{"id":"1af43dc8-9036-48fd-b4b9-cf2a6fe29656","arxiv_id":"2607.22516","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Quantum Spectral Models encode each matrix input as a Hamiltonian, letting sample-dependent spectral gaps act as tunable Fourier carriers, and lead mean test accuracy on four benchmarks at depth 32.","lead":"This paper introduces quantum machine-learning models that build the data-encoding operation from the spectrum of each input matrix, so the available frequencies change from sample to sample. On four benchmarks, the models have the highest mean test accuracy at the largest tested circuit depth, but the comparison uses unequal qubit and parameter counts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth-32 QSM leadership may reflect narrower, more trainable circuits rather than spectral inductive bias; no same-resource control exists for the 2–4 qubit global QSMs.","rationale":"I read the paper as making two claims: a formal characterization of QSM Fourier support (Eqs. 20/24) and an empirical demonstration that this representation helps. The formal part is elementary and internally consistent; the derivations use standard spectral decomposition and the support inclusions are correct. The empirical part is where the argument is vulnerable. The protocol in Section 5 explicitly compares 'models as implemented' with different qubit counts and parameter counts. The patch-local QSM vs patch-SU(4) comparison is a decent partial control (same 8 qubits, comparable parameter counts), and the ablations are informative. But the global QSMs' leadership on both synthetic tasks uses 2–4 qubits with no same-resource non-spectral control; the 16-qubit rotation encoders suffer from strongly suppressed mixer-gradient variance at depth 32 (Section 6), so the observed ordering is consistent with a trainability/resource explanation. A resource-matched replication would settle this. The dirty-worktree and single-seed diagnostics are secondary reproducibility issues; the main theoretical result is unaffected. Verdict remains CONDITIONAL, matching the reader's assessment.","tokens_in":48595,"tokens_out":11012,"duration_ms":120498,"concrete_test":"Run a resource-matched depth-32 replication on the two synthetic tasks. (a) Implement global Hblock and Hsym on 8 qubits by direct-sum padding H→H⊕0 with the same 8-qubit brick-wall SU(4) mixer used for patch-SU(4); (b) train, under the identical optimizer/splits/seeds, the fixed and trainable patch-SU(4) encoders and an 8-qubit rotation encoder at matched parameter counts; (c) compare final test accuracy over 20 seeds. If any non-spectral 8-qubit control reaches or exceeds Hblock's 85.15% (EIGENGAP) or 92.72% (SINGULAR), the QSM leadership on these tasks is explained by resource width/trainability rather than by the spectral inductive bias. If padded 8-qubit QSMs maintain or improve the margin while non-spectral controls do not, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Fourier-support derivation (Eqs. 17–25) is mathematically sound. The load-bearing weakness is the empirical claim that the input-conditioned construction, rather than resource count/trainability, drives the depth-32 leadership. Section 5 states: 'The results compare the models as implemented, rather than isolating the encoder from all resource differences.' Rotation encoders use 16 qubits and ~7200–7712 parameters at L=32; patch encoders use 8 qubits and 3360–11040; global QSMs use 2–4 qubits and 512–1472. The global QSM leadership on SYNTHETIC EIGENGAP/SINGULAR is therefore confounded with a large reduction in circuit width and parameter count, which the paper's own gradient diagnostics show strongly affects trainability (rotation encoders reach median log10 mixer-gradient variance ≈ −11 to −20 at depth 32 versus ≈ −1 to −4 for QSMs). The patch-SU(4) baselines provide a useful same-width, comparable-parameter control for the patch QSM, but no such control exists for the 2–4 qubit global QSMs. Since the abstract credits 'input-conditioned spectral representations' with the results, this confound is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Quantum Spectral Models (QSMs), a data-reuploading family in which the generator of the encoding unitary is built from each input matrix. Three variants are studied: symmetric Hamiltonian H_sym=(M+M^T)/2, block Hamiltonian H_block=[[0,M],[M^T,0]], and a non-overlapping patch-local block Hamiltonian. The central formal contribution, Eqs. (17)-(25), derives the maximal one-upload candidate Fourier support as eigengaps of H_sym or signed singular-value gaps of H_block, with Fourier coefficients depending on input-derived spectral projectors, the initial state, the trainable mixer, and the readout. Empirically, the paper reports that at reuploading depth 32, a QSM variant attains the largest mean test accuracy among the tested quantum models on all four benchmarks (Pendigits DYN and STA4, and two synthetic spectral tasks), and that value-vs-subspace ablations show a task-dependent reversal: subspace-preserving controls dominate on Pendigits while spectral-value-only controls dominate on the synthetic tasks. The body of the paper is notably candid about the limits of the comparisons, but the abstract and conclusion frame the empirical results as evidence for the input-conditioned spectral inductive bias itself.","tokens_in":48886,"tokens_out":13792,"duration_ms":148887,"significance":"The formal analysis is a genuine contribution: Eqs. (17)-(25) are a clean, self-contained spectral-theorem characterization of a novel encoder family, with no hidden fitted constants in the support derivation. The paper ships code, reports full depth sweeps with 20 seeds and standard deviations, discloses data-provenance issues (git_dirty records, Appendix F), and explicitly labels the latent-state diagnostics as descriptive. The value-versus-subspace ablation template is a useful methodological idea for the QML community. However, the significance is materially tempered by three issues. First, the headline depth-32 empirical claim is confounded with resource count and trainability; the paper itself states in Section 5 that the comparisons are 'as implemented, rather than isolating the encoder from all resource differences.' Second, the Fourier support results, while correct, are expansions in the upload-time variable t for each fixed input, not in the input features; the abstract's phrasing invites the standard QFM input-Fourier reading. Third, on the synthetic benchmarks the value-only ablation success is expected by construction, since the labels are defined by the very spectral","major_comments":[{"comment":"The depth-32 leadership of the global QSMs is confounded with circuit width, parameter count, and trainability. At L=32 the global QSMs use 2-4 qubits and 512-1472 trainable scalars, while the rotation-gate baselines use 16 qubits and 7200-7712 scalars (Appendix E). Section 5 explicitly disclaims resource isolation ('The results compare the models as implemented, rather than isolating the encoder from all resource differences'). The paper's own gradient diagnostics (Section 6, Figure 4a) show median log10 mixer-gradient variance of roughly -11 to -20 for the 16-qubit rotation encoders versus roughly -1 to -4 for the QSMs, so a narrower-more-trainable-circuit explanation is fully consistent with the accuracy ordering. No same-width control exists for the 2-4 qubit global QSMs (the 8-qubit patch-SU(4) baseline controls the patch QSM, but not the global variants). Since the abstract credits","section":"§5, Table 1, Appendix E; Figure 4a"},{"comment":"The Fourier support results are expansions in the upload-time variable t for each fixed input M, not in the input features. The paper states this correctly in Appendix C.1 ('the matrix M is fixed and the upload time is the Fourier variable'), but the abstract and introduction present 'input-conditioned frequency support' in language that naturally reads as a QFM-style input-Fourier statement (Schuld et al., Eq. (4) of this paper). At inference t is a fixed trained parameter, so the support characterization describes the frequency content of the unitary family along a variational parameter; the input-dependence of the trained output enters through the M-dependence of both the phases and the coefficients, and is not itself a truncated Fourier representation. This is a legitimate analytic framework, but the delimitation should be stated in the abstract-level claims, otherwise the central th","section":"§4, Eqs. (17)-(25); Appendix C.1"},{"comment":"The 'task-dependent reversal' in the ablations is partially a consequence of benchmark construction. On SYNTHETIC EIGENGAP and SYNTHETIC SINGULAR, the labels are defined by the largest eigengap and the leading singular-value sum respectively - exactly the quantities that the spectral-value-only controls feed directly - and the classical MLP trained on those descriptors reaches 98.01% and 98.50% (Table 6). The value-only controls reaching 97.92% and 98.42% (Table 2) is therefore a consistency check of the data-generation rule rather than an empirical discovery about QSM representations. The paper does acknowledge this ('This reversal agrees with the data-generating rules'), but the abstract presents the reversal as shedding new light. The Pendigits half of the reversal is the informative empirical evidence; the synthetic half should be framed as a calibration check.","section":"Table 2, Appendix D.2, Table 6"}],"minor_comments":[{"comment":"The depth-32 'leads' claim is based on point estimates. Several QSM-vs-QSM gaps are within one standard deviation (e.g., DYN: 92.75±0.84 vs 92.35±5.95; EIGENGAP: 85.15±2.40 vs 82.81±3.16). The QSM-vs-rotation gaps are large and robust, but pairwise significance tests or bootstrapped confidence intervals over the 20 seeds would substantiate the ranking claims.","section":"Table 1, Figure 3"},{"comment":"Please state explicitly in the abstract that the Fourier representation is in the upload-time variable, e.g., 'for each input matrix, the output as a function of the upload time admits a truncated Fourier representation whose candidate carriers are input-dependent spectral gaps.'","section":"Abstract"},{"comment":"Typo: 'Lloydet al.' should be 'Lloyd et al.'","section":"Appendix C.3"},{"comment":"Consider defining 'maximal candidate support' once near Eq. (7) and distinguishing it consistently from 'realised support' throughout; the distinction is used heavily and the current text introduces it somewhat informally.","section":"§4, Eq. (20)"},{"comment":"The description of the synthetic data generation is clear, but the sentence 'In the clean limit, the input-derived symmetric Hamiltonian has eigenvalues λj(Si)' should note that the model sees the noisy Xi, not Si; the current wording could be misread as claiming the model has access to the clean spectrum.","section":"Appendix D.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is unusually honest: it discloses the resource limitations in Section 5, the descriptive nature of the diagnostics in Section 6, and even the dirty-worktree provenance in Appendix F. The gap is between that disclosed protocol and the abstract's causal framing. The theory is sound and publishable; the empirical section either needs a same-resource control for the 2-4 qubit global QSMs, or the headline claim must be reframed to 'as implemented' performance with the inductive-bias attribution deferred. I would be willing to accept after such a revision; I do not see this as a reject. The paper's fit for the journal is good, but the abstract should not promise more than the protocol delivers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The genuinely new thing is the block-Hamiltonian embedding and the explicit sample-conditioned Fourier-support characterization: eigengaps for the symmetric embedding, signed singular-value gaps (including zero-mode terms) for the block embedding. The derivations in Eqs (17)-(25) are standard spectral-theorem expansions and they are correct. The patch-local variant is a natural extension, and the value-vs-subspace ablations produce a clean task-dependent reversal. I also want to give credit where it's due: the paper is unusually candid. It says in Section 5 that the comparisons are \"as implemented\" rather than isolating the encoder from resource differences. It discloses that 319 of 3840 archive records came from a dirty worktree, while the headline depth-32 winner aggregates are clean. It flags that latent diagnostics are single-seed. That honesty is real.\n\nNow the soft spots, in proportion. The depth-32 leadership claim is load-bearing and it is not resource-normalized. Global QSMs use 2-4 qubits and 512-1472 parameters; rotation encoders use 16 qubits and ~7200-7700 parameters; patch encoders use 8 qubits and 3360-11040. The gradient diagnostics show that the rotation encoders have median log10 mixer-gradient variance around -11 to -20 while QSMs are -1 to -4. So the QSM wins could reflect narrower, more trainable circuits rather than spectral inductive bias. The patch-SU(4) controls give a same-width, comparable-parameter control for the patch QSM, but there is no such control for the global QSMs. The abstract credits \"input-conditioned spectral representations\" with the results; that attribution needs a same-resource experiment or the claim should be softened.\n\nOther issues: depth 32 is where QSMs happen to lead, but the largest observed mean often occurs at depth 16; there are no pairwise significance tests; the synthetic tasks are generated from spectral statistics, so the value-only ablation doing well there is partly built-in. The dirty records affect secondary comparisons (e.g., the trainable patch-SU(4) Fisher-rank example in Appendix G is entirely dirty), and the aggregate initialisation diagnostics don't store the git_dirty field at all. The latent diagnostics are genuinely single-seed; the paper says so, but that limits interpretation.\n\nBottom line: the Fourier-support theory is a real contribution and the empirical work is honest but overclaimed. This paper deserves a serious referee and likely major revision. The fix is not hard: add a resource-matched control for the global QSMs (same qubit/parameter budget with a coordinate-wise or SU(4) encoder), run significance tests, and either validate depth selection or drop the 'largest depth' emphasis. I'd bring it to a reading group, and I'd cite it if I were working on QML encodings.","headline":"Solid new Fourier-support theory for Hamiltonian-based encoders; the empirical leadership claim is honest but confounded by resource differences, so referees should require a same-resource control.","tokens_in":49332,"tokens_out":2567,"would_cite":true,"duration_ms":25838,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","15A18"],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"Given matrix input M, the QSM data-encoding unitary is exp(-iH(M)t/2) with H_sym(M)=(M+M^T)/2 or H_block(M)=[[0,M],[M^T,0]]. The paper shows that, for one upload, the output is a truncated Fourier series in t with maximal support consisting","keywords":["quantum machine learning","data reuploading","Hamiltonian embedding","spectral inductive bias","Fourier representation","singular value decomposition","eigengaps","Pendigits"],"falsifier":"Take a small matrix with known singular values, train a block-Hamiltonian QSM, and numerically Fourier-transform the observable output in the upload time; if any frequency appears that is not a signed singular-value gap (half-difference, half-sum, or half-singular-value term) of that matrix, Equation (24)'s support characterisation is wrong. Alternatively, re-run the depth-32 comparisons with matched qubit and parameter counts; if a rotation encoder then reaches or exceeds the QSM means, the inductive-bias attribution is unsupported.","tokens_in":48518,"feed_emoji":"⚛️","tokens_out":10586,"duration_ms":104886,"temperature":0.7,"pith_summary":"Quantum data encoding can carry an inductive bias matched to matrix-structured inputs: instead of encoding features through coordinate-wise rotations, a Quantum Spectral Model (QSM) constructs the generator of the data-encoding unitary directly from each input matrix. The paper claims that the observable output of such a model becomes a truncated Fourier series in the upload time whose candidate frequencies are input-dependent spectral gaps, while input-dependent spectral subspaces help set the coefficients. It further claims that at the largest depth tested (32), a QSM variant records the highest mean test accuracy among all tested quantum models on all four benchmarks, and that ablations show a task-dependent reversal between spectral values and spectral subspaces. A sympathetic reader should care because this turns data encoding into a principled choice of representation: the model can expose the spectral structure that the task rewards.","feed_headline":"Lead four benchmarks with input-tuned quantum tones","feed_subtitle":"At depth 32, a QSM variant posts the best mean test accuracy among all tested quantum models.","key_machinery":"The central object is the sample-conditioned Hamiltonian H(M), used as the generator of the data-encoding unitary within a standard upload-mixer data-reuploading circuit. Its spectral decomposition supplies the phase carriers: exponentials exp(-i lambda_a t/2) for a symmetric Hamiltonian, and exp(-i sigma_j t/2) for a block Hamiltonian. Taking pairwise differences of the generator eigenvalues yields the maximal one-upload Fourier support—eigengaps for the symmetric variant, signed singular-value gaps for the block variant—while the spectral projectors, together with the initial state, mixer, and observable, determine which of these carriers are actually realized and with what coefficients. T","core_discovery":"Given matrix input M, the QSM data-encoding unitary is exp(-iH(M)t/2) with H_sym(M)=(M+M^T)/2 or H_block(M)=[[0,M],[M^T,0]]. The paper shows that, for one upload, the output is a truncated Fourier series in t with maximal support consisting of the half-eigengaps of H_sym(M), or the signed half-differences and half-sums of singular values of M (plus half-singular-value terms if zero modes exist). Coefficients are formed by the spectral projectors, the initial state, the mixer, and the readout observable. Because H(M) depends on M, the candidate frequencies vary per sample; the patch-local variant uses four 2x2 block Hamiltonians and gets a Cartesian-product support.","pith_inferences":["My inference beyond the paper: if this construction is taken as a template, the design lesson is to pick the generator whose spectrum matches the informative statistics of the data family; one could test this by building QSMs from other input-derived operators (covariance matrices, graph Laplacians of patch graphs) and checking whether sample-conditioned carriers continue to beat rotation gates.","My inference: the paper leaves the causal role of locality unresolved; a natural follow-up is a resource-matched re-run—same qubit count and comparable parameter count for global vs patch-local QSMs—to see whether the depth-32 ordering survives once width and parameter totals are held fixed.","My inference: the value-versus-subspace reversal suggests an adaptive model that chooses the spectral component to preserve based on task statistics; this could be formalised as a meta-learning or routing problem, but the paper does not propose it.","Manuscript-disclosed limitation worth flagging: Section 5 explicitly says the comparisons are 'as implemented' rather than isolating the encoder from resource differences, and Appendix F/G disclose that some archived diagnostics came from a dirty worktree (the four depth-32 winner aggregates and global-Hblock maxima are marked clean). A clean-tree, resource-matched reproduction is needed before th"],"forward_implications":["If the Fourier support characterisation is right, then no trained symmetric-QSM output can contain a frequency outside the half-eigengaps of the input matrix, and no block-QSM output can contain a frequency outside the signed singular-value gaps; this gives a testable, sample-by-sample spectral fingerprint of the model.","Data encoders no longer need to be limited to a globally shared or globally learned frequency grid; each input matrix can determine its own available phase carriers, so the inductive bias is adapted at the sample level.","At depth 32 the paper reports a QSM variant as top mean test accuracy among the tested quantum models on all four benchmarks (patch-local block-Hamiltonian on the two Pendigits representations; global block-Hamiltonian on the two synthetic spectral tasks).","The ablation result implies that no single spectral component is universally sufficient: for Pendigits, sample-dependent spectral subspaces carry most of the useful information, while for the synthetic eigengap and singular-value tasks the spectral values themselves are the decisive component."],"fun_headline_variants":["Input-tuned quantum spectra lead four benchmarks","Matrix-encoded quantum model beats tested rivals","Quantum frequencies from data shape winning model","Patch-local quantum spectral model tops Pendigits","Input-dependent quantum Fourier support wins tasks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The empirical conclusion rests on comparing the full encoder-mixer models as implemented, with qubit counts ranging from 2-4 for global QSMs to 16 for rotation encoders and parameter counts differing by an order of magnitude; if unequal resource counts, not the encoder construction, explain the accuracy ordering, the paper's main empirical claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Input-tuned quantum spectra lead four benchmarks","Matrix-encoded quantum model beats tested rivals","Quantum frequencies from data shape winning model","Patch-local quantum spectral model tops Pendigits","Input-dependent quantum Fourier support wins tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00016,"raw_usage":{"total_tokens":1117,"prompt_tokens":838,"completion_tokens":279,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":215}},"tokens_in":582,"tokens_out":279,"duration_ms":3501,"temperature":1.0,"reasoning_tokens":215,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:29:25.232885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small matrix with known singular values, train a block-Hamiltonian QSM, and numerically Fourier-transform the observable output in the upload time; if any frequency appears that is not a signed singular-value gap (half-difference, half-sum, or half-singular-value term) of that matrix, Equation (24)'s support characterisation is wrong. Alternatively, re-run the depth-32 comparisons with matched qubit and parameter counts; if a rotation encoder then reaches or exceeds the QSM means, the inductive-bias attribution is unsupported.","supporting_citations":[],"review_version":1}