Wide parameter-space O3 search for continuous gravitational waves from unknown neutron stars in binary systems

"Continuous gravitational waves, i.e., persistent and nearly-monochromatic signals emitted by asymmetric spinning neutron stars, remain elusive. Searches for these signals from unknown binary systems are the most computationally challenging, but they are essential, given that binary accretion provides a natural mechanism for creating the required asymmetry, and around half of the known pulsars rotating above 25 Hz are part of a binary system. Here we report on a search of a large uncharted parameter-space region: for the first time we cover gravitational-wave frequencies above 520 Hz (from 50 to 1 000 Hz), and, for the first time with advanced detectors, orbital periods lower than 3 days are explored. No signal is detected, and we set the most stringent constraints to date on the amplitude of signals of this kind. Our results exclude with 95% confidence neutron stars within 100 pc and rotating faster than ∼ 495 Hz from having ellipticities above 5.2 × 10−8. Within the same distance our results also exclude r-mode amplitudes above 1.5 × 10−6 for stars rotating faster than ∼ 740 Hz."

"In this paper we present a search using the public O3 data from the Advanced LIGO detectors (Abbott et al. 2023). Our search covers signal frequencies up to 1 000 Hz (see figure 1), almost doubling the maximum frequency previously investigated by a search of this kind. Allowing for the spin frequency of the objects to be high is important due to the recycling scenario, where neutron stars gain angular momentum due to the accretion of matter from their companion (Patruno & Watts 2020). Furthermore, some studies suggest that gravitational-wave emission could help to explain the observed spins of the galactic pulsar population (Gittins & Andersson 2019), which is particularly important for neutron stars with high rotational frequencies like the ones targeted by this search."

"We use the BinarySkyHouF (Covas & Prix 2022a) semi-coherent search method with Nseg = 31 675 segments of length Tseg = 900 s. This search method uses templates with Porb = 0 at the coherent stage, and calculates the FAB;l detection statistic of Covas & Prix (2022b) with l = 1,..., Nseg over a grid in f0 and sky. We take a linear combination (with weights wl) of the FAB;l, each tracking the time-frequency pattern produced by a signal. The template grid spacings for the initial stage search are shown in table 2. We exclude from the analysis the segments within the lowest 25th percentile of the wl. This strategy results in a sensitivity loss smaller than ∼ 5%, while reducing the computational cost of the search by ∼ 25%."

"We follow-up these selected candidates by searching a volume of parameter space around each of them with a coherent time Tseg = 2 700 s, three times longer than that of the original search. On these reduced volumes we can leverage a nested sampling algorithm (Ashton et al. 2019) to calculate the detection statistic (Covas et al. 2024). We now explicitly search the f1, e, and ω ranges given in table 1 due to the increased resolution."

"To distinguish potential gravitational-wave signals from noise, we require the evolution of a candidate’s detection statistic to be consistent with the behavior of an astrophysical signal. This expected behavior is measured on simulated test-signals added to the O3 data (see appendix B for their distribution). In particular we consider two aspects: the consistent increase in significance s1 (s0) and in the robust detection statistic log10 B̂1;S/GLtL (s0) (Keitel 2016) across stages 0 and 1. Based on the results from the simulated signals we define acceptance/rejection regions in the s0 − s1 and s0 − log10 B̂1;S/GLtL planes for the candidates. Our criteria has a false dismissal probability of about 11/7 200 ≈ 0.15% from 11 missed simulated signals and results in 15 candidates above both thresholds. 12 of these candidates can be associated with fake signals present in the data for validation purposes (the so-called “hardware injections”, see Biwer et al. 2017). The name “hardware injection” stems from these signals being produced by moving the detectors’ mirrors, rather than being added in software to the calibrated data. There are no hardware injection signals from the orbital parameter-space that we are investigating, but there are signals at the frequencies that we search from isolated neutron stars. Even though their waveforms do not match our binary ones, these signals are so loud, that our searches still identify them as unlikely to be due to noise. In particular 9 of our surviving candidates are related to the hardware injection at ∼ 52.8 Hz and 3 candidates to the one at ∼ 848.9 Hz. The other 3 surviving candidates (at 507.9 Hz, 517.2 Hz, and 998.3 Hz) can be associated with non-Gaussian disturbances, since they accumulate more than 80% of their signal-to-noise ratio from very few segments. We conclude that none of these candidates can be associated with a defensible CW signal."

"We estimate the 95% confidence upper limits on the gravitational-wave amplitude h0 95% in every 0.1 Hz band, which is the amplitude such that 95% of a population of signals with frequency in that band and with the same distribution for the other parameters as the simulated test-signals (see appendix B) would have been detected. The detection criterion for a simulated signal is that it generates a cluster with a higher significance than the 3rd cluster in its 0.1 Hz band and its orbital parameter-space region."

"The estimated upper limits are shown in the left plot of figure 5. The lowest upper limit is h0 95% = 2.63×10−25 at f0 = 211.1 Hz. These results are available in machinereadable format at Covas et al. (2026). Because the sensitivity depth is estimated based on detection efficiency studies carried out in noise not obviously affected by disturbances, upper limits estimates in disturbed bands are not trustworthy. We identify disturbed bands based on the expected value of the maximum significance s0. When this exceeds the mean by 5 standard deviations, we assume that the band is disturbed and do not place an upper limit in it. Out of 9 500 0.1 Hz bands we identify 1 159 as disturbed."

"If the CWs are sourced by a mass asymmetry, the gravitational-wave amplitude is given by (Bonazzola & Gourgoulhon 1996) ε Izz h0 ≃ 10−26 (f0 / 100 Hz) (d / 1 kpc)−1 (Izz / 10 38 kg m 2)−1. On the other hand, if the CWs are sourced by r-modes, the gravitational-wave amplitude is given by (Owen 2010) h0 ≃ 10−26 α cubed (M / 1.4 M⊙)−1 (R / 11.7 km)−3 (f0 / 100 Hz) (d / 1 kpc)−1."

"The upper right plot of figure 5 shows the ellipticity upper limits for two distances and two values of the moment of inertia. The tightest constraint for the ellipticity is achieved at f0 = 989.9 Hz, which for sources at 100 pc with Izz = 10 38 kg m squared is ε < 5.2 × 10−8. If instead we assume Izz = 3 × 10 38 kg m squared, the upper limits are more stringent, as shown by the dashed traces. For example, for sources within 100 pc at 989.9 Hz, we find ε < 1.7 × 10−8, which is well within the predicted maximum ellipticities of neutron stars, that range between 10−8 and 10−5 (Rossetto et al. 2025; Gittins & Andersson 2021). The lower right plot shows the r-mode amplitude upper limits for two different distances and two mass-radius combinations. The tightest constraint for the r-mode amplitude is achieved at f0 = 993.3 Hz, which for sources at 100 pc with M = 1.4 M⊙ and R = 11.7 km is α < 1.5 × 10−6. If instead we assume M = 1.9 M⊙ and R = 13 km, our constraint becomes more stringent: α < 8.1 × 10−7."

"In this paper we present the most extensive search to date for CWs from unknown neutron stars in binary systems, with gravitational-wave frequencies between 50 and 1 000 Hz and orbital periods between 0.2 and 150 days. The main objective of this search is to cover the widest possible parameter-space region with a certain computational budget. For this reason we use coarser template grids and a shorter coherent timebaseline compared to our other recent search (Covas et al. 2025), whose main objective was to achieve the best possible sensitivity. We use the BinarySkyHouF pipeline, which is the most efficient and sensitive pipeline to carry out this type of search. We do not detect any astrophysical signal. The main search has taken ∼ 10.15×10 6 CPU corehours to complete, running on a combination of AMD EPYC 7402 CPUs and NVIDIA A100 GPUs. The computational cost of the follow-up is negligible."

"Previous searches such as Abbott et al. (2021); Abac et al. (2025) are ≈ 20% more sensitive than this search but cover a much smaller parameter-space. Specifically this search reaches frequencies ∼ 3 times larger, which is particularly difficult because the computational cost scales ∝ f0 5. Additionally this search covers large regions of the orbital parameter-space, some of which have not been searched before. Using the “breadth” B to quantify the extent of the investigated parameter space, from equation (75) of Wette (2023), we find that our parameter space is about 4 orders of magnitude larger than ever investigated with Advanced LIGO data."

"Vast regions of parameter space remain unexplored. This search demonstrates what state-of the art search techniques can achieve in the challenging high-frequency range and may be used as blueprint for investigating new and more sensitive datasets."

Based on the technical methodologies described in this paper, here are specific improvements for AI systems, particularly those involved in signal processing, high-dimensional optimization, and large-scale scientific computing:

1. (5/5) - This is a highly specialized application.

Proposed AI Improvement Specific Implementation Mechanism Enhanced System Capability

:---:---:---

Implement a multi-stage, semi-coherent to coherent hierarchical search architecture for high-dimensional parameter spaces. Transition from a low-resolution initial stage (using short segments and coarse grids) to a high-resolution follow-up stage (using nested sampling and longer coherent baselines) triggered by significance thresholds. The AI can perform exhaustive searches across massive, computationally prohibitive parameter spaces (e.g., combining frequency, sky position, and orbital dynamics) without requiring immediate full-scale coherence, drastically reducing initial compute costs while maintaining high sensitivity.

Develop a Significance-Based Clustering algorithm for anomaly detection in non-Gaussian noise environments. Integrate a clustering procedure that groups candidate signals based on spatial/parameter proximity and filters them using a ranking statistic derived from the expected mean and variance of Gaussian noise. The system can autonomously distinguish between true astrophysical/physical anomalies and transient glitches or non-Gaussian disturbances by evaluating the consistency of signal accumulation across temporal segments.

Integrate Hardware-Injection Aware Vetoing/Validation modules in training loops. Incorporate known synthetic signals (hardware injections) into the validation dataset to train a classifier that distinguishes between real physical phenomena and instrumental artifacts (e.g., power mains or detector mirror movements). The AI can achieve extremely low false-alarm rates in dirty data environments by learning to recognize the specific signature of instrumental noise versus genuine signal evolution.

Implement Adaptive Grid Spacing based on frequency-dependent complexity. Dynamically adjust the density of the search template grid (e.g., increasing density as a function of frequency or orbital period) to optimize the trade-off between mismatch (error) and computational overhead. The AI can optimize its own resource allocation, focusing high-density computation only where parameter sensitivity is highest, preventing over-computation in low-sensitivity regions.

Develop Robustness Estimators for non-ideal signal models (Phase Coherence Protections). Implement a verification layer that tests the robustness of detected signals against deviations from ideal models (e.g., timing noise, glitches, or spin-wandering) using shorter coherent timeframes. The AI system becomes resilient to model mismatch, allowing it to detect real-world signals that do not perfectly conform to theoretical mathematical templates (e.g., non-stationary or stochastic signals).

Continuous gravitational waves, i.e., persistent and nearly-monochromatic signals emitted by asymmetric spinning neutron stars, remain elusive. Searches for these signals from unknown binary systems are the most computationally challenging, but they are essential, given that binary accretion provides a natural mechanism for creating the required asymmetry, and around half of the known pulsars rotating above 25 Hz are part of a binary system. Here we report on a search of a large uncharted parameter-space region: for the first time we cover gravitational-wave frequencies above 520 Hz (from 50 to 1000 Hz), and, for the first time with advanced detectors, orbital periods lower than 3 days are explored. No signal is detected, and we set the most stringent constraints to date on the amplitude of signals of this kind. Our results exclude with 95% confidence neutron stars within 100 pc and rotating faster than about 495 Hz from having ellipticities above 5.2 times 10-8. Within the same distance our results also exclude r-mode amplitudes above 1.5 times 10-6 for stars rotating faster than about 740 Hz.