Speaker
Description
The Non-Poissonian Template Fitting (NPTF) likelihood is widely used for inferring properties of unresolved point sources (PSs) in counts based data, such as those observed by the Fermi Gamma-Ray Space Telescope. I will show that the NPTF likelihood is generically overconfident, i.e. producing narrower confidence intervals / posteriors than expected. I will demonstrate that this effect is primarily due to NPTF not capturing correlations between neighboring pixels caused by PSs under finite point spread functions, and can be understood and corrected in simple cases. For realistic fits that such those concerning the Galactic Center Gamma-Ray Excess, I will argue that NPTF's overconfidence persists, and that Simulation-Based Inference can be a well-calibrated alternative.
