© ESO, 2017

1. Introduction

Deep multi-wavelength imaging of the sky has provided a tremendous amount of information on galaxies in the distant Universe since the Hubble Deep Field North was published (Williams et al. 1996). The tools to efficiently and accurately exploit these data by fitting their spectral energy distributions (SEDs) have also evolved in step and have now reached a fairly high level of maturity (see Conroy 2013; Walcher et al. 2010, for reviews).

The development of photometric redshift (photo-z) estimation techniques has been particularly notable. The number of objects with photometric information is much larger than can be efficiently followed-up spectroscopically and this means that large-scale multi-band surveys of the sky have to rely on photo-zs to determine distances to galaxies. This central role for photo-zs has also led to the development of a wide range of techniques for photo-z estimation. These fall basically into two categories: machine learning techniques which aim to empirically determine the map between colours and redshift, and the template fitting techniques which take a set of physically motivated SEDs and find the best match of a (combination of) these SEDs to the data. An up-to-date overview of the photo-z methods can be found in the introduction of Sadeh et al. (2016) and more extensive comparisons of codes can be found in for instance Hildebrandt et al. (2010), Abdalla et al. (2011), Acquaviva et al. (2015).

The requirements on photometric redshifts from cosmological weak lensing survey such as the Kilo Degree Survey (KiDS; Hildebrandt et al. 2017), the Dark Energy Survey (DES; the Dark Energy Collaboration 2005), and the Large Synoptic Survey Telescope (LSST, Ivezic et al. 2008) ground-based surveys, and the Euclid (Laureijs et al. 2011) and Wide-Field InfraRed Survey Telescope (WFIRST, Spergel et al. 2015) space missions, are stringent. The requirement on individual redshift of σ_pz< 0.05(1 + z) is non-trivial but this requirement is not the biggest challenge. The preferential way to carry out the weak lensing surveys is to do this in redshift bins which leads to strict constraints on the accuracy of the mean redshift in each bin. In the case of future surveys, the mean redshift must be constrained to better than 2 × 10^-3 (1 + z) which is a very challenging requirement for future surveys (e.g. Newman et al. 2015).

As a consequence of these needs, several studies have explored the performance of different photo-z codes on data appropriate for cosmological studies (e.g. Hildebrandt et al. 2008, 2010; Abdalla et al. 2011; Bonnett et al. 2016; Beck et al. 2017). These studies typically find that the required constraints on individual photo-z estimates of σ_pz< 0.05(1 + z) is an achievable goal for the large missions. The more stringent constraint (e.g. Zhan 2006) is however on the bias of the mean redshifts in a particular redshift bin, which must be < 2 × 10^-3(1 + z) to reach the goals of the upcoming surveys.

The majority of the photometric redshift tests have focused on relatively bright galaxies (i_AB or r_AB< 24) since those are the galaxies targeted by weak lensing surveys, but also because this is typically the magnitude limit to which most spectroscopic surveys target galaxies. Among the deepest comparisons to date is the study by Dahlen et al. (2013) of photo-zs in the Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey (CANDELS) which focuses on the GOODS-S field, the PHAT1 photo-z accuracy test by Hildebrandt et al. (2010) which uses data from the GOODS-N field. Both of these studies tested multiple photometric redshift codes on deep HST data, with spectroscopic reference samples mostly extending to F160W = 24 with a small tail extending to fainter magnitudes. The more recent study by Dahlen et al finds a mean bias of ⟨ (z_spec−z_phot) / (1 + z_spec) ⟩ = −0.008 and an outlier fraction of 2.9% when combining all photo-z considered, while the rates for the best indicator in the PHAT1 test gave a bias of 0.009 and an outlier fraction of 4.5% for the R< 24 subsample. While these quantities vary from study to study and between photometric redshift estimators, outlier fractions well below 10% and biases < 0.01 are frequently seen. In these cases, it is reasonable to use photometric redshifts to study trends of the galaxy population as the mean predicted properties are not likely to be strongly influenced by the errors in the photo-zs. However, it is worth noting that the results in Hildebrandt et al. (2010) clearly improve when limiting the study to R< 24, and Dahlen et al. (2013) show a clearly degraded performance of photometric redshifts when artifically dimming their spec-z sample, but as this was not a real test against spectra of faint objects, its relevance to real data is harder to assess.

The dependence of photo-z performance on magnitude, also primarily down to JH_IR = 24, was explored in detail by Bezanson et al. (2016). Those authors compared 3D-HST grism redshifts (Momcheva et al. 2016) to photometric redshifts from Skelton et al. (2014). This is a different kind of test since grism redshifts can vary significantly in accuracy and also depend on the photometric information as discussed in Bezanson et al., but a clear advantage is that the grism redshifts are available in a fairly unbiased way across the galaxy population. They focus on the scatter in the redshift comparison as well as the outlier fraction, and find comparable outlier fractions (1.9−4.9%) to the studies cited above. They also show that the scatter between photometric and grism redshifts increases strongly towards fainter magnitudes and were able to use the test of photo-z accuracy using galaxy pairs developed by Quadri & Williams (2010) to show that this extends to F160W = 26.

The decreasing performance of photo-zs at faint magnitudes highlights an important point that is well-known but often not stated explicitly: the quality of photometric redshifts depends on the quality of the photometry. When it is stated that photo-zs are improved by adding filter X, then it is implicitly assumed that the photometric quality of that band is high. When this is not the case, adding this band might in fact decrease photo-z performance. While not a key point of the present paper, we discuss this in the context of adding IRAC photometry in Appendix A where we will show that for faint galaxies adding IRAC photometry worsens photo-z performance.

In contrast to these preceding papers, here we focus on the performance of photometric redshifts for faint galaxies which have spectroscopic redshifts from the Multi-Unit Spectroscopic Explorer (MUSE) instrument (Bacon et al. 2010). Since we are exploring to fainter magnitudes than normally used for machine learning techniques for photo-z estimation we ignore these completely (see Sadeh et al. 2016, for a discussion). We also do not explore the full range of template fitting methods, thus widely used codes, such as Le PHARE (Arnouts et al. 1999; Ilbert et al. 2006), ZEBRA Feldmann et al. (2006), and HyperZ (Bolzonella et al. 2000) are not discussed further. This is because our aim is somewhat different from the recent tests of photo-zs for cosmological surveys. We wish to present a first exploration of the performance of the codes at magnitudes F775W> 24 so we are limiting our attention to codes that have already been run on the Rafelski et al. (2015, R15 hereafter) catalogue which was the starting point for our source extractions. This amounts to two established template fitting codes: BPZ (Benítez 2000) and EAZY (Brammer et al. 2008), as well as the new, fully Bayesian fitting code BEAGLE (Chevallard & Charlot 2016). The advantage of template fitting codes over the machine learning methods that might be preferred for cosmological applications (e.g. Newman et al. 2015), is that they provide the user with the possibility to also extract physical parameters for the galaxies in question. Indeed, the main difference between BEAGLE relative to BPZ and EAZY is the fact that it was optimised for physical parameter estimation rather than photometric redshifts − making a comparison of the different codes particularly interesting.

In Sect. 2 below we provide a brief discussion of the data used in the study. The reduction and redshift analysis of these are described in detail in Inami et al. (2017,Paper II hereafter). In Appendix A we justify our focus on the 11 band HST-only catalogue from R15 by showing that for our faint galaxies, the use of the 11 HST bands only results in better photo-zs than does the use of 44 bands from Skelton et al. (2014). In Sect. 3 we compare our spectroscopic redshifts to the photometric ones. We will find that the EAZY photo-zs in the R15 catalogue are surprisingly discrepant and re-calibrate these in Sect. 4. The nature of the galaxies for which the photometric and spectroscopic redshifts are clearly discrepant is discussed in Sect. 5 with some complementary information in Appendix B. The effect of redshift incompleteness is discussed in Sect. 6. We explore the impact of object superpositions on photometric redshift estimators, including machine learning ones, in Sect. 7. We discuss our results and conclude in Sects. 8 and 9.

2. Data

The data used in the present paper come from MUSE guaranteed time observing (GTO) observations of a 3′ × 3′ field of the Hubble Ultra Deep Field (UDF, Beckwith et al. 2006). This amounts to a total integration time of 116 hours in the autumns of 2014 and 2015. The details of the survey strategy and data reduction are given in Bacon et al. (2017,Paper I hereafter). For the present paper the most imporant aspect of the observations is that they were carried out to two different depths: a 3′ × 3′ medium deep field, mosaic, with an effective integration time of approximately 10 h, and nested within this a 1′ × 1′ ultra-deep field, udf-10, which contains data with a total of ≈ 31 h of integration time. The data reduction followed broadly the process used for the Hubble Deep Field South (HDF-S) observations described in Bacon et al. (2015) with several enhancements. The main improvements were to the self-calibration procedure which now uses a polychromatic correction and works directly on the pixtables output by the MUSE data reduction pipeline Weilbacher et al. (DRS, 2012). Inter-stack defects were removed by using a bad pixel mask projected onto the 3D cube, see Paper I for details, and a realistic variance cube was derived as described there. The resulting data cube is considerably better calibrated spectrally and spatially than the one produced for the HDF-S used in Bacon et al. (2015).

We will also make use of the photometric information from the R15 catalogue. This has photometric measurements in up to 11 bands. F225W, F275W, and F336W from Teplitz et al. (2013), F435W, F606W, F775W, and F850LP from (Beckwith et al. 2006) and finally F105W, F125W, F140W, and F160W mostly from the UDF09 and UDF12 programs (Bouwens et al. 2011; Oesch et al. 2010; Koekemoer et al. 2013; Ellis et al. 2013) with shallower F105W, F125W, and F160W data from Koekemoer et al. (2011) and Grogin et al. (2011). In addition to providing photometry, R15 also ran the BPZ and EAZY photo-z codes on their catalogue. We here use the results as reported by R15, the interested reader can consult their paper for details of how the codes were run and the templates used. We will refer to the IDs from this paper as for example RAF 4471, where the number is the ID number in the R15 catalogue.

We do not seek to add further photometric information to the R15 catalogue. The comprehensive compilation of photometry for 3D-HST by Skelton et al. (2014, S14 hereafter) does provide a total of 44 bands for our field, but it has photometry for fewer objects with z_MUSE than the R15 catalogue (see Fig. A.1), and it was not used for the object definition and spectrum extraction. This means that the association of spectroscopic redshifts with photometric object is less secure. Finally, in Appendix A we show that although the large number of bands might lead to better performing photo-zs at bright magnitudes and an overall smaller bias, the S14 catalogue leads to a higher number of outliers at faint magnitudes than the R15 catalogue. We therefore do not use these data here, although we will return to discuss the performance of EAZY run on these data in Sect. 8.

2.1. Object definitions and spectrum extraction

Paper I explains the method for object detection and the spectrum extraction in described in detail in Paper II. In short we use two distinct approaches to define objects. The first takes the existing segmentation map for the HST catalogue of Rafelski et al. (2015, R15 hereafter) and convolves this with the MUSE Point Spread Function (PSF) to get a segmentation map suitable for MUSE. Following Paper II we refer to these as continuum selected sources. The second approach uses the matched filter detection method ORIGIN (Mary et al. in prep.; see Paper II for an overview of the process) to find emission line sources in the cube and we will refer to these as emission line selected objects. These two approaches provide partially overlapping object lists and these have been consolidated by inspection case-by-case.

Both approaches produce a mask defining the spatial extent of a source and the resulting object masks have then been used to extract spectra using both a straight sum and a weighted extraction. The redshifts used in the present paper were obtained from the higher signal-to-noise weighted extractions. For objects with a full-width half-maximum size in the HST F775W image the white-light image of the object was used as a weight, while for smaller objects the estimated PSF as a function of wavelength was used as a weight. The process of spectrum extraction is described further in Paper II but has no impact on the results presented here.

2.2. Redshift determinations

The process of redshift determination is detailed in Paper II, but it is important for the present paper to summarise the steps. The redshift determination for the continuum selected objects was done in a semi-automatic manner using a modified version of the MARZ redshift determination software (Hinton et al. 2016). The software provides redshift estimates using cross-correlation with a set of templates. The redshift solutions are visually inspected by at least three researchers. The inspection step looks both at the 1D spectrum, as well as narrow-band images over the putative spectral features. For the emission line selected objects the main challenge is to identify which line is detected and this is done by two researchers for both the udf-10 and the mosaic.

For each redshift determination we assigned a confidence, where confidence 3 corresponds to a secure redshift, determined by multiple features, confidence 2 to a secure redshift, determined by a single feature (frequently Ly-α where the asymmetry of the line in obvious), and confidence 1 are considered possible redshifts which are determined by a single feature with uncertain identification. We will almost exclusively show results based on spectra with confidence ≥ 2.

For the present paper the number of sources is dominated by the UDF mosaic, but we also use the deeper udf-10 data. The process to determine redshifts was slightly different in these two fields. In the udf-10 the spectra of all objects from the R15 catalogue were inspected visually to attempt to determine redshifts, whereas in the UDF mosaic only objects with F775W< 27 were inspected visually. The ORIGIN code was run in both fields. This mixture of selection criteria complicates the selection function of the sample but this does in general not affect our results significantly. One consequence is however the fact that at F775W> 27, we have mostly included objects that have been detected by blind emission line detection codes directly in the cube with no recourse to the HST images (see Paper I for details). The exception is the udf-10 and a few “split” objects (see Paper II for details).

One aspect of this process is important to underline for this paper: the redshifts were initially determined without any knowledge of the photometric redshift of the objects. The HST images of the source were used for instance to help the identification of lines but not beyond that. In a later update to the catalogue in Paper II photo-zs were occasionally consulted, although they were not directly used to determine redshifts. To avoid any possible biases caused by this we do not use these updated redshifts here. The catalogue used here therefore differs slightly from that of Paper II.

In order to compare the results below with other results in the literature, it is also important to have an understanding of the type of objects for which we have redshifts. Since MUSE obtains spectra without photometric pre-selection, we are not biased towards continuum bright objects, and a consequence of this is that most redshifts are determined on the basis of emission lines. This selection leads to a different mix of spectral types than commonly seen in magnitude limited spectroscopic surveys.

The catalogue discussed in Paper II has 1,329 spectra with redshift > 0 and a confidence ≥ 2. Out of these 1329 redshifts only 63 (4.5%) are determined purely on the basis of absorption lines, with the faintest such galaxy having F775 = 26.2. This is however a strong function of magnitude, reaching ≈ 20% near F775W = 24. It should also be noted that most of these absorption line galaxies are expected to be star-forming as they are bright in the UV; there are only 11 sources with absorption line redshifts at z< 1.5 that are likely to be passive galaxies, most of which are at F775W< 22. The remainder of the spectra have one or more strong emission lines. The majority of redshifts are determined on the basis of Ly-α and [Oii] λ3727, with most F775W> 27 sources with secure redshifts being Ly-α-emitters. A detailed breakdown of the types of emission lines sources is given in Table 2 in Paper II but for the present paper this is of little relevance.

The overall redshift completeness is > 50% at F775W< 25.5 (Paper II) but fainter than this we lose absorption line galaxies. In Fig. 1 we show the spectroscopic completeness as a function of observed F775W−F160W colour where we have only included galaxies with CONFID ≥ 2. To put this in context, a single stellar population from Bruzual & Charlot (2003) with > 20% solar metallicity forming at z = 10 will have F775W−F160W> 2 for z> 1.5. Thus the grey histogram that is inset in the figure which shows the number of objects in the R15 catalogue at each colour, shows that very few truly red sources are in the parent catalogue. Each line shows the completeness in bins of F775W as indicated by the labels on the left of the lines. It is clear that down to F775W = 24 our spectroscopic sample is complete but going fainter the incompleteness increases with a clear colour-dependence. The consequence for the following is that we will mostly test the performance of photometric redshifts on star forming galaxies and we will not be able to say anything about the performance of the codes on passive or very dusty galaxies. We explore the effects of the spectroscopic incompleteness in Sect. 6 below.

2.3. Resolving blends and final sample definition

In the present paper our focus is on determining the redshifts of objects detected in the HST images to be able to compare with the photometric redshifts of these sources. The procedure outlined above is not specifically designed for this as the object identification step convolves the HST detection mask with the MUSE PSF. Hence, although we base ourselves on the catalogue in Paper II, we spent considerable care on establishing the association with objects in the HST images defined in the R15 catalogue.

While the association of spectroscopic features to a particular photometric object is in most cases straightforward, in a small subset this is not so. These have been inspected in detail, using the HST imaging to help resolve the assocation in a number of cases. To do this, a combination of the centroid of the narrow band images over the main spectroscopic features in a spectrum and the visibility of the object in different HST bands was used.

Figure 2 shows an example of this process, see also Paper II for more discussion of the redshift determination process. In a first iteration two Ly-α lines were identified based on their characteristic line shapes and coherent narrow-band images with clearly different centroids. The source mask was adjusted to create two MUSE objects (MUSE IDs 3052 and 7053 at z = 3.71 and z = 3.55). However the association of a z = 3.71 redshift yet a very clear detection in F435W was problematic, while for z = 3.55 it is still feasible. Subsequent examination then showed that the Ly-α at z = 3.71 appears to be associated with a small point source just below the central galaxy (see the blue arrows in the HST images in Fig. 2) which disappears between F775W and F435W. Re-examination of the spectrum also showed tentative Hα and Hβ at z = 0.24. In this case, then, the single HST object RAF 4919 is a blend of a z = 3.71 Ly-α-emitter and a star-forming z = 0.24 galaxy. These are not easily separated at ground-based resolution, although their narrow-band image centroids are slightly offset (second and third image from the left in the lower panel). The segmentation map is shown as the black.

In most cases this careful examination leads to a unique association of an emission line to an HST object, but for a total of 58 MUSE objects this has proven impossible with the existing data so they have multiple RAF IDs associated to them. For these objects we have a redshift for the spectrum but we are unable to ascertain which of the HST objects the spectroscopic feature pertains to. An example of the latter case is MUSE ID 2277, shown in Fig. 3. The MUSE spectrum shows a strong Ly-α line but there is no easy way to determine which object this line belongs to (indeed it could be associated with both). We refer to these as blended objects, and they are, unless otherwise stated, excluded from the plots that follow. However, it is important to realise that in a study without the exquisite HST images available in the UDF, these subtleties might not be noticed. This has important consequences for the accuracy with which we can assign a redshift to an object and we explore this further in Sect. 7 below.

The end product of this procedure is a catalogue with 182 objects with spectroscopic redshift confidence (defined in Paper II) of 1602 with confidence 2 and 557 with confidence 3 (the most secure redshift), adding up to a total of 1341 HST objects with redshifts. This can be contrasted with the R15 catalogue in the same spatial region which has a total of 6362 objects brigther than F775W = 30 and 1181 objects with F775W< 27. There are in total 126 HST objects for where the redshift is clear but the association of the redshift to the object is impossible due to blending, corresponding to 58 MUSE sources. This can be compared with the 160 spectroscopic redshifts known in this area before – an increase of a factor of 8.2.

3. Comparison to photometric redshift estimates

We have opted to focus our comparison on photometric redshift estimates for the UDF data from the recent literature. R15 provide photometric redshift estimates for their detected objects, one using the Bayesian Photometric redshift code BPZ (Benítez 2000) and one using the EAZY code described by Brammer et al. (2008), in addition to these, we will also test the recent photo-z estimates from the BEAGLE code (Chevallard & Charlot 2016). We have verified that the results below are also found when using the photo-zs from S14 which we discuss in more detail in Appendix A.

The main difference from previous comparisons in the GOODS-S area is that we now can test these predictions outside the magnitude range where they have been validated earlier. As we will show shortly, the base photometric redshift estimates in the R15 catalogue and from BEALGE are systematically biased at faint magnitudes/high redshift, and the EAZY estimates in particular are significantly off. However we will also see that we can improve the situation for EAZY considerably by adopting other spectral templates.

3.1. Sample definitions

We are interested in characterising how well photometric redshifts work, including at very faint limits. As outlined above, this presents some challenges for the association between spectroscopic redshift and photometric counter-part. Firstly we will only consider galaxies for which we have a clear HST counterpart, so we require that they are not flagged as blended. Our spectroscopic redshift determinations are considerably less accurate for galaxies with confidence 1. In the main, therefore, we will limit our studies to galaxies with spectroscopic redshift confidence ≥ 2. With these cuts we are confident that the spectroscopic redshifts are correct and we expect that any discrepancy with photometric redshifts is due to an incorrect photometric redshift estimate. We will examine this assumption and the impact of relaxing the confidence requirement in the next section when we discuss the results.

We will adopt a notation throughout where we denote photometric redshifts as pz and spectroscopic redshifts from MUSE as z_MUSE. The difference between the spectroscopic and photometric redshift is denoted Δz and is defined as (1)throughout, and we will frequently follow the literature and normalise this by 1 + z_MUSE.

Figure 4 shows the difference between the spectroscopic and photometric redshifts using the BPZ code in R15. The left panel shows a view of all galaxies that are not blended. The confidence level 2 and 3 galaxies are shown as black points with error bars while the confidence level 1 objects are shown as smaller pink symbols.

One immediately sees the outliers in this figure. At high redshift these lie along the degeneracy line between the 4000 Å and Lyman-breaks. This degeneracy ought to be detected by photo-z codes but for many of the objects that show clear Ly-α lines this is not the case and the photo-zs are strongly in disagreement with the spectroscopic redshifts and for most cases the probability distribution function (PDF) has no second peak at the spectroscopic redshift.

We will return to the outliers, but first we will turn our attention to the overall bias between photo-zs and spectroscopic redshifts. While the agreement at low redshift is reasonable in this difference metric, there appears to be a significant offset at higher redshift as previously noted in Oyarzún et al. (2016) and Herenz et al. (2017). This is made clearer in the right panel, which zooms the y-axis and suppresses the error-bars to show the offsets more clearly. The shaded grey region shows the median trend with the shading corresponding to the 68% confidence limit on the median determined by bootstrap resampling plus Monte Carlo resampling of the photo-zs within their uncertainties. It is clear there is a weak bias for photometric redshift to overestimate the true redshift in 0.5 <z< 1.5 or even to z = 3 for BPZ and BEAGLE, while there is a clear offset towards a systematic underestimate of the true redshift at z> 3. It is also worth nothing that the latter offset is considerably larger than that induced by the deviation from Ly-α redshifts from the true systemic redshift − as an example a 1000 km s^-1 offset between the Ly-α redshift and the systemic redshift of the galaxy would lead to a normalised difference (z_MUSE−pz) / (1 + z_MUSE) = −0.0008 at z = 3 where the effect is the largest for MUSE. For cosmological applications this could be crucial but the amplitude is too small for what is seen here.

These offsets are not unique to the BPZ photometric redshifts, but are also seen in the EAZY and BEAGLE photo-zs as shown in Fig. 5. The solid lines show the median offsets as a function of spectroscopic redshift and the shaded regions show the 68% uncertainty region on the median from bootstrap resampling including resampling the photo-zs within their errors. The systematic offset at high redshift is clear and since we have here normalised by 1 + z_MUSE, we see the deviations at low redshift also relatively clearly. The left panel shows the results using the published photo-zs and it can be seen that the EAZY predictions have a higher bias than the other two. We return to this in the next section.

The bias as a function of redshift is also reflected in the trend as a function of magnitude, shown in Fig. 6. The lines and shading show the same as in Fig. 5, but the offset between EAZY and the other two methods is even clearer.

These figures show two separate issues: firstly the EAZY predictions from R15 have a clearly different, and larger, bias than the other two codes, and secondly, there are systematic biases with redshift that are common to all three codes. It is natural to ask what the reasons are for these offsets and whether they can be reduced. Since the predictions using EAZY are the most strongly discrepant, we will focus on these first.

4. Reducing the bias in the EAZY photometric redshifts

There are at least two immediate possibitilities for the substantial offsets seen at high redshift: template mismatch and an incorrect treatment of intergalactic medium (IGM) absorption. Here we will first focus on the former because this also has the potential of resolving the low redshift discrepancies.

4.1. Template mismatch

To explore this we have run EAZY² with the seven different template-sets provided with the code detailed in Table 1. For each template set we used three of the combination methods provided by EAZY: we used each template on its own, any pair of templates and all templates simultaneously. This provides a total of 21 photo-z runs for the UDF data. We do not attempt to deal with strong AGNs here, as these present a separate set of complications (e.g. Salvato et al. 2011) and we have very few strong AGNs in our sample. The version of EAZY used by us does not support iterative adjustments of the zeropoints but we did this by calling EAZY iteratively. However this did not lead to noticeable improvements in photo-z performance, indicating that the main problem with the current data is not photometric calibration, this also matches the finding by S14 that adjustments for HST bands are so small as to be ignored. Given the lack of improvement we show the results without the iterative adjustments here for simplicity.

Focusing first on the overall performance, we summarise the global agreement as the median of Δz/ (1 + z_MUSE) over all redshifts. This is shown in the left panel of Fig. 7 and the great variation is clear to see. It is of course to be expected that the bias will decrease as more flexibility in the template combination is allowed since there will be a higher probability to match the actual SED of the galaxy, in agreement with the arguments in Acquaviva et al. (2015). This is also borne out in the figure where the single template fitting results in significantly higher bias than pairs of templates which again has a larger bias than fitting all templates together. This is however not true for the cww+kin set of templates where the different combination results are consistent with each other within the errors. The strong deviation for the br07_goods template set when using only one template is likely due to the way it was optimised for higher redshift work − Table 2 demonstrates that it performs well at high redshift. It is also worth noting that the bias is positive in all cases, this is caused by the systematic bias at high redshift which we will discuss further below.

Indeed breaking down the numbers by redshift is also enlightening. We do this in Table 2. This shows (100 times) the bias for three bins in spectroscopic redshift and the total. What is striking to see is that in the high redshift case, the ability to combine multiple templates is of little importance except for the eazy_1.3 template set, while at low redshift this is crucial and most template sets perform similarly here. This suggests that a crucial advantage of the eazy_1.3 template set is the addition of the one extra SED from Erb et al. (2010) which allows a more flexible fit to the high-redshift galaxies. The change with redshift is otherwise as expected: for the low redshift galaxies the photometry probes the rest optical to near-IR where a mixture of stellar populations has strong effect on the observed colours, while at high redshift the photometry mostly probe the shorter wavelength UV where the age range of stars contributing to the spectrum is fairly narrow and thus mixtures of SEDs have less effect. The addition of high-quality (observed-frame) K and MIR photometry would undoubtedly alter the picture, but as we show in Appendix A, the addition of ground-based K-band and Spitzer IRAC fluxes does not currently improve the photo-z estimates for the very faint galaxies studied in this paper. This will certainly change with the advent of JWST NIRCam photometry.

The right panel of Fig. 7 shows the fraction of significant outliers in the sample. We here define a galaxy to be an outlier if Δz/ (1 + z_MUSE) > 0.15, but the conclusions are not very sensitive to this choice. What is notable here is that the absolute outlier fraction is high. We will return to this in Sect. 5 below but for now we merely note that the relative trends are as expected and similar to the trends for the bias. The best-performing setup uses the eazy_1.3 template set with simultaneous combination of templates. From now on, when we refer to the “best EAZY” photo-zs, we mean this particular run of EAZY. This combination has a median (z_MUSE−pz) / (1 + z_MUSE) of 0.008, with a median absolute deviation (MAD) of 0.045 and an outlier fraction of 0.10, using z_peak as the point estimate of the photo-z and |(z_MUSE−pz) / (1 + z_MUSE)| > 0.15 as the outlier criterion. This is quite comparable to the BPZ photo-z estimates from R15 which for the same sample have a median bias of 0.00, a MAD of 0.045 and a median outlier fraction of 0.08, and to the BEAGLE photo-zs which have a median bias of 0.002 and a MAD of 0.053 with a median outlier fraction of 0.10.

The improvement in the EAZY photo-zs is clearly seen when comparing the left and right panels in Figs. 5 and 6. The left-hand panels use the EAZY photo-zs from R15, while the right-hand panels show the result with the best EAZY photo-zs. It also clear that after this improvement, the three photo-z codes examined all show almost identical trends for the median bias.

Figures 5 and 6 show only the median trends and not the individual data points since in that case it would not be possible to overplot multiple photo-z codes and still have a readable figure. The figures also show the uncertainty on the median, and not the scatter in each redshift bin. This information is instead shown in Figs. 8−10. Each figure here shows the trends of the normalised redshift difference against magnitude (top panel), observed colour (middle panel) and spectroscopic redshift (bottom panel). The bins along the x-axis were chosen to have similar number of objects per bin. The error-bars show the range containing 68% of the data points in that bin both in the x and y direction and bins with more than 20% outliers (defined as having |(z_MUSE−pz) / (1 + z_MUSE)| > 0.15) are plotted in red. These figures show that the offset at high redshift persists also for EAZY with the eazy_v1.3 template set, and also that there are systematic trends with redshift at the few percent level for all three photo-z codes. At magnitudes fainter than F775W = 27, the scatter goes down but this is almost certainly a selection effect because we are only able to securely determine redshifts for the subset of strong Ly-α-emitters at those redshifts. It is also of particular note that in the 0.4 <z< 0.6 bin, which contains 186 galaxies, the bias is significantly larger. We do not know why this is. Finally, we note that there is only a very minor residual trend with colour. In contrast, the EAZY estimates in the R15 catalogue show very significant trends with colour − strongly indicating that they suffer from a mismatch in the template set used.

4.2. IGM absorption modelling

The absorption of the neutral medium at high redshift is an important ingredient in photometric redshift codes for objects with photometry shortwards of rest-frame Ly-α. The widely used model of Madau (1995, M95 hereafter) was recently revised by Inoue et al. (2014, I14 hereafter) who demonstrated that the updates in IGM modelling could lead to modest, but systematic, changes in the photometric redshift estimates of Δz ~ 0.05 in the redshift range of interest to us.

The default EAZY operation uses a hybrid approach where the absorption longwards of the Lyman limit is treated using the M95 model, while the Lyman continuum opacitiy is treated using the I14 model. However the code contains the options to use the I14 model throughout and we have compared both.

The IGM treatment is statistical in nature as the IGM seen by a given galaxy will differ depending on the local conditions. Thus ideally a photo-z code should also marginalise over this unknown but this is not normally done. To approximately treat this we have modified EAZY to include a scale factor for the attenuation shortwards of the Lyman limit, s_LC, and another for the Lyman-forest attenuation, s_LAF. We then ran EASY varying s_LC and s_LAF between 0.1 and 2.5 in steps of 0.1 for both the default IGM treatment using a mixture of M95 and I14, and only using the I14 IGM treatment.

We found that the dependence on s_LC is signficantly weaker than that of s_LAF. This is natural since the attenuation bluewards of the Lyman limit is very large and any small change in its value has little effect. To assess the effect of the scale factors we calculate a running median of (z_MUSE−pz) / (1 + z_MUSE) versus z_MUSE between z = 3.0 and z = 6.5. We calculate the median with 31 galaxies per bin and calculate the uncertainty on the median using 999 bootstrap repetitions. We then fit a linear function (z_MUSE−pz) / (1 + z_MUSE) = a + b(z_MUSE−3) and minimise |a| + |b|, that is, we attempt to jointly minimise the bias at z = 3 and the slope. This gave us a minimum for s_LAF = 0.8 and s_LC = 1.1 for the default IGM treatment in EAZY.

The results are summarised in Fig. 11. The black solid line shows the normalised bias as a function of z_MUSE for the optimised EAZY model, similar to the bottom panel of Fig. 8 but with finer sampling in z_MUSE and a zoomed y-axis. The dashed blue line shows the bias for the I14 IGM model and this shows a slightly stronger redshift trend but overall a slightly smaller bias. The orange solid line shows the best adjusted IGM model with the shading indicating the 68% confidence region around the median.

The main point to note here is that the optimised IGM model works somewhat better but still fails to correct all the bias at high redshift. This should not be taken as a failure of the IGM model, however. As mentioned above this is statistical in nature and the scaling factors we have introduced are also mean quantities for the whole redshift range. By letting the scale factors vary with redshift, we can remove the trend also at high redshift but we do not have enough data points to carry out this minimisation in a statistically meaningful way.

It is also interesting to ask whether changes in the IGM model is sufficient to explain all scatter in the normalised bias. The answer to this is a qualified “yes”. For objects for which |(z_MUSE−pz) / (1 + z_MUSE)| > 0.2, the effect of the IGM is insufficient to change the photometric redshifts to agree with the spectroscopic. However for galaxies with |(z_MUSE−pz) / (1 + z_MUSE)| < 0.1, we can find values for s_LAF and s_LC that bring photometric and spectroscopic redshifts in agreement. However, these solutions show two clear problems: the scatter in the implied Ly-α to Ly-β flux depression is much larger than that inferred from Ly-α-forest studies (e.g. Fan et al. 2006; Faucher-Giguère et al. 2008), and secondly, the fits tend to prefer the extremes of the scales (s_LAF = 0.1 or s_LAF = 2.5), rather than something in the middle. This strongly hints that while the IGM treatment might be part of the problem, the template spectra are also not fully covering the range of real galaxy spectra at these redshifts.

To summarise this section, we wanted to compare the three photo-z codes used here and found, in good agreement with previous work (e.g. Hildebrandt et al. 2010), that these all agree well. Overall the BPZ estimates performs very slightly better than the re-calibrated EAZY estimates, except perhaps at high redshift, and show less scatter than BEAGLE although the median trend is very similar. In the following we will therefore mostly use the BPZ photo-zs but occasionally also show the results of using EAZY or BEAGLE, however we stress that for practical use these can be used interchangeably − although we do not recommend using the EAZY estimates from the R15 catalogue.

5. Redshift outliers

It is customary, albeit arbitrary, to define outliers as those galaxies for which Δz/ (1 + z_MUSE) >x, where x is often taken to be 0.1 or 0.15, or as Δz/ (1 + z_MUSE) >nσ_MAD, where σ_MAD is the median absolute deviation around the median, adjusted to correspond to a standard deviation for a Gaussian distribution, (2)We take n = 3 in this section since this gives a reasonable number of galaxies and hence avoids the problems of small-number statistics. The statistics of truly significant outliers, with x = 0.5 or n = 5, is also interesting and we discuss these catastrophic outliers in some detail in Appendix B. We show there that BEAGLE and BPZ have the lowest fraction of catastrophic outliers, with EAZY having a factor of 2–3 more catastrophic outliers, depending on the definition.

In Fig. 12 we show the outlier fraction for the BPZ photo-zs against F775W magnitude (top), F775W−F160W colour (middle) and spectroscopic redshift (bottom).

There are some points worth noting in Fig. 12. The first is that the outlier fractions are relatively high compared with what is normally reported in the literature. In the most similar studies to our work, Dahlen et al. (2013), Hildebrandt et al. (2010) and R15 all found outlier fractions that were < 5% in the best cases with a cut-off of 0.15 (see Fig. 14 below for a quantitative illustration). We are generally above this level at all magnitudes. The second point of note is that for the more stringent 0.15 case (black symbols), the outlier fraction rises somewhat towards fainter magnitudes and lower redshifts. For the catastrophic outliers (orange symbols) the trend is much less pronounced and more a tendency for catastrophic outliers to be more prevalent at high redshift (see Appendix B for more details on catastrophic outliers).

The increase in outlier fraction with fainter magnitude is real, but it is in part a reflection of an increased scatter in the relationship between photometric and spectroscopic redshifts and in part a reflection of the systematic bias in the photometric redshifts, particularly at high redshift, that we saw above. This is made explicit in Fig. 13 where we have defined outliers to be those that have |Δz/ (1 + z_MUSE)| > 3σ_MAD, where σ_MAD was calculated locally in each bin, choosing 5σ_MAD would lead to very similar trends but the points would be shifted down by on average a factor of 1.7−1.8. The panels are the same as in Fig. 12 and the contrast is clear: there is no increase in the number of outliers with magnitude or redshift, however the trend with colour is approximately maintained.

To put these results in context, Fig. 14 shows the outlier fraction in our sample down to F775W = 30, as a function of the (z_MUSE−pz) / (1 + z_MUSE) cut-off. The strong trend with the cut-off value is clear and it is also clear that the outlier fraction for EAZY (dashed line) is consistently somewhat higher than for BPZ (solid blue line) and BEAGLE (red dot-dashed line) but recall that the points here are not independent. We can also estimate a limit to how well we can reduce the outlier fraction by combining photo-z estimates. We do this by calculating the fraction of objects that are outliers for EAZY, BPZ and BEAGLE while at the same time all three photo-z codes give redshift estimates in agreement with each other. The resulting values (green dashed line) lie below the other curves but remains above 1% for cut-off values < 0.15.

A comparison of these results to other studies is not entirely straightforward. Firstly, the photometry might be of different quality and not all objects might be included. To reduce the effect of this we will compare to studies that use well-studied photometric data sets, including HST data. Secondly, different studies have different limiting magnitudes and this has significant effects on outliers, both due to the photometric quality and because the mean redshift will differ. And finally, the methods used are different in different surveys which might make comparisons hard. With those caveats we have chosen four different surveys to compare against.

We show the results from the comprehensive set of photo-zs in the COSMOS field by Ilbert et al. (2009) as the green squares. This survey is considerably brighter on average than ours but their photo-zs use 30 bands in contrast to our 11. They report outlier fractions for various subsamples. The fraction found for their bright sample, i⁺< 22.5, is very low (< 1%) while that for the faintest sample 24 <i⁺< 25, and also for their 1 <z< 3 sample, reaches 20% − considerably higher than our estimates at the same threshold but of course for less deep photometry. We plot the outlier fraction reported for the GOODS-S field by Bezanson et al. (2016) as a grey rectangle with error-bars extending from outliers in the grism-z vs. photo-z comparison (lowest) to the spec-z vs. photo-z comparison (highest). We also report results for Hildebrandt et al. (2010) and Dahlen et al. (2013) who both use a mixture of space- and ground-based photometry which is somewhat shallower than the photometry in the UDF. They present results for a range of codes and we show this as an error bar with a symbol placed at the median outlier fraction. It is reasonable to compare to the best results from their studies which is shown as the lower end of the error bars. It is obvious that the best-performing codes in these studies give lower outlier fractions than we are finding, by a factor of approximately two. However, Dahlen et al. also report the performance of the best codes in two faint magnitude bins (24 <F160W< 26 and 26 <F160W< 28). This is shown by the open triangles and it is clear that the outlier fractions for these fainter subsamples are higher than what we find for our sample down to F775W = 30.

The CANDELS data are of course shallower on average than the HUDF data used here. The median F160W for the galaxies with redshift confidence level ≥ 2 in our sample is 25.89, while for the Dahlen et al. (2013) it is close to 22 and for Hildebrandt et al. (2010) is it somewhat fainter. Thus we argue that our outlier fraction estimates are as good or lower than what has been reported in the literature at comparable depth.

That said, however, it is important to realise when using photo-zs, that the outlier fraction at faint magnitudes, even with ultra-deep photometry, is going to be considerably higher than what is found for bright galaxy samples such as the i⁺< 22.5 COSMOS sample.

6. The impact of redshift completeness

Our statements above about photo-z accuracy are, by necessity, only made about galaxies for which we have spectroscopic redshifts. However, it is also of major interest to know how well the photometric redshifts work also for galaxies for which we have not managed to secure a redshift. An accurate assessment of this will of course have to wait until a deeper data set becomes available, but we can already make some progress with the data in hand.

Figure 15 shows the redshift incompleteness in the redshift-m_F775W plane, assuming perfect photometric redshifts (the pz_BPZ were used). It shows very clearly the preponderance of Ly-α-emitters at faint magnitudes and the lack of redshifts in the 1.5 <z< 2.9 MUSE redshift desert. While the completeness at bright magnitudes is close to 1, it is not everywhere at that level. The first thing to ask is therefore whether this incompleteness can affect our conclusions about the performance of the photometric redshifts at bright magnitudes.

There are only two galaxies with m_F775W ≤ 24 for which we have no redshift in the catalogue. These are RAF 2231 and 24499. The first of these does in fact have a clear absorption line redshift (z = 0.66708) but was missed in the first iteration of redshifts (cf. Sect. 2.2) because of the strong [Oii] λ3727 line from RAF 2236 next to it. However it is clear from narrow-band images over Ca K that the absorption line redshift is associated to RAF 2231. Since we recognised this discrepant redshift during the photo-z comparison, we have not added these updated redshifts here, nor are they in the catalogue in Paper II but will be in future updates.

RAF 24499 is very close to a much brighter star (RAF 24467) which would require a more sophisticated spectrum extraction than we have adopted here to have a chance to determine its redshift (it clearly has no very strong emission lines). Thus, the failure to determine a redshift for this object has nothing to do with its intrinsic properties.

Faintwards of F775W = 24 however, a few objects have no redshift at present purely due to their intrinsic properties. There are 10 objects with F775W ≤ 25 (out of 310 or 3.2%) that have no spectroscopic redshift estimate. While we have been unable to determine a precise redshift, the sources all have clearly detected continua that are consistent with their photo-z estimates, which fall mostly in the 1.5 <z< 2.8 redshift range where the main diagnostic feature in the MUSE spectra is the relatively weak Ciii ] λ1909 doublet (Maseda et al. 2017).

Turning next to the fainter galaxies, we would like to ask whether the galaxies for which we have no spectroscopic redshift fall in distinct regions in the space of SEDs. To do this, we then need to partition this space or project it down to a smaller-dimensional space. The latter approach has been taken by Masters et al. (2015) who use a self-organising map to project photometry from the COSMOS survey on to a 2D plane. While this approach shows significant promise for photo-z calibration, given our significantly smaller sample, we here adopt the first option and use k-means clustering to partition the space spanned by F435W−F606W, F606W−F775W, F775W−F850LP, and F606W−F125W colours. This is very similar to, and was inspired by, the approach taken by Bonnett et al. (2016) for their investigation of the photo-zs in the DES shear catalogue.

The k-means algorithm assigns points in space to k clusters with the clusters located so that the distances to the cluster centres is minimal. In two dimensions this is more commonly called a Voronoi binning of the data and the clusters are called Voronoi cells. Here we divide our 4-dimensional colour-space in to 12 and 25 bins respectively. For k = 12 this leads to bins with 334−1176 galaxies per bin, with an average of 650, while for k = 25 the bins have from 82 to 591 galaxies with a mean of 312. The bins are chosen by the k-means algorithm and each galaxy is assigned to one of these bins. The process creates bins that are adapted to the distribution of data which is very convenient for small data sets such as ours. To illustrate the properties of the sample, Fig. 16 shows the median redshift (right) and redshift completeness (left) for k = 12 (top) and k = 25 (bottom) bin segmentation of four-colour space. For these figures we only include galaxies with F775W ≤ 27, but the qualitative appearance remains when including all galaxies to F775W = 30 although the reshift completeness goes down and the mean redshift goes up.

One main point to note is that while there is variation in completeness between the different bins, the quantitative variation is modest. It is also true that with 12 bins, the median redshift in each bin is fairly close to the overall mean of the sample (median(z) = 1.091), while 25 bins clearly segregate the galaxies somewhat more but at the obvious cost of fewer galaxies per bin.

Moving now to the overall trends of bias in redshift determinations, it is natural to ask whether the bias we found in Sect. 3 is artificially large or small given that we only considered galaxies with secure spectroscopic redshifts. To tackle this question, we need to have a way to assign likely redshifts to the objects for which no spectroscopic redshift was possible, and we would need to do this independent of our photometric redshift codes. For this we will use our k-means bins. The galaxies in each of these bins will have very similar colours so our assumption is that the distribution of redshifts for the sample with MUSE redshifts is similar to the distribution of redshifts for all galaxies in that bin. Given this assumption, we assign each galaxy without a secure spectroscopic redshift a redshift at random from the galaxies with a spectroscopic redshift in the k-means bin in which the galaxy falls. This is of course to some extent a very simple machine-learning method for photometric redshifts, but we choose to use the k-means clusters rather than a more sophisticated algorithm as the method is simple and its limitations are therefore easier to identify. This approach is rather conservative as the k-means bins are not particularly small and any residual degeneracy in colour-redshift space is therefore not resolved.

Figure 17 shows the resulting trend in normalised photo-z bias as a function of magnitude (top) and F775W−F160W colour (bottom). The black points show the result when only galaxies with F775W< 27 are used, while the orange points result when the full sample of galaxies with F775W< 30 is used. The photometric redshifts used here are BPZ ones, but the conclusions are not dependent on this choice.

The blue line in each panel is the trend found when only considering galaxies with Confid ≥ 2 which we showed in Fig. 9. At bright magnitudes we have very low spectroscopic incompleteness and as expected the blue line and the black points fall on top of each other in the top panel. At fainter magnitudes we predict comparable, but not identical biases in the redshift determination. To guide the eye a pair of dotted lines at a bias of ± 10^-2 are included − recall that for cosmological weak lensing applications a bias < 2 × 10^-3 is desired. It is clear that for most redshift bins the photo-z codes considered here are not able to determine the mean redshift to that level, as commented above, but they are able to constrain it to within 5%. From this analysis we can tentatively conclude that spectroscopic incompleteness is unlikely to change this conclusion significantly.

The assignment of redshifts is done in bins of colour so it is not surprising that the predicted bias shows a dependence on colour, but we note that the colour used here is not used in the k-means cluster definition. The bias changes clearly in the redder bins when including faint galaxies. This is expected and reflects the mean colour with redshift, but the important point is that at no colour do we see a mean bias above 6%, although that should be tempered by noting that in many bins the bias is larger than 1%.

We do not show the comparison as a function of redshift because the method is inherently biased in this case. Each k-means bin has an upper and lower redshift cut-off given by the sample we have redshifts for. This is not corrected for in the Monte Carlo approach used above and as a consequence the mean redshift at high redshift is biased low and the mean redshift at low redshift is biased high.

7. The impact of galaxy superpositions and blends

As discussed in the introduction, for cosmological uses of photometric redshifts the requirements on biases are now very strict. A frequently considered approach is therefore to calibrate photometric redshifts using an existing spectroscopic redshift catalogue. A comprehensive overview of the different methods is given in Newman et al. (2015) and references therein, while Hildebrandt et al. (2017) show the impact of different calibration methods on derived cosmological parameters.

This approach places considerable constraints on the properties of the spectroscopic training/calibration sample, and this has been discussed extensively in the literature as reviewed by Newman et al. While the size of the spectroscopic sample has got much attention, Cunha et al. (2014) showed convincingly that the fraction of incorrect redshifts might have a severe effect on cosmological parameter determinations and that it is necessary to limit this to ~ 1%.

In Cunha et al. and other work in the literature, the focus is on redshifts that are incorrectly determined from the spectrum but this is only one half of the problem. To carry out such an endeavour, it is also necessary to be able to map reliably between spectroscopic redshift and photometric object. At bright magnitudes this presents few challenges as comparatively bright objects are well separated on the sky, but at fainter magnitudes this is not the case. Thus we might have cases where the redshift is very certain, but that the assignment of the redshift to a photometric object might be uncertain − or indeed in many cases it might appear certain when in fact it is wrong. We have already discussed this fact above in Sect. 3.1, but we did not quantify the effect.

The challenge when using a fibre or long-slit spectrograph is that other, fainter, sources within the aperture sampled by the spectrograph can contribute emission lines that lead to incorrect redshift determinations. In the MUSE data this is frequently seen at faint magnitudes and requires careful inspection of narrow band images to ensure correct assignment of redshifts to objects.

Here we seek to quantify this effect in an approximate way for the sample as a whole, we will discuss the shortcomings of this approach below. Specifically we ask, for each galaxy, whether there is a galaxy with a fainter F775W within a circular region of radius r that has a confidence ≥ 2 emission line redshift, and where the strongest line in the fainter galaxy has a higher flux than the strongest line in the galaxy considered. If this is the case, we consider that this is a possible contamination source for an observation taken with an aperture or seeing comparable to the size r. This effectively assumes that the contamination source is a point source whereas in truth they are extended. In addition we also keep the sources that we are unable to resolve, even with MUSE, as a possible source of contamination. For these we lack the neccessary information to treat these on the same footing, so we assume the line detected comes from the faintest object but would be assigned to the brightest. We also do not know the spatial separation for these and set it to zero − this will then give a minimum level of contamination at all resolutions which is clearly unphysical but with the present data we cannot do better.

Figure 18 shows the fraction of galaxies with a potential contaminant as a function of separation for two magnitude ranges. The contamination potential for objects with magnitude 24 <m_F775W< 25 (blue line) is lower than that for objects one magnitude fainter (red line). The general shape in both cases is as expected, with the larger search area increasing the chance of finding a contaminant. The zero level at small radii is caused by the blended objects as discussed above.

The magnitude dependence clearly indicates that we should also inspect this as a function of magnitude. This is shown in Fig. 19 which shows the trends as a function of magnitude in three different aperture sizes. There are two noticeable trends: a drop at bright magnitudes and another at faint magnitudes. The drop at faint magnitudes is a consequence of the declining redshift completeness at fainter magnitudes. The blue solid line shows the effect of ignoring all blended objects − since these truly are confused this is of course an over-simplification but it shows that the majority of the blended objects at bright magnitudes are impossible to deblend even with MUSE.

For a spectroscopic observation with a spatial resolution of 0.̋5, one can expect significant contamination in ≈ 1% of all spectra, while for a larger aperture, say a fibre spectrograph with 2′′ aperture, typically ≈ 3% of all spectra will have stronger emission lines from a galaxy that is not the brightest source in the aperture, at least for galaxies fainter than F775W = 24 which is where the UDF sample has significant number of galaxies. That this is a real concern can also be seen in the fact reported in Paper II, that 2 of 12 redshifts in this area from the VIMOS Ultra Deep Survey (VUDS) data release 1 (Tasca et al. 2017), are incorrectly associated with HST objects due to contamination. VUDS primarily selects objects with i_AB ≥ 25 so this rate of 17% can be contrasted with our estimate of ~ 5% at a similar depth for a 1″ radius.

That the VUDS contamination is larger is not entirely surprising as our our estimate is in fact likely to be an understimate at radii less than a couple of arcseconds. This is because the contaminating sources are typically extended, as spectacularly seen for Ly-α emitters in MUSE data (e.g. Wisotzki et al. 2016; Leclercq et al. 2017). Thus in reality the contaminants will not be delta functions in spatial extent and this will boost the number of contaminants at small separations. However as this depends on the line flux profiles and might be mitigated by examining spatial line profiles in slit spectra, a precise quantification of this effect is beyond the present work.

The importance of this contamination depends of course on whether it leads to incorrect redshift determinations for the galaxy. This will clearly depend on the depth of spectroscopy. With a well-detected continuum it is more likely that one can recognise that the emission line does not belong to that galaxy or not, but for very faint galaxies it is probable that this contamination will lead to incorrect redshift determinations and hence biases in the calibration of photometric redshifts.

One might also imagine that the blends might be identified as outliers in photo-z versus spectroscopic redshift comparisons. To test the effect of this, we assigned each blended object the redshift implied by the strongest line, but in the case of multiple contaminating sources we chose the one that would minimise the discrepancy with the photometric redshift. We then calculated the fraction of contaminants only counting those objects with |(z_MUSE−pz) / (1 + z_MUSE)| < 0.15. The result of this test is shown by the dashed green line in Fig. 19 and as can be seen this reduces the contamination problem somewhat but does not remove it, the effect is similar if blends are removed. At faint magnitudes the overall effect of the (z_MUSE−pz) / (1 + z_MUSE) cut is very small because almost all sources involved are Ly-α emitters.

8. Discussion

The requirements on photometric redshift accuracy naturally depend on the scientific question that they are needed for. We have found that the mean bias in redshift determination is always below 0.05(1 + z) even down to F775W = 30 with a σ_MAD< 0.06 in (z_MUSE−pz) / (1 + z_MUSE) for all but the 0.4 <z< 0.6 bin. While this might appear to be a small bias, it is clearly too high for cosmological needs. For many galaxy evolution studies, however, it is an acceptable bias. That said, the clear systematic trends with both magnitude, redshifts and colour are important to note and also be aware that the scatter varies somewhat less with these properties.

We find two regions in redshift space where spectroscopic and photometric redshifts differ significantly. These are at low redshift, 0.5 <z< 1.5, and at higher redshift z> 3.5. Armed with the MUSE data it is reasonable to ask whether the discrepancies we see at low redshift can be due to the spectral templates being systematically unrepresentative. To explore this, Fig. 20 shows the location of our MUSE sources in two different line ratio diagrams where we have used the reference (weighted) line measurements although this particular choice has no significant influence on the results shown here. The figure on the left shows the classical BPT [Nii] λ6584/Hα vs. [Oiii] λ5007/Hβ diagram. This is appropriate for the lowest redshift galaxies where Hα is still available. To give context to the MUSE galaxies, we make use of line flux measurements for galaxies from the Sloan Digital Sky Survey (SDSS) Data Release 7 (Abazajian et al. 2009) from the MPA-JHU compilation³ which is an update of that discussed in Brinchmann et al. (2004). The distribution of the SDSS galaxies with S/N in each line > 3 are shown by the grey contours where the outermost contour encloses 99% of galaxies. It is immediately obvious that our objects fall mostly within the locus of SDSS galaxies albeit with a relative overabundance of lower metallicity (low [Nii] λ6584/Hα ratio) galaxies. The colour gives the relative discrepancy between BPZ photometric and spectroscopic redshift, Δz/ (1 + z). There is no evidence of an increased discrepancy away from the bulk of the SDSS galaxies in this figure.

The right-hand panel shows the same but now plotting [Oiii] λ5007/[Oii] λ3727 on the x-axis. This line ratio is more comprehensively explored, for a larger sample, in Paalvast et al (in prep), here we simply use it as a way to explore the spectroscopic properties of the higher redshift part of the sample. The meaning of the symbols and lines are the same and there is perhaps some evidence that the most extreme [Oiii] λ5007/[Oii] λ3727 galaxies show increased photo-z discrepancy and this might reflect an incomplete template sample. However, the bulk does not show such a correlation so this is unlikely to explain the offset in the median between photometric and spectroscopic redshifts.

In the literature, line ratios are either predicted by the code providing the underlying SEDs for photo-z fitting (e.g. Brammer et al. 2008), but more commonly line ratios are taken from some compilation and the flux of the lines is tied to the star formation rate of the SED template. The most widely used line ratios are those provided by Ilbert et al. (2009). Those authors compiled observed line ratios (relative to [Oii] λ3727) provided in the literature for a mixture of galaxies at z< 0.1. The resulting line ratios are plotted as the dashed purple line in the left-hand panel and as the purple square in the right-hand panel. It is clear that these deviate from the bulk of the galaxies in the MUSE sample. This is not entirely surprising given that the line ratios used by Ilbert et al are more representative for low-z massive galaxies than our sample of faint, star-forming galaxies. For application to fainter sources it would therefore be desirable to have a more comprehensive treatment of emission lines such as that taken by Salmon et al. (2015) who use a library of simulated emission line spectra for metal-poor galaxies from Inoue (2011), or the BEAGLE (Chevallard & Charlot 2016) which exploits a large library of photoionisation models from Gutkin et al. (2016).

The importance of emission line ratios does however appear to be relatively minor given the lack of clear correlations seen for our sample. One might instead worry that the offsets would correlate with the relative strength of the emission lines. We have explored this by looking for correlations between Δz and the equivalent width of the strongest lines. Our assumption here is that this line is representative for all emission lines in the spectrum. Except for tentative evidence of the 2–3 systems with the highest Ly-α equivalent width to have slightly higher Δz, we have not found any robust evidence for the equivalent width to have an impact on the photo-z estimates. Since the strengths of lines are tied to the star formation rate of each template in the template fitting codes, this is not entirely unexpected.

Taken together with the lack of clear correlation between bias and location in the k-means bins in Sect. 6, we find no obvious “second-parameter” that can be used to reduce the photometric redshifts errors. A logical next step would be to combine multiple photometric redshift estimators to create a consensus photo-z (e.g. Carrasco Kind & Brunner 2014; Cavuoti et al. 2017; Süveges et al. 2017), but we postpone this to a future work.

The addition of more photometric bands might offer another way to reduce bias and indeed Bezanson et al. (2016) showed convincingly that increasing the number of bands, and in particular adding IRAC photometry does improve photo-z performance significantly for brighter galaxies. In Appendix A we explore the effect of using the 44 bands in the S14 photometric catalogue for the fainter galaxies considered here. We find there that indeed the bias is lower when all 44 bands are included, but at a cost of a significiantly increased scatter, particularly at faint magnitudes. Thus for faint F775W> 25 galaxies the current photometric catalogues from the ground and Spitzer are not of sufficient quality to help improve photo-z determinations.

While reducing the bias for individual galaxies is important, re-calibration methods either using spectroscopic redshifts directly (e.g. Lima et al. 2008) or using spatial cross-correlations (Newman 2008) show great potential to reduce the bias in the estimated mean redshift of a sample of galaxies. However these all need correct redshifts, and we have demonstrated above that for a few percent of galaxies the redshifts might be of high confidence, but may be assigned to the wrong photometric object. This presents significant challenges for future surveys (Cunha et al. 2014) and ways must be found to counter this − particularly for studies that go fainter than 24th magnitude, such as the LSST weak lensing survey. The most obvious, but less practical, way to reduce this problem is to use spatially resolved spectroscopy for the fainter galaxy population. The various data sets obtained by MUSE are particularly well suited for this. While not quite as powerful because of the loss of information, slit spectra also provide a way to at least address some of these problems and spectra should ideally have not just a quality flag for the redshift, but also one for the association to photometric objects. There is no easy way to tackle this problem in surveys using fibre-based spectrographs, and the residual uncertainy will remain and must be included as a possible source of error in subsequent analysis.

9. Conclusions

Our aim with this study was to quantify the performance of photometric redshifts in galaxies down to F775W = 30 with predominantly emission line galaxies at F775W> 25.5. We have extended the validation of photometric redshifts to fainter magnitude limits than previous studies and found that they do in general perform well throughout. The median redshift in bins of redshifts, colour and magnitude is always determined to |(z_MUSE−pz) / (1 + z_MUSE)| < 0.05 for the galaxies for which we have spectroscopic redshifts from MUSE as shown in Figs. 8−10. This conclusion appears insensitive to the choice of photo-z estimation code with EAZY, BPZ and BEAGLE all performing comparably well. This does give us some confidence in results based on photo-zs at faint magnitudes. We do, however find systematic trends with redshift and colour: the photometric redshifts being systematically biased high for 0.4 <z< 1.5 by (z_MUSE−pz) / (1 + z_MUSE) = −0.02 to − 0.04, and biased low by a similar to slightly higher amount at z> 3.

While the cause of the deviation at low redshift is unclear, we showed in Fig. 11 that allowing for free scaling of the IGM model for Ly-α-forest and Lyman continuum absorption does reduce the discrepancy and that it is possible to remove all bias from galaxies with |(z_MUSE−pz) / (1 + z_MUSE)| < 0.1 by adjusting the amount of IGM absorption for each galaxy individually, although the amount required appear to be inconsistent with the scatter seen in Ly-α-forest absorption data indicating that the changing the IGM modelling is unlikely to be the full explanation.

The number of objects with strongly discrepant photometric redshifts is considerably higher than that found in brighter subsamples of galaxies in the literature. However we also demonstrate in Fig. 13 that the fraction is a strong function of how an outlier is defined, but one should at least expect an outlier fraction of 10–20% for these faint galaxies, where an outlier is defined as having |(z_MUSE−pz) / (1 + z_MUSE)| > 0.15. The sample is too small to make strong statements on catastrophic outliers, defined as those satisfying |(z_MUSE−pz) / (1 + z_MUSE)| > 0.5 or |(z_MUSE−pz) / (1 + z_MUSE)| > 5σ_MAD, but we did show that EAZY has 2–3 times more catastrophic outliers than BEAGLE or BPZ. Encouragingly, only < 0.6% of the galaxies are catastrophic outliers for all codes simultaneously (Fig. 14 and Table B.1), indicating that comparisons of codes might help identify possible problem cases.

Finally, we found that the fraction of spatially overlapping galaxies is sufficiently high to be of concern for cosmological surveys using faint galaxies for weak lensing studies. We showed in Fig. 19 that at F775W = 25, 1–2% of all galaxies will have a fainter source within 0.5′′ that has a stronger emission line than the strongest line in the galaxy in question. While the redshift determined from the emission line(s) might be a very secure redshift, the association of the redshift to the photometric object will be incorrect, leading to incorrect redshifts in photo-z calibration samples at this percentile level. This problem is expected to become considerably more severe for spectroscopic surveys of faint galaxies using fibres, where > 5% of the galaxies might be affected by this, depending on the fibre diameter, and it is clearly necessary to develop mitigating strategies for this problem. Unfortunately a simple constraint on the difference between photometric and spectroscopic redshift is not enough as we showed in the same figure, thus a more sophisticated approach is called for and ultimately a statistical approach to correct for the effect might be needed.

---

Acknowledgments

We thank the referee for a constructive and helpful report that helped improve the paper. J.B. acknowledges useful discussions with Peter Capak. This work was in part supported by Fundação para a Ciência e a Tecnologia (FCT) through national funds (UID/FIS/04434/2013) and by FEDER through COMPETE2020 (POCI-01-0145-FEDER-007672). During part of this work, J.B. was supported by FCT through Investigador FCT contract IF/01654/2014/CP1215/CT0003. This work is supported by the ERC advanced grant 339659-MUSICOS (R. Bacon). T.C. acknowledges support of the ANR FOGHAR (ANR-13-BS05-0010-02), the OCEVU Labex (ANR-11- LABX-0060) and the A*MIDEX project (ANR-11-IDEX-0001-02) funded by the “Investissements d’avenir” French government program managed by the ANR. J.R. acknowledges support from the ERC starting grant 336736-CALENDS. J.S. acknowledges support from ERC Grant 278594-GasAroundGalaxies. L.W. acknowledges funding by the Competitive Fund of the Leibniz Association through grant SAW-2015-AIP-2. R.A.M. acknowledges support by the Swiss National Science Foundation. S.C. acknowledges support from the ERC advanced grant 321323-NEOGAL.

References

Appendix A: Adding Spitzer/IRAC and ground-based photometry

The paper focuses on photometry from R15, but this is limited to HST imaging and while up to 11 bands are available, the reddest filter is F160W and there are no intermediate width filters. This can be contrasted with the study by S14 which has photometry in a total of 44 bands, including the K_s-band and Spitzer/IRAC photometry in 3.6, 4.5, 5.8, and 8 μm. We have not used this catalogue as our master catalogue for three main reasons: 1) it does not go as deep as R15, thus many more galaxies with a MUSE redshift have no corresponding photometry in the catalogue; 2) the source extraction and redshift determinations were done on the R15 catalogue (see Papers I and II for details); and 3) the S14 catalogue does not include the UV photometry from Teplitz et al. (2013).

However the S14 catalogue offers the possibility to test how the photo-z estimates are influenced by the photometric catalogue used. To do this, we have matched the photometric catalogues from S14⁴ to the master catalogue used here, requiring that the catalogue positions differ by less than 0.̋2. This results in 3956 matches for which 3899 have five or more colours, so we can calculate reliable photometric redshifts. For these 3899 we have MUSE redshifts with confidence ≥ 2 (1) for 905 (1071). It is important to keep in mind that due to difference in image segmentation, two matched detections might differ rather significantly although for most objects we expect good correspondence. This is borne out in a direct comparison of photometry with a slight difference in F435W photometry with S14 being slightly brighter than R15. This is shown indirectly in Fig. A.1 which compares the F435W−F606W, F606W−F775W, and F775W−F850lp colours calculated using the S14 photometry to that calculated using the R15 photometry (the conclusions are unchanged when using flux ratios rather than magnitudes). The solid line shows the median trend and it can be seen that as a function of the F775W magnitude from R15, the colours are broadly in good agreement with execption of the aforementioned offset in F435W which could be caused by the iterative adjustment of photometry done by S14. The scatter is substantial at faint magnitudes but consistent with that expected from the photometric uncertainties. The final panel in the figure shows the number of objects in the R15 catalogue within the MUSE mosaic field of the view (black) as a function of F775W magnitude. This can be contrasted to the joint catalogue of S14 and R15 (blue) which one can see is relatively deficient in the faintest objects since both histograms are normalised to unit area.

We run EAZY using the same settings that we found gave the best results in Sect. 4 but we will also show the results from S14 below. We will refer to these re-runs of EAZY as the reprocessed S14 photo-zs. Our reference run uses all 44 filters and is plotted as a thicker black line in Figs. A.2– A.4. To test the sensitivity of the predictions to the photometry available, we also run EAZY excluding the 14 medium band filters (dash-dotted orange line in the figures), and excluding all ground-based photometry (dashed blue line). To first order, removing filters can be viewed as removing information, one would therefore expect that the runs with a larger number of filters will result in photo-zs that better match the spectroscopic redshifts.

This basic assumption is borne out in Fig. A.2, where we show the median (z_MUSE−pz) / (1 + z_MUSE) as a function of F775W magnitude (top), F775W−F160W colour (middle), and z_MUSE (bottom).

The first thing to note is that while they perform well, the original S14 photo-zs (thin pink line) are considerably worse than the BPZ (dotted red line) and the reprocessed photo-zs. This demonstrates the same result that we found in Sect. 4 and underlines the importance of the choice of template set. We also see that the reprocessed photo-zs that use all photometry perform best in terms of median bias and shows no magnitude and very weak redshift trends. It is also clear that removing the medium band photometry reduces performance, and removing all ground-based photometry makes the photo-zs considerably worse.

Finally, it is also noteworthy that the BPZ photo-zs using the RAF catalogue which only use HST photometry, perform much better than the corresponding reprocessed photo-zs that only uses space-based (HST and Spitzer) photometry. This might at first glance be somewhat surprising, but there are two main reasons why this might be so: firstly, as mentioned above, the R15 catalogue uses HST UV photometry of the UDF that were not used by S14. This is likely to improve photometry redshift performance, particularly at lower redshift. Secondly, the assignent of MUSE redshift to an HST object was based on the RAF catalogue and their segmentation map, but we have not done a similar scrutiny of the S14 catalogue so some of the S14 photometric objects may refer to an object that has a different z_MUSE, or which includes several distinct objects at different redshift in the RAF catalogue.

The same broad trends can also be seen in Fig. A.3, which shows the scatter around the median (calculated as MAD[(z_MUSE−pz) / (1 + z_MUSE)]). The most notable point here is that the scatter rises sharply towards the faintest magnitudes for all, except the R15 based BPZ photo-zs. We should note that just as for R15, we have used all S14 photometry blindly despite S14 suggesting this is not a good idea, but the scatter at the bright end matches very well the MAD = 0.010 found for this field in Bezanson et al. (2016). Thus the increased scatter at fainter magnitudes might simply be a reflection of the challenge of providing reliable IRAC and ground-based photometric measurements for these sources. This is also indicated by the fact that the reprocessed photo-zs that omit all ground-based photometry perform as well or better than the run with all 44 bands at the very faintest magnitudes (F775W> 28). To show this explicitly, we have run EAZY on the catalogue with only HST photometry, that is excluding IRAC photometry as well. The results of this is shown as the short-dashed grey line in Figs. A.2−A.4, which should be compared to the dashed blue line. The top panel of Fig. A.3, for instance, shows that the scatter is generally lower when IRAC fluxes as included until F775W ≈ 26, where the two curves cross over. The bias is less affected and the inclusion of IRAC fluxes improves the bias slightly down to F775W ≈ 28. At brighter magnitudes the removal of IRAC data worsens the performance, in good agreement with the findings of Bezanson et al. (2016). These findings are not particularly surprising: when IRAC photometry is of good quality it will help constrain the stellar populations of a galaxy, particularly at high redshift, thus adding IRAC fluxes should improve photo-z estimates as long as the quality of the photometry is sufficient. For faint galaxies it is much more challenging to extract reliable IRAC fluxes and hence including them in the photo-z fit might lead to incorrect solutions.

Finally, Fig. A.4 shows the fraction of outliers for the same photo-z runs as in the previous two figures. The outlier fraction has here been defined as galaxies having |(z_MUSE−pz) / (1 + z_MUSE)| > 0.15 and we see a strong increase in the outlier fraction towards fainter magnitudes for the S14 original and reprocessed photo-zs, in good quantitative agreement with Bezanson et al. (2016)’s study using grism redshifts. This appears to be a reflection of the increased scatter at fainter magnitudes seen in Fig. A.4.

In summary then, we can see that adding more bands can lead to a smaller bias, but the increased scatter and related high outlier fraction at faint magnitudes means that we prefer the R15 catalogue for the present study as this performs better with 11 bands than the S14 catalogue with 44 at faint magnitudes. At bright magnitudes we are limited by the number of objects with z_MUSE, but data are consistent with the statement that the increased number of filters in the S14 catalogue leads to better performing photo-zs than the R15 catalogue. This would also agree well with the findings of Bezanson et al. (2016) for the same field.

Appendix B: The significant (“catastrophic”) outliers

Finally, it of interest to comment on the galaxies for which the photo-zs are in strong disagreement with the spectroscopic redshift. Here we will define two different classes of these outliers. We will refer to the galaxies for which |(z_MUSE−pz) / (1 + z_MUSE)| > 0.5 as catastrophic absolute outliers and those galaxies for which |(z_MUSE−pz) / (1 + z_MUSE)| > 5σ_MAD(x), where σ_MAD(x) is the local standard deviation at auxiliary quantity x, as catastrophic relative outliers.

The set of relative outliers does depend on the quantity which you calculate the local scatter against, x in the defintion above. In practice, however, this is not a strong effect,

In Table B.1 we provide a summary of the two types of catastropic outliers. For the absolute outliers, we took all galaxies brighter than F775W = 30 and 27 respectively, whereas for relative outliers we used all galaxies to F775W = 30 and we have excluded the 11 galaxies that have no BEAGLE photometric redshift. It is clear from the table that BEAGLE and BPZ perform best with a catastrophic outlier fraction of < 1.3% for galaxies brighter than F775W = 27 although we caution that low-number statistics start to play a significant role here. The last two columns show the number and fraction of outliers where all codes report photo-zs strongly in disagreement with the spectroscopic redshift and it is clear that this is lower by a factor of ~ 4 than the absolute outlier fraction for each code. The implication of this is that combining different estimators for identification of catastrophic outliers ought to be a good way to identify these.

The catastrophic relative outliers present much the same picture. Again BEAGLE and BPZ perform clearly better than BPZ and clearly better than EAZY, and again we see a clear potential for a gain from combining methods.

Finally, let us briefly discuss the nature of the outliers. For concreteness, we will focus on the most problematic objects, namely those that are always both catastrophic relative and absolute outliers, regardless of code. There are a total of 6 of these objects and we are interested in exploring whether the failure of all codes to fit these are due to problems with the data, or to intrisically add properties of the sources. The narrow-band images over the strongest line and the HST images are shown in Fig. B.1.

RAF 4510, MUSE ID 7213. This is a F775W = 27.9 object with a clear narrow-band image over Ly-α and z_MUSE = 6.385. The photometric redshift estimates from BPZ (pz_BPZ = 1.28), EAZY (pz_EAZY = 1.28), and BEAGLE (pz_BEAGLE = 1.36) are all in close agreement but far away from the spectroscopic redshift. The Ly-α source is however not overlapping with the photometric object, and the only photometric source (RAF 4510) close to the Ly-α peak is visible also in F435W, ruling out a z> 6. Thus in this case, the most reasonable explanation is that the photometric object is not associated to the fairly clear Ly-α line.

RAF 4892, MUSE ID 7285. This object has F775W = 29.0 and z_MUSE = 5.496. This is in strong disagreement with the photometric redshifts which are pz_BPZ = 0.76, pz_EAZY = 0.86, and pz_BEAGLE = 0.99. In this case the photometric object co-incident with the Ly-α narrow-band image is clearly seen in the F775W image but is not convincingly seen in the F606W image, although it has a catalogue flux of F606W = 30.31, and it is entirely absent in F453W. The source is in a crowded region and the F606W photometry might be is incorrect, leading to the mismatched photometric redshift.

RAF 9916, MUSE ID 1504. This F775W = 26.24 object has z_MUSE = 3.609 and pz_BPZ = 0.25, pz_EAZY = 0.41, and pz_BEAGLE = 0.46. The most prominent feature in the narrow-band image over Ly-α is the strong emitter with MUSE ID 6878 (RAF 7843) in the lower left. There is a clear line also over RAF 9916 and it is aligned with the object seen in the HST image. The bright source in the image is MUSE ID 6877 (RAF 9958) at z_MUSE = 0.734 In this case it is entirely possible that the Ly-α seen over RAF 9916 belongs to an extended emission region around MUSE ID 6678 and is unrelated to RAF 9916.

RAF 10185, MUSE ID 399. This object has F775W = 28.89 and the Ly-α narrow-band image is well aligned with the HST source. It has z_MUSE = 5.136 and pz_BPZ = 0.61, pz_EAZY = 0.56, and pz_BEAGLE = 0.59. The line is strong with a pronounced asymmetric shape and seems highly unlikely to be [Oii] λ3727 (which also would be in disagreement with the photo-zs). There is residual light in the F606W image which in the rest-frame of the source would have a central wavelength of 960 Å. The emission seen in F606W, assuming the spectroscopic redshift is correct, would seem to require a very transparent IGM for this source and this might be the cause of the photometric redshift discrepancy.

RAF 10218, MUSE ID 83. This object has F775W = 26.09 and has been given a redshift of z_MUSE = 1.149 in the udf-10 spectrum. The photometric redshifts all cluster near z = 3, with pz_BPZ = 2.86, pz_EAZY = 3.02, and pz_BEAGLE = 2.87. The narrow-band image over the putative [Oii] λ3727 line is not convincing and upon inspection of the spectrum, it seems to have Ly-α absorption just at the blue edge of the spectrum and a tentative detection of Si ii λ1526 and C iv λ1548,1550 which would give z = 2.937 in much better agreement with the photo-zs. Thus in this case it is arguably the spectroscopic redshift that is incorrect.

RAF 24953, MUSE ID 6702. This object has z_MUSE = 3.713 and pz_BPZ = 0.12, pz_EAZY = 0.10, and pz_BEAGLE = 0.86. It is fairly faint at F775W = 28.86 and has a slightly asymmetric emission line that has been identified as Ly-α. Identifying this line as [Oiii] λ5007 would bring the spectroscopic redshift in reasonable agreement with BPZ and EAZY but there is no sign of [Oiii] λ4959 nor any other lines and they should be easily observed given the strength of the one line seen, this seems highly unlikely. Thus the spectroscopic redshift is almost certainly correct and it is unclear what has caused this strong discrepancy.

All Tables

All Figures