{"id":"f7ee5d24-0994-4a83-a7e2-b4ac41d4bb56","arxiv_id":"2502.01704","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SubsCoRe adaptively chooses the number of quantum measurement shots per VQE optimization step to keep GP uncertainty over the updated subspace below a threshold, cutting total shot cost versus fixed-shot baselines.","lead":"This paper presents SubsCoRe, a method that adaptively sets the number of measurement shots used at each step of a variational quantum eigensolver so that the Gaussian-process uncertainty over the optimized subspace stays below a target threshold. It matters because quantum measurements dominate VQE cost, and this approach spends fewer shots early in optimization and more near convergence.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed minimum-shot optimality of SubsCoRe is not established: Theorem 3.1 proves only uniform posterior variance for equidistant no-prior observations, not global minimality, and Algorithm 2 applies it heuristically with previous data.","rationale":"The paper is trying to do something specific and useful: adaptively control the number of quantum measurement shots in SMO-based VQE so that the total quantum budget is minimized while the updated subspace remains in the GP confident region. For the central claim to hold, two things must be true: (i) the proposed equidistant, equal-variance observation scheme must actually solve or reliably approximate the constrained minimization in Eq. (13), and (ii) the CoRe threshold must translate into a meaningful accuracy guarantee for the subspace optimization. The second point is weaker than it appears: a posterior-variance threshold bounds predictive uncertainty, but it does not by itself bound the error of the argmin of the GP mean without additional smoothness or curvature assumptions. The first point is the more structural gap. Theorem 3.1 is internally sound for the limited statement it proves: with only 1+2V_d equidistant observations, equal observation variance, and no previous training data, the posterior variance is uniform along the subspace. However, the text goes further and asserts min-max optimality, and no extremal comparison supports that assertion. Moreover, the actual SubsCoRe algorithm always operates with previous GP training data, and its shot-allocation subroutine performs a grid search over observation variances rather than optimizing Eq. (13) directly. Therefore, the 'minimum total number of measurement shots' claim is not established by the presented theory. This is a correctness risk in the paper's central contribution, not merely a missing experiment. The empirical demonstration on one Ising Hamiltonian with 100 seeded trials is encouraging, and the released code is a real asset, but a single benchmark cannot establish optimality. The reader's weakest-assumption concern about unknown single-shot variance eta^2 is also legitimate and practically important, since a misestimated eta^2 would make the computed shot counts inaccurate; however, it is an implementation-parameter issue that could be addressed by variance estimation, whereas the missing optimality proof is a structural gap in the argument. Accordingly, the verdict should remain CONDITIONAL: the method may well work well, but the paper should either prove the minimality claim under its stated assumptions or explicitly reframe it as a heuristic, and it should clarify that the 'guaranteed accuracy' is a posterior-variance guarantee, not a direct bound on optimization error.","tokens_in":24194,"tokens_out":5092,"duration_ms":50761,"concrete_test":"For a single SMO step with V_d = 1 and no prior observations, fix a required CoRe threshold kappa^2 and compute, by direct evaluation of the GP posterior variance (Eq. 5), the minimal total number of shots over a fine grid of observation placements and shot splits. Compare this optimum against the equidistant equal-variance allocation prescribed by SubsCoRe. If any non-equidistant placement or unequal split satisfies the CoRe constraint with fewer total shots, the 'minimum' claim fails. Repeat the same comparison with one or two prior observations on the subspace (as in Algorithm 2); the posterior variance will no longer be uniform, so this tests whether the no-prior theorem transfers and whether the Algorithm 2 grid search actually finds the minimal shot count.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion is that SubsCoRe 'uses the minimum total number of measurement shots and distributes them optimally' so that each SMO subspace lies in the CoRe (Section 3, Eq. 12). The proof supplied (Theorem 3.1, Appendix C) establishes something narrower: for observations at 1+2V_d equidistant points with equal variance, and with no earlier training points, the GP posterior variance along the subspace is uniform and given by Eq. (14). Uniformity does not imply min-max optimality: no comparison is made to any other placement of the same number of shots, to unequal splits, or to configurations that include previous data. The footnote in Section 3 asserts that the equidistant design is 'min-max optimal', but the appendix contains no extremal argument. In Algorithm 2, the actual method uses previous training data X_t (lines 6 and 21) and a grid search over observation variances, so neither Theorem 3.1 nor Corollary 3.2 applies to the posterior that determines feasibility; the resulting shot count is a heuristic feasible solution to the relaxed problem (15), not a proven minimum of Eq. (13). This matters because the paper's claimed advantage over EMICoRe is precisely adaptive cost minimization: if an alternative shot allocation meets the CoRe constraint with fewer total shots, the headline efficiency result is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SubsCoRe, an adaptive shot-control method for sequential minimal optimization (SMO) in variational quantum eigensolvers. In each SMO step, SubsCoRe measures the fixed set of 1+2V_d equidistant points along the updated axis and chooses the per-point number of quantum measurement shots so that the GP posterior variance along the entire 1D subspace stays below an adaptively updated threshold κ_t^2. The theoretical core is Theorem 3.1, which gives a closed-form expression for the posterior variance after observing 1+2V_d equidistant points with equal variance and no prior training points, showing that this variance is uniform along the subspace. The paper then proposes SubsCoRe-Bound and SubsCoRe-Center variants that use this theory to set observation noise levels, and it reports experiments on a 5-qubit Ising Hamiltonian comparing SubsCoRe-Center with NFT, EMICoRe, and SGLBO over 100 seeded trials, with statistically significant improvements in energy and fidelity at a fixed cumulative shot budget.","tokens_in":24479,"tokens_out":3601,"duration_ms":36416,"significance":"If the claims are correct, the paper makes a useful practical contribution: an adaptive shot-allocation rule for SMO that reduces total quantum measurement cost while maintaining a controlled GP uncertainty level. The strengths are concrete: the analytic derivation in Appendix C is detailed, the experimental comparison uses 100 independent trials and an external significance test, the code is publicly linked, and the comparison includes three state-of-the-art baselines. However, the headline claim of using the minimum total number of shots with an optimal distribution is not established by the supplied proof, and the actual Algorithm 2 is a heuristic feasibility search. The significance is therefore conditional on revising these claims or supplying the missing optimality argument.","major_comments":[{"comment":"The theorem proves that, for a fixed set of 1+2V_d equidistant points with equal observation variance and no previous training data, the GP posterior variance is uniform along the subspace. Uniformity alone does not imply min-max optimality: no comparison is made to other placements, unequal shot splits, or configurations with previous data. The footnote in Section 3 asserts that the equidistant design is 'min-max optimal', but Appendix C contains no extremal argument. Either provide a proof of this optimality or remove/soften the optimality claim.","section":"Section 3, Theorem 3.1 and Eq. (14)"},{"comment":"The actual method does not solve Eq. (13): Algorithm 2 uses previous training data X_t (lines 6 and 21) and performs a grid search over observation variances, so Theorem 3.1 and Corollary 3.2, which assume no prior points and equal variances, do not apply to the posterior that determines feasibility. The resulting shot count is therefore a feasible solution to the relaxed problem (15), not a proven minimum. The paper's central assertion that SubsCoRe 'uses the minimum total number of measurement shots and distributes them optimally' is unsupported by the supplied theory and should be revised, or supported by an additional optimality argument or a numerical optimality-gap study.","section":"Section 3, Eq. (13) and Algorithm 2"},{"comment":"The phrase 'guaranteed accuracy' overstates what the CoRe condition provides: the constraint S_d(x) ⊆ Z(κ_t^2) controls the posterior variance, not the error of the resulting SMO minimum. The energy error also depends on the GP mean bias, kernel hyperparameters, the threshold schedule, and the fact that the threshold κ_t is updated heuristically via Eq. (16). In addition, Algorithm 2 assumes the single-shot variance η^2 is known exactly; if η^2 is misestimated, the computed shot counts do not realize the intended observation noise, and the posterior variance condition is not actually enforced. These caveats should be stated explicitly where the accuracy guarantee is claimed.","section":"Section 3, Eq. (12) and abstract"}],"minor_comments":[{"comment":"The quantum-circuit call uses 'shots= ¯N0' rather than the newly computed per-point shot counts from Algorithm 2; this appears to be a typo, since lines 8-12 of Algorithm 2 return the vector ¯N^{t+1} with potentially different values for the center and the shifted points.","section":"Appendix D.2, Algorithm 1, line 10"},{"comment":"The statistical test is reported only as 'p-value < 0.05 according to the Wilcoxon signed-rank test'; please specify the exact test setup (comparison at the final budget or at each cumulative budget, one-sided or two-sided, and what is being compared, final ΔEnergy/ΔFidelity or trajectories).","section":"Section 4.2, Figure 3"},{"comment":"SubsCoRe-Center is described as solving the problem by a 2D search over ˘σ_0^2 and ˘σ^2, but Algorithm 2 performs two sequential 1D grid searches: first for the shifted points, then for the center. Please clarify whether this is meant to be an approximate 2D search or a specific decoupled heuristic.","section":"Section 3, text after Eq. (15)"},{"comment":"The abstract says SubsCoRe 'outperforms the state-of-the-art methods' without qualification; the experiments cover a single Ising benchmark, one circuit family, and one qubit/layer setting. Please qualify the empirical claim accordingly, e.g., 'on the tested Ising benchmark'.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper's core method and experiments are credible, and the proof in Appendix C is careful. The main issue is that the paper's most prominent claim—minimum total shot count and optimal distribution—is not supported by the theorem as stated, since Theorem 3.1 establishes uniformity, not global minimality, and Algorithm 2 is a heuristic feasibility search. This is fixable by rewording and adding caveats, but it is load-bearing for the stated contribution, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dan,\n\nQuick summary: this paper gives SMO-style VQE an adaptive measurement-shot controller. The new piece is Theorem 3.1, which shows that if you observe 1+2V_d equidistant points along an axis with equal noise variance, the GP posterior variance (using their VQE kernel) is constant along that whole line. That is a real, non-obvious derivation, and it connects nicely to Fourier analysis. The empirical work is honest: they compare against NFT, SGLBO, and EMICoRe on a 5-qubit Ising model, 100 trials, and show consistent improvement in energy and fidelity under a fixed shot budget. The code is public. So the core idea is sound and the experiments back it.\n\nBut the headline is stronger than what is proven. The abstract and Section 3 say SubsCoRe uses \"the minimum total number of measurement shots and distributes them optimally.\" Theorem 3.1 gives uniform posterior variance for a specific design with no previous training data. It does not prove that design is min-max optimal among all designs, nor that it minimizes shot count for the constrained problem in Eq. (13). The footnote in Section 3 asserts min-max optimality, but there is no extremal argument anywhere. Meanwhile, Algorithm 2 actually uses previous training data and a grid search over variances, so the theorem does not apply to the posterior that determines feasibility. The resulting shot count is a heuristic feasible solution, not a proven minimum. That matters because adaptive cost minimization is the entire selling point over EMICoRe.\n\nThere is also a pseudocode bug: Algorithm 1 line 10 measures with the initial ¯N0 instead of the newly chosen ¯N^{t+1}_i. And the phrase \"guaranteed accuracy\" is too strong—the CoRe condition is a posterior-variance threshold, not a bound on optimization error. The reliance on knowing the single-shot variance eta^2 exactly is a real weakness; misestimate it and the subspace may not actually lie in the CoRe. These are fixable, but they need to be addressed in revision.\n\nThe limitations section is honest about hardware noise, and the experiments cover only one Hamiltonian. That is a scope limitation rather than a fatal flaw.\n\nBottom line: the mathematical core is solid, the idea is useful, and the empirical trend is credible. But the central optimality claim is unsupported, and the pseudocode needs correcting. I would send this to peer review, and I would expect referees to ask for the min-max claim to be either proven or softened, and the pseudocode fixed.\n\nWould I cite it? Probably, for the uniformity theorem, once it is cleaned up.","headline":"A clean uniformity theorem and a sensible adaptive shot-count heuristic for SMO-based VQE, but the 'minimum shots' claim is not actually proven and the paper overstates its guarantee.","tokens_in":25089,"tokens_out":1500,"would_cite":true,"duration_ms":14965,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adaptive shot-count control—keeping the optimized subspace inside a Gaussian-process confident region—can substantially reduce the total quantum measurement budget of a variational quantum eigensolver without…","keywords":["variational quantum eigensolver","sequential minimal optimization","Gaussian process","confident region","adaptive shot control","shot noise","quantum measurement cost","VQE kernel"],"falsifier":"Run SubsCoRe on a simulator while tracking the empirical GP posterior variance at points in the subspace: if, with the shots chosen from a deliberately misestimated $η^{2}$ (say, half the true value), the posterior variance exceeds $κ_t^{2}$ on any tested point, the claimed guarantee fails. The same check with the true $η^{2}$ should keep the subspace inside the CoRe.","tokens_in":23962,"feed_emoji":"⚛️","tokens_out":3897,"duration_ms":33079,"temperature":0.7,"pith_summary":"The paper claims that the quantum measurement cost of running a variational quantum eigensolver (VQE) can be substantially reduced by adapting, in every sequential-minimal-optimization step, the number of measurement shots to the current need for accuracy. Its method, SubsCoRe, keeps the whole one-dimensional subspace being optimized inside the 'confident region' of a Gaussian-process surrogate, where the surrogate's uncertainty stays below a threshold $κ_t^{2}$. The threshold is tuned to the optimization progress: coarse early, fine near convergence. If the claim holds, users of VQE-type algorithms can spend far fewer total shots—the dominant cost on quantum hardware—without losing energy accuracy or state fidelity.","feed_headline":"Adaptive measurement budgets shrink VQE quantum cost","feed_subtitle":"SubsCoRe spends the minimum shots per SMO step to keep the optimizer inside a confidence region, beating fixed-shot baselines.","key_machinery":"The confident region (CoRe), Equation (11), is the set of inputs where GP posterior variance is below a threshold $κ^{2}$. Theorem 3.1 shows that with the VQE kernel and 1+2V_d equidistant measurements along one axis, the posterior variance is a constant (Eq. 14) independent of the test point along that axis. That constant, together with Corollary 3.2 (posterior variance < observation noise $σ^{2}$), turns the hard shot-allocation problem (Eq. 13) into a simple choice of per-point noise variances σ˘, solved by the SubsCoRe-Center subroutine (Algorithm 2). The threshold κ_t is updated by a linear regression on the recent best values (Eq. 16), tying the required accuracy to the measured progress.","core_discovery":"The central discovery is that, for VQE objective functions, observing 1+2V_d equidistant points along the axis being optimized makes the GP posterior variance exactly uniform along the subspace and independent of the test position α'. That uniformity is the key to the cost allocation: to guarantee the entire subspace lies inside the CoRe at accuracy $κ_t^{2}$, it suffices to set the per-point observation noise to a value tied to $κ_t^{2}$ and distribute shots among the 1+2V_d points, because uniform uncertainty is min-max optimal. The paper's SubsCoRe-Center variant exploits that the previous best point is already inside the CoRe from earlier iterations, thus spending fewer shots there and the minimum total number of shots overall while satisfying Equation (12), the 'subspace in CoRe' constraint. The consequence is an SMO loop whose quantum cost adapts to progress, and which the paper reports to outperform fixed-shot NFT, EMICoRe, and the adaptive-shot SGLBO on the Ising benchmark.","pith_inferences":["If the single-shot variance η^2 is state-dependent and only approximately known, the guarantee in Eq. (12) degrades gracefully: the actual uncertainty scales with η^2/σ^2, so misestimation by a factor of four changes the required shots by the same factor. A robust variant would estimate η^2 online from repeated measurements of the same point.","The min-max argument suggests the equidistant point choice α=2π/(1+2V_d) is the right default for any cost-controlled SMO, not just for the Ising Hamiltonian, and could be tested on molecular Hamiltonians where V_d > 1.","A testable extension: monitor the GP's leave-one-out predictive error against the claimed CoRe guarantee; if the observed error regularly exceeds κ, the assumed η^2 is too small, and the algorithm should raise its shot allocation."],"forward_implications":["If SubsCoRe is right, fixed-shot SMO methods waste a large fraction of their quantum budget: shots spent early, where accuracy is not yet needed, can be moved to convergence without degrading the final energy.","The same CoRe-criterion applies to any VQE ansatz built from parameterized rotations, since the trigonometric-polynomial form (Eq. 9) holds for any such circuit.","SubsCoRe-Center combined with EMICoRe's point-selection would, by the paper's own account, be possible and likely further reduce total shots (the authors list it as future work).","The per-step shot count becomes a direct function of the required accuracy κ_t^2 times the known single-shot variance, which gives practitioners a principled way to set measurement budgets rather than tuning fixed shot counts."],"supporting_citations":[{"why":"Establishes the low-order trigonometric polynomial form of the VQE objective and the NFT sequential minimal optimization that fits a 1D subspace from 1+2V_d observations.","marker":"Nakanishi et al. (2020)"},{"why":"Introduces the VQE kernel and the notions of confident region and EMICoRe that SubsCoRe builds upon and extends with adaptive shot control.","marker":"Nicoli et al. (2023a)"},{"why":"Supplies the Gaussian-process posterior mean and covariance formulas (Eqs. 3–5) that the theory and the shot-allocation subroutines rely on.","marker":"Rasmussen & Williams (2006)"},{"why":"SGLBO is the adaptive-shot gradient-based baseline that SubsCoRe must outperform, and it provides the earlier adaptive-shots idea for gradient-based methods.","marker":"Tamiya & Yamasaki (2022)"}],"fun_headline_variants":["VQE shots adapt per step to keep subspace in confidence","GP uniform variance enables minimal-shot VQE updates","Subspace confidence regions set adaptive VQE observation budgets","Cost-efficient VQE via uniform-uncertainty shot allocation","Adaptive measurement distribution cuts VQE quantum cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole shot-allocation rule assumes the single-shot measurement variance $η^{2}$ is known exactly; if it is misestimated, the computed shot counts will not actually place the subspace inside the confident region and the accuracy guarantee is void.","fun_headline_variants_meta":{"raw":{"variants":["VQE shots adapt per step to keep subspace in confidence","GP uniform variance enables minimal-shot VQE updates","Subspace confidence regions set adaptive VQE observation budgets","Cost-efficient VQE via uniform-uncertainty shot allocation","Adaptive measurement distribution cuts VQE quantum cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000773,"raw_usage":{"total_tokens":3416,"prompt_tokens":932,"completion_tokens":2484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2407}},"tokens_in":548,"tokens_out":2484,"duration_ms":16939,"temperature":1.0,"reasoning_tokens":2407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:39:12.933902+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SubsCoRe on a simulator while tracking the empirical GP posterior variance at points in the subspace: if, with the shots chosen from a deliberately misestimated $η^{2}$ (say, half the true value), the posterior variance exceeds $κ_t^{2}$ on any tested point, the claimed guarantee fails. The same check with the true $η^{2}$ should keep the subspace inside the CoRe.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian-process posterior mean and covariance formulas (Eqs. 3–5) that the theory and the shot-allocation subroutines rely on."}],"review_version":1}