• multivariate analysis
• higgs boson
• CMS
• Vector Boson Fusion

Understanding the mechanism that breaks the electroweak symmetry and generates the masses of the known elementary particles has been one of the fundamental endeavors in particle physics. The breaking of the electroweak symmetry is allowed if at least one new particle with well defined properties is added to the ensemble of the elementary particles. Such a particle has long been know as the Higgs boson. Its discovery at the Large Hadron Collider (LHC) at Cern in 2012 by the ATLAS and CMS collaborations represents therefore a major achievement in the field. Starting in 2012, the properties of the Higgs boson have been measured in many of the accessible final states originating from its decay. The mass of the Higgs boson has been determined to be 125.09 ± 0.21 (stat) ±0.11 (syst) GeV, from a combination of the ATLAS and CMS measurements. Several results from both experiments established that its measured properties, including its spin, CP properties, and coupling strengths to fermions and bosons, are consistent with the Standard Model (SM) expectations.
As new data is collected, the properties of the Higgs boson can be measured with increasing precision and rarer decay modes become accesible. Such measurements are interesting because any deviation from the prediction of the theory might be a hint of new physics beyond the SM. Among the rare decay modes currently under investigation, the Higgs boson decay into two muons (H → μμ) is the object of study of this thesis.
For a Higgs boson with mass of approximately 125 GeV, the probability to decay into a muon pair is expected to be B(H → μμ) = 2.2 × 10 −4, making it one of the smallest accessible at the LHC. On the other hand, the H → μμ signature is one of the cleanest to detect experimentally. Higgs boson decays in two muons are of particular importance because they extend the study of its couplings from the third generation to the second generation of fermions, where deviations from the SM predictions, due to new physics are predicted to be larger.
The search for H → μμ presented in this work is performed selecting the vector- boson fusion (VBF) production mode. The cross section is about 10% of the cross section for the gluon-gluon fusion, which is the most important production mode.
However, the VBF process gives a cleaner experimental signature. In fact, in the VBF process, a quark coming from each colliding proton radiates a W or Z bosons vvi that subsequently interacts. The two quarks therefore slightly deviated from their original flight direction and typically fall inside the detector acceptance, while a Higgs is emitted.
Restricting the scope of the search to the VBF production mode, makes the process even rarer but the peculiar signature of the VBF production mode can be exploited to effectively reduce the experimental backgrounds. The VBF quarks are revealed as jets: two back to back high momentum narrow cones of hadrons and other particles produced by the hadronization of a quark or gluon. Generally the two VBF jets are expected to have high pseudorapidity and large invariant mass while the Higgs decay products are expected to be in the central region of the detector.
Imposing the constraints to the invariant mass and the rapidity of the jets as additional cut one reaches an impressive improvement of the signal-to-background ratio.
The data used for this search were collected using proton-proton collision at sqrt(s) = 13 TeV by the CMS experiment in 2016, corresponding to an integrated luminosity of 35.9 fb −1 . Only 30 event are expected during the entire data taking period. It is therefore essential to have a high signal efficiency, both in the online and the offline selections, while greatly reducing the backgrounds. The dominant sources of background in these studies are production of top quark pairs (tt) and Drell-Yan leptons with associated jets (referred to as DY+jets). These have a good probability to decay into muons, whose tracks risk to be misclassified as coming from a Higgs decay. The DY+jets background is the hardest to discriminate because it is characterized by two real prompt leptons from a virtual Z or γ boson in addition to two jets, either from initial state radiation.
A multivariate approach is used to further discriminate signal from background.
As background processes are many orders of magnitude larger than the signal, a Machine Learning (ML) classifier with an extremely good signal acceptance versus background rejection performance is required. For this purpose two different ma- chine learning techniques are used: Boosted Decision Trees (BDTs) and Deep Neural Networks (DNNs). Such systems "learn" (i.e. progressively improve performance on tasks) by considering examples, generally without task-specific programming.
The toolkits used in this thesis to implement the multivariate classifier algorithm are TMVA [6] for the BDT method and the Keras library, running on top of Theano, for the NN one. Both are integrated into the ROOT analysis framework.
My personal contribution has been the development of these dedicated multivariate techniques, including the search and selection of the most discriminant variables.
In order to improve the suppression of the background sources and to obtain the maximum sensitivity a particular attention was given to the choice of the variables starting with the definition of an extensive set of kinematic observables. Several tests were made to search the best discriminant variables checking also the correlation between all the features. Seven variables are considered as the inputs of the BDT. The same input variables are sent to the NN with the addition of other five. After several network configurations the best one results using a pretraining step without the muon invariant mass mll (that is the most discriminant variable) and then a training with the previous weights with the complete features set. In this way is possible to exploit the discriminating power of all the selected variables.
The expected final goal is an improvement of the branching ratio upper limit of the process. The current results are still preliminary but encouraging: for a Higgs boson decaying to two muons, the upper limit on the decay rate at 95% confidence level (CL) is expected to be approximately 2.5 times the SM value.