Phase Retrieval Algorithms: A Brief Overview

Phase retrieval is the problem of reconstructing a signal from measurements of its intensity (magnitude) while the phase information is lost. This challenge arises across many fields including X-ray crystallography, astronomy, microscopy, and optics, where detectors can only measure intensity, not the complex-valued field with both amplitude and phase.

The Fundamental Problem

When we measure the Fourier transform of a signal, detectors typically capture only |F(ω)|², losing the phase φ(ω). Recovering the original signal from intensity-only measurements is generally non-unique without additional constraints, as different signals can produce identical intensity patterns. This ambiguity makes phase retrieval an ill-posed inverse problem.

Classical Algorithms

Gerchberg-Saxton (1972) alternates between two domains (typically real space and Fourier space), imposing known constraints in each. Starting with a random phase guess, it iteratively applies constraints: enforce known intensity in Fourier space, transform back to real space, apply spatial support constraints, and repeat. While intuitive and widely used, it can stagnate in local minima.

Fienup’s Algorithms (1980s) improved upon Gerchberg-Saxton by introducing better real-space constraints. The Hybrid Input-Output (HIO) algorithm uses a feedback parameter to escape stagnation, making it more robust for difficult problems. These methods became standard in astronomical imaging and diffractive imaging.

Modern Approaches

Ptychography uses multiple overlapping illumination positions, providing redundant measurements that constrain the solution more tightly. This has revolutionized coherent X-ray imaging by enabling high-resolution reconstructions even without perfect experimental conditions.

Convex Relaxation Methods reformulate phase retrieval as optimization problems using techniques like PhaseLift, which lifts the problem into a higher-dimensional space where it becomes convex and can be solved via semidefinite programming.

Deep Learning approaches have emerged recently, using neural networks either to directly predict phases from intensities or to improve traditional iterative algorithms. These methods can be fast and handle noise well but require extensive training data.

Key Considerations

Success depends critically on having sufficient constraints—such as known support regions, oversampling, multiple measurements, or diversity in illumination or sample positions. The choice of algorithm depends on the specific application, available computational resources, noise levels, and whether real-time reconstruction is needed.

Current Research

Helping with development of dual-plane imaging systems that improve phase retrieval performance when looking through high turbulence and the target is actively illuminated. .