{"id":"26ba1362-9e39-437b-b20b-f89bfa6493fc","arxiv_id":"2608.07822","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Quaternion networks, which use the same SU(2) rotation structure as single-qubit quantum gates, match or outperform depth-3 variational quantum circuits on MNIST, FashionMNIST, and CIFAR-10.","lead":"This paper compares ordinary neural networks, quaternion-based networks, and small quantum circuits on standard image classification tasks, using the same precomputed features for all models. It finds that classical quaternion networks match regular networks and beat the small quantum circuits, suggesting shallow quantum circuits offer no practical edge on these tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The VQC's 12-observable measurement bottleneck, not SU(2) geometry, may explain the gap; a richer observable set could restore parity.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing concern I find: the VQC's information bottleneck (512-D to 8 angles to 12 observables) and the use of features trained for the real baseline. The first part is the more fundamental issue, because the paper's own discussion states the bottleneck may disadvantage quantum models, and the abstract's 'shared local SU(2) geometry and shallow entanglement are insufficient' claim depends on the evaluated scheme being representative. The paper is honest about this limitation and the conclusion is explicitly bounded to 'shallow, measurement-limited variational circuits,' so the experiments are internally consistent and the reported numbers support the narrow claim. The CONDITIONAL verdict is appropriate: the paper should be accepted only with the caveat that the headline generalizes only to the specific encoding/measurement regime. I see no internal inconsistency, p-hacking, or overreach beyond the acknowledged bottleneck, so no change to the reader's verdict is warranted.","tokens_in":17329,"tokens_out":8134,"duration_ms":77154,"concrete_test":"Re-run the CIFAR-10 frozen-ResNet18 comparison with the same depth-3, 8-qubit circuits (both NoEnt and Ent), but replace the 12 fixed Pauli observables with a larger measurement set—e.g., all 64 Z-expectations on single qubits and nearest-neighbor pairs, or classical-shadow estimates of ~128 Pauli expectations—feeding these to the same linear classifier (with appropriately scaled weights). If Quantum-NoEnt/Ent accuracy rises from 41.7%/32.5% to near QuatNet's 45.8%, the gap is driven by measurement dimensionality, not SU(2) geometry. As a control, insert a 12-dimensional linear bottleneck before QuatNet's final layer; if QuatNet then drops to ~32%, the comparison is dominated by representation compression rather than model class.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is externally valid only if the chosen encoding and measurement scheme is representative of shallow VQCs. In the ResNet18 regime, the quantum head compresses 512-D features to 8 encoding angles and then to 12 Pauli observables, whereas QuatNet retains a 128-dimensional hidden representation (H^32) before classification. This is not a resource-matched comparison: the quantum head's final pre-classification representation has 12 real dimensions versus 128 for QuatNet, and the parameter counts differ by 4x (4,282 vs 17,832). The paper explicitly acknowledges the compression 'creates an information bottleneck that may affect the comparison between classical and quantum models' (Discussion). If enlarging the observable set (e.g., classical shadows or learned POVMs) restored parity, the conclusion that 'shared local SU(2) geometry and shallow entanglement are insufficient' would not generalize to shallow VQCs; it would describe only the specific 6/12-observable scheme tested. The title and abstract state the broader claim, so the bottleneck is load-bearing for the paper's headline conclusion, even though the authors carefully bound the detailed conclusion to the evaluated regime.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a controlled benchmark comparing real-valued MLPs, quaternion-valued networks, and depth-3 variational quantum classifiers on identical frozen features from MNIST, FashionMNIST, and CIFAR-10, under both a 16-dimensional learned bottleneck and 512-dimensional frozen ResNet18 embeddings. Across all settings, QuatNet closely matches RealNet and outperforms both product-state and ring-entangled VQCs; the entangled circuit degrades sharply under pretrained CNN features. The authors add robustness checks (five-seed MNIST, sample scaling, an extra Quantum-NoEnt sample-size point) and FS/QFI optimization diagnostics. The paper concludes that shared local SU(2) geometry and shallow entanglement are insufficient for practical quantum advantage on these classical vision tasks, with conclusions explicitly bounded to the shallow, measurement-limited regime studied.","tokens_in":17563,"tokens_out":5628,"duration_ms":52967,"significance":"If the conclusion holds, the paper provides a useful classical baseline and a cautionary data point for quantum machine learning claims on classical vision benchmarks. Strengths include the frozen-feature protocol that isolates head architecture, explicit accuracy and variance tables, effect-size reporting, a five-seed MNIST extension, and an unusually frank limitations section. The main caveat is external validity: the VQC's 6- or 12-observable measurement bottleneck and the RealNet-trained learned bottleneck mean the headline claim is established for a narrow implementation, not for shallow VQCs generally. The paper's own text acknowledges the missing observable-count ablation, which is the key experiment needed to support the broader title claim.","major_comments":[{"comment":"The 512D-to-8-qubit/12-observable compression is an information bottleneck that is not matched in the classical heads. In Table 2, Quantum-NoEnt uses 4,282 trainable parameters versus 17,832 for QuatNet and 66,954 for RealNet, and the final pre-classification representation has 12 real dimensions versus 128 hidden units for QuatNet. The authors state that expanding the measurement space would introduce an additional classical post-processing layer, and later acknowledge that the compression 'may disadvantage the quantum models.' This makes the title and abstract claim about 'shallow variational quantum circuits' broader than what the experiments establish. Please either add an experiment with a richer observable set (e.g., a full set of Pauli observables or classical-shadow-style measurements) or scope the title, abstract, and conclusion to 'shallow, measurement-limited variational quantum circuits.' This is load-bearing for the central claim.","section":"Quantum classifiers; Discussion, Scope, limitations"},{"comment":"The 16-dimensional learned bottleneck is 'trained once as part of the real-valued baseline (seed 42) and then frozen.' QuatNet and the VQC heads are therefore evaluated on features selected to optimize the real baseline's objective, which can systematically disadvantage the non-real heads in all MNIST, FashionMNIST, and 16-D CIFAR-10 experiments. The ResNet18 regime avoids this concern because those features are pretrained and not optimized for any head, but the bottleneck comparisons do not. Please either train the feature extractor jointly with each head, use a representation-neutral preprocessor for the bottleneck regime, or explicitly analyze how the real-baseline-trained extractor could bias the observed head ordering.","section":"Materials and Methods, Frozen feature extraction"},{"comment":"The paper concedes that 'the interpretation of performance gaps in terms of measurement compression and shallow-circuit expressivity limits ... has not been experimentally isolated through ablations on observable count or circuit expressivity.' This missing ablation is exactly what separates the effect of measurement compression from the effect of SU(2) geometry, entanglement, or circuit depth. Without it, the data support the narrow claim that the specific depth-3, 6/12-observable architectures tested underperform the classical heads; the broader explanatory claim that shared local SU(2) geometry and shallow entanglement are insufficient is not directly tested. Adding an observable-count sweep or a comparable-capacity classical head would strengthen the paper considerably.","section":"Discussion, Measurement bottlenecks and shallow-circuit limits"}],"minor_comments":[{"comment":"The single-seed entries at 2,000 and 2,500 samples per class for Quantum-Ent are described only in the table caption; please add an explicit footnote or marker in the table body itself so that readers do not mistake them for three-seed means.","section":"Table 3"},{"comment":"The FS/QFI diagnostic uses K=10 updates, and the text notes this is deliberately short-horizon. Please restate that horizon in the sentence summarizing the 'statistically indistinguishable' loss reduction, since the current wording in the Results could be read as a general optimizer equivalence claim.","section":"Table 1 and Optimization geometry diagnostics"},{"comment":"The lower panels show log-scaled training time, but the axis label reads 'Training Time (seconds)' without indicating the log scale; please add '(log scale)' to the axis label.","section":"Figure 5"},{"comment":"There are minor typographical issues, including 'F ashionMNIST' in Figure 5 labels and inconsistent spacing in 'T able 1' and 'T able 2'; these should be corrected in a final pass.","section":"Throughout"},{"comment":"The phrase that the Quantum-Ent degradation 'persists' under increased data exposure is based on single-seed runs at 2,000 and 2,500 samples, as the paper itself notes; consider using 'is observed at' or 'is indicative at' to reflect the reduced statistical strength of those points.","section":"Robustness across seeds and sample sizes"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript is internally consistent and unusually honest about its limitations, and the code availability statement is a strength. The primary issue is that the title and abstract generalize beyond what the experimental design can support because of the measurement bottleneck and the real-baseline-trained feature extractor. A major revision that either adds an observable-count ablation or re-scopes the headline claims would make this a solid empirical contribution. I see no signs of citation manipulation or overlap concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, transparent empirical comparison that adds another data point against shallow VQC advantage on classical vision, and the genuinely new bit is the quaternion network as a classical SU(2)-matched baseline. The entanglement reversal under pretrained ResNet features is also worth noting.\n\nThe paper does a lot right: identical frozen features across heads, five-seed MNIST, sample-scaling checks, explicit reporting of variance and wall-clock time, and a Discussion that directly names the measurement bottleneck. The code is available, the tables are clear, and the authors are careful to bound the conclusion to the shallow, measurement-limited regime. I think the data support the ordering: QuatNet ≈ RealNet > product-state VQC > entangled VQC in the ResNet18 regime. The Friedman test at n=5 is fine, and the effect sizes are large.\n\nThe main soft spot is the one the stress test flags: the quantum arm compresses 512-D features to 8 angles and then to 12 Pauli expectations, while QuatNet keeps a 128-dimensional hidden representation. That is not a resource-matched comparison, and the title/abstract lean on a broader claim than the experiments can support. If a richer observable set restored parity, the headline would not generalize. That said, the authors explicitly acknowledge this bottleneck and say the results characterize the evaluated regime. The bottleneck is load-bearing for the abstract, but it is not hidden.\n\nA minor additional concern: the frozen features are trained as part of the real baseline, which is a standard transfer setup but technically gives the classical heads a small advantage in feature choice. I do not think it changes the ordering, but it is a small bias in the comparison.\n\nWho is this for? People working on QML benchmarking, variational circuits, and classical baselines for quantum advantage. It is not a theoretical contribution, but it is a worthwhile empirical benchmark. I would send it to peer review; it deserves a referee's time, and a revision should tighten the title and abstract to match the measurement-limited scope.","headline":"A careful, honest QML benchmark with a genuinely useful classical SU(2) baseline; the measurement bottleneck is a real caveat but the authors mostly own it.","tokens_in":18061,"tokens_out":2464,"would_cite":true,"duration_ms":22812,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Classical quaternion networks match or beat shallow variational quantum circuits on vision benchmarks","keywords":["quaternion neural networks","variational quantum circuits","SU(2) geometry","quantum machine learning","vision benchmarks","entanglement","measurement bottleneck","natural gradients"],"falsifier":"Run the same frozen ResNet18 CIFAR-10 comparison with an amplitude-based encoding or a richer measurement strategy (e.g., classical shadow tomography or a much larger observable set) on the same depth-3 circuits; if the entangled VQC then matches or exceeds quaternion accuracy, the paper's conclusion that shallow SU(2) geometry and entanglement are insufficient would be refuted.","tokens_in":17121,"feed_emoji":"⚛️","tokens_out":4614,"duration_ms":35057,"temperature":0.7,"pith_summary":"This paper asks whether the shared local SU(2) geometry of quaternion-valued neural networks and single-qubit variational quantum circuits translates into comparable learning behavior on classical image classification. On identical frozen features from MNIST, FashionMNIST, and CIFAR-10, quaternion classification heads match or closely approach real-valued MLPs while using fewer parameters, and they consistently outperform depth-3 product-state and entangled VQCs. The entangled circuit, which adds a ring of CNOT gates, gives only modest gains on grayscale data and degrades sharply under frozen ResNet18 features, falling 9.25 percentage points below the product-state circuit on CIFAR-10. The authors conclude that shared local SU(2) geometry and shallow ring entanglement are insufficient, within this shallow, measurement-limited regime, to confer practical quantum advantage on these tasks.","feed_headline":"Quaternion nets beat shallow quantum circuits on vision tasks","feed_subtitle":"Classical SU(2) models match real baselines while depth-3 VQCs lag by up to 9 points on CIFAR-10.","key_machinery":"The load-bearing object is the Lie group isomorphism between unit quaternions and SU(2): unit quaternions form the three-sphere $S^3$, which maps onto single-qubit rotations via a double cover of $SO(3)$. Quaternion layers implement this geometry as Hamilton products with renormalized unit-quaternion weights, while the VQCs implement the same local rotations as $R_Y$ and $R_Z$ gates with data re-uploading and an optional ring of CNOT gates for entanglement. The experimental design isolates this geometry by freezing feature extraction, a shared 16-dimensional learned bottleneck or a frozen 512-dimensional ResNet18 embedding, so that only the classification head differs; quantum heads then compress features into 4 or 8 qubits and read out 6 or 12 Pauli expectation values before a linear classifier.","core_discovery":"On the paper's own terms, the central discovery is that a purely classical implementation of local SU(2) structure, quaternion-valued layers applying Hamilton products with unit-quaternion weights, reproduces the behavior of real-valued MLP heads and outperforms the shallow variational quantum classifiers evaluated here across all datasets and feature regimes. QuatNet matches RealNet within 0.2 percentage points on MNIST and FashionMNIST (93.64% vs 93.54% and 84.47% vs 84.60%) and retains 94–97% of real-valued accuracy on CIFAR-10, while the product-state VQC trails by 2.4–6.1 points and the entangled VQC drops 9.25 points below the product-state circuit under frozen ResNet18 features. The paper interprets this as evidence that measurement compression, shallow depth, and optimization instability, rather than the SU(2) rotation group itself, limit the quantum models.","pith_inferences":["A fair test of the paper's negative conclusion would require richer quantum encodings, such as amplitude encoding or larger observable sets (e.g., classical shadows), because the current design compresses 512 features into 8 qubits and 12 expectations, an information bottleneck the classical heads do not face.","The results suggest an empirical rule for near-term quantum machine learning: when classical data has no intrinsic quantum structure, the burden of proof shifts to showing that the encoding and measurement strategy, not the ansatz family, is what makes quantum features useful.","If the bottleneck explanation is right, one testable prediction is that enlarging the measurement space should narrow the gap between the entangled circuit and QuatNet on ResNet18 features, while deepening the circuit would mainly add barren-plateau difficulty."],"forward_implications":["Claims of practical quantum advantage on classical vision benchmarks should be benchmarked against parameter-efficient quaternion networks, not only real-valued MLPs, since both classical heads share the SU(2) local geometry.","The benefit of shallow entanglement depends on feature quality: it helps slightly on simple grayscale inputs, does nothing on a learned bottleneck, and actively hurts under rich pretrained CNN features.","Fubini–Study and quantum-Fisher-information natural gradients align with the natural-gradient direction but do not reduce short-horizon loss compared with Adam, indicating optimizer geometry is not the bottleneck.","Measurement compression is a prime suspect for the quantum gap: product-state circuits with 6 or 12 observables cannot carry the information that quaternion layers preserve in their continuous 4D representations."],"supporting_citations":[{"why":"Supplies the Lie group isomorphism between unit quaternions and SU(2) that motivates the comparison.","marker":"[5]"},{"why":"Defines quaternion recurrent layer constructions with Hamilton products that QuatNet adapts.","marker":"[2]"},{"why":"Provides the deep quaternion network layer formulation used for the classification heads.","marker":"[3]"},{"why":"Source of the MNIST benchmark and its train/test split.","marker":"[12]"},{"why":"Source of the FashionMNIST benchmark and its train/test split.","marker":"[13]"},{"why":"Source of the CIFAR-10 benchmark and its train/test split.","marker":"[14]"},{"why":"Establishes barren plateaus as the reason for restricting circuits to depth 3.","marker":"[15]"},{"why":"Supplies the data re-uploading circuit strategy used in all VQC variants.","marker":"[18]"},{"why":"Defines the Fubini–Study and quantum Fisher information natural-gradient method used in optimization diagnostics.","marker":"[19]"}],"fun_headline_variants":["Quaternion nets beat shallow quantum circuits on vision","Quaternion nets outpace shallow VQCs on vision","Classical SU(2) nets beat quantum on vision","Quaternion networks outperform shallow VQCs on vision","No quantum advantage: quaternion nets win on vision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the particular shallow-VQC encoding and measurement scheme used here, compressing 512 features to 8 qubits and reading out only 12 Pauli observables, is representative enough of shallow variational quantum classifiers that its underperformance reflects a structural limit rather than an information bottleneck.","fun_headline_variants_meta":{"raw":{"variants":["Quaternion nets beat shallow quantum circuits on vision","Quaternion nets outpace shallow VQCs on vision","Classical SU(2) nets beat quantum on vision","Quaternion networks outperform shallow VQCs on vision","No quantum advantage: quaternion nets win on vision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000727,"raw_usage":{"total_tokens":3349,"prompt_tokens":1127,"completion_tokens":2222,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":743,"completion_tokens_details":{"reasoning_tokens":2145}},"tokens_in":743,"tokens_out":2222,"duration_ms":13695,"temperature":1.0,"reasoning_tokens":2145,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:12:51.727167+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same frozen ResNet18 CIFAR-10 comparison with an amplitude-based encoding or a richer measurement strategy (e.g., classical shadow tomography or a much larger observable set) on the same depth-3 circuits; if the entangled VQC then matches or exceeds quaternion accuracy, the paper's conclusion that shallow SU(2) geometry and entanglement are insufficient would be refuted.","supporting_citations":[{"cited_title":"Princeton University Press, Princeton, NJ (1999)","cited_arxiv_id":null,"evidence_quote":"Supplies the Lie group isomorphism between unit quaternions and SU(2) that motivates the comparison."},{"cited_title":"In: Proceedings of the International Conference on Learning Representations (ICLR) (2019)","cited_arxiv_id":null,"evidence_quote":"Defines quaternion recurrent layer constructions with Hamilton products that QuatNet adapts."},{"cited_title":"In: Proceedings of the IEEE International Joint Conference on Neural Networks (IJCNN), pp","cited_arxiv_id":null,"evidence_quote":"Provides the deep quaternion network layer formulation used for the classification heads."},{"cited_title":"Technical Report CSRG-TR-2009-1, University of Toronto, Toronto, Canada (2009)","cited_arxiv_id":null,"evidence_quote":"Source of the CIFAR-10 benchmark and its train/test split."}],"review_version":1}