{"id":"30fdd436-24fb-4542-a3ce-a08dec0b5e22","arxiv_id":"2411.16970","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"On a synthetic financial dataset, one-class SVMs using projected quantum kernels achieved higher mean F1 scores than a classical rbf-kernel baseline at every tested anomaly ratio, both in simulation and on quantum hardware, though the statistical evidence is weak.","lead":"The authors tested whether quantum-computer-based kernels can detect credit-card fraud better than a classical machine-learning method. They simulated the algorithm and ran it on five different quantum processors, finding a modest but consistent F1-score improvement over the classical baseline in simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fixed hyperparameters chosen at r=10^-2 and unreported feature scaling likely under-tune the classical rbf baseline; with only one split for most r and 65-75% win probabilities at r=0.2, the 'all anomaly regimes' claim is not yet established.","rationale":"The reader identified the same weak point, and I agree it is the load-bearing one. The paper's central claim is comparative and empirical: quantum-enhanced OCSVMs generalize better than classical rbf across all tested anomaly ratios. Any such comparison is only meaningful if both models are given equivalent tuning opportunities. The fixed hyperparameter pair breaks that symmetry because the two kernels operate in different feature spaces with different distance scales, and because the optimal RBF bandwidth depends on the anomaly ratio via dataset size and geometry. The fact that the grid was run only at r=10^-2 and then frozen makes it especially hard to rule out a tuned classical baseline closing the gap. The statistical evidence is a further part of the same concern: a single split per r cannot justify 'all anomaly regimes', and the only multi-split estimate (r=0.2) gives a 65-75% probability, not a decisive advantage. A per-regime nested tuning protocol is the natural, low-cost way to settle this; if the advantage survives, the paper's claim is much stronger. I would not reject the paper: the experiments are useful and honestly reported, and the hardware results are a bonus. But the headline claim should be conditional on a fairer hyperparameter protocol and more splits. Note that I am not raising classical simulability as a separate fatal objection: the paper does not claim computational speedup, though the phrase 'quantum-enhanced' would be clearer if the comparison included an rbf kernel on the same RDM features.","tokens_in":13104,"tokens_out":10437,"duration_ms":105050,"concrete_test":"Run a controlled benchmark where, for each anomaly ratio r in {0.01, 0.02, 0.05, 0.1, 0.2}, (nu, gamma) are chosen independently for rbf-OCSVM and for each qrbf-OCSVM via 5-fold cross-validated grid search on each of 10 train/test splits, reporting best-mean-F1 and paired 95% CIs. Include the rbf-OCSVM trained on the same 1Q-RDM feature vectors as a third arm. If the best-tuned rbf-OCSVM (on raw or RDM features) matches or beats the best qrbf-OCSVM at any r, the 'all anomaly regimes' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the fairness of the comparison between qrbf-OCSVM and rbf-OCSVM. In Sec. III.A the authors fix (nu,gamma)=(0.1,0.1) for every anomaly ratio r and for both kernel types, where this pair was selected by one grid search at r=10^-2 across all feature maps jointly. This is not a controlled comparison. The qrbf kernel (Eq. 12) is an RBF on 40 bounded RDM coordinates, while the classical rbf is an RBF on 20 raw features with heterogeneous scales (dollar, lat/long, counts); no standardization is reported in App. B. The same gamma can therefore be appropriate for one kernel and pathological for the other. Since RBF bandwidth and nu should scale with dataset size, feature-space dimension, and pairwise distance distribution, the classical baseline may be under-tuned at r>0.01. The evidence at other r is also thin: apart from r=0.2 (10 splits, Table 1, Pr=64.8-74.7%), the 'all anomaly regimes' claim is supported by a single train/test split per r. The abstract's strongest claim therefore outruns the controlled evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an unsupervised anomaly detection pipeline based on one-class SVM with projected quantum kernels (qrbf) defined via one-qubit reduced density matrices. Using a synthetic credit card fraud dataset, the authors compare qrbf-OCSVM with classical RBF-OCSVM across anomaly ratios in simulation, and test the qrbf models on trapped-ion and superconducting quantum processors. They report that quantum-enhanced models achieve better F1 scores in all anomaly regimes and that the advantage persists on hardware.","tokens_in":13371,"tokens_out":10525,"duration_ms":89201,"significance":"If the claims are established, the paper would provide a valuable empirical benchmark for projected quantum kernels in a realistic unsupervised learning task, including one of the first multi-platform hardware comparisons. The systematic treatment of partial state tomography and the explicit acknowledgement of hardware-related limitations are commendable. However, the statistical methodology for the central comparison is flawed, and the experimental evidence is too limited to support the strong claims in the abstract.","major_comments":[{"comment":"Equation (14) is mathematically inconsistent: with the normalization constant N as defined, the right-hand side equals 1 for any μ̃ and σ̃, because the integral of the Gaussian numerator over [0,1] is exactly √(2πσ̃²)[Φ((1−μ̃)/σ̃)−Φ(−μ̃/σ̃)]. The reported probabilities in Table 1 therefore cannot stem from this formula. Moreover, the computation uses only 10 data splits, and no confidence intervals or hypothesis tests are provided. The evidence for the superiority of qrbf-OCSVM over rbf-OCSVM at r=0.2 is thus not statistically quantified as claimed.","section":"Section III.A, Eq. (14) and Table 1"},{"comment":"The comparison between qrbf-OCSVM and rbf-OCSVM is not controlled. The hyperparameters (ν,γ)=(0.1,0.1) are selected from a single grid search at r=10⁻² (Sec. III.A) and applied to every anomaly ratio and both kernel types. The two kernels operate on different feature spaces: the qrbf kernel uses the Frobenius norm of 1Q-RDM differences (Eq. 12), which gives a bounded, roughly homogeneous coordinate set, whereas the classical rbf kernel uses the raw 20 features listed in Table 3, which include dollar amounts, latitudes, longitudes, and counts on very different scales. No feature standardization is reported in Appendix B. The same γ can therefore be appropriate for one kernel and pathological for the other. The authors should tune kernels separately per r (or justify the common γ), and should report whether the classical baseline's performance changes with scaling.","section":"Section III.A, hyperparameter selection and feature scaling"},{"comment":"For all anomaly ratios except r=0.2, the claimed advantage is based on a single train/test split (Figure 1). No error bars or repeated-split statistics are given for these regimes, so the 'all anomaly regimes' claim is not supported by the data. The paper should either provide repeated splits or cross-validation for every r, or limit the claim to the regime where statistical support exists.","section":"Section III.A, Figure 1"},{"comment":"The abstract states that the quantum-enhanced OCSVMs lead to better generalization 'in all anomaly regimes' and that results were 'experimentally validated'. However, the hardware experiments (Sec. III.C) cover only r=0.2, with a single run per platform, and the text explicitly notes that 'this is just one experiment execution ... we can't properly evaluate the models.' The abstract should be aligned with the actual scope of the evidence.","section":"Abstract and Section III.C"}],"minor_comments":[{"comment":"The phrase 'as we we do not optimize' contains a duplicated word 'we' that should be corrected.","section":"Section II.B"},{"comment":"The sentence 'Further decreasing r the F1 score saturates towards F1≈ 0.95.1.' contains a stray '1' at the end.","section":"Section III.A"},{"comment":"The caption refers to 'operator 2-norm differences (cf. Eq. (16))', but Eq. (16) defines the Frobenius norm, not the operator 2-norm.","section":"Figure 2 caption"},{"comment":"No statement is provided about data or code availability; making the implementation and the split-specific results publicly available would strengthen reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The statistical flaw in Eq. (14) is serious and should be corrected before publication. The hyperparameter selection issue also risks biasing the central comparison. I suggest the editor ask the authors to re-analyze the data with a valid probability calculation, to tune the classical baseline fairly, and to temper the abstract's claims in line with the actual evidence. The paper is within the journal's scope but needs major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is a genuine empirical benchmark: OCSVM with projected quantum kernels (qrbf) against classical rbf on a credit-card transaction dataset, across anomaly ratios and validated on several QPUs. The systematic scan over anomaly ratios is new, and the partial-tomography precision study in Sec. III.B is a practical contribution. The authors are also honest about limits: they note the hardware results are single runs and that further statistical analysis is needed.\n\nThe soft spot is the central claim. The hyperparameters nu=0.1 and gamma=0.1 were picked at r=1e-2 and then held fixed for all anomaly regimes and both kernels. The qrbf kernel acts on bounded 1Q-RDM coordinates while the classical rbf sees raw features with wildly different scales (dollar, lat/long, counts) — no standardization is reported. So the comparison may not be apples-to-apples. The evidence at other r values rests on a single train/test split; only r=0.2 gets 10 splits, and the win probabilities there are 65–75%. The abstract's claim that quantum-enhanced OCSVMs lead to \"better generalization properties\" in all anomaly regimes therefore outruns the data. This is a fixable empirical weakness rather than a conceptual error.\n\nWhat I liked: the tomography results quantify the calibration-drift limit, which is genuinely useful for anyone planning QPU experiments. The feature engineering details are also transparent.\n\nWho is this for? Groups working on quantum kernel benchmarking or QML applications with small data will want to read it. It deserves a serious referee: the question is meaningful, the method is standard, and the limitations are visible. My recommendation would be to engage with it in review, but explicitly require per-regime hyperparameter tuning (or at least a sensitivity analysis), more cross-validation splits, and reporting of the feature scaling used for the classical baseline.\n\nBest.","headline":"A useful empirical benchmark of qrbf-OCSVM on financial transaction data, but the 'all anomaly regimes' claim is under-powered and the classical baseline may be under-tuned.","tokens_in":13929,"tokens_out":2164,"would_cite":false,"duration_ms":20327,"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":"Quantum kernel one-class SVMs beat the classical baseline on every fraud ratio in a credit-card dataset, in simulation and on noisy hardware.","keywords":["quantum machine learning","one-class support vector machine","anomaly detection","quantum kernels","projected quantum kernels","partial state tomography","financial fraud detection","NISQ hardware"],"falsifier":"Run the identical dataset and splits but perform a grid search over gamma and nu separately at each anomaly ratio for both the classical rbf-OCSVM and the qrbf-OCSVM; if any classical hyperparameter setting yields a test F1 equal to or higher than the best quantum model, the paper's all-regime generalization claim fails on this dataset.","tokens_in":12915,"feed_emoji":"⚛️","tokens_out":9012,"duration_ms":75230,"temperature":0.7,"pith_summary":"The paper sets out to show that a one-class support vector machine (OCSVM) equipped with a projected quantum radial basis function kernel generalizes better than the classical rbf-OCSVM on a realistic financial fraud-detection dataset, across every anomaly ratio tested. The authors simulate the quantum kernels from one-qubit reduced density matrices and then estimate them on trapped-ion and superconducting quantum processors using partial state tomography, finding that the quantum model's F1 advantage persists on hardware, with fluctuations dominated by calibration drift rather than finite sampling. The result matters because it is a concrete, unsupervised use case where quantum kernel methods appear to deliver better generalization in a small-data, high-anomaly regime on today's noisy devices.","feed_headline":"Quantum OCSVM beats classical SVM at every fraud ratio","feed_subtitle":"Projected-quantum-kernel SVMs generalize better across all anomaly regimes, even on noisy hardware.","key_machinery":"The central object is the projected quantum radial basis function kernel, defined as $K_\\gamma(x_i,x_j) = \\exp\\left(-\\gamma \\sum_{k=1}^N \\|\\rho_k(x_i)-\\rho_k(x_j)\\|_F^2\\right)$, where $\\rho_k$ is the one-qubit reduced density matrix of the kth qubit after encoding data point $x$ into a hardware-efficient feature map. These 1Q-RDMs are estimated by partial state tomography using Pauli basis measurements (classical shadows), which requires only single-qubit rotations and is compatible with current gate-based hardware. The kernel is then used in the standard OCSVM dual problem. The tomography protocol is the workhorse of the paper: its precision on real devices, limited by shot count and calibration drift, determines how faithfully the quantum kernel approximates the noise-free one.","core_discovery":"On the 20-dimensional feature-engineered Sparkov credit-card dataset, the paper finds that the qrbf-OCSVM achieves higher test F1 scores than the classical rbf-OCSVM for all anomaly ratios from r=0.01 to r=0.2, with estimated probabilities of beating the classical model above 64% at the highest anomaly ratio. The same ordering is observed in a single hardware execution at r=0.2 on ion-trap and superconducting QPUs, where the quantum models' F1 scores hover around the noise-free simulation values and sometimes exceed them on the test split; the authors explicitly caution that a single execution does not permit rigorous statistical evaluation. They attribute the advantage to the structure of the projected quantum feature map, since finite-sampling variance is small, and they identify calibration drift during long experiments as the main source of hardware-induced fluctuation.","pith_inferences":["The comparison fixes one hyperparameter pair (nu=0.1, gamma=0.1) for all anomaly ratios and both kernels; since the pair was selected at r=10^-2, a per-ratio grid search for the classical baseline could close or reverse the gap, so the universal-superiority claim is conditional on this baseline choice.","The observed advantage may be tied to the synthetic dataset's structure, including step-function fraud rates and strong categorical signals; re-running the protocol on real transaction data or on datasets with different correlation geometry would delimit how generic the gain is.","Because the 1Q-RDMs are measured once and the kernel is constructed a posteriori, the same tomography data supports hyperparameter search over gamma without rerunning the circuits, which is a practical cost advantage for deployment.","The calibration-drift result implies that reducing wall-clock experiment time—by measurement-efficient tomography or by less circuit repetition—may matter more than incremental improvements in gate fidelity for quantum kernel methods on current hardware."],"forward_implications":["On the tested dataset, the quantum kernel's test F1 advantage holds at every anomaly ratio, so the benefit appears across the whole operating range rather than at one tuned point.","The advantage is present in noise-free simulation, so it originates from the quantum feature map's geometry, not from hardware noise or sampling artifacts.","Real-hardware F1 scores stay close to simulation, meaning current NISQ devices can run the full pipeline without destroying the quantum kernel's signal.","Tomography precision saturates beyond roughly 10^3 shots on the superconducting QPU, and total experiment duration matters because calibration drifts; shorter, faster experiments are a practical requirement for reliable quantum kernels.","Because the OCSVM is unsupervised and the projected kernel uses only single-qubit reduced density matrices, the approach is directly applicable to other rare-event detection problems, such as new physics searches."],"supporting_citations":[{"why":"Defines the one-class SVM that the paper extends with a quantum kernel; provides the decision function and the nu-property.","marker":"[7]"},{"why":"Introduces projected quantum kernels built from reduced density matrices, the basis of the qrbf kernel.","marker":"[14]"},{"why":"Documents exponential concentration of quantum kernels, the overfitting risk that motivates projecting back to classical space.","marker":"[15]"},{"why":"Supplies the classical-shadow partial tomography protocol used to estimate one-qubit reduced density matrices on hardware.","marker":"[45]"},{"why":"Source of the synthetic credit-card transaction dataset used in all anomaly detection experiments.","marker":"[44]"},{"why":"The generator that produced the dataset, controlling its fraud structure and label balance.","marker":"[49]"},{"why":"Provides the few-training-data generalization theory the paper invokes to explain the small-data quantum advantage.","marker":"[50]"},{"why":"The LIBSVM solver used to train the OCSVM dual problem in practice.","marker":"[48]"}],"fun_headline_variants":["Quantum OCSVM beats classical at every fraud ratio","Fraud detection improves with quantum kernels on noisy QPUs","Projected quantum kernels top classical SVM in all anomaly regimes","Noisy quantum processors still deliver quantum advantage in fraud detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on comparing both models under a single fixed hyperparameter pair (nu=0.1, gamma=0.1) chosen by one grid search at one anomaly ratio; if the classical rbf kernel were tuned at every anomaly ratio, it might match or beat the quantum model.","fun_headline_variants_meta":{"raw":{"variants":["Quantum OCSVM beats classical at every fraud ratio","Fraud detection improves with quantum kernels on noisy QPUs","Projected quantum kernels top classical SVM in all anomaly regimes","Noisy quantum processors still deliver quantum advantage in fraud detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1692,"prompt_tokens":887,"completion_tokens":805,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":738}},"tokens_in":503,"tokens_out":805,"duration_ms":7118,"temperature":1.0,"reasoning_tokens":738,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:40:42.448270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical dataset and splits but perform a grid search over gamma and nu separately at each anomaly ratio for both the classical rbf-OCSVM and the qrbf-OCSVM; if any classical hyperparameter setting yields a test F1 equal to or higher than the best quantum model, the paper's all-regime generalization claim fails on this dataset.","supporting_citations":[{"cited_title":"LIBSVM: a li- brary for support vector machines","cited_arxiv_id":null,"evidence_quote":"The LIBSVM solver used to train the OCSVM dual problem in practice."},{"cited_title":"Estimating the support of a high-dimensional distribution","cited_arxiv_id":null,"evidence_quote":"Defines the one-class SVM that the paper extends with a quantum kernel; provides the decision function and the nu-property."},{"cited_title":"Power of data in quantum ma- chine learning","cited_arxiv_id":null,"evidence_quote":"Introduces projected quantum kernels built from reduced density matrices, the basis of the qrbf kernel."},{"cited_title":"Predicting many properties of a quantum system from very few measurements","cited_arxiv_id":null,"evidence_quote":"Supplies the classical-shadow partial tomography protocol used to estimate one-qubit reduced density matrices on hardware."},{"cited_title":"Credit Card Transactions Fraud Detec- tion Dataset: Simulated Credit Card Transactions gener- ated using Sparkov","cited_arxiv_id":null,"evidence_quote":"Source of the synthetic credit-card transaction dataset used in all anomaly detection experiments."},{"cited_title":"Sparkov Data Generation: Generate Fake Credit Card Transaction Data, Including Fraudulent Transactions","cited_arxiv_id":null,"evidence_quote":"The generator that produced the dataset, controlling its fraud structure and label balance."},{"cited_title":"Generalization in quantum ma- chine learning from few training data","cited_arxiv_id":null,"evidence_quote":"Provides the few-training-data generalization theory the paper invokes to explain the small-data quantum advantage."}],"review_version":1}