A New Estimate of the Galaxy Luminosity Function, Using Machine Learning and a Mock Catalogue

Abstract

A new measurement of the galaxy luminosity function (LF) is presented over the redshift range using data from the Physics of the Accelerating Universe Survey (PAUS). Leveraging the high photometric redshift precision () made possible by PAUS's 40 narrow-band optical filters, rest-frame magnitudes are derived and the LF is estimated in multiple redshift bins using the method. The analysis is supported by a realistic mock catalogue constructed from the GALFORM semi-analytic galaxy formation model. To compute rest-frame magnitudes for observed galaxies, a Random Forest regression model was trained to predict -corrections from observable properties. The mock was used to investigate the impact of photometric uncertainties and redshift errors. This revealed that these observational effects significantly modify the shape of the LF—especially at the bright end. A detailed error analysis revealed that photometric and photometric redshift errors dominate the LF uncertainty, contributing errors nearly an order of magnitude larger than those from large-scale structure across all redshift bins. Importantly, the observed faint-end turnover in the LF—driven by selection in the observed -band—can still be used to constrain galaxy formation models when the same selection is applied to the simulated data. The resulting luminosity functions in the -band and other rest-frame bands show good agreement with theoretical predictions from GALFORM. The LF is also measured separately for red and blue galaxies, revealing distinct evolutionary trends. Symbolic regression models developed by collaborators are used to estimate stellar masses and measured the stellar mass function from both PAUS and the mock sample. This thesis highlights the value of combining narrow-band photometric surveys and machine learning methods to probe galaxy evolution with precision, and demonstrates the usefulness of mocks to model selection effects.