Source-linked AI summary
Deterministic denoising of long-duration gravitational-wave signal candidates with $\alpha$-(de)blending
Przemysław Figura, Michał Bejger
TL;DR
All-sky continuous-wave searches must handle weak, long-duration signals across unknown source parameters, and use shorter segments followed by coincidence analysis. This paper applies deterministic α-(de)blending to time-domain F-statistic candidate patterns; it generally recovers the correct declination pattern around ρ ≳ 4, with recovery dependent on morphology.
Problem
All-sky searches must detect weak, long-duration continuous waves across unknown sky positions and frequency parameters using computationally demanding searches and semi-coherent coincidence across shorter segments.
Source: §1 · p. 1Method
The paper post-processes time-domain F-statistic candidate images with deterministic α-(de)blending to recover signal structure and assess consistency with the assumed signal model.
Source: §1 · p. 2, §3 · p. 8, §3.1 · p. 9Results
The method reliably clusters around the correct declination for ρ ≳ 4 except in specific cases, and recovery near that threshold depends primarily on signal morphology.
Source: §5 · p. 18Takeaways & Limitations
The method may serve as a consistency or veto tool for candidates before the semi-coherent coincidence stage.
Source: §1 · p. 2The study assumes complete data segments, although real detector gaps can reduce SNR and make matched-filter recovery suboptimal.
Source: §2.3 · p. 7
Abstract
from arXiv · showhide
We present a deterministic machine-learning method for denoising candidate signals from all-sky searches for continuous gravitational waves from rotating neutron stars. Using the time-domain $F$-statistic search pipeline, we post-process the resulting $F(f,\dot{f})$ candidate patterns with iterative $\alpha$-(de)blending, a deterministic diffusion-type generative model implemented as a U-Net, well suited to the non-Gaussian, correlated noise of the $F$-statistic output. Two models, trained on data based on simulated 6- and 12-day time-domain segments with software-injected signals similar to the LVK hardware injections, are tested by comparing deblended images to the library of expected signal patterns via the Structural Similarity Index Measure. In general the method recovers the correct sky-position-dependent pattern for signal-to-noise ratios $\rho \gtrsim 4$, moderately below typical all-sky detection thresholds, with recovery depending strongly on pattern morphology, demonstrating that deterministic diffusion-based denoising may serve as a consistency/veto tool ahead of the semi-coherent coincidence stage. We also discuss limitations and possible improvements of the method.
1. Introduction
Continuous gravitational waves are weaker than transient signals but persist for observing-time-scale durations, requiring long data segments for detection. Because all-sky searches cover unknown sky positions and frequency parameters, this work post-processes time-domain F-statistic candidates with deterministic α-(de)blending before semi-coherent coincidence analysis.
1. Introduction
- Continuous gravitational waves have smaller amplitudes than detected transient signals but persist for durations comparable with observing time, so detection requires long data segments.The section contrasts this long-duration signal class with targeted, directed, and all-sky search strategies.Source: §1 · p. 1
- All-sky searches cover unknown sky positions and broad rotational-parameter ranges, making them computationally demanding and motivating shorter segments followed by coincidence searches.The coincidence stage compares candidates across segments in a semi-coherent search.Source: §1 · p. 2
- The paper presents deterministic α-(de)blending to recover or veto continuous-wave patterns in coherent time-domain F-statistic output before semi-coherent coincidence analysis.The stated aim is to increase candidate significance or falsify/veto a candidate for the next search step.Source: §1 · p. 2
2. Time domain F-statistic search
The time-domain F-statistic is used as a coherent matched-filter search output for continuous-wave candidates, represented as frequency–spindown patterns whose morphology depends on source parameters and sky position. The study prepares these patterns from simulated, hardware-injection-based signals and Gaussian noise in 6- and 12-day segments, while assuming complete data segments.
2.1. The method
- The F-statistic is a matched-filter detection statistic that analytically maximizes over amplitude parameters while searching frequency, spindown, and sky position.For signal-free data, 2F follows a central χ² distribution with four degrees of freedom; with a signal, it follows a non-central χ² distribution whose noncentrality is ρ².Source: §2.1 · pp. 3–4
2.2. F-statistic pattern dependence on source parameters and sky position
- For multi-sidereal-day segments, the F-statistic pattern is independent of initial phase and right ascension, but depends on declination and orientation.The section states the symmetry F(δ, ι, ψ) = F(−δ, ι + π, −ψ); intermediate inclinations can flip the pattern and reduce its value.Source: §2.2 · p. 4
2.3. Data preparation
- The simulations use six software injections based on LVK hardware-injection parameters, added to Gaussian noise and analyzed in 6- and 12-day segments.The injection amplitudes are rescaled to match expected hardware-injection SNRs, and the study evaluates 0.5-Hz bands.Source: §2.3 · pp. 6–7, Table 1 · p. 7
- The data preparation assumes complete segments, although real detector gaps can reduce signal-to-noise ratio and make matched-filter recovery suboptimal.The effect of gaps is identified as important but outside the study’s scope.Source: §2.3 · p. 7
- F(f, ˙f) images are centered on the signal-only maximum and normalized to [0, 1], with 108×108 pixels for 6-day data and 216×216 pixels for 12-day data.The noisy maximum is not used for centering because noise alters the F-statistic distribution; the noise retains spatial correlations while its pixel values remain compatible with χ²₄.Source: §2.3 · pp. 7–8
3. Noise removal with machine learning
The section presents iterative α-(de)blending as a deterministic denoising approach for structured, correlated F-statistic noise, then describes its U-Net implementation and evaluation. The method estimates a starting blend level for each candidate and uses image similarity to assess recovered patterns.
3.1. Machine learning model description
- Iterative α-(de)blending is chosen over probabilistic diffusion to handle χ2 noise and correlated image points while remaining computationally straightforward and stable.It builds a deterministic trajectory by linearly interpolating source- and target-distribution samples while a neural network predicts the path’s local tangent.Source: §3.1 · p. 9
- α-blending linearly interpolates samples from two distributions, and α-deblending recovers a sample pair capable of producing a given blended sample.For a random blended sample, the possible contributing subsets merge across the source and target distributions.Source: §3.1 · p. 9
- The network predicts the average direction of small α steps; chaining these predictions averages out randomness and yields a deterministic mapping between distributions.Training uses random sample pairs and blend levels, with the network learning to predict the segment between paired samples.Source: §3.1 · pp. 9–10
3.2. Neural network implementation
- Models A and B use a U-Net with maxpool preprocessing to map both image resolutions to 108×108 inputs, and its spatial scales reflect the F-statistic peak spacing.The U-Net includes an attention layer at 18×18; inference starts at an α chosen to weight signal and noise images, with larger signal-to-noise ratio corresponding to larger α.Source: §3.2 · p. 11, Figure 4 · p. 11
- The starting α estimate uses the noise maximum, the signal’s central peak, and a scaling factor b estimated from 4,000 noise instances.For real data, the text proposes generating noise instances using the detector’s amplitude spectral density to estimate b.Source: §3.2 · p. 12
- The evaluation uses SSIM, which compares image luminance, contrast, and structure and ranges from −1 to 1, with 1 indicating a perfect match.Unlike LPIPS, SSIM is not human-perception biased.Source: §3.2 · p. 12
- High-SNR software injections P05 and P10 require fewer deblending steps than lower-SNR injections.Figure 5 compares step-count distributions for Models A and B; dashed lines mark their averages.Source: Figure 5 · p. 13
4. Results
Noise-corrected SSIM profiles support recovery of most tested signal patterns, but P02 and P03 remain difficult cases. Recovery generally improves with longer segments, while morphology and signal strength affect the outcome.
4. Results
- Higher-SNR injections P05 and P10 required fewer than the maximum 1,000 deblending steps.The number of steps depended on injection amplitude.Source: §4 · p. 13
- Robustness tests compared deblended nonblank images with signal templates using SSIM, and corrected signal-plus-noise distributions by subtracting the noise-only distribution.Clustering around the highlighted true declination indicates correct deblending; the template library contains no exact image for each injection.Source: §4 · pp. 13–14
- Noise-only deblending produced roughly flat SSIM-versus-declination profiles and about 20% blank outputs, while signal-image nonblank fractions could reach 100% without correct identification.The noise-only blank fraction was consistent with the blank fraction in training templates.Source: §4 · pp. 14–15, Figure 7 · p. 15, Table 2 · p. 15
- After noise-background correction, SSIM profiles matched pure-signal patterns for the remaining injections, while P02 and P03 stayed roughly flat near zero for both models.The figure compares signal-plus-noise and pure-signal distributions; larger SSIM values indicate greater similarity.Source: Figure 6 · p. 14, §4 · p. 16
- Longer time segments provided higher SSIM values and greater contrast between peak and lowest distribution values in most cases.This pattern was reported for both model distributions and pure-signal distributions.Source: Figure 8 · p. 15, §4 · p. 16
- P03 performance was substandard because its near-zero cos ι produces many comparably saturated peaks, making its morphology nearly out of distribution for the template bank.The closest training templates match its declination but use ι = 0 or π, unlike the injection’s ι ≈ π/2.Source: §4 · p. 16, Figure 10 · p. 17
- Increasing SNR sharpened ground-truth declination clustering for P10-like signals, but P02’s solitary-peak pattern still failed to recover the correct declination.The solitary peak may be confused with matched-filter noise artifacts or parts of patterns at other declinations.Source: §4 · pp. 16–17, Figure 11 · p. 17
5. Discussion and conclusions
The TDFstat output supports deterministic generative denoising despite its correlated, non-Gaussian noise, and deblending-step counts track injection SNR. Recovery is generally reliable for ρ ≳4 but depends on signal morphology and does not improve uniformly with longer segments.
5. Discussion and conclusions
- TDFstat all-sky F(f, ˙f) images suit deterministic generative denoising because they retain spatial correlations while their pixel values remain consistent with χ2_4 noise.This supports the IADB approach over probabilistic diffusion models relying on Gaussian-noise assumptions.Source: §5 · p. 18
- High-SNR injections converge in fewer deblending steps than weak injections, so step count tracks signal strength.The authors propose exploring step count alongside similarity metrics as a joint significance proxy.Source: §5 · pp. 18–19
- For ρ ≳4, SSIM comparisons generally cluster around the correct declination, but equal-SNR injections show that recovery also depends on morphology.At ρ ≈4, only P00 and P11 of four injections were correctly recovered; weaker signals and pure noise yield flat SSIM distributions.Source: §5 · p. 18
- Recovery has morphology-dependent boundaries: the P02 single-peak pattern is confused with noise artifacts, while P03 has a distinct image associated with its cos ι value.The authors identify these cases as limits to extrapolation and propose richer parameter coverage, architectural changes, or additional conditioning.Source: §5 · p. 18
- Model A recovered P05 more completely than Model B, with non-blank fractions of 1.00 versus 0.79 despite its shorter 6-day segments.The comparison shows that 12-day coherence does not uniformly improve recovery; factors beyond duration and amplitude may matter.Source: §5 · p. 18
- The method is best suited to directed searches or all-sky outlier follow-up because evaluating every frequency-spindown point is computationally prohibitive.Future work includes sequential analysis across segments, testing realistic detector gaps and frequency-dependent noise, and broadening the six-pulsar study.Source: §5 · pp. 18–19
Appendix A. F-statistic dependency on angle parameters
For segments spanning multiple sidereal days, the F-statistic pattern loses dependence on right ascension and starting phase. A simultaneous declination and inclination transformation also leaves the signal and its F-statistic invariant.
Appendix A. F-statistic dependency on angle parameters
- For multiple sidereal days, Earth’s rotation averages out right-ascension dependence, making signals at fixed declination but different right ascensions phase-shifted versions with equivalent patterns.The pattern is also independent of arbitrary starting phase.Source: Appendix A · p. 19
- The signal is h(t) = F+(t)h+(t) + F×(t)h×(t), combining the detector antenna patterns with two polarization waveforms independent of declination.The appendix defines F+ and F× using a and b, whose declination dependences are respectively even and odd.Source: Appendix A · pp. 19–20
- Under δ → −δ and ι → ι + π, with ψ → −ψ, h(t) and consequently the signal’s F-statistic remain invariant.The antenna-pattern transformations preserve F+ while transforming F× as shown in the derivation.Source: Appendix A · p. 20
Appendix B. Neural network training procedure
The training procedure uses separate noise and template-signal image sets for the two models, incorporates SSIM into the loss, and selects checkpoints by validation loss. Training and validation losses stabilize during the 500-epoch runs.
Appendix B. Neural network training procedure
- Model A and Model B use 52,388 and 48,956 noise images, respectively, alongside 356 signal and 89 blank template images each.Blank images make up 20% of each 445-image template set.Source: Appendix B · p. 20
- The training loss weights root mean square error using SSIM, which is used to compare deblended and template images.The described weight ranges from 0 for identical images to 1 for opposite images.Source: Appendix B · pp. 20–21
- Both models were trained for 500 epochs; inference used the lowest-validation-loss checkpoints at epochs 483 and 470 for Models A and B.Validation used 6,000 noise images unseen during training, and the loss evolution is shown for both models in Figure B1.Source: Appendix B · p. 21, Figure B1 · p. 21