Improving 3x2pt Cosmology Constraints: Training Sample Augmentation, Optimal Binning, and Neural Network Classifiers

Large imaging surveys of galaxies rely on photometric redshifts (photo-z’s) and tomographic binning for 3 × 2 pt analyses that combine galaxy clustering and weak lensing. We divide simulated galaxy catalogs into training and application sets, where the spectroscopic training set is non-representative in a realistic way, and then estimate photometric redshifts for the application set. Spectroscopic training samples for the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) will be biased towards redder, brighter, lower-redshift galaxies, leading to photo-z estimates with outlier fractions nearly 4 times larger than for a representative training sample. Training sample augmentation allows us to add simulated galaxies possessing otherwise unrepresented features to our mock spectroscopic training sample, reducing the outlier fraction of the photo-z estimates by 50% and the scatter by 56%. We sort the galaxies into redshift bins chosen to maximize the 3x2pt signal using a novel generalized binning parameterization introduced by Moskowitz et al. (2023, ApJ 950, 49). Applying a neural network classifier trained to identify galaxies that are highly likely to be sorted into the correct redshift bin improves the figure of merit by ∼13%, equivalent to a 28% increase in data volume.