{"id":"94275466-9b96-465b-a76b-ffdadf9108af","arxiv_id":"2608.02075","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Reweighting quasi-probability decomposition samples from a reference circuit reduces sampling overhead for estimating fidelities and the quantum geometric tensor across nearby parameter settings.","lead":"This paper introduces a way to reuse quantum circuit samples from one parameter setting to estimate properties at nearby settings, using quasi-probability decompositions. It applies the method to fidelity and quantum geometric tensor estimation, showing gains over standard methods under simulated hardware noise.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finite-sample bias of the self-normalized estimator is not covered by the chi-based variance bound; the QGT error analysis assumes the unbiased estimator.","rationale":"The reader's identified weakest assumption—support inclusion ξ' ⊆ ξ and χ ≤ 1 for the SPSA reference—is actually provably satisfied for the specific construction in Sec. III-A. The reference has parameter difference 2h in every component, while every SPSA target has differences in {−2h, 0, 2h}. For the CRZ QPD in Eq. (8), the support for a nonzero angle is identical for ±θ, and for θ = 0 only the identity channel remains; hence every target support is a subset of the reference support. Furthermore, χ factorizes over gates: for gates with target difference ±2h, χ_j = 1; for gates with difference 0, χ_j = 1/(γ(2h) cos⁴(h/2)) < 1 for small h (e.g., h = 0.1). So the support/χ issue is not load-bearing. A more serious gap is the mismatch between the estimator analyzed in the theory (un-normalized, Eq. (19)) and the estimator used in the numerics (self-normalized, Eq. (30)). The paper provides no finite-sample bias bound for the latter, and Appendix D's error analysis implicitly assumes the unbiased fidelity estimator. This is a real gap that could affect the claimed variance control and the QGT error scaling, though the numerical results may still be valid. The paper remains conditionally acceptable, but the missing analysis of the self-normalized estimator should be addressed.","tokens_in":14807,"tokens_out":23222,"duration_ms":179472,"concrete_test":"Run a noiseless simulation reproducing the QGT experiment of Fig. 6(a) for the 3-qubit ansatz with M = 5×10^5. For each of the 4K target fidelities, compute the exact fidelity and the self-normalized estimate (30), then estimate the bias B = E[F̂ − F_true] and the standard deviation σ over many independent trials. Check whether |B|/σ ≤ 0.1 and whether σ is consistent with the χγ²/M bound from Eq. (23). If the bias is non-negligible, compare the QGT error using estimator (30) versus the bias-corrected estimator (40); this will determine whether the practical estimator needs correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central variance guarantee (Eqs. (22)–(23)) is derived for the un-normalized reweighted estimator in Eq. (19), but the implemented QGT protocol uses the self-normalized importance-sampling estimator (24) and (30), where the denominator Σ p̂_i w_i is itself random. For self-normalized estimators, the leading variance is similar, but there is an additional finite-sample bias of order O(1/M) that is not bounded anywhere in the manuscript. Appendix D's error analysis uses the fidelity RMSE ε_F from Appendix B, which is derived for the bias-corrected unbiased estimator (40); the actual implementation does not apply this correction and uses (30) directly. If M is not large enough, or if γ becomes large (exponentially in the number of cut gates), the bias can be non-negligible and is not controlled by χ. Thus the theoretical statements about variance control do not directly apply to the estimator used in the numerical QGT results, leaving the actual error scaling unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a reweighting strategy for quasi-probability decompositions (QPDs) that reuses samples from a single reference parameter setting to estimate expectation values at nearby parameter settings, using self-normalized importance sampling. The authors derive variance formulas for the reweighted estimator, introduce a factor χ that controls the variance, and apply the method to fidelity estimation via a compressed Hadamard test and to SPSA-based estimation of the real part of the quantum geometric tensor (QGT). Numerical simulations on a 3-qubit layered ansatz with noiseless and noisy Qiskit backends are used to support the claim that the QPD-based approach outperforms compute–uncompute and Hadamard-test methods under realistic noise.","tokens_in":15088,"tokens_out":11994,"duration_ms":101492,"significance":"If the theoretical guarantees are made fully rigorous, the idea of reusing QPD samples across a family of parameterized circuits is a genuinely useful contribution to circuit knitting and to variational quantum algorithms that require many nearby fidelity evaluations. The paper contains a concrete QPD decomposition with angle-independent channels, explicit variance formulas (Eqs. (19)–(23)), and numerical comparisons under a fixed circuit-execution budget, which are strengths. However, two load-bearing points need attention: the support-inclusion/χ≤1 condition is only numerically illustrated for two CRZ gates, and the self-normalized estimator actually used in the QGT protocol is not covered by the χ-based variance bound used in the error analysis.","major_comments":[{"comment":"The SPSA application requires that for every target fidelity in Eq. (29) the reference QPD satisfies ξ'⊆ξ and χ≤1. The choice of θ+2h1 is justified only by the numerical two-CRZ-gate example in Fig. 5; no proof is given that the CRZ QPD of Eq. (8) has this property for arbitrary θ, h, and number of gates. Since the total χ factorizes over gates, a per-gate proof would suffice, but it is missing. If a target parameter falls outside the reference support, Eq. (21) shows the estimator is biased and Eq. (22) no longer bounds the error. This condition is load-bearing for the central claim that a single reference QPD can estimate all 4K SPSA fidelity terms.","section":"Section III-A1 and Fig. 5"},{"comment":"The χ-based variance guarantee (Eqs. (22)–(23)) is derived for the un-normalized reweighted estimator, while the QGT protocol uses the self-normalized estimator (24) and (30). For self-normalized estimators the finite-sample bias is O(1/M) and the variance differs from Eq. (22); the paper only states asymptotic unbiasedness. Appendix D's ε_F comes from Appendix B, which analyzes the un-normalized estimator (33) and the bias-corrected fidelity estimator (40), not the estimator used in Fig. 6. The theoretical error scaling of the QGT results is therefore not directly supported. The authors should either provide a finite-sample bias/variance analysis for Eq. (30) or use the un-normalized estimator with the bias correction (40), in which case the χ bound would apply.","section":"Section III-1 and Appendix D"},{"comment":"The statement 'the normalization condition Σ_i p̂_i w_i = 1 holds only in expectation' is not correct in general: E[Σ_i p̂_i w_i] = Σ_i p_i w_i = γ'/γ, not unity. This suggests a confusion between the un-normalized estimator of Eq. (19), the sample version of Eq. (22), and the self-normalized estimator of Eq. (24). The definitions and the variance formulas should be attached consistently to the estimator actually used in the numerics.","section":"Section III-1, text near Eq. (24)"}],"minor_comments":[{"comment":"In Fig. 4(b), the horizontal axis is γ, but since h varies, each point corresponds to a different target state; a brief explanation of the relationship between h and γ would help the reader interpret the RMSE-vs-γ scaling.","section":"Fig. 4"},{"comment":"The sentence 'if the qubit remains in |0⟩ or |1⟩ at measurement, then ⟨Z⟩=1 or −1' should be clarified: this statement applies after the final Hadamard gate, with the pre-Hadamard state being |+⟩ or |−⟩. As written, it appears to contradict the preceding sentence.","section":"Appendix C"},{"comment":"The unbiased fidelity estimator F^⋆ is introduced but it is not stated whether Fig. 4 reports F̂ or F^⋆. Since the unbiased correction is not used in the QGT implementation, the text should explicitly say which estimator is used in each numerical result.","section":"Eqs. (39)–(40) and Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the algebraic derivations are mostly consistent, but the manuscript currently has a gap between the estimator analyzed and the estimator implemented in the main application. The support-inclusion/χ≤1 assumption also needs a rigorous statement or at least a much more systematic numerical verification. These issues are fixable within the scope of a revision, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThe reweighting idea is the real contribution: instead of resampling a QPD for every new parameter value, reuse samples from one reference decomposition and reweight. The support condition and chi factor give a clean criterion for when this is efficient. The SPSA application to QGT, reusing a single reference QPD for all 4K fidelity terms, is a nice trick and could make QGT estimation practical in variational loops. I also appreciate that the authors are upfront that CU wins in the noiseless case; the advantage is specifically in noisy settings.\n\nThe theory is mostly solid. Eqs. (19)-(23) are standard importance sampling algebra, and Appendix B is consistent. The angle-independent CRZ decomposition is clever and necessary for the reweighting to work.\n\nThe soft spots are real but fixable. The support-inclusion assumption is only checked numerically for two CRZ gates; for the SPSA reference theta+2h1, the target fidelities can be a distance 4h away in some components, and there's no proof that chi <= 1 or support inclusion holds for the whole circuit. A more systematic analysis of reference selection would strengthen the paper.\n\nThe more technical problem, which the stress-test note correctly identifies, is that the formal error analysis does not cover the estimator actually used in the QGT numerics. Appendix B derives the variance and the bias-corrected fidelity estimator (40) for the un-normalized estimator. The implementation uses the self-normalized estimator (30), whose finite-sample bias is only acknowledged as 'asymptotically unbiased' in the text. No bound on that O(1/M) bias is given, and it could become non-negligible when gamma is large. This means the claimed error scaling for QGT is not strictly supported by the derivations. I don't think this is fatal – the numerics may be fine – but it needs to be closed, either by deriving a bias bound for the self-normalized estimator or by running the QGT experiments with the bias-corrected estimator.\n\nAlso, the numerical validation is limited: 3 qubits, fake backends, no error bars, no code release. That limits confidence but not the conceptual contribution.\n\nWho is this for? People working on circuit knitting, variational quantum algorithms, or QGT estimation. It deserves a serious referee; the gaps are addressable and the core idea is likely to be useful. I'd send it to peer review and ask for a revised version that aligns the theory with the implemented estimator and adds more thorough validation.\n\nBest.","headline":"Reweighting QPDs across parameter shifts is a sound and potentially useful idea, but the paper's error analysis doesn't match the estimator used in the numerical QGT experiments.","tokens_in":15486,"tokens_out":3617,"would_cite":true,"duration_ms":31689,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"This paper introduces a reweighting strategy for quasi-probability decompositions that reuses samples from one reference circuit to estimate expectation values at many nearby parameter settings, cutting circuit evaluations in fidelity and q","keywords":["quasi-probability decomposition","circuit knitting","importance sampling","fidelity estimation","quantum geometric tensor","SPSA","variational quantum algorithms","sample reuse"],"falsifier":"Choose a target parameter $\\theta'$ outside the reference support, where $a_i(\\theta)=0$ but $a_i(\\theta')$ is nonzero for some channel in a single CRZ QPD; the reweighted estimator then has bias equal to $\\sum_{i \\in \\xi'\\setminus \\xi} a_i(\\theta') \\mathrm{tr}(O u_i)$. A direct numerical check of this bias, or a scan of $\\chi$ over the SPSA displacement set for a 3-qubit layered ansatz, would settle whether the claimed variance bound holds.","tokens_in":14768,"feed_emoji":"⚛️","tokens_out":3790,"duration_ms":31344,"temperature":0.7,"texified_at":"2026-08-05T21:59:02.434078+00:00","pith_summary":"The paper's central claim is that when two parameterized quantum circuits share the same decomposition basis and the target's support lies inside the reference's, samples drawn once from a reference quasi-probability decomposition (QPD) can be reweighted via self-normalized importance sampling to estimate expectation values at other parameter settings, without running new circuits. If correct, this tames the exponential sampling overhead of QPDs in regimes where many related expectation values are needed at nearby parameters. The authors demonstrate the method on fidelity estimation between variational states and on the real part of the quantum geometric tensor via SPSA, showing under realistic noise fewer circuit executions than compute-uncompute or Hadamard-test approaches. A sympathetic reader would care because variational algorithms routinely need many such evaluations at nearby points.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":5510,"prompt_tokens":749,"completion_tokens":4761,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":749,"completion_tokens_details":{"reasoning_tokens":4033}},"feed_headline":"Reweighting lets one quantum sample set serve many parameter settings","feed_subtitle":"Fidelity and quantum-geometric-tensor estimation cut circuit runs by reusing a single reference decomposition.","key_machinery":"The central object is the quasi-probability decomposition (QPD), which expresses a quantum channel $U$ as $U = \\sum_i a_i u_i$, with real coefficients summing to 1 and sampling overhead $\\gamma = \\sum |a_i|$. The decomposition is chosen with angle-independent local channels so that the same sampled channels can serve multiple parameter values. The reweighted estimator uses self-normalized importance sampling with weights $w_i = \\left| \\frac{a_i(\\theta')}{a_i(\\theta)} \\right|$, and the variance indicator $\\chi$ (Eq. 23) tells when sample reuse is efficient. The support-inclusion condition $\\xi' \\subset \\xi$ is the gate that makes the reweighted estimator unbiased.","core_discovery":"Quasi-probability decompositions express difficult gates as signed mixtures of local channels, but sampling overhead $\\gamma$ grows exponentially with the number of cut gates. The paper shows that for a family of circuits $U(\\theta)$ sharing the same decomposition basis, a sample from a reference QPD can be reused for a target $U(\\theta')$ as long as the target's support is contained in the reference's. The reweighted estimator is unbiased under that support-inclusion condition, and its variance is controlled by $\\chi(\\theta,\\theta') = \\sum_i \\frac{|a_i(\\theta')|^2}{\\gamma |a_i(\\theta)|}$. For the CRZ decomposition with angle-independent channels, parameters within an $\\ell_\\infty$ neighborhood of the reference sati","pith_inferences":["The chi criterion is more informative than comparing gamma alone: the paper shows parameter points with gamma < gamma_r but chi > 1, so support overlap, not just overhead, determines whether reuse is efficient. A natural next step is choosing the reference point to minimize variance across a whole parameter family.","If reweighting is combined with adaptive reference selection, the method might extend to parameter trajectories in variational time evolution, where consecutive states are close; this is not demonstrated in the paper.","The paper's numerical support for the required support-inclusion condition rests on a two-CRZ example, so a testable extension is to check chi <= 1 over the full SPSA displacement set for larger ansatze with more layers and qubits.","Because the classical weight recomputation costs up to M x n_c per parameter set but is parallelizable, the practical bottleneck may shift from quantum circuits to classical post-processing as M grows."],"forward_implications":["Fidelity between U(theta)|0> and U(theta+h1)|0> can be estimated by sampling only the reference QPD; the root-mean-square error scales as O(gamma/sqrt(M)).","The number of circuit evaluations for SPSA-based QGT estimation drops from 4KN (compute-uncompute) or 8KN (Hadamard test) to at most M, independent of the number of SPSA samples K.","Under realistic hardware noise, the shallow QPD circuits give lower relative QGT error than compute-uncompute at equal circuit-execution budget, with the advantage growing as CNOT error rates increase.","A parameter-independent decomposition of CRZ enables reweighting; the standard optimal decomposition with angle-dependent channels cannot be reweighted in this way.","The same reweighting framework applies to any variational algorithm that needs many expectation values at nearby parameter configurations, such as variational time evolution and kernel methods."],"fun_headline_variants":["Reweighted samples slash QPD overhead across parameter settings","One reference decomposition, many parameter settings: sample reuse via reweighting","Cut exponential sampling cost by reusing QPD samples across parameters","Sample reuse in quasi-probability decompositions tames exponential overhead","Reweighting makes QPD samples portable across variational parameters"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method requires that the target QPD's support lies inside the reference QPD's support (and $\\chi \\le 1$ for variance control); for the SPSA application this is assumed for all target fidelities and is justified only by a numerical example with two CRZ gates, not by a proof.","fun_headline_variants_meta":{"raw":{"variants":["Reweighted samples slash QPD overhead across parameter settings","One reference decomposition, many parameter settings: sample reuse via reweighting","Cut exponential sampling cost by reusing QPD samples across parameters","Sample reuse in quasi-probability decompositions tames exponential overhead","Reweighting makes QPD samples portable across variational parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1557,"prompt_tokens":719,"completion_tokens":838,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":751}},"tokens_in":463,"tokens_out":838,"duration_ms":5982,"temperature":1.0,"reasoning_tokens":751,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:42:05.098592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a target parameter $\\theta'$ outside the reference support, where $a_i(\\theta)=0$ but $a_i(\\theta')$ is nonzero for some channel in a single CRZ QPD; the reweighted estimator then has bias equal to $\\sum_{i \\in \\xi'\\setminus \\xi} a_i(\\theta') \\mathrm{tr}(O u_i)$. A direct numerical check of this bias, or a scan of $\\chi$ over the SPSA displacement set for a 3-qubit layered ansatz, would settle whether the claimed variance bound holds.","supporting_citations":[],"review_version":1}