Separating signal from combinatorial jets in a high background environment

We study procedures for discriminating combinatorial jets in a high background environment, such as a heavy ion collision, from signal jets arising from a hard-scattering. We investigate a population of jets clustered from a combined PYTHIA+TennGen event, focusing on jets which can unambiguously be classified as signal or combinatorial jets. By selecting jets based on their kinematic properties, we investigate whether it is possible to separate signal and combinatorial jets without biasing the signal population significantly. We find that, after a loose selection on the jet area, surviving combinatorial jets are dominantly imposters, combinatorial jets with properties indistinguishable from signal jets. We also find that, after a loose selection on the leading hadron momentum, surviving combinatorial jets are still dominantly imposters. We use rule extraction, a machine learning technique, to extract an optimal kinematic selection from a random forest trained on our population of jets. In general, this technique found a stricter kinematic selection on the jet's leading hadron momentum to be optimal. We find that it is possible to suppress combinatorial jets significantly using this machine learning based selection, but that some signal is removed as well. Due to this stricter kinematic selection, we find that the surviving signal is biased towards quark-like jets. Since similar selections are used in many measurements, this indicates that those measurements are biased towards quark-like jets as well. These studies should motivate an increased emphasis on assumptions made when suppressing and subtracting combinatorial background and the biases introduced by methods for doing so.