In many physical imaging systems—from X-ray crystallography to ptychography—we can measure the intensity of a signal's transform but not its phase. The mathematical challenge of recovering the original signal from these magnitude-only measurements is known as phase retrieval.
While we have long known that signals can often be uniquely recovered from their continuous spectrograms, a troubling gap exists between theory and practice. In the digital world, we must sample. Recent research has uncovered a fundamental "discretization barrier": for many standard signal classes, sampling on a regular grid makes unique recovery mathematically impossible, no matter how high the sampling density.
The Failure of Uniform Sampling
In classical signal processing, the Nyquist-Shannon sampling theorem provides a comfort zone: if you sample a band-limited signal densely enough on a regular lattice, you can reconstruct it perfectly. However, Lukas Liehr has demonstrated that this intuition fails for phase retrieval.
In Non-uniqueness theory in sampled STFT phase retrieval, Liehr proves that for the space of square-integrable functions, there is no "Nyquist-like" critical density for phaseless measurements. Even if you sample a spectrogram on an infinitely fine regular lattice, there will always exist distinct signals that produce the exact same magnitude samples.
This is not a limitation of noise or algorithmic power, but a structural property of the sampling geometry. Regular lattices create algebraic symmetries—specifically, the ability to complex-conjugate coefficients within shift-invariant spaces—that allow different signals to "mask" as one another when the phase is discarded. As established in On Foundational Discretization Barriers in STFT Phase Retrieval, this barrier persists even if we restrict ourselves to real-valued signals and windows.
Breaking the Symmetry with Signal Priors
To bypass this barrier, researchers have identified two primary levers: changing what we know about the signal or changing how we sample.
The first approach involves restricting the signal space to exclude the ambiguous "twins" created by lattice sampling. Kishore Jaganathan, Babak Hassibi, and Samet Oymak showed in Sparse Phase Retrieval that sparsity is a powerful tool. If a signal is sparse and lacks periodic structure, it can be uniquely identified from its Fourier magnitude.
Similarly, Lukas Liehr and Irina Shafkulovska proved in From completeness of discrete translates to phaseless sampling of the short-time Fourier transform that signals with compact support (signals that are zero outside a finite interval) can be uniquely recovered from lattice samples using a wide variety of window functions, such as Gaussians or Hermite functions. By forcing the signal to be "small" in the time domain, we break the shift-invariance that the discretization barrier relies on.
The Geometry of Uniqueness
The second, more radical approach is to abandon the regular lattice entirely. If the problem is the rigid, periodic structure of the grid, the solution is to introduce irregularity.
In Phase retrieval in Fock space and perturbation of Liouville sets, Lukas Liehr and Martin Rathmair showed that while regular lattices fail, "Liouville sets"—formed by slightly perturbing the points of a lattice—can guarantee uniqueness. These irregular sampling patterns break the algebraic obstructions that prevent a unique global phase from being identified.
Another geometric solution is the square-root lattice. In Phaseless sampling on square-root lattices, Liehr demonstrates that if sampling points cluster more densely as they move away from the origin (following a square-root distribution), the discretization barrier vanishes. This non-uniform density provides enough "local" information to lock the phase of the signal across the entire domain.
Generalizing the Framework
The implications of these discretization limits extend beyond the standard Fourier transform. Robert Beinert has extended the study of these ambiguities to the Linear Canonical Transform (LCT), a framework that includes Fresnel and fractional Fourier transforms. This suggests that the discretization barrier is a universal feature of phaseless reconstruction in any transform domain.
Furthermore, the need for principled discretization is appearing in deep learning. Md Ashiqur Rahman and Raymond A. Yeh recently addressed similar themes in Group Downsampling with Equivariant Anti-aliasing, noting that without a generalized sampling theorem for discrete group structures, signal reduction leads to permanent information loss.
Summary
The research direction pioneered by Liehr and colleagues shifts our understanding of phase retrieval from an algorithmic problem to a geometric one. We now know that:
- Uniformity is the enemy: Regular grids create symmetries that make phase retrieval inherently ambiguous.
- Density isn't enough: Simply sampling faster on a grid does not resolve these ambiguities.
- Structure is the solution: Uniqueness is restored by either restricting the signal (sparsity, compact support) or by using non-uniform sampling geometries (square-root lattices, perturbed Liouville sets).