{"id":"5e66e9d7-3656-4cba-a017-9240423c4ffe","arxiv_id":"1908.06942","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A weight-sorted insertion algorithm and two Clifford circuit constructions reduce the number of measurements and two-qubit gates needed to estimate Pauli expectation values in VQE.","lead":"This paper shows how to group the terms of a quantum Hamiltonian so that fewer experimental repetitions are needed to estimate its energy, and provides a rule for choosing the groups based on term weights. It also constructs two circuits that let a quantum computer measure each group all at once, with lower two-qubit gate counts than previous methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing concern is that R-hat, the metric used to design and benchmark Sorted Insertion, is validated against the true measurement saving R only for nine small molecules and one ansatz family; if this proxy fails on realistic VQE states, the practical performance claims lose force.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: the practical force of the paper depends on R-hat being a reliable proxy for the state-dependent measurement saving R, and the evidence for that proxy is limited to the nine smallest molecules with a single hardware-efficient ansatz at depth 1. The mathematical core—Theorem 1 and the gate-count bounds—is independently supported by the supplied proofs and worked examples, so the concern does not invalidate the main theoretical contributions. It does, however, justify the CONDITIONAL verdict: the practical claims about Sorted Insertion's advantage and the 10-to-60-fold improvement range should not be taken as established beyond the tested regime. A single additional validation on a mid-sized molecule with a different ansatz, using exact R, would settle whether the proxy concern actually lands. The Intro's '10 to 60 fold improvement' is also an overstatement relative to Table 4, but it is secondary to the proxy issue.","tokens_in":25939,"tokens_out":14385,"duration_ms":166672,"concrete_test":"Classically simulate SiH4 (24 qubits, 18713 Pauli terms) with 100 parameter sets of a chemically motivated ansatz such as UCCSD or k-UpCCGSD; compute the exact R metric of Eq. (10) for the Sorted Insertion arrangement and for the Independent Set arrangement from Table 3. If the mean R for Sorted Insertion does not exceed that of Independent Set whenever R-hat does, or if R-hat differs from mean R by more than about 20% per molecule, then the R-hat-based comparison does not support the paper's practical conclusion that Sorted Insertion reduces VQE measurement counts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mathematical results—Theorem 1 and the two rotation-circuit constructions—appear sound. The load-bearing weakness is the empirical bridge from the state-independent metric R-hat (Eq. 20, Appendix B) to the actual measurement saving R (Eq. 10). R-hat is obtained by replacing all variances and covariances in R by their uniform-spherical expectation; because R is nonlinear in those variances, R-hat is not literally E[R], and for the structured states produced by VQE ansatze there is no a priori reason that the two orderings agree. The only validation is Table 2, which compares R-hat to the mean of R over 100 random hardware-efficient depth-1 ansatz states for the nine smallest molecules (up to 16 qubits). For every larger molecule in Table 4, only R-hat is reported, with no R values given. Since the abstract's central practical claim is that Sorted Insertion reduces the number of measurements and outperforms greedy colouring algorithms, that claim rests on an unvalidated proxy precisely in the regime—larger, chemically relevant molecules—where the proxy is needed. The paper itself notes that further investigation of the R-hat/R relationship is needed if a different ansatz is considered (Section 4), leaving the practical claim conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the problem of reducing the number of measurements required to estimate expectation values of a weighted sum of Pauli operators, in the setting where mutually commuting Paulis are measured simultaneously and the available measurements are optimally allocated among collections. It introduces two metrics, R (state-dependent) and Rhat (an analytic, coefficient-only approximation to the uniform-spherical average of R), proves that merging two commuting collections never decreases R or Rhat (Theorem 1), and proposes a coefficient-weighted heuristic called Sorted Insertion for forming collections. It also gives two Clifford rotation-circuit constructions that rotate any collection of k independent commuting n-qubit Paulis to the computational basis with worst-case two-qubit gate counts kn - k(k+1)/2 and O(kn/log k), respectively. The methods are benchmarked on molecular Hamiltonians up to 38 qubits.","tokens_in":26174,"tokens_out":12712,"duration_ms":140360,"significance":"The central mathematical contributions are convincing and useful. Theorem 1 is proved cleanly through a positive-semidefinite covariance argument and Cauchy-Schwarz, and it correctly identifies that earlier counterexamples to collection merging relied on a suboptimal uniform allocation of measurements. The rotation constructions are a genuine advance because they explicitly treat the k < n case, and the stated gate-count bounds are concrete and appear correct. The numerical study is extensive, and the paper is careful to distinguish R from Rhat for the smaller molecules where both are computed. The main reservation is that the practical measurement-reduction claims for the larger molecules are carried by Rhat, a state-independent proxy that is validated against R only for the nine smallest molecules and a single depth-1 hardware-efficient ansatz; this limits the strength of the VQE-specific conclusions as currently stated.","major_comments":[{"comment":"The practical claim that Sorted Insertion yields a 10- to 60-fold reduction in the number of measurements is supported for the 18- to 38-qubit molecules only by Rhat, not by the state-dependent metric R. Because Rhat is obtained by moving the uniform-spherical expectation inside the nonlinear square roots of Eq. (10), it is not E[R], and for structured VQE ansatz states there is no a priori reason that Rhat tracks R. The paper validates Rhat against the average of R over 100 random depth-1 hardware-efficient ansatz states only for the nine molecules up to 16 qubits; Table 4 reports no R values for the remaining molecules. This gap is load-bearing because the abstract's headline improvement and the practical superiority of Sorted Insertion are phrased in terms of the number of measurements required. I ask the authors either to (i) add R validation for representative states on at least some of the larger molecules, or (ii) explicitly rephrase the practical claims as statements about Rhat, with a clear caveat that the true saving on a given VQE state may differ.","section":"Section 4, Table 4, Eq. (20)"},{"comment":"The abstract's statement that Sorted Insertion outperforms four conventional greedy colouring algorithms is based on Table 3, which compares only Rhat and the number of collections N. The toy example in Section 2 already shows that the arrangement with the largest Rhat can have more collections than the minimum-clique-cover arrangement, and the same need not hold for R. Since Rhat is not shown to be a reliable proxy for R on realistic VQE states, the claim that Sorted Insertion is better in practice should be qualified as 'better according to Rhat' unless R-based comparisons are provided.","section":"Abstract and Table 3"}],"minor_comments":[{"comment":"The summation limits in Eq. (43) appear to contain a typo (N_i instead of m_i); please correct them.","section":"Appendix B, Eq. (43)"},{"comment":"The definition of Rhat should state explicitly that identity Pauli terms are excluded or have their variance expectation set to zero, since the derivation in Appendix B applies only to non-identity Paulis.","section":"Section 2, Eq. (20)"},{"comment":"The ansatz used to generate the 100 random states is described only as a hardware-efficient ansatz of depth 1; a circuit diagram or reference would improve reproducibility.","section":"Section 4, Table 2"},{"comment":"The phrase 'ratio of worst-case maximum number of two-qubit gates to actual number' in Fig. 3(c) is ambiguous because both quantities are maxima; I suggest rewording it as 'ratio of the theoretical worst-case gate count to the largest true gate count over all collections'.","section":"Fig. 3(c) and Section 4"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper is within scope and the theoretical contributions are solid. The main uncertainty is the empirical bridge from Rhat to R; if the authors either extend the validation to larger molecules or qualify the practical claims, I would support publication. I have no concerns about citation practices or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is a solid contribution to the VQE measurement-overhead literature. The new pieces are real: the R/Rhat metrics under optimal shot allocation, Theorem 1 (splitting never helps under optimal allocation), Sorted Insertion, and the two rotation-circuit constructions that exploit k < n. The proof of Theorem 1 is a clean Cauchy-Schwarz argument on a PSD covariance matrix; the gate-count bounds are defended with detailed constructions and a Four-Russians-style argument for the CNOT version. I checked the worked CZ example enough to convince myself the machinery hangs together.\n\nThe empirical story is the soft spot. Rhat is a state-independent proxy obtained by replacing variances and covariances with their uniform-spherical expectations. Because R is nonlinear in those variances, Rhat is not E[R]. For the nine smallest molecules, Table 2 shows Rhat close to mean R over 100 random hardware-efficient depth-1 states. That is decent but narrow: one ansatz family, small systems, and for every larger molecule Table 4 reports only Rhat, no R values. So the headline \"10-60x improvement\" and the comparisons against greedy coloring rest on Rhat exactly in the regime—larger, chemically relevant molecules—where it is least validated. The authors acknowledge this at the end of Section 4, which is honest, but it does make the practical claim conditional rather than established.\n\nTwo smaller quibbles. The abstract says \"10 to 60 fold improvement\" but H2 gets 1.76; the claim should be qualified. And no code or data artifacts are released, so the numerical tables cannot be independently reproduced without reimplementation; that is addressable but worth noting.\n\nCitation pattern: nothing load-bearing is self-citational. Related work is cited sensibly, and the NP-hardness reduction to Min-Clique-Cover is a reasonable addition.\n\nWho is this for? Anyone working on measurement reduction for VQE or Pauli grouping. The metrics are worth adopting, and the k < n rotation circuits fill a real gap. I would send this to a serious referee. The Rhat validation issue is exactly what a referee should push on, but it does not undermine the theoretical core.\n\nRecommendation: accept peer review, with a request for either broader Rhat validation or a softer statement of the practical claims.","headline":"Solid, publishable VQE measurement paper with a genuinely useful Theorem 1 and new k