Title: Phase Retrieval Without Small-Ball Probability Assumptions

Abstract: This paper addresses the problem of phase retrieval using random subgaussian measurements. Even in this simple setting, the question of which signals can be reconstructed from phaseless measurements is somewhat delicate. It is known that one must impose additional conditions on the measurement distribution D, for instance by controlling its small-ball probabilities, in order to recover arbitrary signals.
This paper shows a complementary result: even when one does not make any additional assumptions about D, there is still a large class of signals that can be successfully recovered. This class consists of all non-peaky signals x \in R^n, which satisfy ||x||_\infty <= \mu ||x||_2, where \mu \in (0,1) is a constant that does not depend on n. This phenomenon is illustrated by two results: first, a proof of stability and uniqueness, extending previous work by Eldar and Mendelson; and second, a uniform recovery guarantee for the PhaseLift algorithm, extending previous work by Candes and Li. In all of these cases, the number of measurements m approaches the information-theoretic lower bound.
Finally, for certain special choices of D (e.g., random Bernoulli measurements with erasures), it is shown that PhaseLift achieves uniform recovery of all signals in R^n (including peaky ones).