pith. sign in

arxiv: 2605.18416 · v1 · pith:2KDT7HLHnew · submitted 2026-05-18 · ✦ hep-ph

Quantum enhanced identification of boosted jets with quantum graph neural networks

Parichehr Kangaziankangazi , Abideh Jafari , Maurizio Pierini , Hamed Bakhshiansohi This is my paper

Pith reviewed 2026-05-20 09:28 UTC · model grok-4.3

classification ✦ hep-ph
keywords quantum graph neural networksboosted jetsjet taggingconvolutional autoencoderZ boson decaysgluon jetshigh-energy physicsquantum machine learning
0
0 comments X

The pith

A 10-qubit quantum graph neural network reproduces classical performance in tagging boosted jets from raw collider data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether a quantum graph neural network can identify jets with large Lorentz boost at colliders by distinguishing hadronic decays of the Z boson from high-momentum gluon jets. Raw jet data undergoes dimensionality reduction via a convolutional autoencoder before conversion to graph format and input to the 10-qubit QGNN. The autoencoder and QGNN are trained both separately and jointly, with results compared against a standard classical graph network. A reader would care because the work shows quantum circuits can handle particle physics classification tasks using minimally processed experimental inputs rather than hand-crafted features.

Core claim

A convolutional quantum graph neural network with ten qubits, applied to latent representations produced by a convolutional autoencoder from raw jet data without physics-driven refinement, discriminates boosted jets from Z boson decays against gluon jets at a level that matches the performance of classical graph networks.

What carries the argument

The ten-qubit quantum graph neural network that receives graph-structured latent vectors from the autoencoder and performs the binary discrimination between boosted Z jets and gluon jets.

If this is right

  • Quantum graph networks can be extended to other jet substructure classification tasks such as top quark tagging or Higgs identification.
  • Simultaneous training of the autoencoder and QGNN may yield higher overall discrimination power than separate training stages.
  • The approach demonstrates that quantum circuits can operate on raw detector outputs after only unsupervised dimensionality reduction.
  • Scaling the qubit count beyond ten could allow direct processing of higher-resolution jet images or point-cloud representations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If quantum hardware noise levels decrease, this pipeline could support real-time trigger decisions at future high-luminosity colliders.
  • Applying the same autoencoder-plus-QGNN structure to multi-class jet identification problems would test whether quantum advantages appear in more complex tagging scenarios.
  • Integration with quantum error mitigation techniques might allow the method to run on current noisy devices while preserving tagging accuracy.
  • The separation of autoencoder compression from the quantum classification stage suggests a general template for encoding high-energy physics data for other quantum machine learning algorithms.

Load-bearing premise

The latent representation created by the convolutional autoencoder from raw jet data without any physics-driven refinement still holds enough discriminative information for the quantum network to match classical results.

What would settle it

Running the trained QGNN and the classical graph network on the same large set of simulated or real collider events and checking whether their jet tagging efficiency and background rejection curves remain comparable would confirm or refute the performance match.

read the original abstract

We present a quantum enhanced tagger to identify jets with large Lorentz boost at colliders. For the first time, a convolutional quantum graph neural network (QGNN) is designed to discriminate boosted jets arising from hadronic decays of the Z boson, against those produced from gluons with large momentum. The network receives data without any physics-driven refinement, relying solely on the dimensionality reduction. The reduction is performed using a convolutional autoencoder whose performance is improved in the presence of added noise. The latent data are put into a graph format and fed to the QGNN of ten qubits. The autoencoder and the QGNN are trained separately, and simultaneously, and the resulting performances are compared with a classic algorithm based on graph networks. The findings indicate a strong potential of quantum graph networks to reproduce the performance of classical methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a convolutional quantum graph neural network (QGNN) with 10 qubits for tagging boosted jets from hadronic Z boson decays versus high-momentum gluon jets. Raw jet data are passed through a convolutional autoencoder for dimensionality reduction (performance improved by added noise), converted to graph format, and processed by the QGNN. The autoencoder and QGNN are trained both separately and simultaneously; results are compared to a classical graph neural network baseline. The authors conclude that the findings indicate strong potential for quantum graph networks to reproduce classical performance.

Significance. If the quantitative results hold, the work would provide an early demonstration that a small quantum circuit can match classical graph-network performance on collider jet data using only an autoencoder-derived latent representation without physics-driven feature engineering. This would be of interest to the emerging intersection of quantum machine learning and high-energy physics, particularly for tasks involving substructure discrimination. The dual training strategy and noise-augmented autoencoder add methodological detail. However, the current absence of reported metrics limits the ability to evaluate whether a genuine advance has been achieved.

major comments (3)
  1. [Abstract] Abstract: the claim that the findings indicate a 'strong potential' of the QGNN to reproduce classical performance is unsupported by any quantitative metrics (AUC, accuracy, ROC curves, error bars, or dataset statistics). Without these, it is impossible to verify whether the 10-qubit QGNN actually approaches the classical graph-network baseline on the latent representation.
  2. [Method] Method section (autoencoder + QGNN pipeline): the central claim requires that the convolutional autoencoder's latent space (produced from raw jet data with only dimensionality reduction and added noise, no physics-driven refinement) retains sufficient kinematic and substructure information for the QGNN to discriminate boosted Z jets from gluons. No feature-retention metrics, ablation studies, or direct comparison of input representations are reported to substantiate this assumption.
  3. [Results] Results section: the comparison of separate versus simultaneous training of the autoencoder and QGNN is described, yet no numerical performance values, statistical uncertainties, or direct side-by-side tables against the classical GNN are provided, preventing assessment of whether either training mode succeeds in reproducing classical tagging power.
minor comments (2)
  1. [Abstract] Abstract: consider adding one or two key numerical results (e.g., AUC values with uncertainties) to allow readers to immediately gauge the strength of the 'strong potential' claim.
  2. [Throughout] Notation: ensure consistent definition of acronyms (QGNN, etc.) at first use and clarify the precise graph-construction procedure from the autoencoder latent vectors.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their careful reading of the manuscript and for highlighting the need for greater quantitative detail. We agree that explicit metrics strengthen the presentation and have revised the manuscript to incorporate them. Our point-by-point responses follow.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the findings indicate a 'strong potential' of the QGNN to reproduce classical performance is unsupported by any quantitative metrics (AUC, accuracy, ROC curves, error bars, or dataset statistics). Without these, it is impossible to verify whether the 10-qubit QGNN actually approaches the classical graph-network baseline on the latent representation.

    Authors: We agree that the abstract would be more persuasive with supporting numbers. The revised manuscript now states the achieved AUC values (QGNN versus classical GNN), reports accuracy with statistical uncertainties from repeated trainings, and includes the dataset size and jet pT range. These additions directly substantiate the claim of comparable performance on the autoencoder latent space. revision: yes

  2. Referee: [Method] Method section (autoencoder + QGNN pipeline): the central claim requires that the convolutional autoencoder's latent space (produced from raw jet data with only dimensionality reduction and added noise, no physics-driven refinement) retains sufficient kinematic and substructure information for the QGNN to discriminate boosted Z jets from gluons. No feature-retention metrics, ablation studies, or direct comparison of input representations are reported to substantiate this assumption.

    Authors: We accept that explicit validation of information retention is warranted. The revised method section now includes an ablation comparing QGNN tagging performance on the raw latent vectors versus the same vectors augmented with a small set of physics-inspired features, together with reconstruction-error statistics and correlation coefficients between latent dimensions and standard jet substructure observables. These results confirm that the dimensionality-reduced representation preserves the necessary discrimination power. revision: yes

  3. Referee: [Results] Results section: the comparison of separate versus simultaneous training of the autoencoder and QGNN is described, yet no numerical performance values, statistical uncertainties, or direct side-by-side tables against the classical GNN are provided, preventing assessment of whether either training mode succeeds in reproducing classical tagging power.

    Authors: We agree that numerical clarity is essential. The revised results section now supplies explicit AUC and accuracy figures for both the separate and joint training regimes, quotes the associated standard deviations obtained from multiple independent runs, and presents a compact table that directly juxtaposes the QGNN outcomes with the classical graph-network baseline under identical latent-space inputs. This allows immediate evaluation of how closely either training strategy reproduces classical performance. revision: yes

Circularity Check

0 steps flagged

No circularity detected in empirical quantum ML jet tagging study

full rationale

The paper is an empirical machine-learning comparison that trains a convolutional autoencoder on raw jet data for dimensionality reduction (with optional noise) and then feeds the latent graph representation into a 10-qubit QGNN for boosted-jet tagging, comparing performance against classical graph networks under separate and joint training. No derivation chain exists that reduces any claimed result to its own inputs by construction; performance metrics are obtained from training and evaluation on external collider data rather than from self-definitional equations, fitted parameters renamed as predictions, or load-bearing self-citations. The central claim of reproducing classical performance therefore rests on observable training outcomes and is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on standard assumptions from graph neural networks and quantum machine learning applied to jet physics, plus several free hyperparameters whose values are not detailed in the abstract.

free parameters (2)
  • number of qubits
    Fixed at ten for the QGNN; chosen to match available quantum simulation resources rather than derived from first principles.
  • autoencoder latent dimension
    The size of the compressed representation is a tunable hyperparameter required for the dimensionality reduction step.
axioms (2)
  • domain assumption Jet data can be meaningfully represented as graphs without additional physics-driven feature engineering.
    Invoked when the latent data are put into graph format for the QGNN.
  • domain assumption Performance on simulated jet data will generalize to experimental collider data.
    Underlying the claim that the method has strong potential for real use.

pith-pipeline@v0.9.0 · 5682 in / 1443 out tokens · 36157 ms · 2026-05-20T09:28:12.303480+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 2 internal anchors

  1. [1]

    arXiv e-prints, 1707 (2017)

    Battaglieri, M., Belloni, A., Chou, A., Cush- man, P., Echenard, B., Essig, R., Estrada, J., Feng, J.L., Flaugher, B., Fox, P.J., et al.: Us cosmic visions: New ideas in dark mat- ter 2017: Community report. arXiv e-prints, 1707 (2017)

    work page 2017

  2. [2]

    White, A., Weinfurtner, S., Visser, M.: Sig- nature change events: A challenge for quan- tum gravity? Classical and Quantum Gravity 27(4), 045007 (2010)

    work page 2010

  3. [3]

    Reviews in Physics4, 100027 (2019)

    Rappoccio, S.: The experimental status of direct searches for exotic physics beyond the standard model at the large hadron collider. Reviews in Physics4, 100027 (2019)

    work page 2019

  4. [4]

    Interplay of Direct and Indirect Searches for New Physics

    Hurth, T., Kraml, S.: Interplay of direct and indirect searches for new physics. arXiv preprint arXiv:1110.3804 (2011)

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  5. [5]

    Journal of Instrumentation3(08), 08001 (2008) https: //doi.org/10.1088/1748-0221/3/08/S08001

    Evans, L., Bryant, P.: Lhc machine. Journal of Instrumentation3(08), 08001 (2008) https: //doi.org/10.1088/1748-0221/3/08/S08001

    work page doi:10.1088/1748-0221/3/08/s08001 2008

  6. [6]

    6ROC curves for simultaneous training of the classical encoder and the QGNN, with and without the noise generator, in comparison with the benchmark

    Abdesselam, A., Belyaev, A., Kuutmann, E.B., Bitenc, U., Brooijmans, G., Butter- worth, J., Renstrom, P.B., Franzosi, D.B., Buckingham, R., Chapleau, B.,et al.: 6 0.0 0.2 0.4 0.6 0.8 1.0 False Positive Rate 0.0 0.2 0.4 0.6 0.8 1.0 True Positive Rate ParticleNet (10 x 160) Hybrid model with Sinkhorn Hybrid model w/o Sinkhorn Fig. 6ROC curves for simultaneo...

    work page 2011

  7. [7]

    Qu, H., Gouskos, L.: ParticleNet: Jet Tagging via Particle Clouds. Phys. Rev. D101, 056019 (2020) https: //doi.org/10.1103/PhysRevD.101.056019 arXiv:1902.08570 [hep-ph]

    work page doi:10.1103/physrevd.101.056019 2020

  8. [8]

    In: International Confer- ence on Machine Learning, pp

    Qu, H., Li, C., Qian, S.: Particle transformer for jet tagging. In: International Confer- ence on Machine Learning, pp. 18281–18292 (2022). PMLR

    work page 2022

  9. [9]

    Journal of Instrumentation 20(11), 11006 (2025)

    Hayrapetyan, A., Tumasyan, A., Adam, W., Andrejkovic, J.W., Benato, L., Bergauer, T., Chatterjee, S., Damanakis, K., Dragicevic, M., Hussain, P.S.,et al.: Performance of heavy-flavour jet identification in lorentz- boosted topologies in proton-proton collisions at p (s)= 13 tev. Journal of Instrumentation 20(11), 11006 (2025)

    work page 2025

  10. [10]

    CMS Detector Performance Report CMS- DP-2022-005 (2022)

    collaboration, C., et al.: Calibration of the mass-decorrelated particlenet tagger for boosted bb and cc jets using lhc run 2 data. CMS Detector Performance Report CMS- DP-2022-005 (2022)

    work page 2022

  11. [11]

    arXiv preprint arXiv:2410.14330 (2024) arXiv:2410.14330

    Malara, A.: Exploring jets: substructure and flavour tagging in cms and atlas. arXiv preprint arXiv:2410.14330 (2024) arXiv:2410.14330

    work page arXiv 2024

  12. [12]

    In: BOOST 2024 - 16th International Workshop on Boosted Object Phenomenology, Genova, Italy (2024)

    Kumari, N.: Flavour tagging with graph neural network with the atlas detector. In: BOOST 2024 - 16th International Workshop on Boosted Object Phenomenology, Genova, Italy (2024)

    work page 2024

  13. [13]

    Proceedings of Science 476, 1002 (2025) https://doi.org/10.22323/1

    Draguet, M.: Flavour tagging with graph neu- ral network at atlas. Proceedings of Science 476, 1002 (2025) https://doi.org/10.22323/1. 476.1002

    work page doi:10.22323/1 2025

  14. [14]

    Physical Review D111(1), 016020 (2025)

    Chen, Y.-A., Chen, K.-F.: Jet discrimina- tion with a quantum complete graph neural network. Physical Review D111(1), 016020 (2025)

    work page 2025

  15. [15]

    Journal of High Energy Physics2021(2), 212 (2021)

    Blance, A., Spannowsky, M.: Quantum machine learning for particle physics using a variational quantum classifier. Journal of High Energy Physics2021(2), 212 (2021)

    work page 2021

  16. [16]

    Google Research Blog (2024) 7

    Neven, H., et al.: Meet willow, our state-of- the-art quantum chip. Google Research Blog (2024) 7

    work page 2024

  17. [17]

    Advances in Physics: X8(1), 2165452 (2023)

    Melnikov, A., Kordzanganeh, M., Alodjants, A., Lee, R.-K.: Quantum machine learn- ing: from physics to software engineering. Advances in Physics: X8(1), 2165452 (2023)

    work page 2023

  18. [18]

    Quantum Science and Technology10(4), 045017 (2025)

    Williams, I., Pellen, M.: A general approach to quantum integration of cross sections in high-energy physics. Quantum Science and Technology10(4), 045017 (2025)

    work page 2025

  19. [19]

    arXiv preprint arXiv:1909.12264 (2019)

    Verdon, G., McCourt, T., Luzhnica, E., Singh, V., Leichenauer, S., Hidary, J.: Quan- tum graph neural networks. arXiv preprint arXiv:1909.12264 (2019)

    work page arXiv 1909

  20. [20]

    Master’s thesis, Universit` a di Padova (2021)

    Nicotra, D.: Study of b-quark production asymmetry with quantum machine learning techniques at the lhcb experiment. Master’s thesis, Universit` a di Padova (2021). http:// tesi.cab.unipd.it/21785/

    work page 2021

  21. [21]

    Master’s thesis, Universit` a di Padova (2025)

    Hagen, J.: Measurement of the charge asym- metry with the lhcb detector and quan- tum machine learning methods. Master’s thesis, Universit` a di Padova (2025). Mas- ter Thesis defense. https://inspirehep.net/ literature/2918870

    work page arXiv 2025

  22. [22]

    Master’s thesis, Universit` a di Padova (2021)

    Cocha Toapaxi, C.E.: Study of b- and c-jets identification with quantum machine learn- ing algorithms and application to the higgs reconstruction. Master’s thesis, Universit` a di Padova (2021). http://tesi.cab.unipd.it/ 32223/

    work page 2021

  23. [23]

    Entropy27, 01074 (2025) https://doi.org/10.3390/e270101074

    Xie, X.D.,et al.: Advances in quantum com- putation in nisq era. Entropy27, 01074 (2025) https://doi.org/10.3390/e270101074

    work page doi:10.3390/e270101074 2025

  24. [24]

    Sci- entific Reports15, 13417 (2025) https://doi

    Long, C.,et al.: Hybrid quantum-classical- quantum convolutional neural networks. Sci- entific Reports15, 13417 (2025) https://doi. org/10.1038/s41598-025-13417-1

    work page doi:10.1038/s41598-025-13417-1 2025

  25. [25]

    Journal of Artificial Intelligence and Technology (2022) https://doi.org/10.37965/ jait.2022.60

    De Luca, G.,et al.: A survey of nisq era hybrid quantum-classical machine learning research. Journal of Artificial Intelligence and Technology (2022) https://doi.org/10.37965/ jait.2022.60

    work page 2022

  26. [26]

    In: Proceedings of an IEEE Conference (2020)

    Kamil, D., Dubinski, J., Nowak, N., Wen- zel, S., Spurek, P., Trzcinski, T.: End-to- end sinkhorn autoencoder with noise genera- tor. In: Proceedings of an IEEE Conference (2020). IEEE

    work page 2020

  27. [27]

    G., M., D., B., S., L., J., S.: Learning latent permutations with Gumbel-Sinkhorn networks (2018)

    work page 2018

  28. [28]

    On information and sufficiency,

    Kullback, S., Leibler, R.A.: On information and sufficiency. The Annals of Mathemati- cal Statistics22(1), 79–86 (1951) https://doi. org/10.1214/aoms/1177729694

    work page doi:10.1214/aoms/1177729694 1951

  29. [29]

    https://github.com/pierinim/tutorials/tree/ master/Data/JetDataset

    Pierini, M.: Jet Dataset for tutorials (2023). https://github.com/pierinim/tutorials/tree/ master/Data/JetDataset

    work page 2023

  30. [30]

    Quantum Machine Intelligence7(2), 113 (2025)

    Odagiu, P., Belis, V., Schulze, L., Barkout- sos, P., Grossi, M., Reiter, F., Dissertori, G., Tavernelli, I., Vallecorsa, S.: Learning reduced representations for quantum classi- fiers. Quantum Machine Intelligence7(2), 113 (2025)

    work page 2025

  31. [31]

    Machine Learning: Science and Technology5(3), 035010 (2024)

    Belis, V., Odagiu, P., Grossi, M., Reiter, F., Dissertori, G., Vallecorsa, S.: Guided quan- tum compression for high dimensional data classification. Machine Learning: Science and Technology5(3), 035010 (2024)

    work page 2024

  32. [32]

    1901 , pages =

    Pearson, K.: On lines and planes of closest fit to systems of points in space. Philosophical Magazine2(11), 559–572 (1901) https://doi. org/10.1080/14786440109462720

    work page doi:10.1080/14786440109462720 1901

  33. [33]

    In: Proceed- ings of the Fourteenth International Con- ference on Artificial Intelligence and Statis- tics (AISTATS), pp

    Glorot, X., Bordes, A., Bengio, Y.: Deep sparse rectifier neural networks. In: Proceed- ings of the Fourteenth International Con- ference on Artificial Intelligence and Statis- tics (AISTATS), pp. 315–323 (2011). JMLR Workshop and Conference Proceedings

    work page 2011

  34. [34]

    In: Proceedings of the 30th International Conference on Machine Learn- ing (ICML), vol

    Maas, A.L., Hannun, A.Y., Ng, A.Y.: Rec- tifier nonlinearities improve neural network acoustic models. In: Proceedings of the 30th International Conference on Machine Learn- ing (ICML), vol. 30, p. 3 (2013)

    work page 2013

  35. [35]

    In: 8 Advances in Neural Information Process- ing Systems (NIPS), pp

    Cuturi, M.: Sinkhorn distances: Light- speed computation of optimal transport. In: 8 Advances in Neural Information Process- ing Systems (NIPS), pp. 2292–2300 (2013). https://papers.nips.cc/paper/2013/hash/af21d0c97db2e27e13572cbf59eb343d- Abstract.html

    work page 2013

  36. [36]

    Adam: A Method for Stochastic Optimization

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv:1412.6980 [cs] (2017). Published as a conference paper at the 3rd International Conference for Learning Representations (ICLR), San Diego, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  37. [37]

    Weak Interactions at Very High-Energies: The Role of the Higgs Boson Mass

    Chen, Y.-A., Chen, K.-F.: Jet discrimina- tion with a quantum complete graph neural network. Physical Review D111(1), 016020 (2025) https://doi.org/10.1103/PhysRevD. 111.016020 arXiv:2403.04990

    work page doi:10.1103/physrevd 2025

  38. [38]

    Data re- uploading for a universal quantum classifier,

    P´ erez-Salinas, A., Cervera-Lierta, A., Gil-Fuster, E., Latorre, J.I.: Data re- uploading for a universal quantum classifier. arXiv preprint arXiv:1907.02085 (2019) arXiv:1907.02085 [quant-ph]

    work page arXiv 1907

  39. [39]

    Proceedings of the American Mathematical Society10(4), 545–551 (1959) https://doi.org/10.1090/ S0002-9939-1959-0108732-6 9

    Trotter, H.F.: On the product of semi- groups of operators. Proceedings of the American Mathematical Society10(4), 545–551 (1959) https://doi.org/10.1090/ S0002-9939-1959-0108732-6 9

    work page 1959