Wing Sheung Chan

Object reconstruction and identification 53 is a correction factor to p comb T . The predicted transverse momentum p MVA T is then used together with the directions ( φ and η ) from the substructure reconstruction to form the final calibrated τ had - vis momentum. The expected distributions of the response R = p reco T p truth T , (3.6) where p reco T is the reconstructed p T in one of the TESs, are shown in Figure 3.2 for τ had - vis in different modes. From the figures, it can be seen that the MVA TES incorporates both the better core resolution of the pantau TES and the better tail behaviour of the calo TES. It also corrects for the biases of the calo TES. As a figure of merit, the relative resolution (of 68% coverage) is defined as Resolution (68% coverage) = Q 0 . 84 ( R ) − Q 0 . 16 ( R ) 2 , (3.7) where Q x ( R ) is the x th quantile of R . The expected relative resolutions of the calo, pantau, combined, and MVA TESs are shown in Figure 3.3 as functions of p truth T . At p T < 100 GeV , the MVA TES offers a significantly better resolution than both the calo and pantau TESs. The resolution is improved by a factor of two with respect to the calo TES in this region. At p T 100 GeV , the resolution of the MVA TES approaches that of the calo TES. The improved energy resolution provided by the MVA TES directly impacts the sensitivity of the search for Z → `τ decays as it affects our ability to separate the signal from backgrounds like Z → τ τ . 3.2.4. Jet rejection The reconstruction of τ had - vis candidates provides very little rejection against quark- or gluon-initiated jets. Algorithms have been developed to distinguish real τ had - vis from other jets. These algorithms use MVA to exploit features of the tracks and clusters in both the core and isolation regions, as well as the reconstructed π 0 ’s. In particular, the most recently developed algorithm uses recurrent neural network (RNN) to fully harness the discriminating power in low-level input variables of individual tracks and clusters. This algorithm, known as the RNN ID [82] , doubles the rejection against jets with respect to its predecessor, which uses only high-level variables that collectively describe the tracks and clusters. This significant improvement in jet rejection is one of the important factors that make the sensitivity of the search presented in this thesis achievable. Isolated τ had - vis typically have one or three associated tracks in the core region and very few energetic tracks in the isolation region. Their showers are on average more collimated than jets, especially gluon-initiated ones. τ had - vis with π 0 ’s also have distinctive energy deposit patterns in the EMCal. Furthermore, given the relatively long lifetime of τ leptons,