STELLAR POPULATIONS OF ULTRAVIOLET-SELECTED ACTIVE GALACTIC NUCLEI HOST GALAXIES AT z ∼ 2–3^*

As described in Section 2, there is ${\it UG}\cal {R}{\it JK}$ photometry for the majority of our sample, but only 16 AGNs have IRAC measurements. Of these 16 objects, 11 are characterized by "power-law" emission in the rest-frame near-IR, where the fluxes in the IRAC bands increase monotonically toward longer wavelengths. This emission is hypothesized to arise in an AGN from thermal or non-thermal emission near the dusty active nucleus (Rees et al. 1969; Neugebauer et al. 1979; Elvis et al. 1994). An additional two objects with IRAC coverage (Q1623-BX454 and Westphal-MM47, only including Channels 1 and 2) have flat mid-IR slopes, indicating a weak AGN as compared to the stellar population. As these objects do not have K-band data, we fit their SEDs solely with a stellar population. The final three objects (Q1217-BX46, Q1623-BX747, and Q2233-MD21) only have IRAC data in Channel 1, which was not sufficient for the purpose of discerning the presence of a power law in the rest-frame near-IR. We fit the SEDs for these objects, including the IRAC Channel 1 data, with a stellar population only.

To place our sample of UV-selected IRAC AGNs in context, we compare it to other galaxy samples with coverage in all four IRAC channels. In Figure 3, we plot the 11 objects with IRAC power-law slopes on the color–color diagrams from Lacy et al. (2004) and Stern et al. (2005), which were initially used to select galaxies with red IR colors as possible AGNs. The AGNs with detections in all four IRAC bands are plotted as red points, and, in the left figure, we use dashed lines to indicate the positions of three AGNs with only Channel 4 and Channel 2 photometry. In this left panel, we show the AGN selection box from Lacy et al. (2004) with long dashed lines. Each of our AGNs lies in the detection area. Donley et al. (2012) revised the selection criteria to account for contamination due to star formation, and we indicate the revised AGN selection area with a dark line. All but one of our objects fall within this revised area. On the right, we show the AGN selection criteria from Stern et al. (2005) with long dashed lines, and all but one object falls within this area. HDF-BMZ1156 falls outside of the Stern et al. and Donley et al. selection areas (although it is consistent with being included within the Donley et al. selection area when its photometric uncertainty is taken into account).⁸ The gray points on this diagram are those objects from the Spitzer First Look Survey with clean detections in all four IRAC channels across all redshifts, and illustrate the full range in color–color space spanned by extragalactic sources (Fadda et al. 2004). We have also plotted those objects from our non-active UV-selected z ∼ 2–3 comparison sample with IRAC coverage in all four channels. These points are shown as dark gray squares. Quite strikingly, UV-selected non-AGNs and AGNs occupy largely distinct regions in IRAC color–color space. Only a small fraction of the UV-selected non-AGNs fall within the Donley et al. (2012) selection area, and these objects may comprise a sample of additional AGNs without strong high-ionization emission lines in their rest-frame UV spectra.

Zoom In Zoom Out Reset image size

To compare to high-redshift power-law galaxies, we have plotted a sample of z > 1.4 power-law sources from Park et al. (2010) on both diagrams as blue points. Our active sample spans a large range in IRAC power laws. The specific values for the power-law slope for our objects (α, where f_ν∝ν^α) range between −0.62 and −3.50, with an average (median) of α = −1.75 (− 1.79). These values show the same range as those presented in Park et al. (2010), although the objects in our sample with the most negative slope values, Q1623-BX663 and Q1700-MD174, would be among the reddest of the Spitzer power-law galaxies (the power law for Q1700-MD174 is fit from only the IRAC Channel 2 and Channel 4 points). Our galaxies are thus representative of the full range of high-redshift Spitzer power-law galaxies, and should illustrate AGN properties for a range of SED types. It is notable that the majority of the UV-selected AGNs in our sample with IRAC data have power laws, supporting their identification as AGNs, consistent with the observation in Reddy et al. (2006a) that all but one UV-identified AGN in the GOODS-N field had significant excess flux at 8 μm.

In order to accurately describe the SEDs for our sample, we require a suite of stellar population models and an AGN template. For our sample of z ∼ 2–3 AGNs we chose Bruzual & Charlot (2003, hereafter BC03) stellar population models. Recently, much attention has been focused on the degree to which thermally pulsating asymptotic giant branch (TP-AGB) stars contribute to the near-IR luminosity of a population of 0.5–2 Gyr age stars (Maraston 2005; Maraston et al. 2006; Eminian et al. 2008). Different treatments of TP-AGB stars systematically affect derived stellar mass and ages (Maraston 2005). In order to gauge how the presence of TP-AGB stars affects our inferred stellar parameters, we also applied the stellar population models of Maraston (2005, hereafter Mar05) and S. Charlot & G. Bruzual (2012, private communication, hereafter CB12) to our SEDs. For the BC03 and CB12 models, we adopted the Padova 1994 stellar evolution tracks. We selected solar metallicity models and a Chabrier (2003) IMF extending from 0.1 to 100 M_☉. We adopted solar metallicity models as non-active galaxies in a similar mass range appear to have metallicities only slightly lower than solar (Shapley et al. 2004). It should be noted that using a Salpeter IMF increases the best-fit stellar masses and SFRs by a factor of 1.8 with little effect on the predicted spectral shape. To account for dust extinction, we used a Calzetti et al. (2000) starburst attenuation law, which on average provides an accurate description of the reddening and attenuation of the UV stellar continuum in both nearby and distant starburst galaxies (Reddy & Steidel 2004; Reddy et al. 2006b).

For the BC03 and CB12 models, we use constant star formation (CSF) models. For the Mar05 models, we used an exponentially declining SF model of the form SFR(t)∝exp(− t/τ), with an e-folding time of τ = 20 Gyr as the closest approximation of a CSF model. We chose not to model with more complex star formation histories, such as exponentially declining star-forming models, short-duration bursts superposed on top of an exponentially declining star formation history, or exponentially increasing star formation models. Recent work presented in Reddy et al. (2012) indicates that exponentially declining star formation models result in SED-derived SFR values that are lower on average than those derived from a combination of UV and IR data.⁹ Exponentially increasing star formation models have been shown to apply to galaxies at z > 2, and naturally arise in hydrodynamic and semi-analytic models of star formation, and are suggested by the relationship between SFR and stellar mass observed at high redshift (Maraston et al. 2010; Papovich et al. 2011; Reddy et al. 2012). These histories give SFRs that are similar to the SFRs derived from CSF models, along with older ages (Reddy et al. 2012).

Stellar population models, even those with large contributions from old, red stars, cannot successfully model the infrared emission seen in an IRAC power-law object. In Figure 4, we show in red the distribution in reduced χ² (χ²_red) found by modeling our IRAC AGNs with only a stellar population. We also plot the χ²_red distribution for the non-AGN comparison sample in gray, using the subsample of 231 non-AGNs with SED coverage extending out to 8 μm. The AGNs have higher χ²_red values than the bulk of the non-AGN comparison population, indicating poor fits. For the non-AGNs, 〈χ²_red〉 = 2.3 ± 1.8, which is lower than all but 1 of our IRAC AGNs (Q0142-BX256, which does not have IRAC Channel 4 data, but has large errors on the Channel 3 data point). In the inset we show a CSF fit for one of our objects, HDF-MMD12, an AGN at z = 2.648. The best fit is very poor, with a χ²_red = 36, due to the presence of the strong AGN power-law emission in the IRAC bands. Often, the SEDs of Type II AGN hosts are modeled excluding points redward of the rest-frame optical, as the AGN emission for narrow-lined objects should be minimal at rest-frame UV and optical wavelengths (Barger et al. 2005; Bundy et al. 2008). With our sample of IRAC AGNs, we can model the power-law AGN emission and the stellar emission simultaneously to see exactly how the presence of an AGN affects the best-fit stellar population parameters. We also note that there exists a tail of galaxies without signs of AGN activity in their rest-frame UV spectra with large χ²_red values from SPS-only model fits. We will discuss these objects in Section 4.1, and in a forthcoming paper.

Zoom In Zoom Out Reset image size

We then require a prescription for the non-stellar AGN emission in our dual-component modeling. Often, the non-stellar continuum contribution to the near-IR emission is removed by assuming that this portion of the spectrum can be represented by a power law of the form f_ν∝ν^−α, where f_ν is the flux density at a frequency ν (Alonso-Herrero et al. 2003; Hainline et al. 2011b). Since this method of modeling the AGN emission has no explicit basis in the physics of the AGN, we instead opt for the AGN model from Assef et al. (2010) to describe our AGN infrared emission.¹⁰ The Assef et al. AGN model was created using an iterative method, starting with the Richards et al. (2006) average quasar SED template, modified to better replicate the behavior in the UV predicted by an accretion disk model (in an attempt to minimize, in part, contamination by host galaxy emission). This template was applied to the SEDs of AGNs from the NOAO Deep Wide-Field Survey in the Boötes field, and iteratively updated to fit the full sample of data. The final AGN template is thus physically motivated, and, as the template has been both normalized to a distance of 10 pc and created to have an integrated luminosity of 10¹⁰ L_☉, its application to our data gives us an estimate of the bolometric luminosity of the AGN. The Assef et al. (2010) AGN template describes a Type I AGN, with strong emission in the rest-frame ultraviolet thought to be emitted from the accretion disk, as well as emission in the rest-frame near- to mid-IR from warm and hot dust emission. At the same time, the AGNs that we model with the Assef et al. template were selected to be Type II by virtue of narrow emission lines seen in the rest-frame UV, where accretion disk emission should be obscured by dust. To model our Type II AGNs, following the analysis of Assef et al. (2010), we apply dust extinction to the AGN template using a reddening curve derived from the Small Magellanic Cloud for λ < 3300 Å (Gordon & Clayton 1998), and following the Galactic curve at longer wavelengths (Cardelli et al. 1989). The resulting template should then consist of strong IR emission from hot dust near the active nucleus, which we use to model the IRAC power law.

Since the objects we are modeling were selected to be Type II AGNs, we made a prior assumption of E(B − V)_AGN >1.0. We model these objects by taking a Type I AGN template and applying extinction in order to produce a Type II AGN template to fit the SEDs. For the Type II AGNs discussed here, the rest-frame UV and optical light should be dominated by stellar emission, given the narrow emission lines in the rest-frame UV suggesting significant obscuration of the central engine (Hainline et al. 2011a). Low values of AGN extinction, on the other hand, would result in SEDs with prominent AGN continuum emission even at rest-frame UV wavelengths, as well as broad-line region (BLR) emission. Although the initial modeling results for two systems, HDF-BMZ1384 and Q2343-BX333, indicated best-fit values of E(B − V)_AGN less than 1.0, we rejected such fits on the basis that these are Type II objects. Furthermore, the χ² values for these low E(B − V)_AGN fits were not significantly lower than for fits with E(B − V)_AGN at values similar to the bulk of the sample (E(B − V)_AGN ∼ 2–7). The same lower limit on E(B − V)_AGN is also adopted in Mainieri et al. (2011) for a similar SPS+AGN fitting procedure used to fit X-ray-selected Type II quasars at z > 0.8. Throughout this paper, we report only those fits where E(B − V)_AGN >1.0. We also limit E(B − V)_AGN < 8.0, as larger values for the AGN extinction did not improve the fits for the objects we examined.

We used the Assef et al. AGN template, along with the stellar population models, to fit the observed SEDs of our IRAC subsample. The overall procedure we used is similar to that of Shapley et al. (2001, 2004, 2005), but here the predicted colors are the sum of stellar emission and non-stellar emission from our AGN model. The best-fit model therefore yields constraints on both stellar population and AGN parameters. The stellar parameters that we modeled included dust extinction (parameterized as E(B − V)_SF), stellar population age (t_SF), SFR, stellar mass (M_*), and in the case of the Mar05 models, star formation history (τ). The ages that we used ranged between 50 Myr and the age of the universe at the redshift of the modeled galaxy. This lower limit of 50 Myr was chosen to exclude models with ages younger than the dynamical timescales estimated for typical z ∼ 2–3 LBGs (Erb et al. 2006a; Law et al. 2007). We considered stellar extinctions between E(B − V)_SF = 0.0 and 0.7.¹¹ The AGN parameters included a central source extinction, E(B − V)_AGN, and a normalization constant, N_AGN.

In the first step of the SED fitting procedure, we created total SPS+AGN models for each potential combination of input parameters. The stellar model flux at each wavelength (with a specific star formation history, age, and extinction value) was summed with the AGN model flux (with a specific AGN extinction value, and multiplied by the normalization factor N_AGN). In order to add the AGN template to our SPS models in a meaningful way, we used values of N_AGN that reflected how the SPS models are normalized. The Assef et al. AGN template is normalized to an AGN bolometric luminosity L_bol = 10¹⁰ L_☉, where this luminosity is calculated by integrating the entire template longward of Lyα. In terms of the SPS models, the BC03 and CB12 CSF templates are normalized at 1 M_☉ yr⁻¹ of star formation. Thus, for the BC03 and CB12 CSF fits, N_AGN = L_bol/(SFR × 10¹⁰). The Mar05 models are all normalized to 1 M_☉ of stars, and so N_AGN = L_bol/(M_* × 10¹⁰) for these fits. While we report the best-fitting N_AGN values, it is useful also to compare L_bol between fits, which indicates the strength of the AGN for each object.

The combined AGN and stellar model spectrum was further attenuated by IGM absorption from neutral hydrogen (Madau 1995), which only affects the predicted U and/or G magnitudes, depending on the redshift. These output SPS+AGN templates for a given set of parameters were compared in magnitude space to the optical through mid-IR SED for a specific object (though, not including Spitzer MIPS data, see Section 4.1), and a value of χ² for the fit was calculated. In these fits, we only modeled photometric detections. In a few cases, there were upper limits in IRAC channels that were not modeled (Channel 4 for Q0142-BX256, Channel 3 and Channel 4 for Q1623-BX454, Channel 2 for Q1623-BX74, and Channel 3 for Q2233-MD21). In all cases except that of Q0142-BX256, the IRAC upper limits are entirely consistent with the best-fit SED model. We sought to scale our total models to the observed SEDs such that the calculated χ² was minimized, and this normalization naturally led to an estimation of the SFR and stellar mass. The parameters (E(B − V)_SF, t_SF, SFR, M_*, E(B − V)_AGN, and N_AGN) resulting in the minimum χ² with respect to the measured photometry for an object are referred to as the "best-fit" parameters.

Flux contribution from strong nebular emission lines excited by the presence of an AGN can be a significant source of contamination in broadband SEDs. As described in Hainline et al. (2011a), the rest-frame UV spectra for the z ∼ 2–3 Type II AGNs described here include emission lines such as Lyα, with an average rest-frame EW of 80 Å. This strong emission line, as well as others in the rest-frame UV, such as C ivλ1549 and He iiλ1640, can enhance the flux in the U or G bands (depending on redshift), and alter the observed U − G or $G-\cal {R}$ colors. Rest-frame optical features such as Hα or [O iii]λ5007 can similarly add flux in the near-IR, and subsequently affect the measured $\cal {R}-K$ colors. We adopt a method of correcting for these emission lines outlined in Papovich et al. (2001), where a rest-frame EW, EW₀, for an object at a redshift, z, introduces a change in the magnitude:

$$\begin{equation} \Delta m = -2.5\, \mathrm{log} \left[1 + \frac{{\rm EW}_0 (1+z)}{\Delta \lambda } \right]. \end{equation} \tag{ 1 }$$

In this equation, Δλ is the width of the filter bandpass. We calculated Δm for each measured emission line, and, where its absolute value was larger than the photometric error, we corrected for this magnitude difference to produce the final photometry.

Strong emission lines do not affect the U or G bands for 19 out of 33 AGNs in our sample at 2.17 ⩽ z ⩽ 2.48. For the remainder, we used UV spectra (described in Hainline et al. 2011a) to measure EW values for the strong emission lines such as Lyα, C iv, and He ii. Lyα was the chief contaminant to the G band for this redshift range, with the average (median) value of the magnitude difference Δm = 0.36 (0.34). The other UV lines were not strong enough to significantly contaminate the photometry: in objects where C iv was detected, the EW was on average (median) 48% (17%) of the EW of Lyα, and similarly, in objects where He ii was detected, the EW was on average (median) 25% (13%) of the EW of Lyα. However, for one object (HDF-BMZ1156), very strong C iv and He ii contaminated the G band (at the redshift of this object, Lyα fell blueward of the G band), with a combined Δm = 0.37.

The K band contains Hα for 20 out of 33 AGNs, and [O iii]λ5007 for an additional 4 objects. Near-IR spectra have only been obtained for a subset of our AGNs, as part of an ongoing project to measure the rest-frame optical spectroscopic properties for this sample (Erb et al. 2006b; K. N. Hainline et al., in preparation). For this analysis, there are nine objects with near-IR spectral coverage (eight objects with Hα measurements and one with a measurement of [O iii]), and we have determined the Hα and [O iii] fluxes for the available lines. These spectra do not show a significant continuum, so we used the K-band magnitude in order to estimate the continuum flux density (corrected for line flux), and therefore the EW for Hα and [O iii]. The average (median) magnitude difference from the Hα emission lines is Δm = 0.18 (0.12). It should be cautioned that the majority of our objects do not have rest-frame optical spectra, and Hα or [O iii] could contribute to the K band photometry for these AGNs. However, we expect that the effect should not be large compared to the K-band uncertainties, based on objects with such measurements.

While the uncertainties relating to single-component stellar population fitting are well described in Shapley et al. (2005), the addition of an AGN component to the fits introduces new model degeneracies and systematics. For example, without an AGN component, the inferred extinction, E(B − V), depends on the chosen star formation history and age. For CSF models, a given set of galaxy $G-\cal {R}$ and $\cal {R}-K$ colors can be produced by models with older ages and less dust extinction, or younger ages and more dust extinction.

With the addition of an AGN template to the SPS modeling, we find covariances between AGN model extinction (E(B − V)_AGN) and normalization (N_AGN), as well as between the AGN parameters and the stellar parameters. Within the IRAC sample described in Section 3.1, a given set of photometric values for our individual AGNs can typically be explained by model fits with lower values of E(B − V)_AGN along with lower values of N_AGN, or higher values of E(B − V)_AGN along with higher values of N_AGN. This degeneracy has an effect on the best-fit stellar parameters. For a given object, lower E(B − V)_AGN and N_AGN model fits result in lower values of E(B − V)_SF and correspondingly lower SFRs. As described above, lower E(B − V)_SF fits are also accompanied by older ages. We can explain these covariances qualitatively. For a given intrinsic AGN power law, lower values of E(B − V)_AGN indicate more AGN contamination at shorter wavelengths, and these fits have low values of N_AGN in order to fit the observed SED without overpredicting the $UG\cal {R}$ flux. The increased flux at shorter wavelengths from a low E(B − V)_AGN AGN template is accompanied by lower E(B − V)_SF values, as the increased AGN flux requires less dust to account for the red UV-to-optical continuum ($\cal {R}-K$ color). Our modeling procedure works by adding the AGN template (modified by some dust obscuration, and multiplied by N_AGN) to a stellar population model and then normalizing the result to the observed SED to determine the SFR. As a low-E(B − V)_AGN, low-N_AGN dual-component model has more flux from the AGN at shorter wavelengths, such a model predicted that young stars contribute less flux in the rest-frame UV, which results in lower SFR values. We demonstrate this covariance using Q0100-BX172 as an example in Figure 5.

Zoom In Zoom Out Reset image size

In order to account for covariance among the best-fit stellar and AGN parameters, errors in best-fit parameters were estimated using a Monte Carlo analysis. We created a large set (N = 500) of fake SEDs by varying the observed colors and magnitudes in a manner consistent with the photometric errors. Each fake SED was modeled using the same method as for the actual data, where the best-fit stellar and AGN parameters were those that resulted in a minimization of χ². We estimated the 1σ errors for the best-fit model parameters by calculating the standard deviation of the distribution on each parameter from our Monte Carlo simulation results.

The dual-component SPS+AGN fits successfully reproduce the range in SEDs found in the IRAC AGN subsample. We show the SED fits for the 11 IRAC AGNs in Figure 6. In each panel, the gray curve represents the best-fitting dual-component model, which is decomposed into the sum of fluxes from a stellar population (blue curve) and an Assef et al. AGN template (red curve). Also shown are the χ² and χ²_red values output from the modeling program. The poorest fits, with the highest values of χ²_red, are for those objects where the K-band flux was greater than predicted from the fit, as in HDF-BMZ1156, HDF-BMZ1384, and to a lesser extent Q1700-MD174 and Q2343-BX333. The underprediction of K-band flux could be a result of the choice of star formation history, or the contamination from Hα emission to the flux in the K band. It should be noted that we have a rest-frame optical spectrum for HDF-BMZ1156, and have made a correction for the presence of Hα in this object (Δm = 0.07), and yet the best-fitting model still underpredicts the K band. We have also calculated the fraction of the total emission in the best-fit SPS+AGN model at rest-frame 1 μm that is contributed by the AGN template. These values of f_{AGN, 1 μm} are reported in the bottom right corner of each panel in Figure 6, and express the relative strength between the emission from the AGN and the stellar population. The values of f_{AGN, 1 μm} range from <0.01 for HDF-BX160 and Q1700-MD157 to 0.79 for Q1623-BX663. The best-fitting parameters for the IRAC AGNs are reported in Table 2. The IRAC AGNs have average (median) values and standard deviations for the best-fit parameters of $\langle \mathrm{log(Mass)_{{\rm CSF}}} \rangle = 10.82\,(10.76) \pm 0.22$, $\langle \mathrm{SFR_{{\rm CSF}}} \rangle = 56\,(37) \pm 51$ M_☉ yr⁻¹, 〈E(B − V)_{SF, CSF}〉 = 0.23 (0.24) ± 0.08, and $\langle \mathrm{log(Age)_{{\rm CSF}}} \rangle = 3.23\,(3.21) \pm 0.25$.

Zoom In Zoom Out Reset image size

Based on the range in best-fitting parameters, we can split the IRAC AGN host galaxies into two populations based on the slope of the IRAC power law (Figure 3). The first population contains those objects in our IRAC subsample with shallower observed power-law slopes. Most of the galaxies that comprise this population (Q2343-BX333, Q0100-BX172, Q0142-BX195, HDF-BX160, and Q1700-MD157) have dual-component fits where the AGN only becomes the dominant source of emission at 6–8 μm observed frame (Figure 6). These objects are best fit by models with low values of N_AGN along with a range of E(B − V)_AGN values. The one exception is HDF-MMD12, which has a low value for N_AGN accompanied by strong AGN emission at observed-frame wavelengths shorter than 8 μm. HDF-MMD12 is best fit by a relatively young stellar population (t_SF ∼ 200 Myr). This model SED has a weak Balmer break with respect to those of the other objects, and less flux from stars at longer wavelengths, in comparison to the Assef et al. (2010) AGN template.

The second population of IRAC AGNs consists of objects with much steeper (redder) power-law slopes. These objects (HDF-BMZ1384, HDF-BMZ1156, Q1700-MD174, Q0142-BX256, and Q1623-BX663) are best fit by models with larger values of N_AGN, in which the AGN emission already becomes dominant at wavelengths as short as 5–6 μm in the observed frame. Correspondingly, the rest-frame UV AGN emission lines in the objects with steeper power-law slopes are stronger on average than those measured in the shallow power-law objects. The AGN bolometric luminosities of the steep power-law objects are also larger, on average, than those of the shallow power-law objects. Q1623-BX663 is the most extreme example. This galaxy has a steep power law (α = −2.24), with a large value of N_AGN accompanied by a low value for the AGN extinction. As AGN emission in this object is dominant even to the K band for CSF models, Q1623-BX663 represents a "worst-case scenario" for modeling the AGNs using ${\it UG}\cal {R}{\it JK}$ photometry alone and not including an AGN component. Q1623-BX663 has one of the steepest slopes observed in our IRAC subsample, and it is unlikely that many objects in the full AGN sample will have similar AGN-dominated fits. Together, these AGN populations show a wide range of AGN contamination, and should provide a representative subset for correcting the full sample for AGN emission.

As a sanity check, we apply our dual-component modeling procedure to the subsample of non-AGNs (discussed in Sections 2 and 3.2) that have coverage extending to 8 μm. Half of these objects have best-fit models with N_AGN = 0.00, and the full distribution is highly skewed toward either zero or extremely low values of N_AGN, in contrast to our UV-selected AGNs with IRAC coverage, all 11 of which are fit by models with a non-zero value for N_AGN. We assessed the probability of drawing a sample of 11 non-zero values from the non-AGN N_AGN distribution, using 10⁶ random trials. This test revealed that the likelihood of randomly drawing such a sample from the non-AGN distribution is only 0.05%, again demonstrating the distinct nature of the SEDs of the UV-selected AGNs. We do note that there is a small population of objects in the non-AGN sample that is best fit by models with N_AGN significantly greater than zero. While a discussion of these objects is outside of the scope of the current work, their strong IRAC power-law emission indicates that they may be hosting an AGN, but do not show significant emission lines in their UV spectra. These objects will be followed up in a paper analyzing LBG/BX/BM AGNs selected on the basis of their infrared emission. As these objects comprise only a small fraction of the total non-AGN sample, they will not significantly affect our conclusions regarding the distribution of non-AGN stellar populations.

As discussed in Section 3.2, we also used the CB12 and Mar05 models in our dual-component SPS+AGN fitting. Here we report the differences in derived parameters for the CSF models. For the CB12 modeling, the stellar masses and ages from the fitting are typically 70% of those derived using the BC03 models, while the CB12 SFR and E(B − V)_SF values are statistically equivalent to the BC03 values. The Mar05 models produced similar differences. For the Mar05 modeling, the stellar masses from the fitting are typically 60% of those derived using the BC03 models, and the ages are 50% of those derived using BC03. The E(B − V)_SFR values derived using both models were statistically the same, and the Mar05 SFR values were larger than the BC03 values by 25%. Despite these small systematic differences among stellar population models, our main conclusions regarding the AGN contamination do not depend on the specific stellar population model adopted.

We have MIPS photometry for eight of our objects with IRAC data. We do not use the MIPS data to constrain our SPS modeling, and find that for the majority of the objects with MIPS data, the best-fit AGN template overpredicts the MIPS flux in our objects by a factor of two on average. Examining the range of stellar and AGN parameters that best fit our optical through IRAC data, we find that attempting to fit for the MIPS data leads to best-fit models that systematically underpredict the IRAC fluxes, due to the shape of the reddened Assef et al. template. Analysis of the mid-IR SED of HDF-BMZ1156 by Alexander et al. (2008b) sheds light on the origin of this discrepancy. Alexander et al. find that a Type II AGN template (i.e., the SED of NGC 1068) provides a good fit to the photometry of HDF-BMZ1156 at wavelengths from 16 through 70 μm. However, an additional hot (∼1000 K) dust component, often found in obscured AGNs (e.g., Alonso-Herrero et al. 2001), is required to fit the IRAC data points at shorter wavelengths. If the other AGNs in our sample are similar to HDF-BMZ1156, our use of a single component to describe both IRAC and longer-wavelength data will naturally overpredict the MIPS 24 μm flux, since we have forced a Type II template to fit the IRAC data. At the same time, this difference is not significant for most of the sample. The range of SPS+AGN templates that fit the observed ${\it UG}\cal {R}{\it JK}$+IRAC data within the photometric errors is within 1σ of the MIPS uncertainty for the majority (five out of the eight) of our objects. Since we do not use the MIPS data in our SED fitting, the discrepancies between predicted and observed fluxes suggest that the L_bol values reported in Table 2 may slightly overpredict the true values.

To understand the true AGN host galaxy stellar populations, it is very important to correct for AGN emission in the SED of those objects without IRAC data, where we cannot directly quantify the AGN contribution. Many authors fit AGN host galaxy SEDs using SPS models alone (see Bundy et al. 2008; Bluck et al. 2011 for discussions), and find that while the presence of an AGN does affect stellar-mass estimates, it is not a significant difference. Here we adopt a different approach, and use our results from the SPS+AGN modeling to quantify how much the presence of an AGN affects the stellar modeling parameters. For the 11 objects in the IRAC AGN sample, we refit the SED but only through the K band, and using an SPS-only model. The best-fitting parameters for the IRAC AGNs fitted in this manner are reported in Table 3. The difference in best-fit parameters based on K-band and SPS-only versus K-band+IRAC and SPS+AGN modeling indicates how the majority of the objects in our sample without IRAC coverage must be corrected. It is valid to apply the correction factors derived from the IRAC AGN sample to our entire sample, given that the objects with IRAC coverage have similar average $\cal {R}$ magnitudes and $\cal {R}-K$ colors to those without IRAC coverage. The procedure, star formation models, and parameters are the same as those discussed in Section 3.2, but here we do not include the AGN template in the K-band-only fits.

The resulting stellar masses and SFRs calculated with an AGN component are lower than those calculated without the AGN component. The best-fit ages and values for E(B − V)_SF are consistent between fits with and without the AGN component. For a few objects, such as Q1623-BX663, the best-fit parameters are quite different, as this galaxy suffers from the most AGN contamination at the shorter wavelengths (Figure 6). The results for HDF-BMZ1156 and HDF-BMZ1384 are also discrepant, due to the bright K magnitude for these objects, which is not well fit in the dual-component modeling extending to IRAC wavelengths.

We analyze the AGN contamination by calculating the ratio between the parameters (M_*, SFR, E(B − V)_SF, and age) resulting from fits with the SPS+AGN modeling to the average parameters resulting from the SPS-only modeling, which we refer to as "correction factors." We calculated the ratios for the best-fit parameters, and we also analyze the AGN contamination using results from the Monte Carlo error analysis (Section 3.4). In this method, we calculated the average value for the full distribution of Monte Carlo output parameters, and then divided these averages for our SPS+AGN modeling by the averages from the SPS-only modeling. This second method gives a better understanding of the full set of models that fit an object's SED within the error on the photometry and therefore we only report correction factors calculated in this manner. We note however that the correction factors do not depend significantly on the method used. The correction factors calculated for the 11 IRAC AGNs are shown in Table 4. While our results suggest that the $UG\cal {R}JK$ photometry for Type II AGNs is largely free from AGN emission, the best-fit masses and SFR values derived without including an AGN component are larger by a factor of 1.4 (although, on an object-by-object basis, these factors for mass and SFR are not always the same). The values for E(B − V)_SF are only larger by a factor of 1.1, and the ages are consistent. We use these factors to correct the best-fit parameters based on SED modeling results for those objects where we do not have IRAC data under the assumption that the SPS+AGN model is more appropriate.

Table 4. AGN Correction Factors

Object	Massa	SFRa	Agea	E(B − V)SFa
$\mathrm{Q0100{\rm -}BX172}$	1.0 ± 0.5	1.0 ± 0.4	1.0 ± 0.6	1.0 ± 0.3
Q0142-BX195	1.2 ± 0.6	1.1 ± 0.7	1.3 ± 1.0	1.0 ± 0.3
Q0142-BX256	0.7 ± 0.2	0.8 ± 0.3	0.8 ± 0.2	0.9 ± 0.1
HDF-MMD12	0.9 ± 0.7	0.2 ± 1.1	2.1 ± 1.4	0.8 ± 0.2
HDF-BMZ1156	0.5 ± 0.3	0.6 ± 0.3	0.8 ± 0.2	0.9 ± 0.1
HDF-BMZ1384	0.6 ± 0.3	0.6 ± 0.4	1.0 ± 0.0	0.8 ± 0.2
HDF-BX160	0.6 ± 0.4	0.8 ± 0.4	0.8 ± 0.5	1.0 ± 0.2
Q1623-BX663	0.3 ± 1.0	0.8 ± 0.3	0.4 ± 0.9	0.8 ± 0.1
Q1700-MD157	0.5 ± 0.3	0.8 ± 0.3	0.6 ± 0.4	1.0 ± 0.1
Q1700-MD174	0.9 ± 0.5	0.9 ± 0.4	1.0 ± 0.0	1.0 ± 0.2
Q2343-BX333	1.0 ± 0.3	1.0 ± 0.3	1.0 ± 0.0	1.0 ± 0.1
Averageb	0.7 ± 0.3	0.8 ± 0.2	1.0 ± 0.4	0.92 ± 0.07

Notes. ^aCorrection values represent the derived parameters from the AGN+SF modeling fit divided by the SF modeling fit through the K-band data. ^bUncertainties for the average correction factors represent the standard deviation of the distributions.

Download table as:  ASCIITypeset image

We also calculated AGN correction factors using the CB12 and Mar05 modeling. The correction values calculated for both the CB12 and Mar05 modeling are similar to those calculated using the BC03 modeling. While the mass and age correction values are smaller than those derived from the BC03 models, they differ by less than one standard deviation of the distribution of BC03 correction factors. The corrected best-fit parameters are not significantly dependent on the choice of model (e.g., BC03, CB12, or Mar05).

To fit the objects in our AGN sample lacking multiple IRAC photometric measurements, we used SPS-only models (using the same procedure described in Section 3.2, but without the addition of an AGN model). For each of these objects, we corrected the best-fit mass, SFR, age, and E(B − V)_SF for the presence of an AGN using the average correction factors reported in the previous section, and list the best-fit values (both uncorrected and corrected) in Table 5. The errors on the corrected parameters reflect both the errors derived using the Monte Carlo bootstrapping described in Section 3.4 as well as the standard deviation of the correction factors, given in Table 4. Objects with only three photometric data points have reduced χ² values given by a dash. SED fits for each of these objects are shown in Figure 7. The masses indicated in these figures are the corrected best-fit masses. The χ² values indicate that many of the objects are well fit by the SPS models. In the case of Q0000-C7 and Q1623-BX747, the SEDs were not well fit by the CSF models. For Q0000-C7, a blue $G-\cal {R}$ color leads to models that are unable simultaneously to fit both the UV and near-IR portions of the SED. For Q1623-BX747, the J, K, and Channel 1 photometric data¹² show a decline to larger wavelengths, which is not reflected in the best-fit stellar models.

Zoom In Zoom Out Reset image size

We summarize the full statistics for our sample of UV-selected AGNs in Figure 8. All 28 of the AGNs with fitted SEDs are represented in these histograms. The AGNs with IRAC coverage are plotted with their best-fit parameters from the dual-component modeling, and the AGNs without IRAC coverage are plotted with the corrected best-fit parameters, calculated as described in Section 4.3. The average values for the best-fit parameters and standard deviations are indicated in the figure with vertical lines. There is a large spread in E(B − V)_SF values, with an average (median) of 〈E(B − V)_SF〉 = 0.22 (0.23) ± 0.11, and a range of 0.00–0.40. The best-fit SFRs, with a mean of 〈SFR〉 = 63 (37) ± 67 M_☉ yr⁻¹ similarly span a large range of values between 5.2–309 M_☉ yr⁻¹. The most striking results, however, are the ages and masses for this sample. As suggested by the $\cal {R}-K$ colors shown in Figure 2, the best-fit masses are high: log(〈M_*/M_☉〉) = 10.85 (10.71) ± 0.36. The AGNs that comprise the sample have old ages, with an average age of log(〈age/Myr〉) = 3.19 (3.18) ± 0.26. These average stellar masses and ages are large compared to the typical values quoted for UV-selected non-active galaxies (Erb et al. 2006b; Shapley et al. 2001), and will be discussed further in Sections 5 and 6.

Zoom In Zoom Out Reset image size

We can use the L_bol values derived from the dual-component fitting to estimate accretion rates for the IRAC AGNs. It should be noted that, as described in Section 4.1, the values of L_bol for our sample are potentially too high as a result of the overprediction of the observed MIPS flux, so the derived accretion rates and Eddington ratios should be viewed as upper limits. The accretion rate is described by the equation $\dot{M} = L_{{\rm bol}} / \epsilon _{{\rm rad}} c^{2}$, where we assume a standard accretion efficiency _rad = 0.1. Using the values for L_bol from our fitting, we calculate an average (median) accretion rate for our sample of 0.3(0.1) M_☉ yr⁻¹ with a range of 0.04–0.9 M_☉ yr⁻¹. These rates are similar to the range of accretion rates found for optically selected Type I AGNs at 1 < z < 2.2 from Merloni et al. (2010) (with a median accretion rate of $\dot{M} \sim 0.4\, M_{\odot } \, \mathrm{yr}^{-1}$), and a factor of several lower than those estimated for X-ray-selected Type II QSOs from Mainieri et al. (2011) (with a median accretion rate of $\dot{M} \sim 1$ M_☉ yr⁻¹).

The results from the SED fitting were also used to derive the properties of the central AGN in each of our galaxies. Black hole masses are typically estimated in Type I AGNs using broad emission line widths and AGN continuum luminosities by setting M_BH∝v²r, where r, the size of the AGN BLR, is estimated from the continuum luminosity using the relationship between luminosity and BLR size based on reverberation mapping studies (e.g., Bentz et al. 2009; Barth et al. 2011). For our sample of Type II AGNs, black hole masses were estimated using the M_BH − M_bulge scaling relation from Häring & Rix (2004):

$$\begin{eqnarray} \mathrm{log}(M_{\rm{BH}} / M_{\odot }) &=& (8.20 \pm 0.10)\nonumber\\ && + (1.12 \pm 0.06) \, \mathrm{log}(M_{{\rm bulge}} / 10^{11}\, M_{\odot }).\quad \end{eqnarray} \tag{ 2 }$$

For this calculation, we assumed that the stellar masses derived from our SED fitting roughly corresponded to the bulge masses for these objects. We adjusted the black hole masses for redshift evolution following the parameterization derived for Type I AGNs out to z = 2.2 in Merloni et al. (2010) (but see Lauer et al. 2007; Alexander et al. 2008a):

$$\begin{equation} \Delta \mathrm{log}(M_{\rm{BH}}/M_{*})(z) = (0.68 \pm 0.12)\, \mathrm{log}(1 + z). \end{equation} \tag{ 3 }$$

The values for log(M_BH) for our sample of AGNs have a range of 7.7–8.7, with an average (median) of log(M_BH/M_☉)=8.36(8.29). The typical uncertainty in M_BH is ∼36%. For the 11 AGNs with IRAC data and L_bol measurements, we calculated Eddington ratios (λ_Edd = L_bol/L_Edd, where L_Edd/L_☉ = 3.2 × 10⁴ M_BH/M_☉).¹³ The range is lower than that presented for X-ray-selected Type II QSOs in Mainieri et al. (2011) (in which λ_Edd values were calculated in the same manner). The median Eddington ratio for our sample is 0.03 (with a typical uncertainty of ∼45%), while the median Eddington ratio for the Mainieri et al. sample is 0.1. Therefore the black holes hosted by the AGNs in our sample are accreting at significantly sub-Eddington rates.

A careful examination of the demographics of the UV-selected AGNs requires the selection of a comparison sample of non-active galaxies. In Section 2, we highlighted a sample of non-AGNs that lie in the same redshift range. Recent results have demonstrated the importance of controlling for stellar mass in understanding the relationship between AGNs and their host galaxies (Silverman et al. 2009; Xue et al. 2010; Aird et al. 2012), and so we also constructed a comparison sample that spans a similar mass range to that of the AGN host galaxies. For each AGN in our sample, we chose six non-active sources at random, without any duplicate objects, and characterized by masses within the uncertainty of the selected AGN host mass. The size of the mass-matched sample was limited by the nature of the stellar-mass distribution of the parent non-AGN sample, in which there are only six unique objects for each AGN at the massive end. Thus, the final mass-matched sample has 168 objects, as compared to the sample of 28 AGNs with stellar population fits, and has similar redshift properties to the AGN and non-AGN samples. The redshift distribution of the mass-matched sample is characterized by 〈z〉 = 2.52 ± 0.37, which closely follows that of the UV-selected AGN sample. We will refer to the larger sample of non-AGNs as the "full non-active comparison sample," and the subsample matched in stellar mass as the "mass-matched comparison sample." The average mass for the mass-matched comparison sample is log(〈M_*〉) = 10.85 ± 0.36, which, by construction, is the same as for the AGNs. We modeled the photometry of non-AGNs using BC03 CSF SPS-only models as described in Section 4.3.

We used the full non-active comparison sample to calculate the observed fraction of AGNs as a function of stellar mass. In Figure 9, we show the results, with errors derived from the mass errors for the AGNs that went into each bin. These results indicate that the fraction of AGNs detected rises steadily as a function of stellar mass, peaking at 19% for log(M_*/M_☉) = 11.25. Above this mass, the number of non-AGNs in the full non-active sample drops to only 4 objects, preventing the determination of tight constraints on the AGN fraction. We observe AGNs predominantly in higher-mass galaxies, although as described in Section 6, this trend likely reflects our increasing incompleteness toward lower stellar masses. Accordingly, these results may indicate that at least 19% of UV-selected star-forming galaxies to $\cal {R} = 25.5$ host actively accreting black holes. We return to this discussion in Section 6.

Zoom In Zoom Out Reset image size

Star-forming galaxies follow a tight positive correlation between SFR and stellar mass, which has been observed locally (Brinchmann et al. 2004), as well as out to intermediate (Noeske et al. 2007; Elbaz et al. 2007, z ∼ 1) and high redshifts (Reddy et al. 2006a, 2012; Daddi et al. 2007, z ∼ 2–3). This trend is used to argue for the smooth star formation histories of the majority of star-forming galaxies in the universe (Noeske et al. 2007; Daddi et al. 2007). This relation is characterized by roughly the same slope across a wide redshift range (although results from Noeske et al. (2007) indicate a non-linear relation for star-forming galaxies at intermediate redshifts), while the normalization has decreased by an order of magnitude from z ∼ 2, reflecting the overall decrease in the global star formation rate density with time (Daddi et al. 2007). We plot the SFR and stellar mass for our AGNs with red points in Figure 10. We find a positive correlation for the AGN host galaxies with a best-fit slope of 0.20 ± 0.04, shallower than derived for non-AGNs.

Zoom In Zoom Out Reset image size

We compared the AGN population to our non-active comparison samples in order to examine the origin of the shallow slope. In Figure 10 we plot log(SFR) against log(M_*) for the non-active comparison sample in gray and for the mass-matched comparison sample in blue. The non-active galaxies display a positive trend with a large scatter. The ridge of points seen at the high SFR side of the relation is an artifact of imposing a minimum age of t_SF = 50 Myr on our fits. Given the linear relation between M_* and SFR for CSF models, all points with best-fit ages of t_SF = 50 Myr will fall on a line of constant M_*/SFR. We note that the sharp cutoff for galaxies with low SFRs at a fixed stellar mass is due to the restriction on the maximum allowed age, which must be younger than the age of the universe at the redshift of each object. The lack of points at the low-mass, low-SFR end is due to a bias in the $UG\cal {R}$ selection technique toward selecting those galaxies with the highest SFR at a given mass, causing the slope of the relation to become shallower at the low-mass end (Reddy et al. 2012). At the high-mass end, the shallow slope reported for the AGNs is similar to what is observed among the non-active sources. Above log(M_*/M_☉) = 10.5, the $UG\cal {R}$ selection is biased against selecting those galaxies with the largest SFR values at a given stellar mass, as such objects will have large amounts of dust extinction and red $UG\cal {R}$ colors, placing them outside the $UG\cal {R}$ selection window. This bias tends to flatten the log(M_*) − log(SFR) relation at the high-mass end. Of particular interest, the SFR values spanned by the UV-selected AGN sample ((〈SFR〉) = 63 (37) ± 67 M_☉ yr⁻¹) are similar to those in the mass-matched comparison sample, where the average (median) SFR is 〈SFR〉 = 45 (35) ± 38 M_☉ yr⁻¹. These similarities are observed over the full AGN and mass-matched comparison samples, as well as in bins of stellar mass. We quantify them using the Kolmogorov–Smirnov two-sample test, which yields a 44% probability that the AGN and mass-matched non-AGN SFR distributions are drawn from the same parent population.

The above trends can also be considered in terms of the specific star formation rate (sSFR = SFR/M_*, which, for CSF models is equal to 1/t_SF). We calculated an average (median) sSFR =1.36 (0.66) Gyr⁻¹ for the AGN hosts, with a range of sSFR: 0.4–9.2 Gyr⁻¹, indicating active star formation.¹⁴ We find that the sSFR values for our mass-matched non-active sample span a similar range, with an average (median) sSFR =1.10 (0.47) Gyr⁻¹, and a range of sSFR: 0.33–8.78 Gyr⁻¹. The sSFR values for our Type II UV-selected AGNs are similar to those obtained for a sample of X-ray-selected Type II QSOs with active star formation at z ∼ 2 from Mainieri et al. (2011) (sSFR = 0.8–3 Gyr⁻¹ for the high-mass objects at z ∼ 2). These authors found consistent evolution for the sSFRs of the QSO host galaxies and those of non-active BzK-selected star-forming galaxies from Pannella et al. (2009). Similarly, once we control for stellar mass, the SFR and sSFRs of AGNs are indistinguishable from those of non-AGNs.

A basic probe of the demographics of AGN activity is provided by the optical color magnitude diagram, which can also be cast in terms of galaxy color as a function of stellar mass. In the local universe, galaxies display a bimodality in their observed colors, separating into a population of lower-mass blue star-forming galaxies and a sequence of higher-mass red quiescent galaxies (e.g., Baldry et al. 2004). This same bimodality has been observed in galaxies at intermediate (z ∼ 1, Bell et al. 2004) and higher redshifts (z ∼ 2, Cassata et al. 2008; Kriek et al. 2008; Williams et al. 2009). The existence of these two populations out to z ∼ 2 suggests that the mechanisms driving the evolution of galaxies from blue to red began early in the history of the universe. AGNs have been proposed as a potential driver of the evolution in the color–mass relation by providing energy to the ISM that quenches star formation, causing a galaxy to transition from the blue population of galaxies to the red (Croton et al. 2006; Hopkins et al. 2008). In support of this scenario, X-ray and optically selected AGNs at low and intermediate redshifts are found to reside in galaxies in the transition region in optical color space between the blue and red sequences (Nandra et al. 2007; Salim et al. 2007; Coil et al. 2009; Hickox et al. 2009; Schawinski et al. 2010). However, recent AGN host galaxy studies indicate that both mass selection and dust obscuration effects may bias conclusions about the role of AGNs in galaxy evolution. Silverman et al. (2008) demonstrated that a sample of X-ray- and mass-selected AGNs were preferentially found in galaxies with ongoing star formation and blue optical colors. Similarly, Xue et al. (2010), using a mass-matched sample of X-ray-selected AGNs, showed that the fraction of AGNs remains constant as a function of color. Aird et al. (2012) found that, when stellar-mass selection effects are taken into account, AGN hosts are preferentially found in galaxies with blue and green optical colors, but can exist in galaxies of any color. Cardamone et al. (2010) corrected AGN host colors for the presence of dust, revealing that AGN hosts cleanly separate into a population of young, dusty, star-forming galaxies, and older, red, passive galaxies.

We used the results from AGN SED modeling (Section 4) to examine the positions of the UV-selected AGNs in color–mass space. We calculated rest-frame colors of AGN host galaxies using the best-fit SPS models, along with the Johnson UBV transmission curves from Maíz Apellániz (2006). The estimation of the rest-frame colors requires an assumption that the models accurately represent the stellar populations in the rest-frame UV and optical, which is supported by the good SPS+AGN dual-component modeling fits to the IRAC AGNs. In order to compare the AGN host colors to those of the non-active UV-selected star-forming galaxies, we also calculated the rest-frame colors for the non-active comparison samples described in the previous section. The (U − V)_{rest, AB} versus M_* relation is shown in Figure 11, where the AGNs are plotted in red and the non-active galaxies in gray. The non-AGN hosts follow a trend where the most massive galaxies are also the reddest, similar to the blue sequence observed for galaxies out to z ∼ 3 (Labbé et al. 2007). The sample of UV-selected AGNs is preferentially hosted by high-mass galaxies and these objects are found to span the same range of (U − V)_{rest, AB} colors as those of the non-AGNs in the mass-matched sample. The average (median) color for the AGN sample is (U − V)_{rest, AB} = 0.9 (0.9) ± 0.2, while for the mass-matched sample, it is (U − V)_{rest, AB} = 1.0 (1.0) ± 0.3 (inset of Figure 11). Xue et al. (2010), Mainieri et al. (2011), and Aird et al. (2012) found similar results for the demographics of their AGN host galaxy samples. All of these results demonstrate the importance of mass selection in the comparison of AGN hosts to non-active galaxies. The redder colors of the AGN hosts relative to the full non-AGN comparison sample do not indicate a lack of ongoing star formation, but, rather, as we show below, more dust and perhaps older stellar populations characteristic of the massive end of the UV-selected star-forming population. The UV-selected AGNs were initially chosen based on their $UG\cal {R}$ colors to be star forming, and, as discussed in the previous section, we found that all of the AGNs have sSFR >0.4 Gyr⁻¹, indicating active star formation. Figure 12 demonstrates that the SFR, age, color, and dust extinction properties are similar between the AGN sample and the mass-matched comparison sample. The presence of an AGN does not seem to affect the global properties of their host galaxies in comparison to the mass-matched sample. The lack of UV-selected AGNs in host galaxies below 10¹⁰ M_☉ will be discussed in Section 6.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We can use the mass-matched sample to explore the $\cal {R}-K$ distribution shown in Figure 2 in more detail. The fact that AGNs occupy the red portion of this distribution can be explained by their massive host galaxies, and the relation between rest-UV/optical color and stellar mass. The average $\cal {R}-K$ color for the UV-selected AGN sample is 4.0 ± 0.7, while for the mass-matched sample, the average is 3.6 ± 0.6. We used the results from our dual-component modeling to estimate that the presence of an AGN caused the $\cal {R}-K$ color to be 0.3 mag redder than it would be in the absence of an AGN. If we subtract this difference from those AGNs where we did not use dual-component modeling, the average UV-selected AGN host galaxy $\cal {R}-K$ color becomes 3.6 ± 0.7. Therefore, at fixed stellar mass, the AGN host galaxies are typical in terms of their UV/optical (i.e., $\cal {R}-K$) colors.

In Hainline et al. (2011a), we investigated the UV spectral properties for our sample of UV-selected AGNs. The UV portion of the galaxy spectrum contains strong emission and absorption features that we used to probe both AGN and outflow activity. At z ∼ 2–3, the individual UV spectra have low signal-to-noise ratio (S/N), and, to overcome this limitation, we averaged the individual spectra to create higher S/N composite spectra. One of the most striking results from Hainline et al. (2011a) consisted of the extremely red rest-frame UV continuum of the AGN composite spectrum, relative to that of the z ∼ 3 non-AGN composite spectrum from Shapley et al. (2003). We can now use the results from SED modeling to investigate the physical origin of this difference. Furthermore, we can analyze how the rest-frame UV spectroscopic properties of AGNs vary as a function of stellar population properties.

For this analysis, we created composite spectra from subsamples of AGNs separated by the host-galaxy properties. Due to the small sample size of objects with SED fitting parameters (28 objects) and our desire to maximize the S/N of the resulting subsample composite spectra, we simply divided the sample in half for these analyses. We constructed composite spectra following the methodology of Hainline et al. (2011a). Each individual spectrum was first shifted to the rest frame using an estimate of the systemic redshift from He ii λ1640. We converted each spectrum from the units of flux density to those of luminosity density, scaled each one to a common median in the wavelength range of 1250–1380 Å, and combined the spectra, with the four highest and lowest outliers removed at each wavelength.

In order to investigate the red UV continuum of the AGN composite spectrum, we consider the properties of the mass-matched comparison sample. Using existing rest-frame UV spectroscopy for the full non-active sample, we created a UV composite spectrum from the objects in the mass-matched comparison sample. We used the same method to create this non-AGN composite as for the AGN composite. In Figure 13, we compare the full AGN composite spectrum to the mass-matched non-AGN composite as well as the non-AGN LBG composite at z ∼ 3 from Shapley et al. (2003). The UV continuum shape is commonly described by β, the slope of a power law of the form L_λ∝λ^β, which is fit to the continuum. For the AGN composite spectrum, β = −0.3 ± 0.2 (Hainline et al. 2011a), which is a much redder slope compared to the LBG composite from Shapley et al. (2003), where β = −1.5. We measure a value of β = −0.7 ± 0.1 for the mass-matched composite, redder than the full non-AGN composite.¹⁵ The comparably red slopes of the AGN and mass-matched UV composite spectra (in contrast to the significantly bluer slope of the total LBG composite from Shapley et al. 2003) confirm that the underlying UV continuum seen for the AGN sample is predominantly due to stellar light, and not AGN emission. Furthermore, since scattered light from a buried AGN would have an intrinsically bluer rest-frame UV spectrum than that of a stellar population (Zakamska et al. 2006), the slight difference in UV slopes cannot be attributed to AGN emission. We can naturally explain the redder slopes observed in the high-mass active and inactive samples by looking at the distributions of reddening for each subsample. In Figure 12, we show histograms of the best-fit properties of the AGN sample along with the full non-AGN and mass-matched comparison samples. As discussed in Section 4.3, 〈E(B − V)_{SF, AGN}〉 = 0.22 ± 0.11 for the AGNs, while for the mass-matched comparison sample, 〈E(B − V)_{SF, MM}〉 = 0.20 ± 0.09. For the sample of non-AGNs at z ∼ 3 from Shapley et al. (2003), the estimated 〈E(B − V)_{SF, Shapley}〉 = 0.15 ± 0.09.¹⁶ The errors on these values represent the standard deviation of the distribution of values. The average extinction values for the AGN and mass-matched samples are significantly different from the value for the non-AGNs in Shapley et al. (2003), reflecting what is observed for the UV power-law slopes.

Zoom In Zoom Out Reset image size

We further investigated the role that extinction plays in the slope of the UV spectrum by separating our objects into two bins based on E(B − V)_SF. The AGNs were separated at the median extinction from our fits, E(B − V)_SF = 0.242, and both composites are plotted in Figure 14. In red, we plot the composite spectrum for those objects with E(B − V)_SF > 0.242, and in black, the composite spectrum for those objects with E(B − V)_SF < 0.242. The composite spectrum for objects with higher E(B − V)_SF is much redder, with β = 0.0 ± 0.3. For the AGNs with lower E(B − V)_SF, the composite spectrum has a β = −1.0 ± 0.3. This trend is also seen in the mass-matched comparison sample separated by E(B − V)_SF, where the high-extinction mass-matched composite is redder, with β = −0.4 ± 0.2, while the low-extinction mass-matched composite has β = −1.1 ± 0.1. The slope of the low-extinction mass-matched composite spectrum (〈E(B − V)_{SF, MM}〉 = 0.12 ± 0.05) is similar to the slope of the low-extinction AGN composite (〈E(B − V)_{SF, AGN}〉 = 0.13 ± 0.08), which provides more evidence that dust extinction significantly modulates the UV power-law slope in our sample of AGNs.

Zoom In Zoom Out Reset image size

We also consider how rest-frame UV AGN spectroscopic properties vary with stellar mass. The continuum normalized composite spectra for the AGNs separated by stellar mass are shown in Figure 15.¹⁷ We measured the luminosities and EW for the strongest UV emission lines (H i Lyα, N v λ1240, C iv λ1549, and He ii λ1640) in both the high-mass and low-mass composites. Uncertainties on these values were calculated following a bootstrap technique where 500 fake composite spectra were constructed from the sample of spectra used in creating the real composite spectra (see Shapley et al. 2003; Hainline et al. 2011a). The results are shown in Table 6. We find that the EW values for the C iv and He ii emission lines are stronger in the high-mass composite than the low-mass composite, while the N v and Lyα lines have statistically equivalent EW values. The EW of an emission line was calculated by integrating the luminosity in the line and dividing by the average of the luminosity density of the continuum on either side of the line. As these are Type II AGNs, the observed UV continuum arises from stellar emission, which is supported by the similar UV slopes seen in the AGN UV composite and the mass-matched UV composite. We find that the emission lines used to identify the presence of an AGN (C iv and He ii) are stronger in the high-mass composite spectrum. This trend of decreasing high-ionization EW with decreasing stellar mass may explain the dearth of low-mass UV-selected AGNs in our sample. We will explore the origins and implications for this trend in the following section.

Zoom In Zoom Out Reset image size

Table 6. UV Emission Features for Stellar-mass Subsamples

	log(M*/M☉) > 10.7	log(M*/M☉) < 10.7
WaLyα	56 ± 15	58 ± 18
$W_{\mathrm{{\rm N\,\mathsc{v}},1240}}^{\rm a}$	4.6 ± 1.4	3.7 ± 1.6
$W_{\mathrm{{\rm C\,\mathsc{iv}},1549}}^{\rm a}$	17.6 ± 3.3	7.4 ± 3.3
$W_{\mathrm{{\rm He\,\mathsc{ii}},1640}}^{\rm a}$	10.2 ± 2.4	5.0 ± 1.1

Note. ^aRest-frame EW in Å, measured from the composite spectra. Uncertainties are calculated as described in Section 5.4.

Download table as:  ASCIITypeset image