79
32

Distributed Low Rank Approximation of Implicit Functions of a Matrix

Abstract

We study distributed low rank approximation in which the matrix to be approximated is only implicitly represented across the different servers. For example, each of ss servers may have an n×dn \times d matrix AtA^t, and we may be interested in computing a low rank approximation to A=f(t=1sAt)A = f(\sum_{t=1}^s A^t), where ff is a function which is applied entrywise to the matrix t=1sAt\sum_{t=1}^s A^t. We show for a wide class of functions ff it is possible to efficiently compute a d×dd \times d rank-kk projection matrix PP for which AAPF2A[A]kF2+εAF2\|A - AP\|_F^2 \leq \|A - [A]_k\|_F^2 + \varepsilon \|A\|_F^2, where APAP denotes the projection of AA onto the row span of PP, and [A]k[A]_k denotes the best rank-kk approximation to AA given by the singular value decomposition. The communication cost of our protocols is d(sk/ε)O(1)d \cdot (sk/\varepsilon)^{O(1)}, and they succeed with high probability. Our framework allows us to efficiently compute a low rank approximation to an entry-wise softmax, to a Gaussian kernel expansion, and to MM-Estimators applied entrywise (i.e., forms of robust low rank approximation). We also show that our additive error approximation is best possible, in the sense that any protocol achieving relative error for these problems requires significantly more communication. Finally, we experimentally validate our algorithms on real datasets.

View on arXiv
Comments on this paper