Discriminating between a 'physics' signal and a background is a common task in high energy physics and various methods have been designed to solve it. To classify a single event as signal or background it is necessary to know the probability that this event is a background event and or signal event, i.e. one has to estimate the probability density around each accessible phase space point. This is usually done by Monte Carlo techniques.
The best approximation to the local probability density at each point in phase space of the event which is to be classified, can be obtained by counting all the events in the vicinity of this phase space point. This calculation is, however, very CPU and memory consuming and difficult to perform. To speed up the probability estimation it is crucial to use a sophisticated search method for the events lying in the vicinity of the event to classify.

We will present a search algorithm based on a binary tree search which allows to find all events lying in a given phase space box by consecutively subdividing the event samples. This algorithm needs only a time proportional to N log(N) ... Since this method is fast and stable against correlations in the input variables, it can also be used to perform a scan for the most promising variables for event classification. Contrary to neural networks where learning and quality checking can be time consuming and can only be finalised by human intervention, this multivariate discrimination technique can easily be automated.

In the talk we will present how this discrimination technique is applied to the search for Instanton induced events in deep-inelastic scattering (DIS) at HERA. Instanton-induced processes can be distinguished from normal DIS events by their characteristic hadronic final state. The predicted cross section of such process is about a factor 100 smaller that the inclusive DIS cross section. In order to discover instantons a huge suppression of the background based on several discriminating observables is therefore necessary. The results of the application of the multivariate technique are compared to those produced with neural networks.

Back to Program

Back to ACAT Web page