Impact of noise transients on gravitational-wave burst detection efficiency of the BayesWave pipeline with multidetector networks

Detection confidence of the source-agnostic gravitational-wave burst search pipeline BayesWave is quantified by the log signal-versus-glitch Bayes factor, $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$. A recent study shows that $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$ increases with the number of detectors. However, the increasing frequency of non-Gaussian noise transients (glitches) in expanded detector networks is not accounted for in the study. Glitches can mimic or mask burst signals resulting in false alarm detections, consequently reducing detection confidence. This paper presents an empirical study on the impact of false alarms on the overall performance of BayesWave, with expanded detector networks. The noise background of BayesWave for the Hanford-Livingston (HL, two-detector) and Hanford-Livingston-Virgo (HLV, three-detector) networks are measured using a set of nonastrophysical background triggers from the first half of Advanced LIGO and Advanced Virgo’s Third Observing Run (O3a). Efficiency curves are constructed by combining $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$ of simulated binary black hole signals with the background measurements, to characterize BayesWaves’s detection efficiency as a function of the per-trigger false alarm probability. The HL and HLV network efficiency curves are shown to be similar. A separate analysis finds that detection significance of O3 gravitational-wave candidates as measured by BayesWave are also comparable for the HL and HLV networks. Consistent results from the two independent analyses suggests that the overall burst detection performance of BayesWave does not improve with the addition of Virgo at O3a sensitivity, because the increased false alarm probability offsets the advantage of higher $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$.

Figure 1

Log signal-to-glitch Bayes factor $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$ versus network signal-to-noise ratio ${\mathrm{SNR}}_{\mathrm{net}}$ for IS1. The blue circles (orange stars) correspond to HL (HLV) network injections; each data point corresponds to a single injection. Gaussian-noise-like events with $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}=-500$ are not shown.

Figure 2

$\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$ versus ${\mathrm{SNR}}_{\mathrm{net}}$ for IS2 off-source injections of GW190512_180714 (pink), GW190408_181802 (green) and GW190412 (purple). The vertical dashed lines in the respective colors at ${\mathrm{SNR}}_{\mathrm{net}}=12.2$, 15.3, 18.9 indicate the median HLV network match-filter SNRs of the GW events (from Table 1). The circles and stars correspond to HL and HLV injections respectively. Gaussian-noise-like events with $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}=-500$ are not shown.

Figure 3

Background measurements for the BW algorithm. The blue (orange) curve corresponds to the HL (HLV) background measured using the downselected O3a background triggers described in Sec. 3a. The shaded bands show the $1\text{−}\sigma$ Poisson uncertainty regions for each network in corresponding colors.

Figure 4

Worked example: computing one representative point on the efficiency curve for a significance threshold $P_{\mathrm{FA}}=0.2$. Top: histogram of $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$ for the O3a background triggers described in Sec. 3a. Bottom: histogram of $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$ for IS1. The HL and HLV network histograms are color-coded blue and orange respectively. In both panels, the vertical dashed lines at $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}=18.1$ (HL) and 34.2 (HLV) indicates the threshold for $P_{\mathrm{FA}}=0.2$. The fraction of injections to the right of the thresholds in the bottom panel yields $P_{\det }=0.74$ (HL) and 0.71 (HLV).

Figure 5

BW efficiency curves constructed using IS1 for the HL (blue) and HLV (orange) networks. The shaded bands with matching colors are the $1\text{−}\sigma$ Poisson uncertainty regions for $P_{\mathrm{FA}}$, same as in Fig. 3. The region where $P_{\mathrm{FA}}\le 0.4$ is shaded green to indicate astrophysical relevance.

Figure 6

Log signal-to-glitch Bayes factor, $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$ of the HLV network versus the HL network for IS1. The color bar shows ${\mathrm{SNR}}_{\mathrm{net}}$ in HLV for each injection. The diagonal line indicates equal $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}$ for both networks. The dashed lines at $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}(\mathrm{HL})=18.1$ and $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}(\mathrm{HLV})=34.2$ indicate the thresholds for $P_{\mathrm{FA}}\le 0.2$ with the respective networks. Gaussian-noise-like events with $\ln {\mathcal{B}}_{\mathcal{S},\mathcal{G}}=-500$ are excluded in this plot.

Figure 7

$P_{\mathrm{FA}}$ of the HLV network versus the HL network for O3 GW events in IS2. Each point represents a single GW event as shown in the legend and is color coded by the HLV ${\mathrm{SNR}}_{\mathrm{net}}$. The $P_{\mathrm{FA}}$ and ${\mathrm{SNR}}_{\mathrm{net}}$ shown are the medians of the off-source injections of the corresponding event; the horizontal (vertical) gray bars span the interquartile range of the HL (HLV) $P_{\mathrm{FA}}$ measurements. The diagonal line indicates equal $P_{\mathrm{FA}}$ for both networks.

Figure 8

Cumulative number of events versus per-trigger false alarm probability, $P_{\mathrm{FA}}$ for the BW background measurements of the HL (left, 1008 triggers) and HLV (right, 1134 triggers) network.