Rank-Optimal Sample Complexity of Riemannian Gauss-Newton for Low-Rank Matrix Recovery

Abstract

In this paper, we study the recovery of an unknown rank-$r$ matrix $\boldsymbol{X}_\star \in \mathbb{R}^{n_1\times n_2}$ from Gaussian linear measurements using the Riemannian Gauss-Newton (RGN) algorithm. Each iteration of RGN solves the RGN equation followed by a fixed-rank retraction. We prove that, with high probability, RGN with spectral initialization converges quadratically to $\boldsymbol{X}_\star$ using $\mathcal{O} ( r(n_1+n_2)κ^2 )$ measurements, where $κ$ is the condition number of $\boldsymbol{X}_\star$. This is the first result to simultaneously achieve rank-optimal sample complexity and quadratic convergence for Gaussian low-rank matrix sensing. The key to the proof is a sharp spectral-norm analysis of the iterates, which requires controlling the inverse operator arising from the RGN equation. Our analysis develops an exact finite Neumann expansion of this inverse operator and combines it with a virtual-sequence decoupling argument and the restricted isometry property. Extensive numerical experiments support our theoretical findings and demonstrate the advantages of RGN.