How to Unfold Top Decays

Luigi Favaro, Roman Kogler, A. Paasch, Sofia Palacios Schweitzer, Tilman Plehn, Dennis Schwarz

Abstract

Using unfolded top-quark decay data we can measure the top quark mass, as well as search for unexpected kinematic effects. We show how generative unfolding enables both tasks and how both benefit from unbinned, high-dimensional unfolding. Our method includes an unbiasing step with respect to the training data and promises significant advantages over standard methods, in terms of flexibility and precision.

Introduction

In the high-stakes world of particle physics, "unfolding" is the ultimate reality check. It is the mathematical process of stripping away the distorting effects of a particle detector to reveal the underlying "truth" of subatomic interactions. Traditionally, this has been a low-dimensional, binned affair. However, a new approach introduced in "How to Unfold Top Decays" leverages generative machine learning to make unfolding unbinned, high-dimensional, and—crucially—unbiased.

The Challenge of Detector Effects

In experiments like those at the Large Hadron Collider (LHC), we don't observe quarks directly. Instead, we see "detector signatures"—showers of particles and energy deposits. To compare experimental results with theoretical predictions (made at the "parton level"), physicists must perform unfolding.

Unfolding is an "ill-posed" inverse problem because detectors have finite resolution and efficiency; multiple "true" physics states can lead to the same observed state. Standard methods, such as Iterative Bayesian Unfolding, require binning the data, which inevitably discards valuable information. While modern machine learning methods can handle unbinned data, they often suffer from model dependence. If a neural network is trained on a simulation assuming a specific top-quark mass, it tends to "pull" the unfolded results toward that assumed mass, biasing the final measurement.

Comparison of generator-level (gen) and reconstructed (reco) distributions for various top-quark decay observables, illustrating the smearing effects of the detector.

Generative Unfolding with Conditional Flow Matching

The researchers employ a state-of-the-art generative framework to learn the conditional probability of a "true" event given an "observed" event. The core of this method is Conditional Flow Matching (CFM), which learns a vector field to transform a simple Gaussian distribution into the complex distribution of physics events. To handle the intricate correlations between decay products, they utilize a Transfusion Architecture—a Transformer-based encoder-decoder setup.

To solve the training bias problem, the authors introduced a two-part unbiasing strategy:

  1. Batch-wise Conditioning: A global estimator of the top mass is added as an input to the network.
  2. Multi-mass Training: The network is trained on a range of top masses simultaneously without explicit labels, forcing it to learn sensitivity to the mass directly from kinematic distributions.

Precision Mass Measurement

The study validated this approach using a simulated dataset mimicking a CMS analysis. The method successfully unfolded data generated with a top mass of 173.5 GeV using a network trained on a range centered at 172.5 GeV, accurately recovering the true peak without being pulled toward the training center.

Sensitivity of the jet mass distribution to the top-quark mass (mt) in CMS data, serving as a benchmark for unfolding precision.

In mock mass measurements, the generative method achieved a statistical precision of $\pm 0.17$ GeV, outperforming the standard binned TUnfold method ($\pm 0.21$ GeV). Furthermore, the authors demonstrated the first successful 12-dimensional unbinned unfolding of top-quark decays, capturing complex correlations across the entire decay chain.

Conclusion and Future Outlook

While the results are promising, challenges remain. The generative network currently lacks inherent knowledge of physical bounds (e.g., angular limits), requiring post-processing acceptance cuts. Future work will focus on integrating background subtraction and applying the method to high-fidelity GEANT4 simulations used by major collaborations like CMS. By moving beyond traditional binning, this generative approach opens new doors for precision measurements and searches for new physics at the LHC.

Reference: Favaro, L., et al. (2025). How to Unfold Top Decays. arXiv:2501.12363v3 [hep-ph].

Summary updated

Abstract

Using unfolded top-quark decay data we can measure the top quark mass, as well as search for unexpected kinematic effects. We show how generative unfolding enables both tasks and how both benefit from unbinned,high-dimensional unfolding. Our method includes an unbiasing step with respect to the training data and promises significant advantages over standard methods, in terms of flexibility and precision.
BibTeX
@inproceedings{favaro2025how,
  author = {Luigi Favaro and Roman Kogler and A. Paasch and Sofia Palacios Schweitzer and Tilman Plehn and Dennis Schwarz},
  title = {{How to Unfold Top Decays}},
  booktitle = {SciPost},
  year = {2025},
  doi = {10.21468/scipostphyscore.8.3.053},
}

Loading related content...

Figure