Classifying blazar candidates from the 3FGL unassociated catalog into BL Lacs and FSRQs using Swift and WISE data
DDraft version December 15, 2020
Typeset using L A TEX twocolumn style in AASTeX63
Classifying blazar candidates from the 3FGL unassociated catalog into BL Lacs and FSRQs using
Swift and WISE data
Amanpreet Kaur , Abraham D. Falcone , and Michael C. Stroh Department of Astronomy and AstrophysicsPennsylvania State University University Park, PA 16802, USA Center for Interdisciplinary Exploration and Research in Astrophysics (CIERA), Northwestern University, Evanston, IL 60201, USA
Submitted to AJABSTRACTWe utilize machine learning methods to distinguish BL Lacertae objects (BL Lac) from Flat SpectrumRadio Quasars (FSRQ) within a sample of likely X-ray blazar counterparts to Fermi 3FGL unassociatedgamma-ray sources. From our previous work, we have extracted 84 sources that were classified as ≥
99% likley to be blazars. We then utilize Swift − XRT, Fermi, and WISE (The Wide-field InfraredSurvey Explorer) data together to distinguish the specific type of blazar, FSRQs or BL Lacs. VariousX-ray and Gamma-ray parameters can be used to differentiate between these subclasses. These arealso known to occupy different parameter space on the WISE color-color diagram. Using all these datatogether would provide more robust results for the classified sources. We utilized a Random ForestClassifier to calculate the probability for each blazar to be associated with a BL Lac or an FSRQ.Based on P bll , which is the probability for each source to be a BL Lac, we placed our sources into fivedifferent categories based on this value as follows; P bll ≥ bll ≥ bll ≤ bll ≤ < P bll < Keywords: catalogs — surveys INTRODUCTIONBlazars are a subclass of the Active Galactic Nucleiwhich have their jets pointing along our line of sight(Blandford & Rees 1978). They are further divided intotwo categories; Flat Spectrum Radio Quasars (FSRQs)and BL Lacertae Objects (BL Lacs) based on their op-tical spectra. The FSRQs display broad emission lines,whereas the BL Lacs display no lines or narrow lines(equivalent width < a r X i v : . [ a s t r o - ph . H E ] D ec Kaur et al.
Gamma Ray Observatory has revealed more than 5000sources since its launch in 2008. Blazars constitute thebulk of the overall known extragalactic gamma-ray pop-ulation ( > ∼
8% ofthe total gamma-ray population). However, each cata-log represents approximately one-third of its sources asunassociated or unknown. Finding associations to thesesources or classifying these is a multi year task which of-ten requires multiwavelength observations for confirma-tions. In the past few years, various studies have beenconducted on these gamma-ray sources where machinelearning methods were employed to classify the unasso-ciated sources; e.g., (Ackermann et al. 2012; Saz Parkin-son et al. 2016; Marchesini et al. 2020a). In particular,Marchesini et al. (2020b,a) first explored the connec-tion between gamma-rays and X-rays for blazars andlater utilized X-ray data from
Swift in conjunction withthe
Fermi gamma-ray data to find BL Lacs among theunassociated sources.. We (Falcone A. D. & et al. Stroh2015; Kaur et al. 2019) have conducted an X-ray surveytargeting the Fermi unassociated source fields and foundvarious possible X-ray associations . Since the majorityof the known gamma-ray sources are blazars and pul-sars, it is highly likely that a rather large population ofthe unassociated sources could belong to these two pop-ulations. It should be noted that the sensitivity for Swift -XRT is ∼ × e − erg/cm /s for a 4 ksec exposure,which was the average exposure for this survey. Most ofthe known Fermi blazars, as well as some of the knownFermi pulsars, are detectable with high signal-to-noiseratio within this exposure time, as shown in Fig. 1-4in Kaur et al. (2019). These authors utilized machinelearning methods on these X-ray counterparts to findblazars and pulsars. Their results yielded 134 blazarsand 8 pulsars with high probabilities based on the ma-chine learning methods.In this work, the objective is to classify the highly prob-able blazar candidates revealed from our previous workinto subclassifications of BL Lacs and FSRQs using themethods of Machine Learning. The paper is dividedas follows. Section 2 describes the process of the finalsample selection and Section 3 explains the overall anal-ysis method. The results and the conclusions of the thisstudy are published in Sections 4 and 5, respectively. SAMPLE SELECTION The
Swift -Xray Telescope (XRT) (Burrows et al.2005) conducted observations for 803 3FGL unassoci-ated sources in order to search for their X-ray counter-parts. These were chosen at random from the completesample of ∼ Swift fieldof view. All these X-ray observations were completedthrough a
Swift
Swift -XRT sensitivity for detect-ing an X-ray source during the randomly distributed
Swift -XRT exposures on the 3FGL fields (the 3FGLfollow-up fields were distributed across the whole skyand randomly chosen for
Swift -XRT follow-up as partof a fill-in observation program) coupled with the LATerror ellipse for the sources in our sample. We foundthis sensitivity by using ’empty’
Swift -XRT fields dis-tributed across the sky, using GRB fields with the GRBmasked out, and calculating the spurious source detec-tion density in these fields as a function of exposure time,using the same source detection criteria that is utilizedfor finding possible X-ray counterparts to Fermi unasso-ciated sources. The majority of the
Swift -XRT follow-up observations of the
Fermi unassociated sources inour sample were of roughly 4 ksec exposure time. To beincluded in our sample, a potential X-ray counterparthad to be detected at the > σ signal-to-noise threshold.For a 4 ksec Swift -XRT exposure, less than ∼ σ threshold thatwe used for X-ray source selection and when using atypical Fermi -LAT 95% confidence ellipse for an unas-sociated gamma-ray source. However, a subset of ourobservations received significantly longer exposures andsome of the Fermi-LAT error ellipses are larger than typ-ical, thus increasing the chance coincidence probabilityin those cases. By using the actual
Swift -XRT exposureand the Fermi-LAT error ellipse of each follow-up field,we found that the median chance probability of a spuri-ous 4 σ X-ray source detection was less than 0.01; asidefrom a few outliers, there was generally a low chancethat the given X-ray source counterpart candidate is notactually associated with the gamma-ray source. These aur et al. 2020 ≥
4. Kaur et al. (2019, hereafter, K19)performed a machine learning analysis (random forest)to find pulsar and blazar candidates among these 217X-ray counterparts to the 3FGL unassociated sources.According to the random forest classifier method usedin K19, 134 sources from the sample resulted in classifi-cations as highly likely blazars, i.e. the sources for whichthe probabilities to be associated with the blazar classwas ≥ Fermi unassociated sources were observed with
Swift − XRT,and 25 of these were found to have exactly one sourcewithin the
Fermi uncertainty circle with SNR ≥
4. Weapplied the K19 criteria and ML methods to these 25X-ray sources in order to form a more complete list ofblazar candidates for this work. P bzr was defined asthe probability for a source to be a blazar, which wereyielded by the RF classifier. As was done in K19, thesources were classified into one of the following cate-gories: pulsar (P bzr ≤ bzr ≤ bzr ≥ bzr ≥ ≤ P bzr ≥ Swift -XRT positions ofthese sources were utilized to search for any positionalassociations in the AllWISE catalog (Cutri & al. 2013)within 5 arcsec positional uncertainty. An average un-certainty associated with an XRT position for these datais less than ∼ Swift -XRT blazar catalog which comprises of 2831 blazars us-ing 15 years of data (Giommi et al. 2019). We searchedfor WISE sources corresponding to these blazar posi-tions within the 5” uncertainity region which is the av-erage
Swift -XRT positional uncertainty. Aside from the WISE counterparts to these blazars, we found thata secondary source within 5” was found for 22 positionswhich suggests that there is a ∼ ANALYSISMassaro et al. (2012) introduced a method to findblazars among other sources using WISE colors; W1,W2, W3 and W4 corresponding to 3.4, 4.6, 12 and 22 µ m, respectively. These authors showed that blazars oc-cupy a particular region on a color-color plot (W1-W2 vsW2-W3) in the IR regime (WISE in this case) which sep-arates them from other source types. They termed thisregion as the WISE Blazar Strip (WBS). Moreover, thetwo classes of blazars; FSRQs and BL Lacs occupy differ-ent regions within this strip. Therefore, these color in-dices could be utilized to separate one blazar class fromanother, which is the immediate objective of this work.See Fig. 1,2 in Massaro et al. (2012) for details. Fig. 2shows the blazar strip for Fermi blazars along with the84 unassociated sources. It is quite apparent from thisfigure that these are highly likely blazars (also predictedby our ML methods in K19), as they follow the patternof the known blazars. In addition, the fact that the BLLacs and FSRQs occupy different parameter space onthis color-color plot is clearly shown. In our previouswork, we used both X-ray and gamma-ray properties ofthe
Fermi unassociated sources to distinguish pulsarsfrom blazars using the random forest classifier machinelearning algorithm (Breiman 2001). Here we employ thesame method, with additional WISE parameters, to thesubsample of highly likely blazar candidates which are84 in number, as described in Section 2. While some dis-tinction between the two classes of blazars can be seenwhen the five considered X-ray and gamma-ray prop-erties are compared (e.g. X-ray flux, gamma-ray flux,gamma-ray variability index, gamma-ray spectral index,and curvature), the addition of WISE color parameters
Kaur et al. is expected to enhance this distinction. In this work, weutilize these five X-ray and gamma-ray parameters alongwith two WISE color indices; W1-W2 and W2-W3. Wecompare them simultaneously using the random forestclassifier. In order to proceed with this algorithm, asample that includes both known classes of blazars wasrequired, with each of the blazars in this sample hav-ing known values for all of these above mentioned sevenparameters. 3.1.
Training and Test data
A total of 501 known blazars were extracted fromAckermann et al. (3LAC, 2015) for which gamma-ray,X-ray and WISE data were available. The gamma-rayand X-ray properties of these known blazars were ob-tained from Acero et al. (3FGL, 2015) and Ackermannet al. (3LAC, 2015), respectively. Fig 4 displays thecomparison of two subclasses of blazars along with theunassociated sources. It should be noted that the X-rayfluxes for blazars in 3LAC catalog were extracted fromthe RASS survey (Voges et al. 1999, 2000). These fluxvalues are provided in the energy range 0.1-2.4 keV. Forthe 84 unassociated sources, the X-ray fluxes were de-rived using
Swift -XRT in the energy range from 0.1to 2.4 keV for the consistency. The WISE color in-dices were obtained from the AllWISE catalog. Outof these 501 blazars, 162 were FSRQs and 339 were BLLacs. The unbalanced proportion of these two classescould lead to biased results towards one particular class,therefore this was corrected by using a class balancingalgorithm, SMOTE (Chawla et al. 2002). SMOTE usesthe k nearest neighbors method to synthetically gener-ate sources for the underrepresented class to match itwith the number of sources in the other class. Herewe employed the SMOTE algorithm provided in the scikit-learn library of python which utilizes 5 near-est neighbors to create one synthetic data point. In thiscase, since FSRQs were 162 by number as compared to339 BL Lacs. SMOTE algorithm added 177 syntheticdata points mimicking the properties of the original FS-RQs. This led to our final sample of 339 BL Lacs and 339FSRQs. An example displaying the results of SMOTEanalysis are shown in Fig. 1. Here these results areshown for gamma-ray flux vs the spectral index, butthese FSRQs mimic the real FSRQs for all the parame-ters which would be used to train the classifier.3.1.1.
Random Forest Parameter Selection and AccuracyCalculation Method
We employed the random forest classifier from sklearn using python 3.6 , which is a supervisedmethod of machine learning based on the method of de-cision trees (Breiman 2001). A complete description of this method and the details of its implementation aredescribed in section 3.1 in K19. A parameter tuning al-gorithm
GridSearchCV in sklearn was employed to findthe optimum parameters for the random forest classifier.Based on this optimization, 1000 decision trees with amaximum depth of 10 splits(nodes) in each tree were em-ployed to obtain the final classification for each sourcein this method.Generally in a machine learning algorithm, a majorityof the complete sample is reserved for the training setand a smaller subset is utilized as the test sample tocheck the accuracy of the underlying classifier. How-ever, the accuracy obtained from this method is clearlybiased since it is based on one given test sample. There-fore, in this work, we employed a 10-fold cross-validationmethod using sklearn which divided the original sam-ple into 10 equal size subsamples such that one out ofthese 10 samples was chosen as a test sample (one at atime) and the rest combined were considered a trainingsample. The trained classifier was then applied to thegiven test subsample to obtain the accuracy value. Thisprocedure was repeated 10 times to obtain accuraciesfrom each test sample. The overall accuracy was cal-culated as an average of accuracies obtained from theseiterations. This accuracy calculation method has theadvantage that it iterates through the complete samplewhich results in less sample bias, relative to calculat-ing accuracy based on a single test sample. Based onthe procedure explained above, the random forest clas-sifier was trained and then cross validated which yieldedan average accuracy of 93.5%. For an additional cross-check, we also conducted an experiment where only asingle test sample example was chosen to check the ac-curacy of this classifier. Based on a random one testsample of 106 known BL Lacs and 107 known FSRQS,our classifier returns 98 true FSRQs and 102 true BLLacs. In other words, this wrongly classified 4 BL Lacsand 9 FSRQs. This yielded an accuracy of ∼ RESULTS aur et al. 2020 G a mm a - r a y Sp e c t r a l I n d e x after SMOTEoriginal FSRQs Figure 1.
An example two dimensional plot displaying thedistribution of the original FSRQs (blue dots) and artifi-cially created FSRQs (red pluses) using SMOTE. The x-axis represents the logarithm of gamma-ray flux (0.1-100)(GeV)[erg/cm /s] and the y-axis displays the gamma-rayspectral index. These parameters were obtained from the3FGL catalog. The trained classifier using the X-ray, gamma-ray andWISE parameters was then applied to the sample of 84blazar candidate counterpart sources. Since the RF clas-sifier provides a probability value for each source to be-long to a particular class, we define the following classesbased on the predicted probabilities: BL Lac (bll): P bll ≥ bll ≥ bll ≤ bll ≤ ≥ P bll ≤ bll is the probability for a source to be a BLLac. Using these definitions, we found that among thesample of 84 highly likely blazar candidates, 50 are likelyBL Lacs and 34 are ambiguous. None of the sources werepredicted to be FSRQs, nor did any of them fall into thecategory of likely FSRQs (See Table. 2). This is consis-tent with a visual inspection of Fig. 2, which shows thatour newly identified blazar candidate/counterpart sam-ple seems to be constituted primarily of BL Lac classblazars, with a relatively small number of outliers thatare ambiguously consistent with either the BL Lac orFSRQ classification and another small group of outliersthat are ambiguous in the sense that they seem to falloutside of both the Bl Lac and the FSRQ distribution.The respective significances (percentage importances) ofeach parameter employed in the classifier are as follows:X-ray flux: 0.058, Curvature: 0.062, Spectral Index:0.241, Variability Index: 0.094, Gamma ray flux: 0.055,W1-W2: 0.312 and W2-W3: 0.175.4.1. Miscellaneous-Outliers
A few of these unassociated source candidates seem todiverge from the usual WBS, as displayed in Fig. 3 by us- ing black circular regions around them. These are sevenin number out of which one belongs to the likely BLLac and the rest to the ambiguous category. After fur-ther inspection, it was found that the positions of threeof these were coincident with stars within 5 arcsec posi-tional uncertainties of Swift − XRT positions. TYC4199-1248, which was also reported in K19, corresponds to apositional coincidence with 3FGL J1729.0+6049. Simi-larly, 3FGL J0748.8-2208 is spatially coincident with astar, TYX 5993-3722-1 and 3FGL JJ1801.5-7825 withHD 162298. No further information about these starswas found in literature. It is clear that the WISEposition match yielded the colors for these non-blazarsources, which would explain their placement far leftfrom the WBS in Fig. 3. This is consistent with the factthat our ML method classified them as ”ambiguous.” Inthese few cases, it is also possible that the stellar sys-tems could be associated with the source of gamma raysfrom the corresponding Fermi source. However, one ofthese outliers, namely 3FGL J1958.1+2436, is a con-firmed BL Lac despite its position on the WISE color-color plot. Another interesting case is that of 3FGLJ2035.8+4902, for which the position of the only X-raysource in the 3FGL error circle is positionally coincidentwith an eclipsing binary, V* V2552 Cyg. Some of theseoutliers may not belong to the blazar population and/ormay not be the actual X-ray counterpart to the Fermiunassociated source, and various direct methods such asoptical spectroscopy could be used to verify their truenature. However, in some of the outlier cases, it is alsopossible that the detected X-ray source is a counterpartblazar which deviates from the usual gamma-ray blazarpopulation and therefore should be further investigatedfor its interesting behavior. It should be noted that al-though the position of the few outliers on the WISEcolor-color plot does not mimic the other gamma-rayblazars, these parameters lie within the limits of themore general blazar population, particularly as an ex-tension of the BL Lac distribution; please see Fig. 1 inMassaro et al. (2011). Since the latest Fermi catalogData Release 2 (Lott et al. 2020, 4FGL-DR2) was pub-lished recently, we compared our list of 84 sources tothe source classifications in this release, for the caseswhere they are available. We found that 52 of the 84sources are identified as bcus, and 7 are identified as BLLacs. All except one (3FGL J0427.9-6704 classified asa ”Binary” in the 4FGL catalog) of our results matchwith the 4FGL predictions as seen in Table 2, althoughthree identified BL Lacs in the 4FGL catalog were char-acterized as ”ambiguous” by our RF classifier (note: ourclassifier did find that these 3 sources were >
81% likelyto be BL Lacs). In addition, we compared our classifi-
Kaur et al. cation results with a similar study conducted by March-esini et al. (2020a), in which the authors searched for BLLac candidates among the Fermi Unassociated sources.These authors selected their sample using a slightly dif-ferent criteria such as SNR > >
4. Moreover, these authors selected X-ray coun-terparts where more than one X-ray source was foundwithin the
Fermi uncertainty region. Regardless, theseauthors found 19 sources in which are highly likely BLLacs. Among these 19, we found 7 which are also presentin our sample. 4 out of these 7 are identified as BL Lacsin our classification, whereas 3 are identified as ambigu-ous. Furthermore, it should be noted that two of theseambiguous sources have blazar probability >
89% in ourclassification, which makes them consistent with the BLLac classification. Therefore, we consider our results tobe consistent with this independent study conducted bythese authors on this subset of the sources in our study. CONCLUSIONSThe immediate objective of this work is to classify, aseither FSRQ or BL Lac, a sample of 84 highly likely X-ray blazar candidates that were drawn from a list of X-ray counterparts to 3FGL unassociated sources. This isa step forward towards the completeness of finding thegamma-ray emitting blazars. Finding associations tothe unassociated gamma-ray emitting sources has beena necessary step in order to understand the gamma-ray emitting population in the Universe. Most of thegamma-ray sources belong to the category of blazars.Classifying these blazars is an important step towardsunderstanding their evolution and their role in galaxyevolution. Understanding the distribution of the sub-classes for these gamma-ray emitting blazars plays a vi-tal role in putting constraints on the blazar sequence(Fossati et al. 1998; Ghisellini et al. 2017). In previouswork, we contributed to this task by finding counterpartsand classifying the blazars among these unassociatedsources, while this paper focuses on sub-classificationof these blazar sources. Using the methods of machinelearning, we find that 50 out of these 84 sources are ≥
90% likely to be BL Lacs, and the other 34 are not ableto categorized (i.e. we categorize them as ’ambiguous’).However, since all these outliers/ambiguous sources area subset of our blazar sample, these could be consid-ered ”bcus” (blazars of uncertain type). This impliesthat these sources are mostly likely either peculiar BLLacs, FSRQs, or transitional blazars. Various follow upmultiwavelength campaigns would be required to dis-cern their nature. We don’t find any of these sources tobe clearly labeled as FSRQs. There could be multiplereasons for the paucity of sources categorized as FS- RQs; e.g. (i) most of the X-ray counterparts to Fermiunassociated sources are indeed BL Lacs, which couldbe caused by inherent selection biases such as the factthat BL Lacs are more likely to have a synchrotron peakin the UV to X-ray band, (ii) The blazar componentof Fermi unassociated sources has a selection bias thatmakes it more likely for an unassociated source to be aBL Lac, or, (iii) Since FSRQs are more bright and oftenhave spectra available via various surveys, it is highlylikely that most of the unassociated blazar populationin the Fermi catalog are indeed BL Lacs. This patternhas also been seen in various optical spectroscopic sur-veys of unassociated Fermi sources, e.g., ´Alvarez Crespoet al. (2016); Crespo et al. (2016); Pe˜na-Herazo et al.(2017); Paiano et al. (2017) In the future, optical spec-troscopic techniques can be utilized to find the natureof the 24 ambiguous sources, and to further investigatethe properties of the classified blazars. Our study pro-vides likely blazar targets for these spectroscopic opti-cal observations, providing another avenue for localizingand characterizing possible counterparts with high pre-cision.The ongoing questions regarding the Fermi blazarsequence and Fermi blazar divide require the redshift es-timates for blazars. One should be able to confirm anddetermine the redshifts for the FSRQs found within the24 sources by using various 4m class optical facilities,e.g. ´Alvarez Crespo et al. (2016); Crespo et al. (2016).For the case of BL Lacs, traditionally 8-10m class tele-scopes (Shaw et al. 2009, 2013; Paiano et al. 2019) areutilized. This method of estimating the redshifts for BLLacs is highly effective, but it is time and cost consum-ing. Recently, Rau et al. (2012) devised a photometricmethod to determine the redshifts (z), or find an up-per limit, for BL Lacs. One caveat of this method isthat it works for sources with z > aur et al. 2020 Kaur et al. ¡ W3 ¡ : : : : W ¡ W FSRQ(WISE)BLL(WISE)Unassoc
Figure 2.
WISE blazar strip for the known
Fermi
BL Lacs (blue) and FSRQs (red). Overplotted are the unassociated blazarcandidates from this work (green). The W1, W2 and W3 correspond to the WISE filters; 3.4, 4.6 and 12 µ m, respectively.Please see Section 3 for complete details. ¡ : : : : : : : : : : : ¡ W3 ¡ : : : : W ¡ W FSRQ(WISE)BLL(WISE)OutliersLikelyBLL(unassoc)Ambiguous(unassoc)
Figure 3.
WISE blazar strip for the known
Fermi
BL Lacs (blue) and FSRQs (red). Overplotted are the unassociated blazarcandidates from this work as displayed in Fig. 2. The subcategories displayed are based on the probabilities obtained with ourmachine learning algorithm dividing these into likely BL Lacs (magenta, P bll ≥ ≥ P bll ≤ black circles. See the discussion in Section 4.1 for a complete details onthese sources. The W1, W2 and W3 correspond to the WISE filters; 3.4, 4.6 and 12 µ m, respectively. aur et al. 2020 F X S i gn i f _ C u r v e S pe c t r a l _ I nde x V a r i ab ili t y _ I nde x F G w - w
12 10 FX w - w Signif_Curve
Spectral_Index
Variability_Index
12 10 FG w1-w2 w2-w3 bll fsrq unassoc Figure 4.
A comparison pairplot of seven parameters used to distinguish BL Lacs (bll; blue ) from FSRQs (fsrq; orange ). The84 unassociated sources (unassoc; green ) are also plotted. The plotted parameters are defined as follows:
FX, Signif Curve,Spectral Index, Variability Index, FG, w1-w2, w2-w3 represent log (X-ray Flux), Gamma-ray Curvature, Gamma-raySpectral Index, log (Gamma-ray Variability Index), log (Gamma-ray Flux), WISE Color Index (W1-W2) and WISE ColorIndex (W2-W3), respectively. Kaur et al.
Table 1 . Additional sample of blazar and pulsar candidates from the3FGL unassociated sources since Kaur et al. (2019)
Swift Name a b Class c Random Forest d SwF3 3FGL BLL Probability J . J . . J . J . c ambiguous 0 . J . − J . − c blazar 0 . J . − J . − . J . − J . − . J . − J . − . J . − J . − . J . − J . − . J . − J . − . J . J . . J . − J . − . J . − J . − c ambiguous 0 . J . − J . − . J . − J . − . J . J . . J . − J . − . J . − J . − . J . J . . J . − J . − c ambiguous 0 . J . − J . − . J . J . . J . − J . − . J . J . . J . − J . − . J . − J . − . a The name of the
Swift discovered X-ray source within the 95%
Fermi uncertaintyregion of the corresponding 3FGL source as defined in K19. b The Fermi source name as defined in 3FGL catalog. c The Classification based on the probability of a given source to be called ablazar/pulsar/ambiguous as defined in K19 and also described in Section 2 d The probability of a given source to be identified a BL Lac, derived from the randomforest classifier.
Table 2 . Classification using Machine Learning : BLLs and FSRQs
Swift Name a b Expected c Class d Random Forest e NotesSwF3 3FGL chance X-ray sources BLL Probability Classification in literature J . J . .
955 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
957 bcu(4FGL, Lott et al. 2020) J . J . .
93 bcu(4FGL, Lott et al. 2020) J . J . . J . − J . − .
512 bcu(4FGL, Lott et al. 2020) J . J . .
954 bcu(4FGL, Lott et al. 2020)
Table 2 continued aur et al. 2020 Table 2 (continued)
Swift Name a b Expected c Class d Random Forest e NotesSwF3 3FGL chance X-ray sources BLL Probability Classification in literature J . − J . − .
912 bcu(4FGL, Lott et al. 2020) J . − J . − .
958 bcu(4FGL, Lott et al. 2020) J . − J . − .
805 LMXB (4FGL, Lott et al. 2020) J . − J . − .
939 bcu(4FGL, Lott et al. 2020) J . − J . − .
934 bcu(4FGL, Lott et al. 2020) J . J . .
958 bcu(4FGL, Lott et al. 2020) J . J . . J . − J . − .
956 bcu(4FGL, Lott et al. 2020) J . − J . − .
898 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
894 bcu(4FGL, Lott et al. 2020) J . − J . − .
949 bcu(4FGL, Lott et al. 2020) J . − J . − .
871 bll (4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
888 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
954 bcu(4FGL, Lott et al. 2020) J . J . .
953 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
957 bcu(4FGL, Lott et al. 2020) J . J . .
882 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
899 bcu(4FGL, Lott et al. 2020) J . − J . − .
937 bcu(4FGL, Lott et al. 2020) J . − J . − .
906 bcu(4FGL, Lott et al. 2020) J . − J . − .
904 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − . J . J . .
847 bll(4FGL, Lott et al. 2020) J . − J . − .
955 bcu(4FGL, Lott et al. 2020) J . − J . − .
955 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
955 bcu(4FGL, Lott et al. 2020) J . − J . − .
946 bcu(4FGL, Lott et al. 2020) J . − J . − .
755 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
952 bcu(4FGL, Lott et al. 2020) J . − J . − .
955 bcu(4FGL, Lott et al. 2020) J . J . . J . J . .
501 bcu(4FGL, Lott et al. 2020) J . − J . − .
946 bcu(4FGL, Lott et al. 2020) J . − J . − .
833 bcu(4FGL, Lott et al. 2020) J . J . .
955 bll(4FGL, Lott et al. 2020) J . − J . − .
903 bcu(4FGL, Lott et al. 2020) J . J . .
837 bll(4FGL, Lott et al. 2020) J . − J . − . J . − J . − . J . J . . J . − J . − . J . − J . − .
901 bcu(4FGL, Lott et al. 2020) J . J . .
897 bcu(4FGL, Lott et al. 2020) J . J . . J . J . .
941 bcu(4FGL, Lott et al. 2020)
Table 2 continued Kaur et al.
Table 2 (continued)
Swift Name a b Expected c Class d Random Forest e NotesSwF3 3FGL chance X-ray sources BLL Probability Classification in literature J . − J . − .
956 bll(4FGL, Lott et al. 2020) J . − J . − .
951 bcu(4FGL, Lott et al. 2020) J . − J . − .
946 bcu(4FGL, Lott et al. 2020) J . (cid:63) J . .
877 bcu(4FGL, Lott et al. 2020) J . J . . J . J . . J . J . .
931 bll(4FGL, Lott et al. 2020) J . J . .
917 bcu(4FGL, Lott et al. 2020) J . J . .
936 bcu(4FGL, Lott et al. 2020) J . J . .
949 bll(4FGL, Lott et al. 2020) J . J . .
912 bcu(4FGL, Lott et al. 2020) J . J . .
85 bcu(4FGL, Lott et al. 2020) J . J . .
702 bcu(4FGL, Lott et al. 2020) J . J . .
949 bll(4FGL, Lott et al. 2020) J . J . . J . J . . J . J . .
887 bcu(4FGL, Lott et al. 2020) J . − J . − .
925 bcu(4FGL, Lott et al. 2020) J . − J . − . J . J . .
931 bcu(4FGL, Lott et al. 2020) J . − J . − . J . − J . − .
959 bll(4FGL, Lott et al. 2020) J . J . .
954 bll(4FGL, Lott et al. 2020) J . − J . − .
936 bcu(4FGL, Lott et al. 2020) a The name of the
Swift discovered X-ray source within the 95%
Fermi uncertainty region of the corresponding 3FGL source as defined inK19. b The Fermi source name as defined in 3FGL catalog. c The expected number of X-ray sources found within the Fermi uncertainty ellipse using
Swift -XRT. See section 2 for further details d The Classification based on the probability of a given source to be called a BL Lac/FSRQ/ambiguous as defined in this work. See Section 4. e The probability of a given source to be identified a BL Lac, derived from the random forest classifier.
Note — (cid:63) : The Swift X-ray position of this source is positionally coincident with an eclipsing binary, V* V2552 Cyg Facilities:
Swift(XRT), WISEREFERENCES ????Abdo, A. A., Ackermann, M., Ajello, M., et al. 2010,Astrophysical Journal, Supplement Series,doi: 10.1088/0067-0049/188/2/405Abdollahi, S., Acero, F., Ackermann, M., et al. 2020, TheAstrophysical Journal Supplement Series, 247, 33,doi: 10.3847/1538-4365/ab6bcbAcero, F., Ackermann, M., Ajello, M., et al. 2015, TheAstrophysical Journal Supplement Series, 218, 23,doi: 10.1088/0067-0049/218/2/23Ackermann, M., Ajello, M., Allafort, A., et al. 2012,Astrophysical Journal, 753, 83,doi: 10.1088/0004-637X/753/1/83 Ackermann, M., Ajello, M., Atwood, W. B., et al. 2015,The Astrophysical Journal, 810, 14,doi: 10.1088/0004-637X/810/1/14´Alvarez Crespo, N., Masetti, N., Ricci, F., et al. 2016, TheAstronomical Journal, 151, 32,doi: 10.3847/0004-6256/151/2/32Blandford, R. D., & Rees, M. J. 1978, Physica Scripta, 17,265, doi: 10.1088/0031-8949/17/3/020Breiman, L. 2001, Machine Learning,doi: 10.1023/A:1010933404324Burrows, D. N., Hill, J. E., Nousek, J. A., et al. 2005, SpaceScience Reviews, 120, 165,doi: 10.1007/s11214-005-5097-2 aur et al. 2020 Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer,W. P. 2002, Journal of Artificial Intelligence Research,16, 321, doi: 10.1613/jair.953Crespo, N. ´A., Massaro, F., Milisavljevic, D., et al. 2016,The Astronomical Journal, Volume 151, Issue 4, articleid. 95, 10 pp. (2016)., 151, 95,doi: 10.3847/0004-6256/151/4/95Cutri, R., & al., E. 2013, VizieR Online Data Catalog,2328, 0Falcone A. D., P. M., & et al. Stroh, M. 2015, Proceedingsof The 6th Fermi Symposium, Arlington, VAFossati, G., Maraschi, L., Celotti, A., Comastri, A., &Ghisellini, G. 1998, Monthly Notices of the RoyalAstronomical Society, 299, 433,doi: 10.1046/j.1365-8711.1998.01828.xGhisellini, G., Celotti, A., Fossati, G., Maraschi, L., &Comastri, A. 1998, Monthly Notices of the RoyalAstronomical Society, 301, 451,doi: 10.1046/j.1365-8711.1998.02032.xGhisellini, G., Righi, C., Costamante, L., & Tavecchio, F.2017, Monthly Notices of the Royal AstronomicalSociety, Volume 469, Issue 1, p.255-266, 469, 255,doi: 10.1093/mnras/stx806Ghisellini, G., & Tavecchio, F. 2008, Monthly Notices of theRoyal Astronomical Society,doi: 10.1111/j.1365-2966.2008.13360.xGiommi, P., Brandt, C. H., Barres de Almeida, U., et al.2019, Astronomy & Astrophysics, 631, A116,doi: 10.1051/0004-6361/201935646Kaur, A., Falcone, A. D., Stroh, M. D., Kennea, J. A., &Ferrara, E. C. 2019, The Astrophysical Journal, 887, 18,doi: 10.3847/1538-4357/ab4cebKaur, A., Rau, A., Ajello, M., et al. 2018, TheAstrophysical Journal, doi: 10.3847/1538-4357/aabdec—. 2017, The Astrophysical Journal, Volume 834, Issue 1,article id. 41, 10 pp. (2017)., 834, 41,doi: 10.3847/1538-4357/834/1/41Lott, B., Gasparrini, D., & Ciprini, S. 2020.https://arxiv.org/abs/2010.08406Marchesini, E. J., Paggi, A., Massaro, F., et al. 2020a,Astronomy and Astrophysics, 638, A128,doi: 10.1051/0004-6361/201936928—. 2020b, Astronomy and Astrophysics, 631, A150,doi: 10.1051/0004-6361/201936285Massaro, F., D’Abrusco, R., Ajello, M., Grindlay, J. E., &Smith, H. A. 2011, Astrophysical Journal Letters, 740,48, doi: 10.1088/2041-8205/740/2/L48 Massaro, F., D’Abrusco, R., Tosti, G., et al. 2012,Astrophysical Journal, 750, 138,doi: 10.1088/0004-637X/750/2/138Nolan, P. L., Abdo, A. A., Ackermann, M., et al. 2012,Astrophysical Journal, Supplement Series, 199,doi: 10.1088/0067-0049/199/2/31Padovani, P., Giommi, P., & Rau, A. 2012, MonthlyNotices of the Royal Astronomical Society: Letters, 422,L48, doi: 10.1111/j.1745-3933.2012.01234.xPaiano, S., Falomo, R., Treves, A., Franceschini, A., &Scarpa, R. 2019, The Astrophysical Journal, 871, 162,doi: 10.3847/1538-4357/aaf6e4Paiano, S., Franceschini, A., & Stamerra, A. 2017, MonthlyNotices of the Royal Astronomical Society, Volume 468,Issue 4, p.4902-4937, 468, 4902,doi: 10.1093/mnras/stx749Pe˜na-Herazo, H. A., Marchesini, E. J., ´Alvarez Crespo, N.,et al. 2017, Astrophysics and Space Science, 362, 228,doi: 10.1007/s10509-017-3208-7Rajagopal, M., Kaur, A., Ajello, A., et al. 2020.https://arxiv.org/abs/2002.07740Rau, A., Schady, P., Greiner, J., et al. 2012, Astronomy &Astrophysics, 538, A26,doi: 10.1051/0004-6361/201118159Saz Parkinson, P. M., Xu, H., Yu, P. L. H., et al. 2016, TheAstrophysical Journal, 820, 8,doi: 10.3847/0004-637X/820/1/8Shaw, M. S., Romani, R. W., Healey, S. E., et al. 2009,Astrophysical Journal,doi: 10.1088/0004-637X/704/1/477Shaw, M. S., Romani, R. W., Cotter, G., et al. 2013,Astrophysical Journal,doi: 10.1088/0004-637X/764/2/135Taylor, M. B. 2005, in Astronomical Data AnalysisSoftware and Systems XIV, Vol. 347, 29–+Voges, W., Aschenbach, B., Boller, T., et al. 1999,Astronomy and Astrophysics, v.349, p.389-405 (1999),349, 389. https://arxiv.org/abs/9909315—. 2000, International Astronomical Union Circular, 7432,1. https://ui.adsabs.harvard.edu/? { }}