Quantum interference between signal and background Feynman diagrams produce non-linear effects that challenge core assumptions going into the statistical analysis methodology in particle physics. I show that for such cases, no single observable can capture all the relevant information needed to perform optimal inference of theory parameters from data collected in our experiments. I then discuss an optimal data analysis strategy for the Higgs width measurement with high-dimensional statistical inference enabled by neural networks. This Neural Simulation-Based Inference (NSBI) approach naturally handles high dimensional data, avoiding the need to design low-dimensional summary histograms. Next, we design a general purpose statistical framework in the ATLAS experiment that enables the application of NSBI to a full-scale physics analysis, leading to the most precise measurement of the Higgs width by the experiment to date. This work develops several innovative solutions to introduce uncertainty quantification and enhance robustness and interpretability in NSBI. The developed method is a generalisation of the standard frequentist statistical inference framework used in particle physics, and is applicable to a wide range of physics analysis. This method enables the design of analysis that were not feasible before due to the computational complexities and investment in human time required.Join from PC, Mac, Linux, iOS or Android: https://stanford.zoom.us/j/98973156241?pwd=cEU5RFdlVXoyc0JTeTlDMkozKzQ5UT09
Peter Gaemers (pgaemers@slac), Sayan Ghosh (sghosh92@slac), Jamie Ryan (jlryan@slac)
