Wavelet Design with Optimally Localized Ambiguity Function: a Variational Approach

Ron Levie, Efrat Krimer Avraham, Nir A. SochenarXiv 2021 · arXiv

Introduction

The Continuous Wavelet Transform (CWT) is a cornerstone of signal processing, providing a multi-resolution analysis of non-stationary signals by mapping a one-dimensional time-domain signal into a two-dimensional time-scale phase space. However, this mapping is inherently limited by the "blur" introduced by the wavelet's ambiguity function. This paper presents a novel variational framework to design mother wavelets that achieve optimal localization in the time-scale plane, effectively sharpening the "lens" through which we observe signal characteristics.

The Problem: Ambiguity and Blur in Phase Space

In CWT analysis, the ambiguity function serves as a point spread function. It defines how a single point in the ideal phase space is smeared across adjacent time and scale coordinates. The wider this function, the more correlated the resulting wavelet coefficients become, leading to a loss of resolution and reduced sparsity in the representation.

Traditionally, researchers have relied on the Heisenberg uncertainty principle to minimize this spread. However, the authors argue that this approach is often misapplied to wavelets. Unlike the Short-Time Fourier Transform (STFT)—where the mathematical generators of the transform align with the physical observables of time and frequency—the CWT's group structure does not naturally lend itself to standard variance-based optimization. Furthermore, optimizing a 2D ambiguity function directly is computationally prohibitive and involves complex constraints to ensure the resulting function remains a valid wavelet.

The Variational Framework: The Power of the Pull-Back

The core innovation of this work is the application of "wavelet-Plancherel theory" to simplify the optimization problem. The authors utilize a mathematical isomorphism to perform a "pull-back," translating the 2D variances of the ambiguity function in phase space into 1D operations acting directly on the mother wavelet.

Window-Signal Space

The authors define a "window-signal space" ($W \otimes S$), allowing the wavelet to be treated simultaneously as the analyzing window and the signal itself. This dual role is critical for defining an uncertainty functional $\mathcal{L}(\hat{f})$ that represents the total spread in time and scale.

Euler-Lagrange Equations

By formulating the problem using the calculus of variations, the authors derive Euler-Lagrange equations that describe the "optimally sharp" wavelet. This transforms the task from a complex 2D search into a 1D variational problem. The resulting functional $\mathcal{L}(\hat{f})$ is minimized subject to specific constraints: the wavelet must remain centered at zero time and unit scale. These constraints prevent the optimization algorithm from converging to trivial solutions, such as the wavelet "drifting" or scaling infinitely to reduce one variance at the expense of another.

Numerical Optimization and Results

To validate the theory, the authors implemented a constrained gradient descent algorithm. The optimization was initialized with a truncated Gaussian, a common "well-behaved" starting point in signal processing.

The results demonstrate a significant quantitative improvement in localization. Over 8,000 iterations, the uncertainty value dropped from an initial 11.427 to 8.0619. This reduction in the "blurriness" of the phase space suggests that the derived wavelets are significantly more efficient at capturing signal features with minimal overlap. The gradient descent showed smooth, monotonic convergence, indicating that the variational formulation is stable and well-posed.

Discussion and Future Outlook

The "optimally sharp" wavelets produced by this method offer immediate benefits for tasks requiring high-resolution feature extraction and signal compression. By minimizing correlation between coefficients, these wavelets promote sparser representations.

One of the most significant theoretical insights is the identification of the "STFT Accident." The authors clarify that the success of standard uncertainty principles in STFT is a specific consequence of its group structure—a mathematical convenience that does not extend to the affine group of wavelets. This necessitates the more sophisticated observable-based approach developed in this paper.

While the current work focuses on the 1D CWT and finds local minima, the underlying "pull-back" machinery is generic. Future research could extend this variational approach to higher-dimensional transforms such as Shearlets or Curvelets, which are essential for edge detection in image processing. Additionally, the framework could be adapted to design signal-specific wavelets, optimizing localization for particular data classes like seismic or biomedical signals.

Conclusion

This paper provides a rigorous mathematical foundation for designing wavelets with optimal phase-space localization. By shifting the optimization from the 2D ambiguity function to 1D variational operations, the authors provide an efficient and effective path toward sharper, more precise signal analysis.

Summary updated

Abstract

In this paper, we design mother wavelets for the 1D continuous wavelet transform with some optimality properties. An optimal mother wavelet here is one that has an ambiguity function with minimal spread in the continuous coefficient space (also called phase space). Since the ambiguity function is the reproducing kernel of the coefficient space, optimal windows lead to phase space representations which are "optimally sharp." Namely, the wavelet coefficients have minimal correlations with each other. Such a construction also promotes sparsity in phase space. The spread of the ambiguity function is modeled as the sum of variances along the axes in phase space. In order to optimize the mother wavelet directly as a 1D signal, we pull-back the variances, defined on the 2D phase space, to the so called window-signal space. This is done using the recently developed wavelet-Plancharel theory. The approach allows formulating the optimization problem of the 2D ambiguity function as a minimization problem of the 1D mother wavelet. The resulting 1D formulation is more efficient and does not involve complicated constraints on the 2D ambiguity function. We optimize the mother wavelet using gradient descent, which yields a locally optimal mother wavelet.
BibTeX
@misc{levie2021wavelet,
  author = {Ron Levie and Efrat Krimer Avraham and Nir A. Sochen},
  title = {{Wavelet Design with Optimally Localized Ambiguity Function: a Variational Approach}},
  year = {2021},
  eprint = {2104.01654},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.01654},
}
Figure