⚛️ general relativity

The WDM Time-Frequency Transform in Gravitational-Wave Data Analysis I: Formalism

This paper provides a self-contained formalism and practical guide for gravitational-wave analysts on applying the Wilson-Daubechies-Meyer (WDM) time-frequency transform, detailing its mathematical foundations, implementation nuances, and statistical framework for analyzing slowly-varying noise.

Original authors: Aaron Johnson, Katerina Chatziioannou, Jake Summers

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aaron Johnson, Katerina Chatziioannou, Jake Summers

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a specific instrument in a chaotic orchestra. The music is the "signal" (like a gravitational wave from colliding black holes), and the chaotic crowd noise is the "noise." To hear the instrument, you need a way to separate the sound not just by what note is being played (frequency), but also by when it is being played (time).

This paper introduces a new, highly efficient tool for doing exactly that, called the Wilson-Daubechies-Meyer (WDM) Transform. Here is how it works, using simple analogies:

1. The Problem: The "Too Fast" vs. "Too Slow" Dilemma

Scientists usually analyze data in two ways:

The Frequency Domain (The "Whole Song" approach): You look at the entire recording and ask, "What notes are in this song?" It's fast and efficient, but you lose track of when the notes happened. It's like knowing the recipe for a cake but not knowing the order in which the ingredients were mixed.
The Time Domain (The "Every Second" approach): You look at the sound wave second by second. You know exactly when things happen, but it's computationally heavy and messy to find specific patterns in the noise.

Time-Frequency Analysis tries to find a middle ground: a grid that shows you both when and what note is playing. However, a famous mathematical rule (the Balian-Low theorem) says you can't have a perfect grid that is both perfectly sharp in time and perfectly sharp in frequency without running into mathematical glitches.

2. The Solution: The WDM "Smart Grid"

The authors propose the WDM transform as a clever workaround. Imagine the time-frequency plane as a floor you need to tile perfectly with tiles (atoms) to cover the whole area without gaps or overlaps.

The Old Way (Gabor): You try to use square tiles. But to make them fit perfectly without gaps, they end up being blurry on the edges.
The WDM Way: Instead of using complex, blurry tiles, the authors use a special set of real, smooth tiles (based on sine and cosine waves) that are shifted slightly off-center.
- The "Meyer" Filter: Think of this as the shape of the tile. It's a smooth, rounded shape that fits perfectly with its neighbors. When one tile fades out, the next one fades in, so the total "power" of the music is never lost or double-counted.
- The "Wilson" Trick: By using real numbers (sines and cosines) instead of complex numbers, they can pack these tiles tightly together (critical sampling) without breaking the mathematical rules.

3. How It Works in Practice

The paper details how to take a raw sound recording and turn it into this grid of tiles:

The Grid: They divide time into chunks and frequency into bands.
The Tiles: Each "cell" in the grid holds a number (a coefficient) representing how much of the signal exists in that specific time and frequency.
The Magic of Real Numbers: In standard math, you often need complex numbers (imaginary numbers) to handle phase (timing shifts). The WDM transform is clever because it keeps everything in real numbers.
- Analogy: If you have a pure tone (a single note), a standard method might store it as one complex number (a vector with direction and length). The WDM method stores it as two real numbers in two neighboring time slots. One neighbor might hold the "cosine" part, and the next holds the "sine" part. Together, they perfectly reconstruct the note and its timing without needing "imaginary" math.

4. Why This Matters for Gravitational Waves

Gravitational waves are like a "chirp"—a sound that changes pitch very quickly as two black holes spiral into each other.

Sparse Representation: Because the WDM tiles are so well-localized, a chirp signal only lights up a few specific tiles on the grid. It's like a spotlight hitting a few specific spots on a stage, while the rest of the stage (the noise) remains dark.
Efficiency: The math allows scientists to calculate this transformation very quickly (using a method similar to the Fast Fourier Transform, or FFT), making it practical for real-time analysis.
Handling Noise: The paper explains how to calculate the "noise" in this new grid. Even if the background noise isn't perfectly steady (it changes over time), the WDM method provides a way to weigh the data correctly so the signal stands out.

5. Key Takeaways from the Paper

No Information Lost: The transformation is "lossless." You can convert the data from time to this grid and back again perfectly. The number of data points stays the same.
Edge Cases: The paper carefully explains what happens at the very beginning (DC) and very end (Nyquist) of the frequency range. These are like the "corners" of the room where the tiles are half-sized, and the math adjusts for that automatically.
A New Standard: The authors argue that while time-frequency methods are becoming popular, they haven't been well-documented. This paper provides the "instruction manual" for the WDM transform, showing exactly how to build it, how to reverse it, and how to use it to find signals in noisy data.

In short, the WDM transform is a mathematical tiling system that lets scientists look at gravitational wave data with a "microscope" that is sharp in both time and frequency, allowing them to spot the faint "chirps" of colliding black holes against the loud noise of the universe.

Technical Summary: The WDM Time-Frequency Transform in Gravitational-Wave Data Analysis I: Formalism

Problem Statement
Gravitational-wave (GW) data analysis traditionally relies on the frequency domain for stationary noise due to its computational efficiency and the diagonal nature of the noise covariance matrix. However, for slowly-varying or non-stationary noise, time-frequency methods offer a necessary middle ground between the efficiency of the frequency domain and the generality of the time domain. Despite their growing adoption, these methods remain less documented and less familiar in GW literature compared to standard Fourier techniques. Existing treatments of specific time-frequency bases often leave critical derivations, discretization details, edge effects, and covariance normalizations implicit. There is a need for a self-contained, reproducible account of a specific basis that balances orthogonality, critical sampling, and good time-frequency localization without violating the Balian-Low theorem.

Methodology
This paper provides a rigorous formalism for the Wilson-Daubechies-Meyer (WDM) transform, a time-frequency representation designed for GW data analysis. The methodology is built upon three theoretical pillars:

Wilson Bases: Replacing the complex exponentials of Gabor systems with real trigonometric functions (sine and cosine) on a half-shifted lattice to evade the Balian-Low theorem while maintaining critical sampling and orthonormality.
Daubechies Construction: Utilizing a window function that, when generating a Gabor system at twice the critical sampling density, yields an equivalent orthonormal Wilson basis with rapid time-domain decay.
Meyer Scaling Function: Employing a smooth, compactly supported window in the frequency domain (the Meyer function) to ensure the basis functions have compact frequency support, reducing false alarms in excess power searches.

The paper details the mathematical construction of the WDM basis, including the generalized Meyer filter function characterized by parameters $(\alpha, \beta, d)$ that satisfy a partition-of-unity (POU) property. It derives the forward and inverse transforms for discrete data, explicitly handling the discretization from the time domain to the frequency domain and finally to the WDM domain. The authors derive the transformation of the noise covariance matrix, addressing the specific structure of the WDM covariance (block tri-diagonal with parity-modulated Toeplitz blocks) under both generic and stationary noise assumptions. Finally, the paper derives the Gaussian likelihood function and the noise-weighted inner product in the WDM domain, demonstrating their equivalence to time- and frequency-domain formulations.

Key Contributions

Self-Contained Derivation: The paper explicitly documents derivations previously left implicit, including the discretization of the transform, the handling of DC and Nyquist edge bands, and the normalization of the noise covariance matrix.
Edge Effects: It clarifies the behavior of edge frequency bins (DC and Nyquist), showing that half the cells in these bands vanish identically due to the properties of the sine/cosine basis, preserving the total degrees of freedom ( $N$ ) matching the time-domain samples.
Covariance Structure: The authors derive the structure of the WDM noise covariance matrix, showing it is block tri-diagonal (due to filter overlap) and possesses a parity-modulated Toeplitz structure. They provide explicit formulas for transforming between the frequency-domain Power Spectral Density (PSD) and the WDM covariance matrix.
Likelihood Formulation: The paper derives the log-likelihood and inner product in the WDM domain, showing that for stationary noise, the WDM likelihood reduces to a form analogous to the Whittle likelihood but with an evolutionary PSD $S(t, f)$ that accounts for time dependence.
Phase Encoding: It explains how phase information, which is encoded in the complex argument of Fourier coefficients, is instead encoded in the relative amplitudes of adjacent real WDM coefficients with opposite parity.

Results

Information Preservation: The WDM transform is proven to be a complete, orthonormal mapping that preserves all information ( $N$ degrees of freedom) without loss, satisfying Parseval's identity.
Computational Efficiency: The transform requires $O(N \log N)$ operations, comparable to the Fast Fourier Transform (FFT), making it computationally feasible for large datasets.
Localization: The basis functions exhibit good localization in both time and frequency, with the Meyer window ensuring compact frequency support.
Covariance Properties: Under the assumption of slowly varying noise, the WDM covariance matrix becomes approximately diagonal, with diagonal elements proportional to the PSD evaluated at the bin midpoint. The paper notes that off-diagonal terms exist even for stationary noise due to filter overlap but can be neglected under specific "locally flat" noise assumptions.
Likelihood Equivalence: The derived WDM likelihood is shown to be mathematically equivalent to the time-domain and frequency-domain likelihoods, provided the noise covariance is treated self-consistently.

Significance
The paper positions the WDM transform as a practical and theoretically sound tool for analyzing gravitational-wave data, particularly in scenarios involving non-stationary or slowly varying noise where standard frequency-domain methods may be suboptimal. By providing a reproducible formalism, the authors aim to facilitate the implementation of WDM-based analyses in GW pipelines. The work serves as the first part of a series, establishing the mathematical foundation and handling stationary noise, while promising to address evolutionary noise and physical examples in a companion paper. The significance lies in demystifying the WDM transform, clarifying its relationship to familiar Fourier and wavelet concepts, and providing the necessary tools for analysts to implement it for inference and signal detection.

1. The Problem: The "Too Fast" vs. "Too Slow" Dilemma

2. The Solution: The WDM "Smart Grid"

3. How It Works in Practice

4. Why This Matters for Gravitational Waves

5. Key Takeaways from the Paper

More like this