In many physical systems, we can measure the "how much" but not the "when." Whether it is X-ray crystallography, astronomical imaging, or holographic microscopy, our sensors—like CCD cameras or photon counters—are typically only sensitive to the intensity (magnitude) of a wave, completely discarding its phase.
This loss of phase information is catastrophic because the phase often encodes the structural essence of the signal. In microscopy, for instance, transparent biological samples don't absorb much light (low intensity contrast) but they do shift the light's phase. Without that phase data, the image looks like a blank sheet. The challenge of reconstructing a signal from only these magnitude-only measurements is known as the Phase Retrieval problem.
The Geometry of the Phase Problem
Mathematically, we want to recover a signal $x$ from measurements of the form $y_i = | \langle a_i, x \rangle |^2$. If $x$ is a vector in a complex space, losing the phase means we only know the distance of $x$ from the origin along various projection axes $a_i$, but we don't know the "angle."
This is fundamentally a non-convex problem. If you imagine the set of all signals that could produce a specific magnitude measurement, they form a circle (or a hypersphere) in space. Finding a signal that satisfies many such measurements is like finding the intersection of many circles. This intersection is not a "nice" convex set; it is a collection of isolated points, making traditional optimization difficult.
Lifting into Higher Dimensions
One of the most influential breakthroughs in making this problem tractable was the concept of PhaseLift, introduced by Emmanuel J. Candès, Thomas Strohmer, and Vladislav Voroninski in PhaseLift: Exact and Stable Signal Recovery from Magnitude Measurements via Convex Programming.
The "intuition" behind PhaseLift is a technique called lifting. Instead of searching for a vector $x$, we look for a matrix $X = xx^*$. The measurement equation then becomes linear: $y_i = \text{Tr}(A_i X)$, where $A_i = a_i a_i^*$. While we have gained linearity, we have also increased the dimensionality significantly. To ensure we find a valid signal, $X$ must be a rank-1, positive semi-definite matrix. By relaxing the "rank-1" constraint to a "trace" constraint (similar to how $\ell_1$ norm relaxes sparsity), the authors transformed phase retrieval into a convex semidefinite program that can be solved reliably.
Leveraging Sparsity
If we know the signal we are looking for is "sparse"—meaning most of its values are zero—we can recover it with far fewer measurements. This is the domain of Compressive Phase Retrieval (CPR).
Mehmet Akcakaya and Vahid Tarokh established in New Conditions for Sparse Phase Retrieval that for a signal with $k$ non-zero entries, we only need roughly $2k$ measurements for real signals and $4k$ for complex ones. To actually perform this recovery, researchers like Zai Yang and Cishen Zhang combined $\ell_1$ minimization with iterative solvers like the Alternating Direction Method (ADM) to reconstruct images from as little as 20% of their Fourier data Robust Compressive Phase Retrieval via L1 Minimization With Application to Image Reconstruction.
Smoothing the Optimization Landscape
While convex relaxation is powerful, it is computationally expensive for large images. This has led to a focus on "first-order" methods—like gradient descent—that operate directly on the non-convex landscape.
The danger with non-convexity is getting stuck in local minima (false solutions). Jian-Feng Cai and colleagues have shown that by using "Perturbed Amplitude Models," we can mathematically "smooth" the landscape so that almost any random starting point will slide down into the global optimum The global landscape of phase retrieval I: perturbed amplitude models.
Furthermore, the "starting point" itself can be optimized. Antoine Maillard and Lenka Zdeborova developed automated frameworks for designing spectral methods—mathematical "warm-ups" that provide a near-perfect initial guess for the signal before the main optimization begins Construction of optimal spectral methods in phase retrieval.
Physics-Informed Deep Learning
The most recent shift in the field involves moving away from pure hand-coded algorithms toward neural networks that "understand" physics.
In holographic microscopy, a major hurdle is the "twin-image" problem, where the reconstruction process creates a ghost-like artifact. Mazen Mel and the team behind HoloADMM solved this by "unrolling" a physics-based optimization algorithm into a deep neural network HoloADMM: High-Quality Holographic Complex Field Recovery. This allows the system to use the rigor of physics to maintain accuracy while using the pattern-recognition power of CNNs to clean up noise and resolve fine details.
Similarly, frameworks like SelfPhish by Xiaogang Yang use self-supervised learning to recover phase from a single X-ray hologram without needing any prior training data, essentially "learning" the physics of the specific sample during the reconstruction process Self-supervised physics-informed generative networks for phase retrieval from a single X-ray hologram.