English

Matrix Compression using the Nystro\"om Method

Numerical Analysis 2013-05-02 v1

Abstract

The Nystr\"{o}m method is routinely used for out-of-sample extension of kernel matrices. We describe how this method can be applied to find the singular value decomposition (SVD) of general matrices and the eigenvalue decomposition (EVD) of square matrices. We take as an input a matrix MRm×nM\in \mathbb{R}^{m\times n}, a user defined integer smin(m,n)s\leq min(m,n) and AMRs×sA_M \in \mathbb{R}^{s\times s}, a matrix sampled from the columns and rows of MM. These are used to construct an approximate rank-ss SVD of MM in O(s2(m+n))O\left(s^2\left(m+n\right)\right) operations. If MM is square, the rank-ss EVD can be similarly constructed in O(s2n)O\left(s^2 n\right) operations. Thus, the matrix AMA_M is a compressed version of MM. We discuss the choice of AMA_M and propose an algorithm that selects a good initial sample for a pivoted version of MM. The proposed algorithm performs well for general matrices and kernel matrices whose spectra exhibit fast decay.

Keywords

Cite

@article{arxiv.1305.0203,
  title  = {Matrix Compression using the Nystro\"om Method},
  author = {Arik Nemtsov and Amir Averbuch and Alon Schclar},
  journal= {arXiv preprint arXiv:1305.0203},
  year   = {2013}
}

Comments

31 pages, 3 figures, submitted to Linear Algebra and its Applications

R2 v1 2026-06-22T00:09:39.239Z