PhD Defense

“Bayesian Computational Statistics for the study of galaxy Spectral Energy Distributions” by Grégoire Aufort 

Abstract :
The exploitation of new observation tools in astrophysics requires the development of new statistical tools in order to be effective and precise. We are particularly interested Bayesian statistics for the study of the Spectral Energy Distributions of galaxies.
After an introduction to the analysis of spectral energy distributions, the first part of this thesis proposes a Machine Learning based Approximate Bayesian Computation algorithm (ABC) for the choice of Star Formation History models from data photometric. The method is applied to data from the COSMOS survey for the identification of galaxies whose star formation rate has undergone a violent alteration in the near past.
The second part of the thesis proposes a new Multiple Adaptive Importance Sampling algorithm: TAMIS (Tempered Anti-Truncated Multiple Importance Sampling). By introducing a sequence of self-calibrated auxiliary target distributions TAMIS is robust to the curse of dimensionality as well as poor initialization, and requires relatively few evaluations of the target density, without using its gradient.
The third part presents the CIGALE code used for modeling the spectral distributions of galaxies from physical models. The calculation of the expected SED with known parameters is done in successive stages corresponding to the different physical processes at play. To speed up the computation, we propose the replacement of the explicit computations of certain steps by a neural network approximation.
Finally we presents a complete statistical model for the Bayesian parameter inference and model choice using both photometric and spectroscopic data, and the application of TAMIS to this specific problem.

Jury:
Nicolas CHOPIN, ENSAE, Président du jury
Sylvain LE CORFF Institut Polytechnique de Paris, Rapporteur
Marc HUERTAS-COMPANY Observatoire de Paris, Rapporteur
Véronique BUAT Aix-Marseille Université, Membre du jury
Florence FORBES INRIA, Membre du jury
Clotilde LAIGLE Institut d’Astrophysique de Paris, Membre du jury
Pierre PUDLO Aix-Marseille Université, Co-directeur
Denis BURGARELLA Laboratoire d’Astrophysique de Marseille, Co-directeur