TWO BRIGHT SUBMILLIMETER GALAXIES IN A z = 4.05 PROTOCLUSTER IN GOODS-NORTH, AND ACCURATE RADIO-INFRARED PHOTOMETRIC REDSHIFTS

GN20 is very securely detected in both D- and B-configuration observations (Figure 1 (left)). Most of the signal appears to be due to a CO emission line. Averaging the signal from 150 km s⁻¹ to 900 km s⁻¹ (here and in the following all velocities are defined with respect to the central frequency of the observations at 91.375 GHz) results in a detection with signal-to-noise ratio (S/N) = 10 for the D-configuration data (Figure 1 (left) and Figure 3 (top)). The B-configuration observations within the same velocity range result in a detection with S/N = 13 and a position of α(J2000) = 12^h37^m1190, δ(J2000) = 62^d22^m121 (Figure 1 (left)). This position is consistent with the 7'' submillimeter error box of the SMG GN20 (P06), and is within 05 of the accurate interferometric Iono et al. (2006) submillimeter-array (SMA) detection of this source and of the radio counterpart in 20 cm VLA data (G. Morrison et al. 2009, in preparation; Figure 4). The PdBI position coincides well with a faint galaxy that P06 already identified with the optical counterpart of GN20 (Figure 5). GN20 is the brightest SMG in the GOODS-N area. The radio, CO, and SMA submillimeter positions actually appear to be slightly offset to the east of the source seen in the optical (by ∼05 or so), as also noted by Iono et al. (2006). It is not clear if this is due to some additional galactic structure being present but extremely obscured, or just a random fluctuation across the data sets.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The PdBI D- and B-configuration data sets were independently fitted with a point source and the resulting spectra were corrected for PBA and co-added with appropriate weighting (the D-configuration observations are noisier on an absolute flux scale, also due to the larger correction for PBA). We performed Gaussian fitting of the resulting spectrum (Figure 3) to determine the CO line properties, allowing for the presence of a faint underlying continuum. The best-fitting Gaussian line is offset by 580 ± 40 km s⁻¹ relative to the central frequency, and has an velocity integrated flux of I_CO = 1.5 ± 0.2 Jy km s⁻¹, with both errors given at the 1σ level and following Avni (1976) for the case of two parameters of interest (line velocity and integrated emission line flux). There is evidence for faint 3 mm continuum as well, as shown in Figure 1 (right panels) and Figure 3. To confirm this result, we averaged the signal in UV space for velocities outside a 1200 km s⁻¹ range centered on the line (about 1.7×FWHM), where negligible line contribution should be present. The resulting continuum map has a faint source at a position consistent within the errors with that of the emission line and corresponding to a signal of 0.33 mJy, significant at the 5σ level. To our knowledge, this is the first secure dust continuum detection of an SMG in the 3 mm band.

Averaging the velocity channels between −300 km s⁻¹ and 825 km s⁻¹ a source is detected in the PdBI map with S/N = 5.8 at the position of α(J2000) = 12^h37^m0877, δ(J2000) = 62^d22^m017 in the B-configuration observations (Figure 2 (left) and Figure 3 (right)). Positive flux with S/N just below 3 is also observed in the D-configuration data at this position (Figure 2 (left)). We extracted the spectra by fitting point sources in the D- and B-configuration data sets independently and co-added the spectra with weighting after correcting for PBA. Averaging the total resulting spectrum within −300 km s⁻¹ and 825 km s⁻¹, we find a positive signal significant at the 7σ level.

The position of this source is within 02 of the relatively bright radio source that P06 identified with the counterpart of the SMG GN20.2, close but well distinguished from GN20 that is 24'' to the NE. The position corresponds to a faint galaxy in the deep Hubble Space Telescope (HST) images (Figure 5). Also in this case the signal appears to be mostly due to an emission line. The best-fitting Gaussian line is offset by 330 ± 150 km s⁻¹ relative to the central frequency, and has a velocity integrated flux of I_CO = 0.9 ± 0.3 Jy km s⁻¹. The largest contribution to the flux error is the uncertainty in the continuum level. Some positive continuum appears to be present, although the S/N is not yet sufficient for a reliable measurement (Table 2). Accounting for the possible continuum emission and its uncertainty, the CO emission line is still detected at the 6σ level.

Inspection of the newly reprocessed 1.4 GHz radio map around the position of GN20.2 (Figure 4; G. Morrison et al. 2009, in preparation) shows an additional faint radio source with a flux density of 32.2 μJy (5σ) at α(J2000) = 12^h37^m0973, δ(J2000) = 62^d22^m026 (radio frame), 34 from the GN20.2 SCUBA position (P06) and 68 from the CO-detected counterpart discussed above. We refer to the primary counterpart of GN20.2 as GN20.2a, and this newly identified faint radio source is referred to as GN20.2b. GN20.2b is a plausible additional counterpart to the submillimeter emission.

Interestingly, positive signal is found in the PdBI map close to the position of GN20.2b. The PdBI D- and B-configuration data sets were independently fitted with a point source fixed at the VLA position of GN20.2b (a secure position cannot be derived from the PdBI data alone, due to the low S/N) and the resulting spectra were corrected for PBA and co-added, weighting according to the noise, similarly to what was done for the other galaxies. Averaging over the whole 1 GHz bandwidth in the co-added spectrum results in a ∼3σ signal. If we average the spectrum between ≈0 km s⁻¹ and ≈700 km s⁻¹, we find a >4σ signal. Although longer integrations at PdBI would be required for a more reliable conclusion, the data are again consistent with CO emission, quite close in frequency to the other securely detected lines. If real, that would correspond to an integrated flux of I_CO ≈ 0.45 Jy km s⁻¹.

3.1.1. Photometric Redshift from Stellar Emission

Figure 5 shows HST+ACS imaging in the four bands available in GOODS (Giavalisco et al. 2004a). As already noted by P06, the optical counterpart of GN20 is a B-dropout Lyman break galaxy according to the definition of Giavalisco et al. (2004b). This definition returns galaxy samples with z = 3.78 ± 0.34, favoring the identification of the PdBI line as CO[4–3] at z = 4.055. Circumstantial evidence for the likelihood of a redshift around z = 4.05 had been even earlier presumed from the presence of a bright B-band dropout galaxy at a 16'' separation (Figure 5) with fairly similar colors and a known spectroscopic redshift of z = 4.058 measured from Keck spectroscopy (D. Stern et al. 2009, in preparation; we refer to this object as BD29079). The IRAC colors of the GN20 B-band dropout counterpart are very red, with the brightest IRAC band being the 8.0 μm (see e.g., P06). If due to the emission of stars, this implies that the 1.6 μm rest-frame bump is beyond the 5.8 μm channel, or z ≳ 3.5. However, the IRAC emission can also be red due to, or affected by, the presence of obscured active galactic nuclei (AGNs; e.g., Stern et al. 2005; Daddi et al. 2007b).

Stellar photometric redshifts were thus estimated using only the four HST bands and running the hyperz code (Bolzonella et al. 2000). The Maraston (2005) models were used, with a variety of star formation histories and allowing for reddening using the Calzetti et al. (2000) law. The near-IR bands are not used because the available data sets, described, e.g., in Daddi et al. (2007a), are not sufficiently deep to place meaningful constraints on the spectral energy distribution (SED) of this faint and relatively blue galaxy. The distribution of χ² values versus redshift (Figure 7) is consistent with 3.4 < z < 4.2 (95% confidence range). Figure 7 shows the observed SED and the best-fitting model for the case of z = 4.055.

Zoom In Zoom Out Reset image size

3.1.2. Photometric Redshift from the Infrared and Radio Emission

An independent estimate of the redshift of GN20 can be derived from its FIR emission. The ratio between mid-IR, FIR, mm, and radio flux densities is redshift dependent. Locally, the radio luminosity is observed to be proportional to the bolometric IR luminosity (Condon 1992; Yun et al. 2001), and this seems to hold at high redshift as well (Elbaz et al. 2002; Garrett 2002; Appleton et al. 2004). However, the K-corrections at radio and FIR wavelengths are very different, and thus the radio/FIR flux ratio depends strongly on redshift. Originally this idea was exploited by Carilli & Yun (1999, 2000) to estimate photometric redshifts of distant SMGs, also attempted by several other studies (Hughes et al. 2002; Wiklind 2003; Aretxaga et al. 2003, 2005, 2007; Clements et al. 2008). In addition, one can exploit the fact that the mid to FIR emission has a generally well-defined peak at 50–200 μm, affecting directly the ratio of mid-IR to submillimeter and millimeter emission.

For this purpose, we use the available photometry at 24 μm, 850 μm, 3 mm, and 20 cm. Some of the Spitzer+MIPS 24 μm and VLA 20 cm measurements are updated from P06 using the most up-to-date GOODS data sets. We measure a 24 μm flux density of 65.5 ± 3.5 μJy and a 20 cm flux density of 75.8 ± 7.9 μJy. We use a 850 μm flux density of 20.3 ± 2.3 mJy from P06 and a 3.3 mm flux density of 0.33 ± 0.07 mJy (Section 2).

The observed flux densities have been compared, as a function of redshift, to the predictions of a suite of 105 template SEDs that were built following the luminosity correlations observed for local galaxies as described in Chary & Elbaz (2001). The template dust temperatures get warmer with increasing luminosity from a minimum template luminosity of Log L_IR/L_☉ = 8.41 up to a maximum luminosity of Log L_IR/L_☉ = 13.55. Radio continuum emission is added to each template following the radio–IR correlation (Yun et al. 2001). For each template of a given luminosity, we compute the expected flux densities in the available bands (24 μm, 850 μm, 3 mm, and 20 cm) as a function of redshift and perform a χ² minimization over the template total IR luminosity as a function of redshift, without allowing the template normalizations to vary. In order to avoid biasing the fitting to a particular band, we increased the formal uncertainty in the 24 μm flux density in order to reduce the S/N from 20 to 9, comparable to the radio and submillimeter bands.

Figure 8 shows the results. The best-fitting redshift is z_phot = 3.7 and the 99% confidence interval is 3.3 < z < 4.3. Therefore, we conclude that the radio–IR SED also independently supports the z = 4.055 redshift determination. We note that this analysis allows us to reject with high confidence any redshift identification with z < 3 or z > 5.

Zoom In Zoom Out Reset image size

Figure 8 (right panel) also shows the comparison of the observed GN20 SED from UV to radio to the best-fitting (and brightest) template in the library of Chary & Elbaz (2001). The SED is remarkably well reproduced in the mid- to far-infrared and radio, without any rescaling or ad hoc normalizations. The UV–optical part of the template SED also happens to match the shape of the observed data quite well, although there is a factor of ∼2 offset over the IRAC bands that implies a different L(FIR)/L(optical) ratio, suggesting a higher specific star formation rate (SSFR). A similarly good agreement, including the optical/UV portion of the SED, was seen by Daddi et al. (2005b) when comparing the average SED of z = 2 BzK-selected ultraluminous infrared galaxies (ULIRGs) to Chary & Elbaz (2001) templates. The optical/UV emission was simply added to the models using a synthetic stellar population model matched to the Arp220 SED and scaled to reproduce the local broad correlation between bolometric luminosity and B-band rest-frame luminosity.

3.1.3. Keck Spectroscopy

A nearby (24'' separation; 169 kpc proper distance at z = 4.05) SMG companion to GN20 is GN20.2, first discovered by Chapman et al. (2001).¹³ GN20.2 has an 850 μm flux density of 9.9 mJy as reported by P06, who also identifies the most likely optical counterpart with a relatively bright radio source (S(20 cm) = 180 μJy). This galaxy is also a B-dropout (P06; Figure 5) with a faint 24 μm flux density of 30.2 μJy.

An analysis similar to the one performed in Section 3.1 results in a photometric redshift of z_phot,opt = 3.93 and a 90% confidence range within 3.55 < z < 4.05. Similarly, the analysis of the IR SED suggests a redshift of 3.4 < z < 4.2 at the 99% confidence. This confirms that the overall properties of this source are compatible with it being at the same redshift of GN20. This is further demonstrated by deep Keck+DEIMOS spectroscopy (2.5 hr) of this galaxy that was obtained in 2005 April during cirrus conditions, and reduced in the same way as for the GN20 observations (see Section 3.1.3). Figure 9 shows that a strong spectral break is detected at about 6150 Å, consistent with the location expected for the Lyα forest break at a redshift of 4.055. The observed break corresponds to a Lyα forest decrement D_A = 0.67 ± 0.05, larger than the average expected value but still within the observed scatter in the intergalactic medium transmission at z = 4.05 (Songaila 2004). We constrain the redshift of this source to be z = 4.059 ± 0.007 (1σ), based on fitting the Keck spectroscopy with the Maraston (2005) model shown in Figure 9, demonstrating that this source is at a similar redshift to that of its brighter companion GN20. We conclude that in this case the detected emission is also CO[4–3] at z = 4.051 ± 0.003, in good agreement with the Keck spectroscopy.

Zoom In Zoom Out Reset image size

The GN20.2b radio source is 34 from the GN20.2 SCUBA position of P06, closer than GN20.2a which is 60 away (although, we caution that the SCUBA position of GN20.2 might be less accurate than that for the average SMG, due to the brighter and nearby companion, GN20). The radio position is close (only 05 to the north) to a very faint galaxy (z = 27.34) detected in the HST+ACS imaging (Figure 5). It is possible that this faint radio galaxy could be at the same z ∼ 4.05 redshift of GN20 and GN20.2a, and that it could be contributing some of the submillimeter flux detected at the position of GN20.2. We will discuss these two issues here in turn.

Although undetected in the HST/ACS F435W band, the faint ACS counterpart to GN20.2b is not formally classified as a B-band dropout due to the upper limit in F435W not being stringent enough, but its blue (i − z) and red (v − i) colors support the hypothesis that this object is also at z ∼ 4. The galaxy is also fairly bright and red at IRAC wavelengths (SED peak at 8.0 μm with 20.93 mag) and is selected as a z > 3.5 massive galaxy candidate in the work of Mancini et al. (2009). Inspecting the highest resolution IRAC imaging at 3.6 μm and 4.5 μm, it appears that blending with a neighboring source is somewhat affecting the Spitzer IRAC measurements and biasing the flux densities high (likely by at most a factor of 1.5–2). We used the radio–IR photometric redshift technique to constrain its redshift, under the assumption that the emission is dominated by star formation and not an AGN. We find that by itself, the large radio to 24 μm flux density ratio for this source strongly constrains its redshift to z > 3.2 (99% confidence level), independently of how much it contributes to the 9.9 mJy flux density at 850 μm. This is consistent with this object being an intrinsically very luminous, highly star-forming galaxy at high redshift (although, in principle, the high ratio could also be due to radio-loud AGN emission at a lower redshift). As discussed in Section 2.3, there is possible evidence for CO emission from this source, where a positive signal is found at the >4σ level. Although the situation is not secure as for the other two sources, we conclude that there is some supporting evidence for a similar z ∼ 4.05 redshift also for GN20.2b. Taking the CO signal at face value, we would assign a redshift of z = 4.052 ± 0.006 to GN20.2b, where the error corresponds to half the frequency range spanned by the tentative positive emission, if we identify that as CO[4–3]. This is within Δz = 0.003 of the redshift of GN20 and GN20.2a, while our data are sensitive in principle to emission from a redshift range that is 17 times wider, from 4.02 < z < 4.07. We conclude that it is at least plausible that GN20.2b could be lying at a redshift similar to that of GN20.2a. In any case, this appear to be a fairly high-redshift (z > 3.5) galaxy.

Could GN20.2b be contributing some major fraction of the submillimeter flux of GN20.2? Finding multicomponent systems is not a new situation for SMG identification. For example, SMMJ094303 (Tacconi et al. 2006) and GN19 (HDF242; Tacconi et al. 2008) are two cases where two-component systems, both radio detected, have been confirmed through CO observations. Dannerbauer et al. (2004), P06, Ivison et al. (2007), and Younger et al. (2008a) report other cases of submillimeter/millimeter galaxies with two radio counterparts. If this is also happening for GN20.2, it is not easy to assess quantitatively. Below we present circumstantial evidences in favor or against this hypothesis.

The main supporting evidence is that its VLA 1.4 GHz radio continuum flux density is 42%± 10% that of GN20, which is similar to the ratio between the submillimeter fluxes of the same two galaxies, suggesting that GN20.2b might be contributing some substantial flux, unless the radio emission is affected by an AGN. We note though that GN20.2b is 5.5 times fainter in the VLA 1.4 GHz radio continuum than GN20.2a, but this is likely due to the fact that the radio emission of GN20.2a (more than twice as bright as that of GN20) is powered by an AGN, as discussed in the following section.

There is, however, evidence against a major submillimeter flux contribution from GN20.2b. Comparing the 24 μm flux density measurements, we find that GN20.2b (only tentatively detected at 24 μm) is at least 2–3 times fainter than GN20.2a and at least 5.5 times fainter than GN20. Accounting also for blending, the S(24 μm)/S(4.5 μm) flux density ratios are at most a factor of 2 for GN20.2b and about factors of 6–7 for GN20 and GN20.2a (Table 2). This is evidence for lower SSFR in GN20.2b, suggesting that this galaxy has already passed its major peak of activity. Similarly, the ratio of the CO fluxes detected for GN20.2 and GN20.2a is roughly a factor of 2, similar to the ratio of submillimeter fluxes between the two galaxies (Table 2). Even if the tentative CO emission of GN20.2b is real, this would still be at most half of the CO emission of GN20.2a (but, of course, there could be stronger CO emission if the galaxy lies outside of the redshift range 4.02 < z < 4.07). Given that the CO and bolometric luminosities of SMGs are known to correlate (Greve et al. 2005; Solomon & Vanden Bout 2005; see also Section 4), this could suggest that GN20.2a does account for most of the submillimeter emission of GN20.2 from P06, and that GN20.2b could contribute at most some 30% of it but likely not much more. A final argument is that GN20.2 was originally discovered by Chapman et al. (2001) in photometry mode with SCUBA at JCMT, pointing at the position of the GN20.2a radio galaxy. This resulted in a submillimeter flux density of S(850 μm) = 10.2 ± 2.7 that is fully consistent with the flux in the P06 map. GN20.2b is 68 away from GN20.2a and any emission from its position would have been close to the edge of the telescope primary beam, resulting attenuated by a factor of 2. Nevertheless, the Chapman et al. (2001) flux is actually slightly brighter than the flux in P06, although consistent within the errors. Accounting for the errors in the P06 and Chapman et al. (2001) 850 μm flux density measurements, this suggests that GN20.2a does contribute at least half of the flux of GN20.2, and possibly most of it.

For completeness, we note that an additional faint radio-detected galaxy is present just 1'' north outside the edge of the SCUBA beam. This galaxy is a hard X-ray source in the catalog of Alexander et al. (2003), for which we derive a photometric redshift of z_phot = 1.6. The radio and mid-IR emission of this object, if entirely due to star formation, would correspond to about 100 M_☉ yr⁻¹. This in turn would produce a submillimeter flux of order of 1 mJy or less at 850 μm. At 8'' from the P06 position and at 14'' from the Chapman et al. (2001) pointed observations, any potential submillimeter emission from this galaxy would be negligible for the GN20.2 system 850 μm photometry.

In conclusion, our results suggest a scenario in which two z ∼ 4.05 counterparts might be contributing to the submillimeter emission of GN20.2, although GN20.2a is likely to be the dominant component.

Following the analysis described in Section 3.1.2, the radio–IR SED of GN20 is best reproduced with a CE01 library template corresponding to a luminosity of L_IR = 3.5 × 10¹³ L_☉, redshifted to z = 4.055. However, the CE01 library implies a correlation between dust temperature and luminosity that is calibrated on observations in the local universe but that might not necessarily hold at higher redshifts. In order to explore the uncertainties in the determination of GN20's total IR luminosity due to possible SED temperature variations, we again used the CE01 library, but allowing this time for a free normalization of the templates when comparing to the observed fluxes, thus spanning the full range of dust temperatures in the library. For this exercise, we used the measurements or upper limits at 24 μm, 850 μm, 3.3 mm, and 20 cm and also included the 1.1 mm and 1.2 mm measurements from Perera et al (2008) and Greve et al. (2008), and the 70 μm upper limit from Frayer et al. (2006). Integrating over the best-fitting template yields in this way L_IR = 2.4 × 10¹³ L_☉ and a formal uncertainty (1σ) from the fit of less than 0.05 dex (the luminosity step in the CE01 library). The best-fitting template has an intrinsic L_IR = 1.5 × 10¹² L_☉ and is scaled up in luminosity by a factor of 16. This would correspond to an SED temperature of about 57 K, lower than those of CE01 templates of comparable luminosities L_IR ≳ 10¹³ L_☉. The same effect can be appreciated more directly in Figure 10 where we show the 850 μm to 24 μm and radio flux density ratios. The GN20 flux density ratios are reasonably in line with CE01 predictions, although the 850 μm to radio flux density ratio is 80% higher than expected for the L_IR = 10^13.5 L_☉ galaxy in the CE01 library. The effect is significant at the 2σ level only, but indeed its SED is closer to the expected ratio for L_IR = 10¹² L_☉ expected ratio. SMGs have been generally found to be somewhat colder than CE01 models, as a result of the selection at 850 μm (P06), so this result is not too surprising.

Zoom In Zoom Out Reset image size

As a third estimate of the IR luminosity of GN20 we might use directly the radio–IR correlation. The other two estimates also used this correlation to fit the radio, although this was weighted together with all the other measurements. For its measured 1.4 GHz flux density of 75.8 μJy, assuming a radio continuum (f_ν ∝ ν^−α) with α = 0.8, we would derive L_IR = 2.8 × 10¹³ L_☉. It is not obvious which of the three estimates is more reliable but fortunately the measurements are very close, all within a range of less than 0.2 dex. In summary, we adopt the average value of L_IR = 2.9 × 10¹³ L_☉ for our best estimate and an uncertainty of 0.2 dex.

The 850 μm to VLA radio flux density ratio for GN20.2a is about 5 times lower than that for GN20 (Figure 10), implying a strong radio excess of a factor of 4–5. GN20.2a has a radio luminosity of 2 × 10²⁵ W Hz⁻¹ at 1.4 GHz rest frame, again assuming a radio continuum index α = 0.8. This demonstrates that a powerful radio-loud AGN is hidden in this z = 4.05 source. Therefore, we cannot use the 1.4 GHz flux density to estimate its total IR luminosity. We used the CE01 library templates, with and without allowing for a free normalization, to fit the 24 μm, 70 μm, 850 μm, and 3.3 mm measurements or upper limits. Similarly to GN20, we find L_IR = 2.2 × 10¹³ L_☉ and L_IR = 1.0 × 10¹³ L_☉, the latter being the "free-normalization" case that also prefers a lower temperature template. These estimates use the full 850 μm flux density measurements although some part of this might be due to GN20.2b. This also assumes that the AGN inside GN20.2a is not contributing substantially to the measured 24 μm and 850 μm flux densities. Pope et al. 2008 showed that this is generally the case for z ∼ 2–3 SMGs. Given the uncertainties, we adopt the average of these two estimates, L_IR = 1.6 × 10¹³ L_☉, and an uncertainty of about 0.3 dex.

This analysis implies that GN20 and GN20.2 are extremely luminous galaxies, with L_IR > 10¹³ L_☉. Besides being among the most distant known SMGs, these are also some of the most luminous galaxies known so far. We use a Kennicutt et al. (1998) conversion of SFR[M_☉ yr⁻¹] = L_IR[L_☉]/10¹⁰, expressed for the Chabrier 2003 IMF adopted in this paper. The IR luminosities correspond to an SFR of >1000 M_☉ yr⁻¹ if the IR emission is dominated by star formation. We note that neither the GN20 nor GN20.2a (or GN20.2b) counterparts are detected in the X-rays in the catalog of Alexander et al. (2003), although the radio emission does suggest the presence of an AGN inside GN20.2a. There is no clear evidence, instead, for the presence of an AGN inside GN20. If a top-heavy IMF is adopted, the implied SFR could also become significantly smaller.

The observed CO[4–3] fluxes convert to luminosities of L'_CO[4-3] = 6.2 × 10¹⁰ L_☉ and L'_CO[4-3] = 3.7 × 10¹⁰ L_☉, for GN20 and GN20.2, respectively. Figure 11 shows that these galaxies appear to lie close to the (nonlinear) correlation between L'_CO and L_IR traced by local ULIRGs and distant SMGs. Under the assumption that the CO[1–0] and CO[4–3] transitions correspond to the same brightness temperature and assuming X_CO = 0.8M_☉ (K km s⁻¹ pc²)⁻¹ (Downes & Solomon 1998; Solomon & Vanden Bout 2005), the CO luminosities convert to total molecular gas masses of 5 × 10¹⁰ M_☉ and 3 × 10¹⁰ M_☉ for GN20 and GN20.2a, respectively. These might be regarded as lower limits if the CO[4–3] transitions observed were not thermalized, the total CO luminosities and gas masses would be higher. VLA observations of CO[1–0] will be able to address this point.

Zoom In Zoom Out Reset image size

At the observed SFR levels, the gas reservoirs would be exhausted in roughly 20–30 Myr (Figure 11). Similar timescales have been derived for z ∼ 2–3 SMGs by Greve et al. (2005).

The CO[4–3] emission from GN20 has an FWHM of 710 ± 120 km s⁻¹, matching well to the typical value of CO-detected SMGs at lower redshifts (Greve et al. 2005). From the higher-resolution B configuration PdBI data alone we find that the CO line is marginally resolved spatially, with an FWHM of 042 ± 018 from Gaussian fitting of the uv visibilities performed within GILDAS. The possible extension of the source is supported by the fact that for point source extractions the flux measured with the D-configuration data is higher than the one from the B-configuration data, a 2σ effect suggesting that the CO emission is starting to be resolved at the 13 resolution of our B-configuration data. This size matches well to the typical CO sizes of r_e ∼ 2 kpc for SMGs (Tacconi et al. 2006, 2008), and is roughly consistent with its UV extent as well (Figure 5). Recently, Younger et al. (2008b) measured a size of 06 ± 015 for GN20 (Gaussian fitting), from the far-IR dust continuum emission at 890 μm with the SMA, which is fully consistent with our CO size estimate. The velocity and size correspond to a total dynamical mass of ∼2.3 × 10¹¹ sin i⁻² M_☉ including gas, stars, and dark matter.

Given the molecular gas mass estimates, this corresponds to a molecular gas fraction of about 20%, similar to what is typically found for lower redshift SMGs (Greve et al. 2005). The stellar mass derived from SED fitting of the ACS to IRAC photometry is ∼2.3 × 10¹¹ M_☉, expressed for a Chabrier (2003) IMF (see e.g., Maraston et al. 2006 for more details on the range of models and assumptions implied). The combined molecular gas and stellar masses, only roughly half of which presumably fall within r_e, add up to roughly 70%–80% of the estimated dynamical mass, close enough within the large uncertainties of such estimates. This implies that the stellar mass is not highly overestimated for the choice of a Chabrier (2003) IMF and the case of a top-heavy IMF is disfavored by the data.

For GN20.2a, using a similar method we infer a stellar mass of ∼0.5 × 10¹¹ M_☉.

The intrinsic SFRs and luminosities of these sources might be lower if there is amplification by gravitational lensing (see Paciga & Scott 2009). Lensing by foreground structure could be an alternative way to interpret the fact that colder SEDs, typical of lower luminosity galaxies in the local universe, seem to reproduce better colors of the GN20 and GN20.2 galaxies. In that case, the luminosity would also match those of colder galaxies without the need to advocate evolution in the temperature–luminosity relation of starburst galaxies. However, some evidence against lensing is provided by the FIR to CO properties of GN20 and GN20.2 (Figure 11). If lensed by a factor of 10 or more, the de-lensed properties would place them in a region of the diagram corresponding to a very high star formation efficiency, not commonly observed for L_IR ∼ 10¹² L_☉ sources at high or low redshifts. Following the analysis discussed in Tacconi et al. (2008), the agreement between the dynamical mass estimate and the stellar and gas mass estimates also disfavors the possibility that GN20 is lensed by a very large factor, given that emission line FWHMs should be independent of magnification. Qualitatively, the same picture also holds for GN20.2a, for which the data suggest a CO[4–3] line FWHM ∼1000 km s⁻¹ (albeit with a large uncertainty), not uncommon amongst SMGs (Greve et al. 2005).

A tight correlation exists between the SFR and stellar mass of star-forming galaxies from z = 0 to z ∼ 2, after excluding the locus of quiescent/passive galaxies (Elbaz et al. 2007; Daddi et al. 2007a; Noeske et al. 2007). The implication of this correlation remains subject to debate but suggests that, on average, the processes regulating the star formation activity of a galaxy are the same over a large range of stellar masses. Outliers do exist at both extremes in this correlation. On the low SFR side, the cloud of red-dead galaxies have SFRs lower than expected by several orders of magnitudes. On the high SFR side, ULIRGs and z ∼ 2–3 SMGs exhibit approximately 1 order of magnitude larger SFRs for their stellar mass than the average galaxy of equivalent stellar mass (Daddi et al. 2007a; Elbaz et al. 2007). Here, following Daddi et al 2007a, we use the Tacconi et al. 2008 estimate of SSFR for z = 2–3 SMGs that is based on dynamical masses, which should more robustly represent the total galaxy mass than is possible from stellar mass estimates. We converted the average L_IR = 10^13.1 L_☉ of the sample into SFR ∼ 1200 M_☉ yr⁻¹. Takagi et al. (2008) confirm higher SSFRs in SMGs, with respect to normal galaxies of the same masses, using stellar mass estimates, although they find SSFRs that are only a factor of 3 higher on average than those of the ordinary BzK galaxies, with more overlap between samples (see also Dannerbauer et al. 2006 for overlapping objects between the BzK and SMG samples). Their SFRs are based on the radio fluxes and assume a radio–IR correlation with a different normalization respect to the local one by a factor of about 2.

On the low side, the red-dead galaxies appear to fall short of SFR because of a lack of molecular gas fuel, possibly due to negative feedback. However, the excess SFR of z ∼ 2–3 SMGs remains to be understood. A possible explanation is external triggering, e.g., by major mergers (Tacconi et al. 2006; 2008) as is seen for local ULIRGs.

When compared to their closer z ∼ 2–3 siblings, z ∼ 4 SMGs appear to be forming stars with an equivalently high SSFR. However, the stellar mass–SFR relation shows a continuous increase of SSFR with increasing redshift from z = 0 to 2, which is expected due to the higher gas mass fractions at higher redshifts. The question therefore arises as to whether an SMG could represent the typical star-forming galaxy at z ∼ 4, unlike their z ∼ 2 siblings which are atypical compared to the average galaxy at their cosmic epoch (see Figure 12).

Zoom In Zoom Out Reset image size

In order to address this question, one needs to identify the locus of average star-forming galaxies at z ∼ 4 and then compare it to the locus of z ∼ 4 SMGs. We consider the subsample of B-band dropout Lyman break galaxies in GOODS-N with IRAC magnitudes at 5.8 and 8.0 μm brighter than 24. The requirement of an IRAC detection is necessary to reliably estimate the stellar masses. The use of B-band dropouts ensures that we are concentrating on star-forming galaxies, with a negligible contribution from passive galaxies. Within this limit, about 10% of the B-band dropouts to z_AB < 27 are selected, or 77 sources. We fitted the ACS to IRAC SEDs of these B-band dropouts using Maraston (2005) models with constant star formation, allowing the redshift to vary freely within 3.1 < z < 4.4 (the 95% confidence region expected for the redshift distribution of B-band dropouts) and also allowing for dust reddening following a Calzetti et al. (2000) law. In this way, we estimated stellar masses and (basically UV-driven) SFRs for each of the B-band dropout Lyman break galaxies. Typical uncertainties from the fit amount to factors of about 2 for both quantities. From this we derive an average stellar mass of 1.5 × 10¹⁰ M_☉ and an average SFR of 40 M_☉ yr⁻¹. We performed stacking at 850 μm using the SCUBA supermap of Pope et al. (2005) and with a similar technique to that used in Daddi et al. (2005b, 2007a). We do not detect the galaxies, deriving a 3σ flux density limit of 0.54 mJy, corresponding to 57 M_☉ yr⁻¹. From radio stacking, again performed with a similar technique to Daddi et al. (2005b, 2007a), we get a tentative detection with a peak flux of 1.7 ± 0.6μJy. The 3σ radio upper limit corresponds to an SFR of 54 M_☉ yr⁻¹, consistent with the limit from submillimeter stacking. Overall, the stacking limits appear to be consistent with the SFRs derived from SED fitting in these galaxies. This is similar to what found at z ∼ 2 where analogous estimates based on the UV luminosity of galaxies appear to agree well on average with mid- and far-IR estimates (Daddi et al. 2007a).

The locus of typical, massive z ∼ 4B-band dropouts does not support a continuous increase of SSFR with redshift, suggesting instead a plateau of the SSFR for z ≳ 2. The average SFR of 40 M_☉ yr⁻¹ places these "typical" z ∼ 4 galaxies fairly close to the z ∼ 2 correlation.

The B-band dropout Lyman break galaxies exhibit an average stellar mass lower than that of the z ∼ 4 SMGs. This implies that not only do z ∼ 4 SMGs have similar SSFRs to those of their z ∼ 2 siblings, but they are also found to be outliers with respect to the average star-forming galaxy at their epoch. Note that the similarity of z = 2–3 and z = 4 SMGs is also supported by their equivalently large FIR to CO ratios, corresponding to high star formation efficiencies and rapid gas consumption timescales.

We hypothesize that this favors a triggering mechanism for the activity of the z ∼ 4 SMGs, such as major mergers. Indeed, GN20.2a has a very nearby (05 or 3.5 proper kpc at z = 4.05) B-band dropout companion (Figure 5). GN20 has a clumpy morphology that could be reminiscent of a merger (although, see Bournaud et al. 2008). The 05 offset between the optical light peak and submillimeter/radio and CO position might also point to a complex geometrical situation often found among mergers (e.g., the Antennae; Wang et al. 2004). The high gas densities of a merging event could also explain why, contrary to what is found for typical massive galaxies at 1.4 < z < 2.5 (Daddi et al. 2007a; Dannerbauer et al. 2006), the emitting region is completely opaque in the UV, similar to z ∼ 2–3 SMGs. Indeed, we find that the radio–IR-based SFR estimates of the z ∼ 4 SMGs are a factor of >10 times larger than the UV-based ones.

For the most reliable z ≳ 4 starbursts (GN10, GN20, GN20.2a) we note that all of these have very high 20 cm to 24 μm flux density ratios, greater than ∼ 1 (Figure 15). GN20.2b itself satisfies such a criterion and might also be at z = 4.05. This suggests that a criterion S_1.4 GHz ≳ S_24 μm could be useful at singling out the very high redshift tail of SMGs. Applying this criterion to the P06 sample we recover the SMGs listed above, plus GN12, GN16 and GN19a, all sources at 2.5 < z ≲ 3.2, a distribution clearly skewed to higher redshifts than typical SMGs. The z = 4.547 galaxy from Capak et al. (2008) also satisfies this criterion.

In principle, this simple criterion could be applied regardless of a submillimeter detection, and could avoid the inherent bias toward intrinsically cold sources. Of course, care should be taken and a careful analysis of all multiwavelength information should be performed on targets selected in this way, as radio emission from radio-loud AGNs would bias the radio to 24 μm flux density ratio. Mid-IR emission can also be boosted by AGNs, which could make genuine z > 4 objects fail this criterion. More importantly, given the relatively small dynamic range expected over redshift for the 20 cm to 24 μm flux density ratios in galaxies (Figure 15), scatter due to SED variations can imply a substantial contamination from lower redshift sources.

We have performed a preliminary test of this criterion using a sample of radio selected galaxies with S_1.4 GHz > 30μJy in GOODS-N (G. Morrison et al. 2009, in preparation), applying a simple but very conservative approach. By requiring S_1.4 GHz > 0.8 × S_24 μm, 63 radio sources are retained. Only half of these (34) are undetected in the ACS F435W band (S/N < 2), a necessary condition for being at z ≳ 4. Of these, 16 are also consistent with having a red IRAC SED, with a peak beyond 7 μm, as expected for z ≳ 4.¹⁶ We further require relatively blue Spitzer+MIPS to IRAC flux density ratios: S_24 μm/S_4.5 μm < 13 and S_24 μm/S_8.0 μm < 5, as observed in the bona-fide z ≳ 4 sample (GN20, GN20.2a, GN20.2b, GN10), leaving us with 11 galaxies. This excludes from the sample the z = 4.424 radio source from Waddington et al. (1999), where the near-IR to mid-IR spectral range is likely dominated by a powerful AGN. Finally, we exclude all sources detected in the X-ray catalog of Alexander et al. (2003) which, if at z ≳ 4, must be AGN dominated. This leaves us with seven sources, including the three reliable z > 4 galaxies already known (GN20, GN20.2a, GN10) plus GN20.2b. We consider the properties of the three new candidates in detail.

• 1.  
  VLA 512: this object is undetected in a relatively shallow region of the SCUBA supermap. However, some signal is found at its position with S_1.1 mm = 3.2 mJy (S/N = 3.4) in an AzTEC map of GOODS-N (G. Wilson, private communication). This source has no counterpart at any ACS band, and is a good high-redshift starburst candidate.
• 2.  
  VLA 812: this galaxy is very close to the edge of the SCUBA supermap, where the noise is fairly high. However, this source is listed in Chapman et al. (2005) as a SCUBA detection in photometry mode (S_850 μm = 8.8 ± 2.1 mJy), with a redshift of z = 2.098. The source is also detected by Perera et al. 2008 and Greve et al. 2008 at 1.1 mm and 1.2 mm with MAMBO and AzTEC, respectively. We find that a relatively bright z = 2.098 galaxy is present in the area, but 2'' away from the IRAC and radio position, which corresponds instead to a faint galaxy with ACS photometry of i_AB = 27.6 and z_AB = 27.3, undetected in the B and V bands. We suggest that the Chapman et al. (2005) redshift identification does not correspond to the actual counterpart of the radio and submillimeter emission. This radio source is a reliable candidate for a very high redshift starburst (z ≳ 4).
• 3.  
  VLA 966: we place a 3σ upper limit of S_850 μm < 8.1 mJy from the SCUBA supermap at the position of this source. Its optical counterpart is a B-band dropout with ACS magnitude z₈₅₀ = 26.6 and red IRAC SED, confirming that this is a reliable z ∼ 4 candidate. The VLA radio flux density of this source is comparable to that of GN20. The lack of submillimeter emission could signal a warmer SED (see e.g., Chapman et al. 2008).

We note that, due to the very conservative requirements applied, we finally retained only three galaxies of the 63 selected down to 30 μJy in GOODS-N with our radio/24 μm criterion as bona-fide z ≳ 3.5 starburst galaxy candidates. We believe that we efficiently rejected in this way all contamination from AGN dominated radio galaxies. Several of the discarded galaxies are most likely to be starbursts at 2.5 ≲ z ≲ 3.5, still very interesting objects but beyond the scope of the simple exercise described in this Section where we focus on the z > 3.5 redshift tail. We are probably also missing several other genuine z ≳ 3.5 starbursts due to the very strict requirements that we have applied. For example, we required a distinct photometric peak in IRAC beyond 7 μm, but this can be weak or washed out due to the S/N level of the IRAC data. We required X-ray nondetections, but powerful starbursts might also contain an unobscured AGN that does not necessarily affect the radio emission to a significant degree. Given the promising results of this attempt, we will defer to future work an investigation of the properties of the full sample of galaxies selected with this radio/24 μm criterion.

Although we by no means claim to have selected a complete sample, the existence of these three additional z > 3.5 starburst candidates reinforces our conclusion that a substantial population of vigorous star-forming galaxies was already present at early epochs.

We now consider expectations for forthcoming Herschel surveys to detect very high redshift starbursts. Herschel will obtain FIR imaging with an unprecedented sensitivity at 70 μm, 100 μm and 160 μm with PACS (Poglitsch et al. 2006) and at 250 μm, 350 μm and 520 μm with SPIRE (Griffin et al. 2007).

We computed the expected flux densities of GN20 in the Herschel bands, based on various assumptions. Using the best-fitting z = 4.055 unscaled CE01 model shown in Figure 8, we predict flux densities of about 40–50 mJy in the SPIRE bands, 24 mJy at 160 μm, 4.8 mJy at 100 μm and 1.8 mJy at 70 μm. The latter value is comparable, but still consistent, with the Spitzer 70 μm upper limit. If these predictions are correct, galaxies similar to GN20 should be very easily detectable with Herschel, e.g., over the full 70 deg² survey of the Hermes guaranteed time consortium,¹⁷ which could yield hundreds z > 4 hyper-luminous infrared galaxies. Even if we used a colder template following Figure 10, normalized conservatively at a 1σ lower 850 μm flux density of 17 mJy, we would still predict flux densities of about 30 mJy at 350 μm and 520 μm, which is still above the expected flux limit of the 70 deg² Hermes survey.

This suggests that, while strongly limited by confusion to quite bright flux densities (∼15 mJy; Le Borgne et al. 2009), the Herschel SPIRE observations will efficiently detect a significant tail of very high redshift (z > 4) starbursts. Due to the relation between redshift and 850 μm flux density (Ivison et al. 2002; Pope et al. 2006), Herschel might detect the bulk of the most active merger-powered galaxies in the distant universe, possibly accounting for an important fraction of the SFRD in the early universe. Clearly, a real challenge will be to recognize such very high redshift galaxies in the Herschel-selected samples, which will require deep multi-wavelength information.