X-RAY SOURCES IN THE DWARF SPHEROIDAL GALAXY DRACO

The 87 ks image of Draco (Figure 1; R.A. = 17^h20^m124 and decl. = +57°54'553) was obtained by the European Space Agency's (ESA) X-ray Multi-Mirror Mission—Newton (XMM-Newton) between 2009 August 4 and 28 (PI: K. Dhuga; see Table 1 for details).

Zoom In Zoom Out Reset image size

We used the 3XMM-DR5⁴ (3XMM) catalog (Rosen et al. 2015) for the automated multiwavelength (MW) classification described in Section 3.1. The parameters obtained from the 3XMM catalog are the source coordinates and X-ray fluxes in four different energy bands.

For several of the brighter X-ray sources and for the diffuse background emission (see Section 4), we performed spectral analyses and extracted spectra from the EPIC data. The observation data files (ODFs) were processed using standard procedures of the XMM-Newton Science Analysis System (XMM-SAS version 13.0.1). The standard SAS tool edetect-chain was used to perform source detections. We extracted spectra using circular (r = 17'') regions centered on the X-ray sources for each observation. Background subtraction was performed using source-free apertures located nearby. A standard event screening described in the multixmmselect manual⁵ was applied. The spectral analysis of the X-ray data was performed using XSPEC version 12.7. The pn and MOS spectral energy channels were grouped to have at least 10 counts per bin. Each spectrum was fit in the 0.2–10 keV energy range.

The Hubble Space Telescope (HST) partly observed Draco with the WFPC2, ACS/WFC, and WFC3/UVIS cameras (Table 2 lists the images used for this study). Figure 1 shows the HST pointings, in white, overlaid on top of the EPIC image. Figure 2 shows 8'' × 8'' cutouts centered on the X-ray sources observed by HST. We use the high-resolution HST images to distinguish between active galactic nuclei (AGNs) and stars.

Zoom In Zoom Out Reset image size

Traditional MW classification includes heuristic examination of the MW source properties, such as X-ray hardness ratios, X-ray/optical/NIR flux ratios, as well as color–color and color–magnitude diagrams (CMDs) in the optical and IR, such as those shown in Figures 3–5, respectively. These diagrams allow one to devise crude criteria with which to discriminate between sources of different natures (e.g., Kaplan et al. 2006; Misanovic et al. 2010; Kargaltsev et al. 2012; Lin et al. 2012). While the MW analysis is often the only reliable way to unveil the nature of unknown X-ray sources, traditional classification, based on drawing dividing lines in the MW diagrams, is a very laborious process lacking any quantitative classification confidence criterium.

Zoom In Zoom Out Reset image size

There are over 100 X-ray sources in the 3XMM catalog⁶ in the Draco field. We visually inspected all of the sources, and after excluding source duplications (in five separate observations of Draco) and spurious detections (e.g., edge of the chip, bad CCD columns), we selected 35 bright X-ray sources with a signal-to-noise ratio $({\rm{S}}/{\rm{N}})\gt 6$ for further investigation. Table 3 lists the source parameters: X-ray flux ${F}_{0.5\mbox{--}12}$ in the 0.5–12 keV range and two hardness ratios HR2 and HR4 defined as $\mathrm{HR}2=({F}_{(1.2\mbox{--}2}-{F}_{0.2\mbox{--}1.2})/({F}_{1.2\mbox{--}2}+{F}_{0.2\mbox{--}1.2})$ and $\mathrm{HR}4=({F}_{2\mbox{--}7}-{F}_{0.5\mbox{--}2})/({F}_{2\mbox{--}7}+{F}_{0.5\mbox{--}2})$, where ${F}_{x\mbox{--}y}$ are the observed fluxes in the respective energy bands. Figure 3 shows the hardness ratio diagram for the sources in Draco compared to different classes of literature-verified sources used for automated classification (see A.1 for details). The two chosen hardness ratios allow for some separation between AGNs and the rest of the sources, although there is an expected overlap between the AGNs, CVs, and pulsars.

Table 3.  Properties and Automated Results for Classification X-Ray Sources in the Draco Field

#	R.A.	decl.	δra	Fluxb	HR2c	HR4c	S/Nd	ge	ie	W1f	Classg (Prob.)
1	260.0917	57.9740	0.40	86(2)	−0.43(1)	−0.86(1)	115.4	12.80	11.81	7.73	GM Drah
2	260.4930	57.8228	0.56	157(16)	−0.26(2)	−0.31(6)	51.3	11.21	10.10	8.25	Star/CV ?
3	260.1406	58.1417	0.58	74(8)	−0.36(2)	−0.11(6)	48.4	17.34	16.24	12.83	AGN (84%)
4	260.1575	58.0366	0.49	21(2)	−0.24(3)	−0.68(6)	43.4	11.39	12.19	7.80	Stari (99%)
5	260.2180	57.9202	0.47	18(2)	−0.20(4)	0.08(7)	31.8	20.29	19.96	14.84	AGN (99%)
6	259.8937	57.9806	0.47	15(3)	−0.33(4)	−0.1(1)	26.1	19.34	19.02	13.90	AGN (99%)
7	259.8409	57.8290	0.52	5(2)	−0.08(8)	−0.1(2)	16.1	22.76	21.69	15.98	?
8	259.7567	58.0081	0.67	6(2)	0.0(1)	0.2(2)	14.4	18.79	18.68	15.18	AGN (97%)
9	260.1548	57.8155	0.79	12(3)	−0.73(6)	0.5(1)	14.3	19.32	18.28	13.97	AGNj (99%)
10	259.9750	57.9977	0.71	1(2)m	0.91(1)	0.78(4)	14.2	22.97	21.88	⋯	?
11	260.1798	57.9119	0.69	5(1)	0.5(1)	0.4(1)	14	20.93	20.22	15.93	CV/AGN
12	259.7936	57.8301	0.75	10(2)	0.69(9)	0.44(9)	13.4	18.96	17.46	13.87	AGNj (88%)
13	259.7689	57.7914	0.81	5(2)	−0.4(1)	0.3(2)	12.5	20.49	19.12	15.75	AGN (99%)
14	260.2161	57.6997	0.86	15(5)	0.1(1)	0.3(2)	12.5	19.85	19.72	15.41	AGN (99%)
15	260.0755	57.8516	1.65	2.6(8)	−0.3(1)	−0.2(2)	12.4	19.48	16.85	13.88	AGN (84%)
16	259.8326	57.9952	0.85	9(2)	0.9(1)	0.61(8)	12	18.80	17.68	14.97	AGN ?
17	259.9902	57.8351	0.86	0.1(4)m	−1.0(${}_{-0}^{+0.3}$)n	0.9(${}_{-0.9}^{+0.1}$)n	10.6	17.75	16.19	13.26	C1k
18	260.3452	57.8417	1.02	5(2)	−0.0(1)	0.1(2)	10.4	18.05	17.81	14.89	AGN (84%)
19	259.7686	58.0586	0.86	6(3)	−0.2(1)	−0.2(3)	10	⋯	⋯	⋯	?
20	260.4225	57.8769	1.11	9(3)	−0.3(1)	0.3(2)	10	⋯	⋯	15.75	AGN?
21	259.9869	57.7715	1.17	5(2)	0.2(1)	0.2(2)	9	⋯	⋯	⋯	?
22	259.7530	57.8628	0.99	2(1)	0.0(1)	−0.2(3)	8.8	20.27	19.18	16.19	AGN (99%)
23	259.9394	57.8825	1.05	0.9(5)	−0.1(1)	−0.2(3)	8.6	23.20	21.33	16.01	?
24	259.8536	57.7479	1.14	2(1)	0.1(2)	−0.3(4)	8.3	⋯	⋯	⋯	?
25	260.1971	57.8889	1.17	1.3(9)	−0.0(2)	0.0(4)	7.9	⋯	⋯	⋯	?
26	259.9040	57.7832	1.16	1.3(9)	−0.4(2)	0.1(3)	7.9	22.84	21.35	⋯	AGN?
27	260.2515	57.8649	1.05	1(1)	−0.0(4)	−0.1(2)	7.6	21.45	21.53	⋯	AGN?
28	260.3006	57.8342	1.06	3(2)	−0.2(2)	0.4(2)	7.6	22.99	23.18	⋯	?
29	260.1186	57.9400	1.20	0.6(5)	0.4(2)	0.0(3)	7.3	⋯	⋯	⋯	?
30	260.4516	57.9685	2.00	2(3)m	−0.4(2)	0.1(5)	7.2	20.22	19.81	16.19	AGN (99%)l
31	260.2698	57.8927	1.21	2(2)	−0.2(2)	−0.1(4)	7.2	15.97	15.03	13.17	Star (99%)
32	260.1072	57.8848	1.16	0.3(3)	−0.9(1)	−1.0(${}_{-0}^{+0.5}$)n	7.1	14.30	11.30	10.16	Star (95%)
33	259.9014	57.8166	1.18	0.9(6)	−0.0(2)	−0.2(3)	6.8	⋯	⋯	⋯	?
34	259.6866	57.8509	6.27	6(6)	−0.5(2)	0.3(4)	6.2	⋯	⋯	⋯	?
35	260.1380	58.1042	1.69	1(1)	0.0(2)	−0.4(6)	6.1	⋯	⋯	⋯	?

Notes.

^a1σ positional uncertainty in arcseconds (from 3XMM-DR5 catalog). ^bObserved X-ray fluxes in the 0.5–12 keV range in units of 10⁻¹⁴ erg s⁻¹ cm⁻². The value in brackets is the measurement uncertainty in the last digit. ^cHardness ratio calculated as $\mathrm{HR}2=({F}_{(1.2-2}-{F}_{0.2-1.2})/({F}_{1.2-2}+{F}_{0.2-1.2})$ and $\mathrm{HR}4=({F}_{2-7}-{F}_{0.5-2})/({F}_{2-7}+{F}_{0.5-2})$, where ${F}_{X-Y}$ is the observed flux in the X − Y keV energy band. The value in brackets is the measurement uncertainty in the last digit. ^dSignal-to-noise ratio. ^eg and i magnitudes front the USDSS DR9 catalog. ^fW1 magnitudes front the WISE catalog. ^gClassification confidence (see text). ^hGM Dra, variable star in Tycho 2 catalog (Høg et al. 2000). ⁱA cool K9-type star with X-ray luminosity of 3.7 × 10²⁹ erg s⁻¹ (at a distance of 125 pc; Ammons et al. 2006). ^jVisually resolved in the SDSS images. ^kSymbiotic star in Draco, see Belczyński et al. (2000). ^lSource #30 also shows complex morphology (likely, superposition of star and galaxy as can be seen in Figure 2. ^mThree sources have the uncertainties of fluxes in the individual bands (1−2 keV, 2−4.5 keV, etc.) larger than the fluxes themselves (according to 3XMM-DR5), which results in the large total flux uncertainty. ⁿIn these cases, the uncertainty is strongly asymmetric because the value of HR hits the boundary, and therefore two uncertainties are provided.

Only a portion of this table is shown here to demonstrate its form and content. A machine-readable version of the full table is available.

Download table as:  DataTypeset image

Similarly, we used color–color diagrams to compare the sources in Draco to known sources of different classes (AGNs, CVs, pulsars etc.). The left panel of Figure 4 shows a good degree of separation between the stars and AGNs in terms of IR to X-ray flux ratio⁷ , which is often used to separate the two classes. The color–color diagram in the right panel of Figure 4 also shows clear separation between the AGNs and stars but it also allows us to see the noticeable separation between the AGNs and CVs. We also utilize CMDs (see Figure 5) and visually inspect the HST and SDSS images. The former allow us to estimate the age of stellar sources, which in turn helps to determine whether a given star is part of Draco. If a star whose absolute magnitude is calculated for the Draco distance happens to land on a $\lt 1\;{\rm{Gyr}}$ isochrone, it is unlikely to be part of Draco because the stellar population of the dSph is at least several billion years old (Segall et al. 2007). AGNs have a spectral energy distribution (SED) that is different from that of stars, and therefore on different CMDs, AGNs "jump" between different isochrones or do not land on one at all.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Table 4.  Best-fit Spectral Results for Selected Point Sources in the Draco Field

Source #	Model	Normalizationa	nHb	kT, Tine	Γ	${\chi }_{\nu }^{2}$	νf
			(1022 cm−2)	(keV)
#10	PL	(1.53 ± 0.48) × 10−5	1.35 ± 0.27	⋯	1.72 ± 0.23	0.85	171
	diskbb	(1.92 ± 1.15) × 10−4	0.83 ± 0.16	2.09 ± 0.35	⋯	0.85	171
	bbodyrad	(3.35 ± 0.89) × 10−3	0.35 ± 0.14	1.11 ± 0.08	⋯	0.88	171
#11	bbodyrad+PL	(5.64 ± 2.94) × 10−3c	<0.027	0.75 ± 0.11	1.24 ± 0.48	1.05	171
	⋯	(2.20 ± 1.35) × 10−6d	⋯	⋯	⋯	⋯	⋯
#16	PL	(1.85 ± 0.50) × 10−5	0.92 ± 0.17	⋯	1.82 ± 0.21	1.16	60
#19	PL	(1.01 ± 0.16) × 10−5	0.06 ± 0.04	⋯	2.91 ± 0.38	0.77	43
	bbody	(3.32 ± 0.28) × 10−7	<0.027	0.18 ± 0.01	⋯	1.06	43
	bbodyrad	2.21 ± 0.63	<0.027	0.18 ± 0.001	⋯	1.06	43
#23	PL	(3.47 ± 1.98) × 10−6	0.23 ± 0.18	⋯	3.6 ± 1.3	0.89	75
	bremss	(5.63 ± 5.56) × 10−6	< 0.027	0.83 ± 0.55	⋯	0.88	75
#25	diskbb	(2.12 ± 0.14)×10−3	<0.027	0.64 ± 0.11	⋯	0.83	197

Notes.

^aNormalization of the bbodyrad model is ${R}_{\mathrm{km}}^{2}/{D}_{10}^{2}$, where R_km is the source radius in km, D₁₀ is the distance to the source in units of 10 kpc and the normalization of the diskbb model given as ${(({R}_{\mathrm{in}}/{\rm{km}})/(D/10{\rm{kpc}}))}^{2}$ cosθ, where R_in is the is "an apparent" inner disk radius. ^bGalactic extinction is assumed as 2.7 × 10²⁰ cm⁻² for each source. ^cNormalization of the bbody component. ^dNormalization of the PL component. ^ekT; temperature keV from bbody and bbodyrad models. T_in; temperature at inner disk radius (keV) from diskbb model. ^fNumber of degrees of freedom.

Download table as:  ASCIITypeset image

Since we are interested in objects that belong to Draco, we attempt to eliminate Galactic stars. One helpful constraint is the optical brightness of these objects at the known distance to Draco. The distance modulus of Draco (80 kpc) is ∼19.5. Figure 5 clearly shows that we do not expect to see any stars brighter than ${M}_{g}\approx -1$ (a very conservative estimate, which translates to apparent magnitude m_g ≈ 18), assuming the age of Draco to be a few Gyr. This means that stars brighter than m_g ≈ 18 likely belong to our Galaxy rather than Draco. (This criteria does not apply to QSOs/AGNs, which can be brighter.) Sources #1, #2, #4, #11, #31, and #33 appear to coincide with stars associated with our Galaxy. Inspection of the HST images confirms that source #1 also contains a very bright foreground star.

In addition, we utilize the superb angular resolution of HST to identify AGNs (we also check the lower-resolution SDSS images when HST image is not available). Unfortunately, only very limited (often single-band) HST coverage of Draco currently exists (see Figure 1, left panel). Nonetheless, high-resolution HST images allow us to eliminate several foreground and background sources to Draco (see Table 3).

Sources #5 and #6 appear to be point-like in the HST images but, given their optical magnitudes (of 19–20), we cannot exclude the possibility of QSOs (which could also appear point-like). Based on the SDSS images and photometry (see Figure 4, right panel), we also conclude that #2 and #32 are likely (bright) foreground stars. The SDSS images suggest that #12 and #9 are AGNs (in agreement with Figure 4).

We also attempted to verify the classification using the SIMBAD catalog.⁸ However, we found that the SIMBAD classifications are incorrect for several faint sources. For example, we find AGNs (sources #9 and #12, resolved with HST) to be classified as stars. Also, a few stars are referred to as QSOs/AGNs in SIMBAD. Therefore, we do not rely on the SIMBAD for classifications of the X-ray sources in Draco.

As an aid to help with the classification, we have developed an automated machine-learning classification pipeline (MUWCLASS; Brehm et al. 2014) that we subsequently applied to the X-ray sources in Draco. The pipeline uses a learning decision tree algorithm to perform the X-ray source classification (see A.1 for details). The automated procedure produced 20 classifications with confidence $\gt 70\%$, including 14 AGNs, 5 stars, and 1 CV (see Table 3). The classifications with lower confidence are indicated by a question mark. Note that the calculated confidences for each type do not include the uncertainties associated with the X-ray flux determination, which are substantial for faint X-ray sources and also affect HRs. They also do not include the possibility of assigning false MW counterparts to the X-ray sources (confusion), however, this group is expected to be small (see Section 3.2). For a number of interesting X-ray sources which could belong to Draco or be XRBs or compact objects in our Galaxy, we have carried out a more detailed analysis (see Section 4.1) by extracting the X-ray spectra from the original data and performing spectral fits (not used in the automated classifications). The locations of the sources in the diagrams shown in Figures 3–5 are mostly consistent with the automated classifications or, whenever this is not the case, the reasons for the discrepancy are discussed in Section 4.1 on a source-by-source basis.

After we submitted this paper, we became aware of two other papers (Manni et al. 2015 and Saeedi et al. 2015) discussing X-ray source classification in Draco that were  submitted roughly at the same time as our paper. The Saeedi et al. (2015) paper presents a more comprehensive analysis than that of Manni et al. (2015) and, bearing in mind differences in class definitions, nearly all of our classifications agree with Saeedi et al. (2015), except for source 15 for which our automated tool preferred an AGN (84%) classification, while Saeedi et al. (2015) suggest a binary star (not an XRB) classification (but uncertain). We also note that the machine-learning method used in our paper is different from those used by Saeedi et al. (2015) and Manni et al. (2015), who relied on previously established heuristic prescriptions from multiple publications. However, the very good agreement between our classifications and those of Saeedi et al. (2015) shows that both methods worked well for this field and similar results were obtained with different classification approaches, adding credibility to both independent analyses.

We cross-correlated all of the X-ray sources listed in Table 3 with the USNO-B1, SDSS, 2MASS, and WISE catalogs. We find a single optical/NIR/IR counterpart within the 2'' search radius⁹ for all but one X-ray source¹⁰ . The only sources with multiple counterparts within the search radius are source #5, which has two USNO-B1 counterparts, and source #32, which also has two counterparts in SDSS. For those two sources, we only consider the brightest (which also happens to be the closest) counterpart to be the "true" counterpart below. We have verified that all of the optical/NIR/IR counterparts are within ∼05 of one another. We see no systematic offset between the X-ray and 2MASS sources and the rms offset between the X-ray and 2MASS/SDSS positions for the 35 sources listed in Table 3 is ≈1'', i.e., well within the distribution of uncertainties for the X-ray source positions (see Table 3). We note that while the HST images reveal multiple counterparts for some of the sources, the surveys are not as deep, and therefore pick up only the brightest objects in the 2'' search radius around each source. Unfortunately, only five X-ray sources are located within the HST fields, and even for them multiband HST photometry is lacking. We cannot exclude that some of the brighter X-ray sources are associated with the much fainter optical sources that can only be seen in the HST images. However, these would have to be sources of very rare types (e.g., solitary neutron stars (NSs) or quiescent BH binaries¹¹ in Draco or extended Galactic halo) or as-yet unknown classes. As for the fainter X-ray sources, we can hardly say much without deep XMM-Newton and HST observations.

For each survey, we calculate the chance superposition using three different techniques. First, based on the average optical/NIR/IR source densities in the field (ρ = 0.0048, 0.01, 0.0004, and 0.001 stars arcsec⁻² in the USNO-B1, SDSS, 2MASS, and WISE surveys, respectively), we calculate the probability of finding zero field sources in the circle of radius r = 2'', $P=\mathrm{exp}(-\rho \pi {r}^{2})$. This leads to chance coincidence probabilities of 1 − P = 6%, 12%, 0.5%, and 1% (for the USNO-B1, SDSS, 2MASS, and WISE surveys, respectively). For the second method, we run 1000 simulations where we randomly populate the optical/NIR/IR field of Draco with 35 synthetic X-ray sources (similar to, e.g., Rangelov et al. 2012). We then count the average number of optical/NIR/IR counterparts to the synthetic X-ray sources within the r = 2'' radius. This method results in 0.9, 1.9, 0.08, and 0.27 (for the USNO-B1, SDSS, 2MASS, and WISE surveys, respectively) spurious detections of the 35 sources. The third method we employed was to offset each X-ray source in a random direction (within 100'' of its original position). This led to 1, 1.8, 0.1, and 0.4 (for the USNO-B1, SDSS, 2MASS, and WISE surveys, respectively) spurious detections in 1000 simulations. Therefore, all three methods give very similar results suggesting only 1–2 spurious detections, compared to the 26 X-ray sources for which we find optical/NIR/IR counterparts (see Table 3).

Below, we discuss individually several sources that do not appear to be AGNs or foreground Galactic stars based on the color–color and CMDs, HST/SDSS images, and automated classification. For the brightest of these sources, we performed spectral fits with blackbody radiation (bbodyrad), disk blackbody (diskbb), and power-law (PL) models modified by interstellar absorption (phabs model in XSPEC). In general, these models provide an adequate description of the thermal and non-thermal components of X-ray spectra typically observed in the 0.2–10 keV range for objects such as CVs, isolated NSs, and LMXBs. The best-fit parameters for these sources are given in Table 4.

Source #10: The optical-IR properties of this source exhibit characteristics of an old (few Gyr), evolved star. The X-ray spectrum is well fit with the bbodyrad model with a temperature of of 1.11 ± 0.08 keV. Normalization of the bbodyrad model corresponds to an effective emission region of 0.44 km at 80 kpc. The absorption n_H = 3.5 × 10²¹ cm⁻² exceeds the Galactic n_H = 2.69 × 10²⁰ cm⁻² in this direction (Chandra Colden toolkit¹² ). The source diskbb model gives a similar n_H and temperature of 2.1 ± 0.3 keV. If the source is in Draco, then the normalization of the diskbb model, ${(({R}_{\mathrm{in}}/\mathrm{km})/(D/10\mathrm{kpc}))}^{2}$ cosθ, where θ is the angle of the disk (assumed face-on), corresponds to ${r}_{\mathrm{in}}=0.11$ km, which seems to be too small for an accretion disk in an XRB. The X-ray luminosity, L_X ≃ 8 × 10³⁴ erg s⁻¹, is too high for a redback or black widow type binary unless it goes through an outburst episode and switches to an accretion state (see e.g., Stappers et al. 2014). However, the X-ray light curve does not show any significant variability during any of the five XMM-Newton observations. Alternatively, it is possible that #10 is a non-accreting MSP in a Galactic binary with a late-type companion at a distance of ∼1 kpc with the X-rays being emitted from the hot polar cap of the NS. However, the large (compared to the Galactic) absorbing column would be puzzling in this scenario.

Source #11: The X-ray spectrum requires two components, bbody (BB)+PL, as PL or BB models alone do not provide acceptable fits to the data. Both the temperature (${kT}=0.7\pm 0.1$ keV) and the photon index (Γ = 1.2 ± 0.5) are not uncommon for several object types (e.g., AGN and CV), but even for the two-component model the data show systematic excess below 0.4 keV over the best fit, suggesting that an even softer component is needed. Although the luminosity of 4.5 × 10³² erg s⁻¹ at d = 80 kpc is typical of a redback or black widow system, the spectrum is unusually soft for these kinds of binaries. The absorbing column is uncertain and low, consistent with the Galactic value. The HST observations (Figure 2) show a point-like source and an extended structure, which may or may not be related to the point source. Given the high density of background galaxies in the deep HST images, it is possible that we are seeing the superposition of a star (either in Draco or in the Galaxy) and a background galaxy/AGN. This could explain why the automated classification is confused (split between AGN and CV). We found no variability between five XMM-Newton exposures.

Source #16: This source is strongly (intrinsically) absorbed with an odd jump $\gt 8\;{\rm{keV}}$ in the X-ray spectrum. The source's optical/NIR properties are consistent with those of an evolved low-mass single star, but the apparently high absorption in X-rays rules out an active corona in a nearby star as a source of X-rays. It is possible that this source is a symbiotic star or CV in Draco. In fact, the X-ray spectrum shows signs of emission line(s) from the iron complex (Figure 6). The spectrum also shows hints of other lines at lower (1−2 keV) energies. This could be indicative of a magnetic CV (Fujimoto & Ishida 1997). We would like to note that symbiotic stars are not a separate class in our current training data set for MUWCLASS, and therefore such a system would be classified as another object type. This source is classified as an AGN by the automated algorithm but with low confidence (71%). The X-ray light curve does not exhibit any significant variability.

Zoom In Zoom Out Reset image size

Source #17: This source has been reported as a candidate symbiotic star (Belczyński et al. 2000). Aaronson et al. (1982) identified this source as a carbon star with an unusual SED showing strong emission lines. The authors suggest that the optical colors can be explained by a symbiotic binary (a red giant with a hot main-sequence companion). Aaronson et al. (1982) also claim that #17 is in Draco based on radial velocity measurements. However, the observed X-ray flux implies a high X-ray luminosity of L_X = 1.2 × 10³³ erg s⁻¹ at the distance of Draco, which can hardly be produced in a non-degenerate binary with a red giant and a late-type main-sequence star. Therefore, we consider this source to be a good candidate for a quiescent XRB in Draco. The automated classification did not produce a confident result. The confusion could be due to the unusual SED noticed by Aaronson et al. (1982), who also suggested that the star could be in a binary with a (pre-)degenerate object or that it can be a star caught very early in the process of ejecting its outer layers and forming a planetary nebula. If the object does not fit any of the predefined object classes, then the automated algorithm is expected to be confused.

Source #19: Given its X-ray flux and the lack of an optical counterpart (arguing against binary nature), it is unlikely that this source would belong to Draco (X-ray luminosity, ${L}_{{\rm{X}}}=4.6\times {10}^{35}$ erg s⁻¹, for d = 80 kpc, is too high for isolated NSs), and so it could be an old recycled or non-recycled pulsar in a non-accreting binary in our Galaxy. However, the source is imaged strongly off-axis in the XMM-Newton image, which increases the positional uncertainty, even though source #19 has one of the smallest positional uncertainties in Table 3 (possibly underestimated by the automated source detection tool in 3XMM). There is an SDSS source within 3'' with u = 20.6, g = 20.4, r = 20.3, i = 20.3, and z = 20.3. If this SDSS source is a counterpart of the X-ray source, then it would exclude the possibility of #19 being a nearby isolated pulsar/NS, which otherwise would be possible based on its X-ray spectrum and lack of optical/IR counterparts. The lack of optical/IR data prevents our algorithm from confidently classifying this source.

Source #23: This source is faint with a soft X-ray spectrum, too soft for a typical AGN. There is some evidence of even further hardening beyond 5−6 keV in pn. PL and bremss provide a reasonable fit to the data. Our automated algorithm does not provide a confident classification, likely due to incomplete MW parameters (a very faint counterpart found in SDSS, but not in 2MASS).

Note that nine sources lack counterparts in SDSS. These could be AGNs that are too faint to be detected in SDSS or unusual objects that produce little emission in the optical, such as isolated NSs or quiescent BH binaries in Draco. Of those nine, only source #25 was imaged by HST (Figure 2). Two point-like objects are seen within the r = 2'' circle, one of which could be the counterpart of X-ray source. If this is the case, then it could be LMXB/CV in Draco (based on the optical brightness). Unfortunately, no multiband photometry exist for this field, which prevents us from classifying the possible optical counterpart of the source.

After careful examination of all of the X-ray sources, we find four potentially interesting sources (#10, #16, #17, and #25) that could belong to Draco and be potential LMXBs, CVs, and/or symbiotic candidates. We note in passing that Sculptor is reported to host five LMXBs (Maccarone et al. 2005) and Fornax to have two to three "field" LMXBs (i.e., not part of the Fornax globular clusters, which have another two LMXBs; Nucita et al. 2013b). Given the small number statistics and lack of certainty associated with Draco (except, perhaps, source #16 and #17), our findings are roughly on par with those for Sculptor and Fornax (both of which are Milky Way satellites and at roughly the same distance as Draco). However, we caution the reader that the respective analyses use different techniques and rely on different MW data, and therefore a direct comparison is speculative at best.

Walker et al. (2015) recently presented optical spectroscopy for stars in the Draco field. They fit stellar models to all of the sources and produced a data set with the derived properties (line-of-sight velocity, V_los, effective temperature, T_eff, surface gravity, g, metallicity, [Fe/H], etc.). We cross-match the spectroscopic data set with the positions of the 35 X-ray sources and find five matches—sources #9, #16, #17, #22, and #30. The optical spectroscopy for sources #9, #22, and #30 has a very low signal-to-noise ratio (S/N < 3), which resulted in very uncertain fit parameters, and hence no classifications are provided in Walker et al. (2015). These low signal-to-noise sources were classified (by our algorithm) as AGNs (in agreement with Figure 4). Sources #16 and #17 have ${V}_{\mathrm{los}}=-295\pm 1$, which is consistent with the V_los of Draco, and therefore suggests the association of the two sources with the dSph. Spectroscopy suggests that these two sources are evolved, low-mass stars, which is consistent with the CV/LMXB scenarios discussed above.

It has been speculated that dSphs can host IMBHs (Maccarone et al. 2005; Nucita et al. 2013a and references therein). We estimate the X-ray luminosity for a putative IMBH in Draco due to Bondi–Hoyle accretion as

$$\begin{eqnarray}{L}_{{\rm{X}}} & \simeq & 3.4\times {10}^{37}\epsilon {\left(\displaystyle \frac{{M}_{\mathrm{BH}}}{{10}^{3}{M}_{\odot }}\right)}^{2}{\left(\displaystyle \frac{{c}_{s}}{10\mathrm{km}{{\rm{s}}}^{-1}}\right)}^{-3}\\ & & \times \left(\displaystyle \frac{n}{0.1\;{\mathrm{cm}}^{-3}}\right)\;\mathrm{erg}\;{{\rm{s}}}^{-1},\end{eqnarray} \tag{ 1 }$$

where is the radiative efficiency and n is the hydrogen number density in Draco. Based on observational results (Mateo 1998; Grcevich & Putman 2009), we estimate an upper limit of n < 0.02 cm⁻³ and a sound speed of c_s ≈ 10 km s⁻¹ for dSphs, which should be of the order of the stellar velocity dispersion (Hargreaves et al. 1996; Grebel et al. 2003). Therefore, the corresponding flux at the distance of Draco is

$$\begin{eqnarray}{F}_{{\rm{X}}} & \simeq & 8.8\times {10}^{-12}\epsilon {\left(\displaystyle \frac{{M}_{\mathrm{BH}}}{{10}^{3}{M}_{\odot }}\right)}^{2}{\left(\displaystyle \frac{{c}_{s}}{10\mathrm{km}{{\rm{s}}}^{-1}}\right)}^{-3}\\ & & \times \left(\displaystyle \frac{n}{0.02\;{\mathrm{cm}}^{-3}}\right)\;\mathrm{erg}\;{{\rm{s}}}^{-1}\;{\mathrm{cm}}^{-2}.\end{eqnarray} \tag{ 2 }$$

According to the 3XMM catalog, there are four X-ray sources with ${\rm{S}}/{\rm{N}}\approx 3\mbox{--}5$ within ∼3' of the center of Draco (R.A. = 17^h20^m132 and decl. = +57°54'553; Rave et al. 2003). The measured flux for the nearest source is ${F}_{{\rm{X}}}\approx 4.5\times {10}^{-15}$ erg s⁻¹ cm⁻² (the other three sources have similar fluxes and S/N). According to Beskin & Karpov (2005), the expected radiative efficiency for synchrotron radiation (presumably the dominant radiation mechanism for a slowly accreting BH) is $\epsilon \sim {10}^{-5}$. Assuming that one of these X-ray sources is the putative IMBH, we can then estimate its mass. However, it is possible that none of these faint X-ray sources represent the putative IMBH. Therefore, we also can determine the lower limit on the IMBH flux by measuring the diffuse flux fluctuations (∼10⁻¹⁵ erg s⁻¹ cm⁻² at 3σ level) in the local background in the combined EPIC image. This leads to the following constraint:

$$\begin{eqnarray}&&{M}_{\mathrm{BH}}{\left(\displaystyle \frac{n}{0.02{\mathrm{cm}}^{-3}}\right)}^{1/2}\lt 3.4\times {10}^{3}{\left(\displaystyle \frac{{F}_{{\rm{X}}}}{{10}^{-15}{\rm{cgs}}}\right)}^{1/2}\\ &&\quad \times {\left(\displaystyle \frac{\epsilon }{{10}^{-5}}\right)}^{-1/2}{\left(\displaystyle \frac{{c}_{s}}{10\mathrm{km}{{\rm{s}}}^{-1}}\right)}^{3/2}\;{M}_{\odot }.\end{eqnarray} \tag{ 3 }$$

Theoretical models and observational constraints (e.g., Gebhardt et al. 2000) show that the central BH mass would depend on the velocity dispersion as

$$\begin{eqnarray}&&{M}_{\mathrm{BH}}=1.6\times {10}^{3}{\left(\displaystyle \frac{\sigma }{10\mathrm{km}{{\rm{s}}}^{-1}}\right)}^{3.75}\;{M}_{\odot },\end{eqnarray} \tag{ 4 }$$

based on the Draco velocity dispersion measurement by Hargreaves et al. (1996). This mass estimate is comparable to the limit calculated above. We note, however, that the mass estimate is based on the use of the upper limit on the hydrogen density and a rather optimistic efficiency of 10⁻⁵. More conservative values for the density and (especially) the efficiency (where the uncertainty is the largest) could easily reduce the mass limit by several orders of magnitude.

We also searched for the ∼3.5 keV line that was previously reported in the Perseus galaxy cluster (Boyarsky et al. 2014) and M31 (Bulbul et al. 2014). Figure 7 shows the X-ray spectrum of the diffuse EPIC-pn emission from 0.5 to 10.0 keV. No significant spectral features consistent with the previously reported line at ∼3.5 keV are seen.

Zoom In Zoom Out Reset image size

Our automated classification tool, MUWCLASS (Brehm et al. 2014; J. Hare et al. 2016, in preparation), uses C5.0¹³ , which is a supervised decision tree learning algorithm requiring a training data set with already classified sources (see below). C5.0 is an updated version of C4.5 (Quinlan 1993) that is based on the ID3 algorithm created by Quinlan (1986). We ran the C5.0 classification algorithm with the default parameters as described in https://www.rulequest.com/see5-info.html. We have adopted a Laplace prescription for the estimation of the classification confidences (Chawla 2003) in each leaf, $P=(\mathrm{TP}+1)/(\mathrm{TP}+\mathrm{FP}+C)$, where TP, FP, and C are true positives, false positives, and the number of classes, respectively, for the leaf where the source in question has ended up. Currently, MUWCLASS does not take into account the uncertainties of the MW parameters while performing the classification. Our machine-learning classification approach in this paper closely follows that of Lo et al. (2014) and Farrell et al. (2015), except that we do not use the X-ray variability parameters since none of the Draco sources appear to be variable at the statistically significant level.

The training data set is used to evaluate the parameters of objects from known classes and build the decision tree. The decision tree, calculated from the training data set, is applied to a set of unclassified objects. We have invested significant effort to investigate the catalogs and related literature to create a training data set of ≈8000 confidently classified X-ray sources detected by either Chandra or XMM-Newton (J. Hare et al. 2016, in preparation). This data set includes different numbers of sources from various classes. To compensate for the imbalance, we use the SMOTE technique (Chawla et al. 2011) to create an expanded balanced data set. The data set we used here has nine predefined objects classes: (1) main-sequence stars (General Catalog of Variable Stars; Samus et al. 2009), (2) young stellar objects (Chandra Orion Ultradeep Point Source Catalog and PAN-Carina; Getman et al. 2005; Povich et al. 2011), (3) AGNs (Veron Catalog of Quasars & AGN; Véron-Cetty & Véron 2010), (4) LMXBs (Low-Mass X-Ray Binary Catalog, 2007; Liu et al. 2007), (5) HMXBs (Catalog of High-Mass X-Ray Binaries in the Galaxy; Liu et al. 2006), (6) CVs (CVs Catalog, 2006, Downes et al. 2001), (7) isolated NSs (ATNF Pulsar Catalog; Manchester et al. 2005), (8) binary non-accreting NSs (ATNF Pulsar Catalog), and (9) Wolf–Rayet stars (The VIIth Catalog of Galactic Wolf–Rayet Stars; van der Hucht 2001). We cross-match all objects from the training data set with MW data and extract the following parameters: X-ray fluxes in four bands from the 3XMM-DR5 catalog (the same as those defined in Section 3.2), optical ugriz magnitudes form SDSS, NIR jhk magnitudes from the Two Micron All-Sky Survey (2MASS), and IR W1, W2, and W3 magnitudes from Wide-field Infrared Survey Explorer (WISE).

Figure 8 shows the cross-validation matrix for the SMOTEed training data set which has the true object class on the X-axis and the inferred class on Y-axis. The more diagonal the matrix, the better the performance of the classification algorithm. We have also verified the classification performance by dividing the data set into two parts and using the first part for the classification tree construction and the second half for the validation. The data set is divided into parts before the SMOTE procedure is applied to the part that is used for training. We typically find that ≈93% of the sources are classified correctly, with the AGNs and stars being found more confidently than the other sources.

Zoom In Zoom Out Reset image size