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

Simulations and observations now probe comparable redshift regimes with unprecedented accuracy. This overlap allows us to test their consistency under the same observational conditions using the forward-modeling technique. In our previous work, we identified a faint-end discrepancy between the observed and simulated near-infrared galaxy counts in CANDELS GOODS-South. Here, we assess whether this mismatch arises from the forward-modeling process or from intrinsic limitations of the underlying simulations, and we characterize the galaxy populations that give rise to the difference. Using the FORECAST forward-modeling code, we generated ten independent realizations of light cones and mock images emulating the CANDELS fields from the TNG100 and EAGLE simulations. We compared the light-cone catalogs, that is, the input universe and the detections on mock images with observations, testing dependences on fields and redshift. We performed validation tests to assess the robustness of the forward-modeling pipeline, including checks on stellar mass, comparisons with independent mock datasets, and a multiband analysis of the spectral energy distribution modeling incorporating a CEERS band. The faint-end deficit was found in all CANDELS fields; it emerges at z≳3 in both simulations. After we corrected GOODS-South counts for completeness, they exceeded input universe counts already where the survey is 50% complete, indicating that the missing population is not simply hidden below the detection threshold. An artificial deepening of the mock images recovered the observed counts near the peak, but overpredicted the faintest counts, showing that depth alone cannot resolve the difference. The analysis of the structural parameters revealed that the faint compact galaxies with bright central cores in GOODS-South are underproduced in simulations, which are biased toward diffuse systems with a low surface brightness. The shortfall of faint sources partly originates from detection losses of diffuse galaxies with a low surface brightness, but more fundamentally, from the inability of current hydrodynamical simulations to produce enough faint compact galaxies with bright cores at z>3. This difference highlights the need to refine the modeling of early galaxy formation, in particular, the treatment of star formation and feedback, and to improve the treatment of dust in the forward-modeling process.