Main

We have obtained rest-frame far-infrared (FIR; 3–1,000 μm) measurements of nine ‘normal’ (∼1–4 times average luminosity, L*) galaxies and one low-luminosity quasar at redshifts of z = 5–6, about one billion years after the Big Bang. These data were taken with an early 20-antenna version of the Atacama Large Millimeter Array (ALMA) in band 7 (1,090–800 μm) continuum mode with 7.5 GHz of bandwidth, ∼20 min of integration, and with the compact array yielding a resolution of ∼0.6 arcsec. One of the four side bands was always centred on the 158 μm [C ii] line, which is the dominant FIR cooling line for neutral gas in normal star-forming galaxies13,14 and hence is a good indicator of galaxy dynamics and the spatial extent of the interstellar medium14,15 (ISM). The other three spectral bands yield a measurement of the dust continuum emission at λ ≈ 150 μm, which is a good indicator of the total infrared luminosity for normal FIR spectral energy distributions (SEDs; see Methods section ‘Derivation of physical parameters’).

The objects we studied were Lyman break galaxies (LBGs) selected from the 2 square degree Cosmic Evolution Survey (COSMOS) field16; these LBGs (with one object later discovered to be a low-luminosity quasar) have spectroscopically determined absorption line redshifts measured by the Deep Extragalactic Imaging Multi-Object Spectrograph (DEIMOS) on the W. M. Keck-II Observatory in Hawaii. The objects were selected to be ‘characteristic’ with luminosities between 1 and 4 L* and with ultraviolet spectral slopes (β) between −1.4 and −0.7. Three objects (1, 2 and 10) also meet the selection criteria for Lyman-α (Lyα) emitters at z ≈ 5.7. When possible, ultraviolet morphologies are measured using the Hubble Space Telescope (HST) Advanced Camera For Surveys (ACS) F814W (0.8 μm) data17. Stellar masses were determined by fitting SED models to existing COSMOS data and new 3–5 μm photometry from the Spitzer-SPLASH survey18 (see Methods section ‘Optical and NIR data’). Rest-frame infrared luminosities were determined by assuming the range of FIR properties measured at z ≈ 2–5 (ref. 19) and have a systematic uncertainty of 0.3 dex due to the uncertainty in the shape of the FIR SED (see Methods section ‘Derivation of physical parameters’).

We detect four of nine galaxies in the dust continuum (Fig. 1, Extended Data Table 1), and a mean combination (stacking) of the data on the remaining six objects yields a continuum detection of 35 ± 13 μJy. This is less than 1/12 of that expected for systems with similar ultraviolet properties at z < 3 assuming the worst-case systematics in our FIR determination (see Methods section ‘Robustness of upper limits on dust emission’). This can be quantified with the infrared-excess to ultraviolet-slope (IRX–β) relation4 (Fig. 2), which relates the amount of dust absorption measured in the ultraviolet to the amount of infrared re-emission and is sensitive to both dust properties and column density. We find that the measured IRX and β for our continuum detected objects and A1689-zD111 are consistent with the IRX–β relation for the Small Magellanic Cloud (SMC)20, which has lower-metallicity gas, and less thermal FIR emission21 than typical galaxies at z < 3. Our undetected systems appear to have even less IRX than the SMC, but could be consistent given the large systematic uncertainties. Hence, we conclude z ≈ 6 galaxies are significantly less dust-obscured than more-evolved systems at z < 3, confirming previous results1,2,3,6.

Figure 1: Optical, [C ii] and continuum maps of the sources HZ1–HZ10.
figure 1

The [C ii] line detections (red contours) and weak ∼158 μm FIR continuum detections (blue contours) are shown with the rest-frame ultraviolet images as the background. The images are 5″ × 5″ (scale bar at bottom left) and the contours are 2, 6 and 10 σ with [C ii] line profiles for each source shown in Fig. 3. The background images are from HST-ACS in the F814W17 band where the morphologies will be affected by Lyα, except for HZ10, which is Subaru z′ band. All objects are detected in [C ii], showing that a large amount of gas is present in these systems, but only four are detected in continuum.

PowerPoint slide

Full size image
Figure 2: IRX–β measurements of z > 5 objects.
figure 2

The deficit of infrared emission in our sample is evident in the presented infrared excess versus ultraviolet slope (IRX–β) relation when compared with models. Detections are indicated in red, upper limits in orange, the mean IRX ratio (obtained by combining undetected sources) in blue, and A1689-zD111 in maroon. Error bars are 1σ, and include standard measurement error and systematic uncertainty added in quadrature. The Meurer29 relation, which assumes Calzetti-like dust and is consistent with typical galaxies at z < 3, is shown as a black solid line, while a model for lower-metallicity SMC-like dust model20 is shown as a dashed line.

PowerPoint slide

Full size image

The very low IRX values at β ≈ −1.2 implied by our non-detections are difficult to explain in the context of the very young systems expected at z = 5–6 (ref. 22). The simplest explanation is the systematics in our estimates of β and total infrared luminosity (LIR), which would imply relatively un-obscured systems with ∼45 K dust (see Methods section ‘Derivation of physical parameters’). But we cannot exclude changes in the dust geometry or rapid disruption of the molecular clouds that could lead to lower IRX vales at a given β with more-normal dust temperatures23. Higher signal-to-noise near-infrared photometry and shorter wavelength FIR data will ultimately be needed to understand these sources.

A corollary of this result is that ultraviolet derived star-formation rates (SFRs) at high redshift (z > 5) should use the SMC-like IRX–β relation or assume no dust rather than the currently assumed Meurer IRX–β relation. This will decrease the ultraviolet derived SFRs by a typical factor of ∼2–4 for individual galaxies from those implied by the Meurer relation for similar values of β. The effect on the global star-formation history is much smaller, <40%, because the majority of star-formation is in low luminosity (<L*) galaxies that were already assumed to have little or no dust extinction (see Methods section ‘Effects of evolving dust on the global star-formation history’, Extended Data Fig. 4).

In contrast to the dust emission, we find >3σ detections of the [C ii] line in all nine normal galaxies (Fig. 3). The line emission is spectrally resolved in all cases, with [C ii] velocity dispersions of 63–163 km s−1, and marginally resolved at our spatial resolution of ∼0.5–0.9 arcsec, indicating galaxies with dynamic masses (Mdyn) of ∼109–1011 solar masses (M⊙; see Methods sections ‘Reduction of ALMA data’ and ‘Derivation of physical parameters’). We also detect two optically faint [C ii] emitters at redshifts consistent with the targeted objects. HZ5a is detected near HZ5 at a redshift consistent with the in-falling gas seen in the optical spectra of HZ524. HZ8W corresponds to an optically faint companion to HZ8 and has a similar redshift. Taken together, the direct and serendipitous detections suggest ubiquitous and enhanced [C ii] emission in early galaxies similar to that seen in local low-metallicity systems12 (Fig. 4).

Figure 3: [C ii] line profiles for the sources.
figure 3

The line velocity profiles of our [C ii] measurements are shown for our 10 sources and two blind detections, HZ5a and HZ8W. All sources but HZ5, a quasar, are clearly detected. For objects HZ8 and HZ8W, a dotted line showing the integrated flux over both sources is also plotted. The consistency between the integrated and individual profiles indicates they are at similar redshifts.

PowerPoint slide

Full size image
Figure 4: L[C ii] over LIR ratio as a function of LIR.
figure 4

Our sample has a clear excess of [C ii] emission relative to IR emission when compared with normal lower-redshift systems (z = 0–3)6,12,15,30. Detections are indicated in red, upper limits in orange, and the mean of undetected sources is shown in blue. The high L[C ii]/LIR ratio we measure indicates a lower dust-to-gas ratio, more-diffuse clouds, and a more-diffuse ultraviolet radiation field than seen in normal z ≈ 0 galaxies12,13,15,30. Error bars are 1σ, including standard measurement error and systematics added in quadrature.

PowerPoint slide

Full size image

The [C ii] enhancement in local systems is caused by a lower dust-to-gas ratio which allows the ultraviolet radiation field to penetrate a larger volume of molecular cloud12. Our significantly lower IRX values and enhanced [C ii]/FIR ratios would suggest a similar effect is happening in high redshift galaxies. But other possible causes of the transition in obscuration properties with redshift have been suggested25: evolution in metal abundances that changes the intrinsic ultraviolet slope; changes in the dust properties; and differences in the dust geometry. The systems in this study were selected to have broad ultraviolet absorption features in their spectra that indicate a relatively homogeneous metal abundance of ∼0.25 times the solar value. At this metallicity, the ultraviolet spectral slopes are expected to be similar to those of solar metallicity systems26. Furthermore, the population from which these objects were selected has dust attenuation properties similar to those of lower-redshift objects27. So changes in the dust properties that would be measureable in the attenuation curve are probably not the primary factor. When compared with objects of the same stellar masses, our sample has physical properties (gas velocity dispersions, sizes, and SFRs) that imply geometries consistent with z ≈ 1–3 galaxies that exhibit dust properties like those found in the Meurer model28. Changes in the average cloud geometry are therefore not obvious, but we cannot fully exclude this possibility as it would explain the observed IRX properties. So we conclude that a significant decrease in the dust-to-gas ratio is the most likely explanation for the evolution in extinction properties.

These results are seemingly at odds with previous measurements6,7,8,9 that failed to detect [C ii] emission at high redshift. Our sample contains objects with both emission and absorption of Lyα, and the [C ii] emission strength is un-correlated with the Lyα properties (Extended Data Table 1), which means that the use of Lyα redshifts is not the cause. The most likely explanation for the lack of detections is observational uncertainty in the expected line flux, line width, and frequency. Even obscuration uncorrected ultraviolet SFRs will over-estimate the expected FIR continuum flux, because less than 20% of the star formation is leading to FIR emission in our sample6 (Fig. 2). This means the FIR emission of galaxies will be more difficult to detect than previously expected based on z < 3 scaling relations (see Methods sections ‘Robustness of upper limits on dust emission’ and ‘Quantifying [C ii] non-detections in the literature’), so longer integration times are required. That said, these data show [C ii] is enhanced and readily detectable with ALMA at z ≈ 6 if moderate integration times are used. Finally, we note that the [C ii] lines are resolved in velocity space, so single channel sensitivity should not be used for limits (see Methods section ‘Quantifying [C ii] non-detections in the literature’).

Methods

Cosmology

Throughout this paper we assume a simplified cosmology of ΩM = 0.3, ΩV = 0.7 and Ho = 0.7 when calculating physical parameters. These values approximate the 2013 Plank31 and 9-year WMAP32 values, and are commonly assumed in the current literature to make comparisons easier over time as the exact measured values evolve.

Reduction of ALMA data

The ALMA observatory staff, as part of standard data processing and delivery, performed initial data calibration. The calibrated visibility data were then re-analysed, and additional flagging of bad time periods in the data and bad channels was performed. The data were re-imaged using the CASA Briggs flexible weighting of the u-v visibility data with Robust = 1 to create continuum, moment zero, and channel maps. Gaussian fitting of the spectral and spatial data was performed using the CASA viewer and CASA fit tool.

The centroid frequency of these models was compared to the rest frequency of [C ii] to measure the redshift, which was compared to the previously determined redshifts from optical spectroscopy. Since these measures were identical, within our precision, the reference frequency of each cube was set to the centroid frequency, allowing the FWHM of the Gaussian to be read in km s−1 and the line flux in Jy km s−1. The [C ii] line contribution was subtracted from the measured continuum fluxes.

The measurements derived from the ALMA data are given in Extended Data Table 1 and Gaussian fits done with CASA to the image components are given in Extended Data Tables 2 and 4 for the [C ii] and continuum data, respectively. The fits were not constrained by the beam size, and so they can be slightly smaller due to the data errors. The errors quoted are those produced by the CASA Gaussian fitting routine and are sensitive to the box size chosen for the fit, especially in cases of low signal to noise. Deconvolved size measurements for the [C ii] lines are given in Extended Data Table 3. Only objects 3 and 5a are unresolved in [C ii]. In contrast, only object 10 is resolved in the continuum measurement. The baseline configuration used will resolve out spatial scales larger than 5–7″ (29–43 kpc) but none of our sources appear to have emission on that large a scale.

Optical and NIR data

These were obtained from the COSMOS photometric redshift catalogue33 augmented with additional data from the Spitzer-Large Area Survey with Hyper-Suprime-Cam (SPLASH)34. Sizes were taken from the COSMOS Hubble Space Telescope Advanced Camera for Surveys (HST-ACS) weak lensing catalogue16 and are given in Extended Data Table 5. Object 10 falls outside the area covered by HST-ACS.

Derivation of physical parameters

Since only one FIR data point was available, total infrared (3–1,100 μm) luminosity was derived using the range of grey body models covering the full range of objects observed at z ≈ 2–519. The grey body model parameters, α the blue power-law, β the long wavelength slope, and T the black body temperature, ranged between α = 1.5–2.5, β = 1.2–2.0 and T = 25–45 K. The range of allowed infrared luminosities was determined by scaling these grey bodies to the observed rest frame 158 μm flux. This range of parameters resulted in a systematic uncertainty of 0.3 dex in the derived luminosity. This relatively small error is due to the fact that 158 μm is on the flat part of the SED close to the emission peak for the range of observed physical parameters. The derived infrared luminosities are given in Extended Data Table 5.

The rest frame ultraviolet properties were determined by fitting a power-law to the rest frame 0.13–0.3 μm photometry from the COSMOS and Ultra-Vista surveys. This power-law fit was then used to estimate the ultraviolet spectral slope (β) and the luminosity at 0.16 μm, where LUV = ν1,600Lν1,600. The power-law fits appear to be biased high by the bluest bands in several cases which may be due to the intrinsic SED flatting at short wavelengths35,36. To account for this possible systematic, we also fit the slope using only the rest-frame 0.15–0.3 μm photometry and include the shifts in our measurement error bars. Finally, simulations2 of observational biases from blending and signal-to-noise indicate that an additional ±0.3 systematic error should be included. These estimates are tabulated in Extended Data Table 5.

For consistency, we re-analysed A1689-zD111 with the same methods used for our sample. We find LIR = (6.8 ± 1.2) × 1010L⊙, LUV = (1.8 ± 0.4) × 1010L⊙ and β = −0.67 ± 0.78 assuming a lensing magnification of 9.3 versus LIR = (6.2 ± 0.8) × 1010L⊙ and LUV = 1.8 × 1010L⊙. No value was published for β derived from the photometry, but our analysis would not have produced a consistent LUV if the two analyses of β were inconsistent because LUV is derived from the β fit.

Stellar masses were derived using data from the Spitzer Large Area Survey with Hyper-Suprime-Cam (SPLASH) in the same way described in previous work33. In brief, the COSMOS and SPLASH 0.1–5 μm photometry was fitted with Bruzual and Charlot templates37 with a Chabrier IMF allowing for a range of emission line strengths and extinction laws. The estimated stellar masses are given in Extended Data Table 5.

Dynamical masses were estimated based on the [C ii] velocity dispersion (σ[C ii]) and size estimates in Extended Data Table 3. We used the method outlined in Wang et al.38 to estimate the dynamical masses. In this approximation Mdyn = 1.16 × 105 Vcir2 D, where Vcir is the circular velocity in km s−1 and D is the diameter in kpc. Vcir = 1.763 σ[C ii]/sin(i), and i, the disk inclination angle, is approximated by the axis ratio of the object as i = cos−1(b/a).

This method yields an approximate answer but is very uncertain because of our low spatial resolution that leads to a large uncertainty in both i and the object size. The errors quoted in Extended Data Table 5 include the error in the semi-major axis and the velocity dispersion and assumes sin(i) = 0.45–1. This range of sin(i) approximates the range of values from a dispersion dominated system to an edge-on disk, but could be significantly larger.

We find the dynamical masses are typically a factor of ∼3 greater than the stellar masses (see Extended Data Fig. 3). This is higher than the factor of 1.2–1.7 found at z ≈ 1–3 from similar Hα based dynamical mass estimates28, but consistent given the very large errors in our measurements.

Robustness of upper limits on dust emission

The conclusion that the dust properties, and specifically the dust mass and dust to gas ratio, of galaxies are evolving relies on two quantities: the measured deficit of rest-frame 158 μm flux, and the models that translate this into a deficit of thermal emission. Here we quantify the robustness of this result.

If we assume a Meurer et al. like IRX–β relation and the mean of our grey body models, the predicted mean ALMA flux of the undetected sources is 833–7,610 μJy (given the uncertainty in β), which is >61σ discrepant from the measured mean of 35 ± 13 μJy. Individual sources are discrepant from the Meurer et al. relation by >17–490σ except for objects 9 and 10 which are consistent given the large errors in β. If we assume the maximal range of the systematic error, 45 K dust, β = 1.5, α = 1.2, the predicted mean ALMA flux is 447–4,087 μJy, which is >32σ discrepant from the measured mean. Finally we consider an extreme model with 100 K dust, β = 1.5, α = 1.2; we predict an ALMA flux of 78–710 μJy, which is 3σ discrepant from the measured mean. In this extreme case, objects 4, 9 and 10 are consistent with the Meurer et al. relation, and the remaining sources are discrepant at the >1–30σ level.

We also performed a completeness analysis following Casey et al.39 and find the low dust emission is unlikely to be due to selection effects in our ultraviolet sample. However, we note the result only holds for ultraviolet selected samples. The general galaxy population may contain heavily obscured objects that may not be easily found by ultraviolet selected surveys.

Quantifying [C ii] non-detections in the literature

Our sample predicts a peak [C ii] line flux of ∼0.2–0.7 mJy for an L* (ref. 40) galaxy with an SFR ≈ 6.5 at z ≈ 6–7, below the detection limit of most previous studies of z ≈ 6–7 objects6,7,8,9.

In addition to sensitivity, we note Lyα redshifts are systematically offset by ∼400 km s−1 and up to 1,200 km s−1 from the [C ii] (and systemic absorption line) redshifts41, which is a significant fraction of the ALMA band pass.

Finally, caution must be used when interpreting the limits given in the SFR–L[C ii] relation because the estimated line luminosity limits scales as the square root of the assumed line width. A 36–100 km s−1 FWHM channel width6,7,8,9 is typically used for these limits, but our measured line widths in Extended Data Table 1 are significantly broader. If we use our measured line widths this would predict for [C ii] a line limit 0.2–1.0 dex higher than those typically quoted, making them consistent with the SFR–L[C ii] relation.

Comparison of gas and SFR properties to general redshift scaling relations

This sample is enhanced in L[C ii] relative to LIR compared with lower-redshift samples6 (Fig. 4). These sources have ∼1% of their luminosity emitted in the [C ii], consistent with low metallicity and extended star forming objects12,15.

This sample appears to follow the local SFR–L[C ii] relation if the SFR is determined by summing the ultraviolet and FIR measurements6 rather than using a dust corrected ultraviolet estimate (Extended Data Fig. 1). However, if an ultraviolet derived dust correction is used, the points shift ∼0.5–1 dex higher.

Finally we turn to the stellar-mass–SFR or ‘main-sequence’ relation (Extended Data Fig. 2)18. We find our objects generally agree with this relation, but several objects fall below the relation. Specifically, objects HZ1, HZ2 and HZ3 are below it, but HZ1 has an unusually low dynamical to stellar mass ratio, indicating the stellar mass may be over-estimated.

Effects of evolving dust on the global star-formation history

Our results indicate the properties of dust and the amount of extinction evolve significantly between z ≈ 3, where it is well measured, and z ≈ 5–6, where we measure it. To estimate the effect on the global star-formation history, we adopt the luminosity functions and ultraviolet slope measurements from the series of papers by Bouwens et al.2,42, which are largely supported by other work43. We then assume the dust correction can be described by the Meurer29 curve at z < 4, the SMC20 curve at z ≈ 5, and the SMC curve scaled to our measured IRX at z ≈ 6–8. The results of this analysis are shown in Extended Data Fig. 4 and result in ∼30–40% less star-formation at z ≈ 6 than previously thought. For consistency with the literature, the SFR density is shown assuming a Salpeter IMF; a factor of 0.68 needs to be applied to convert to a Chabrier IMF.

Sample size

No statistical methods were used to predetermine sample size.

Randomization

The experiments were not randomized.

Blinding

The investigators were not blinded to allocation during experiments and outcome assessment.