Developing and testing Unfolding Algorithms

Abstract

The observed kinematic distributions of the particles produced in proton-proton collision events are generally different from what is expected for an ideal detector mainly because of the detector effects like imperfect energy resolution, acceptance of the detector, and reconstruction inefficiency. We need to consider these effects so that we can compare the collider observables distributions at the ``truth level" with theoretical predictions and with measurements from the different experiments. These comparisons can enhance the understanding of the Standard Model, tune Monte Carlo event generator parameters and enable precision searches for new physics. Unfolding algorithms are used to obtain these truth distributions from the detector's measured information by correcting these detector effects.

Performing high-dimensional measurements at particle colliders are essential as these are the measurements that keep as much information as possible. They are the key to understanding the correlations across different measurements and how they impact interpretations of differential cross-section, and Wilson coefficients in top physics, among other things. Current unfolding methods depends on the binning chosen for the histogram of the measured observable, which causes problems when unfolding is done on several variables simultaneously. All possible auxiliary features that control the detector response are not taken into account by the traditional unfolding algorithms. Hence, to extract as much information as possible using the high-dimensional measurements, there is a requirement of an unfolding method which do not depend on the binning chosen and performs well for multi-dimensional measures.

We explore different unfolding algorithms to develop and test an unfolding method that can be performed in an unbinned, multi-dimensional fashion preserving as much event information as possible. An unfolding method that performs unfolding using the Energy Mover's distance metric is explored and is compared with OmniFold, which is a new deep learning-based unbinned unfolding method. A specific case study is used to show specific gains and advantages of these new unbinned unfolding methods.