English

Low Rank Approximation and Regression in Input Sparsity Time

Data Structures and Algorithms 2013-04-08 v4

Abstract

We design a new distribution over \poly(r\eps1)×n\poly(r \eps^{-1}) \times n matrices SS so that for any fixed n×dn \times d matrix AA of rank rr, with probability at least 9/10, \normSAx2=(1±\eps)\normAx2\norm{SAx}_2 = (1 \pm \eps)\norm{Ax}_2 simultaneously for all xRdx \in \mathbb{R}^d. Such a matrix SS is called a \emph{subspace embedding}. Furthermore, SASA can be computed in \nnz(A)+\poly(d\eps1)\nnz(A) + \poly(d \eps^{-1}) time, where \nnz(A)\nnz(A) is the number of non-zero entries of AA. This improves over all previous subspace embeddings, which required at least Ω(ndlogd)\Omega(nd \log d) time to achieve this property. We call our matrices SS \emph{sparse embedding matrices}. Using our sparse embedding matrices, we obtain the fastest known algorithms for (1+\eps)(1+\eps)-approximation for overconstrained least-squares regression, low-rank approximation, approximating all leverage scores, and p\ell_p-regression. The leading order term in the time complexity of our algorithms is O(\nnz(A))O(\nnz(A)) or O(\nnz(A)logn)O(\nnz(A)\log n). We optimize the low-order \poly(d/\eps)\poly(d/\eps) terms in our running times (or for rank-kk approximation, the n\poly(k/eps)n*\poly(k/eps) term), and show various tradeoffs. For instance, we also use our methods to design new preconditioners that improve the dependence on \eps\eps in least squares regression to log1/\eps\log 1/\eps. Finally, we provide preliminary experimental results which suggest that our algorithms are competitive in practice.

Keywords

Cite

@article{arxiv.1207.6365,
  title  = {Low Rank Approximation and Regression in Input Sparsity Time},
  author = {Kenneth L. Clarkson and David P. Woodruff},
  journal= {arXiv preprint arXiv:1207.6365},
  year   = {2013}
}

Comments

Included optimization of subspace embedding dimension from (d/eps)^4 to O~(d/eps)^2 in Section 4, by more careful analysis of perfect hashing, and minor improvements to regression / low rank approximation because of it

R2 v1 2026-06-21T21:42:11.780Z