Optimizing the shape of photometric redshift distributions with clustering cross-correlations

We present an optimization method for the assignment of photometric galaxies to a chosen set of redshift bins. This is achieved by combining simulated annealing, an optimization algorithm inspired by solid-state physics, with an unsupervised machine learning method, a self-organizing map (SOM) of the observed colours of galaxies. Starting with a sample of galaxies that is divided into redshift bins based on a photometric redshift point estimate, the simulated annealing algorithm repeatedly reassigns SOM-selected subsamples of galaxies, which are close in colour, to alternative redshift bins. We optimize the clustering cross-correlation signal between photometric galaxies and a reference sample of galaxies with well-calibrated redshifts. Depending on the effect on the clustering signal, the reassignment is either accepted or rejected. By dynamically increasing the resolution of the SOM, the algorithm eventually converges to a solution that minimizes the number of mismatched galaxies in each tomographic redshift bin and thus improves the compactness of their corresponding redshift distribution. This method is demonstrated on the synthetic Legacy Survey of Space and Time cosmoDC2 catalogue. We find a significant decrease in the fraction of catastrophic outliers in the redshift distribution in all tomographic bins, most notably in the highest redshift bin with a decrease in the outlier fraction from 57 per cent to 16 per cent.

Automated summary

We propose an optimization method that combines simulated annealing with a self-organizing map to assign photometric galaxies to redshift bins. This method optimizes the clustering cross-correlation signal between photometric galaxies and a reference sample of galaxies with well-calibrated redshifts. By dynamically increasing the resolution of the self-organizing map, the algorithm converges to a solution that minimizes the number of mismatched galaxies in each redshift bin and improves the compactness of their redshift distribution. We demonstrate the effectiveness of this method on a synthetic catalog, finding a significant decrease in the fraction of catastrophic outliers in all redshift bins.