Identifying Sagittarius Stream Stars by Their APOGEE Chemical Abundance Signatures

The APOGEE survey was part of Sloan Digital Sky Survey III (Eisenstein et al. 2011), and observed 146,000 stars in the MW galaxy (Majewski et al. 2017) from 2011 to 2014. APOGEE-2 began observations in 2014 as part of the Sloan Digital Sky Survey IV (Blanton et al. 2017), and the latest data release, DR14, contains APOGEE observations of an additional ∼100,000 stars in the Northern Hemisphere (Abolfathi et al. 2018). The APOGEE instrument is a high-resolution (R ∼ 22,500) near-infrared (1.51–1.70 μm) spectrograph described in detail in Wilson et al. (2019, in preparation). For the main survey the instrument was connected to the Sloan 2.5 m telescope (Gunn et al. 2006). Targeting strategies for APOGEE and APOGEE-2 are described in Zasowski et al. (2013, 2017), respectively.

The APOGEE data are reduced through methods described by Nidever et al. (2015), and stellar parameters/chemical abundances are extracted using the APOGEE Stellar Parameters and Chemical Abundances Pipeline (ASPCAP, García Pérez et al. 2016). ASPCAP interpolates in a grid of synthetic spectra (Zamora et al. 2015) generated from the APOGEE H-band line list (Shetrone et al. 2015) to find the best fit (through χ² minimization) to the observed spectrum by varying T_eff, surface gravity, microturbulence, metallicity, carbon abundance, nitrogen abundance, and α-element abundance. In this analysis we use results from the 14th data release of SDSS (DR14, Abolfathi et al. 2017; Holtzman et al. 2018; Jönsson et al. 2018).

To identify Sgr stream stars in the DR14 sample, we use the Sgr core sample from Hasselquist et al. (2017) that was selected using the methods described in Majewski et al. (2013) as a representation of the chemical abundance patterns of Sgr. This sample contains 158 stars with spectra having signal-to-noise ratios (S/Ns) > 70 per half resolution element, which the APOGEE DR14 documentation defines as the lower threshold for which the APOGEE detailed chemical abundances are reliably measured. Because the bulk of the Sgr members are in the low-temperature regime (T_eff < 4200 K) where some elemental abundance determinations may exhibit systematic differences from higher temperature stars, we limit our MW sample in which we search for Sgr stream stars to the same stellar parameters of APOGEE Sgr stars. This sample is defined as follows:

• 1.  
  3550 < T_eff < 4200.
• 2.  
  S/N > 70 per half resolution element.
• 3.  
  No "ASPCAPBAD" flag set.²¹

Because the APOGEE Sgr core sample does not contain stars with [Fe/H] < −1.2, we also only consider MW stars with [Fe/H] > −1.2. Distances to the APOGEE stars are derived by the methods described in Hayden et al. (2014), and are accurate to ∼20%. These distances can be found as part of a value-added catalog (VAC) from DR14.²² While the Gaia distances from Bailer-Jones et al. (2018) are more precise at close distances, at d > 5 kpc, the distance uncertainties become larger than ∼20%, and the distances become prior-dominated. More than half of our APOGEE sample on which we perform the clustering analysis consists of stars with distances >5 kpc, so we adopt the spectrophotometric distances from the DR14 VAC to ensure our distance source is homogeneous across the entire distance range we are able to probe with the APOGEE stars.

We supplement the APOGEE radial velocities and distance estimates with proper motions from Gaia Data Release 2 (DR2; Gaia Collaboration et al. 2018). Using the nearest neighbors matching tool within TOPCAT (Taylor 2005), we identify the best Gaia match (nearest neighbor within 1'') for each APOGEE star. Of the 68 chemical candidates we identify in Section 3, 62 have Gaia proper motion matches. All but two of the sources matched within 04 (the other two are at 073 and 086). We then use the full 6D phase-space data (i.e., 3D positions and velocities) to integrate orbits in a model Galactic potential, and further check the membership of our APOGEE-selected sample in the Sgr stream (described in further detail in Section 4).

To further investigate whether we have recovered true Sgr stream stars, we use the proper motions from Gaia and the distances and radial velocities from APOGEE to calculate the orbits of the Sgr stream candidates. Orbits were integrated in the MWPotential2014 potential that is implemented in the galpy (Bovy 2015) suite of Python routines.²³ This potential is a composite of a spheroidal bulge, a Miyamoto–Nagai disk, and a Navarro–Frenk–White halo (see the default parameters in Bovy 2015). We note that while many recent works suggest that the halo mass of MWPotential2014 may be too low to match MW observations (e.g., Fritz et al. 2018; Gaia Collaboration et al. 2018; Watkins et al. 2018), we are mainly interested in the relative orbital properties of stars, so the absolute scale of the potential is unimportant. Throughout this work, we assume a circular velocity of 220 km s⁻¹ at a Solar radius of 8.0 kpc from the Galactic center, and the Solar motion as measured by Schönrich et al. (2010). We follow the galpy usage of a left-handed Galactic Cartesian coordinate (velocity) system with X(U) positive toward the Galactic center, Y(V) along Galactic rotation, and Z(W) in the direction of the north Galactic pole. We initialize the orbit of each star at its current position and velocity, then integrate its orbit both forward and backward for 1.5 Gyr. From this, we estimate the orbital parameters, including the eccentricity, the pericenter and apocenter distances (r_peri and r_apo; relative to the Galactic center), and orbital energy and angular momentum.

The goal of this work is to find Sgr members through chemistry alone, i.e., without any knowledge of kinematics or positions. However, to verify that the clustering algorithm used in this work is properly returning Sgr stream members by recovering the most obvious Sgr stream members, we select a "kinematic" control sample of likely Sgr stream stars from the APOGEE data. Sgr stream stars in the direction out of the MW plane can be easily distinguished from MW stars based on velocities and color.

The method for selecting the kinematic control sample is as follows:

• 1.  
  We begin by applying cuts on distance, color, extinction, and proper motion to the entire APOGEE DR14 catalog (using distances from the VAC's "NMSU_dist" estimator). We transform the coordinates into the Sagittarius stream-aligned system devised by Majewski et al. (2003),²⁴ and only select stars with $\left|B\right|\lt 20^\circ$ (must be within 20° of the Sgr debris plane). To avoid nearby disk stars, we require d > 5 kpc. We further limit the sample to red objects with ${(J-K)}_{0}\gt 0.8$, with low total extinction $E(B-V)\lt 0.75$, and minimal proper motion (to remove nearby high proper motion stars) $\left|{\mu }_{\alpha \cos \delta },{\mu }_{\delta }\right|\lt 15$ mas yr⁻¹. The resulting sample is plotted as magenta points in all three panels of Figure 1.
• 2.  
  From this set of APOGEE candidates, we use the known properties of the Sgr stream to select candidates (also retaining only stars whose APOGEE spectra have S/N > 30). The selections are based on velocity trends derived from SDSS members (Belokurov et al. 2014, dashed lines in the lower panel of Figure 1), and RR Lyrae distances from PanSTARRS PS1 (Hernitschek et al. 2017, dashed lines in center panel). For each of these sub-selections, we include a range of ±3σ about the mean values from each study, using their reported 1σ errors. This yields a final sample of 58 kinematically selected Sgr stream candidates, which are shown as yellow/orange diamonds in Figure 1. Of these 58 stars, 21 have sufficient S/N and stellar parameters set by our criteria above, which we adopt as the "kinematic" sample.

Zoom In Zoom Out Reset image size

This is not a complete sample of Sgr candidates from the APOGEE DR14 database, as the selection criteria are limited to regions with previous measurements. However, it is a reliable set of candidates to use as a test of our chemical tagging selection. We have confirmed that these selection criteria recover the 42 Sgr stream stars observed at high spectral resolution in the optical by Carlin et al. (2018).

To conduct our search for Sgr stream members in chemical abundance space, we use the k-means "CLUST_WTS" function²⁵ in IDL. This function randomly chooses the starting points of k clusters and then moves stars in and out of clusters to minimize variance within each cluster and maximize variance between clusters. We set this function to run with 100 different random starting guesses, which we find consistently returns the same clusters for k < 50. We choose the APOGEE elemental abundances in which Sgr exhibits clear deficiencies, and have low reported uncertainties from the APOGEE pipeline (∼0.05 dex or less). The N-dimensional space in which we cluster is then defined by [O/Fe], [(C+N)/Fe], [Mg/Fe], [Al/Fe], [Mn/Fe], and [Ni/Fe]. The uncertainties for each element are shown in Figure 2. The median uncertainties of each element are used as weights in the k-means clustering analysis such that elements with lower uncertainties, such as Ni and O, are given larger weight than elements with higher uncertainties, such as Al and C+N.

Zoom In Zoom Out Reset image size

With a kinematic sample of expected stream members and the core sample defining the abundance patterns of Sgr, we determine the number of clusters, k, in which to separate the data. To select an appropriate value for k, we run the clustering algorithm for multiple values of k from 3 to 50. For each value of k, we track the Sgr core sample from Hasselquist et al. (2017) and the kinematic control sample defined above. The cluster that contains the bulk of the Sgr core sample is referred to as the "Sgr cluster," and should contain the majority of the kinematic control sample if we are to identify potential stream members. We choose the ideal value of k to be the number of clusters that results in a marginal splitting (10%) of the Sgr core sample from the Sgr cluster. In this way, we are "tuning" the k-means algorithm.

The motivation for our choice of k is shown graphically in Figure 3, where we plot the fraction of stars retained in the Sgr cluster as a function of k for each sample. The blue line, which represents the fraction of Sgr core members retained in the Sgr cluster, exhibits a decrease at k = 16, which corresponds with a similar decrease in the fraction of kinematic sample retained (red line). We adopt k = 16 as our best number of clusters in which to separate the data, but note that 18 and 19 are just as viable choices based on our criteria set here. However, these choices result in very little difference in the amount of "chemical" stream candidates; the fraction of the Sgr cluster that is made up of stream candidates (stars that are not Sgr core members) is denoted by the green line in Figure 3.

Zoom In Zoom Out Reset image size

We find that 35 stars (including 19 stars from the kinematic control sample) of the 62 chemical candidates with Gaia proper motions follow the expected positions and kinematics of the Law & Majewski (2010) model of the system. Of these, 21 stars belong to the trailing arm and 14 stars belong to the leading arm, including all ∼9 of the low-eccentricity "maybe Sgr" stars. This is indicated by the coloring in the top right panel of Figure 6. According to this model, these stars were all stripped in the most recent pericenter passage. In Figure 6 we find that the trailing debris exhibit a range of apocenters from 60 to 120 kpc, and the leading arm debris have apocenters of 40–60 kpc. These are more in line with the apocenters found by Belokurov et al. (2014) and Fardal et al. (2018; ∼48 kpc and ∼100 kpc, respectively).

We find Sgr stream stars as metal-rich as [Fe/H] = −0.2, with the majority of stream stars exhibiting [Fe/H] < −0.6. The mean metallicity of our Sgr stream sample is more metal-poor than the Sgr core sample, suggesting that there was a radial metallicity gradient across the Sgr progenitor when it began merging with the MW such that the more metal-poor stars were stripped first.

In Figure 6, we define the group of stars that have highly eccentric orbits and median r_apo = 20.2 kpc as the "accreted halo" stars. These stars are similar to the accreted halo stars described in the literature (e.g., Schuster et al. 2012; Belokurov et al. 2018; Deason et al. 2018), where multiple studies concluded that the pileup of metal-rich ([Fe/H] > −1.5) halo stars at apocenters of ∼20–25 kpc originates from a major merger some 8–11 Gyr ago. They argued that a massive satellite interacting with the disk can deposit stars on highly radial orbits (see, e.g., Belokurov et al. 2018). Further support for this interpretation can be found elsewhere in the literature (see, e.g., Helmi et al. 2018; Kruijssen et al. 2018; Mackereth et al. 2019, and Chiba & Beers 2000; Brook et al. 2003 for earlier reports), although we explore other potential origins of these stars in Section 5.2.

The accreted halo stars are unlikely to be Sgr stream stars, as they generally have $\left|B\right|$ values that put them well outside the predicted plane of Sgr debris. However, given the nature of our analysis, these stars have the same chemistry as the Sgr stream stars. This suggests that the dwarf galaxy that merged with the MW some 8–11 Gyr ago may have a very similar star formation history (SFH) as Sgr, at least until an enrichment of [Fe/H] ∼ −0.6, the highest metallicity among "accreted halo" stars.

There are four Sgr chemical candidates that actually reside in the Galactic bulge and are on bulge-like orbits. Thus, they are unlikely to be actual stream members. We find that by increasing k clusters, we cannot remove these stars from the Sgr chemical candidate cluster without removing likely stream members, suggesting a potential limitation to using the k-means technique to conduct chemical tagging analysis.

These stars are likely grouped with Sgr stream stars due to their exceptionally deficient [O/Fe] abundance ratios, as well as deficient [Mg/Fe] and [Al/Fe], as seen in Figure 7. They may not be Sgr stream stars, but they also are not typical bulge stars. In Figure 8 we compare the bulge stars (red) to a strictly "bulge" MW sample, which is a subsample of our previous MW comparison sample (MW stars with the same parameters as the Sgr core sample), but with R_Gal < 3 kpc. While there are some bulge stars with similarly deficient abundance ratios, the bulge Sgr stream candidates are deficient by ∼0.5 dex in [O/Fe], by ∼0.4 dex in [Mg/Fe], and by ∼0.7 dex in [Al/Fe] compared to the bulk of the bulge stars. They appear to have the same [Mn/Fe], and are slightly deficient in [(C+N)/Fe].

Zoom In Zoom Out Reset image size

The bulge Sgr stream chemical candidates actually have [O/Fe] ∼ 0.2 dex lower than the actual Sgr stream candidates. They are also enhanced in Ni and Mn relative to the Sgr stream candidates, with Ni and Mn abundances that more closely track the other stars in the bulge. Two of these stars actually have enhanced [Ni/Fe] relative to the bulge stars. This suggests that these may be some class of stars that formed from material containing excess SNe Ia yields as compared to the majority of the MW bulge stars.

While determining the origin of these stars is beyond the scope of this work, the presence of these stars in the Galactic bulge suggests that either there was an epoch of star formation in the bulge that formed stars from gas that was unusually enriched in SNe Ia, or the MW underwent a merger of a dwarf galaxy that was as metal-rich as Sgr, but formed stars from gas that had a higher ratio of Type Ia/Type II ejecta than Sgr.

One star in the Sgr chemical sample is at a Galactocentric distance of ∼54 kpc, near the position of the most recent apocenter of the LM10 Sgr orbit, which happened ∼450 Myr ago. Contrary to expectations for tidal debris at apocenter, this star has a total Galactocentric space velocity of $\sim {577}_{-129}^{+209}$ km s⁻¹. In fact, an orbit integration suggests that this star's closest approach will occur in ∼20 Myr, at r_peri ∼ 53 kpc, and that it is completely unbound from the MW. If this HVS's orbit does not trace back to the Galactic center, then what might its origin be? The backward integration of its orbit carries it from its current location near the Galactic anticenter in the third quadrant of the North Galactic Hemisphere, meaning that its orbit passes nowhere near the LMC or M31. Thus, if this star is an interloper from beyond the MW, its origin will require further study.

Given that this HVS candidate is an early M-giant, it is unlikely to originate as a "runaway" in a binary ejection scenario (e.g., Blaauw 1961), because this formation mechanism requires young, massive stars. We have already ruled out scenarios involving ejection from a black hole at either the center of the MW, M31, or the LMC. Thus, the most likely origin of this distant HVS is a dynamical interaction (e.g., between the components of a triple system, or between a binary and a passing massive star; e.g., Leonard & Duncan 1990; Gvaramadze et al. 2009). However, these mechanisms often require a very massive star to perturb the system; given the old, metal-poor nature of the Sgr debris (and the surrounding halo stars) in the region of sky where the HVS candidate is located, this mechanism also seems unlikely to have created this HVS candidate.

There is some support in the literature for HVSs originating from mergers. Abadi et al. (2009) found that disrupting dwarf galaxies can contribute HVSs; however, the mechanism requires that the HVSs were stripped from their host dwarf on a recent pericentric passage (i.e., near the Galactic center). Given that the star we have found never goes within ∼50 kpc of the Galactic center, this scenario is unlikely. Similarly, Piffl et al. (2011) suggested that massive dwarf galaxy mergers (M > 10⁹ M_⊙) can result in a population of HVS that is essentially the high-velocity tail of the debris. However, at the apocenter of the Sgr orbit, where we find the HVS candidate, the mean Galactocentric velocity should be roughly zero. A high-velocity tail relative to the mean velocity is unlikely to produce a star with velocity >600 km s⁻¹. Again, the simulations discussed by Piffl et al. (2011) require the stars to be debris stripped on the most recent pericentric passage, which seems to be ruled out because our candidate HVSś orbit does not trace back along the Sgr orbit.

It is possible that this candidate HVS simply has a systematic error in its proper motion, which contributes most of the star's 3D space velocity (though its v_gsr = −111 km s⁻¹ alone is much larger than expected at the Sgr orbital apocenter). This star's Gaia proper motions are (μ_αcosδ, μ_δ) = (−1.91, −0.10) ± (0.62, 0.44) mas yr⁻¹, and its distance error is ∼15% (the radial velocity error from APOGEE is negligible). We integrated orbits for the min/max space velocities based on the uncertainties in all measured parameters, and in neither case was the orbit bound to the MW. Nonetheless, if the proper motions are systematically offset, rather than a large random error, we must await future improvements in Gaia proper motions to address this question. Future investigation into the chemistry of HVSs across the Galaxy may shed light on the likely origin of these stars.

Using this k-means clustering technique, we find that 35 of the 62 Sgr stream candidates with reliable Gaia proper motions are likely Sgr stream members. An additional 20 stars are accreted halo stars with indistinguishable abundance patterns from the Sgr tidal debris. Through our "weak" chemical tagging technique, we are able to recover 55 out of 62 accreted dwarf galaxy stars in the MW halo, identified by clustering on chemical space defined by six chemical elements. Because we do not have a good idea of what the chemical abundance patterns of the Sgr core stars with [Fe/H]  < −1.2 in the APOGEE sample look like as compared to the MW, we did not extend down our analysis. However, the Sgr stream contains stars at least as metal-poor as [Fe/H] = −1.6 (Chou et al. 2007; Carlin et al. 2018), and the core contains stars as metal-poor as [Fe/H] ∼ −2.0 (Mucciarelli et al. 2017; Hansen et al. 2018). Future observations of the Sgr core using APOGEE-2 that target more metal-poor K giants will help demonstrate whether this technique can separate Sgr core stars from the MW stars across a much wider metallicity range to fully discover Sgr stream stars.

Thus far, we have only analyzed 1 cluster out of 16. While a comprehensive analysis of each individual cluster is beyond the scope of this paper, we did search the other clusters for the more metal-poor "accreted halo" population, identified in the APOGEE sample by Hayes et al. (2018) and Mackereth et al. (2019). We analyzed the [Mg/Fe] versus [Fe/H] abundance space for each individual cluster and found that one cluster was apparently dominated by the "Low-Mg" population ("LMg") identified and described by Hayes et al. (2018) (but also described in APOGEE by Hawkins et al. 2015 and Fernández-Alvar et al. 2017) as halo stars that exhibit relatively low [Mg/Fe] abundances and halo-like kinematics. This is shown in the left panel of Figure 9.

Zoom In Zoom Out Reset image size

Mackereth et al. (2019) was able to recover this population by conducting a k-means clustering analysis on five-dimensional space defined by four chemical elements and orbital eccentricity. In the right panel of Figure 9, we show that the LMg population that naturally occupies one of our clusters does indeed exhibit high orbital eccentricity and median r_apo = 20.8 kpc, which are both characteristics of the "accreted halo" population thought to dominate the inner halo (Belokurov et al. 2018; Deason et al. 2018; Helmi et al. 2018). We show in Figure 10 that these stars are not in a particular region of the sky, nor are they obviously spatially clustered, as described by Deason et al. (2018). The recovery of this population suggests that the chemical tagging technique applied in this work is useful for identifying stars formed in relatively massive dwarf galaxies. The APOGEE IDs for the stars that fall into our LMg cluster are presented in Table 2.

Zoom In Zoom Out Reset image size

Because we find no stream stars with [Fe/H] > −0.2, which are found in abundance in the APOGEE Sgr core sample, we offer a constraint on how chemically evolved Sgr was during its most recent pericenter passage. The SFH of the Sgr core from Siegel et al. (2007) suggests that these stars formed ∼2 Gyr ago, which might imply that we may expect to see some of these stars stripped as we are looking at debris that Law & Majewski (2010) declared was stripped as early as 1 Gyr ago. However, ff the starburst that formed this population with [Fe/H] > −0.2 only occurred in the very central regions some 2 Gyr ago, then we might not see any in the stream structure.

Alternatively, Tepper-García & Bland-Hawthorn (2018) argue that Sgr came into the MW gas-rich and was fully stripped of gas during its last disk passage approximately 1 Gyr ago. In this scenario, the most metal-rich stars were formed during a last burst of SF coincident with this passage. The lack of stars with [Fe/H] > −0.2 in our stream sample is also consistent with this scenario. This would require a slight tweak to the Siegel et al. (2007) SFH of Sgr, such that stars with [Fe/H] > −0.2 formed 1 Gyr ago rather than 2–3 Gyr ago.

The "accreted halo" stars, which exhibit the same abundance patterns as Sgr, but are on different orbits, are likely the metal-rich end of the accreted halo population described in Hayes et al. (2018), Deason et al. (2018), Helmi et al. (2018), and Mackereth et al. (2019). We have shown that one of our clusters contains the more metal-poor debris. We find that these "accreted halo" stars, identified through chemistry alone, do indeed exhibit apocenters of ∼20 kpc along with orbital eccentricity >0.8. The chemical abundance patterns of the "accreted halo" stars appear to be indistinguishable from the Sgr core and stream stars at −1.0 < [Fe/H] < −0.6. There are potential differences in the stars with [Fe/H] < −1.0, but we only have a small handful of Sgr stars to compare to.

The indistinguishable chemical abundance patterns suggest that Sgr and the progenitor of the "accreted halo" debris exhibited very similar star formation histories. The most metal-rich of the "accreted halo" stars in our sample have [Fe/H] = −0.6. If we assume that the star formation histories must have been the same, we can use the known SFH of Sgr to time when the progenitor of the "accreted halo" debris merged with the MW. Based on the SFH of Sgr, as presented by Siegel et al. (2007), Sgr did not form stars as metal-rich as [Fe/H] ∼ −0.6 until 8–9 Gyr ago. This is in good agreement with simulations of the "accreted halo" debris, which find that the progenitor was likely an LMC-sized galaxy that merged with the MW some 8–11 Gyr ago (e.g., Belokurov et al. 2018; Mackereth et al. 2019).

However, from our data, we cannot determine whether this accreted halo population is from one major accretion event or a handful of smaller accretion events. While we do find stars as metal-rich as [Fe/H] ∼ −0.6, the mean [Fe/H] of the accreted halo population is much more metal-poor than this. If we assume that the accreted halo stars are indeed the metal-rich extension of the LMg population, then we find a mean metallicity of this population of $\langle [\mathrm{Fe}/{\rm{H}}]\rangle \sim -1.0$. This is an upper limit, as the true mean metallicity is likely lower than this because the LMg population extends to metallicities more metal-poor than our [Fe/H] = −1.2 cutoff (Hayes et al. 2018). Based on the stellar mass–stellar metallicity relationship described by Kirby et al. (2013), a mean metallicity of ≲−1.0 suggests the progenitor galaxy had a stellar mass of ∼10⁷ M_⊙, similar to that of Fornax. While good arguments have been made as to why this progenitor galaxy need be massive (e.g., Belokurov et al. 2018; Mackereth et al. 2019), such a galaxy would fall considerably off the stellar mass–stellar metallicity relationship, being too metal-poor for its mass.