A simulation-based inference pipeline for cosmic shear with the Kilo-Degree Survey

Abstract

The standard approach to inference from cosmic large-scale structure data employs summary statistics that are compared to analytic models in a Gaussian likelihood with pre-computed covariance. To overcome the idealising assumptions about the form of the likelihood and the complexity of the data inherent to the standard approach, we investigate simulation-based inference (SBI), which learns the likelihood as a probability density parameterised by a neural network. We construct suites of simulated, exactly Gaussian-distributed data vectors for the most recent Kilo-Degree Survey (KiDS) weak gravitational lensing analysis and demonstrate that SBI recovers the full 12-dimensional KiDS posterior distribution with just under $10^4$ simulations. We optimise the simulation strategy by initially covering the parameter space by a hypercube, followed by batches of actively learnt additional points. The data compression in our SBI implementation is robust to suboptimal choices of fiducial parameter values and of data covariance. Together with a fast simulator, SBI is therefore a competitive and more versatile alternative to standard inference.