{"id":"626409ef-e086-4807-8caf-d63bfcc5c29a","arxiv_id":"2607.17804","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A stochastic Pauli-path simulator with importance reweighting and path automatic differentiation produces unbiased gradient estimates for quantum optimization, demonstrated on VQE at 100 qubits and QNN at 40 qubits.","lead":"This paper introduces SPPS, a classical simulator that estimates unbiased gradients of quantum circuits by randomly sampling Pauli propagation paths rather than truncating them. It offers fast, faithful pre-training of variational quantum eigensolvers with up to 100 qubits and quantum neural networks with up to 40 qubits, all without a quantum computer.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Adaptive A/B stopping in App. D5 is not covered by the fixed-budget unbiasedness theorem and can bias the deployed gradient oracle.","rationale":"The fixed-budget core of SPPS is sound: the estimator g = (nabla Psi)/Pr is unbiased by direct expectation, and the Bernstein-based sample complexity in App. E is reasonable. The value-gradient separation theorem for Tb-PBS is a genuine conceptual contribution with explicit CT/FT/WT constructions. The soft spot is not the fixed-budget algorithm but the practical adaptive scheduling, which is used everywhere in the experiments and is explicitly not covered by the theorem-level confidence bound. This is exactly the reader's weakest_assumption. My independent check of the PAD derivation found a notational slip in App. D3/Eq. D22, but the estimator as defined in Eq. (6) is unbiased, so that slip is not load-bearing. The adaptive stopping concern is load-bearing because unbiasedness is the property that distinguishes SPPS from Tb-PBS and drives Corollary 2; if the deployed estimator is biased, the convergence guarantee no longer applies and the experimental advantage becomes an empirical observation rather than a proven property. The proposed test directly isolates the stopping-rule bias by comparing the adaptive estimator against its own fixed-budget counterpart at a known gradient. Since the reader already made the same identification and issued a conditional verdict, no verdict change is needed; the paper should either prove unbiasedness for the adaptive rule or rerun headline experiments with a fixed, theorem-compliant budget.","tokens_in":40000,"tokens_out":8855,"duration_ms":90813,"concrete_test":"Implement the exact App. D5 adaptive rule on a one- or two-parameter circuit with analytically known gradient, e.g. f(theta) = cos(theta_1) sin(theta_2) with a Pauli observable and product input. Run 10^5 independent adaptive trials with the same delta used in the paper, record the mean final gradient estimate, and compare it with the exact gradient; repeat with fixed-budget SPPS using the average final budget as a control. If the adaptive mean deviates from the exact gradient by more than three standard errors while the fixed-budget control is within noise, the deployed adaptive estimator is biased and Theorem 2 does not cover the implementation. For a direct experimental check, also re-run the 15-qubit TFIM trajectory at one fixed iterate with the adaptive rule versus fixed-budget SPPS and compare to the exact PennyLane gradient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"App. D5 describes the practical SPPS stopping rule: two independent macro-replicates are drawn, and the sample budget is doubled until the proxy hat-Delta_m is below delta, after which the replicates are averaged. Theorem 2 and the theorem-level proof in App. E apply only to a fixed budget B chosen independently of the data: Eq. (E27) fixes B before sampling, and no result bounds estimators whose stopping time is a function of the sampled paths. This gap is load-bearing because the final estimate is a conditional average: E[g | hat-Delta <= delta] is not generically equal to E[g]. For an importance-reweighted path derivative the distribution is skewed and heavy-tailed, so the selection event can shift the conditional mean; the paper's own statement that the proxy is 'not used as a theorem-level confidence bound' concedes the absence of a coverage proof. Every experimental headline, including trajectory tracking, Table I, the 20-100-qubit VQE scaling, and the 40-qubit QNN result, is produced by this adaptive estimator rather than by fixed-budget SPPS. If the adaptive rule is biased, Corollary 2's convergence guarantee does not apply to the reported simulator, and the central distinction from Tb-PBS, which the paper itself attacks for gradient bias, is not established for the deployed algorithm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces SPPS, a stochastic Pauli-path simulator that samples Heisenberg-propagation paths and uses importance reweighting plus 'path automatic differentiation' to estimate gradients of expectation values. It proves a value-gradient separation result for truncation-based PBS, establishes unbiasedness and a high-probability sample-complexity bound for a fixed-budget version of SPPS, and derives an SGD convergence corollary. The experiments cover VQE pre-training up to 100 qubits, QNN pre-training up to 40 qubits, and small state-preparation circuits, reporting trajectory tracking and runtime advantages over CT-, FT-, and WT-PBS baselines.","tokens_in":40237,"tokens_out":7624,"duration_ms":73070,"significance":"If the guarantees hold for the deployed algorithm, this is a meaningful step: Pauli-based simulation would move from forward estimation to optimization, with clear practical value for low-magic variational circuits. The paper has real strengths: the adversarial value-gradient separation in Theorem 1 is explicit and clean; the fixed-budget unbiasedness proof in App. E is self-contained; and the experiments are broad, including a 100-qubit VQE benchmark and a reproducibility link. The main caveat is that the estimator actually used in the experiments includes an adaptive stopping rule that lies outside the theorem-level guarantees, and the main-text sample-complexity statement is internally inconsistent with the appendix. These issues need to be resolved before the headline claims are fully supported.","major_comments":[{"comment":"The main-text Theorem 2 states a sample complexity of O-tilde((1+2a)^L kappa(theta)^L/(a epsilon^2)), but the formal Theorem 5 in Appendix E, Eq. (E27), gives B = O(L kappa(theta)(1+2a)^L log(L/delta)/(a epsilon^2)). Since kappa(theta) is already a product over the L layers, the extra exponent L in the main text is material and makes the two statements disagree. This internal inconsistency must be corrected, and the corrected version should be used in the discussion of when the sampling budget is polynomial.","section":"Section IV.B, Theorem 2; Appendix E, Eq. (E27)"},{"comment":"The adaptive stopping rule described in App. D.5 is not covered by the theoretical guarantees. Theorem 2 and Appendix E fix the sample budget B before sampling, whereas the deployed estimator doubles the budget until the data-dependent proxy Delta-hat falls below delta and then returns a conditional average of the two macro-replicates. The identity E[Delta-hat^2] = E[||g~-g||^2] in Eqs. (D30)-(D31) is a fixed-budget statement; it does not imply E[g~ | Delta-hat <= delta] = g. For the skewed, heavy-tailed importance-reweighted path estimators used here, the selection event can bias the conditional mean. Because all reported experiments use the adaptive rule, Corollary 2's convergence guarantee does not currently apply to the evaluated simulator. The authors should either prove a bias or concentration statement for the stopped estimator or rerun the headline benchmarks with fixed budgets and report whether the results change.","section":"Appendix D.5, Eqs. (D27)-(D31); Section IV.B, Theorem 2"},{"comment":"The derivation of path automatic differentiation in App. D.3 differentiates the reweighted contribution h~_omega(theta) = Psi_omega(theta)/Pr(omega) with respect to theta_j while treating Pr(omega) as constant. However, Pr(omega) depends on theta through q_j in Eq. (D5), so Eq. (D17) is not the derivative of h~_omega. The formal estimator in Appendix E, Eq. (E1), is defined directly as nabla Psi_omega/Pr(omega), which is unbiased, and the score identity in Eq. (6) follows from that definition. The appendix should be corrected so the derivation and the formal estimator agree; otherwise a reader following Eq. (D17) cannot reproduce the main-text formula.","section":"Appendix D.3, Eq. (D17); Section IV.A, Eq. (6)"}],"minor_comments":[{"comment":"The main text should state the restriction epsilon <= sqrt(L kappa(theta)) that appears in the formal Appendix E statement of the high-probability bound, since the stated formula only guarantees the epsilon error in that regime.","section":"Section IV.B, Theorem 2"},{"comment":"The state-preparation benchmark uses a truncated Pauli expansion of the projector observable (keeping the largest 100 non-identity coefficients). This is an additional approximation beyond the full-O expansion in Theorem 2; the text should clarify that this benchmark demonstrates the estimator under observable truncation rather than the untruncated theorem.","section":"Appendix G.3"},{"comment":"The text contains numerous spacing and formatting inconsistencies involving 'Tb-PBS', 'CT-PBS', and related acronyms, and some figure labels (especially Fig. 1) are very hard to read in the provided version. A careful copyedit would improve readability.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the gap between the fixed-budget theorem and the adaptive estimator used in practice. If the authors can close this gap, or convincingly rerun the experiments under fixed budgets, the paper would be substantially stronger. The Theorem 2 vs. Appendix E sample-complexity mismatch should be fixed promptly, as it is the first thing a careful reader will check."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one-sentence version: this is the first PBS I have seen that takes gradient estimation seriously rather than treating truncation as a value-estimation problem, and the fixed-budget estimator is honestly derived; but the implemented adaptive sampling rule is not covered by that derivation, and the headline guarantees do not transfer to the code that actually produced the numbers.\n\nWhat is genuinely new: the value-gradient separation theorem says a truncated Pauli path set can approximate f to ε while keeping gradient error of order one, and the constructions show CT, FT, and WT all realize this separation. That is a clean, useful result. The second contribution, SPPS itself, replaces deterministic truncation with hierarchical path sampling, importance reweighting, and path automatic differentiation. The unbiasedness of the averaged estimator in Eq. (5) is immediate, and the App. E proof of the variance bound for a fixed sample budget B is credible. The small-scale validation against exact PennyLane gradients is the right sanity check, and the runtime comparisons against CT/FT/WT are convincing: on the 15-qubit VQE, SPPS tracks exact GD and lands orders of magnitude closer in energy.\n\nThe soft spots, in order.\n\n1. The adaptive A/B stopping in App. D5 is not covered. The theorem-level unbiasedness and concentration results assume B is fixed before sampling. The deployed rule draws two replicates, checks a proxy Δ̂_m, doubles the budget until the proxy is below δ, and then averages the two replicates. That makes the stopping time a function of the sampled paths, and the conditional mean E[g | Δ̂ ≤ δ] can differ from the true gradient for a heavy-tailed importance-reweighted estimator. The paper explicitly says the proxy is 'not used as a theorem-level confidence bound,' which is an honest admission but also a concession that the experiments are running an algorithm with no guarantee. Every large-scale headline — the 100-qubit VQE and 40-qubit QNN — comes from this adaptive version. This is a real gap, not a nitpick.\n\n2. The main-text Theorem 2 reports complexity O((1+2a)^L κ(θ)^L/(a ε²)), while App. E derives (1+2a)^L κ(θ)/(a ε²) (with log factors). The appendix form is what the proof supports; the main text needs to match. Minor but confusing.\n\n3. Reproducibility. The paper says code is on GitHub but gives no URL, and the large-qubit reference energies are approximate (thermodynamic-limit expression, CT-PBS evaluation with τ=2^-12). The QNN labels are filtered to |y|>0.1. Those choices are defensible, but they weaken the 'faithful at 100 qubits' rhetoric.\n\nWho should read it: anyone working on Pauli-based simulation, classical surrogates for variational algorithms, or warm-starting VQEs. It deserves a serious referee. I would suggest the authors either prove the stopping-rule estimator stays unbiased or restore fixed-budget sampling in the experiments; the rest is solid enough to build on. Yes, send it to peer review.","headline":"Strong, novel PBS-gradient framework with a clean value-gradient separation theorem, but the deployed adaptive stopping rule escapes the unbiasedness proof, so the guarantees apply to a different algorithm than the one that produced the experiments.","tokens_in":40791,"tokens_out":2963,"would_cite":true,"duration_ms":28398,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68Q12"],"pacs":["03.67.Lx","03.67.Ac"],"model":"deepseek-v4-flash","headline":"By sampling instead of truncating Pauli propagation paths, SPPS obtains unbiased gradients and provably convergent optimization trajectories.","keywords":["stochastic Pauli-path simulation","unbiased gradient estimation","variational quantum eigensolver","quantum neural networks","importance reweighting","path automatic differentiation","Pauli propagation","classical simulation of quantum optimization"],"falsifier":"Fix a circuit and parameter point with known exact gradient, run the adaptive-stopping SPPS many times from independent seeds, and compare the distribution of the stopped averaged estimator to the exact gradient: a systematic offset larger than the Monte Carlo standard error, or a correlation between stopping time and estimator magnitude, would show the practical rule breaks the unbiasedness the theorems assume.","tokens_in":39784,"feed_emoji":"⚛️","tokens_out":6805,"duration_ms":57444,"temperature":0.7,"pith_summary":"This paper argues that the standard way to make Pauli-based quantum simulators practical—truncating the explosion of propagation paths—is fundamentally incompatible with gradient-based optimization: a path set can approximate the objective value to any accuracy while its gradient is still order-one wrong (Theorem 1), and the error accumulates across descent steps into a macroscopic gap (Corollary 1). To fix this, the paper introduces the stochastic Pauli-path simulator (SPPS), which samples paths from the full path space, reweights each contribution by its sampling probability, and differentiates each path's trigonometric weight to form stochastic gradients. The central theoretical result is that these gradients are unbiased, with sample complexity governed by an effective branching factor $\\kappa(\\theta)=\\prod_{j=1}^L(1+|\\sin 2\\theta_j|)$, and that SPPS-driven gradient descent reaches an $\\epsilon$-stationary point in $T=O(L^3\\|O\\|_2^4/\\epsilon^2)$ iterations. Numerically, SPPS tracks exact gradient descent on variational quantum eigensolver pre-training up to 100 qubits and quantum neural network pre-training up to 40 qubits, while running in minutes and beating truncation-based alternatives in both accuracy and runtime. A sympathetic reader should take the paper as establishing that stochastic path sampling, not deterministic truncation, is the right way to bring Pauli-based simulation into optimization tasks.","feed_headline":"Unbiased gradients make Pauli-path simulation faithful to 100 qubits","feed_subtitle":"It reproduces exact gradient descent, beating truncation-based simulators in accuracy and speed.","key_machinery":"The load-bearing object is the Pauli propagation path $\\omega\\in\\{0,\\pm1\\}^L$: each entry records whether a rotation gate branches into its cosine or sine component, and the observable expectation expands as a sum over paths of $\\Psi_\\omega(\\theta)\\mathrm{Tr}[P_\\omega(O)\\rho]$. SPPS's sampling distribution assigns the two branches probabilities $q_j=(|\\cos\\theta_j|+a)/(|\\cos\\theta_j|+|\\sin\\theta_j|+2a)$ for a smoothing parameter $a>0$, so derivative-sensitive branches with tiny coefficients are still visited. Each sampled path is reweighted by $1/\\Pr(\\omega)$, and path automatic differentiation multiplies the reweighted contribution by score factors $-\\tan\\theta_j$ for $\\omega_j=1$ and $\\cot\\theta_j$ for $\\omega_j=-1$, yielding all gradient entries from the same samples. The quantity that controls cost is the effective branching factor $\\kappa(\\theta)=\\prod_{j=1}^L(1+|\\sin 2\\theta_j|)$, which appears in the variance, sample complexity, and convergence bounds.","core_discovery":"The paper's central claim is that the gradient bias of truncation-based Pauli simulators is not incidental but structural. Theorem 1 exhibits, for any target accuracy, a truncated path set that reproduces the function value to within $\\epsilon$ while missing a path whose derivative is order one, so the gradient error stays above $1-\\epsilon$; Corollary 1 shows this bias pushes gradient descent to a wrong limit point separated by a macroscopic gap. Against that, SPPS samples legal Pauli propagation paths with a smoothed distribution, divides each sampled contribution by its probability, and computes all gradient components at once through path automatic differentiation, giving an unbiased stochastic gradient. Theorem 2 bounds the sample budget to $\\tilde{O}((1+2a)^L L \\kappa(\\theta)/(a\\epsilon^2))$ for an $\\epsilon$-accurate gradient with high probability, and Corollary 2 converts unbiasedness plus a variance bound into a polynomial convergence guarantee for SPPS-driven descent. The experiments are the concrete payoff: faithful tracking of exact optimization trajectories on VQE and QNN benchmarks, with 100-qubit VQE pre-training completed in about a minute.","pith_inferences":["Editorial inference: the same unbiased-sampling strategy should extend to other propagation bases, such as Majorana fermions or Clifford perturbation theory, because the argument only needs a finite path expansion and a differentiable trigonometric weight per path.","Editorial inference: SPPS's unbiased gradient noise gives a direct classical probe of landscape flatness; the variance bound proportional to $\\kappa(\\theta)$ could be used to detect barren-plateau regimes before expensive quantum training.","Editorial inference: a clean testable consequence is that fixed-budget SPPS and adaptive-stopping SPPS should agree statistically; if they do not, the practical implementation carries a selection bias that the theorems do not cover.","Editorial inference: the value-gradient separation theorem suggests a new reporting standard for classical simulators: validate gradients, not just energies, on optimization benchmarks."],"forward_implications":["Truncation-based Pauli simulators should not be trusted for gradient-driven optimization even when their energy estimates look accurate; a small value error can hide an order-one gradient error.","When the effective branching factor $\\kappa(\\theta)$ stays moderate along the optimization trajectory, SPPS provides polynomial sample budgets, so optimization dynamics join the class of classically simulable quantum processes.","SPPS-driven gradient descent converges to an approximate stationary point in $O(L^3\\|O\\|_2^4/\\epsilon^2)$ iterations, giving the first convergence guarantee for a non-exact classical simulator of quantum optimization.","The same sampled paths yield both expectation values and all gradient components, so the cost of a gradient step is comparable to one forward simulation rather than a parameter-shift factor of $L$.","In practice the method pre-trains VQE on 100-qubit transverse-field Ising models in about one minute and a 40-qubit quantum neural network in under ten minutes while following exact optimization dynamics."],"supporting_citations":[{"why":"Documents empirically biased gradients of low-weight Pauli propagation, the motivation the paper sets out to fix.","marker":"[42]"},{"why":"Exemplifies coefficient-truncation PBS, a target of Theorem 1's bias construction.","marker":"[22]"},{"why":"Exemplifies frequency-truncation PBS, another truncation rule shown to be biased for gradients.","marker":"[24]"},{"why":"Exemplifies weight-truncation PBS for noisy variational circuits, shown here to be biased for optimization.","marker":"[27]"},{"why":"Shows how Pauli-path simulation is used for utility-scale state preparation, the application class SPPS extends.","marker":"[29]"},{"why":"Supplies the Pauli-propagation implementation used to realize the Tb-PBS baselines.","marker":"[31]"},{"why":"Defines the VQE task used as the main benchmark.","marker":"[50]"},{"why":"Provides the exact automatic-differentiation baseline that SPPS is compared against on small systems.","marker":"[72]"},{"why":"Gives the thermodynamic-limit ground-state energy used to evaluate large-system VQE results.","marker":"[91]"}],"fun_headline_variants":["Pauli-path sampling unlocks unbiased quantum optimization","100-qubit quantum optimization: unbiased gradients in minutes","SPPS: unbiased gradient descent for large-scale quantum tasks","Provable convergence via unbiased Pauli-path gradients","Unbiased gradients: Pauli-path simulator scales to 100 qubits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the implemented adaptive sample-budget rule, which stops when two independent macro-replicates agree, is free of systematic error; the paper explicitly calls that stopping proxy practical rather than a theorem-level confidence bound, so if stopping bias sneaks into the average, the unbiased-gradient and convergence guarantees no longer cover the deployed code.","fun_headline_variants_meta":{"raw":{"variants":["Pauli-path sampling unlocks unbiased quantum optimization","100-qubit quantum optimization: unbiased gradients in minutes","SPPS: unbiased gradient descent for large-scale quantum tasks","Provable convergence via unbiased Pauli-path gradients","Unbiased gradients: Pauli-path simulator scales to 100 qubits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000836,"raw_usage":{"total_tokens":3641,"prompt_tokens":931,"completion_tokens":2710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":2632}},"tokens_in":547,"tokens_out":2710,"duration_ms":17654,"temperature":1.0,"reasoning_tokens":2632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:34:52.274529+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a circuit and parameter point with known exact gradient, run the adaptive-stopping SPPS many times from independent seeds, and compare the distribution of the stopped averaged estimator to the exact gradient: a systematic offset larger than the Monte Carlo standard error, or a correlation between stopping time and estimator magnitude, would show the practical rule breaks the unbiasedness the theorems assume.","supporting_citations":[{"cited_title":"The one-dimensional ising model with a transverse field.Annals of Physics, 57(1):79–90, 1970","cited_arxiv_id":null,"evidence_quote":"Gives the thermodynamic-limit ground-state energy used to evaluate large-system VQE results."}],"review_version":2}