{"id":"7cf0e3d5-9fa4-4d07-aab6-18d34ee24f2a","arxiv_id":"2501.07986","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper claims a quantum graph Hamiltonian neural network learns graphs with 99.8% cosine similarity, but the circuit decomposition is invalid and the evaluation is circular.","lead":"This paper introduces a quantum graph Hamiltonian neural network (QGHNN) that maps graphs to Hamiltonians and trains quantum circuits with gradient descent. The core circuit derivation contains a mathematical error, and the reported experimental success is weakened by an evaluation that uses the target graph as the training loss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (10) factorizes exp(-iθH_c/ℏ) into a product of exponentials despite non-commuting Pauli terms, so the circuit in Eq. (12) does not implement the intended unitary. The QGHL/QGHNN construction therefore rests on an invalid derivation, and the numerical benchmarks do not rescue it.","rationale":"The reader's weakest-assumption identification is correct, and it is the decisive defect. Equations (7)-(12) are the derivation of the entire QGHL circuit, and the factorization in Eq. (10) is exact only when all exponentiated operators commute. Here the local σ_x, σ_y, and σ_z⊗σ_z terms on shared qubits do not commute, so the equality is false. This is an internal mathematical error, not a dispute with external consensus. The final gate sequence in Eq. (12) therefore does not implement the intended Hamiltonian evolution. The PennyLane experiments are statevector simulations with no described noise model, and the loss function is built from the target graph's Hamiltonian, so the reported accuracy cannot validate the claimed Hamiltonian-to-graph mapping. Because the core construction is invalid, rejection is warranted. My read does not change the reader's verdict, so I recommend keeping the existing REJECT outcome.","tokens_in":16877,"tokens_out":6359,"duration_ms":65704,"concrete_test":"Independently verify Eq. (10) numerically: take N=2, ℏ=1, θ=0.1, and build H_c from Eq. (8). Compute U_exact = exp(-iθH_c) as a matrix exponential, and compute P = exp(-iθ/ℏΣσ_y) · exp(-iθ/ℏΣI/2) · exp(iπθ/4ℏΣσ_z⊗σ_z) · exp(-iθ/ℏΣσ_x). Evaluate the Frobenius norm ‖U_exact - P‖_F. If the factorization were exact, the norm would be at machine precision; instead it will be of order θ² because of the non-vanishing commutators. Repeating at θ=0.01 should show the discrepancy scaling quadratically in θ. This directly disproves the equality claimed in Eq. (10).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (10) is the load-bearing step that connects the graph-derived Hamiltonian H_c in Eq. (8) to the implementable gate sequence in Eq. (12). It asserts an exact equality of the form exp(-iθ/ℏ [Σσ_y + Σ(I/2 - π/4 σ_z⊗σ_z) + Σσ_x]) = exp(-iθ/ℏΣσ_y) · exp(-iθ/ℏΣI/2) · exp(iπθ/4ℏΣσ_z⊗σ_z) · exp(-iθ/ℏΣσ_x). This is valid only if the non-identity operators in the exponent commute. They do not: for any qubit n, [σ_y^n, σ_x^n] = 2iσ_z^n, and both σ_y^n and σ_x^n fail to commute with the σ_z^n⊗σ_z^{n+1} term. The correct expansion contains commutator corrections of order θ², which Eq. (12) omits. At best Eq. (12) is a first-order Trotter approximation, not the exact unitary evolution generated by H_c. Since the claimed mapping between graphs and Hamiltonians and the derivation of the QGHL circuit both depend on Eq. (10), the central quantum construction is invalid as stated. The numerical results, obtained by training against the same target Hamiltonian used in the loss, cannot repair this derivation failure; they only show that some parameterized circuit can fit a known H_m.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a quantum graph Hamiltonian neural network (QGHNN) built on a quantum graph Hamiltonian learning method (QGHL). QGHL maps a graph G to a Hamiltonian H_m of a topological quantum system, constructs a parameterized quantum circuit from a circuit Hamiltonian H_c by factorizing exp(-iθH_c/ℏ) into rotations and CNOT gates, and then trains the circuit parameters by minimizing the expectation value of H_m. The authors report experiments on the PennyLane platform for 4-, 5-, and 6-qubit graphs and claim that QGHNN outperforms VQE, QAOA, and QNN on MSE, cosine similarity, Frobenius norm, and correlation coefficient, with the best MSE of 0.004 and cosine similarity of 99.8%.","tokens_in":17210,"tokens_out":3443,"duration_ms":37639,"significance":"If the construction were valid, the paper would offer an explicit quantum circuit ansatz for representing graph structure through Hamiltonian learning, with potential relevance to quantum machine learning on NISQ devices. The paper also provides concrete comparison tables and convergence plots, which is useful presentation material. However, the central derivation connecting H_c to the implementable circuit is mathematically incorrect, and the experimental evaluation is circular because the target graph is used both to construct the loss Hamiltonian and to score the learned graph. These are load-bearing issues for the paper's main claims, so the current manuscript is not publishable in its present form.","major_comments":[{"comment":"Equation (10) asserts an exact equality between exp(-iθH_c/ℏ) and a product of exponentials of the individual terms in H_c from Eq. (8). This equality is mathematically false because the operators in H_c do not commute: for each qubit n, [σ_y^n, σ_x^n] = 2iσ_z^n, and the single-qubit σ_y and σ_x terms do not commute with the neighboring σ_z^n⊗σ_z^{n+1} terms. The product of exponentials differs from exp(-iθH_c/ℏ) by commutator corrections of order θ². Therefore the circuit in Eq. (12) does not implement the unitary evolution generated by H_c; at best it is a first-order Trotterization. Since the claimed graph-to-Hamiltonian mapping and the QGHL circuit both rely on Eq. (10), the central quantum construction is invalid as stated.","section":"§III-C, Eq. (10)"},{"comment":"The experimental evaluation is circular. The loss function in Eq. (13) is the expectation value of H_m, which is constructed from the target graph G_t via Eq. (6). The same target graph's adjacency matrix is used as the input Hamiltonian, and the learned graph G* is extracted from the optimized quantum state and then compared against that same target graph. There is no training/test split, no held-out graph, and no task in which the model predicts an unseen target. The reported metrics in Tables III–V therefore quantify fitting to the training label, not graph learning or generalization.","section":"§IV-A and §V-A, Eqs. (6) and (13)"},{"comment":"The claimed noise robustness is not supported by the experiments. No noise model, shot count, or hardware backend is described; the tables report single aggregate values without error bars or repeated runs. The abstract and Section V conclude that QGHNN 'reduces the impact of quantum noise' and has 'high robustness ability,' but no noisy simulation or statistical analysis is presented. Additionally, the comparisons to VQE, QAOA, and QNN lack controlled hyperparameter settings and repeated trials, so the statement that QGHNN 'outperforms all assessment metrics' is not substantiated as a fair comparison.","section":"§V-A, Tables III–V"},{"comment":"The gradient update step is written incorrectly as ∂Loss(θ_j)/∂θ_j = [∂Loss(θ_j+Δ_j) − ∂Loss(θ_j−Δ_j)]/(2Δ_j), which mixes derivatives with function values. The correct central difference should use Loss(θ_j+Δ_j) and Loss(θ_j−Δ_j), not ∂Loss at those points. Also, the stopping criterion Gap = |H_t − H_m| in §III-A never defines how H_t is computed from |ψ_t(θ)⟩, so the proposed learning procedure is not fully specified.","section":"§IV-B, Algorithm 1"}],"minor_comments":[{"comment":"There are several typos and inconsistent labels: 'QVE' appears for VQE in Tables VI–VIII and in the text, and the rotation angle notation alternates between R_z(2ℏθ/π) and R_z(2ℏ/π θ). These should be unified.","section":"General notation"},{"comment":"The procedure for extracting the learned graph G* from the final quantum state |ψ_out⟩ is not described: it is unclear how measured probabilities are converted back into an adjacency matrix and how normalization by Eq. (15) is inverted.","section":"§V-A, Fig. 5"},{"comment":"The comparison criteria in Tables IX and X, such as 'Noise Resistant' and 'Implementability on Quantum Computers,' are not defined operationally, so the qualitative superiority claims cannot be assessed.","section":"§V-C, Tables IX–X"},{"comment":"The abstract claims the method 'shows high robustness ability' and 'can reduce the impact of quantum noise,' but no noisy experiment is reported; the claim should be limited to what the evidence supports.","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"The central derivation error and the circular experimental design are fundamental to the paper's contribution. The Trotterization issue could in principle be repaired by reframing the circuit as an approximate Trotterized evolution, but the current manuscript does not do so, and the empirical claims would still require a non-circular benchmark with noise and error bars. I do not see a limited revision that would make the present claims valid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper proposes a quantum graph Hamiltonian neural network (QGHNN) that maps a graph's adjacency matrix to a parameterized Ising Hamiltonian, trains a small variational circuit on it, and reads back a learned graph. The headline numbers (MSE 0.004, cosine similarity 99.8% on 4-qubit graphs) are plausible for a fitting task, but the central construction has a load-bearing mathematical error.\n\nThe trouble is Eq. (10). The paper factorizes exp(-iθH_c/ℏ) into a product of exponentials of the individual Pauli sums in H_c. Those sums do not commute: σ_y and σ_x anticommute on the same qubit, and neither commutes with the σ_z⊗σ_z interaction. So the circuit in Eq. (12) is not the exact unitary generated by H_c; at best it is a first-order Trotter approximation. That invalidates the claimed QGHL derivation, and the rest of the paper builds on it.\n\nThe paper does have strengths. It is clearly organized, gives a concrete circuit ansatz, includes pseudocode, and the experiment is transparent enough to reproduce in PennyLane. The mapping from adjacency matrix to Hamiltonian is standard in QAOA and quantum graph networks, but the paper cites the relevant literature. The comparison to VQE, QAOA, and QNN is reasonable, even if the baselines are not deeply tuned.\n\nThe other soft spots are evaluation rather than math. The loss is the expectation value of H_m, which is built from the target graph, and the learned graph is scored against that same target graph. That is a memorization or reconstruction task, not a test of generalization; there are no held-out graphs or error bars. The noise-resilience claim is unsupported because the PennyLane experiments are noiseless simulations. The self-assessment tables (IX and X) use entries like 'Excellent' and 'Completely,' which are not evidence.\n\nBottom line: the paper is clearly written but the central derivation fails. A serious referee would catch the non-commutation issue immediately. The numerics don't rescue it, because they only show that a parameterized circuit can fit a known Hamiltonian. I would not cite it, but it could serve as a useful negative example in a reading group about Trotterization and circular evaluation.","headline":"A clearly written graph-learning ansatz that is undone by a load-bearing non-commutation error in Eq. (10), which the numeric results cannot repair.","tokens_in":17731,"tokens_out":4834,"would_cite":false,"duration_ms":48927,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Encoding a graph as a spin-lattice Hamiltonian and training the derived circuit recovers the graph with up to 99.8% cosine similarity, outperforming VQE, QAOA, and QNN baselines.","keywords":["quantum graph Hamiltonian learning","quantum neural network","graph representation learning","Hamiltonian learning","parameterized quantum circuit","NISQ devices","variational quantum algorithm","adjacency matrix encoding"],"falsifier":"Take the two-qubit case, fix $\\theta$ to a generic value, compute $\\exp(-iH_c\\theta/\\hbar)$ directly from Eq. (8), and compute the product $R_y(\\hbar\\theta)\\cdot \\mathrm{CNOT}\\cdot R_z(2\\hbar\\theta/\\pi)\\cdot \\mathrm{CNOT}\\cdot R_x(\\hbar\\theta)$ from Eq. (12); any difference beyond machine precision means Eq. (10) does not hold and the circuit does not implement the intended Hamiltonian evolution.","tokens_in":16645,"feed_emoji":"⚛️","tokens_out":10477,"duration_ms":89259,"temperature":0.7,"pith_summary":"This paper proposes a way to represent a classical graph as the Hamiltonian of a topological (lattice) quantum system, with the adjacency matrix appearing as coupling coefficients of Pauli operators. The paper then builds a quantum graph Hamiltonian neural network (QGHNN) whose layered circuit is derived by decomposing a time-evolution operator into rotation and CNOT gates, and trains the circuit parameters by gradient descent on the expectation value of the graph-derived Hamiltonian. Experiments on graphs with four to six nodes report that the trained circuit recovers the target adjacency matrix with mean squared error as low as 0.004 and cosine similarity as high as 99.8%, and that these numbers beat the VQE, QAOA, and QNN baselines under the same noise conditions. The intended significance is that graph learning can be carried out on near-term quantum devices through Hamiltonian learning, providing a route toward quantum knowledge graphs and recommendation systems.","feed_headline":"Quantum circuit learns graph structure with 99.8% cosine similarity","feed_subtitle":"By encoding adjacency as a spin Hamiltonian, a trained parameterized circuit outperforms VQE, QAOA, and QNN baselines.","key_machinery":"The load-bearing object is the pair consisting of the graph-derived Hamiltonian $H_m$ and the circuit ansatz $U(\\theta)$. $H_m$ encodes the graph through the adjacency matrix $A_{ij}$ in front of Pauli string operators $\\sigma_x\\otimes\\sigma_x$, $\\sigma_y\\otimes\\sigma_y$, $\\sigma_z\\otimes\\sigma_z$, so that the graph structure becomes the coupling constants of a spin-lattice system. The circuit $U(\\theta)$ is obtained by decomposing $\\exp(-iH_c\\theta/\\hbar)$, with $H_c = \\sum_n \\sigma_n^y + \\sum_n (\\frac{1}{2}I - \\frac{\\pi}{4} \\sigma_n^z\\otimes\\sigma_{n+1}^z) + \\sum_n \\sigma_n^x$, into the product $R_y(\\hbar\\theta)\\cdot \\mathrm{CNOT}\\cdot R_z(2\\hbar\\theta/\\pi)\\cdot \\mathrm{CNOT}\\cdot R_x(\\hbar\\theta)$; this decomposition is what allows the method to turn Hamiltonian learning into a parameterized circuit whose angles are trained by gradient descent. The graph information is injected twice: into the target observable $H_m$ used to define the loss, and into the structure of the layered circuit that produces the intermediate state.","core_discovery":"The central claim, stated on the paper's own terms, is that QGHL establishes a mapping link between a graph $G=(V,A)$ and the Hamiltonian $H_m = \\sum_{i,j} A_{ij}(J_x \\sigma_i^x \\sigma_j^x + J_y \\sigma_i^y \\sigma_j^y + J_z \\sigma_i^z \\sigma_j^z)$ of a topological quantum system, and that QGHNN learns this Hamiltonian by training the parameterized circuit $U(\\theta)=R_y(\\hbar\\theta)\\cdot \\mathrm{CNOT}\\cdot R_z(2\\hbar\\theta/\\pi)\\cdot \\mathrm{CNOT}\\cdot R_x(\\hbar\\theta)$ to minimize the loss $\\mathrm{Loss}(\\theta)=\\langle\\psi_t(\\theta)|H_m|\\psi_t(\\theta)\\rangle$. The paper reports that after gradient-descent training, the Hamiltonian distribution of the learned graph approaches the target graph's distribution, with the best result on a four-node graph (MSE $0.004$, cosine similarity $99.8\\%$) and slightly lower but still leading accuracy on five- and six-node graphs, which it attributes to noise resilience inherited from the topological system.","pith_inferences":["Because the claimed circuit derivation depends on the validity of Eq. (10), a direct check of that identity would show whether QGHNN is truly simulating the stated Hamiltonian or simply learning as a generic variational ansatz; the reported numbers would then need reinterpretation.","The graph-to-Hamiltonian mapping is independent of the particular circuit ansatz, so the same encoding could in principle be paired with a different, exactly implementable ansatz if the current decomposition fails.","A natural next experiment is to test the same training scheme on weighted or directed graphs, where the adjacency entries carry more information than the unweighted examples used here.","The method's practical value on near-term hardware will depend on how the circuit depth and measurement cost scale with the number of nodes, since the experiments here are limited to six qubits."],"forward_implications":["For any graph with a known adjacency matrix, QGHL gives a concrete recipe to transcribe the graph into a spin-lattice Hamiltonian and then into a trainable quantum circuit.","The reported loss curves converge to about $-1.0$ within 200–300 steps for all three graph sizes, indicating stable training on the tested examples.","QGHNN uses fewer trainable parameters than the VQE, QAOA, and QNN baselines while achieving lower error on all four evaluation metrics.","The accuracy gaps over baselines persist as the number of qubits increases from 4 to 6, consistent with the paper's claim of noise-resilient graph learning."],"supporting_citations":[{"why":"Supplies the quantum Hamiltonian learning paradigm that QGHL extends to graphs.","marker":"[16]"},{"why":"Provides the parameterized quantum circuit and loss-based training recipe on which QGHNN is based.","marker":"[18]"},{"why":"Underpins the claim that topological (lattice) quantum systems are robust to quantum noise.","marker":"[22]"},{"why":"Defines the QNN baseline and the gradient-descent parameter update rule used by QGHNN.","marker":"[11]"},{"why":"Supplies the VQE baseline that QGHNN is compared against in the experiments.","marker":"[34]"},{"why":"Supplies the QAOA baseline and its graph-optimization context.","marker":"[35]"}],"fun_headline_variants":["Quantum graph Hamiltonian net outperforms VQE and QAOA","Noise-resilient quantum graph learner hits 99.8% cosine similarity","Topological graph encoding beats standard quantum models","MSE 0.004 and 99.8% similarity: quantum graph Hamiltonian learns","QGHNN: robust graph learning via quantum Hamiltonian encoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the unitary evolution with the circuit Hamiltonian can be split into separate rotations for each Pauli term; the terms involved in Eq. (8) do not commute, so the split is mathematically unjustified.","fun_headline_variants_meta":{"raw":{"variants":["Quantum graph Hamiltonian net outperforms VQE and QAOA","Noise-resilient quantum graph learner hits 99.8% cosine similarity","Topological graph encoding beats standard quantum models","MSE 0.004 and 99.8% similarity: quantum graph Hamiltonian learns","QGHNN: robust graph learning via quantum Hamiltonian encoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3416,"prompt_tokens":1054,"completion_tokens":2362,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":2272}},"tokens_in":670,"tokens_out":2362,"duration_ms":16019,"temperature":1.0,"reasoning_tokens":2272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:29:27.005906+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the two-qubit case, fix $\\theta$ to a generic value, compute $\\exp(-iH_c\\theta/\\hbar)$ directly from Eq. (8), and compute the product $R_y(\\hbar\\theta)\\cdot \\mathrm{CNOT}\\cdot R_z(2\\hbar\\theta/\\pi)\\cdot \\mathrm{CNOT}\\cdot R_x(\\hbar\\theta)$ from Eq. (12); any difference beyond machine precision means Eq. (10) does not hold and the circuit does not implement the intended Hamiltonian evolution.","supporting_citations":[{"cited_title":"Quantum hamiltonian learning using imperfect quantum resources,","cited_arxiv_id":null,"evidence_quote":"Supplies the quantum Hamiltonian learning paradigm that QGHL extends to graphs."},{"cited_title":"Parameterized hamil- tonian learning with quantum circuit,","cited_arxiv_id":null,"evidence_quote":"Provides the parameterized quantum circuit and loss-based training recipe on which QGHNN is based."},{"cited_title":"Learning quantum systems,","cited_arxiv_id":null,"evidence_quote":"Underpins the claim that topological (lattice) quantum systems are robust to quantum noise."},{"cited_title":"Qksan: A quantum kernel self-attention network,","cited_arxiv_id":null,"evidence_quote":"Defines the QNN baseline and the gradient-descent parameter update rule used by QGHNN."},{"cited_title":"Qaoa for max-cut requires hundreds of qubits for quantum speed-up,","cited_arxiv_id":null,"evidence_quote":"Supplies the QAOA baseline and its graph-optimization context."}],"review_version":1}