This story draft by @escholar has not been reviewed by an editor, YET.

Discussion, Acknowledgments, and References

EScholar: Electronic Academic Papers for Scholars HackerNoon profile picture
0-item

Table of Links

Abstract and 1. Introduction

  1. Unfolding

    2.1 Posing the Unfolding Problem

    2.2 Our Unfolding Approach

  2. Denoising Diffusion Probabilistic Models

    3.1 Conditional DDPM

  3. Unfolding with cDDPMs

  4. Results

    5.1 Toy models

    5.2 Physics Results

  5. Discussion, Acknowledgments, and References


Appendices

A. Conditional DDPM Loss Derivation

B. Physics Simulations

C. Detector Simulation and Jet Matching

D. Toy Model Results

E. Complete Physics Results

6 Discussion

Our results demonstrate that a single cDDPM can successfully unfold detector effects on particle jets from a variety of physics processes, including those not seen during training. The distinguishing feature of our method is its non-iterative and flexible posterior sampling approach, which exhibits a strong inductive bias that allows the cDDPM to generalize to unseen processes without explicitly assuming the underlying physics distribution, setting it apart from other unfolding techniques so far.


The cDDPM’s ability to unfold data from combined processes presents a new opportunity to handle detector data prior to background subtraction, streamlining the unfolding process. These results are consistent for both the DELPHES CMS and the analytical data-driven detector simulations, indicating that the cDDPM’s performance is not drastically limited by the degree of detector smearing. We expect this approach to be applicable to other particles, detector-level observables, and event-wise quantities, enabling the reconstruction of full events after unfolding.


Several open questions remain regarding the implementation of the conditioning on the moments. These include optimal selection of priors and the number of moments required for the best unfolding performance. Further investigation is needed to determine the extent of the cDDPM’s inductive bias and its tolerance to variations in the underlying physics processes. Understanding these aspects will be crucial for refining the method and ensuring its robustness across a wide range of scenarios.


While this approach shows promise, improvements such as uncertainty estimation, accounting for systematic and experimental uncertainties, and handling particles falling outside detector thresholds are necessary to fully realize its potential. Incorporating these enhancements will be essential for making the cDDPM a reliable and comprehensive unfolding tool in high-energy physics. We leave these developments for future work.




Acknowledgments

This work has been made possible thanks to the support of the Department of Energy Office of Science through the Grant DE-SC0023964. Shuchin Aeron and Taritree Wonhjirad would also like to acknowledge support by the National Science Foundation under Cooperative Agreement PHY-2019786 (The NSF AI Institute for Artificial Intelligence and Fundamental Interactions, http://iaifi.org/).


References

[1] Anders Andreassen et al. “OmniFold: A Method to Simultaneously Unfold All Observables”. In: Phys. Rev. Lett. 124 (18 May 2020), p. 182001. doi: 10 . 1103 / PhysRevLett . 124 . 182001. url: https : //link.aps.org/doi/10.1103/PhysRevLett.124.182001.


[2] Mathias Backes et al. An unfolding method based on conditional Invertible Neural Networks (cINN) using iterative training. 2024. arXiv: 2212.08674 [hep-ph].


[3] Christian Bierlich et al. A comprehensive guide to the physics and usage of PYTHIA 8.3. 2022. arXiv: 2203.11601 [hep-ph].


[4] Volker Blobel. An Unfolding Method for High Energy Physics Experiments. 2002. arXiv: hep - ex / 0208022 [hep-ex].


[5] Lydia Brenner et al. “Comparison of unfolding methods using RooFitUnfold”. In: Int. J. Mod. Phys. A 35.24 (2020), p. 2050145. doi: 10.1142/S0217751X20501456. arXiv: 1910.14654 [physics.data-an].


[6] Matteo Cacciari, Gavin P. Salam, and Gregory Soyez. “FastJet user manual: (for version 3.0.2)”. In: The European Physical Journal C 72.3 (Mar. 2012). issn: 1434-6052. doi: 10.1140/epjc/s10052- 012-1896-2. url: http://dx.doi.org/10.1140/epjc/s10052-012-1896-2.


[7] Jooyoung Choi et al. ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models. 2021. arXiv: 2108.02938 [cs.CV].


[8] ATLAS Collaboration. “Determination of jet calibration and energy resolution in proton–proton collisions at √ s = 8 TeV using the ATLAS detector”. In: The European Physical Journal C 80.12 (Dec. 2020). issn: 1434-6052. doi: 10.1140/epjc/s10052-020-08477-8. url: http://dx.doi.org/10. 1140/epjc/s10052-020-08477-8.


[9] Kaustuv Datta, Deepak Kar, and Debarati Roy. Unfolding with Generative Adversarial Networks. 2018. arXiv: 1806.00433 [physics.data-an].


[10] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. 2021. arXiv: 2105.05233 [cs.LG].


[11] Sascha Diefenbacher et al. Improving Generative Model-based Unfolding with Schr¨odinger Bridges. 2023. arXiv: 2308.12351 [hep-ph].


[12] J. de Favereau et al. “DELPHES 3: a modular framework for fast simulation of a generic collider experiment”. In: Journal of High Energy Physics 2014.2 (Feb. 2014). issn: 1029-8479. doi: 10.1007/ jhep02(2014)057. url: http://dx.doi.org/10.1007/JHEP02(2014)057.


[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. 2020. arXiv: 2006.11239 [cs.LG].


[14] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. 2022. arXiv: 2207.12598 [cs.LG].


[15] Nathan Huetsch et al. “The Landscape of Unfolding with Machine Learning”. In: arXiv preprint arXiv:2404.18807 (2024). [16] Alexander Shmakov et al. End-To-End Latent Variational Diffusion Models for Inverse Problems in High Energy Physics. 2023. arXiv: 2305.10399 [hep-ex].


A Conditional DDPM Loss Derivation

In the proposed conditional DDPM, the forward process is a Markov chain that gradually adds Gaussian noise to the data according to a variance schedule β.


Training is performed by optimizing the variational bound on negative log likelihood:


Following the similar derivation provided in [13], this loss can then be rewritten using the KL-divergence



Table 1: List of physics simulations generated for the DELPHES CMS detector simulation, along with the corresponding parton distribution functions (PDFs), phase space biases, and their inclusion in the training dataset.


and the reverse process posterior as


B Physics Simulations


Table 2: List of physics simulations generated for the data-driven detector simulation, along with the corresponding parton distribution functions (PDFs), parton shower models, phase space biases, and their inclusion in the training dataset. Simulations not included in the training dataset were used as test datasets.


C Detector Simulation and Jet Matching

Our results present the unfolding of detector effects from two different detector simulation frameworks: DELPHES and a data-driven approach. DELPHES is a framework developed for the simulation of multipurpose detectors for physics studies [12]. Specifically, the DELPHES CMS configuration is frequently used as the detector simulation of choice in recent machine-learning based unfolding studies.








D Toy Model Results

The full unfolding results of the toy model tests are shown here, including the η, ϕ, and E distributions of the data 4-vectors. In fig. 9 we show the full results of multidimensional unfolding tests as well as the tests on the cDDPM dependence on the training prior when learning the posterior. In fig. 10 the full results for the moments-based unfolding are shown.


Figure 9: Complete unfolding results for the multidimensional toy model tests, demonstrating the cDDPM’s ability to learn the posterior P(x|y) given a dataset of pairs {x, y}. The bottom row shows the unfolding performance when the cDDPM is trained on an alternative dataset with different marginal distributions but the same posterior, confirming that the cDDPM’s sampling is based on the learned posterior without significant dependence on the training prior.


In fig. 10 we present the complete unfolding results for the multidimensional toy model tests using a class of exponential functions. The successful unfolding of all particle properties for the test datasets underscores the cDDPM’s capacity to generalize beyond the specific distributions seen during training.


E Complete Physics Results

In fig. 11 and fig. 12 we present the unfolding results for the remaining particle properties (η, ϕ, E, px, py, pz) that were not shown in the main text. These additional plots demonstrate the cDDPM’s ability to successfully unfold the full particle vector, providing a comprehensive view of its performance across all dimensions.


The tables in this appendix (table 3 and table 4) provide a detailed breakdown of the unfolding performance metrics for each particle property and dataset. Furthermore, these metrics (Energy distance, Wasserstein distance between histograms, and KL divergence between histograms) offer a more in-depth perspective on the unfolding quality.


Figure 10: Complete unfolding results for the multidimensional toy model tests using a class of exponential functions. The cDDPM is trained on datasets with different exponential distributions (characterized by βi) and conditioned on the moments of the pT distributions. The successful unfolding of the test datasets demonstrates the cDDPM’s ability to interpolate and extrapolate within the class of distributions based on the provided moments.


Figure 11: Unfolding performance of a single cDDPM on the remaining particle properties (η, ϕ, px, py, pz) for various simulated physics processes with detector effects simulated using DELPHES CMS. These plots show the particle vector properties that were not included in the main results in fig. 5.


Figure 12: Unfolding performance of a single cDDPM on the remaining particle properties (η, ϕ, px, py, pz) for various simulated physics processes with detector effects simulated using an analystical data-driven approach. These plots show the particle vector properties that were not included in the main results in fig. 6.


Table 3: Metrics for evaluating the unfolding performance on datasets with the DELPHES CMS detector simulation. The Wasserstein distance, Energy distance, Wasserstein distance between histograms, and KL divergence between histograms are reported for each observable (pT , η, ϕ, E, px, py, pz) of the unfolded and detector-level datasets, compared to the truth-level dataset.


Authors:

(1) Camila Pazos, Department of Physics and Astronomy, Tufts University, Medford, Massachusetts;

(2) Shuchin Aeron, Department of Electrical and Computer Engineering, Tufts University, Medford, Massachusetts and The NSF AI Institute for Artificial Intelligence and Fundamental Interactions;

(3) Pierre-Hugues Beauchemin, Department of Physics and Astronomy, Tufts University, Medford, Massachusetts and The NSF AI Institute for Artificial Intelligence and Fundamental Interactions;

(4) Vincent Croft, Leiden Institute for Advanced Computer Science LIACS, Leiden University, The Netherlands;

(5) Martin Klassen, Department of Physics and Astronomy, Tufts University, Medford, Massachusetts;

(6) Taritree Wongjirad, Department of Physics and Astronomy, Tufts University, Medford, Massachusetts and The NSF AI Institute for Artificial Intelligence and Fundamental Interactions.


This paper is available on arxiv under CC BY 4.0 DEED license.


L O A D I N G
. . . comments & more!

About Author

EScholar: Electronic Academic Papers for Scholars HackerNoon profile picture
EScholar: Electronic Academic Papers for Scholars@escholar
We publish the best academic work (that's too often lost to peer reviews & the TA's desk) to the global tech community

Topics

Around The Web...

Trending Topics

blockchaincryptocurrencyhackernoon-top-storyprogrammingsoftware-developmenttechnologystartuphackernoon-booksBitcoinbooks