{"id":"85930afc-e990-4729-8405-5087527a2004","arxiv_id":"2412.19176","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The hybrid optimizer QN-SPSA+PSR, mixing SPSA metric estimates with parameter-shift gradients, converges faster and more stably than QN-SPSA+SPSA on a 12-spin Ising model.","lead":"Researchers tested classical and quantum optimizers for the Variational Quantum Eigensolver on the Ising model and introduced a hybrid optimizer called QN-SPSA+PSR. It combines a stochastic approximation of the quantum metric with exact parameter-shift gradients, and the simulations suggest faster, more stable convergence than the all-stochastic alternative.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-iteration convergence plots ignore the 2p PSR gradient cost, so QN-SPSA+PSR's claimed low-overhead advantage over Cobyla and QN-SPSA+SPSA is not established.","rationale":"The reader's CONDITIONAL verdict already captures much of this: the single benchmark and missing variance weaken generalization, and the per-iteration comparison was mentioned in the rationale. My stress test makes the cost-normalization issue the primary concern because it directly targets the phrase 'low computational consumption' in the central claim. The method itself is a straightforward composition of existing techniques (QN-SPSA metric, PSR gradient), with code provided, so the concern is not fatal; an author response could add cost-normalized plots and seed statistics. For these reasons I do not move the verdict; it should remain CONDITIONAL pending those checks.","tokens_in":17986,"tokens_out":6709,"duration_ms":65549,"concrete_test":"Use the supplied GitHub code to rerun the 12-spin RealAmplitudes benchmark and plot relative error against cumulative quantum circuit evaluations, with per-iteration costs: Cobyla=1, SPSA=2, FD=2p, QN-SPSA+SPSA=6, QN-SPSA+PSR=2p+4, QN-BDA+PSR=2p+L. If QN-SPSA+PSR no longer dominates Cobyla or QN-SPSA+SPSA at equal evaluation budgets, the low-overhead advantage fails; also report mean±std over at least 30 seeds to quantify stability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests on an unfair resource comparison. Table I lists PSR gradient cost as 2p and QN-SPSA metric cost as 4, so QN-SPSA+PSR costs about 2p+4 expectation evaluations per iteration; for the 12-qubit, two-layer RealAmplitudes benchmark p=36, that is about 76 evaluations per update versus 6 for QN-SPSA+SPSA and 1 for Cobyla. Figures 4 and 5 plot relative error only against iterations, so faster apparent convergence per iteration can simply mean spending many more circuit evaluations per step. Because VQE cost is dominated by circuit executions, the abstract's 'low computational consumption' claim and the stated outperformance of Cobyla are not supported without cost-normalized curves. The hybrid is plausible and no internal mathematical error is apparent, but this is a load-bearing gap in the empirical argument. The absence of error bars on seven stochastic samples additionally leaves 'stability' unquantified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies VQE for the 1D transverse Ising model on up to 12 qubits, comparing classical optimizers (COBYLA, finite differences, SPSA) with quantum natural-gradient-based optimizers (QN-BDA+PSR, QN-SPSA+SPSA) and proposing a hybrid optimizer, QN-SPSA+PSR, which approximates the Fubini-Study metric via 2-SPSA while computing the cost-function gradient exactly via the parameter-shift rule. The authors motivate a RealAmplitudes ansatz from the model's real-representation, local-interaction, and Z2 symmetry properties, and they report that QN-SPSA+PSR converges faster per iteration and more stably than QN-SPSA+SPSA and matches QN-BDA+PSR at lower estimated cost. The numerical evidence is a noiseless benchmark at h=2, J=1, with a fixed two-layer ansatz and seven samples for stochastic methods.","tokens_in":18164,"tokens_out":3555,"duration_ms":36105,"significance":"The proposed QN-SPSA+PSR method is a plausible and simple combination of existing ideas; if its practical advantage were established with a fair resource comparison, it would be a useful addition to the VQE optimizer toolbox. The paper also gives a transparent complexity table (Table I), releases source code, and uses physical symmetries of the transverse Ising model to justify the ansatz, which is a helpful methodological point. However, as presented, the central empirical claim is not yet established: the convergence plots compare methods per iteration rather than per quantum-circuit evaluation, and the stochastic results lack error bars. The significance of the work therefore depends on the outcome of the requested cost-normalized and statistically grounded revision.","major_comments":[{"comment":"The central claim that QN-SPSA+PSR has 'low computational consumption' and outperforms COBYLA is not supported by the reported convergence plots, because the x-axis is iterations rather than quantum-circuit evaluations. From Table I, one QN-SPSA+PSR iteration costs 2p expectation evaluations for the PSR gradient plus 4 for the QN-SPSA metric; for the 12-site, two-layer RealAmplitudes benchmark (p=36) this is about 76 circuit evaluations, versus 6 for QN-SPSA+SPSA and 1 for COBYLA. Figures 4 and 5 should be replotted against a resource measure such as total circuit executions or objective-function evaluations, or the algorithmic claim should be restricted to per-iteration convergence speed rather than total computational cost.","section":"§IV.B, Figs. 4–5; Table I"},{"comment":"Stochastic methods are evaluated on only seven samples, and no error bars, seed information, or per-run spread is reported anywhere in the manuscript. Since Section IV.B explicitly states that stochastic methods are 'evaluated based on seven samples,' the qualitative claims that QN-SPSA+PSR is 'more stable' and 'consistently delivered reliable and accurate results' (Figs. 6–7) are unquantified; a statement about stability requires at least variance or confidence intervals over random SPSA perturbations and over random initializations.","section":"§IV.B, Figs. 4–7"},{"comment":"The ansatz layer count L=2 is selected after 'conducting the experimental survey,' and the same benchmark is then used to demonstrate the optimizer comparison. This creates a selection-on-the-benchmark concern: the reported results do not show that L=2 was chosen by a predetermined rule, nor do they show that the optimizer conclusions survive across other L values, system sizes, or field strengths. Please report the layer survey and at least one out-of-sample check (for example, a different N or h) to support the generalization claims made in the conclusion.","section":"§II.C.2, Eq. (13), and §IV.B"}],"minor_comments":[{"comment":"Phrases such as 'quantum supremacy' and 'potential quantum supremacy' overstate what a noiseless 12-qubit comparative study can establish; replace them with 'potential advantage' or similar calibrated wording.","section":"Abstract and Conclusion"},{"comment":"The keyword 'Ansazt' and the phrase 'Fubini-study metric' should be corrected, and the first sentence of Section II.A has an apparent grammatical issue that should be reworded.","section":"Keywords and Section II.A"},{"comment":"Figure 3 is described as 'illustrative' and partly based on 'bias-informed conjectures'; please state explicitly in the caption that this figure is a schematic and not a numerical result, so that it is not mistaken for simulation data.","section":"Fig. 3 caption"},{"comment":"The counting argument leading from 2^{N+1}-2 to 2^{N-1}-1 is compressed; please expand the steps and state explicitly which ansatz parameterization and symmetry constraints are being assumed so that Eq. (13) can be checked.","section":"Eq. (12) and Eq. (13)"},{"comment":"References [6] and [8] are the same Nature paper and should be consolidated to avoid duplicate citation.","section":"References"},{"comment":"The data availability statement refers to 'datasets' while the linked repository appears to contain code; please clarify whether raw outputs, processed data, or only scripts are archived, and consider using a versioned DOI for reproducibility.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of a quantum-computing journal and the proposed hybrid is a reasonable incremental contribution. The main blocking issue is the resource accounting: Figs. 4 and 5 plot per-iteration convergence even though the methods differ by an order of magnitude in per-iteration circuit evaluations, so the headline 'low computational consumption' claim is not currently supported. The absence of error bars on the seven-sample stochastic runs is a second load-bearing gap. I would encourage the editor to request a revision with cost-normalized convergence curves, explicit statistical measures, and at least one out-of-sample benchmark rather than rejecting the work, since the proposed method is plausible and the code is available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the QN-SPSA+PSR scheme is a real but small hybrid—replace the SPSA gradient in QN-SPSA with the parameter-shift rule—and the paper is clear about what it does. The main claim, though, rests on a comparison that ignores the cost of the exact gradient. The authors' own Table I gives PSR cost 2p; for the 12-site, two-layer RealAmplitudes ansatz, p=36, so QN-SPSA+PSR uses about 76 expectation evaluations per update versus 6 for QN-SPSA+SPSA and 1 for COBYLA. Plotting only against iterations makes it look cheap, but the abstract's 'low computational consumption' claim and the stated outperformance of COBYLA are not established. That is a load-bearing gap, not a style issue.\n\nWhat's good: the paper ships code and data links, the ansatz discussion using TIM symmetries is a useful bit of context, and the hybrid itself is plausible—no mathematical red flags. The external exact-energy target is fine; there is no circularity here. The complexity table is a handy summary.\n\nOther soft spots: the benchmark is a single noiseless 12-site Ising with h=2, J=1, seven samples per stochastic run, and no error bars. 'Stability' is asserted but not quantified. The layer number L=2 appears chosen after seeing the data. And the abstract's 'quantum supremacy' phrase should go; that is not what is shown and it will only draw ridicule.\n\nWho it's for: someone picking an optimizer for small VQE benchmarks might find this a useful data point, and as a teaching case for cost-normalized comparisons it is excellent. It deserves a serious referee—the idea is legitimate and the code is there—but I'd ask for major revision: cost-normalized convergence curves, error bars, a second test case, and a careful rewrite of the abstract. As it stands, I wouldn't rely on the performance claims.","headline":"Useful hybrid, but the cost-comparison flaw and an overreaching abstract mean the central claims don't hold yet.","tokens_in":18754,"tokens_out":2512,"would_cite":false,"duration_ms":25325,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68"],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"The paper claims that a hybrid optimizer, QN-SPSA+PSR, improves VQE convergence on the transverse Ising model by keeping the cheap stochastic Fubini-Study metric of QN-SPSA while computing the cost gradient exactly with the…","keywords":["Variational Quantum Eigensolver","Quantum Natural Gradient","Parameter-Shift Rule","Simultaneous Perturbation Stochastic Approximation","Transverse Ising model","NISQ","QN-SPSA+PSR","Ansatz construction"],"falsifier":"Run QN-SPSA+PSR against QN-SPSA+SPSA and COBYLA on a noisy simulator or real device with shot noise, or on a larger lattice and other values of $h$, and report the distribution of final energies over many random seeds; if the stability margin over QN-SPSA+SPSA vanishes or COBYLA wins, the claimed advantage does not generalize.","tokens_in":17749,"feed_emoji":"⚛️","tokens_out":7232,"duration_ms":59068,"temperature":0.7,"pith_summary":"The paper claims that a hybrid optimizer, QN-SPSA+PSR, improves the optimization subroutine of the Variational Quantum Eigensolver (VQE) on the transverse Ising model. The method keeps the cheap Fubini-Study metric estimate of QN-SPSA, which costs four quantum-circuit evaluations per update, but replaces the stochastic gradient of the original scheme with the exact Parameter-Shift Rule gradient. On a 12-spin noiseless benchmark with a two-layer RealAmplitudes ansatz, the authors report that it converges more stably and faster than QN-SPSA+SPSA, outperforms COBYLA and Finite Difference, and comes close to the much costlier QN-BDA+PSR. The same study argues that a symmetry-informed RealAmplitudes ansatz with linear entanglement is sufficient for the Ising model. A sympathetic reader would care because the claimed combination targets the main practical bottleneck of VQE on near-term quantum devices: convergence quality per quantum-circuit evaluation.","feed_headline":"Hybrid quantum optimizer beats Cobyla and Finite Difference in VQE","feed_subtitle":"Cheap stochastic metric plus exact parameter-shift gradients matches costlier results on 12 spins.","key_machinery":"The central object is the QN-SPSA+PSR update rule, $\\theta_{k+1} = \\theta_k - \\eta_k \\tilde{g}^+(\\theta_k) \\nabla f(\\theta_k)$, where $\\tilde{g}$ is the QN-SPSA stochastic approximation to the Fubini-Study metric obtained from four circuit evaluations per iteration and $\\nabla f$ is the exact Parameter-Shift Rule gradient. The stability gain comes from replacing the stochastic SPSA gradient of the predecessor with the exact PSR gradient, leaving stochasticity only in the metric. The argument also leans on the symmetry-informed RealAmplitudes ansatz, whose real parameters exploit the transverse Ising model's real eigenstates and total spin-flip symmetry, reducing the parameter count.","core_discovery":"QN-SPSA+PSR is proposed as an extension of QN-SPSA+SPSA: the Fubini-Study metric is approximated stochastically via second-order SPSA using four quantum expectations per iteration, while the outer gradient of the cost function is computed analytically by the Parameter-Shift Rule. This removes the gradient stochasticity that makes QN-SPSA+SPSA unstable, while retaining the metric's low computational cost. Numerical experiments on the 12-spin transverse Ising model at $h=2$ and $J=1$ show it reaching the ground-state energy with relative error declining faster than COBYLA and Finite Difference, closely approaching QN-BDA+PSR's accuracy, and with more stable convergence than QN-SPSA+SPSA. Additional scans over external field and qubit count show the method tracking the exact average ground-state energy alongside QN-BDA+PSR.","pith_inferences":["The paper's noiseless, seven-sample benchmark does not probe shot-noise-dominated or flat-landscape regimes, so the stability advantage may shrink exactly where stochastic gradients usually help.","A direct extension would be to swap the SPSA gradient back in only when noise is high, producing a noise-adaptive interpolation between QN-SPSA+SPSA and QN-SPSA+PSR.","The symmetry-reduction argument for the ansatz likely transfers to any Hamiltonian that is real and symmetric under a global spin flip, so the same construction could be tested on other spin models.","Reporting per-seed convergence curves rather than only averaged relative errors would let practitioners check whether the improvement is consistent or driven by a few runs."],"forward_implications":["VQE on near-term devices could obtain near-natural-gradient convergence at a cost of roughly four quantum evaluations for the metric plus two parameter-shift evaluations per gradient, independent of the number of parameters.","The symmetry-guided RealAmplitudes ansatz with linear entanglement is put forward as sufficient for the transverse Ising model, simplifying the circuit without losing accuracy.","QN-SPSA+PSR's reported accuracy on average ground-state energies across external fields and qubit numbers positions it as a practical optimizer choice for Ising-type Hamiltonians.","The method's per-iteration overhead is small enough that the authors suggest it can be carried over to other variational quantum algorithms, including QAOA and quantum machine learning training."],"supporting_citations":[{"why":"provides the SPSA approximation of the quantum Fisher information used to estimate the Fubini-Study metric.","marker":"[51]"},{"why":"supplies the parameter-shift rule that computes the exact gradient in the proposed optimizer.","marker":"[46]"},{"why":"introduces the quantum natural gradient and the block-diagonal approximation used by the QN-BDA+PSR baseline.","marker":"[50]"},{"why":"supplies the simultaneous perturbation stochastic approximation concept behind the metric estimator and the SPSA baseline.","marker":"[45]"},{"why":"introduces the variational quantum eigensolver that forms the algorithmic setting.","marker":"[10]"},{"why":"motivates hardware-efficient ansatz construction, supporting the RealAmplitudes choice.","marker":"[22]"},{"why":"is the COBYLA derivative-free optimizer treated as a classical baseline.","marker":"[41]"}],"fun_headline_variants":["Stochastic metric plus exact gradient stabilizes VQE optimizer","QN-SPSA+PSR: low-cost metric, exact gradients for faster VQE","New VQE optimizer uses hybrid gradient: stochastic metric + exact rule","Hybrid optimizer: stochastic metric, exact gradients, beats classics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed advantage rests on a single noiseless 12-spin transverse Ising benchmark at $h=2$, $J=1$, with a two-layer RealAmplitudes ansatz and only seven samples per stochastic run, with no variance reported; whether the improvement persists across system sizes, fields, ansätze, or under device noise is assumed rather than demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Stochastic metric plus exact gradient stabilizes VQE optimizer","QN-SPSA+PSR: low-cost metric, exact gradients for faster VQE","New VQE optimizer uses hybrid gradient: stochastic metric + exact rule","Hybrid optimizer: stochastic metric, exact gradients, beats classics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0009,"raw_usage":{"total_tokens":3897,"prompt_tokens":990,"completion_tokens":2907,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2829}},"tokens_in":606,"tokens_out":2907,"duration_ms":21291,"temperature":1.0,"reasoning_tokens":2829,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:52:17.316020+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run QN-SPSA+PSR against QN-SPSA+SPSA and COBYLA on a noisy simulator or real device with shot noise, or on a larger lattice and other values of $h$, and report the distribution of final energies over many random seeds; if the stability margin over QN-SPSA+SPSA vanishes or COBYLA wins, the claimed advantage does not generalize.","supporting_citations":[{"cited_title":"The newuoa software for uncon- strained optimization without derivatives","cited_arxiv_id":null,"evidence_quote":"supplies the parameter-shift rule that computes the exact gradient in the proposed optimizer."},{"cited_title":"General parameter-shift rules for quantum gra- dients","cited_arxiv_id":null,"evidence_quote":"provides the SPSA approximation of the quantum Fisher information used to estimate the Fubini-Study metric."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the quantum natural gradient and the block-diagonal approximation used by the QN-BDA+PSR baseline."},{"cited_title":"Uobyqa: unconstrained optimiza- tion by quadratic approximation.Mathematical Program- ming, 92(3):555–582, 2002","cited_arxiv_id":null,"evidence_quote":"supplies the simultaneous perturbation stochastic approximation concept behind the metric estimator and the SPSA baseline."},{"cited_title":"Strong quan- tum computational advantage using a superconducting quantum processor","cited_arxiv_id":null,"evidence_quote":"introduces the variational quantum eigensolver that forms the algorithmic setting."},{"cited_title":"Besides that, applications of VQE in drug discovery [24, 25] and materials [26–28] are consid- erably concerned","cited_arxiv_id":null,"evidence_quote":"motivates hardware-efficient ansatz construction, supporting the RealAmplitudes choice."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the COBYLA derivative-free optimizer treated as a classical baseline."}],"review_version":1}