TractorJAX: GPU-Accelerated SPHEREx Forced Photometry and the Trade-offs of Regularization and External Priors

Abstract

We present TractorJAX, an open-source GPU reimplementation of the catalog-forced mode of The Tractor. We use it to test whether regularization or external flux priors enable reliable fitting of reference catalogs that extend below the depth directly constrained by undersampled SPHEREx images. For full-depth regularized fits, TractorJAX fits $100\times100$-pixel cutouts at 11.9 ms each on a single NVIDIA H100 GPU. In simulated images of the Abell 2537 galaxy-cluster field, extending the fitted catalog from $m_z<21$ to the full DESI Legacy Imaging Surveys DR10 catalog increases the scatter in fractional flux residuals by 34%, even with regularization. We also test the methods using SPHEREx observations in nine fields, assessing the number of reliably measured sources and photometric-redshift accuracy. Requiring S/N $>5$ in more than half of the 102 channels, we find that the regularized full-depth fit yields 40% fewer sources than the truncated fit. For sources meeting this requirement in both fits, the normalized median absolute deviation of the photometric redshifts, $σ_{\rm NMAD}$, is 17% larger in the full-depth fit. External flux priors recover nearly as many sources as the $m_z<21$ fit, but increase $σ_{\rm NMAD}$ by 0.0015 for the common calibrated sample. Neither regularization nor the external flux priors tested here simultaneously preserve the sample size and photometric-redshift accuracy achieved by the truncated fit. Our results suggest that the optimal fitted-catalog depth cannot be determined from image-level performance alone, but should be selected through end-to-end tests tailored to the intended science measurement.