{"id":"1d4a73a0-a671-487f-a02d-37a75bf9c3d5","arxiv_id":"2507.19505","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In simulated GRB detection, quantum fingerprinting encoding consistently outperforms data reuploading, confirming that data encoding choice dominates QML performance.","lead":"This paper compares four methods for encoding astrophysical data into quantum machine learning circuits, testing them on simulated gamma-ray burst detection. It finds that encoding choice strongly affects accuracy, with a quantum fingerprinting method beating data reuploading in most settings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fingerprinting-vs-reuploading comparison is inequitable: §VI-B4 sets DR layers l=f/(2q), so DR encodes only half of each sample's features while FP encodes all; this may explain FP's accuracy edge.","rationale":"The reader selected the runtime model (50 ns per gate) as the weakest assumption. That is a valid side concern about the speed narrative, but it is not the load-bearing point for the paper's central claim. The central claim is that the choice of data encoding changes QML accuracy, and the main evidence is FP beating DR in Table V. Before that claim can be accepted, the two methods must receive the same information. The manuscript's own formulas show they do not: DR as configured in the experiments encodes only f/2 features. If DR is given all f features and still underperforms, the paper's qualitative conclusion is supported; if not, the result reduces to 'using more features helps', which is not about encoding. This is checkable from the stated equations, and the fix is straightforward. I keep the reader's CONDITIONAL verdict because the current manuscript is not acceptable as-is, but the flaw is falsifiable and repairable rather than fatal to the research direction.","tokens_in":16862,"tokens_out":9984,"duration_ms":119807,"concrete_test":"Re-run the Table V comparisons with DR layers set to l = f/q (so all features are encoded), keeping the same qubit counts, ansatz, COBYLA iterations, and 10 random initializations. If DR accuracy at f=12 remains near 66%, the encoding-effect claim survives; if it rises toward FP's 97.5%, the reported FP-over-DR result is an artifact of incomplete feature encoding, and the central claim needs re-qualification.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The core claim that data encoding choice drives QML performance rests on the FP-over-DR accuracy comparison of Table V. For that comparison to be meaningful, the two arms must differ only in encoding. They do not. Section III-C defines DR to use q features per layer and requires f = lq, with Eq. (1) and Fig. 3a showing one RY rotation per qubit per layer. Section VI-B4, however, sets l = f/(2q), and every Table V configuration follows that formula (f=12,q=6,l=1; f=24,q=6,l=2; f=120,q=6,l=10 and q=10,l=6). Consequently the DR circuit receives only q·l = f/2 of the input features, while FP uses all f features. The reported FP advantage (97.5% vs 66.1% at f=12) could therefore be an artifact of feature coverage rather than of the encoding technique itself. This is an internal inconsistency, not merely a stylistic choice, and it directly undercuts the abstract's conclusion that data encoding significantly affects QML performance.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a four-step quantum machine learning (QML) workflow—data encoding, circuit design, training, and testing—and applies it to the detection of gamma-ray burst (GRB) signals in simulated Cherenkov Telescope Array (CTAO) light curves. The authors compare several encoding techniques, focusing on quantum fingerprinting (FP) and data reuploading (DR), and report that FP achieves higher accuracy than DR, that a random fingerprinting kernel with the Pegasos classifier matches a classical RBF kernel in accuracy, and that an estimated real-hardware runtime for the FP kernel is comparable to or faster than classical ML for small training sets. The paper also documents a negative result: optimizing FP parameters does not improve over random parameter initialization.","tokens_in":17054,"tokens_out":5681,"duration_ms":66363,"significance":"If the central comparison were controlled, the paper would offer a practically useful empirical message for QML in astrophysics: encoding choice can dominate accuracy, random fingerprinting is a cheap and robust option, and quantum kernels can match classical kernels on small datasets. The strengths are the realistic simulation setup (Gammapy with CTAO instrument response functions), the systematic sweep over feature dimensions and training sizes, the explicit negative result on FP parameter optimization, and the comparison against a classical RBF baseline. However, the paper's main quantitative claims are currently weakened by a confounded encoding comparison and an unvalidated runtime model, so the significance of the results as presented is limited.","major_comments":[{"comment":"The FP-versus-DR comparison is not a controlled comparison of encoding techniques. Section III-C defines DR by choosing qubits q and layers l with f=lq, so l layers encode lq features (Eq. (1)). In the experiments, however, the authors set l=f/(2q), meaning DR encodes only f/2 of the f features, while FP encodes all f features through the superposition in Eq. (2). For f=12, q=6, l=1, DR encodes 6 features; for f=120, q=6, l=10 or q=10, l=6, DR encodes 60 features. The reported FP advantage (97.5% vs 66.1% at f=12) could therefore be an artifact of incomplete feature coverage rather than of the encoding method itself. To support the abstract's claim that data encoding significantly affects QML performance, the DR experiments should be rerun with l=f/q (or otherwise use the same number of features as FP) and the accuracy comparison repeated.","section":"§VI-B4, Table V"},{"comment":"The runtime comparison rests on an unvalidated linear model: T_sim = Nreps·d·t_sim,1 and T_real = Nreps·d·t_real,1, with t_real,1=50 ns. State-vector simulation wall time is not a product of a constant per-gate time and gate count; it depends on memory bandwidth, circuit layout, and implementation overhead. Replacing t_sim,1 with a 50 ns hardware gate time assumes no measurement, reset, control, error-correction, or noise-induced repetition overhead. This model is the sole basis for the claim that QML could be faster than classical ML for small training sets, and it should either be removed or presented explicitly as an idealized lower bound with its limitations stated. In addition, the text says 'the quantum method Treal value is smaller than the classical one up to ntr=12', but Table IV shows FP T_real=1.061 ms versus RBF 0.457 ms at ntr=12; the data support the claim only for ntr=6.","section":"§VI-B3, Table IV"}],"minor_comments":[{"comment":"There is a typo: 'inps ts' should be 'inputs' in the sentence introducing the fingerprinting protocol.","section":"§III-D"},{"comment":"References [29] and [30] appear to contain placeholder DOIs and nonstandard journal details; they should be corrected or removed.","section":"References [29] and [30]"},{"comment":"The source code is said to be 'available on request' but no repository or public link is provided; for reproducibility, the authors should make the code available in a permanent archive.","section":"§VI-B1"},{"comment":"The label 'FP' changes meaning between Table II (optimized fingerprinting) and the rest of the paper (Pegasos random fingerprinting). The captions and text should define the acronym consistently, e.g., FP-Rand for the random variant.","section":"Table II and §VI-B2"},{"comment":"The statement that DR 'has some flexibility in the number of qubits' is vague; the authors should specify the exact constraints on q and l that were used to generate the configurations in Table V.","section":"§VI-B4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central comparison is confounded and the runtime estimate is unvalidated, but the underlying workflow and simulated dataset are suitable for a solid empirical study after major revisions. I also recommend verifying references [29] and [30], whose DOIs look like placeholders; this may indicate that the reference list needs careful checking before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The main result in the abstract — \"data encoding significantly affects QML performance\" — is not supported by the paper's headline comparison, because the authors compare fingerprinting (FP) against a data-reuploading (DR) configuration that only sees half the input features. And the runtime advantage claimed for quantum execution rests on a naive 50 ns-per-gate linear extrapolation that no real hardware would satisfy.\n\nWhat's genuinely useful here is the workflow description and the negative result about FP parameter optimization: Table II shows random FP coefficients match optimized ones, and Pegasos is much faster than QSVC. That is a concrete, reproducible-in-principle finding. The Gammapy-generated GRB dataset is a sensible testbed for this kind of benchmark, and the paper explains the circuit construction carefully.\n\nThe soft spots are serious. Section III-C states that DR requires f = lq (Eq. 1, Fig. 3a). Section VI-B4 then sets l = f/(2q). As a result, every DR row in Table V encodes q·l = f/2 features, half as many as FP. The 97.5% vs 66.1% gap at f=12 is just as likely a feature-coverage effect as an encoding-technique effect. That is an internal inconsistency, not a stylistic choice, and it undercuts the central claim. The 'Classical vs Quantum ML' section also needs a strong caveat: the T_real numbers are a back-of-the-envelope product of N_reps, depth, and 50 ns, with no accounting for readout, control, or error correction. Treating milliseconds from that model as a speed advantage is premature. Additionally, references [29] and [30] have implausible DOIs and look fabricated; the C hyperparameter is tuned on the same test split later used for reporting; and neither code nor data are public (\"available on request\"), which hurts reproducibility.\n\nIf I were editing this journal, I'd send it to a serious referee — there is enough solid empirical work here to warrant a full review — but I'd expect major revisions. The authors need to either run DR with all f features (adjusting qubits per layer as the theory requires) or explain and justify the half-feature choice; tone down or heavily qualify the runtime claims; clean up the references; and release the code and data. As it stands, the central conclusion does not survive the FP-vs-DR analysis, and the rest is an incremental but potentially useful benchmark.\n\nWho should read it? People actively working on QML for astrophysics, as a cautionary example of how easy it is to make an unequal comparison. I would not cite it as evidence that encoding choice matters.","headline":"The FP-vs-DR comparison that drives the abstract is unfair (DR sees half the input features), so the central claim is unsupported; still, enough solid empirical work exists to merit a serious referee.","tokens_in":17618,"tokens_out":3483,"would_cite":false,"duration_ms":39206,"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":"Data encoding, not just circuit size, decides how well quantum ML detects gamma-ray bursts.","keywords":["quantum machine learning","data encoding","quantum fingerprinting","data reuploading","gamma-ray bursts","quantum kernels","quantum neural networks","astrophysical signal detection"],"falsifier":"Train both fingerprinting and data reuploading models on the same 12-feature GRB light curves but run them on a real noisy quantum processor, measuring wall-clock time per inference and accuracy; if fingerprinting's accuracy drops below the classical RBF baseline or the real time exceeds the estimated $N_{\\mathrm{reps}}\\cdot d\\cdot 50\\,\\mathrm{ns}$, the paper's central encoding and speed claims would be contradicted.","tokens_in":16637,"feed_emoji":"📡","tokens_out":8785,"duration_ms":95960,"temperature":0.7,"pith_summary":"This paper tries to establish that the way classical astrophysical data is mapped into quantum states—the encoding step—is the main factor controlling how well quantum machine learning models perform, and it tests this on simulated gamma-ray burst light curves. The authors lay out a four-step workflow of data encoding, circuit design, training, and testing, then compare amplitude, angle, data reuploading, and quantum fingerprinting encodings. Their experiments show that quantum fingerprinting with randomly chosen parameters reaches roughly 97.5% accuracy on the 12-feature GRB task and matches a classical support-vector classifier from just 12 training samples, while data reuploading stays near chance until about 60 samples. If the paper is right, encoding design should be the first thing a QML practitioner tunes, and quantum methods can be competitive where training data are scarce.","feed_headline":"Quantum fingerprinting beats data reuploading on gamma-ray burst data","feed_subtitle":"Random fingerprint encoding matches classical accuracy from just 12 training samples; data reuploading needs 60.","key_machinery":"The load-bearing objects are two data-encoding circuits plus the four-step workflow that frames them. Quantum fingerprinting encoding maps each feature $x_{i,j}$ into a separate subspace, producing $|\\psi(x_i)\\rangle = \\frac{1}{\\sqrt{f}}\\sum_{j=1}^f |j\\rangle(\\cos(\\theta_j x_{i,j})|0\\rangle + \\sin(\\theta_j x_{i,j})|1\\rangle)$, and compresses a long input into $\\lceil\\log f\\rceil+1$ qubits. Data reuploading interleaves single-qubit rotations with entangling layers so the same input enters the circuit repeatedly, giving the model a partial Fourier-series structure whose frequencies are set by the encoding gates. The comparison of these two encodings, with the same ansatz and optimizer, is what carries the argument that encoding choice controls performance.","core_discovery":"The central claim is that data encoding significantly affects QML model performance, and in the GRB case study fingerprinting encoding can achieve better accuracy than data reuploading. On the 12-feature simulated light curves, fingerprinting with random coefficients reaches 97.5% accuracy and matches a classical RBF kernel from 12 training samples onward, whereas data reuploading stays near 50% accuracy until the training set has at least 60 samples. The paper also argues that, under a noiseless model where each gate costs a constant 50 ns, the estimated real-hardware runtime of the quantum circuits is comparable to or shorter than the classical RBF kernel for small training sets, suggesting quantum methods are most promising in the small-data regime.","pith_inferences":["If encoding, not circuit depth, is the dominant lever, then QML benchmarks that vary only the ansatz while keeping one encoding fixed may miss the main design choice; a natural follow-up is an encoding-by-ansatz grid on the same GRB data.","The success of random fingerprinting coefficients suggests the subspace structure of the encoding, rather than the specific rotation angles, carries the discriminative information; this could be tested by comparing fingerprinting against amplitude and angle encodings at equal qubit counts.","Fingerprinting is a compression technique, so the same workflow could be tried on other sparse astrophysical transients, such as supernova or fast radio burst light curves, which the paper does not test.","The timing model ignores noise and readout overhead, so a direct hardware run could change the speed ranking; that is the paper's unstated boundary condition."],"forward_implications":["Encoding choice should be treated as a first-class hyperparameter in QML pipelines for astrophysical data, since the same ansatz and optimizer give very different results depending on how data is encoded.","Quantum fingerprinting with random coefficients is a practical default: it reaches about 97.5% accuracy on the 12-feature GRB dataset while avoiding hours of parameter optimization, and with Pegasos training takes minutes rather than hours.","Data reuploading is not competitive for very small training sets: below about 60 training samples it stays near chance accuracy on this task, so it needs more data or a different setup.","Quantum kernel methods with fingerprinting encoding can match a classical RBF kernel in accuracy once 12 training samples are available, so the quantum approach does not sacrifice accuracy in the small-data regime.","The estimated hardware runtime of the tested quantum circuits is comparable to or smaller than classical RBF up to $n_{tr}=12$, so the case for QML is strongest when training data are scarce."],"supporting_citations":[{"why":"Supplies the preliminary GRB dataset and the quantum convolutional benchmark this paper extends to a general workflow.","marker":"[7]"},{"why":"Defines data reuploading encoding and motivates it as a universal classifier, the method compared against fingerprinting.","marker":"[38]"},{"why":"Shows that data reuploading circuits act as partial Fourier series, grounding the theoretical description of DR encoding.","marker":"[39]"},{"why":"Introduces quantum fingerprinting, the compression idea behind the FP encoding.","marker":"[40]"},{"why":"Provides the two fingerprinting encoding variants (optimized and random coefficients) used for FP and FP-Random.","marker":"[41]"},{"why":"Supplies the ansatz structure used for both quantum neural networks in the comparison.","marker":"[46]"},{"why":"Describes the Pegasos solver that makes random fingerprinting training fast in the experiments.","marker":"[49]"},{"why":"Generates the simulated gamma-ray light curves and instrument response functions used as the testbed.","marker":"[52]"},{"why":"Provides the 20–40 ns single-gate time used to estimate real-hardware runtime from simulation time.","marker":"[57]"}],"fun_headline_variants":["Quantum fingerprinting wins over data reuploading in GRB detection","Fingerprint encoding outperforms reuploading for quantum GRB analysis","Quantum ML: fingerprint encoding needs far fewer samples on GRBs","Small-data quantum edge: fingerprint encoding on gamma-ray bursts","Fingerprint beats reuploading: quantum signal detection with less data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed comparison rests on assuming that executing a circuit on real hardware takes exactly the number of repetitions times the circuit depth times a constant 50 ns per gate, with no noise, readout, or control overhead; if those overheads are significant, the claimed quantum speed advantage for small training sets disappears.","fun_headline_variants_meta":{"raw":{"variants":["Quantum fingerprinting wins over data reuploading in GRB detection","Fingerprint encoding outperforms reuploading for quantum GRB analysis","Quantum ML: fingerprint encoding needs far fewer samples on GRBs","Small-data quantum edge: fingerprint encoding on gamma-ray bursts","Fingerprint beats reuploading: quantum signal detection with less data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1472,"prompt_tokens":860,"completion_tokens":612,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":522}},"tokens_in":476,"tokens_out":612,"duration_ms":6908,"temperature":1.0,"reasoning_tokens":522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:31:15.094538+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train both fingerprinting and data reuploading models on the same 12-feature GRB light curves but run them on a real noisy quantum processor, measuring wall-clock time per inference and accuracy; if fingerprinting's accuracy drops below the classical RBF baseline or the real time exceeds the estimated $N_{\\mathrm{reps}}\\cdot d\\cdot 50\\,\\mathrm{ns}$, the paper's central encoding and speed claims would be contradicted.","supporting_citations":[{"cited_title":"Benchmarking Quantum Convolutional Neural Networks for Signal Classification in Simulated Gamma-Ray Burst Detection","cited_arxiv_id":"2501.17041","evidence_quote":"Supplies the preliminary GRB dataset and the quantum convolutional benchmark this paper extends to a general workflow."},{"cited_title":"Schuld and F","cited_arxiv_id":null,"evidence_quote":"Introduces quantum fingerprinting, the compression idea behind the FP encoding."},{"cited_title":"Quantum Fingerprinting and Quantum Hashing. Computational and Cryptographical Aspects,","cited_arxiv_id":null,"evidence_quote":"Provides the two fingerprinting encoding variants (optimized and random coefficients) used for FP and FP-Random."},{"cited_title":"Efficient implementation of amplitude form of quantum hashing using state-of-the-art quantum processors,","cited_arxiv_id":null,"evidence_quote":"Supplies the ansatz structure used for both quantum neural networks in the comparison."},{"cited_title":"Ex- pressibility and entangling capability of parameter- ized quantum circuits for hybrid quantum-classical al- gorithms,","cited_arxiv_id":null,"evidence_quote":"Describes the Pegasos solver that makes random fingerprinting training fast in the experiments."}],"review_version":1}