Nonconvex matrix factorization from rank-one measurements

We consider the problem of recovering low-rank matrices from random rank-one measurements, which spans numerous applications including phase retrieval, quantum state tomography, and learning shallow neural networks with quadratic activations, among others. Our approach is to directly estimate the low-rank factor by minimizing a nonconvex least-squares 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 and computational complexities with respect to the problem size. To the best of our knowledge, this is the first theoretical 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.