Speaker
Description
SBI wih Euclid gives us the potential to perform accurate parameter estimation and model comparison in the field-level setting, allowing us to better determine the true nature of dark energy (Spurio Mancini et al. 2024).
SBI, however, hinges upon both accurate forward modelling as well as compression due to the high-dimensional nature of fields. On the forward modelling side, we present work where we found that the inclusion of anisotropic variable seeing depth in the forward model can have up to a 10% shift on the value of the maximum posteriori (MvWK et al. 2024).
On the compression side, methodologies have typically been based on neural compression (e.g. CNNs) or statistical compression (e.g. scattering transforms) (Gatti et al. 2024, Jeffrey et al. 2024, Cheng et al. 2024). For SBI however, often two stages of compression are applied. This sometimes vastly increases the simulation budget required to perform SBI. I will touch upon recent work (Lin et al. 2024) that makes use of scattering transforms applied directly to cold dark matter fields without any further compression for SBI.
Furthermore, Euclid will provide wide field data whilst past SBI works and associated compression methodologies have focussed largely on flat sky data. I will present recent developments and corresponding codes for spherical CNNs and scattering transforms applicable to wide field Euclid data.
