Improving the Threshold for Finding Rank-1 Matrices in a Subspace
Data Structures and Algorithms
2025-04-28 v1
Abstract
We consider a basic computational task of finding planted rank-1 matrices in a linear subspace where . The work of Johnston-Lovitz-Vijayaraghavan (FOCS 2023) gave a polynomial-time algorithm for this task and proved that it succeeds when , under minimal genericity assumptions on the input. Aiming to precisely characterize the performance of this algorithm, we improve the bound to and also prove that the algorithm fails when . Numerical experiments indicate that the true breaking point is . Our work implies new algorithmic results for tensor decomposition, for instance, decomposing order-4 tensors with twice as many components as before.
Cite
@article{arxiv.2504.17947,
title = {Improving the Threshold for Finding Rank-1 Matrices in a Subspace},
author = {Jeshu Dastidar and Tait Weicht and Alexander S. Wein},
journal= {arXiv preprint arXiv:2504.17947},
year = {2025}
}
Comments
37 pages