Giorgos Saloustros

The promise of Simulation-Based Inference for cosmology

On the verge of the step-change in cosmological analysis represented by LSST/Vera Rubin Observatory, traditional inference methods require overhauling in order to deal with the large number of objects and subtle statistical and modelling effects that will otherwise dominate systematics. 

I present the case for Neural Ratio Estimation (NRE), a type of Simulation-Based Inference, in the context of supernova type Ia (SNIa) cosmology, showing that NRE matches traditional likelihood-based hierarchical Bayesian modeling on real data; removes systematics offsets due to linearization in large (~100,000) samples; performs Bayesian model selection at almost no additional computational cost; deals effortlessly with complex selection effects; enables sophisticated calibrations of posterior intervals and confidence regions thanks to its amortized nature. I will demonstrate how the power of SBI can be harnessed by conducting joint inference on supernovae type Ia and their host, obtaining much superior constraining power. Once fully integrated into the data analysis pipeline, NRE has the potential of becoming the tool of choice for SNIa cosmology in the 21st century.