Efficient Search for Detection Candidates Using a Peak Finder Strategy for All-Sky-All-Frequency Gravitational Wave Radiometer

TL;DR

The paper tackles the challenge of correlated sky samples in the all-sky-all-frequency (ASAF) GW radiometer search by introducing a Peak Finder that clusters correlated candidates into representative peaks. It develops the Peak SNR statistic and demonstrates via Monte Carlo simulations that this method can significantly reduce trial factors and false dismissals, especially at lower frequencies, while preserving detection sensitivity. The results show improved candidate selection for follow-up analyses and provide a framework to adapt follow-up strategies based on frequency and injection strength. The approach promises practical gains in computational efficiency and follow-up effectiveness for persistent GW searches across broad frequency bands.

Abstract

The first all-sky-all-frequency (ASAF) radiometer search was conducted using data from the first three observing runs of the Advanced LIGO and Advanced Virgo detectors. The significance of this search lies in its fast and unmodeled approach, leveraging a cross-correlation technique to identify common signals across the detector network. As a result, this method serves as an excellent alternative to search for unknown or poorly modeled continuous wave sources and narrowband components of the gravitational wave (GW) background. For continuous wave sources whose waveform can be modeled, this method can serve as the first stage of a hierarchical scheme by identifying sub-threshold candidates to be followed up with more optimal but computationally expensive searches. The ASAF search, however, presently suffers from beam smearing, where multiple candidates may arise due to the same noise fluctuations, detector artifact, or a GW source. This can reduce the detection probability in follow-up analyzes, especially with limited computing resources. To mitigate this issue and reduce the number of correlated and unnecessary candidates, we introduce a novel Peak Finder algorithm. This algorithm helps identifying the most representative candidates while preserving detection sensitivity, thereby allowing follow up of a much larger number of independent candidates. The reduction in correlated samples leads to a significant reduction in False Dismissal Rate (FDR) using the Peak Finder method compared to the Full-sky method. For instance, following up 2 Peak Finder candidates at 30 Hz reduces FDR by a factor of 3.

Figures (7)

• Figure 1: Simulated SNR sky maps at two frequencies, 30 Hz and 200 Hz, with and without a source. The maps illustrate the point spread function of pixels, showing the spatial distribution of SNR values and the effect of source injection on the observed signal. The source is injected with an SNR of 3 at the sky coordinates (RA, Dec) = $(6.2\,{\rm h}, 45^\circ)$, marked by the red star. The yellow triangles highlight the "peaks" (representative of correlated pixels) identified using the peak finder algorithm introduced in Sec. \ref{['sec:peak_finder']}.
• Figure 2: The flow chart illustrates the Peak Finder algorithm for a single frequency bin. For more details, refer to the main text in Sec. \ref{['sec:peak_finder']}.
• Figure 3: Mean number of candidates as a function of SNR threshold. Plots are provided for both the All-sky (with dashed line) and Peak Finder (with solid line) methods across a range of frequencies (30-700 Hz). We note that there is a significant reduction in the number of candidates at lower frequencies with latter method.
• Figure 4: Number of peaks for different frequencies with the mean and standard deviation calculated over $10^4$ maps
• Figure 5: Left panel: The distribution of Peak SNR for noise-only (blue dashed) and noise+injection source (orange solid). Middle panel: Variation of FAR (solid) and FDR (dashed) as a function of the threshold on the Peak SNR for a set of frequencies. The error bars depict $1\sigma$ Poisson fluctuations in FAR and FDR. Right panel: Frequency-direction-averaged ROC curve for a set of injected SNR values. The threshold decreases from left to right. The error bars represent $1\sigma$ fluctuations across seven frequency bins and four sky directions. See main text in Sec. \ref{['sec:results2']} for details.