{"id":"08519971-e259-4d75-acd1-60f1e4362d41","arxiv_id":"2411.17726","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A rescaled feature map (2*x-1.5 on normalized inputs) combined with a standard variational circuit is claimed to reach 100% accuracy on a two-class mobile data usage task while using fewer gates than the IBM tutorial baseline.","lead":"This paper proposes a quantum neural network whose input encoder rescales data with a simple linear formula before mapping it into quantum states. The authors report perfect accuracy on a phone-data plan prediction task with fewer gates than a standard IBM benchmark, but the result is not backed by error bars or public data.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (19) maps [0,1] inputs to [-1.5,0.5], not [-1.5,1.5], so the EFM as written is not the circuit motivated in Sec. III; the reported advantage is therefore tied to a formula that contradicts the paper's own design argument.","rationale":"The reader's weakest assumption—that Eq. (19) does not implement the claimed mapping—is precisely the most load-bearing internal inconsistency I find. The paper's novelty and contribution rest on the EFM, and the EFM is defined by Eq. (19). The arithmetic error directly severs the link between the design motivation (Sec. III) and the implemented circuit (Sec. IV). This is not a matter of disagreeing with a community consensus; it is a correctness defect in the paper's own specification. I also note that the convergence claim is contradicted by the paper's own description of Model 1, and the absence of code, data, error bars, and train/test details means that the reported 100% versus 97.9-99.9% differences cannot be statistically distinguished. I agree with the reader that REJECT is appropriate at moderate confidence: the empirical finding may survive correction of Eq. (19), but the manuscript as written does not support its central claim. I would not escalate to high confidence without attempting the concrete re-run, because the formula discrepancy could plausibly be a typographical error rather than a flaw in the actual experiments.","tokens_in":10112,"tokens_out":6347,"duration_ms":63784,"concrete_test":"Obtain or re-implement the exact Qiskit circuits described in Sec. IV and run them on the same 1000-sample mobile-data dataset with a fixed train/test split and 10 independent seeds for each optimizer (COBYLA, SPSA, AQGD). Compare three variants: (a) literal EFM with 2*x - 1.5, (b) corrected EFM with 3*x - 1.5, which actually maps [0,1] to [-1.5,1.5], and (c) the benchmark. Report mean plus/minus standard deviation of test accuracy and the iteration number at which loss plateaus. If only (b) reproduces the 100% accuracy and faster convergence, Eq. (19) is a load-bearing error; if (a) also reproduces the reported results, the range argument is not essential and the core empirical claim may still hold; if neither reproduces them, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the EFM-based EQNN achieves higher accuracy with fewer gates and faster convergence. The design rationale in Sec. III.C-D is that input values in [-1.5,1.5] are especially suitable for the simplified QNN. Section IV.A then states that inputs are normalized to [0,1] and mapped by Eq. (19), 2*x - 1.5, 'to the range [-1.5,1.5]'. This is arithmetically false: for x in [0,1], 2*x - 1.5 lies in [-1.5,0.5]. The circuit in Eq. (20) and Fig. 9 therefore executes RY gates with angles in [-1.5,0.5], not the interval that the preceding analysis identifies as beneficial. Because the claimed 100% accuracy and gate-count advantage are attributed to this exact EFM, the error is load-bearing: if 3*x - 1.5 was actually used, the manuscript does not describe the evaluated model; if 2*x - 1.5 was actually used, the stated motivation is false. This is not a minor wording issue because no code, data, train/test split, or error bars are provided, so the reader cannot determine which mapping was implemented. A second internal contradiction compounds the problem: Sec. V.B states that the benchmark converges faster than EQNN Model 1 under AQGD, while the abstract claims the EQNN 'converges faster to the optimal solution under different optimization algorithms'. The paper shows convergence curves only for AQGD and none for COBYLA or SPSA, despite the broader claim. The empirical conclusion may be true, but as written the central argument is internally inconsistent and unverifiable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Enhanced Quantum Neural Network (EQNN) built from an Enhanced Feature Map (EFM) in which normalized input features are linearly transformed and fed into RY gates, followed by a variational RealAmplitudes circuit. The method is evaluated on a mobile data usage classification task with 1000 samples, comparing accuracy, quantum-gate count, and convergence speed against an IBM Qiskit tutorial benchmark across COBYLA, SPSA, and AQGD optimizers. The authors claim that the EQNN achieves higher accuracy with fewer quantum logic gates and converges faster under different optimization algorithms.","tokens_in":10491,"tokens_out":5227,"duration_ms":47265,"significance":"If fully supported, the contribution would be a pragmatic but modest feature-map design with a small empirical demonstration. The paper does provide some theoretical motivation by fitting linear, sigmoid, and tanh functions with a simplified single-qubit QNN, and it reports accuracy tables. However, the central empirical claims are not verifiable as written: the manuscript gives no code, no data split, no random seeds, no repeated runs, and no error bars. More importantly, the paper contains an arithmetic error in the feature-map formula that contradicts its own design rationale, and the convergence claim in the abstract is contradicted by the paper's own Figure 14. Thus the claimed advantages are not established.","major_comments":[{"comment":"The manuscript states that the transformation 2*x - 1.5 maps normalized inputs x in [0,1] to the range [-1.5, 1.5], but the actual range is [-1.5, 0.5]. Since Section III.C and III.D identify [-1.5, 1.5] as the suitable input range based on the fitting experiments, the EFM circuit in Eq. (20) and Figure 9, which uses 2*x - 1.5 as RY rotation angles, is not the circuit the design argument motivates. The paper does not provide code or data, so it is impossible to determine whether the experiments used 2*x - 1.5 (contradicting the stated design) or some other mapping such as 3*x - 1.5 (which would make the text wrong). This error is load-bearing because the claimed accuracy and gate-count advantages are attributed to the EFM.","section":"IV.A, Eq. (19)-(20)"},{"comment":"The abstract claims that the EQNN 'converges faster to the optimal solution under different optimization algorithms,' but in Section V.B the authors explicitly state that for AQGD the benchmark converges faster than EQNN Model 1 (benchmark flattens at iteration 20, Model 1 at iteration 60). The convergence curves in Figures 13-16 are only for AQGD; no convergence results are reported for COBYLA or SPSA. Thus the central convergence claim is directly contradicted for Model 1 and unsupported for the other optimizers. The phrase 'different optimization algorithms' in the abstract should be constrained or removed.","section":"Abstract and V.B (Figs. 13-16)"},{"comment":"The experimental section provides no information on how the 1000 samples are split into training and test sets, how random initial weights are chosen, how many independent runs are performed, or what the variance of the accuracy estimates is. All EQNN models are reported to achieve 100.0% accuracy under every optimizer, while the benchmark ranges from 97.9% to 99.9%, but without error bars or repeated trials the difference could easily be within statistical fluctuation. This lack of reproducibility information undermines the empirical basis for the claimed advantages.","section":"V.A-V.B, Tables II-III"},{"comment":"The gate-count numbers used to support the 'fewer quantum logic gates' claim are internally inconsistent with the circuits described. Section II.D states that one iteration of RealAmplitudes on two qubits (Figure 3) consists of two RY gates and one CNOT gate, i.e., 3 gates, and that three iterations (Figure 4) use 11 gates, but the natural count from the described structure is 9. Similarly, Section IV.B reports that EQNN Model 1's variational model uses 5 gates, but one iteration of RealAmplitudes as defined would use 3 gates, making the total EFM+variational count 8 rather than 10. These discrepancies affect the headline quantitative claim, even if the qualitative ordering (EQNN models use fewer gates than the benchmark) might still hold under corrected counts.","section":"II.D and IV.B, gate-count calculations"}],"minor_comments":[{"comment":"The second EQNN column in both tables is labeled 'EQNN Model 1' a second time; it should be 'EQNN Model 2'. This makes the tables difficult to interpret.","section":"Tables II and III"},{"comment":"The text says there are 8 weight variables W = {w0, w1, …, w8}, but the list contains 9 elements; either the count or the index range should be corrected.","section":"II.D"},{"comment":"The text refers to 'GY Gate' in the description of the proposed feature map; this appears to be a typo for 'RY Gate.'","section":"IV.A"},{"comment":"The sentence stating that EQNN Model 1 uses 5 gates for the variational model plus 5 for the feature map totals 10 gates is inconsistent with the earlier definition of one RealAmplitudes iteration as 3 gates; the text should resolve this discrepancy.","section":"IV.B"},{"comment":"Reference [15] lists the author as 'O. A. Warga,' which may be a misspelling of 'Warqa' as it appears in the original conference publication; please verify.","section":"References"}],"recommendation":"reject","confidential_remarks":"This is a very lightweight empirical paper whose central claims rest on an arithmetic error (Eq. 19) and an internal contradiction between the abstract and the reported convergence behavior. The absence of code, data splits, seeds, and repeated runs means the results cannot be independently verified. Even if the EFM formula were corrected, the paper would need substantial additional experimental rigor and a more careful treatment of the convergence claim to meet the standards of a serious quantum-computing journal. The paper also appears to be a translation of a traditional Chinese conference-style manuscript; the journal should consider whether this fits its scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a routine angle-encoding variant with a clean comparison setup, but the load-bearing formula in Eq. (19) is arithmetically wrong as written, and the convergence claim in the abstract contradicts the paper's own AQGD results. I would not send it to peer review in its current form.\n\nWhat's actually new: the specific circuit—H, RY with 2x-1.5, CNOT—applied to a two-qubit mobile-data classification task and compared against the IBM Qiskit tutorial baseline. That combination is probably new, but it's a run-of-the-mill variant of angle encoding; the Sec. III activation-function fits are one-qubit demonstrations of known single-qubit expressibility, not a theoretical contribution.\n\nWhat the paper does well: it reports gate counts and accuracy for three optimizers, uses a real dataset, and makes the comparison to a public benchmark concrete. If the empirical result holds, it's a modest data point.\n\nThe soft spots are serious. Eq. (19) says 2*x-1.5 maps normalized inputs to [-1.5,1.5]. For x in [0,1] it actually maps to [-1.5,0.5]. The whole design rationale in Sec. III is that [-1.5,1.5] is the favorable range, so the EFM as written is not the circuit the paper motivated. If the intended map was 3*x-1.5, the manuscript doesn't describe what was actually run. Since the accuracy and gate-count results are attributed to this EFM, the error is load-bearing, not cosmetic.\n\nThe convergence claim is also overreach. The abstract says EQNN converges faster 'under different optimization algorithms,' but Sec. V.B reports that under AQGD the benchmark converges faster than Model 1 (benchmark loss flattens around iteration 20, Model 1 around 60). Only AQGD curves are shown; there is no evidence for COBYLA or SPSA convergence speed.\n\nThere are also no error bars, no repeated runs, no explicit train/test split or seeds, and no code or data. With 1000 samples and 100 iterations, a 97.9% vs 100% gap could easily be noise. The empirical conclusion may be true, but the paper as written is internally inconsistent and unverifiable.\n\nWho this is for: a reader collecting feature-map recipes might skim it, but I would not cite it in this state. The fixable parts are fixable, but they require the authors to correct the formula, narrow the convergence claim, and release data/code with statistics.\n\nRecommendation: desk reject for now; if the authors resubmit with the formula corrected and the claims matched to evidence, it becomes a borderline minor empirical paper.","headline":"Routine angle-encoding variant with a load-bearing arithmetic error in Eq. (19) and a convergence claim contradicted by its own results; not referee-ready.","tokens_in":11020,"tokens_out":3796,"would_cite":false,"duration_ms":32911,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that replacing the baseline feature map with a five-gate Enhanced Feature Map lets a quantum neural network classify mobile data plan choices with 100% accuracy using fewer quantum gates and converging faster under three…","keywords":["quantum neural network","Enhanced Feature Map","feature map","variational model","classification","mobile data usage prediction","quantum optimization"],"falsifier":"Compute the output range of Eq. (19): for normalized $x\\in[0,1]$, the expression $2x-1.5$ covers $[-1.5,0.5]$, not the intended $[-1.5,1.5]$. Re-running the classification benchmark with the EFM exactly as written, across several random seeds on the same 1000-sample dataset, would show whether the reported 100% accuracy belongs to the circuit the design argument describes or to a circuit with an asymmetric encoding range.","tokens_in":9898,"feed_emoji":"⚛️","tokens_out":10971,"duration_ms":100384,"temperature":0.7,"pith_summary":"The paper proposes an Enhanced Quantum Neural Network whose key ingredient is a new Enhanced Feature Map (EFM) that encodes each normalized input feature with a small fixed circuit before the trainable part of the network. On a two-class task that recommends mobile data plans from two usage measurements, the paper claims all three EFM-based models reach 100% classification accuracy while the tutorial baseline reaches at most 99.9%, and they do so with 10, 13, or 16 quantum logic gates instead of the baseline's 18. The paper further claims the gains hold across three quantum optimizers and that one variant converges to a flat loss curve faster than the baseline. The paper argues this demonstrates that input encoding is a key lever for making quantum classifiers both cheaper and more accurate.","feed_headline":"EQNN hits 100% accuracy with 10-16 quantum gates","feed_subtitle":"A redesigned feature map beats an 18-gate baseline on mobile plan choice while using fewer gates and converging faster.","key_machinery":"The load-bearing object is the Enhanced Feature Map (EFM): a two-qubit encoding circuit of the form $\\mathbf{C}(q_0)(R(2x_0-1.5)\\otimes R(2x_1-1.5))(H\\otimes H)|q_1q_0\\rangle$, where $R$ is the Y-axis rotation gate and $\\mathbf{C}(q_0)$ is a CNOT controlled by $q_0$. The paper intends this five-gate circuit to place normalized features in the $[-1.5,1.5]$ range that its activation-function study favors, replacing the seven-gate ZZFeatureMap of the baseline. The variational model is the standard RealAmplitudes parameterized circuit, repeated one, two, or three times, giving total gate counts of 10, 13, and 16 respectively. The mechanism carries the argument by showing that a cheaper encoding, not a deeper ansatz, is what the accuracy gain is attributed to.","core_discovery":"The core claim, stated on the paper's own terms, is that the choice of feature map determines how quickly and how accurately a quantum neural network trains. Section III compares RY-gate circuits with three classical activation functions and concludes that inputs in the range $[-1.5,1.5]$ are especially suitable. Based on that, the paper designs the Enhanced Feature Map (EFM), a five-gate, two-qubit circuit that applies Hadamard gates, then RY rotations parameterized by an affine transform of each normalized feature, then a CNOT. With the EFM feeding one, two, or three repeats of the RealAmplitudes variational circuit, the resulting EQNN models achieve 100% classification accuracy on the mobile-data-plan task under COBYLA, SPSA, and AQGD, whereas the tutorial baseline (ZZFeatureMap plus three repeats of RealAmplitudes, 18 gates) reaches 97.9%, 99.3%, and 99.9%. The paper also reports that the 16-gate EQNN model flattens its loss curve within about five AQGD iterations, which it presents as faster convergence.","pith_inferences":["An implication left implicit is that the EFM's benefit is demonstrated only for a two-feature, two-class benchmark; extending it to more qubits would require deciding how to entangle additional features and would test whether the five-gate advantage persists.","A direct test of the design rationale would be to benchmark the circuit with a transform that genuinely maps $[0,1]$ onto $[-1.5,1.5]$ (for instance $3x-1.5$) against the written $2x-1.5$; this isolates whether the advertised value range is the actual source of the accuracy gain.","If the result is stable under re-implementation, feature-map design should be treated as a first-class hyperparameter in quantum neural network pipelines, on par with ansatz depth and optimizer choice."],"forward_implications":["A smaller feature map plus a shallower variational circuit can replace a larger standard circuit on at least one real classification task, lowering the gate count from 18 to as few as 10 gates.","The accuracy gains hold under COBYLA, SPSA, and AQGD, so the reported improvement is not tied to one optimizer's behavior.","Reaching 100% with one repeat of RealAmplitudes suggests that the bottleneck in this benchmark is the input encoding, not the depth of the trainable variational model.","Because the full-accuracy models use fewer gates, they are correspondingly cheaper to run on near-term quantum hardware, where circuit depth directly affects noise and decoherence."],"supporting_citations":[{"why":"Provides the ZZFeatureMap and RealAmplitudes operators that the baseline uses and that the proposed EQNN reuses for its variational model.","marker":"[6]"},{"why":"Defines the tutorial-based benchmark circuit (ZZFeatureMap plus three RealAmplitudes repetitions) whose 18-gate count, accuracies, and convergence curves the paper compares against.","marker":"[16]"},{"why":"The citation the paper associates with the COBYLA optimizer used in the accuracy and convergence experiments.","marker":"[10]"},{"why":"The citation the paper associates with the SPSA optimizer used in the experiments.","marker":"[11]"},{"why":"The citation the paper associates with the AQGD optimizer used in the experiments.","marker":"[12]"},{"why":"One of the two sources of the mobile data usage measurements that supply the input features x0 and x1.","marker":"[17]"},{"why":"The second source of the mobile data usage dataset, used together with [17] to build the two-class mobile-data-plan task.","marker":"[18]"}],"fun_headline_variants":["EQNN's enhanced feature map hits 100% accuracy with 16 gates","Smarter encoding: EQNN reaches 100% accuracy with fewer gates","EQNN: 100% accuracy and faster convergence via EFM design","New feature map cuts gates, boosts QNN accuracy to 100%","EQNN uses 16 gates for perfect 100% accuracy on plan tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The design assumes that the preprocessing formula places normalized inputs in the interval the author chose, but that formula actually sends $[0,1]$ to $[-1.5,0.5]$, so the trained circuit may not be the one whose design is justified.","fun_headline_variants_meta":{"raw":{"variants":["EQNN's enhanced feature map hits 100% accuracy with 16 gates","Smarter encoding: EQNN reaches 100% accuracy with fewer gates","EQNN: 100% accuracy and faster convergence via EFM design","New feature map cuts gates, boosts QNN accuracy to 100%","EQNN uses 16 gates for perfect 100% accuracy on plan tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000306,"raw_usage":{"total_tokens":1771,"prompt_tokens":977,"completion_tokens":794,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":696}},"tokens_in":593,"tokens_out":794,"duration_ms":7690,"temperature":1.0,"reasoning_tokens":696,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:32:48.225709+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the output range of Eq. (19): for normalized $x\\in[0,1]$, the expression $2x-1.5$ covers $[-1.5,0.5]$, not the intended $[-1.5,1.5]$. Re-running the classification benchmark with the EFM exactly as written, across several random seeds on the same 1000-sample dataset, would show whether the reported 100% accuracy belongs to the circuit the design argument describes or to a circuit with an asymmetric encoding range.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ZZFeatureMap and RealAmplitudes operators that the baseline uses and that the proposed EQNN reuses for its variational model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The citation the paper associates with the COBYLA optimizer used in the accuracy and convergence experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The citation the paper associates with the SPSA optimizer used in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The citation the paper associates with the AQGD optimizer used in the experiments."}],"review_version":1}