3D-HST: A WIDE-FIELD GRISM SPECTROSCOPIC SURVEY WITH THE HUBBLE SPACE TELESCOPE^*

The WFC3 G141 grism is the primary spectral element used for the 3D-HST survey. The combined transmission of the HST optical telescope assembly and the primary spectral order of the G141 grism ("+1st") is greater than 30% from 1.10 to 1.65 μm, reaching a peak of nearly 50% at 1.45 μm. The mean dispersion of the +1st order is 46.5 Å pixel⁻¹ (R ∼ 130) and varies by a few percent across the field of view. The uncertainties of the wavelength zero point and dispersion of the G141 grism are 8 Å and 0.06 Å pixel⁻¹, respectively. For a full description of the calibration of the WFC3/G141 grism, see Kuntschner et al. (2010). Spectral features covered by the G141 grism include Hα at 0.7 < z < 1.5, [O iii] λ5007 at 1.2 < z < 2.3, [O ii] λ3727 at 2.0 < z < 3.4, and the Balmer/4000 Å break at 1.8 < z < 3.1 (Figure 1). The nominal G141 dispersion corresponds to ∼1000 km s⁻¹ for Hα at z > 1; however, in practice the resolution of the slitless grism spectra is determined by the physical extent of a given object (see Section 4.2).

Zoom In Zoom Out Reset image size

Observations with the HST grisms typically require an accompanying image taken with an imaging filter to establish the wavelength zero point of the spectra (see, e.g., Kümmel et al. 2009). For 3D-HST, we obtain these so-called direct images in the broad F140W filter that spans the gap between the standard J and H passbands and lies roughly in the center of the G141 sensitivity (Figure 1). While CANDELS will eventually provide significantly deeper imaging of the 3D-HST fields in the F125W and F160W WFC3 filters, the 3D-HST F140W direct images can be useful for scientific analysis in addition to calibrating the grism, as they reach depths competitive with even the deepest ground-based surveys (H ∼ 26.1, 5σ) with spatial resolution ∼013 (see Section 3.2 and also van Dokkum et al. 2011).

In addition to the primary WFC3 observations, 3D-HST obtains parallel ACS F814W imaging and G800L grism spectroscopy. The G800L grism covers wavelengths 0.55–1.0 μm with a dispersion of 40 Å pixel⁻¹ and a resolution of 80 Å for point-like sources (Kümmel et al. 2011b). The parallel spectroscopy extends the Hα line sensitivity of the survey to z = 0 and will provide coverage of additional lines at certain redshift intervals where only a single line is visible in WFC3/G141, for example, [O iii] in G800L and Hα in G141 at 0.7 < z < 0.8 (Figure 1).

The 3D-HST survey is allocated 248 primary+parallel orbits over cycles 18 and 19. 3D-HST will cover roughly 75% of the area imaged by the CANDELS survey in the EGS/AEGIS, COSMOS, UKIDSS-UDS, and GOODS-South fields (Figure 2). Furthermore, G141 grism coverage of most of the GOODS-N field from program GO-11600 (PI: B. Weiner) is incorporated into 3D-HST as the observational strategy of the GOODS-N observations is nearly identical to that of 3D-HST. These parts of the sky are the best-studied extragalactic survey fields, offering a wealth of deep, multi-wavelength imaging and spectroscopic observations from a large number of previous ground-based and space-based surveys. Grogin et al. (2011) provide a detailed summary of the ancillary data sets available in these survey fields, from X-ray to radio wavelengths, which are crucial for the interpretation of both the WFC3 imaging and grism spectroscopy data sets (see also Section 4.2).

Zoom In Zoom Out Reset image size

The 248 3D-HST orbits are divided among 124 individual visits of two orbits each. The layout of the 3D-HST pointings is shown in Figure 2 and summarized in Table 1. In order to schedule 3D-HST concurrently with CANDELS observations of the same fields over cycles 18 and 19, initially no ORIENT constraints were imposed on any of the visits. After scheduling, the ORIENTs were fixed and the positions of the individual pointings were optimized to provide contiguous mosaics and maximum overlap between the primary WFC3 G141 and parallel ACS G800L observations, as shown in Figure 2. Owing to this optimization, fully 90% of the G141 mosaic (excluding GOODS-N) will be covered by between two and four orbits of the ACS grism. Within the GOODS-South mosaic, four additional two-orbit visits at the same orientation are centered on the Ultra Deep Field (UDF). The GOODS-South pointings outside of the area with CANDELS coverage provide WFC3 grism spectroscopy of the HUDF09¹⁵ and WFC3-ERS fields (see Bouwens et al. 2011).

The first 3D-HST exposures were obtained in 2010 October, and the survey is nearly completed as of 2012 March: all pointings in the COSMOS, GOODS-S, and UDS fields have been observed, and two failed pointings in AEGIS will be re-observed by the end of 2012. The 28 grism pointings in GOODS-N were completed in 2011 April.¹⁶ Due to scheduling constraints, a given field will not have both complete CANDELS and 3D-HST coverage before both surveys are finished. For example, the two epochs of CANDELS observations in the UDS were completed in 2011 January, while the 3D-HST coverage of that field was only completed in 2012 February.

Each of the 3D-HST two-orbit visits with WFC3 is structured in an identical fashion: four pairs of a short F140W direct image followed by a longer G141 grism exposure. The four pairs of direct+grism exposures are separated by small telescope offsets to enable the rejection of hot pixels and pixels affected by cosmic rays, as well as dithering over some WFC3 cosmetic defects such as the "IR-blobs" (Pirzkal et al. 2010). The dither pattern is shown schematically in Figure 3. The sub-pixel offsets of the dither pattern are chosen to improve sampling of the WFC3 point-spread function (PSF), which enables some recovery of the image quality lost by the pixels that undersample the PSF by a factor of two (see, e.g., Fruchter & Hook 2002; Koekemoer et al. 2011). Due to the offset and rotation of ACS with respect to WFC3, the pixel subsampling of the ACS parallel exposures is somewhat less than optimal (Figure 3).

Zoom In Zoom Out Reset image size

All of the WFC3/IR exposures are obtained in MULTIACCUM mode using either a SPARS50 (F140W) or SPARS100 (G141) read-out sequence. All of the direct exposures have NSAMP = 5, corresponding to individual exposure times of 203 s and a total of 812 s for each visit/pointing. Depending on the varying usable length of an orbit, the G141 exposures have NSAMP = 12–15 and the total grism exposure times range from 4511 to 5111 s per visit/pointing. The ACS F814W direct images all have total exposure times of 480 s per visit, and the G800L grism images have exposure times ranging from 2925 to 3523 s. The ACS exposure times are somewhat less than those of WFC3 as a result of the larger overheads for reading out the ACS/WFC detectors.

We use as a starting point the standard calibrated data products (images with the flt extension) provided by the HST archive that have been processed by the calwf3 and calacs reduction pipelines for the WFC3 and ACS 3D-HST exposures, respectively. The calibrated WFC3/IR images have 1014 × 1014 pixels, with roughly 0128 pixel⁻¹. The ACS/WFC images consist of two 4096 × 2048 pixel extensions with roughly 005 pixel⁻¹. Briefly, the calibration pipelines flag known bad pixels in the data quality image extensions, subtract the bias structure determined from zeroth reads and overscan regions, subtract dark current, and apply multiplicative corrections for the detector gain. Additionally, the pipelines apply a multiplicative flat-field correction to the WFC3/F140W and ACS/F814W direct images. The flat fielding of the grism exposures is discussed below in Section 3.2.2. Finally, bias striping (Grogin et al. 2010) and charge-transfer efficiency (Anderson & Bedin 2010) corrections are applied to the ACS flt pipeline products. A more comprehensive description of the WFC3 and ACS reduction pipelines is given by Koekemoer et al. (2011).

With the calibrated images obtained as described above, we perform a number of additional preparation steps in order to produce mosaics of each visit independently, which are composed of the four-exposure sequences described in Section 2.3. First, we use the MultiDrizzle software (Koekemoer et al. 2002) to identify any hot pixels or cosmic rays not flagged by the instrument calibration pipelines. We use essentially the same strategy as described by Koekemoer et al. (2011), though we note that we had to increase the first-pass detection threshold of the cosmic-ray rejection to avoid rejecting the central pixels of stars. The final data product after applying the alignment and sky-subtraction steps described below is an undistorted mosaic with a final scale of 006 pixel⁻¹ (with PIXFRAC = 0.8), which is justified by the optimal sampling of the WFC3 PSF by the 3D-HST dither pattern (Figure 3). The 5σ depth of the F140W detection images is H₁₄₀ ≈ 26.1 for point sources within a 05 diameter aperture, varying by ±0.3 mag due to the field-dependent background levels (see Table 1).

3.2.1. WCS Alignment

3.2.2. Background Subtraction and Grism Flat Fielding

While the near-IR background is much lower in low-Earth orbit than it is from the ground, the background flux is still a significant component that must be subtracted from both the direct and grism exposures. Subtracting the background from the direct images is done in the following way: we subtract a second-order polynomial fit to each exposure after aggressively masking objects detected in the MultiDrizzle mosaic and mapped back to the distorted frame using the PyRAF blot routine. A polynomial fit with order greater than zero is necessary as some exposures show structure in the background likely caused by enhanced airglow for a particular pointing orientation with respect to the Earth limb.

Subtracting the background from the grism exposures is more difficult than for the direct images. There is significant structure in the grism background that is a result of a superposition of multiple dispersed grism orders and the fact that the finite field of view of the entrance window causes different regions of the detector to see different combinations of the spectral orders. Effort has been made to produce a "master" grism background image created from the (masked) average of many grism exposures (Kümmel et al. 2011a), which can then be scaled to and subtracted from a particular grism exposure. This is the approach followed by the aXe software package (Kümmel et al. 2009).

The 3D-HST fields span a wide range of celestial coordinates that result in distinct orientations of the instrument with respect to the zodiacal and Earth glow sources of the background light. From the large number of 3D-HST grism observations that have already been obtained, we find that a single master G141 background image is insufficient to explain the observed variety of structure in the background. We therefore correct the background in the following way. We produced four representative master background images that are median (masked) combinations of subsets of 3D-HST grism exposures that have similar overall structure, determined by eye. These images are shown in Figure 4. The primary features in the grism background are the dark vertical bands spanning roughly 100 pixels at the left and right sides of the image. The relative intensity of these bands and the sharpness of their edge are correlated with the overall background level, but they can also vary at a given background level for images in different survey fields. There are dark horizontal bands in some of the master images that look like negative grism spectra. These features are the result of the grism "dispersing" the IR-blob features of decreased sensitivity, which are not physically located on the detector itself but rather on the Channel Select Mechanism in the WFC3/IR optical path (Pirzkal et al. 2010).

Zoom In Zoom Out Reset image size

The master background images shown in Figure 4 look very different from the master background image produced by Kümmel et al. (2011a) and distributed with the G141 calibration files. For the images in Figure 4 we have divided out the F140W imaging flat field before creating the image combinations in order to separate the pixel-to-pixel variations of the flat field from the more smoothly varying background. We find that the wavelength dependence of the flat field is less than ∼1% across the G141 sensitivity (Appendix A). This is similar to the uncertainties of the overall G141 flux calibration, and we therefore opt for a simplified approach by dividing out the single F140W flat field from all of the G141 grism exposures. We then subtract the scaled master background image whose structure best matches that of a particular exposure, as determined from a χ² test. Note that it is possible that two subsequent grism exposures within a given visit require different master background images depending on the varying background within an orbit (Appendix B). Finally, we subtract the median level of the resulting (masked) background pixels averaged along image columns. While adopting the variable background images greatly improves the overall background subtraction compared to using the single image, we find that this last step is necessary to remove any residual structure in the background, which tends to be parallel to image columns. The complete process of first dividing by the flat field, then subtracting the master background image, and finally removing the x-dependent residuals is demonstrated in Figure 5. The background in the final corrected grism images is typically flat to better than 1%.

Zoom In Zoom Out Reset image size

We have described the flat fielding and background subtraction in considerable detail as these steps are crucial for minimizing systematic effects in the spectra for this blank-field spectroscopic survey. Local background subtraction is not feasible for the relatively deep 3D-HST grism exposures because it is generally impossible to identify pixels adjacent to a given object that will be entirely free of flux contributed by nearby overlapping spectra. Small uncorrected errors in the background subtraction can later be manifested as continuum break features in the extracted spectra, particularly at fainter magnitudes where 3D-HST can provide truly unique near-IR spectra currently unobtainable from the ground.

We use the aXe software package (Kümmel et al. 2009) to extract the grism spectra in much the same way as described by van Dokkum et al. (2010) and Atek et al. (2010). The primary inputs to the extraction software are a detection image mosaic (F140W or F814W) and the individual background-subtracted grism exposures generated as described in Section 3.2. An object catalog is generated from the detection image with the SExtractor software (Bertin & Arnouts 1996), which also produces a segmentation map that indicates which pixels in the direct image are assigned to each object. For a given pixel within a particular object's segmentation map, the calibration of the HST grisms determines where the dispersed light from that pixel will fall on the grism exposures, with the pixel in the direct image defining the wavelength zero point for the spectrum. Thus, the grism spectrum of an object is the sum of all of the dispersed pixels within that object's segmentation image, or rather, the spectrum is a superposition of the object profile at different wavelengths offset by the grism dispersion. The effective spectral resolution is therefore a combination of the grism dispersion and the object profile in the sense that the effective resolution decreases with increasing object size in the dispersion direction (see also Section 4.2 and Figure 10).

Because no slit mask is used and the length of the dispersed spectra is larger than the average separation of galaxies down to the detection limit of the 3D-HST survey, the spectra of nearby objects can overlap. This "contamination" of an object's spectrum by flux from its neighbors must be carefully accounted for in the analysis of the grism spectra. We use aXe to produce a full quantitative model of the grism exposures using the information in the direct image as described above. This is the aXe "fluxcube" contamination model, which makes use of the spatial information contained in the high-resolution HST images to model the two-dimensional (2D) grism spectrum. In order to generate a model spectrum based on the direct image, we take the observed F140W flux and full spatial profile within a given object's segmentation map and assume a constant profile and a flat spectrum in units of f_λ. As additional HST photometric bands become available, for example, the CANDELS F125W and F160W imaging covering the 3D-HST survey fields, they will be added to the aXe fluxcube to incorporate the wavelength dependence of both the flux (i.e., the color) and spatial profile into the model.

The relationship between the direct image and the grism spectra and a demonstration of the fluxcube spectral model are shown in Figure 6. The images are oriented as in the individual flt exposures, with the grism spectra offset in the positive x-direction with respect to the direct image and with wavelength increasing toward the right. The main horizontal feature in Figure 6(b) is the "+1st" spectral order, which has the greatest sensitivity (Kuntschner et al. 2010). Compact, point-like features seen in both the observed and model images are zeroth-order spectra. The residuals of the model subtracted from the grism image are shown in Figure 6(d). The model is generally a reliable quantitative representation of the data. It bears mention that for the present reduction there has been no fit to optimize the model—the relatively low level of the residuals demonstrates the quality and stability of the G141 grism calibration. Most of the compact features in the residuals are in fact emission lines, which are not included in the model. The zeroth-order spectra do not always subtract completely, but their presence in the model allows them to be distinguished from emission lines. In future versions of the reduction, we will implement an iterative scheme to refine the spectral model based on the observed spectra.

Zoom In Zoom Out Reset image size

We use aXeDrizzle to combine the four grism exposures (in the original distorted flt frame) of each visit/pointing and extract a 2D spectrum for each object with perpendicular spatial and dispersion axes. We adopt output 006 × 22 Å spectral pixels, which are roughly square with respect to the drizzled pixels in the direct image mosaics. These 2D spectra with HST spatial resolution (∼013 for WFC3/G141) are one of the truly unique products of the 3D-HST survey. As there is no slit defining a spatial axis, an "effective slit" running roughly parallel to the major axis of each object is not generally parallel to the y pixel direction in the undistorted frame. The aXe software includes an option to account for the orientation of this effective slit to optimize the wavelength resolution of the extracted spectra (see Kümmel et al. 2009); however, we simply adopt the spatial axis as perpendicular to the spectral trace and account for the orientation and shape of the object profiles in post-processing analysis (see Section 4.2). Finally, we use aXe to extract optimally weighted 1D spectra from the drizzled 2D spectra, where the relative weights are determined from the object profile perpendicular to the dispersion axis.

We have described above how the non-negligible background emission must be subtracted from the grism exposures. In fact, this background flux from a combination of zodiacal light, Earth glow, and low-level thermal emission is the limiting factor for the sensitivity of the 3D-HST grism exposures: the read noise of the WFC3/IR detector is ∼20 e⁻, while the number of background electrons per pixel in a typical grism exposure is 1.4 e⁻ s⁻¹× 1300 s = 1820 e⁻. The average background level varies for the different survey fields, as shown in Table 1, depending on the field location relative to the ecliptic plane and the date of observation. In general, the observed background levels are consistent with those predicted by the WFC3 Exposure Time Calculator (ETC). The angle of the bright Earth limb can vary within an orbit, and the two grism exposures taken within an orbit can have background levels that differ by as much as 50% (Appendix B) and also different background structure (Section 3.2.2).

We evaluate the effective continuum and emission-line sensitivities of 3D-HST using a suite of simulations that is tied closely to the observed F140W direct and G141 grism exposures. We use a custom developed software package modeled closely after the aXeSIM package (Kümmel et al. 2009) to generate a 2D model spectrum based on (1) the spatial distribution of flux as determined in the F140W direct image and (2) an assumed input (1D) spectrum, normalized to the F140W flux. The 2D grism spectrum is then determined uniquely by the grism configuration files provided by STScI that specify how the flux in a given pixel of the direct image is dispersed into the spectrum of the grism image. These scripts will be described in more detail and released to the community in a subsequent publication; for a simple spectral model of a flat continuum, they produce nearly identical results to aXeSIM and the aXe fluxcube.

The primary advantage of the custom software is that we can easily modify the full input spectrum used to generate the model: here we assume a simple continuum, flat in units of f_λ, combined with a single emission line at 1.3 μm (i.e., Hα at z = 1). The emission line has a fixed equivalent width of (arbitrarily) 130 Å, observed frame, and the overall normalization of the spectrum (and thus the integrated line flux) is set to the flux_auto flux measured by SExtractor on the F140W image. The result is very much like the aXe "fluxcube" model shown in Figure 6; however, each modeled spectrum has the same line+continuum shape. We add realistic noise to the simulation using the WFC3/IR noise model in the error extension of the flt images, which includes terms for the Poisson error of the source counts and the read noise and dark current of the WFC3 IR detector.¹⁷ Thus, the simulations fully account for noise variations as a function of background level across all of the available pointings, and, most importantly, for the true distribution of source morphologies as a function of brightness within the 3D-HST survey. There are approximately 13,000 objects in the simulation.

After computing the full grism image models, we extract individual spectra with the standard optimal extraction weighting (Horne 1986) and measure the median continuum signal-to-noise ratio (S/N) between 1.4 and 1.6 μm, averaged over a typical G141 resolution element of 92 Å. The emission-line strengths are extracted using the technique described in Section 4.2, and uncertainties on the line fluxes are determined with an MCMC fit of the line + continuum template combination. Thus, the continuum and line S/N is measured in a very similar way as is done in the analysis of the observed spectra. Because of the optimal weighting in the spectral extraction, the effective "aperture" of the extraction is likely somewhat larger (and variable) compared to the 1 × 3 pixel extraction window used by the WFC3 ETC.

The result of these simulations is shown in Figure 7. The continuum and emission-line S/N depends, clearly, on the signal itself (S), the morphological properties of the objects, and the noise properties of the grism exposures. We explore the dependence of the grism S/N on the signal itself, the object half-light radius, R₅₀ (SExtractor 50% flux_radius), the background count rate, and a concentration index defined as C = R₅₀/R₉₀. The dependence of S/N on all parameters but the one shown on the ordinal axis is removed in each panel to isolate the contribution of each individually. After the strength of the signal itself, the S/N depends strongly on the object size, R₅₀, and background level, with only a weak, if any, dependence on concentration. The brightness where the average curves shown in black in the right panels cross the S/N = 5 threshold—m₁₄₀ = 23.1, f_λ = 5.5 × 10⁻¹⁷ erg s⁻¹ cm⁻²—represents an average, not the absolute, sensitivity of the survey. The properties of the "average" galaxy are indicated by where the curves cross δ = 1 in the right three sets of panels in Figure 7: (R₅₀, background, C) ∼ (036, 1.4 e⁻ s⁻¹, 0.44). For the optimal case of point sources and the minimum background, one can read the S/N scaling directly off the figure panels: the 5σ continuum and line limits are approximately 1.4 × 1.2 = 1.7 and 2 × 1.2 = 2.4 times fainter, respectively, or 23.7 mag and 2.3 × 10⁻¹⁷ erg s⁻¹ cm⁻². The sensitivity limits in the high-background COSMOS pointings (>2 e⁻ s⁻¹, Table 1 and Appendix B) are brighter than the average limits by a factor of ∼1/0.8 (0.24 mag).

Zoom In Zoom Out Reset image size

The second column of panels in Figure 7 shows that the continuum S/N is roughly ${\propto} 1/\sqrt{R_{50}}$ while the S/N of the emission lines goes more like ∝1/R₅₀. This effect is expected in the sense that the continuum sensitivity only depends on the object extent in the spatial direction as any extent along the spectral axis is averaged out by the dispersion of the grism. In the case of emission lines, however, it is the pixel area of the emission-line extent that determines their S/N as more extended lines are spread out over more noise from the background in both dimensions. Both the line and continuum S/Ns roughly follow the square root of the background counts (third column, Figure 7). The result that the observed trends are slightly flatter than $1/\sqrt{x}$, where appropriate, likely reflects additional higher-order effects that affect the S/N on the level of ∼5%, whose full characterization is beyond the scope of the present work.

Given the limiting line flux determined as in Figure 7, one can calculate analytically the limiting line equivalent width as a function of observed broadband magnitude, where the broadband passband contains the line and assumptions are made about the continuum shape. Figure 8 shows this limiting equivalent width as a function of magnitude in the F140W filter computed for different integrated line fluxes assuming a flat continuum (f_λ). We note that the input equivalent width assumed in the simulations described above does not figure into this calculation. At the faint end, m₁₄₀ > 24, 3D-HST is sensitive to (observed frame) equivalent widths ≳ 500 Å. At the bright end, the equivalent width limit will reach a floor when the Poisson noise of the continuum flux is similar to the background noise, which occurs only at m₁₄₀ ≲ 19.

Zoom In Zoom Out Reset image size

Figure 9 shows the number of objects brighter than a given magnitude found in each 2 × 2 arcmin WFC3 pointing of the 3D-HST survey. The number counts shown are taken from 70 pointings taken as of 2011 August, and the number of a particular type of object expected in the full 149 pointings of the 3D-HST survey (including the GOODS-N pointings) is shown on the right axis. The large majority of 3D-HST objects are matched within 10 to objects in the ancillary photometric catalogs (>90% at H < 23), though the fraction of matched objects begins to decrease significantly at fainter magnitudes where the shallower ancillary catalogs are incomplete. The fraction of objects whose spectra are significantly contaminated by flux from their neighbors increases with magnitude because the object surface density increases and faint galaxies are more likely to be close to other faint galaxies and also because even the extended envelopes of nearby brighter galaxies can have fluxes similar to faint galaxies. However, the fraction of largely uncontaminated objects is still ∼50% even at the faint limits of the survey. Objects with radii less than ∼018 (SExtractor 50% flux_radius) are clearly separated from resolved objects down to H₁₄₀ ∼ 23, and these point sources (i.e., stars) represent the majority of the brightest objects. The cumulative number counts of galaxies with estimated redshifts z > 1 and z > 2 (Section 4.2) are also indicated in Figure 9. Based on the available pointings, some 7000 galaxies with z > 1 and H₁₄₀ < 23.8 are expected within the full survey. For most magnitudes fainter than this practical continuum limit, the redshift estimates will converge to the limit of pure photometric redshifts limited by the depth of the ancillary photometry. However, 3D-HST will provide precise redshifts and line fluxes for a smaller sample of galaxies with extreme equivalent widths, such as the starbursting dwarf galaxies at z > 1.5 discovered recently by van der Wel et al. (2011; see also Figure 8).

Zoom In Zoom Out Reset image size

The 3D-HST/CANDELS fields are among the best-studied extragalactic survey fields in the sky. All of the fields have extensive multi-wavelength observations from X-ray to radio wavelengths, the details of which can be found in Grogin et al. (2011). For the redshift and SED fitting analysis of the grism spectra, we require deep optical and near-IR photometric catalogs selected at near-IR wavelengths, corresponding to the rest-frame optical for the redshifts of interest. Currently, we simply match the F140W-selected objects within the 3D-HST fields to the deepest K-selected catalogs available with broad multi-wavelength coverage: AEGIS, COSMOS, Whitaker et al. (2011); GOODS-S, Wuyts et al. (2008); GOODS-N, Kajisawa et al. (2011); and UDS, Williams et al. (2009). These catalogs have between 8 (UDS) and 35 (COSMOS) photometric bands from U through the Spitzer-IRAC channels at 3–8 μm. The image quality of the ground-based catalog detection bands (07–10) is not well matched to that of the 3D-HST imaging (013). As a result, multiple 3D-HST objects can be matched to single objects in the photometric catalogs (cf. ∼15% at H = 23), which, among other effects, can produce artificial 3D-HST pairs in the current analysis. Furthermore, even the relatively short F140W exposures are somewhat deeper than the ground-based K-band catalog detection images (e.g., K < 22.8 for the Whitaker et al. 2011 COSMOS catalog; see Figure 9). For these reasons, we are developing PSF-matched and deblended photometric catalogs, following Labbé et al. (2006), selected from the high-resolution, deep 3D-HST F140W and CANDELS images that will provide a one-to-one correspondence between objects with extracted spectra and their measured ancillary photometry (R. Skelton 2012, in preparation).

We estimate the redshifts of 3D-HST galaxies using a modified version of the Eazy code (Brammer et al. 2008). Two examples of the fits are shown in Figure 10. The inputs to the fit are the 1D grism spectra, calibrated and extracted with aXe, combined with broadband and medium-band photometry from the catalogs described above when there is an object in the catalog matched within 10 of a 3D-HST object. We subtract the quantitative contamination model directly from the spectra (Figure 6). Finally, we compute a normalization and a linear tilt to scale the G141 spectra to the J and H broadband and medium-band photometry, in order to minimize residuals affecting the continuum shape that remain after the contamination correction and background subtraction. At a minimum, a zeroth-order normalization is necessary to match the flux-calibrated spectrum to the total photometric fluxes that have been measured within an effective infinite aperture (see, e.g., Whitaker et al. 2011).

Zoom In Zoom Out Reset image size

In the standard operation of the Eazy photometric redshift code, the model fluxes of the high-resolution spectral templates at a given redshift are computed by convolving the templates with the filter transmission curve of each photometric band. We define effective "filter curves" for each grism spectral bin (of each object individually) by convolving a boxcar with the width of the spectral bin with the object profile. This profile is taken from the direct image and averaged along the spectral dimension. This accounts for both the possible tilt of the effective slit defined roughly along the object major axis, which is not necessarily perpendicular to the dispersion axis of the slitless spectra (Section 3.3), and higher-order features in the object profile (e.g., spiral arms). In principle, the template fitting could be done on the full 2D spectra directly, though we opt for the relative computational simplicity of using the extracted 1D spectra for the present analysis. An implicit assumption of this method is that the 2D profile is "gray," i.e., that the profile is the same in the line as in the continuum across the spectral range covered by the grism. Investigating the validity of this assumption is in fact one of the primary goals of the 3D-HST survey (see, e.g., Nelson et al. 2012), made possible by the superb spatial resolution provided by the HST.

The default set of the Eazy spectral templates contains emission lines added by determining an effective SFR for each template and then adding lines with empirically calibrated line ratios (Brammer et al. 2011; see also Ilbert et al. 2010). In order to fit the 3D-HST spectra + photometry, we remove the emission lines from the galaxy templates and provide separate templates for emission lines of Hα, Hβ, [O iii] λλ4959+5007, and [O ii] λ3727 individually. The non-negative normalizations of the galaxy templates and emission lines are computed simultaneously by the code, providing an implicit measurement of the line strengths and equivalent widths along with the redshift fit. Also by construction, the equivalent widths of the hydrogen Balmer lines are corrected for underlying absorption.

For all but the most extreme velocity-broadened lines, the line shape is determined by the object profile. For example, the spectral resolution of an object with (spatial) FWHM = 05 is approximately 180 Å, or 1500 km s⁻¹ for Hα at z = 1.3. Due to the low spectral resolution, Hα and N ii λλ6550+6584 are not resolved and the Hα line measurements represent the sum of these line species. The G141 spectral resolution tends to be just sufficient to produce an asymmetrical profile for the [O iii] λλ4959+5007 doublet, which can help in differentiating it from Hα assuming that the profile of the line-emitting region is roughly symmetric (see also Atek et al. 2010; Trump et al. 2011 for examples of Hα and [O iii] line profiles).

A comparison of redshifts determined from the grism+photometry fits to ground-based spectroscopic redshift measurements taken from the literature is shown in Figure 11. The precision of the grism redshifts is excellent (0.0035 × (1 + z) at z > 0.7) and is an order of magnitude better than is typically possible with high-quality broadband photometry alone (cf. 0.034 × (1 + z); Brammer et al. 2008). Furthermore, the redshift precision is nearly constant over the full redshift range of interest 1 < z < 3, whereas the precision of typical photometric redshifts is generally lower at these redshifts as relatively fewer observed photometric bands sample rest-frame optical wavelengths and the Balmer break (Brammer et al. 2011).

Zoom In Zoom Out Reset image size

The precision of the redshift measurements depends somewhat on the strength and availability of emission lines in the grism spectra. This is demonstrated in Figure 12, which shows the deviations of the grism redshifts from the ground-based spectroscopic measurements as a function of F140W magnitude and the S/N of the Hα emission line for objects at 0.8 < z < 1.4, where the line lies within the coverage of the G141 grism. While the redshift precision is nearly constant as a function of F140W magnitude, the redshift precision increases from only σ ∼ 0.0025 for galaxies with strong emission lines to σ ∼ 0.005 for galaxies without detected emission lines (S/N ≲ 5). The sample shown in Figure 12 is dominated by objects from the extensive catalog of Barger et al. (2008) in the GOODS-N field (see also Figure 13), which has fewer ancillary photometric bands than the other 3D-HST fields. Therefore, it is notable that the precision of the continuum-only grism redshifts is still significantly better than has been achieved for surveys using many intermediate-width photometric filters (e.g., Wolf et al. 2004; Ilbert et al. 2009; Whitaker et al. 2011).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Of the galaxies shown in the comparison of Figure 11, some 4% have redshift measurements that differ by more than 0.1 × (1 + z) from the spectroscopic value. The causes of these outliers include misidentification of single emission lines, poor background, and/or contamination subtraction of the G141 spectra and apparent errors of the spectroscopic redshifts themselves, where a clear emission line (or lines) is observed in the grism spectrum and the overall SED (spectrum+photometry) is well fit by the Eazy templates. In the current analysis, some significant systematic redshift errors remain at z < 0.7, where the WFC3/G141 grism does not sample strong spectral features. As we do not currently impose any constraints on the emission-line ratios, these errors are typically caused by the fitting algorithm incorrectly adopting emission lines with extreme equivalent widths to fit optical continuum features sampled by individual, non-overlapping broadband filters. These effects will be removed by future improvements to the fitting code and, in particular, by the incorporation of the ACS/G800L optical grism spectra.

For the z > 0.7 selection shown in Figure 11, 30% of the objects in the full 3D-HST sample have a measured ground-based spectroscopic redshift. However, this ratio varies greatly as a function of object magnitude and color among the survey fields depending on the selection criteria of the spectroscopic surveys that cover them. Figure 13 shows the distribution of galaxies with measured spectroscopic redshifts as a function of H (m₁₄₀) magnitude and observed (i − H) color. Spectroscopic surveys selected in the optical are most complete for bright, blue galaxies, typically at z ≲ 1. For galaxies with H < 21 and (i − H) < 2.2, 50% in the COSMOS and AEGIS fields have a measured spectroscopic redshift, and this fraction reaches 80% and nearly 100% for the deeper surveys of the GOODS-South and North fields. At fainter magnitudes, optically selected surveys only observe the bluest galaxies and do not sample the significant population of redder galaxies, typically at z > 1. Measuring spectroscopic redshifts for this population generally requires near-infrared spectra, which are challenging and time-consuming to obtain from the ground for large samples (e.g., Kriek et al. 2008). 3D-HST provides near-infrared spectra of all of the galaxies in Figure 11, which yield both crucial spectroscopic coverage of galaxies missing from typical redshift surveys and rest-frame optical spectra of galaxies at z ≳ 1 that may have been observed previously only at rest-frame UV wavelengths (see, e.g., van Dokkum et al. 2011).

As a blind survey of every object that falls within a given WFC3 (and/or ACS) grism pointing, 3D-HST produces high-quality spectra for a broad diversity of celestial objects. Some examples of the NIR grism spectra, chosen not as a representative sample but rather to demonstrate some of this diversity, are shown in Figure 14.

• 1.  
  Figures 14(a) and (b) show examples of line-emitting galaxies. The object in panel (a) has multiple components that are all strong [O iii] λλ4959+5007 emitters with large [O iii]/Hβ line ratios. The spatial extent of the [O iii] λλ4959+5007 is clearly visible in the 2D grism spectra and closely follows the spatial profile in the direct image. A typical magnitude-limited spectroscopic survey would likely only be able to place a slit on one of the two objects, and their physical association would be unknown. There are many such line-emitting pairs and groups within the 3D-HST survey that will be used to study the merger history of the universe out to z ∼ 2.Panel (b) shows a quasar in the COSMOS field at z = 4.656, identified from strong emission lines of Mg ii and C ii and a strong Lyman break in the ancillary optical photometry (not shown). The template fit of the rest-frame UV continuum is also in remarkable agreement with the G141 spectrum, showing numerous low-level absorption features. The galaxy is identified as an X-ray source by Elvis et al. (2009), but its distant redshift has not otherwise been identified; it has z_phot = 0.48 in the catalog from Ilbert et al. (2009) and is too faint (i = 22.6) to be included in the zCOSMOS-bright sample (Lilly et al. 2007). In this sense, among other things 3D-HST is a near-IR spectroscopic survey of X-ray (or radio, or 24 μm, etc.) selected sources within the survey fields.
• 2.  
  Figures 14(c) and (d) show NIR spectra of extremely massive (M > 10^11.2 M_☉) galaxies in the COSMOS and GOODS-S fields. The grism spectra covering the Balmer/4000 Å break, even with their low resolution, can greatly improve the constraints on the star formation histories of massive galaxies such as these (van Dokkum & Brammer 2010). The remarkable galaxy in panel (d) shows a strong 4000 Å break characteristic of a significantly evolved stellar population. The object in panel (d) is taken from the CANDELS grism pointing within the UDF (see Trump et al. 2011). The spectrum is somewhat deeper (t_exp = 15 ks) than typical 3D-HST G141 spectra but was fully reduced with the 3D-HST pipeline.
• 3.  
  The bottom two panels of Figure 14 show spectra of two more nearby objects, cool brown dwarf stars with spectral types T5/6 and L4. The spectral types are estimated from template fits from the SpeX spectral library (Burgasser et al. 2010), with the two best fits shown plotted along with the spectra. The T5/6 dwarf of panel (e) is the faintest (i.e., most distant) T dwarf by almost three magnitudes in H (F140W).¹⁸ The T5/6 dwarf has a photometric distance of 300–400 pc using the M_H relations tabulated by Vrba et al. (2004), with the broad range resulting from the crude estimate of its spectral type. Recently, three similarly faint, distant, cool brown dwarfs have also been found in the WISP grism survey by Masters et al. (2012).

Zoom In Zoom Out Reset image size

Admittedly, the objects shown in Figure 14 are somewhat brighter than typical galaxies found in the full 3D-HST survey. We provide additional example spectra in Appendix C down to the effective sensitivity limits of the survey. Continuum breaks can indeed be detected down to the continuum limit (H₁₄₀ ∼ 23), though the redshift precision of such faint, line-free, continuum objects will likely be intermediate between that shown in Figure 12 and that of medium-band photometric surveys such as the NEWFIRM Medium Band Survey (see Figure 12). Robust 5σ emission-line detections are found at considerably fainter magnitudes at line fluxes consistent with the limits described in Section 3.4. For more examples of 3D-HST grism spectra, see van Dokkum et al. (2011) for the 1D spectra of a complete, physically defined sample of massive galaxies at 1 < z < 1.5, and a sample of Hα emission-line morphologies is presented by Nelson et al. (2012).

In the low-redshift universe many galaxies are observed to be quiescent, with current SFRs only ∼1% of their past average (e.g., Pasquali et al. 2006). These quiescent galaxies tend to be massive early-type galaxies, forming the "red sequence" in the color–mass distribution of galaxies. Recent work has shown that at z ∼ 2 many massive galaxies (M ≳ 10¹¹ M_☉) exhibit spectacularly high SFRs of hundreds of solar masses per year, whereas others were already quiescent, particularly those that are extremely compact for their mass (Kriek et al. 2006; van Dokkum et al. 2008; Brammer et al. 2009). Active galactic nucleus (AGN) feedback is a possible mechanism to suppress gas cooling and star formation (e.g., Croton et al. 2006), but direct evidence is scarce.

Diagnostics of quiescence can be correlated with stellar mass, surface density (i.e., compactness), and the environment of galaxies on Mpc scales. These diagnostics are most reliably identified spectroscopically, through the strength of the Balmer or 4000 Å break (D₄₀₀₀; see Figures 14(c) and (d)) and/or the absence of emission lines such as Hα. For a 5σ limiting emission-line flux of 5 × 10⁻¹⁷ erg s⁻¹ cm⁻² (Section 3.4), 3D-HST will reach Hα and [O ii] λλ3727 SFR limits (Kennicutt 1998) of 2.5 and 25 M_☉ yr⁻¹ at z = 1 and z = 2, respectively. If the simultaneous presence of quiescent and starbursting massive galaxies at z ∼ 2 is the result of their surface density or their environment, correlations should exist between the SFR and these parameters.

The morphology–density relation (Dressler 1980) states that "early-type" galaxies (mostly massive quiescent galaxies) are relatively abundant in dense environments such as groups and clusters. However, at low redshift most massive galaxies are quiescent regardless of environment (Kauffmann et al. 2003; Balogh et al. 2004), so it is difficult to determine whether the environment provides a physical mechanism that alters the galaxies (e.g., through gas stripping), or whether dense environments simply are the place where quiescent galaxies tend to end up. There remains some tension in the recent literature whether galaxy-mass-driven effects (Peng et al. 2010) or the properties of their host dark matter halos (Wake et al. 2012) are the dominant factors shaping the star formation histories of galaxies.

To further disentangle the roles of mass and environment, epochs should be considered when "massive" did not yet directly imply "quiescent" (van Dokkum et al. 2011). The 3D-HST sample will be sufficiently large to determine the relation between SFR and environment in bins of fixed mass and redshift. If no environmental dependence is observed at fixed mass, then the relations between galaxy properties and environment are simply a by-product of the underlying relations of both quantities with mass. Even excellent photometric redshifts with errors δz ≈ 0.04(1 + z) have a radial error on the comoving distance of >150 Mpc at z = 2, which is larger than the distance of the Milky Way to the Coma cluster. With redshift errors more than an order of magnitude smaller (Figure 11), 3D-HST will be able to provide a sensible definition of the environmental galaxy density on the scale of a few comoving Mpc, as well as spectroscopic diagnostics of galaxy "quiescence."

The epoch 1 < z < 3 saw the wholesale transition from small, star-forming clumps evident in the deepest HST images to the ordered "realm of galaxies" seen today (e.g., Elmegreen et al. 2007; Wuyts et al. 2012). Since z ∼ 1, most star formation has taken place in large spiral disks, but different modes may have been prevalent at earlier times, such as disks made of star bursting clumps that coalesce to form compact spheroids directly (Dekel et al. 2009).

If bulges formed before disks, then objects with stellar masses (1–5) × 10¹⁰ M_☉ and star formation concentrated on 1 kpc scales should be seen at z ∼ 3. These should evolve into young compact bulges surrounded by disk-like (∼5 kpc) star formation at z ∼ 2, and then to old bulges and regular disks by z ∼ 1. By contrast, if bulges formed mostly from subsequent merging of disks, extended star-forming disks should be pervasive at z ∼ 3. The relative ages of the subcomponents of galaxies and the spatial extent of line-emitting star-forming regions can be measured from the spatially resolved HST grism spectra (van Dokkum & Brammer 2010). Ground-based integral-field spectroscopy has demonstrated the power of spatially resolved spectroscopy for galaxies at z > 1 (e.g., Förster Schreiber et al. 2009; Law et al. 2009), but it has been typically limited to more luminous, rare objects or relatively small samples.

Although the merger-driven growth of massive galaxies is a common prediction of galaxy formation models, it has been difficult to test at higher redshift where mergers should be most common (e.g., Guo & White 2008). The merger rate can be determined from physical pair statistics, but these have been difficult to measure due to insufficient spatial resolution and contamination by chance superpositions of unassociated galaxies (e.g., Williams et al. 2011; Man et al. 2012).

With its spatial resolution of 013 and spectral resolution of δv ≈ 1000 km s⁻¹ (Figure 11), 3D-HST can spectroscopically identify true physical pairs and groups down to separations ≲ 5 kpc, weeding out projected galaxy pairs (see also Figure 14(a)). Within the 3D-HST sample, the pair fraction can be measured as a function of mass and redshift, and the fractions can be turned into a merger rate using models (e.g., Kitzbichler & White 2008; Lotz et al. 2011; Williams et al. 2011). The mass growth of galaxies due to mergers can be compared to the growth due to star formation, and the sizes, densities, and AGN content of the merger components can be determined. The SFRs of the spectroscopically identified merging pairs can be compared to predictions of hydrodynamical simulations, such as that mergers should play a large role in driving star formation activity and black hole accretion (e.g., Cox et al. 2006).