The discrete sign problem: uniqueness, recovery algorithms and phase retrieval applications

TL;DR

This paper investigates a specific 1-D phase retrieval problem focusing on sign recovery, proving uniqueness under over-sampling, and developing an efficient algorithm with applications in vectorial phase retrieval and object separation.

Contribution

It establishes the uniqueness of the sign problem with over-sampling, models it as a piecewise constant phase problem, and proposes a robust, efficient recovery algorithm.

Findings

Unique solution with sufficient over-sampling.

Algorithm guarantees sign recovery in noise-free, high sampling scenarios.

Applications demonstrated in vectorial phase retrieval and object separation.

Abstract

In this paper we consider the following real-valued and finite dimensional specific instance of the 1-D classical phase retrieval problem. Let ${\bf F}\in\mathbb{R}^N$ be an $N$-dimensional vector, whose discrete Fourier transform has a compact support. The sign problem is to recover ${\bf F}$ from its magnitude $|{\bf F}|$. First, in contrast to the classical 1-D phase problem which in general has multiple solutions, we prove that with sufficient over-sampling, the sign problem admits a unique solution. Next, we show that the sign problem can be viewed as a special case of a more general piecewise constant phase problem. Relying on this result, we derive a computationally efficient and robust to noise sign recovery algorithm. In the noise-free case and with a sufficiently high sampling rate, our algorithm is guaranteed to recover the true sign pattern. Finally, we present two phase retrieval applications of the sign problem: (i) vectorial phase retrieval with three measurement vectors; and (ii) recovery of two well separated 1-D objects.