Stable Recovery and Benign Overparameterized Landscapes for Phase Retrieval from Coded Diffraction Patterns
In the authors' words
Coded diffraction patterns (CDPs) provide a structured and physically relevant model for phase retrieval, but the dependence among Fourier measurements generated by a common mask makes sharp stability analysis challenging. For a fixed unit-norm signal , let , and let be the lifted CDP measurement operator. We prove that, with random masks, the following uniform lower isometry holds with high probability: from which we derive two consequences. First, for , PhaseLift-type convex programs achieve the Gaussian-type stable recovery bound Second, in the noiseless case, the nonconvex factorized loss has a benign landscape when the factor width satisfies : every second-order critical point satisfies . The key ingredient is a uniform operator-norm bound over row subsets of the dependent CDP measurement matrix, which permits the removal of a controlled set of adaptively selected rows while preserving tangent injectivity.
Appeared: Monday, September 28. arXiv. Preprint, not yet peer-reviewed.
Authors' comment: 38 pages