Mitigation of Incoherent Spectral Lines via Adaptive Coherence Analysis for Continuous Gravitational-Wave Searches

This paper presents an unsupervised adaptive network coherence analysis framework that effectively mitigates non-Gaussian spectral artifacts in continuous gravitational-wave searches using Advanced LIGO O3 data, significantly reducing spectral lines while preserving astrophysical signal integrity and detector sensitivity.

The Gist

The Great Gravitational Wave "Noise-Canceling" Headset

Imagine you are trying to hear a single, very quiet violin playing a specific note in a massive, crowded concert hall. This violin represents a Continuous Gravitational Wave (CW)—a signal from a spinning neutron star that has been playing for years.

Now, imagine the concert hall is filled with thousands of people. Some are whispering (the natural background noise of the universe), but many are making specific, annoying, repetitive noises: a dripping faucet, a humming refrigerator, a squeaky door, and someone tapping a rhythm on a table. These are the spectral lines or "noise artifacts" in the data collected by LIGO (the gravitational wave detectors).

Because the violin is so quiet, these annoying noises drown it out. In fact, the "dripping faucet" is so loud that the computer searching for the violin thinks the faucet is the violin. This is why, despite years of listening, we haven't found these continuous signals yet.

The Problem:
Traditionally, when scientists heard a "dripping faucet" (a known noise line), they would just throw away that entire section of the recording. If the faucet was at 60 Hz, they would delete everything between 59 Hz and 61 Hz.

• The Flaw: What if the real violin was playing a note right at 60 Hz? By deleting the whole band, you might accidentally throw away the very signal you are looking for. It's like silencing the whole room just to stop one person humming, potentially missing the singer in the corner.

The Solution: The "Adaptive Coherence" Framework
The paper by Ye Zhou and Karl Wette introduces a smarter way to clean the data. Instead of throwing away whole sections of the recording, they built a smart, unsupervised filter that acts like a super-intelligent sound engineer.

Here is how it works, using simple analogies:

1. The "Two-Ear" Test (Coherence Analysis)

The most important trick is that LIGO has two detectors: one in Hanford, Washington (H1) and one in Livingston, Louisiana (L1). They are thousands of miles apart.

• Real Signals (The Violin): If a real gravitational wave hits Earth, it will vibrate both detectors almost exactly the same way, at the exact same time.
• Fake Signals (The Faucet): A local noise source (like a power line hum or a vibrating machine) will usually only affect one detector. It won't be heard in the other one.

The new pipeline acts like a bouncer checking IDs at two different doors. If a noise appears in Hanford but not in Livingston, the bouncer says, "That's just local noise; let's fix it." If the noise appears in both perfectly synchronized, the bouncer says, "That might be a real signal; leave it alone."

2. The "Shape-Shifter" Classification

Not all noises look the same. Some are sharp and thin (like a laser pointer); others are wide and fuzzy (like a foggy window).

• The pipeline looks at the shape of the noise.
• Ultra-narrow lines: These are usually very stable electronic hums (like the 60 Hz power line). The system checks if they are in both detectors with extreme precision.
• Wide lines: These might be mechanical vibrations. The system checks if they overlap in both detectors.

It's like sorting a pile of leaves. Some leaves are perfectly identical twins (real signals or global environmental noise), and some are just random, unique leaves (local glitches). The system sorts them out based on how "twin-like" they are.

3. The "Surgical Scalpel" (Mitigation)

This is the magic part. Old methods were like using a sledgehammer to remove a weed; they cut out the whole patch of grass.

• The New Method: This pipeline uses a surgical scalpel.
• When it identifies a "weed" (a local noise line) in one detector, it doesn't delete the data. Instead, it gently rescales the volume of that specific noise down to the level of the background whisper.
• Imagine a person shouting in a library. Instead of kicking them out of the building (deleting the data), the system gently lowers their voice until they are just whispering like everyone else. The library (the data) remains intact, but the shouting stops.

The Results: A Cleaner Concert Hall

The authors tested this on data from the LIGO "O3" run. Here is what they found:

• High Success Rate: They successfully identified and "quieted" 89% of the annoying noises in the Hanford detector and 77% in the Livingston detector.
• Minimal Damage: They only had to touch less than 7% of the total frequency range. This means 93% of the data was left completely untouched.
• Safety: Crucially, they proved that they didn't accidentally mute any real "violins." The signals that looked like they belonged in both detectors (the real candidates) were left exactly as they were.

Why This Matters

Think of this framework as a noise-canceling headset for the entire universe.

• Before this, scientists had to ignore huge chunks of the radio spectrum because they were too noisy.
• Now, they can listen to those "noisy" parts again, knowing the local glitches have been surgically removed.
• This opens up new "real estate" in the search for gravitational waves. It increases the chances of finally hearing that elusive, continuous violin note from a spinning neutron star, which could tell us secrets about the inside of stars that we can never learn any other way.

In short: They built a smart filter that knows the difference between a local glitch and a cosmic signal, allowing them to clean the data without throwing away the treasure.

Technical Summary

1. Problem Statement

Continuous Gravitational Wave (CW) searches aim to detect persistent, quasi-sinusoidal signals from rapidly rotating neutron stars. However, the sensitivity of these searches is severely limited by non-Gaussian spectral artefacts (narrow-band "lines") present in detector data.

• The Challenge: Unlike transient signals, CW searches integrate data over months or years. Even low-amplitude, stationary instrumental or environmental lines accumulate significant coherent power over these long baselines, masquerading as astrophysical signals.
• Limitations of Current Methods:
  • Traditional Vetoing: Existing methods often discard entire frequency bands where known lines exist. This "frequency masking" destroys any potential astrophysical signal that might overlap with the line.
  • Machine Learning: While deep learning shows promise, it often relies on training priors that may fail when encountering "out-of-distribution" artefacts in evolving detector states.
  • Linear Subtraction: Feed-forward loops using auxiliary sensors struggle with nonlinear and transient noise couplings.
• Goal: Develop a robust, unsupervised framework that selectively suppresses local, incoherent noise while preserving global, coherent signals (both astrophysical and global environmental lines) without discarding large portions of the frequency spectrum.

2. Methodology

The authors propose a generalised, unsupervised pipeline based on adaptive network coherence analysis. The framework exploits the physical principle that astrophysical signals and global environmental noise appear coherently across multiple detectors (e.g., LIGO Hanford and Livingston), whereas local instrumental artefacts are typically incoherent (present in only one detector).

The pipeline consists of four primary stages:

A. Robust Baseline Fitting and Normalisation

• Input: Short Fourier Transforms (SFTs) of GW data (typically 1800s segments).
• Process: A robust iterative algorithm fits the noise floor (Power Spectral Density) using a Chebyshev polynomial (degree 10) in log-frequency–log-PSD space.
• Technique: It employs a running-median filter and iterative peak rejection (masking windows around peaks) to ensure the baseline fit is unbiased by spectral lines. This separates the fundamental noise floor from the artefacts.

B. Individual Line Detection

• Peak Identification: Peaks are identified based on amplitude (2.5x local baseline) and prominence (10% above surroundings).
• Boundary Determination: Boundaries are defined where spectral power decays to 1.05x the baseline.
• Logarithmic Clustering: Adjacent peaks are merged on a logarithmic frequency scale to treat broad, complex structures as single physical entities, accounting for frequency-dependent resolution.

C. Adaptive Network Coherence Analysis

This is the core innovation. Detected lines are classified into four morphological categories based on bandwidth ($W$), and coherence criteria are adaptively applied to each class:

1. Very Narrow ($W < 5$ mHz): Strict absolute frequency tolerance ($\approx 0.28$ mHz) but relaxed amplitude tolerance (up to 90% difference). Targets ultra-stable global lines.
2. Narrow ($5$ mHz $\le W < 2$ Hz): Relative frequency tolerance of 0.1%.
3. Medium ($2$ Hz $\le W < 10$ Hz): Relative frequency tolerance of 0.5% + 5% boundary overlap.
4. Wide ($W \ge 10$ Hz): 2% frequency tolerance + 10% boundary overlap.

• Logic: Lines appearing in at least two detectors are flagged as coherent and preserved. Lines appearing in only one detector are flagged as incoherent and targeted for mitigation.

D. Spectral Artefact Mitigation

• Strategy: Instead of zeroing out data or replacing it with Gaussian noise (which destroys phase information), the pipeline employs frequency-domain substitution.
• Implementation: For incoherent lines, the complex SFT coefficients are rescaled so that the Power Spectral Density (PSD) matches the fitted noise baseline:
  $$\tilde{x}_{cleaned} = \tilde{x}_{original} \sqrt{\frac{S_{baseline}}{S_{original}}}$$
• Result: This surgically suppresses excess power to the Gaussian noise floor while preserving the original phase information, ensuring that any potential CW signal within the band is not degraded by artificial incoherence.

3. Key Contributions

• Unsupervised Adaptive Framework: A novel pipeline that does not require training data or auxiliary sensors, relying instead on the physical coherence between detectors.
• Morphological Classification: The introduction of adaptive coherence criteria based on line width (narrow vs. wide) significantly improves the accuracy of distinguishing between stable global lines and wandering local noise.
• Phase-Preserving Mitigation: A rescaling technique that reduces noise power without introducing spectral discontinuities or destroying phase coherence, a critical improvement over simple masking or Gaussian replacement.
• High Selectivity: The method modifies less than 7% of the analysis bandwidth (20–2000 Hz) while removing the majority of spectral lines.

4. Results (Validation on LIGO O3 Data)

The pipeline was tested on Advanced LIGO O3 data using 1-day, 3-day, and 5-day integration subsets.

• Line Removal Efficiency:
  • For the 5-day dataset, the pipeline identified and mitigated 89% of spectral lines in the Hanford (H1) detector and 77% in the Livingston (L1) detector.
  • It successfully preserved 7 coherent features in each detector (e.g., 60 Hz mains harmonics, 510 Hz violin modes).
• Spectral Selectivity:
  • The cleaned bandwidth fraction was 5.79% (H1) and 6.87% (L1), demonstrating that the method targets specific artefacts without broad data loss.
  • As integration time increased (1 to 5 days), the cleaned bandwidth fraction decreased, indicating higher precision in targeting stationary lines as spectral resolution improved.
• Morphological Validation:
  • Retained (Coherent): Showed sharp, discrete peaks at known frequencies (60 Hz, 510 Hz) and narrow bandwidths, consistent with global resonances.
  • Removed (Incoherent): Formed a broad, quasi-continuous background with wider bandwidths, consistent with local mechanical/electronic noise.
• Statistical Verification (Q-Q Analysis):
  • Coherent Signals: The F-statistic distribution for coherent bins remained invariant ($r \approx 1$), confirming that potential astrophysical signals were preserved with high fidelity.
  • Incoherent Noise: The tail slope of the F-statistic distribution for incoherent bins dropped significantly ($s \approx 0.69$), confirming effective suppression of non-Gaussian outliers without creating artificial nulls.
• Comparison to O3 Lines List: The pipeline found a significant fraction of lines listed in the official O3 catalogue but also identified many transient or short-duration lines missed by long-term averaging, while avoiding the over-masking of the entire frequency band.

5. Significance and Outlook

• Recovering Sensitivity: By removing the "spectral forest" of incoherent lines, the framework recovers detector sensitivity in parameter spaces previously discarded by traditional veto methods. This allows for the detection of weaker CW signals that were previously obscured.
• Robustness: The method is robust against the non-stationary evolution of detector noise and does not suffer from the "out-of-distribution" failures common in supervised machine learning.
• Future Applications:
  • The authors propose using this pipeline as a pre-processing stage for deep learning-based denoising algorithms, simplifying the feature space for neural networks.
  • It is particularly beneficial for Hidden Markov Model (HMM) tracking methods, which are highly susceptible to wandering instrumental lines.
  • Future work will involve "mock data challenges" with injected simulated CW signals to rigorously prove that the pipeline does not falsely veto real signals and can enhance recovery rates.

In conclusion, this paper presents a critical advancement in GW data analysis, offering a precise, unsupervised tool to clean the "spectral forest," thereby significantly enhancing the prospects for the first detection of continuous gravitational waves.