Wing Sheung Chan

Conclusion and outlook “Would you tell me, please, which way I ought to go from here?” “That depends a good deal on where you want to get to.” “I don’t much care where —” “Then it doesn’t matter which way you go.” “— so long as I get SOMEWHERE” “Oh, you’re sure to do that, if you only walk long enough.” — Lewis Carroll, “Alice” and “the Cheshire Cat”, Alice’s Adventures in Wonderland In this thesis, a search for lepton-flavour-violating Z → `τ decays is presented. The search is motivated by the fact that lepton flavour symmetry in the Standard Model is only accidental and approximate, and that any observation of Z → `τ decays would be an unambiguous signal of physics beyond the Standard Model. The large amount of high- energy proton–proton collision data collected by the LHC and the ATLAS detector, with approximately eight billion Z bosons produced in the second operational run, created the opportunity to search for these rare hypothetical decays with an unprecedented sensitivity. Although no significant evidence of Z → `τ decays has been found, stringent upper limits on the LFV branching fractions B ( Z → `τ ) have been set, which superseded the otherwise most stringent limits set by the LEP experiments more than two decades ago. Under the assumption that the LFV Z`τ interaction is parity-conserving, the following upper limits at 95% confidence level are set: B ( Z → eτ ) < 8 . 1 × 10 − 6 and B ( Z → µτ ) < 9 . 5 × 10 − 6 . Similar constraints are also set under different scenarios where the Z`τ interaction is assumed to be parity-violating, which, to the best of the author’s knowledge, has never been done before in similar experiments. The high sensitivity of the presented search does not only rely on the availability of a large data set, but also hinges on the analysis techniques. In the heart of the search are the neural network classifiers. While using neural network for event classification is not a novel idea in and of itself, it is rather unconventional for a rare-decay search like the one presented her e † to be heavily based on neural network classification. In the analysis, neural networks are used to exploit correlations in low-level kinematic variables, which are often neglected or not fully explored in conventional “cut-based” analyses or analyses that utilise boosted decision trees. An original approach of having multiple classifiers, † which is similar to a precision measurement 121