Background Modeling in Partial-Wave Analysis of $\tau^- \to \pi^+\pi^-\pi^- \nu_{\tau}$ Decays at Belle II

Sumitted to PubDB: 2026-06-24

Category: Master Thesis

Tags:

Principal Authors Miriam Weiskopf, Stefan Wallner, Hans-Günther Moser, Stephan Paul, Godo Kurten
Date 2026-05-18
Belle II Number BELLE2-MTHESIS-2026-051
Abstract The decay $\tau^- \to \pi^+\pi^-\pi^- \nu_{\tau}$ provides an important laboratory for studying the strong interaction and light-meson resonances using the technique of partial-wave analysis (PWA). A major challenge in this analysis is the treatment of background contribu- tions, which exhibit highly correlated multidimensional phase-space distributions. In this thesis, we investigate neural-network-based background modeling methods and in- tegrate them into the PWA framework. We study different neural network approaches and evaluate their performance in partial-wave decomposition fits on simulated pseudo data. We find that strong regularization and kinematic restrictions to signal-populated phase-space regions are necessary to avoid unstable neural-network predictions. Further- more, we compare different training strategies and examine the impact of dividing the full invariant-mass range into multiple training regions for the neural networks. We find that training on overlapping invariant-mass regions (so-called sliding-window approach) yields the most stable and accurate background description. In addition, the analyzed data contains several distinct background contributions. To model these, two model- ing approaches are investigated: a combined approach, where all background contribu- tions are modeled simultaneously, and a split approach, where individual background contributions are modeled as separate components. Their performance is validated through complementary input-output studies. We find that the combined approach yields stable background descriptions and reproduces the input in the signal partial waves well. In contrast, the split approach exhibits significant leakage into the dominant $\tau^- \to \pi^+\pi^-\pi^- \nu_{\tau}$ background component. However, this effect can be mitigated by fixing its yield to the true Monte Carlo value. Additional supporting studies show that neglecting detector resolution introduces only small biases in the signal description. We further investigate biases caused by overlapping partial waves and find that these effects can be reduced by excluding individual partial waves in weakly constrained invariant- mass regions. Overall, the presented studies demonstrate that neural-network-based background modeling can successfully be integrated into the partial-wave decomposition framework and provide a foundation for future partial-wave analyses at Belle II.
Institute MPP

Files