TL;DR
This paper establishes a fundamental link between the "completeness of discrete translates" and the uniqueness of phase retrieval in the Short-Time Fourier Transform (STFT). The authors prove that signals with compact support can be uniquely recovered from phaseless lattice samples using a wide variety of window functions—including Gaussians, bandlimited functions, and Hermite functions—moving significantly beyond previous restrictions to Gaussian-only regimes.
What Problem Does This Paper Address?
In many physical applications, such as diffraction imaging or quantum mechanics, sensors can only measure the intensity (magnitude) of a signal's transform, losing the phase information. This is the phase retrieval problem. Specifically, this paper looks at the STFT, where a signal is "windowed" before being Fourier transformed.
The core question is one of uniqueness: If we only have the absolute values of STFT samples taken on a discrete lattice, is that enough to uniquely identify the original signal (up to a trivial global phase)? Previous research showed that for general signals in $L^2(\mathbb{R}^d)$, lattices usually fail to provide uniqueness. However, if we restrict ourselves to signals with compact support (signals that are zero outside a certain range), uniqueness becomes possible. Until now, most proofs for this relied heavily on the specific properties of Gaussian window functions. This paper seeks to generalize those results to a much broader, more practical class of windows.
Methods
The authors bridge two seemingly different areas of mathematics:
- Phase Retrieval: Recovering a function from its transform's magnitude.
- Completeness of Translates: Determining if the set of all shifted versions of a function $g$ (translates) can "span" or approximate any other function in a given space.
The key technical innovation is Theorem 2.1, which provides a general recipe for uniqueness. It states that if the complex exponentials are "complete" on the signal's support difference set ($K-K$), and if the translates of a specific derived window function are "complete" on the signal's support ($K$), then the phaseless samples on a lattice are sufficient for uniqueness.
To prove this for various windows, the authors employ tools from:
- Functional Analysis: Working within Banach spaces and using the Riesz representation theorem.
- Complex Analysis: Utilizing Carlson’s theorem on the uniqueness of entire functions of exponential type to handle windows like Hermite functions.
- Time-Frequency Analysis: Leveraging the structural properties of the STFT and its relation to the ambiguity function.
Datasets
Not discussed in the paper. As a theoretical work in mathematical analysis, the paper relies on formal proofs rather than empirical testing on datasets.
Key Insights and Ideas
- The "Completeness" Bridge: The most significant insight is that the "uniqueness problem" in phase retrieval is essentially a "completeness problem" in disguise. If your window function is "rich" enough that its shifted versions can describe any function on the signal's support, you have enough information to ignore the phase.
- Beyond the Gaussian Regime: The paper breaks the "Gaussian bottleneck." It shows that windows commonly used in physics and engineering—like bandlimited functions or Hermite functions—work just as well for ensuring uniqueness.
- The Airy Disk: The authors specifically highlight the Airy disk (the pattern formed by light passing through a circular aperture). They demonstrate that their theory applies to this physically ubiquitous window, which is critical for applications in optics and astronomy.
- Sparse Sampling: Interestingly, the results suggest that the sampling set doesn't necessarily need to be a dense lattice; in some cases, even "very sparse" irregular sets (with zero point-density) can theoretically provide uniqueness.
Key Results
- General Uniqueness Condition: Theorem 2.1 provides the first general framework linking translate completeness to phaseless STFT sampling.
- Multivariate Generalization: The authors extend previous 1D results to any dimension $d$, proving that multivariate Gaussians on any lattice provide uniqueness for compactly supported signals.
- Bandlimited Windows: Proposition 2.3 proves that if a window is bandlimited, a lattice $U$ provides uniqueness as long as the signal's support fits within the "fundamental domain" of the dual lattice.
- Hermite and $\mathcal{P}_{\alpha,\beta}$ Class: Proposition 2.4 establishes uniqueness for a broad class of functions that includes Hermite functions, provided the sampling lattice is sufficiently dense relative to the window's parameters.
Limitations and Future Work
- Compact Support Requirement: The proofs rely strictly on the signal $f$ having compact support ($L^2(K)$). Extending these results to signals that merely decay quickly (like Schwartz functions) remains an open challenge.
- Stability vs. Uniqueness: This paper focuses on uniqueness (is the solution mathematically certain?). It does not address stability (how much does a small amount of noise in the measurements change the result?). In practical engineering, stability is often more important than theoretical uniqueness.
- Algorithmic Implementation: While the paper proves that a unique solution exists, it does not provide a specific algorithm to reconstruct the signal from the phaseless samples. Developing efficient, convergent algorithms based on these completeness properties is a natural next step.
Summary updated