We investigate the event topology of the
The top quark decays into a b-quark and a W boson
(tbW). The W boson decays into a lepton
(electron or muon)
and a neutrino about 22% (2/9) of the time and into a
quark-antiquark pair about 67% (6/9) of the time.
This implies that when top-pairs are produced in
hadron-hadron collisions,
(a) Top-Pair Production |
(b) Event Topology |
Figure 1. (a) Illustration of top-pair production in proton-
antiproton collisions in which one of the W bosons decays
leptonically and the other decays hadronically resulting in a
final state consisting of a lepton, a neutrino,
plus a |
In this paper, we concentrate on the
(a) W+Jets Background |
(b) b+Jets Background |
Figure 2. Illustration the (a) W+jets background and the (b) b+jets background to the top-pair production in proton-antiproton collisions shown in Fig. 1. |
ISAJET version 7.06 is used to generate top quarks
with a mass of 175 GeV in 1.8 Tev proton-antiproton
collisions. At this energy, 175 Gev top-pairs are produced
via quark-antiquark annihilation,
We do not attempt to do a detailed simulation of the CDF or
D0 detector [3,4]. Events are analyzed by dividing the solid
angle into "calorimeter'' cells having size
The "zero-level'' trigger is designed to select large transverse momentum W bosons that have decayed into a charged lepton and a neutrino. This first cut is made by demanding that the event contain at least one isolated high transverse momentum charged lepton (electron or muon) in the central region satisfying:
Table 1. 175 GeV Top quark pairs produced in
1.8 TeV proton-antiproton collisions. The table shows the
number of events (with L=100/pb) for the top-pair
signal and the W+jets background.
The |
|||||||||
Top Signal | W+jets Background | ||||||||
Efficiency | Efficiency | Enhancement | |||||||
Cut or Selection | Events | Relative | Overall | Events | Relative | Overall | Relative | Overall | Sig/Bak |
"Zero-Level" Lepton plus Missing ET Trigger: PT(l) > 15 GeV ET(miss) > 20 GeV PT(lv) > 25 GeV |
165 | 100% | 100% | 7,044 (470) | 100% | 100% | 1.0 | 1.0 | 0.0234 |
Calorimeter Cell Cuts: N(cell) > 7 (ET(cell) > 5 GeV) ET(sum) > 100 GeV |
113 | 69% | 69% | 49 (3) | 0.7% | 0.7% | 99.7 | 99.7 | 2.3 |
Fisher Cut: F > 0.75 |
49 | 44% | 30% | 6 | 12% | 0.1% | 3.7 | 373 | 8.7 |
This selection of PT(l±) > 15 GeV,
ET(miss) > 20 GeV,
and PT(lv) > 25 GeV is referred to as the lepton plus
missing ET trigger. Table 1 shows that about 165
top-pair events survive this selection criterion (22% of the
overall top signal). For illustration, we take the integrated
luminosity to be 100/pb. Table 1 also shows that about
7,000 W+jets and roughly 500
In order to quantify how various additional cuts enhance the signal above the background, we define the enhancement, F(enh), and the efficiency, F(eff), of a given set of cuts as follows:
F(enh) = (% of signal surviving cut)/(% of background surviving cut )We define the "zero-level" trigger to be the reference point and the fraction of events escaping this cut is set to 100% in Table 1. Similarly, all "enhancement'' factors are set to one at this level as we measure the effectiveness of all other additional cuts from this point. The overall enhancement and efficiency is determined relative to the "zero-level" trigger, while relative enhancements and efficiencies are determined by examining the number of events before and after the particular cut.
F(eff) = % of signal surviving cut .
Figure 3. Shows the multiplicity of calorimeter cells containing at least 5 GeV of transverse energy for the of the top-pair signal and the W+jets background. In all cases, the events have survived the "zero-level" lepton plus missing ET trigger. The plot shown the percentage of events with N cells with ET(cell) > 5 GeV. The position of our cell cut is marked by the dotted line. |
Figure 4. Shows the total transverse energy of all the calorimeter cells with ET(cell) > 5 GeV for the of the top-pair signal and the W+jets background. In all cases, the events have survived the "zero-level" lepton plus missing ET trigger. The plot shown the percentage of events with ET(sum) within a 25 GeV bin. The position of our cell cut is marked by the dotted line. |
At this stage in the analysis, one normally demands that the event contain at least four jets [3-5]. Cutting on the number of jets is a way to preferencially select the top-pair signal over the background. However, we have found that it is faster and better to simply cut on the number of calorimeter cells, N(cell), with transverse energy greater than some minimum, ET(cell-min). Fig. 3 shows the calorimeter cell multiplicity with ET(cell-min)=5 GeV for the top-pair signal and the W+jets background. On the average, the top-pair signal populates a larger number of cells than does the background. Obviously this is because the top-pair signal produces more jets, however, one does not have to define a "jet" to see this topology. The top-pair signal produces transverse energy flying out in all directions and this can be seen directly from the calorimeter cell multiplicity.
The top-pair signal also produces more global transverse energy than the background. This is shown in Fig. 4, where we define ET(sum) to be the sum of the transverse energy of all the calorimeter cells with ET > ET(cell-min). As illustrated in Fig. 3 and Fig. 4, we make the following calorimeter cell cuts:
Table 1 shows that of the 165 top-pair events passing the "zero-level" lepton plus missing ET cut roughly 69% also pass the calorimeter cell cuts. On the other hand, less than 1% of the W+jets background events survive the cell cuts. The calorimeter cell cuts produce an enhancement of 0.69/0.007 or about 100 over the W+jets background with a 69% efficiency, resulting in a signal to background ratio of about 2. The N(cell) > 7 with ET(cell-min)=5 GeV cut produces more than a factor of two better enhancement than the traditional "jet cuts'' (i.e. N(jet) > 3). Adding the ET(sum) > 100 GeV cut gives an additional relative enhancement of more than a factor of three.
Figure 5. Shows the multiplicity of jets with transverse energy greater than 15 GeV for the of the top-pair signal and the W+jets background. In all cases the events have survived the "zero-level" lepton plus missing ET trigger and the calorimeter cells cuts. The plot shown the percentage of events with N jets with ET(jet) > 15 GeV. |
Ideally one would like to reconstruct the invariant mass of the top-pair from its decay products: lepton, neutrino, and four jets. However, the neutrino is not detected and its presence must be inferred by examining the missing transverse momentum. If we set the transverse momentum components of the neutrino equal to the missing transverse momentum,
We have not used jets in our event trigger, however, we do use jets to reconstruct the top-pair invariant mass. In addition, we use the jet topology to help further distinguish the signal from the backgrounds. Jets are defined using a simple algorithm. One first considers the "hot'' cells (those with transverse energy greater than 5 GeV). Cells are combined to form a jet if they lie within a specified "radius'',
Rj2 = 2 + 2,in - space from each other. Jets have an energy given by the sum of the energy of each cell in the cluster and a momentum given by the vector sum of the momentums of each cell. The invariant mass of a jet is simply
Mj2 = Ej2-pj2.In this analysis, we take the jet radius to be Rj=0.4 and require jets to have at least 15 GeV of transverse energy. Namely,
Figure 6. Shows the reconstructed top-pair invariant mass, M(t), for 175 GeV top quarks produced in 1.8 TeV proton-antiproton collisions (solid curve). The plot contains only the top-pair signal and corresponds to the number of events per year (with L=100/pb) in a 50 GeV. The events have survived the "zero-level" lepton plus missing ET trigger and the calorimeter cell cuts. Also shown in the true parton-parton CM energy of the event (not directly observable experimentally). |
Figure 7. Shows the reconstructed top-pair invariant mass, M(t), for 175 GeV top quarks produced in 1.8 TeV proton-antiproton collisions together with the W+jets background. The plot shown the sum of the signal plus background and corresponds to the number of events per year (with L=100/pb) in a 50 GeV. The events have survived the "zero-level" lepton plus missing ET trigger and the calorimeter cell cuts. |
The top-pair invariant mass, M(t), corresponds to the center-of-mass energy, Ecm, of the underlying parton-parton two-to-two subprocess which has a threshold at twice the mass of the top quark, Ecm > 2Mtop. This is seen clearly in Fig. 6 which compares the true t and ggt CM energy, Ecm (not experimentally observable), with the reconstructed top-pair invariant mass, M(t). If the neutrino momentum could be precisely determined from the missing ET and if we knew exactly which particles to include in the jets then the two curves in Fig. 6 would agree. Although one cannot precisely reconstruct the parton-parton CM energy, there still remains a nice peak in the reconstructed top-pair invariant mass at twice the top mass. In this paper, we use the observation of this peak as a measure of how well one can determine the top quark mass and we would like to remove as much background as possible from the peak.
Fig. 6 includes only the top-pair signal with no background. Fig. 7 shows the reconstructed parton-parton CM energy for the top-pair signal and the W+jets background after the "zero level" lepton plus missing ET trigger and the calorimeter cell cuts. The plot shown the sum of the signal and the background. At this stage the signal is about twice the background. However, the signal to background ratio can be improved by examining in more detail the "shape" of the events.
In 1979 Geoffrey Fox and Stephen Wolfram [6] constructed a complete set of rotationally invariant observables, Hl which can be used to characterize the "shapes" of the final states in electron-positron annihilations. They are constructed from the momentum vectors of all the final state particles as follows,
The Fox-Wolfram observables (or moments) constitute a complete set of shape parameters. For example, the collinear "two-jet" final state results in Hl near 1 for even l and Hl near 0 for odd l. Events that are completely spherically symmetric give Hl near 0 for all l.
Table 2. Shows the mean value and standard deviation from the mean of six of the modified Fox-Wolfram moments applied to the jets in the event with transverse energy greater than 15 GeV. Results are shown for the top-pair signal and the W+jets background. Also shown are the resulting Fisher coefficients. | |||||
Top Signal | W+jets Background | ||||
Moment | Mean | Stdev | Mean | Stdev | Fisher Coefficient |
H1 | 0.2382 | 0.1775 | 0.3598 | 0.2457 | -0.500 |
H2 | 0.2815 | 0.1563 | 0.4381 | 0.2211 | -1.282 |
H3 | 0.2757 | 0.1367 | 0.4031 | 0.1941 | -1.088 |
H4 | 0.2833 | 0.1279 | 0.4099 | 0.1838 | -0.978 |
H5 | 0.2856 | 0.1207 | 0.3994 | 0.1744 | -0.544 |
H6 | 0.2891 | 0.1156 | 0.3989 | 0.1735 | -0.069 |
In hadron-hadron collisions spherical symmetry is lost and we are interested more in the shape of events in the transverse plane. For example, the Fox-Wolfram moments when applied directly to hadron-hadron collisions would interpret a minimum bias event as a "two-jet" event. Whereas, we would like to have a minimum bias event treated more like a spherically symmetric e+e- final state (i.e. no structure). To accomplish this, we define the following modified Fox-Wolfram moments for hadron-hadron collisions,
Figure 8. Shows the modified Fox-Wolfram moment, 1, calculated using the jets in the event with transverse energy greater than 15 GeV for top-pair signal and for the W+jets background. The plot showns the percentage of events in a 0.05 bin. The events have survived the "zero-level" lepton plus missing ET trigger and the calorimeter cell cuts. (If the vector sum of the monentum of all the jets in the events is zero then 1=0.) |
Figure 9. Shows the modified Fox-Wolfram moment, 2, calculated using all the jets in the event with transverse energy greater than 15 GeV for top-pair signal and for the W+jets background. The plot showns the percentage of events in a 0.05 bin. The events have survived the "zero- level" lepton plus missing ET trigger and the calorimeter cell cuts. |
Figure 10. Shows the modified Fox-Wolfram moment, 4, calculated using all the jets in the event with transverse energy greater than 15 GeV for top-pair signal and for the W+jets background. The plot showns the percentage of events in a 0.05 bin. The events have survived the "zero-level" lepton plus missing ET trigger and the calorimeter cell cuts. |
Table 2 shows the mean values and standard deviations for six of the modified Fox-Wolfram moments calculated using all jets with ET(jet) > 15 GeV for events that have survived the "zero-level" lepton and missing ET trigger and the calorimeter cell cuts. There are clearly still some differences between the jet topologies of the top-pair signal and the W+jets background. The mean values of the six moments 1,... , 6 are smaller for the signal than the background indicating that the jets originating from the top- pair signal form a more cylindrically symmetric pattern when they emerge from the event than do the background jets. The top-pair jets are more spread out in - space. This can be seen in Figs. 8, 9 ,and 10 which show the 1, 2, and 4 distributions, respectively, for the signal and background. At this stage one could simply make a linear cut on, for example, 2. Requiring 2 < 0.3 gives an additional signal to background enhancement of about 2 with a relative efficiency of around 60%. One can do a little better, however, by using the information of all six of the l's simultaneously. This can be done by constructing a neural network or by using Fisher discriminates.
Neural networks are an excellent tool for separating patterns into categories (e.g. signal and background). Our neural networks [7] consist of a set of Nin inputs, {x}, which can have any value and one output, znet, which is restricted to the range, 0 < znet < 1. The net output is a function of the input set {x} and the network "memory" parameters as follows:
znet = Fnet({x},{w},{T}),where the network memory consists of a set of weights, {w}, and a set of thresholds {T}. The goal is to construct a network that can distinguish between two patterns of input data, "signal" events and "background" events, where each event is characterized by the Nin variables. A "perfect" network responds with znet near one for a signal input and with znet near zero for a background input and a single cut can be made on this network output which will enhance the signal over the background.
Figure 11. Shows the network response, znet, for the sample of signal and background events used in the training. The plot corresponds to the percentage of events with znet within a 0.05 bin for the top-pair signal and the W+jets background. The events have survived the "zero-level" lepton plus missing ET trigger and the calorimeter cell cuts. |
Figure 12. Shows the enhancement versus the efficiency for the training sample of events for a 6-12-1 neural network with 97 memory parameters. Each point in the plot corresponds to a different choice for the network cut-off with the lower efficiencies and higher enhancements corresponding to larger values of zcut. The network enhancements are compared with the enhancements arrived at by the use of Fisher discriminates. |
Of course, the key to a good network lies in the selection of the input variables. These variables must characterize the differences between the signal and the background. We choose the first six modified Fox-Wolfram variables ( applied to the jets) as the network inputs:
x1=1, x2=2, x3=3, x4=4, x5=5, x6=6.The network is trained on a sample of 4,000 top-pair signal and 3,814 W+jets background events using the six inputs shown above and where both signal and background events have already satisfied the "zero-level" lepton plus missing ET trigger and the calorimeter cell cuts. To get this training sample, it was necessary to generate 50,000 top-pair events and 1,200,000 W+jet events.
There is no systematic procedure that provides the best network topology for a given problem. One looks for the simplest network that can discriminate signal from background. Here we use a simple network with only one "hidden layer". We use a 6-12-1 net which has 97 memory parameters (see Ref. [7] for notation). Fig. 11 shows the network response (i.e. znet) for the sample of signal and background events used in the training. The situation is far from the ideal. There are some events around znet=0.5 for which the net cannot distinguish between signal and background. Nevertheless, the net does allow for some separation of signal and background. The net clearly recognizes some events as signal or background, while for other events there is an overlap and the net cannot distinguish between the two.
The next step is to perform a network cut-off and assign any event with znet > zcut to be signal and events with znet < zcut to be background. The enhancement and efficiency of the network cut-off depends on the value chosen for zcut, where the network enhancement and efficiency are defined as follows:
Fnet(enh) = (% of signal with znet > zcut)/(% of background with znet > zcut)The overall network performance can be characterized by the single curve of the network enhancement versus the network efficiency shown in Fig. 12. Each point in Fig. 12 corresponds to a different choice for the network cut-off with the lower efficiencies and higher enhancements corresponding to larger values of zcut. For example, a net cut of zcut=0.75 corresponds to an additional enhancement of about 4 with a relative efficiency of about 47%.
Fnet(eff) = % of signal with znet > zcut.
Figure 13. Shows the "shifted" Fisher response, , for the sample of signal and background events used in the training of the neural network. The plot corresponds to the percentage of events with within a 0.05 bin for the top-pair signal and the W+jets background. The events have survived the "zero-level" lepton plus missing ET trigger and the calorimeter cell cuts. The position of our "Fisher cut" is marked by the dotted line. |
A simplier method of separating signal and background is to use Fisher discriminates. This method is analogous to a neural network with no hidden layers. Here as with the network, one inputs a set of Nin variables, xi, and there is one output, F. However, in this case F is a linear function of the inputs,
In this case training consists of calculating the Fisher coefficients which involves inverting an Ninx Nin matrix, but this is much easier than training a network. Once this is done the situation is similar to the network (with F replacing znet), for each input of Nin variables there is one output F. We have determined the Fisher coefficients for the sample of signal and background events used to train our network and the Fisher response for these events is shown in Fig. 13. In plotting the Fisher responce in Fig. 13, we have shifted, F, to lie between zero and one as follows:
Figure 14. Shows the reconstructed top-pair invariant mass, M(t), for 175 GeV top quarks produced in 1.8 TeV proton-antiproton collisions together with the W+jets background. The plot shown the sum of the signal plus background and corresponds to the number of events per year (with L=100/pb) in a 50 GeV. The events have survived the "zero-level" lepton plus missing ET trigger, the calorimeter cell cuts, and the Fisher cut-off. |
Fig. 12 shown that Fisher discriminates have essencialy the same performance curve as does the neural network and since it is simplier to calculate the Fisher function, we complete our analysis by making a cut on as follows:
Fig. 14 shows the reconstructed parton-parton CM energy for the top-pair signal and the W+jets background after the "zero level" lepton plus missing ET trigger and the calorimeter cell cuts and the Fisher cut. The plot shows the sum of the signal and the background.
We have developed a procedure that enhances the top quark
signal
In addition, we use Neural networks and Fisher discriminates in conjunction with modified Fox-Wolfram "shape" variables to further distinguish the top-pair signal from background. For example, using the first six Fox-Wolfram moments ( applied to the jets) together with a Fisher cut-off, > 0.75, provides an additional enhancement of around 4 with a relative efficiency of about 44%. By combining the calorimeter cell cuts with the event shape information, we are able to obtain an overall signal to background enhancement of around 370 with an efficiency of 30%, and a signal to background ratio of around nine.
Click here to return to Recent Publications.