THE INTERSTELLAR MEDIUM AND FEEDBACK IN THE PROGENITORS OF THE COMPACT PASSIVE GALAXIES AT z  ∼  2

The galaxies discussed here are the LBG samples at z  ∼  3.4 presented in Williams et al. (2014). These LBGs are z-band selected from the GOODS survey (Giavalisco et al. 2004), and are found over an area of 113 arcmin² from the GOODS-South field HST/WFC3 imaging from CANDELS (Grogin et al. 2011; Koekemoer et al. 2011). Through simulations we established our sample is 90% complete with z  ⩽  26.5 for galaxies of <03 half-light radius (Huang et al. 2013a). We measure multi-wavelength photometry using the object template-fitting method (TFIT; Laidler et al. 2007) software, which includes U-band imaging from the Visible Multiobject Spectrograph (VIMOS) on the Very Large Telescope (VLT; Nonino et al. 2009), HST/ACS B, V, i, z-band, HST/WFC3 J, H-band, VLT/ISAAC K_s photometry (Retzlaff et al. 2010), and Spitzer/IRAC 3.6, 4.5, and 5.7 μm imaging (M. Dickinson et al., in preparation). The details of the construction of multi-wavelength photometry are discussed in Guo et al. (2013). Measurements of galaxy properties for the LBG samples using the multi-wavelength photometry assume a Salpeter initial mass function (IMF), and is described in Guo et al. (2012a) and Williams et al. (2014). Note that the specific assumption of an IMF here is relatively inconsequential, since the comparison between the two types of galaxies is done at fixed stellar mass. Changing the shape of the IMF will result in different absolute values of stellar mass, but comparisons between the two flavors of galaxies would remain unchanged. We use GALFIT (Peng et al. 2002) to fit Sersic models to the WFC3 H-band images of the galaxies (rest-frame 3600 Å at z  ∼  3.4) from which we measure half-light radii (the sample are bright, H < 25, which ensures robust morphological measurements). A detailed description of these measurements and also the selection of the LBG samples are presented in Sections 2 and 3 of Williams et al. (2014), and we briefly outline the procedure here.

This sample of LBGs was split into plausible "candidate" progenitors of CPGs, and "non-candidate" LBGs as described in Williams et al. (2014). The selection of candidate progenitors is based on evidence that any star-forming progenitors must also be compact, since merging and accretion tend to increase the sizes of galaxies (e.g., Hopkins et al. 2008). Therefore to identify candidate progenitors, we require that they be of similar size and morphology as the CPGs, and also have high enough SFR that if they quench soon after observation they will meet the CPG selection criteria (Cassata et al. 2013). This sample of CPGs is selected from CANDELS photometry to have low specific SFRs (<10⁻² Gyr⁻¹), high masses (>10¹⁰M_☉), non-detections in 24 and 100 μm maps inconsistent with dust obscured SF, and visually identified in the H-band to be "early-type." It is impossible to know the future star-formation histories (SFHs) of galaxies; however, CPGs have strong constraints on their SFHs, because their low sSFRs implies they must have quenched star-formation >1 Gyr prior to the epoch of observation (Kriek et al. 2006, 2009b; Onodera et al. 2012; Kaviraj et al. 2013; Gobat et al. 2012; van de Sande et al. 2013; Whitaker et al. 2013). Therefore, we assumed a projected (future) SFH for all galaxies (exponentially declining with τ = 100 Myr so that all galaxies would be passive according to the sSFR criteria of Cassata et al. 2011, 2013) with the observed initial SFR, and estimated how much stellar mass would be accumulated from the time of observation until the epoch of CPGs, which we took to be 〈z〉 = 1.6. We additionally assumed no growth in size. Those LBGs that would meet the CPG selection of Cassata et al. (2011, 2013) are considered candidate progenitors (from now on referred to as candidates). The candidate selection is primarily based on SFR and size, and to a lesser extent, the stellar mass. In other words, candidates tend to be more compact and a milder tendency to higher SFR. The combination of these two characteristics means that candidates have higher surface density of SFR (Σ_SFR; see Figure 1 for the properties of the subset of candidates and non-candidates with spectroscopy studied in this paper, and Figure 2 for their HST color composites). The average SFRs, stellar masses, and circularized half-light radii, for candidates (non-candidates) in the spectroscopic sample are 520 (170) M_☉ yr⁻¹, 9.8 (9.6) log₁₀M_☉, and 0.9 (1.9) kpc.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We use spectra in GOODS-South obtained at the European Southern Observatory's VLT as part of the GOODS spectroscopy program with FORS2 (Vanzella et al. 2005, 2006, 2008), VIMOS (Popesso et al. 2009), and the Galaxy Mass Assembly Ultra-deep Spectroscopic Survey with FORS2 (Kurk et al. 2009, 2013). At z  ∼  3.4, these optical spectra probe the rest-frame UV spectrum of the LBGs, a spectral region rich with diagnostics of the physical state of the ISM. The spectral resolution of these surveys varies between R  ∼  250–1000, corresponding to velocity resolutions of 300–1200 km s⁻¹. The majority (60%) of the spectra we use in our analysis have R  ∼  660–1000 (velocity resolution 300–450 km s⁻¹).

The scientific goal of the surveys above were to identify the redshifts of large samples (tens of thousands) of galaxies at z >1, and while there are often multiple identifiable emission or absorption lines, generally the spectra do not have adequate signal-to-noise ratios for detailed studies of the galaxy properties on an individual basis. In the following sections we therefore study the average properties using stacked spectra to maximize signal to noise. We visually inspected each individual spectrum, retaining only those of high quality for our final spectroscopic sample. This quality screening is qualitative, including visual identification of absorption and emission lines at the cataloged redshift, and also to some extent taking into account the quality flag published by each survey, whose redshifts may have been estimated using cross-correlations with galaxy templates rather than obvious lines. We additionally require "good" spectra to have identifiable features, which we use to estimate systemic redshifts (see next paragraph). Our final "good" quality spectroscopy sample includes 12 candidate progenitors and 20 non-candidates (27% and 15% of the original samples from Williams et al. (2014), respectively). The total exposure time in the stacks of each are ∼103 and ∼98 hr, respectively.

The cataloged redshifts that are published in the spectroscopic surveys were typically measured using spectral lines that are either resonant lines (Lyα) or strong interstellar lines that may be heavily affected by outflows (for example Si ii, Si iv and C iv). Therefore these redshifts are unlikely to be systemic (tracing the exact redshift of the galaxy) but rather reflect the redshift of the galaxy, modulated by additional velocities due to the kinematic properties of the ISM. Interstellar absorption lines are often observed blue shifted with respect to the rest transition wavelength due to large-scale outflows in the ISM of high-redshift galaxies. Lyα emission is often observed redshifted, since when redshifted out of resonance Lyα photons may travel freely, and therefore the Lyα may also reflect the properties of outflows (Shapley et al. 2003; Hansen & Oh 2006; Verhamme et al. 2008; Schaerer & Verhamme 2008; Vanzella et al. 2009; Steidel et al. 2010; Kulas et al. 2012). Lines that trace the systemic redshift, for example photospheric lines that originate in the atmospheres of stars, are generally too faint to be detected in our spectra.

In this section, we document our procedure to estimate systemic redshifts using the UV spectroscopy. We also quantify our expected uncertainty in the systemic redshifts using a subset of galaxies for which systemic redshifts have been published based on rest-frame optical nebular emission lines, using NIR spectra. These NIR spectra and systemic redshifts come from Keck/MOSFIRE (Holden et al. 2014) and the Assessing the Mass-Abundance redshift[-Z] Evolution survey (Maiolino et al. 2008) with VLT/SINFONI. We use a comparison of 15 UV spectra of LBGs with published NIR-derived redshifts.

Previous studies have published methods to derive systemic redshifts from rest-frame UV spectroscopy, when no systemic lines are available (e.g., Adelberger et al. 2005; Steidel et al. 2010). In particular, Adelberger et al. (2005) have used a large sample of LBG at z  ∼  3 with near-IR spectra to estimate systemic redshifts from the UV lines only, for the case where only Lyα is present, both ISM lines and Lyα are detected, and when only ISM lines are detected. This method is based primarily on the observed offsets in velocity between the various UV lines and NIR observed nebular emission lines. These velocity offsets are direct measurements of velocities in the ISM, and therefore these conversions from UV-line redshifts to estimated systemic redshifts will reflect the average outflowing ISM properties of their LBG samples. This means, by using their conversion, we artificially imprint the average kinematic properties of their LBG samples on our two galaxy samples, and as we are interested in studying outflow velocities, we approach this method with caution. We seek the best method to derive systemic redshifts from our samples, without any assumptions about the properties of the outflowing ISM.

Our individual spectra often exhibit emission lines immediately redward of broad interstellar absorption. These are most often seen in C iv λλ1548,1550 and Lyα, which often exhibit P-Cygni profiles, but are also observed in other absorption lines (e.g., C ii λ1334, Si ii λ1526). Motivated by the need for an entirely empirical systemic redshift estimator that is independent of assumptions about the average properties of LBGs, we test how the redshift measured by the transition point between the absorption and emission components in various elemental species relates to the NIR-derived systemic redshift where available.

We measure these redshifts using a method based on the midpoint between the centroid wavelengths of emission and absorption components. We measure the wavelength of the midpoint between centroids for each observed feature, assuming this "midpoint" wavelength to be at rest with respect to the galaxy, and measure its redshift. Note that we are not suggesting a physical scenario to motivate the use of this mid-point method to estimate systemic redshifts and rather the method is entirely empirical. The left panel of Figure 3 shows the velocity offset between the predicted systemic redshift from the most common individual features (e.g., Lyα, Si ii λ1260, O i/Si ii λ1303, C ii λ1334, Si ii λ1526, C iv λλ1548,1550) compared with the systemic redshift of the galaxy measured from the NIR. We supplement these measurements with measurements from the Keck/HIRES spectrum of D6 (N. Reddy & C. Steidel 2014, private communication; Shapley et al. 2003; Steidel et al. 2010), whose systemic redshift has been derived from NIR spectroscopy (Pettini et al. 1998). We find that Lyα is the poorest estimator of systemic redshift among our LBG sample, not surprising since the resonant nature of the escape of Lyα photons from a galaxy may mean its transition point is not at rest, and also due to the complication of identifying the absorption component due to the Lyα forest. The measure from the Lyα in the HIRES spectrum of D6 may be better due to the more accurate identification of the P-Cygni absorption component. We also find that transitions of Si ii may be poor estimators, possibly due to confusion between P-Cygni emission (if it exists) with fluorescent emission Si ii* (if present), which exists at slightly longer wavelengths than would be expected of P-Cygni emission. C iv appears to be a good estimator. The midpoint derived redshift from C iv is relatively accurate, with the mean systemic redshift centered at zero and never deviating more than 500 km s⁻¹ from the rest transition, and is the most frequently observed in our spectra of all features tested here.

Zoom In Zoom Out Reset image size

In the right panel of Figure 3, we show how the combined information of all available transitions in individual spectra (i.e., the average redshift from each available midpoint) compares with the NIR-derived systemic redshift. The average redshift appears to agree remarkably well with the NIR-derived one. The mean of the distribution of velocity offsets between midpoint-derived redshifts and systemic redshifts is −42 km s⁻¹, with standard deviation of 283 km s⁻¹. No galaxy deviates larger than 490 km s⁻¹. Since Lyα is the most deviant in the left panel, we repeat the distribution of velocity offsets with the Lyα midpoint redshift removed from the average. We find that although removing Lyα tightens the distribution to 190 km s⁻¹, the offset from 0 increases to a mean offset of −193 km s⁻¹. We therefore conclude that we do not gain a closer estimate to systemic by excluding the Lyα, and it should be included where available. Examples of individual spectra that enter the stack can be found in Figure 4, which illustrates the derivation of the systenic redshift derived from these midpoints.

Zoom In Zoom Out Reset image size

In addition to testing the midpoint between emission and absorption, we also tested using the transition point between where the emission and absorption components meet the continuum level, and found that this continuum transition point did not trace systemic as accurately as the midpoint, and this may be due to the difficulty in accurately measuring the continuum level in noisy spectra.

For comparison, we have additionally computed systemic redshifts estimated using the Adelberger et al. (2005) method for comparison with the observed NIR redshifts. These are also plotted in red in the right panel of Figure 3, and find that the Adelberger et al. (2005) have a mean offset of 165 km s⁻¹, with a standard deviation of 260 km s⁻¹. This average offset is larger than that derived using the mid-point method with a similar distribution width, and results in multiple galaxies with offsets larger than 550 km s⁻¹.

Therefore, in the remainder of our analysis, we estimate systemic redshifts based on the midpoint method we have outlined here, which we have characterized to be a tracer of systemic with an expected uncertainty of ∼280 km s⁻¹. Since the main purpose of this paper is to study the relative properties of candidate progenitors, compared to ordinary LBGs, it is therefore very important to use a uniform way to estimate redshift from all galaxies studied here, independent of any assumptions on the properties of outflows in the ISM (e.g., Adelberger et al. 2005; Steidel et al. 2010). Although individual outflow velocities may not be accurate to more than this velocity uncertainty, we choose this option rather than mixing redshift estimators for different galaxies, which could smear away differences between the samples. Importantly, however, we found that the results of this paper remain if we instead adopt either the Adelberger et al. (2005) method, or, if we combine the mid-point method with the NIR-derived redshift where available when producing the stacks. We found that all qualitative comparisons of the stacks are independent of the choice of the above methods used to convert spectra to the rest frame, and we would reach the same conclusions were we to adopt any of them.

We stack our two samples of galaxies using the following procedure. We first individually de-redshift each spectrum using the estimate of systemic redshift from the mid-point of P-Cygni-like profiles discussed above. We then normalize each spectrum using the median value of the continuum measured between rest frame 1400 Å  <  λ  <   1500 Å. Finally, we stack using the scombine package in IRAF, where we rebinned to a common dispersion (dλ = 0.6 Å pixel⁻¹, rest frame), and performed a 3σ clipping during the stack (although we note that the clipping minimally affected the resulting stack, indicating that there are few significant outliers in terms of the features within our samples). Our final stacks are presented in the top panel of Figure 5. The bottom panel shows the spectral ratio (candidate stack divided by non-candidate stack). The ratio should be unity for identical features between the two samples, and therefore large deviations from this indicate features that differ. An indication of significance of the spectral differences can be see in the pink region, which is defined by unity plus or minus the combined (in quadrature) sample standard deviations of candidates and non-candidates. These standard deviations are estimated as follows. We repeat the stacks of each sample using jackknife resampling, each time removing one spectrum and stacking the rest. The final sample error on each spectral point is then the standard deviation of the jackknifed stacks. We then combine in quadrature the standard deviations of candidates and non-candidates around unity to make the pink region. We note that the combined region redward of 1250 Å is remarkably uniform, indicating the variation in those spectral features from galaxy to galaxy is relatively small. In contrast, the pink region near Lyα is much larger, indicating a diversity of Lyα properties among our samples. The differences in stacked Lyα between the candidates and non-candidates, along with error bars indicating the standard deviation of each stack, can be seen in Figure 6.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

In the composite spectrum shown in Figure 5 there are many features that are generically visible among LBGs at z  ∼  3–4 (e.g., Shapley et al. 2003; Jones et al. 2012). There are a variety of high- and low-ionization lines visible in absorption. First, by comparing the rest-wavelength of these lines (dotted lines) it is obvious that both candidates and non-candidates exhibit outflows, as indicated by the blue shifts in absorption troughs relative to the rest wavelength. In particular, Si ii, C ii, Si iv, C iv, and Fe ii are visible in both spectra. We also see He ii in emission, detected in the candidates. The UV spectral slope is redder for candidates than non-candidates, similar to the redder UV observed photometrically in Williams et al. (2014). Both of the samples show strong Lyα in emission (although it extends beyond the range shown in Figure 5 to emphasize the weaker features in other parts of the UV spectrum, we show Lyα individually in Figure 6). The candidates exhibit stronger absorption in Lyα, and their Lyα emission is also more redshifted. As these are stacks and therefore represents the average properties of galaxies selected according to our criteria, we do not focus on specific quantities from each stack. Rather, in the following sections, we will focus on comparing the spectral properties in a relative way.

Because of this relative comparison, it is necessary to test if mass differences between the two samples could produce the relative differences we observe. For example, it is possible that the kinematic properties of the absorbing gas may be a function of luminosity or mass. Additionally, if the mass–metallicity relation persists to z > 3 (e.g., Maiolino et al. 2008), samples that are offset from each other in mass may be more likely to show offsets in metallicity. Both Steidel et al. (2010) and Lee et al. (2013b) have found that velocity offsets between interstellar absorption lines and the systemic redshifts of galaxies are dependent on the baryonic mass of the galaxy. In the top left panel of Figure 1, the non-candidate sample has a larger low-mass tail than the candidates, although their overall mass range and high-mass distribution are very similar. To explore the dependence of any of our main conclusions on this low-mass tail, we repeat the non-candidate stack with a subsample of non-candidates that are more similar to candidates in mass, such that the average mass of the distributions match. We remove all non-candidates whose log₁₀ M*  <  9.6, resulting in a sample of non-candidates with the same average stellar mass as the candidate sample, 〈M_star/M_☉〉  ∼  9.8. This high-mass sample includes 14 of the 20 non-candidates. Due to the smaller number of non-candidates that meet this criteria in mass, the noise in the stack is larger. The second (lighter) blue distributions in each panel of Figure 1 shows the sample differences when excluding these lowest mass galaxies. The stack of the high-mass sample compared with the full sample of non-candidates is presented in Figure 7. We find a slight increase in the equivalent width of some of the interstellar absorption lines (e.g., Si iv λ1393), and a decrease in Lyα emission strength, but otherwise, the high-mass stack is qualitatively very similar to that of the stack of the full sample of non-candidates. We stress that the main results of this study, reported in the next sections, remain unchanged if we used measurements made with the high-mass stack of the non-candidate galaxies. Therefore, our results are independent of any offsets in their mass distributions, and rather reflect intrinsic differences between the candidates and non-candidates.

Zoom In Zoom Out Reset image size

In this section, we focus on the relative kinematic properties of the ISM in the two samples of galaxies, which can be derived from the following observed interstellar absorption lines: Si ii λ1260, the O i/Si ii λ1303 blend, C ii λ1334, Si iv λ1393 and λ1402, Si ii λ1526, and Fe ii λ1608. As can be seen in Figure 5, the candidates exhibit stronger absorption in the majority of these lines. Visual inspection of several lines shows the candidates have larger velocity offsets. Both of these are also true of the C iv λλ1548,1550 doublet, however, due to the asymmetric nature of the P-Cygni profile, which is a combination of stellar absorption and emission, in combination with further absorption by outflowing material in the ISM, we defer analysis of this line to the next section.

We fit Gaussian profiles to the seven interstellar lines mentioned above, where the line center, width, continuum level and peak flux density are left as free parameters. The observed and fitted absorption profiles are shown in Figure 8, and the centroids of these fits are listed in Table 1. (For comparison, the same fits using the high-mass only non-candidate stack is shown in Figure 9). We also calculate equivalent widths of each line by directly integrating the line flux in the observed stack. For each ISM line, we fit a line to two continuum regions bracketing the ISM line that are devoid of other features, interpolate this continuum fit at the ISM line center, and then integrate the equivalent width with respect to the continuum. We estimate the error on the equivalent width using the standard deviation of 1000 Monte Carlo simulations generated from simulated stacks produced from Gaussian variates of each point according to the rms values in the continuum regions around each line, as measured by IRAF.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We compare the line centroid relative to the rest wavelength of the line, and the line equivalent width in Figure 10. This figure shows the equivalent width versus blueshift of each individual line fit in Figure 8 for both candidates and non-candidates. In the figure we show low- and high-ionization lines color coded separately, although we do not see that there is any trend with ionization state. It is apparent, however, that there is a tendency for the absorption lines from the candidate stack to exhibit a stronger blueshift with respect to the rest wavelength than the non-candidates.

Zoom In Zoom Out Reset image size

We also plot in Figure 11 the same measurements with a different representation. The left panel shows the centroid blueshift of each individual line compared directly between the candidate and non-candidate stack. In other words, the blue shift of any given line (i.e., Si ii λ1260) in the candidate stack is compared directly with the blueshift of that same line in the non-candidate stack. Similarly we compare directly the absolute values of equivalent width in absorption of any given line in the right panel of Figure 11. The one-to-one line would indicate the exact same centroid blueshift, or equivalent width, was measured in candidates and non-candidates for any given absorption line. This figure shows more clearly that in nearly all measurements of blueshift and equivalent width, the candidates show a larger measurement than the non-candidates, indicating that on a line by line basis, as well as globally, candidates exhibit larger centroid blueshifts and stronger equivalent widths than non-candidates.

Zoom In Zoom Out Reset image size

This is indicative that the kinematics in the ISM of these two samples of galaxies may be distinct from each other. First, this offset in blueshift could in principle be an indication that the velocity spread of the outflowing component is higher on average among the candidate spectroscopic sample than that of the non-candidates, with more high-velocity gas contributing to the growth of the absorption lines in the former than in the latter. Another way to state this is that the smaller blueshift of the centroid of the non-candidates could be the result of a larger component of zero-velocity gas component, i.e., their ISM is on average less energized than that of the candidate galaxies. The right panel of Figure 11 compares the equivalent width of each absorption line from the candidate stack with that same line in the non-candidate stack. These lines are saturated (e.g., Pettini et al. 2002; Shapley et al. 2003), and therefore any increase in equivalent width is only weakly related to column density of that gas in the ISM. Rather, the measured equivalent width is instead a reflection of the combined effect of the spread in velocities of the absorbing gas clouds in the ISM, and the covering fraction of gas. With this low-resolution data, it is not possible to distinguish these two contributions. The data presented in Figure 11 are consistent with a scenario where the outflowing gas of the candidates includes faster components and also a larger spread of velocities of its gas clouds in the ISM, compared with non-candidates. They also are consistent, however, with a lower covering fraction of the ISM in the non-candidates relative to the candidates, which is a distinct possibility since the morphology of the stellar component of the former is much more extended than that of the latter. It is not possible to clearly disentangle these two contribution with our data. As we discuss later in Section 3.3, however, the fact that the candidates have much stronger Lyα emission, given the strength of their interstellar absorption, compared to the non-candidates suggests that the gas covering fraction is unlikely playing a key role here, since a larger covering fraction would generally tend to depress the strength of Lyα emission, not enhance it.

Since the candidates on average have higher SFRs than non-candidates, we investigate how these velocities relate to each other in terms of the differences in SFR. Relations between outflow velocity and SFR have been observed over many orders of magnitude in galaxy SFR, and Martin (2005) identified a relation between outflow velocity and SFR, such that V ∝ SFR^0.35.

This relation is based on the upper envelope of the distribution of velocity versus SFR for galaxies at z  ∼  0 over four orders of magnitude in SFR (the full distribution studied there exhibits a lot of scatter on an individual galaxy basis).

Other studies have found similar relationships between SFR and outflow velocity for a variety of galaxy populations (Weiner et al. 2009; Banerji et al. 2011; Song et al. 2014; although still other studies have found that a more significant relationship exists between outflow velocity and surface density of star-formation, see a more complete discussion in Section 4.2). Here, we use the observed relation from Martin (2005) to see if our observed average velocity differences and SFRs are consistent with this relation. The relation is given by logV = (0.35 ± 0.06)log(SFR)+(1.56 ± 0.13), where V is outflow velocity in km s⁻¹ and SFR is in M_☉yr⁻¹. We plot this relation in Figure 10 (dashed line), where we use the relation to predict the ratio of the average outflow velocities, given the average SFRs of the two samples. The gray region outlines the region within the uncertainties of the fit parameters for this relation (Martin 2005). We see that the average low-ionization line blueshift of the two stacks follow this empirical scaling relation between SFR and outflow velocity for galaxies at lower redshifts. The weighted averages and standard deviations of the low-ionization lines, which generally trace the same gas and therefore may have similar kinematics, are also presented. Candidate low-ionization lines exhibit on average centroid velocities and standard deviations of −426 ± 134 km s⁻¹, and that of non-candidates is −262 ± 202 km s⁻¹. Martin (2005) notes that although the upper envelope of the velocity-SFR distribution for the lower redshift sample of galaxies studied there follow this relation, at SFR > 10 the scatter in velocity at a given SFR is huge, and the slope may in fact be shallower. If such a scaling relation exists, and is flatter than that suggested by Martin (2005) and Weiner et al. (2009), the difference in centroid velocities between our two samples may be more significant than that expected purely from SFR differences using the scaling relation of the upper envelope. Regardless, we find that these velocities are consistent with what is expected given the differences in their SFRs.

These blueshift measurements obviously rely on a robust measurement of systemic redshift in each galaxy, and as discussed previously, the systemic redshift we measure is subject to some uncertainty. With these data this is unavoidable. However, it is possible to get a first-order check on our conclusion of the difference in relative centroid velocities between the two samples by comparing the net offsets in velocity between the absorption lines and the Lyα. Because Lyα is a resonant line, it is generally seen redshifted with respect to systemic in high-redshift galaxies, because Lyα photons that scatter off outflows for example allows the line to redshift out of resonance (Shapley et al. 2003; Steidel et al. 2011). Indeed, LBGs generically show a difference in velocity between Lyα and interstellar absorption lines that is smaller than the difference between rest wavelengths, suggesting Lyα scatters off outflows in the background of galaxies (i.e., moving away) and interstellar lines absorbed by outflowing gas in the foreground (i.e., moving toward us).

As can be seen in Figure 6, both samples exhibit Lyα profiles that are redshifted, as would be expected if it traces outflowing gas and must be redshifted out of resonance in order to travel freely, or, if it is absorbed by neutral gas close to rest with respect to the galaxy. It is clear that the candidate Lyα profile exhibits a larger redshift, whereas the non-candidate Lyα profile is closer to the rest wavelength. In Figure 12, we plot a histogram of the velocity difference between the peak wavelength of the Lyα profile, and the Gaussian centroid of the interstellar absorption lines. This figure shows that the net offsets in velocity differ significantly between the two samples, with the candidates showing larger relative velocities between interstellar absorption and Lyα emission, independent of systemic velocities.

Zoom In Zoom Out Reset image size

As is clear from the large jackknifed error bars on the stacks in Figure 6, which reflect the variation among the individual spectra, the LBGs making up both our samples exhibit a variety of Lyα properties. In the candidate stack, three galaxies lack Lyα coverage, and of the remaining, about half exhibit Lyα in emission (one weak detection, two with detected Lyα in excess of the average and two with detected Lyα below the average). Of the non-candidates, 1 galaxy lacks Lyα coverage, and 9 show well detected Lyα emission, 4 have weak detections, and 6 are not detected. To summarize, roughly half of each sample with Lyα spectral coverage has Lyα in emission and the other half in absorption. This is in good agreement with the global statistics observed among LBG at z  ∼  3 in general, in which roughly half have Lyα detected in emission (e.g., Shapley et al. 2003). Thus, although the two samples of spectra that enter in both stacks are relatively small, we do not believe that small-number statistics are significantly affecting the average properties of each LBG sample as represented by the two stacks. We further note that removing the 4 galaxies without Lyα coverage from the stacks does not change our results.

Since the SFR and overall UV spectra of the two samples are similar, the differences between their Lyα lines very likely reflect differences in the coverage and conditions of neutral Hydrogen in the ISM. As discussed previously, we observe the candidate Lyα profile to be more redshifted than that of the non-candidates. However, the presence of Lyα emission blue shifted with respect to the rest wavelength may indicate the escape of Lyα photons through patchy or ionized foreground ISM (e.g., Heckman et al. 2011) or be an indication of infall (e.g., Kulas et al. 2012). In this section, we describe our measurement of the equivalent width of the Lyα line, and present analysis based on this measurement.

In addition to being characterized by redshifted emission, the Lyα line in our samples, like many galaxies at similar high-redshift, is characterized by asymmetric profiles, including some blue shifted emission, as well as blue shifted absorption. Measurement of the intrinsic equivalent width of Lyα at high-redshift is complicated by additional intervening absorption due to neutral Hydrogen in the circum-galactic medium of the galaxy and by the Lyα forest. This extrinsic absorption, i.e., not intrinsic to the galaxy itself, makes defining the blue ward edge and shape of the Lyα profile highly uncertain. To ensure an equal comparison with previous measures in the literature, we adopt a similar procedure to that by Shapley et al. (2003); Stark et al. (2010); Jones et al. (2012) when we measure the equivalent width of the Lyα. Using the stacked spectrum of z  ∼  3 of Shapley et al. (2003), we have reproduced their measurement of Lyα equivalent width of 14.3 Å using a continuum level defined by the mean value between rest wavelengths 1225–1235 Å, and where the bounds for the line are defined as follows. The red ward edge of the Lyα profiles is defined by where the flux in the line meets the continuum level, and the blue ward edge is defined to be as far from the rest wavelength of Lyα as the point where the red side of the profile hits the continuum, i.e., we use for the blue side of the line the same half-width at zero level observed for the red side. We found that this procedure reproduced the documented equivalent width in Shapley et al. (2003) for their LBG stack within our error on the equivalent width. We estimate our error on the equivalent width as described in Section 3.2, except we use the jackknifed error on the stack to produce Gaussian variates on the stack, since this error is much larger than the continuum rms in the vicinity of Lyα. Following this procedure, we find that the candidate stack is characterized by much stronger Lyα equivalent width (20.8 ± 2.1 Å) than the non-candidate stack (13.3 ± 0.8 Å). Furthermore, we note that non-candidates exhibit a larger equivalent width of blue shifted Lyα than the candidates, and makes up a higher fraction of the total equivalent width in Lyα (about 15%, compared to 7% in the candidate stack).

We compare our measurement of Lyα equivalent widths to the general trend of equivalent widths in Lyα and low-ionization interstellar absorption lines found among z > 3 LBGs in Figure 13, which is adapted from Jones et al. (2012). The figure shows that, in general, the Lyα equivalent width correlates with the equivalent width of low-ionization interstellar absorption lines. The sense of the correlation is that the larger the equivalent width of low-ionization interstellar absorption lines the smaller the emission equivalent width of Lyα, or the larger the absorption equivalent width of Lyα (Shapley et al. 2003; Vanzella et al. 2009; Jones et al. 2012; Berry et al. 2012). This is generally interpreted to indicate that Lyα is present in galaxies with patchy neutral covering fraction of gas in the ISM, or clumpy ISM (e.g., Shapley et al. 2003; Quider et al. 2009; Jones et al. 2012, 2013), since for Lyα to escape the galaxy it must travel freely without neutral hydrogen in the foreground. Alternatively, Lyα can escape if it is redshifted out of resonance (e.g., Verhamme et al. 2008; Steidel et al. 2010).

Zoom In Zoom Out Reset image size

Figure 13 also shows the general relation between the equivalent widths of the low-ionization ISM lines and Lyα for z  ∼  4 LBGs that have been split into bins according to each of the following properties; equivalent width of Lyα (W_Lyα), UV luminosity (M_UV), stellar mass (M*), UV slope (β), and SExtractor half-light radius (r_h). The cluster of colored points at low Lyα equivalent width are half of the z  ∼  4 galaxies, which are each the brightest and more massive, and have redder UV slopes and larger radii. Jones et al. (2012) note that all trends in Figure 13 with galaxy properties listed above are due to their correlations with either the neutral gas covering fraction, or the kinematics in the ISM.

The non-candidate stack has interstellar absorption and Lyα emission consistent with the general trend of z  ∼  3 LBGs. However, the candidate stack does not, and instead has stronger Lyα emission than expected given the large (negative equivalent width) absorption.

Differences in morphology appear to correlate with this trend, as can be seen by the purple points in this figure. When split by half-light radius, larger galaxies make up the low equivalent width of Lyα part of the trend, while small galaxies have stronger Lyα. Vanzella et al. (2009); Law et al. (2012) find a relationship between Lyα and half-light radius, such that smaller galaxies are more likely to have Lyα in emission (see also Malhotra et al. 2012), although Pentericci et al. (2010) find no dependence of Lyα properties and UV morphology. Our two samples are not explicitly split in size, they are roughly split in the mass-size diagram (see Figure 1) and therefore comparison in the context of the trend with galaxy size may be complicated by the overlap in their size distributions. However, if we compare the candidates to galaxies in the small size bin, the discrepancy between the candidates and the LBG trend of Lyα and low-ionization lines becomes even larger. Given their smaller size on average, it might be expected that candidates should exhibit weaker interstellar absorption than non-candidates. The candidates, however, are not just smaller, they are denser and have a generally larger stellar mass and SFR such that, if they quench soon after z  ∼  3, they are passively evolving by z  ∼  2 and have stellar mass M > 10¹⁰ M_☉. We will discuss plausible interpretations of this strong deviation among candidates in Section 4.

In this section, we investigate any evidence in the data that either sample is affected by AGN activity. In Williams et al. (2014), the X-ray and infrared properties of the parent samples of candidate and non-candidate LBGs were studied. Only three non-candidates were detected in the Chandra 4 Ms data, and X-ray stacking indicated no statistical difference between the two samples. Additionally stacks in the Spitzer 24 μm and Herschel 100 μm images indicated no significant average IR emission among candidates, and a minor (4σ) 24 μm detection among non-candidates, i.e., possible evidence of marginally larger dust obscuration. For the spectroscopic subsamples studied here, no galaxies are X-ray detected, three galaxies are detected in the 24 μm images (two candidates, one non-candidate) and one of those two candidates has a 100 μm detection based on the 24 μm position. The high number of non-detections may very well be due to the fact that fluxes at these wavelengths for z > 3 galaxies are below the detection limits of the survey, rather than indicating a lack of (faint) AGN presence. To test for this possibility we have repeated the X-ray stacking in the Chandra 4 Ms imaging for this subsample of LBGs as in Williams et al. (2014). We find that the spectroscopic candidates show a very marginal detection in the soft band only, with only 1.6σ significance. The average luminosity implied is 3.8×10⁴¹ erg sec⁻¹cm⁻², too low to be the result of AGN activity. There is no significant detection in the hard band, suggesting that the soft-band detection, if real, is most likely due to star formation rather than AGN activity. The non-candidate sample do not show significant stacked signal in either band.

Figure 14 shows that candidates have larger equivalent width of C iv both in the absorption and emission components of the P-Cygni profile, as well as He ii in emission. The non-candidate spectrum, however, exhibits smaller equivalent widths in both absorption and emission components of C iv and have no detected He ii in emission. We measure the equivalent widths as in Section 3.2, and list the measured equivalent widths for each sample in Table 2. In this section, we present analysis of AGN contribution and constraints on the properties of their stellar populations based on these two features in the spectra of the two samples.

Zoom In Zoom Out Reset image size

3.5.1. Emission Lines and AGNs

3.5.2. Contribution from Massive Stars and Their Properties

Both the C iv P-Cygni feature and the He ii emission line, which originate from Wolf–Rayet (WR) and other massive stars, are specifically related to the stellar metallicity, because the higher opacity results in increased radiation pressure (Vink & de Koter 2005). Additionally, the mass loss rates for WR stars appear to be metallicity dependent, with higher mass loss rates for increased metallicity (Vink & de Koter 2005). Thus, the ratio of WR stars to O type stars correlates with metallicity, in the sense that at higher metallicity, more O stars evolve through a WR phase because the mass threshold for evolving through the WR phase is lower. At lower metallicity the threshold for a WR phase is at a much higher mass, and therefore, for a given IMF, fewer stars will become WR (Maeder 1991; Meynet 1995; Crowther et al. 2002; Gräfener & Hamann 2005; Meynet & Maeder 2005; López-Sánchez & Esteban 2010). Therefore, an increase in the metallicity of a galaxy increases the likelihood of observing WR-associated features (Brinchmann et al. 2008b, 2008a).

The presence and luminosity of the He ii emission line in the spectra of WR stars also depends on the metallicity (Crowther & Hadfield 2006), and in low-metallicity WR stars, the line may be absent altogether. Brinchmann et al. (2008b) find that in an (instantaneous) starburst model, the He ii emission from O-type stars will increase until the emission peaks during the WR phase of the stars formed in the burst. In the case of continuous star-formation, however, the equivalent width of He ii will eventually plateau at equivalent width ∼1.5 Å until star-formation ceases, provided the metallicity is at least half solar (Brinchmann et al. 2008b). It is interesting to note that in Table 2 we measure an equivalent width in He ii of 1.9 ± 0.3 Å, roughly consistent with their value. In any case, the presence of the line in the candidate average spectrum and its absence in that of the non-candidates suggests that the former have higher metallicity than the latter, in agreement with the relative strength of the C iv feature.

In the rest of this section, we investigate evidence for metallicity differences using what indicators we have available in the rest-frame UV. Rix et al. (2004) established a variety of UV metallicity indicators, based on photospheric blends of iron in massive stars (the 1978 Å index). Due to the high-redshift of our galaxies and relatively noisy spectra, we do not have good signal to noise in this spectral region of our stacks. Other metallicity indices exist based on photospheric blends between 1370–1500 Å (Leitherer et al. 2001; Rix et al. 2004; Sommariva et al. 2012), in regions of our stacks with higher signal-to-noise. These indices have been used as metallicity indicators using bright lensed LBGs (Quider et al. 2009, 2010; Dessauges-Zavadsky et al. 2010), and also on a stack of 75 spectra at z  ∼  2 (Halliday et al. 2008). It is unclear if with the noisiness of the continuum in our stacks and the redshift uncertainties that these would provide robust measures of metallicity. However, as mentioned above, stellar wind lines in the UV are metallicity dependent, and Keel et al. (2004) also noted the metallicity dependence of the C iv P-Cygni profile properties in observations, very similar to those in spectral synthesis codes (Rix et al. 2004; Eldridge & Stanway 2009), such that both the absorption and emission components increase their equivalent width with increasing metallicity. Such properties of the winds of hot massive stars, and their variation with metallicity, have been incorporated into the Starburst99 library of theoretical UV spectra (Leitherer et al. 1999, 2010).

Although it is not possible to use the exact equivalent width to estimate the average metallicity of the galaxies in the stack using these models, we can use them in a relative way to try to estimate how much more metal rich the candidates may be than the non-candidates, based on their stellar wind signature. To illustrate further the dependence on metallicity of both the emission and absorption component of the C iv P-Cygni profile in (purely stellar, excluding interstellar absorption) spectra, we have calculated the equivalent width of both components in Starburst99 models of varying metallicity. We have generated models at the five metallicity intervals available (0.05, 0.2, 0.4, 1, 2 Z_☉) with continuous star-formation at an age of 60 Myr. We convolve the model spectra with a Gaussian of FWHM matching that of our observed VLT spectra, and calculate the equivalent width in both the emission and absorption components of C iv from each model. To illustrate the variation in C iv properties between any two models of differing metallicity, we calculate the equivalent width ratio (i.e., equivalent width of C iv₁/equivalent width of C iv₂) and plot it as a function of metallicity ratio, z₁/z₂ in Figure 15. The left panel of Figure 15 shows this for the stellar C iv absorption component from the models (points), where it is obvious that the equivalent width ratio in stellar absorption is strongly correlated with the metallicity ratio. The equivalent width ratio is related to the metallicity ratio as a well-fit power-law. The C iv absorption in our spectra, however, are a blend of both stellar absorption as in the Starburst99 models, but also interstellar absorption, and from these data it is impossible to separate the two components. Therefore, we cannot use the measured C iv absorption as an estimate of metallicity.

Zoom In Zoom Out Reset image size

The emission component, however, both in the models and our observed stacks is indeed stellar in origin. In the right panel of Figure 15, we repeat the same comparison for the emission component. The variation in equivalent width of the emission component is a bit more complicated than the absorption, but a trend of increasing equivalent width with metallicity is still apparent. We measure the observed ratio of equivalent width in the C iv emission in the candidate stack to that of the non-candidates, and include this as a horizontal line in the right panel of Figure 15. The dashed lines represent the propagated errors from the equivalent width measurements. Although the complicated nature of the equivalent width ratio of emission as a function of metallicity ratio precludes a robust translation of equivalent width ratio to value of metallicity ratio (as for example might be possible using the linear fit to the absorption component), the intersection of the observed ratio with the relation from the models clearly corresponds to a ratio larger than 1, indicative that the candidates have larger metallicity than non-candidates.

Although the emission components of C iv are stellar in origin, their equivalent width may be eroded by interstellar absorption near systemic, and it is impossible to tell if it preferentially does so in one sample versus the other. Therefore, we see if the C iv absorption component at least agrees with our conclusion of increased metallicity after some simple assumptions.

In both stacks, C iv absorption equivalent width exceeds that of the other ISM lines on average by a factor of two, indicative of the additional contribution of stellar absorption. To compare the measured absorption component of C iv equivalent width from the stacks to the Starburst99 models, we first use the other ISM absorption lines as proxies for the interstellar component in the absorption of C iv. We do this two conservative ways. We first subtract an estimate of the contribution to the equivalent width from interstellar absorption using the largest equivalent width for the high-ionization line Si iv, which comes from the Si iv λ1393 line in both samples. We then calculate the equivalent width ratio of candidates to non-candidates, using these ISM-corrected stellar C iv equivalent widths. This result, with the propagated errors on the equivalent width measurement, is shown as purple horizontal lines in the left panel of Figure 15. We repeat the ISM-corrected equivalent width ratio using a more conservative estimate of the ISM contribution, by using the largest equivalent width of all ISM lines we measure, that of the blend O i/Si ii λ1303 (orange). We find that in both cases, the ratio of equivalent widths imply that candidates are more metal enriched relative to non-candidates, by at least a factor of two, and quite possibly larger, in agreement with the findings from the emission component.

Figures 6 and 13 show that on average the candidates exhibit strong, redshifted Lyα emission, in combination with large equivalent width of interstellar absorption lines, characteristics that strongly deviate from the overall trend observed in LBGs at z  ∼  3. Strong Lyα emission may escape a star-forming galaxy if the ISM has patchy distribution around the production regions, i.e., if there are holes punched in the neutral medium by ionizing sources, or, if the line is generated in the receding part of the outflow and is redshifted to an extent that its wavelength is too long for resonant scattering by the intervening ISM. In this section, we discuss the evidence for each of these processes in our samples, and the implications they have for energizing the ISM.

4.1.1. Holes in Neutral Gas and the Covering Fraction

4.1.2. Redshifted Lyα

The stronger interstellar absorption of the candidates implies a larger spread of velocities in the absorbing trough, which in turn suggest that their gas kinematics is more turbulent than the non-candidates, likely as the result of a more extreme feedback in their denser environment. This is also in qualitative agreement with the faster velocity spread of the outflowing component of the candidates. The relatively exceptional ISM of the candidates is illustrated in Figure 13, which shows that while the equivalent widths of Lyα emission and interstellar line absorption of the non-candidates follow the general trend observed for LBG at z  ∼  3–4, the candidates markedly deviate from it because their Lyα emission is too strong given the amount of interstellar absorption. In fact, given the equivalent width of the interstellar lines of the candidates, their Lyα should have been observed in absorption had they followed the general trend.

Another evidence that the Lyα emission from the candidates is, on average, significantly different from that of the general LBG population is illustrated in Figure 16 (adapted from Shapley et al. 2003; Shibuya et al. 2014). This figure shows another trend characteristic of both LBGs and also Lyα emitters at 2  <  z  <  3, which is that the velocity offset between Lyα and ISM absorption lines decreases as the equivalent width of the Lyα increases. Galaxies with stronger Lyα emission exhibit lower velocity offsets, whereas galaxies with Lyα in absorption exhibit the larger velocity offsets. Other studies find similar trends (Vanzella et al. 2009; Jones et al. 2012; Song et al. 2014), although Berry et al. (2012) do not find this correlation very significant on an individual galaxy basis, presumably due to the larger scatter in the measurements from individual spectra. Our two average spectra show that while the non-candidates are in very good agreement with this general trend, the candidates strongly deviate from it, in the sense that their average velocity offset (the mean value of all lines from Figure 12 is ∼970 km s⁻¹) is much larger than that of the other galaxies given the observed Lyα strength. Again, given the observed velocity spread, the Lyα of the candidates should have been observed in absorption had the general trend be followed.¹⁷ We suggest that winds accelerated to very high speed very close to the galactic center, in combination with fast outflows could produce the deviation from the global trend of LBGs.

Zoom In Zoom Out Reset image size

The redshifted, high equivalent width Lyα seen among candidates could also be a result of Lyα scattering off H i in the immediate circumgalactic medium, and in particular, the outflowing gas. Diffuse Lyα halos have been seen among LBG at z  ∼  3, regardless of their spectroscopically determined Lyα equivalent width (i.e., emission or absorption; Steidel et al. 2011). These diffuse halos are faint relative to the UV continuum emitted by the central galaxy but appear to be generically present, independent of the morphological properties of LBG. In some cases, the total Lyα flux from halo regions between 2 and 4 arcsec around the galaxies is similar to that being emitted from the central 2'' (see Hayashino et al. 2004). The main driver of the Lyα emission in the halos is likely the fraction of photons that are able to escape the central regions (Steidel et al. 2011).

The faintness of these diffuse halos can to some extent be attributed to a decreased density of the outflowing neutral gas with radius as the outflow increases volume, and the fact that escaping Lyα photons end up distributed over the much larger projected area of the halo. A simple model of such outflowing media was outlined in Steidel et al. (2010), who showed that scattering of Lyα photons off circumgalactic gas in a high-velocity, spherically symmetric outflow with a uniform velocity field (i.e., independent of galactocentric radius) and a covering fraction that is radially dependent could explain the redshifted Lyα emission seen in LBGs. In this context, it is possible that the high-velocity outflowing medium immediately surrounding the candidate galaxies at small galactocentric radius could be capable of redshifting the Lyα we observe. As discussed in Steidel et al. (2011), if strong Lyα is observed in combination with strong interstellar absorption, the two must be originating from physically distinct regions, or the Lyα would be modified (e.g., the global trend in Figure 13 seen by Shapley et al. 2003; Jones et al. 2012). It is possible that, due to the strong outflowing medium in combination with the compact sizes of our galaxies, that dense, outflowing circumgalactic material at small galactocentric radius may scatter sufficient Lyα to reproduce the Lyα profiles we observe.

Mapping scattered light from outflowing material outside the ISM of the galaxy may confirm or reject this interpretation either through narrow band imaging or spectroscopy (e.g., Steidel et al. 2011; Martin et al. 2013). We conclude based on this data that the deviations of candidates from the trends of LBGs in terms of their Lyα properties and other ISM properties is an indication of a larger fraction of gas moving at higher speed, suggestive that feedback is more effective in compact LBGs.

4.1.3. Implications

In Section 3 we presented evidence that the compact LBG studied here have fast (weighted mean velocity −426 km s⁻¹) blueshifted velocity offsets measured from ISM absorption lines, with significantly larger equivalent width seen among candidates. Understanding the origin of the large velocity spread in the wind is important in the context of star-formation quenching. The presence of these strong winds in candidates, which are clearly more powerful than those in more normal SFGs of similar mass at similar epoch, i.e., the non-candidates (weighted mean velocity −262 km s⁻¹), is an indication that the conditions in the ISM of candidates may be more hostile to cold gas, and lends support to the hypothesis of Williams et al. (2014) that the candidates may quench sooner than normal SFGs. In this section, we discuss the plausible origins for this outflowing component.

Conventionally the generation of fast winds (∼1000 km s⁻¹) have been attributed to AGNs (Trump et al. 2006; Tremonti et al. 2007; Sturm et al. 2011; Rupke & Veilleux 2013; Gabor & Bournaud 2014; Genzel et al. 2014), but see also Coil et al. (2011). For candidates, the properties of emission lines from our limited data do not indicate evidence of a detectable AGN. The rest-frame UV spectroscopic properties of AGNs among LBGs at z  ∼  2–3 have been studied by Steidel et al. (2002); Hainline et al. (2011), and do not share similar properties with our candidate sample. If an AGN is present but obscured, this could prevent its presence from being detected in the He ii/C iv emission line ratio. In a recent analysis of a compact post-starburst galaxy from the sample of candidate progenitors of local passive galaxies (Marchesini et al. 2014), Cemile Marsan et al. (2014) do find evidence of an obscured AGN using rest-frame optical emission lines. Their galaxy, like ours, is not X-ray detected, and has a similarly small C iv/He ii emission line ratio (∼0.8). However, Rupke et al. (2005) and Krug et al. (2010) found that obscured AGNs do not launch high-speed large-scale outflows, and rather only launch local outflows near the central black hole. Further analysis including rest-frame-optical spectroscopy will be necessary to investigate further the presence of AGNs and its effect in driving winds in compact LBGs.

Whether or not star-formation driven (or supernova (SN) driven) winds are capable of halting star-formation on galaxy scales is still a matter of debate. Starburst driven winds often occur with lower velocities (∼500 km s⁻¹; e.g., Fabian 2012). However, evidence that star-formation can launch high-velocity winds is accumulating, especially at high SFR surface densities, (Heckman et al. 2011; Diamond-Stanic et al. 2012; Bradshaw et al. 2013; Sell et al. 2014).

Our candidates have on average larger SFRs (although overlapping distributions in SFR), but have distinctly higher surface density of star-formation (see Figure 1). As discussed in Section 4.1, their surface density of SFR are similar to those of the DCO, whose outflows were determined to be star-formation driven (Overzier et al. 2009; Heckman et al. 2011). Therefore it is possible that the surface density of star-formation is related to our detection of differences in outflow velocities. (Nevertheless, as we showed in Section 3, our data is also consistent with the scaling relation presented in Martin (2005) between velocity and SFR.) Based on conflicting evidence in the literature, it is unclear if outflow velocities in SFGs are universally correlated with SFR or SFR surface density (Martin 2005; Rupke et al. 2005; Erb et al. 2012; Weiner et al. 2009; Sato et al. 2009; Rubin et al. 2010; Steidel et al. 2010; Chen et al. 2010; Kornei et al. 2012; Law et al. 2012; Bordoloi et al. 2014; Lee et al. 2013b; Song et al. 2014). With this data it is not possible to identify whether star-formation or surface density of star-formation is more strongly correlated with gas velocity in our galaxies.

Our detection of significant populations of massive evolving stars (i.e., the WR population), which currently affect the ISM with their strong winds, is consistent with this idea that the outflowing gas we detect are star-formation driven. The feedback from radiation and winds of the massive stars themselves (while on the main sequence) is especially significant at high stellar surface densities, leading (Hopkins et al. 2010) to conclude that feedback from massive stars sets a limit on stellar surface densities in galaxies (Σ_max  ∼  10¹¹ M_☉ kpc⁻²; similar to that observed in z  ∼  2 ultra-compact passive galaxies; Cassata et al. 2011, 2013). These WR stars are expected to form SNe on short timescales. It has been shown that SNe from concentrated groups of young stars, such as that in large star-clusters, may be more effective at evacuating gas than isolated SNe (e.g., Murray et al. 2011; Nath & Shchekinov 2013; Sharma et al. 2014). Therefore, in such compact galaxies as our candidate sample, the compact stellar configurations of both the massive stars and SNe could thus be very effective at evacuating gas, compared to more extended distributions of stars.

The implication that candidates exhibit higher metallicities may be important for the efficacy of feedback in the near future (i.e., over the next Gyr) as these galaxies evolve. If a significant portion of the feedback in these galaxies comes from a dense concentration of high-mass stars with strong stellar winds affecting a very concentrated distribution of cold ISM gas, this feedback may increase its impact as the metallicity of the galaxy continues to increase.

The strength of winds from WR stars in particular, the features of which we have observed in the candidate sample, is dependent strongly on metallicity (Vink & de Koter 2005), with wind velocity and mass loss (and therefore momentum) increasing with increasing metallicity (e.g., Leitherer et al. 1992; Kudritzki & Puls 2000; Crowther 2000, 2007; Nugis & Lamers 2000; Crowther & Hadfield 2006; Mokiem et al. 2007). Over time, therefore, as more metal enriched young stars produce stellar winds, the energy deposition to the surrounding dense ISM is expected to increase.

Although calibrating our metallicity measures onto an absolute scale is not possible with our data, and therefore it is unclear that candidates are already at metallicities high enough to be consistent with those of passive galaxies, the suggestion of enhanced metallicities does provide evidence for a more advanced evolutionary state. It is not currently possible to identify differences in the star-formation timescale of the two samples. Future robust measurements of metallicities and abundance ratios may be able to differentiate between a scenario where the candidate sample form stars on a faster timescale and quench earlier, or, form earlier on a similar timescale (but formed stars longer).

The presence of strong WR features, which arise during relatively short phases (10⁵ yr) relative to the much larger timescale probed by the stack, argues for a relatively continuous SFH, perhaps driven by cold accretion, rather than burst-like star-formation from mergers that could elevate their star-formation during short periods of time. We argue that that the enhanced metallicity of the candidates relative to the non-candidates is consistent with the idea that the former form their stars on a faster timescale than the latter during a period of sustained dissipative accretion of gas. Direct constraints on this interpretation will be possible with future NIR spectroscopic observations that would not only allow robust systemic redshifts to be measured, but also metallicity measurements, including α/Fe abundances, which could directly test our interpretation that these galaxies follow a faster evolution than normal main sequence galaxies.

As detailed in Williams et al. (2014), the candidates have been selected based on their morphology, stellar mass and SFR to reproduce the observed properties of the CPG at z  ∼  2, assuming that they quench their star formation activity shortly after the epoch of observation. Is this assumption well posed and is there any evidence that supports it?

We have provided evidence that the kinematics of the ISM of the candidates shows that they are subject to a more energetic feedback than the non candidates. The half-light radius of the candidates is ≈2 × smaller than the non-candidates, and since their average stellar mass is the about the same, the stellar density is ≈8 × higher. The velocity centroid of the outflowing component of the gas in candidates also is ≈2 × faster. Thus, the mechanical energy density of the ISM of the candidates is roughly 32 × larger than that of the non candidates. Since the SFR of the candidate also is larger by about 2 ×, the radiation energy density contributed by massive stars is also larger by the same amount.

The larger energy deposited in the ISM of the candidates is likely to increase its turbulence, temperature and ionization state, and this is broadly consistent with the observed Lyα properties, strength of the IS absorption lines and presence of He ii emission. The larger SFR is also likely to result in large metallicity, which is also consistent with the stronger C iv emission of the candidates. Their larger SFRs would also produce more dust, except that the larger radiation field and higher ISM temperature are likely to be conducive to destruction of dust grains more effectively than in the non-candidates, a scenario that is supported by the far-IR observations that show that dust-reprocessed emission is detected among the non-candidates but not among candidates, despite their larger SFR (Williams et al. 2014).

There is no evidence for AGN activity either in the X-ray emission or in the UV spectra of the candidates (as well as the non-candidates). So, if the quenching has to come by feedback, then this is likely to be provided by star formation through heating and removal of gas. The outflow velocity of both candidates and non candidates are very likely larger than the escape velocity from the regions where the stars are forming, $v_{esc}\,{\sim}\,\sqrt{2\, G\, M_{{\rm star}}/r_e}\approx 300$ km s⁻¹ for the candidates and ≈200km s⁻¹ for the non candidates if the dark matter gravity can be neglected within the radius where star formation is observed. We have no direct information about the mass loading of the outflows, but if it is roughly proportional to the SFR (Heckman et al. 2011), then over a similar period of time, e.g., from z  ∼  3 to z  ∼  2, the candidates will have ejected one order of magnitude more gas in ≈1/5 of the time than the non candidates, because their SFR is higher (2.5 ×), the velocity of outflowing component of ISM gas is larger (2 ×)and their size is smaller (2 ×). Combined with heating the gas to a larger temperature, this might be sufficient to quench them.

Thus, it seems to us physically motivated to think that the candidates might quench their SFR sooner and in a shorter time scale than the non-candidates. This supports the validity of the selection criteria (which was based solely on morphology, stellar mass and SFR) that we adopted to identify them as progenitors of z  ∼  2 CPGs. Unfortunately, we are not aware of any way at this time to quantitatively predict, based on the observed properties of the candidates, e.g., their ≈30 × energy density relative to the non-candidates, when and how fast they will quench. An additional question is whether the feedback is sufficient to prevent ejected gas recycling, such that cooling gas is prevented from refueling future star-formation. Future observations at (sub)millimeter wavelengths of the dust content and cold gas morphology and kinematics, e.g., with ALMA, as well as future refinement of theoretical modeling of feedback, will provide a path to progress.