Cosmology in the Einstein Telescope era: comparing traditional and simulation-based methods for population inference
The next generation of gravitational wave detectors, such as the Einstein Telescope (ET), will observe orders of magnitude more binary black hole mergers than current facilities. Most of these events will lack an electromagnetic counterpart, also known as dark siren events, yet will still enable percent-level cosmological constraints. However, the likelihood traditionally used in Hierarchical Bayesian Inference (HBI) for population-level analyses becomes computationally prohibitive as the size of dark siren catalogues and population parameters grow. In this work we compare HBI against simulation-based inference (SBI) as a scalable alternative for cosmological population inference in the ET era. Studying a proof-of-concept example of a mock ET inference, we build a catalogue of $O(10^4)$ binary black hole events, then perform inference on the Hubble constant $H_0$ and matter density $\Omega_m$ in a flat $\Lambda$CDM cosmology, using both a hierarchical analytical likelihood and Marginal Neural Ratio Estimation (MNRE). We find excellent agreement between the two approaches, with SBI reproducing the HBI posteriors to high accuracy, while requiring orders of magnitude less computation once the simulation and training cost is amortized. We further demonstrate that SBI extends straightforwardly to a joint cosmology-plus-astrophysics analysis, simultaneously constraining $(H_0,\Omega_m)$ together with the parameters of the star formation rate density, at negligible additional cost compared to the significant increase in complexity such an extension would require within the HBI framework. Our results indicate that SBI is a promising and scalable tool for population inference with third-generation GW detectors.