FIRST SEARCHES FOR OPTICAL COUNTERPARTS TO GRAVITATIONAL-WAVE CANDIDATE EVENTS

Triggers for this search were identified with a collection of low-latency pipelines designed to find transient GW events in data from the three site LIGO/Virgo network. Here, we provide a brief summary of the trigger production and selection, while a more detailed description is described in Abadie et al. (2012c, 2012d). During the winter run, two pipelines were used to identify generic short-duration transients of significant signal power, or "bursts," and estimate their source positions: the Omega (Ω) Pipeline (Searle et al. 2008; Abadie et al. 2010a) and the coherent WaveBurst (cWB) pipeline (Klimenko et al. 2011). For the autumn run, a third trigger pipeline was added: the Multi-Band Template Analysis (MBTA; Beauville et al. 2008; Abadie et al. 2012c), which sought inspiral waveforms from coalescing compact objects. The autumn run also added a second instance of cWB, configured to target linearly polarized GW signals, as might be expected from supernovae.

To compare triggers from different pipelines and identify the ones suitable for observation, follow-up software made event candidate selections based on the estimated false alarm rate (FAR) of each trigger. The rate of background false alarms was estimated by forming a distribution of artificial triggers from data with one or more data streams shifted by at least several seconds. Time-shifting data removes correlations of possible GW signals between detectors, so this distribution was considered to be free from any putative signals and represented the rate of triggers not due to transient GWs (Abadie et al. 2012a, 2012e). During the winter run, a FAR threshold of 1 trigger day⁻¹ was applied to triggers, and a less significant FAR was accepted in the last week to exercise the system. For the autumn run, the FAR threshold was set to 0.25 day⁻¹. Triggers which passed the automated threshold received attention from an on-call follow-up team. The on-call team checked that the trigger occurred in high quality data in each interferometer. In addition, the criteria for manual validation in the winter run included demands that the three suggested (see below) QUEST fields covered a sky area corresponding to a greater than 50% probability of containing the GW source and that follow-up requests were sent at a rate of less than one per 24 hr.

The trigger pipelines reported the estimated position of each candidate GW event as a skymap, a list of probability densities assigned to pixels in a grid covering the sky. The grid used pixels approximately 04 on a side, selected to be similar to the degree-scale resolving power of the GW network (for example, Fairhurst 2011; Klimenko et al. 2011; Vitale et al. 2012; Nissanke et al. 2011). The large angular size of the skymaps required a choice of where within the uncertainty region to observe. To observe the regions most likely to contain an observable GW source, we used a catalog of galaxies within 50 Mpc and Milky Way globular clusters (GWGC; White et al. 2011), thought to be around 70% complete to 50 Mpc by B-band luminosity. Each pixel in the skymap was given a weight P according to the formula

$$\begin{equation} P \propto L \left(\frac{M}{D} \right), \end{equation} \tag{ 1 }$$

where L is the probability of the pixel derived from the GW data alone; M is the blue light luminosity of the galaxy or galaxies contained in the pixel, which is used as a proxy for the star formation rate; and D is the distance to the galaxy (Nuttall & Sutton 2010). For MBTA triggers, a slightly modified version of this approach was applied, using the maximum distance consistent with the apparent inspiral signal (Abadie et al. 2012c). The suggested fields for each telescope were those that maximized the sum of P within the respective FOV. Unless unobservable due to daylight or geometrical constraints, the suggested fields were passed to each optical telescope for every GW event candidate that passed manual validation. However, a more stringent selection was applied for PTF, and only one GW trigger was sent to PTF.

In the winter run, the on-call team was alerted a total of nine times. Three of these triggers were vetoed by the on-call team. Six triggers were approved by the on-call team and sent to the QUEST and TAROT telescopes with roughly 30 minutes of latency. Of the six requests, four were rejected as unobservable by the scheduling software of both telescopes and two triggers were followed-up with the QUEST telescope. In addition, two triggers that did not pass the automated FAR threshold were selected by the on-call team and passed to the partner observatories in an effort to expand the winter run data set (see Table 3).

In the autumn run, only one trigger was manually rejected due to data quality concerns. Six triggers resulted in alerts to the observing partners, four of which resulted in follow-up observations¹⁴⁷ (see Table 4). Two of the triggers are worth special note. The September 16 trigger was recognized by the on-call team as having a special significance: in addition to a small estimated FAR, spectrograms of the GW data revealed frequency evolution characteristic of the late inspiral and merger of two compact objects. This event was later revealed to be a blind hardware injection, a simulated signal secretly added to the data to test the end-to-end system. The September 26 event candidate was also discovered with a low FAR estimate. In subsequent GW data analysis, this trigger was found to be the most significant cWB trigger above 200 Hz in the time period where H1, L1, and V1 were running in coincidence in this science run, though was removed from the analysis based on data quality concerns. The FAR was measured to be 0.023 events per day, or one such trigger expected for every 44 days of network livetime. Since these detectors ran in coincidence for a total of 52.2 days throughout the Virgo science run, this trigger was consistent with expectations for detector noise.

This section describes the image analysis pipeline developed specifically for the TAROT, Zadko Telescope, and QUEST observations. Unlike other approaches presented in this work, the pipeline did not use image subtraction but it extracted a source catalog from each image, and sought transients by comparing the set of catalogs to a reference. For this reason, we refer to this pipeline as the "catalog-based search."

4.1.1. Analysis Pipeline

4.1.2. Background Estimation

The background was estimated by running the analysis over a series of images obtained from random time permutations of the real observation images. The first night observations were excluded from being selected as the first image in each permuted sequence to remove any astrophysical electromagnetic counterparts from the data set. The background simulation was repeated 100 times for TAROT and the Zadko Telescope and for all the permutations allowed by the observations for QUEST.

Genuine optical transients would have lost their regularly fading light curve in the scrambled image set. Random sequencing thus erased them while artifacts such as CCD noise, pixel saturation, bad pixels, errors in the de-blending and source association, etc., were just as likely to pass the pipeline's selection cuts as with the true sequencing. This procedure allowed a measurement of the rate of false positives due to "technical" noise. However, this procedure did not permit a valuable estimate of the "astrophysical" background since the randomization reduced the number of identified astrophysical transients that actually dimmed over time. A statistically significant estimate of the astrophysical background would require the study of survey data not associated with GW triggers, which was not available at this time.

An example of the distribution of technical background events (after the removal of rapid transients) detected in the FOV of TAROT for trigger G19377 is shown in Figure 1. The cumulative distribution of their initial magnitude is shown in the left plot, and the FAP as a function of the slope index is in the central plot. The on-source analysis showed a greatly reduced background level compared to the whole-field analysis, since only objects near a local galaxy were included. In this example, the nominal slope index threshold of 0.5 reduced the FAP to less than 1% in the on-source analysis. For the whole-field analysis, in addition to the same cut on slope index, a requirement that objects showed an initial flux brighter than magnitude 14 was needed to reduce the FAP below the 10% objective.

Zoom In Zoom Out Reset image size

The "technical background" rate varied significantly between different instruments due to different fields of view, limiting magnitudes, image quality, and star crowding. For TAROT and Zadko, the number per square degree of "technical false positives" brighter than a reference magnitude of 14.5 mag for TAROT and 15.5 mag for Zadko was evaluated to be less than 1 deg⁻² using a slope index threshold of 0.5. For QUEST, the background study was performed CCD by CCD to account for the different density of false positives on each CCD. Compared to TAROT and Zadko, the deeper sensitivity observations of QUEST led to a higher number of false positives: an average value of 6 deg⁻² brighter than 18 mag and with magnitude variation larger than 0.5. Reducing the analysis to the on-source regions allowed us to lower the density of background transients to less than 1 deg⁻².

4.1.3. Analysis Tuning

The ROTSE-III network consists of four robotic telescopes at various locations around the world. For each GW trigger in the autumn run, the telescopes repeatedly observed a single field. Each field was observed in a series of 30 exposures on the first night after the trigger time. Follow-up images were collected over the next 30 nights, with observations spaced an average of every 2 nights. Each follow-up observation included 8 exposures, each 20 or 60 s.

We used the existing ROTSE pipeline to analyze the images taken with the network. Based on the ISIS package,¹⁴⁹ which uses a single convolution algorithm described in Alard & Lupton (1998) and Alard (2000), the ROTSE pipeline was adapted to use cross correlation to improve image subtraction results. The details of this method can be found in Yuan & Akerlof (2008). The pipeline was implemented for our analysis to require minimal user interaction and for large scale processing which enabled characterization of the background, as described in Nuttall et al. (2013).

The pipeline began by stacking images from the same night on top of one another to form a coadded image. SExtractor was used to produce a list of objects and their coordinates for each coadded image. These images were then subtracted from the coadded reference image, and several criteria were imposed on any objects found in the subtracted image. Selection criteria included requiring a full width at half-maximum consistent with a point source, seeking a minimum fractional flux variation between images and a signal-to-noise ratio (S/N) greater than some amount. The specific criteria depended on the location of the source in an image. For example, if a source matched a star or an unknown object a flux change of 60% was required, whereas if a source was within 20% of the semi-major axis length from the center of a galaxy, but not consistent with a core, only a 3% flux change was required. The result was several lists of candidates (one from each night), which we combined to produce a single list of unique candidates which appeared in the images, and generated light curves for all candidates.

The vast majority of these candidates were due to poor subtraction, with a fraction of real but uninteresting transients (such as variable stars or asteroids). In order to remove contaminants from the list of candidate transients, each object was subjected to a series of cuts. In order to be of interest, the transient must have appeared on more than one night, shown a sufficiently decaying light curve 48 hr after the trigger, and not have been coincident with a known variable source (from the SIMBAD catalog¹⁵⁰) or with a minor planet (Minor Planet Checker¹⁵¹). These cuts proved efficient at rejecting the majority of the background. Candidates were then highlighted if they overlapped with known galaxies or if their light curves were consistent with a target theoretical light curve (Metzger et al. 2010; Kann et al. 2011, 2010). They were also assigned an ad hoc ranking statistic, R, defined as:

$$\begin{equation} R \equiv \sum _i (18 - m_i)\Theta (18 - m_i) \times w_i \,. \end{equation} \tag{ 2 }$$

Here Θ(x) is the step function, m_i is the background-subtracted magnitude of the transient in image i, and w_i is a weight factor defined by

$$\begin{equation} w_i = \left\lbrace \begin{array}{@{}c l@{}}1 & \hspace{11.38109pt} t_i - t_\mathrm{GW} < 1\,\mathrm{day} \\ \left(1 + \log _{10}\frac{t_\mathrm{i} - t_\mathrm{GW}}{1\,\mathrm{day}} \right)^{-3} & \hspace{11.38109pt} t_i - t_\mathrm{GW} \ge 1\,\mathrm{day} \end{array} \right. \end{equation} \tag{ 3 }$$

where t_GW is the time of the GW trigger, t_i is the time of image i. The ranking statistic was designed to prefer events which were bright within a day of the trigger time and which appear in multiple images.

The ROTSE FAR was investigated by processing sets of images for each of 100 random field locations selected from the ROTSE archive. Each set contained ∼240 images of the field from a month of nominally nightly observing. The FAP for each GW candidate was estimated by counting the number of transient objects visible in archived images with a similar cadence as the images collected for that GW candidate. The ranking statistic for each such transient object was calculated using Equation (2). These studies allowed us to set thresholds on the ranking statistic to keep the target light curves, while rejecting contaminants.

POTS has an unusually wide FOV of 20° × 20°, with a typical limiting magnitude of 11.5 for a 10 s exposure. This allowed the telescope to image a large part of the sky in response to one LIGO/Virgo trigger, over 40° × 40° on most nights. We used the standard POTS pipeline to analyze the images taken by the telescope. A detailed description may be found in Malek et al. (2009) and Sokolowski (2008). The full analysis was carried out in two steps. First, in each image taken by the telescope, the Guide Star Catalog (Jenkner et al. 1990) was used to identify previously unknown sources. Second, POTS's nova recognition algorithm was applied to the list of unknown sources. To separate optical transients from contaminating sources, the algorithm utilized several types of vetoes, including checks on background saturation, nearby bright objects, satellite databases, and the Guided Star Catalog. Objects that passed the cuts were then visually inspected.

During the human inspection stage, every candidate that was not identified as a satellite or background fluctuation was checked against lists of known sources. First, we queried the POTS, INTA (Spain) site for observations made in 2011. Due to the long time (∼1 yr) between the autumn science run and observations from the INTA site, any objects observed by INTA were likely unrelated to the GW trigger.¹⁵² Finally, objects were cross-correlated with the SIMBAD catalog, and sources that appeared nearer than 150'' to the position of any known star or infrared source were rejected.

SkyMapper obtained two epochs of an eight image mosaic covering a total of ∼42 deg² in response to the 2010 September 16 trigger. An image subtraction technique was applied to identify possible transients. The SkyMapper images were reduced via the normal bias subtraction, overscan correction and flat fielding using a custom made Python-based pipeline. Thereafter, frames from the two epochs were aligned with the WCSREMAP¹⁵³ routine and subtracted with HOTPANTS¹⁵⁴ to create residuals images. SExtractor was used to identify sources with S/N greater than three. Then, a series of cuts was applied to the SExtractor output parameters to identify noise and bad subtractions. These included using the ellipticity parameter, photometry from different size apertures, and catalog matching of variable stars. In addition, a study of the point-spread function (PSF) of each object was performed on the subtracted images by fitting the detection with a two-dimensional Gaussian and comparing the fit parameters to the expected, known, PSF. The remaining objects were then examined manually to verify they correspond to an object which was visible in the first epoch and not detectable/fainter in the second. The light curves were then measured using differential photometry with nearby stars.

The PTF accepted the trigger of 2010 September 26. Nine PTF fields, each covering 7.26 deg², were schedule automatically for observations, and they were observed beginning ≈6 hr after the trigger time (since the trigger occurred during day-time on the Pacific Coast). PTF then repeated the observations on several subsequent nights. The number of follow-up observations was mainly limited by full moon constraints.

The imaged fields were searched for candidate transients using the image subtraction pipeline hosted at LBNL (P. E. Nugent et al. 2014, in preparation; Gal-Yam et al. 2011). Only three of the fields imaged by PTF had previously constructed reference images. For the rest of the fields, image subtraction was performed using a reference image constructed by co-adding several images taken during the first night of observations. Image differencing inherently produces a large number of spurious candidates, and only a small fraction (less than few percent) of these are real events. As described in Bloom et al. (2012), in a typical PTF night of order 10⁵ residual sources are found per 100–200 deg² of imaging, after performing subtraction of the reference image.

To distinguish between astrophysical objects and "bogus" image subtraction residuals, we made use of a classification parameter named the "realbogus" parameter (RB; Bloom et al. 2012), which was assigned by a machine-learned (ML) classifier so as to reasonably mimic the human scanning decision of real or bogus. The RB parameter ranged from 0 (definitely bogus) to 1 (definitely real), and was constructed from 28 SExtractor output parameters, including magnitude, ellipticity of the source, and distance from the candidate to reference source.

To maximize the chances of identifying a potential optical counterpart to G21852, the images collected by PTF were analyzed using two different procedures for transient identification, both based on the RB parameter as a starting point (P. E. Nugent et al. 2014, in preparation). While the first procedure (hereafter, the "automated" approach) was largely based on automated ML techniques and optimized for fast transients, the second (hereafter, the "citizen-based" approach) was largely based on a citizen project (Smith et al. 2011) and optimized for supernova searches. In what follows, we describe these two approaches in more detail.

4.5.1. Automated Approach

4.5.2. Citizen-based Approach

4.5.3. Selection for LIGO/Virgo Event Candidates

The Liverpool Telescope observed the G23004 trigger using both the 4.6 arcmin FOV RATCam instrument and the 1° FOV SkyCamZ camera. This produced a total of 22 SDSS r'-band RATCam images and 121 "clear" filter SkyCamZ images from two nights 29 days apart. In addition, 3 RATCam and 17 SkyCamZ images were taken in early 2012 to serve as reference images for image subtraction. The analysis made use of several freely available software packages, and was split into several sections written in Python.

First, we combined the images from 2012 to create our reference images. This was done by aligning the images using the WCSRemap¹⁵⁵ package and combining them using the SWarp¹⁵⁶ package. We also combined sets of five SkyCamZ images on each night to improve image quality and provide a similar cadence to the RATCam images. We removed one RATCam image and two SkyCamZ images due to quality issues.

Second, as the SkyCamZ images used a non-standard filter,¹⁵⁷ they were calibrated using the USNO-B catalog of stars to determine the zero point offset required to calculate correct magnitudes, in the same way ROTSE and TAROT images were calibrated (see Section 4.1). This was done by comparing the USNO-B R-band magnitude of stars in the combined SkyCamZ fields with those same stars found using SExtractor.

The images were then aligned individually to the reference images, again using WCSRemap, and the reference image was subtracted using the HOTPANTS¹⁵⁸ image subtraction package. SExtractor was then used to detect potential candidates in each individual field with a minimum of 4 pixels each with a flux greater than 4σ above the background noise of the image. This reduced the frequency of detecting uninteresting objects, such as cosmic rays, extremely faint stars and noise from the image subtraction process while allowing us to achieve a sensitivity around 20th magnitude in the narrow-field RATCam images.

Using the output of SExtractor from each of the subtracted images, a Python script combined the objects found into a master list containing every unique candidate found in those images, along with useful parameters from SExtractor. From this data, a series of cuts were made to find candidates interesting to this analysis. First, candidates found to be near an image edge (or a bad pixel strip in the case of RATCam images) were rejected. Second, a cut was made to remove artifacts due to bad subtraction. This was achieved by examining the region in the subtracted image around the candidate and calculating the total flux more than 4σ below the median noise of the image. Since bad subtractions are usually caused by poor alignment or convolution, they typically produce a large amount of "negative" flux in the residual image. If the total amount of flux below this threshold was the equivalent required for detection of candidates (4 pixels above 4σ) then the candidate was rejected. The next cut removed candidates not seen in at least half of the images available on the first night, to ensure candidates were visible long enough to be used in our analysis. We also rejected candidates that appeared close to known variable stars and minor planets. Finally, we required that a candidate must decrease in brightness by more than 5σ of the median error on the magnitude measurements from SExtractor, from the first night to the second night 29 days later. Since the pipeline is designed to work with images from two telescopes for this analysis which may have different magnitude errors for the same trigger, we used a threshold based on the noise in the image rather than a fixed magnitude variation in the same way as ROTSE and TAROT.

Any objects that remained after these cuts were considered likely candidates, and looked at in more detail. This was done by plotting the light curves of each object across both nights and inspecting images of the candidates in both the original and subtracted images. This allowed us to gauge whether any transients warranted further investigation.

For each winter run trigger, images were collected only during one night. The absence of a second night's observations prevented the construction of variability measures and limited the analyses to only identify "unknown objects," i.e., those not listed in the USNO catalog or with a magnitude significantly different from USNO, but visible in all the collected images. For both the TAROT and QUEST image analysis procedure, at least one observation on another night would have been required to identify a unique electromagnetic counterpart.

In the winter run, TAROT responded to one trigger, CWB1, and collected six images starting the single night observation at T+3d11h. The QUEST camera responded to three triggers, G3821, G4202, CWB2, starting the observations at T+9h46m, T+24m, and T+16h12m, respectively. For each trigger it collected images corresponding to three fields. Each field was observed twice within 20 minutes during the same night.

The TAROT observation associated with CWB1 reached a sensitivity of 15.8 mag. Fifteen galaxies with a distance smaller than 50 Mpc were in the FOV. The analysis found 9 unknown objects in the on-source region and 46 in the entire FOV up to the limiting magnitude. No unknown objects were found with magnitude brighter than 11.8 in the on-source region and brighter than 10.7 mag in the entire FOV.

The three QUEST fields associated with G3821 included a total of 34 galaxies with a distance smaller than 50 Mpc. Only 14 of the galaxies were analyzed due to the exclusion of galaxies observed only one time or lying in CCDs that did not work or had calibration problems. The average limiting magnitude was about 18.6 mag.

For trigger G4202 the three fields included a total of 17 galaxies with a distance smaller than 50 Mpc. Ten galaxies were removed from the analysis because they were observed only one time or associated with poor image quality (impacted by bad lines and pixels or by background subtraction artifacts) or calibration problems of the CCDs (astrometric calibration or flat-field problems). An average limiting magnitude of 19.2 mag was reached during the observations.

For trigger CWB2 the three fields included a total of 12 galaxies with a distance smaller than 50 Mpc. Two of the galaxies were not analyzed due to poor image quality or CCD calibration problems. An average limiting magnitude of 19.8 mag was reached during the observations.

The QUEST analysis found 9, 1 and 1 unknown objects in the on-source region with a magnitude brighter than 14 mag for the triggers G3821, G4202 and CWB2, respectively. The number of unknown objects increased to 140, 35, 6 for magnitudes brighter than 18 mag. The number of unknown objects showed a stronger dependence on the density of image artifacts and stars in the FOV than on the on-source area. No "unknown objects" were found for magnitude brighter than 9, 12, 7 mag for G3821, G4202 and CWB2, respectively.

Event G19377 was a simulated signal added to the GW detector data in order to test our data analysis pipelines. The ROTSE-IIIc telescope responded at T+∼12 hr when 30, 20 s exposure images were taken within ∼15 minutes. On subsequent follow-up nights (6–29) both ROTSE-IIIa and c telescopes gathered 80, 20 s exposure images. The images from the two scopes varied vastly in terms of image quality, which posed difficulties for injection studies. We discarded the lower quality images from the 3c telescope, leaving just the 3a images, with an average limiting magnitude of 15.1. Two galaxies at ∼24 Mpc (PGC 078144 and PGC 078133) were visible within the FOV. The ROTSE image processing pipeline revealed 68 unique objects, one of which passed the candidate validation. Further tests found this candidate was consistent with background, with a FAP of 7%. This left no significant candidates. At the location of this background transient there is a known star (red magnitude of 13.1 from the USNO catalog), which shows no significant magnitude variation in the TAROT images associated with the same GW trigger. This location was not covered by the Swift observations taken for G19377. We also tried analyzing images from both the 3a and 3c telescopes together, and found no additional candidates.

SkyMapper observed an eight tile mosaic, 7 days after the initial alert. An analysis was performed, but no plausible transients were discovered.

TAROT took images starting at T+43m and repeated the observations at T+2d, T+3d and T+4d. Observations from the four nights displayed an average limiting magnitude of 15.1. The on-source analysis was performed on the two same galaxies observed by ROTSE and identified no transient counterpart. The whole-field analysis was performed with an initial magnitude threshold of 14 mag, and identified one transient candidate with a slope index of 0.6. A deeper analysis showed that this candidate resulted from an artifact of the de-blending in crowded images.

The Zadko telescope observed the regions around the five galaxies evaluated to be the most likely hosts of the G19377 trigger: NGC 2380, ESO 560-004, ESO 429-012, PGC 078133, and PGC 078144; the last two being in common with ROTSE and TAROT. The observations started at T+1d12.6h and were repeated 5 months later for reference. The average limiting magnitude for both the early and reference images was 16.5 mag. No electromagnetic counterparts were identified by either the on-source on whole-field analysis.

All four ROTSE-III telescopes responded to this GW trigger, taking images spanning T+34h38m to T+29d, centered on the region around the galaxy UGC 11944. However, all images taken with the ROTSE-IIIa, b and d telescopes were discarded because of defocusing factors in addition to weather conditions at those sites being less than optimal. This resulted in 56 images being used for the analysis, with an average limiting magnitude of 15.5. The ROTSE image subtraction pipeline found 77 potential candidates, none of which passed the candidate validation procedure.

The TAROT telescope collected three images in association with G20190. Due to the full moon only an average limiting magnitude of 14.6 mag was reached. Nine months later 18 images were taken by TAROT in the same region of the sky as reference. A mean limiting magnitude of 17 mag was reached during this second observation. No counterpart with a FAP less than 10% was identified by the on-source analysis. The whole-field analysis was performed with a threshold of 10 mag on the initial magnitude and the required presence in the first three images and absence in the reference images. It resulted in four identified candidates. The candidates were seen to be image artifacts linked to the spikes of saturated stars.

The Zadko telescope was pointed toward two Galactic globular clusters: NGC 7078 and NGC 7089, and three galaxies UGC 11868, NGC 7177, and NGC 7241, evaluated to be the most likely hosts of the GW source. Observations of galaxies UGC 11868 and NGC 7241 were taken about 50 minutes after the GW trigger. All five fields were observed subsequently during at least two nights between T+1d and T+4d. The observations were repeated 11 months later for reference. The average limiting magnitudes were 16.4 mag and 17.3 mag for the very first and reference observations, respectively. The on-source analysis identified three transient candidates associated with NGC 7078 and 15 associated with the center of NGC 7089. The candidates were found to be due to problematic de-blending in the central region of globular clusters. No transient was identified by the on-source analysis associated with the three galaxies. The whole-field analysis required a magnitude brighter than 10 and the presence during the first nights and absence in the reference images. This resulted in no detected transient.

The QUEST observations started at T+12h3m. Each field was observed twice within 15 minutes as pairs of images dithered to fill the gaps between rows of CCDs. The entire observation sequence was repeated at T+1.5d. A total of 10 galaxies with a distance smaller than 30 Mpc were identified in the three fields. Three of the galaxies were not analyzed due to poor image quality CCDs or calibration problems. The observation was taken during a full moon night that allowed an average limiting magnitude of 17.6 mag. The on-source analysis¹⁵⁹ identified one possible transient with a FAP less than 10% (see Section 6) associated with the galaxy UGC 11916. A deeper analysis of the candidate showed this to be artificial. The analysis pipeline identified the possible GW host galaxy itself as a transient due to variations in the estimate of its surface photometry over the two nights. An estimate using fixed photometry apertures indicated magnitudes in agreement within the errors with no flux decrease.

ROTSE-IIIb took images spanning T+11h53m to T+29d centered on a region containing both M31 and M110. One follow-up night had to be ignored due to defocusing issues. The average limiting magnitude of the images was 16.6, with 81% of them having an exposure times of 60s. The subtraction pipeline found 187 objects, which resulted in four candidates after candidate validation. All four candidates overlapped with one of the galaxies mentioned, however all were consistent with background. The highest ranked candidate had a FAP of 9%. Consequently, we found no significant candidates. Within the 2 arcsec positional accuracy of PTF, the ROTSE background events are all coincident with known stars, and according to the PTF analysis criteria applied, these sources are not considered candidates.

PTF observed nine different fields on five nights, beginning at T+6h37m. The median limiting magnitude reached in the observed fields over the observation time (and over the 11 CCDs that make the core of the PTF imager) was in the range R ≈ 20.2–19.2. The images collected by PTF were analyzed using two different procedures for transient identification, one entirely based on automated selection criteria for fast transients, and the other largely based on a citizen project targeting supernovae (see Section 4.5 for more details). These procedures for transient identification were routinely used by the PTF survey (P. E. Nugent et al. 2014, in preparation). By applying the selection criteria for fast transients (automated approach; see Section 4.5.1) on the images that were taken for follow-up of trigger G21852, we obtained a list of 172 candidates, none of which passed the vetting for "LIGO/Virgo interesting" transients performed according to the criteria described in Section 4.5.3. We also applied these last criteria to the candidates obtained via the citizen-based approach (optimized for supernova searches—see Section 4.5.2). Of the 218 candidates selected according to criteria (1)–(4) in Section 4.5.2 and sent out to the citizens for scanning, 28 were saved by the citizens and assigned an official PTF name. However, none of these 28 candidates passed the additional vetting described in Section 4.5.3. We also took a closer look at 55 other candidates that were not saved by the citizens, but that had a SN_zoo predictor score >0.025 or a RB2 > 0.3 (see Section 4.5.3). We vetted these candidates according to criteria (1)–(5) in Section 4.5.3, and none of them passed our screening.

The ROTSE-IIIb, c and d telescopes responded to G23004 at T+6h25m and collected data up to T+29d. These images contained one galaxy (NGC 1518) at 11.5 Mpc within the FOV. Around 75% of the data was of poor quality; many of the images were out of focus and cloud cover was also a factor. This resulted in the analysis of 30 images with an average limiting magnitude of 16.7. The ROTSE subtraction pipeline found 124 potential candidates of which none survived the candidate validation tests.

The Liverpool Telescope observed a single field centered on the location of the galaxy NGC 1507, with one hour of observations taken at T+9h and a further one hour at T+30d. The limiting magnitude of the RATCam images was r' ≈ 20.5, averaged over all images, with the calibrated limiting magnitude of the SkyCamZ images averaging R ≈ 17.5. We found 406 unique objects in the RATCam images and 163 unique objects in the SkyCamZ images. After applying cuts described in Section 4 we found no candidates in either the RATCam or SkyCamZ images that met our criteria.

The POTS telescope responded at T+6h56m after the alert. On the first night the telescope used 10 different pointing locations to cover an area containing 40% of the G23004 probability map. Each location was imaged twice. The limiting magnitudes for the first night's observations spanned 10.5–11.0 mag. On the first night there were over 700 cases that were recognized by the pipeline as possible optical transients, but all of them were either already included in the database of weak stars or were noise due to ice crystals on the camera. There were no real optical transients found. The same fields were followed up on the nights of October 5, 6, 7, 11, and 30. Each follow-up night's observed area was covered by nine pointing locations, with each location imaged at least three times. Images from the first four nights were searched by the pipeline for optical transients, and 40 objects were identified as existing in images over multiple nights and have been present on all frames that were taken of that field. Each of these was manually investigated, and none were found to be linked to the GW trigger. Most of the 40 objects were traced to variable stars or were caused by ice crystals on the camera.

The QUEST follow-up for this GW trigger consisted of three nights of observations over three different fields. The first observation began at T+11h32m and then observations were repeated at T+2.4d and T+32.4d. Each night's observations included two visits to each of two dithered positions for each of the three field locations. A total of 32 galaxies with a distance smaller than 50 Mpc were identified in the three fields. Due to inoperative CCDs or CCD calibration problems the regions occupied by four galaxies were not analyzed. The average limiting magnitude for the three night observations was 19.7 mag. The on-source¹⁶⁰ analysis identified one possible transient with an "on source" FAP of less than 10% (see Section 6). The candidate transient overlaid the extended emission of the galaxy IC0402. A deeper analysis indicated no flux change for the object: the point source immersed in the fainter galaxy edge emission has a similar neighboring object that biased its photometry. Using a suitable fixed photometry aperture the magnitudes of the object agree within the errors in all the images. The object could be a foreground star not listed in the USNO catalog or a bright knot of one of the galaxy's arms.

For each set of images collected by TAROT and the Zadko telescopes, 100 simulated transients were added to the data for each counterpart model and distance. To model PSF variations in the wide-field images, reference model stars were identified in each image, and the PSF of the reference star closest to the injection position was used for each simulated object. For the GRB afterglows, we used a range of magnitudes uniformly distributed between the brightest and faintest GRBs (see normalization in Table 1). The results are presented in Figure 8. Long GRB afterglows/short GRB afterglows/kilonovae were recovered with 50% efficiency in TAROT observations to distances of 400 Mpc/18 Mpc/6.5 Mpc respectively for trigger G19377 and 355 Mpc/16 Mpc/13 Mpc for trigger G20190. For Zadko Telescope observations, we obtained 195 Mpc/8 Mpc/4 Mpc for G19377, and 505 Mpc/ 25 Mpc 13 Mpc for G20190. As expected, the results showed some dependence on the depth of the observations, the observation time after the GW trigger, and the density of stars in the field.

Zoom In Zoom Out Reset image size

The QUEST pipeline's recovery efficiency was evaluated separately for each on-source galaxy region. As for TAROT and Zadko, 100 simulated transients were added to the images for each model (kilonova, short and long GRBs) and distance. Randomly distributed magnitudes between the brightest and faintest GRBs (see normalization in Table 1) were used. Figures 9–11 show some representative examples of the achieved recovery efficiencies. The wide range in the recovery efficiencies reflects variations in CCD sensitivity and rates of contaminating artifacts. In addition, bright galaxy extended emission prevented the recovery of some injections, even at close distances. A similar efficiency loss was found when a large part of the on-source region was occupied by foreground stars or image problems like bad pixels and bad lines. The results for the QUEST observations can be characterized by the mean and the standard deviation of the distances corresponding to 50% efficiency to recover injections. For trigger G20190, we found mean distances of 33 Mpc (σ = 7 Mpc) for kilonovae, 30 Mpc (σ = 6 Mpc) for short GRBs, and 820 Mpc (σ ≈ 180 Mpc) for long GRBs. For G23004, a mean distance of 64 Mpc (σ = 25 Mpc) for kilonovae, 63 Mpc (σ = 30 Mpc) for short GRBs, and 1530 Mpc (σ ≈ 700 Mpc) for long GRBs were found.¹⁶¹ The larger spreads for QUEST reflect CCD-to-CCD variations. For both GW triggers, the 50% efficiency distances for long GRB afterglows were well beyond the maximum distance that the LIGO and Virgo detectors could have detected signals coming from NS binary coalescences, while the kilonova and short GRB distances were comparable. However, the result obtained for the kilonova transients is dependent on the adopted model and relies on the fact that the QUEST observations were made around the peak time of the light curve model used for this study.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

For each set of images collected by ROTSE, 140 simulated transients were added to the data for each counterpart model for 10 different distances. The PSFs for the injected transients were modeled on "good" objects PSFs within each image, as described in White et al. (2012). The GRB models used the brightest normalizations shown in Table 1; i.e., assuming magnitude 16 (23) at 1 day from z = 1 for LGRB (SGRB) afterglows. The results are presented in Figure 13. For each GW trigger, the efficiencies for the different counterpart models are very similar as functions of the injection magnitude. The efficiencies peak at ∼70%–80% for triggers G19377 and G20190, and at ∼55% for G21852. Trigger G23004 (not shown) contained images of very poor quality and the injection efficiency only reached a maximum of ∼20%. Long GRB afterglows/short GRB afterglows/kilonovae were recovered with 50% detection efficiency to distances of 400 Mpc/16 Mpc/2 Mpc for trigger G19377, 1000 Mpc/40 Mpc/5 Mpc for trigger G20190, and 1000 Mpc/90 Mpc/5 Mpc for trigger G21852. The maximum sensitive distances correspond to transient magnitudes of approximately 15 on the second night. This was typical of the average limiting magnitude of ROTSE over the FOV. Since the pipeline required transients to be seen on at least two nights, the magnitude on the second night was the primary factor determining the sensitivity to each model. Transients at much smaller distances tended to suffer from saturation and were discarded in the image subtraction. The maximum detection efficiency was less than 100% because the pipeline was not always able to produce the background-subtracted lightcurve for a transient; this depended on the position in the image and on the image quality, as 16 reference stars were needed in the region around the transient for accurate image subtraction. Variations in efficiency between triggers were due mainly to differences in image quality and also differences in CCD performance between the different telescopes in the ROTSE network.

An example of the distribution of injections against the background can be seen in Figure 12. This figure shows that of all the injections that produced a nonzero ranking statistic with the specific distance scales shown, more than 60% of the injections were recovered with a rank comparable to the most highly ranked background event. However, none of the injections were found with a ranking statistic higher than loudest background event. As the injection distances increased, the injections fell more and more within the background.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The efficiency of the Liverpool Telescope pipeline was measured with the same methods used for ROTSE. A Python script was written to inject 100 transient objects per 10 Mpc bin per model, with light curves following the three models described in Table 1, assuming the brightest normalization for the GRB models. These images were then analyzed using the pipeline, and a script used to find and flag injections found in the pipeline output. Figure 14 shows that we obtained efficiencies around 90% for injections brighter than the limiting magnitude, including saturated objects normally discarded in other image subtraction methods. For RATCam, any of the tested models would have been observable out to 100 Mpc or more—well beyond the initial LIGO/Virgo horizon distance for NS mergers. For SkyCamZ, we found similar efficiencies, over smaller distance ranges.

Zoom In Zoom Out Reset image size

The efficiency of the POTS transient search was investigated by adding simulated stars to existing images and reprocessing them. The objects that were injected had different magnitudes and were chosen from real observed stars during the autumn science run. Unlike the other simulations described in this paper, objects added to POTS data did not follow model light curves, but instead measured the ability of the pipeline to recover a transient of a given magnitude using data from a single night. Stars injected in one image were also injected in subsequent images of that field taken during the same night. Only injections that were made to the inner part of the CCD chip, at least 150 pixels from CCD borders, were considered to estimate transient detection efficiency. The border part of the CCD was rejected by the off-line optical transient recognition algorithm due to the possibility of CCD anomalies that might be mistaken as short optical transients. Also, only injections starting on a good quality image were considered in efficiency estimation. This means that the effective FOV for optical transient recognition corresponds to 15° × 15°. At each stage of the processing it was determined how many of the injected objects were detected.

Figure 15 shows two curves demonstrating the efficiency of the POTS pipeline. The first one describes how many of the injected objects were detected in at least one image and the second curve shows how many of the injections were detected in five or more images. The first case corresponds to the minimal criterion that was required for the candidate to be classified as an optical transient and be inspected by a human. The second case reflects the criteria used for an optical transient to have been automatically classified as a nearly certain real event. On both curves we see that the maximal efficiency did not reach near 100%, even for very bright sources. This can be attributed to several causes. An important loss of efficiency came from areas excluded from the search due to the presence of previously discovered stars. Objects injected within a radius of 150'' of stars listed in the POTS star catalogue were not recognized as optical transients and discarded by the pipeline, resulting in a 12%–15% impact to the injection recovery rate. Additional sources were lost to structure in the CCD: 10%–15% of the CCD area consisted of wire guiding electric charge. A significant part of the losses also came from quality checks in the algorithm preprocessing. At this stage transients that were fainter than 11th magnitude, or observed on multiple low quality frames, were discarded. This impacted the efficiency by 10% for bright transients, and up to 30% for faint transients injected with brightness around magnitude 11. Other cuts in the data processing pipeline resulted in an additional 3%–10% loss of efficiency.

Zoom In Zoom Out Reset image size