{"id":"2a2affea-81c0-4005-a6dd-11facb2ba462","arxiv_id":"2605.22275","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An adaptive shot-allocation rule that spends noisy kernel-estimation measurements on SVM decision-critical entries beats uniform allocation and can stop early.","lead":"This paper proposes an adaptive way to spend a limited measurement budget when estimating kernel matrices for SVM classifiers from noisy observations, directing shots to the entries that matter most to the decision boundary. It reports better classifier fidelity than uniform allocation in synthetic and simulated quantum-kernel experiments, plus an early-stopping rule that uses a fraction of the budget.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 12 misapplies the envelope theorem: the true margin derivative is not α_iα_j y_i y_j, so the claimed Neyman-optimal allocation lacks a valid proof.","rationale":"The central claim is that the Neyman-like allocation (Eq. 23) plus active-set instability yields an SVM closer to the true-kernel SVM than uniform allocation. A necessary condition for the theoretical half of this claim is that the sensitivity used to define w_ij is actually the derivative of the quantity being variance-minimized. Eq. 12 is derived by an envelope-theorem argument that cannot work: the theorem applies to the value function, and the margin norm is not the value function. The minimal two-point example shows Eq. 12 has the wrong sign, and the three-point bound example shows the discrepancy is not a global constant. This means the paper's proof of Proposition 1 and the optimality of Eq. 23 are unsupported. I am not claiming the adaptive heuristic is useless; the active-set instability term is a reasonable addition, and the empirical results could still be valid. But without code/data and without a corrected derivation, the theoretical grounding of the central claim fails. The reader's CONDITIONAL verdict is therefore the right call; the condition should be a corrected sensitivity analysis (or an explicit statement that the allocation is heuristic) plus release of code/data. Secondary issues (no AQKA baseline, no reproducibility artifacts) reinforce the conditional status but are not the load-bearing flaw.","tokens_in":22019,"tokens_out":22809,"duration_ms":241822,"concrete_test":"Take a small fixed SVM with a mixture of free and C-bound support vectors (e.g., n=3, y=(+,+,−), K_ii=1, K_12=K_23=0.5, vary K_13=l, set C so α=(C/2,C/2,C)). For each off-diagonal entry, numerically perturb K_ij by ±10^-6, retrain the SVM, and compute Δ||w||²/ΔK_ij. Compare these finite-difference derivatives with Eq. 12. If the ratio (true derivative)/(α_iα_j y_i y_j) is not constant across entries, the sensitivity formula is invalid and the allocation in Eq. 23 is not the oracle margin-variance minimizer; a corrected KKT sensitivity would then be needed to support the optimality claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing defect is the sensitivity formula (Eq. 12), used to derive the variance proxy (Eq. 19) and the Neyman allocation (Eq. 23). The envelope theorem differentiates the dual value V(K)=g(α*(K),K), not the squared margin W(K)=||w||². Since W=2(Σα_i − V), the derivative of W includes 2Σ∂α_i/∂K_ij, which the paper drops. In the minimal two-point free-SV case (K_11=K_22=1, K_12=k, y=(+1,−1)), W=2/(1−k), so dW/dk=+2/(1−k)², whereas Eq. 12 gives α₁α₂y₁y₂=−1/(1−k)². The sign is wrong, and the correct magnitude differs by a factor of 2. The error is not a harmless global constant: with bound SVs the discrepancy is entry-dependent. For a 3-point kernel with α=(C/2,C/2,C) (third point at C), a direct calculation gives dW/dK_13=−2C² while α₁α₃y₁y₃=−C²/2 (ratio 4), and the ratio for K_12 is different. Thus (α_iα_j)² is not proportional to the true squared sensitivity, so Eq. 23 is not the minimizer of the stated margin-variance objective. The sparsity conclusion (entries with α_i=0 or α_j=0 have zero first-order influence) may survive, but the precise weighting and the optimality claims (Prop. 1, Eq. 23) do not follow from the given derivation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies learning kernelized SVMs when the Gram matrix is observed through noisy Bernoulli measurements, as in quantum kernel estimation. It proposes to allocate a fixed shot budget non-uniformly across kernel entries, using an oracle Neyman-type allocation derived from a margin-variance proxy, and then an adaptive algorithm that combines estimated geometric sensitivity with active-set instability, plus an early-stopping rule based on dual-coefficient stability. The empirical section compares the adaptive scheme with uniform allocation on synthetic and quantum-kernel datasets and reports improved support-vector recovery, margin accuracy, and decision-function RMSE, with substantial measurement savings from early stopping.","tokens_in":22532,"tokens_out":12026,"duration_ms":136624,"significance":"The problem is timely and practically relevant: shot allocation in quantum kernel methods is usually uniform, and a principled task-aware allocation could yield real resource savings. The paper contains a broad experimental study, an honest treatment of regimes where uniform allocation wins, and a simple early-stopping criterion that appears useful. However, the theoretical core is currently not reliable. The margin-sensitivity formula (Eq. 12) misapplies the envelope theorem, so the oracle allocation Eq. (23), Proposition 1, and the claim that Eq. (19) is the margin variance are not established. As a result, the paper presently supports a promising heuristic rather than the formal optimality claims it states. The empirical findings may survive a corrected theory, but the formal claims must be repaired or substantially reframed.","major_comments":[{"comment":"The envelope theorem is applied to the wrong quantity. Appendix B correctly shows that for the dual value V(K)=max_α g(α,K), one has ∂V/∂K_ij = -α_iα_j y_i y_j. But Eq. (12) claims ∂||w||²/∂K_ij = α_iα_j y_i y_j. Since ||w||² = 2(Σ_i α_i - V), the correct expression contains 2Σ_l ∂α_l/∂K_ij plus the direct term; the dual variables are not locally constant when the active set is fixed. In the minimal two-point model with K_11=K_22=1, K_12=k, y=(+1,-1), the exact solution gives ||w||²=2/(1-k), so d||w||²/dk = 2/(1-k)², whereas Eq. (12) gives -1/(1-k)². The sign and magnitude are wrong, and the discrepancy is entry-dependent in larger examples with bound support vectors, so it is not absorbed by the squared weight in Eq. (23). Consequently Eq. (19) is not the stated margin variance, Eq. (23) does not solve (18), and Prop. 1 does not establish the claimed oracle comparison. The sparsity conc","section":"Sec. II-C, Eq. (12)"},{"comment":"The oracle allocation Eq. (23) is defined in terms of the true dual solution α and true kernel K. Algorithm 1 replaces these by estimates from the noisy kernel bK and the current model. The paper acknowledges this in the discussion after Prop. 1 but provides no convergence result, consistency argument, or finite-sample bound showing that the adaptive score approaches Eq. (23), or that the misallocation error is controlled. The central claim that the adaptive scheme approximates the oracle-optimal Neyman allocation is therefore an assertion rather than a theorem. Moreover, the experiments compare Algorithm 1's heuristic score (Eq. 28) against uniform allocation, not against the oracle allocation Eq. (23); they do not directly validate the margin-variance optimality claimed in Sec. III-B.","section":"Secs. III-B and IV-A, Eq. (23) vs Algorithm 1"},{"comment":"The two propositions are formally correct for the proxy objective V = Σ_i<j w_ij/N_ij if the weights w_ij are the ones in Eq. (24). But because Eq. (12) is incorrect, the weights do not correspond to the true squared sensitivity of ||w||². The inequality V* ≤ V_unif is therefore a statement about an auxiliary objective, not about the geometric margin of the SVM. The practical interpretation in Sec. III-C and Fig. 2 — that heterogeneity in the true kernel importance governs the advantage of adaptive allocation — may be qualitatively right, but it is not supported by the derivation as written.","section":"Sec. III-C, Prop. 1 and Prop. 2"}],"minor_comments":[{"comment":"The pseudocode defines s_ij ← (1−λ)|α_iα_j y_i y_j| + λ P_i P_j, while Eq. (28) in the text has an additional C² factor in the active-set instability term. The two are inconsistent; if the C² scaling is intentional, it should appear in the pseudocode.","section":"Algorithm 1, line 11 vs Eq. (28)"},{"comment":"The measurement model writes K^(k) = K + ε^(k) with zero mean but does not state that K^(k) must lie in [0,1] for the Bernoulli sampling in Eq. (7) to be well-defined. Either the noise should be interpreted as clipped to [0,1], or a bounded-noise assumption should be stated.","section":"Sec. II-B, Eq. (4)"},{"comment":"The text states that the decision-function RMSE decreases monotonically, but Fig. 6 shows only medians with an interquartile ribbon; no per-run monotonicity is demonstrated. Please soften the wording or add a supporting statement.","section":"Sec. V-C and Fig. 6"},{"comment":"The comparison between margin-based and decision-function allocations uses the inequality |α_iα_j| ≤ (α_i²+α_j²)/2 ≤ α_i²+α_j². This shows the two weights have the same support-vector concentration, but it does not establish that the resulting allocations are 'equivalent up to constant factors', as stated; the sqrt-based allocation differs by more than a constant factor.","section":"Sec. III-D"}],"recommendation":"major_revision","confidential_remarks":"I agree with the stress-test concern: Eq. (12) is a genuine load-bearing error, not a presentation issue. The revision should either re-derive the sensitivity by differentiating the KKT system and update Eq. (19), Eq. (23), and Prop. 1, or explicitly reframe the method as a heuristic and remove the optimality claims. The experimental study is substantial and may still support a useful heuristic, but the current formal framing is not acceptable. I would not recommend rejection outright because the core idea is plausible and the empirical evidence is suggestive; however, the theory must be corrected or withdrawn before the claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a look, but its headline theoretical claim doesn't survive contact with the math. The problem it tackles is real: when kernel entries are estimated from a limited shot budget, allocating uniformly across the Gram matrix is obviously wasteful for SVM-type classifiers. The adaptive scheme—margin sensitivity plus active-set instability, with early stopping on dual-coefficient stability—is a sensible heuristic, and the experiments on synthetic and quantum-derived kernels show the expected regime behavior. The paper is clearly written and honestly discusses when adaptivity helps and when it doesn't.\n\nThe soft spot is load-bearing. Eq. 12 claims that d||w||^2/dK_ij = α_i α_j y_i y_j by the envelope theorem. That is not what the envelope theorem says. The theorem differentiates the dual value function V(K) = g(α*(K), K), not the squared margin. A minimal two-point SVM (K11=K22=1, K12=k, balanced labels) gives the true derivative as +2/(1−k)^2, while Eq. 12 gives −1/(1−k)^2. Wrong sign, wrong magnitude, and the discrepancy is entry-dependent when bounds are active. So the weights w_ij = (α_i α_j)^2 K_ij(1−K_ij) are not the correct sensitivity weights for the margin-variance objective, and Proposition 1 plus Eq. 23 do not follow from the derivation. The sparsity claim (non-support-vector entries have zero first-order influence) also doesn't follow, because α varies continuously with K.\n\nThe adaptive loop also has a circularity issue: the oracle weights are defined in terms of the true α of the very kernel being estimated. The paper acknowledges this and uses an iterative approximation, but the justification remains heuristic.\n\nWhat the paper does well is the empirical framing. The experiments are internally consistent, the regime analysis (low/high structure, kernel concentration) is plausible, and the early-stopping results at ~16% budget are interesting. But there is no comparison to AQKA, and no code or data are provided.\n\nMy take: the empirical method may survive a correction, but the paper as written overclaims. It deserves a serious referee—not a desk reject—because the problem is relevant and the heuristic is worth testing. A referee should push the authors to either fix the sensitivity derivation or reframe the allocation rule as a heuristic without optimality claims. I would not cite it in its current form.","headline":"Useful empirical heuristic, but the core derivation is wrong: Eq. 12's envelope-theorem claim does not hold, so the stated optimality of the allocation formula does not follow.","tokens_in":22904,"tokens_out":2762,"would_cite":false,"duration_ms":32249,"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":"Under noisy observations of a Gram matrix, this paper claims SVM learning should allocate measurement shots according to classifier sensitivity, not uniformly: shots proportional to |α_i α_j| sqrt(K_ij(1−K_ij)), with an active-set instabili","keywords":["kernel methods","support vector machines","measurement allocation","noisy kernels","quantum kernel estimation","Neyman allocation","adaptive sampling","early stopping"],"falsifier":"Compute the exact derivative of ||w||² with respect to K_12 in a two-point SVM with K_11=K_22=1, K_12=k, opposite labels, and no box constraint. The paper's sensitivity formula gives −1/(1−k)², while direct differentiation of the exact margin ||w||²=2/(1−k) gives +2/(1−k)²; this single toy calculation settles whether the sensitivity identity at the core of the allocation rule is correct.","tokens_in":21940,"feed_emoji":"🎯","tokens_out":6085,"duration_ms":63381,"temperature":0.7,"pith_summary":"The paper argues that when a kernel matrix must be estimated from noisy Bernoulli measurements—as in quantum machine learning—spreading the measurement budget evenly across entries is the wrong default for SVMs. It claims that the right rule is to allocate shots in proportion to each entry's influence on the classifier margin, which yields a Neyman-type allocation N*_ij ∝ |α_i α_j| sqrt(K_ij(1−K_ij)), augmented by an active-set-instability term. The resulting multi-round adaptive scheme is claimed to recover the true-kernel SVM more accurately than uniform allocation at the same budget, and a dual-coefficient stability criterion lets it stop early using only about 16% of the budget. If correct, this turns shot allocation from a generic estimation problem into a task-aware design problem and offers a practical replacement for uniform sampling in noisy and quantum kernel learning.","feed_headline":"Adaptive shot allocation beats uniform sampling for noisy kernel SVMs","feed_subtitle":"A Neyman-style rule spends shots on support-vector kernel entries; early stopping uses ~16% of the budget.","key_machinery":"The load-bearing object is the score function s_ij = (1−λ)|α_i α_j y_i y_j| + λ P_i P_j C², modulated by the Bernoulli variance estimate sqrt(p_ij(1−p_ij)). Here |α_i α_j| comes from the claimed margin-sensitivity identity ∂||w||²/∂K_ij = α_i α_j y_i y_j, and P_i = Φ(−Δ_i/σ_{f,i}) estimates the probability that point i crosses the margin threshold under decision-function noise. These scores are normalized into a multinomial distribution over independent kernel entries, so each adaptive round concentrates shots on support-vector and near-margin entries. The same optimization also yields the closed-form Neyman allocation N*_ij ∝ sqrt(w_ij), which is the theoretical target the adaptive loop app","core_discovery":"The paper's central claim is that measurement allocation for kernelized SVMs should follow the classifier's geometry, not equalize entry-wise variance. Minimizing a first-order proxy for Var(||w||²) under a total-shot constraint gives an oracle allocation N*_ij ∝ |α_i α_j| sqrt(K_ij(1−K_ij))—a Neyman allocation in which shots go to pairs with large dual products and uncertain kernel values. Because the true α and K are unknown, the paper wraps this formula in an adaptive procedure: a uniform pilot, repeated rounds that combine margin sensitivity with active-set instability P_i P_j, and early stopping when the dual vector stops changing. The claimed result is that this procedure yields an SVM","pith_inferences":["Because the decision-function objective leads to the same Neyman structure (weights sqrt(α_i²+α_j²) instead of |α_i α_j|), the allocation principle is likely to transfer to other kernel-based losses whose relevant entries are concentrated on large dual coefficients.","The active-set instability term is introduced heuristically; a sharper probabilistic model of threshold crossings could yield a formal prior over support-vector membership and improve the early rounds.","The early-stopping criterion could also serve as an online change detector: in non-stationary environments, a rise in dual instability would signal when the kernel estimate needs fresh measurement.","The paper's regime analysis suggests that a practical system should estimate the Gini or heterogeneity of α before choosing adaptive versus uniform allocation; this is a testable decision rule the authors do not explicitly propose."],"forward_implications":["With a fixed total shot budget, adaptive allocation improves decision-function and margin estimation relative to uniform sampling on synthetic and quantum-kernel tasks, and reconstructs the support-vector block of the kernel much more accurately even though global kernel RMSE degrades.","The dual-coefficient stability criterion δ_r < ε is an observable stopping rule: at the critical threshold ε*≈0.5 it terminates in a median of six rounds using about 16% of the shot budget while still beating uniform allocation.","The gain from adaptivity is controlled by weight heterogeneity: in low-structure problems (small Gini of α) uniform allocation can tie or win, while in high-structure problems the adaptive rule gives large reductions in decision error.","On quantum kernels from real data, adaptive allocation yields significant gains at intermediate qubit numbers and budgets, and extends feasible learning before kernel concentration degrades both methods.","Under a quantum/classical cost model with typical small τ, the adaptive scheme is cost-efficient: the break-even τ* grows with n and favors adaptivity once measurements dominate classical retraining costs."],"fun_headline_variants":["Spend shots where they matter: adaptive kernel measurements boost SVM","Neyman-style kernel sampling cuts shots 84% while keeping SVM fidelity","Margin-sensitive measurement allocation trumps uniform for noisy kernels","Adaptive allocation for kernel SVMs: more accuracy, fewer measurements","Task-aware shot budgeting for noisy kernel SVM learning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole allocation rule stands on the claim that a small perturbation of a kernel entry leaves the SVM's optimal weights on training points unchanged, so an entry's influence on the margin is just the product of those weights.","fun_headline_variants_meta":{"raw":{"variants":["Spend shots where they matter: adaptive kernel measurements boost SVM","Neyman-style kernel sampling cuts shots 84% while keeping SVM fidelity","Margin-sensitive measurement allocation trumps uniform for noisy kernels","Adaptive allocation for kernel SVMs: more accuracy, fewer measurements","Task-aware shot budgeting for noisy kernel SVM learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000431,"raw_usage":{"total_tokens":2058,"prompt_tokens":784,"completion_tokens":1274,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":1189}},"tokens_in":528,"tokens_out":1274,"duration_ms":11882,"temperature":1.0,"reasoning_tokens":1189,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:25:18.188500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact derivative of ||w||² with respect to K_12 in a two-point SVM with K_11=K_22=1, K_12=k, opposite labels, and no box constraint. The paper's sensitivity formula gives −1/(1−k)², while direct differentiation of the exact margin ||w||²=2/(1−k) gives +2/(1−k)²; this single toy calculation settles whether the sensitivity identity at the core of the allocation rule is correct.","supporting_citations":[],"review_version":2}