Practical relevance of relaxed support assumptions in generative unfolding

Determine whether the weaker Monte Carlo support assumptions of generative-model-based unfolding methods provide a substantive performance advantage over the classifier-based \Omnifold algorithm in realistic unfolding applications.

Background

Generative unfolding methods can make lighter assumptions about the support of the Monte Carlo simulations than classifier-based \Omnifold, but they require additional generative-model training and have not yet been applied to experimental data in the thesis’s survey. The practical importance of this difference remains unresolved: it is unknown whether the support assumptions made by \Omnifold are limiting in realistic cases and how much generative approaches would improve performance if such a case occurred.

References

These methods do have lighter assumptions than \Omnifold on the support of the MC simulations used in the training, but it is unclear whether these assumptions are ever actually limiting for \Omnifold and how much better the generative-model-based approaches would perform if such a case were found.

— A High- and Variable-Dimensional Measurement of the $Z$+jets Differential Cross Section with the ATLAS Experiment and Artificial Intelligence  (2608.28449 - Greif, 28 Aug 2026) in Section 5.4.2, “Generative Unfolding Methods”