Numerical Linear Algebra in Linear Space

TL;DR

This paper introduces a novel randomized linear-space solver for general linear systems over the rationals that operates efficiently without condition number assumptions, enabling applications in various numerical linear algebra problems.

Contribution

It presents the first linear-space, randomized solver for rational linear systems with near-quadratic time complexity, applicable to a broad class of problems.

Findings

First linear-space solver for rational systems with near-quadratic time

Efficient algorithms for linear regression, LP, eigenvalues, and SVD

Works without condition number assumptions

Abstract

We present a randomized linear-space solver for general linear systems $\mathbf{A} \mathbf{x} = \mathbf{b}$ with $\mathbf{A} \in \mathbb{Z}^{n \times n}$ and $\mathbf{b} \in \mathbb{Z}^n$, without any assumption on the condition number of $\mathbf{A}$. For matrices whose entries are bounded by $\mathrm{poly}(n)$, the solver returns a $(1+\epsilon)$-multiplicative entry-wise approximation to vector $\mathbf{x} \in \mathbb{Q}^{n}$ using $\widetilde{O}(n^2 \cdot \mathrm{nnz}(\mathbf{A}))$ bit operations and $O(n \log n)$ bits of working space (i.e., linear in the size of a vector), where $\mathrm{nnz}$ denotes the number of nonzero entries. Our solver works for right-hand vector $\mathbf{b}$ with entries up to $n^{O(n)}$. To our knowledge, this is the first linear-space linear system solver over the rationals that runs in $\widetilde{O}(n^2 \cdot \mathrm{nnz}(\mathbf{A}))$ time. We also present several applications of our solver to numerical linear algebra problems, for which we provide algorithms with efficient polynomial running time and near-linear space. In particular, we present results for linear regression, linear programming, eigenvalues and eigenvectors, and singular value decomposition.