{"id":"78314167-baff-4295-adf2-2d0e7a10003a","arxiv_id":"2411.15315","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Lorentz-equivariant quantum graph neural network matches the classical LorentzNet on quark-gluon jet discrimination in noiseless simulations.","lead":"This paper builds a quantum version of LorentzNet, a classical symmetry-aware neural network, and tests it on telling quark jets from gluon jets. It reports that the quantum model matches the classical model's accuracy in noiseless simulations, a step toward quantum machine learning for particle physics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim of performance on par with or slightly better than LorentzNet is not statistically supported: no error bars, repeated seeds, or numerical test metrics are reported for the central comparison.","rationale":"The reader's conditional verdict already flags missing error bars and an overclaim about real-world viability, and my stress-test reinforces that the central comparison is not quantitatively established. The reader's named weakest assumption—unspecified measurement/readout of the quantum modules—is genuine and affects reproducibility, but it is secondary to the statistical indefensibility of the performance claim: even a perfectly equivariant architecture would fail to demonstrate 'comparable or slightly better' performance from a single unquantified run. The public code is a positive reproducibility signal and makes the recommended multi-seed test feasible. Since the reader's recommendation already calls for conditional acceptance with those concerns, no verdict change is needed.","tokens_in":8344,"tokens_out":4540,"duration_ms":47023,"concrete_test":"Using the released repository (https://github.com/ML4SCI/QMLHEP/tree/main/Lie_EQGNN_for_HEP_Jogi_Suda_Neto), train the full-quantum Lie-EQGNN and LorentzNet with identical data splits, optimizer, and hyperparameters for K=10 random seeds, and report the mean and standard deviation of validation/test accuracy and AUC for each model. If the difference between the two mean metrics is smaller than the pooled standard error, the 'comparable or slightly better' claim is not statistically supported and should be downgraded; if the separation exceeds several standard errors, the concern is resolved. As a secondary check, record the observable and post-processing used to read out phi_e, phi_x, phi_h, and phi_m to confirm that each circuit output is a scalar expectation value in a fixed Lorentz frame.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in Section 6, is that 'the performance of Lie-EQNNs is comparable to or slightly better than that of their classical counterparts, LorentzNet.' Section 5 reports only training/validation accuracy and loss curves from what appears to be a single noiseless run. No final test-set accuracy, AUC, standard deviation, or repeated-seed statistics are given anywhere in the text or tables. Because the variational circuits are randomly initialized and the full-quantum model has 592 trainable parameters versus LorentzNet's 1088, the observed separation in the learning curves could be attributable to initialization luck, optimization details, or early stopping rather than to the quantum architecture itself. Without a distribution over runs, the 'slightly better' component of the claim is not meaningful. The reader's identified weakest assumption—that the quantum modules must output deterministic scalars—is a real but secondary reproducibility gap: Lorentz equivariance would be preserved if the circuits output expectation values of a fixed observable, and the public code can settle that question. The load-bearing failure is that the empirical comparison at the heart of the paper is not quantified, so the central claim cannot be assessed from the manuscript alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Lie-EQGNN, a quantum graph neural network for quark-gluon jet tagging, obtained by replacing the classical multilayer perceptrons (phi_e, phi_x, phi_h, phi_m) in LorentzNet with variational quantum circuits. It reports training and validation accuracy/loss curves for several partially and fully quantized variants and claims that the performance is comparable to or slightly better than the classical LorentzNet in noiseless simulation. The paper includes a public code repository and frames the work as a step toward symmetry-preserving quantum machine learning for LHC physics.","tokens_in":8617,"tokens_out":3945,"duration_ms":36688,"significance":"If substantiated, the result would be a useful existence proof that variational circuits can substitute for classical MLPs inside an equivariant jet-tagging architecture without degrading accuracy, and the public code release is a concrete strength. However, the empirical claim currently rests on unquantified single-run curves, and the quantum readout is underspecified, so the significance cannot be fully assessed from the manuscript alone. The equivariance construction itself is a straightforward adaptation of LorentzNet and is plausible.","major_comments":[{"comment":"The central conclusion in Section 6 ('comparable to or slightly better than LorentzNet') is not supported by the reported experiments. Only training and validation accuracy and loss curves from what appears to be a single noiseless run are shown; no test-set accuracy, AUC, standard deviation, or repeated-seed statistics are reported anywhere. Since the variational circuits are randomly initialized and the full-quantum model has 592 trainable parameters versus LorentzNet's 1088, the observed separation in the learning curves could be attributable to initialization luck or optimization details rather than to the quantum architecture. The authors should provide final test metrics with multiple seeds and error bars (or at least a table of mean plus standard deviation) before the claim of parity can be assessed.","section":"Section 5, Figures 4 and 5, Table 1"},{"comment":"For the Lorentz-equivariant update to hold, the quantum circuits replacing phi_e, phi_x, phi_h, and phi_m must behave as deterministic functions returning real scalars (or vectors of scalars), because the coordinate update in Eq. (2) multiplies x_j by the scalar phi_x(m_ij). The manuscript never specifies the measurement observable, the readout strategy, or how the 6-qubit circuit output is post-processed into the scalar used in the update and into the multi-component outputs of phi_e and phi_h. Without this specification, the equivariance claim is not fully verifiable and the experiments are not reproducible from the text; the paper should state the measurement and post-processing explicitly.","section":"Section 4.4, Figure 2, Equations (2)-(4)"}],"minor_comments":[{"comment":"The summation indices in Eqs. (2) and (3) are inconsistent: Eq. (2) sums over i while using x_j on the right-hand side, and Eq. (3) contains an undefined summation index. The intended update over neighboring particles should be written with a clear index (for example, summing over j in the neighborhood of i) to remove ambiguity.","section":"Section 4.4, Equations (2) and (3)"},{"comment":"The manuscript does not specify how raw particle features (pT, eta, phi, and scalar features such as particle ID and mass) are preprocessed before being encoded as RY rotation angles. Since RY angles are periodic and the feature scales vary widely, a normalization or scaling step is needed for the encoding to be meaningful; the authors should state the exact preprocessing used.","section":"Section 3 and Figure 2"},{"comment":"The parameter counts in Table 1 are reported without a description of what is counted (for example, whether bias terms are included and how the quantum circuit parameters are counted). This makes it difficult to interpret the comparison between the quantum and classical models.","section":"Section 4.4, Table 1"},{"comment":"There is a typographical error in the sentence 'the f holds the desirable property of equivariance'; it should read 'f holds the desirable property'.","section":"Section 2, paragraph 2"},{"comment":"The paper does not report any resource estimates such as the number of circuit executions, wall-clock time, or qubit counts used in the simulations, despite discussing quantum utility and the NISQ era in the introduction; adding these details would strengthen the practical claims.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The two major concerns are fixable within the scope of the manuscript: the authors should add repeated-seed test metrics (or soften the parity claim) and clearly specify the quantum circuit readout. The paper is quite short for a full journal article and reads more like a workshop contribution; the editor may wish to consider whether the expected length and depth of the journal require a more thorough empirical study, including baselines with matched parameter counts and resource estimates. The self-citation pattern in the references is not problematic in itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is a genuine novelty — first Lorentz-equivariant quantum GNN for jet tagging — and the paper is honest about its scope. But the central claim that Lie-EQGNN performs on par with or slightly better than LorentzNet is not supported by the evidence as presented: single run, learning curves only, no test metrics, no error bars, no repeated seeds. That's a load-bearing gap, not a stylistic one.\n\nWhat's new and good: substituting variational circuits for the MLP modules (phi_e, phi_x, phi_h, phi_m) in LorentzNet is a clean idea, and the ablation replacing each module one at a time is the right way to localize where quantum substitution helps or hurts. The equivariant update structure is inherited from LorentzNet, so the symmetry argument is plausible if the circuits output deterministic scalars (e.g., expectation values). Code is public, which makes the reproducibility signal real.\n\nSoft spots, in order of severity. First, the empirical comparison. Section 6 claims parity or slight improvement, but Figures 4 and 5 show only training/validation accuracy and loss from what appears to be one noiseless run. No final test-set accuracy, AUC, or standard deviation is reported anywhere. With random initialization and 592 trainable parameters in the full quantum model versus 1088 in LorentzNet, the observed separation could be initialization luck. This needs repeated seeds, error bars, and explicit test metrics. Second, the paper never specifies how the circuit output becomes the scalar used in the coordinate update. An expectation value of a fixed observable would preserve equivariance, but the reader shouldn't have to infer that from the code. Third, the dataset usage is a small subsample (12.5k jets, ≥10 constituents), which is fine for proof-of-principle but should be stated as such; the abstract's 'viable alternative' is stronger than the evidence warrants. Minor: Eq. (2) has an index typo (sum over i inside an update for i). Also, the paper mentions 'data efficient' but doesn't demonstrate it with a learning-curve analysis.\n\nOverall: this is a legitimate proof-of-principle for QML+HEP, and the code and ablation make it worth engaging with. It needs a serious referee, but the referee should push hard on the empirical section before acceptance. My recommendation: send to peer review, with the expectation of substantial revision.","headline":"A useful proof-of-principle for Lorentz-equivariant quantum GNNs, but the central parity claim rests on a single unquantified run and needs a proper empirical section before it can be endorsed.","tokens_in":9099,"tokens_out":1913,"would_cite":false,"duration_ms":17705,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Lorentz-equivariant quantum graph neural network matches or slightly beats the classical LorentzNet at quark-gluon jet discrimination in noiseless simulation.","keywords":["Lorentz equivariance","quantum graph neural networks","jet tagging","quark-gluon discrimination","variational quantum circuits","geometric deep learning","particle physics","invariant theory"],"falsifier":"Run the trained LEQB on a jet and on the same jet with every constituent four-momentum transformed by a fixed Lorentz boost; compare the updated coordinates $x_i^{l+1}$. If the two outputs are not related by the same boost, the claimed equivariance is false. This is a purely classical check: with fixed trained parameters, the variational circuits are deterministic scalar functions, so the check can be done in exact simulation.","tokens_in":8201,"feed_emoji":"⚛️","tokens_out":8593,"duration_ms":76791,"temperature":0.7,"pith_summary":"This paper tries to establish that a quantum graph neural network can respect the same Lorentz symmetry that makes the classical LorentzNet effective for jet tagging, without losing accuracy. The authors build a Lie-equivariant quantum GNN by replacing the four learned functions inside LorentzNet with small variational quantum circuits, creating the Lorentz-Equivariant Quantum Block. Their experiments on quark-gluon jet discrimination show the quantum model performing on par with or slightly better than the classical LorentzNet in noiseless simulation. The significance is practical: if the result holds, symmetry-preserving quantum models become a credible alternative for LHC data analysis and for resource-constrained quantum machine learning.","feed_headline":"Quantum GNN matches classical LorentzNet on quark-gluon jets","feed_subtitle":"A Lorentz-equivariant quantum network reaches the same jet-tagging accuracy with 592 parameters against 1088.","key_machinery":"The central object is the Lorentz-Equivariant Quantum Block (LEQB): a graph block identical in form to LorentzNet's update equations, but with the multilayer perceptrons replaced by six-qubit variational circuits consisting of a Hadamard layer, $R_Y$ angle encoding, and two trainable layers of entangling gates with parameterized $R_Y$ rotations. It carries the argument by keeping coordinate updates linear in the input coordinates with scalar coefficients, so that the required equivariance follows from the fact that the message depends only on Lorentz-invariant scalars. Invariant theory, the principle that equivariant maps can be built from scalar invariants, is the theoretical basis for treating the noncompact Lorentz group this way.","core_discovery":"The paper's central discovery is that the Lorentz-Equivariant Quantum Block (LEQB), a graph update in which the four learned functions $\\phi_e$, $\\phi_x$, $\\phi_h$, and $\\phi_m$ are variational quantum circuits, reproduces the performance of the classical LorentzNet for quark-gluon jet tagging. The coordinate update takes the form $x_i^{l+1}=x_i^l+c\\sum_j \\phi_x(m_{ij}^l)\\,x_j^l$, and the edge message $m_{ij}^l$ is computed only from Lorentz-invariant quantities: the squared Minkowski norm $\\|x_i-x_j\\|^2$ and the Minkowski inner product $\\langle x_i,x_j\\rangle$. Since the quantum modules are treated as scalar-valued functions of these invariants, the block remains equivariant under Lorentz transformations by construction. On the Pythia8 quark-gluon jet dataset (12,500 jets), the fully quantum model reaches accuracy comparable to or slightly better than LorentzNet, with 592 trainable parameters against LorentzNet's 1088, in idealized noiseless infinite-shot simulation.","pith_inferences":["The same quantum substitution should transfer to other Lie-equivariant GNNs whose updates multiply coordinates by scalar functions, for example E(n)-equivariant networks, because the scalar-output structure is all that is needed for equivariance.","A clean stress test would train the fully quantum model on jets transformed by random Lorentz boosts and check that test accuracy is unchanged; exact equivariance predicts no change in noiseless simulation.","Because the simulations assume infinite shots and zero noise, the parity with LorentzNet is a ceiling: real-hardware operation with finite shots and noise would need error mitigation to approach the same accuracy."],"forward_implications":["A Lorentz-equivariant quantum GNN can match the classical LorentzNet on quark-gluon jet discrimination, so symmetry-preserving quantum models do not necessarily sacrifice accuracy.","The fully quantum model does so with fewer trainable parameters (592 versus 1088), making it a candidate for resource-constrained quantum settings.","Because the message function uses only Lorentz-invariant inputs (Minkowski norms and inner products), the equivariance proof of the classical architecture carries over to the quantum modules whenever those modules output scalars.","The noiseless infinite-shot results establish the baseline performance that any real-hardware implementation of the LEQB would need to approach."],"supporting_citations":[{"why":"LorentzNet: the classical architecture whose four neural modules are replaced; supplies the update equations and the performance baseline.","marker":"[25]"},{"why":"The Pythia8 quark-gluon jet dataset used for training, validation, and testing (12,500-jet subset).","marker":"[36]"},{"why":"Prior invariant and equivariant classical and quantum GNN comparison; provides the equivariant coordinate-update construction the LEQB builds on.","marker":"[19]"},{"why":"Shows equivariant maps can be constructed from invariant scalars, the basis of the LEQB update strategy.","marker":"[41]"},{"why":"Machine-learning and invariant-theory framework used to handle noncompact Lie groups such as the Lorentz group.","marker":"[42]"}],"fun_headline_variants":["Quantum GNN matches classical jet tagging with half the parameters","Lorentz-equivariant quantum net ties LorentzNet on jet tagging","592-parameter quantum model equals 1088-parameter classical","Quantum graph net rivals classical on quark-gluon jets","Lie-equivariant quantum GNN matches classical accuracy on jets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The equivariance argument assumes each variational circuit returns one deterministic real number used to rescale a particle momentum, and the paper never states the measurement or post-processing step that produces that number.","fun_headline_variants_meta":{"raw":{"variants":["Quantum GNN matches classical jet tagging with half the parameters","Lorentz-equivariant quantum net ties LorentzNet on jet tagging","592-parameter quantum model equals 1088-parameter classical","Quantum graph net rivals classical on quark-gluon jets","Lie-equivariant quantum GNN matches classical accuracy on jets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1413,"prompt_tokens":883,"completion_tokens":530,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":446}},"tokens_in":499,"tokens_out":530,"duration_ms":4606,"temperature":1.0,"reasoning_tokens":446,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:26:08.351357+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained LEQB on a jet and on the same jet with every constituent four-momentum transformed by a fixed Lorentz boost; compare the updated coordinates $x_i^{l+1}$. If the two outputs are not related by the same boost, the claimed equivariance is false. This is a purely classical check: with fixed trained parameters, the variational circuits are deterministic scalar functions, so the check can be done in exact simulation.","supporting_citations":[{"cited_title":"Forestano et al","cited_arxiv_id":null,"evidence_quote":"Prior invariant and equivariant classical and quantum GNN comparison; provides the equivariant coordinate-update construction the LEQB builds on."}],"review_version":1}