TL;DR
This paper investigates the stability of phase retrieval from discrete Gabor measurements, a crucial problem in signal processing. The authors demonstrate that by relaxing the recovery to a "semi-global" phase, they can achieve stable reconstruction for bandlimited signals, with stability constants that grow polynomially with the signal dimension, a significant improvement over previous results.
What Problem Does This Paper Address?
Phase retrieval is the problem of reconstructing a signal from measurements of its magnitude only, discarding its phase information. This is a fundamental challenge in many scientific fields, including coherent diffraction imaging (like ptychography) and audio processing. While phase retrieval is known to be stable for general frames in finite-dimensional spaces, it becomes unstable in infinite-dimensional settings. Even in finite dimensions, using specific frames like the Gabor transform can lead to severe ill-conditioning, where the stability constant (a measure of how much small errors in measurements can amplify into large errors in the reconstructed signal) grows exponentially with the signal dimension. This instability is particularly problematic for signals with small or vanishing magnitudes. Recent work has shown that a "semi-global" phase recovery regime can restore stability in some cases, but this paper aims to provide a more rigorous and general analysis of this regime for discrete Gabor measurements, specifically for bandlimited signals.
Methods
The paper leverages two key reconstruction formulae derived from Gabor measurements:
- Autocorrelation Relation (Lemma 2.1): This relates the Gabor measurements to the autocorrelation of a time-shifted version of the signal.
- Ambiguity Function Relation (Lemma 2.2): This connects the Fourier transform of the Gabor measurements to the ambiguity functions of the signal and the Gabor window.
The authors then build upon these relations and existing uniqueness results (Corollaries 2.3-2.8) to derive stability estimates. A core technique involves analyzing the "essential support graph" of a signal, which captures only the larger magnitude entries of the signal. This allows them to focus on stable phase propagation between these significant entries, effectively bypassing the instability issues associated with small-magnitude components. They develop theorems for both magnitude retrieval (Lemma 3.4) and phase retrieval (Theorem 3.6, Theorem 3.10, Theorem 3.11, Theorem 3.13), considering scenarios with single "temporal islands" (connected components in the essential support graph) and multiple islands, as well as frequency gaps.
Datasets
Not discussed in the paper. The paper focuses on theoretical analysis and stability estimates for signals in $\mathbb{C}^L$, rather than empirical evaluation on specific datasets.
Key Insights and Ideas
- Semi-global phase recovery is key: The paper strongly advocates for and rigorously analyzes a "semi-global" phase recovery approach. This is crucial because it allows for stable reconstruction even when individual phase values are ill-defined or unstable.
- Essential support graph for stability: By focusing on the "essential support graph" (which considers only signal entries above a certain magnitude threshold $\delta$), the authors effectively isolate the parts of the signal where phase retrieval is more robust. This is a clever way to handle signals with small or vanishing entries.
- Polynomial scaling of stability constants: A major contribution is showing that for bandlimited signals and specific Gabor windows, the stability constant scales polynomially (specifically, linearly in $L$ for certain cases) with the signal dimension, a significant improvement over the exponential scaling seen in some previous works.
- Trade-offs between signal and window properties: The paper highlights how the stability depends on both the signal's properties (bandlimited, magnitude distribution) and the window function's ambiguity function. Different window functions lead to different stability guarantees.
- Duality between time and frequency: The authors demonstrate dual results for phase retrieval in the time domain and the frequency domain, suggesting a unified theoretical framework.
Key Results
- Stable magnitude retrieval: Lemma 3.4 establishes stable magnitude retrieval for B-bandlimited signals, with the stability constant depending on the essential support of the signal and the Gabor window.
- Stable phase retrieval for temporal islands: Theorem 3.6 and Theorem 3.10 provide stability estimates for phase retrieval on single and multiple "temporal islands" (connected components of the essential support graph), respectively. The stability constants involve factors related to the size of the essential support and the signal's magnitude.
- General stability for bandlimited signals: Theorem 3.11 extends these results to a more general case using the essential support graph with a time-separation parameter $\Delta$. It shows that stability constants scale polynomially with $L$ and depend on $\Delta$.
- Stability for frequency gaps: Theorem 3.13 provides stability estimates for phase retrieval in the frequency domain, analogous to the time-domain results.
- Polynomial scaling: For specific cases, like the rectangular window, the stability constant is shown to scale linearly with the signal dimension $L$.
Limitations and Future Work
- Bandlimited signals: The derived stability results are primarily for bandlimited signals. Extending these to more general signal classes would be valuable.
- Time-frequency atolls: The paper acknowledges that extending the current approach to handle "time-frequency atolls" (a more general concept of signal structure where phase propagation might be needed in both time and frequency simultaneously) is not straightforward and likely requires different techniques.
- Deterministic measurements: The results are for deterministic measurements, which is a strong assumption. Investigating stability under noisy or probabilistic measurement models could be a future direction.
- Specific window properties: While the paper relaxes some conditions on the window function compared to prior work, the stability still depends on certain properties of the window's ambiguity function. Exploring windows that offer optimal stability guarantees remains an area of interest.
Summary updated