Beschreibung
One of the major frontiers for Stage IV galaxy surveys is photometric redshift estimation. For weak lensing cosmology, the required characterization of redshift distributions is an order of magnitude more precise than current state-of-the-art. Because the true population is not directly accessible to us, every photo-z method relies - implicitly or explicitly - on a prior over galaxy SEDs, whether through physical modeling of the galaxy population or as imprinted in the training data of machine learning algorithms.
While advances are being made both in population modeling (on the physically motivated side, as in recent e.g. SPS-based forward models, as well as in advanced ML algorithms) and the gathering of large, representative spectroscopic calibration sets (e.g. with DESI and 4MOST), current methods still fail to reach Stage IV precision even under ideal conditions. Closing this gap will need a prior that is as flexible and agnostic to the population as possible, as any restrictive parameterised assumption ultimately limits the achievable accuracy. Generative models can fulfill this role as an expressive realization of such a prior in the context of a full forward-model of the galaxy population.
To address this challenge, we propose an entirely data-driven approach for forward-modeling the galaxy population. We train a generative model to learn the joint distribution of compact representations of rest frame galaxy spectra and redshifts. As a proof-of-concept, we train and validate on simulations, which enables direct comparison to the ground truth, and show that we reproduce the color-redshift relation to the required accuracy. Specifically, the learned population can reach a mean redshift bias below the per-mille level in realistically color-selected tomographic bins. This demonstrates that generative modelling is a promising path towards a Stage IV galaxy population prior.