DISCOVERY OF BRIGHT GALACTIC R CORONAE BOREALIS AND DY PERSEI VARIABLES: RARE GEMS MINED FROM ACVS

Candidate selection of possible RCB stars was performed using version 2.3 of the machine-learned ACVS classification catalog (MACC; Richards et al. 2012b) of variable sources cataloged from All-Sky Automated Survey (ASAS; Pojmański 1997, 2001). Full details of the classification procedure can be found in Richards et al. (2012a, 2012b). Briefly, we employ a Random Forest (RF) classifier, which has been shown to provide the most robust results for variable star classification (see, e.g., Richards et al. 2011; Dubath et al. 2011), to provide probabilistic classifications for all of the 50,124 sources in the ASAS Catalog of Variable Stars (ACVS; Pojmański 2000). The classification procedure proceeds as follows: for each source in the ACVS 71 features are computed, 66 from the ASAS light curves (e.g., period, amplitude, skew, etc.; for the full list of features we refer the reader to Richards et al. 2012b and references therein) and 5 color features from optical and near-infrared (NIR) catalogs. A training set, upon which the RF classifications will be based, is constructed using light curves from 28 separate science classes, most of which are defined using well-studied stars with high-precision light curves from the Hipparcos and Optical Gravitational Lensing Experiment surveys (Debosscher et al. 2007; Richards et al. 2011), as well as some visually classified sources from ACVS for a few of the classes that are not well represented in Hipparcos (Richards et al. 2012a). The same 71 features are calculated for all the sources in the training set, and the RF classifier uses the separation of the 28 science classes in the multi-dimensional feature space to assign probabilistic classifications to each source in the ACVS. In the end, the probability of belonging to each individual science class is provided for each ACVS source and a post-RF procedure is used to calibrate these probabilities (meaning that a source with P(Mira) = 0.5 has a ∼50% chance of actually being a Mira).

When searching for RCB stars in time-domain survey data, RF classification provides a number of advantages relative to the more commonly used method of placing hard cuts on a limited set of a few features. Many studies have focused on light curves with large-amplitude variations and a lack of periodic signal (e.g., Alcock et al. 2001; Zaniewski et al. 2005; Tisserand et al. 2008). A few recent studies have noted that additional cuts on NIR and mid-infrared colors can improve selection efficiency (Soszyński et al. 2009; Tisserand et al. 2011; Tisserand 2012). While these surveys have all proven successful, the use of hard cuts may eliminate actual RCB stars from their candidate lists.

Hard cuts are not necessary, however, when using a multi-feature RF classifier, which is capable (in principle) of capturing most of the photometric behavior of RCB stars (including the large-amplitude, aperiodic fades from maximum light as well as the periodic variations that occur near maximum light). Another general disadvantage in the use of hard cuts for candidate selection of rare sources is that the hard cuts are typically defined by known members of the class of objects for which the search is being conducted. Any biases present in the discovery of the known members of a particular class will then be encoded into the absolute (i.e., hard cuts) classification schema. This can exclude subclasses of sources that differ slightly from the defining members of a class. Furthermore, new discoveries will be unable to refine the selection criteria since, by construction, they will fall within the same portion of feature space as previously known examples.

The RF classifier produces an estimate of the posterior probability that a source is an RCB star given its light curve and colors. This allows us to construct a relative ranking of the RCB likelihood for all the sources in ACVS. Instead of making cuts in feature space, we can search down the ordered list of candidates. In this sense the RF classifier identifies the sources that are closest to the RCB training set relative to the other classes. The RF classifier finds the class boundaries in a completely data-driven way, allowing for the optimal use of known objects to search for new candidates in multi-dimensional feature spaces. This helps to mitigate against biases present in the training set, as classifications are performed using the location of an individual source in the multi-dimensional feature phase-space volume relative to defined classes in the training set.

The MACC RCB training set was constructed using high-confidence positional matches between ACVS sources and known RCB stars identified in SIMBAD⁷ and the literature. In total there are 18 cataloged RCB stars that are included in the ACVS, which we summarize in Table 1. The light curves of the known RCB stars were visually examined for the defining characteristic of the class: sudden, aperiodic drops in brightness followed by a gradual recovery to pre-decline flux levels. All of the known RCB stars but one, ASAS 054503–6424.4, showed evidence for such behavior. ASAS 054503–6424.4 is one of the brightest RCB stars in the LMC (V_max ≈ 13.75 mag), which during quiescence is barely above the ASAS detection threshold. The light curve for ASAS 054503–6424.4 does not show a convincing decline from maximum light, and as such we do not include it in the training set.

In addition to the 18 RCB stars in ACVS, 7 additional RCB stars are detected in ASAS with the characteristic variability of the class.⁸ These sources all have clearly variable ASAS light curves; their exclusion from the ACVS means there is some bias in the construction of that catalog. In order to keep this bias self-consistent the training set for the MACC only included sources from Richards et al. (2011) and supplements from ACVS (see Richards et al. 2012a). We note that a future paper to classify all ∼12 million sources detected by ASAS will include all ASAS RCB stars in its training set (J. W. Richards et al., in preparation). Therefore, the training set includes 17 RCB stars, which is limited by the coverage and depth of ASAS, the selection criteria of the ACVS, and the paucity of known RCB stars in the Galaxy. There are no known DYPers in ACVS: only two are known in the Galaxy and the DYPers in the MCs are fainter than the ASAS detection limits. Nevertheless, the similarity in the photometric behavior of RCB stars and DYPers allows us to use the RCB training set to search for both types of star. As more Galactic RCB stars and DYPers are discovered, we will be able to supplement the training set and improve the ability of future iterations of the RF classifier (see Section 6.2).

In order to determine our ability to recover known RCB stars using the RF classifier we perform a leave-one-out cross-validation (CV) procedure. For the 17 sources in the RCB training set, we remove one source and re-run the RF classifier in an identical fashion to that used in Richards et al. (2012b). We then record the RF-determined probability that the removed source belongs to the RCB class, P(RCB), and the ranked value of P(RCB) relative to all other stars that are not included in the training set, R(RCB). We repeat the CV procedure for each star included in the training set, and the results are shown in Table 1. Since the training set is being altered in each run of the CV, R(RCB) provides a better measure of the quality of each candidate; R(RCB) is a relative quantity, whereas the calibration of P(RCB) will differ slightly from run to run. Eight of the 17 sources in the training set have R(RCB) ⩽ 3, implying that ∼50% of the training set would be a top three candidate RCB star had we previously not known about it. Fifteen of the 17 RCB stars in the training set would be in the top 0.8% of the 50,124 sources in the ACVS, while all the known RCB stars in ACVS, including ASAS 054503−6424.4 (which is not in the training set), are in the top ∼6% of RCB candidates. Two sources in the training set, SV Sge and ES Aql, are not listed near the top of R(RCB) ranking during CV. For ES Aql this occurs because the star is highly active during the ASAS observations showing evidence for at least six separate declines during the ∼10 yr observing period. As a result the light curve folds fairly well on a period of ∼397 days, and ES Aql becomes confused with Mira and semi-regular periodic variables (see Figure 2). SV Sge, on the other hand, shows significant periodicity at the parasite frequency of 1 day, which precludes it from having a high R(RCB). The CV procedure allows us to roughly tune the efficiency of our selection criteria; the purity of the selection criteria cannot be evaluated until candidates have been spectroscopically confirmed.

Due to the relative rarity of RCB stars, we elected to generate a candidate list with high efficiency while sacrificing the possibility of high purity. With only ∼50 Galactic RCB stars known to date, every new discovery has the potential to add to our knowledge of their population and characteristics. To generate our candidate list, we selected all sources from the MACC with P(RCB) > 0.1, which resulted in a total of 472 candidates. The selection criterion was motivated by the CV experiment, which indicates that our candidate list should have an efficiency ≳80%. To obtain an efficiency close to 1 would require visual examination of roughly 3000 sources.

Since the expected purity of our sample is small by design, we examine the light curves of all sources within our candidate list by eye to remove sources that are clearly not RCB stars. These interlopers are typically semi-regular pulsating variables or Mira variables, often with minimum brightness levels below the detection threshold. We use the ALLSTARS Web interface (Richards et al. 2012a) to examine candidates, which in addition to light curves provides summary statistics (period, amplitude, color, etc.) for each source, as well as links to external resources, such as SIMBAD. We also remove any sources from the candidate list that are spectroscopically confirmed as non-carbon stars.

Following the removal of these stars the candidate list was culled from 472 to 15 candidates we considered likely RCB stars, for which we obtained spectroscopic follow-up observations. The general properties of the 15 spectroscopically observed candidates, including their names, coordinates, and RF probabilities, are summarized in Table 2. Finding charts using images from the Digitized Sky Survey⁹ (DSS) for the spectroscopically confirmed RCB and DYPer candidates can be found in Figure 1. Six of the selected candidates for spectroscopic observations are known carbon stars listed in the General Catalog of Galactic Carbon Stars (CGCS; Alksnis et al. 2001; see Table 2).

Zoom In Zoom Out Reset image size

RF classifiers can provide quantitative feedback about the relative importance of each feature used for classification. The RF feature importance measure describes the decrease in the overall classifier performance following the replacement of a single feature with a random permutation of its values (see Breiman 2001 for further details). We measure the importance of each feature using the average importance from a one-versus-one classifier whereby the RCB class is iteratively classified against each of the 27 other science classes on an individual basis. This procedure is run five times and the average of all runs is taken to reduce the variance present in any single run. Unsurprisingly, we find that amplitude is the most important feature. The importance measure does not properly capture the covariance between features and as a result the majority of the important features have to do with amplitude. The second and third most important features that are not highly covariant with amplitude are qso_log_chi2nuNULL_chi2nu,¹⁰ a measure of the dissimilarity between the photometric variations of the source and a typical quasar, and freq1_harmonics_freq_0, the best-fit period. Interestingly, freq_signif, the significance of the best-fit period of the light curve, ranks as only the 31st most important out of the 71 features.

We summarize the results of these findings with two-dimensional cuts through the multi-dimensional feature space showing amplitude versus period significance, χ²_{QSO, False}, and period in Figure 2. We also show amplitude versus P(RCB). In each panel we show the location of the RCB stars in the training set as well as the newly discovered RCBs and DYPers presented in this paper, and we use the P(RCB) values from the CV experiment from Section 2.2 for the RCB stars in the training set. We also show the location of cuts necessary to achieve ∼80% efficiency (blue dashed line) when selecting candidates using only two features, as well as the cuts necessary to achieve ∼100% efficiency (red dashed line). As would be expected based on the results presented above, it is clear that χ²_{QSO, False} and period are far more discriminating than period significance when selecting RCB candidates. To achieve an efficiency near 100%, P(RCB) is vastly superior to any two-dimensional slice through feature space. We note that the discretization seen in the distribution of P(RCB) is the result of using a finite number of trials within the RF classifier. The probability of belonging to a class is defined as the total number of times a source is classified within that class divided by the total number of trials. These discrete values are then smeared following the calibration procedure described in Section 2.1.

Zoom In Zoom Out Reset image size

Many of the known and new RCB stars have very similar measured best periods clustered near ∼2400 and 5300 days, which for each corresponds to the largest period searched during the Lomb–Scargle analysis in Richards et al. (2012b). Folding these light curves on the adopted periods clearly shows that they are not periodic on the adopted periods, despite the relatively high period significance scores (see the upper left panel of Figure 2), which suggests some peculiarity in the feature generation process for these sources. We are exploring improved metrics for periodicity to be used in future catalogs. Nevertheless, despite these spurious period measurements, the ML classifier has correctly identified that this feature tends to be erroneous for RCB stars, and as such it is a powerful discriminant for finding new examples of the class.

All optical photometric observations were obtained during ASAS-3, which was an extension of ASAS, conducted at the Las Campanas Observatory (for further details on ASAS and ASAS-3 see Pojmański 1997, 2001). Light curves were downloaded from the ACVS,¹¹ and imported into our DotAstro.org (http://dotastro.org) astronomical light curve warehouse for visualization and used with internal frameworks (Brewer et al. 2009). The ACVS provides V-band measurements for a set of 50,124 pre-selected ASAS variables, measured in five different apertures of varying size (Pojmański 2002). For each star in the catalog an optimal aperture selection procedure is used to determine the final light curve, as described in Richards et al. (2012b).¹² The ASAS-3 V-band light curves for the eight new RCB stars and DYPers are shown in Figure 3.

Zoom In Zoom Out Reset image size

Optical spectra of the candidate RCB stars were obtained between 2011 September and 2012 May with the Kast spectrograph on the Lick 3 m Shane telescope on Mt. Hamilton, California (Miller & Stone 1993), the Low-Resolution Imaging Spectrometer (LRIS) on the 10 m Keck I telescope on Mauna Kea (Oke et al. 1995), and the RC Spectrograph on the SMARTS 1.5 m telescope at the Cerro Tololo Inter-American Observatory (Subasavage et al. 2010). All spectra were obtained via long-slit observations, and the data were reduced and calibrated using standard procedures (e.g., Matheson et al. 2000; Silverman et al. 2012). On each night of observations, we obtained spectra of spectrophotometric standards to provide relative flux calibration for our targets. For queue-scheduled observations on the RC spectrograph, all observations in a single night are conducted with the slit at the same position angle. Thus, the standard stars and targets were not all observed at the parallactic angle, leading to an uncertain flux correction (Filippenko 1982). We note, however, that the uncertainty in the flux correction does not alter any of the conclusions discussed below. A summary of our observations is given in Table 3, while the blue portion of the optical spectra are shown in Figures 4 and 5.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

While the unique photometric behavior of RCB stars makes them readily identifiable in well-sampled light curves taken over the course of several years, there are several examples of high-amplitude variables being classified as RCB stars that are later refuted by spectroscopic observations. Most of the misidentified candidates are either cataclysmic, symbiotic, or semi-regular variables (see, e.g., Lawson & Cottrell 1990; Tisserand et al. 2008). RCB stars are a subclass of the HdC stars. For an RCB candidate to be confirmed as a true member of the class, its spectrum must show the two prominent features of HdC stars: anomalously strong carbon absorption and a lack of atomic and molecular H features.

To confirm the RCB candidates found in the ACVS, we obtained low-resolution spectra of the 15 candidates presented in Section 2. Candidates observable from the northern hemisphere were observed with Kast and LRIS, while those only accessible from the southern hemisphere were observed with the RC spectrograph. For some of the southern hemisphere targets very low resolution spectra were obtained first to confirm the presence of C₂ before slightly higher resolution observations were obtained (see Table 3).

We searched the spectra for the presence of strong carbon features, primarily C₂ and CN, and a lack of Balmer absorption to confirm the RCB classification for the ACVS candidates. We find these characteristics in eight of the spectroscopically observed stars (see Figures 4 and 5), which we consider good RCB and DYPer candidates as summarized in Table 4. The remaining candidates were rejected as possible RCB stars based on their spectra, which typically showed strong TiO and VO absorption or clear evidence for H. The properties of the rejected candidates are summarized in Table 5. In the remainder of this paper, we no longer consider these stars candidates and restrict our discussion to the eight good candidates listed in Table 4.

In addition to the hallmark traits of overabundant carbon and a lack of Balmer absorption, RCB stars show a number of other unique spectroscopic characteristics. In particular, they show a very high ratio of ¹²C/¹³C and no evidence for G-band absorption. To search for the presence of ¹³C, we examined the spectra for the λ4744 band head of ¹²C¹³C, which is typically very weak or absent in the spectra of RCB stars. We find evidence for ¹²C¹³C in ASAS 191909−1554.4, ASAS 162232−5349.2, ASAS 065113+0222.1, and ASAS 182658+0109.0 while ASAS 162232−5349.2 shows possible evidence for the ¹³C¹³C band at λ4752. The presence of ¹³C suggests that these four stars are likely DYPers. We consider these four stars closer analogs to DY Per and the DYPers found in the LMC and SMC (Alcock et al. 2001; Tisserand et al. 2009) than they are to classical RCB stars. One of the DYPers, ASAS 065113+0222.1, shows weak evidence for CH λ4300 (G band) absorption and possible evidence for Hγ, which is sometimes seen in the spectra of DYPers. We note that the signal-to-noise ratio (S/N) of all our spectra in the range between ∼4300–4350 is relatively low, making definitive statements about the presence or lack of both CH and Hγ challenging. Finally, we note that we see evidence for the Merrill-Sanford bands of SiC₂ in three of our candidates: ASAS 162232−5349.2, ASAS 065113+0222.1, and ASAS 182658+0109.0. To our knowledge this is the first identification of SiC₂ in a DYPer spectrum, though the presence of this molecule should not come as a surprise as RCB stars are both C and Si rich (Clayton 1996).

In addition to spectral differences, RCB stars and DYPers show some dissimilarities in their photometric evolution as well. The first-order behavior is the same: both show deep, irregular declines in their light curves which can take anywhere from a few months to a several years to recover to maximum brightness. Beyond that generic behavior, however, the shape of the decline tends to differ: RCB stars show fast declines with slow recoveries whereas DYPers tend to show a more symmetric decline and recovery.

The photometric properties of our candidates, including decline rates for the most prominent and well-sampled declines, are summarized in Table 4. As previously noted in the caption of Figure 1, the full amplitude of the variations of these stars are likely underestimated due to the large point spread function (PSF) on ASAS images. This means that the decline rates should be treated as lower limits, since the true brightness of the star may be below that measured in a large aperture. Nevertheless, the decline rates for the four RCB stars are relatively fast and consistent with those given in Tisserand et al. (2009) for RCB stars in the MCs, ∼0.04 mag day⁻¹. The most telling feature of the light curves is the shape of the declines, however. For the four spectroscopic RCB stars, ASAS 170541−2650.1, ASAS 162229−4835.7, ASAS 165444−4925.9, and ASAS 203005−6208.0 the declines are very rapid. While we do not detect ASAS 165444−4925.9 after its sharp decline on TJD ≈ 5000, the other three show slow asymmetric recoveries to maximum light. The four spectroscopic DYPers generally show a slower decline with a roughly symmetric recovery, though we note that the full recoveries of ASAS 065113+0222.1 and ASAS 162232−5349.2 are not observed.

All RCB stars are variable near maximum light, with most and possibly all of the variations thought to be due to pulsation (Lawson et al. 1990; Clayton 1996). Typical periods are ∼40–100 days, and the amplitudes are a few tenths of a magnitude. The pulsational properties of DYPers are not as well constrained because the sample is both small and only recently identified. Each of the four DYPers identified in Alcock et al. (2001) shows evidence for periodic variability near maximum light, with typical periods of ∼100–200 days.

To search for the presence of pulsations in our candidate RCB stars, we use a generalized Lomb–Scargle periodogram (Lomb 1976; Scargle 1982; Zechmeister & Kürster 2009) to analyze each star (see Richards et al. 2011 for more details on our Lomb–Scargle periodogram implementation). Our analysis only examines data that are well separated from decline phases, and we focus on the portions of light curves where the secular trend is slowly changing relative to the periodic variability. For each star we simultaneously fit for the harmonic plus linear or quadratic long-term trend in the data. The frequency that produces the largest peak in the periodogram, after masking out the one-day alias, is adopted as the best-fit period.

We find evidence for periodicity in the light curves of each star, except for ASAS 162229−4835.7. For most of the observing window ASAS 162229−4835.7 was in or near a decline, and we predict that additional observations of ASAS 162229−4835.7, be they historical or in the future, will show periodic variability near maximum light. The trend-removed, phase-folded light curves of the remaining seven stars are shown in Figure 6. Insets in each panel list the range of dates that were included in the Lomb–Scargle analysis, as well as the best-fit period for the data.

Zoom In Zoom Out Reset image size

Some RCB stars are known to have more than a single dominant period (see Clayton 1996 and references therein). We find evidence for multiple periods in ASAS 191909−1554.4, with periods of 120, 175, and 221 days that appear to change every ∼1–2 years. Evidence for multiple periods also appears to be present in ASAS 165444−4925.9. The best-fit periods in this case are 27 and 56 days, which differ by roughly a factor of two. The longer period may in this case simply be a harmonic of the shorter period. Finally, we note that the best-fit period for ASAS 162232−5349.2, 359 days, is very close to one year, and it is possible that the data are beating against the yearly observation cycle. The folded light curve appears to traverse a full cycle over ∼half the full phase cycle. The slight upturn in the folded data around phase 0.15 suggests that the true period is likely ∼180 days, half the best-fit period.

All RCB stars are known to have an IR excess due to the presence of circumstellar dust (Feast 1997; Clayton 1996), and all the known DYPers in the MCs also show evidence for excess IR emission (Tisserand et al. 2009). To check for a similar excess in the new ACVS RCB stars and DYPers, broadband spectral energy distributions (SEDs) were constructed with catalog data obtained from USNO-B1 (Monet et al. 2003), the Two Micron All Sky Survey (2MASS; Skrutskie et al. 2006), the Wide-Field Infrared Survey Explorer (WISE; Wright et al. 2010), AKARI (Murakami et al. 2007), the Midcourse Space Experiment (MSX; Mill et al. 1994), and the Infrared Astronomical Satellite (IRAS; Neugebauer et al. 1984).¹³ The USNO-B1 catalog contains measurements made on digital scans of photographic plates corresponding roughly to the B, R, and I bands. Repeated B and R plates were taken typically more than a decade apart. To convert the five separate USNO-B1 magnitude measurements to the standard g'r'i' system of the Sloan Digital Sky Survey (SDSS; Fukugita et al. 1996), we invert the filter transformations from Monet et al. (2003; see also Sesar et al. 2006). The two measurements each for the g' and the r' band are then averaged to get the reported SDSS g' and r' magnitudes, unless the two measurements differ by >1 mag, in which case it is assumed that the fainter observation occurred during a fading episode of the star. Then the final adopted SDSS magnitude is that of the brighter measurement. There is a large scatter in the transformations from USNO-B1 to SDSS (Monet et al. 2003; Sesar et al. 2006), which leads us to adopt a conservative 1σ uncertainty of 40% in flux density on each of the transformed SDSS flux measurements. The 2MASS magnitude measurements are converted to fluxes via the calibration of Cohen et al. (2003) and the WISE magnitudes are converted to fluxes via the calibration in Cutri et al. (2011). The remaining catalogs provide flux density measurements in Jy rather than using the Vega magnitude system.

The full SEDs extending from the optical to the mid-infrared for each of the new RCB stars and DYPers are shown in Figure 7. All of the candidates but ASAS 203005−6208.0 saturate the W1 (3.5 μm) and W2 (4.6 μm) bands of WISE, while of those all but ASAS 170541−2650.1 and ASAS 165444−4925.9 saturate W3 (11.6 μm) as well. ASAS 191909−1554.4 and ASAS 182658+0109.0 saturate all three of the 2MASS filters and are the only candidates detected at either 60 and/or 100 μm by IRAS. ASAS 162229−4835.7 and ASAS 191909−1554.4 were the only candidates detected at 90 μm by AKARI.

Zoom In Zoom Out Reset image size

Using RCB stars and DYPers in the MCs, Tisserand et al. (2009) find that the RCB stars typically have SEDs with two distinct peaks, whereas DYPers typically have a single peak. It is argued in Tisserand et al. that the SEDs of both can be understood as emission from a stellar photosphere and surrounding dust shell; the cooler photospheric temperatures of DYPers are less distinct relative to the dust emission leading to a single broad peak rather than two. We caution that the reddening toward each of the new Galactic candidates is unknown, which makes a detailed analysis of their SEDs challenging. Furthermore, the observations were not taken simultaneously in each of the various bandpasses. Nevertheless, a few interesting trends can be gleaned from the data. The four stars that spectroscopically resemble RCB stars, ASAS 170541−2650.1, ASAS 162229−4835.7, ASAS 165444−4925.9, and ASAS 203005−6208.0 all show clear evidence for a mid-infrared excess relative to their optical brightness. The peak of emission from ASAS 191909−1554.4 and ASAS 182658+0109.0 is not well constrained because they saturate the detectors between 1 and 6 μm, yet interestingly both show evidence for an infrared excess redward of 50 μm. This suggests that there might be some very cool (T < 100 K) dust in the circumstellar environment of these stars, which is observed in some of the bright, nearby RCB stars. For instance, Spitzer and Herschel observations of R CrB show evidence for a large, ∼4 pc, cool, T ∼ 25 K, and diffuse shell of gas that is detected in the far-IR (Clayton et al. 2011). ASAS 162232−5349.2 and ASAS 065113+0222.1 show evidence for a single broad peak in their SED occurring around ∼2 μm, which is similar to the SEDs of the DYPers observed in the MCs.

An overlap in the survey fields between 2MASS and the DEep Near Infrared Souther Sky Survey (DENIS; Epchtein et al. 1994) allows measurements of the NIR variability of four of the newly discovered RCB stars and DYPers. Photometric measurements from 2MASS and the two epochs of DENIS observations for these stars are summarized in Table 6. Unfortunately, the 2MASS and DENIS observations proceeded the ASAS monitoring, and so we cannot provide contextual information such as the state of the star (near maximum, on decline, during deep minimum, etc.) at the time of the NIR observations. To within a few tenths of a magnitude, ASAS 162229−4835.7 is not variable between the 2MASS and DENIS observations. On timescales of a few weeks to months both ASAS 162232−5349.2 and ASAS 170541−2650.1 show evidence for variations ≳1 mag in the NIR. Similar variations have been observed for several of the RCB stars and DYPers in the MCs (e.g., Tisserand et al. 2004, 2009). The largest variations were observed in ASAS 203005−6208.0, which changed by ∼4 mag in the J band during the ∼4 yr between the DENIS and 2MASS observations. ASAS 203005−6208.0 also shows a large variation between the DENIS I-band measurement and the I-band measurement from USNO-B1, with Δm ≈ 7 mag. This star is clearly a large-amplitude variable, which likely explains its unusual SED. In the DENIS observations, which provide simultaneous optical and NIR measurements, ASAS 203005−6208.0 is always fainter in the optical, suggesting that the unusual shape to its SED (see Figure 7) is the result of non-coeval observations.

This star was first identified as a variable source on Harvard photographic plates with the name Harvard Variable 4368 and was cataloged as a likely long-period variable based on the large amplitude of variations from 13.9 mag to below the photographic limit of ∼16.5 mag (Swope 1928). It was later named GV Oph in the General Catalog of Variable Stars (GCVS; Kukarkin et al. 1971; Samus et al. 2008) as a variable of unknown type with rapid variations. The light curve, spectrum, and SED of this star are consistent with it being an RCB star.

This star is listed as a Mira variable in the GCVS with the name IO Nor. In Clarke et al. (2005) it is identified as a star with an IR excess based on MSX observations. On the basis of its NIR and mid-IR colors, it is identified as an RCB candidate in Tisserand (2012) and considered a likely RCB star on the basis of its ASAS light curve.¹⁴ We independently identified IO Nor as a likely RCB star on the basis of its light curve (in the MACC it is the most likely RCB in ACVS), and our spectrum confirms that it is a genuine RCB star. The previous classification as a Mira variable is likely based on the late spectral type and large amplitude of variability, but Figure 3 clearly shows that ASAS 162229−4835.7 is not a long-period variable.

The variability of this star has not been cataloged to date, and it is listed in the CGCS as C* 2377 (Alksnis et al. 2001). The spectrum, SED, and pulsations exhibited by this star are consistent with RCB stars. There may be evidence for weak CH absorption, though we caution that the S/N is low near ∼4300 Å. The light curve shows a sharp decline, similar to RCB stars, but the recovery is not observed. Nevertheless, the evidence points to ASAS 165444−4925.9 being a new member of the RCB class.

This star was first identified as variable by Luyten (1932) with a maximum brightness of 14 mag and a minimum >18 mag. Luyten assigned it the name AN 141.1932, and it was later cataloged as a possible variable star in the GCVS as NSV 13098. The light curve, spectrum, and SED are all consistent with an RCB classification. Higher resolution and S/N spectra are needed to confirm if H absorption is present, though we note that some RCB stars do show evidence of H in their spectra (e.g., V854 Cen; Kilkenny & Marang 1989), leading us to conclude that ASAS 203005−6208.0 is an RCB star.

This star is listed as a slow irregular variable of late spectral type with the name V1942 Sgr in the GCVS. It is the brightest star among our candidates, and as such it is one of the best-studied carbon stars to date. According to SIMBAD ASAS 191909−1554.4 is discussed in more than 50 papers in the literature. In the CGCS it is listed as C* 2721. ASAS 191909−1554.4 is detected by Hipparcos (HIP 94940) and has a measured parallax of 2.52 ± 0.82 mas (Perryman et al. 1997). This corresponds to a distance d = 397 ± 115 pc and a distance modulus μ ≈ 8 mag. ASAS 191909−1554.4 is one of the few Galactic carbon stars with a measured parallax, and it is important for constraining the luminosity function of carbon stars (Wallerstein & Knapp 1998). In their spectral atlas of carbon stars, Barnbaum et al. (1996) identify ASAS 191909−1554.4 as having a spectral type of N5+ C₂5.5. Relative abundance measurements by Abia & Isern (1997) show that ¹²C/¹³C =30, which is low relative to classical RCB stars. The proximity of ASAS 191909−1554.4 allows its circumstellar dust shell to be resolved in IRAS images (e.g., Young et al. 1993), and Egan & Leung (1991) use ASAS 191909−1554.4 and other carbon stars with resolved dust shells and 100 μm excess to statistically argue that each of these stars must be surrounded by two dust shells, one that is old, ∼10⁴ yr, and the other that is produced by a current episode of mass loss. Recent H i observations by Libert et al. (2010) have shown evidence for the presence of H in the circumstellar shell of ASAS 191909−1554.4. The shallow, symmetric fade of the light curve, along with the N-type carbon star spectrum and the presence of ¹³C in the spectrum, leads us to conclude that ASAS 191909−1554.4 is a DYPer. This is the only candidate within our sample for which we can measure the absolute magnitude, since we have the Hipparcos parallax measurement. Adopting a maximum light brightness of V_max = 6.8 mag, we find that ASAS 191909−1554.4 has M_V ≈ −1.2 mag. This is roughly 0.4 mag fainter than the faintest DYPers in the MCs (Tisserand et al. 2009), suggesting that either the luminosity function extends fainter than that observed in the MCs or there is unaccounted for dust extinction toward ASAS 191909−1554.4.

The variability of this star has not been cataloged to date, and it is listed in the CGCS as C* 2322. The relatively slow, symmetric decline and recovery in the light curve of ASAS 162232−5349.2 lead us to classify it as a DYPer variable. The presence of ¹³C in the spectrum and the single peak in the SED support this classification.

The variability of this star has not been cataloged to date, and it is listed in the CGCS as C* 596. The presence of ¹³C in the spectrum and the single peak in the SED lead us to classify ASAS 065113+0222.1 as a DYPer variable.

The variability of this star has not been cataloged to date, and it is listed in the CGCS as C* 2586. Based on weekly averages of DIRBE NIR observations taken over 3.6 yr, Price et al. (2010) list ASAS 182658+0109.0 as a non-variable source. Low-resolution spectra taken with IRAS show 11 μm SiC dust emission, which typically indicates significant mass loss from a carbon star (Kwok et al. 1997). The light curve shows a ∼5–6 yr symmetric decline, and there is clear evidence for ¹³C in the spectrum. The IRAS detection at 60 μm shows a clear IR excess relative to a single temperature blackbody. While there is no evidence for Hα, the S/N in our spectrum is low blueward of ∼4700 Å. We consider ASAS 182658+0109.0 a likely DYPer, though higher S/N spectra are required for a detailed abundance analysis to confirm this classification.

As mentioned in Section 2.1, one of the major strengths of ML classification is that new discoveries may be fed back into the machinery in order to improve future iterations of the classifier. In an attempt to recover more RCB stars and DYPers in the ACVS that were missed in our initial search of the MACC, we created an augmented RCB training set by adding the eight new RCB stars and DYPers identified in this paper to the 17 sources already included in the training set. This augmented training set should increase the likelihood of discovering new candidates, particularly DYPers, of which there were no examples in the original training set.

Using the augmented training set we re-ran the RF classifier from Richards et al. (2012b) on all the ACVS light curves to search for any additional good candidates. We focus our new search on candidates with a significant change in R(RCB), which were not examined in the initial search of the MACC. In particular, we visually examine the light curves of all sources with R(RCB)_new < 500, P(RCB)_MACC < 0.1, and R(RCB)_new < R(RCB)_MACC. There are a total of 96 sources that meet these criteria, which were not included in the 472 visually inspected sources from the original MACC. Of these 96, we conservatively select 6 as candidate RCB stars or DYPers. One is a highly likely RCB star with multiple declines and asymmetric recoveries, three show evidence for a single decline that is only partially sampled, and two are known carbon stars that are likely semi-regular periodic variables. We list the candidates in Table 7 with brief comments on each and show their light curves in Figure 8.

Zoom In Zoom Out Reset image size

Restricting our search for bright RCB stars to only those sources in the ACVS has biased the results of our search. As was mentioned in Section 2.2, there are seven known RCB stars that show clear variability in their ASAS light curves yet were not selected for inclusion in the ACVS. This suggests that several large-amplitude ASAS variables are missing from the ACVS, presumably including a few unknown RCB stars. This bias can easily be corrected by searching all of ASAS for RCB stars; however, such a search would include significant new challenges as there are ∼12 million sources in ASAS. In addition, our existing classification framework is not designed to deal with a catalog where the overwhelming majority of sources are not in fact variable. Nevertheless, both of these challenges must be addressed prior to the LSST era. We have developed frameworks that can ingest millions of light curves and are currently experimenting with methods to deal with non-variable sources, the results of which will be presented in a catalog with classifications for all ∼12 million sources in ASAS (J. W. Richards et al., in preparation). Furthermore, it has been shown that the use of mid-infrared colors is a powerful discriminant when trying to select RCB stars (Tisserand et al. 2011; Tisserand 2012). While our use of NIR colors is important for selecting RCB stars (see, e.g., Soszyński et al. 2009), adding the mid-infrared measurements from the all-sky WISE survey will dramatically improve our purity when selecting RCB candidates as several of the Mira and semi-regular variables that served as interlopers in the current search (Section 2.3) would be eliminated with the use of mid-IR colors.