A Neural networks search for single top quark production in CDF Run I data

In the proton-antiproton collisions a t the Ferm ilab Tevatron collider, individual top quarks are expected to be produced by an electroweak mechanism referred to as single top production. We present here a Neural Networks m ethod of searching for single top quarks in the 106 pb~l of d a ta collected by the Collider Detector a t Fermilab (CDF) from 1992-1996. The prospects for observing single top in the current run a t the Tevatron are also discussed. In searching for single top, improvement to signal to noise can be obtained by requiring a charged lepton and missing energy from the IT' decay, as well as a Btagged je t. We conduct our search in the IT' -I1,2, and 3 je ts channels, where we impose additional selections to further reduce background contributions from QCD m ultijet (IT '+jets) processes and top pair production tt respectively. Specifically, in the W + 1 je t channel we require exactly one additional je t w ith £V > 8 G eV and [77I < 2.4, while in the \V + 3 je ts channel we require no such additional jets. After applying these selection cuts to Monte Carlo sam ples, we are able to estim ate the R un I expected contributions as: 4.2 single top events, 43.3 ± 8.4 IF + je ts events, and 7.4 ± 2.2 t t events. The signal purity is therefore 8 %. which makes the Neural Networks approach particularly suitable for the single top analysis. To distinguish between signal and background we select seven event variables w ith good separating power. These variables are: the transverse energies of the leading two jets, the lepton, ii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. and the neutrino: E ^ n , E ^ t2, E ^ pton, the transverse m om entum of the leading two-jet system P^3, the to ta l transverse energy in the event H r, and the product Q x r] between the prim ary lepton charge and the pseudorapidity of the h ighest-lsr non S -tagged jet. The Neural Networks used is a three-output perceptron with one hidden layer trained on M onte Carlo generated events. We use the M onte Carlo signal and backgrounds o u tp u t tem plates to estim ate the com position of the 64event R un I dataset passing our above selection. By using a simple binned-likelihood function w ith Gaussian background constrain ts to fit the data, we ob tain the following contributions: ^signal = 23.6 ± 7.7, uqCd = 36.2 ± 6.2, rtti = 7.6 ± 2.0 (events) We note th a t the fit result is approxim ately 2.5 standard deviations away from the expected value of 4.2 signal events. VVe further use this result to ex trac t a 95% C.L. upper lim it on single top cross section of 24.4 pb, roughly 10 tim es higher than the S tandard Model prediction of 2.43 pb. For Run II, we expect a single top cross section m easurem ent a t 4cr precision level w ith roughly 2 f b ~ l of data.