PHYSTAT Seminar: Profiling systematic uncertainties in Simulation-Based Inference with Factorizable Normalizing Flows

by Davide Valsecchi (ETH Zurich (CH))

Wednesday 1 Jul 2026, 16:00 → 17:15 Europe/Zurich

This 45'+30' seminar is about the paper https://arxiv.org/abs/2602.13184

 

Generative machine learning has opened new frontiers for statistical inference and high-dimensional analysis in High-Energy Physics. Yet a central computational obstacle remains. Fully unbinned likelihood fits are often prohibitive, because profiling continuous systematic uncertainties across highly multidimensional phase spaces is both statistically and computationally demanding. Moreover, current methods still target a single parameter of interest rather than full differential distributions. In this talk I present a Simulation-Based Inference framework that addresses both problems (arXiv:2602.13184). Its core ingredient is Factorizable Normalizing Flows (arXiv:2606.30489), a density-morphing technique that represents systematic variations as parametric deformations of a nominal density while avoiding the combinatorial explosion of the joint parameter space. Building on this, an amortized training strategy learns in a single pass how the multivariate Distribution of Interest (DoI) changes with the nuisance parameters. This removes the need for repetitive network retraining during likelihood scans, making fully unbinned profiled inference feasible under realistic conditions with many systematic variations. We demonstrate the approach on a synthetic HEP-style analysis workflow that delivers a complete statistical and systematic uncertainty budget within a single unbinned likelihood.

---

26-07-01 PHYSTAT - Profiling SBI with FNF.pdf