Presentation #101.04 in the session Cosmology.
Bayesian parameter estimation via nested sampling is a powerful tool for mapping the full posterior distributions of model parameters when fitting data. In the context of the global 21-cm signal, it has been shown that nested sampling analyses can utilize neural-network-based emulators for the signal to greatly speed up the sampling process, which can require a significant number of likelihood evaluations to achieve convergence. In this work, we fit simulated global 21-cm data generated by the semi-analytical model ARES and show for the first time that when using an emulator, in particular globalemu, as the signal model in the likelihood, the posteriors obtained are largely consistent with those obtained when using the full signal model in the likelihood. We train globalemu using a large training set of ARES signals in which we vary the eight astrophysical parameters of interest. We employ two leading nested sampling algorithms – MultiNest and PolyChord – and show that they converge on the same result, but with PolyChord requiring 28 times more likelihood evaluations than MultiNest for our 8-dimensional parameter space. Furthermore, we show that the posteriors significantly improve when jointly-fitting the global 21-cm signal and high-z galaxy UVLF mock data. We find that jointly-fitting the UVLF along with the global 21-cm signal successfully recovers the input parameter values used to generate the mock data, while reducing important covariances between parameters relevant for both experiments.