English

Input-Sparsity Low Rank Approximation in Schatten Norm

Data Structures and Algorithms 2020-07-01 v3

Abstract

We give the first input-sparsity time algorithms for the rank-kk low rank approximation problem in every Schatten norm. Specifically, for a given n×nn\times n matrix AA, our algorithm computes Y,ZRn×kY,Z\in \mathbb{R}^{n\times k}, which, with high probability, satisfy AYZTp(1+ϵ)AAkp\|A-YZ^T\|_p \leq (1+\epsilon)\|A-A_k\|_p, where Mp=(i=1nσi(M)p)1/p\|M\|_p = \left (\sum_{i=1}^n \sigma_i(M)^p \right )^{1/p} is the Schatten pp-norm of a matrix MM with singular values σ1(M),,σn(M)\sigma_1(M), \ldots, \sigma_n(M), and where AkA_k is the best rank-kk approximation to AA. Our algorithm runs in time O~(nnz(A)+mnαppoly(k/ϵ))\tilde{O}(\operatorname{nnz}(A) + mn^{\alpha_p}\operatorname{poly}(k/\epsilon)), where αp=0\alpha_p = 0 for p[1,2)p\in [1,2) and αp=(ω1)(12/p)\alpha_p = (\omega-1)(1-2/p) for p>2p>2 and ω2.374\omega \approx 2.374 is the exponent of matrix multiplication. For the important case of p=1p = 1, which corresponds to the more "robust" nuclear norm, we obtain O~(nnz(A)+mpoly(k/ϵ))\tilde{O}(\operatorname{nnz}(A) + m \cdot \operatorname{poly}(k/\epsilon)) time, which was previously only known for the Frobenius norm (p=2p = 2). Moreover, since αp<ω1\alpha_p < \omega - 1 for every pp, our algorithm has a better dependence on nn than that in the singular value decomposition for every pp. Crucial to our analysis is the use of dimensionality reduction for Ky-Fan pp-norms.

Keywords

Cite

@article{arxiv.2004.12646,
  title  = {Input-Sparsity Low Rank Approximation in Schatten Norm},
  author = {Yi Li and David Woodruff},
  journal= {arXiv preprint arXiv:2004.12646},
  year   = {2020}
}

Comments

Submitted to ICML 2020. In this version, we state the result for rectangular matrices in the abstract and introduction and give an example comparing to low-rank approximation to A^TA

R2 v1 2026-06-23T15:06:59.153Z