STAR FORMATION AT 4 < z < 6 FROM THE SPITZER LARGE AREA SURVEY WITH HYPER-SUPRIME-CAM (SPLASH)

Studies of star-forming galaxies at a wide range of cosmic epochs have revealed a strong correlation at fixed redshift between star formation rate (SFR) and stellar mass (M_*), of the form

$$\begin{equation} \log {\rm SFR} (\,M_\odot\, \hbox{yr}^{-1}) = \alpha \times (\log M_*/M_\odot - 10.5) + \beta, \end{equation} \tag{ 1 }$$

with α and β likely time-dependent. This relationship has been shown to hold over 4–5 orders of magnitude in mass (Santini et al. 2009) and from z = 0 to z ∼ 6 (Noeske et al. 2007; Daddi et al. 2007; Karim et al. 2011; Kashino et al. 2013; Speagle et al. 2014). It is a tight correlation, with only ∼.25–.35 dex of scatter at any redshift (Daddi et al. 2007; Whitaker et al. 2012).

This Letter extends the star-forming main sequence (SFMS) to massive galaxies at higher redshifts than previous studies by using the first data from the Spitzer Large Area Survey with Hyper-Suprime-Cam (SPLASH; Takada 2010). This survey is obtaining 2475h (>6h/pointing) of Spitzer IRAC 3.6 μm and 4.5 μm observations on the two Hyper-Suprime-Cam ultra-deep fields COSMOS (Scoville et al. 2007; Sanders et al. 2007) and SXDS (Ueda et al. 2008). Here we use the first 50% of the data on the COSMOS field (∼4h/pointing) along with previously published 0.15–2.5 μm data (Ilbert et al. 2013) to probe star forming galaxies with masses >10¹⁰ M_☉ to z ∼ 6.

SPLASH is a photometric survey that improves upon existing studies at 4 < z < 6. For these galaxies, the Lyman break allows quality photometric redshift determination, while SPLASH is complete and deep enough to greatly reduce biases that occur in studies specifically selecting for Lyman break galaxies (Lee et al. 2012; Speagle et al. 2014).

In Section 2, we describe the SPLASH catalog, as well as the spectral energy distribution (SED) template fitting we use to produce photometric redshifts (photo-z's), stellar masses, and SFRs. Section 3 describes the SPLASH view of the SFMS, including a lack of observed quenching at the high-mass end. Possible explanations for this lack of a turnoff are discussed in Section 4, and in more detail in Steinhardt & Speagle (2014).

Throughout this work, we assume a standard $(H_0,\Omega _m,\Omega _\Lambda)$ = (70, 0.3, 0.7) cosmology, AB magnitudes, and a Chabrier (Chabrier 2003) initial mass function (integrated from 0.1 M_☉ to 100 M_☉).

This work uses a revised version of the Subaru i⁺ band selected Cosmological Evolution Survey (COSMOS) catalog from Capak et al. (2007) to provide 0.15–2.5 μm photometry, with 10⁶ spurious sources removed (Salvato et al. 2009), updated with intermediate band data, photo-z's, and physical parameters as described in Ilbert et al. (2010). The catalog was further augmented with significantly (1 mag) deeper z⁺ band data taken with an updated Subaru-Suprime-Cam, Ultra-Vista (McCracken et al. 2012) imaging in Y, J, H, and K bands, and importantly the SPLASH IRAC 3.6 μm and 4.5 μm data. A full description of SPLASH is provided in P. Capak et al. (in preparation).

In this Letter we use the first 50% of the SPLASH IRAC data, which covers a 1.2 degree diameter circle centered on the COSMOS field to 5σ ∼25.3 mag_AB at both wavelengths. To reduce active galactic nucleus contamination, X-ray detected sources (Brusa et al. 2010; Civano et al. 2012) were removed from the sample. To overcome source confusion (blending) we extracted the photometry using the i⁺ band catalog as input for the IRACLEAN procedure described in Hsieh et al. (2012). The IRACLEAN photometry was compared with a T-FIT (Laidler et al. 2007) catalog in the CANDELS (Grogin et al. 2011; Koekemoer et al. 2011) area using an HST-WFC3 H band image as a prior and an EMPHOT (Conseil et al. 2011) catalog of the whole COSMOS field using the Ultra-Vista K band image as a prior. No systematic bias as a function of flux or position was found due to the photometric method.

The primary selection effect in our sample is the i-band catalog, which is used as a prior for the Spitzer photometry and as the basis for the photometric catalog. The use of this catalog limits us to z ≲ 6, when the Lyman break leaves the i-band. The depth limits us to unobscured star formation rates >5–10 M_☉ yr⁻¹ at z ∼ 4–6 (typical excitations are A_V ∼ 0.4 mag for objects in our sample; cf. Bouwens et al. 2012). The typical depth of i-band selection is 26.5 mag (Capak et al. 2007; Ilbert et al. 2009). The second limit is the depth of the IRAC data, which introduces an age-dependent mass cutoff. For zero-age galaxies with a flat spectral energy distribution the IRAC data will limit the lowest measurable masses. For older galaxies, however, the depth of the Subaru i⁺ image will limit the lowest measurable masses since they would not be in the i-band catalog. For stellar populations with approximately the age of the universe, our joint mass cutoff is noted in Figure 2. A third, more subtle effect is the method used to extract the IRAC photometry. If the photometry is not properly deblended, fainter galaxies will have systematically increased fluxes due to photometric crowding. The IRACLEAN algorithm corrects for this at least as well as other methods (TFIT, EMPHOT), but we note that this may be affecting a small number of objects crowded by a source not in the i-band catalog.

2.1. Photometric Redshifts

We determine photo-z's using Le_PHARE (Arnouts & Ilbert 2011) with the methodology described in Ilbert et al. (2013). Stellar masses were estimated with Bruzual & Charlot (2003, BC03) models including strong emission lines with a mix of exponentially declining (τ/100 Myr = [0.1, 0.3, 1, 3, 10, 30]) and delayed-τ (Δt_peak/Gyr = [1, 3]) star formation histories (SFHs) with solar and half-solar metallicities at a range of ages spanning 0.05–13.5 Gyr to account for the effects of low metallicities and increasing SFHs on estimated physical parameters. Dust attenuation is modeled using the starburst Calzetti et al. (2000) curve and the λ^−0.9 curve from Arnouts et al. (2013) with E(B − V) up to 0.7 mag.

Although the presence of a Lyman break should result in precise photo-z's, the 4000 Å break and Lyman break can be confused resulting in catastrophic redshift failures. To check the quality of the photo-z's we cross-reference them with spectroscopically confirmed sources (spec-z's), primarily from DEIMOS observations selected to be representative of z > 4 galaxies (Capak et al. 2010; Mallery et al. 2012). Out of 139 galaxies with spec-z > 4 and robust IRAC fluxes, 87 (63%) have successful photo-z determinations (Figure 1) to within 15%, with most of the rest mistakenly fit with at much lower redshift. However, there are very few false positives in the SPLASH catalog. 6% of objects with photo-z > 4 and measured spec-z's lie at spec-z < 2 (indicating confusion between the two breaks), and after correcting for the significantly larger fraction of objects with measured spectroscopic redshifts at z < 2 (10% at z < 2 versus 4% at z > 4 with ch1<25.3), this implies a contamination rate of 2%. So, the current photo-z analysis is preferentially scattering objects down from z ∼ 4–6 to z ∼ 0.5, (Figure 1) with approximately 40% of SPLASH star-forming galaxies erroneously fit as low-redshift and removed from the sample. Tests using the two-point correlation and cross-correlation functions confirm these fractions (Coupon et al. 2013, private communication).

Zoom In Zoom Out Reset image size

We note that typical SFRs are far lower at z ∼ 0.5 than z > 4, so up-scattered objects fall well off the SFMS at z > 4 where they will not affect our further analysis.

2.2. Stellar Masses and Star Formation Rates

2.3. Sample Selection

The main result of this Letter, the SFMS from SPLASH at 4 < z < 4.8 and 4.8 < z < 6 (bins of equal time interval), are shown in Figure 2. In both redshift ranges, there is a strong correlation between stellar mass and SFR, qualitatively similar to that seen in approximately two dozen previous studies (Speagle et al. 2014) but with a different locus than at lower redshift. Almost all objects lie near this SFMS, with a small number of clear outliers at low SFR. Because of the limits of broad-band SED fitting, it is unclear without follow-up observations whether these outliers are candidates for quenched galaxies, catastrophic errors in photo-z's, and/or SED fitting, or have some alternative explanation.

Zoom In Zoom Out Reset image size

The main sequence in each panel is divided by stellar mass into bins of width 0.1 dex, with the median for each bin shown in Figure 3. For ease of comparison with other studies, medians above the mass completeness threshold that encompass >25 objects are then fit with a power law (Equation (1)), with best-fit α and β indicated in Table 1. Using both fits, individual objects exhibit 1σ scatter of ∼0.24 dex about the main sequence, and agree well with Lyman-break selected objects from Lee et al. (2012) at similar/slightly lower redshifts and lower stellar masses.

Zoom In Zoom Out Reset image size

Previous studies have disagreed on whether α is increasing (e.g., Santini et al. 2009), constant (e.g., Karim et al. 2011), or decreasing (e.g., Whitaker et al. 2012) with redshift. We find a high-redshift slope of 0.78 ± 0.02 (Table 1), in good agreement with Lee et al. (2012) and the increasing-α prediction from the Speagle et al. (2014) literature compilation. The uncertainties given are statistical, but this good agreement may imply that the unknown systematic uncertainties are smaller than might otherwise have been expected.

3.1. High-mass Galaxies and Quenching

Several previous studies have examined the redshift evolution of the SFMS, finding that the slope α = d(log SFR)/d(log M_*) may decrease toward higher redshift. SPLASH allows us to check whether this trend continues to z ∼ 6. At lower redshift the evolution of α has been fit as a function of redshift (Whitaker et al. 2012) and of time (Speagle et al. 2014). We find the time-based fit is better at high redshift as (1 + z)∝t^2/3 becomes more clearly non-linear (Table 1), which makes sense because the physical quantity affecting SFR is likely time, not redshift.

It is surprising that the SFMS at 4 < z < 6 has almost exactly the properties that one would extrapolate from lower-redshift behavior, with increasing specific SFRs (≡ ψ/M_*) for M > 10¹¹ M_☉ galaxies. This trend cannot continue indefinitely because the virialization time for a galaxy depends upon its initial overdensity, and a typical galaxy with total baryonic mass of M > 10¹¹ M_☉ (M ⩾ M_*) will not virialize until z ≲ 10, with more massive galaxies having even lower virialization redshifts (Haiman & Loeb 1997). Furthermore, there is an effective Eddington limit on star formation for a given galaxy mass before star formation destroys the interstellar medium (Younger et al. 2008) of

$$\begin{equation} \psi _{\rm{max}} = 600 \sigma _{400}^2D_{\rm{kpc}}\kappa _{100}^{-1}\,M_\odot\, \hbox{yr}^{-1}, \end{equation} \tag{ 2 }$$

where the line-of-sight velocity dispersion in units of 400 km s⁻¹ (σ₄₀₀), the interstellar medium opacity in units of 100 cm g⁻² (κ₁₀₀), and characteristic physical scale of the starburst in kpc (D_kpc) are likely of order unity. Thus, even under ideal conditions, it takes at least an additional ∼10⁸–10⁹ yr beyond virialization for the largest protogalaxies observed in SPLASH to attain M_* > 10^11.5 M_☉. The universe at z ∼ 6 is slightly less than 10⁹ yr old.

At z ∼ 6 the formation of these galaxies is possible if they spend most of their time at near-maximum SFRs. Such galaxies would then not have a history of sporadic starbursts, but rather near-continuous high-rate star formation from the moment of collapse until z = 6. This calculation, however, is only constrained to order of magnitude, so some variability in their SFHs is possible. However, even under ideal conditions this trend cannot continue far beyond z > 6, as there is insufficient time to build up stellar mass. Thus, we should expect that upcoming HSC data, which is anticipated to extend the SPLASH catalog to z ∼ 8, should find a turnover redshift where the most massive galaxies are not observed, indicating we are seeing the rapid buildup of galaxies that have just completed their primordial formation and virialization. Otherwise, a substantially modified theory for early-universe structure formation will be required.

HSC observations will also provide a substantial improvement in completeness in selecting high-mass, low-SFR galaxies. Such galaxies lying only slightly below the SFMS would be present in the current data, but improved Y- and Z-band detection using HSC is necessary to select dusty and/or more quiescent high-mass, high-redshift galaxies. The current SPLASH catalog is sufficient to demonstrate that there is no population of turnoff galaxies lying just below the main sequence, but cannot determine whether this is because no galaxies have completed their star formation by z ∼ 6 or because these galaxies have rapidly changed their SED and are lost due to selection. Such galaxies have been observed to z ∼ 3.5 (Ilbert et al. 2013) and likely form in Eddington limited starbursts (Toft et al. 2014), but it is unclear when the quenching process began. Finding these first turnoff galaxies is essential to understanding when and how the most massive systems in the universe are quenched.

Taken together, our results provide strong hints that the most massive galaxies at high redshift all have a very similar history, one where they are forming stars at a nearly maximal rate from the first moment of virialization until a rapid turnoff at high redshift. This might turn out to be incompatible with the current consensus view that most galaxies form their stars through a series of individual starbursts, likely triggered by local events and environment and individual to the history of each galaxy. However, SPLASH is reporting only on the most massive galaxies, formed at the highest redshifts and very overdense initial fluctuations, so such galaxies are not necessarily typical. A natural next step is to follow up on this analysis, combining it with observational evidence about the formation process for galactic nuclei to try and build a consistent framework for galaxy evolution (cf. Steinhardt & Speagle 2014), but such a model is beyond the scope of this Letter.

The authors would like to thank Steve Bickerton, Sean Carey, Martin Elvis, and Brian Feldstein for helpful comments.