Impact of methodological uncertainties

Figure 5.10 of [Gomez Ambrosio et al., 2022].

Pair-wise 95% CL contours for the Wilson coefficients entering \(hZ\) production in the dilepton channel.

We compare the bounds obtained by calculating the median across the \(N_{\rm rep}\) MC replicas used in training with those obtained by taking into account the impact of the 2-\(\sigma\) methodological uncertainties. Both ML models are trained on the full set of 7 kinematic features relevant for particle level Higgs production in association with a \(Z\) boson in the dilepton channel.

By comparison with the corresponding bounds displayed on the previous page, we confirm that these methodological uncertainties, as estimated with the replica method, do not modify the qualitative results obtained from the median of the likelihood ratio. In particular, the observation that bounds obtained using multivariate unbinned observables are much stronger than those based on unbinned models based on one or two kinematic features remains valid once replica uncertainties are accounted for.