A Catalog of Emission-line Galaxies from the Faint Infrared Grism Survey: Studying Environmental Influence on Star Formation

FIGS (HST Cycle 22, PID:13779, PI S. Malhotra) used the HST WFC3 G102 infrared grism (see Figure 1) to obtain deep slitless spectroscopy of ∼6000 galaxies. FIGS achieved 40-orbit depth in four fields within the greater GOODS-North and GOODS-South fields (Giavalisco et al. 2004), designated GN1, GN2, GS1 (HUDF), and GS2 (HDF-PAR2). Objects in each field were observed in five different eight-orbit PAs in order to mitigate the contamination of spectra from overlapping spectra from nearby objects. The grism image at each PA covers a 205 × 227 field of view, and the G102 grism has a resolution of 24.5 Å pixel⁻¹. The total area of coverage over all fields is 17.7 arcmin².

Zoom In Zoom Out Reset image size

In this paper, we used 1D spectra of individual PAs that were generated using the methods described in Pirzkal et al. (2017). Here we briefly summarize this process. We reduced FIGS data in a manner that loosely follows the method used for GRAPES and PEARS, the previous deep HST grism surveys (Pirzkal et al. 2004, 2013; Malhotra et al. 2005; Xu et al. 2007; Rhoads et al. 2009; Straughn et al. 2009; Xia et al. 2012). First, we generated a master catalog of sources from deep CANDELS survey mosaics in the F850LP filter in ACS and the F125W and F160W filters in WFC3 (approximately the z, J, and H bands; Grogin et al. 2011; Koekemoer et al. 2011). We astrometrically corrected the data to match the absolute astrometry of the GOODS V2.0 catalogs. We estimated the background levels of the grism observations by using a two-component model, which included a constant zodiacal light background, as well as a varying He i light background. To generate the individual spectra, we used a simulation-based extraction approach that accounts for spectral contamination from overlapping spectra, as well as allowing the use of an optimal extraction approach (Horne 1986) when generating 1D spectra from 2D spectra. We refer the reader to Pirzkal et al. (2017) for a complete description of these processes. Pirzkal et al. (2017) also describe a method for combining individual PAs into a single 1D spectrum, though we do not use the combined spectra here. Individual PAs are subject to different levels of contamination, and depending on the location of the emission-line region, they may not all exhibit the same line emission. For these reasons, we use the individual PA spectra for our ELG search and analysis.

We initially extracted all sources down to a continuum level of F105W < 30 mag. When the extractions were complete, we had an average of ∼1700 spectra per field, with a typical 3σ continuum detection limit of m_F105W = 26 mag and an emission-line sensitivity of 10⁻¹⁷ erg cm⁻² s⁻¹. The middle panel of Figure 1 shows the throughput curve for the G102 grism compared to other spectral and broadband curves. We restricted use of the G102 spectra to wavelengths between 8500 and 11500 Å, where the grism throughput is greater than 20%.

2.3.1. MUSE/VLT

2.3.2. G800 Grism Data

2.3.3. G141 Grism Data

2.3.4. Spectroscopic Redshifts

We conducted a blind search for ELGs among the FIGS 1D spectra. Because we obtained our infrared spectra via slitless grism spectroscopy, there was no preselection of ELG candidates before the search via the placement of slits or by broadband magnitude cutoffs beyond the survey depth. This had the advantage of enabling the detection of ELGs with potentially very low continuum levels and so might allow for the identification and study of smaller and/or fainter galaxies with nebular line emission. However, this did require an efficient method for selecting ELG candidates from the total sample of FIGS spectra. In order to search the ∼6000 FIGS spectra for emission lines, we developed a code to automatically search for and identify significant peaks in a 1D spectrum.

First, the level of the continuum flux had to be estimated at each wavelength element in the 1D spectrum. To measure this, we used a median-flux filter, which assumes a prospective line width and calculates the local continuum from the median flux outside that line width, in wavelength regions on either side of the line. A given wavelength λ₀ is taken to be the center of a potential line. The potential line flux is measured as all the flux contained within a line width 2Δλ₁, so that the algorithm defines the potential line as the region covered by

$$\begin{eqnarray}&&{\lambda }_{0}-{\rm{\Delta }}{\lambda }_{1}\lt \lambda \lt {\lambda }_{0}+{\rm{\Delta }}{\lambda }_{1}.\end{eqnarray} \tag{ 1 }$$

Then, the code estimates the nearby continuum by looking at regions to either side of the line with width Δλ₂. The nearby continuum is defined then as the regions contained in

$$\begin{eqnarray}&&\begin{array}{l}({\lambda }_{0}-{\rm{\Delta }}{\lambda }_{1})-{\rm{\Delta }}{\lambda }_{2}\lt \lambda \lt ({\lambda }_{0}-{\rm{\Delta }}{\lambda }_{1})\ \ \mathrm{and}\ \ \\ \qquad ({\lambda }_{0}+{\rm{\Delta }}{\lambda }_{1})\lt \lambda \lt (\lambda +{\rm{\Delta }}{\lambda }_{1})+{\rm{\Delta }}{\lambda }_{2}.\end{array}\end{eqnarray} \tag{ 2 }$$

The algorithm then takes the median flux of the wavelength pixels constrained by Equation (2) as an estimate of the local continuum around the hypothetical line and subtracts this flux from the flux at λ₀ in order to obtain the continuum-subtracted or residual flux at that point. If there is a line present at λ₀, this method allows for measurement of the level of the continuum without influence from the line flux. The code takes the standard deviation among this set of continuum fluxes as an estimate of the flux error at λ₀. If λ₀ is too near the edge of the spectrum to measure Δλ₂ on both sides, we estimate the continuum based on the fluxes from just the complete side. The algorithm repeats this process, iterating over each wavelength element in a given spectrum, estimating the continuum flux at that wavelength, and subtracting it. We refer to Figure 2 for an example continuum-subtracted spectrum. We were able to best minimize false detections while retaining real ones with 2Δλ₁ = 122.5 Å and Δλ₂ = 147 Å, based on tests of variable Δλ₁ and Δλ₂ with a preliminary subsample of ELGs with matching spectroscopic redshifts.

Zoom In Zoom Out Reset image size

After the spectrum is continuum-subtracted, the code calculates the signal-to-noise ratio (S/N) at each wavelength with the residual flux and the flux error, once more iterating through the list of wavelength elements. The sum of the fluxes constrained by Equation (1) determines the hypothetical line signal, and the estimated flux errors added in quadrature measure the noise of the hypothetical line. After this calculation is complete for all wavelengths, the algorithm identifies the location with maximum line S/N in the spectrum. If S/N > 5, we fit a Gaussian at the central wavelength element from which we obtain a measure of the continuum-subtracted integrated line flux. The code then subtracts the fit line from the residual flux spectrum and checks the next-highest S/N. If the S/N still exceeds 5, the routine repeats until the peak S/N is below the detection threshold.

We run this routine on the individual PA spectra in each field and record all instances of S/N > 5. If the code finds a peak in at least two PAs with centroids at the same or adjacent wavelength elements (24.5 Å in either direction), it declares a detection. Lower S/N thresholds produced numerous false positives, so we used the S/N > 5 cutoff to maintain a more robust sample. We avoided using simultaneous fits of all PAs in order to avoid including contaminated PAs in a combined significance measurement. With individual PA fits, contaminated detections could more easily be identified and removed. After running the routine over all galaxies in the field, the list of detections forms an ELG candidate list.

Once we had obtained lists of candidate detections for each field, we next attempted to identify the type of emission line detected in each spectrum. First, we matched the candidate lists to our photometric redshift (photo-z) catalog (Pharo et al. 2018) and assigned a preliminary line ID based on the likely redshift. For the purposes of this result, we focused on three strong line IDs that could be robustly detected at FIGS resolution and sensitivity: Hα λ6563, [O iii] λλ4959, 5007, and [O ii] λ3727. We did this because these lines typically have the strongest emission, and therefore can be detected robustly, and because they are common features of star-forming galaxies. Hβ λ4861 could theoretically be resolved and detected alongside [O iii], but it was typically faint enough that it was difficult to detect at a significant level. Other FIGS studies have looked at Lyα line emission at higher redshifts (Tilvi et al. 2016; Larson et al. 2018). We measure detections of GS2-1406 at levels of significance ≥3.5σ, consistent with those measured by Larson et al. (2018), though these are lower than the 5σ cut used in this work. We detect GN1-1292 significantly in only one of the two PAs where it is measured in Tilvi et al. (2016). Pirzkal et al. (2018) also detect GS2-1406 but not GN1-1292.

After the preliminary photo-z identification, we sought to confirm the existence and type of the line by checking the detection against ancillary data. The most straightforward way to do this was to check for other emission lines. Since the wavelength ratio between a given pair of emission lines is invariant across redshift, the detection of two strong lines is a useful check. For eight candidates, two strong lines were measured in the FIGS G102 spectra alone, and for 59 others we identified pairs by checking matched ACS/G800L, MUSE/VLT, and WFC/G141 spectra (described in Section 2.3). This most commonly involved finding Hα–[O iii] and [O iii]–[O ii] pairs. Occasionally, we were able to identify another spectral feature, such as a strong 4000 Å break, in order to confirm the redshift. We note that while finding a matching line can confirm a line detection, not finding a matching line does not necessarily mean that the detection is false, since the true relative line strengths are not known ahead of time. The matching line may be sufficiently weaker than the FIGS line, or the matching spectra sufficiently shallow, that the matching line is not detected.

If matching spectra were not available, or a strong line was not identified, we next checked for a matching spectroscopic redshift (spec-z). If a spec-z assigned the line peak wavelength a rest-frame wavelength that matched an emission line within the wavelength range of an FIGS grism element, we assigned the line the spec-z ID. If neither matching lines nor spec-z IDs were available, then we let the photo-z identification stand.

In each field, there were a handful of objects with a significant detection but no good redshift measurement. These were almost all very continuum-faint (F105W > 27.5 mag) objects, which both reduced the availability of broadband fluxes to use for photometric redshift fits and made the spectra more susceptible to contamination from nearby sources. Consequently, most of these candidates were removed through visual inspection, leaving four likely ELGs with no redshift: GS1-1062, GS2-532, GS2-838, and GS2-1624.

With the lines identified, we compared our results to the line list derived from Pirzkal et al. (2018), a study of FIGS strong line emitters using a distinct identification method with FIGS 2D spectra. In this paper, we do an independent selection and measurement of ELGs so as not to bias the findings of different search methods. However, we have compared our line candidates with those found in Pirzkal et al. (2018) and find them to be in close agreement, with a 90% match in identifications. A complete match would have been highly unlikely, as the methods have different strengths. The 2D method performs better at identifying broader lines that are wider than the median filter used with the 1D search. As described in Section 3.1, we experimented with different median filter widths when designing the 1D search and found that expanding the filter too broadly tended to introduce false detections and make it more difficult to detect fainter objects, since the detection becomes more susceptible to changes in the continuum. The 2D method, however, requires detections in at least three PAs. For a given FIGS field, the different roll angles do not overlap completely, so near the edges of the field there are regions without the full five-PA coverage. In these regions, it is easier to get a detection with the 1D method, as it requires detections in only two PAs.

In Pharo et al. (2018) we used redshift catalogs derived from combined FIGS grism spectroscopy and broadband photometry to search for significant overdensities of galaxies in the FIGS fields. First, we divided each field into slices of redshift with Δz = 0.03(1 + z_min), where z_min is the lower bound of the redshift slice. This Δz is based on the limiting accuracy of the Pharo et al. (2018) photometric redshift catalogs, which enabled the most significant detections of physically associated systems. In each redshift slice, we conducted a seventh-nearest-neighbor density search for a grid of points in the field. This is defined as

$$\begin{eqnarray}&&{n}_{7}=\displaystyle \frac{N}{\pi {R}_{7}^{2}},\end{eqnarray} \tag{ 3 }$$

where N = 7 and R₇ is the angular distance to the seventh-nearest galaxy in that redshift slice.

We then checked for points of significant overdensity with two different metrics. First, we calculated ${ \mathcal M }$, the largest value of n₇ in the slice normalized to the slice's median n₇. We also calculated ${ \mathcal S }$, the peak nearest-neighbor density in the redshift slice divided by the standard deviation of densities in the adjacent redshift slices. We counted peaks with ${ \mathcal M }=10$ or ${ \mathcal S }=10$ as significant detections, based on comparisons with other nearest-neighbor density searches (Spitler et al. 2012) and the values for spectroscopically identified clusters (see Pharo et al. 2018 for more detail). Across the four FIGS fields, we identified 24 such overdensities, as well as determining the R₇ values for individual FIGS galaxies. We make use of both the proximity to a detected overdensity and the R₇ distance of a galaxy to study environmental effects in the subsequent sections.

We also used the redshifts and angular separations of galaxies to determine the physical local surface density Σ for FIGS ELGs. For the purposes of this discussion, we will use terms such that field galaxies have Σ < 6 Mpc⁻², rich fields have 6 < Σ < 10, groups have 10 < Σ < 30, and rich groups have Σ > 30. These definitions are adapted from those used in Sobral et al. (2011), where they were based on derivations from correlation length studies of galaxies in Mo et al. (1996) and Yang et al. (2005). We measure ELGs in the field, rich field, and group density ranges, but not in the range of rich groups or above.

Figures 8 and 9 show the nearest-neighbor density plots of significant overdensities in the FIGS fields. The figures also show the locations of identified ELGs in each redshift slice, with triangles representing Hα emitters, circles for [O iii] emitters, and stars for [O ii] emitters. There appear to be several overdensities with associated ELGs, but this does not appear to be a consistent trend visually. The spectroscopically identified clusters at z = 0.85 and z = 1.84 have two and three ELGs in the same redshift slice, respectively. All but one of these ELGs appear near the overdensities, but not especially near the density peaks. Neither cluster appears to show an excess of identified ELGs compared to other slices. For the z = 1.84 cluster, this may seem in contradiction to the results of Tran et al. (2010), which found increased star formation activity near a cluster core at similar redshift. It is possible that this is because the relatively faint [O ii] is simply less complete compared to the lower-redshift sources, or because we do not find many galaxies with surface densities of Σ ≥ 20 Mpc⁻², the density at which Tran et al. (2010) begin to find the increased fraction of star-forming galaxies. We discuss the SFRs themselves and their implications in Sections 5.3–5.5.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

In order to systematically study a possible relationship between strong line emission and galaxy environment, we first looked at the R₇ distance for both ELGs and regular galaxies. If ELGs have a preferred relationship with overdensities, then the distribution of R₇ distances could be distinct from nonemitting galaxies, since, for example, a preference for ELGs to be close to overdensities should result in a distribution that peaks more at low R₇.

Figure 10 shows the probability density distributions of R₇ distances for ELGs compared to the whole set of galaxies in our redshift catalog. The distributions are broken down into six subsamples, in order to make meaningful comparisons of distance and stellar mass: first, by redshift ranges corresponding to the three strong emitters, and then by bright and faint F105W continuum magnitudes as a proxy for mass. To judge the significance of the differences between a given pair of distributions, we applied a two-sample Kolmogorov–Smirnov (K-S) test, a statistical test to determine whether two underlying one-dimensional probability distributions differ, to each subsample pair. The test produces a p-value determined by the sizes and differences of the two distributions, and this p-value gives the level of significance at which the two may be considered distinct. A lower p-value corresponds to a more significant determination that the two distributions are different.

Zoom In Zoom Out Reset image size

This test showed no significant difference in the R₇ distribution of Hα emitters compared to other galaxies in either the bright or faint bins. For [O iii] λλ4959, 5007 emitters, the test does find a significant difference in distributions for both the bright (p = 3 × 10⁻⁷) and faint (p = 0.02) bins, with [O iii] λλ4959, 5007 emitters having a higher probability of appearing at midrange R₇ distances compared to other galaxies. For [O ii] λ3727 emitters, the test finds a significant difference only in the bright bin (p = 0.002). This measurement for [O iii] supports previous studies that find line emitters preferentially at intermediate distances around clusters (Darvish et al. 2014) at z ≃ 1. We explore this result in more detail in Section 5.5.

Studying the R₇ distribution by itself gives insight into only the relationship between the locations of ELGs and those of overdensities, while ignoring the other properties of the ELGs. With the flux catalog, we were also able to investigate how an ELG's environment might influence its emission-line luminosity and recent SFR.

To account for the effects of dust extinction in measuring the SFR, we used a dust calibration developed by Sobral et al. (2012) using rest-frame u–z colors. The calibration was developed and tested using Hα and [O ii] emitters at z = 0.1 and z = 1.47. It is given by

$$\begin{eqnarray}\begin{array}{rcl}{A}_{{\rm{H}}\alpha } & = & -0.092{\left(u-z\right)}^{3}+0.671{\left(u-z\right)}^{2}\\ & & -\ 0.952(u-z)+0.875.\end{array}\end{eqnarray} \tag{ 4 }$$

Sobral et al. (2012) find that this relation holds across redshift epochs, covering most of the redshift range of our sample and for both kinds of emitters. To convert the A_Hα calculation to A_{[O iii]} and A_{[O ii]}, we applied the Calzetti et al. (2000) reddening law. For the few objects for which one of rest-frame u or z was unavailable, we assigned the median reddening value from the rest of the sample. We also measured A_Hα using the stellar mass dust parameterization developed in Garn & Best (2010) (see Section 5.4 for discussion of stellar masses). This typically produced similar dust measures to the color calibration, with a divergence at masses greater than $10.5\mathrm{log}({M}_{\odot })$, as was also observed in Ramraj et al. (2017). Such objects make up a very small fraction of our ELG sample, so this does not change any overall trends.

We calculated SFRs for the ELGs using the following equations:

$$\begin{eqnarray}&&{\mathrm{SFR}}_{{\rm{H}}\alpha }({M}_{\odot }\,{\mathrm{yr}}^{-1})=7.9\times {10}^{-42}L({\rm{H}}\alpha )\ (\mathrm{erg}\ {{\rm{s}}}^{-1})\end{eqnarray} \tag{ 5 }$$

$$\begin{eqnarray}&&{\mathrm{SFR}}_{[{\rm{O}}{\rm{II}}]}({M}_{\odot }\,{\mathrm{yr}}^{-1})=1.4\pm 0.4\times {10}^{-41}L([{\rm{O}}\,{\rm{II}}])\ (\mathrm{erg}\ {{\rm{s}}}^{-1})\end{eqnarray} \tag{ 6 }$$

$$\begin{eqnarray}&&{\mathrm{SFR}}_{[{\rm{O}}{\rm{III}}]}({M}_{\odot }\,{\mathrm{yr}}^{-1})=6.4\pm 4.0\times {10}^{-42}L([{\rm{O}}\,{\rm{III}}])\ (\mathrm{erg}\ {{\rm{s}}}^{-1}).\end{eqnarray} \tag{ 7 }$$

Equations (5) and (6) are calibrations from Kennicutt (1998), derived with a Salpeter IMF. Equation (7) was derived by Straughn et al. (2009) using [O iii]/Hα ratios from star-forming galaxy knots where both emission lines were present.

We obtained stellar masses for our ELG sample by applying our EAZY SED catalogs to the SED fitting code FAST (Kriek et al. 2009), using a Chabrier (2003) initial mass function, Bruzual & Charlot (2003) templates, and an exponential star formation history. We used spec-zs for the fits where high-quality matching redshifts were available in our compilation, and we used the best-fit photometric redshift otherwise. We checked the results for the GS1 field against the GOODS-South catalogs compiled by Santini et al. (2015), which largely exclude our GS2 parallel field. For the galaxies with existing measurements, we found our mass results consistent with the Santini et al. (2015) catalogs, with a median difference of less than 0.1 dex. For the other fields, we checked against matching masses from Skelton et al. (2014) and found a similar level of agreement.

With the stellar masses calculated, we were able to determine each ELG's SFR per stellar mass, or specific SFR (sSFR). The relation between sSFR and mass for star-forming galaxies is typically called the galaxy main sequence, and it suggests an evolution of star formation with redshift and stellar mass (Noeske et al. 2007). In star-forming galaxies, as redshift decreases, ongoing star formation builds up increased stellar mass, and as this happens, sSFRs decline as the galaxies exhaust their supplies of gas.

This smooth relation does not account for cases of rapid quenching and does not address the influence of environment on how galaxies evolve. In Figure 11, we show the sSFR as a function of the stellar mass and compare our ELGs to the results from Noeske et al. (2007), who measured this in four redshift bins up to z = 1.1. We show the best-fit staged-tau models of star formation history from Noeske et al. (2007) in the redshift bins most closely matching our Hα and [O iii] emitters (the [O ii] sample is at too high a redshift) and find that our ELGs typically have higher sSFR for a given stellar mass.

Zoom In Zoom Out Reset image size

This may be due to the luminosity limit of our sample. Since the limiting luminosity is redshift dependent, we determined the lower-bound sSFR based on the lower and upper redshift limits of each ELG sample. We show these lines in Figure 11. For [O iii and [O ii], the limiting sSFRs are relatively high, indicating that we are likely sampling the upper region of the galaxy main sequence.

We were still able to investigate the possible environmental effects on the main sequence, especially for Hα, where we probe the main sequence most closely. In Figure 11, we also show each galaxy's local surface density Σ. There is no significant relationship between either Σ and the stellar mass or Σ and position on the sSFR–mass relation to the densities probed, which at most get only as dense as galaxy groups (10 < Σ < 30, as defined in Section 5.1.). The same holds for the ELGs' R₇ measurements. This suggests that environmental effects do not play a systematic role in either quenching or triggering star formation, since they do not appear to disrupt the smooth star formation relation of the main sequence.

We also studied the relationship between emission-line strength and galaxy clustering more directly. First, we used the galaxies' redshifts and angular separations to compute the local physical surface density Σ in units of Mpc⁻² for each ELG. Then, we split the sample of ELGs into two subsamples: those located in a redshift slice where a significant overdensity is detected ("In OD"), and those in a redshift slice with no overdensity detection ("No OD"). By looking at these subsamples, we could check whether ELG properties differ for galaxies near a range of peak densities. This also gives us a subsample to compare directly to studies focused only on known clusters. This result can be seen in Figure 12, which shows the line luminosity as a function of Σ for Hα, [O iii], and [O ii] emitters. In each panel, the horizontal dashed lines give the median line luminosity for each subsample, and the vertical dashed lines give the median Σ. We inspected the SFR–Σ and sSFR–Σ relations as well (see Figure 14), and the distributions remained essentially unchanged for each emitter. Thus, for this discussion we will refer simply to the L–Σ relation shown in Figure 12, as that requires the fewest additional assumptions.

Zoom In Zoom Out Reset image size

The Hα L–Σ distribution (top panel) shows little indication of a preferred relationship with density. Even in the OD slices, Hα emitters are found at a range of local densities and with a range of luminosity values. The median luminosity for emitters near overdensities is 0.3 dex higher than for those in non-OD slices, which is about twice the typical error size for the Hα emitters. Figure 13 shows the same L–Σ relation with the points shaded by sSFR and stellar mass, but no clear trends with density emerge.

Zoom In Zoom Out Reset image size

In Figure 14, we compared the derived Hα SFRs from this luminosity sample to a narrowband-selected sample of star-forming (SFR > 3 M_⊙ yr⁻¹) Hα emitters at z = 0.845 (Sobral et al. 2011), using the same SFR diagnostic. Our distribution differs from the result in Sobral et al. (2011), which shows SFR increasing with density up to Σ ≃ 50 Mpc⁻². The discrepancy may be at least partially explained by their selection of SFR > 3 M_⊙ yr⁻¹ emitters, which would exclude much of our sample. SFR measurements of FIGS ELGs go down to 0.1 M_⊙ yr⁻¹, but our sample probes to lower stellar mass, yielding comparable sSFRs. Nevertheless, this does limit the utility of a comparison with the Sobral et al. (2011) results. Darvish et al. (2014) conducted a similar study down to a limit of SFR > 1.5 M_⊙ yr⁻¹ and find only a small difference in median SFR between field and cluster galaxies. However, Darvish et al. (2014) also find that at intermediate densities, comparable to the rich field or group densities described earlier, a higher fraction of galaxies exhibit star formation compared to fields and rich clusters.

Zoom In Zoom Out Reset image size

To get a broader sense of environmental effects at different redshifts, we need to look at non-narrowband studies. Grützbauch et al. (2011a) found a weak correlation between overdensity and U–B color, used as a proxy for star formation, using data from the POWIR and DEEP2 surveys. Their study showed declining star formation with increasing density at 0.4 ≤ z ≤ 0.7, roughly matching the redshift range of our Hα emitters, which becomes a flat relation for 0.85 ≤ z ≤ 1. In their paper, they do not attempt to convert the color to SFR or sSFR, but in Grützbauch et al. (2011b) the change in observed color is shown to correlate with a half-dex decrease in sSFR at 1.5 < z < 2. Grützbauch et al. (2011a) also measure these effects in terms of overdensity, not physical density, so a direct comparison with our results and that of Sobral et al. (2011) is not possible. However, their largest overdensity bin corresponds to the value of our overdensity cutoff; in this case, one might expect to see a decline in Hα line luminosity among those galaxies marked "In OD," but no such decline is apparent. The result of Grützbauch et al. (2011a) was limited to stellar masses down to only $\mathrm{log}({M}_{\star }/{M}_{\odot })\gt 10.25$, so they may not have detected the low-luminosity scatter we observe at low Σ, while observing a decline in high SFR as Σ increases.

Scoville et al. (2013) studied near-UV-continuum-derived SFRs versus density in redshift slices up to z < 3 in COSMOS, finding a flat SFR–Σ relationship for 0.8 < z < 2. Scoville et al. (2013) studied the SFR–Σ relation in terms of density percentiles, of which the highest encompassed Σ > 10 Mpc⁻² and the lowest Σ < 0.1 Mpc⁻². At z < 1, they measure a flat relationship up to Σ of a few for the lower- and medium-density percentiles, after which the SFR declines with increasing density in the highest percentiles. At 0.35 < z < 0.6, the redshift bin most comparable to our Hα sample, the decline is notable in only the highest-density bin, where the median SFR drops from 1.5 to 0.3 M_⊙. We do not observe this drop in the line luminosity (or in dust-corrected SFRs), but the high-density percentile from Scoville et al. (2013) includes densities up to ∼100 Mpc⁻², and perhaps the drop in SFR is concentrated in these very high density sources. If so, it could be the case that we do not detect ELGs at such high density because their star formation has dropped too low, leading to fainter (or no) line emission. The trends from Scoville et al. (2013), combined with the results of Darvish et al. (2014), could suggest a transition period from peak star formation and merger interactions at z ≃ 2 (Madau et al. 1998) to the local universe. After the merger peak, higher-density environments may have already quenched star formation in local galaxies through strangulation or ram pressure stripping (Muzzin et al. 2012, 2014), depleting their reserves of star-forming material and increasingly relegating star formation to intermediate and field densities. This could be the case with our Hα sample, if the lack of high-Σ emitters is due to the depletion of star formation at high density.

The [O iii] λλ4959, 5007 L–Σ distribution (middle panel) shows an essentially flat relationship, with high scatter at the lowest densities. In slices with an overdensity, the emitters we find are much more likely to occupy densities in the range 5 < Σ < 15, with 12 out of 20 emitters in OD slices found in this region. The median Σ is a factor of a few higher for [O iii] emitters in overdensity slices compared to those not near overdensities. This is distinct from the Hα emitters, which do not seem to have a distinct density preference when near overdensities.

The density where the [O iii] emitters are preferentially located corresponds to the rich fields and galaxy groups at the outskirts of a denser cluster, corroborating what we see in Figure 10. This result also matches the findings in Grützbauch et al. (2011a) and Scoville et al. (2013), as we find a flat SFR–density relationship at 0.8 < z < 1.3. We do not see the higher SFR at intermediate density that Darvish et al. (2014) measured, but our results do corroborate their finding of a higher fraction of star-forming galaxies at those densities. Compared to the Hα distribution, which shows emitters at all density ranges near overdensities, this could suggest an evolution with redshift in the preferred locations of group star-forming galaxies near overdensities, which are found nearer to rich group densities among the FIGS Hα emitters.

The [O ii] λ3727 SFR–Σ distribution (bottom panel) shows no relationship, with all the emitters found at low Σ. This is likely due in part to the limits of our overdensity search near z ≃ 2, where our sample of fainter galaxies is less complete (see Section 5.4 for further discussion). One can see this effect in the range of SFRs calculated for the [O ii] emitters, which is restricted to much higher star formation than the other two samples. Of the 24 overdensity candidates, only 4 are in the redshift range where we might find [O ii], and these are of lesser significance and based on fewer galaxies compared to the overdensity candidates at lower redshift. These caveats aside, this could suggest that at higher redshift field galaxies exhibit higher star formation, at least among the brightest galaxies. Patel et al. (2009) found in a study of z ≃ 0.8 galaxies with $\mathrm{log}({M}_{\star }/{M}_{\odot })\gt 10$ that specific star formation (SFR per stellar mass) declined with increasing density. Since the limits of our completeness at this redshift select more massive galaxies, this could indeed explain our findings.

We also investigated the behavior of ELGs with close companion galaxies, in order to study overdensities and environmental effects on a smaller scale. If nebular emission and related star formation are triggered by interactions with companion galaxies (Kennicutt et al. 1987; Alonso et al. 2004), then we could observe a difference in the number of nearby galaxies between ELGs and galaxies. Ellison et al. (2010) studied the effects on environment of interacting galaxy pairs selected from Sloan Digital Sky Survey DR4, finding a small increase in sSFR for the closest pairs at low Σ relative to both more distant pairs and pairs found at higher galaxy densities. Their distance criterion for identifying a pair required a projected distance of ${R}_{p}\lt 80\times {h}_{70}^{-1}$ kpc between the two galaxies. Using this projected distance as the range for possible companions, we find that the fractions of ELGs (31%) and non-ELGs (32%) that form a near pair are essentially the same. Kocevski et al. (2012) used a much narrower allowable pair range (12 kpc) to search for interactions near active galactic nucleus hosts. Applying this much stricter cut yields a 3% pair rate in both ELGs and non-ELGs, suggesting that line emission and star formation are not necessarily directly connected to the presence of a nearby companion.