Q: What is the Track-based Discriminant?

A:  This is a neural network specifically designed to separate events with real missing Energy (high energy neutrinos) from those in which missing energy is generated by mismeasurement. The single largest background in this analysis involves multi-jet QCD heavy flavor production, which can fake the signature of a Dijet event recoiling against missing transverse energy. Mismeasured energy in calorimeters is uncorrelated to the momenta of charged particles measured in the tracking chamber, making tracking information particularly useful in this anaysis. Many track based quantities are utilized as inputs into a neural network, which is trained to seperate QCD background from a ZH signal.

Q: How is the Track-based Discriminant used in the analysis?

A:  It is used as an input into the final Neural Network Discriminant, which combines this tracking information with calorimeter based quantities.

Q: How is the heavy flavor QCD modeled?

A:  The heavy flavor QCD model is taken directly from the single tag data. The shapes of the electroweak, top and mistagged light flavor backgrounds are subtracted out of this data at their expected contributions, yielding a shape which represents true heavy flavor QCD production for a given distribution. Additionally, biases introduced by tagging are applied to this data to successfully model the double tag data.

Q: What samples are used in training the Neural Network Discriminant?

A:  The final neural network in this analysis utilizes 5 input variables, including the track-based NN. ZH and WH Monte Carlo is trained against tt Monte Carlo and untagged data. The neural network provides excellent seperation of Higgs from the QCD and Top backgrounds. The neural network also provides good seperation of Higgs signal from electroweak backgrounds, although not as dramatic.

Q: How does the neural network compare to a mass fit?

A:  We have not performed limit calculations for the Dijet Mass distribution, as there is a dedicated Dijet Mass analysis at CDF, which uses tighter event selection to optimize S/B. The latest results from that search can be found here.