Phase Retrieval from Sampled Gabor Transform Magnitudes: Counterexamples

Rima Alaifari, Matthias WellershoffJournal of Fourier Analysis and Applications 2021 · DOI

TL;DR

This paper demonstrates that, contrary to what might be hoped, you generally cannot uniquely recover a signal (up to a simple phase shift) from just the magnitudes of its sampled Gabor transforms, even with very fine sampling. The authors construct explicit counterexamples showing that different signals can produce identical magnitude measurements on any given sampling lattice.

What Problem Does This Paper Address?

The problem of "phase retrieval" is fundamental in signal processing and many scientific fields. It asks: if you only have access to the magnitude of a signal's transform (like its Fourier transform or, in this case, its Gabor transform), can you recover the original signal? For the standard Fourier transform, phase retrieval is generally impossible. However, for the Gabor transform, which provides a time-frequency representation of a signal, there have been hopes that phase retrieval might be more successful, especially when sampling is involved. Specifically, researchers have investigated whether sampling the magnitude of the Gabor transform on certain discrete sets (lattices) could uniquely determine the signal (up to a global phase factor, which is often acceptable). This paper tackles the question of whether there exists any sampling lattice for the Gabor transform magnitude that guarantees unique signal recovery for all possible signals.

Methods

The authors construct counterexamples by leveraging the relationship between the Gabor transform and the Bargmann transform. They utilize the Hadamard factorization theorem and properties of the Gaussian window function. A key technique involves constructing pairs of functions, denoted $f_+$ and $f_-$, which are designed to have identical Gabor transform magnitudes on specific sampling sets. These functions are built using fractional Fourier transforms and time-shift/modulation operators, allowing them to be rotated and shifted in the time-frequency plane. The core of their argument relies on showing that these constructed functions $f_+$ and $f_-$ do not agree up to a global phase, yet their sampled Gabor transform magnitudes are identical.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Gabor Phase Retrieval is Not Universally Solvable: The most significant insight is that no matter how densely you sample the Gabor transform magnitude, you cannot guarantee unique signal recovery for all signals. This is a strong negative result for the field.
  • Counterexamples are Constructible: The paper provides a constructive method to generate pairs of distinct signals that are indistinguishable based on sampled Gabor transform magnitudes. This is crucial for proving the impossibility result.
  • Time-Frequency Concentration Matters: The constructed counterexamples are well-concentrated in both time and frequency. This suggests that signals with desirable properties (like being localized) are precisely the ones that can lead to ambiguity.
  • Real-Valued Signals are Possible Counterexamples: For rectangular sampling lattices, the authors show that these counterexamples can even be constructed as real-valued functions, making them more practically relevant.

Key Results

The paper's main result is a negative answer to the question of whether there exists a sampling lattice for Gabor transform magnitudes that allows for unique signal recovery (up to global phase) for all $L^2(\mathbb{R})$ functions. They demonstrate that for any lattice $S \subset \mathbb{R}^2$, one can find two distinct functions $f_+$ and $f_-$ such that $|V_\phi f_+(x, \omega)| = |V_\phi f_-(x, \omega)|$ for all $(x, \omega) \in S$, where $V_\phi$ denotes the Gabor transform with a Gaussian window $\phi$. The constructed functions are shown to be well-localized in time and frequency.

Limitations and Future Work

The paper focuses on demonstrating the impossibility of universal Gabor phase retrieval from sampled magnitudes. The authors acknowledge that their construction provides counterexamples for any lattice. Future work could explore specific types of lattices or signal classes for which phase retrieval is possible, building on their earlier work on bandlimited signals. The paper also hints at the generality of their findings by mentioning that similar results have been shown for general short-time Fourier transforms.

Summary updated

Abstract

Abstract We consider the recovery of square-integrable signals from discrete, equidistant samples of their Gabor transform magnitude and show that, in general, signals can not be recovered from such samples. In particular, we show that for any lattice, one can construct functions in $$L^2({\mathbb {R}})$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msup><mml:mi>L</mml:mi><mml:mn>2</mml:mn></mml:msup><mml:mrow><mml:mo>(</mml:mo><mml:mi>R</mml:mi><mml:mo>)</mml:mo></mml:mrow></mml:mrow></mml:math> which do not agree up to global phase but whose Gabor transform magnitudes sampled on the lattice agree. These functions have good concentration in both time and frequency and can be constructed to be real-valued for rectangular lattices.
BibTeX
@article{alaifari2021phase,
  author = {Rima Alaifari and Matthias Wellershoff},
  title = {{Phase Retrieval from Sampled Gabor Transform Magnitudes: Counterexamples}},
  journal = {Journal of Fourier Analysis and Applications},
  year = {2021},
  doi = {10.1007/s00041-021-09901-7},
}
Figure