Probabilistic redshift estimation of unresolved galaxies from multi-band background light maps
Accurate knowledge of the redshift distributions of unresolved galaxy populations is essential for extracting the cosmological information encoded in the cosmic optical background. We present the first framework for estimating these distributions directly from multi-band maps of unresolved background light, using conditional normalising flows trained on realistic image simulations. For imaging qualities comparable to those expected from the synergy between the Euclid and Rubin LSST surveys, our model achieves sub-per cent accuracy in both the mean and standard deviation of the flux-weighted redshift distributions when the training and target samples are statistically well matched. Furthermore, the network proves highly resilient to incomplete photometric coverage when combined with tailored data imputation strategies. To address the challenge of statistical mismatches between simulations and real observations, we propose incorporating observational quantities into the target variables, providing a built-in diagnostic to assess model reliability and uncertainty from the predictions themselves. We also demonstrate the feasibility of map-based tomographic analyses, showing that the model retains sufficient pixel-level detail to recover redshift distributions for tomographically defined subsamples with near sub-per cent accuracy across all bins. These results establish conditional normalising flows as a viable tool for redshift estimation of unresolved sources, opening new opportunities for component separation, signal interpretation, and cosmological inference using multi-band optical background fluctuations.