Online Pipelines

A number of search pipelines run in a low latency, online mode. These can be divided into two groups, modeled and unmodeled. The modeled (CBC) searches specifically look for signals from compact binary mergers of neutron stars and black holes (BNS, NSBH, and BBH systems). The unmodeled (Burst) searches on the other hand, are capable of detecting signals from a wide variety of astrophysical sources in addition to compact binary mergers: core-collapse of massive stars, magnetar star-quakes, and more speculative sources such as intersecting cosmic strings or as-yet unknown GW sources.

Why does LIGO/Virgo/KAGRA operate multiple search pipelines?

Several analysis pipelines contribute to each type of search. Searching with multiple pipelines has advantages and disadvantages compared to operating a single search of each type.On the practical and operational sides, greater redundancy results in higher overall coverage and reliability in low latency, though with the disadvantage of higher overall resource requirements. In addition, studies with simulated signals indicate that combining the results of two or more pipelines searching for the same signal type can increase the expected number of detections over any single pipeline.

False alarm rate and significance

Each search produces a set of candidate events time-stamped at or close to the estimated peak of GW strain amplitude. For binary merger candidates, this would be the time of merger.

Each candidate event is assigned a ranking statistic value by the search pipeline that produced it: higher statistic values correspond to a higher probability of astrophysical (signal), as opposed to terrestrial (noise) origin. The statistical significance of a candidate produced by a given pipeline is quantified by its false alarm rate. This is the expected number of events of noise origin produced by the pipeline with a higher ranking statistic than the candidate, per unit of time searched. Since each search pipeline has an independent method of generating and ranking events, and of estimating the noise background, the false alarm rates assigned for events in the same superevent will in general be different. For an alert to be sent automatically, we require at least one event to have a false alarm rate below the alert threshold.

GstLAL, MBTA, PyCBC Live and SPIIR are matched-filtering based analysis pipelines that rapidly identify compact binary merger events, with 1 minute latencies. They use discrete banks of waveform templates to cover the target parameter space of compact binaries, with all pipelines covering the mass ranges corresponding to BNS, NSBH, and BBH systems.

A coincident analysis is performed by all pipelines, where candidate events are extracted separately from each detector via matched-filtering and later combined across detectors. SPIIR extracts candidates from each detector via matched-filtering and looks for coherent responses from the other detectors to provide source localization. Of the four pipelines, GstLAL and MBTA use several banks of matched filters to cover the detectors bandwidth, i.e., the templates are split across multiple frequency bands. All pipelines also implement different kinds of signal-based vetoes to reject instrumental transients that cause large SNR values but can otherwise be easily distinguished from compact binary coalescence signals.

GstLAL [1] [2] is a matched-filter pipeline designed to find gravitational waves from compact binaries in low-latency. It uses a likelihood ratio, which increases monotonically with signal probability, to rank candidates, and then uses Monte Carlo sampling methods to estimate the distribution of likelihood-ratios in noise. This distribution can then be used to compute a FAR and a p-value. The template bank being used by GstLAL in O4 is described in [3].

MBTA [6] [7] constructs its background by making every possible coincidence from single detector triggers over a few hours of recent data. It then folds in the probability of a pair of triggers passing the time coincidence test.

PyCBC Live [8] [9] estimates the noise background by performing time-shifted analyses using triggers from a few hours of recent data. Single-detector triggers from one detector are time shifted by every possible multiple of 100 ms, thus any resulting coincidence must be unphysical given the 10 ms light travel time between detectors. All such coincidences are recorded and assigned a ranking statistic. The false alarm rate is then estimated by counting accidental coincidences ranked higher than a given candidate, i.e., with a higher statistic value. When three detectors are observing at the time of a particular candidate, the most significant double coincidence is selected, and its false alarm rate is modified to take into account the data from the remaining detector.

SPIIR [4] [5] applies summed parallel infinite impulse response (IIR) filters to approximate matched-filtering results. It selects high-SNR events from each detector and finds coherent responses from other detectors. It constructs a background statistical distribution by time-shifting detector data one hundred times over a week to evaluate foreground candidate significance.

cWB [10] [11] searches for and reconstructs gravitational-wave transient signals without relying on a specific waveform model. cWB searches for signals with durations of up to a few seconds that are coincident in multiple detectors. The analysis is performed on the time-frequency data obtained with a wavelet transform. cWB selects wavelet amplitudes above the fluctuations of the detector noise and groups them into clusters. For clusters correlated in multiple detectors, cWB reconstructs the direction to the source and the signal waveforms with the constrained maximum likelihood method. To assign detection significance to the found events, cWB ranks them by the coherent signal-to-noise ratio obtained from cross-correlation of the signal waveforms reconstructed in different detectors. A tuned version of the search (cWB-BBH) chooses time-frequency patterns with frequency increasing in time to better match the signal associated with binary black holes mergers.

oLIB [12] uses the Q transform to decompose GW strain data into several time-frequency planes of constant quality factors Q, where Qτf0. The pipeline flags data segments containing excess power and searches for clusters of these segments with identical f0 and Q spaced within 100 ms of each other. Coincidences among the detector network of clusters within a 10 ms light travel time window are then analyzed with a coherent (i.e., correlated across the detector network) signal model to identify possible GW candidate events.

MLy (“Emily”) [13] is a machine-learning-based search for generic sub-second-duration transient gravitational wave signals in the 20 Hz to 512 Hz frequency band. MLy works by passing data from the Hanford-Livingston-Virgo or Hanford-Livingston networks through a pair of convolutional neural networks (CNNs) trained to recognize signals that are simultaneous (up to the light travel time across the network) and coherent between detectors. The training uses randomly generated test signals rather than specific signal models, giving MLy sensitivity to a wide range of signal morphologies. Events are ranked by their combined CNN scores. MLy constructs a sky map of the direction to the source using a maximum-likelihood approach.

RAVEN [14] [15] In addition, we operate the Rapid On-Source VOEvent Coincidence Monitor (RAVEN), a fast search for coincidences between GW and non-GW events. RAVEN will process alerts for gamma-ray bursts (GRBs) from the Gamma-ray Burst Monitor (GBM) onboard Fermi, the Burst Alert Telescope (BAT) onboard the Neil Gehrels Swift Observatory, and the Mini-Calorimeter (MCAL) onboard AGILE, as well as galactic supernova alerts from the SNEWS collaboration. Two astronomical events are considered coincident if they are within a particular time window of each other, which varies depending on which two types of events are being considered (see the table below). Note that these time windows are centered on the GW, e.g., [-1,5] s means we consider GRBs up to one second before or up to 5 seconds after the GW.

Event Type

Time window (s)

Notice Type Considered (see full list)

CBC

Burst

GRB
(Fermi, Swift, INTEGRAL, AGILE)

[-1,5]

[-60,600]

FERMI_GBM_ALERT
FERMI_GBM_FIN_POS
FERMI_GBM_FLT_POS
FERMI_GBM_GND_POS
SWIFT_BAT_GRB_ALERT
SWIFT_BAT_GRB_LC
INTEGRAL_WAKEUP
INTEGRAL_REFINED
INTEGRAL_OFFLINE
AGILE_MCAL_ALERT
SubGRB
(Fermi)

[-1,11]

[-1,11]

FERMI_GBM_SUBTHRESH
SubGRBTargeted
(Fermi)

[-1,11]

[-1,11]

via Kafka alert
SubGRBTargeted
(Swift)

[-10,20]

[-10,20]

via Kafka alert
Low-energy Neutrinos
(SNEWS)

[-10,10]

[-10,10]

SNEWS

In addition, RAVEN calculates coincident FARs, one including only timing information (temporal) and one including GRB/GW sky map information (space-time) as well; we require the latter to publish an alert. Both the GRB and SubGRB searches uses external candidates published via GCN independently, checking for potentially coincident GW candidates around these. Due to the high significance of these GRB candidates, we use the untargeted search method where the astrophysical rate dominates. For the SubGRBTargeted search, our search partners analyze their sub-threshold data around our low-significance alerts. These GRB candidates typically are not significant, so their false alarm rate dominates and the targeted search method is more appropriate. Regardless of method, every search uses the same format for their alert contents.

LLAMA [16] [17] The Low-Latency Algorithm for Multi-messenger Astrophysics is a an online search pipeline combining LIGO/Virgo/KAGRA GW triggers with High Energy Neutrino (HEN) triggers from IceCube. It finds temporally-coincident sub-threshold IceCube neutrinos and performs a detailed Bayesian significance calculation to find joint GW+HEN triggers.