{"id":"2d063e3d-a34f-4f03-bf01-34df1c1d08cd","arxiv_id":"2607.13847","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"QTDE encodes higher-order topological structure into quantum states via evolution under the combinatorial Laplacian; on clique-complex benchmarks it edges out a Laplacian-comparison baseline only in easy, high-dimensional regimes, with QSVT-filter gains largely fitted.","lead":"This work proposes turning the 'shape' of a dataset—encoded as a topological Laplacian matrix—into a quantum state by letting the state evolve under that matrix, then using these quantum states as features for machine learning. It tests whether this quantum representation beats simply comparing the matrices, on synthetic random-graph data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unreported evolution time t (Eqs. 7, 9, 14) is the load-bearing assumption: with t only 'some sufficiently large choice', the claimed consistent outperformance could be fitting; the paper's own Sec. V.D admits comparable selection bias in the QSVT sweep.","rationale":"The paper proposes a coherent and mathematically sound framework: combinatorial Laplacians are Hermitian, e^{-iL_k t} is unitary, the fidelity kernel is positive semidefinite by the Schur product theorem, and the survival features are legitimate spectral fingerprints. The proofs and derivations are internally consistent, and the authors explicitly acknowledge several limitations (small benchmarks, no formal guarantees, computational cost). However, the central empirical claim — that QTDE 'consistently outperform[s]' the Laplacian-comparison baseline — is not adequately supported because the evolution time t, which is a free parameter of every representation, is never reported nor justified. Eq. (7), Eq. (9), and Eq. (14) all depend on t; without a specified time value or a principled selection rule, the reported accuracies cannot be reproduced or audited, and the comparison to a baseline with no analogous free parameter is unfair. The paper's own Section V.D demonstrates that per-dimension selection over a family of QSVT polynomials inflates reported values and that the residual gains fall within cross-validation spread; this is precisely the risk for an unreported t choice. My concern is therefore about the empirical comparison, not the mathematical construction. The reader's conditional verdict already captures this: revisions should require reporting t and a sensitivity/nested-CV analysis. Thus no change to the verdict is needed; the paper should be accepted only if these conditions are met.","tokens_in":17723,"tokens_out":9641,"duration_ms":95063,"concrete_test":"Obtain the public repository (github.com/DimitThanos/QTDE) and extract the exact t values (and {t_j} grids) used for Figures 2 and 5. Then rerun the benchmark for at least two representative tasks (e.g., G(50,0.60) vs G(50,0.70) and G(50,0.5) vs G(50,0.52)) across a t grid such as t ∈ {0.1, 0.5, 1, 2, 5, 10, 20}, selecting t on training folds only via nested 5-fold CV (or fixing t a priori), and compare QTDE accuracy against the Frobenius baseline for each k. If the QTDE advantage disappears, changes sign, or is not outside fold-to-fold variance under this protocol, then the reported 'consistent' outperformance is an artifact of a favorable, unreported t choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — QTDE consistently beats the Frobenius-Laplacian baseline — stands or falls on the choice of evolution time t. Eq. (7) defines |Ψ_k(X)> = e^{iL_k t}|s>, and both Eq. (9) (fidelity kernel) and Eq. (14) (survival features) depend on t and on the time grid {t_j}; Sec. III.B specifies only 'some sufficiently large choice of t' and no t values appear anywhere in the text or figure captions. Since F(t) = |<s|e^{-iL_1 t}e^{iL_2 t}|s>|^2 is quasiperiodic in t, different t values can change which classes appear similar; an unreported t selected per dataset or per k is indistinguishable from fitting to the validation labels. The baseline has no analogous free parameter: its bandwidth γ is set by the median heuristic (Sec. IV.D). The paper's own Sec. V.D states that taking the maximum accuracy over polynomial filters 'inflates the reported values' and that 'residual advantages fall within the cross-validation spread' — the same selection risk applies to choosing t from the data. Without reporting t and without a sensitivity or nested-cross-validation analysis, the central empirical claim is underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Quantum Topological Data Encoding (QTDE), a framework that maps a simplicial complex to a quantum state by evolving a uniform superposition over k-simplices under the unitary generated by the combinatorial Laplacian L_k. Two readouts are proposed: an implicit fidelity kernel between evolved states and an explicit feature vector formed from survival amplitudes sampled at multiple times. A generalization based on QSVT replaces the exponential by polynomial spectral filters. The method is evaluated on binary classification of clique complexes of Erdős–Rényi graphs, against a Gaussian kernel on Frobenius distances between Laplacians. The paper claims that the quantum representations 'consistently outperform' this classical baseline and that higher-dimensional Laplacians provide discriminative information not available from direct matrix comparisons.","tokens_in":18073,"tokens_out":4018,"duration_ms":39923,"significance":"If the empirical claims were fully supported, QTDE would be a useful proof-of-concept for topology-driven quantum encodings: the fidelity kernel is valid (PSD by the Schur product theorem), the survival features are a compact spectral fingerprint, and the extension to QSVT is technically natural. The manuscript also includes a welcome degree of self-criticism, especially in Sec. V.D and Sec. VII, and promises a public code repository, which would aid reproducibility. However, the central advantage claim currently rests on unreported evolution times and on per-dimension best-filter selection, both of which can turn a fitted curve into an apparent prediction. The paper's own admission that the QSVT advantages fall within the cross-validation spread further weakens the 'consistently outperform' statement. The contribution is therefore a promising framework whose empirical validation is not yet complete.","major_comments":[{"comment":"The evolution time t and the time grid {t_j} are never reported. The text only states that t is 'some sufficiently large choice' (Sec. III.B). Both the fidelity kernel (Eq. 9) and the survival features (Eq. 14) depend on t and on the sampled times. Since quantities like |<s|e^{-iL_1 t}e^{iL_2 t}|s>|^2 are quasiperiodic in t, different t values can change which classes appear similar. If t was chosen per dataset or per dimension to maximize cross-validated accuracy, the reported advantage is fitted rather than predicted. Please report the exact t and grid used for every experiment, and provide either a sensitivity analysis over t or a nested cross-validation procedure that selects t on the training folds only. Without this, the central empirical claim is underdetermined.","section":"Sec. III.B, Eq. (7) and Eq. (14); Sec. IV.C, Eq. (16)"},{"comment":"The abstract and Sec. VII state that QTDE 'consistently outperform[s]' the Laplacian-comparison baseline, but Sec. V.A reports that on the two hard ensembles, G(30,0.78 vs 0.80) and G(50,0.69 vs 0.70), the fidelity kernel, survival features, and matrix-difference baseline interleave within the 0.5–0.7 band and the separations are comparable to the cross-validation spread. The visible advantage occurs mainly on the easily separable G(50,0.60 vs 0.70) panel at higher k. The claim should be qualified to match the evidence, e.g., 'in some regimes', and supported by a statistical comparison across folds or a paired significance test rather than by inspection of overlapping error bars.","section":"Sec. V.A and Abstract"},{"comment":"The QSVT 'best poly' curves in Fig. 4(a) are obtained by selecting, at each dimension k, the best-performing polynomial from a fixed library. Sec. V.D explicitly acknowledges that taking the maximum over the family 'inflates the reported values' and that the residual advantages 'fall within the cross-validation spread'. This is a selection-bias problem of exactly the kind that affects the choice of t. To substantiate the Sec. V.C claim that polynomial filters 'sit consistently above' the exponential evolution and the baseline, the evaluation must select the polynomial within the training folds (nested CV) or pre-register the polynomial family, and report the resulting distribution rather than the per-dimension maximum.","section":"Sec. V.C and Sec. V.D"},{"comment":"The zero-padding procedure for complexes with different numbers of k-simplices makes the representation dataset-dependent. The states are embedded in a common Hilbert space whose dimension is determined by the union of simplices across the whole dataset, so the kernel value K_k(X1,X2) in Eq. (9) is not a function of the two data points alone but depends on which other graphs happen to be in the sample. The Frobenius baseline (Eq. 17) shares the same dataset-dependent common basis, so the comparison is fair, but the manuscript should acknowledge this dependence and test its stability, for example by subsampling the dataset and checking whether the kernel geometry and classification results change.","section":"Sec. III.B1"}],"minor_comments":[{"comment":"The phrase 'consistently outperform' is stronger than the evidence in Sec. V.A and should be softened to reflect the regimes in which the advantage actually appears.","section":"Abstract / Sec. VII"},{"comment":"The sign convention for the evolution operator is inconsistent: Eq. (5) defines U_k(t)=e^{iL_k t} and Eq. (7) uses e^{iL_k t}|s>, while the introductory Eq. (1) and the kernel Eq. (9) write e^{-iL_k t}. The states are related by time reversal, but the inconsistency should be fixed for clarity.","section":"Sec. II.C and Sec. III.B"},{"comment":"The caption says 'cyan lines indicate the class boundary' while the Appendix captions for Figs. 6–7 refer to a 'cyan cross'. Please make the marker description consistent.","section":"Fig. 3 caption"},{"comment":"The normalization constant c_k = max_G(max diag(L_k(G))) is described verbally but the actual values are not reported. Since all Laplacians are scaled by this dataset-wide constant, reporting c_k for each benchmark would aid reproducibility.","section":"Sec. IV.C"}],"recommendation":"major_revision","confidential_remarks":"The paper contains an unusually candid admission of its own selection bias in Sec. V.D, which I take as a sign of good faith rather than as a reason for rejection. The fixable issue is that the abstract and Sec. VII do not carry that caveat. In addition to the required reporting of t and the QSVT selection procedure, I would encourage the editor to ask for the code repository to be made available at revision time, since the numerical comparisons are central to the paper's contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clean, honest incremental extension of Henry et al.'s quantum evolution kernel to higher-order combinatorial Laplacians. The math is sound: L_k is Hermitian, e^{-iL_k t} is unitary, and the fidelity kernel is PSD via the Schur product theorem. The genuinely new bits are the use of L_k with k>0 as the generator, the survival-feature readout, and the QSVT polynomial filter extension. As a proof of concept on clique complexes of Erdős–Rényi graphs, it is a reasonable contribution.\n\nWhere it gets soft: the abstract says the representations 'consistently outperform' the Laplacian baseline, but Figure 2's two hard ensembles show all methods interleaved within cross-validation spread. The only clear win is the easy, separable panel at high k. That overstatement should be fixed before publication.\n\nMore importantly, the evolution time t is never reported. Equation (7), the fidelity kernel (9), and the survival features (14) all depend on t and the time grid, and the text only says 'some sufficiently large choice of t.' The stress-test concern holds: F(t) is quasiperiodic, and different t values can change which classes look similar. Without reporting t or a sensitivity analysis, the central empirical claim is underdetermined—we cannot distinguish a genuine advantage from a fitted hyperparameter. The QSVT section has the same selection problem, and the authors actually admit it in Sec. V.D: taking the maximum over polynomial filters 'inflates the reported values' and 'residual advantages fall within the cross-validation spread.' That self-admission is good, but the same logic should be applied to the exponential evolution time.\n\nThe zero-padding to the union of simplices is a minor structural caveat: the common embedding depends on the sample, so the representation is not fully intrinsic. Probably acceptable for this benchmark, but worth stating explicitly.\n\nOn the positive side, the code is promised as open source, and the baseline comparison is fair—both methods consume the same L_k. The authors are also explicit about limitations: small benchmarks, no quantum advantage claim. This is a proof of concept, not a breakthrough.\n\nWho is this for? Researchers working on quantum kernels or topological data encoding. It deserves a serious referee: the framework is coherent, and the empirical issues are fixable by reporting t and running nested cross-validation. I would send it to review, with a clear request to report t and add a sensitivity analysis before acceptance.","headline":"A coherent but modest extension of quantum evolution kernels to higher-order Laplacians; the empirical advantage is underdetermined because the evolution time t is never reported and the QSVT gains are admitted to be selection-biased.","tokens_in":18584,"tokens_out":2526,"would_cite":true,"duration_ms":25755,"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":"The paper proposes encoding a dataset's higher-order topology into a quantum state by evolving a uniform simplex state under the combinatorial Laplacian, and reports that this representation separates random-graph clique complexes more accu","keywords":["quantum machine learning","topological data encoding","combinatorial Laplacian","simplicial complexes","clique complexes","quantum kernels","survival features","spectral filters"],"falsifier":"Re-run the benchmark with a fixed, disclosed evolution-time schedule (same t and time grid for every graph in both classes) and with the classical Laplacian baseline given the same hyperparameter tuning; if the quantum representations no longer lead, the central comparison is not robust.","tokens_in":17610,"feed_emoji":"⚛️","tokens_out":7218,"duration_ms":85272,"temperature":0.7,"pith_summary":"The paper's central claim is that the combinatorial Laplacian of a simplicial complex can serve directly as the generator of a quantum feature map. Representing each dataset by the state obtained after evolving a uniform simplex superposition under e^{-iL_k t}, and comparing these states through fidelity kernels or survival amplitudes, yields discriminative topological information that direct comparisons of Laplacian matrices miss. The authors test this on classification of clique complexes built from random graphs at different densities, across several simplicial dimensions, and report that the quantum representations consistently outperform the matrix-comparison baseline. They also show that replacing the exponential evolution with polynomial spectral filters implemented by quantum singular value transformation can tune which part of the Laplacian spectrum dominates. The authors present this as preliminary evidence that higher-order topology can guide quantum encodings without first reducing it to Betti numbers or persistence summaries.","feed_headline":"Topology-driven quantum states beat matrix baseline in graph tests","feed_subtitle":"Evolving a simplex state under the combinatorial Laplacian keeps higher-order structure a matrix-distance baseline misses.","key_machinery":"The key object is the k-th combinatorial Laplacian L_k = B_k^T B_k + B_{k+1}B_{k+1}^T on a simplicial complex: the down term couples k-simplices through shared (k−1)-faces and the up term through shared (k+1)-cofaces. Because L_k is Hermitian and sparse, the exponential e^{-iL_k t} is a unitary driven by the topology, and the uniform state |s⟩ evolved under it carries the Laplacian's spectrum. The fidelity kernel |⟨Ψ(X_1)|Ψ(X_2)⟩|^2 and the survival features ⟨s|e^{-iL_k t_j}|s⟩ turn that spectrum into similarity scores; quantum singular value transformation lets the exponential be replaced by a trainable polynomial p(L_k) to emphasize selected spectral regions.","core_discovery":"On its own terms, the paper claims that a data structure's k-th order topology can be encoded as |Ψ_k(X)⟩ = e^{-i L_k t}|s⟩, where L_k is the combinatorial Laplacian on k-simplices and |s⟩ is the uniform state over simplices. It defines an implicit quantum kernel from state fidelity and an explicit feature vector from survival amplitudes a(t) = ⟨s|e^{-iL_k t}|s⟩ sampled over time. On clique complexes of random graphs with different edge densities, the authors report that these quantum representations separate classes more accurately than a Gaussian kernel on the Euclidean distance between Laplacian matrices, with the largest advantage appearing at higher simplicial dimensions, and that the e","pith_inferences":["Not claimed by the paper: because the fidelity kernel and survival features depend on the Laplacian spectrum, two complexes with identical spectra but different homology would look identical to QTDE; a spectral-invariant test would clarify how much topology beyond the spectrum is actually being captured.","Not claimed by the paper: the reported advantage may be sensitive to how the evolution time and time grid are set; if those are optimized per dataset, the comparison is fitted, and a fixed, disclosed schedule is needed to confirm the qualitative ranking.","Not claimed by the paper: the zero-padding used to align complexes of different sizes introduces a dependence on the particular sample fleet; a normalization independent of the dataset's composition would make the kernel more portable.","Not claimed by the paper: the framework suggests a natural extension to persistent Laplacians—evolving over a filtration and reading survival amplitudes at multiple scales could give a quantum persistence fingerprint."],"forward_implications":["Higher-order topological structure can be injected into quantum kernels and feature vectors without first computing Betti numbers or persistence diagrams.","The explicit survival features give a low-dimensional, measurement-based readout that may be implementable with fewer resources than full state tomography.","The discriminative topology of a dataset lives at a task-dependent simplicial dimension; QTDE motivates evaluating multiple k rather than a single fixed scale.","Trainable polynomial spectral filters through quantum singular value transformation provide a mechanism to amplify or suppress spectral regions of the Laplacian to improve classification.","Because the quantum representations are built from the Laplacian spectrum, they inherit invariance to vertex relabeling once the simplicial complex is fixed."],"fun_headline_variants":["Quantum states encode graph topology better than Laplacian distances","Topology-driven quantum encoding beats matrix baseline in graph tasks","Quantum topological encoding captures higher-order graph structure","New quantum encoding method outperforms Laplacian comparisons","Simplex states under quantum evolution separate graph classes better"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The encoding's usefulness rests on the choice of evolution time t, which the paper refers to as 'some sufficiently large choice' and never fixes; if t is tuned per dataset to maximize accuracy, the claimed advantage is a fitting artifact rather than a property of the representation.","fun_headline_variants_meta":{"raw":{"variants":["Quantum states encode graph topology better than Laplacian distances","Topology-driven quantum encoding beats matrix baseline in graph tasks","Quantum topological encoding captures higher-order graph structure","New quantum encoding method outperforms Laplacian comparisons","Simplex states under quantum evolution separate graph classes better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000616,"raw_usage":{"total_tokens":2670,"prompt_tokens":690,"completion_tokens":1980,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":1904}},"tokens_in":434,"tokens_out":1980,"duration_ms":15719,"temperature":1.0,"reasoning_tokens":1904,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T03:31:57.618807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the benchmark with a fixed, disclosed evolution-time schedule (same t and time grid for every graph in both classes) and with the classical Laplacian baseline given the same hyperparameter tuning; if the quantum representations no longer lead, the central comparison is not robust.","supporting_citations":[],"review_version":1}