{"id":"96d7c96d-e241-4563-8ddd-22a8ee623d8e","arxiv_id":"2504.19760","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"SURGE-VQE constructs compact VQE ansatze from seniority-zero paired doubles and generalized singles ranked by one-parameter energy tests, reaching chemical accuracy on BH, LiH, and BeH2 with low CNOT counts.","lead":"The authors introduce SURGE-VQE, a method that builds a compact quantum chemistry circuit from paired electron excitations and single excitations, ranked by cheap one-parameter energy tests. It reports chemically accurate energies for three stretched molecules with far fewer two-qubit gates than a standard adaptive VQE at the tested sizes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed O(N^2) parameter bound is inconsistent with the ansatz construction, undermining the shallow-gate-depth claim.","rationale":"The reader's weakest-assumption analysis focuses on the heuristic reliability of the sequential energy-deviation ordering for operator selection. I agree that the ordering heuristic lacks stress testing, but a more load-bearing problem appears in the complexity analysis: the claimed O(N^2) parameter bound for the final ansatz is inconsistent with Eq. 19, where each of the O(N^2) seniority-zero blocks can contain up to O(N^2) singles, and singles may repeat across blocks. This directly undermines the 'shallow (linear or at worst, sub-quadratic) gate depth' claim that is central to the paper's contribution. The empirical statevector results for BH, LiH, and BeH2 are plausible and correctly reported, but they do not establish the asymptotic scaling. The measurement-overhead claim is also underquantified, but the operator-count inconsistency is more fundamental because it affects the core resource-scaling promise of the method. Since the issue is fixable in principle by either proving a bound on M_α or revising the complexity claim, the verdict remains conditional rather than a rejection; the condition should now include a corrected scaling analysis with numerical evidence on larger systems. I therefore recommend leaving the reader's CONDITIONAL verdict unchanged, while noting that the specific condition should be the resolution of the O(N^2) versus O(N^4) inconsistency, not only the ordering heuristic.","tokens_in":15879,"tokens_out":7978,"duration_ms":81178,"concrete_test":"Run the SURGE-VQE prescreening (with ε = 1e-6) on a sequence of increasing active spaces, e.g., LiH STO-3G (N=12 spin-orbitals), BeH2 STO-3G (N=12), H4 STO-3G (N=8), H6 STO-3G (N=12), N2 STO-3G (N=20), and record the total number of single operators Σ_α M_α and the total CNOT count for the final fermionic ansatz at a strongly correlated geometry. Plot log(Σ M_α) and log(CNOT) versus log(N) and fit the scaling exponent. If the exponent exceeds 2, the Sec. II D O(N^2) parameter bound is refuted and the gate-depth claim fails. As a complementary analytical check, count the worst-case Σ M_α when all singles satisfying the threshold are included; if this scales as O(N^4), the paper's asymptotic claim is contradicted by its own construction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim rests on the assertion in Sec. II D that the final ansatz Eq. 19 contains at most O(N^2) parameters and hence, via the k-UpCCGSD analysis, has linear or sub-quadratic gate depth. This bound is not supported by the construction. Eq. 19 is a product over α = 1..nonv blocks, where nonv is the number of paired double excitations (O(N^2)), and within each block up to M_α generalized singles are included, with M_α ≤ N^2. The paper explicitly notes that a given single may reappear in different blocks. The total number of operators in the ansatz is therefore Σ_α M_α, which in the worst case is O(N^4), not O(N^2). The statement that the overall parameter scaling is 'upper-bounded at O(N^2)' is internally inconsistent with this construction. The linear-depth bound borrowed from k-UpCCGSD assumes a fixed number of operator copies (k), whereas SURGE-VQE permits each single to appear in every block; the non-commuting repetitions increase both parameter count and circuit depth. Without an argument or numerical evidence that the threshold ε = 1e-6 keeps every M_α = O(1), the advertised 'shallow (linear or at worst, sub-quadratic) gate depth' is unestablished. This is a load-bearing flaw because the paper's headline contribution is gate efficiency for scalable systems, not just the three small benchmark molecules.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SURGE-VQE, a variational quantum eigensolver ansatz construction that selects and orders a pool of seniority-zero paired double excitations and generalized single excitations using one-parameter energy-deviation screening against Hartree–Fock and seniority-zero reference states. The final ansatz is a product of blocks, each containing a paired double and a threshold-pruned set of singles ordered by their single-parameter energy deviations. The authors claim that this construction yields chemically accurate ground-state energies for strongly correlated molecules (BH, LiH, and BeH2 in STO-3G) with low CNOT counts, shallow (linear or at worst sub-quadratic) gate depth, and minimal pre-circuit measurement overhead. The manuscript reports statevector simulations with errors relative to FCI, CNOT counts per function evaluation, comparisons with UCCSD, k-UpCCGSD, ADAPT-SD, and ADAPT-gSpD, energy-vs-CNOT convergence curves, a CNOT-efficient qubit-excitation variant, and a noisy simulation for LiH using a FakeMelbourne noise model.","tokens_in":16120,"tokens_out":8199,"duration_ms":87193,"significance":"If the resource claims are correct, SURGE-VQE would be a useful gate-efficient VQE ansatz design: it provides a heuristic, gradient-free alternative to ADAPT-style operator selection, with explicit numerical evidence of chemical accuracy and order-of-magnitude CNOT reductions on three benchmark molecules. The manuscript is honest about the heuristic nature of the selection in Sec. IV and compares against several standard baselines. The numerical benchmarks themselves (errors vs. FCI, CNOT counts, noise robustness) are a strength, as are the explicit acknowledgments of the method's limitations. However, the central scaling claims about parameter count, gate depth, and measurement overhead are not established by the current analysis, and the claimed minimal pre-circuit overhead is not actually measured.","major_comments":[{"comment":"The claimed O(N^2) parameter upper bound is not supported by the ansatz construction. In Eq. (19), each of the nonv blocks can contain up to M_alpha <= N^2 singles, and the text explicitly allows a given single to reappear in different blocks. The total number of variational parameters is therefore nonv + sum_alpha M_alpha, which in the worst case scales as O(N^4), not O(N^2). The statement that \"the overall parameter scaling is governed by the leading contribution from generalized single excitations, ensuring that the scaling remains upper-bounded at O(N^2)\" is internally inconsistent with this construction. Since the linear gate-depth argument is borrowed from k-UpCCGSD, which has a fixed number k of operator layers, it does not transfer to an ansatz with unbounded repetitions across blocks. The authors must either prove (or numerically demonstrate) that the threshold epsilon keeps every M_alpha = O(1), or revise the abstract and conclusion claims of shallow linear/sub-quadratic gate depth to refer only to the specific small systems tested.","section":"Sec. II D, Eq. (19)"},{"comment":"The claim of \"minimal pre-circuit measurement overhead\" is not supported by any measurement count, shot count, or total overhead estimate. The prescreening procedure requires one one-parameter VQE for each of the nonv paired doubles (Step 2) and, for each block, for each candidate single (Step 4). With nonv ~ O(N^2) and ~O(N^2) singles, this is O(N^4) one-parameter optimizations, each requiring repeated energy measurements of the O(N^4) Hamiltonian terms; the worst-case total measurement scaling is therefore O(N^8), which does not obviously constitute \"minimal\" overhead. The manuscript only states the per-prescreening-cycle scaling O(N^4) and does not count the number of cycles. Please report the actual number of one-parameter optimizations performed for the benchmark molecules, the total number of circuit executions or shots used in the prescreening phase, and, ideally, an apples-to-apples comparison with the measurement budget of ADAPT-VQE under matched convergence criteria.","section":"Sec. II C and II D"},{"comment":"The gate-count comparisons to ADAPT-VQE are not made under matched stopping criteria. In Fig. 1, ADAPT-SD and ADAPT-gSpD use eigenvalue and gradient thresholds of 1e-8, while SURGE-VQE uses epsilon = 1e-6 for its preselection. The reported CNOT advantage of SURGE-VQE over ADAPT may therefore be partially attributable to the different thresholds rather than to the operator ordering or seniority-based selection. In Fig. 2, the eigenvalue threshold for ADAPT-gSpD is relaxed to 1e-6, but this is not consistent across all comparisons. To substantiate the \"gate-efficient\" claim, the authors should include ADAPT runs truncated at a comparable accuracy target (e.g., where each method first reaches chemical accuracy) and compare CNOT counts at that point, or at least report the sensitivity of the ADAPT CNOT counts to the chosen thresholds.","section":"Sec. III A, Figs. 1 and 2"},{"comment":"The threshold epsilon is a free parameter of the method, and no sensitivity analysis is provided. All reported results use epsilon = 1e-6 for all three molecules, and the authors explicitly state in Sec. IV that the algorithm is heuristic and that the ordering may not be optimal. Without tests varying epsilon (for example, 1e-4, 1e-6, 1e-8), it is unclear whether the reported accuracy/CNOT trade-off is robust or a fine-tuned artifact of the chosen threshold. This is particularly relevant because the operator ordering—the load-bearing step of the method—depends directly on the relative magnitudes of the one-parameter energy deviations, which are thresholded by epsilon.","section":"Sec. II C.4, Sec. IV"}],"minor_comments":[{"comment":"The sign convention for the anti-Hermitian operator is inconsistent: Eq. (13) writes e^{theta_alpha tau_alpha}, while Eq. (14) writes e^{-theta_alpha tau_alpha} H e^{theta_alpha tau_alpha}. Please state explicitly that tau_alpha is anti-Hermitian and clarify the convention used in the circuit implementation.","section":"Eqs. (13)-(14)"},{"comment":"The symbol N in Eq. (17) (\"I in [1, N^2), N refers to the total number of spin-orbitals\") is not defined until Sec. II D. Please define N and the notation no/nv at first use.","section":"Sec. II C, Eq. (17)"},{"comment":"Reference 39 is cited for the Jordan-Wigner transformation, but the cited paper describes the Bravyi-Kitaev transformation. Either the citation or the mapping name should be corrected.","section":"Sec. III A"},{"comment":"The text says ADAPT-gSpD uses \"an operator pool identical to ours,\" but Step 1 of the SURGE construction applies orbital-symmetry pruning to the singles. Please clarify whether the ADAPT-gSpD calculations use the same pruned pool or the full generalized-singles-plus-paired-doubles pool.","section":"Sec. III A, Sec. II C.1"},{"comment":"The conclusion states that the CNOT-efficient formulation brings the CNOT count \"under 180 CNOTs even for challenging cases such as LiH at highly stretched bond lengths,\" but this specific number is not easily located in Fig. 3. Please report the exact value or add a table with the CNOT counts for all systems and geometries.","section":"Sec. IV"},{"comment":"Please specify what happens when no singles in a block satisfy Delta E_alpha I > epsilon: is the block left with only the paired double, or are the singles included regardless?","section":"Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"The three benchmark systems are all small (4 electrons, 10-12 qubits) and the scaling claims rest on an unproven parameter-count bound. The numerical results are credible and could support a revised paper that either proves the scaling or restricts the claims to the demonstrated small-system regime. If the authors can provide even a numerical scaling study for larger systems (e.g., N = 16-24 qubits) showing that sum_alpha M_alpha remains O(N^2) in practice, the central efficiency claim would be substantially strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: SURGE-VQE is a legitimate addition to the compact-ansatz toolbox, and the three benchmark molecules are honestly evaluated. But the paper's central asymptotic claim—shallow linear/sub-quadratic depth via O(N^2) parameters—does not follow from its own construction. Eq. 19 is a product over nonv blocks, and each block can contain up to M_α singles, with M_α ≤ N^2. Since a single may reappear in different blocks, the total parameter count is Σ_α M_α, worst-case O(N^4), not O(N^2). The text in Sec. II D appears to count distinct single operators rather than their occurrences across blocks. The k-UpCCGSD linear-depth argument assumes a fixed number of operator copies; here the number of copies is not fixed. This is a real flaw in the scaling claim, not a minor typo.\n\nWhat is genuinely good: the paper explicitly credits k-UpCCGSD, DISCO-VQE, and tUPS for the same generalized-singles-plus-paired-doubles operator pool, so the novelty claim is appropriately scoped to the heuristic ordering and prescreening. The statevector results for BH, LiH, and BeH2 show chemical accuracy with CNOT counts far below ADAPT-SD, and the energy-vs-CNOT trajectories in Fig. 2 make the resource advantage concrete. The noisy simulation with FakeMelbourne is a nice addition, and the CNOT-efficient qubit-excitation implementation is a sensible practical step. The authors also state plainly in Sec. IV that the ordering is heuristic and may fail for strongly multireference systems—that honesty deserves credit.\n\nThe soft spots beyond the scaling argument: the pre-circuit measurement overhead claim is never actually measured. They say each prescreening cycle costs O(N^4) and that total overhead is low, but no numerical accounting is given. The baselines are also missing natural competitors: qubit-ADAPT-VQE and DISCO-VQE are cited but not benchmarked, which weakens the comparative case. No code or data are released, which is increasingly expected for this kind of numerics. The stress test is correct: the ε=1e-6 threshold may keep M_α small for these molecules, but there is no evidence it does in general.\n\nBottom line: the numerical results are credible and the heuristic is plausible, but the asymptotic gate-depth claim needs correction. This deserves peer review—a referee can demand the scaling fix and the missing baselines. The paper is worth engaging with, but not in its current form.\n\nRecommendation: send to review, with a request for major revision.","headline":"SURGE-VQE is a credible compact-ansatz heuristic with honest benchmarks, but the asymptotic O(N^2)-parameter and shallow-depth claim is internally inconsistent and needs to be fixed before the paper's main selling point holds.","tokens_in":16654,"tokens_out":2521,"would_cite":false,"duration_ms":23843,"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":"The paper proposes SURGE-VQE, an ansatz built from paired double and generalized single excitations ordered by one-parameter energy-deviation screening, and claims chemically accurate ground states for stretched small molecules at shallow…","keywords":["variational quantum eigensolver","unitary coupled cluster","seniority-zero excitations","operator ordering","one-parameter prescreening","CNOT-efficient circuits","strong correlation","shallow quantum circuits"],"falsifier":"Run the same ansatz-construction protocol on a small molecule in a basis where static correlation is stronger than in STO-3G, such as a stretched symmetric molecule or a twisted double bond, and compare the final energy to full configuration interaction: if a single excitation that the one-parameter screening discarded contributes substantially to the energy once later operators are present, or if the greedy order costs more CNOTs than a random order for the same accuracy, the central premise fails.","tokens_in":15629,"feed_emoji":"⚛️","tokens_out":14679,"duration_ms":134674,"temperature":0.7,"pith_summary":"This paper puts forward an ansatz-construction algorithm for near-term quantum chemistry that tries to obtain both accuracy and shallow circuits from the same ingredient: an ordered set of computationally cheap excitations. Working only with generalized single excitations and seniority-preserving paired double excitations, it ranks each operator by the energy it recovers in a one-parameter quantum circuit, first against the Hartree-Fock state and then against each paired reference state. The paper claims this ordering yields a dynamic ansatz with at most $O(N^2)$ parameters and at worst linear gate depth, and it demonstrates chemically accurate potential curves for BH, LiH, and asymmetric BeH$_2$ while using an order of magnitude fewer two-qubit CNOT gates, the dominant source of circuit error, than gradient-based adaptive ansatze. The practical stake is that, if the ordering premise holds, strongly correlated small molecules can be prepared on noisy hardware with far less two-qubit gate error and with pre-circuit measurement overhead that stays at $O(N^4)$ instead of growing with the number of variational parameters.","feed_headline":"Gate-light VQE keeps chemical accuracy on stretched bonds","feed_subtitle":"SURGE-VQE ranks simple excitations by one-parameter energy gains, cutting two-qubit gates below 180 on LiH.","key_machinery":"The central machinery is a two-tier operator pool built on seniority, the number of unpaired electrons in a determinant: seniority-zero paired double excitations $\\tau_{2,p}^\\alpha$, which move one electron pair from an occupied spatial orbital to a virtual spatial orbital and leave that number unchanged, and generalized single excitations $\\tau_{1,g}^I$, which move one electron between arbitrary spin-orbitals. The argument runs on one-parameter VQE prescreening: Eq. 14 optimizes each paired double alone against Hartree-Fock, Eq. 15 converts its energy gain into a block-ordering score, Eq. 17 scores each candidate single against the paired reference $|\\Psi_\\alpha\\rangle$, and the threshold $\\epsilon$ in Eq. 18 prunes inactive singles. This replaces the per-iteration gradient measurement of adaptive methods with a fixed $O(N^4)$ measurement cost per candidate, and the decomposition identities quoted in Eqs. 10 and 11 are what license the claim that this restricted pool can span arbitrary correlation.","core_discovery":"The paper's central claim is that strong molecular correlation can be captured by a dynamic ansatz built only from generalized rank-one single excitations and seniority-preserving paired double excitations, provided the operators are ordered by their one-parameter energy gains. The ordering protocol first optimizes each paired double alone against the Hartree-Fock state and computes $\\Delta E^\\alpha = E_\\mathrm{HF} - E^\\alpha$ (Eq. 15); doubles are then placed into blocks in descending order of this deviation. Inside each block, candidate singles are optimized against the corresponding seniority-zero reference $|\\Psi_\\alpha\\rangle$, ranked by $\\Delta E^{\\alpha I} = E_\\alpha - E_{\\alpha I}$, and kept only when this gain exceeds a threshold $\\epsilon$. The final ansatz of Eq. 19 therefore has at most $O(N^2)$ parameters, total Jordan-Wigner gate count $O(N^3)$, and worst-case linear circuit depth by the scaling analysis inherited from the generalized UCC construction. On BH, LiH, and asymmetric BeH$_2$ in the minimal STO-3G basis, the paper reports errors relative to full configuration interaction within chemical accuracy (about 1.6 millihartree) across the potential curves, with two-qubit CNOT counts an order of magnitude below the adaptive singles-and-doubles ansatz and, for LiH at one stretch, 1120 CNOTs versus 2024 for the adaptive generalized-singles-paired-doubles ansatz at similar accuracy. With particle-preserving exchange circuits for qubit excitations, CNOT counts drop below 180 for LiH at large bond length while the energy error stays within tens of microhartree.","pith_inferences":["If the one-parameter ordering transfers to larger active spaces, the same prescreening could be used as a cheap first filter before a gradient-based adaptive step, shrinking the operator pool that needs full gradient evaluation.","The reported 'burrowing' energy-versus-CNOT trajectories are heuristic evidence that local one-parameter energy gains agree with the global variational landscape; a direct test would compare the greedy order with the order given by full gradient norms for the same operator set.","Because the CNOT-efficient variant drops fermionic parity terms, its slight accuracy loss could be partially recovered by symmetry post-selection or by restoring the discarded terms for only a few dominant operators; the paper notes the trade-off but does not pursue it.","The screening shortcut of optimizing only one spin sector assumes spin-complementary singles behave identically; in open-shell or spin-symmetry-broken systems that shortcut would need to be revisited, and the measurement overhead would roughly double."],"forward_implications":["On all three tested molecules, the ansatz stays within chemical accuracy of full configuration interaction across the whole potential curve, including stretched geometries where fixed coupled-cluster-style ansatze fail.","The two-qubit CNOT count per energy evaluation is roughly an order of magnitude smaller than for the gradient-based adaptive singles-and-doubles ansatz; at one stretched LiH geometry the count is 1120 versus 2024 for the adaptive generalized-singles-paired-doubles ansatz at comparable accuracy.","Replacing fermionic excitations with qubit excitations implemented by particle-preserving exchange circuits cuts CNOT counts by nearly an order of magnitude for LiH and BeH$_2$, bringing stretched LiH below 180 CNOTs while keeping the error within tens of microhartree.","Because the final circuit is shallow, a noisy-backend simulation of LiH at a bond length of 3 Angstroms shows better inherent accuracy than the fixed and adaptive ansatze compared without error mitigation, implying lower error-mitigation cost when chemical accuracy is sought.","Pre-circuit prescreening costs $O(N^4)$ measurements per operator and can be parallelized, whereas the gradient-based adaptive protocol pays $O(N^4 N_p)$ plus operator-pool gradient measurements across iterations."],"supporting_citations":[{"why":"Supplies the generalized UCC operator pool and the linear gate-depth scaling analysis that the paper adapts to its ordered ansatz.","marker":"[9]"},{"why":"Defines the gradient-based adaptive baseline whose accuracy and measurement overhead the paper compares against.","marker":"[12]"},{"why":"Shows unpaired two-electron excitations decompose into nested commutators of paired doubles and generalized singles, justifying the operator pool.","marker":"[13]"},{"why":"Provides the exponential CNOT scaling of higher-rank fermionic and qubit excitations that motivates restricting the pool to rank-one and paired rank-two operators.","marker":"[29]"},{"why":"Proves arbitrary fermionic operators decompose into nested commutators of rank-one and rank-two excitation operators, grounding the expressiveness of the pool.","marker":"[35]"},{"why":"Introduces particle-preserving exchange circuits whose low two-qubit gate depth underpins the CNOT-efficient implementation.","marker":"[41]"},{"why":"Adapts qubit-excitation circuits to variational eigensolvers, supporting the qubit-excitation variant used in the efficiency comparison.","marker":"[42]"}],"fun_headline_variants":["Seniority-ordered ansatz cuts VQE gates deep","Operator ordering yields shallow VQE circuits","VQE with ordered excitations hits chemical accuracy","Dynamic ansatz slashes CNOTs for strong correlation","Shallow-depth VQE via seniority-driven selection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that ordering operators by the energy each recovers on its own, first against Hartree-Fock and then against paired reference states, produces a near-optimal final ordering; the paper explicitly calls the algorithm heuristic and tests it on only three small molecules, so a failure of that local-to-global extrapolation would erode both the accuracy and the gate savings.","fun_headline_variants_meta":{"raw":{"variants":["Seniority-ordered ansatz cuts VQE gates deep","Operator ordering yields shallow VQE circuits","VQE with ordered excitations hits chemical accuracy","Dynamic ansatz slashes CNOTs for strong correlation","Shallow-depth VQE via seniority-driven selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1711,"prompt_tokens":1134,"completion_tokens":577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":750,"completion_tokens_details":{"reasoning_tokens":503}},"tokens_in":750,"tokens_out":577,"duration_ms":5522,"temperature":1.0,"reasoning_tokens":503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:44:23.434414+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same ansatz-construction protocol on a small molecule in a basis where static correlation is stronger than in STO-3G, such as a stretched symmetric molecule or a twisted double bond, and compare the final energy to full configuration interaction: if a single excitation that the one-parameter screening discarded contributes substantially to the energy once later operators are present, or if the greedy order costs more CNOTs than a random order for the same accuracy, the central premise fails.","supporting_citations":[],"review_version":1}