English

Nonconvex Matrix Factorization from Rank-One Measurements

Information Theory 2018-12-04 v2 Machine Learning math.IT Machine Learning

Abstract

We consider the problem of recovering low-rank matrices from random rank-one measurements, which spans numerous applications including covariance sketching, phase retrieval, quantum state tomography, and learning shallow polynomial neural networks, among others. Our approach is to directly estimate the low-rank factor by minimizing a nonconvex quadratic loss function via vanilla gradient descent, following a tailored spectral initialization. When the true rank is small, this algorithm is guaranteed to converge to the ground truth (up to global ambiguity) with near-optimal sample complexity and computational complexity. To the best of our knowledge, this is the first guarantee that achieves near-optimality in both metrics. In particular, the key enabler of near-optimal computational guarantees is an implicit regularization phenomenon: without explicit regularization, both spectral initialization and the gradient descent iterates automatically stay within a region incoherent with the measurement vectors. This feature allows one to employ much more aggressive step sizes compared with the ones suggested in prior literature, without the need of sample splitting.

Keywords

Cite

@article{arxiv.1802.06286,
  title  = {Nonconvex Matrix Factorization from Rank-One Measurements},
  author = {Yuanxin Li and Cong Ma and Yuxin Chen and Yuejie Chi},
  journal= {arXiv preprint arXiv:1802.06286},
  year   = {2018}
}

Comments

34 pages

R2 v1 2026-06-23T00:25:28.526Z