Compressed Sensing from Phaseless Gaussian Measurements via Linear Programming in the Natural Parameter Space

TL;DR

This paper introduces SparsePhaseMax, a linear program for phaseless compressed sensing that guarantees recovery of sparse signals from Gaussian measurements with optimal sample complexity, leveraging a novel connection to 1-bit compressed sensing.

Contribution

It presents the first provable method for phaseless compressed sensing with optimal measurements using a linear program in the natural parameter space.

Findings

Recovers k-sparse vectors from O(k log(n/k)) measurements

Establishes a connection between phaseless and 1-bit compressed sensing

Achieves recovery with high probability from Gaussian measurements

Abstract

We consider faithfully combining phase retrieval with classical compressed sensing. Inspired by the recent novel formulation for phase retrieval called PhaseMax, we present and analyze SparsePhaseMax, a linear program for phaseless compressed sensing in the natural parameter space. We establish that when provided with an initialization that correlates with an arbitrary $k$-sparse $n$-vector, SparsePhaseMax recovers this vector up to global sign with high probability from $O(k \log \frac{n}{k})$ magnitude measurements against i.i.d. Gaussian random vectors. Our proof of this fact exploits a curious newfound connection between phaseless and 1-bit compressed sensing. This is the first result to establish bootstrapped compressed sensing from phaseless Gaussian measurements under optimal sample complexity.