Abstract
We prove that there exists no window function $g \in L^2(\mathbb{R})$ and no lattice $\mathcal{L} \subset \mathbb{R}^2$ such that every $f \in L^2(\mathbb{R})$ is determined up to a global phase by spectrogram samples $|V_g f(\mathcal{L})|$, where $V_g f$ denotes the short-time Fourier transform of $f$ with respect to $g$. Consequently, the forward operator mapping a square-integrable function to its spectrogram samples on a lattice is never injective on the quotient space identifying functions that agree up to a multiplicative constant of modulus one. This result highlights that in the discretization of the STFT phase retrieval problem from lattice measurements, a prior restriction of the underlying signal space to a proper subspace of $L^2(\mathbb{R})$ is inevitable.
Introduction
Phase retrieval is the challenge of recovering a signal from the magnitude of its transform, having lost the phase information. This problem is ubiquitous in applications such as ptychography, X-ray crystallography, and audio processing. In these fields, the Short-Time Fourier Transform (STFT) is a primary tool, providing a "spectrogram"—a time-frequency representation of the signal.
While it is well-established that a signal can be uniquely recovered (up to a global phase factor) from its continuous spectrogram, practical implementations rely on discrete samples. Most often, these samples are taken on a lattice (a regular grid in the time-frequency plane). This paper addresses a fundamental theoretical question: Does there exist a "perfect" combination of a window function and a lattice that allows for the unique identification of any signal in the space of square-integrable functions ($L^2(\mathbb{R})$)?
The Discretization Barrier
The core contribution of this work is a definitive "no-go" theorem for discrete phase retrieval. The authors establish that the transition from a continuous domain to a discrete lattice represents a fundamental loss of injectivity. No matter how dense the lattice is, there will always be "blind spots" where different signals appear identical to the spectrogram.
Methodology: Constructing Ambiguities
The authors employ a sophisticated blend of Time-Frequency Analysis, Fractional Fourier Transforms (FrFT), and the theory of Shift-Invariant Spaces to prove this barrier. The key technique involves constructing pairs of non-equivalent signals—signals that are not just global phase-shifted versions of each other—that yield identical spectrogram samples.
By using the FrFT, the researchers relate STFT measurements on rotated lines in the time-frequency plane to standard shifts and modulations. They define specific shift-invariant spaces associated with fractional Fourier shifts and, by carefully choosing coefficients, build signals $f$ and $h$ that produce identical magnitudes on sets of parallel lines—and consequently, on any lattice contained within those lines.
Key Results and Insights
The Geometry of Parallel Lines
The non-uniqueness is deeply tied to the geometry of parallel lines in the time-frequency plane. Since lattices are essentially intersections of such lines, the ambiguities inherent to parallel line sampling propagate to all lattice-based sampling. Theorem 1.3 formalizes this by showing that uniqueness fails for any set of shifted parallel lines.
Real-Valued Signals
One might hope that restricting signals to be real-valued would solve the ambiguity. However, Theorems 1.4 and 3.13 show that for most common lattices (such as rectangular ones or those with rational dependencies), non-uniqueness persists even for real signals and real window functions.
Inevitable Subspace Restriction
The most significant implication for signal processing is that to make phase retrieval work in practice, we must give up on recovering every possible $L^2$ signal. We must assume the signal belongs to a "proper subspace," such as functions with compact support or specific bandwidth constraints. The set of signals producing these ambiguities forms an infinite-dimensional "cone," suggesting that non-uniqueness is a pervasive feature rather than a rare edge case.
Conclusion and Future Work
This paper proves that for any window $g \in L^2(\mathbb{R})$ and any lattice $\mathcal{L}$, the forward operator $f \mapsto |V_g f(\mathcal{L})|$ is never injective. While this establishes a foundational barrier, several questions remain open, such as the behavior of real-valued signals on hexagonal lattices and whether uniqueness fails for any discrete set with a minimum distance (separated sets). Future research will also investigate "deficiency"—the number of additional points required to potentially achieve uniqueness on a given lattice.
Summary updated