Imagine trying to reconstruct the three-dimensional shape of a glass sculpture by looking only at the flat shadows it casts on a wall. You can see the outlines and varying brightness, but the precise depth—the timing of how light waves slowed down and bent as they passed through the glass contours—is lost.
This is the fundamental challenge of phase retrieval. In fields ranging from optical microscopy and X-ray crystallography to astronomical imaging, physical sensors (such as CMOS cameras or electron detectors) can only record the intensity of a wave. Because the frequency of light oscillates far too quickly for electronics to track directly, detectors simply count photons over time. This process captures the squared magnitude of the complex wave field ($|Ax|^2$) but completely discards the phase—the relative timing of the wave's peaks and troughs.
The Physical Loss of Phase
Why does losing the phase matter so much? In wave physics, the phase contains the primary structural information of an object. As highlighted across domains from holographic microscopy by Mazen Mel et al. in HoloADMM: High-Quality Holographic Complex Field Recovery to diffractive imaging by Simon Welker et al. in Deep Iterative Phase Retrieval for Ptychography, phase shifts describe how a wave is delayed as it travels through different materials.
For transparent biological specimens, the amplitude of light remains virtually unchanged as it passes through the sample; only the phase shifts. Without recovering this phase information, these specimens remain invisible to standard intensity-only sensors.
The Mathematical Ambiguity
Mathematically, recovering a complex-valued signal from magnitude-only constraints is a severely ill-posed, non-convex problem. If we multiply a target signal $x$ by a global phase factor $e^{i\phi}$, the measured intensity $|Ax|^2$ remains completely unchanged. As demonstrated by Praneeth Netrapalli et al. in Phase Retrieval using Alternating Minimization, this global phase ambiguity means a signal can only ever be recovered up to a constant complex phase shift.
In scattering applications, this loss of information manifests as severe spatial ambiguities. For instance, Xiaoxu Xu et al. in Uniqueness in Inverse Scattering Problems with Phaseless Far-Field Data at a Fixed Frequency and Bo Zhang 0052 et al. in Uniqueness and Direct Imaging Method for Inverse Scattering by Locally Rough Surfaces with Phaseless Near-Field Data show that phaseless data exhibits "translation invariance." Shifting the absolute position of an object does not alter the modulus of its far-field scattering pattern. Because multiple distinct physical structures can yield identical intensity profiles, optimization algorithms easily get trapped in incorrect local minima.
Classical and Convex Optimization Strategies
To recover the complex field, algorithms must find a vector $x$ that satisfies the known physical constraints of the system.
Alternating Minimization
The classical approach relies on alternating projections. The algorithm bounces back and forth between the spatial domain (where it enforces constraints like object boundaries, positivity, or sparsity) and the frequency domain (where it replaces the calculated wave magnitudes with the actual measured intensities). The authors of Phase Retrieval using Alternating Minimization provided some of the first rigorous theoretical guarantees for these heuristics, proving that a smart initialization allows alternating minimization to converge to the true signal.
Convex Lifting
To bypass the non-convexity of the phase problem entirely, researchers developed a technique called "lifting." The core insight is to transform a non-convex vector problem into a convex matrix problem. The quadratic measurement can be rewritten using the trace operator:
$$|a_i^* x|^2 = a_i^* x x^* a_i = \text{Tr}(a_i a_i^* X)$$
where $X = xx^*$ is a rank-one, positive semidefinite matrix. By optimizing for the matrix $X$ rather than the vector $x$, the measurements become linear. Frameworks like PhaseCut, introduced by Fogel et al. in Phase retrieval for imaging problems, and CPRL, proposed by Henrik Ohlsson et al. in CPRL -- An Extension of Compressive Sensing to the Phase Retrieval Problem, drop the difficult rank-one constraint to solve the problem via convex semidefinite programming.
Direct Non-Linear Optimization
While lifting guarantees global convergence, scaling a matrix $X$ of size $N \times N$ becomes computationally prohibitive for large images. As an alternative, Zalan Fabian et al. in 3D Phase Retrieval at Nano-Scale via Accelerated Wirtinger Flow utilize Wirtinger Flow—a method that applies a specialized gradient descent directly to the non-convex objective in the original vector space, using regularizers like Total Variation to stabilize 3D structural recovery.
Physics-Informed Deep Learning
Purely classical optimization often struggles when measurements are corrupted by heavy noise or contain large missing regions. Modern phase retrieval architectures address this by embedding the physical forward model directly into deep learning frameworks.
Rather than using neural networks as black-box end-to-end mapping functions, the authors of HoloADMM: High-Quality Holographic Complex Field Recovery unroll classical optimization algorithms (like the Alternating Direction Method of Multipliers) into neural network layers. By interleaving physics-based wave propagation steps with convolutional neural network (CNN) denoising blocks, they eliminate structural artifacts like the "twin-image" problem in digital holography.
+-------------------------------------------------------+
| |
v |
[Physics Step: Wave Propagation] ---> [CNN Prior: Denoising]
(Enforces Intensity Constraints) (Enforces Structural Realism)
When explicit training data is unavailable, untrained neural networks can serve as powerful structural priors. In Phase Retrieval with Holography and Untrained Priors: Tackling the Challenges of Low-Photon Nanoscale Imaging, Hannah Lawrence et al. demonstrate that optimizing the weights of an untrained network against a physics-informed Poisson likelihood function prevents the algorithm from fitting noise, enabling high-fidelity phase recovery even in extreme, low-photon regimes.
Hardware-Software Co-Design
Computational algorithms perform best when the physical measurement process itself is designed to break the underlying mathematical ambiguities.
One prominent strategy is multi-frame optical modulation. In WISH: wavefront imaging sensor with high resolution, Yicheng Wu 0001 et al. introduce a system that replaces physical microlens arrays with a programmable Spatial Light Modulator (SLM). By actively changing the phase profile of the SLM across multiple exposures, they capture a diverse set of modulated intensity patterns, providing the mathematical diversity required to uniquely reconstruct the complex field at high resolutions.
Alternatively, single-shot phase retrieval can be achieved by engineering the sensor itself. The authors of Asymmetric metasurface photodetectors for single-shot quantitative phase imaging. design specialized, asymmetric metasurfaces directly on top of photodetectors. These nanostructures transform local phase variations into measurable intensity differences in a single capture, bypassing the need for bulky, vibration-sensitive interferometric setups.