Work

Advanced statistical methods for data analysis in particle physics

Year: 2018
Type: dissertation
Abstract: The thesis has been developed focusing on the use of multivariate statistical methods in the High Energy Physics framework. Stemming from the framework described by the current dominant physical theory, known as the Standard Model, the thesis has been developed by following two directions, associated with two different physical research questions. \n \nThe first route takes the steps from the need of improving the knowledge within the Standard Model. From a statistical point of view, such improvement refers to the aim of obtaining more accurate estimates of the parameters describing the Standard Model in order to gain a better knowledge of the probability distribution of the underlying physical process, known as the background. In practice, estimation of such probability distribution builds on the use of Monte Carlo simulated data, which, in turn, can be costly and imprecise. To prevent these problems, the physical community has developed a novel procedure to generate artificial background data from the experimental ones. Within the thesis, a formal validation of the physical procedure is performed by means of introducing a statistical permutation-based two-sample test for density equality. The test relies on kernel density estimation and is suitably adjusted to be applied to high dimensional data. \n \nThe second direction of research derives from the incompleteness of the Standard Model, known to be unable to fully describe the Universe and the interactions among its characterising forces. \nThe goal of going beyond the Standard Model is reached through model-independent searches of new physics which aim at looking for new possible particles not predicted by the Standard Model. Such particles, referred to as a signal, are expected to behave as a deviation from the known background. From a statistical perspective, the problem is recasted to a peculiar classification one where only partial information is available. Therefore a semi-supervised approach shall be adopted, either by strengthening or by relaxing assumptions underlying clustering or classification methods respectively. Within this context, the thesis follows two distinct approaches. The first approach consists of developing a parametric semi-supervised method which originates from the framework of model-based clustering. A dimensionality reduction technique is proposed by resorting to penalised methods to circumvent issues related to parameters estimation and the curse of dimensionality. The proposed variable selection approach is extended from the unsupervised to the semi-supervised context with attention to features exhibiting anomalous properties. The second approach followed with the aim of new physics searches consists of suitably adjusting and statistically validating an existing procedure, developed within the physical community. Some improvements to the algorithm are also proposed regarding, among others, cases of high dimensional and correlated data.
Language: English
Open Access status: green
OpenAlex ID: W3006431200
Cites: 1
Cited by: 1
Related to: 10