Interface definition for online detection pipelines

The BAYESTAR rapid localization algorithm is designed as a post-processing stage for compact binary merger search pipelines. The script bayestar-localize-lvalert provides rapid sky localization as a service.

Note

Other available modes of operation for BAYESTAR that are documented elsewhere include the script bayestar-localize-coincs for offline batch processing and the method ligo.skymap.bayestar.localize() for directly calling BAYESTAR from Python code.

Sequence diagram

In online operation, search pipelines upload candidates to the Gravitational-Wave Candidate Event Database (GraceDB). The script bayestar-localize-lvalert (or the equivalent Celery task in GWCelery) listens for and intercepts LVAlert pubsub messages. For each event created, the script downloads the pipeline’s output data products from GraceDB, performs rapid sky localization, and uploads the resulting FITS file back to GraceDB.

The interactions between the search pipeline, GraceDB, and BAYESTAR are illustrated in the sequence diagram below. Line styles have the following meanings:

  • Solid lines directed into GraceDB represent HTTP requests.

  • Solid lines directed out of GraceDB represent HTTP responses.

  • Dashed lines represent LVAlert pubsub messages.

Sequence diagram for unified coinc.xml file

sequenceDiagram note over Search: New detection Search ->>+ GraceDB: Upload coinc.xml note over GraceDB: Create event G123 GraceDB -->>+ BAYESTAR: LVAlert: coinc.xml added to G123 deactivate GraceDB BAYESTAR ->>+ GraceDB: Get coinc.xml from G123 GraceDB ->>- BAYESTAR: coinc.xml note over BAYESTAR: Perform sky localization BAYESTAR ->> GraceDB: Upload bayestar.fits to G123 deactivate BAYESTAR

Sequence diagram for separate coinc.xml and psd.xml.gz files

sequenceDiagram note over Search: New detection Search ->>+ GraceDB: Upload coinc.xml activate Search note over GraceDB: Create event G123 GraceDB ->>- Search: GraceDB ID: G123 Search ->>+ GraceDB: Upload psd.xml.gz to G123 deactivate Search GraceDB -->>- BAYESTAR: LVAlert: psd.xml.gz added to G123 activate BAYESTAR BAYESTAR ->>+ GraceDB: Get coinc.xml from G123 GraceDB ->>- BAYESTAR: coinc.xml BAYESTAR ->>+ GraceDB: Get psd.xml.gz from G123 GraceDB ->>- BAYESTAR: psd.xml.gz note over BAYESTAR: Perform sky localization BAYESTAR ->> GraceDB: Upload bayestar.fits to G123 deactivate BAYESTAR

Input files

This section describes the interface between search pipelines and BAYESTAR. The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in RFC 2119.

The following file MUST be uploaded to GraceDB:

  • coinc.xml: The event file, which SHOULD be the initial upload that creates the event.

The contents of the coinc.xml file MUST conform to the Event data section below. The coinc.xml file SHOULD also contain the data described in the PSD data section below. If the coinc.xml file does not include the PSD data, then the following additional file MUST be uploaded to GraceDB:

  • psd.xml.gz: The power spectral density data file, which MUST be uploaded with the psd tag.

If the psd.xml.gz is uploaded, then its contents MUST conform to the PSD data section below.

The format of both files MUST be LIGO-LW (see LIGO-T990023). LIGO-LW is a legacy XML-based format used by a variety of LIGO/Virgo/KAGRA software and services for storing tabular datasets.

Unfortunately, LIGO-LW is a rather complicated format. We recommend using either the ligo.lw module or GWPy’s tabular LIGO-LW I/O feature to simplify reading and writing LIGO-LW files.

Note

There are two variants of the LIGO-LW format, an old format implemented by glue.ligolw that uses string (“ilwdchar”) row IDs, and a new format implemented by ligo.lw that uses integer row IDs. GraceDB and BAYESTAR can accept either format, but pipelines SHOULD upload files in the new format.

The ligolw_no_ilwdchar command-line tool provided by ligo.lw can convert from the new format to the old format.

Event data

This event data describes the search pipeline’s matched filter output. It MUST include the point estimates of the time, phase, and amplitude on arrival in each detector. It MUST provide the intrinsic template parameters (masses and spins). It SHOULD include a signal-to-noise time series for each detector.

The event data MUST include at least the following LIGO-LW tables (in any order):

process
  • The process table MUST contain at least one row with the process_id and program columns populated in order to identify the search pipeline.

  • The value of those rows’ program column MUST be one of pycbc, gstlal_inspiral, gstlal_inspiral_postcohspiir_online, MBTAOnline, bayestar_realize_coincs, or bayestar-realize-coincs.

  • Additional valid columns of this table MAY be populated in order to identify the pipeline software version or include other metadata. Additional unrelated rows (e.g. to identify prior analysis steps such as template bank generation) MAY be included and will be ignored.

sngl_inspiral
  • The sngl_inspiral table MUST contain exactly one row per detector that the search analyzed.

  • The values of the event_id column MUST be distinct across all rows.

  • The values of the following columns that specify the intrinsic template parameters MUST be identical across all rows: mass1, mass2, f_final, spin1x, spin1y, spin1z, spin2x, spin2y, and spin2z.

  • If the template has zero spin, then the spin columns MAY be left blank. If the template has aligned spins, then the _x_ and _y_ spin components MAY be left blank.

  • The end_time and end_time_ns columns MUST report the seconds and nanoseconds parts of the GPS time at which the same fiducial reference part of the signal (e.g., the time of merger, or the time at which the inspiral reaches reference frequency) is received in each detector. It SHOULD record the merger time. If the event is an “early warning” or pre-merger event, then it SHOULD record the predicted time of merger.

  • If the event is an early warning event, then the high-frequency cutoff frequency MUST be recorded in the f_final column.

  • The snr column MUST report the absolute value of the complex matched filter SNR of the best-matching template. It MUST NOT report a modified SNR-like quantity such as newSNR.

  • The coa_phase column MUST report the argument of the complex matched filter SNR of the best-matching template.

  • If the search pipeline as identified by the program column in the process table is pycbc, then phase convention of the coa_phase column MUST be that the matched filter output is linear in terms of the data. Otherwise, the phase convention MUST be that the matched filter output is antilinear in terms of the data.

  • The end_time, end_time_ns, snr, and coa_phase columns MAY be blank for any row for which there is a corresponding SNR time series (see below).

  • Due to a bug in GraceDB, all columns of the sngl_inspiral table (including blank ones) must be present.

coinc
  • There MUST be exactly one row in the coinc table with at least the coinc_event_id column populated.

  • The value of the process_id column of the coinc tale MUST match the value of the process_id column in the:class:process table that identifies the search pipeline.

  • Note that due to another bug in GraceDB, the time_slide_id column MUST be populated. It MAY have a legal dummy value such as time_slide:time_slide_id:0.

coinc_event_map
  • There MUST be exactly one row in the coinc_event_map table for each row in the sngl_inspiral table.

  • The value in each row’s event_id column must be set to the value of the event_id column in the corresponding row of the sngl_inspiral table.

  • The value in each row’s table_name column must be set sngl_inspiral.

  • Each row MUST have the coinc_event_id column set to the value of the coinc_event_id column in the one row of the coinc table.

coinc_inspiral
  • The coinc_inspiral table MUST be present because it is required by GraceDB (although it is ignored by BAYESTAR).

  • It MUST have exactly one row.

  • The value in the coinc_event_id column MUST match the value in the corresponding column in the coinc table.

  • The following columns MUST be populated: coinc_event_id, combined_far, end_time, end_time_ns, ifos, and snr.

  • The mass and mchirp columns SHOULD be populated.

The coinc.xml file SHOULD also provide SNR time series for each detector.

  • Each SNR time series MUST be stored inside a LIGO_LW element as a serialized COMPLEX8TimeSeries. The function lal.sereries.build_COMPLEX8TimeSeries() can be used to serialize a COMPLEX8TimeSeries.

  • Each of the LIGO_LW elements for serialized SNR time series MUST contain a Param element to link it to a row in the sngl_inspiral. The param name MUST be event_id:param and the param’s type and value must match the event_id column in the corresponding sngl_inspiral row.

  • The SNR time series MUST have an odd number of samples, e.g., the length must be \(2 * n + 1\) for some integer \(n\).

  • The timestamp of the central sample (e.g. \(n\) times the sample interval plus the epoch) MUST differ from the corresponding sngl_inspiral row’s time (if present) by no more than one sample interval.

  • The timestamps of the samples of the SNR time series MUST correspond to sample boundaries. The timestamps MUST NOT have any sub-sample time shift applied to them.

  • For any detector that lacks an SNR time series, sub-sample interpolation SHOULD be applied by the search pipeline to obtain the values for the snr, coa_phase, end_time, and end_time_ns columns in the corresponding row of the sngl_inspiral table.

PSD data

The PSD data consists of each analyzed detectors’ estimated noise power spectral density (PSD) series.

  • There MUST be exactly one PSD per detector analyzed.

  • Each PSD MUST be stored inside a LIGO_LW element as a serialized REAL8FrequencySeries. The lal.sereries.build_COMPLEX8TimeSeries() function or the lal.sereries.make_psd_xmldoc() function can be used to serialize REAL8FrequencySeries.

  • Each LIGO_LW element MUST contain a Param element to link it to a detector. The param’s name MUST be instrument:param, its type MUST be instrument:param, and its value should be a detector prefix such (e.g. one of H1, L1, V1, K1, I1, etc.)

  • Any samples that are invalid because their frequencies are outside of the range analyzed by the search MUST be absent or have their values set to positive infinity. Invalid values MUST NOT be set to zero.

Example files

For a minimal example, see the mock coinc.xml file.