Speaker
Description
Extracting a signal of interest (LSS, CMB, Galactic emission…) from contaminated observations is a central challenge in many astrophysical and cosmological analysis. However, instrumental systematics, Galactic foregrounds or other contaminants usually makes the forward process stochastic and non-invertible, leaving the inverse problem ill-posed. A probabilistic framework is therefore required to recover a distribution of signals compatible with the observed data. In a Bayesian setting this require to introduce a prior distribution for the signal of interest. However, specifying a physically-driven prior of complex non-gaussian processes is often difficult. In this work, we propose to use a maximum entropy generative model parametrised by scattering transform (ST) statistics for the signal of interest. Leveraging these ST-based generative models, we develop an iterative algorithm to estimate a posterior distribution of maps which are solution of the inverse problem. We validate our approach on large-scale structure, weak lensing and turbulence fields, under a challenging forward model including noise, beam, and masks. We show that our approach recovers key astrophysical statistics like the power spectrum, one point PDF and Minkowski functionals.