Korea Microlensing Telescope Network Microlensing Events from 2015: Event-Finding Algorithm, Vetting, and Photometry
D.-J. Kim, H.-W. Kim, K.-H. Hwang, M. D. Albrow, S.-J.Chung, A. Gould, C. Han, Y.K. Jung, Y.-H. Ryu, I.-G. Shin, J. C. Yee, W. Zhu, S.-M. Cha, S.-L. Kim, C.-U. Lee, Y. Lee, B.-G. Park, R. W. Pogge
aa r X i v : . [ a s t r o - ph . E P ] D ec Korea Microlensing Telescope Network Microlensing Events from2015: Event-Finding Algorithm, Vetting, and Photometry
D.-J. Kim , H.-W. Kim , K.-H. Hwang andM. D. Albrow , S.-J.Chung , A. Gould , , , C. Han , Y. K. Jung , Y.-H. Ryu ,I.-G. Shin , J. C. Yee , W. Zhu , S.-M. Cha , S.-L. Kim , C.-U. Lee , D.-J.Lee , Y. Lee , B.-G. Park , R. W. Pogge (The KMTNet Collaboration) Korea Astronomy and Space Science Institute, Daejeon 34055, Republic of Korea University of Canterbury, Department of Physics and Astronomy, Private Bag 3800,Christchurch 8020, New Zealand Max-Planck-Institute for Astronomy, K¨onigstuhl 16, 59116 Heidelberg, Germany Department of Astronomy, Ohio State University, 130 W. 18th Ave., Columbus, OH32210, USA Department of Physics, Chungbuk National University, Cheongju 28644, Republic ofKorea Harvard-Smithsonian Center for Astrophysics, 50 Garden St., Cambridge, MA 02128,USA
ABSTRACT
We present microlensing events in the 2015 Korea Microlensing TelescopeNetwork (KMTNet) data and our procedure for identifying these events. In par-ticular, candidates were detected with a novel “completed event” microlensingevent-finder algorithm. The algorithm works by making linear fits to a ( t , t eff , u )grid of point-lens microlensing models. This approach is rendered computation-ally efficient by restricting u to just two values (0 and 1), which we show isquite adequate. The implementation presented here is specifically tailored to thecommission-year character of the 2015 data, but the algorithm is quite generaland has already been applied to a completely different (non-KMTNet) data set.We outline expected improvements for 2016 and future KMTNet data. The lightcurves of the 660 “clear microlensing” and 182 “possible microlensing” eventsthat were found in 2015 are presented along with our policy for their publicrelease. Subject headings: gravitational lensing: micro – methods: numerical – planetarysystems 2 –
1. Introduction
Gould & Loeb (1992) originally advocated a two-step approach for finding planets bygravitational microlensing. In the first step, a wide area survey would monitor 10s or 100s ofmillion of stars roughly once per night using a wide-angle camera in order to find microlensingevents. Then, in the second step, individual events found in the first step would be monitoredat high cadence from several longitudes using narrow angle cameras. That is, the cadence ofeach survey would be matched to the timescale of the effects being sought: a few dozen pointsduring the roughly month-long microlensing events from the first survey and a few dozenpoints over the day-long (or shorter) planetary perturbations from the second survey. AsGould & Loeb (1992) specifically pointed out, this approach required “microlensing alerts”,i.e., the recognition and public notification of microlensing events while they were still inprogress, and preferably before they peaked.This approach was in fact adopted by the microlensing community and led to the de-tection of several dozen planets. Right from the beginning, however, (including the veryfirst microlensing planet OGLE-2003-BLG-235, Bond et al. 2004) planets were detected bythe surveys themselves, without any “second step” follow-up observations. As the surveyswent through several generations of improvements, such survey-only detections became morecommon, e.g., Poleski et al. (2014). Nevertheless, given that many of the planets found dorequire a second step, the same “microlensing alert” mode of event detection, pioneered bythe Optical Gravitational Lensing Experiment (OGLE) group (Udalski et al. 1994; Udalski2003), remained the main practical method by which events were discovered.Although the “microlensing alert” system is the main path to event detection, the alter-nate approach of finding “completed events” in the data set has been present from the birthof the field. Both the MACHO and EROS collaborations developed algorithms for searchingthrough archival data to find microlensing events (Alcock et al. 1997; Afonso et al. 2003;Hamadache et al. 2006), and in particular, the very first microlensing event MACHO-LMC-1 was found this way (Alcock et al. 1993). One important advantage of this approach is thatit has relatively little reliance on human input and therefore can (mostly) be modeled as anobjective algorithm, which then permits objective estimates of microlensing event rates (orplanet rates). Such algorithms have been used to measure the microlensing optical depthin the LMC and SMC (Wyrzykowski, L., et al. 2009; Wyrzykowski et al. 2010, 2011a,b).Wyrzykowski et al. (2015) specifically applied this approach to OGLE-III observations ofthe Galactic bulge, which is the main (so far, only) field where microlensing planets arediscovered.Directly opposite the Gould & Loeb (1992) observing regime would be a very high-cadence survey with multiple sites allowing continuous monitoring of the Galactic bulge 3 –without the need for any followup observations. The first survey of this type was a col-laboration between the OGLE, MOA, and Wise observatories (Shvartzvald & Maoz 2012;Shvartzvald et al. 2014).As originally conceived, the Korea Microlensing Telescope Network (KMTNet, Kim et al.2016) would also lie in this regime. KMTNet consists of three 1.6m telescopes, each equippedwith 4 deg cameras, and located on three southern continents, CTIO (KMTC, South Amer-ica), SAAO (KMTS, Africa), and SSO (KMTA, Australia). According to the original plan,which was basically implemented in 2015, it would observe four fields (16 deg ) continuouslywith a cadence of Γ = 6 hr − . Hence, there would be virtually no point in follow-up obser-vations, and therefore no point in microlensing alerts. This in turn implied that KMTNetshould focus on finding completed events, both because it is easier than finding events inreal time and because (as noted above) of the potential of such an event-finding algorithmfor measuring rates.In fact, the above paragraph notwithstanding, there would be many potential applica-tions for KMTNet alerts. Most of these stem from the fact that KMTNet has abandoned itsoriginal strategy as implemented in 2015, in favor of the layered approach pioneered by theOGLE group. Currently, (3,7,11,3) fields are observed at cadences Γ = (4 , , . , .
2) hr − .In all but the highest cadence fields, high-magnification events could be profitably followed-up at substantially higher cadence. Moreover, if anomalies could be alerted in real time inthese fields, then detected planets could be characterized much better. Actually, alerts forthe highest cadences fields would be exceptionally important for the next few years due tothe emergence of Spitzer microlensing, which was not at all anticipated at the time KMT-Net was conceived in 2004. Because
Spitzer is in solar orbit, synoptic
Spitzer observationscan yield “microlens parallaxes”, which are critical for characterizing the microlens and anyplanets it may have (Udalski et al. 2015; Yee et al. 2015; Gould et al. 2013, 2014, 2015a,b,2016). However, in order for
Spitzer to take useful observations, it must be alerted to themicrolens target before the event ends, and usually before peak. Hence, in sharp contrast tothe situation envisaged a decade ago, both a “completed-event” finder and alert capabilityare important for KMTNet.Ideally, therefore, both forms of event-finder would have been basically ready when thefirst KMTNet data began arriving from the telescope in early 2015, or certainly by early2016 following the first year of commissioning data. Unfortunately, work on event-findersdid not begin until mid-2016. By the time the event finder was first tested, it was late 2016,meaning that two full years of data were already taken. This fact alone implied that muchhigher priority had to be given to constructing an event finder that worked on completedevents than to developing alert capability. Combined with the facts that the original core 4 –of KMTNet science was built around a pure-survey detection strategy, and that fine-tuningand operating a completed-event finder is much easier and less time-consuming than an alertsystem, this made the completed-event finder the obvious first choice for development.In this paper, we present the first microlensing events detected directly from the KMT-Net survey data. In Sections 2 and 3, we present the basic algorithm for identifying mi-crolensing event candidates and discuss its robustness. In Section 4, we detail the specificprocedure for applying this algorithm to the 2015 KMTNet data starting with the photo-metric pipeline and proceeding through the application of this algorithm and the vetting ofthe event candidates. We also assess the efficacy of the event-finding process by comparisonto OGLE-IV EWS alerts and by checking the detectability of binaries with our algorithm. Asummary of the detected events is given in Section 4.10. Information on accessing the finalsample of events is given in Section 5 along with our data policy. We discuss some potentialimprovements to the algorithm in Section 6 and summarize our findings in Section 7.
2. The Basic Algorithm
The main challenge of an event-finder is that it must be both quick and robust. “Quick”here has two senses. First, it must sift though 10 photometric observations each year,spread over 3 × light curves, in a reasonable amount of computer time. Second, it mustshow a restricted subsample to an operator who can vet these candidates in a reasonableamount of time. And it must find all candidates that might plausibly harbor a planetarysignal.By far, the fastest method of evaluating light curves is a linear fit, which applies asimple deterministic formula twice: once to determine the parameters and the second timeto evaluate χ . Unfortunately, microlensing events are described by five parameters, onlytwo of which are linear: F ( t ) = f s A [ u ( t ; t , u , t E )] + f b ; u ( t ) = s ( t − t ) t + u ; A ( u ) = u + 2 u √ u + 4 . (1)Here F is the observed flux, A is the magnification, t is the time of maximum magnifica-tion, u is the impact parameter (normalized to the Einstein radius θ E ), t E is the Einsteintimescale, f s is the source flux, and f b is any blended flux that does not participate in theevent.Hence, if one wishes to use a linear fit, one must work on a 3-dimensional (3D) grid of( t , u , t E ), which is prohibitive. We therefore begin with the insight of Gould (1996) that 5 –in the high-magnification limit, microlensing events can be described by just two non-linearparameters F ( t ) = f max p t − t ) /t + f b (2)where t eff → u t E and f max → f s /u . This formula is approximately valid only for u . . t eff around its peak time t . Gould (1996) was only concerned with high-magnification events because that was allhe thought could be detected in M31, the subject of his paper. Whether or not this istrue of M31, it is certainly not true of the Galactic bulge, where we expect to detect manylow magnification events as well. Therefore, we augment the Gould (1996) approach byconsidering also a representative low magnification event, namely Equation (1) with u = 1.That is, we consider F ( t ) = f A j [ Q ( t ; t , t eff )] + f ; Q ( t ; t , t eff ) ≡ (cid:16) t − t t eff (cid:17) ; ( j = 1 ,
2) (3)where A j =1 ( Q ) = Q − / ; A j =2 ( Q ) = Q + 2 p Q ( Q + 4) = [1 − ( Q/ − ] − / (4)Note that we no longer call the two flux parameters ( f s , f b ), but rather ( f , f ). This isbecause no physical meaning can be ascribed to these parameters. In the high-magnificationlimit, f → f max and at u = 1, ( f , f ) → ( f s , f b ), but in the general case, these fluxparameters are not identifiable with any specific physical quantities.The set of t eff ,k are a geometric series t eff ,k +1 = (1 + δ t eff ) t eff ,k (5)and for each t eff ,k , we choose an arithmetic series for t ,k,l t ,k,l +1 = t ,k,l + δ t t eff ,k . (6)We adopt δ t = δ t eff = 1 / , (7)which we show below to be quite conservative. We set t eff , = 1 day (so t eff , ≃ .
56 day), witha maximum t eff ≃
99 days. It is not possible to reliably detect events that are substantially 6 –longer than this upper limit in a single season. The lower limit is also conditioned by thecharacteristics of the first season’s data. We discuss extending these limits in Section 6. Weconsider values of t from δ t before the first epoch of the 2015 season until δ t after the lastepoch.We restrict the fits to data within t ± Zt eff where Z = 5. We require that this intervalcontain at least N min = 50 points. For 2015 data, we consider only data taken at CTIO inthe initial, automated phase of the search. We discuss expanding the search to include otherobservatories in Section 6.There are therefore about 2∆ T / ( δ t t eff ,k ) trial fits for each value of t eff ,k and these trialswill contain a total of 4 ZN/δ t = 240 ,
000 data points (with repeats). Here ∆ T ∼
250 days isthe duration of the 2015 season and N ∼ t eff trial takes about the same amount of time. The total number of trials is roughly2(1 + δ − t eff )∆ T / ( δ t t eff , min ) ∼ ,
3. Analytic Characterization of Robustness
Since we are sampling a continuous function of three variables ( t , t eff , u ) on a discretegrid, the ∆ χ of the fit (relative to a flat light-curve model) will inevitably be smaller thanfor the optimal values of ( t , t eff , u ). This is not in itself of any concern because the onlypurpose of these fits is to find which of the light curves have sufficient deviation to warrantshowing them to the operator. If, for example, the discrete-model ∆ χ could be as muchas 10% smaller than the true (continuous model) ∆ χ , then one simply has to set the ∆ χ threshold 10% lower than the level that one wishes to investigate. Here we show how thediscretization of the parameter space impacts the fits.We would expect the largest ∆ χ deviation from the most coarsely sampled parameter u . As outlined in Section 2, u is represented by just two discrete values: “high-mag”( u ≪
1, in effect, “ u = 0”) and u = 1. To test the robustness of this approach, weconsider light curves with u = 4 , , , / , / , / t , t eff ) to get the best fit. Figure 1 shows that for u = 1 /
4, thehigh-magnification formula A ( Q ) = Q − / already works extremely well, while at u = 1 / u = 2 and u = 4 inputs are quite well fit by the u = 1 (i.e., A ) model.Note that u = 4 ( A max ∼ . u = 2. This differs from both model curves 7 –by a noticeable amount, but these differences are still quite small. Since these are purelytheoretical investigations, the conclusion that one of A or A will fit the data acceptablyregardless of the true u is independent of the magnitude of the event.Next, we ask how ∆ χ is affected if the true value of t differs from the model t by ǫt eff .Note that, by construction, ǫ ≤ δ t /
2. We can quantify the offset between the grid-based andtrue-model light curves by δχ /χ , where δχ is the difference between the best ∆ χ and theone that is obtained from a fit forced to u = 0 (or u = 1), while χ ≡ P i [ f s ( A i − /σ i ] .Note that (because the algorithm compares the observed lightcurve to its mean – rather thana true baseline) ∆ χ ∼ χ /
3. Assuming uniform cadence and below-sky errors, one findsthat for high-magnification light curves A ( Q ) = Q − / , δχ χ ≃ − R ∞−∞ dx [(1 + ( x − ǫ/ )(1 + ( x + ǫ/ )] − / R ∞−∞ dx (1 + x ) − → ǫ . (8)Since, as mentioned above, ǫ ≤ δ t / /
6, this ratio is negligibly small.The ∆ χ deficit due to discretely chosen t eff leads to a similar calculation with the result: δχ /χ = ǫ /
4. Implementation for the KMTNet 2015 Data
In addition to introducing the algorithm itself, this paper has two main goals. The first,addressed in this section, is to identify the strengths and weaknesses of the algorithm byapplying it to a real data set. The second is to make available the results of the event searchto the general microlensing community, which we will address further in Section 5. Both ofthese goals require that the algorithm be specifically applied to our data set through a seriesof concrete choices. One choice, for example, is whether to search for events in the data froma single observatory or whether to conduct the search by simultaneously fitting data fromall three (or perhaps just two) observatories. Another choice is whether to initially vet forvariable stars and artifacts by some sort of algorithm or whether to reject these totally inthe operator stage “by eye”.The actual choices made for the 2015 data were strongly conditioned by the operationalenvironment, including both the nature of the 2015 data and the operational pressures 8 –affecting the reduction and analysis of these data. It is therefore necessary to summarizethese operational constraints to understand both the empirical evaluation of the algorithmconducted below and the data products, as well as to develop ideas on how to improveapplication of the algorithm in the future. The last point is discussed in Section 6.
Much of the 2015 data is of very high scientific quality, and they have already been thebasis of many published and submitted scientific papers (Shvartzvald et al. 2015; Bozza et al.2016; Zhu et al. 2016; Shin et al. 2016; Han et al. 2016a,b; Zhu et al. 2017; Chung et al.2017) However, 2015 was the commissioning year and as such it was heavily impacted byengineering tests as well as physical problems, refinements, and repairs. For example, theelectronics of the KMTC camera were changed substantially on HJD ′ ≡ HJD − ′ ∼ ′ = 7180. There were many other issues as well,including problems with the polar alignment and pointing model. Problems of this order arecompletely normal for the commissioning of relatively complex systems, but they also implythat algorithmic ideas that are designed for basically homogeneous data must be significantlyadjusted if they are to function in this environment.The first step was to write a pipeline to extract photometry from the images. Thispipeline construction impacted the event-finder in several ways, both direct and indirect.This will be discussed in the next several subsections. However, from the present standpointthe main impact was to delay the development and testing of the event-finder. It requiredabout 8 months to develop and test the pipeline and another 3 months, using all availablecomputer resources, to extract light curves just for KMTC. Testing of the event-finder couldproceed in parallel with light curve extraction, but full implementation could only be carriedout after the light curves were extracted. Hence, event finding for 2015 did not begin inearnest until the end of the 2016 season.The combined effect of these considerations was to drive us toward event-finding proce-dures that could be implemented quickly and would be robust in the face of inhomogeneousdata. Work on more sophisticated versions was deferred, although this was subsequently 9 –initiated. See Section 6 for a discussion of the progress on a more advanced version. The most straightforward way to extract light curves is to measure photometry at thepositions of known stars. This requires a catalog of stars matched to the detector coordinatesystem.The original approach that had been taken was to carry out point-spread-function (PSF)photometry using the DoPhot package (Schechter et al. 1993), which extracts a catalog di-rectly from the images themselves. Then, those DoPhot light curves would be searched forevents, and finally difference image analysis (DIA, Alard & Lupton 1998) re-reductions usingthe Albrow et al. (2009) pySIS package would be used for events of particular interest. Thisplan was well-matched to the available computer resources but suffered from what turnedout to be a fatal flaw: the DoPhot-based catalogs contained a factor ∼ .A new pipeline was written based on publicly available DIA code from Wo´zniak (2000).The central problem remained, however, of constructing a deep input catalog. While variouspackages were tried, it proved impossible to even come close to matching the depth of theexisting OGLE-III (Szyma´nski et al. 2011) catalog. The reasons for this are not completelyclear. No doubt, this is partly explained by the fact that OGLE pixels are much smallerthan KMTNet pixels (0 . ′′ versus 0 . ′′ ), but it is difficult to believe that this is the fullexplanation. Probably the biggest factor is just greater photometric experience.The OGLE-III catalog provides stellar positions in RA and Dec, so it was astromet-rically matched to the template frame, i.e., the image coordinates, by cross-matching withthe DoPhot catalog. The pipeline uses the OGLE-III catalog stars and photometric scalewhenever possible. However, the OGLE-III catalog does not cover the full extent of the2015 KMTNet fields. Some KMTNet fields cover fairly broad continuous regions that arenot covered by the OGLE-III survey. In addition, the OGLE-III catalog has non-trivialtopology, due to gaps of various sizes and shapes between OGLE-III fields. Approximately10.3% of the area covered by the four prime KMTNet fields is not covered by OGLE-III This factor is less than the naive “five” because, for identical event geometries, brighter sources havesmaller magnitude errors, making it is easier to detect and measure subtle anomalies.
10 –(see Section 4.7). In these regions, the input catalog is derived from DoPhot templates andthe photometric scale is calibrated by cross-matching with the OGLE-III catalog in nearbyregions. In total, our OGLE-III + DoPhot catalog consists of 71,555,640 stars.
Once the light curves for all catalog stars are extracted, they must be searched formicrolensing events. Note that for 2015 data, only KMTC light curves were used to searchfor events (as indeed these were the only data reduced). The basic algorithm for searchingfor the best microlensing model was described in Section 2. Its basic features were to performlinear fits to the data on a grid of ( t , t eff , u ) and to fit only those data from the interval( t − t eff < t < t + 5 t eff ). In practice, those linear fits were performed in four phases.First, all the points were included in the fits to evaluate the best-fit parameters, ( f , f ).Then, the exact procedure was repeated to determine the naive χ of each data point, usingthe ( f , f ) from the previous step. Then the 10% worst- χ points were eliminated, and thedata were refitted to redetermine ( f , f ). Finally, these ( f , f ) were used to determine theoverall χ µ lens .To determine the significance of the microlensing signal, the final sample of points(i.e., with 10% points removed) was then fitted to a flat line, i.e., a constant flux level.Nominally, this would imply a single parameter, i.e., the mean flux during the intervalprobed ( t − t eff < t < t + 5 t eff ). In fact, however, to allow for breaks in the light curvedue to firmware adjustments at KMTC, each of the three intervals described in Section 4.1was permitted a separate mean flux level. Hence, there could be one, two, or three suchparameters depending on how many of these three intervals were included in ( t − t eff
500 then the candidate “event” is shown to the operator in a four-panel display,together with some auxiliary information (Figures 3–6 and Figure 11). Each panel containsall the data (not just the data surviving the 10% rejection). One reason to show all of thedata is that caustic crossings in binaries are likely to be removed by the 10% rejection cutas “outliers” from the point lens fit. The bottom two panels show the whole season of data,while the top two show only the 10 t eff that were fitted to the model. The left two panelsare restricted to the flux range predicted by the model (plus small border regions at top andbottom), while the right two panels show the full range of data values within the temporalrange of the diagram.These four displays allow the operator to rapidly determine whether this is plausiblya microlensing event. For example, the upper left panel gives a very clear impression ofwhether the event is reasonably well fit by point lens microlensing over the 10 t eff range thatis being fit (see Figure 3). If, on the other hand, the event has pronounced binary structure(e.g., a pronounced caustic near the peak of the event), then this structure will mostly not bevisible on this plot, and the light curve may appear to deviate from point-lens microlensingin an unfamiliar way. However, this caustic structure will appear clearly on the upper rightpanel, even if the main part of the microlensing event (easily visible on the upper left) nowis so small that it is barely visible (see Figure 4).This approach also permits rapid rejection of variables. The algorithm itself quite easilyfits wide classes of variables to microlensing light curves because it “censors” data fromoutside the 10 t eff fitting interval. However, this is readily apparent from the two full-seasonlight curve panels (see Figure 5). 12 –Above the upper-right panel are displayed four numbers (∆ χ , I cat , RMS(mag) , RMS(flux)).Here I cat is the magnitude of the catalog star, RMS(mag) is the amplitude of scatter of thisstar derived from the OGLE-III catalog, and RMS(flux) ≡ RMS(mag) × . − I cat ) . Thelast number then gives the expected variability in flux units. For OGLE-III stars, I cat isderived from the OGLE-III catalog. For non-OGLE-III stars, it is derived by calibratingthe DoPhot photometry based on the overlap with OGLE-III areas. In this case, there is noinformation on variability, so the final two numbers are set to arbitrary negative numbers,meaning “ignore”.The operator then classifies the candidate into one of four categories: “clear microlens-ing”, “possible microlensing”, “variable”, and “artifact”. All “clear” or “possible” microlensing events are then further evaluated using KMTSand KMTA data. As mentioned above, for 2015, only the KMTC data were reduced enmasse. For candidate events, KMTS and KMTA light curves were extracted from individual(256 × χ > χ > A less obvious path to “false positive confirmation” comes from artifacts. Based on thecommissioning issues described in Section 4.1, one can well imagine that there are a largenumber of artifacts in KMTC data that trigger the event finder. In the overwhelming ma-jority of cases, these are easily recognized as such (or perhaps “misclassified” as variables)and so do not make it to the stage of vetting by KMTS and KMTA comparisons. In manyof the remaining cases, these artifacts are easily removed because they are not duplicated indata from one or both of the other observatories. However, a large fraction of candidates ini-tially classified as “possible microlensing” and a handful of those initially classified as “clearmicrolensing”, turn out to be examples of one of two types artifacts: “bleeding columns”and “displaced variables”. We developed simple procedures for identifying these.For each of the events that was classified as either “clear” or “possible” microlensing,we display side-by-side the finding chart and a difference image from near the peak of the“event”. Below these are the set of light-curve displays described in Section 4.6, and belowthese panels are a set of four difference images: the two nearest the peak, and two othergood seeing images relatively near peak.We place cross-hairs at the catalog position of the source in all images. From the top-level difference image, one can easily see whether or not the cross hairs are at the positionof the variation. If variation is not at the catalog (cross hair) position, this is not in itself aproblem: many microlensing events occur on stars that are too faint to be cataloged, and are 14 –recognized by the flux variations that they induce at the position of a neighboring catalogstar due to the finite width of the PSF. However, if the finding chart contains a bright star atthe position corresponding to the flux variation on the difference image, then the “event” isalmost certainly an artifact, i.e., an echo of variations in this bright star. Of course, a givenbright star might actually undergo a microlensing event, and this event would likewise be“echoed” at the positions of neighboring cataloged stars. But in this case we would expectthat the bright star would itself be cataloged and would yield a microlensing light curvewith higher ∆ χ . For variable sources, however, the “cleaner” light curve from the variableis much more easily excluded as a “variable” than is the “echo”, so only the echo winds upin the list of candidates.Moreover, it is almost always the case that these echos last for most or all of theseason, simply because repeating variations would have been recognized as variables, evenin their echos. In the course of this vetting, we noted a handful (2–3) cases for which thesource position was offset and coincident with a bright star, but the event was neverthelessshort and well contained within the season. We accepted these as microlensing events andassumed that either the catalog or the event finder had failed to function properly in thesecases. This means that there were probably also a few long microlensing events with similarcharacteristics that we misclassified as variables, but this problem cannot be addressed withonly a single year of data.Another quite common artifact that easily shows up in the difference images are fakeevents due to bleeding columns. In fact, the level of bleeding is far too low to be noticeddirectly in the images, and can only be perceived in difference images, which permit a muchstronger “stretch”. Time variable bleeds can be generated by variable stars, in which casethey are typically either long-timescale or repeating. However, they can also be generatedby shorter-period irregular variables that only rise above the bleed threshold for a relativelyshort period, once in a season. In addition, they can be caused by changes in seeing andbackground. The former is not likely to be correlated among observatories, but the latter is,i.e., if it is due to lunar phase. In any case, as mentioned above, such bleeding columns areeasily recognized by examining the four images that we routinely display.The event finder yielded 955,659 candidates, which were automatically grouped into385,565 groups to be shown to the operator. Of these, 148,010 were classified as “variables”,and 236,698 as “artifacts”. Our final catalog of events consists of 673 “clear microlensing”and 184 “possible microlensing” events (with some duplicates). See Figure 7. Section 4.10summarizes the properties of our final sample.It is interesting to compare this map of detections with a map of the underlying catalog(Figure 8). Note in particular that BLG02N has a density only about 1.35 times higher than 15 –BLG04N but has an order of magnitude more “clear microlensing” events. This is broadlyconsistent with our expectations based on Poleski (2016), which shows that the microlensingevent rate should be highest near the center of BLG02N. The best external check on the event-detection procedures described above is to comparewith OGLE-IV. Since OGLE operates from a very nearby site in Chile, it has very similarfield visibility and weather to KMTC, i.e., the KMTNet site used for the initial selections.In fact, the visibility is not identical because OGLE observes to greater hour angle thanKMTC and also longer into austral spring. On the other hand, KMTNet observes duringall lunar phases, while OGLE typically halts observations when the Moon is in the GalacticBulge. Similarly, KMTNet observes in all atmospheric conditions, provided that these donot endanger the telescope and provided that the sky is not opaque. Further, for abouta quarter of the KMTNet prime area, OGLE-IV observes with cadence Γ = 3 hr − , whichis not qualitatively different from the KMTNet cadence. And, for most of the remainingarea, OGLE-IV observes with cadence Γ = 1 hr − , which still should be sufficient to detecta substantial majority of microlensing events that are accessible to KMTNet.For this comparison, we use the set of OGLE-IV events announced by their Early Warn-ing System (EWS Udalski 2003). These events are identified in real-time based on only apartial light curve rather than a full microlensing fit to a completed event. As such, thiscomparison sample may not be complete and it may also contain some false positives. How-ever, it represents the most complete, publicly available, independent list of microlensingdiscoveries for the 2015 season.
428 OGLE-IV events (after eliminating duplicates) in the KMTNet field were not recov-ered by our event finding procedure, neither as “clear” nor “possible” microlensing events.These represent potential false negatives and may indicate problems with our procedure.We first assess the fraction of the OGLE-IV alerts discovered as a function of brightness.We define I peak = 18 − . ∗ log[ f S ( A max − f S and A max are determined fromthe OGLE EWS table. For the following ranges of I peak , we find the recovery percentagesare I peak ( <
14 mag = 63%, [14–15] = 91%, [15–16] = 76%, [16–17] = 72%, [17–18] = 72%,[18–19] = 61%, [19–20] = 43%, and >
20 = 16%), i.e., the events missed by our procedure 16 –tend to be biased toward the faint end.To explore this in more detail, we examined one out of each 20 OGLE-IV events inthe KMTNet fields that were not found by the procedures outlined above. This test wasconducted after selection based on KMTC data, and before vetting based on KMTS andKMTA data, in order to provide the closest basis of comparison. One result of this test wasthat we found that three of the 22 tested events lay in image patches that had not beenprocessed by the event-finder due to a failure in the computer architecture (and so havingnothing to do with the event finder itself). These computer problems were fixed and thesepatches were re-run. We report only on the remaining 19 tested events.Of these 19 “missing events”, three were clear failures, three were judged to be “marginallyacceptable” false negatives, and remaining 13 were “acceptable” false negatives.Two of the clear failures (OGLE-2015-BLG-0693 and OGLE-2015-BLG-1589) were dueto operator error. The first was mislabeled as “CV”, and the second should have been called“possible microlensing” (and then further investigated using KMTS and KMTA data). Thethird failure (OGLE-2015-BLG-1015) lay in a thin band between two patches. Each 1 deg chip of the 2015 data was analyzed in a (40 ×
40) grid of “overlapping” (256 × χ = 436. If it had been shown, it probably would have been classified as “possiblemicrolensing”. Rectification of this “problem” would require setting the threshold lower andso greatly increasing the already large number of events to be vetted.To better understand the issues involved in possibly setting a lower threshold, we showin Figure 10 cumulative distributions by ∆ χ of i) all event groups, ii) “clear microlensing”events, and iii) “possible microlensing” events. Of particular note: while 64% of all (385,565)event groups had 500 < ∆ χ < . ′′ and 2 . ′′ , which led to very weak and no signal respectively.Neither passed the ∆ χ = 500 threshold, but if they had, neither would have been chosenby the operator. Both events lie in regions not covered by the OGLE-III catalog, and forwhich we are dependent on the much shallower DoPhot catalogs.One event peaked midway between 2014 and 2015 seasons, and so was not recognizablebased on 2015 data. Two were long events that were (correctly) judged by the operator tobe consistent with being variables based on 2015 data. Two events were affected by artifactsin the 2015 data. The remaining eight had low S/N and so were not shown to the operator,but if they had been they would have been judged as too noisy for reasonable identificationas microlensing events. (One of these, OGLE-2015-BLG-1835, may be a CV.)In brief, this test shows that of order 3 ×
20 = 60 genuine microlensing events thatcould have been detected in this region were not. These failures are mostly due to “operatorfatigue” from reviewing hundreds of thousands of light curves. In future years, this maybe ameliorated by eliminating most variables in advance and by reduction in the number ofartifacts now that commissioning is complete.
For “clear microlensing” events found by KMTNet but not found by the OGLE-IVEWS, we would like to try to understand why they were not detected by OGLE-IV. Forthese events there are three possibilities: OGLE could not detect (or would have extremedifficulty detecting) the event due to insufficient data, OGLE excluded the event as a falsepositive, or the OGLE EWS missed the event. Since OGLE has a different camera, a differentsite, and different observing protocols, the data on a particular event might be insufficientdue to the location of chip gaps or observability gaps. The OGLE survey has also beenin operation for many years and so can identify (and exclude) false positives due to long-timescale, sporadic variables. We examine a representative sample of the KMTNet eventsnot found by the OGLE EWS in greater detail to assess whether or not we expect OGLEto have detected them. If it appears that OGLE ought to have detected an event, then thatevent is a candidate for being a false positive, i.e. it is possible OGLE did not alert the eventbecause they had reason to believe it was a variable star rather than microlensing. At thesame time, the absence of an alert is not conclusive evidence of a false positive; the eventcould simply have been missed. On the other hand, if the synoptic observing informationsuggests that OGLE would have difficulty detecting an event, than the absence of an alert 18 –does not provide any information about whether or not an event is a false positive or realmicrolensing.An unpublished study by Gould & Udalski has already determined that events detectedby MOA but not by OGLE are overwhelmingly undetectable by OGLE because they arein chip gaps, outside the OGLE fields, during gaps in the data, etc. A few are variablestars that have been masked from the OGLE catalog (in this case, two of 40, P. Mr´oz, 2017,private communication). Thus, if an event is detected by MOA but not by OGLE, we donot investigate it further because there is likely a plausible explanation as to why it was notalerted. We then reviewed a random 10% (i.e., 18) of the 177 remaining events to assesswhether or not there was sufficient data for OGLE to alert them.We find that four of these randomly chosen 18 could not plausibly have been detectedby OGLE since they were either in chip gaps or would have had very few magnified points(BLG01N.3928.2251, BLG02M.0326.3862, BLG02T.0123.2655, BLG03M.0939.0577). Fouradditional events were very likely missed either because they lay very near a chip edge (thoughformally within the chip) or because they would have had only 3 or 4 significantly magnifiedpoints. (BLG01T.2412.0089, BLG02N.2218.3192, BLG03K.3511.3254, BLG04N.3638 2631).Thus, it is not possible to draw conclusions based the absence of an OGLE alert for these4 + 4 = 8 events.There were four other events that appeared to us to be probably detectable, but becausethey contained only 6–8 magnified points, could have plausibly been missed, perhaps dueto mildly adverse conditions (BLG02T.2218.3029, BLG03N.2123.2010, BLG04K.2626.3153,BLG04T.2822.2817). Here we should keep in mind that by selecting non-OGLE events, weare biased toward such adverse circumstances.Finally, there were six events that we thought should clearly have been detected byOGLE (BLG01K.0126.0793, BLG01M.1125.1711, BLG01N.0633.3015, BLG02N.0726.2197,BLG02N.3209.0444, BLG03T.1911.1388). Five of these six lie in OGLE fields BLG505,BLG506, or BLG512, which had OGLE cadences of Γ = 3 hr − in 2015. If there are falsepositives in our sample, these should be good candidates. In reviewing these events, we findthat all look plausibly like microlensing. In all cases the t eff <
10 days and the remainderof the light curve is consistent with being flat. In all cases, all three data sets basicallyagree on the form of the light curve (except for the first, for which there are no KMTAdata). In one case (BLG01M.1125.1711) the variation is offset from the catalog star andat the position of another resolved star, which we would generally regard as a warning signthat the resolved star is a variable and the “event” is an echo. However, as discussed inSection 4.7, we override this concern when the event is of short duration and the remainderof the season is flat. This is particularly true in the present case, for which the star that is 19 –varying is relatively faint. Hence, while any of these candidates could in principle be falsepositives, there is no evidence that this is the case.Subsequent to the posting of this paper on arXiv, P. Mr´oz (2017, private communication)kindly provided a list of all events in OGLE high cadence fields that were recognized asvariables by OGLE. This list contained one of the four we thought “likely missed” by OGLE,two of the four events that we considered “probably” should have been detected, but noneof the six events that we considered should “definitely” have been detected.
We expect that our approach will be efficient at detecting planetary events simplybecause these typically look similar to point-lens events but with relatively brief anomalies.One might be concerned that these anomalies could reduce ∆ χ and so prevent their beingshown to the operator. However, recall that the 10% “worst outliers” to the point-lens-likefit are eliminated from the comparison with a flat light curve. Hence, we actually do notexpect this to be an issue.The situation is, however, quite different for binary events, many of which look nothinglike point-lens events. When we devised the event-finder, we had no definite expectation ofhow well it would perform on binaries.We conducted a test to evaluate this performance after the fact. We manually examinedall OGLE events lying in the KMTNet fields to find those that either were, or plausiblycould be binaries. We found 57 binary events announced by the OGLE EWS that lie in theKMTNet fields. Thirty-four are included in our list of microlensing events (all of them areclassified as “clear”. Of the 23 that were missed, only one of these both plausibly looks likemicrolensing in KMTC data and failed to be shown to the operator (i.e., had ∆ χ < χ = 265. The reason for this low ∆ χ is that the “effective time scale” of this riseis much shorter than the 0.56 day minimum of the current search. This limit was in turn setby the large number of artifact-driven “events” that would be found in the 2015 data, whichwould have greatly multiplied the work of the operator.Of the other 22 that were missed, three were due to operator error: one (OGLE-2015-BLG-0095) should have been classified as “clear” microlensing and two (OGLE-2015-BLG-1346 and OGLE-2015-BLG-2017) as “possible microlensing”. For the remaining 19, thesignal in the data (whether shown to the operator or not) was not sufficient to count themas plausible microlensing events. From the perspective of understanding the algorithm, the 20 –key point is that only one known binary was missed because the algorithm itself failed todetect it.We consider it somewhat surprising that the algorithm is working as well as it is onbinaries, given that it is specifically tailored to find point-lens events. One reason for this isthat the algorithm will detect pretty much any sort of bump provided that it is sufficientlypronounced and not too sharp, as illustrated by Figure 11. We suspect that the good bi-nary sensitivity then derives from the sampling density of the 3-parameter point-lens griddescribed in Section 3. While this sampling density is overly conservative for detecting pointlens events, it likely permits matching with sharp caustic features, allowing the detection ofmicrolensing binaries. This conjecture can be tested in the context of the improved eventfinder (Section 6) by repeating the fitting procedure on known binaries with a progressivelyless dense grid. However, because the event finding procedure is significantly less compu-tationally intensive than the photometry pipeline, there is no strong driver to scale backthe sampling density. If the computational situation changes, this test should be performedbefore scaling back the sampling density. In total, this procedure for identifying microlensing events from the 2015 KMTNetcommissioning data resulted in 857 microlensing event candidates; 15 of which are duplicatesleaving 842 event candidates. Of these, we classify 660 as “clear” microlensing events and182 as “possible” microlensing. Based on Mr´oz et al. (2015) and P. Mr´oz (2017, privatecommunication), we find that 40 of the “clear” microlensing events and 54 of the “possible”microlensing events were classified as variables by the OGLE collaboration. Of the remainingevents, 483 “clear” events and 38 “possible” events were detected by either the OGLE EWSor the MOA alert system (or both). Thus, KMTNet has detected 137 “clear” microlensingevents and 90 “possible” events not found by any other survey. Based on our investigationdescribed in Section 4.8.2 and P. Mr´oz (2017, private communication), there is no evidenceindicating these are false positives.The distributions of t eff and I base for these events are shown in Figure 9. The readershould bear in mind that t eff is discretely defined as described in Section 2 and therefore isnot a perfect representation of the event timescale. 21 –
5. Data Policy and Data Releases
KMTNet data policy is framed by several goals, which are basically compatible butare subject to some mutual tension. First, we seek to give adequate time for the KMTNetteam to publish results based on the huge amount of work required to obtain and processthese data. Second, we want the KMTNet data to be exploited to the maximum extentpossible, whether by KMTNet team members or others. Third, we want to promote themicrolensing field by making data available to as many workers as possible. Fourth, we wantto avoid overworking our team on tasks that are auxiliary to the team’s own effort to publishour results. Finally, we want to avoiding infringing on the data rights of others, whetherexplicitly or implicitly.
KMTNet data are available at http://kmtnet.kasi.re.kr/ ∼ ulens/event/2015/ . The mainpage is ordered by KMT event number in the form KMT-2015-BLG-[NNNN]. These numbersare in turn sorted by field BLG[NN] (01, 02, 03, 04), chip (K,M,N,T), patch, and star number.The “clear microlensing” events are listed first, so that all entries ≤ ≥ t , t eff , u ) (Equations (3) and (4)), the baseline flux, and the RA and Dec. There are alsocross references to OGLE and MOA discoveries. By clicking on the event name, one can seethe finding chart as well as the same 16 panels of light curve displays shown to the operator.These pages also contain links to the light curve data. Guided by the above principles, we have formulated the following policies for KMTNetdata for 2015. We then outline issues that are under consideration for future years in thenext section.1) All 2015 KMTNet data remain proprietary until the acceptance for publication ofthis paper and all papers based on 2015 KMTNet data that are explicitly mentioned inSection 4.1.2) Beginning immediately (i.e., before the end of the proprietary period), all light curvesthat are posted together with this paper (see below) can be used by anyone to prepare futurepapers for publication. However, during the proprietary period (see (1)), they cannot be 22 –submitted for publication, nor posted on arXiv.3) Once the priority period ends (see (1)), papers can be submitted based on the releaseddata. We welcome collaboration with the KMTNet team, but do not require it as a conditionfor publication. For the cases that we collaborate, the KMTNet team does not demand any“editorial control”, but rather simply reserves the right to withdraw from papers with whichwe disagree strongly enough to warrant such withdrawal.4) If additional data processing is needed, we will either provide re-reduced light curvesin exchange for co-authorship or we will provide flat-fielded “stamps” surrounding the event,provided that sufficient evidence is given to us that there is substantive progress towardpublication of a paper.5) No data will be given out for events discovered by other teams but not independentlydiscovered by KMTNet using the algorithm presented here, unless there is explicit agreementwith those teams.The data that will be made available are1) All the data used by the event finder for all events that were determined to be either“clear microlensing” or “possible microlensing”. Recall that these data are based on Wo´zniak(2000) DIA using input positions derived from a pre-constructed catalog. Hence, they arenot usually of the highest quality, although sometimes they are in fact close to optimal.2) Automated Albrow et al. (2009) pySIS reductions of all light curves . In roughly halfof all cases, these reductions are better or substantially better than the reductions from (1).Unfortunately, in the other half of cases, these automated reductions either basically fail ortotally fail. This means that, for a substantial fraction of events, obtaining very good orexcellent light curves requires additional, by hand, reductions.No effort will be made to rectify these or any other residual problems for the 2015 datarelease as a whole. As discussed in Section 4.1, the 2015 data have intrinsic problems thatare not likely to propagate into future years, and our main data reduction efforts will beapplied to those future-year data. Note that as of the time of submission, these reductions are not yet complete, but will be uploaded tothe website as they are completed.
23 –
Future data policy and releases will be governed by the same principles, but will bemodified based both on the experience of the 2015 release and the improved quality of2016+ data. For 2015, we are posting light curves about 18 months after the close of the2015 season. We hope that this time lag will be substantially improved for 2016 and furtherimproved for 2017. Since we do not know exactly what problems we will encounter, wecannot guarantee it. However, at this point we believe that a 6-month delay is a plausibletarget for 2017+ data because if a paper cannot be drafted within 6 months of time the dataare taken, it is likely to fall down the list of priorities as the next season of data start tobecome available.For 2016, we hope to have an earlier release of the events in the
Kepler
K2 C9 field.(See Henderson et al. 2016 for a description of the K2 C9 project.) For KMTNet-discoveredevents, these data will have no proprietary period. We will also have a public release ofKMTNet data on events discovered by other teams but not discovered by KMTNet. However,for these data, we will require that prospective authors obtain permission from the otherteams before using the KMTNet data for publications.
6. Future Improvements
Following the work reported here creating and implementing the 2015 event-finder algo-rithm, and based on this experience, numerous improvements are already being implementedfor the 2016, 2017 and 2018 versions of this algorithm. They will be described more fully inthe 2016 (and subsequent data release paper(s).The first improvement is simultaneous fitting to data from all three observatories. This isstraightforward in principle because all three observatories have identical input star catalogs,so one can easily cross-identify light curves of the same physical “star” (really “catalog star”,which very frequently is actually a multi-star asterism).The second is simultaneous fitting of overlapping fields. This was completely unnec-essary in 2015 because there were no overlapping fields. However, in 2016, each of thethree “prime fields” BLG01, BLG02, BLG03, was also observed, slightly shifted, as BLG41,BLG42, BLG43, in order to cover the gaps between chips. Further, fields BLG02 andBLG03 are now somewhat overlapping in order to obtain a very high cadence on a smallarea Γ = 8 hr − . Hence, there can be as many as four overlapping fields (see Figure 12). Andthis also means (taken together with the previous paragraph) that there can be as many as12 light curves that might be combined. In order to reliably cross-identify stars, however, 24 –this improvement will only be implemented for OGLE-III catalog stars.In 2015, we did not consider events with t eff >
99 days because we had no out-of-yearbaseline data. This will remain so in 2016 because the fields have been changed between2015 and 2016, but also because the artifacts in the 2015 data make them unsuitable as abaseline of comparison. Such longer events can only be investigated beginning 2017.On the other hand, it should be possible to extend the search to much shorter timescalesthan the 2015 limit, t eff ≥ .
56 days. In 2015, such a search was not possible because of thelarge number of short timescale artifacts. It is likely that in 2016, the search will be conductedindependently in the t eff ≥ t eff < χ > χ was significantly higher. For2016+, we are already using all available data for 1) OGLE-III catalog stars in all fields and2) all stars outside of the six fields mentioned in the previous paragraph. For these, we willtherefore demand (∆ χ > χ >
7. Summary
We have presented a new algorithm for finding “completed” microlensing events and de-scribed its specific application to the four, primary 2015 commissioning fields of the KMTNetmicrolensing survey. We find 660 “clear microlensing” events and 182 “possible microlensing”events discovered and assessed by our method. The light curves of these events will be madepublicly available on our website (see Section 5.1) according to the data policy described inSection 5.2. In Section 5, we have presented our general approach to KMTNet data releases,in addition to the specific implementation of this approach for 2015 data. Feedback on thedata products, data policy, etc., based on the use of these data will be helpful in preparingfuture releases.Our procedure for vetting microlensing candidates still relies heavily on human inspec-tion of the microlensing candidates. Thus, we also discuss potential modifications for futureyears when we expect the data to be substantially cleaner. In fact, work is currently un-derway to improve this algorithm for the 2016 data. Finally, we note that although theprocedure for eliminating false positives will remain dependent on the specific characteristicsof the data, the algorithm itself as presented in Section 2 is completely general and couldbe applied to diverse microlensing data sets. Indeed, Shvartzvald et al. (2017) have alreadyapplied this algorithm to find the first infrared-only microlensing events from their UKIRTsurvey of heavily extincted bulge regions.We thank the OGLE collaboration for making available full information about their 2015event detections via their website http://ogle.astrouw.edu.pl/, which enabled systematicchecks of our event-finder. Work by YKJ, WZ and AG was supported by AST-1516842 fromthe US NSF. Work by IGS and AG was supported by JPL grant 1500811. This researchhas made use of the KMTNet system operated by the Korea Astronomy and Space ScienceInstitute (KASI) and the data were obtained at three host sites of CTIO in Chile, SAAO inSouth Africa, and SSO in Australia.
REFERENCES
Afonso, C., Albert, J. N., Alard, C., et al. 2003, A&A, 404, 145Alcock, C., Akerlof, C.W., Alssman, R.A., et al. 1993, Nature, 365, 621Alcock, C., Allsman, R. A., Alves, D., et al. 1997, ApJ, 479, 119Alard, C. & Lupton, R.H.,1998, ApJ, 503, 325 26 –Albrow, M. D., Horne, K., Bramich, D. M., et al. 2009, MNRAS, 397, 2099Bond, I.A., Udalski, A., Jaroszy´nski, M. et al. 2004, ApJ, 606, L155Bozza, V., Shvartzvald, Y., Udalski, A., et al. 2016, ApJ, 820, 79Chung, S.-J., Zhu, W., Udalski, A., et al. 2017, ApJ, in pressGould, A. 1996, ApJ, 470, 201Gould, A. & Loeb, A. 1992, ApJ, 396, 104Gould, A., Dorsher, S., Gaudi, B. S., Udalski, A., 2006, Acta Astron., 56, 1Gould, A., Carey, S., & Yee, J. 2013, 2013spitz.prop.10036Gould, A., Carey, S., & Yee, J. 2014, 2014spitz.prop.11006Gould, A., Yee, J., & Carey, S., 2015a, 2015spitz.prop.12013Gould, A., Yee, J., & Carey, S., 2015b, 2015spitz.prop.12015Gould, A., Yee, J., & Carey, S., 2016, 2016spitz.prop.13005Hamadache, C., Le Guillou, L., Tisserand, et al. 2006, A&A, 454, 185Han, C., Udalski, A., Lee, C.-U., et al. 2016a, ApJ, 827, 11Han, C., Udalski, A., Gould, A., et al. 2016b, AJ, 152, 95Henderson, C.B., Poleski, R., Penny, M. et al. 2016 PASP128, 124401Kim, S.-L., Lee, C.-U., Park, B.-G., et al. 2016, JKAS, 49, 37Mr´oz, P., Udalski, A., Poleski, R., et al. 2015, Acta Astron., 65, 313Poleski, R., Udalski, A., Dong, S. et al. 2014, ApJ, 782, 47Poleski, R., 2016, MNRAS, 455, 3656Schechter, P.L., Mateo, M., & Saha, A. 1993, PASP, 105, 1342Shin, I.-G., Ryu, Y.H, Udalski, A. et al. 2016, JKAS, 49, 73Shvartzvald, Y., Maoz, D. 2012, MNRAS, 419, 3631Shvartzvald, Y., Maoz, D., Kaspi, S., et al. 2014, MNRAS, 439, 604 27 –Shvartzvald, Y., Udalski, A., Gould, A. et al. 2015, ApJ, 814, 111Shvartzvald, Y., Bryden, G., Gould, A. et al. 2017, AJ, 153, 61Szyma´nski, M.K., Udalski, A., Soszy´nski, I., et al. 2011, Acta Astron., 61, 83Udalski, A. 2003, Acta Astron., 53, 291Udalski, A., Szymanski, M., Kaluzny, J., Kubiak, M., Mateo, M., Krzeminski, W., &Paczy´nski, B. 1994, Acta Astron., 44, 317Udalski, A., Yee, J.C., Gould, A., et al. 2015, ApJ, 799, 237Wo´zniak, P. R. 2000, Acta Astron., 50, 421Wyrzykowski, L., Koz lowski, S., Skowron, J., et al. 2009, MNRAS, 397, 1228Wyrzykowski, L., Koz lowski, S., Skowron, J., et al. 2010, MNRAS, 407, 189Wyrzykowski, L., Koz lowski, S., Skowron, J., et al. 2011, MNRAS, 413, 493Wyrzykowski, L., Skowron, J., Koz lowski, et al. 2011, MNRAS, 416, 2949Wyrzykowski, L., Rynkiewicz, A.E., Skowron, J., et al. 2015, ApJS, 216, 12Yee, J.C., Gould, A., Beichman, C., 2015, ApJ, 810, 155Zhu, Wei; Calchi Novati, S.; Gould, A. et al., 2016, ApJ, 825, 60Zhu, W., Udalski, A., Calchi Novati, S, et al. 2017, AJ, in press, arXiv:1701.05191
This preprint was prepared with the AAS L A TEX macros v5.2.
28 – (t−t )/t eff A −4 −2 0 2 40.2.4.6.81 u = 4u = 2u = 1u = 1/2u = 1/4u = 1/8 Fig. 1.— Best-fit “high-magnification” ( u ≪
1) models for various values of u =4 , , , / , / , / t , t eff ). Surprisingly, even very lowmag models fit reasonably well. Moreover u ≤ / u = 1 / −4 −2 0 2 411.11.21.31.4 (t−t )/t eff A u = 4u = 2u = 1u = 1/2u = 1/4u = 1/8 Fig. 2.— Best-fit “low-magnification” ( u = 1) models for various values of u =4 , , , / , / , / t , t eff ). Surprisingly, even very highmag models fit reasonably well. Moreover u ≥ u = 1 / d f l u x BLG01K0734.0458 2000 4000 6000 8000 10000 12000 14000 7068 7074 7080 7086 7092 709815376 17.63 -9.00 -202470 2000 4000 6000 8000 10000 12000 14000 16000 7050 7100 7150 7200 7250 7300 7350 d f l u x time (HJD-2450000) -80000 -60000 -40000 -20000 0 20000 7050 7100 7150 7200 7250 7300 7350time (HJD-2450000) Fig. 3.— Example of candidate light curve (ultimately judged to be “clear microlensing”)as initially shown to the operator. Top panels show only the 10 t eff ∼
30 days that werefitted to the model (dashed curve), while bottom panels show full season. Left panels showonly data in the range suggested by the model curve (with small buffers at top and bottom)while right panels show the full range of data. Four numbers above the upper right panelare ∆ χ , I base , variability in magnitudes, and variability in flux. Since the last two numbersare derived from the OGLE-III-catalog “rms” column, while this event comes from an areanot covered by OGLE-III, these entries are negative, meaning: “ignore”. 31 – d f l u x BLG02M1119.3178 -10000 0 10000 20000 30000 40000 50000 60000 70000 7218 7220 7222 7224 7226 7228 72303994 19.82 0.19 568 2000 4000 6000 8000 10000 12000 14000 16000 7050 7100 7150 7200 7250 7300 7350 d f l u x time (HJD-2450000) -40000 -20000 0 20000 40000 60000 7050 7100 7150 7200 7250 7300 7350time (HJD-2450000) Fig. 4.— Example of candidate light curve (ultimately judged to be binary “clear microlens-ing”) as initially shown to the operator. Panel displays are the same as in Figure 3. In thiscase, however, the upper right panel is crucial to realizing that the apparent “noise” near thetop of the model in the upper-left panel, is actually due to a strong binary caustic entrance.Also note that since the source star is in the OGLE-III catalog, the OGLE-III “rms” (0.19mag) is reported, and this is translated into an estimate of 568 flux units in the figures. 32 – d f l u x BLG01T0401.0003 0 10000 20000 30000 40000 50000 60000 70000 7260 7271 7282 7293 7304 73151742 13.80 0.02 15316 10000 15000 20000 25000 30000 35000 40000 45000 7050 7100 7150 7200 7250 7300 7350 d f l u x time (HJD-2450000) -3500000-3000000-2500000-2000000-1500000-1000000-500000 0 7050 7100 7150 7200 7250 7300 7350time (HJD-2450000) Fig. 5.— Example of candidate light curve (ultimately judged to be to be a “variable”)as initially shown to the operator. The upper left panel looks plausibly like microlensing.However, it is immediately obvious from the lower-left panel that there are several othervariations during the season of similar duration, albeit of somewhat lower amplitude. Theassessment of “variable” is further confirmed by the OGLE-III based “rms” of 15316 fluxunits. This is only slightly smaller than the rms one would measure from the “microlensingevent” that is modeled in the upper-left panel. 33 – d f l u x BLG01K0734.0458 -5000 0 5000 10000 15000 7068 7074 7080 7086 7092 709815376 17.63 -9.00 -202470 2000 4000 6000 8000 10000 12000 14000 16000 7050 7100 7150 7200 7250 7300 7350 d f l u x time (HJD-2450000) -100000 -80000 -60000 -40000 -20000 0 20000 7050 7100 7150 7200 7250 7300 7350time (HJD-2450000) Fig. 6.— Example of three-observatory light curve that the operator reviewed after the sameevent (shown in Figure 3) was judged to be either “clear” or “possible” microlensing. TheSAAO data (blue) confirm the microlensing character previously indicated by the CTIOdata (green). The SSO data start too late in the season to serve as a check in this particularcase. The operator would also have been shown additional 4-panel displays for each obser-vatory separately at this stage. Based on this inspection, the event was judged to be “clearmicrolensing”. 34 –
273 272 271 270 269 268 267R.A. (deg)−32−31−30−29−28−27 D e c l . ( deg ) Fig. 7.— Four fields observed by KMTNet in 2015, BLG01, BLG02, BLG03, BLG04 eachwith four 1 deg chips, T, K, M, N (both counterclockwise from lower right). Red and blueindicate “clear” and “possible” microlensing. Star symbols are events found previously byOGLE and/or MOA, while circles were not previously alerted. 35 – R.A. (deg) D E C ( deg )
272 270 268−32−30−28
Fig. 8.— Star catalog density of four fields observed by KMTNet in 2015, BLG01, BLG02,BLG03, BLG04 each with four 1 deg chips, T, K, M, N (both counterclockwise from lowerright). One per 500 catalog stars is plotted. Sharp rectangular boundaries in density aredue to regions not covered by the OGLE-III catalog, where DoPhot finds only about 1/5as many stars. Note that the catalog density is only about 1.35 times higher in BLG02Ncompared to BLG04N, but that there are an order of magnitude more “clear microlensing”events. See Figure 7. 36 – eff )020406080100120 N All ClearPossible
13 14 15 16 17 18 19 20 21 22I base N AllClearPossible
Fig. 9.— The distribution of event detections as a function of log t eff (left) and I base (right).The black solid line shows the histogram for all events; “clear microlensing” events are shownas the magenta line, and “possilbe microlensing” events are shown as the blule line. 37 – log( ∆χ ) C u m u l a t i v e F r a c t i on µ lensingPossible µ lensing ∆χ =1000 Fig. 10.— Cumulative distribution functions in log(∆ χ ) for all 385,565 event groups (black),the 660 “clear microlensing” events (red) and the 182 “possible microlensing” events (green).Almost 2/3 of the event groups have 500 < ∆ χ < d f l u x BLG02N2215.1660 0 50000 100000 150000 200000 250000 7125 7144 7163 7182 72018783 18.93 0.20 1358 10000 20000 30000 40000 50000 60000 70000 80000 7050 7100 7150 7200 7250 7300 7350 d f l u x time (HJD-2450000) -50000 0 50000 100000 150000 200000 250000 7050 7100 7150 7200 7250 7300 7350time (HJD-2450000) Fig. 11.— Example of candidate light curve (ultimately judged to be binary “clear microlens-ing”). Although the event bears little resemblance to the form of point-lens microlensing(upon which the event-finder algorithm is based), it is easily selected by the algorithm(∆ χ = 8783) to be shown to the operator, who in turn easily recognized it as binarymicrolensing. In fact, very few recognizable binaries are rejected by the algorithm. See text. 39 – KMTNet BLG Fields D ec l. ( d e g ) R.A. (deg) -38-37-36-35-34-33-32-31-30-29-28-27-26-25-24-23-22-21-20 260262264266268270272274276
Fig. 12.— 27 fields observed by KMTNet in 2016, color-coded by cadence. Note thatBLG(41,42,43) are shifted by 6 ′ relative to BLG(01,02,03), which enables Γ = 2 hr − coverageof the chip gaps while still preserving Γ = 4 hr − cadence over most of this prime area. Notealso that a small area is covered by four fields BLG(02,03,42,43) and so has Γ = 8 hr −1