The Shape of Eccentricity: Rapid GW Classification via Wavelet Scattering

Executive Summary

TL;DR: This work proposes a high-speed filtering pipeline using the Wavelet Scattering Transform (WST) and a 1D-CNN to identify eccentric binary black holes in gravitational-wave (GW) data. It achieves an AUC of 0.844, effectively distinguishing eccentricity from spin-precession—a notorious "look-alike" in GW signals—with minimal computational cost.

Background: As the LIGO-Virgo-KAGRA (LVK) network detects hundreds of mergers, identifying the ~10% that are eccentric is vital for proving dynamical formation theories (e.g., in dense star clusters). However, traditional eccentric inference is too slow for every trigger; this WST approach acts as the "high-speed sorter" for the next generation of catalogs.

Problem & Motivation: The High Cost of Eccentricity

Most GW searches assume quasi-circular orbits, as isolation in the field tends to circularize binaries. However, binaries formed via dynamical interactions (field triples or clusters) often retain measurable eccentricity.

The technical bottleneck is twofold:

1. Computational Complexity: Eccentric waveform models (like SEOBNRv5EHM) are significantly more expensive than circular ones.
2. Degeneracy: Amplitude and phase modulations from eccentricity can look remarkably similar to those caused by spin-precession.

Current pipelines need a middle-man—a fast, reliable method to flag "potentially eccentric" signals before committing heavy GPU hours to Bayesian parameter estimation.

Methodology: The Wavelet Scattering Transform (WST)

Instead of feeding raw strain or 2D "Q-scans" into a deep model, the authors use the WST.

Why WST? Standard Fourier Transforms lose time-localization, and Wavelets can be unstable to small time-shifts. WST solves this by applying a hierarchy of wavelet convolutions followed by modulus operations and low-pass filtering.

• Invariance: It is stable to small deformations (time-warping).
• Compression: It reduces the 2048 Hz signal into a set of informative "scattering coefficients" that capture multi-scale dependencies.

The 1D-CNN Architecture

The researchers found that a 1D-CNN operating on these coefficients outperformed both linear Logistic Regression and complex 2D-CNNs. The architecture extracts detector-independent temporal signatures and fuses them to capture coherent signal features across the LIGO/Virgo network.

Figure 1: Comparison of WST coefficients for circular (left) vs. eccentric (right) signals. Note the additional power in higher scattering orders for eccentric systems. (Caption based on paper Section II.B)

Results: Speed Meets Precision

The reference 1D-CNN model (with parameters $J=7,Q=2$) delivered impressive results:

• Performance: ~64% True Positive Rate at a fixed 10% False Alarm Rate.
• Robustness: When tested on purely precessing (but circular) binaries, the false positive rate did not increase. This suggests WST captures unique "fingerprints" of eccentricity that are distinct from precession modulations.
• Dominant Factors: The Network SNR (NSNR) is the primary driver of success. Below an NSNR of 30, classification becomes significantly harder.

Figure 2: ROC curve showing the trade-off between False Alarm Rate and True Positive Rate. The 1D-CNN consistently stays in the upper-left quadrant (AUC 0.844). (Caption based on paper Figure 4)

Deep Insight: Why Not Use Higher Multipoles?

Interestingly, the study found that including higher order multipoles (modes beyond the dominant $(2,2)$ quadrupole) did not significantly boost classification accuracy for the mass range of 10-40 $M_{\odot }$. This implies that for standard-mass stellar black holes, the "eccentricity signature" is primarily encoded in the non-linear amplitude modulations of the primary signal, which the WST is exceptionally good at extracting.

Conclusion & Future Outlook

The WST-CNN framework provides a "lightweight champion" for GW data analysis. It is:

1. Precession-Aware: It doesn't get easily fooled by spinning black holes.
2. Efficient: Trained in ~10 minutes on a standard laptop.
3. Scalable: Ready for the high-cadence detections expected in the O4 and O5 observing runs.

Limitations: The model is currently sensitive to the reference frequency used to define eccentricity (10 Hz). Future iterations may use cycle-based definitions to better handle signals that merger quickly after entering the detector band.

Final Takeaway: This work moves us closer to a "triage" system for Gravitational Wave astronomy, where AI handles the massive influx of data to point human researchers toward the most exotic, eccentric systems in the universe.