Coverage is not enough: Frequentist tests of simulation-based inference for primordial non-Gaussianity
The study tested standard validation methods for extracting primordial non-Gaussianity from dark matter simulations using simulation-based inference (SBI) and likelihood-based inference (LBI). Findings show that current coverage tests are insufficient because they don't capture the posterior shape at fixed parameter values. To achieve robust uncertainty quantification, diagnostics sensitive to how the posterior behaves at specific points are required.