The Stability Gap in Discrete Phase Retrieval

This direction is merged into another direction: phase retrieval, discrete phase retrieval ambiguity. Its last version stays readable and citable here.

Version 1 of 134 papers51 conceptsfrozen 2026-10-04

In many physical sensing applications—from X-ray crystallography to audio processing—we can measure the intensity of a signal but not its phase. The challenge of reconstructing the original signal from these magnitude-only measurements is known as phase retrieval.

While researchers have long sought "Nyquist-like" theorems to determine how many samples are needed for a unique reconstruction, recent work has revealed a troubling disconnect: proving that a signal is mathematically unique does not mean it is practically recoverable. This "stability gap" suggests that in discrete settings, the very measurements that guarantee uniqueness often leave the reconstruction process dangerously sensitive to noise.

The Failure of Regular Lattices

For decades, the Short-Time Fourier Transform (STFT) has been a workhorse for phase retrieval. By sliding a window across a signal and taking the Fourier transform of the pieces, we create a spectrogram. Intuitively, if we sample this spectrogram densely enough on a regular grid (a lattice), we should be able to recover the signal.

However, Lukas Liehr demonstrated a fundamental "discretization barrier." In Non-uniqueness theory in sampled STFT phase retrieval, Liehr proved that no matter how densely you sample a spectrogram on a lattice, there will always be distinct signals that produce identical samples. This means regular lattices are fundamentally "blind" to certain phase differences, shattering the hope for a simple sampling density threshold for uniqueness.

Fragility and the Density of Counterexamples

Even when we find specific signals that can be uniquely recovered from lattice samples—such as the standard Gaussian—the victory is often hollow. Rima Alaifari, Francesca Bartolucci, and Matthias Wellershoff have explored the "fragility" of this uniqueness.

In Uncovering the limits of uniqueness in sampled Gabor phase retrieval, they proved that the set of "bad" signals (those that cannot be uniquely recovered) is dense in the space of all possible signals. This means that for any signal you wish to recover, there is a non-unique counterexample arbitrarily close to it.

This proximity creates a nightmare for numerical stability. As noted in On the connection between uniqueness from samples and stability in Gabor phase retrieval, a signal might be theoretically unique, but if it sits right next to a "bad" signal, a tiny amount of measurement noise can make the two indistinguishable. The reconstruction algorithm, unable to tell them apart, may produce a result that is wildly different from the truth.

Breaking Symmetry with Irregularity

If regular lattices fail, the solution may lie in breaking the symmetry of the sampling pattern. Lukas Liehr and Martin Rathmair have shown that moving from regular grids to irregular sets can restore uniqueness.

In Phase retrieval in Fock space and perturbation of Liouville sets, they demonstrated that by slightly perturbing the points of a lattice—specifically using "Liouville sets"—one can achieve uniqueness even with a finite sampling density. Similarly, Liehr showed in Phaseless sampling on square-root lattices that "square-root lattices," where points cluster more densely as they move away from the origin, provide a path to unique recovery that standard lattices cannot.

The intuition here is that irregular sampling prevents the "aliasing" of phases that occurs on a grid. By sampling at "awkward" intervals, we force the signal to reveal its phase in a way that a symmetric grid allows it to hide.

The Persistence of Instability

Despite these breakthroughs in achieving uniqueness, the problem of stability remains largely unsolved. A recurring theme across the research of Fanny Yang, Volker Pohl, and Rima Alaifari is that mathematical injectivity (uniqueness) does not imply a well-posed inverse problem.

In Phase Retrieval In The General Setting Of Continuous Frames For Banach Spaces, Alaifari proved that phase retrieval is fundamentally unstable in infinite-dimensional spaces. Even with highly redundant measurements, the mapping from magnitudes to signals cannot be made robustly continuous.

This suggests that the future of the field lies not just in finding new sets of uniqueness, but in developing "weakly stable" or "locally stable" frameworks. As explored by Tal Amir and Nadav Dym in The stability of generalized phase retrieval problem over compact groups, we may need to restrict our signals to specific structures (like those found in neural networks or sparse sets) to achieve the bi-Lipschitz stability required for real-world, noisy environments.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_the_stability_gap_in_discrete_phase_retr_40221bc1,
  title = {The Stability Gap in Discrete Phase Retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-stability-gap-in-discrete-phase-retrieval/rdr_40221bc1381d40f285c1b11eb1ba4a60},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_the_stability_gap_in_discrete_phase_retr_40221bc1_v1,
  title = {The Stability Gap in Discrete Phase Retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-stability-gap-in-discrete-phase-retrieval/rdr_40221bc1381d40f285c1b11eb1ba4a60/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}