{"id":"12191093-14f5-4636-af34-4271c9e01149","arxiv_id":"2509.02374","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A single-layer complex-valued neural network constrained to the Stiefel manifold via Cayley updates is used to learn and transpile quantum circuit unitaries, with fidelity reaching 1 on reported toy examples.","lead":"The authors train a simple complex-valued network whose internal matrix is kept unitary during training, then turn that matrix into a quantum circuit. The work is a minimal demonstration of an existing optimization method applied to quantum circuit design, with most experimental details missing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training data underspecified: no number of input-output pairs, no held-out evaluation; F=1 on training data does not establish implementation of target unitary.","rationale":"I re-read Sections 2, 4, 5, and the Appendix. The Stiefel-manifold/Cayley-update machinery is standard and appears internally consistent, so I do not object to the mathematics itself. The decisive weakness is empirical: the paper never specifies how many input-output pairs define the learning problem, how they are sampled, or how generalization is measured. Identifying a 2^n-dimensional unitary requires at least N=2^n linearly independent input states; otherwise infinitely many unitaries fit the training data. The reported fidelity F=1 is computed against the target unitary during training, with no held-out evaluation, so it cannot support the claim that the learned circuit implements the target operation. This is exactly the reader's weakest assumption, and it is load-bearing: without sufficient training data and out-of-sample testing, the central circuit-design claim is not established. The reader's rejection remains appropriate, so I recommend no change to the verdict.","tokens_in":5680,"tokens_out":9984,"duration_ms":120539,"concrete_test":"Ask the authors to disclose the exact training protocol for the 5-qubit experiment (number k of input-output pairs, how states are sampled, loss definition). Then re-run training with a specified k and evaluate the learned unitary on computational-basis states not in the training set, reporting average output fidelity. If k < N=32 or held-out fidelity is materially below 1, the F=1 result is an in-sample artifact and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the learned unitary Wfinal reproduces the target circuit Utrue, and the only evidence is in-sample training curves with F=1. For this to be sound, the supplied input-output pairs must uniquely determine Utrue. With N=2^n, k pairs (|psi_i>, Utrue|psi_i>) determine a unitary only on the span of the inputs; if k<N or the inputs are linearly dependent, infinitely many unitaries fit the data. Section 2 specifies only that 'many input and output quantum states are known'; Section 5 reports a 5-qubit experiment without stating k, the state distribution, the loss, or any train/test split. Figure 2b shows fidelity reaching 1 during training, but no held-out states are evaluated. Thus the empirical claim that Wfinal implements the target circuit is unsupported. The Cayley/Stiefel mathematics in the Appendix is standard and likely correct, but it only ensures a unitary descent step; it does not ensure the learned unitary is the target one. This is the load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes learning a quantum circuit's unitary transformation by training a single-layer complex-valued neural network whose weight matrix is constrained to the Stiefel manifold (i.e., unitary) via the Cayley update of Eq. (4). The authors claim this update preserves unitarity at every step and guarantees a globally decreasing loss, and they report numerical experiments on a 5-qubit circuit and a 2-qubit entangling operation, with training loss decreasing and fidelity (Eq. 5) reaching 1, after which the learned unitary is transpiled into a quantum circuit. The appendix contains proofs of unitarity preservation and of a local decrease property for sufficiently small learning rates.","tokens_in":5954,"tokens_out":4572,"duration_ms":57739,"significance":"If fully validated, the proposal would be a simple baseline for learning unitary maps from input--output quantum state pairs, with the useful property that the learned weight matrix is exactly unitary throughout training. The mathematical appendix correctly proves the key feasibility of the Cayley/Stiefel update: the update (A7) stays on the Stiefel manifold, and the directional derivative at λ=0 is non-positive. The method also has the practical advantage of being directly expressible in existing transpilation toolchains. However, the central empirical claim—that the learned Wfinal implements the target quantum circuit—is not established by the reported experiments, and the claimed global monotone decrease is stronger than what the proof supports. These issues are load-bearing for the paper's main contribution.","major_comments":[{"comment":"The paper states as a guaranteed property that \"the loss function decreases globally, i.e., f(W(k+1)) ≤ f(W(k))\" for the update (4). The appendix proof, Eqs. (A12)–(A15), only shows that if the Riemannian gradient AW is nonzero, then for sufficiently small λ one has f(Y(λ)) < f(W). At a stationary point (AW=0) the derivative is zero and no decrease is proved. Moreover, no rule is given for choosing λ that satisfies the bound (A15). The monotone-decrease claim used to motivate the method is therefore not established; please replace it with the local statement and provide a concrete step-size procedure or line search.","section":"§4 and Appendix, Claim 2"},{"comment":"The empirical claim that the trained Wfinal reproduces the target unitary Utrue is not supported by the experiments as reported. A training set of k pairs (|ψ_i>, Utrue|ψ_i>) determines a unitary only on the span of the input states; for k < 2^n or for linearly dependent inputs, infinitely many unitaries fit the training data. Section 2 only says \"many input and output quantum states are known,\" and Section 5 reports a 5-qubit experiment without stating the number of training pairs, their distribution, or their linear independence. Figure 2b shows fidelity reaching 1, but no held-out fidelity on states not used in training is reported. Please specify the training data and add held-out evaluation before the fidelity claim can be assessed.","section":"§2, §5, Fig. 2"},{"comment":"The loss f(W) and its Euclidean gradient G are never concretely specified for the reported experiments. Equation (1) defines the constrained optimization abstractly, but the experiments require a concrete loss—mean squared error over output states, trace fidelity, or another—to compute G in Eq. (3) and the update (4). The paper also omits the learning-rate schedule, initialization, number of epochs, and the exact 5-qubit target circuit. Without these details, the numerical results are not reproducible and the fidelity curves cannot be interpreted.","section":"§5, Eqs. (1)–(4)"},{"comment":"The 2-qubit example in the appendix is also under-specified: the target entangling operation is not named, the training data are not described, and Figure 4 shows a transpiled circuit without any comparison to the target or a correctness metric beyond the stated unitary error. This example should either be integrated into a fully specified experimental section or removed.","section":"Appendix, 'Additional Result'"}],"minor_comments":[{"comment":"Reference [1] is listed as \"Anonymous\" and reference [6] contains \"and 1 others\"; these are incomplete citations and should be corrected.","section":"References"},{"comment":"There are numerous typos: \"varius,\" \"ouput,\" \"Caley\" (should be Cayley), \"uniatry,\" and \"Steifel\" in figure captions. Please proofread carefully.","section":"Throughout"},{"comment":"The axes in Figures 2a and 2b are not fully described. In particular, the fidelity plot should state whether the fidelity is evaluated on the training data or on a separate test set, and the loss plot should state which loss function is displayed.","section":"§5, Fig. 2"},{"comment":"The claim that this is \"one of the first applications of Cayley-transform-based manifold optimization in quantum computing\" is vague, especially given reference [17] on unitary neural networks for quantum circuit design. Please clarify the precise novelty with a more systematic comparison to prior work.","section":"§1, Literature Study"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not ready in its present form. The central idea is a direct application of known unitary RNN training to quantum circuit synthesis, and the mathematical core is standard. The main gaps are experimental: no loss function, no training data specification, and no held-out evaluation. These are fixable in a revision, but the authors should also soften the claims from \"global decrease\" to the local decrease that the appendix actually proves. I would be willing to look at a substantially revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper applies the Cayley-transform update on the Stiefel manifold—originally from Wisdom et al.'s unitary RNN work—to quantum circuit synthesis. The mathematical core is standard and, for the most part, correctly presented. The unitarity-preservation proof (Claim 1) is clean. And the idea of avoiding the intermittent re-orthogonalization of Gram-Schmidt by staying on the manifold throughout training is sensible and worth a sentence of credit.\n\nThe problems are in the claims and the experiments. Section 4 says the loss 'decreases globally,' but the appendix proof only establishes a decrease for sufficiently small learning rate—that is a local, not global, statement. Overclaiming to 'global' needs correcting. More important, the empirical section is underspecified: we don't know how many input-output pairs were used, what states they were, what loss was minimized, or whether any held-out states were evaluated. The fidelity (Eq. 5) is reported between U_learned and U_true, but without saying whether the training set uniquely determines the unitary, F=1 on training data is just convergence of a fit. The 2-qubit appendix example adds little: it is the same procedure without additional details. No code or data is provided, so nothing can be reproduced.\n\nThere is also a citation problem: reference [1] is 'Anonymous' and not verifiable, and the literature review is thin.\n\nIn short, the approach is reasonable but the paper does not make a convincing case that the learned unitary implements the target circuit. The core idea warrants a revision that states the training set size, the loss function, and a train/test split, and that corrects the 'global decrease' claim. As it stands, I'd desk-reject; a revised version with these details might deserve referee time.\n\nHope this helps.","headline":"Standard optimizer applied to a new domain, with a local-decrease proof overclaimed as global and an experiment too underspecified to support the central claim.","tokens_in":6349,"tokens_out":2985,"would_cite":false,"duration_ms":33609,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68T07"],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"The paper claims that a single-layer complex-valued neural network trained on the Stiefel manifold can learn the unitary of a quantum circuit and reach fidelity 1 on the tested 2-qubit and 5-qubit examples.","keywords":["quantum circuit design","complex-valued neural network","Stiefel manifold","Cayley transform","unitary matrix","quantum state fidelity","circuit transpilation","quantum machine learning"],"falsifier":"Take a fixed 2-qubit circuit U_true and provide only one state pair, say |00⟩→U_true|00⟩, to the network. Train until fidelity F=1 on that pair, then compute F on the held-out basis states |01⟩, |10⟩, and |11⟩. If the held-out fidelity is measurably below 1 because many unitaries agree on |00⟩, the paper's claim that F=1 implies U_learned=U_true fails in that setting.","tokens_in":5619,"feed_emoji":"⚛️","tokens_out":8548,"duration_ms":98178,"temperature":0.7,"pith_summary":"The paper is trying to show that a single-layer complex-valued neural network, trained on pairs of input and output quantum states, can recover the unitary transformation of a quantum circuit while keeping the weight matrix unitary at every step. Training is constrained to the Stiefel manifold of unitary matrices, and each update uses a Cayley transform of a skew-Hermitian direction. The authors report that this gives a globally decreasing loss, avoids the loss spikes of periodic Gram-Schmidt re-orthogonalization, and yields a learned unitary with fidelity 1 on the tested examples. If the claim transfers to larger circuits, circuit synthesis could be done by supplying input-output states and letting the network find the unitary.","feed_headline":"Neural net learns 5-qubit circuit at fidelity 1","feed_subtitle":"Single-layer complex-valued network keeps weights unitary throughout training, then turns the learned unitary into a transpiled circuit.","key_machinery":"The central object is the Stiefel manifold V_N(C^N)={W∈C^{N×N}: W†W=I}, the set of unitary matrices. The carrying mechanism is the Cayley update W(k+1)=(I+λA/2)^{-1}(I−λA/2)W(k) with A=GW†−WG†, where G is the Euclidean gradient of the loss. Because A is skew-Hermitian, the Cayley factor is unitary and maps one unitary matrix to another; the proof shows the curve's tangent at λ=0 is −AW and the directional derivative of f is −‖AW‖²_c, so the loss descends along the manifold. This is what guarantees unitarity at every step rather than only after intermittent re-orthogonalization.","core_discovery":"The paper's central claim is that solving the constrained problem min f(W) subject to W†W=I by gradient descent on the Stiefel manifold preserves unitarity throughout training and produces a learned unitary that matches the target circuit. With W as the network weight matrix and G the Euclidean gradient, the paper constructs A=GW†−WG†, which is skew-Hermitian, and updates W(k+1)=(I+λA/2)^{-1}(I−λA/2)W(k). The appendix proves that each update stays on the Stiefel manifold and that, for small enough λ, the loss strictly decreases unless the Riemannian gradient vanishes. On a 5-qubit sample circuit the unitary error ‖WW†−I‖² reaches about 10⁻¹¹, and the fidelity F=(1/d)|Tr(U_true† U_learned)| r","pith_inferences":["The paper leaves open how many input-output state pairs are used and does not test on held-out states; the claim that reaching fidelity 1 means reproducing the target circuit depends on the training set uniquely determining the unitary, which is not demonstrated.","A natural extension is to train on a single basis-state pair for a known random unitary and then evaluate fidelity on the remaining basis states; if many unitaries fit the one pair, the learned matrix will not match the target on the held-out states.","The method could be applied to learning unknown processes: given experimentally measured input-output state pairs, the same Stiefel-constrained optimization would return the unitary closest to the data, with unitarity acting as a built-in regularizer."],"forward_implications":["Because every updated weight matrix is unitary, the learned matrix is always physically valid as a quantum gate and can be handed to a transpiler to produce an executable circuit.","The loss decreases monotonically for sufficiently small learning rates, so training avoids the loss spikes that intermittent Gram-Schmidt orthogonalization introduces.","If the fidelity result generalizes, a user would only need known input-output quantum state pairs to obtain a circuit, bypassing manual gate decomposition.","The same update rule could be applied to any complex-valued neural network that needs a unitary weight matrix, not just quantum circuit synthesis."],"supporting_citations":[{"why":"Supplies the Cayley-transform update rule and the proof that each step stays on the Stiefel manifold and decreases the loss.","marker":"[16]"},{"why":"Defines the Stiefel manifold V_N(C^N) and the tangent-space setup used for the constrained optimization.","marker":"[14]"},{"why":"States the physical postulate that quantum time evolution is unitary, which is the constraint the method preserves.","marker":"[10]"},{"why":"Describes a unitary-neural-network approach to quantum circuit design that the paper contrasts with its own method.","marker":"[17]"},{"why":"Proposes a quantum neural network framework for approximating unitary matrices, a related approach the method builds on or differentiates from.","marker":"[11]"},{"why":"Earlier work enforcing unitarity in recurrent network weights, motivating the use of unitary matrices as trainable parameters.","marker":"[2]"},{"why":"Representative of orthogonalization schemes that enforce unitarity only intermittently, the baseline whose loss spikes motivate the Stiefel approach.","marker":"[4]"}],"fun_headline_variants":["Single-layer complex net achieves perfect quantum circuits","Stiefel manifold optimization yields unit-fidelity circuits","Complex-valued neural net learns 5-qubit gate at fidelity 1","Quantum circuit design via complex net on Stiefel manifold","Neural network preserves unitarity to learn quantum circuits"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper assumes the input-output state pairs used for training uniquely determine the target unitary, so that reaching fidelity 1 on those pairs means the learned circuit is the target circuit.","fun_headline_variants_meta":{"raw":{"variants":["Single-layer complex net achieves perfect quantum circuits","Stiefel manifold optimization yields unit-fidelity circuits","Complex-valued neural net learns 5-qubit gate at fidelity 1","Quantum circuit design via complex net on Stiefel manifold","Neural network preserves unitarity to learn quantum circuits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000763,"raw_usage":{"total_tokens":3167,"prompt_tokens":634,"completion_tokens":2533,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":378,"completion_tokens_details":{"reasoning_tokens":2454}},"tokens_in":378,"tokens_out":2533,"duration_ms":21485,"temperature":1.0,"reasoning_tokens":2454,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:35:15.624798+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed 2-qubit circuit U_true and provide only one state pair, say |00⟩→U_true|00⟩, to the network. Train until fidelity F=1 on that pair, then compute F on the held-out basis states |01⟩, |10⟩, and |11⟩. If the held-out fidelity is measurably below 1 because many unitaries agree on |00⟩, the paper's claim that F=1 implies U_learned=U_true fails in that setting.","supporting_citations":[{"cited_title":"Hershey, Jonathan Le Roux, and Les Atlas","cited_arxiv_id":null,"evidence_quote":"Supplies the Cayley-transform update rule and the proof that each step stays on the Stiefel manifold and decreases the loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Stiefel manifold V_N(C^N) and the tangent-space setup used for the constrained optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"States the physical postulate that quantum time evolution is unitary, which is the constraint the method preserves."},{"cited_title":"A quantum neural network framework for scalable quantum circuit approximation of unitary matrices","cited_arxiv_id":"2405.00012","evidence_quote":"Proposes a quantum neural network framework for approximating unitary matrices, a related approach the method builds on or differentiates from."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier work enforcing unitarity in recurrent network weights, motivating the use of unitary matrices as trainable parameters."},{"cited_title":"On some orthogonalization schemes in Tensor Train format","cited_arxiv_id":"2211.08770","evidence_quote":"Representative of orthogonalization schemes that enforce unitarity only intermittently, the baseline whose loss spikes motivate the Stiefel approach."}],"review_version":1}