arxiv: 2604.05986 · v1 · submitted 2026-04-07 · 🪐 quant-ph · hep-lat

Recognition: no theorem link

Quantum Machine Learning for particle scattering entanglement classification

Hala Elhag , Yahui Chai

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:48 UTC · model grok-4.3

classification 🪐 quant-ph hep-lat

keywords quantum machine learningentanglement classificationparticle scatteringQCNNThirring modelfermion density profilesquantum correlationsclassification task

0 comments

The pith

Quantum convolutional neural networks classify particle scattering entanglement from fermion density profiles, matching or exceeding classical CNN performance especially at small sizes.

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

The paper tests whether fermion density profiles, easier to access than full entanglement measures, can proxy for entanglement in particle scattering by framing it as classification across thresholds. It uses the Thirring model of fermion scattering as the test case and pits Quantum Convolutional Neural Networks against classical CNNs with comparable parameters. The quantum models reach competitive or higher accuracy, converge faster, and show lower variance across training runs. Within the tested architectures, a compact 4-qubit QCNN outperforms larger versions, indicating that encoding quality and trainability matter more than simply adding qubits or layers. If this holds, quantum machine learning offers a practical route to extract quantum correlation information from measurable observables in high-energy physics and many-body systems.

Core claim

In the Thirring model of fermion scattering, fermion density profiles can serve as input features for classifying the degree of entanglement across various thresholds. Quantum Convolutional Neural Networks achieve consistently competitive or superior accuracy compared to classical Convolutional Neural Networks with similar parameter counts, along with faster convergence and lower variance across runs. The study finds that larger model sizes do not enhance performance and that bigger models are more sensitive to encoding choices, with a compact 4-qubit QCNN yielding the optimal results. This indicates that trainability and encoding strategies are more critical than model scaling for this task

What carries the argument

Quantum Convolutional Neural Networks (QCNNs), quantum circuit architectures that apply convolutional-like operations to quantum states, used to process fermion density profiles and output classifications of entanglement thresholds.

If this is right

QCNNs can extract nontrivial quantum information from accessible observables in particle scattering processes.
Compact quantum models can outperform larger ones for entanglement-related tasks due to better trainability.
Encoding choices critically affect the success of quantum machine learning architectures in physics applications.
These methods carry implications for analyzing quantum correlations in high-energy physics and many-body systems.

Where Pith is reading between the lines

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

The same density-to-entanglement mapping could be tested in other quantum field theory models or lattice gauge theories.
If the approach scales, it may allow entanglement studies on quantum hardware without full state reconstruction.
Similar quantum ML pipelines might classify other hard-to-measure quantities such as topological invariants from local density data.
Real experimental data from quantum simulators or particle detectors could directly test whether the observed accuracy advantage persists outside numerical simulations.

Load-bearing premise

Fermion density profiles contain enough information to serve as reliable proxies for entanglement when the task is framed as classification across multiple entanglement thresholds.

What would settle it

In a new set of Thirring-model scattering configurations with independently computed entanglement values, the 4-qubit QCNN classification accuracy drops below that of a comparable classical CNN or shows no correlation with the true entanglement thresholds.

Figures

Figures reproduced from arXiv: 2604.05986 by Hala Elhag, Yahui Chai.

**Figure 3.** Figure 3: FIG. 3. Basic structure of a QCNN model [ [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Test accuracy results averaged over 10 [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5. Test accuracy results averaged over 10 [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

read the original abstract

Entanglement is a key quantity for characterizing quantum correlations in particle scattering processes, but its direct evaluation is computationally demanding on quantum hardware. In this work, we investigate whether fermion density profiles, which are easier to access, can serve as proxies for entanglement by framing the problem as a classification task across multiple entanglement thresholds. Using the fermion scattering in the Thirring model as a test bed, we compare Quantum Convolutional Neural Networks (QCNNs) with classical CNNs of comparable parameter counts, and find that QCNNs achieve consistently competitive or superior accuracy with faster convergence and lower variance. Notably, we observe that increasing the model size does not improve the performance within the architectures studied here, and larger models appear to be more sensitive to the choice of encoding. Instead, a compact 4-qubits QCNN provides the best results, suggesting the importance of trainability and encoding choices over model scaling. These findings demonstrate the potential of quantum and quantum-inspired machine learning models for extracting nontrivial quantum information from accessible observables, with implications for high-energy physics and quantum many-body systems.

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.

Desk Editor's Note private letter to a colleague

A compact 4-qubit QCNN beats larger quantum and classical models on entanglement-threshold classification from Thirring-model density profiles, but the work provides no checks that those profiles actually carry the entanglement signal.

read the letter

The paper reports that QCNNs, especially a small 4-qubit version, reach higher accuracy, faster convergence, and lower variance than classical CNNs of similar size when classifying fermion density profiles into multiple entanglement-threshold bins for Thirring-model scattering. It also finds that increasing quantum model size hurts rather than helps and makes performance more sensitive to the choice of data encoding. Those two observations are the concrete new pieces: a direct head-to-head on this specific lattice scattering task plus the scaling result that runs counter to the usual expectation that bigger is better in quantum ML. The comparison itself is straightforward and the implication for high-energy and many-body work is clear enough on its face. Credit for running the experiment and surfacing the encoding sensitivity. The soft spot is exactly the one the stress-test flags. The abstract never shows a correlation, mutual-information, or ablation check between the local densities and the chosen entanglement measure, nor does it give dataset size, how the labels were generated, error bars, or training details. Without those, the performance numbers could reflect learning of classical patterns rather than quantum information, and the claim that density profiles serve as useful proxies stays untested. That gap is not minor for the central argument. The paper is aimed at people already working on quantum ML for lattice models or scattering processes who want a quick empirical comparison. A reader looking for a ready-to-use method or a verified proxy will come away disappointed; someone scanning for ideas on when small quantum circuits can be preferable might still pick up the scaling note. It is coherent on its own terms and engages the literature, so it clears the bar for serious refereeing even though the current version would need substantial methods and diagnostics additions before publication.

Referee Report

3 major / 1 minor

Summary. The manuscript frames entanglement classification in Thirring-model fermion scattering as a multi-threshold task using fermion density profiles as input features. It compares quantum convolutional neural networks (QCNNs) against classical CNNs of comparable parameter count and reports that QCNNs achieve competitive or superior accuracy, faster convergence, and lower variance, with a compact 4-qubit QCNN emerging as optimal; the work concludes that trainability and encoding choices matter more than model scaling for extracting nontrivial quantum information from accessible observables.

Significance. If the empirical comparison holds under proper statistical controls, the result would indicate that compact quantum machine-learning architectures can extract entanglement information from local density profiles in quantum field theory simulations more efficiently than classical counterparts, with direct relevance to high-energy physics and many-body systems. The reported observation that increasing model size does not improve performance and that larger models are more sensitive to encoding is a concrete, useful datum for understanding QML trainability limits.

major comments (3)

[Abstract] Abstract: the central claim that QCNNs achieve 'consistently competitive or superior accuracy with faster convergence and lower variance' is presented without any quantitative support (dataset size, number of independent runs, error bars, exact baseline accuracies, or training protocols), rendering the performance comparison unverifiable from the text.
[Introduction / Methods] The assumption that fermion density profiles serve as reliable proxies for entanglement thresholds (von Neumann entropy across bipartitions) is load-bearing for the entire classification task, yet no correlation, mutual-information, or ablation diagnostic is supplied to confirm that the input features carry nontrivial signal rather than spurious correlations.
[Results] Results section: the assertion that a 4-qubit QCNN is optimal and that larger models are more sensitive to encoding lacks tabulated accuracy values, variance measures, or statistical significance tests against the classical CNN baselines, so the model-size and encoding conclusions cannot be assessed.

minor comments (1)

[Abstract] Clarify whether 'quantum-inspired' models beyond QCNNs are actually evaluated, as the abstract mentions them but the comparison appears limited to QCNNs versus classical CNNs.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have helped us identify areas where the manuscript can be strengthened for greater clarity, verifiability, and rigor. We have revised the manuscript to address all three major comments by adding the requested quantitative details, validation analyses, and statistical measures. Our point-by-point responses follow.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that QCNNs achieve 'consistently competitive or superior accuracy with faster convergence and lower variance' is presented without any quantitative support (dataset size, number of independent runs, error bars, exact baseline accuracies, or training protocols), rendering the performance comparison unverifiable from the text.

Authors: We agree that the abstract requires explicit quantitative support to make the performance claims verifiable. In the revised manuscript we have updated the abstract to state the dataset size, the number of independent runs, the reported accuracies with error bars, and direct references to the baseline values and training protocols now detailed in the Methods section. These additions preserve the abstract's brevity while ensuring the comparison is fully supported by the text. revision: yes
Referee: [Introduction / Methods] The assumption that fermion density profiles serve as reliable proxies for entanglement thresholds (von Neumann entropy across bipartitions) is load-bearing for the entire classification task, yet no correlation, mutual-information, or ablation diagnostic is supplied to confirm that the input features carry nontrivial signal rather than spurious correlations.

Authors: We acknowledge that a direct diagnostic of the proxy relationship strengthens the foundation of the work. Although the classification results provide supporting evidence, we have added a new subsection to the Methods section containing Pearson correlation coefficients between the density-profile features and the target entanglement-entropy thresholds, together with an ablation study that systematically removes feature subsets. These analyses confirm that the inputs carry nontrivial signal. revision: yes
Referee: [Results] Results section: the assertion that a 4-qubit QCNN is optimal and that larger models are more sensitive to encoding lacks tabulated accuracy values, variance measures, or statistical significance tests against the classical CNN baselines, so the model-size and encoding conclusions cannot be assessed.

Authors: We agree that tabulated values and statistical tests are necessary to substantiate the model-size and encoding conclusions. In the revised Results section we have inserted a new table that reports accuracy, convergence metrics, and variance (with standard deviations across runs) for every QCNN size and encoding variant, together with p-values from two-sample t-tests against the classical CNN baselines. This table directly supports the optimality of the 4-qubit architecture and the greater encoding sensitivity of larger models. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML comparison with externally verifiable results

full rationale

The paper reports numerical experiments training QCNNs and CNNs to classify entanglement thresholds from fermion density profiles in the Thirring model. No equations, derivations, or fitted parameters are presented whose outputs reduce by construction to the inputs (e.g., no self-definitional ratios or predictions forced by subset fits). Reported accuracies, convergence, and variance are direct experimental outcomes on held-out data and can be reproduced independently. The proxy assumption (density profiles as entanglement indicators) is an empirical hypothesis evaluated by classification performance rather than a definitional tautology. No load-bearing self-citations or uniqueness theorems appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions of quantum machine learning and lattice field theory rather than new postulates. No free parameters or invented entities are introduced beyond conventional neural-network hyperparameters.

axioms (1)

domain assumption Fermion density profiles contain sufficient information to classify entanglement levels when framed as a multi-threshold classification task
Invoked when the authors state that density profiles can serve as proxies for entanglement.

pith-pipeline@v0.9.0 · 5475 in / 1320 out tokens · 52550 ms · 2026-05-10T19:48:47.718725+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 21 canonical work pages · 1 internal anchor

[1]

W. Guan, G. Perdue, A. Pesah, M. Schuld, K. Terashi, S. Vallecorsa, and J.-R. Vlimant, Mach. Learn. Sci. Tech. 2, 011003 (2021), arXiv:2005.08582 [quant-ph]

work page arXiv 2021
[2]

Classification with Quantum Neural Networks on Near Term Processors

E. Farhi and H. Neven, arXiv preprint arXiv:1802.06002 (2018), arXiv:1802.06002

work page Pith review arXiv 2018
[3]

Mitarai, M

K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, Phys- ical Review A98, 10.1103/physreva.98.032309 (2018)

work page doi:10.1103/physreva.98.032309 2018
[4]

Parameterized quantum circuits as machine learning models,

M. Benedetti, E. Lloyd, S. Sack, and M. Fioren- tini, Quantum Sci. Technol.4, 043001 (2019), arXiv:1906.07682 [quant-ph]

work page arXiv 2019
[5]

Schuld, A

M. Schuld, A. Bocharov, K. M. Svore, and N. Wiebe, Phys. Rev. A101, 032308 (2020)

2020
[6]

Elhag, T

H. Elhag, T. Hartung, K. Jansen, L. Nagano, G. M. Pirina, and A. D. Tucci, Quantum convolutional neural networks for jet images classification (2025), arXiv:2408.08701 [quant-ph]

work page arXiv 2025
[7]

Felser, M

T. Felser, M. Trenti, L. Sestini, A. Gianelle, D. Zu- liani, D. Lucchesi, and S. Montangero, Quantum-inspired machine learning on high-energy physics data (2021), arXiv:2004.13747 [stat.ML]

work page arXiv 2021
[8]

J. Y. Araz and M. Spannowsky, Journal of High Energy Physics2021, 10.1007/jhep08(2021)112 (2021)

work page doi:10.1007/jhep08(2021)112 2021
[9]

J. Y. Araz and M. Spannowsky, Physical Review A106, 10.1103/physreva.106.062423 (2022)

work page doi:10.1103/physreva.106.062423 2022
[10]

Gianelle, P

A. Gianelle, P. Koppenburg, D. Lucchesi, D. Nico- tra, E. Rodrigues, L. Sestini, J. de Vries, and D. Zuliani, Journal of High Energy Physics2022, 10.1007/jhep08(2022)014 (2022)

work page doi:10.1007/jhep08(2022)014 2022
[11]

Belis, P

V. Belis, P. Odagiu, and T. K. Aarrestad, Reviews in Physics12, 100091 (2024)

2024
[12]

Y. Chai, Y. Guo, and S. K¨ uhn, Towards Quantum Simu- lation of Meson Scattering in aZ 2 Lattice Gauge Theory (2025), arXiv:2505.21240 [quant-ph]

work page arXiv 2025
[13]

Y. Chai, J. Gibbs, V. R. Pascuzzi, Z. Holmes, S. K¨ uhn, F. Tacchino, and I. Tavernelli, Resource-efficient simula- tions of particle scattering on a digital quantum computer (2025), arXiv:2507.17832 [quant-ph]

work page arXiv 2025
[14]

Y. Chai, A. Crippa, K. Jansen, S. K¨ uhn, V. R. Pascuzzi, F. Tacchino, and I. Tavernelli, Quantum9, 1638 (2025)

2025
[15]

Davoudi, C.-C

Z. Davoudi, C.-C. Hsieh, and S. V. Kadam, Quantum8, 1520 (2024)

2024
[16]

Davoudi, C.-C

Z. Davoudi, C.-C. Hsieh, and S. V. Kadam, Quantum computation of hadron scattering in a lattice gauge the- ory (2025), arXiv:2505.20408 [quant-ph]

work page arXiv 2025
[17]

R. C. Farrell, M. Illa, A. N. Ciavarella, and M. J. Savage, Phys. Rev. D109, 114510 (2024)

2024
[18]

R. C. Farrell, N. A. Zemlevskiy, M. Illa, and J. Preskill, Digital quantum simulations of scattering in quantum field theories using w states (2025), arXiv:2505.03111 [quant-ph]

work page arXiv 2025
[19]

N. A. Zemlevskiy, Physical Review D112, 10.1103/qr72- 51v1 (2025)

work page doi:10.1103/qr72- 2025
[20]

N. A. Zemlevskiy, Exclusive scattering channels from entanglement structure in real-time simulations (2026), arXiv:2603.15621 [quant-ph]

work page arXiv 2026
[21]

Schuhmacher, G.-X

J. Schuhmacher, G.-X. Su, J. J. Osborne, A. Gandon, J. C. Halimeh, and I. Tavernelli, Observation of hadron scattering in a lattice gauge theory on a quantum com- puter (2025), arXiv:2505.20387 [quant-ph]

work page arXiv 2025
[22]

I. Cong, S. Choi, and M. D. Lukin, Nature Physics (2019) https://doi.org/10.1038/s41567-019-0648-8 (2019)

work page doi:10.1038/s41567-019-0648-8 2019
[23]

T. Hur, L. Kim, and D. K. Park, Quantum Machine Intel- ligence 4, 3 (2022) https://doi.org/10.1007/s42484-021- 00061-x (2021)

work page doi:10.1007/s42484-021- 2022
[24]

S. Y.-C. Chen, T.-C. Wei, C. Zhang, H. Yu, and S. Yoo, Phys. Rev. Res.4, 013231 (2022), arXiv:2012.12177 [cs.LG]

work page arXiv 2022
[25]

W. E. Thirring, Annals of Physics3, 91 (1958)

1958
[26]

Kogut and L

J. Kogut and L. Susskind, Phys. Rev. D11, 395 (1975)

1975
[27]

Susskind, Phys

L. Susskind, Phys. Rev. D16, 3031 (1977)

1977
[28]

M. C. Ba˜ nuls, K. Cichy, Y.-J. Kao, C.-J. D. Lin, Y.-P. Lin, and D. T.-L. Tan, Phys. Rev. D100, 094504 (2019)

2019
[29]

Rigobello, S

M. Rigobello, S. Notarnicola, G. Magnifico, and S. Mon- tangero, Phys. Rev. D104, 114501 (2021)

2021
[30]

Thanasilp, S

S. Thanasilp, S. Wang, N. A. Nghiem, P. J. Coles, and M. Cerezo, arXiv e-prints , arXiv:2110.14753 (2021), arXiv:2110.14753 [quant-ph]

work page arXiv 2021
[31]

D. P. Kingma and J. Ba, arXiv e-prints , arXiv:1412.6980 (2014), arXiv:1412.6980 [cs.LG]

work page internal anchor Pith review Pith/arXiv arXiv 2014