract

We study the $\ell_p$ regression problem, which requires finding $\mathbf{x}\in\mathbb R^{d}$ that minimizes $\|\mathbf{A}\mathbf{x}-\mathbf{b}\|_p$ for a matrix $\mathbf{A}\in\mathbb R^{n \times d}$ and response vector $\mathbf{b}\in\mathbb R^{n}$. There has been recent interest in developing subsampling methods for this problem that can outperform standard techniques when $n$ is very large. However, all known subsampling approaches have run time that depends exponentially on $p$, typically, $d^{\mathcal{O}(p)}$, which can be prohibitively expensive. We improve on this work by showing that for a large class of common \emph{structured matrices}, such as combinations of low-rank matrices, sparse matrices, and Vandermonde matrices, there are subsampling based methods for $\ell_p$ regression that depend polynomially on $p$. For example, we give an algorithm for $\ell_p$ regression on Vandermonde matrices that runs in time $\mathcal{O}(n\log^3 n+(dp^2)^{0.5+\omega}\cdot\text{polylog}\,n)$, where $\omega$ is the exponent of matrix multiplication. The polynomial dependence on $p$ crucially allows our algorithms to extend naturally to efficient algorithms for $\ell_\infty$ regression, via approximation of $\ell_\infty$ by $\ell_{\mathcal{O}(\log n)}$. Of practical interest, we also develop a new subsampling algorithm for $\ell_p$ regression for arbitrary matrices, which is simpler than previous approaches for $p \ge 4$.