KiDS-SBI: Simulation-based inference analysis of KiDS-1000 cosmic shear

Abstract: We present a simulation-based inference (SBI) cosmological analysis of cosmic shear two-point statistics from the fourth weak gravitational lensing data release of the ESO Kilo-Degree Survey (KiDS-1000). KiDS-SBI efficiently performs non-Limber projection of the matter power spectrum via Levin's method, and constructs log-normal random matter fields on the curved sky for arbitrary cosmologies, including effective prescriptions for intrinsic alignments and baryonic feedback. The forward model samples realistic galaxy positions and shapes based on the observational characteristics, incorporating shear measurement and redshift calibration uncertainties, as well as angular anisotropies due to variations in depth and point-spread function. To enable direct comparison with standard inference, we limit our analysis to pseudo-angular power spectra. The SBI is based on sequential neural likelihood estimation to infer the posterior distribution of spatially-flat $\Lambda$CDM cosmological parameters from 18,000 realisations. We infer a mean marginal of the growth of structure parameter $S_{8} \equiv \sigma_8 (\Omega_\mathrm{m} / 0.3)^{0.5} = 0.731\pm 0.033$ ($68 \%$). We present a measure of goodness-of-fit for SBI and determine that the forward model fits the data well with a probability-to-exceed of $0.42$. For fixed cosmology, the learnt likelihood is approximately Gaussian, while constraints widen compared to a Gaussian likelihood analysis due to cosmology dependence in the covariance. Neglecting variable depth and anisotropies in the point spread function in the model can cause $S_{8}$ to be overestimated by ${\sim}5\%$. Our results are in agreement with previous analysis of KiDS-1000 and reinforce a $2.9 \sigma$ tension with constraints from cosmic microwave background measurements. This work highlights the importance of forward-modelling systematic effects in upcoming galaxy surveys.