Phase Retrieval via Smooth Amplitude Flow

TL;DR

This paper introduces a smooth amplitude flow method (SAF) for phase retrieval that avoids non-smooth loss functions, guarantees global convergence, and outperforms existing methods in recovery accuracy and speed.

Contribution

SAF is a novel smooth loss function approach for phase retrieval that ensures geometric convergence without extra modifications, improving over prior non-smooth methods.

Findings

SAF converges geometrically to a global optimum with high probability.

SAF outperforms state-of-the-art methods in recovery rate and speed.

SAF can recover signals below the information-theoretic measurement limit.

Abstract

Phase retrieval (PR) is an inverse problem about recovering a signal from phaseless linear measurements. This problem can be effectively solved by minimizing a nonconvex amplitude-based loss function. However, this loss function is non-smooth. To address the non-smoothness, a series of methods have been proposed by adding truncating, reweighting and smoothing operations to adjust the gradient or the loss function and achieved better performance. But these operations bring about extra rules and parameters that need to be carefully designed. Unlike previous works, we present a smooth amplitude flow method (SAF) which minimizes a novel loss function, without additionally modifying the gradient or the loss function during gradient descending. Such a new heuristic can be regarded as a smooth version of the original non-smooth amplitude-based loss function. We prove that SAF can converge geometrically to a global optimal point via the gradient algorithm with an elaborate initialization stage with a high probability. Substantial numerical tests empirically illustrate that the proposed heuristic is significantly superior to the original amplitude-based loss function and SAF also outperforms other state-of-the-art methods in terms of the recovery rate and the converging speed. Specially, it is numerically shown that SAF can stably recover the original signal when number of measurements is smaller than the information-theoretic limit for both the real and the complex Gaussian models.