Disentangling the cosmic dipole with Bayesian and simulation-based inference
by
The cosmic dipole measured in surveys of distant extragalactic sources consistently disagrees with the expectation set by the cosmic microwave background, challenging the standard model of cosmology. The fidelity of these measurements rests on our treatment of survey systematics, many of which have an unknown analytical effect on the source counts, rendering the likelihood intractable. Applying neural ratio estimation to the cosmic dipole for the first time, we use simulation-based inference (SBI) to quantify tensions between Planck, NVSS, RACS, and CatWISE. We measure a ≈5.7σ tension between CatWISE and Planck, rising to ≈6.7σ once the CatWISE Eddington bias is forward-modelled. Our SBI architecture provides machinery for future analyses with LSST, Euclid, and the SKA, where modelling complex observational systematics will be essential.