In many physical sensing systems—from optical imaging to wireless communications—sensors can only measure the intensity (magnitude) of a signal, while the crucial phase information is lost. This "phase retrieval" problem is traditionally ill-posed because many different signals can produce the same intensity pattern. However, a growing body of research demonstrates that by introducing redundancy and interference, we can turn this impossible problem into a solvable engineering task with rigorous guarantees.
The Power of Overlap
The most intuitive way to recover phase is to ensure that no part of the signal is measured in isolation. In the context of the Short-Time Fourier Transform (STFT), Kishore Jaganathan shows in STFT Phase Retrieval: Uniqueness Guarantees and Recovery Algorithms that overlapping windows provide the necessary constraints to resolve ambiguity.
Imagine sliding a window across a signal. If the shift between windows is smaller than the window length, adjacent segments share a "common history." This shared information acts as a bridge: the phase of one segment constrains the phase of the next. As Jaganathan notes in later work, if the overlap is sufficient (typically more than half the window length), almost all signals can be uniquely recovered. The redundancy allows phase information to "propagate" across the entire signal domain.
Filling the Gaps with Multiple Windows
While a single overlapping window is powerful, it can be numerically unstable if the signal has "gaps" or if the window function decays to zero. To solve this, researchers like Yunan Yang propose using a diverse collection of window functions. In Multi-Window Approaches for Direct and Stable STFT Phase Retrieval, Yang demonstrates that by strategically combining different windows—such as Gaussians and Hermite functions—we can "fill in" the time-frequency plane.
This multi-window approach ensures that even if one window is insensitive to a specific part of the signal, another window will capture it. Cheng Cheng and colleagues frame this using graph theory in Phase Retrieval From Multiple-Window Short-Time Fourier Measurements. They show that multiple windows create a "connected graph" of measurements, ensuring that the phase of every signal component is linked to every other component, preventing the signal from breaking into unresolvable, independent pieces.
Turning Phase into Amplitude via Interference
Another fundamental mechanism for phase recovery is interference. If you cannot measure phase directly, you can mix the unknown signal with a known reference. This is the principle behind holography and is explored mathematically by Robert Beinert and Gerlind Plonka in One-Dimensional Discrete-Time Phase Retrieval.
The intuition follows the Law of Cosines. If you know the magnitude of signal $A$, the magnitude of a reference $B$, and the magnitude of their sum $|A+B|$, you can calculate the angle (phase) between them. Yi Li uses this combinatorial trick in Sublinear-Time Algorithms for Compressive Phase Retrieval to achieve incredibly fast reconstruction. Similarly, Marzieh Hasannasab and Beinert show in Phase Retrieval via Polarization in Dynamical Sampling that "polarized" measurements—measuring the sums of different signal states—provide a mathematical bridge to recover the complex-valued signal from intensity data alone.
From Theory to 6G Sensing
These mathematical insights are now enabling new hardware paradigms. In 6G communications, Reconfigurable Intelligent Surfaces (RIS) use thousands of elements, making traditional phase-aware channel estimation prohibitively expensive. Jieao Zhu and co-authors describe a "sensing" approach where multiple sources create an interference field.
By using simple, low-cost power detectors to measure this field, the system can recover the phase information necessary for beamforming without complex radio frequency chains. As Shicong Liu explains in Sensing-Enhanced Channel Estimation for Near-Field XL-MIMO Systems, mixing incoming signals with a controlled reference wave allows for environment mapping and user localization using only power measurements, drastically reducing the cost and energy consumption of next-generation wireless networks.