Speaker
Description
Stellar streams are very old, dynamical objects consisting of a collection of stars that originate from tidal disruptions of a globular cluster. In a galaxy like the Milky Way, these systems have the potential to be an extremely sensitive probe of dark matter substructure, baryonic physics, and the evolution history of the stream. In principle, this can be achieved by combining high precision observations at facilities such as Gaia, and consistent modelling of these stellar orbits - which typically trace out a large portion of the galactic gravitational potential.
On the other hand, making statistically robust statements about quantities of interest - e.g. the dark matter subhalo mass function - is extremely challenging. To do so requires some sort of marginalisation over the dynamical history and initial conditions of the stream, its stochastic interactions with dark matter substructures, as well as a reasonable model for foreground effects in the observations. Purely as a result of the huge number of free parameters this introduces, classical statistical methods scale very poorly and must rely instead on constructing bespoke summary statistics or ignoring a subset of effects in the modelling.
In this talk, I will present preliminary results for the GD-1 stellar stream using a powerful new approach within the framework of simulation-based inference: truncated marginal neural ratio estimation (TMNRE). This approach is based on the targeted training of high-precision neural networks for parameter inference, and scales to highly complex models. After introducing the methodology, we will show how this can be applied concretely to real Gaia data for GD-1 to obtain limits on fundamental properties of dark matter and the formation of substructure in galactic halos.
