Inference¶
For CBC events, we calculate a number of quantities that are inferred from the signal. In preliminary alerts, these quantities are based on the candidate significance and the matched-filter estimates of the source parameters. Once parameter estimation has been completed, updated values will be provided based on samples drawn from the posterior probability distribution.
Classification¶
The classification consists of four numbers, summing to unity, that give the probability that the source is either a BNS, NSBH, BBH merger, or is of Terrestrial (i.e. a background fluctuation or a glitch) origin.
Our search pipelines use independent methods to
assign those source probabilities to their own triggers. These methods all
start from the common assumptions that terrestrial and astrophysical events
occur as independent Poisson processes, and that NSs and BHs only exist below
and above
For details of the GstLAL method, see [1] (especially Equation 20). For MBTA, see [2]. For PyCBC Live, see [3]. For SPIIR, see [4] and [3].
RapidPE-RIFT [9] [10] provides source classification by estimating source probabilities assuming astrophysical origin and then combining them with the Terrestrial probability estimated by the search pipelines for the preferred event. The RapidPE-RIFT source classification, if available, will be used in the second preliminary alert and/or initial notice. If RapidPE-RIFT is available after the initial notice, then the RapidPE-RIFT source classification results may be included in the Update GCN Circular notice.
Properties¶
The source properties consist of a set of numbers, each between zero and unity, that give the probabilities that the source satisfies certain conditions. These conditions are:
HasNS: At least one of the compact objects in the binary (that is, the less
massive or secondary compact object) has a mass that is consistent with a
neutron star. Specifically, we define this as the probability that the
secondary mass satisfies
HasRemnant: The source formed a nonzero mass outside the final compact object. Specifically, the probability is calculated using the disk mass fitting formula from [6] (Equation 4). Several neutron star EOSs are considered to compute the remnant mass. The value is marginalized by weighting based on Bayes factors in reference mentioned above.
HasMassGap: At least one of the compact objects in the binary has a mass in
the hypothetical “mass gap” between neutron stars and black holes, defined here
as
The performance of these quantities across online CBC pipelines is shown below.
The ROC curves shown above were constructed using pipeline recovered injections from the O3 Mock Data Challenge (MDC). For details see [8].
The mass values mentioned in this section are source-frame mass. The value reported in the preliminary alert is calculated using a supervised machine learning classifier on a feature space consisting of the masses, spins, and SNR of the best-matching template, described in [7]. This is to account for the uncertainty in the reported template parameters compared to the true parameters. The value reported in the update alerts uses the online parameter estimation to compute the value.