LOR: The Galaxies Science Collaboration Photo-Z Use Case

Title: Galaxies Science Collaboration Photo-z Letter of Recommendation

Co-signers: Sam Schmidt, Peter Hatfield, Sugata Kaviraj, Manda Banerji, Brant Robertson, Vivienne Wild, Will Hartley, Katarzyna Małek, Michael Rutkowski, Ricardo Demarco, Henry Ferguson, Kanak Saha, Eric Gawiser, Biprateep Dey, for the Galaxies Science Collaboration

0: Summary Statement

This LoR summarizes the needs of the Galaxies Science Collaboration from Rubin Observatory-provided photo-z data products. A wide range of science cases will be enabled by diverse photometric redshift estimates, and the Galaxies SC efforts via in-kind contributions can help contribute to the goal of maximizing science returns from the Rubin PZ catalogs.

1: Scientific Utility

The Galaxies Science Collaboration will undertake a broad variety of science cases spanning a full range from local galaxies to some of the earliest galaxies to form in the Universe. Meeting the needs of such a diverse set of analyses with a general purpose photo-z table will be difficult. The 10-year WFD survey will extend galaxy samples to higher redshifts and fainter populations than previous ground-based surveys, including regimes which are not likely to be covered by spectroscopy with any near-future instruments. Thus, much of the discovery space will be reliant on the photo-z table, and any associated physical quantities included within; for novel explorations of these populations an algorithm that can provide reasonable estimates beyond the range covered by training data would be highly valuable.

Many studies do not care about the galaxy redshift in isolation, rather they will study the joint distribution of redshift and other properties, e.g. stellar mass, stellar populations, star formation rate, etc… In such cases, a joint estimation of the redshift and these parameters is needed, as they are often highly degenerate, and subject to bias if estimated separately. As such, the catalog is more useful if there is some joint estimate of the SED/”type” of each object. For example, including an SED/type that enables the selection of red sequence galaxies would enable “stacking” of this population at a series of redshifts, revealing low surface brightness features. Red galaxies are also particularly useful in identification of clusters, which can be employed in a variety of science cases. Cluster identification requires accurate photo-z’s in order to identify overdensities as candidate clusters. As another example, galaxy evolution models require accurate estimates of small-scale spatial clustering to constrain HOD and baryonic feedback models. Hatfield et al (2019) estimate the impact of template-based redshift and physical parameter estimates on the clustering estimates, finding that they can bias the HOD and halo mass estimates.

The Galaxies SC has several codes that will be developed via in-kind contributions that show promise in producing catalogs that fulfill the scientific needs of the Collaboration. DEmP (Hsieh & Yee 2014) and GPz (Almosallam et al. 2016) are machine learning-based codes that are able to produce robust estimates of redshift and associated physical parameters (when provided with adequate training data). Coordination of in-kind efforts with the planned Photo-z Validation Cooperative should provide benefits to both groups, and such cooperation will be the aim of the in-kind efforts as they are developed. Separate LoRs should be submitted for the two specific codes mentioned above.

2: Outputs

The bare minimum set of outputs needed to execute a portion of the Galaxies SC-related science programs is a point estimate redshift and a confidence interval. A more desirable set of outputs that enables the majority of the Galaxies science programs is a photo-z PDF or likelihood distribution for each galaxy, along with a point estimate with uncertainty bounds. Even in the absence of physical parameter estimates, PDF estimates may be used with outside codes to roughly estimate physical parameters (stellar mass, SFR, rest frame colors, host E(B-V), etc…) as a post-processing stage. Even more desirable is a set of outputs that would include physical parameter estimates and error intervals derived self-consistently and simultaneously using the same code and assumptions that were used to derive the redshift PDF. This avoids many potential biases that may arise from the approximation that comes from estimating a marginalized redshift PDF first and then backing out the physical parameters separately. The ideal outputs would be full joint PDFs for redshift, stellar mass, and other physical parameters, though limited compute and storage space make this ideal scenario unlikely for the DM Table. A potential solution to storage issues is to include a method such as Mucesh et al (2021) to generate multidimensional PDFs on the fly. In the end, some method to encode the template/model that was used to derive the physical parameters would be almost essential to users, and some encoding of the specific SED/model should be included in the outputs for each galaxy, along with any photometric offsets, corrections to filters, etc…, i.e. all metadata that was used in constructing the photo-z estimate should be considered as a required photo-z output, and including at least a best fit template.

Many template-based codes that produce robust estimates of galaxy physical properties with error intervals are very computationally expensive, likely limiting their utility to produce estimates for full WFD catalogs; however, a possible two-stage setup where template codes are run on a smaller, but likely representative set of data in order to generate parameter estimates that can be fed into a faster machine learning-based estimator could produce estimates of M*, SFR, and other physical parameters at a manageable computational cost. Such a setup is being considered in at least one other Stage IV survey. There is ample expertise in the Galaxies SC with codes such as BAGPIPES and CIGALE for joint photo-z/physical parameter fitting, generation of supplemental training data, and/or verification tests on subsets of data, and can likely assist the Photo-z Validation Cooperative in utilizing these more computationally intensive codes on a subset of data within the DM pipeline. In addition to the photo-z estimates themselves, access to some of the data products used internal to the algorithms may prove to be very useful: access to template sets, priors, and training data employed by the photo-z codes could enable additional validation checks, extensions to analyses, and avoid potential biases.

If computational and storage costs allow for multiple algorithms to be stored in the DM PZ Table, diversity of algorithms chosen would be highly preferable: for example, a template fitting code and a ML-based code, or one code with strong priors and another with fairly minimal priors. Given the wide range of science, diversity in the outputs should enable a wider range of use cases.

With billions of objects in the Table, any flags that enable efficient selection of desired samples are extremely helpful. Star/galaxy separation flags (particularly if stellar templates at z=0 are included in redshift estimates), flags associated with corrupted photometry (due to satellite streaks, scattered light, etc…), and any flags that can indicate the potential for catastrophic redshift failure, would be very useful. A “probable AGN” flag to identify active galaxies is desirable, and may become more feasible later in the survey when variability enables better identification (the AGN SC should have input on the optimal way to construct such a flag). A “probable blend” flag to eliminate potentially corrupted galaxy colors is also desired. Boolean/integer flags are a minimal requirement for functional selections, though probabilistic flags may be preferable if they are available and not computationally expensive. Finally, several members brought up the fact that, given the steep number counts relation, many galaxies in the WFD releases will be very faint and have low quality photo-z. Some form of minimal “good quality” flag (which could be as simple as being defined by a magnitude cut) could be very useful for those who are not as closely affiliated with a large SC and are grabbing Rubin data for individual projects. We should do our best to make it easy for users to identify and remove low quality photo-z’s from samples.

3: Performance

The Galaxies SC will not list any quantitative requirements on the photo-z outputs, though we note many science cases have a higher tolerance for bias than the strict requirements needed for cosmology measures. The Galaxies SC has a general preference for meaningful uncertainty estimates over even specific metric performance, i.e. codes that score slightly worse in terms of bias, scatter, and even catastrophic outlier rate would be preferred if the redshift uncertainties were more accurate than codes that report better metrics.

4: Technical Aspects

A portion of the discussion while formulating this LoR was on the subject of how DM could include mechanisms for updating the PZ algorithms and procedures as the survey progresses. As with all projects, as more data begin to flow, we expect to find oddities in the data, and potentially identify new flags or new algorithm improvements. Having mechanisms in place to receive such feedback and have science users contribute to incremental pipeline improvements will be beneficial for all involved.

While likely beyond the scope of the LoR, we note that other wide field surveys to be undertaken by the Roman Space Telescope and Euclid have the possibility of supplementing multiwavelength data that could greatly improve both redshift and physical parameter estimates, as well as allow for selections of more distinct populations (e.g. AGN or strong lens candidates). Incorporating such data where available for large portions of the survey could add scientific utility. It may also be valuable to have a code available on the Rubin Science Platform that enables users to include supplemental wavelength data in running a custom photo-z for a small subset of data. In short any considerations that facilitate the combination of Rubin data and supplemental bands for redshift and physical parameter estimation would likely prove beneficial.