Efficient coordinate-wise leading eigenvector computation
Numerical Analysis
2017-02-28 v1 Machine Learning
Machine Learning
Abstract
We develop and analyze efficient "coordinate-wise" methods for finding the leading eigenvector, where each step involves only a vector-vector product. We establish global convergence with overall runtime guarantees that are at least as good as Lanczos's method and dominate it for slowly decaying spectrum. Our methods are based on combining a shift-and-invert approach with coordinate-wise algorithms for linear regression.
Cite
@article{arxiv.1702.07834,
title = {Efficient coordinate-wise leading eigenvector computation},
author = {Jialei Wang and Weiran Wang and Dan Garber and Nathan Srebro},
journal= {arXiv preprint arXiv:1702.07834},
year = {2017}
}