16
6

Krylov Methods are (nearly) Optimal for Low-Rank Approximation

Abstract

We consider the problem of rank-11 low-rank approximation (LRA) in the matrix-vector product model under various Schatten norms: \min_{\|u\|_2=1} \|A (I - u u^\top)\|_{\mathcal{S}_p} , where MSp\|M\|_{\mathcal{S}_p} denotes the p\ell_p norm of the singular values of MM. Given ε>0\varepsilon>0, our goal is to output a unit vector vv such that \|A(I - vv^\top)\|_{\mathcal{S}_p} \leq (1+\varepsilon) \min_{\|u\|_2=1}\|A(I - u u^\top)\|_{\mathcal{S}_p}. Our main result shows that Krylov methods (nearly) achieve the information-theoretically optimal number of matrix-vector products for Spectral (p=p=\infty), Frobenius (p=2p=2) and Nuclear (p=1p=1) LRA. In particular, for Spectral LRA, we show that any algorithm requires Ω(log(n)/ε1/2)\Omega\left(\log(n)/\varepsilon^{1/2}\right) matrix-vector products, exactly matching the upper bound obtained by Krylov methods [MM15, BCW22]. Our lower bound addresses Open Question 1 in [Woo14], providing evidence for the lack of progress on algorithms for Spectral LRA and resolves Open Question 1.2 in [BCW22]. Next, we show that for any fixed constant pp, i.e. 1p=O(1)1\leq p =O(1), there is an upper bound of O(log(1/ε)/ε1/3)O\left(\log(1/\varepsilon)/\varepsilon^{1/3}\right) matrix-vector products, implying that the complexity does not grow as a function of input size. This improves the O(log(n/ε)/ε1/3)O\left(\log(n/\varepsilon)/\varepsilon^{1/3}\right) bound recently obtained in [BCW22], and matches their Ω(1/ε1/3)\Omega\left(1/\varepsilon^{1/3}\right) lower bound, to a log(1/ε)\log(1/\varepsilon) factor.

View on arXiv
Comments on this paper