How to unfold top decays

Luigi Favaro, R. Kogler, A. Paasch, Sofia Palacios Schweitzer, Tilman Plehn, D. Schwarz

TL;DR

This paper introduces a novel generative unfolding method for analyzing high-dimensional particle physics data, specifically top quark decays. By addressing a critical bias issue inherent in current methods, they demonstrate significantly improved accuracy and flexibility in measuring fundamental particle properties and searching for new physics.

What Problem Does This Paper Address?

In particle physics, experiments observe particles after they've interacted with detectors, which distort their original properties. "Unfolding" is the process of mathematically reversing these detector effects to reconstruct the original particle behavior. This is crucial for precise measurements and discovering new physics. However, traditional unfolding methods struggle with high-dimensional data (many variables) and are susceptible to "training bias" – where the simulation used to train the unfolding algorithm has a different assumed value for a key parameter (like the top quark mass) than the actual data. This bias can lead to inaccurate measurements. This paper tackles the challenge of performing unbinned, high-dimensional unfolding for complex processes like top quark decays, while mitigating this critical training bias.

Methods

The authors propose a new generative unfolding method that leverages deep learning. Unlike existing methods that rely on reweighting or iterative approaches, their method incorporates a "targeted unbiasing" with respect to the training data. This involves a combination of techniques:

  • Generative Networks: They employ Conditional Flow Matching (CFM) networks, a state-of-the-art architecture for generative modeling, to learn the complex mapping between simulated detector-level (reco) and particle-level (gen) data.
  • Phase Space Parametrization: To handle the high dimensionality of top quark decays (12 dimensions), they strategically select a reduced set of key kinematic variables (like jet masses) for initial unfolding, and then extend to the full phase space.
  • Bias Mitigation: A key innovation is addressing the training bias. This is achieved by:
  • Batch-wise Conditioning: Including an estimator of the top quark mass (e.g., the median of jet masses) directly in the network's input during training.
  • Combined Training: Training the network on simulated data generated with a range of top quark masses, rather than a single value.
  • Transfusion Network: For higher dimensional unfolding, they utilize a "transfusion network," a transformer-based architecture designed to handle complex correlations between input variables.

Datasets

The study uses simulated events for top pair production, generated with Madgraph 5 and processed with Pythia 8 for parton showering and hadronization, and Delphes 3.5.0 for detector simulation. This dataset is based on a previous CMS analysis [34]. They generate samples with different top quark masses (e.g., 169.5, 171.5, 172.5, 173.5, 175.5 GeV) to train and test their bias mitigation strategies.

Key Insights and Ideas

  • Generative Unfolding is Powerful for High Dimensions: The paper demonstrates that modern generative networks can effectively handle the complexity of unfolding in high-dimensional phase spaces, a significant leap from traditional binned methods.
  • Training Bias is a Major Hurdle: The study highlights how a mismatch in the assumed top quark mass between simulation and data can completely dominate the unfolded results, rendering measurements unreliable.
  • Targeted Unbiasing is Key: Simply using iterative methods isn't enough. The proposed combination of batch-wise conditioning and training on a range of masses is crucial for overcoming the training bias.
  • Trade-offs in High-Dimensionality: While unfolding more dimensions improves mass measurement accuracy, it can introduce subtle issues in other kinematic correlations (like azimuthal angles), requiring careful handling.

Key Results

  • The proposed generative unfolding method achieves percent-level precision in unfolding key kinematic distributions (like the 3-jet mass) in a 4- and 6-dimensional phase space.
  • The bias mitigation techniques successfully reduce the impact of the training mass uncertainty, making it a much smaller contributor to the overall top mass uncertainty compared to previous analyses.
  • The method allows for the extraction of the top quark mass with competitive statistical precision, comparable to or better than traditional methods, while offering greater flexibility.
  • The full 12-dimensional unfolding, while more complex, also demonstrates good performance in reproducing key kinematic features, with minor adjustments needed for angular correlations.

Limitations and Future Work

  • Computational Cost: High-dimensional generative unfolding can be computationally intensive, although the use of fast simulations is a step towards addressing this.
  • Angular Correlations: The paper notes that while masses are well-reproduced, the unfolding of angular correlations (like azimuthal angles) can be challenging and may require additional cuts or further refinement.
  • Real Data Performance: The study primarily uses simulated data. While promising, the performance on actual experimental data with full CMS simulations is expected to be slightly better and is available from CMS members.
  • Future Work: The authors suggest that this work serves as a blueprint for future CMS analyses, not only for top mass measurements but also for broader applications of generative unfolding. They also point to the potential for further improvements by exploring more advanced network architectures and training strategies.

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 present a new generative unfolding method for the two tasks and show how they both benefit from unbinned, high-dimensional unfolding. Unlike weight-based or iterative generative methods we include a targeted unbiasing with respect to the training data. This shows significant advantages over standard, iterative methods, in terms of applicability, flexibility and accuracy.
BibTeX
@inproceedings{favaro2025how,
  author = {Luigi Favaro and R. Kogler and A. Paasch and Sofia Palacios Schweitzer and Tilman Plehn and D. Schwarz},
  title = {{How to unfold top decays}},
  booktitle = {SciPost},
  year = {2025},
  doi = {10.21468/scipostphyscore.8.3.053},
}

Loading related content...

Figure