A Long Time Ago in a Galaxy Far, Far Away: A Candidate z ∼ 12 Galaxy in Early JWST CEERS Imaging

CEERS is one of 13 early release science surveys designed to obtain data covering all areas of astronomy early in Cycle 1. CEERS is based around a mosaic of 10 NIRCam (Rieke et al. 2005) pointings, with six obtaining NIRSpec (Jakobsen et al. 2022) in parallel and four with MIRI (Rieke et al. 2015) in parallel (four of these pointings also include NIRCam wide-field slitless grism spectroscopy; Greene et al. 2016). Here we make use of the first four CEERS NIRCam pointings, obtained on 2022 June 21, known as CEERS1, CEERS2, CEERS3, and CEERS6.

In each pointing, data were obtained in the short-wavelength (SW) channel F115W, F150W, and F200W filters and long-wavelength (LW) channel F277W, F356W, F410M, and F444W filters. The total exposure time for pixels observed in all three dithers was typically 2835 s filter^–1. The exception is F115W, which obtained double the exposure time to increase the depth on the filter covering the wavelength range below the Lyα break at z > 10. The full details of the readout and dither patterns will be available in the CEERS overview paper (S. Finkelstein et al. 2022, in preparation).

We performed a careful initial reduction of the NIRCam images in all four pointings using version 1.6.2 of the JWST Calibration Pipeline ⁹⁴ with some custom modifications. We used the current (2022 July 29) set of NIRCam reference files, ⁹⁵ which includes in-flight read noise, superbias, distortion, and photometric flux calibration references. We note that the flats were created preflight. We describe our reduction steps below and present more details in M. Bagley et al. (2022, in preparation).

Beginning with the raw data, we used Stage 1 of the pipeline with all default parameters to apply detector-level corrections, fit the ramps in each integration, and output count-rate maps. We next subtracted the "wisp" features, stray light that is reflected off the secondary mirror supports, from detectors A3, B3, and B4 for filters F150W and F200W. For each image I, we scaled the corresponding wisp template ⁹⁶ W (available as of 2022 July 8) by the coefficient a that minimized Var (I − aW) and subtracted the scaled template. We then performed a custom step to remove 1/f noise, which is correlated noise introduced in the images during the detector readout that presents as horizontal and vertical striping patterns (Schlawin et al. 2020). We applied the flat field to the count-rate maps to ensure we were measuring the 1/f noise pattern on a flat image. We masked all bad pixels and source flux using Photutils (Bradley et al. 2020) to detect sources and implementing a tiered approach to source masking. This approach convolves the image with progressively smaller kernels, identifying sources at each step. We use four tiers with Gaussian kernels of σ = 25, 15, 5, and 2 pixels (on the original 0031 and 0063 pixel⁻¹ scales for the SW and LW channels, respectively). These values were chosen after experimenting with several filter kernels to aggressively mask as much source flux as possible. First, for each row and then each column, we measured a σ-clipped median value and subtracted this value from the un-flat-fielded count-rate map. This correction was performed amplifier by amplifier in all filters except F444W, for which we did not perform any correction because the significant residual flat-field structure present in the images dominated any 1/f pattern.

After processing the cleaned count-rate maps through Stage 2 of the pipeline, we performed an astrometric calibration using an edited version of the TweakReg step of the pipeline. The TweakReg step detects sources in each input image, identifies their counterparts in the reference catalog, and calculates a transformation to correct the image WCS. We ran SExtractor (SE; Bertin & Arnouts 1996) on each individual image to replace the internal TweakReg source identification, finding that SE did a better job of identifying and deblending real sources. In lieu of using the default options that allow for alignment to Gaia DR2, we used a reference catalog derived from an HST F160W 003 pixel⁻¹ mosaic ⁹⁷ in the EGS field with astrometry tied to Gaia-EDR3 (see Koekemoer et al. 2011, for details). We first determined the relative offsets between images of the same detector, allowing for shifts in x and y. The rms of this relative astrometry is ∼3–6 mas. We then aligned all images to our HST F160W reference catalog, allowing for shifts in x and y, rotations, and, in the LW images only, a scaling to account for any additional distortion (though we note that the scaling factor is ∼1 × 10⁻⁵). The rms of this absolute alignment is ∼12–15 mas, and the alignment between NIRCam images in different filters has an rms of 5−10 mas.

We fit and removed a single value in MJy sr⁻¹ from each calibrated detector image separately before coadding the images onto a common output grid. For each image, we additionally calculate a scaling factor for the read-noise variance array (VAR_RNOISE) from the background-subtracted sky pixels, again avoiding source flux and bad pixels. We apply this scaling factor to the variance arrays so that they include the robustly measured sky variance. The coadding was performed using the drizzle algorithm with an inverse variance map weighting (Casertano et al. 2000; Fruchter & Hook 2002) via the Resample step in the pipeline. The output mosaics have pixel scales of 003 pixel⁻¹. The usable total area covered by these observations, calculated from the number of pixels with low effective error map values in F115W, F150W, F200W, F277W, and the detection image (see below), is 34.5 arcmin².

We note that our data reduction represents a preliminary version, with several aspects that will be improved with the release of updated NIRCam reference files. We also have not removed the features known as "snowballs" from the mosaics at this time. However, we have carefully inspected all input exposures to ensure that the fluxes in all filters at the positions of galaxies of interest are unaffected by snowballs (see Section 4.2).

The full details of our photometric analysis will be presented in S. Finkelstein et al. (2022d, in preparation); here we briefly summarize our procedures (many of which are similar to Finkelstein et al. 2022). The data products from our modified data reduction pipeline come in the form of multi-extension "i2d" files. We first estimate and subtract any residual background using a custom Python-based algorithm. This routine iteratively convolves the image with Gaussian kernels of progressively smaller sizes (with σ = 25, 15, 5, and 2 pixels), then uses Photutils to mask pixels identified with sources in four iterations to mask progressively smaller sources, dilating the masks in between iterations (by 33, 25, 21, and 19 pixels), then measuring the background after masking with photutils.Background2D. This final background image-construction step used the BkgZoomInterpolator algorithm to construct a smooth background based on the robust σ-clipped means measured within boxes of 10 × 10 pixels and median-filtered over 5 × 5 adjacent boxes.

The i2d file was split into separate extensions, subtracting this background from the SCI extension. Empirical point-spread functions (PSFs) were made by stacking stars, and the F115W, F150W, F200W, and F277W images were PSF-matched to the F356W image using pypher. Photometry was computed on the PSF-matched images using SE (Bertin & Arnouts 1996) v2.25.0 in two-image mode, with an inverse variance–weighted combination of the PSF-matched F277W and F356W images as the detection image, with photometry measured on all seven bands.

Colors were measured in small Kron apertures with a Kron factor of 0.8 and a Kron minimum radius of 1.1 pixels; this is smaller than previous studies, which we found necessary to keep the elliptical aperture close to the significant isophotes of small, faint galaxies. An aperture correction was derived in the F356W catalog as the ratio between the flux measured in the default Kron aperture (with PHOT_AUTOPARAMS 2.5, 3.5) and that in our small Kron aperture. This correction was applied to all fluxes and uncertainties. We use the CEERS simulated imaging ⁹⁸ to test the accuracy of this procedure, finding that after this aperture correction, the total fluxes were underestimated by ∼10%–15%, rising to 22% in F444W (understandable due to the larger PSF in F444W as the photometric apertures were defined on F356W). We apply these simulation-based corrections (comparable to similar corrections applied in HST studies; e.g., Finkelstein et al. 2022) to all fluxes and uncertainties to complete our total flux measurements. All fluxes and uncertainties were corrected for Galactic attenuation assuming a field-averaged E(B − V) =0.006 and a Cardelli et al. (1989) Milky Way attenuation curve. We also measure fluxes in a range of circular apertures; as these are used for detection significance tests, we do not correct them to total fluxes (though they are still corrected for Galactic attenuation).

We derive flux uncertainties directly from the data, following Finkelstein et al. (2022), based on previous methodology outlined in Papovich et al. (2016). We fit for the noise as a function of aperture size by measuring the fluxes at ∼5 × 10³ randomly placed positions in 15 circular apertures with diameters ranging from 1 to 100 pixels, fitting a polynomial function to the standard deviation in aperture fluxes as a function of the number of pixels in each aperture. We then use this function to calculate the photometric uncertainties for each object for a given aperture area. These values were scaled by the ratio of the error image value at the central position of a given source to the median error value of the whole map. All aperture and Galactic attenuation corrections were applied to these uncertainties. Finally, around each source in our catalog, we calculate a "local" noise estimate as the standard deviation in flux values from the 200 closest of these previously placed random apertures.

We use the eazy (Brammer et al. 2008) software package to estimate photometric redshifts for all sources in our photometric catalog. eazy fits nonnegative linear combinations of user-supplied templates to derive probability distribution functions (PDFs) for the redshift based on the quality of fit of the various template combinations to the observed photometry for a given source. The template set we use includes the "tweak_fsps_QSF_12_v3" set of 12 FSPS (Conroy & Gunn 2010) templates recommended by the eazy documentation. To this, we add a set of six additional templates spanning bluer colors than the FSPS models, as R. Larson et al. (2022, in preparation) found that these improve the accuracy of photometric redshift fits for the expected blue colors of z > 9 galaxies.

The new templates were created using stellar population models created with BPASS (Eldridge & Stanway 2009). To generate bluer rest-UV colors than the initial set of FSPS templates, we selected BPASS templates that had low metallicities (5% solar) and young stellar populations (log stellar ages of 6, 6.5, and 7 Myr) and were inclusive of binary stars. We added an additional set of these models inclusive of nebular emission lines derived with CLOUDY (Ferland et al. 1998) using a high ionization parameter (log U = −2), a similarly low metallicity, and nebular continuum emission. R. Larson et al. (2022, in preparation) validated the efficacy of including these models by testing the recovery of photometric redshifts from a mock catalog derived by a semianalytic model (Yung et al. 2022), finding that the inclusion of these six additional templates significantly improved the photometric redshift estimates for blue high-redshift galaxies.

We do not use a luminosity prior (e.g., a flat prior is assumed), as the epoch in question is completely unexplored, and we include a systematic error of 5% of the observed flux values. Our fiducial eazy run uses our total fluxes derived from our Kron aperture–measured colors, and we use the measured fluxes and errors even in the case of nondetections (as opposed to using upper limits). We also perform two ancillary runs, which we use for later vetting. One uses fluxes measured in 03 circular apertures (to cover the possibility that a Kron ellipse was drawn inaccurately, which happens in the presence of bright neighbors). A second run had a maximum redshift of z = 7 to allow the exploration of secondary redshift solutions.

To select our sample of candidate very high redshift galaxies, we follow previous work done by our team (Finkelstein et al. 2015, 2022; Rojas-Ruiz et al. 2020; Bagley et al. 2022). We make use of photometric signal-to-noise (S/N) criteria to ensure robust photometric detections (to minimize the chance of a spurious signal) and nondetections below the Lyα break. We add to these several criteria based off of the full eazy redshift PDF (denoted ${ \mathcal P }(z)$). We note that the criteria imposed here are fairly conservative; we wish to identify the most robust highest-redshift candidates. Future work will explore how to relax some of these criteria to improve sample completeness and push to z < 12 without introducing unacceptable levels of contamination.

To derive an initial sample of z ≥ 12 galaxies, we first impose all of following requirements.

• 1.  
  An S/N in both F200W and F277W of >6 in conservatively small 02 (6.7 pixel) diameter apertures for these measurements using both the fiducial (global) and local noise estimates. We note that enforcing an S/N cut in F200W effectively limits this sample to z ≲ 15, though by enforcing detections in both an SW and an LW detector, we limit the inclusion of detector-specific spurious sources.
• 2.  
  Error map values of <1000 (indicating coverage by the majority of exposures) in F115W, F150W, F200W, F277W, and the detection image.
• 3.  
  Initial more inclusive photometric redshift cuts of $\int { \mathcal P }(z\gt 8)$ ≥ 0.9, z_best > 8.5, and ${\chi }_{\mathrm{eazy}}^{2}\,\lt$ 20 (to reject poor eazy fits), and that the intergrated ${ \mathcal P }(z)$ in a Δz = 1 width bin centered on integer redshift values (z_sample) with the largest integrated ${ \mathcal P }(z)$ to be at z_sample ≥ 9.
• 4.  
  Objects at z_sample > 10 must have S/N ≤ 2.0 in F115W, while objects at z_sample > 13 must have S/N ≤ 2.0 in both F115W and F150W (in both the global and local noise in 02 diameter apertures). These redshifts correspond to the wavelength of the Lyα break leaving a given dropout filter.
• 5.  
  An F200W magnitude of <29, to focus on well-detected objects regardless of formal S/N.

After this initial set of selection criteria, we examined the resulting objects. We inspected their spectral energy distributions (SEDs), image stamps, and ${ \mathcal P }(z)$ plots. We noticed several low-confidence sources that could be identified with further automated cuts. We thus implemented the following additional set of selection criteria.

• 1.  
  We additionally implement all of the above detection significance criteria, in both the detection and dropout bands, in a 03 diameter aperture to account for situations where faint flux was visible slightly off-center of the source barycenter.
• 2.  
  We require the χ² from an additional eazy run with a maximum redshift of 7 to have a significantly worse fit than our fiducial run via ${\chi }_{\mathrm{Low}-z}^{2}-{\chi }_{\mathrm{fiducial}}^{2}\gt 4$.
• 3.  
  We impose a single color cut of F200W − F444W < 1 for objects with z_sample ≤ 13 to reduce the incidence of red low-redshift interlopers. This is similar to the color cuts simulated by Hainline et al. (2020) and implemented by Castellano et al. (2022). For objects with z_sample = 13–18, the Lyα break falls in the F200W filter; thus, we require F277W − F444W < 1.
• 4.  
  To account for situations where the Kron aperture could be affected by nearby bright sources, we also require $\int { \mathcal P }(z\gt 8)$ ≥ 0.5 from an independent eazy run performed with colors measured in 03 circular apertures.

As our focus here is on the highest-redshift sources, we limit our analysis to objects with z_sample ≥ 12. Running the above selection process on all four fields, we find a single galaxy candidate that satisfies all of the above criteria. We perform an initial visual inspection of this candidate, inspecting 15 image stamps in all filters and 5'' image cutouts in F200W and the detection image, and find that this object appears astrophysical in origin and is not an artifact.

This source, CEERS J141946.36+525632.8, hereafter known as "Maisie's Galaxy," ⁹⁹ was detected in the CEERS2 field. Promisingly, it was first identified in the earliest (v0.02) internal CEERS reduction in this field, being the first z ≳ 12 candidate viewed on 2022 July 18. In each subsequent reduction, this source continued to satisfy all selection criteria, becoming progressively more robust as the data became cleaner. Several CEERS team members viewed this potential z ≳ 12 candidate on 2022 July 22 and agreed on the robustness of this source. The photometric redshift of this source with NIRCam photometry alone is z = ${12.0}_{-0.4}^{+0.2}$.

Given the much greater sensitivity of NIRCam, we do not expect significant changes to our interpretation based on the inclusion of HST images. These images were not included in our fiducial SE analysis, as they had not yet been pixel-aligned given the short time since NIRCam data acquisition. However, upon inspection, we find a hint of a positive signal at the position of the source in the F160W image. Indeed, while the source is not in the published Stefanon et al. (2017) and Skelton et al. (2014) catalogs, there is a 3.5σ detection at a separation of 015 in the Finkelstein et al. (2022) catalog (this object has S/N < 2 in all other HST filters in this catalog).

Using SE, we perform forced photometry at this position on the CANDELS (e.g., Grogin et al. 2011; Koekemoer et al. 2011) 30 mas images (using the updated images provided by the CEERS team, which has registered the astrometry to Gaia) in the ACS F606W and F814W and WFC3 F125W, F140W, and F160W bands. Following Finkelstein et al. (2022), we apply an additional flux correction of 1.2 to all HST bands to account for the missing wings of the PSF in the larger Kron aperture. We find S/N < 2 in all bands except WFC3 F160W, which has an ∼4σ detection, with a flux consistent with the weak F150W detection for this source. Including this photometry (listed in Table 3) in our photometric redshift fit slightly changes our photometric redshift estimate to z = ${11.8}_{-0.3}^{+0.2}$, which (for our assumed cosmology) corresponds to an age of the universe of ${373}_{-8}^{+16}$ Myr. We use these results inclusive of HST photometry as our fiducial values. The properties of this galaxy are summarized in Table 1, and we list its photometry in Tables 2 and 3.

Table 1. Properties of Maisie's Galaxy

Property	Value
Source ID	CEERS J141946.36+525632.8
R.A. (J2000 [deg])	214.943153
Decl. (J2000 [deg])	52.942442
zeazy	${11.8}_{-0.3}^{+0.2}$
${{ \mathcal T }}_{\mathrm{Big}\ \mathrm{Bang}}$	${373}_{-8}^{+16}$ Myr
MUV (mag)	−20.32${}_{-0.06}^{+0.08}$
β	−2.47${}_{-0.09}^{+0.09}$
log (M*/M⊙)	${8.50}_{-0.44}^{+0.29}$
Av (mag)	${0.07}_{-0.06}^{+0.23}$
SFR10 Myr (M⊙ yr−1)	${2.1}_{-2.0}^{+4.8}$
log sSFR10 Myr (yr−1)	−8.2${}_{-1.6}^{+1.0}$
Mass-weighted age (Myr)	${18}_{-9}^{+18}$

Note. Here ${{ \mathcal T }}_{\mathrm{Big}\ \mathrm{Bang}}$ is the time elapsed from the Big Bang to the photometric redshift for our assumed cosmology, and M_UV and β were computed from the Prospector models using the same techniques as in Tacchella et al. (2022). The physical properties listed below the horizontal line were derived with Prospector.

Download table as:  ASCIITypeset image

Figure 1 shows cutouts of this candidate galaxy in the NIRCam bands, while Figure 2 shows two color composites. Figure 3 shows the observed SED of our candidate with photometric redshift fits. The confidence of this source as a robust very high redshift galaxy is easy to see from all three of these figures. The Lyα break color, here F115W − F200W, is >1.9 mag (2σ lower limit), completely eliminating any known low-redshift interloper (the F150W − F200W color is 1.3 mag, still strong, though smaller in amplitude due to the Lyα break being present at the very red edge of F150W). Such a model would need to have an extremely red color to match our F115W − F200W > 1.9 mag break but then have a very blue color. While lower-redshift passive or dusty galaxies can mimic high-redshift Lyα breaks, the observed >1.9 mag break is much larger than known populations of low-redshift galaxies. Such galaxies would also be fairly red redward of the break.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Though differential geometry could accommodate UV spectral slopes as blue as β ∼ −1, this object has β ∼ −2.5 (see Section 5 of Casey et al. 2014). The significant detection in four broadband filters also rules out low-redshift extreme emission line galaxies. We show as the orange curve in Figure 3 eazy's best-fitting low-redshift model, which is ruled out at high confidence. Based on the nondetection in F115W and strong detection in F200W, the implied redshift is z > 11. This is confirmed by the eazy fit, shown as the blue line, which prefers z ∼ 11.8.

Figure 1 shows 18 cutout images of this source at multiple wavelengths. This source shows the expected pattern for a high-redshift galaxy, with no significant flux in multiple dropout bands and robust flux in redder bands. The very sharp break between F115W and F200W is consistent with a redshift of z > 11 (the faint detection in F150W pushes the redshift solution to z ∼ 12, as opposed to higher redshift). The advantage of JWST is clear here, as this source is well detected in all five NIRCam filters redward of the break. This multiband detection essentially eliminates the possibility of a spurious nature. Of note is that while persistence from previous observations affected several HST programs (see discussion in Bagley et al. 2022 and Finkelstein et al. 2022), CEERS observed with the bluest filters first; thus, any flux from persistence would be most apparent in F115W.

To further rule out a spurious nature, the science, error, and data-quality images were visually inspected at the position(s) of the best candidate(s). This is to ensure that the detected sources in the coadded images are not just chance superpositions of regions that were affected by cosmic rays or other artifacts. In the case of the z ∼ 12 candidate, the source is visible in all of the individual F200W, F277W, and F356W exposures and overlaps with a cosmic ray in only a few images. Even in those cases, the cosmic rays that are masked in the data-quality array are of the typical size that is cleanly rejected in the jump-detection step of the pipeline. There were no overlaps with the larger "snowball" charged-particle events.

As an additional check, we measured photometry at the position of this source on our images without our postprocessing residual background subtraction step to ensure that any systematic effects at this source position did not affect our results. The images already have a pedestal background-subtracted in the pipeline, so the relative colors should be secure when measured in this way. We found that this set of photometry was consistent with our fiducial photometry, and eazy returns $\int { \mathcal P }(z\gt 11)$ = 0.995. Therefore, it is unlikely that our sky-subtraction routine negatively affected our result.

Dust-reddened foreground galaxies are another potential source of contamination. However, Maisie's Galaxy is not significantly detected in the deepest mid- and far-infrared, submillimeter, and radio data available for this sky region, including Spitzer MIPS 24 μm (Magnelli et al. 2009); Herschel PACS 100 and 160 μm (Lutz et al. 2011); Herschel SPIRE 250, 350, and 500 μm (Oliver et al. 2012); JCMT SCUBA2 850 μm (Geach et al. 2017); and VLA 10 cm (M. Dickinson 2022, private communication). The SCUBA2 nondetection is explored in more detail in Zavala et al. (2022).

4.2.1. Stellar Screening

4.2.2. Photometric Accuracy

4.2.3. Photometric Redshift Accuracy

4.2.4. Contamination Estimation

The five photometric detections afforded by NIRCam allow us the unprecedented opportunity to study the physical properties of a galaxy potentially only ∼400 Myr after the Big Bang. Our fiducial stellar population modeling is done with the Prospector Bayesian SED-fitting code (Johnson et al. 2021). We follow the same procedures as in Tacchella et al. (2022), and we refer the reader there for more details. Briefly, we model the SED with a 13-parameter model that includes redshift (where here prior is set to the posterior of eazy, unlike in Section 4.2.3, where it was free), stellar mass, stellar and gas-phase metallicities, dust attenuation (two-component dust model including birth-cloud dust attenuating young stars (<10 Myr) and nebular emission, a diffuse component for the whole galaxy with a flexible attenuation law; three parameters), and an ionization parameter for the nebular emission. We adopt a flexible star formation history (SFH) prescription with six time bins (the first two look-back time bins are spaced at 0–5 and 5–10 Myr, while the other four are log-spaced out to z = 20; five free parameters) and the bursty continuity prior. Furthermore, we assume the MIST stellar models (Choi et al. 2017) and a Chabrier (2003) initial mass function (IMF).

To explore how robust these properties are, we perform an independent fit with the Bayesian Bagpipes (Carnall et al. 2018), Cigale (Burgarella et al. 2005; Noll et al. 2009; Boquien et al. 2019), and Dense Basis (Iyer & Gawiser 2017; Iyer et al. 2019) SED-fitting codes. For Bagpipes, we assumed a simple exponential SFH with a Chabrier IMF and a Calzetti dust attenuation law and included nebular emission with an ionization parameter of 10⁻³, with Bruzual & Charlot (2003) stellar population models. For Cigale, we assume a delayed SFH after checking that adding a burst does not significantly modify the results. Bruzual & Charlot (2003) models with a Chabrier IMF and were used. Dense Basis was run using the flexible nonparametric SFH model and priors described in Iyer et al. (2019) assuming a Calzetti dust law and a Chabrier IMF. Metallicities in all four codes were allowed to vary to allow the uncertainty in metallicity to be included in the uncertainties on other parameters, though this parameter is not well constrained with photometry alone. We note that Prospector and Cigale assume an error floor (of 5% and 10% of the flux, respectively), and that Prospector, Dense Basis, and Bagpipes always fit the measured fluxes, while Cigale uses the 1σ uncertainty as both the flux and error when the measured fluxes have S/N < 1.

The marginalized posterior values of the inferred physical properties from Prospector are summarized in Table 1 and Figure 4. We infer a stellar mass of $\mathrm{log}({M}_{* }/{M}_{\odot })={8.5}_{-0.4}^{+0.3}$. The attenuation in this galaxy is rather low, with ${A}_{V}\,={0.07}_{-0.06}^{+0.22}\,\mathrm{mag}$, though we stress that this is not well constrained because we only fit the rest UV, and it is degenerate with the slope of the attenuation law (which is variable in this fit). However, the low dust attenuation is in agreement with the measured UV spectral slope β = −2.47${}_{-0.09}^{+0.09}$ (measured using the same techniques as in Tacchella et al. 2022). This blue color implies little dust, though it does not require extremely low metallicities (e.g., Finkelstein et al. 2012; Dunlop et al. 2013; Bouwens et al. 2014). Interestingly, this galaxy is about as blue as z ∼ 7 galaxies of similar mass (Finkelstein et al. 2012), implying little evolution in chemical enrichment between these two epochs for similar-mass galaxies.

Zoom In Zoom Out Reset image size

We infer an SFR₁₀ (average of the past 10 Myr ¹⁰⁰ ) of 2.1 M_⊙ yr⁻¹, and the corresponding sSFR₁₀ is 10^−8.2 yr⁻¹. By looking at the posterior distribution of the SFH, it becomes apparent that the model for this galaxy had an episode of elevated star formation 10–20 Myr ago with an SFR of ${8}_{-7}^{+17}\,{M}_{\odot }\,{\mathrm{yr}}^{-1}$; i.e., the SFR has been slightly decreasing in the recent 10 Myr. This explains the mass-weighted age of ${18}_{-9}^{+18}$ Myr. This is also consistent with the half-mass formation time of dark matter halos at z ∼ 12 of a few tens of megayears (Tacchella et al. 2018).

These Prospector-based posterior distributions are consistent with the ones from Bagpipes, Cigale, and Dense Basis (see Figure 4), though the difference in age is large (age is defined at half-mass time, t₅₀, which is close to the mass-weighted age). Bagpipes, Cigale, and Dense Basis prefer higher age values (although the posterior distributions are also broader) of ${67}_{-37}^{+56}$, 34 ± 20, and ${64}_{-54}^{+63}$ Myr, respectively. The SFH inferred from Dense Basis shows a recent burst of star formation in the last ∼30 Myr. The larger mass-weighted age comes from the long tail of low-level star formation in the galaxy leading up to the recent burst. The spread in these results could be explained by differences in SFHs (e.g., nonparametric versus parametric), as well as the lack of observational constraints in the rest-frame optical. These differences also lead to Prospector having the lowest SFR₁₀ value.

Several pre-JWST studies have focused on inferring SFHs and stellar ages of z ≈ 8–10 galaxies (e.g., Hashimoto et al. 2018; Laporte et al. 2021; Stefanon et al. 2022). Specifically, Tacchella et al. (2022)—using Prospector with the same bursty continuity prior—found a diversity of stellar ages, ranging from 10 to 260 Myr, and stellar masses (10⁹–10¹¹ M_⊙), with more massive galaxies being older. In particular, the galaxies at z ≈ 9–10 with stellar masses at the higher end and older ages (t₅₀ ≈ 100 Myr) are consistent with being the descendants of Maisie's Galaxy. Recently, Naidu et al. (2022) inferred the properties of two galaxies at z ≈ 10.6 and 12.4 (see also Castellano et al. 2022) with Prospector and a similar setup, allowing us to do a useful comparison. Their two galaxies have $\mathrm{log}({M}_{* }/{M}_{\odot })={9.4}_{-0.3}^{+0.3}$ and ${9.0}_{-0.4}^{+0.3}$, ${{\rm{S}}{\rm{F}}{\rm{R}}}_{50{\rm{M}}{\rm{y}}{\rm{r}}}\,={12}_{-4}^{+9}$ and ${7}_{-3}^{+4}\,{M}_{\odot }\,{\mathrm{yr}}^{-1}$, and ${t}_{50}={111}_{-54}^{+43}$ and ${71}_{-32}^{+33}\,\mathrm{Myr}$, respectively. This is older than what we infer for our galaxy, though this age difference could be explained by the stellar mass difference, along with the higher preferred redshift for Maisie's Galaxy. Importantly, detailed stellar population analyses of early galaxies will advance significantly with JWST, particularly when including spectroscopic information.

We derive the size of Maisie's Galaxy using two morphological fitting codes, GalfitM ¹⁰¹ (Häußler et al. 2013) and statmorph ¹⁰² (Rodriguez-Gomez et al. 2019). GalfitM is a modified version of Galfit ¹⁰³ (Peng et al. 2002, 2010), a least-squares fitting algorithm that finds the optimum Sérsic fit to a galaxy's light profile. We perform fits using GalfitM by allowing the Sérsic index to vary between 0.01 and 8, the magnitude of the galaxy between 0 and 45, and r_half between 0.3 and 200 pixels (on our 003 pixel scale). As input, we use a 100 × 100 pixel cutout of the F277W science image, the segmentation map created by SE, and the empirical PSF measured from our CEERS2 pointing, which we allow GalfitM to oversample relative to the data by a factor of 9. We estimate the uncertainty on our fits by conducting a Monte Carlo analysis where we modify the input F277W science image to randomly vary the pixel-to-pixel noise, recompute the parameters, and then repeat this analysis 40 times.

Following this procedure, we measure a half-light radius of 3.0 ± 0.12 pixels (009 ± 00036), which corresponds to a physical size of 340 ± 14 pc at z = 11.8. We check these results using the standard configuration of Statmorph, a Python package developed to calculate the nonparametric morphology of galaxies, as well as compute single Sérsic fits. Using the same images as input, we find a half-light radius of 2.9 pixels, in good agreement with the measurement from GalfitM. We repeat this measurement for the F200W filter and a stacked F200W + F277W image and find consistent results. The measured half-light radius of 3.0 ± 0.1 pixels is significantly larger than that expected for a point source (the median r_h for our PSF stars is 1.95 ± 0.22 pixels in F277W), further ruling out a stellar origin for this source.

In Figure 5, we present predictions from a range of theoretical models: the First Light And Reionisation Epoch Simulations (FLARES; Lovell et al. 2021; Vijayan et al. 2021; Wilkins et al. 2022), a suite of hydrodynamical cosmological zoom simulations; the large periodic volume hydrodynamical simulation Bluetides (Feng et al. 2016; Wilkins et al. 2017); the Delphi (Dayal et al. 2014, 2022) and Santa Cruz (Yung et al. 2019, 2020) semianalytical models; the semiempirical UniverseMachine (Behroozi et al. 2020), Mason et al. (2015); and Behroozi & Silk (2015) models. For the FLARES, Delphi, and Behroozi & Silk (2015) models, we show both the attenuated and unattenuated (intrinsic) predictions. These predictions were made by interpolating and integrating either the binned or Schechter luminosity functions across z = 15 → 12, taking account of the areal size of the CEERS observations. Almost all of these models predict an expected source density much less than 1, making the observation of even a single object at this redshift and magnitude surprising and potentially hinting at significant differences between the physical assumptions in these models and the real early universe.

Zoom In Zoom Out Reset image size

The exception is the Behroozi & Silk (2015) model, which extrapolated galaxy formation to high redshifts by assuming that the ratio between galaxies' sSFRs and their host halos' specific accretion rates remained constant, which they showed was equivalent to assuming that galaxies' stellar masses are proportional to a power of their host halo masses. This model was constrained only with observational data at z ≤ 8 and predicted no change in the slope of the CSFR relation with redshift at z > 8. As a result, it predicted many more high-redshift galaxies than later models that were constrained to match z ∼ 9 and ∼10 data from HST that suggested more rapid declines in the number densities of early galaxies. We caution against overinterpretation, as the current sample contains only a single object with a consequently large Poisson error in addition to additional uncertainty due to cosmic variance. Nevertheless, if confirmed, the existence of this object places informative constraints on galaxy formation models in this epoch.

We are now only just getting our first glimpse into this epoch with the first JWST data. Nonetheless, we can compare our observed number density to a few recent observations. We calculate a rough number density for M_UV = −20.3 galaxies assuming a top-hat selection function over 11.5 < z < 12.5. This is overly simplistic and does not account for incompleteness (which, although this is a >10σ detection, certainly is nonunity due to our stringent selection criteria). Nonetheless, it is illustrative of a rough number density. We find a maximum volume over the CEERS first-epoch area of 5.0 × 10⁴ Mpc³ for a number density for our singular source of 2.0${}_{-1.7}^{+4.6}$ × 10⁻⁵ Mpc⁻³ (where the uncertainties are Poisson based on our detection of one object).

We illustrate this number density in Figure 6. Our derived number density is not inconsistent with a variety of observational constraints at z ∼ 10, as well as recent results at z ∼ 12–13. The solid gray line shows the predicted z = 12 Schechter function from Finkelstein (2016), which is extrapolated from an empirical fit to observations at z = 4–8, assuming smooth redshift evolution. Interestingly, our rough number density measurement is above this prediction, though consistent within the 1σ uncertainties, which would support its assumption of a smooth decline in the luminosity function (and the corresponding SFRD). Our results are also consistent with the smoothly evolving double power-law (DPL) model at z = 12 from Finkelstein & Bagley (2022), again supporting a smooth decline in the rest-UV luminosity function to z > 10.

Zoom In Zoom Out Reset image size

As noted by several previous studies, the bright end of the luminosity function at z ≥ 9 exhibits an interesting excess over predicted levels (e.g., Morishita et al. 2018; Bowler et al. 2020; Rojas-Ruiz et al. 2020; Bagley et al. 2022; Finkelstein et al. 2022). While our survey area does not yet probe the volume densities needed to reach these brighter potential z = 12 galaxies, if the high-redshift luminosity function follows a DPL form, the forthcoming 0.6 deg² COSMOS-Web survey (PIs: Kartaltepe & Casey) should be able to discover this population. In combination with the full Cycle 1 slate of surveys, including the completed CEERS imaging, it will afford a more complete view of the z = 12 universe.