THE STELLAR HALOS OF MASSIVE ELLIPTICAL GALAXIES. II. DETAILED ABUNDANCE RATIOS AT LARGE RADIUS

In Paper I we adopted the SDSS model radius (the de Vaucouleurs fit) as the effective radius (R_e). While there is considerable evidence that the shape of the light profile changes systematically with galaxy mass (e.g., Caon et al. 1993; Kormendy et al. 2009), fitting the galaxies with a fixed Sérsic index of four has the benefit that we are less sensitive to both sky subtraction errors (Mandelbaum et al. 2005; Bernardi et al. 2007) and to the detailed shape of the light profile in the very faint wings (e.g., Lackner & Gunn 2012). In the effort to have a uniform analysis, we have therefore adopted the effective radii published by the SDSS. The galaxy NGC 6482 is not in the SDSS, and we have adopted R_e from NED in this case. Below we will examine bins in physical as well as R_e-scaled radii to mitigate uncertainties in the measurement and meaning of R_e (e.g., Kormendy et al. 2009; Huang et al. 2013a).

Individual spectra, with the exception of those at the very center of the IFU, have inadequate signal for stellar population analysis. Therefore, all of our analysis is performed on binned spectra. We utilize four binning schemes here, in all cases defining elliptical annuli based on the axis ratio measured by the SDSS. First, for the maximum spatial resolution, we create bins with radial width of 4'', the width of an individual fiber. Second, we make bins of width 0.5R_e. In Paper I, these two binning schemes were nearly identical, given that most of the galaxies had effective radii of ∼8''. However, in our larger sample, many of the galaxies have effective radii of 10''–25'', requiring a finer binning scheme. In the end, our highest spatial resolution corresponds to 0.2–0.5R_e depending on the galaxy. Third, we make bins of width R_e, from which we measure σ_*. Fourth, we make bins with fixed physical sizes of 0–15 kpc in 3 kpc increments. For reference, in Table 1 we include the number of spectra that are combined in the 1.5–2R_e bin, and the surface brightness of each galaxy at 2R_e. We typically achieve a signal-to-noise ratio (S/N) of >30 per pixel at a surface brightness brighter than r < 23 mag arcsec⁻².

As in Greene et al. (2012), we use lick_ew (Graves & Schiavon 2008) to measure the Lick indices. First, however, we must correct for low-level emission that can fill in the absorption lines and artificially lower their equivalent widths (EWs). Low-EW emission from warm ionized gas is very common in the centers of elliptical galaxies (Sarzi et al. 2010; Yan & Blanton 2012). In particular, the emission from Hβ is weak in all cases, but even a 0.1 Å error in the Hβ EW can lead to errors of ∼2 Gyr in the modeling (e.g., Schiavon 2007).

In Paper I, we utilized pPXF+GANDALF developed by M. Sarzi (Sarzi et al. 2006) and M. Cappellari (Cappellari & Emsellem 2004) to simultaneously model the stellar absorption and emission lines. Here, instead, we fit each spectrum with an empirical template drawn from the composite spectra of Graves et al. (2010). We then fit the [O iii] emission in the residual spectrum, and subtract both [O iii] and Hβ, assuming that the Hβ emission is 70% of the [O iii] flux (good to within a factor of two, Trager et al. 2000b; Graves et al. 2007). In addition, we fit Gaussians to residuals around strong sky lines at 5200 and 5460 Å, and subtract them.

The strongest emission (≳ 0.2 Å in more than one radial bin) is found in NGC 426, NGC 661, NGC 677, NGC 7509, NGC 7684, and UGC 1382. We found that our GANDALF results were sensitive to template mismatch, and so adopted this iterative approach that gives us more control over the templates but less control over the line strengths. The results are reasonably consistent between the two techniques for sources with Hβ > 0.1 Å in the GANDALF fits: 75% of these are detected in our iterative method, which returns Hβ EWs that are ∼60% weaker than those derived from GANDALF. In the future our goal is to refine the GANDALF measurements using models with a range in [α/Fe] (e.g., Coelho et al. 2007; Vazdekis et al. 2010; Conroy et al. 2013).

We then use lick_ew (Graves & Schiavon 2008) on the emission-line-corrected spectra. This code corrects for both the instrumental and intrinsic velocity dispersion, the latter measured using pPXF. The indices are put onto a modified Lick system presented by Schiavon (2007) based on flux-calibrated spectra. In order to demonstrate that we are on the same system, we compare the Lick indices from the flux-calibrated SDSS spectra (the inner 3'') with those from the central 4'' fiber in our data, but we exclude NGC 219, NGC 426, NGC 677, NGC 1267, and IC 301 from the comparison due to the presence of bright stars near the nucleus. There is no net offset between the two sets of indices in any case, with 〈(EW_MS − EW_S)/EW_S〉 = 0.01 ± 0.09. Hβ and 〈Fe〉 each have a scatter of ∼10% while Mgb and the G-band each have a scatter of ∼5%.

As in Paper I, we use the stellar population modeling code EZ_Ages (Graves & Schiavon 2008) to convert the Lick indices from the composite spectra to physical parameters (age, [Fe/H], [α/Fe]). The code works on a hierarchy of index pairs, starting with Hβ and 〈Fe〉, and iteratively solves for the age, abundance and abundance ratios. The models of Schiavon (2007) include abundance ratio differences using the methodology of Trager et al. (2000b) and the response functions of Korn et al. (2005). In our default runs, we utilize the α-enhanced isochrone from Salasnich et al. (2000) and the default assumption that [O/Fe] = 0.5. We revisit this final assumption in Section 6.1.

4.1.1. Uncertainties

While measuring Lick indices is a very powerful technique at high S/N, at the large radii that we are working systematic effects such as small errors in sky subtraction and flux calibration can cause large uncertainties in the Lick indices measured from individual objects. With our sample size, we benefit from averaging over multiple galaxies at each radial bin. The composite spectra will suffer less from the vagaries of sky subtraction and flux calibration, which occur at different wavelengths in each galaxy rest-frame (e.g., Graves et al. 2009; Yan 2011). We are able to examine radial variations in the composite spectra at the percent level.

We first divide the galaxies into three bins of central σ_* of 150 < σ_* < 200, 200 < σ_* < 250, and 250 < σ_* < 300 km s⁻¹, since we know that the stellar population properties are a strong function of σ_* (e.g., Worthey et al. 1992; Bender et al. 1993; Trager et al. 2000a; Graves et al. 2009). Note that the bins would not change if we used σ_* within R_e, but that in this way our bins are more consistent with the large literature based on the SDSS measurements. We combine spectra that have already been emission-line-corrected. We interpolate the rest-frame spectra onto the same wavelength grid and then smooth each galaxy to the highest dispersion in the stack (300 km s⁻¹ for the high-dispersion bin and 200 km s⁻¹ for the low-dispersion bin). However, to increase the contrast, we will focus exclusively on the highest- and lowest-dispersion bins here. We remove the continuum by dividing each spectrum by a heavily smoothed version of itself. This step simultaneously normalizes all spectra to the same level and ensures that differences in continuum shape (whether real or due to small errors in sky subtraction or flux calibration) do not impact the final line strengths. We then combine all pixels at each wavelength using the biweight estimator, which should be robust even with limited statistics (Beers et al. 1990). We experiment with multiplying the composite spectrum by the median continuum before measuring indices, but the changes to the Lick indices are negligible in all cases.

The composite spectra are shown in Figure 2. Again we focus only the lowest- and highest-dispersion bins, which contain 11 galaxies each (excluding NGC 219 due to severe night sky contamination and NGC 6482 due to contamination from a neighboring galaxy). The spectra are strikingly similar as a function of radius, so we use ratio spectra to highlight the percent-level radial variations (Figure 2, right). Specifically, we divide each composite spectrum, at each radial position, by the composite spectrum at 0.5–1 R_e. We choose this radius, rather than the central bin, because as we will see below the largest variance in spectral properties occurs in the very center. We will examine the radial trends in these spectra in detail in Section 5. For now, we note only that the strongest variations are seen in carbon, Mgb, and perhaps nitrogen.

Zoom In Zoom Out Reset image size

To determine the level of variation in the composite spectra, we generate 100 trial composite spectra by randomly drawing from the total list of galaxies within that σ_* bin, with replacement. We measure Lick indices from each of these 100 trial spectra. We then assign errors on the Lick indices measured from the stack derived to enclose 68% of the Lick indices measured from the 100 trials.

We first investigate the average radial profile in the Lick indices as a function of radius in the high-dispersion (250 < σ_* < 300 km s⁻¹) and low-dispersion (150 < σ_* < 200 km s⁻¹) galaxies. We use Lick indices measured from the composite spectra (Section 4.2) and plot them as filled points as a function of radius in Figure 3. As a check on the composite spectra, we also calculate the median indices at each radius from the individual galaxy measurements, divided into the same high- and low-dispersion groups. These median profiles are shown as lines in Figure 3. The consistency between the composite and median measurements gives us confidence in the measurements from our composite spectra.

Zoom In Zoom Out Reset image size

We show radial profiles both as a function of radius scaled to R_e and in physical units of kpc. Using bins scaled to the effective radius is convenient when comparing galaxies of varying size. However, there are two problems with using R_e-normalized units. First, R_e is difficult to measure, and becomes more so for high-mass galaxies that have an extended low surface-brightness halo (e.g., Kormendy et al. 2009). Second, R_e grows with cosmic time. Therefore, if we are searching for changes in stellar populations that correspond to different epochs in galaxy growth, we may want to look at physical as well as R_e-scaled radii.

Interestingly, most of the variation in the Lick indices occurs within the central ∼7 kpc or ∼1.5R_e. From photometric fitting, Huang et al. (2013a, 2013b) find evidence for a distinct outer (∼10 kpc) component, perhaps formed via accretion. We see a tantalizing hint that the index values are converging beyond ≳ 2R_e between the low- and high-dispersion galaxies. If indeed the central compact regions of elliptical galaxies were formed in a very rapid event at high redshift, with the outer parts accreted later (e.g., Thomas et al. 2005; Oser et al. 2010), then we might be reaching a radius where a large fraction of the mass has been accreted from smaller galaxies. As our next means of studying radial stellar population trends, we examine the stacked spectra as a function of radius.

In addition to measuring more reliable Lick indices at large radius, we examine percent-level variations directly from the stacked spectra. While we do not derive any quantitative conclusions from this exercise, it is nevertheless revealing to see where the largest radial variance occurs in the spectra. We create composite spectra for two groups of galaxies divided by high (σ_* = 250–300 km s⁻¹) and low (σ_* = 150–200 km s⁻¹) stellar velocity dispersion, and we add them radially in units of R_e (Figure 2) and kpc (Figure 4). Since the largest variations occur within the central bin, we have made ratio spectra (Section 4.2) dividing by the composite spectrum in the 0.5–1R_e or the 3–6 kpc bin respectively. "Emission" features in the ratio spectra are manifestations of a declining EW relative to ∼R_e, while "absorption" indicates increasing EW.

Zoom In Zoom Out Reset image size

The most prominent radial changes are seen in the molecular bands CN (the most prominent band is at 4150 Å), C₂ (the Swan band at 4668 Å, and another Swan band at ∼5100 Å on the wing of the Mgb line), and MgH (Mgb at 5200 Å). While the gentle decline in [Z/H] is known to be the dominant change in the stellar populations as a function of radius (Paper I, and references therein), we see no strong variability in the pervasive atomic Fe absorption features. Also, both the CN and C₂ lines show strong radial gradients, but we see no corresponding variation in the G-band (CH) at 4300 Å. According to Tripicco & Bell (1995), the G-band is more sensitive to microturbulent velocity (and thus effective temperature) than abundances. Therefore, we prefer to rely on the C₂4668 index to infer carbon abundances (e.g., Tripicco & Bell 1995; Worthey et al. 1994; Schiavon 2007). Finally, we note that there also may be a shift in the centroid of the C₂4668 index with radius, but higher S/N is needed to be certain.

The ratio spectra demonstrate very clearly what was already apparent in the radial Lick index gradients (Figure 3). The largest spectral variations occur in the central regions, perhaps reflecting a two-phase formation mode for these galaxies. The strong radial decrease in the C₂, CN, and MgH molecular bands reflects the decreasing [Z/H]. At lower-metallicity molecules do not form as effectively. However, as we show below, we also find compelling evidence for true abundance ratio gradients as a function of radius, particularly in [C/Fe].

Finally, we derive radial stellar population trends based on the composite Lick index measurements presented in Section 5.1. The resulting radial profiles in age, [Fe/H], [Mg/Fe], [C/Fe], and [N/Fe] are shown in Figure 5. The stellar population trends inform us directly about when, where, and how rapidly the stars were formed and thus provide some clues to the assembly of the outer parts of massive elliptical galaxies. Encouragingly, we recover well-known trends as a function of velocity dispersion. In their centers, higher-dispersion galaxies are older, and have higher [Mg/Fe], [C/Fe] and [N/Fe] ratios (e.g., Trager et al. 2000b; Worthey 2004; Thomas et al. 2005; Sánchez-Blázquez et al. 2006; Graves et al. 2007; Smith et al. 2009; Price et al. 2011; Johansson et al. 2012; Worthey et al. 2013; Conroy et al. 2013).

Zoom In Zoom Out Reset image size

Turning to the radial trends, we see first that the metallicity [Fe/H] drops gently as a function of radius (e.g., Davies et al. 1993). The gradient Δ[Fe/H]/Δ log R_e has a similar slope for both the high-dispersion and low-dispersion galaxies. Previous work has found interesting trends between [Fe/H] gradients and σ_*, but mostly at lower σ_* than probed here (e.g., Carollo et al. 1993; Spolaor et al. 2010). The [Mg/Fe] ratio stays high or even rises slightly for the low-dispersion galaxies (Paper I). Only in age might we see tentative differences between high- and low-dispersion galaxies: the low-dispersion galaxies show a weak negative age gradient (get older) with radius, while the high-dispersion galaxies are old everywhere. New to our analysis from Paper I, we also consider the [C/Fe], [N/Fe], and [Ca/Fe] ratios as a function of radius. We see a strikingly strong trend in the radial decline of [C/Fe] with radius. In contrast, [N/Fe] and [Ca/Fe] more closely track the behavior of [Mg/Fe] and remain constant with R.

As we already noted in regard to the Lick index gradients in Section 5.1, the stellar population properties of the high-dispersion and low-dispersion galaxies begin to converge beyond ∼1.5R_e. At large radius, the typical stars in all bins are old, and have high α-abundances and low [Fe/H] (a trend seen in spiral bulges as well; Jablonka et al. 2007). If galaxies are built in two phases, then we might expect the largest variations in properties to occur in their centers, where the formation timescales and metal retention depend on the depth of the potential well of the final galaxy. In contrast, if stars at large radius were accreted from smaller systems that formed their stars rapidly and early (e.g., Oser et al. 2010), we expect stars in the outskirts to be uniformly old, α-enhanced, and metal-poor, as we find for our galaxies.

We note three caveats in interpreting the more detailed abundance ratios, such as [C/Fe]. First, our CN measurements are potentially problematic, due to the steep continuum shape around the 4000 Å break and the difficulties of accurate flux calibration and sky subtraction at the blue edge of our spectra. The [N/Fe] and [Ca/Fe] abundances, because they depend on both CN and the C₂ measurements, are both impacted by the uncertainty in measuring CN. However, the CN index measured from our central spectra shows no systematic offset from the CN index measured from the SDSS spectra, with (CN_center − CN_SDSS)/CN_SDSS = 0.07 ± 0.2. This agreement gives us confidence that our CN index measurement is not driven by systematics in our reductions.

Second, we do not directly model [O/Fe], but the assumed oxygen abundance directly impacts the [N/Fe] and [C/Fe] values (Graves et al. 2007). Because most of the C is locked up in CO, a slightly super-solar [C/O] leads to a large increase in the strength of the C₂ Swan bands (see also Tripicco & Bell 1995; Korn et al. 2005). To bracket the uncertainty in [O/Fe], we run a second set of stellar population models with [O/Fe] = 0.1 (rather than the default [O/Fe] = 0.5; dotted lines in Figure 5). We make the reasonable assumption that as an α element [O/Fe] tracks [Mg/Fe], as has been seen in recent studies of elliptical galaxy centers. Conroy et al. (2013) model oxygen abundance in SDSS galaxies using full spectral fitting and find that the O/Mg ratio is constant to within 0.05 dex. Likewise, Johansson et al. (2012) find O/Mg ∼ 1 for all σ_*. Thus, adopting a range of [O/Mg] spanning ±0.2 dex should bracket the range of allowed [C/Fe], [N/Fe], and [Ca/Fe]. Figure 5 shows that as expected, only [C/Fe] and [N/Fe] are strongly impacted by different values of [O/Fe]. At lower [O/Fe], the absolute value of [C/Fe] and [N/Fe] are both correspondingly lower, and we note that a value of [N/Fe] ≈ 0.5 dex is consistent with values reported for the most massive SDSS galaxies (Johansson et al. 2012; Conroy et al. 2013). On the other hand, the radial behavior of C and N relative to each other are not impacted by the overall O abundance.

Third, we note that the models do not include carbon stars (Schiavon 2007). However, the incidence of carbon stars is low at the metallicities considered here, and then increases strongly at yet lower metallicity (e.g., Blanco & McCarthy 1983; Groenewegen 1999; Mouhcine & Lançon 2003). Thus, we do not believe carbon stars can be dominating the observed [C/Fe] trends.

We find that stellar population gradients are strongest within ∼R_e, while at larger radius the gradients begin to flatten. Furthermore, the differences between the high- and low-dispersion galaxies in terms of age and α-abundance decrease with R; at all dispersions, stars at large radius are old, α-enhanced, and relatively metal-poor. Note that we include S0s in this analysis; we do not yet have a large enough sample to separate them, but will do so in future work. Finally, we infer strong negative gradients in [C/Fe] with radius, while the [N/Fe] abundances are high (at least [N/Fe] ≈ 0.5) and flat. We now discuss the ramifications of our results for the formation histories of massive elliptical galaxies.

Each stellar population property shown in Figure 5 tells us something about the provenance of these stars. The stellar age obviously provides one important clue. The metallicity presumably maps onto the depth of the potential well in which the star formed (e.g., Larson 1974). Finally, [α/Fe] and the more detailed abundance ratios we consider here reflect how rapidly the stellar population was formed. High [α/Fe] indicates a preponderance of Type II supernova relative to Type Ia, and thus rapid formation timescales (e.g., Matteucci & Greggio 1986). Finally, although their origins are less clear (e.g., Renzini & Voli 1981; Cescutti et al. 2009), [C/Fe] and [N/Fe] provide complementary information about star formation timescales, as they are likely produced at least partially in intermediate-mass stars (e.g., Graves et al. 2007; Johansson et al. 2012).

To a large extent, the stellar populations in the centers of elliptical galaxies scale with the stellar velocity dispersion. Luminosity-weighted mean age, total metallicity [Z/H] and formation timescale [α/Fe] all correlate with σ_* (e.g., Worthey et al. 1994; Trager et al. 2000b; Graves et al. 2007). The more detailed abundance ratio patterns are likewise monotonic functions of σ_* (e.g., Trager et al. 2000b; Graves et al. 2007; Conroy et al. 2013; Worthey et al. 2013).

As we discussed in Paper I, the stars at large radius do not share the detailed stellar population properties of the stars at the center of any present-day galaxy. The stellar ages are old and [Fe/H] is 0.3–0.4 dex subsolar. However, despite the low-metallicity, the [α/Fe] abundance ratios are high (e.g., Spolaor et al. 2010, Paper I). The low [Fe/H] values are best explained by formation in shallow potential wells, either of small galaxies as advocated in the two-phase model of galaxy formation, or the outer parts of big galaxies. However, given the old ages and high [α/Fe] ratios, these stars must have formed early and over short timescales. To estimate how rapidly, we adopt the scaling from Thomas et al. (2005) between [α/Fe] and star formation timescale, which is based on a simple closed-box model. Assuming that [O/Fe] tracks [Mg/Fe], we find a timescale of 250 Myr for [Mg/Fe] ≈ 0.3.

Likewise, the ratio of C/N apparently falls at large radius, while it remains constant in the centers of elliptical galaxies. As we argued above, while the absolute values of carbon and nitrogen are uncertain due to the unknown [O/Fe], their ratio as a function of radius should be robust. We will first discuss the radial behavior of [C/Fe], and then turn to the more puzzling [N/Fe] trends.

Carbon is made in the triple alpha process in intermediate-mass (1–8 M_☉; e.g., Renzini & Voli 1981) stars. Massive, metal-rich stars also release significant C through stellar winds (e.g., Maeder 1992). The rising [C/Fe] seen in the centers of massive elliptical galaxies, where the star formation timescales are short, seems to require that a significant fraction of the carbon actually is made in massive stars (e.g., Graves et al. 2007). Carbon from massive stars is also needed to explain the correlation between the C/O and O/H ratios observed in H ii regions and individual Milky Way stars (e.g., Cescutti et al. 2009; Garnett et al. 1999). Yield calculations find that massive metal-rich stars can provide enough carbon to explain the Galactic observations (Maeder 1992; Marigo et al. 1998; Henry et al. 2000). Based on their high [α/Fe], stars at large radius were likely formed rapidly, so the declining [C/Fe] presumably reflects the declining carbon yields from lower-metallicity massive stars, and supports our inference of rapid star formation timescales.

Nitrogen is more complicated. We must explain both the very super-solar [N/Fe] values that we observe (at least [N/Fe] ∼ 0.5) and the falling C/N with radius. Nitrogen is produced in intermediate-mass stars as part of the CNO cycle (e.g., Matteucci 1986), but C is required before N can be produced. The C is either synthesized in the star through the triple alpha process ("primary") or present in the star to begin with ("secondary"). While some primary N is required from low-metallicity massive stars to explain the floor in N abundances seen in H ii regions (e.g., Izotov et al. 1999; Meynet & Maeder 2002), the yields from massive stars alone are not high enough to produce the super-solar enrichment that we observe. As pointed out in Johansson et al. (2012), high [N/Fe] ratios then provide a lower limit on the star formation timescale of at least a few Myr, the lifetimes of intermediate-mass stars. Therefore, the star formation timescales inferred from [α/Fe] of 250 Myr are consistent with the timescales required by [N/Fe].

Since synthesizing N requires existing C (Henry et al. 2000), it is possible that the high [N/Fe] results from the super-solar [C/Fe]. However, we would not expect [N/Fe] to remain high at large radius where [C/Fe] is falling. Longer star formation timescales could provide higher [N/Fe] ratios, but presumably would flatten [C/Fe] as well. It could be that N is more effectively released by lower-metallicity stars, although we do not have a proposed mechanism here. A changing initial mass function (IMF), in particular a bottom-heavy IMF as have been invoked for the most massive elliptical galaxies (Conroy & van Dokkum 2012; Thomas et al. 2011; Dutton et al. 2012) cannot help, since the C is made by massive stars. We do note that the nucleosynthesis of N is a puzzle in globular clusters as well (e.g., Cohen et al. 2005), and it may be that these problems have a similar solution. Intriguingly, an ultra-compact dwarf in Virgo recently discussed by Strader et al. (2013) also appears to have solar carbon abundances but super-solar [N/Fe] ∼ 0.6 dex, perhaps providing an additional link between small stripped galaxies and elliptical stellar halos. Finally, there is always the possibility that our [N/Fe] values are biased high due to uncertainties in the CN measurements, although as mentioned above, we find no systematic differences between our measurements and the SDSS measurements. We have no clean explanation for the observed radial decrease in C/N.

The conclusions for galaxy assembly are as follows. We have seen that stellar populations at radius ≳ 2R_e do not look like those found at the centers of any elliptical galaxies today (Paper I, and this work). According to our stellar population modeling, stars at ∼2R_e are similar at all σ_* (see how they converge at large radius in Figure 6). The typical star is old (∼10 Gyr), relatively metal-poor ([Fe/H] ≈ −0.5), and α-enhanced ([Mg/Fe] ≈ 0.3). We infer that stars at large radius are formed at z ≈ 1.5–2 in shallow potential wells over ∼250 Myr timescales. Declining [C/Fe] ratios support this picture: with rapid star formation timescales, the declining C yields from massive stars at low-metallicity leads to a decline in this ratio with radius. On the other hand, the high and flat [N/Fe] ratios are unexpected, but possibly also seen in stripped dwarf galaxies.

Zoom In Zoom Out Reset image size

Given our observed radial trends in stellar population properties, we return to the assembly history of massive elliptical galaxies. Based on the striking average size growth of elliptical galaxies between redshift two and the present, numerous papers have proposed a two-phase model for their growth (e.g., Naab et al. 2009; Bezanson et al. 2009; Oser et al. 2010, 2012; Hilz et al. 2012). Our observations are consistent with this picture. The stars are converging toward similar properties at large radius independent of σ_*. The stellar populations at large radius demand that the accreted galaxies must be small (to explain the low [Fe/H]) and form early and rapidly (to explain the old ages and high [α/Fe]). Simulations make similar predictions (Oser et al. 2010).

While the stars in the outskirts of massive ellipticals do not resemble the centers of any elliptical galaxies today, there must be present-day stars with similar metallicities and abundance patterns that did not end up in massive elliptical galaxy outskirts. In Figure 6, we show that the average star in the outskirts of our ellipticals resembles thick-disk Milky Way stars. Perhaps the notion of a thick disk is outdated. Instead, we can say that stars in elliptical galaxy outskirts are similar to the luminosity-weighted mean star in the Milky Way disk (Bovy et al. 2012a, 2012b, J. Bovy 2013, private communication). Since stars form in disks, and low-density disks are easier to disrupt than centrally concentrated ellipticals, we suggest that the outer parts of our elliptical galaxies were built by the shredding of disky galaxies at early times (e.g., Toft et al. 2007; Conselice et al. 2011). Based on the stellar populations, similar stars, formed at z ≈ 1.5–2 in disks but never accreted by a more massive halo, form the thick disk components of present-day spiral galaxies (see Figure 6). Our observations are consistent with a two-phase model for elliptical galaxy growth.

On the other hand, our observations alone do not require such a model. Instead, the average size growth in the elliptical galaxy population may occur as larger galaxies join the red sequence at later times (Valentinuzzi et al. 2010; Cassata et al. 2011; Newman et al. 2012; Carollo et al. 2013; Barro et al. 2013), consistent with the observation that at a given velocity dispersion, more diffuse galaxies quenched later (Graves et al. 2010). If stars are formed in situ at large radius, followed by global star formation quenching, then we might expect stellar population gradients as predicted by classic monolithic collapse models (e.g., Larson 1974; Carlberg 1984). Monolithic collapse naively predicts very steep metallicity gradients as a function of radius (e.g., Larson 1974; Kobayashi 2004). However, the shallower observed [Fe/H] gradients, as well as flat [α/Fe] gradients, can be reproduced in monolithic scenarios by varying the star formation efficiency and gas inflow rate (Pipino et al. 2010). In situ formation at large radius also more naturally explains the observation that Mgb EW correlates strongly with local escape velocity (e.g., Franx & Illingworth 1990; Weijmans et al. 2009; Scott et al. 2009, 2013).

Both assembly paths (two-phase and all in situ) likely occur in nature (e.g., Faber et al. 2007; Toft et al. 2007; Bundy et al. 2010; Szomoru et al. 2012; Barro et al. 2013). In the future, we hope that combining our stellar population measurements with kinematics may provide further clues to the level of dissipation involved in forming stars at large radius (e.g., Hilz et al. 2012). We will grow more sensitive to any such trends as our survey continues (S. Raskutti et al., in preparation).

Eventually, we would like to examine the radial behavior of elliptical galaxies as a function of their fundamental plane location (or compactness) as in Graves et al. (2009, 2010). In the meantime, we can look at one spectacular outlier in our sample: NGC 1270. This galaxy has a half-light radius of ∼2 kpc, and a central stellar velocity dispersion of ∼370 km s⁻¹. van den Bosch et al. (2012) highlight NGC 1277 as the prototype of these compact high-dispersion galaxies, which also appear to contain very massive (>10¹⁰ M_☉) supermassive black holes. van den Bosch and collaborators have found a handful of galaxies like NGC 1277, including NGC 1270 (for low-mass versions see Trujillo et al. 2009; Jiang et al. 2012; Ferré-Mateu et al. 2012). They have disky light profiles and are rotationally dominated at large radius.

NGC 1270 appears to be an extreme example of a compact galaxy that formed at high redshift, and then grew no more. With a central age of ∼11 Gyr, it is one of the oldest galaxies in the sample (Paper I), has a very high central Mgb EW, and shows no gradients as a function of radius (Δ log Age/Δ log R_e = 0.20 ± 0.13). As far as we can tell, NGC 1270 follows the Mgb–σ_* relation (Paper I) but the statistics are very limited for velocity dispersions so high. Based on color gradients and spectroscopy respectively, NGC 1277 and SDSS J151741.75−004217.6 also seem to fit this picture (van den Bosch et al. 2012; Läsker et al. 2013). We will look in more detail at the stellar population gradients for a larger set of the entire van den Bosch sample (A. Yildrim et al., in preparation).