PROBING THE DAWN OF GALAXIES AT z ∼ 9–12: NEW CONSTRAINTS FROM HUDF12/XDF AND CANDELS DATA^*

The HUDF (Beckwith et al. 2006) was imaged with WFC3/IR as part of two large HST programs now. The HUDF09 (PI: Illingworth; Bouwens et al. 2011b) provided one pointing (4.7 arcmin²) of deep imaging in the three filters Y₁₀₅ (24 orbits), J₁₂₅ (34 orbits), and H₁₆₀ (53 orbits). These data were extended recently with the HUDF12 campaign (Ellis et al. 2013; Koekemoer et al. 2012), which imaged the HUDF further in Y₁₀₅ (72 orbits) and H₁₆₀ (26 orbits), and additionally added a deep exposure in JH₁₄₀ (30 orbits). These are the deepest NIR images ever taken, resulting in a final 5σ depth of H₁₆₀ ∼ 29.8 mag (see also Table 1).

Since the acquisition of the original optical HUDF ACS data, several programs have added deeper ACS coverage to this region, mainly as part of parallel imaging. We combined all the available ACS data over the HUDF, which allows us to improve the backgrounds and also to push photometry limits deeper by ∼0.1–0.2 mag. These data, along with the matched WFC3/IR data from all programs, are released publicly as the eXtreme Deep Field (XDF) dataset⁸ and are discussed in more detail in Illingworth et al. (2013).

For longer wavelength constraints we also include the ultra-deep Spitzer/IRAC data from the 262 hr IUDF program (PI: Labbé; see also Labbé et al. 2012), which reach to ∼27 mag AB (5σ total) in both [3.6] and [4.5] channels. These data are extremely important for eliminating lower-redshift contaminating sources, particularly intermediate-redshift dusty and/or evolved galaxies (see Section 3.5.1).

In addition to the HUDF data, we also include the two additional deep parallel fields from the HUDF09 program, as well as all the WFC3/IR data over the GOODS-South field. The latter were taken as part of the ERS program (Windhorst et al. 2011) and the multi-cycle treasury campaign CANDELS (PI: Faber/Ferguson; Grogin et al. 2011; Koekemoer et al. 2011). These data were already used for a z ∼ 10 search in our previous analysis from Oesch et al. (2012a). We therefore refer the reader to that paper for a more detailed discussion. However, since our previous analysis the acquisition of an additional four epochs of CANDELS DEEP data was completed, resulting in deeper data by about 0.2 mag. These are included now in this paper, which will allow us to further tighten our constraints on the z ∼ 10 LF.

In the optical, we make use of all ACS data taken over the GOODS South field, which includes additional imaging from supernova follow-up programs. These images reach ∼0.1–0.3 mag deeper than the v2.0 reductions of GOODS, in particular in the z₈₅₀ band. We also reduce and include all the I₈₁₄ data, which were taken over this field. By combining all these datasets, we have produced what is the deepest optical image to date over the GOODS-S field. Such deep optical data are very important for excluding lower-redshift interlopers in LBG samples.

For constraints from Spitzer/IRAC, we use the public data from the GOODS campaign. These exposures are 23 hr deep and reach to ∼26 mag (M. Dickinson et al., in preparation). All these fields are also outlined in Figure 1.

In this paper, we adopt an LBG selection to identify galaxies at z ≳ 8.5. The major advantages of this approach over a photometric redshift selected sample as is used, e.g., in Ellis et al. (2013) are simplicity and robustness. The Lyman break technique provides a straightforward and robust selection, which is easily reproducible by other teams (e.g., Schenker et al. 2013). Furthermore, the simplicity of the Lyman break color–color criteria also allows for a straightforward estimate of the selection volumes based on simulations.

In contrast, the photometric redshift likelihood functions are heavily dependent on the assumed template set and even on the specific photometric redshift code that is used. Additionally, the photometric redshift likelihood functions depend on largely unknown priors which are needed to account for the number density of intermediate-redshift passive or dusty galaxies. A particular problem is that most high-redshift photometric analyses give equal weight to all templates at all redshift (i.e., they adopt a flat prior). This includes faint galaxies with extreme dust extinction at intermediate redshift or passive sources at z > 5, which are unlikely to be very abundant in reality.

A further uncertainty, in particular for high-redshift sources, is how undetected fluxes are treated in the fitting process. This can have significant influence on the lower-redshift likelihood estimates.

Given all these advantages, we will therefore select high-redshift galaxies using the Lyman break technique and we will determine their photometric redshifts a posteriori using standard template fitting on this pre-selected sample of LBGs.

The addition of deep F140W imaging data over the HUDF gives us the ability to select new samples of z ∼ 9 galaxies over that field. As can be seen in Figure 2, the absorption due to the intergalactic neutral hydrogen shifts in between the Y₁₀₅ and J₁₂₅ filters at z ≳ 9. For a robust Lyman break selection, we thus combine the Y₁₀₅ and J₁₂₅ filter fluxes in which galaxies start to disappear at z ∼ 9. Our adopted selection criteria are

$$\begin{eqnarray} &(Y_{105}+J_{125})/2 - {\it JH}_{140}>0.75 & \end{eqnarray} \tag{ 1 }$$

$$\begin{equation*} (Y_{105}+J_{125})/2 - {\it JH}_{140} > 0.75 + 1.3\times ({\it JH}_{140} - H_{160}) \end{equation*}$$

$$\begin{equation*} {\rm S/N}(B_{435}\ \mathrm{ to }\ z_{850})<2 \; \wedge \; \chi ^2_{{\rm opt}}<2.8. \end{equation*}$$

Zoom In Zoom Out Reset image size

These criteria (shown in Figure 3) are chosen to select sources at z ∼ 8.5–9.5. We additionally use a (J₁₂₅ − H₁₆₀) < 1.2 criterion to cleanly distinguish our z ∼ 9 and z ∼ 10 samples (see next section).

Zoom In Zoom Out Reset image size

We only include sources which are significantly detected in the H₁₆₀ and JH₁₄₀ images with at least 3σ in each filter and with 3.5σ in at least one of the two. As a cross-check we selected sources based on an inverse-variance weighted combination of the J₁₂₅, JH₁₄₀, and H₁₆₀ images at 5σ. Both selections resulted in the same final sample of high-redshift sources, i.e., all selected candidates are >5σ detections.

In addition to the color selections, we require sources to be undetected at shorter wavelengths. In particular, we use 2σ non-detection criteria in all optical bands individually. Furthermore, we adopt an optical pseudo-χ² constraint. This is defined as $\chi ^2_{{\rm opt}} = \sum _i \mathrm{SGN}(f_i) (f_i/\sigma _i)^2$. The summation runs over all the optical filter bands B₄₃₅, V₆₀₆, i₇₇₅, I₈₁₄, and z₈₅₀, and SGN is the sign function, i.e., SGN(x) = −1 if x < 0 and SGN(x) = 1 if x > 0. This measure allows us to make full use of all information in the optical data. We only consider galaxies with $\chi ^2_{{\rm opt}}<2.8$. This cut reduces the contamination rate by a factor ∼3 ×, while it only reduces the selection volume of real sources by 20% (see also Oesch et al. 2012b). This is a powerful tool for providing source lists with low contamination rates (see also Section 3.5.2).

These selection criteria result in seven z ∼ 9 galaxy candidates in the HUDF12/XDF dataset. These sources are listed in Table 2 and their images are shown in Figure 4. In Figure 5, we additionally show the spectral energy distribution (SED) fits and redshift likelihood functions for these sources. For comparison, these are derived from two photometric redshift codes, ZEBRA (Feldmann et al. 2006; Oesch et al. 2010b) as well as EAZY (Brammer et al. 2008). As is evident, the vast majority of sources do show a prominent peak at z ∼ 8–9 together with a secondary, lower likelihood peak at z ∼ 2. The best-fit photometric redshifts of these candidates range between z = 8.1 and 9.0, with the exception of one source (XDFyj-39446317), which has a ZEBRA photometric redshift of only z_phot = 2.2. However, using the EAZY code and template set, the best-fit redshift is found at z = 8.6. The photometric redshift likelihood function for this source is very wide using both codes.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Table 2. Photometry of z > 8 LBG Candidates in the HUDF12/XDF Data

ID	R.A.	Decl.	H160	(YJ) − JH140	J125 − H160	JH140 − H160	S/NH160	S/NJH140	S/NJ125	$\chi ^2_{{\rm opt}}$
	$z_{{\rm phot}}^{{\rm ZEBRA}}$	$z_{{\rm phot}}^{{\rm EAZY}}$				Comments
z ∼ 9 YJ-dropouts
XDFyj-38135540	03:32:38.13	−27:45:54.0	27.95 ± 0.10	0.8 ± 0.1	0.0 ± 0.1	−0.2 ± 0.1	13.1	16.0	9.8	0.2
	$8.4^{+0.1}_{-0.1}$	$8.4^{+0.1}_{-0.2}$	Bouwens UDFy-38125539; McLure HUDF12-3813-5540 (z = 8.3); and in other Y-dropout samples.
XDFyj-39478076	03:32:39.47	−27:48:07.6	28.53 ± 0.14	0.8 ± 0.2	0.4 ± 0.2	−0.0 ± 0.2	8.7	9.7	5.0	−0.6
	$8.1^{+0.3}_{-0.6}$	$8.3^{+0.2}_{-0.5}$	Bouwens UDFy-39468075; Ellis HUDF12-3947-8076 (z = 8.6)
XDFyj-39216322	03:32:39.21	−27:46:32.2	29.49 ± 0.25	1.1 ± 0.5	0.3 ± 0.4	0.1 ± 0.3	5.1	4.9	3.2	−1.4
	$8.8^{+0.5}_{-0.5}$	$8.9^{+0.5}_{-0.4}$	Ellis HUDF12-3921-6322 (z = 8.8)
XDFyj-42647049	03:32:42.64	−27:47:04.9	29.15 ± 0.21	1.5 ± 0.7	0.8 ± 0.6	0.1 ± 0.3	4.4	4.9	2.1	−0.5
	$9.0^{+0.5}_{-0.5}$	$9.2^{+0.5}_{-0.6}$	Ellis HUDF12-4265-7049 (z = 9.5)
XDFyj-40248004	03:32:40.24	−27:48:00.4	29.87 ± 0.30	1.3 ± 0.7	0.3 ± 0.6	−0.1 ± 0.4	3.5	3.3	2.2	0.0
	$8.8^{+0.5}_{-0.5}$	$8.9^{+0.6}_{-0.3}$	Faint source. Not in Ellis et al. (2013) sample.
XDFyj-43456547	03:32:43.45	−27:46:54.7	29.69 ± 0.42	1.3 ± 0.7	0.5 ± 0.7	−0.1 ± 0.4	3.1	3.5	1.8	−3.8
	$8.7^{+0.6}_{-0.5}$	$8.9^{+0.7}_{-0.8}$	Ellis HUDF12-4344-6547 (z = 8.8)
XDFyj-39446317a	03:32:39.44	−27:46:31.7	29.77 ± 0.27	1.1 ± 0.7	>1.0	−0.3 ± 0.5	3.8	3.7	1.3	1.3
	$2.2^{+0.8}_{-0.7}$	$8.6^{+1.0}_{-1.5}$	Faint source. Very wide p(z), with low best-fit redshift. Not in Ellis et al. (2013) sample.
z ∼ 10 J-dropouts
XDFj-38126243	03:32:38.12	−27:46:24.3	29.87 ± 0.40	>1.9	1.4 ± 0.9	0.3 ± 0.4	5.8	3.4	1.2	−0.6
	$9.8^{+0.6}_{-0.6}$	$9.9^{+0.7}_{-0.6}$	This source was selected as a z ∼ 10 candidate in the HUDF09 year 1 data, but did not appear in
			final Bouwens et al. (2011a) sample due to low S/N in second year data (see Figure 6).
z ∼ 10.7 JH-dropouts
XDFjh-39546284	03:32:39.54	−27:46:28.4	28.55 ± 0.14	⋅⋅⋅	>2.3	>2.3	7.3	0.2	−1.6	0.8
	$11.8^{+0.2}_{-0.4}$	$11.9^{+0.2}_{-0.5}$	UDFj-39546284 of Bouwens et al. (2011a), Oesch et al. (2012a); Ellis HUDF12-3954-6284 (z = 11.9)

Notes. S/N are measured in circular apertures of fixed 035 diameter. ^aDue to the low photometric redshift estimate, we do not include the source XDFyj-39446317 in our analysis of the UV LF at z ∼ 9. One contaminating lower-redshift source is expected in our sample due to photometric scatter (see Section 3.5.2).

Download table as:  ASCIITypeset image

From photometric scatter simulations (see Section 3.5.2), we expect to find 0.9–1.1 low-redshift contaminants in our z ∼ 9 sample due to photometric uncertainties. Therefore, finding an LBG candidate with such a low redshift is not necessarily unexpected. We will thus list it as a possible candidate in Table 2. However, we will exclude XDFyj-39446317 from our determination of the UV LF at z ∼ 9.

Note that Ellis et al. (2013) and McLure et al. (2013) performed a photometric redshift selection of z ≳ 8.5 sources over the same field, which we discuss in detail in Section 3.7.3.

Galaxies at redshifts approaching z ∼ 10 start to disappear in the J₁₂₅ filter. Following Bouwens et al. (2011a) and Oesch et al. (2012a), we select z ∼ 10 galaxies based on very red J₁₂₅ − H₁₆₀ colors and we use Spitzer/IRAC photometry to guard this selection against intermediate-redshift extremely dusty and evolved galaxies in a second step. This selection process also used JH₁₄₀ data when available (i.e., over the HUDF12/XDF field), and was used for all the datasets shown in Figure 1.

The HST selection criteria are

$$\begin{equation} (J_{125}-H_{160})>1.2 \; \wedge \; ({\it JH}_{140}-H_{160})<1.0 \end{equation} \tag{ 2 }$$

$$\begin{equation*} {\rm S/N}(B_{435}\ \mathrm{ to }\ Y_{105})<2 \; \wedge \; \chi ^2_{{\rm opt}}<2.8 \end{equation*}$$

in addition to at least 3σ detections in both H₁₆₀ and JH₁₄₀ and >3.5σ in one of these. All sources in our final list also satisfy a >5σ detection criterion in the combined J₁₂₅ + JH₁₄₀ + H₁₆₀ image.

The JH₁₄₀ − H₁₆₀ color criterion was introduced to distinguish z ∼ 10 from z ∼ 11 galaxies over the HUDF12/XDF field. The other fields, which do not have deep JH₁₄₀ imaging, do not include this criterion. We account for this difference in our analysis of the selection functions (Section 3.6).

When applying these selection criteria to the WFC3/IR+ACS data over GOODS-S, we previously identified 16 galaxies which satisfied these criteria. However, these are all extremely bright in the Spitzer/IRAC bands and are even detectable in the shallow [5.8] and [8.0] channel data over GOODS-S (having H₁₆₀ − [5.8] = 2.4–4.0 mag). These sources were therefore excluded from our z ∼ 10 analysis, as their H₁₆₀ to IRAC colors were too red for a genuine z ∼ 10 galaxy. These are most likely z ∼ 2–3 galaxies with significant extinction and possibly evolved stellar populations (see Oesch et al. 2012a).

Even taking advantage of the deeper WFC3/IR data that became available over the CANDELS-South field subsequent to the Oesch et al. (2012a) analysis, no new credible z ∼ 10 source could be found. However, our selection revealed three potential sources in the HUDF12/XDF data. Unfortunately, two of these are very close to a bright, clumpy foreground galaxy. Their photometry is therefore very uncertain, and it is unlikely that they are real high-redshift sources. We nevertheless list these as potential sources in Table 6. However, we will not use them in the subsequent analysis.

This leaves us with only one likely z ∼ 10 galaxy candidate in all the fields we have analyzed here. This is XDFj-38126243, which we had previously identified in the first-epoch data of the HUDF09 as a potential z ∼ 10 candidate (Bouwens et al. 2011a). However, it was not detected at a significant enough level in the subsequent second epoch H₁₆₀ data to indicate at high confidence that it was real. As a result it was not included in our final sample of z ∼ 10 sources from the HUDF09 data.

The source XDFj-38126243 is now clearly detected both in the new H₁₆₀ and in the JH₁₄₀ data from the HUDF12 survey, which clearly confirms its reality. This is demonstrated in Figure 6. As can also be seen from that figure, the source is extremely compact, consistent with being a point source. We can therefore not exclude that this source is powered by an active galactic nucleus (AGN), which could also explain the possible variability over a timescale of 1 yr (see lower panel of Figure 6). However, the low flux measurement in the second-year HUDF09 data is still consistent with expectations from Gaussian noise. Taken together, the flux measurements of all three epochs are consistent with the source showing no time variability (χ² = 2.6).

Zoom In Zoom Out Reset image size

The source is not detected in our ultra-deep IRAC data, and its colors place it at a photometric redshift of $9.8^{+0.6}_{-0.6}$ (see also Table 2 and Figure 7).

Zoom In Zoom Out Reset image size

For completeness, we also list three additional potential candidates in the Appendix (see Table 6). These sources lie very close to bright foreground galaxies. Formally, they show colors consistent with being at very high redshift. However, they are significantly blended with their neighbors, such that the fluxes of these sources cannot be accurately measured with SExtractor, without a more sophisticated neighbor subtraction technique. Additionally, the close proximity to very bright foreground sources casts significant doubt on the reality of these sources, and we therefore do not include these in our analysis.

The addition of JH₁₄₀ imaging from the HUDF12 data also allows for the selection of z ≳ 10.5 galaxies based on a red JH₁₄₀ − H₁₆₀ color, as the intergalactic medium (IGM) absorption shifts through the center of the JH₁₄₀ filter (see Figure 3). We therefore search for galaxies in the HUDF12/XDF data satisfying the following:

$$\begin{equation} ({\it JH}_{140}-H_{160})>1.0 \end{equation} \tag{ 3 }$$

$$\begin{equation*} {\rm S/N}(B_{435}\ \mathrm{ to }\ J_{125})<2 \; \wedge \; \chi ^2_{{\rm opt}}<2.8. \end{equation*}$$

In order to ensure the reality of sources in this single-band detection sample, we require >5σ detections in H₁₆₀.

Only one source satisfies these criteria: XDFjh-39546284. This is our previous highest-redshift candidate from the HUDF09 data (UDFj-39546284; see Bouwens et al. 2011a). Surprisingly, with JH₁₄₀ − H₁₆₀ > 2.3 it has an extremely red color in these largely overlapping filters. It is by far the reddest source in the HUDF12/XDF data, and its rather dramatic color raises some questions.

At face value, the extreme color, together with the non-detection in the optical data and in our deep IRAC imaging, results in a best-fit redshift of z = 11.8 ± 0.3 for this source. However, such a high redshift would imply that the source is ∼20 × brighter than expected for z ∼ 12 sources at the same number density (see Bouwens et al. 2012d). While a strong Lyα emission line could reduce the remarkably high continuum brightness, the lack of Lyα seen in galaxies within the reionization epoch at z ≳ 6 indicates that the high fraction of neutral hydrogen in the universe at early times absorbs the majority of Lyα photons of these galaxies (e.g., Schenker et al. 2012; Pentericci et al. 2011; Caruana et al. 2012; Bunker et al. 2013). Seeing strong Lyα emission from a source at z ∼ 12 during the early phase of reionization is particularly unexpected (although perhaps not impossible).

Alternatively, the source could be a z ∼ 2 extreme line emitter. An emission line at ∼1.6 μm would have to produce the majority of its H₁₆₀ flux, which would require extreme equivalent widths. A possible example of such a source is presented in Brammer et al. (2013). For more extensive discussions of these alternative options for this source see Brammer et al. (2013) and Bouwens et al. (2012d).

Given the uncertain nature of this source, we will treat its detection as an upper limit of ⩽1 source in the following analysis, and we will only derive upper limits on the luminosity and SFRDs at z ∼ 10.7 from this single-source sample.

In the following sections we discuss possible contamination of our z > 8 LBG samples.

3.5.1. Dusty and Evolved Galaxies

3.5.2. Photometric Scatter of Low-z Sources

3.5.3. Additional Sources of Contamination

The expected redshift distributions of our LBG samples are estimated based on extensive simulations of artificial galaxies inserted in the real data, which are then re-selected in the same manner as the original sources (see also Oesch et al. 2007, 2012a). In particular, we estimate the completeness C(m) and selection probabilities S(z, m) as a function of H₁₆₀ magnitude m and redshift z.

Following Bouwens et al. (2003), we use the profiles of z ∼ 4 LBGs from the HUDF and GOODS observations as templates for these simulations. The images of these sources are scaled to account for the difference in angular diameter distance as well as a size scaling of (1 + z)⁻¹. The latter is motivated by observational trends of LBG sizes with redshift across z ∼ 3–8 (see, e.g., Ferguson et al. 2004; Bouwens et al. 2004; Oesch et al. 2010a; Ono et al. 2012).

The colors of the simulated galaxy population are chosen to follow a distribution of UV continuum slopes with β = −2.4 ± 0.4 (see, e.g., Bouwens et al. 2009, 2010; Finkelstein et al. 2010; Stanway et al. 2005), and are modulated by the IGM absorption model of Madau (1995) over a range of redshifts z = 8 to z = 13. Ten thousand galaxies are simulated for each redshift bin in steps of dz = 0.2, which allows for a reliable estimate of the completeness and selection probability taking into account the dispersion between input and output magnitudes.

The result of these simulations enables us to compute the redshift distribution of galaxies after assuming an LF ϕ(M)

$$\begin{equation} p(z) = {\it dN/dz} = \int dm \frac{dV}{dz} S(m,z)C(m)\phi (M[m,z]). \end{equation} \tag{ 4 }$$

We assume a baseline UV LF evolution with α = −1.73 and M_*(z) = −20.29 + 0.33 × (z − 6), consistent with the trends found across z ∼ 4 to z ∼ 8 (Bouwens et al. 2011b). The normalization is not relevant for the relative distributions. For the K-correction in the conversion from absolute to observed magnitude, we use a 100 Myr old, star-forming template of Bruzual & Charlot (2003).

The results of this calculation are shown in Figure 8. It is clear that the redshift selection functions are significantly wider than the target redshift range based on the simple LBG color tracks shown in Figure 3. The reason for this is simply photometric scatter. The mean redshift (and the width) of our samples are 9.0 (±0.5), 10.0 (±0.5), and 10.7 (±0.6) for the YJ-, J- and JH-dropout samples, respectively.

Zoom In Zoom Out Reset image size

In the following sections, we compare our LBG samples with previous selections over these fields in the literature.

3.7.1. Bouwens et al. (2011a) z ≳ 8.5 Y-dropouts

3.7.2. Bouwens et al. (2011a) z ∼ 10 Candidates

3.7.3. Ellis et al. (2013) and HUDF12 Team Papers

The HUDF12 team has recently published a sample of seven z ≳ 8.5 galaxy candidates identified in the HUDF12 data in Ellis et al. (2013). These sources were based on a photometric redshift selection technique (see also McLure et al. 2013), with a >5σ detection in ∼05 diameter apertures. Here, we use smaller apertures of 035 diameter for S/N measurements, which in most cases are more optimal for the very small z ∼ 9–10 sources, and result in a small additional gain of ∼20% in S/N. This is why we could expect to find additional sources in our sample compared to Ellis et al. (2013).

In general, our sample is in very good agreement with the selection of Ellis et al. (2013). With the exception of two, we include all their sources in our z > 8 samples. The discrepant ones are UDF12-3895-7114 and UDF12-4106-7304, which we discuss below. Their photometry is additionally listed in the Appendix (see Table 6).

UDF12-3895-7114. This source certainly shows colors very similar to a z > 8 candidate. However, we measure (YJ) − JH₁₄₀ = 0.5 ± 0.5, which is bluer than our selection color for the z ∼ 9 sample. Hence, it is not included in our z ∼ 9 sample. While Ellis et al. (2013) find a best-fit photometric redshift of 8.6 ± 0.7, the source is not present in the "robust" sample of McLure et al. (2013), and we find a photometric redshift distribution function which is very broad, with a best fit at z ∼ 0.5 (using both ZEBRA or EAZY). This different result compared to the Ellis et al. (2013) redshift estimate may be caused by small uncertainties in the photometry measurements (given that we also use different apertures). Additionally, we perform IRAC flux measurements on a source-by-source basis. These include an additional uncertainty due to the subtraction of neighboring sources (see, e.g., Labbé et al. 2010, 2012). This is not the case in the Ellis et al. (2013) analysis, who note that they use constant upper limits on the IRAC fluxes.

As we have discussed, photometric redshifts are very uncertain for sources this faint and so there is a chance that this source is still at z > 8. Nevertheless, our analysis raises significant doubt about its high-redshift nature.

UDF12-4106-7304. The WFC3/IR PSF shows significant diffraction spikes, which are caused by the mount of the secondary mirror. While typically only seen around bright stars, these diffraction spikes are so strong that they can also emanate from compact foreground galaxies, particularly in the redder WFC3/IR filters. The source UDF12-4106-7304 of Ellis et al. (2013) is located at the edge of such a diffraction spike for both the H₁₆₀ and the JH₁₄₀ filters (the only filters wherein UDF12-4106-7304 is significantly detected). This is shown in Figure 9. The photometry of this source is clearly significantly enhanced by the diffraction spike. The detection significance of UDF12-4106-7304 is critically reduced once the diffraction spike signal is removed. The profile of the bright foreground galaxy is non-trivial to model, but fortunately only its core is relevant for causing the diffraction spikes. We therefore use galfit (Peng et al. 2002) to model the center of this source and subtract the diffraction spikes that were scaled to match the core flux. Doing so results in the flux of UDF12-4106-7304 being reduced by a factor ∼2 ×, in both JH₁₄₀ and H₁₆₀, which makes it only a 2.8σ total NIR detection. This is too low to be included in a robust sample.

Zoom In Zoom Out Reset image size

Additional uncertainty about the reality of this source arises due to its different morphology in the H₁₆₀ and JH₁₄₀ images. The "source" also lies close to another faint foreground galaxy. It is therefore not clear whether the diffraction spike and the neighboring galaxy conspired to lead to the detection of this potential candidate. In any case, for these reasons, and for the low detection flux, the reality of UDF12-4106-7304 remains in question and we do not include this source in our analysis.

The sample of nine z > 8 galaxy candidates we compiled in the previous sections allows us to make some of the first estimates of the z ∼ 9–11 UV LFs. Although limited in area, the HUDF12/XDF data alone provide very useful constraints already at z ∼ 9 and limits at z ∼ 11. Additionally, due to the deeper data over the HUDF and the CANDELS GOODS-S field compared to our previous analysis in Oesch et al. (2012a), we are able to improve our constraints on the z ∼ 10 LF.

These new constraints at z ∼ 9–11 will allow us to test whether the galaxy population underwent accelerated evolution at z > 8 as previously found in Bouwens et al. (2011a) and Oesch et al. (2012a), or whether the UV LF trends from lower redshift continue unchanged to z > 8 (the preferred interpretation of, e.g., Ellis et al. 2013; Coe et al. 2013; Zheng et al. 2012).

In order to test for such accelerated evolution, we start by estimating the number of galaxies we would have expected to see in our z ∼ 9–11 LBG samples, if the lower-redshift trends were to hold unchanged at z > 8. By comparing the observed number of sources with those expected from the extrapolations, we derive a direct estimate of any changes in the evolution of galaxies at z > 8.

To derive the expected numbers, we use our estimates of the selection function and completeness measurements described in Section 3.6. This allows us to compute the number of sources expected as a function of observed magnitude. For an assumed LF ϕ(M), this is given by

$$\begin{equation} N^\mathrm{exp}_i = \int _{\Delta m} dm \int dz \frac{dV}{dz} S(m,z)C(m)\phi (M[m,z]). \end{equation} \tag{ 5 }$$

For the UV LF evolution, we adopt the relations of Bouwens et al. (2011b): ϕ* = 1.14 × 10⁻³ Mpc⁻³ mag⁻¹ = const, α = −1.73 = const, and M*(z) = −20.29 + 0.33 × (z − 6). Note that we assume constant values for the faint-end slope α and the normalization ϕ_*. These relations are used as a baseline, when extrapolated to higher redshifts, to test whether the observed galaxy population at z > 8 is consistent with the trends at later times, i.e., at lower redshifts.

With these assumptions, we find that we would expect to detect a total of 17 ± 4 galaxies in our z ∼ 9 YJ-dropout sample over the HUDF12/XDF field alone. Yet, after correcting for one potential contaminant (see Section 3.5.2), we only detect six sources. This is 2.8 × fewer than expected from the trends from the lower-redshift LFs.

Similarly, with the same assumptions about the evolution of the UV LF from z ∼ 4 to z ∼ 10, we would expect to see a total of 9 ± 3 sources in our z ∼ 10 J-dropout sample. We only find one such source, which suggests that beyond z ∼ 9, the decrement compared to the baseline evolution is even larger than we found previously from the HUDF09 and the six-epoch CANDELS data. We now expect to see three more sources compared to the six that we expected to see in the earlier analysis of Oesch et al. (2012a). Yet, no additional z ∼ 10 sources are found.

The expected magnitude distribution for the z ∼ 10 sample is shown in Figure 10. As indicated in the figure, our search for z ∼ 10 sources in the CANDELS and ERS fields of GOODS-S should have resulted in two detections. Only 0.5 sources were expected in these fields from our previous analysis using somewhat shallower data (Oesch et al. 2012a).

Zoom In Zoom Out Reset image size

Depending on the assumptions about halo occupation, the expected cosmic variance for a single WFC3/IR field is 40%–45% (Trenti & Stiavelli 2008; Robertson 2010). In order to estimate the significance of our finding of a large offset (i.e., decrement) between the expected number of sources and that seen, we have to combine the Poissonian and cosmic variance uncertainties. We estimate the chance of finding ⩽1 source in our full search area using the appropriate expected number counts and cosmic variance estimates for the individual search fields. The latter are based on the cosmic variance calculator of Trenti & Stiavelli (2008). Using simple Monte Carlo simulations, we derive that given that we expected to find nine sources, finding ⩽1 occurs at a probability of only 0.5%. Therefore, our data are inconsistent with a simple extrapolation of the lower-redshift LF evolution at 99.5%.

This new estimate reinforces the conclusion of Bouwens et al. (2011a) and Oesch et al. (2012a) that the evolution in the number density of star-forming galaxies between z ∼ 10 and z ∼ 8 is large, and larger than expected from the rate of increase at later times, i.e., the evolution was accelerated in the ∼200 Myr from z ∼ 10 to z ∼ 8.

The above calculations of the expected number of sources can be used directly to constrain the UV LFs at z ∼ 9–11. Since the number of candidates is small, we need to make some assumptions about how to characterize the evolution. The UV LFs at later times provide a valuable guide. The parameter that evolves the most is the characteristic luminosity L*; the normalization (ϕ*) and the faint-end slope (α) are relatively unchanged from z ∼ 8 to z ∼ 4 (although we do have evidence for evolution toward steeper faint-end slopes). This suggests that we should estimate what evolution in the characteristic luminosity best reproduces the observed number of sources, while keeping both the normalization and the faint-end slope fixed.

Doing so results in a best estimate of the luminosity evolution of dM/dz = 0.49 ± 0.09 from z ∼ 6 to z ∼ 9 and dM/dz = 0.6  ±  0.2 to z ∼ 10. The characteristic magnitudes at these redshifts are thus expected to be M_*(z ∼ 9) = −18.8  ±  0.3 and M_*(z ∼ 10) = −17.7 ± 0.7. The uncertainties on these measurements are still quite large, given the small sample sizes and the small area probed, in particular for the z ∼ 9 search.

We perform the same calculation for the z ∼ 10.7 JH₁₄₀-dropout sample. However, given the uncertain nature of the single candidate source in that sample (XDFjh-39546284), we treat the estimate at z ∼ 10.7 as an upper limit. We therefore compute the evolution in M_* which is needed to produce one source or fewer in the sample. This is found to be dM/dz > 0.4, which is a less stringent constraint than that for our z ∼ 10 estimate, due to the much smaller area probed by our z ∼ 10.7 search. The inferred constraint on the characteristic magnitude is M_*(z = 10.7) > −18.4 mag. All our estimates of the UV LFs are summarized in Table 3.

Table 3. Summary and Comparison of z ≳ 8.5 LF Determinations in the Literature

Reference	Redshift	log ϕ*	$M_{{\rm UV}}^*$	α
		(Mpc−3 mag−1)	(mag)
This work	11	−2.94 (fixed)	> − 18.4 (1σ)	−1.73 (fixed)
This work	10	−2.94 (fixed)	−17.7 ± 0.7	−1.73 (fixed)
This work	9	−2.94 (fixed)	−18.8 ± 0.3	−1.73 (fixed)
Oesch et al. (2012a)	10	−2.96 (fixed)	−18.0 ± 0.5	−1.74 (fixed)
Bouwens et al. (2012a)	9.2	−3.96 ± 0.48	−20.04 (fixed)	−2.06 (fixed)

Download table as:  ASCIITypeset image

The above estimates for the characteristic luminosity of the UV LF can also be compared with the characteristic luminosity from the step-wise determination of the LF using the observed galaxies and limits. The step-wise LF is derived using an approximation of the effective selection volume as a function of observed magnitude $V_{\rm eff}(m) = \int _0^\infty dz ({dV}/{dz}) S(z,m)C(m)$. The LF is then simply $\phi (M_i)dM = N^{\rm obs}_i/V_{\rm eff}(m_i)$. Errorbars on the step-wise LF measurements are computed based on Poissonian statistics, with an addition of 45% cosmic variance uncertainty for a single WFC3/IR pointing (Trenti & Stiavelli 2008).

The above approximation of the effective volume is only valid as long as the absolute magnitude varies slowly with observed magnitude. This is not the case for z > 11, since the IGM absorption affects the H₁₆₀ band such that the relation between luminosity and the observed magnitude becomes strongly redshift dependent. We therefore restrict our analysis of the step-wise LFs to the z < 10.7 samples. In any case, the small area probed by the HUDF12/XDF data currently does not significantly constrain the UV LF at z ∼ 11. The step-wise z ∼ 9 and z ∼ 10 LFs are tabulated in Table 4.

Table 4. Step-wise Determination of the z ∼ 9 and z ∼ 10 UV LF Based on the Present Dataset

MUV	ϕ*
(mag)	(10−3 Mpc−3 mag−1)
z ∼ 9
−20.66	<0.18
−19.66	$0.15^{+0.15}_{-0.13}$
−18.66	$0.35^{+0.24}_{-0.24}$
−17.66	$1.6^{+0.9}_{-0.9}$
z ∼ 10
−20.78	<0.0077a
−20.28	<0.013
−19.78	<0.027
−19.28	<0.083
−18.78	<0.17
−18.28	<0.34
−17.78	$0.58^{+0.58}_{-0.50}$

Note. ^a1σ upper limit for a non-detection.

Download table as:  ASCIITypeset image

Figure 11 shows our constraints on the UV LF at z ∼ 9 and z ∼ 10. The expected UV LFs extrapolated from the lower-redshift trends are shown as dashed lines. Clearly, at both redshifts, the observed LF lies significantly below this extrapolation, as expected from our analysis of the observed number of sources in the previous section.

Zoom In Zoom Out Reset image size

The best-fit z ∼ 9 LF using M_* evolution is a factor 1.5–5 × below the extrapolated LF at M_UV > −20 mag. At the bright end, the small area probed by the single HUDF12/XDF field limits our LF constraints to ≳ 10⁻⁴ mag⁻¹ Mpc⁻³, which is clearly too high to be meaningful at M_UV < −20. It will therefore be very important to cover several fields with F140W imaging in the future in order to push the z ∼ 9 selection volumes to interesting limits.

At z ∼ 10, the use of deeper data both on the HUDF and on the GOODS-S field allows us to push our previous constraints on the UV LF to fainter limits. Since we only detect one z ∼ 10 galaxy candidate, however, our constraints mainly consist of upper limits. Nevertheless, it is evident that at all magnitudes M_UV > −20 mag these upper limits are consistently below the extrapolated UV LF, up to a factor 4 × lower. Additionally, the limits are also clearly below the best-fit z ∼ 9 LF, showing that the UV LF continues to decline at z > 9.

The evolution of the UV LD at z > 8 has received considerable attention in recent papers, triggered by our initial finding of a significant drop in the LD from z ∼ 8 to z ∼ 10 (i.e., a rapid increase in the LD within a short period of time). With the new HUDF12/XDF data, it is now possible to refine this measurement by adding a z ∼ 9 and a z ∼ 11 point, while also allowing us to improve upon our previous measurements at z ∼ 10.

The UV LDs inferred from our z > 8 galaxy samples are shown in Figure 12. The measurements show the LD derived by integrating the best-fit UV LF determined in the previous section. The integration limit is set to M_UV = −17.7 mag, which is the current limit probed by the HUDF12/XDF data. For comparison, we also show the lower-redshift LD measurements from the compilation of Bouwens et al. (2007, 2012d). These were computed in the same manner as the new z > 8 values, and were not corrected for dust extinction. A summary of our measurements for the LD is listed in Table 5.

Zoom In Zoom Out Reset image size

Table 5. Summary of Luminosity Density and Star Formation Rate Density Estimates

Dropout Sample	Redshift	log10ρUVa	log10ρ*
		(erg s−1 Hz−1 Mpc−3)	(M☉ yr−1 Mpc−3)
YJ	9.0	$25.00^{+0.19}_{-0.21}$	$-2.86^{+0.19}_{-0.21}$
J	10.0	$24.1^{+0.7}_{-0.9}$	$-3.7^{+0.7}_{-0.9}$
JH	10.7	<24.6	< − 3.3
B	3.8	26.38 ± 0.05	−1.21 ± 0.05
V	5.0	26.08 ± 0.06	−1.54 ± 0.06
i	5.9	26.02 ± 0.08	−1.72 ± 0.08
z	6.8	25.88 ± 0.10	−1.90 ± 0.10
Y	8.0	25.58 ± 0.11	−2.20 ± 0.11

Notes. The lower-redshift data points are based on the UV LFs from Bouwens et al. (2007, 2011b) and Oesch et al. (2012b). ^aIntegrated down to $0.05\,L_{z=3}^*$ (M₁₄₀₀ = −17.7 mag).

Download table as:  ASCIITypeset image

Table 6. Photometry of Additional Potential z > 8 LBG Candidates Not Used in This Analysis

ID	R.A.	Decl.	H160	(YJ) − JH140	J125 − H160	JH140 − H160	S/NH160	S/NJH140	S/NJ125	$\chi ^2_{{\rm opt}}$
	$z_{{\rm phot}}^{{\rm ZEBRA}}$	$z_{{\rm phot}}^{{\rm EAZY}}$				Comments
42126501	03:32:42.12	−27:46:50.1	28.45 ± 0.05	1.7 ± 0.2	1.4 ± 0.2	0.3 ± 0.1	22.2	15.8	5.2	−0.1
	$9.5^{+0.1}_{-0.1}$	$9.7^{+0.1}_{-0.2}$	This potential source is completely blended with a foreground galaxy.
43246481	03:32:43.24	−27:46:48.1	28.61 ± 0.17	0.8 ± 0.6	1.3 ± 0.7	0.7 ± 0.4	5.4	3.0	2.8	−1.9
	$2.5^{+6.1}_{-0.7}$	$2.4^{+7.7}_{-0.4}$	Close to bright, clumpy foreground galaxy.
43286481	03:32:43.28	−27:46:48.1	28.53 ± 0.17	>1.6	>1.9	0.8 ± 0.4	6.0	2.5	0.5	−0.2
	$10.4^{+0.5}_{-0.5}$	$10.6^{+0.6}_{-0.3}$	Close to bright, clumpy foreground galaxy.
Additional sources from Ellis et al. (2013)
UDF12-4106-7304	03:32:41.06	−27:47:30.4	⋅⋅⋅	⋅⋅⋅	⋅⋅⋅	⋅⋅⋅	⋅⋅⋅	⋅⋅⋅	⋅⋅⋅	⋅⋅⋅
	⋅⋅⋅	⋅⋅⋅	The photometry of this source is significantly affected by a diffraction spike (see Figure 9).
			After subtraction of the expected flux of the spike, the source is only a 2.8σ total NIR detection.
UDF12-3895-7115	03:32:38.95	−27:47:11.5	30.02 ± 0.30	0.5 ± 0.5	0.1 ± 0.5	0.1 ± 0.5	3.5	3.7	3.7	1.2
	$0.5^{+5.8}_{-0.5}$	$0.6^{+8.2}_{-0.5}$	Does not satisfy our color selection: (YJ) − JH140 = 0.5 < 0.75.
			Additionally, the best-fit photometric redshift is only 0.5 using our photometry.
Additional previous z ≳ 8 candidates from Bouwens et al. (2011a)
UDFy-37806001	03:32:37.80	−27:46:00.1	28.39 ± 0.12	0.4 ± 0.1	−0.2 ± 0.2	−0.2 ± 0.1	10.1	13.7	9.4	2.7
	$7.8^{+0.2}_{-0.4}$	$7.9^{+0.3}_{-0.4}$	Too blue in (YJ) − JH140 for our selection.
UDFy-33446598	03:32:33.44	−27:46:59.8	29.00 ± 0.20	0.3 ± 0.2	0.1 ± 0.3	0.1 ± 0.2	5.6	6.0	4.7	−1.9
	$7.7^{+0.4}_{-0.4}$	$7.7^{+0.5}_{-0.4}$	Too blue in (YJ) − JH140 for our selection.

Note. S/N are measured in circular apertures of fixed 035 diameter.

Download table as:  ASCIITypeset image

As can be seen, our new measurements at z > 8 lie significantly below the z ∼ 8 value. The decrement in LD from z ∼ 8 to z ∼ 9 is 0.6 ± 0.2 dex, and it is even larger at 1.5 ± 0.7 dex to z ∼ 10. Therefore, our data confirm our previous finding of more than an order of magnitude increase of the UV LD in the short time period, only 170 Myr, from z ∼ 10 to z ∼ 8.

The gray line and shaded area show the expected LD evolution when extrapolating the z ∼ 4–8 Schechter function trends to higher redshift. All our measurements at z > 8 lie below the extrapolation. Although the offsets individually are not large (they are <2σ), the consistent offset to lower LD supports a hypothesis that significant changes are occurring in the LD evolution at z > 8.

It is interesting to note that this offset to lower LD is not unexpected, as it is also seen in several theoretical models. In Figure 12, we compare our observational results to two conditional LF models from Trenti et al. (2010) and Tacchella et al. (2013), the prediction from a semi-analytical model of Lacey et al. (2011), as well as the results from two hydrodynamical simulations of Finlator et al. (2011) and Jaacks et al. (2012).

All these models are in relatively good agreement with the lower-redshift (z < 8) measurements. As can be seen from the figure, essentially all models do show a steeper evolution at z > 8 than a purely empirical extrapolation of the UV LF further into the epoch of reionization. Since these models are all very different in nature, this strongly suggests that the rapid build-up we observe in the galaxy population is mainly driven by the build-up in the underlying dark matter halo mass function, which is also evolving very rapidly at these epochs.

Our results combined with those of others now provide a substantially larger sample at z ≳ 8 for estimate of the SFRD than was available for Bouwens et al. (2011a) and Oesch et al. (2012a). The SFRD at z > 8 was recently estimated based on four high-redshift galaxies identified in the CLASH survey (Bouwens et al. 2012a; Coe et al. 2013; Zheng et al. 2012), and from seven galaxies identified in the HUDF12 data (Ellis et al. 2013). We present all these results, together with our own measurements in Figure 13, where we plot the SFRD as a function of redshift in star-forming galaxies with SFR >0.7 M_☉ yr⁻¹ (corresponding to a magnitude limit of M_UV = −17.7 mag).

Zoom In Zoom Out Reset image size

The SFRDs are derived from the UV LD estimates after correction for dust extinction. We use the most recent determinations of the UV continuum slopes β as a function of UV luminosity and redshift from Bouwens et al. (2012c) together with the Meurer et al. (1999) β-extinction relation. The dust-corrected LDs are then converted to SFRDs using the conversion factor of Madau et al. (1998), assuming a Salpeter initial mass function.

Across z ∼ 4 to z ∼ 8, the SFRD clearly evolves very uniformly. The evolution is well reproduced by a power law $\dot{\rho }_*\propto (1+z)^{-3.6\pm 0.3}$, which is shown as a dark gray line in Figure 13. Interestingly, all measurements lie below the extrapolation of this trend to higher redshift. Again, individual measurements are within <2σ of the trend, but the offset to smaller SFRDs in the mean is very clear.

Note that each of the three groups finds a consistent decrement of the SFRD from z ∼ 8 to z ∼ 9. Specifically, from our data, we find a drop by 0.6 ± 0.2 dex. This is ∼2σ below the simple extrapolation of the lower-redshift trends.

At z ∼ 9 our SFRD estimate is in excellent agreement with the measurement of Ellis et al. (2013) based on a photometric redshift selection. At z ∼ 10, however, we find a significantly lower value, mainly due to our inclusion of a larger dataset covering a larger area, in which we do not find any additional candidate. Nevertheless, our measurement is within ∼1σ of the result of Ellis et al. (2013), in particular, after we correct their measurement down by a factor two to account for the source that is likely the result of a diffraction spike in their z = 9.5 sample (Figure 9).

Also at z ∼ 10.7, our upper limit is significantly below the z = 11 estimate of Ellis et al. (2013). This is likely due to the wide redshift range probed by our JH₁₄₀-dropout sample, which extends from z ∼ 9.5 to z ∼ 11.8 (see Figure 8). Ellis et al. (2013) consider a strict boundary of z = 10.5 to z = 11.5.

It is interesting to derive a best-fit evolution of the SFRD at z ⩾ 8 which includes all the current datasets available. We do this by combining all z > 8 measurements from CLASH with our improved estimates at z ∼ 9–11, together with the previous z ∼ 8 SFRD measurement (Oesch et al. 2012b). The results of Ellis et al. (2013) are not included due to significant overlap with our dataset here.

The resulting best-fit evolution of the SFRD falls off very rapidly, following $\dot{\rho }_* \propto (1+z)^{-11.4\pm 3.1}$. This is shown as the black line in Figure 13. As can be seen, this is significantly steeper than the z ∼ 4–8 trends. By z ∼ 10, the best-fit evolution is already a factor ∼5 × below the lower-redshift trend. Therefore, the combined constraint on the SFRD evolution from all datasets in the literature clearly points to an accelerated evolution at z > 8.