Core Concepts
The paper establishes a universal lower bound on the condition number of the phase retrieval map, which is asymptotically tight. It also identifies the optimal measurement matrix that achieves the minimal condition number for phase retrieval in R2.
Abstract
The paper focuses on analyzing the stability property of phase retrieval by examining the bi-Lipschitz property of the map ΦA(x) = |Ax| ∈Rm+, where x ∈Hd and A ∈Hm×d is the measurement matrix for H ∈{R, C}.
Key highlights:
The authors define the condition number βA := UA/LA, where LA and UA represent the optimal lower and upper Lipschitz constants, respectively.
They establish a universal lower bound on βA, showing that for any A ∈Hm×d:
βA ≥βH0 := √(π/(π-2)) ≈ 1.659 if H = R, and √(4/(4-π)) ≈ 2.159 if H = C.
They prove that the condition number of a standard Gaussian matrix in Hm×d asymptotically matches the lower bound βH0 for both real and complex cases, indicating the tightness of the bound.
For the real case with d = 2, they show that the harmonic frame Em ∈Rm×2 possesses the minimum condition number among all A ∈Rm×2 when m ≥ 3 is odd.
As an application, the authors utilize the established results to investigate the performance of quadratic models for phase retrieval.
Stats
m ≥ 2d - 1 (or m ≥ 4d - 4) generic measurements are adequate for the precise recovery of x ∈Hd, up to a unimodular constant, where H = R (or H = C).