NEPAL

We report the search of lepton flavour violating \( \tau \to \ell \phi\)decays at the Belle II experiment, based on an integrated luminosity Ldt = 189.88 fb−1 of data collected from 2019 to 2021, at the Υ(4S) meson resonance energy of 10.58 GeV. We look for signatures that could result from \( e^+e^- \to \tau^+ \tau^-\) events, by reconstructing one tau going into a charged lepton and a meson \(\phi \to K^+ K^- \), while the oppositely charged tau is left unconstrained to maximise the signal selection efficiency. Using MonteCarlo siulation, we defined a background suppression strategy that combines cut-based selection and boosted decision trees, taking as input kinematics, topological and other event variables. These selections are tunes in order to mininise as much as possible the expected upper limits on the branching fractions. Although the full data is not yet unblinded, we perform an estimation of the expected limis on \( \tau \to e \phi \) and \( \tau \to \mu \phi \) branching fractions of the order of \(10^{-7}\). At equivalent luminosites, our result on the muon decay mode is comparable to the expected limit of the previous Belle experiment analysis (+4%) and slightly better than the one from the Babar experiement (-13%). In addition, we conduct a measurement of the rate of fake tracks in \( e^+e^- \to \tau^+ \tau^-\) events. To this end, we devise an original method applied on tau pairs decaying into four tracks, with the possible presence of additional fake one. This rate, measure in data and simulation, is found to be around 1% in both cases.

This thesis presents the search for lepton flavour violating \( \tau \to \mu \mu \mu\) decays using \( e^+ e^- \to \tau^+\tau^- \) events at the Belle II experiment. The srtategy is based on an 'untagged' reconstruction of one tau going into three muons, while the opposite charged tau is left unconstrained to maximize the signal efficiency. A three-step strategy is adopted to reject background based on muon identification variables, cut-based selections and bossted decision trees, which take kinematic, topological, and other event-related variables as inputs. The background rejections is optimized using Monte Carlo simulated samples to minimize the upnzi figure-of-merit. After determining the number of expected events in data and the systematic uncertiantes, the expected limit on the \( \tau \to \mu \mu \mu\) branching fraction is expected to be 1.6 \( \times 10^{-8}\) at 90% confidence level using 424 \(fb^{-1} \). The previous Belle limit of 2.1 \( \times 10^{-8}\) is thus expected to be improved with a data sample twice smaller. In addition, the thesis also includes the measurement of the vertex detector cluster spatial resolution using overlapping sensors.