Importance of multivariate analysis

During the hunt for the Higgs boson, scientists had to investigate and study a number of predicted processes.

Science proceeds step by step, looking for the unknown and the unexplored. Higgs production is rare, and as many different processes contribute to the background against which the Higgs signal must be distinguished, physicists have to reduce that background piece by piece to bring it to an acceptable level.

The left plot show the distribution of the dijet mass. The right plot shows the neural network output. Both plots are for the one-tag candidates where events from all lepton categories are added together. The best fit to the data is shown.

The left plot show the distribution of the dijet mass. The right plot shows the neural network output. Both plots are for the one-tag candidates where events from all lepton categories are added together. The best fit to the data is shown.

One of the ways in which the Higgs was hunted is through its associated production with W bosons.

W boson properties are well known and provide a way to select events in which a Higgs boson can be searched through its decays into two b quarks.

Unfortunately there are processes that can mimic our signal. One of them is the production of a W and a Z boson together, with the Z boson decaying into two heavy flavor quarks (two charm or two beauty quarks) and the W decaying into leptons (one charged and one neutral). A second is the production of a WW boson pair, in which the second W decays into heavy flavor quark.

These processes are predicted in the Standard Model, but until now, they escaped a clear observation in this final state.

Scientists at CDF have looked into the full data set collected during the 10 years of Tevatron Run II to identify these processes. During this search, they developed a number of techniques to disentangle events containing real W’s from events in which their presence was mimicked by other processes.

The separation of signal and background is shown in the left plot of the above figure. Among the techniques used as part of this analysis was that of a neural network. Its output is shown on the right side of the figure. This technique is modeled on the central nervous system.

These physicists were responsible for this analysis. From left: Giorgio Chiarelli, Sandra Leone and Federico Sforza, all from INFN (University of Pisa).

These physicists were responsible for this analysis. From left: Giorgio Chiarelli, Sandra Leone and Federico Sforza, all from INFN (University of Pisa).

CDF also looked for events containing a W or a Z decaying into heavy quarks, and we used the know-how developed in the Higgs experiment. The evidence for the Higgs boson was obtained in part thanks to an earlier version of this study.

In the end, the measuring of WZ (W→lepton and neutrino; Z→bb or cc) and WW (W→lepton and neutrino; W→c or s) is an interesting check of the Standard Model prediction, with the two signals never measured separately so far at a hadron collider. CDF measured a production cross section for WW of 9.4 ± 4.2 picobarns and WZ of 3.7 +2.5 -2.2 picobarns with and evidence of 2.87 sigma for the first process and 2.12 for the second one.

And in case you didn’t see this item in the news: Earlier this year AlphaGo defeated Lee Se-dol (a champion go player) four games to one. AlphaGo was developed by Google’s subsidiary Deep-Mind and is based on using neural networks.

Learn more