{"id":"8f548d72-8cbc-4869-a422-d31a0fbfcb5e","arxiv_id":"2412.08207","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A proposed measurement-based version of quantum convolutional neural networks on square-lattice cluster states is trained on two small tasks, but the exact mapping from QCNN circuits to cluster states is invalid as written.","lead":"The paper proposes a quantum convolutional neural network that runs by measuring a pre-entangled cluster state instead of executing a deep quantum circuit, with the measurement angles trained on data. It claims faster and sometimes more accurate learning than circuit-based quantum and classical CNNs, but the central construction contains a basic mathematical error in the measurement basis.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (2) is false: measuring ⟨0|R_z(θ) on |L4⟩ gives |ψ⟩ up to a global phase, not U_H R_z(ξ)R_x(ζ)R_z(θ)|ψ⟩, so the U_ij/V_ij cluster gadgets and the claimed exact QCNN mapping collapse.","rationale":"I read the paper's stated central claim as the exact cluster-state solution to general QCNNs, with Eq. (2) as the load-bearing primitive. That primitive is algebraically false for the measurement bases as written: ⟨0|R_z(θ) is independent of θ up to a phase, so the line cluster cannot implement the tunable single-qubit rotations on which the U_ij and V_ij gadgets rely. This is not a matter of parameter-tuning or numerical accuracy; the mathematical identity stated in Eq. (2) does not hold. The subsequent square-lattice numerical models are independent and may be valid variational ansätze, but they do not rescue the exact mapping promised in the abstract. The reader's weakest-assumption analysis identifies exactly this flaw, and the resulting REJECT verdict is appropriate. I find no reason to adjust the verdict, and I do not see a more load-bearing concern than this one.","tokens_in":19091,"tokens_out":3328,"duration_ms":35800,"concrete_test":"Symbolically evaluate the left-hand side of Eq. (2) for arbitrary input |ψ⟩: compute (⟨0|R_z(θ_1) ⊗ ⟨0|R_z(θ_2) ⊗ ⟨0|R_z(θ_3) ⊗ I)|L4⟩. Because R_z(θ)|0⟩ is proportional to |0⟩, the CZ gates that depend on |1⟩ components are never activated, and the result is a constant phase times |ψ⟩_4, independent of θ_1, θ_2, θ_3. A numerical simulation on four qubits comparing the normalized output with U_H R_z(ξ)R_x(ζ)R_z(θ)|ψ⟩ for random parameters and random |ψ⟩ will confirm this immediately: the fidelity will not be 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the paper is an exact cluster-state implementation of general QCNNs, built on the four-qubit line cluster |L4⟩ in Eq. (1). The key identity is Eq. (2), which states that measuring qubits 1–3 of |L4⟩ in bases ⟨M1|=⟨0|R_z(θ_k), ⟨M2|=⟨0|R_z(ζ_k), and ⟨M3|=⟨0|R_z(ξ_k) produces the single-qubit rotation U_H R_z(ξ_k)R_x(ζ_k)R_z(θ_k)|ψ⟩. This identity fails: R_z(θ)|0⟩ = e^{-iθ/2}|0⟩, so ⟨0|R_z(θ) is θ-independent up to an overall phase. Consequently, the three projectors contain no angle information, and the post-measurement state of qubit 4 is simply |ψ⟩ (up to a global phase), not the claimed rotation. Since the U_ij and V_ij clusters are assembled from repeated |L4⟩ gadgets, as shown in Fig. 1(b)–(c) and Supplementary S1–S2, the invalidity of Eq. (2) invalidates the entire exact QCNN-to-cluster-state construction. The numerical examples are not affected in the same way, because they use square-lattice states with projectors ⟨0|R_y(α)R_z(β) in Eq. (7), which do depend on their parameters; however, those numerics demonstrate a different variational model, not the exact correspondence promised in the abstract and Section II. This is an internal algebraic error, not a disagreement with external consensus, and it directly undermines the paper's strongest claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a measurement-based quantum convolutional neural network (MBQCNN) in which a cluster state is prepared and trained by tuning local measurement bases rather than by applying deep gate sequences. The central theoretical claim is an \"exact cluster state solution to general QCNNs,\" built from |L4> line-cluster gadgets for single-qubit rotations and |E8> cluster gadgets for CNOT operations, assembled into U_ij and V_ij clusters. Numerical experiments are then reported for two tasks: identifying Haldane ground-state phases using a 2-by-5 square-lattice cluster state, and classifying the iris dataset using a 2-by-4 square-lattice cluster state. The numerical model uses parameterized projectors <0|R_y(alpha)R_z(beta), not the <0|R_z(theta) projectors of the exact construction.","tokens_in":19481,"tokens_out":4602,"duration_ms":46838,"significance":"If the exact mapping were correct, the paper would offer a practical route to executing deep QCNNs as static cluster-state preparations followed by local measurements, avoiding deep coherent control. The numerical results, with disjoint training and testing sets and parameter-matched comparison models, are a useful heuristic demonstration of a square-lattice variational model. However, the central identity in Eq. (2) is algebraically false, and the numerical sections do not implement the claimed exact construction. The paper therefore does not establish its advertised theoretical contribution; the numerical evidence is for a different model.","major_comments":[{"comment":"The identity in Eq. (2) is algebraically incorrect: since R_z(theta)|0> = e^{-i theta/2}|0>, the projector <0|R_z(theta) is proportional to <0| and carries no theta dependence. Measuring the first three qubits of |L4> in the stated bases leaves the fourth qubit in |psi> up to a global phase, not in U_H R_z(xi) R_x(zeta) R_z(theta)|psi>. Because the U_ij and V_ij clusters in Fig. 1(b)-(c) and Supplementary S1-S2 are composed of such |L4> gadgets, the claimed exact cluster-state implementation of a general QCNN collapses.","section":"Section II, Eq. (2)"},{"comment":"The numerical demonstrations replace the Section II construction with square-lattice cluster states measured in parameter-dependent bases <0|R_y(alpha_i)R_z(beta_i). These simulations therefore do not validate the exact mapping advertised in the Abstract and Section II; they validate a different variational model. The Introduction explicitly notes that the square-lattice states do not have a one-to-one correspondence with the QCNN circuit, but the paper still claims the numerical results support the exact scheme, which they do not.","section":"Section III, Eq. (7) and Figs. 2-4"},{"comment":"The Discussion states that the paper provides \"a strict cluster state that can realize the function of every detail of the circuit.\" This conclusion depends on the invalid Eq. (2) and is therefore unsupported. The paper's own caveat in Section II that the V_ij-cluster constraints are unnecessary in practice further distances the numerical results from the strict construction, but the decisive issue remains the invalid rotation gadget in Eq. (2).","section":"Discussion and Conclusion"}],"minor_comments":[{"comment":"The text refers to \"the ground states of Eq. (7)\" where the Hamiltonian is defined in Eq. (6); this appears to be a typo.","section":"Section III"},{"comment":"The state in Eq. (7) is written as |S13> although it was defined as |S13^(i)> earlier in the same section; the notation should be made consistent.","section":"Section III, Eq. (7)"},{"comment":"The abbreviation \"MBQCC\" appears once in the Introduction; it should be \"MBQCNN.\"","section":"Section I"},{"comment":"Ref. [18] is cited for the QCNN treatment of Haldane ground states, but the relevant prior work appears to be Cong et al., Ref. [27]; the citation should be checked and corrected.","section":"Section III"},{"comment":"In the definition of the CV CZ operator, the text reads \"q_i and q_i\"; the second index should be j.","section":"Appendix C"}],"recommendation":"reject","confidential_remarks":"The central theoretical identity of the paper is false, and the numerical results pertain to a different model than the one advertised. The paper could potentially be rewritten as a heuristic study of square-lattice measurement-based classifiers, but that would be a substantially different contribution. The citation error regarding Ref. [18] versus Ref. [27] for the Haldane QCNN should also be investigated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThe short version: the paper's headline claim—an exact cluster-state implementation of any QCNN—does not hold as written. Eq. (2) uses projectors ⟨0|R_z(θ), and since R_z(θ)|0⟩ = e^{-iθ/2}|0⟩, those projectors carry no angle information. Measuring qubits 1–3 of |L4⟩ leaves qubit 4 in |ψ⟩ up to a global phase, not U_H R_z(ξ)R_x(ζ)R_z(θ)|ψ⟩. The U_ij and V_ij clusters are assembled from this gadget, so the exact mapping collapses. The supplementary derivations inherit the same error.\n\nWhat is genuinely new is the numerical study: a variational model that trains local measurement angles on square-lattice cluster states for two learning tasks. The Haldane phase identification and iris classification results are plausible, and the authors are transparent that these square-lattice numerics do not have a one-to-one correspondence with the QCNN circuit. That is an honest limitation statement, and the convergence comparison to a QCNN circuit is a reasonable first step.\n\nThe soft spots, in proportion. First, the false Eq. (2) is load-bearing; it invalidates the paper's main theoretical contribution. Second, the numerics demonstrate a different model from the one promised, so they cannot rescue the exact-construction claim. Third, the CNN baseline is non-standard (zeroed kernel entry, random learning rates in [0.25,0.75]/[0.95,1.45]), and the zero-padding encoding of the iris vectors into four qubits is under-specified. These weaken the performance comparison but are not fatal to the variational idea.\n\nMy take: this is a salvageable paper, but not in its current form. The authors should either fix the rotation gadget with a parameter-dependent projector (e.g., R_y) and re-derive the exact mapping, or reframe the paper as a variational MBQCNN proposal and drop the exact-claim language. As it stands, I would not rely on the exact construction, and I would not cite it in that form.\n\nIf it lands on my desk, I'd send it to peer review—the numerical direction is worth checking—but I'd expect major revision or rejection depending on whether the construction can be repaired.\n\nBest,","headline":"The exact cluster-state construction for QCNNs is invalid as written because Eq. (2) is angle-independent, but the variational square-lattice numerics are a salvageable idea.","tokens_in":19991,"tokens_out":2443,"would_cite":false,"duration_ms":25131,"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":"This paper claims that a general QCNN can be implemented exactly by preparing a cluster state and training only local measurement bases, replacing deep coherent gate sequences.","keywords":["quantum convolutional neural network","measurement-based quantum computing","cluster states","quantum machine learning","Haldane phase classification","iris classification"],"falsifier":"Compute the left-hand side of Eq. (2) for a general input state $|\\psi\\rangle$: because $R_z(\\theta)|0\\rangle$ is proportional to $|0\\rangle$ for every $\\theta$, each projector $\\langle 0|R_z(\\theta)\\rangle$ is, up to an overall phase, just $\\langle 0|$, so the post-measurement state cannot depend on $\\theta$, $\\zeta$, or $\\xi$. Checking the output state numerically for two different angle triples would settle whether the claimed rotation identity holds; the calculation predicts no angle dependence, contradicting Eq. (2).","tokens_in":18841,"feed_emoji":"⚛️","tokens_out":12096,"duration_ms":117150,"temperature":0.7,"pith_summary":"The paper sets out to show that a quantum convolutional neural network (QCNN) need not be executed as a deep sequence of coherent gates. Its proposal, measurement-based QCNN (MBQCNN), is to prepare a cluster state whose wiring mirrors the QCNN, attach the data qubits to it, and train only the local measurement bases; the network output is read directly from measurement results. The paper claims this is an exact cluster-state solution for general QCNNs, and it supports the approach with numerical learning experiments on two tasks, one using quantum data (Haldane ground states) and one using classical data (the iris dataset). In those experiments square-lattice cluster states learn the tasks successfully, converge faster than a classical CNN, and in the iris case match or exceed the accuracy of a QCNN with the same number of parameters. If the construction holds, deep quantum learning becomes a static resource-preparation and measurement problem rather than a dynamical control problem.","feed_headline":"Measuring one cluster state can run a quantum convolutional network","feed_subtitle":"Training tunes only local measurement angles, so no deep gate-by-gate control is needed.","key_machinery":"The load-bearing object is the cluster-state gadget pair: the line cluster $|L_4\\rangle$ (four qubits chained, with the first as input and the fourth as output) for single-qubit rotations, and the T-shaped cluster $|E_8\\rangle$ (eight qubits, two inputs and two outputs) for a CNOT. These are the elementary logic blocks of the construction: composing them according to known two-qubit unitary decompositions yields the $U_{ij}$ and $V_{ij}$ clusters, and wiring those in the QCNN pattern produces the full resource state. Training is carried by the local projectors $\\langle 0|R_z(\\theta)\\rangle$ (and, in the square-lattice examples, $\\langle 0|R_y(\\alpha)R_z(\\beta)\\rangle$), whose angles are updated by gradient descent on the mean-square loss defined in Eq. (5).","core_discovery":"On the paper's own terms, the central discovery is that the convolutional and pooling layers of a QCNN can be rebuilt as fixed entangled sub-clusters of a larger cluster state. The elementary blocks are the four-qubit line cluster $|L_4\\rangle$, which is claimed to perform a single-qubit rotation when its first three qubits are measured in bases $\\langle 0|R_z(\\theta)\\rangle$, and the eight-qubit T-shaped cluster $|E_8\\rangle$, which is claimed to perform a CNOT when most of its qubits are measured in $\\langle 0|$. Standard decompositions of two-qubit unitaries into single-qubit rotations and CNOTs let these blocks be wired into a $U_{ij}$-cluster for each convolutional gate and a $V_{ij}$-cluster for each pooling control-rotation; connecting those clusters in the same zig-zag pattern as the QCNN circuit gives a total state whose measurement angles are the trainable parameters. The numerical demonstrations then replace the exact wiring with experimentally convenient square-lattice cluster states, which are not circuit-equivalent but still reproduce the learning behaviour.","pith_inferences":["A natural extension the paper leaves implicit is that the same mapping could translate other layered quantum circuits into cluster-state measurements, trading circuit depth for cluster-state width and pre-computable entanglement.","The local, angle-only training landscape suggests that gradient estimates could be obtained from measurement statistics alone, for example by finite differences across repeated preparations, which is how the paper itself computes gradients numerically.","The square-lattice results raise the possibility that cluster-state geometry itself is a trainable inductive bias; testing the scheme on larger classical datasets would show whether the accuracy gain over circuit QCNNs persists outside the iris benchmark."],"forward_implications":["A QCNN of arbitrary depth can in principle be executed by preparing one fixed entangled resource state and measuring it, so the difficulty moves from coherent multi-gate control to cluster-state preparation and local measurement.","Because the same physical cluster state is reprogrammed by changing measurement angles, one hardware resource could be reused for many learning tasks without changing the entanglement structure.","Square-lattice cluster states, which have been generated at large scale, appear sufficient for the demonstrated learning tasks, giving a concrete path from current experiments to quantum deep learning.","In the reported benchmarks the measurement-based model converges at least as fast as a circuit QCNN and faster than a classical CNN of the same parameter count, and it reaches higher iris-classification accuracy than the circuit QCNN."],"supporting_citations":[{"why":"defines the QCNN circuit architecture the paper maps onto a cluster state.","marker":"[27]"},{"why":"provides the QCNN-for-classical-data baseline and the two-qubit gate form used in the iris comparison.","marker":"[28]"},{"why":"introduces one-way (measurement-based) quantum computation, the computing model the scheme builds on.","marker":"[34]"},{"why":"supplies the |L4> and |E8> cluster-state gadgets used as elementary rotation and CNOT blocks.","marker":"[50]"},{"why":"reports large-scale two-dimensional cluster-state generation, cited as the experimental platform for the square-lattice demonstrations.","marker":"[47]"},{"why":"reports deterministic generation of two-dimensional cluster states, cited alongside [47] for the same platform.","marker":"[48]"},{"why":"defines the Haldane Hamiltonian and string order parameter used in the first learning task.","marker":"[51]"},{"why":"supplies the iris dataset used in the second learning task.","marker":"[52]"}],"fun_headline_variants":["Cluster state measurements run quantum convolution nets","Train quantum CNN by adjusting measurement bases only","Measurement-based QCNN: no deep circuit control needed","Square-lattice cluster states enable practical quantum deep learning","QCNN via cluster states: faster convergence, simpler setup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire construction depends on one identity: measuring three qubits of a four-qubit chain in the stated rotated bases is supposed to produce a controllable rotation on the fourth qubit. If that identity does not hold, the two-qubit gate gadgets built from it do not perform their assigned gates.","fun_headline_variants_meta":{"raw":{"variants":["Cluster state measurements run quantum convolution nets","Train quantum CNN by adjusting measurement bases only","Measurement-based QCNN: no deep circuit control needed","Square-lattice cluster states enable practical quantum deep learning","QCNN via cluster states: faster convergence, simpler setup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000314,"raw_usage":{"total_tokens":1797,"prompt_tokens":974,"completion_tokens":823,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":751}},"tokens_in":590,"tokens_out":823,"duration_ms":7980,"temperature":1.0,"reasoning_tokens":751,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:06:18.732293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the left-hand side of Eq. (2) for a general input state $|\\psi\\rangle$: because $R_z(\\theta)|0\\rangle$ is proportional to $|0\\rangle$ for every $\\theta$, each projector $\\langle 0|R_z(\\theta)\\rangle$ is, up to an overall phase, just $\\langle 0|$, so the post-measurement state cannot depend on $\\theta$, $\\zeta$, or $\\xi$. Checking the output state numerically for two different angle triples would settle whether the claimed rotation identity holds; the calculation predicts no angle dependence, contradicting Eq. (2).","supporting_citations":[],"review_version":1}