A DESGW Search for the Electromagnetic Counterpart to the LIGO/Virgo Gravitational-wave Binary Neutron Star Merger Candidate S190510g

All three LVC detectors (LIGO Livingston, LIGO Hanford, and Virgo) recorded the event, with a 98% initial probability of being a binary neutron star (BNS) event, a 2% probability of having a nonastrophysical origin, and a false alarm rate of 1 per 37 yr. The 50% (90%) confidence regions spanned 575 deg² (3462 deg²) in the initial LVC bayestar localization map. At 10:08:19 UTC on May 10, the LVC released an updated map from the LaLInference pipeline (Veitch et al. 2015), decreasing the 50% and 90% confidence regions to 31 deg² and 1166 deg², respectively, and refined the distance estimate to 227 ± 92 Mpc, or z = 0.05 ± 0.02 (using flat ΛらむだCDM cosmology with H₀ = 70 km s⁻¹ Mpc⁻¹ and Ωおめが_m = 0.3) (LIGO Scientific Collaboration & VIRGO Collaboration 2019a). On May 10, 20:43:51 UTC the classification of the nature of the event was updated to 85% BNS and 15% nonastrophysical. Finally at 20:18:44 UTC on May 11, the LVC updated this probability to being of a nonastrophysical origin at 58% and of a BNS to 42% as well as updating the false alarm rate to 1 in 3.6 yr.

DECam was used for two nights to conduct target-of-opportunity imaging of the LIGO/Virgo GW compact binary merger candidate S190510g (LIGO Scientific Collaboration & VIRGO Collaboration 2019b). Since the initial classification of S190510g was a BNS merger with high probability, the GROWTH collaboration chose to trigger DECam (NOAO proposal 2019A-0205). All exposures from this proposal were immediately made public (Andreoni et al. 2019b). GROWTH initiated EM follow-up on 2019 May 10 at 06:00:25.488 UTC. The observing plan on this evening was based on the original LVC bayestar probability map. The updated LVC LALInference map disfavored most of the region observed on the first night. As a result GROWTH prepared a new observing plan for the second night (Andreoni et al. 2019b). This plan consisted of observing for ∼1.5 hr beginning at 22:51:57 UTC on 2019 May 10. Eighty exposures total were taken in the g, r, and z bands for 40 s each. Each filter visited roughly the same area of the sky, approximately 30 minutes apart, in order to eliminate moving objects. The 10σしぐま depths for each band are m_z = 20.58 mag, m_r = 21.72 mag, and m_g = 21.67 mag, where the average seeing was 133, the average airmass was 1.71, and the average attenuation due to cloud was 4%. These observations covered ∼65% of the probability region, as shown in Figure 1. Plans to follow-up this event for a third night were retracted due to the updated classification probability of this event. Our analysis uses only the exposures from the second night of observations as to include only the high probability region from the LALInference LVC map.

Zoom In Zoom Out Reset image size

Images from DECam were downloaded directly to Fermi National Accelerator Laboratory from Cerro Tololo Inter-American Observatory (CTIO) via the National Center for Supercomputing Applications (NCSA). Once the search exposures became available, we immediately started initial image processing, parallelized to run on a CCD by CCD basis, via the DESGW Search and Discovery Pipeline (Herner et al. 2020). The pipeline consists of three major stages: Single-Epoch (SE) processing, difference imaging, and post-processing.

SE processing (Morganson et al. 2018) consists of image correction and astrometric calibration. This stage uses SExtractor (Bertin & Arnouts 1996) to create a list of bright objects in each image, which is fed into scamp for astrometric calibration (Bertin 2006). We use the GAIA-DR2 catalog (Gaia Collaboration, et al. 2016, 2018) in this stage, which allows reductions of DES astrometric uncertainties to below 0.03^''. The outputs of SE processing are the inputs to the difference imaging (diffimg) stage, designed for the DES supernova diffimg pipeline (Kessler et al. 2015), but modified to work with wide-survey images. Diffimg subtracts one or more template images from the search image. Template images are taken in the same area of sky before the event, or after the event is expected to have faded if preexisting templates are unavailable. Our group is able to use all DES images as templates, including those not yet publicly released, as well as all public non-DES DECam images. Combining public DECam data and not-yet-released DES data, we are able to improve on the depth of template images by roughly $\sim \sqrt{2}$ within the DES footprint. After the diffimg subtraction is complete, post-processing takes candidate objects identified and applies a cut based on the machine-learning algorithm (autoscan) score  (Goldstein et al. 2015a, 2015b). Furthermore, forced photometry (described in Kessler et al. (2015)) is applied in this step, as well as host galaxy matching and additional requirements such as detections in multiple bands and/or on multiple nights. For S190510g, these requirements were only a detection in two or more exposures and a machine-learning score adequate to reject nonastrophysical sources. We eliminated known asteroids using the minor planet center, but since a second night of exposures were not taken in the region of interest, we did not eliminate other asteroids from our candidate list. Finally, the resulting candidate list was vetted via human inspection, as described in Section 3.2. The stamps for each of the DESGW candidates are shown in Figure 2.

Zoom In Zoom Out Reset image size

3.1.1. Pipeline Performance

In total, there were 1165 candidates identified after post-processing. The final candidate list was published in GCN 24480 at 12:24 pm 2019 May 11 UTC (Soares-Santos 2019). The primary cuts for our candidates require no SExtractor errors in image processing, such as masking of objects overlapping the transient or inability to measure the flux, and an autoscan score of at least 0.9 out of 1.0. This cut found 96 candidates (20 with autoscan score >0.95), while the final 11 were selected via visual inspection. The key properties we looked for when performing visual inspection is a host galaxy in the template image, a non-noisy template image, and no regions of over- or under-subtraction. We also took into consideration the possibility that the candidate could be an AGN since we are unable to resolve objects that are close to the center of the host galaxy and therefore disfavored stamps where the candidate is not distinguishable from the host galaxy. Further, we note that no candidate from our pipeline is fully dismissed until there is secondary follow-up or enough evidence to definitively categorize the object. For a single night of observations, our goal is to rapidly identify objects that are the most obvious candidates, then refine our search criteria as we observe more epochs.

Additionally, we matched candidates to hosts and used DES data to measure properties of the host, such as photometric redshift, absolute magnitude, stellar mass, and star formation rate, as well as the separation of the candidate and the host at the redshift of the nearest potential host galaxy. Photometric redshifts have been computed using Directional Neighborhood Fitting (DNF; De Vicente et al. 2016), while the galaxy properties have been computed using the Bayesian Model Averaging method as described in Palmese et al. (2020b). The coordinates and other information about each of our candidates can be found in Table 1, and information about their host galaxies is listed in Table 2.

Table 2.  Candidate Host Galaxies' Information

Name	Host Gal. Name	Sep	z	Log(M⋆)	Log(SFR)	Mi	Rank
		(kpc)		Log((M⊙))	Log((M⊙yr−1))
desgw-190510a	2MASS J06060625-3532351	32.1	0.106 ± 0.004	${11.069}_{-0.05}^{+0.19}$	−0.190	−23.043	3
desgw-190510b
desgw-190510c	DES J061124.4562-363104.494	97.7	0.202 ± 0.098	${9.028}_{-0.060}^{+0.055}$	−0.462	−20.071	9
desgw-190510d	WISEA J054914.81-355724.3	5.0	0.130 ± 0.021	${9.388}_{-0.09}^{+0.14}$	−0.502	−19.33	4
desgw-190510e	WISEA J055624.41-302817.8	24.6	0.163 ± 0.005	${10.829}_{-0.12}^{+0.05}$	−0.390	−21.796	8
desgw-190510f	2MASS J06091226-3452506	68.9	0.168 ± 0.003	${11.200}_{-0.034}^{+0.034}$	−0.156	−22.685	7
desgw-190510g	DES J060952.4784-340540.704	218.3	0.575 ± 0.174	${9.517}_{-0.14}^{+0.17}$	0.259	−20.557	10
desgw-190510h	2MASS J05510277-2757201	4.3	0.049 ± 0.002	${10.004}_{-0.052}^{+0.034}$	−1.388	−20.917	1
desgw-190510i	WISEA J060745.00-304928.7	55.7	0.193 ± 0.019	${10.341}_{-0.041}^{+0.052}$	0.387	−21.539	6
desgw-190510j	WISEA J060914.02-350858.5	3.1	0.134 ± 0.014	${9.69}_{-0.13}^{+0.14}$	−0.461	−19.927	5
desgw-190510k	2MASS J05483537-3559390	1.9	0.067 ± 0.002	${9.829}_{-0.20}^{+0.09}$	−0.272	−20.627	2

Note. Redshifts listed are mean photometric redshifts with 1σしぐま errors. All photometric data, star formation rates (SFR), stellar mass (M_⋆), and magnitudes are computed using DES Year 3 data. Additionally, the separation ("Sep") between candidate and host galaxy is calculated using the redshift of the galaxy. Galaxies are ranked based on their position in the sky and redshift, using the information provided in the skymap. Log indicates a logarithm in base 10.

Download table as:  ASCIITypeset image

The first stage of analysis, performed as exposures became available, presented 11 candidates (of which six were also detected by GROWTH) that were produced via the DESGW Search and Discovery Pipeline discussed in Section 3. Follow-up from other observatories is crucial for determining if a candidate is the GW counterpart through rejection of false positives. The Korea Microlensing Telescope Network (KMTNet) followed-up five of our candidates, desgw-190510a, desgw-190510c, desgw-190510i, desgw-190510j, and desgw-190510k (GCN 24493 and 24529; Im et al. 2019a, 2019b), at the KMTNet South Africa (SAAO), Chile (CTIO), and Australia (SSO) stations showing that each of these candidates did not have significant fading over ∼1 day, but did show very slow or no fading, therefore deeming these candidates likely supernovae. Additionally desgw-190510c was observed by Swift-XRT (GCN 24541; Evans et al. 2019), showing no XRT source found, as well as with Magellan (GCN 24511; Gomez et al. 2019), which found a broad feature consistent with H-αあるふぁ at a redshift of 0.06 and suggests a good match to a Type II SN approximately one week after peak brightness. Finally, desgw-190510h was initially detected by ATLAS on 2019 March 13 and later classified as a Type Ia SN at redshift 0.07 roughly a few days after maximum light by the Spectral Classification of Astronomical Transients (SCAT) survey and desgw190510-b was recorded by Gaia on 2019 January 30 and reported as a "blue hostless transient." This transient can also be seen in previous DES images dating about 2.5 yr ago, though with not enough information to classify it with certainty, thus we provide no host information in Table 2. This leaves only four candidates, desgw-190510d, e, f, and g, that were not classified by secondary follow-up, and thus still potential counterpart candidates.

The remaining information about each candidate that can be used to determine if a candidate is viable can be found in Table 2. The table reports photometric redshift, star formation rate, stellar mass, and absolute magnitude of the hosts, computed using DES Year 3 data (Abbott et al. 2018b). Furthermore, galaxies are ranked based on their probability of association, which can be computed using the skymap information, the galaxies' position, and redshift (Singer et al. 2016), assuming a flat ΛらむだCDM cosmology with H₀ = 70 km s⁻¹ Mpc⁻¹ and Ωおめが_m = 0.3.

Using the same exposures, the GROWTH collaboration reported a list of 13 candidates (GCN 24467; Andreoni et al. 2019a). Seven of the GROWTH candidates were not listed in the initial DESGW candidate list reported in GCN 24480. Candidates DG19bexl and DG19nouo were found in the final stages by our automated pipeline; DG19bexl did not pass the autoScan score cut (≥0.9) and DG19nouo was rejected due to visual inspection. Candidates DS19qcso and DG19llhk both had a detection in a single exposure, where two were required to be picked up as a candidate. The overlapping search exposures for these candidates failed in the HOTPANTS step of our pipeline. Reprocessing of these exposures with an updated (current) version of the DESGW pipeline did identify these candidates. Similarly, candidates DG19ukvo and DG19oahn were not found in our initial processing of the event due to HOTPANTS errors in all exposures. DG19oahn was later found in reprocessing, while DG19ukvo continued to have processing failures in two out of the three exposures. The fraction of missing candidates is consistent with the overall failure rate of 28% for all jobs that were submitted on that night, where ∼15% of total jobs failed due to issues in HOTPANTS. These failures are largely due to the observing conditions described in Section 2.2. Finally, candidate DG19ootl was never found in our pipeline. The templates used for this exposure were taken from not yet publicly available DES images and thus did not show any source in the difference image. Candidates, including those initially detected only by GROWTH, are shown in Figure 2.

To better understand our search efficiency, we performed an offline analysis using SuperNova ANAlysis software suite (SNANA) (Kessler et al. 2009). These simulations produce SN and KN light curves as they would be observed during our observations. Each KN simulation randomly assigns an ejecta mass, ejecta velocity, and lanthanide fraction based on the Kasen KN model (Kasen et al. 2017), as well as host galaxy extinction between 0 and 3 mag (Cardelli et al. 1989). The SN simulations use the SALT2 model for SN Ia (Guy et al. 2010) and templates for the core collapse SN (SN CC) are taken from Kessler et al. (2010) and Jones et al. (2018). Additionally we note that these light curves are modeled assuming face-on emission of the observed component or spherically symmetric emission. While studies have shown the importance of viewing angle dependence on KN brightness (Dhawan et al. 2020; Kawaguchi et al. 2020; Korobkin et al. 2020), this approach was selected for simplicity.

Using these simulations, we computed the detection efficiency for each KN model given our observing conditions, the results of which are shown in Figure 3. The KN simulations used for this analysis produce events that use a distance distribution consistent with that reported by the LVC as well as being located within the 65% probability area that was surveyed. The efficiency of each model, represents the fraction of light curves that are detected to be brighter than our 5σしぐま limiting magnitude at the time of DECam observations.

Zoom In Zoom Out Reset image size

Next, we used these simulations to examine the color–magnitude space for both KN and SN (Figure 4). For this analysis, we use both KN and SN simulations. Here we require the detected object to be brighter than our 5σしぐま limiting magnitude. Additionally, we require the object's host-galaxy photometric redshift to be consistent with the LVC luminosity distance posterior at the 3σしぐま confidence level. Additionally, the simulated SNe were distributed in redshift according to the measured volumetric rates of SNe-Ia and SNe-CC.

Zoom In Zoom Out Reset image size

Figure 3 shows the likelihood that we would have been able to detect a KN produced by this event given the observing conditions and depth of observations. Here we show all possible sets of KN parameters and note that a GW170817-like KN follows a two component model, red and blue, where the blue component is dominant at early times (i.e., up to ∼2.5 days after merger; Kasen et al. 2017) in the light-curve evolution. Assuming S190510g is a GW170817-like KN viewed from the same orientation located within our exposures, our simulations show that we would have a 99% chance of detecting the counterpart KN. However, a wide range of KN models would have been outside of our sensitivity range and thus unobservable.

While we have the ability to detect such a source, it is challenging to determine a candidate to be KN or SN with a single night of observations in the absence of spectroscopic information. To demonstrate the difficulty of this task we examined the color–magnitude space of the simulated KN and SN events. All KN simulations are shown as the blue contours (indicating 50% and 90% density of simulations) in the left panel of Figure 4, with the parameters for the blue component of GW170817 (ejecta velocity = 0.3c, lanthanide fraction = 10⁻⁴, ejecta mass = 0.025M⊙, and assuming spherically symmetric emission) highlighted as orange contours. Meanwhile the color–magnitude distribution of SN simulations is shown by the green contour on the right panel of Figure 4. All DESGW S190510g candidates from this event (depicted as red crosses in Figure 4) fall within the possible 90% color–magnitude regions of SN events. For a KN roughly one day after burst, and given only this color–magnitude information, each of these candidates could be either an SN or a KN.

In the first half of the O3 observing run, most of the events that included a neutron star did not have a good localization (i.e., hundreds of deg²) as well as being far away (>200 Mpc) when compared to GW170817. While it would be ideal to cover 100% of the localization area with multiple filters, limited telescope time and poor localization maps make this very challenging. In the following, we show that prioritizing sufficiently deep images as opposed to covering large areas and/or using multiple filters, will result in a higher chance of detecting counterparts.

To show how many events it would take to have a 50% (99%) chance of detecting one counterpart, we have to consider the cumulative probability inside the LVC localization map that was observed (Σしぐま_spatial), the fraction of DECam that was live during observations (_camera), the probability that the event is astrophysical in nature (_real), and our likelihood of being able to detect a KN at that distance given the observing conditions (i.e., the fraction of simulated light curves that are brighter than our 5σしぐま depth) (_efficiency)

$$\begin{eqnarray}&&{P}_{i}={{\rm{\Sigma }}}_{\mathrm{spatial}}\times {\epsilon }_{\mathrm{camera}}\times {\epsilon }_{\mathrm{efficiency}}\times {\epsilon }_{\mathrm{real}}\end{eqnarray} \tag{ 1 }$$

$$\begin{eqnarray}&&{P}_{\mathrm{one}}=1-\displaystyle \prod _{i}^{N}(1-{P}_{i})\end{eqnarray} \tag{ 2 }$$

Here, _Pi is the probability of being able to detect a KN from a single GW event. P_one is the cumulative probability of being able to detect a single counterpart given N GW events (Annis & Soares-Santos 2016). For this calculation, we find that if we assume there is a kilonova associated with S190510g that is GW170817-like, i.e., _efficiency = 0.993, we would need to observe 3 (19) identical events with Σしぐま_spatial = 0.65, _camera = 0.8, _real = 0.42, and _efficiency = 0.993 in order to have 50% (99%) probability of identifying the event using the current strategy. Since there is no way of knowing that the event will have a light curve similar to GW170817, we also calculate this using the average efficiency value of all KN parameters, _efficiency = 0.553, with all other parameters the same. Here we find that we would need 6 (36) events to reach 50% (99%) likelihood of detecting the counterpart.

We then repeat this calculation assuming the observing strategy uses one filter instead of three. If we conserve the telescope time used and area surveyed per event, we can then increase the exposure time from 40 to 170 s. In this scenario, the efficiency for a GW170817-like KN is 0.995, meaning we would again need 3 (19) events to have 50% (99%) likelihood of detection. Using the average efficiency in this scenario though, 0.742, we would only need 4 (27) events to have 50% (99%) likelihood of detecting a counterpart. By increasing the depth of our observations, we become sensitive to more KN models and will thus need to observe fewer total GW events to have a high probability of making a detection.