Beschreibung
Weak gravitational lensing and photometric galaxy clustering are the two pillars of large-scale-structure cosmology, and both rest on one fragile assumption: that we know the redshift distribution, $n(z)$, of our galaxy samples. In this talk I show that this assumption breaks down in subtle but consequential ways, and present forward-modelling methods to fix it. Galaxies blended in the focal plane carry shear from multiple redshifts, distorting the effective distribution $n_\gamma(z)$ that lensing actually measures; I introduce an emulator of the pairwise shear response to neighbouring galaxies, built from dedicated HSC-like image simulations, that corrects this mixing and recovers a high-redshift tail otherwise biased low by ~0.05. On the clustering side, observational systematics make the lens $n(z)$ spatially anisotropic rather than the isotropic distribution standard analyses assume. Propagating this through a full forward model — realistic systematics maps, an ML-emulated selection function, and tomographic binning — into a 2×2pt analysis, I find the angular signal underestimated by up to ~10% at small scales, and for LSST-like depth $\sigma_8$ biased low by 1.7$\sigma$ and $\Omega_m$ high by 1.4$\sigma$, largely absorbed by a ~4$\sigma$ shift in high-redshift galaxy bias. These redshift-distribution systematics are a minor nuisance today but a serious threat for Stage-IV surveys, and tractable forward-modelling pipelines can mitigate them.