Robust Identification of Active Galactic Nuclei through HST Optical Variability in GOODS-S: Comparison with the X-ray and mid-IR Selected Samples
E. Pouliasis, I. Georgantopoulos, A. Z. Bonanos, M. Yang, K. V. Sokolovsky, D. Hatzidimitriou, G. Mountrichas, P. Gavras, V. Charmandaris, I. Bellas-Velidis, Z. T. Spetsieri, K. Tsinganos
MMNRAS , 1–20 (2019) Preprint 30 May 2019 Compiled using MNRAS L A TEX style file v3.0
Robust Identification of Active Galactic Nuclei through
HST
Optical Variability in GOODS-S: Comparison withthe X-ray and mid-IR Selected Samples
E. Pouliasis , (cid:63) , I. Georgantopoulos , A. Z. Bonanos , M. Yang , K. V. Sokolovsky , , ,D. Hatzidimitriou , , G. Mountrichas , P. Gavras , , V. Charmandaris , , I. Bellas-Velidis ,Z. T. Spetsieri , and K. Tsinganos , IAASARS, National Observatory of Athens, 15236 Penteli, Greece Department of Astrophysics, Astronomy & Mechanics, Faculty of Physics, University of Athens, Zografos, 15783 Athens, Greece Sternberg Astronomical Institute, Moscow State University, Universitetskii pr. 13, 119992 Moscow, Russia Astro Space Center of Lebedev Physical Institute, Profsoyuznaya Str. 84/32, 117997 Moscow, Russia Rhea for ESA/ESAC, Camino bajo del Castillo, s/n, Urbanizacion Villafranca del Castillo, Villanueva de la Ca˜nada, 28692 Madrid, Spain University of Crete, Department of Physics, GR-71003 Heraklion, Greece Institute of Astrophysics, Foundation for Research and ˆId’echnology-Hellas, GR-71110 Heraklion, Greece
Accepted 2019 May 27. Received 2019 May 24; in original form 2019 February 15
ABSTRACT
Identifying Active Galactic Nuclei (AGNs) through their X-ray emission is efficient,but necessarily biased against X-ray-faint objects. We aim to characterize this biasby comparing X-ray-selected AGNs to the ones identified through optical variabil-ity and mid-IR colours. We present a catalogue of AGNs selected through opticalvariability using all publicly available z-band
Hubble
Space Telescope images in theGOODS-South field. For all objects in the catalogue, we compute X-ray upper limitsor discuss detections in the deepest available ∼ Chandra
Deep Field South imagesand present the
Spitzer /IRAC mid-IR colours. For the variability study, we consideronly sources observed over at least five epochs and over a time baseline of up toten years. We adopt the elevated median absolute deviation as a variability indicatorrobust against individual outlier measurements and identify 113 variability-selectedAGN candidates. Among these, 26 have an X-ray counterpart and lie within the con-ventional AGN area in the F X /F opt diagram. The candidates with X-ray upper limitsare on average optically fainter, have higher redshifts compared to the X-ray detectedones and are consistent with low luminosity AGNs. Out of 41 variable optical sourceswith IR detections, 13 fulfill the IR AGN colour selection criteria. Our work empha-sizes the importance of optical variability surveys for constructing complete samplesof AGNs including the ones that remain undetected even by the deepest X-ray andIR surveys. Key words: galaxies: active – galaxies: photometry – galaxies: nuclei – x-rays: galax-ies – methods: observational
It is widely accepted that massive galaxies and a fractionof lower mass galaxies host a supermassive black hole intheir centre (Magorrian et al. 1998; Kormendy & Kennicutt2004; Filippenko & Ho 2003; Barth et al. 2004; Greene &Ho 2004, 2007; Dong et al. 2007; Greene et al. 2008). More- (cid:63)
E-mail: [email protected] over, there is a correlation between the mass of the blackholes and the properties of their host galaxies (Kormendy& Ho 2013), such as the luminosity, the stellar mass, thevelocity dispersion or the bulge rotational velocity (Dressler1989; Kormendy & Richstone 1995; Magorrian et al. 1998;Ferrarese & Merritt 2000; Gebhardt et al. 2000; Tremaineet al. 2002; Marconi & Hunt 2003; H¨aring & Rix 2004; Fer-rarese & Ford 2005; Graham & Driver 2007; G¨ultekin et al.2009). Hence, either the processes that take place in Active © a r X i v : . [ a s t r o - ph . GA ] M a y E. Pouliasis et al.
Galactic Nuclei (AGNs) play an important role in star for-mation and shaping of the galaxy structure, or vice versathe galaxy evolution directly affects the mass and spin ofthe central black hole. To better understand the relationsbetween the central black hole and its host galaxy, it is nec-essary to have complete samples of AGNs, not biased againstredshift, obscuration, luminosity, etc.X-rays penetrate deep into material with high columndensities and therefore are nearly unaffected by moderateobscuration ( N H < cm − ) (Brandt & Alexander 2015;Alexander 2017). Thus wide-field X-ray imaging constitutesthe most common and efficient technique to identify AGN.Other methods make use of infrared (IR), ultraviolet and op-tical single and multiple colour selection criteria to identifyAGNs, as the AGN spectral energy distributions differ fromthose of normal galaxies or stars (Richards et al. 2002, 2005;Lacy et al. 2004; Stern et al. 2005; Alonso-Herrero et al.2006; Donley et al. 2007; Schneider et al. 2007, 2010). How-ever, at these wavelengths, the AGN samples may be con-taminated by foreground stars or biased by the dust emissionfrom the host galaxies compared to X-ray AGN identifica-tion where the dilution by the hosts is not that dominant.An alternative method to identify AGNs is based onthe detection of variability at any wavelengths on timescalesfrom hours to years (Ulrich et al. 1997; Kawaguchi et al.1998; Paolillo et al. 2004; Garc´ıa-Gonz´alez et al. 2014).There is a correlation between the amplitude of the AGNvariability and the timescale of variation (Hook et al. 1994;Trevese et al. 1994; Cristiani et al. 1997; di Clemente et al.1996; Vanden Berk et al. 2004; Kelly et al. 2009; Bauer et al.2009; Middei et al. 2017), the redshift and the black holemass (for timescales longer than 100 days; Cristiani et al.1990; Hook et al. 1994; Trevese et al. 1994; Vanden Berket al. 2004). On the other hand, the variability amplitude isanticorrelated with the rest-frame wavelength (di Clementeet al. 1996; Cristiani et al. 1997; Giveon et al. 1999; Helfandet al. 2001; Vanden Berk et al. 2004; Zuo et al. 2012) and thenuclear luminosity (Trevese et al. 1994; Vagnetti et al. 2016).The latter suggests that low-luminosity AGNs (LLAGNs)will dominate a variability survey compared to more lumi-nous AGNs.Several theories and mechanisms have been proposedto explain the variability. For example, Rees (1984) andKawaguchi et al. (1998) proposed that AGN variability maybe due to the fact that accretion disks are vulnerable to dy-namical instabilities, while Li & Cao (2008) and Zuo et al.(2012) attributed the variability to fluctuations in the accre-tion rate. In the extreme case of blazars (the most luminousclass of AGNs possessing a relativistic jet pointing towardsthe observer), the variability is modulated by changes in theaccretion disk, but also by non-thermal power-law emissionfrom the jets and the (not fully explored) connection be-tween them (Gu & Li 2013; Finke & Becker 2014; Chatterjeeet al. 2018). Other interpretations of AGN variability includegravitational microlensing effects (Hawkins 1993; Alexander1995), tidal disruption events (Komossa 2015) and multipleexplosions of supernovae (SNe) near the nuclei (Kawaguchiet al. 1998; Terlevich et al. 1992).In the last several years, AGN variability has been usedin many studies. In the X-ray region, Young et al. (2012)and Ding et al. (2018) selected LLAGNs in the 4 and 7Ms Chandra
Deep Field South (CDF-S), respectively. In the optical and near IR bands, Sarajedini et al. (2003, 2011);Trevese et al. (2008); Villforth et al. (2010); Simm et al.(2015); Falocco et al. (2015); Graham et al. (2014); De Ciccoet al. (2015); Baldassare et al. (2018); Kim et al. (2018)identified a large sample of AGNs, suggesting that the searchfor variability at short time scales is efficient in selectingLLAGNs that would have been missed by X-ray surveys. AsAGNs exhibit a red-noise behaviour (Lawrence & Papadakis1993; Park & Trippe 2017; i.e. they have more power atlow frequencies in the Fourier space), De Cicco et al. (2015)and Paolillo et al. (2017) pointed out that the longer thetime baseline (e.g. greater than few years), the larger thevariability amplitude and the more complete is the AGNselection.However, while blazars tend to show variability at alltimescales, the power spectrum and structure function anal-ysis of light curves of many radio-quiet AGNs suggest thattheir variability amplitude does not rise indefinitely withlonger timescales. Their power spectrum flattens below somefrequency. Such power spectra were modeled with a dampedrandom walk and continuous auto-regressive moving aver-age models (de Vries et al. 2005; Kelly et al. 2009; MacLeodet al. 2010; MacLeod et al. 2012; Kelly et al. 2014; Kasliwalet al. 2015; Kozlowski 2016; Simm et al. 2016).In order to study AGN variability over cosmic time oneneeds a deep field observed multiple times. The Great Ob-servatories Origins Deep Survey Southern field (GOODS-S Giavalisco et al. 2004) centered at α = δ = − ◦ (cid:48) (cid:48)(cid:48) J2000, covers an area of (cid:48) × (cid:48) . It is the mostdata-rich area of the sky in terms of depth and wavelengthcoverage and as it has been observed by the Hubble
SpaceTelescope (HST) multiple times, it perfectly satisfies the re-quirements of our study.Variability studies in this field based on HST multi-epoch data have been performed by Sarajedini et al. (2011)and Villforth et al. (2010, 2012). Sarajedini et al. (2011)used the V -band (F606W) images over five epochs span-ning almost seven months and the standard deviation, σ ,as the statistical variability indicator to identify 42 variablesources. The authors compared their results with the mid-IR and the 2Ms Chandra
X-ray data. Villforth et al. (2010)identified 88 variable sources (out of ∼ σ , to the expected one σ exp , in this case scaled fromthe estimated photometric errors; de Diego 2010) on z -banddata with the same epochs and time baseline as in Sarajediniet al. (2011). The authors, after removing the false positivedetections and the stellar population, validated the AGN na-ture of 55/88 variable sources through spectral energy dis-tribution fitting, the identification of X-ray counterparts inthe CDF-S 4Ms catalogue and auxiliary radio and IR data(Villforth et al. 2012).The field was also targeted in ground-based optical vari-ability studies. Trevese et al. (2008) studied the variabilityof sources in AXAF, a larger field that includes GOODS-S. They analysed V -band images taken from ground-basedtelescopes and used magnitude differences between eightepochs over two years of observations to identify 132 vari-able AGN candidates. Similarly, Falocco et al. (2015) ap-plied a multi-epoch variability search spanning six monthswith the SUDARE-VOICE survey dataset obtained with theVLT Survey Telescope. They selected 175 variable sources MNRAS000
X-ray data. Villforth et al. (2010)identified 88 variable sources (out of ∼ σ , to the expected one σ exp , in this case scaled fromthe estimated photometric errors; de Diego 2010) on z -banddata with the same epochs and time baseline as in Sarajediniet al. (2011). The authors, after removing the false positivedetections and the stellar population, validated the AGN na-ture of 55/88 variable sources through spectral energy dis-tribution fitting, the identification of X-ray counterparts inthe CDF-S 4Ms catalogue and auxiliary radio and IR data(Villforth et al. 2012).The field was also targeted in ground-based optical vari-ability studies. Trevese et al. (2008) studied the variabilityof sources in AXAF, a larger field that includes GOODS-S. They analysed V -band images taken from ground-basedtelescopes and used magnitude differences between eightepochs over two years of observations to identify 132 vari-able AGN candidates. Similarly, Falocco et al. (2015) ap-plied a multi-epoch variability search spanning six monthswith the SUDARE-VOICE survey dataset obtained with theVLT Survey Telescope. They selected 175 variable sources MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S over an area of 2 de g around CDF-S using σ as the variabil-ity index. They compared the optical variable sample withAGNs selected through optical-NIR and IR colour diagnos-tics and AGNs with X-ray counterparts in the 4Ms CDF-Scatalogue.We extend the previous HST-based studies of Sarajediniet al. (2011) and Villforth et al. (2010, 2012) by using thelatest data, variability detection and IR-colour-based AGNselection techniques. We construct a new catalogue of opti-cally variable AGNs based on HST z -band observations andcomparing it with other selection techniques. We highlightthe following novel aspects of this study: • We expand the time baseline of the deep HST observationsof GOODS-S up to ten years, which should result in a morecomplete AGN selection. • We use the Median Absolute Deviation (MAD) as thevariability-detection statistic, which, unlike σ , is robustagainst individual outlier measurements (Sokolovsky et al.2017b). We expect MAD to yield a cleaner sample of variablesources compared to the previous studies. • We use the new deepest available 7 Ms
Chandra imageto constrain the X-ray brightness of the variability-selectedAGNs. • We compare our variable sample with AGN selected in themid-IR using the Lacy et al. (2007) and Donley et al. (2012)criteria.The HST optical observations and the data reduction(astrometry and photometry) along with ancillary data usedin this work are presented in Section 2, while in Section 3,we describe the method we used to create the list of vari-able sources. We also exclude stars and supernovae from thesample of the AGN candidates. In Section 4, we demonstratethe properties of the AGN candidates (e.g. magnitude andredshift distributions or X-ray luminosities) and constructthe mid-IR AGN samples. In Section 5, we compare our re-sults with other variability studies and we discuss the differ-ences between optically variable, mid-IR and X-ray selectedAGNs, while Section 6 presents the summary of the resultsand conclusions.
We analyze all publicly available images of the GOODS-S region obtained with the Wide Field Channel of the HSTAdvanced Camera for Surveys (ACS, Ford et al. 1998) in theF850LP filter ( z -band). The images are collected from theHubble Legacy Archive (HLA) Data Release 10 . Each im-age corresponds to an individual HST visit and results froma combination of three or more individual exposures withthe purpose of rejecting the cosmic rays. The observationswere collected in the framework of the observing programslisted in Table 1. We analyze totally 437 individual imagesspanning up to 10 years in some regions.We used the code developed by M. Tewes which isbased on P. G. van Dokkum’s L.A.Cosmic algorithm (van http://hla.stsci.edu/
16 18 20 22 24 26 28
F850LP (mag) M a g n i t u d e e rr o r S i g n a l - t o - N o i s e R a t i o Figure 1.
The magnitude errors as a function of the magni-tude for each detection. The measurements are colour coded bythe SNR. The vertical and horizontal lines represent the limitsin magnitude and magnitude error, respectively, after the SNRfiltering.
Dokkum 2001) to further reduce the cosmic-ray contamina-tion of the visit-combined images, especially on the edges ofthe combined frames. This algorithm is based on variationsof the Laplacian edge detection and is capable of rejectingany cosmic ray, regardless of its shape and size, keeping atthe same time the faint point-like sources untouched.Source detection and photometry was performed usingSExtractor (Bertin & Arnouts 1996, 2010). We applied the mexican hat spatial filter for detection and set the mini-mum contrast parameter for deblending ( deblend_mincont )to 0.0075 in order to avoid multiple detections for individualextended sources. This did not affect the unresolved sources,since the GOODS-S field is not crowded. For the photome-try and the variability analysis, we used a circular aperturewith a radius of 0.36 (cid:48)(cid:48) . This was the radius used by Villforthet al. (2010) as for smaller radii the photometry is affectedby changes in the point-spread function and aperture cen-tring issues. In addition, we measured the magnitudes fortwo more radii (0.05 (cid:48)(cid:48) and 0.15 (cid:48)(cid:48) ), which correspond to theones used for the
Hubble
Source Catalogue (HSC, Whitmoreet al. 2016) apertures (
MagAper1 and
MagAper2 ). The latterwere used to compute the concentration index ( CI ) and tovalidate our photometry against the HSC. CI is an indicatorof the extension of a source and is described below.After visually inspecting images associated with outlierpoints appearing in many light curves, we noticed that manyoutliers were situated near the frame edge or the gap be-tween CCD chips in these images. This is related to the back-ground estimation, which is essential for aperture photome-try. Because the images have been resampled to a north-upeast-left orientation, blank areas appear around the actualimage (e.g. CCD gaps and image edges). SExtractor usesthese blank areas and gets incorrect background estimates.To avoid this effect and, consequently, outliers and false-positive variable sources, we used the weight images pro-vided by HLA and excluded all detections located within 10pixels, or ∼ (cid:48)(cid:48) , from the edges ( ∼
1% of all the detections).To ensure the quality of the data but also enable the
MNRAS , 1–20 (2019)
E. Pouliasis et al.
Table 1.
The HST Treasury programs included in this studyProp. ID PI Name Cycle N imga N expb Exp. Time (s) Start Obs c End Obs c Note. – (a): Number of combined (level 2) images. (b): Number of single exposure images. (c):Starting and ending time of the observations. −0.5 0.0 0.5 . . . . Normalized measurement error D en s i t y Figure 2.
The distributions of the normalized measurement er-rors for each magnitude bin (gray). The black dashed line repre-sent the Gaussian fit to the data. The units are given in magni-tudes. search for small amplitude variations, we selected measure-ments with a signal-to-noise ratio (SNR) greater than five.The SNR was calculated for each of the detections for allimages using the fluxes and the corresponding flux errorsderived with SExtractor. In Figure 1, we plot the magni-tude error as a function of the magnitude, colour-coded withSNR. The relation between magnitude and magnitude erroris approximately linear as expected if the errors are mainlystatistical at the faint end. After applying the SNR cut-off,the faintest sources have an average error of about 0.25 mag.We also divided all the detections into magnitude binsand over-plotted the distributions of the normalized mag-nitude errors (after subtracting the median) for each of themagnitude bins in order to check if all detections of the same magnitude have similar errors. Figure 2 shows that thesedistributions are approximately Gaussian. The distributionwould have been skewed, if there was a group of sourcesmeasured with systematically larger uncertainties (e.g. dueto bright local background).Since the astrometric accuracy of the HST is limited bythe positional accuracy of individual Guide Star Cataloguestars (Lasker et al. 2008), we applied a triangle matchingtechnique based on the Valdes et al. (1995) algorithm to findthe astrometric solution. We used the match_v1 program byM. W. Richmond to automatically determine the coordi-nate system corrections using the 50 brightest sources ineach source list. We used the second version of HSC (HSCv2)as the reference catalogue for the astrometry and the result-ing positional errors are less than 0.1”.We then cross-matched the coordinates-correctedsource lists with each other to construct a light curve foreach source. We kept only sources with at least five measure-ments. Figure 3 presents the histogram of the median magni-tude, < F850LP > , of all the sources after the SNR filtering.Due to the drop-off of detected sources beyond 25.7 mag,our sample is photometrically complete down to this mag-nitude. Since the images used to derive the source lists havedifferent depth (Table 1), we over-plot in Figure 3 the com-pleteness curves of different images with extreme exposuretimes. We summed the detections of images with exposuretimes of ∼ ∼ p , as a function of the time baseline, T bas , which is de-fined as the time difference between the first and the lastobservation of a source. A large fraction of sources ( ∼ ) http://spiff.rit.edu/match/match-1.0/ MNRAS000
The distributions of the normalized measurement er-rors for each magnitude bin (gray). The black dashed line repre-sent the Gaussian fit to the data. The units are given in magni-tudes. search for small amplitude variations, we selected measure-ments with a signal-to-noise ratio (SNR) greater than five.The SNR was calculated for each of the detections for allimages using the fluxes and the corresponding flux errorsderived with SExtractor. In Figure 1, we plot the magni-tude error as a function of the magnitude, colour-coded withSNR. The relation between magnitude and magnitude erroris approximately linear as expected if the errors are mainlystatistical at the faint end. After applying the SNR cut-off,the faintest sources have an average error of about 0.25 mag.We also divided all the detections into magnitude binsand over-plotted the distributions of the normalized mag-nitude errors (after subtracting the median) for each of themagnitude bins in order to check if all detections of the same magnitude have similar errors. Figure 2 shows that thesedistributions are approximately Gaussian. The distributionwould have been skewed, if there was a group of sourcesmeasured with systematically larger uncertainties (e.g. dueto bright local background).Since the astrometric accuracy of the HST is limited bythe positional accuracy of individual Guide Star Cataloguestars (Lasker et al. 2008), we applied a triangle matchingtechnique based on the Valdes et al. (1995) algorithm to findthe astrometric solution. We used the match_v1 program byM. W. Richmond to automatically determine the coordi-nate system corrections using the 50 brightest sources ineach source list. We used the second version of HSC (HSCv2)as the reference catalogue for the astrometry and the result-ing positional errors are less than 0.1”.We then cross-matched the coordinates-correctedsource lists with each other to construct a light curve foreach source. We kept only sources with at least five measure-ments. Figure 3 presents the histogram of the median magni-tude, < F850LP > , of all the sources after the SNR filtering.Due to the drop-off of detected sources beyond 25.7 mag,our sample is photometrically complete down to this mag-nitude. Since the images used to derive the source lists havedifferent depth (Table 1), we over-plot in Figure 3 the com-pleteness curves of different images with extreme exposuretimes. We summed the detections of images with exposuretimes of ∼ ∼ p , as a function of the time baseline, T bas , which is de-fined as the time difference between the first and the lastobservation of a source. A large fraction of sources ( ∼ ) http://spiff.rit.edu/match/match-1.0/ MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S
18 20 22 24 26 28 30
F850LP (mag) N u m b e r All sources~2000 s~5000 s
Figure 3.
The distribution of the median magnitude, < F850LP > , of all the sources (gray shaded histogram) along withthe completeness curves for images with exposure times of 2000s (blue hatched) and 5000 s (red), respectively. The dashed lineindicates the completeness limit of our final sample. Figure 4.
The number of data points as a function of the maxi-mum time baseline covered by the light curve. has been observed for more than two and up to ten yearswith a median time baseline of 8.5 years. The average andmedian number of data points in the light curves are 15 and12, respectively.The photometric accuracy was tested by comparingour
MagAper2 magnitudes to those in HSCv2. First, wecross-matched all the sources with HSCv2 using a radiusof 1 (cid:48)(cid:48) , resulting in 7245 matches out of 21,647 final sources(our source list is much deeper than HSCv2 in this re-gion). In Figure 5, we plotted the difference in magni-tude, < F850LP > this work – < F850LP > HSCv2 as a functionof magnitude between this study and HSCv2 and find thevalues to be comparable. The relation is linear through thefull magnitude range as expected. We visually inspected theimages and the light curves of the outliers (marked withblack filled circles on Figure 5) and attributed the magni-
18 19 20 21 22 23 24 25 − . − . . . . MagAper2 (mag) M agn i t ude d i ff e r en c e ( m ag ) median 3 s s Figure 5.
A comparison between the < F850LP > magnitude ofour photometry and HSCv2. The grey points represent all thesources in common. The blue solid line shows the median, whilethe blue dotted and red dashed lines represent the 3 σ and 5 σ values of the magnitude difference, respectively. Sources exceedingthe 5 σ value are highlighted in black. CI (mag) < F L P > ( m a g ) N u m b e r Point-likeExtended
Figure 6.
The median magnitude, < F850LP > , as a functionof the CI for point-like (red circles) and extended sources (bluecrosses). The dashed line represents the chosen threshold at CI = . mag that separates the two populations. The upper panelshows the histogram of the CI . tude discrepancies to multiple detections of the same ex-tended source in the HSCv2.The separation of the extended and point-likesources was performed using CI as defined in the HSC, CI=MagAper1 − MagAper2 (Whitmore et al. 2016). The CI his-togram reveals two well-defined areas (Figure 6, top panel).We fitted two Gaussians to the two populations (point-likeand extended) and the point where these two fits come acrossis at CI =1.33 mag. Adopting it as the separation thresholdresults in 21,022 extended and 625 point-like sources. Fig-ure 6 (bottom panel) shows the CI as a function of magni-tude, where the two populations are plotted using differentcolours. MNRAS , 1–20 (2019)
E. Pouliasis et al.
RA (J2000) -28°00'55'50'45'-27°40' D e c ( J ) Figure 7.
Colour composite image of the exposure correctedand smoothed 7 Ms CDF-S. The white polygon represents theGOODS-S footprint.
GOODS-S is among the deepest and best studied fields inthe sky and over the last decades there is a variety of imagingand spectroscopic data available from radio to X-ray wave-lengths. In this study, apart from the HST optical observa-tions, we utilize X-ray and IR catalogues and, also, photo-metric or spectroscopic data when available to validate thenature of the variable sources.We use four X-ray catalogues from CDF-S and the Ex-tended Chandra Deep field South (ECDF-S) with differentdepths: 250 ks (Xue et al. 2016), 2 Ms (Luo et al. 2008), 4 Ms(Xue et al. 2011) & 7 Ms (Luo et al. 2017). The area studiedin this work partly overlaps with CDF-S, which is centeredat α = δ = − ◦ (cid:48) (cid:48)(cid:48) (J2000) and coversan area of ∼ . ECDF-S covers an area of . deg in the sky centered at α = δ = − ◦ (cid:48) (cid:48)(cid:48) (J2000). Figure 7 shows CDF-S and GOODS-S projected onthe sky. The catalogues are produced from the X-ray imagestaken by the Advanced CCD Imaging Spectrometer cam-era (ACIS, Garmire et al. 2003) aboard the Chandra X-RayObservatory . We also used these four catalogues to test thedependence of the number of optical counterparts on thedepth of the X-ray image.Table 2 presents the number of sources detected in atleast one band (the supplementary sources – lower signifi-cance X-ray sources with bright IR counterparts – of eachcatalogue are also included), the number of observationsand the observation dates with the on-axis sensitivity limitsin the soft (0.5 − − − Chandra images using the
CIAO software v2.0.1 to con-struct the 7 Ms image and calculated the X-ray flux upper-limits. First, we created the 7 Ms co-added images in thethree bands – broad, soft and hard, starting with the level2 event files. We used 99 observations taken from October1999 to March 2016. We kept only the central CCD chips( ccd_id =0,1,2,3) and filtered out flares affecting the back-ground in the light curve of each observation by maskingthe sources and using the deflare tool with the clean_lc option. For each observation, we ran the wavedetect tool tocreate source catalogues, so we could reproject the imagesat the same reference point in the sky to achieve a good ab-solute astrometric solution. The final step was to combinethe event files with the dmmerge tool and create images andexposure maps for all the bands.The smoothed image including all three bands is shownin Figure 7. For the non-detected sources in the X-rays, wemeasured the counts and the exposure effective areas in acircular region centered on the position of the optical coun-terparts. The radius, r i , used for each optical source wascalculated in a way that to enclose a specified fraction ofthe point spread function. The fraction adopted here de-creases from the on-axis (0.95) to off-axis sources (0.5). Thebackground was extracted from 500 circular regions (withaperture half of r i ) at random positions around the opticalsource (within a distance from 1.5 to 5 times of r i ) that donot overlap with other X-ray or the optical variable sources.We derived the count rate for each background region andfinally normalized the mean value of all of them to the areaof the source. Then we derived the upper limits with a con-fidence interval of 99.7%. The count rates were converted tofluxes, using an energy conversion factor equal to . × − , . × − and . × − ergs photon -1 for the broad, softand hard band, respectively by assuming a power-law modelwith photon index of Γ = . .GOODS-S overlaps, also, with the SpitzerIRAC/MUSYC Public Legacy Survey in the ExtendedChandra Deep Field South (PI: Pieter van Dokkum,SIMPLE). SIMPLE covers an area of ∼ surrounding GOODS-S and contains photometry for ∼ µ m , 4.5 µ m , 5.8 µ m and 8.0 µ m ) to construct the IR selected AGN sample(Section 4.2), while the r, J, K and 3.6 µ m bands from thesame catalogue were used to separate the stellar from theextra-galactic objects in Section 3.2. The full description ofthese data can be found in Damen et al. (2011). MNRAS000
CIAO software v2.0.1 to con-struct the 7 Ms image and calculated the X-ray flux upper-limits. First, we created the 7 Ms co-added images in thethree bands – broad, soft and hard, starting with the level2 event files. We used 99 observations taken from October1999 to March 2016. We kept only the central CCD chips( ccd_id =0,1,2,3) and filtered out flares affecting the back-ground in the light curve of each observation by maskingthe sources and using the deflare tool with the clean_lc option. For each observation, we ran the wavedetect tool tocreate source catalogues, so we could reproject the imagesat the same reference point in the sky to achieve a good ab-solute astrometric solution. The final step was to combinethe event files with the dmmerge tool and create images andexposure maps for all the bands.The smoothed image including all three bands is shownin Figure 7. For the non-detected sources in the X-rays, wemeasured the counts and the exposure effective areas in acircular region centered on the position of the optical coun-terparts. The radius, r i , used for each optical source wascalculated in a way that to enclose a specified fraction ofthe point spread function. The fraction adopted here de-creases from the on-axis (0.95) to off-axis sources (0.5). Thebackground was extracted from 500 circular regions (withaperture half of r i ) at random positions around the opticalsource (within a distance from 1.5 to 5 times of r i ) that donot overlap with other X-ray or the optical variable sources.We derived the count rate for each background region andfinally normalized the mean value of all of them to the areaof the source. Then we derived the upper limits with a con-fidence interval of 99.7%. The count rates were converted tofluxes, using an energy conversion factor equal to . × − , . × − and . × − ergs photon -1 for the broad, softand hard band, respectively by assuming a power-law modelwith photon index of Γ = . .GOODS-S overlaps, also, with the SpitzerIRAC/MUSYC Public Legacy Survey in the ExtendedChandra Deep Field South (PI: Pieter van Dokkum,SIMPLE). SIMPLE covers an area of ∼ surrounding GOODS-S and contains photometry for ∼ µ m , 4.5 µ m , 5.8 µ m and 8.0 µ m ) to construct the IR selected AGN sample(Section 4.2), while the r, J, K and 3.6 µ m bands from thesame catalogue were used to separate the stellar from theextra-galactic objects in Section 3.2. The full description ofthese data can be found in Damen et al. (2011). MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S Table 2.
Summary of the basic information of the CDF-S X-ray catalogues.
X-ray Number of Observation F F -2 s -1 ) (ergs cm -2 s -1 ) (ergs cm -2 s -1 ) optical variable250 ks ECDF-S 9 2004 430 > . × − > . × − > . × −
144 142 Ms CDF-S 23 1999 - 2000 578 > . × − > . × − > . × −
298 164 Ms CDF-S 54 1999 - 2010 776 > . × − > . × − > . × −
464 217 Ms CDF-S 102 1999 - 2016 1055 > . × − > . × − > . × −
621 24
Note. – The flux limits of the 7 Ms CDF-S catalogue in the broad and the hard band are derived up to 7 keV.
18 20 22 24 26 . . . . . .
18 20 22 24 26 . . . . . . .
18 20 22 24 26
18 20 22 24 26
Figure 8.
The
MAD (upper panels) and normalized significance
MAD * (lower panels) as a function of the median magnitude ( < F850LP > )for the extended (left) and point-like sources (right). All the sources in our survey are shown with grey points, while the AGN candidatesare shown with red circles, and the confirmed SNe with black triangles. The blue solid and dashed lines represent the median and thethreshold, while the vertical black dashed line the completeness limit of our sample. Sokolovsky et al. (2017b) discussed two classes of statisticalmethods that quantify variability of a source. The first classquantifies the scatter of the magnitudes within a light curve,while the methods of the second class quantify the smooth-ness of a light curve by taking into account the order and http://cxc.harvard.edu/ciao time at which the magnitude measurements were obtained.Regular variability can be detected that way too, if the ob-serving cadence is shorter than the variability timescale (Fer-reira Lopes & Cross 2016), or, if the scatter is higher thanwhat is expected from noise (Ferreira Lopes & Cross 2017).Light curve simulations by Sokolovsky et al. (2017b)suggest that the scatter-based methods are more suitablefor detection of variability in light curves having a smallnumber of points compared to the methods that character- MNRAS , 1–20 (2019)
E. Pouliasis et al. ize the light curve smoothness. Median Absolute Deviation (Rousseeuw & Croux 1993, MAD ) belongs to the first classof methods. It is the most robust to outliers among thevariability indices discussed by Sokolovsky et al. (2017b).In Appendix A1, we compare the performance of variousvariability-detection statistics in the presence of photomet-ric outliers and find that MAD is a reliable method, resistantto individual outlier measurements.
MAD is defined as the median value of the absolutedeviations of the measurements, m j , from the median: MAD = b × median (| m j − median ( m j )|) , (1)where b = /(√ − ( / )) (cid:39) . is the factor scaling themedian absolute deviation to the standard deviation (assum-ing the normal distribution of m j ); erf − is the inverse errorfunction.Specifically, our variability detection algorithm works asfollows: we divided the sources into magnitude bins and byassuming a Gaussian distribution, we calculated for each binthe median magnitude, the median MAD and the standarddeviation ( σ ) of MAD. The bin size is adjusted so as to haveat least 50 sources in each bin. To get a smooth magnitudedependence, we fitted a cubic spline to the median and thethreshold values. We, also, extrapolated toward fainter mag-nitudes to account for the completeness limit (Section 2).Taking into account that the majority of the sources arenormal galaxies and no variations are expected, the variablesources are those that exceed a cut-off above the median.We note that Sarajedini et al. (2011) rely on the same criti-cal assumptions as we do here: that the majority of sourcesare non-variable and that sources of similar brightness havesimilar photometric errors. Villforth et al. (2010) also relyon this assumption indirectly when they derive the scalingfactors for the estimated photometric errors that they useto compute the C -statistic.Following Bershady et al. (1998), or more recently Sara-jedini et al. (2011), we determine the threshold separatelyfor the point-like and the extended sources (except we de-termine the threshold in MAD scaled to σ rather than in σ as Sarajedini et al. 2011). Figure 8 (upper panels) shows thevariability index, MAD , as a function of the median mag-nitude, < F850LP > for the extended (left) and point-like(right) sources. We also calculated the normalized signifi-cance, MAD ∗ , for each source through the following formula: MAD ∗ i = MAD i − median ( MAD ) b σ ( MAD ) b , (2)where MAD i is the MAD for the i th source and b the cor-responding magnitude bin. The significance has units of σ .The plots of MAD * as a function of magnitude are shown inthe lower panels of Figure 8.We set the threshold of . σ in MAD* above which weconsider the sources to be variable. Assuming the normaldistribution of MAD* we estimate the fraction of sourcesthat are expected to have the value of MAD* above the https://en.wikipedia.org/wiki/Median_absolute_deviation . . . . . . MAD* N u m be r Figure 9.
The median cumulative distributions of the normal-ized significance for under-sampled data with different numberof data points in the light-curve (blue thin lines). The red thickline represents the distribution from our final sample, while thehorizontal dashed and dotted line represents the statistical signif-icance of 100% and 99.65%, respectively. The vertical line is our3.5 σ threshold. threshold, i.e. the false positive rate, as: FP rate = − ( + erf ( . √ )) (cid:39) . × − , (3)so out of 21,022 extended and 625 point-like sources of theinitial sample we expect 5 and < false positives among theextended and point-like candidate variable sources, respec-tively. This corresponds to ∼ σ thresholds is 99.65%. However,the true significance should be greater than this value, sincethe distribution also includes the variable sources. We visu-ally inspected the light curves of all the candidate variablesources and the associated images. We checked for diffrac-tion spikes from nearby foreground stars, close neighbors,poorly removed cosmic rays, proximity to a frame edge and MNRAS000
The median cumulative distributions of the normal-ized significance for under-sampled data with different numberof data points in the light-curve (blue thin lines). The red thickline represents the distribution from our final sample, while thehorizontal dashed and dotted line represents the statistical signif-icance of 100% and 99.65%, respectively. The vertical line is our3.5 σ threshold. threshold, i.e. the false positive rate, as: FP rate = − ( + erf ( . √ )) (cid:39) . × − , (3)so out of 21,022 extended and 625 point-like sources of theinitial sample we expect 5 and < false positives among theextended and point-like candidate variable sources, respec-tively. This corresponds to ∼ σ thresholds is 99.65%. However,the true significance should be greater than this value, sincethe distribution also includes the variable sources. We visu-ally inspected the light curves of all the candidate variablesources and the associated images. We checked for diffrac-tion spikes from nearby foreground stars, close neighbors,poorly removed cosmic rays, proximity to a frame edge and MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S saturation or misalignment of individual exposures. All thesefactors may introduce false variability. Following this proce-dure, we classified all the variable sources into three cat-egories. Sources with clear variability in their light curve,far from other sources with no artifacts or potential prob-lems recorded and with accumulated significance higher than99.9%, were assigned grade A (86 sources). Sources with mi-nor problems that may affect the reliability were assigned agrade B (32 sources). This category includes sources thatmight have centring issues, caused by the extension of theobject or that are too faint and dispersed and sources be-tween 3.5 sigma and accumulated significance of 99.9%. Fi-nally, all sources that were found to be affected by satura-tion, diffraction spikes, blending or other significant prob-lems were assigned a grade C (69 sources). In order to separate AGN candidates from stars, we followedRowan-Robinson et al. (2005) and Damen et al. (2011). Ac-cording to Rowan-Robinson et al. (2005), stars can be distin-guished from the extra-galactic objects, such as QSOs andvery distant AGNs by their brightness in the r band (r < . µ m / r flux ratio versus the r − i diagram (Figure 10, right). The two populations occupydifferent regions in the diagram and can be easily separated.This method was also used by Falocco et al. (2015) andRowan-Robinson et al. (2013). On the other hand, Damenet al. (2011) excluded the stellar population using a colourcut-off ( [ J − K ]( AB ) < . ma g ) and applied certain qual-ity criteria to their initial sample: signal-to-noise ratio in Kband ( S / N ) K > and their relative weight in the K bandversus the z- band, w K > . .We cross matched our initial catalogue of 21,647 sourceswith the SIMPLE data (Sec. 2.2) to obtain the colours forour sources and applied the above diagnostics. Figure 10shows the [ J − K ] vs. . µ m and the . µ m / r vs. [ r − i ] di-agrams with the sources colour coded by the CI . The openblack circles presented in the plot are used to show the vari-able sources in each diagram. Both diagnostics indicatedeight classified variable sources to be stars. All of them aregrade C variable sources, as six were saturated and two wereblended sources. We therefore identify no high-confidencecandidate variables among the foreground stars.For the identification of SNe, we relied on visual in-spection of the light curves and the corresponding images.We found three sources with light curves of Grade A re-sembling SNe (ID 7343, 9581 & 14446), which have all beenpreviously reported in the literature by Strolger et al. (2004)and Riess et al. (2007). Their light curves and observationalproperties can be found in Figure B1 and Table B1 in theAppendix B. Out of the SNe catalogues, there are 13 moreSNe that have a counterpart in our initial sample, but theyare all below the variability threshold. Possible explanationscould be the differences in the observation dates (the peakwas not observed), or that our variability algorithm couldnot detect variability if the number of the data points inthe light curve that correspond to the peak was small andthey were considered as outliers by MAD. For these sources,we also calculated the standard deviation, but they werestill below the threshold, thus the first case is more likely tohappen. In the next sections, we proceed with the analysis ofthe remaining 113 variable sources (10 point-like and 103extended), which are presumed to be AGN candidates (Ta-ble 3). Figure 11 illustrates the positions of the variablesources on the sky, while some examples of the AGN lightcurves can be found in Figure 12. We cross-matched the final sample of the 113 AGN can-didates with the four X-ray catalogues described in Sec-tion 2.2. We used a search radius of 2” as the maximal posi-tional error of the X-ray sources reaches values of ∼ F X / F opt diagram (Maccacaro et al. 1988; Barger et al. 2003; Horn-schemeier et al. 2003), where F X and F opt are the X-rayand optical flux, respectively. The conventional AGN pop-ulation lies in the area between log ( F X / F opt ) = ± , whilespectroscopically confirmed AGN have been reported up to log ( F X / F opt ) = ± . The normal galaxies are expected to have log ( F X / F opt ) (cid:54) − . In Figure 13, we plot the optical flux(F850LP) as a function of the X-ray (0.5-8 keV) flux of ouroptically variable AGNs compared with the normal galaxy(filled green squares) and AGN (open black squares) popu-lations from Luo et al. (2017). The optically variable AGNswith X-ray detections (filled red circles) lie over the wholearea within F X / F opt = ± . We could be quite confident thatsources with F X / F opt > − are AGNs, while between -2 and-1 their flux ratio is still consistent with AGNs, but theirnature is confirmed by the combination of variability withX-ray emission.The flux upper limits (open blue circles) are consis-tent with AGNs according to their F X / F opt ratio, thoughdeeper X-ray images are needed to detect them. In particu- MNRAS , 1–20 (2019) E. Pouliasis et al.
Table 3.
Catalogue of the variable AGN candidates.
ID Grade RA Dec N p T bas CI < F850LP > MAD * z z Ref. F x [0.5-8 keV](J2000) (J2000) (years) (mag) (mag) ( σ ) (ergs cm -2 s -1 )(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)347 B 52.94086 -27.85703 5 9.1 2.22 22.56 5.93 0.58 3,c < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < Note. – (1): Identifier. (2): Quality grade. (3): Right ascension. (4): Declination. (5): Number of data points in the lightcurve. (6): Time baseline. (7): Concentration index. (8): Median magnitude in the F850LP filter. (9): Normalized
MAD . (10):Redshift. (11): Method used to compute z (1: spectroscopy, 2: grism and 3: photometry), while the letter refers to the paperwhere the redshift obtained (a: Luo et al. (2017), b: Momcheva et al. (2016), c: Cardamone et al. (2011), d: Straatman et al.(2017), e: Xue et al. (2016), f: Taylor et al. (2009) and g: Wolf et al. (2008)). (12): Flux in the X-ray [0.5-8 keV] band. The’ < ’ symbol represent the flux upper limit. MNRAS000
MAD . (10):Redshift. (11): Method used to compute z (1: spectroscopy, 2: grism and 3: photometry), while the letter refers to the paperwhere the redshift obtained (a: Luo et al. (2017), b: Momcheva et al. (2016), c: Cardamone et al. (2011), d: Straatman et al.(2017), e: Xue et al. (2016), f: Taylor et al. (2009) and g: Wolf et al. (2008)). (12): Flux in the X-ray [0.5-8 keV] band. The’ < ’ symbol represent the flux upper limit. MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S
18 20 22 24 26 [3.6] (AB) [ J - K ] ( A B ) C o n c e n t r a t i o n i n d e x , C I r-i (AB) l o g ( F . / F r ) C o n c e n t r a t i o n i n d e x , C I Figure 10.
The [ J − K ] versus . µ m diagram (left) and the . µ m / r flux ratio versus r − i diagram (right). The crosses represent thecommon sources between our sample and the SIMPLE data. The filled circles are the sources classified as stars. All sources are colourcoded by the CI . The open black circles indicate the variable sources. − . − . − . − . R.A. (J2000) / deg D e c . ( J ) / deg Figure 11.
Spatial distribution of the AGN candidates (red cir-cles). The confirmed SNe are shown with black triangles. Thebackground grey points represent the whole sample consisting of21,647 sources. lar, the majority of the flux upper limits lie below the aver-age F X / F opt = line, indicating that these AGNs are mostlyX-ray weak. Concerning the sources that overlap with thenormal galaxies, Luo et al. (2017) mentioned that a fractionof the normal galaxy population may also include LLAGNs.For the sources detected only in the soft or hard band,we transformed the fluxes into the [0.5-8 keV] band using the WebPIMMS v4.8d software, assuming the photon index Γ = . and the Galactic HI column density n H = cm -2 .The same transformations were applied to the broad band( . − keV) of Luo et al. (2017).To further understand the nature of these sources, wecross-matched our AGN candidates with the catalogue ofMomcheva et al. (2016) to associate each source with thecorresponding redshift of the host galaxy. Their catalogueprovides the best redshift among grism, ground-based spec-troscopic or photometric redshifts (Skelton et al. 2014). Forthe candidates with X-ray counterparts, we used the spec-troscopic redshifts provided by Luo et al. (2017). The red-shifts of some sources that did not have a match in theprevious catalogues, were recovered from Straatman et al.(2017), Cardamone et al. (2011), Taylor et al. (2009) andWolf et al. (2008). We found published redshifts for all thevariability-selected AGN candidates: 63 sources have pho-tometric redshifts, while for the remaining 60 sources theredshifts were derived from spectra.Even though the available spectra from Momcheva et al.(2016) and the other catalogues are capable to provide secureredshifts, the width of the lines is too noisy for the classifi-cation of our optically faint AGN sample (broad or narrowlines). The redshift distribution for the AGN candidates ispresented in Figure 14 with the distribution of all the X-raysample of the 7 Ms CDF-S catalogue that have an opticalcounterpart in GOODS-S for comparison. The AGNs withnot yet detected X-ray emission extend to higher redshifts.Figure 15 shows the magnitude distribution of the can-didate AGNs and also the magnitude distribution of the X-ray sample reported in Luo et al. (2017) with an opticalcounterpart in GOODS-S (both AGNs and normal galaxies).It is very clear that the sample with X-ray upper limits isoptically fainter than the AGN candidates with X-ray detec- https://heasarc.gsfc.nasa.gov/docs/software/tools/pimms.html MNRAS , 1–20 (2019) E. Pouliasis et al. < F L P > ( m a g ) < F L P > ( m a g ) MJD (days)
MJD (days)
Figure 12.
Example light curves of the AGN candidates. The dashed line indicates the median magnitude and the number on the topof each plot indicates the identifier of the source. tions. The redshift and magnitude distributions suggest thatthe optical surveys, such as HST, may be able to identifyfaint high-redshift AGNs through variability. These AGNswould have been missed by current X-ray studies. Since faintnuclear emission can be observed in the optical, there are noobscuration at all or very weak obscuration effect by dust,thus these AGNs are likely LLAGNs. Their position in the F X / F opt diagram suggests that these high-redshifted intrin-sically X-ray weak AGNs lie below the conventional AGNpopulation (around F X / F opt = ), and thus, the dependenceon redshift and X-ray flux should be considered when work-ing with F X / F opt diagrams.We next estimated the X-ray luminosity [2-10 keV] ofall the AGN candidates. We found the sample of variablesources with X-ray detections and upper limits to be dis-tributed over ∼ ∼ . to ergs s − and . to ergs s − , respectively,with mean values of . × and . × ergs -1 s -1 .In Figure 16, we compare our luminosity distribution withthe distribution of different populations of AGNs located in the nearby universe (z ∼ In addition to studying their X-ray properties, we explorewhether our optically variable AGN candidates show evi-dence of accretion onto a supermassive black hole via theirinfrared emission. The mid-IR emission from AGNs, in par-ticular after the advent of sensitive spectrographs in spacetelescopes such as
ISO and
Spitzer , has proven extremelyuseful in revealing the presence of an AGN and character-ising whether it is of type I or type II (Clavel et al. 2000;Verma et al. 2005; Weedman et al. 2005; Wu et al. 2009;Alonso-Herrero et al. 2016). It is now widely accepted that
MNRAS000
MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S Table 3 – continued ID Grade RA Dec N p T bas CI < F850LP > MAD * z z Ref. F x [0.5-8 keV](J2000) (J2000) (years) (mag) (mag) ( σ ) (ergs cm -2 s -1 )(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)10511 A 53.11638 -27.87660 14 9.32 2.21 22.52 5.78 0.38 1,a < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < Note. – (1): Identifier. (2): Quality grade. (3): Right ascension. (4): Declination. (5): Number of data points in the lightcurve. (6): Time baseline. (7): Concentration index. (8): Median magnitude in the F850LP filter. (9): Normalized
MAD . (10):Redshift. (11): Method used to compute z (1: spectroscopy, 2: grism and 3: photometry), while the letter refers to the paperwhere the redshift obtained (a: Luo et al. (2017), b: Momcheva et al. (2016), c: Cardamone et al. (2011), d: Straatman et al.(2017), e: Xue et al. (2016), f: Taylor et al. (2009) and g: Wolf et al. (2008)). (12): Flux in the X-ray [0.5-8 keV] band. The’ < ’ symbol represent the flux upper limit. the AGN continuum emission appears as a power law fromthe 3 to 10 µ m range, since the strong UV and X-ray radi-ation destroys the molecules responsible for the PolycyclicAromatic Hydrocarbon (PAH) emission, while heating thesurrounding dust particles in thermal equilibrium to neardust sublimation temperatures. The mid-IR AGN spectrummay also display absorption features with variable strength(due to astronomical silicates at 9.7 and 18 µ m) dependingon the geometry of the obscuring dust as well as the lumi- nosity of the active nucleus compared with the host galaxy(Nenkova et al. 2008a,b).Even when mid-IR spectra are not available, one mayuse mid-IR broad-band colours to trace this slope. A num-ber of such diagnostics have been proposed using the IRACinstrument (Fazio et al. 2004) on board the Spitzer
SpaceTelescope (Werner et al. 2004) which provided imaging at3.6, 4.5, 5.8 and 8.0 µ m for a large sample of galaxies. Theseinclude the ”Lacy wedge” (Lacy et al. 2004, 2007; Sajinaet al. 2005), the ”Stern wedge” (Stern et al. 2005) and more MNRAS , 1–20 (2019) E. Pouliasis et al. log(F X ) [0.5-8 keV] l o g ( F o p t ) [ F L P ] -2 -1 log(F x / F opt )=0 +1 +2 Luo+17 GalaxyLuo+17 AGNUpper limitsX-ray detected
Figure 13.
Broad (0.5-8 keV) X-ray vs. optical ( < F850LP > ) flux for the 26 AGN candidates with X-ray counterparts (red circles). Openblue circles represent those sources for which only upper limits were derived. The open black and filled green squares in the backgroundrepresent the sources classified as AGN or normal galaxies in Luo et al. (2017), respectively. The dashed line indicates the log ( F X / F opt ) = and the solid lines from left to right correspond to log ( F X / F opt ) = − , − , + , + , respectively. The fluxes are given in units of ergs cm -2 s -1 . Redshift N u m b e r Upper limitsX-ray detectedLuo+17
Figure 14.
Redshift distribution for the candidate AGNs with(filled red) and without (hatch-filled blue) X-ray counterparts.The grey histogram indicates all the X-ray sources with opticalcounterparts in GOODS-S. recently the ”Donley wedge” (Donley et al. 2007, 2012). Sim-ilar methods have also been proposed for sources observedwith WISE (Stern et al. 2012; Assef et al. 2013; Mateoset al. 2012). We examine our candidates using the diagnos-tic of Donley et al. (2012), which has proven to be the mostrobust for a rather wide redshift range. The criteria by Lacyet al. (2007) are also used, for comparison.We used the SIMPLE data mentioned in Section 2.2.This sample is photometrically complete at 5 µ J y , wherethere is a turn-over in the number density plot of the fluxesin the [5.8] band. Furthermore, we restricted our analysis to3,904 mid-IR sources (IR sample) that have detections at allfour IRAC channels as well as an optical HST counterpartwith five or more data points in the light curve identified
18 20 22 24 26 28
Figure 15.
Median F850LP magnitude distribution for the candi-date AGNs with (red filled) and without (blue hatch-filled) X-raycounterparts. The grey histogram indicates all the X-ray sourceswith optical counterparts in GOODS-S. by our survey. Out of the IR sample, there are 41 opticallyvariable sources. Following the AGN selection criteria byDonley et al. (2012):1. x > . and y > . y > ( . ∗ x ) − . and y < ( . ∗ x ) + . f [ . ] > f [ . ] and f [ . ] > f [ . ] and f [ . ] > f [ . ] ,where x=log(f[5.8]/f[3.6]), y=log(f[8.0]/f[4.5]) and f[band] isthe flux of the corresponding band, we found 53 sources(hereafter, Donley IR AGNs). Out of those, 37 have X-raycounterparts, while five sources are optical variables. Thelatter five sources have also been detected in X-rays (in the7Ms image) and are classified as QSOs in the literature. Theoptical variability can be explained by their QSO nature; MNRAS000
Median F850LP magnitude distribution for the candi-date AGNs with (red filled) and without (blue hatch-filled) X-raycounterparts. The grey histogram indicates all the X-ray sourceswith optical counterparts in GOODS-S. by our survey. Out of the IR sample, there are 41 opticallyvariable sources. Following the AGN selection criteria byDonley et al. (2012):1. x > . and y > . y > ( . ∗ x ) − . and y < ( . ∗ x ) + . f [ . ] > f [ . ] and f [ . ] > f [ . ] and f [ . ] > f [ . ] ,where x=log(f[5.8]/f[3.6]), y=log(f[8.0]/f[4.5]) and f[band] isthe flux of the corresponding band, we found 53 sources(hereafter, Donley IR AGNs). Out of those, 37 have X-raycounterparts, while five sources are optical variables. Thelatter five sources have also been detected in X-rays (in the7Ms image) and are classified as QSOs in the literature. Theoptical variability can be explained by their QSO nature; MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S (a) Upper limitsX-ray detected246810 (b)
Panessa+06 Sf1Panessa+06 Sf2246810 (c)
Gonzalez-Martin+15 Sf1Gonzalez-Martin+15 Sf236 37 38 39 40 41 42 43 44 450246810 (d)
Gonzalez-Martin+15 L1Gonzalez-Martin+15 L2 log (L X ) [2-10 keV] N u m b e r Figure 16.
X-ray luminosities in the [2-10 keV] band (panel a) forthe AGN candidates, with (red filled) and 3 σ upper limits (bluehatch-filled) X-ray counterparts. In panel (b) the luminosity dis-tributions of Seyfert galaxies are shown from Panessa et al. (2006).Panel (c) and (d) represent the galaxy population of Seyferts andLINERs derived from Gonz´alez-Mart´ın et al. (2015). The blackand shaded grey histograms indicate the type I and II, respec-tively. i.e. direct view to the central source. At the same time thepowerful AGN heats the dusty torus (seen face on) and itsreprocessed emission dominates the infrared emission fromthe host galaxy.Lacy et al. (2007) used a similar mid-IR colour-colourdiagram with somewhat relaxed limits and without thepower law condition:1. x > − . and y > − . y > ( . ∗ x ) − . .Among the 770 sources that fulfill these criteria (hereafter,Lacy IR AGNs), there are 188 X-ray detections and 13 op-tically variable AGNs according to our analysis. The LacyIR AGNs contains all 53 Donley IR AGNs. Figure 17 showsthe IRAC 4-band colour-colour plot for the IR sample. Thelines represent the wedges as defined in Lacy et al. (2007)and Donley et al. (2012). We also over-plotted the opticallyvariable, Donley IR and the X-ray selected AGNs. We compared our variable sources with other variabilitystudies of GOODS-S, including Villforth et al. (2010) and log(S . /S . ) l o g ( S . / S . ) Figure 17.
IRAC colour-colour diagram of the IR sample (graypoints). The Lacy IR AGNs defined by the solid line. The DonleyIR AGNs are those inside the dashed line and follow an IR power-law (filled orange circles). The optically variable and the X-rayselected AGN samples are represented by open circles and bluecrosses, respectively.
Sarajedini et al. (2011), who also searched for optical vari-ability in this field. Out of the 88 variable sources reportedby Villforth et al. (2010), 86 sources were included in ourinitial sample of 21,647 sources. Out of these, ∼ wereidentified as variable with our method. Similarly, out of the42 variable sources of Sarajedini et al. (2011) in commonwith our initial sample, we recovered ∼ . In a larger fieldof view, Trevese et al. (2008) found 132 variable sources, 23of which lie in the area studied in this work and are includedin our survey; we find eight sources to be variable in our cat-alogue. Regarding the sample of Falocco et al. (2015), thereis only one common source with our survey, which is classi-fied as non-variable by our variability detection algorithm.The main differences between our study and the stud-ies of Villforth et al. (2010) and Sarajedini et al. (2011) liein the source detection algorithm and the larger amount ofdata. Our approach to identifying variable objects amonga set of light curves is similar to the one used by Saraje-dini et al. (2011) with two important modifications: 1) weused MAD as the measure of the light-curve scatter to filterout individual outliers and 2) we used the median instead ofmean to determine the expected value of scatter in a givenmagnitude bin. Villforth et al. (2010) used C statistics thatrely mostly on the estimated photometric uncertainties toselect variables, while Sarajedini et al. (2011) used the clas-sic standard deviation on V -band imaging data. Both useda 3 σ vs. a more secure threshold of 3.5 σ employed in thiswork. In order to check the dependence of the recovery rateand the false-positive contamination by the adopted variabil-ity threshold, we calculated the recovery rate of both studiesand also the percentage of the false positive rate out of the MNRAS , 1–20 (2019) E. Pouliasis et al. . . . . . . . . Threshold in MAD % False−PositiveVillforth+10Sarajedini+11Luo+17
Figure 18.
Percentage of false positive rate (black circles), re-covery rate of variable sources identified by Villforth et al. (2010)(blue triangles) and Sarajedini et al. (2011) (green squares) andpercentage of X-ray sources by Luo et al. (2017) (red crosses)identified as variables in this study as a function of different val-ues of thresholds. The y-axis is in logarithmic scale. variable sources (as described in Section 3.1) for differentvalues of the threshold. In Figure 18, we plot the results. Byrelaxing the threshold to lower values (MAD (cid:39) .Thus, the quality of the images used in this work and, con-sequently, the reliability of our photometry are supposed tobe much higher than previous studies. To facilitate a direct comparison and present the various se-lection methods in a uniform manner, we selected the Don-ley IR, Lacy IR and X-ray detected AGNs that lie inside the http://hla.stsci.edu/hla_faq.html Figure 19.
Venn diagram of the AGN samples selected throughoptical variability (blue), X-rays (orange), Donley et al. (2012)(red) and Lacy et al. (2007) (green) IR criteria. area of GOODS-S along with the optical variable AGN can-didates. In Figure 19, we demonstrate the overlapping of theoptical variability (113), Donley IR (53), Lacy IR (770) andX-ray selected (825) AGN samples with a Venn diagram.The deepest available X-ray catalogue contains 825sources in the area of GOODS-S. 621 sources have opti-cal counterparts with five or more data points in their lightcurve, while 587 have both optical and IR detections. Wefound ∼ ∼ ∼ ∼ ∼ r ≤ mag. Given the samelimits in the magnitude, we derive almost the same com-pleteness with respect to X-rays (Figure 20).Furthermore, in this work we set a 3.5- σ cut-off in MADto identify variables. Given a less conservative value of thecut-off, the percentage of X-ray detected sources in the vari-able sample increases. In particular, a 3- σ cut-off increasesthe X-ray detected sources to 30%. However, at the sametime the false positive variability rate also increases signifi- MNRAS000
Venn diagram of the AGN samples selected throughoptical variability (blue), X-rays (orange), Donley et al. (2012)(red) and Lacy et al. (2007) (green) IR criteria. area of GOODS-S along with the optical variable AGN can-didates. In Figure 19, we demonstrate the overlapping of theoptical variability (113), Donley IR (53), Lacy IR (770) andX-ray selected (825) AGN samples with a Venn diagram.The deepest available X-ray catalogue contains 825sources in the area of GOODS-S. 621 sources have opti-cal counterparts with five or more data points in their lightcurve, while 587 have both optical and IR detections. Wefound ∼ ∼ ∼ ∼ ∼ r ≤ mag. Given the samelimits in the magnitude, we derive almost the same com-pleteness with respect to X-rays (Figure 20).Furthermore, in this work we set a 3.5- σ cut-off in MADto identify variables. Given a less conservative value of thecut-off, the percentage of X-ray detected sources in the vari-able sample increases. In particular, a 3- σ cut-off increasesthe X-ray detected sources to 30%. However, at the sametime the false positive variability rate also increases signifi- MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S <22 (13) 22−23 (13) 23−24 (15) 24−25 (32) >25 (40)Total 7 Ms 4 Ms 2 Ms250 ks
The fraction of AGNs selected through optical vari-ability divided into five magnitude bins that are X-ray detectedin catalogues of different depths. The number of optically variableAGNs in each bin are shown in the parentheses. cantly. To avoid a high incidence of false variables, we nec-essarily miss a population of X-ray detected AGNs whichdisplay lower optical variability. Future surveys with highersensitivities and better sampled data with longer time base-lines will allow us to identify individually variable objectsand fully characterize their variability properties withouttaking into account the whole population statistics. In thatcase, variability will be able to recover the X-ray detectedAGN population with low significance (either low redshiftedor high luminous AGNs).Regarding the IR selected AGNs, the Lacy et al. (2007)method selected a large number of AGN candidates throughcolour-colour criteria (770), comparable to that of X-rayAGNs. 25% of these are X-ray detected, while the majorityof the X-ray sources fall outside the Lacy wedge (Fig. 17).Despite the large number of AGN candidates, the contami-nation of star-forming galaxies is expected to be as high as80% (Donley et al. 2012), as the sensitivity limit of our IRsample is at 5 µ J y . Donley et al. (2012) studied the star-forming contamination of the IR selected AGNs defined byLacy et al. (2007) and Stern et al. (2005) for samples withdifferent depths and revised these criteria by adding an ad-ditional power law criterion. Thus, the Donley IR sample of53 sources is expected to only have ∼
10% contamination asstated by Donley et al. (2012). Out of those, 37 have X-rayemission ( ∼ ∼
12% and ∼ Variability is a basic property of AGNs and has been provento be a reliable method to reveal non-obscured LLAGNs.Many studies assembled multi-epoch data and used variabil-ity to identify AGNs. The need, though, for highest photo-metric accuracy and long-term observational monitoring im-pose limits on the completeness of such surveys. Previouslyin the GOODS South field, Villforth et al. (2010, 2012) andSarajedini et al. (2003, 2011) used a five epoch dataset ( i -and z -band, respectively) spanning six months. In a largerarea and from ground-based telescopes, Falocco et al. (2015)used, also, a six month baseline, while Trevese et al. (2008)used a longer time baseline of two years.In this work, we substantially increased the time base-line by up to ten years using deep HST observations ( z -band).We created a new catalogue of optically variable AGNcandidates in the GOODS-S. We used SExtractor to con-struct the light curves of ∼ σ cut-off was appliedto identify variable sources. Our results can be summarizedas follows: • We identified 116 high confidence variable sources. Afterremoving three known Supernovae, we ended up with 113AGN candidates (103 extended and ten point-like). • We explored the mid-IR properties of our AGN candidates.41 sources have been detected in all four
Spitzer /IRACbands and, out of those, 13 and five sources are classi-fied as AGNs through the colour selection adopted by Lacyet al. (2007) and Donley et al. (2012), respectively. Also, thespace observations compared to ground-based studies iden-tify AGNs deep into the IR region when other methods fail. • We cross-matched our AGN sample with the published X-ray catalogues (CDF-S 2, 4 & 7 Ms and ECDFS 250 ks)and found 26 variable sources with X-ray counterparts. Thiscorresponds to ∼
23% (26/113). • For all the sources without X-ray detections, we used the7 Ms image in CDF-S and estimated the flux upper limitsusing a confidence level of 99.7%. These sources are opticallyfainter with higher redshifts up to z=4.
MNRAS , 1–20 (2019) E. Pouliasis et al. • The X-ray to optical flux ratios revealed that the variablesources are consistent with the AGN population, as they liewithin the area of − < log ( F X / F opt ) < + , but their average log ( F X / F opt ) ratio suggests that high-redshifted intrinsicallyX-ray weak AGNs lie below the conventional log ( F X / F opt ) = area. • The X-ray luminosities of our variable AGN candidatesare comparable to those of LLAGNs in the Local Universe(Panessa et al. 2006; Gonz´alez-Mart´ın et al. 2015). Hence,the variability in deep optical photometric data is a promis-ing method of finding optically low luminosity AGNs, whichthe X-ray observations may miss.We conclude that the different methods (optical vari-ability, IR, X-rays) used to identify AGNs are complemen-tary to each other and equally important to constrain thefull picture of the AGN demographics. In particular, opticalvariability is able to identify a large number of LLAGNs athigh redshifts. These are critical for studying the faint endof the AGN luminosity function and it might be the keybetween normal galaxies and AGNs.This work is part of the European Space Agency (ESA)project “
Hubble
Catalogue of Variables” (HCV, Sokolovskyet al. 2017a; Gavras et al. 2017), which aims to identify vari-able sources (stars, transients, Supernovae, AGN, etc.) fromthe Hubble Source Catalogue (HSC, Whitmore et al. 2016),through different filters and instruments. The HCV targetedfields are more than 250 and the number of sources includedexceeds 3.5 million. Specifically for AGNs, there are morethan 30 bona-fide deep fields covered by multi-wavelengthdata with observing time baselines more than two years.Extrapolating our results to these fields, we expect to iden-tify more than 2,000 new AGNs with a high fraction of thembeing LLAGNs. The variability detection technique used inthis work may be applied not only to HST observations butalso to other surveys such as the Large Synoptic SurveyTelescope (Ivezi´c et al. 2008, LSST).
ACKNOWLEDGMENTS
The authors are grateful to the anonymous referee for valu-able suggestions that significantly improved the manuscript.E. Pouliasis acknowledges financial support by ESA underthe HCV programme, contract no. 4000112940. This re-search has made use of the VizieR catalogue access tool,CDS, Strasbourg, France. The original description of theVizieR service is presented by Ochsenbein et al. (2000). Thisresearch has made use of the SIMBAD database (Wengeret al. 2000), operated at CDS, Strasbourg, France and, also,of NASA’s Astrophysics Data System. This research madeuse of Astropy, a community-developed core Python pack-age for Astronomy (Astropy Collaboration et al. 2013, ). This publication made use of TOP-CAT (Taylor 2005) for all table manipulation. This work isbased (in part) on observations made with the Spitzer SpaceTelescope, which is operated by the Jet Propulsion Labo-ratory, California Institute of Technology under a contractwith NASA. The plots in this publication were producedusing Matplotlib, a Python library for publication quality graphics (Hunter 2007) and R . This work was supported inpart by Michigan State University through computationalresources provided by the Institute for Cyber-Enabled Re-search. REFERENCES
Alexander T., 1995, MNRAS, 274, 909Alexander D. M., 2017, Astronomische Nachrichten, 338, 172Alonso-Herrero A., et al., 2006, ApJ, 640, 167Alonso-Herrero A., et al., 2016, MNRAS, 455, 563Assef R. J., et al., 2013, ApJ, 772, 26Baldassare V. F., Geha M., Greene J., 2018, preprint,( arXiv:1808.09578 )Barger A. J., et al., 2003, AJ, 126, 632Barth A. J., Ho L. C., Rutledge R. E., Sargent W. L. W., 2004,ApJ, 607, 90Bauer A., Baltay C., Coppi P., Ellman N., Jerke J., RabinowitzD., Scalzo R., 2009, ApJ, 696, 1241Bershady M. A., Trevese D., Kron R. G., 1998, ApJ, 496, 103Bertin E., Arnouts S., 1996, A&AS, 117, 393Bertin E., Arnouts S., 2010, SExtractor: Source Extractor, Astro-physics Source Code Library (ascl:1010.064)Brandt W. N., Alexander D. M., 2015, A&ARv, 23, 1Cardamone C. N., et al., 2011, VizieR Online Data Catalog, 218Chatterjee R., Roychowdhury A., Chandra S., Sinha A., 2018,ApJ, 859, L21Clavel J., et al., 2000, A&A, 357, 839Cristiani S., Vio R., Andreani P., 1990, AJ, 100, 56Cristiani S., Trentini S., La Franca F., Andreani P., 1997, A&A,321, 123Damen M., et al., 2011, ApJ, 727, 1De Cicco D., et al., 2015, A&A, 574, A112Ding N., et al., 2018, ApJ, 868, 88Dong X., et al., 2007, ApJ, 657, 700Donley J. L., Rieke G. H., P´erez-Gonz´alez P. G., Rigby J. R.,Alonso-Herrero A., 2007, ApJ, 660, 167Donley J. L., et al., 2012, ApJ, 748, 142Dressler A., 1989, in Osterbrock D. E., Miller J. S., eds, IAUSymposium Vol. 134, Active Galactic Nuclei. p. 217Falocco S., et al., 2015, A&A, 579, A115Falocco S., et al., 2016, The Universe of Digital Sky Surveys, 42,275Fazio G. G., et al., 2004, ApJS, 154, 10Ferrarese L., Ford H., 2005, Space Sci. Rev., 116, 523Ferrarese L., Merritt D., 2000, ApJ, 539, L9Ferreira Lopes C. E., Cross N. J. G., 2016, A&A, 586, A36Ferreira Lopes C. E., Cross N. J. G., 2017, A&A, 604, A121Filippenko A. V., Ho L. C., 2003, ApJ, 588, L13Finke J. D., Becker P. A., 2014, ApJ, 791, 21Ford H. C., et al., 1998, in Bely P. Y., Breckinridge J. B., eds,Proc. SPIEVol. 3356, Space Telescopes and Instruments V.pp 234–248, doi:10.1117/12.324464Fruchter A. S., Hook R. N., 2002, PASP, 114, 144Garc´ıa-Gonz´alez J., Alonso-Herrero A., P´erez-Gonz´alez P. G.,Hern´an-Caballero A., Sarajedini V. L., Villar V., 2014, inHighlights of Spanish Astrophysics VIII, Proc. of the XI Sci-entific Meeting of the Spanish Astronomical Society,Garmire G. P., Bautz M. W., Ford P. G., Nousek J. A., RickerJr. G. R., 2003, in Truemper J. E., Tananbaum H. D., eds,Proc. SPIEVol. 4851, X-Ray and Gamma-Ray Telescopes andInstruments for Astronomy.. pp 28–44, doi:10.1117/12.461599 R Core Team (2016). R: A language and environment for statis-tical computing. R Foundation for Statistical Computing, Vienna,Austria. URL .MNRAS000
Alexander T., 1995, MNRAS, 274, 909Alexander D. M., 2017, Astronomische Nachrichten, 338, 172Alonso-Herrero A., et al., 2006, ApJ, 640, 167Alonso-Herrero A., et al., 2016, MNRAS, 455, 563Assef R. J., et al., 2013, ApJ, 772, 26Baldassare V. F., Geha M., Greene J., 2018, preprint,( arXiv:1808.09578 )Barger A. J., et al., 2003, AJ, 126, 632Barth A. J., Ho L. C., Rutledge R. E., Sargent W. L. W., 2004,ApJ, 607, 90Bauer A., Baltay C., Coppi P., Ellman N., Jerke J., RabinowitzD., Scalzo R., 2009, ApJ, 696, 1241Bershady M. A., Trevese D., Kron R. G., 1998, ApJ, 496, 103Bertin E., Arnouts S., 1996, A&AS, 117, 393Bertin E., Arnouts S., 2010, SExtractor: Source Extractor, Astro-physics Source Code Library (ascl:1010.064)Brandt W. N., Alexander D. M., 2015, A&ARv, 23, 1Cardamone C. N., et al., 2011, VizieR Online Data Catalog, 218Chatterjee R., Roychowdhury A., Chandra S., Sinha A., 2018,ApJ, 859, L21Clavel J., et al., 2000, A&A, 357, 839Cristiani S., Vio R., Andreani P., 1990, AJ, 100, 56Cristiani S., Trentini S., La Franca F., Andreani P., 1997, A&A,321, 123Damen M., et al., 2011, ApJ, 727, 1De Cicco D., et al., 2015, A&A, 574, A112Ding N., et al., 2018, ApJ, 868, 88Dong X., et al., 2007, ApJ, 657, 700Donley J. L., Rieke G. H., P´erez-Gonz´alez P. G., Rigby J. R.,Alonso-Herrero A., 2007, ApJ, 660, 167Donley J. L., et al., 2012, ApJ, 748, 142Dressler A., 1989, in Osterbrock D. E., Miller J. S., eds, IAUSymposium Vol. 134, Active Galactic Nuclei. p. 217Falocco S., et al., 2015, A&A, 579, A115Falocco S., et al., 2016, The Universe of Digital Sky Surveys, 42,275Fazio G. G., et al., 2004, ApJS, 154, 10Ferrarese L., Ford H., 2005, Space Sci. Rev., 116, 523Ferrarese L., Merritt D., 2000, ApJ, 539, L9Ferreira Lopes C. E., Cross N. J. G., 2016, A&A, 586, A36Ferreira Lopes C. E., Cross N. J. G., 2017, A&A, 604, A121Filippenko A. V., Ho L. C., 2003, ApJ, 588, L13Finke J. D., Becker P. A., 2014, ApJ, 791, 21Ford H. C., et al., 1998, in Bely P. Y., Breckinridge J. B., eds,Proc. SPIEVol. 3356, Space Telescopes and Instruments V.pp 234–248, doi:10.1117/12.324464Fruchter A. S., Hook R. N., 2002, PASP, 114, 144Garc´ıa-Gonz´alez J., Alonso-Herrero A., P´erez-Gonz´alez P. G.,Hern´an-Caballero A., Sarajedini V. L., Villar V., 2014, inHighlights of Spanish Astrophysics VIII, Proc. of the XI Sci-entific Meeting of the Spanish Astronomical Society,Garmire G. P., Bautz M. W., Ford P. G., Nousek J. A., RickerJr. G. R., 2003, in Truemper J. E., Tananbaum H. D., eds,Proc. SPIEVol. 4851, X-Ray and Gamma-Ray Telescopes andInstruments for Astronomy.. pp 28–44, doi:10.1117/12.461599 R Core Team (2016). R: A language and environment for statis-tical computing. R Foundation for Statistical Computing, Vienna,Austria. URL .MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S Gavras P., et al., 2017, in IAU Symposium. pp 369–372( arXiv:1703.00258 ), doi:10.1017/S1743921316012746Gebhardt K., et al., 2000, ApJ, 539, L13Giavalisco M., et al., 2004, ApJ, 600, L93Giveon U., Maoz D., Kaspi S., Netzer H., Smith P. S., 1999, MN-RAS, 306, 637Gonz´alez-Mart´ın O., et al., 2015, A&A, 578, A74Graham A. W., Driver S. P., 2007, ApJ, 655, 77Graham M. J., Djorgovski S. G., Drake A. J., Mahabal A. A.,Chang M., Stern D., Donalek C., Glikman E., 2014, MNRAS,439, 703Greene J. E., Ho L. C., 2004, ApJ, 610, 722Greene J. E., Ho L. C., 2007, ApJ, 670, 92Greene J. E., Ho L. C., Barth A. J., 2008, ApJ, 688, 159Gu M. F., Li S.-L., 2013, A&A, 554, A51G¨ultekin K., et al., 2009, ApJ, 698, 198H¨aring N., Rix H.-W., 2004, ApJ, 604, L89Hawkins M. R. S., 1993, Nature, 366, 242Helfand D. J., Stone R. P. S., Willman B., White R. L., BeckerR. H., Price T., Gregg M. D., McMahon R. G., 2001, AJ, 121,1872Hook I. M., McMahon R. G., Boyle B. J., Irwin M. J., 1994,MNRAS, 268, 305Hornschemeier A. E., et al., 2003, AJ, 126, 575Hunter J. D., 2007, Computing In Science & Engineering, 9, 90Ivezi´c ˇZ., et al., 2008, preprint, ( arXiv:0805.2366 )Kasliwal V., Vogeley M., T. Richards G., 2015, Monthly Noticesof the Royal Astronomical Society, 451Kawaguchi T., Mineshige S., Umemura M., Turner E. L., 1998,ApJ, 504, 671Kelly B. C., Bechtold J., Siemiginowska A., 2009, ApJ, 698, 895Kelly B. C., Becker A. C., Sobolewska M., Siemiginowska A.,Uttley P., 2014, ApJ, 788, 33Kim J., Karouzos M., Im M., Choi C., Kim D., Jun H. D., LeeJ. H., Mezcua M., 2018, preprint, ( arXiv:1809.01350 )Komossa S., 2015, Journal of High Energy Astrophysics, 7, 148Kormendy J., Ho L. C., 2013, ARA&A, 51, 511Kormendy J., Kennicutt Jr. R. C., 2004, ARA&A, 42, 603Kormendy J., Richstone D., 1995, ARA&A, 33, 581Kozlowski S., 2016, Monthly Notices of the Royal AstronomicalSociety, 459, stw819Lacy M., et al., 2004, ApJS, 154, 166Lacy M., Petric A. O., Sajina A., Canalizo G., Storrie-LombardiL. J., Armus L., Fadda D., Marleau F. R., 2007, AJ, 133, 186Lasker B. M., et al., 2008, AJ, 136, 735Lawrence A., Papadakis I., 1993, ApJ, 414, L85Li S.-L., Cao X., 2008, MNRAS, 387, L41Luo B., et al., 2008, ApJS, 179, 19Luo B., et al., 2017, ApJS, 228, 2MacLeod C. L., et al., 2010, The Astrophysical Journal, 721, 1014MacLeod C. L., et al., 2012, ApJ, 753, 106Maccacaro T., Gioia I. M., Wolter A., Zamorani G., Stocke J. T.,1988, ApJ, 326, 680Magorrian J., et al., 1998, AJ, 115, 2285Marconi A., Hunt L. K., 2003, ApJ, 589, L21Mateos S., et al., 2012, MNRAS, 426, 3271Middei R., Vagnetti F., Bianchi S., La Franca F., Paolillo M.,Ursini F., 2017, A&A, 599, A82Momcheva I. G., et al., 2016, ApJS, 225, 27Nenkova M., Sirocky M. M., Ivezi´c ˇZ., Elitzur M., 2008a, ApJ,685, 147Nenkova M., Sirocky M. M., Nikutta R., Ivezi´c ˇZ., Elitzur M.,2008b, ApJ, 685, 160Ochsenbein F., Bauer P., Marcout J., 2000, A&AS, 143, 23Panessa F., Bassani L., Cappi M., Dadina M., Barcons X., CarreraF. J., Ho L. C., Iwasawa K., 2006, A&A, 455, 173Paolillo M., Schreier E. J., Giacconi R., Koekemoer A. M., GroginN. A., 2004, ApJ, 611, 93 Paolillo M., et al., 2017, MNRAS, 471, 4398Park J., Trippe S., 2017, ApJ, 834, 157Rees M. J., 1984, ARA&A, 22, 471Richards G. T., et al., 2002, AJ, 123, 2945Richards G. T., et al., 2005, MNRAS, 360, 839Riess A. G., et al., 2007, ApJ, 659, 98Rousseeuw P. J., Croux C., 1993, Journal of the American Sta-tistical Association, 88, 1273Rowan-Robinson M., et al., 2005, AJ, 129, 1183Rowan-Robinson M., Gonzalez-Solares E., Vaccari M., MarchettiL., 2013, Monthly Notices of the Royal Astronomical Society,428, 1958Sajina A., Lacy M., Scott D., 2005, ApJ, 621, 256Sarajedini V. L., Gilliland R. L., Kasm C., 2003, ApJ, 599, 173Sarajedini V. L., Koo D. C., Klesman A. J., Laird E. S., PerezGonzalez P. G., Mozena M., 2011, ApJ, 731, 97Schneider D. P., et al., 2007, AJ, 134, 102Schneider D. P., et al., 2010, AJ, 139, 2360Simm T., et al., 2015, A&A, 584, A106Simm T., Salvato M., Saglia R., Ponti G., Lanzuisi G., Trakhten-brot B., Nandra K., Bender R., 2016, A&A, 585, A129Skelton R. E., et al., 2014, ApJS, 214, 24Sokolovsky K., et al., 2017a, in European Physical Jour-nal Web of Conferences. p. 02005 ( arXiv:1703.02038 ),doi:10.1051/epjconf/201715202005Sokolovsky K. V., et al., 2017b, MNRAS, 464, 274Stern D., et al., 2005, ApJ, 631, 163Stern D., et al., 2012, ApJ, 753, 30Straatman C. M. S., et al., 2017, VizieR Online Data Catalog,183Strolger L.-G., et al., 2004, ApJ, 613, 200Taylor M. B., 2005, in Shopbell P., Britton M., Ebert R., eds,Astronomical Society of the Pacific Conference Series Vol. 347,Astronomical Data Analysis Software and Systems XIV. p. 29Taylor E. N., et al., 2009, ApJS, 183, 295Terlevich R., Tenorio-Tagle G., Franco J., Melnick J., 1992, MN-RAS, 255, 713Tremaine S., et al., 2002, ApJ, 574, 740Trevese D., Kron R. G., Majewski S. R., Bershady M. A., KooD. C., 1994, ApJ, 433, 494Trevese D., Boutsia K., Vagnetti F., Cappellaro E., Puccetti S.,2008, A&A, 488, 73Ulrich M.-H., Maraschi L., Urry C. M., 1997, ARA&A, 35, 445Vagnetti F., Middei R., Antonucci M., Paolillo M., Serafinelli R.,2016, A&A, 593, A55Valdes F. G., Campusano L. E., Velasquez J. D., Stetson P. B.,1995, PASP, 107, 1119Vanden Berk D. E., et al., 2004, ApJ, 601, 692Verma A., Charmandaris V., Klaas U., Lutz D., Haas M., 2005,Space Sci. Rev., 119, 355Villforth C., Koekemoer A. M., Grogin N. A., 2010, ApJ, 723,737Villforth C., Sarajedini V., Koekemoer A., 2012, MNRAS, 426,360Weedman D. W., et al., 2005, ApJ, 633, 706Wenger M., et al., 2000, A&AS, 143, 9Werner M. W., et al., 2004, ApJS, 154, 1Whitmore B. C., et al., 2016, AJ, 151, 134Wolf C., Hildebrandt H., Taylor E. N., Meisenheimer K., 2008,A&A, 492, 933Wu Y., Charmandaris V., Huang J., Spinoglio L., Tommasin S.,2009, ApJ, 701, 658Xue Y. Q., et al., 2011, ApJS, 195, 10Xue Y. Q., Luo B., Brandt W. N., Alexander D. M., Bauer F. E.,Lehmer B. D., Yang G., 2016, ApJS, 224, 15Young M., et al., 2012, ApJ, 748, 124Zuo W., Wu X.-B., Liu Y.-Q., Jiao C.-L., 2012, ApJ, 758, 104de Diego J. A., 2010, AJ, 139, 1269MNRAS , 1–20 (2019) E. Pouliasis et al. de Vries W. H., Becker R. H., White R. L., Loomis C., 2005, AJ,129, 615di Clemente A., Giallongo E., Natali G., Trevese D., Vagnetti F.,1996, ApJ, 463, 466van Dokkum P. G., 2001, PASP, 113, 1420
APPENDIX A: VARIABILITY DETECTIONSIMULATIONSA1 Variability detection in the presence of outliermeasurements
For all but the faintest optical sources, the accuracy oftheir brightness measurements is limited by the poorly con-strained systematic effects rather than the number of col-lected photons (“shot noise”) and uncertainties in the back-ground level estimations. This means we typically do nothave a reliable error bar attached to a photometric measure-ment. Of a particular concern in the context of HST pho-tometry are the residual cosmic rays that were not cleaned-out perfectly in the process of image stacking (“drizzling”;Fruchter & Hook 2002) that overlap with the measuredimage of the object. To circumvent the above issues, wemay assume that in a non-crowded field like GOODS-S (Sec. 1) objects of similar brightness will have similarmeasurement errors and the majority of objects are non-variable; employ a variability detection statistic that isrobust against individual outlier measurements (similar tothose caused by cosmic ray hits).We perform Monte-Carlo modeling to characterize theperformance of various variability-detection statistics inthe presence of photometric outliers. First, we model i = ... light curves each containing N points randomly dis-tributed in time. At each point in the model light curve weassigned a brightness value drawn from the Gaussian distri-bution characterized by the variance e . In addition, 1% ofthe points get a “cosmic ray hit” modeled by the additionalincrease in brightness by a value drawn from a uniform dis-tribution between 0 and e . We, then, compute the me-dian value, I non − var and the standard deviation σ ( I non − var ) scaled from the median absolute deviation of I non − var valuesfor each of the tested variability indices: σ ( I non − var ) = . × median (| I non − var i − median ( I non − var i )|) . After that, we add to each light curve an aperiodic variationcharacterized by a power-law power spectral density with aslope of − and amplitude e (equal to the noise level). Weuse these lightcurves to compute the median value of thevariability index: I var = median ( I var i ) and the typical Signal-to-Noise ratio , SNR, of variabilitydetection (among all the realizations of the noise and vari-ability patterns): SNR = ( I var − I non − var )/ σ ( I non − var ) . The resulting values of SNR as a function of N are pre-sented in Figure A1 for the three variability indices: thestandard deviation σ , the median absolute deviation (MAD)that characterize the scatter of measurements in a light curveand the / η that quantifies the smoothness of a lightcurve. -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 20 40 60 80 100 120 140 160 180 200 S N R NAperiodic signal = 1 ✕ Noise, no outlieresMAD1/ ησ S N R NAperiodic signal = 1 ✕ Noise, 1% of measurements are outlieresMAD1/ ησ Figure A1.
The simulated median signal to noise ratio of vari-ability detection as a function of the number of light curve pointswith no outlier measurements (top panel) and in the presence ofoutliers (bottom panel).
A detailed discussion of these variability indicators can befound in Sokolovsky et al. (2017b).Figure A1 highlights that in the absence of outlier mea-surements (i.e. non-periodic variability is being detected overa pure Gaussian noise) σ and / η typically provide a higherSNR detection for a given number of light curve points thanMAD. If outlier measurements are present in light curves,they dramatically affect the efficiency of σ as a variabilityindicator rendering its useless as soon as each light curvehas so many points that it is likely to contain at least oneoutlier (recall, that in our model the variability amplitudeis lower than the amplitude of outliers). The ability of / η to identify smooth variability is also reduced considerablyby outliers, while MAD maintains the SNR that is steadilyincreasing with N.The simulations described above confirm that MADmay serve as a variability indicator resistant to individualoutlier measurements. It is also apparent that σ is on av-erage a more sensitive variability indicator than MAD aslong as N is sufficiently low that each individual lightcurveis unlikely to contain even one outlier. However, if outliersare present in the data set, the light curves that containoutliers will predominantly be selected with σ as candidatevariables. The use of MAD is still preferred to select a clean MNRAS000
A detailed discussion of these variability indicators can befound in Sokolovsky et al. (2017b).Figure A1 highlights that in the absence of outlier mea-surements (i.e. non-periodic variability is being detected overa pure Gaussian noise) σ and / η typically provide a higherSNR detection for a given number of light curve points thanMAD. If outlier measurements are present in light curves,they dramatically affect the efficiency of σ as a variabilityindicator rendering its useless as soon as each light curvehas so many points that it is likely to contain at least oneoutlier (recall, that in our model the variability amplitudeis lower than the amplitude of outliers). The ability of / η to identify smooth variability is also reduced considerablyby outliers, while MAD maintains the SNR that is steadilyincreasing with N.The simulations described above confirm that MADmay serve as a variability indicator resistant to individualoutlier measurements. It is also apparent that σ is on av-erage a more sensitive variability indicator than MAD aslong as N is sufficiently low that each individual lightcurveis unlikely to contain even one outlier. However, if outliersare present in the data set, the light curves that containoutliers will predominantly be selected with σ as candidatevariables. The use of MAD is still preferred to select a clean MNRAS000 , 1–20 (2019) ptically variable AGNs in GOODS-S
20 21 22 23 24 25 26 . . . . . . .
Figure A2.
MAD as a function of magnitude of all sources inour initial sample. The blue solid line represents the adopted vari-ability threshold of 3.5 σ , while the red dashed lines represent thethresholds for the median under-sampled sets for different numberof data points in the light curves, N. sample of variable objects, even at the cost of a slightly lowerdetection efficiency compared to σ . A2 Variability threshold dependence on thenumber of data points in light curves
In order to test if the 3.5 σ threshold will be at a differentlevel if we consider only light curves with a specific num-ber of data points, we used simulations and checked if thethresholds derived from under-sampled data is lower thanthe adopted threshold in this work. If this is the case, thenthe sample of variable sources derived from applying thesame threshold to all light curves (irrespective of the num-ber of points in them) will not be affected.We simulated ten sets of under-sampled data with thesets differing in the number of data points in the light curves(N=5, 6, 7, 8, 9, 10, 15, 20, 30 & 40). The number of sourcesof the under-sampled data are 21022, 19346, 17428, 15891,14207, 13128, 6790, 3677, 1769 & 1350, respectively. Foreach set, we performed 1000 iterations and at each time,we randomly selected N data points for all the sources ofour initial sample. We then calculated the MAD values andfollowed the procedure described in Section 3.1 to find the3.5 σ thresholds. We took the median values of the thresh-olds with N data points in the light curve and we comparedthe results with different N. In Figure A2, we plot the MADas a function of magnitude for all the sources of our sam-ple and the 3.5 σ threshold adopted in this work (solid blueline). We over-plot the median values of 3.5 σ thresholds for Table B1.
Catalogue of confirmed SNe identified in our survey.ID RA Dec N p T bas < F850LP > MAD *(J2000) (J2000) (years) (mag) ( σ )7343 53.07570 -27.73630 8 0.14 24.10 11.19581 53.10564 -27.75084 17 2.51 22.71 4.7214446 53.15638 -27.77966 6 0.18 24.97 4.20 different number of data points in the light curve, N (dashedblack lines).We find no extreme differences between the thresholdsderived from different N. From the bright end of the magni-tude distribution up to ∼
24 mag, the thresholds follow thesame trend with small scatter with each other. Above ∼ APPENDIX B: OBSERVATIONALPROPERTIES AND LIGHT CURVES OFSUPERNOVAE
Table B1 lists the observational properties, while Figure B1presents our photometry of the identified SNe. All threeSNe have been reported previously in the literature (Strolgeret al. 2004; Riess et al. 2007).
This paper has been typeset from a TEX/L A TEX file prepared bythe author.MNRAS , 1–20 (2019) E. Pouliasis et al.
10 20 30 40 50 +5.335e423.524.024.525.025.526.0 < F L P > ( m a g )
10 20 30 40 50 60 70
MJD (days) +5.253e425.025.526.0
Figure B1.
Light curves of the confirmed SNe identified in thisstudy. The dashed line indicates the median magnitude and thenumber on the top of each plot indicates the identifier of thesource. MNRAS000