{"id":"fbc2da59-397d-4494-afe6-94b75a3a4132","arxiv_id":"2601.23084","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"For QSVMs under local depolarising noise, the paper claims analytic upper and lower bounds on the squared geometric margin, validated (it says) by simulation and on an IBM device, plus empirical evidence that margins predict QSVM generalisation.","lead":"This paper derives upper and lower bounds on how much the geometric margin of a quantum-kernel support vector machine shrinks when kernel values are computed on a device with per-qubit depolarising noise, and reports simulations plus an IBM-hardware test. It also argues that margins are good predictors of QSVM generalisation and that the commonly used global depolarising noise model is overly optimistic.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Results 1A/1B are false, not just unproved: for N=2, p=0.1, states |00⟩,|11⟩, the quadratic-form bound vᵀK̃v ≤ (1−p)^4 vᵀKv fails although the entrywise kernel bound (33) holds, so the margin bound derivations collapse.","rationale":"The paper's headline contribution is a theorem about noisy QSVM margins. For that theorem to be true, the algebraic step from entrywise kernel bounds to quadratic-form bounds must be valid. The reader flagged this as unproved; my stress test shows it is false: there exists a minimal two-point instance satisfying the entrywise bound where the quadratic-form inequality fails. I selected this as the single most load-bearing concern because it is the hinge on which both the upper and lower margin bounds depend. The counterexample is within the theorem's own domain (N=2, p=0.1, pure IQP-type states, valid dual solutions), so it is not a regime the authors can exclude without weakening the theorem. The p=0 lower-bound contradiction is also real and independently fatal, but it afflicts only the lower bound and could perhaps be repaired by restricting p; the Result 1A/1B failure undermines the proof of both bounds. Since this confirms the reader's weakest-assumption in strengthened form, the REJECT verdict stands.","tokens_in":24282,"tokens_out":23471,"duration_ms":210635,"concrete_test":"Use the two-point counterexample: N=2, L=1, p=0.1, pure feature states |00⟩ and |11⟩, labels y=(+1,−1). Compute the ideal kernel K (2×2 identity) and the exact noisy kernel K̃_ij=Tr(Λ^⊗2(ρ_i)Λ^⊗2(ρ_j)) with single-qubit depolarizing parameter λ=1−4p/3. Check: (i) the entrywise bound (33) holds for all entries; (ii) for the optimal noisy dual solution v=α(1,−1) with α=1/(K̃_11−K̃_12) (or any α>0), the claimed inequality vᵀK̃v ≤ (1−p)^4 vᵀKv fails. If (ii) fails, Results 1A/1B are falsified and the derivations of (42) and (56) are invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorem rests on Results 1A/1B (Appendix A.3.1/A.3.2), which convert the entrywise kernel upper bound (33) into the quadratic-form inequalities ∥w̃∥²_ᾱ ≤ (1−p)^{2N}∥w∥²_ᾱ and ∥w̃∥²_α ≤ (1−p)^{2N}∥w∥²_α. This step is not merely missing a proof; it is false. Because the coefficients ᾱ_iᾱ_j y_i y_j have indefinite signs, an entrywise bound cannot be substituted into the quadratic form. The failure is concrete. Take N=2, L=1, p=0.1, two training states |00⟩ and |11⟩ with labels ±1. The single-qubit depolarizing channel gives λ=1−4p/3; the noisy kernel elements are K̃_11=K̃_22=((1+λ²)/2)²≈0.766, K̃_12=((1−λ²)/2)²≈0.0155. The entrywise bound (33) with A=(1−p)^4≈0.656, B≈0.344 is satisfied. But for the SVM dual vector v=α(1,−1) (any α>0), vᵀK̃v=2α²λ²≈1.502α², while A vᵀKv=2α²(1−p)^4≈1.312α². Thus vᵀK̃v > A vᵀKv, contradicting (93) and (95). This is exactly the step used to derive (41) and (55); both margin bounds inherit it. The exact expression (89) is also incomplete—it restricts the remainder sum to m=n≠0, dropping the m≠n terms that cause the violation. Consequently the upper bound (42) and lower bound (56) are unsupported. (The lower bound also independently fails at p=0 under condition (61), but the Result 1A/1B failure is the deeper defect.)","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a margin-based robustness measure for quantum kernel-assisted support vector machines (QSVMs) under local depolarising noise. It derives an entrywise upper bound on noisy kernel elements, then uses SVM KKT and optimality relations to obtain upper and lower bounds on the squared geometric margin of a QSVM trained on the noisy kernel and evaluated against the ideal kernel. The main theoretical results are eqs. (42) and (56). The paper also reports empirical margin-accuracy correlations across several datasets, a comparison of local vs. global depolarising noise, and an experiment on the IBM ibm_fez device.","tokens_in":24715,"tokens_out":17939,"duration_ms":167545,"significance":"If the main theorem were correct, it would be a valuable analytic characterisation of noise-induced margin decay for QSVMs under a realistic, per-qubit noise model, and it would strengthen the case for margin-based generalisation analysis in quantum machine learning. The entrywise kernel bound (33) is a genuine and checkable derivation, and the empirical correlation between median geometric margin and test accuracy (r > 0.9) is interesting. However, the central step that converts the entrywise kernel bound into a quadratic-form bound is invalid, and the lower margin bound is contradicted already at p=0. The theoretical contribution is therefore not established; the useful empirical observations do not rescue the main claim.","major_comments":[{"comment":"The substitution of the entrywise bound K̃_ij ≤ A K_ij + B into the quadratic form is invalid because the coefficients s_i s_j = α_i α_j y_i y_j are not nonnegative. The KKT condition Σα_i y_i=0 removes only the BΣs_i s_j term; it does not control the A term. Concrete counterexample: N=2, L=1, p=0.1, training states |00⟩ and |11⟩ with labels ±1. For the local depolarising channel λ=1−4p/3=0.8667, K̃_11=K̃_22=((1+λ²)/2)²≈0.766, K̃_12=((1−λ²)/2)²≈0.0155, and A=(0.9)^4≈0.656. For the two-point SVM, the optimal noisy dual solution (for C not saturated) is v=α(1,−1) with α=2/(K̃_11+K̃_22−2K̃_12). Then v^T K̃ v ≈ 2.665 while A v^T K v ≈ 2.330, so v^T K̃ v > A v^T K v. This contradicts (93). Since (93) is used to derive (41) and (55), the margin bounds (42) and (56) are unsupported.","section":"Appendix A.3.1/A.3.2, eqs. (91)-(95)"},{"comment":"At p=0 the noisy kernel coincides with the clean kernel, so the noisy margin equals the clean margin: γ̃²_α = γ²_α. Under condition (61), which is allowed when Cmγ²_α < 1/2, the right-hand side of (56) becomes γ²_α / (1 − 2Cmγ²_α) > γ²_α. Thus (56) asserts γ²_α ≥ γ²_α / (1 − 2Cmγ²_α) > γ²_α, a direct contradiction. This failure is independent of the Result 1A/1B issue and shows that the lower margin bound is false in an allowed parameter regime, not merely unproven.","section":"Section 6.3, eq. (56) and conditions (57)–(61)"},{"comment":"Equation (89) is called the exact expression for the noisy kernel, but the remainder sum is restricted to m=n≠0. The m≠n terms are not shown to vanish; nothing in the derivation up to (87) eliminates them, and the original sum in (76) contains no such restriction. The label 'exact' is therefore unjustified, and (89) cannot serve as a repair for the signed-remainder problem that invalidates the quadratic-form step in Results 1A/1B.","section":"Appendix A.2, eq. (89)"},{"comment":"The feasible C′ range used in the numerical validation is not fully a priori: C′_min in (49) contains the noisy margin γ̃², the quantity being bounded. The paper estimates it on a held-out validation set and then selects datasets that have an overlapping feasible C′ range across noise levels. Although the authors acknowledge this, it means the agreement in Figs. 11–13 is a weaker test of the theory than a parameter-free bound would provide, and the dataset-selection step conditions the empirical evidence.","section":"Section 7.1 and eqs. (47)–(49)"}],"minor_comments":[{"comment":"The notation for the noise probability changes from p′_L to p_L to p without an explicit statement of the relation. In particular, p in (15) should be p_L, and the domain p_L∈[0,3/4] should be stated where the Kraus operators are introduced.","section":"Section 3.3, eqs. (12)–(15)"},{"comment":"The symbol C′ is first defined as mC before eq. (46) and later as C_0 m^{1−β} in eq. (63). These are different objects; please use distinct notation or explicitly state the reparametrization.","section":"Eqs. (46)–(49) and Eq. (63)"},{"comment":"The definition of c_{n_i} in Appendix A.2 conflicts with the c_{αβ} product constants used in Section 6.1. This makes the derivation hard to follow; distinct symbols would help.","section":"Appendix A.2, eq. (83)"},{"comment":"The statement that the hardware noise level is 'within the particular noise range' is not quantified. The crossing points of the green horizontal line with the bounds are used to infer a noise level, but no uncertainty or error bar for the hardware margin is given.","section":"Section 7.2, Fig. 13"},{"comment":"The arXiv title mentions a 'cross-kernel margin', but the term is not defined in the manuscript. Either define this quantity explicitly or adjust the title to match the content.","section":"General"}],"recommendation":"reject","confidential_remarks":"The central theoretical claim is invalid due to the indefinite quadratic-form substitution and the independent p=0 contradiction in the lower bound. The empirical sections on margin-accuracy correlation and local-vs-global noise are interesting but do not compensate for the failure of the main result. The authors would need a substantially different argument (for example, a genuine matrix inequality for K̃ − A K − B J, or an entirely different bounding technique) to restore the margin bounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The kernel-element bound under local depolarising noise (eq. 33) is a genuine derivation and appears new. The comparison of local vs global depolarising noise is empirically useful and well motivated, and the margin-generalisation correlation data give QSVMs something analogous to Hur–Park for QNNs. The paper is clearly written and the literature review is sensible. That is the good part.\n\nThe margin bounds do not hold up. Results 1A/1B convert the entrywise kernel bound into a bound on the quadratic form \\|w\\|^2 = \\sum \\alpha_i \\alpha_j y_i y_j K_ij. The coefficients have indefinite signs, so an entrywise upper bound does not imply a quadratic-form bound. This is not just a missing proof. For N=2, p=0.1, states |00> and |11> with labels (+1,-1), take v=(1,1). This satisfies \\sum v_i y_i = 0, the entrywise bound (33) holds, but v^T K\\tilde v is about 1.56 while (1-p)^4 v^T K v is about 1.31. So the claimed quadratic-form inequality is false. Both margin bounds inherit this step, and the exact expression (89) in the appendix drops the m≠n terms that produce the violation.\n\nThere is also a direct internal contradiction. At p=0, the noisy margin equals the clean margin exactly. The lower bound (56), under the paper's own validity condition (61), gives \\gamma\\tilde^2 >= \\gamma^2 / (2Cm\\gamma^2 - 1), with denominator between -1 and 0, so the RHS exceeds \\gamma^2. The bound therefore contradicts the trivial p=0 identity on a region the authors claim is valid.\n\nThe numerical validation is weakened by a calibration loop: C'_min is estimated using a validation-set measurement of the noisy margin, the very quantity being bounded. The reported beta and C0 values also do not match the feasible C' ranges in Table 1 (e.g., Heart Disease with C0=10, beta=0 gives C' ~2270, well above the stated [0,21.1)). The metadata abstract describes a different framing than the full text, which suggests version control issues.\n\nWho is this for? Researchers studying robust quantum kernels might want the kernel bound and the local-vs-global noise data, but they should not rely on the margin theorems. I would not accept this paper in its current form. But I would not desk-reject it either: the kernel bound is checkable and novel, the noise-model comparison is valuable, and the flaws are specific enough that one careful referee report could either kill the margin claims or point to a tractable repair. Send it to referees with a clear request to focus on Results 1A/1B and the p=0 behavior of the lower bound.","headline":"Kernel bound is real, but the margin bounds rest on an invalid quadratic-form step and a p=0 contradiction; salvageable but not acceptable as is.","tokens_in":25332,"tokens_out":8995,"would_cite":false,"duration_ms":76809,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For QSVMs trained under local depolarising noise, the squared geometric margin — measured against the ideal kernel — is bounded above and below by explicit functions of the per-qubit noise survival probability (1−p)^{2N}, the clean margin,","keywords":["quantum kernel methods","quantum support vector machines","cross-kernel margin","geometric margin","local depolarising noise","generalisation bounds","NISQ","kernel perturbation"],"falsifier":"Take any trained QSVM with 0 < Cmγ² < 1/2 and set p=0; the claimed lower bound (56) requires γ² ≥ γ²/(1−2Cmγ²), which is strictly stronger than the exact equality γ̃² = γ². A single such instance settles whether the validity conditions are sufficient. Separately, compute the weighted sum Σ_{i,j} ᾱ_i ᾱ_j y_i y_j [K̃_ij − (1−p)^{2N}K_ij − (1−(1−p)^{2N})] for a two-qubit dataset; any positive value violates the substitution step shared by both margin bounds.","tokens_in":23990,"feed_emoji":"⚛️","tokens_out":6894,"duration_ms":63525,"temperature":0.7,"pith_summary":"This paper introduces the cross-kernel margin for quantum kernel-assisted support vector machines (QSVMs): the geometric margin of a classifier trained on a noisy kernel but evaluated against the ideal kernel. It claims that under local per-qubit depolarising noise with error probability p, each noisy kernel element obeys K̃_ij ≤ (1−p)^{2LN} K_ij + 1 − (1−p)^{2LN}, and that this entrywise bound forces the squared cross-kernel margin into an explicit interval whose endpoints depend on the clean margin, the regularisation parameter C, and the noise survival probability. If correct, this gives a concrete characterisation of noise-induced margin decay for QSVMs and supports margin-based robustness analysis in the NISQ era. The paper also reports that the median geometric margin tracks test accuracy almost perfectly (Pearson r > 0.9) across four datasets, and that global depolarising noise is more optimistic than local noise at intermediate noise levels.","feed_headline":"Quantum-SVM margins shrink like (1−p)^{2N} under noise","feed_subtitle":"Upper and lower bounds on the cross-kernel margin certify how local depolarising noise degrades QSVM generalisation.","key_machinery":"The central object is the cross-kernel margin: the squared geometric margin 1/∥w̃∥², where w̃ uses the dual solution obtained from the noisy kernel but the norm is evaluated against the ideal kernel. The carrying mechanism is the entrywise kernel bound derived from the Kraus decomposition of the local depolarising channel via Schatten/Hölder trace inequalities, together with the KKT identity Σ ᾱ_i y_i = 0, which is used to drop the additive constant when passing from kernel elements to the quadratic form ∥w̃∥² = Σ ᾱ_i ᾱ_j y_i y_j K̃_ij. SVM duality then converts the resulting inequalities into the margin bounds.","core_discovery":"On the paper's own terms: local depolarising noise applied independently to each qubit degrades quantum kernel values by a contraction-plus-additive-constant form, K̃_ij ≤ (1−p)^{2LN}K_ij + (1−(1−p)^{2LN}). Substituting this bound into the SVM dual objective and using optimality relations for the soft-margin problem, the squared geometric margin of the noisy-trained, ideal-evaluated classifier satisfies γ̃² ≤ (1−p)^{2N}γ² / [2(1−mCγ²) − (1−p)^{2N}] and γ̃² ≥ γ²[1−2(1−p)^{2N}] / [2Cmγ² − 1], subject to feasibility constraints on mCγ². These are the upper and lower cross-kernel margin bounds; at p=0 the upper bound is tight. The construction also yields a hard-margin analogue and extends to L","pith_inferences":["A tighter, state-dependent bound could replace the additive constant: the paper's exact expression (89) exhibits a signed remainder depending on the purity and Pauli coefficients of the data states, so the loose entrywise bound may be far from tight for low-purity states.","The label-corruption experiments establish margin–accuracy correlation in the ideal setting; applying the same corruption pipeline under kernel noise would test whether the cross-kernel margin predicts accuracy under device noise as directly as it does under label noise.","The bounds suggest an experimentally testable noise threshold: solving for the p at which the lower bound crosses zero should predict where QSVM test accuracy collapses, and that prediction can be checked on hardware.","The derivation structure should carry to other single-qubit noise channels, such as amplitude damping, by replacing (1−p)^{2N} with the channel-dependent survival factor; disagreement would show where the depolarising assumption is load-bearing."],"forward_implications":["Margin decay under local depolarising noise is controlled by the survival probability (1−p)^{2N}; increasing qubit number or encoding layers compounds the noise effect.","The bounds give a computable range of regularisation parameters for which the noisy margin is guaranteed non-negative, allowing practitioners to certify training stability before running a noisy device.","The noisy-trained, ideal-evaluated margin, not the self-consistent noisy margin, is the informative robustness statistic; the paper argues the latter hides noise-induced decay.","In the noiseless setting, median geometric margin can serve as a proxy for QSVM generalisation, with Pearson correlations above 0.9 across the tested datasets.","Global depolarising noise models understate performance loss; local per-qubit noise should be preferred when modelling QSVM robustness in the NISQ era."],"fun_headline_variants":["Quantum-SVM margins shrink exponentially with qubit count and noise","Noise contracts quantum kernel margins: bound on QSVM stability","Cross-kernel margin: quantifying noise's hit on quantum SVMs","Tight bounds on quantum SVM margin loss under depolarizing noise","Margin-based proof: local noise degrades quantum SVMs predictably"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the entrywise kernel bound can be multiplied by signed SVM weights and summed to bound the squared weight norm; the exact expression (89) shows a state-dependent signed remainder, and the lower bound's validity conditions contradict the p=0 equality unless they are stronger than stated.","fun_headline_variants_meta":{"raw":{"variants":["Quantum-SVM margins shrink exponentially with qubit count and noise","Noise contracts quantum kernel margins: bound on QSVM stability","Cross-kernel margin: quantifying noise's hit on quantum SVMs","Tight bounds on quantum SVM margin loss under depolarizing noise","Margin-based proof: local noise degrades quantum SVMs predictably"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000543,"raw_usage":{"total_tokens":2466,"prompt_tokens":800,"completion_tokens":1666,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1577}},"tokens_in":544,"tokens_out":1666,"duration_ms":13285,"temperature":1.0,"reasoning_tokens":1577,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T06:18:03.062025+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any trained QSVM with 0 < Cmγ² < 1/2 and set p=0; the claimed lower bound (56) requires γ² ≥ γ²/(1−2Cmγ²), which is strictly stronger than the exact equality γ̃² = γ². A single such instance settles whether the validity conditions are sufficient. Separately, compute the weighted sum Σ_{i,j} ᾱ_i ᾱ_j y_i y_j [K̃_ij − (1−p)^{2N}K_ij − (1−(1−p)^{2N})] for a two-qubit dataset; any positive value violates the substitution step shared by both margin bounds.","supporting_citations":[],"review_version":1}