Probing the faint end of simulated galaxy counts at z>3

Abstract

Simulations and observations now probe comparable redshift regimes with unprecedented accuracy, enabling direct consistency tests through forward modeling. In a previous work, we identified a faint-end discrepancy between observed and simulated near-infrared galaxy counts in CANDELS GOODS-South. Here we investigate whether this tension originates from the forward-modeling procedure or from limitations of the underlying simulations, and we characterize the galaxy populations responsible for the tension. Using the FORECAST forward-modeling code, we generated ten independent light-cone realizations and mock CANDELS images from the TNG100 and EAGLE simulations. We compared both the intrinsic light-cone catalogs and the mock-image detections with observations, testing dependencies on field and redshift, and validating the pipeline through stellar mass and multi-band analyses. The faint-end deficit is present in all CANDELS fields and appears at z>3 in both simulations. GOODS-South counts corrected for completeness exceed intrinsic simulation counts already at the 50% completeness limit, indicating that the missing population is not simply hidden below the detection threshold. Increasing the depth of mock images recovers the counts near the peak but overpredicts the faintest sources, showing that depth alone cannot resolve the discrepancy. Structural analyses reveal that compact galaxies with bright central cores observed in GOODS-South are underproduced in simulations, which instead favor diffuse low-surface-brightness systems. We conclude that the discrepancy arises both from detection losses of diffuse galaxies and, more fundamentally, from the inability of current hydrodynamical simulations to produce enough faint compact galaxies at z>3. This tension points to the need for improved modeling of early star formation, feedback, and dust treatment.