In signal processing, phase retrieval centers on a simple but difficult question: can we fully reconstruct a signal if we only know the magnitude of its Short-Time Fourier Transform (STFT)? In a purely continuous setting, the answer is often yes. However, practical applications require us to sample the time-frequency plane at discrete points.
Intuitively, one might assume that if we sample densely enough, we can recover the signal perfectly—mirroring the classic Nyquist-Shannon sampling theorem. Yet, recent mathematical breakthroughs have revealed a striking "no-go" result known as the discretization barrier. When sampling on a standard regular lattice, increasing the sampling density does not help; certain signals remain fundamentally indistinguishable from their magnitudes alone.
Fortunately, a series of recent papers has mapped out exactly why this barrier exists and developed clever workarounds to bypass it.
The Paradox of the Discretization Barrier
The core conflict between continuous and discrete phase retrieval was brought to light by Lukas Liehr in Non-uniqueness theory in sampled STFT phase retrieval (arXiv 2022). The paper proved that regardless of the window function chosen or how densely a regular lattice is sampled, there will always exist distinct, square-integrable ($L^2$) signals that produce identical spectrogram samples.
This phenomenon is fundamentally different from standard aliasing. In classical sampling theory, information loss occurs because the sampling rate is too low to capture high-frequency oscillations. In lattice-based phase retrieval, the loss of injectivity is structural. As established in On Foundational Discretization Barriers in STFT Phase Retrieval, the geometry of a regular lattice relies on the intersections of parallel lines in the time-frequency plane. This rigid, periodic structure introduces deep algebraic symmetries. These symmetries act as a blind spot, allowing completely different signals to yield identical absolute values across every single grid point.
Because the ambiguity is baked into the grid's algebraic structure, making the grid finer simply samples the same ambiguous structural patterns more densely. To achieve unique signal recovery, researchers have had to alter the problem's constraints, the measurement tools, or the geometry of the sampling points themselves.
Regular Lattice (Symmetric) Square-Root / Irregular Grid (Symmetry-Breaking)
• • • • • • • • • •
• • • • • • • • • •
• • • • • • • • • •
[Symmetries cause phase ambiguity] [Symmetry broken; uniqueness restored]
Strategy 1: Restricting the Signal Space
If the entire space of square-integrable functions contains ambiguous signal pairs, the most immediate workaround is to restrict the types of signals we attempt to recover.
In Injectivity of Gabor phase retrieval from lattice measurements, Lukas Liehr demonstrated that the discretization barrier can be bypassed by enforcing prior constraints on the signal domain. By limiting the signal class to specific mathematical subspaces—such as compactly supported functions or functions within Gaussian shift-invariant spaces—the problematic, ambiguous signal pairs are effectively excluded from the search space. Under these restricted domains, standard lattice measurements become sufficient to guarantee unique phase reconstruction.
Strategy 2: Deploying Multi-Window Systems
Instead of changing the signal, another approach changes the measurement setup. The fundamental limitation of the discretization barrier assumes the use of a single window function during the STFT process.
To overcome this, Lukas Liehr et al. proposed a multi-window framework in Multi-window STFT phase retrieval: Lattice uniqueness. The authors proved that by capturing discrete spectrogram samples using multiple distinct window functions, the overlapping, diverse perspectives break the algebraic symmetries inherent to a single-window lattice. This multi-window diversity provides a robust mathematical workaround, restoring uniqueness without requiring any constraints on the underlying signal space.
Strategy 3: Redesigning the Sampling Geometry
Perhaps the most elegant solution to the discretization barrier is to eliminate the root cause: the regular lattice itself. If rigid, uniform grids create the symmetries that cause phase ambiguity, then changing the geometry of the sampling points should restore uniqueness.
Researchers have successfully demonstrated this via two distinct geometric modifications:
- Square-Root Lattices: In Phaseless Sampling on Square-Root Lattices, Lukas Liehr proposed abandoning uniform grids in favor of non-linear spacing, where sampling points are distributed according to $\pm\sqrt{n}$. This simple geometric shift completely shatters the parallel-line symmetries of standard lattices, enabling the unique recovery of any square-integrable function using standard windows like Gaussians.
- Perturbed Liouville Sets: Alternatively, one can introduce irregular perturbations to a standard grid. In Phase retrieval in Fock space and perturbation of Liouville sets, the authors looked at phase retrieval within Fock spaces. By applying deterministic or random perturbations to regular lattices—transforming them into what are known as Liouville sets—the authors broke the algebraic obstructions that render standard grids unsolvable. This provided the first concrete construction of finite-density, irregular uniqueness sets for phase retrieval.
Ultimately, these insights show that the "impossible" problem of discrete phase retrieval is highly dependent on structure. By restricting the signal space, utilizing multiple windows, or distorting the sampling grid, researchers have successfully turned a foundational mathematical barrier into a solvable engineering problem.
Key Foundational Work
Non-uniqueness theory in sampled STFT phase retrieval (2022) — contributed 8 concept elements to this direction.
On Foundational Discretization Barriers in STFT Phase Retrieval (2022) — contributed 8 concept elements to this direction.
Injectivity of Gabor phase retrieval from lattice measurements (2022) — contributed 3 concept elements to this direction.