The Atacama Cosmology Telescope: Systematic Transient Search of 3 Day Maps

We conduct a systematic search for transients in 3 yr of data (2017–2019) from the Atacama Cosmology Telescope (ACT). ACT covers 40% of the sky at three bands spanning from 77–277 GHz. Analysis of 3 day mean-subtracted sky maps, which were match filtered for point sources, yielded 29 transient detections. Eight of these transients are due to known asteroids, and three others were previously published. Four of these events occur in areas with poor noise models and thus we cannot be confident they are real transients. We are left with 14 new transient events occurring at 11 unique locations. All of these events are associated with either rotationally variable stars or cool stars. Ten events have flat or falling spectra indicating radiation from synchrotron emission. One event has a rising spectrum indicating a different engine for the flare.


INTRODUCTION
Transient astronomical sources comprise some of the most dramatic and extreme astrophysical events, such as supernovae, binary star mergers, active galactic nuclei (AGN), gamma ray bursts (GRB), and tidal distruption events (TDE).Continuous monitoring of large sky regions to detect transients has been a major motivation behind the upcoming Rubin Observatory (Ivezić et al. 2019) and several other efforts at optical wavelengths, such as the Zwicky Transient Facility (Masci et al. 2018) and ASAS-SN (Christy et al. 2022).Transient astronomy has also received a large increase in attention with the first gravitational wave sources detected by LIGO (Abbott et al. 2016), and detection of sources in both gravitational waves and multiple electromagnetic wave bands.Wide-field millimeter wavelength surveys have focused on mapping the unchanging microwave background for cosmological applications, but recently some have reached the combination of high angular resolution, wide survey area and sensitivity required to start probing the time-variable millimeter sky.
In recent years there have been a growing number of millimeter transient detections.In 2016, the South Pole Telescope (SPT) reported one candidate event observed during the 2012-2013 season at a 2.6σ significance with a peak flux of 16.5±2.4mJy in the 150 GHz band (Whitehorn et al. 2016).This event was broadly consistent with a GRB orphan afterglow.Kuno et al. (2004) reported a detection at 90 GHz of the afterglow of GRB 030329 (z = 0.17) with a flux around 65mJ.More than a decade later, Laskar et al. (2019) detected a polarized reverse shock from GRB 190114C at a frequency of 97.5 GHz.Other types types of events are also seen in the millimeter.For instance, Yuan et al. (2016) reported millimeter emission from the TDE IGR J12580+0134.Ho et al. (2019) studied the unprecedented transient AT2018cow in radio and submillimeter bands, finding its emission to be inconsistent with synchrotron emission or Compton scattering of UVIOR photons, indicating a new class of millimeter transients.
There are also many examples of stellar flares seen in the millimeter.ACT reported three bright transient sources coincident with stars (Naess et al. 2021a) and SPT reported 15 additional transient events (Guns et al. 2021).Most of the sources reported from SPT were consistent with stellar activity, except for two of their events which may be extragalactic in origin.Millimeter stellar flares have also been reported in other works such as Bower et al. (2003), Brown & Brown (2006), Massi et al. (2006), Mairs et al. (2019) and are most likely due to stellar magnetic field reconnection.
In this work, we complete a systematic search for transient events with three years' of ACT data using 3-day stacked maps.In Section 2, we introduce the ACT survey and the maps used for this analysis.In Section 3, we describe how the transient events are detected and how data cuts applied to the sample.In Section 4, we present our findings and possible counterparts for each transient.After all cuts are applied we are left with 29 events.Three of these events are redetections from Naess et al. (2021a).Eight appear to be from asteroids that were not masked in the 3-day maps and four occur in areas with poor noise models and cannot be confirmed as real transients.This leaves 14 new transient detections at 11 unique positions on the sky.Most of the events have clear stellar associations.In Section 5, we summarize the nature of these counterparts.

DATA
The Atacama Cosmology Telescope (ACT) was a 6 m off-axis Gregorian telescope located in the Atacama Desert in Chile at an altitude of 5200 m (Fowler et al. 2007).AdvACT was the third generation ACT instrument.It housed three optics tubes with a total field of view (FOV) spanning ∼1.5 • (Thornton et al. 2016), each containing a set of lenses, low-pass filters, and a single AdvACT detector array (Ho et al. 2017;Choi et al. 2018).The ACT dataset considered here covers the three bandpasses f090 (77-112 GHz), f150 (124-172 GHz) and f220 (182-277 GHz) using the three dichroic polarization-sensitive transition edge sensor (TES) bolometer arrays PA4 (f150+f220), PA5 (f090+f150) and PA6 (f090+f150).During observations, the telescope scanned the sky in azimuth at a fixed elevation at a scan speed of ∼1.5 deg per second.When the sky was rising, it took ∼6 minutes for a point on the sky to be gradually swept across by the detectors in PA4 and PA5.Then, after ∼3 minutes this process would repeat for the detectors in PA6.This order was reversed when the sky was setting.The bolometer signals were stored in time-ordered data (TOD) files, each containing a roughly 10 minute time series for all detectors in a single array.Detailed data selection and calibration process is described in Aiola et al. (2020).
We search for transient events using 3-day maps originally created to search for Planet 9 (Naess et al. 2021b).The search covered 18000 deg 2 of the sky using ACT data taken from 2017 to 2019.Although these maps were not designed for a systematic transient search, their short timescale is well suited for this purpose.The maps are made by dividing the daytime and nighttime TODs taken by each of the three detector array at each of the two frequency bands into approximately 3-day long data chunks.Therefore, there are twelve maps (day and night time data within three arrays with two bands) for each 3-day period.The maps are made by calculating the maximum likelihood estimate for the CMB temperature from TOD.The map is then matched filtered to look for point sources, which gives an estimation of flux density and S/N at each pixel.Detailed methods are described in Naess et al. (2021b).For each 3-day map, a mean sky subtraction at the corresponding frequency band is implemented so only the time-varying signal remains.Regions around the planets and bright asteroids are masked to avoid contamination from planet sidelobes and false transient detection from known moving astronomical objects.Since the 3-day maps were made as part of a search for Planet 9, some tradeoffs are made that are suitable when performing large stacks of maps, but at times not ideal when analyzing maps in isolation.In particular the maps' matched filter numerator was computed directly in time domain while the denominator was computed in map-space.This approach is fast, but only accurate if the time-domain noise model and map-space noise model are consistent.In practice this is an approximation which introduces a bias in areas where the map hit-density changes rapidly from pixel to pixel.This does not happen when performing large stacks like in the Planet 9 search, but it is quite common when looking for objects in the individual 3-day maps, resulting in many spurious detections (see Section 3.1).Furthermore, the matched filtered 3-day maps are stored at 1 arcminute resolution, which was an acceptable performance tradeoff for the Planet 9 search (where this was a minor contribution to the overall smoothing budget), but is suboptimal for blind transient detection.The ways in which the position uncertainty introduced by the selected resolution are handled is detailed in Section 4.1.We plan to resolve these deficiencies in a future paper using maps tailored for transient detection.Any pixel or group of pixels with S/N > 5 in the 3-day sky-subtracted maps is considered a transient candidate.
For each set of 3-day S/N and flux density maps, we first mask pixels with S/N < 5, then apply this mask to the corresponding flux density map (Figure 1).The positions of candidates are evaluated to be the flux-weighted centroids.This initial detection finds 332,333 candidate events.If a detection is a real transient event and not simply a noise fluctuation, we will likely detect it in more than one array.Thus, we cross match each candidate between detector arrays with a matching distance of 1.5 arcminutes, which is 1.5 times the resolution of the 3-day maps.We only keep candidates that appear in at least two frequency and array combinations.This cuts 76% of the initial detections.
Even with this cut applied, there are still many spurious detections grouped in clusters and along map edges.We perform three additional geometric cuts, requiring a candidate to be cut in all detected frequency and array combinations to exclude it from the analysis.The results of these cuts are summarized in Table 1. 1 We apply these cuts in tandem to the frequency and array requirement, thus the following statistics include all candidates.
The first geometric cut is motivated by the observation that many of the 3-day exhibit a stripy pattern of spurious detections along the scanning direction.These false detections are caused by a systematic underestimation of the variance when the coverage is uneven due to low hit counts.The stripy regions are reliably identified by searching for detections with nearby neighbours within the same 3-day day map.This approach is mo-Figure 2. Left: A histogram of the distance to each detection's nearest neighbor with a binsize of 5 arcminutes.The peak close to zero indicates there are clusters of spurious detections in many of the maps so we cut any candidate with a nearest neighbor of 20 arcminutes or less.Center: A histogram of each candidate's distance from the nearest zero variance contour, defined to be a variance of less than 1.5 × 10 −5 K, with a bin size of 3 pixels.There is a large peak of candidates near zero variance contours which drops after 3 pixels.We mask out to 5 pixels, cutting 74% of all sources.Right: A histogram of each candidate's distance from the map edge in pixel units with a bin size of 3 pixels.As expected, there is an excess number of candidates near the edge of the map as the map edges are noisy and so appear variable when sampled every three days.At a mask size of 5 pixels, we cut off the peak of candidates near the edge.
tivated by the lefthand plot in Figure 2 which shows a histogram of each candidate's nearest neighbour: there is a clear excess of detections with neighbours within ∼1 deg, which would not be expected for real detections since extragalactic transients should be spatially uncorrelated.Based on this histogram, we cut any candidates with a neighbor within 20 arcminutes as there is an overdensity of detections with a closest neighbor within this range.This cut removes 60% of all candidates.
Second, we see a similar stripy pattern along zero variance contours.During the map making process, some approximations result in streaks of spurious detections with vanishing variance and therefore a high S/N .In the center plot in Figure 2 we plot a histogram of the candidates' distances from a zero variance region and cut any candidate within 5 pixels of these features (about 74% of all candidates).Lastly, we cut any candidate near a map edge.Map edges are especially noisy and appear to be variable when sampled every three days.In the right-hand plot in Figure 2 we plot a histogram of the pixel distance of each candidate from the edges of the map.We see a spike in candidates within five pixels from the map edge and so we mask this region.This cuts applies to 17% of all candidates.
After these cuts are applied an internal cross match is performed to find repeating events.Candidates with positions within one arcminute of each other are considered the same object.This leaves us with 667 independent candidates.

Candidate Verification
For each of the 667 candidates, we calculate the mean flux by applying a matched filter to the mean sky map using data from 2017 to 2021, and mask candidates with mean flux > 50 mJy or < −50 mJy.The high-meanflux candidates are likely variable point sources such as AGN or dusty star forming galaxies (DSFG).These high-mean-flux candidates are already detected in our standard point source catalogs, and there are dedicated studies in preparation on the light curves of these candidates.In this paper we will concentrate on candidates that are not detectable in the mean sky maps.The candidates with negative mean fluxes are located close to bright point sources.The flux density around this region is negative due to the effect of the matched filter that acts as high pass filter, and will increase when the flux of the corresponding point source decreases, resulting in a false detection.There are 82 candidates left after applying the cut on mean flux.We then make light curves to confirm these detections, requiring a S/N > 3.5 in at least two arrays or in one array at both frequency bands.These light curves are made from forced photometry, which gives a per-detector flux and flux error, and then evaluating a weighted mean flux as the arraywise result at each frequency band for each scan.This analysis further cuts 41 candidates.
For the rest of the candidates, we made 2 • by 2 • intensity maps (shown in Figure 7) from TODs covering a three day period (or shorter if the scanning cadence is especially high) centered on each event.These maps, referred as "thumbnail maps," have 0.25-arcminute resolution and are centered at the candidate positions.We visually inspected these maps both before and after matched filtering2 and remove candidates with with an extended or irregular shape, since we expect all our transients to be point sources.This issue may be caused by poor data quality causing arcminute scale fluctuations.For a few of the remaining candidates, we observed that the raw and filtered maps appear discrepant with each other, e.g. a supposedly strong detection having no visible counterpart in the maps before filtering; or there was an overdensity of > 3σ peaks near the event in the filtered maps, which should not happen if the noise model is correct.We do not discard these events, but classify them as low quality candidates.All in all, we are left with 29 transient detections.Eight of these events are due to asteroids (see section 4.1), three were previously detected, and four are low quality candidates.

Characterization
Two groups of our transient events appeared to be close in time and space to each other (see Table 2).Using the IAU Minor Planet Center NEOChecker webtool,3 we found three events to be coincident with asteroid 10 Hygiea and five events to be coincident with asteroid 511 Davida.In Figure 4, we plot the paths of these asteroids with our events overlayed.All eight events appear consistent in space and time with the asteroid observations which were not masked in our maps.For future blind transient searches we must be prepared for the possibility of detections from moving objects such as asteroids.This will become especially important for real-time follow-up observations.A paper dedicated to studying asteroids in ACT data is in preparation.The properties of the remaining 21 transient events are shown in Table 3.Three single detections labeled as "N" were previously published in Naess et al. (2021a).Among the newly detected events, 12 are single events and three repeat twice, with time intervals ranging from one month to one year.The low (1 arcminute) resolution of the 3-day maps potentially increases the candidates' position error, so we re-estimate the candidate positions using a method similar to the one used in the initial detection on the higher-resolution thumbnail maps described in Section 3.2.We first subtract the corresponding f220/f150/f090 mean sky map and apply a matched filter.The final position for each candidate is the inverse-variance-weighted average among individual positions, using the ratio between half maximum of the beam and the appropriate S/N as weights.The position accuracy is evaluated by the square root of the variance of weighted mean.As mentioned in Section 3.2, four out of the 12 single events are difficult to determine if they are real transients due to unusual noise patterns on the intensity maps.We present these events at the bottom of the table without assigning a formal name.
Flux densities for each scan are evaluated at the refined positions and used to generate light curves (Figure 5) using the same method described in Section 3.2.Higher-resolution light curves (Figure 6) are made by dividing detectors in each array into four groups by the order of observation time to study the minute by minute change in flux density.When the flux density values are available in both bands of PA4 and PA5, the spectral index α is evaluated by the best-fit of a power law S ν ∝ ν α of frequency ν to the flux measurement S ν at the three different bands, taking into account the variance of each band flux.Under this circumstance, PA6 is omitted because it scans across the candidates approximately 10 minutes earlier (or later when the sky is rising) than PA4 and PA5.Events 2 and 15 do not have f220 measurements, so α is directly calculated by using the f150 and f090 flux measured by PA5 or PA6, depending on which array measures the peak flux.

Counterparts
We searched for counterparts for each transient event in the SIMBAD 5 (Wenger et al. 2000) database.The results of this search are summarized in Table 4.We searched for any Gaia objects (Mou & Webster 2021) 5 https://simbad.cds.unistra.fr/simbad/within one arcminute of each event and list the SIM-BAD identification.We calculate the probability of a chance association given the density of Gaia stars (Mou & Webster 2021) with the counterpart's magnitude or lower within the counterpart's separation of the transient's position.We exclude any counterparts with a chance association probability greater than ten percent.We also searched for counterparts of the transients found in Naess et al. (2021a) that are recovered in this analysis.The associations agree with the original findings.
In addition to the Gaia source catalog, we search for galactic objects in the Gaia extragalactic catalog (Mou & Webster 2021) using the VizieR (Ochsenbein et al. 2000).Only events 10 and 14 have associations from this catalog.Event 10 is coincident with an AGN with a 1.1 percent chance of a false association (see Table 4).Event 14 is coincident with a RR Lyrae star with a 2.4 percent chance of a false association.This association is not listed because there is a much more obvious Gaia star associated with event 14.
In Figure 8 we plot the positions of the transients overlayed on combined ACT data from 2007-2019.This plot includes color images in the optical from the Digitized Sky Surveys (DSS) 67 that show there are bright stellar counterparts for most of the events.

DISCUSSION AND CONCLUSIONS
We have demonstrated a pipeline of transient detection and characterization using ACT data, and presented a catalog of 21 millimeter transients from 3-day maps, including three of which that were previously detected, and four events that cannot be confirmed as real transients.After assessing the spectral indices and counterpart associations of these events, two classes of objects emerge: flares associated with stars with flat or falling spectra, indicating radiation from sychrotron emission, and stellar flares with rising spectra, indicating thermal emission.Performing a statistical analysis on such a small sample size is difficult, but in the future as we discover more millimeter stellar transients, we will be able to better characterize them.For example, we will be able to determine whether these events reside near the galactic plane or scattered across the sky, or whether two classes of transients, rising versus falling spectra, really exist or if there is simply a wide distribution of spectral indices.Tests of these statistics with our small number of events are inconclusive.Events 3, 4a, 4b, 5a, 5b, 11, 12a, 12b, and 15 are associated with rotating variable stars.These stars have large dark spots on their surface from magnetic activity in their chromospheres.These spots cause intrinsic variability in flux in optical and IR wavelengths but also cause flares in radio wavelengths from synchrotron radiation (Hjellming & Gibson 1980).Events 4, 11, 12, and 15 are coincident with a special type of rotating variable stars called RS Canum Venaticorum variables (RSCVn).These are binary systems that also exhibit variability due to dark spots (See Hall 1976, Zeilik et al. 1979).Similar events were seen in the millimeter systematic transient search from SPT-3G (Guns et al. 2021).
Event 1 is also associated with a binary system.Although it is not classified as rotationally variable, it most likely flares in the microwave with a similar mechanism.This association is similar to ACT-T J200758+160954 (N3) which was previously published in Naess et al. (2021a).
Events 7, 8, and 15 are all coincident with cool G, K, and M type stars.Event 14 is associated with a Gaia star but its spectral type is unknown.Since all of these events also have negative spectral indices, these flares are most likely from magnetic activity producing synchrotron radiation.This is consistent with findings from Guns et al. (2021) who reported microwave flares from M and K dwarfs.These cool stars have convective envelopes which cause increased magnetic activity but do not exhibit the same dark spots found in rotationally variable stars (Yang et al. 2017).
Events 6, N1, and N2 all have rising spectra which are associated with thermal emission.In these cases, some mechanism other than synchrotron radiation is driving the flares.N1 and N2 are associated with cool M and K stars respectively but the spectral types of the star associations for 6 and 9 are unknown.Guns et al. (2021) also observed two events with M-dwarf counterparts with rising spectra indicating this is a common class of millimeter transients.
Although we cannot be certain events 2, 9, 10, and 13 are real transients, we still present possible counterparts for them in Table 4.There is a six percent chance that the Gaia counterpart for Event 2 is a chance association.If the association is correct this transient is likely due to synchrotron radiation from stellar magnetic fields.The field around Event 10 is too dense to pinpoint a stellar counterpart.However, there is a small chance this event is associated with an AGN.Eftekhari et al. (2022) predicts ACT will see on the order of ten gamma ray bursts but this event is not associated with any known GRB flares listed in the Fermi All-sky Variability Analysis (Abdollahi et al. 2017).The associations for Events 9 and 13, are also unclear.Both events have rising spectra so they would most likely be similar to events 6, N1, and N2.The Zwicky Transient Facility (ZTF; Masci et al. 2018) detected a flare with the same Gaia counterpart associated with event 13 twelve days after and 23 away from the ACT detection.ZTF also detects repeating flares that are 34 away from Event 9, although they did not occur close in time to the ACT detection.8 .This sample likely only represents a fraction of microwave transient events present in the ACT data.The 3-day maps were not made for a systematic transient search.Recently, single observation maps that span more seasons than the 3-day maps have been made for the purpose of ACT time domain study.Major advantages of the single observation maps include the relative consistency in noise performance in each map and the freedom of stacking single scan maps for different periods of time.We will perform a similar analysis to search for transient events within those maps to gain a more accurate event rate for ACT.This will inform current and future large sky-area millimeter surveys such as The South Pole Telescope Carlstrom et al. (2011), Simons Observatory (Ade et al. 2019), CCAT-prime (CCAT-Prime Collaboration et al. 2023) and CMB-S4 (Abazajian et al. 2019) on how to best detect and study these events.

ACKNOWLEDGEMENTS
The Authors would like to thank Carles Badenes for his helpful discussions and comments.
This work was supported by the U.S. National Science Foundation through awards AST-0408698, AST-0965625, and AST-1440226 for the ACT project, as well as awards PHY-0355328, PHY-0855887 and PHY-1214379.Funding was also provided by Princeton University, the University of Pennsylvania, and a Canada Foundation for Innovation (CFI) award to UBC.ACT operates in the Parque Astronómico Atacama in northern Chile under the auspices of the Agencia Nacional de Investigación y Desarrollo (ANID; formerly Comisión Nacional de Investigación Científica y Tecnológica de Chile, or CONICYT).
Computations were performed using Princeton Research Computing resources at Princeton University.This research has made use of data and/or services provided by the International Astronomical Union's Minor Planet Center.This research has made use of the SIM-BAD database, operated at CDS, Strasbourg, France.Table 3. Properties of the transient events.The candidates are listed in the order of the detection time (or the detection of the first event for the repeating candidates).Three single detections labeled as "N" are the redetections of the transients published in Naess et al. (2021a).It is difficult to fit for rise and fall time using light curves due to uneven scanning cadence.Instead, we examine if the scan right before and after the scans with the peak flux density has a > 5σ detection, and calculate the time interval in between the scans.The rise time of Event 4a is left empty because the time gap between the peak scan and the scan right before is longer than 50 days.Spectral index is evaluated as described in text, using array-wise flux density values, except for candidate 13.This transient event completed the rise and fall process within the time the sky took to drift across the array.We therefore evaluate the spectral index using flux values taken by only one quarter of detectors in each array, to capture the peak flux.Field too dense to identify stellar counterpart Table 4. Possible counterparts for each transient event from the SIMBAD database.If known, the spectral type is given next to the object type.The chance of a false association is calculated using the density of Gaia sources with the same magnitude or brighter of the counterpart.If the Gaia object is not found in the SIMBAD database, then the Gaia identification number is listed.All separations are calculated using the Gaia coordinates.This table also includes counterparts associated with the three previously published ACT transients from Naess et al. (2021a).When two stars from the same system are resolved, as is the case with events 7 and 12, we quote the average separation weighted by one minus the chance association and we list the lowest of the two chance associations.
Initial Detection and Spurious Candidate Cuts

Figure 1 .
Figure 1.Process of the initial detection, with each plot showing a 0.3 deg by 0.3 deg map.The first step is to make a mask (middle) on S/N map (left) selecting pixels that have S/N > 5.The mask is then applied to the flux map (right), and the candidate position, shown as the red cross mark, is evaluated as the center of mass weighted by the flux values within the selected pixels. 4

Figure 3 .
Figure 3. 10 x10 3-day thumbnail maps for each transient.The upper row is the intensity map with ±5000 µK color range.The bottom row is the S/N map after applying a matched filter, with ±5 color range.Due to the conjugate gradient iteration used to solve the maximum-likelihood map-making equation only being run for 10 steps, these maps are effectively mildly highpass filtered.The affected scales have negligible weight in the matched filter.Events 2, 9, 10, 13 at the bottom of the table are the four events that are difficult to determine if they are real transients.

Figure 4 .
Figure 4. Left: The three transient events consistent with Hygiea observations.Right: The five transient events consistent with Davida observations.All of the plotted positions and observation times are consistent with the asteroids' paths.The position errors are on the order of 0.1 arcminutes.

Figure 5 .
Figure 5.Light curves for 18 detected transient events on day time scales from the peak.Each frequency is denoted by a different color.We see that in most cases the peak is correlated with all frequencies.

Figure 8 .
Figure 8. Combined ACT data from 2007-2019 overlayed with the transient positions.Images from DSS at each location, indicated by a in the center of each image, are also shown with a one arcminute contour plotted for scale.The majority of the transient events have bright stellar counterparts seen in the optical.

Table 1 .
This table summarizes the three geometric data cuts applied to each 3-day map after the requirement that a candidate appear in at least two frequency or array combinations is applied.The first column quotes the number of candidates remaining after that cut, the second gives the fraction of all candidates that do not pass these cuts, and the last column gives the fraction of the total number of pixels masked by each cut.Note that these cuts are done in tandem and are independent from each other.The neighbor cut removes candidates with neighbors in the same map within 20 arcminutes, the zero variance contour (ZVC) cut masks candidates within five pixels of zero variance contours, and the edge cut masks candidates within five pixels of the edge.
table are the four events that are difficult to determine if they are real transients.

Table 2 .
Eight events associated with asteroids.
table are the four events that are difficult to determine if they are real transients.