A New Framework for Matrix Discrepancy: Partial Coloring Bounds via Mirror Descent
Abstract
Motivated by the Matrix Spencer conjecture, we study the problem of finding signed sums of matrices with a small matrix norm. A well-known strategy to obtain these signs is to prove, given matrices , a Gaussian measure lower bound of for a scaling of the discrepancy body . We show this is equivalent to covering its polar with translates of the cube , and construct such a cover via mirror descent. As applications of our framework, we show: Matrix Spencer for Low-Rank Matrices. If the matrices satisfy and , we can efficiently find a coloring with discrepancy . This improves upon the naive bound for random coloring and proves the matrix Spencer conjecture when . Matrix Spencer for Block Diagonal Matrices. For block diagonal matrices with and block size , we can efficiently find a coloring with . Using our proof, we reduce the matrix Spencer conjecture to the existence of a quantum relative entropy net on the spectraplex. Matrix Discrepancy for Schatten Norms. We generalize our discrepancy bound for matrix Spencer to Schatten norms . Given and , we can efficiently find a partial coloring with and , where .
Cite
@article{arxiv.2111.03171,
title = {A New Framework for Matrix Discrepancy: Partial Coloring Bounds via Mirror Descent},
author = {Daniel Dadush and Haotian Jiang and Victor Reis},
journal= {arXiv preprint arXiv:2111.03171},
year = {2021}
}
Comments
24 pages