{"id":"b2b839d0-d88f-4279-9dfa-f9158659ab78","arxiv_id":"2607.24725","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Infinite-order response resummation for Pauli controls yields a hyperparameter-light pulse optimizer that converges faster than CRAB on 2- and 3-qubit QFT.","lead":"PEPRino is a gradient-free quantum pulse optimizer that maps the fidelity landscape by resumming response functions to infinite order from only first- and second-order susceptibilities. It reaches high-fidelity two- and three-qubit Quantum Fourier Transform pulses with fewer iterations than CRAB/Nelder-Mead and without a learning-rate hyperparameter.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The reported “full 2-design” is only a product of local one-qubit designs, so the reported gate infidelities are not standard Haar-averaged gate infidelities.","rationale":"The central resummation itself appears secure: for any Pauli-string control B, B²=I implies ad_B³=4ad_B, which is exactly the structure behind Eqs. (13)–(15). The projection of a delta kick onto finitely many sine modes is an approximation, but the paper presents it as such and Fig. 1 checks it qualitatively.\n\nThe deeper weak point is the state average used to define and report gate performance. The reader identified the finite random batch as fragile; this is the same general premise, but the problem persists even without finite-sample noise because the reference 36-state ensemble is not a global two-qubit 2-design. Because PEPRino and CRAB are compared on the same nonstandard objective, the relative iteration comparison may still be meaningful for that objective, and this does not warrant rejection. However, the conventional high-fidelity QFT claim—and the claim that the demonstrated performance transfers to standard gate optimization—should be conditional on recomputing the optimized gates with the exact trace formula or a genuine global 2-design.","tokens_in":17318,"tokens_out":8180,"duration_ms":320265,"concrete_test":"First verify the moment mismatch directly: compute (1/36)Σ_{a,b=1}^6 (P_a⊗P_b)^{⊗2} and compare it with P_sym/10. Then, for representative converged pulses from Figs. 3–5, propagate the full unitary U and compute the exact unitary average-gate infidelity 1−(d+|Tr(V†U)|²)/(d(d+1)). Compare it with the paper’s 36-product-state result and, for an independent check, a true global Clifford/stabilizer 2-design. Material changes in the reported floor or convergence ranking would mean the fidelity claims do not currently land.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Appendix C constructs six one-qubit stabilizer states and then calls their N-fold tensor products an N-qubit 2-design. That is not a projective 2-design for the global d=2^N Hilbert space. For two qubits, the local-product moment is (1/36)Σ(P_a⊗P_b)^{⊗2}=P_sym^(1)⊗P_sym^(2)/9=(I+S_1)(I+S_2)/36, whereas a global two-qubit 2-design must give P_sym/10=(I+S_1S_2)/20. The product ensemble omits the sector odd under each local swap but even under the global swap.\n\nThis is not merely a small-batch sampling issue: Eq. (23), Figs. 3–6, and the CRAB objective use the 36-product-state set as the “full 2-design,” while the three-qubit results use random product states. They therefore measure a local-product-state average, not the conventional Haar/average gate fidelity. Near a target unitary, one- and two-local Pauli errors are weighted differently from the Haar measure, so the quoted 10^-8-scale floors and “high-fidelity QFT” claim may shift. Even taking the stated sampling distribution to infinity would not converge to the global gate-fidelity objective.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript introduces PEPRino, an extension of the authors' earlier PEPR optimal-control method [21], in which the fidelity response to a time-local control kick is resummed to infinite nonlinear order. For Pauli-product control operators, the nested-commutator hierarchy collapses (Eqs. 13–14, App. A), so the full landscape reduces to a closed form, ΔF_PEPRino = (1/2)χ⁽¹⁾ sin(2ε) + (1/4)χ⁽²⁾(1−cos(2ε)) (Eq. 15), whose maximizer ε* is obtained analytically (Eq. 16). The kick is then projected onto a sine-mode pulse basis, giving a multi-parameter update with no learning rate and no gradient computation. The method is benchmarked against CRAB/Nelder-Mead on the 2-qubit QFT (Figs. 3–4) and applied to the 3-qubit QFT (Fig. 5), showing convergence to ~10⁻⁸ infidelity in roughly one-third the iterations of CRAB.","tokens_in":17629,"tokens_out":3681,"duration_ms":132096,"significance":"If the results hold, the paper delivers a genuinely useful contribution: an analytic, closed-form resummation of the response series (not a fit), an update rule whose step size is determined by the landscape rather than a tuned learning rate, and a direct, controlled comparison against a standard baseline (CRAB/Nelder-Mead) with matched per-iteration resources (n_B = 41 vs. 41 simplex vertices). Appendices A–B give a clean, checkable derivation of the Pauli-algebra collapse and the sin/cos resummation, and Fig. 1 demonstrates that the approximate landscape tracks the true ΔF near the chosen ε*. The restriction to Pauli-type controls is honestly stated and is a reasonable domain for many qubit platforms. The main weaknesses are in the evaluation metric (a local-product \"2-design\" that is not a global 2-design) and in overstatement of the \"hyperparameter-free\" and wall-time claims, not in the central derivation.","major_comments":[{"comment":"The N-qubit '2-design' constructed in App. C is a tensor product of one-qubit 2-designs (6^N product states), which is not a projective 2-design on the global d=2^N Hilbert space. For two qubits, the 36-state product ensemble has second moment (I+S_1)(I+S_2)/36, whereas a global two-qubit 2-design gives (I+S_global)/20; the sector odd under each local swap but even under the global swap is missed. Consequently Eq. (23), Figs. 3–6, and the CRAB objective all measure a local-product-state average, not the conventional Haar/average gate fidelity, and one- and two-local Pauli error components are weighted differently than under the Haar measure. The quoted ~10⁻⁸ floors and the 'high-fidelity QFT' claim therefore refer to a non-standard metric, and the 3-qubit evaluation (n_E=100 random product states) inherits the same issue. Since both PEPRino and CRAB were scored on the same metric, the re","section":"Appendix C; Eq. (23); Figs. 3–6"},{"comment":"The abstract and §I advertise the method as 'hyperparameter-free', but Figs. 3 and 5 demonstrate explicit dependence on the number of modes n_m and the batch size n_B, and the initialization scale θ ~ N(0, 1/(n_m√k)) and transformation time t_f are further choices the user must make. What the method actually eliminates is the learning rate α₀ of PEPR (Fig. 1) and the gradient computation — which is a real and worthwhile advance. The claim should be restated accordingly (e.g., 'learning-rate- and gradient-free'), and the remaining sensitivities to n_m and n_B acknowledged in the abstract/conclusion rather than only in §IV.","section":"Abstract; §I; Figs. 3, 5"},{"comment":"The abstract and §IV.A claim faster convergence 'regarding iteration steps and computational time', and §IV.A states PEPRino is 'demanding fewer computational resources', but no wall-clock or per-iteration cost data are reported anywhere. The iteration-count advantage is clear from Fig. 4, but per-iteration cost differs between the methods (PEPRino: n_B × two susceptibility evaluations; CRAB: ~41+ simplex evaluations over the 36-state ensemble), so the wall-time claim is plausible yet currently unsubstantiated. Please add a quantitative comparison (wall time per run, or an explicit count of time evolutions per iteration for both methods).","section":"§IV.A, Fig. 4; Abstract"},{"comment":"The resummation in Eq. (15) and the convergence of the underlying series are exact only because ad_B³ = 4 ad_B for Pauli strings (App. A, Eqs. 13–14). App. A verifies this explicitly only for B = σ_z⊗σ_z and asserts the pattern 'extends to N qubits'. Since the whole method stands on this identity, please give the one-line general proof (any Pauli string B has B²=I, so ad_B² acts as 4·id on the anti-commuting Pauli components and 0 on the commuting ones, hence ad_B³ = 4 ad_B), and state at Eq. (15) precisely which class of control operators is admissible — e.g., whether sums of non-commuting Pauli terms as a single control operator are excluded.","section":"§II, Eqs. (13)–(15); Appendix A"}],"minor_comments":[{"comment":"Sign conventions are inconsistent across Eqs. (10), (11), (12) and (B18)–(B21): Eq. (10) has (−ε)ⁿ while Eq. (11) has +εⁿ, and Eq. (B19) has (iε/ℏ)ⁿ. The signs presumably get absorbed into the definition of χ⁽ⁿ⁾, but this should be made uniform or explicitly noted.","section":"Eqs. (10)–(12), (B18)–(B21)"},{"comment":"In Fig. 1 the true fidelity change is labeled ΔF0 (and ΔF₀ in the text), which reads as 'zeroth-order' in a paper about response orders; a different symbol (e.g., ΔF_true or ΔF_exact) would avoid confusion with the perturbative orders.","section":"Fig. 1"},{"comment":"Fig. 3 caption describes 'thin dotted lines' as individual runs while the main text describes 'thin solid lines with distinct line styles' as the per-batch-size logarithmic averages; please make caption and text consistent. Similarly check Fig. 6, whose legend lists n_B = 5–8 while the body text (§IV.A) discusses n_B ∈ {4,5,6,7} for the main 2-qubit runs.","section":"Figs. 3, 6"},{"comment":"ℏ appears explicitly in Eqs. (8), (12), (B11) etc., but the numerical simulations evidently use ℏ = 1; please state the units convention once.","section":"Eq. (8) onward"},{"comment":"Ref. [32] (Dankert et al.) concerns exact/approximate unitary 2-designs; when revising App. C per the major comment, it would help to cite standard references on average gate fidelity and state 2-designs (e.g., Nielsen; Horodecki et al.; Emerson et al. on gate-fidelity estimation) and to clarify which quantity the chosen ensemble actually estimates.","section":"Appendix C, references"},{"comment":"The choice 'smallest |ε| among maxima' (Eq. 16) is reasonable but unmotivated in the text; a sentence noting that Fig. 1 shows the approximation degrades for |ε| ≳ π/2, hence the smallest-|ε| rule, would connect the criterion to the stated accuracy window.","section":"§II, Eq. (16)"}],"recommendation":"major_revision","confidential_remarks":"The central derivation is sound and the comparison protocol is more careful than is common in this subfield (matched per-iteration resources, 40-run log averages, direct CRAB implementation rather than literature numbers). The one substantive error — passing off a product of local 2-designs as a global N-qubit 2-design — is a real technical mistake but is repairable within the manuscript's scope (d=4 and d=8 Haar evaluation is cheap) and does not touch the algorithm itself. If the authors redo the fidelity evaluation properly and temper the 'hyperparameter-free' and wall-time claims, I would expect to support acceptance. The work builds closely on the group's own PEPR paper [21]; the incremental step (infinite-order resummation) is independently derived here and, in my view, clears the novelty bar."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that they close the response series for Pauli-type controls. Nested commutators stay in a two-dimensional span, so ΔF collapses to a simple sin/cos of χ⁽¹⁾ and χ⁽²⁾, and the maximizing ε* becomes a learning-rate-free update. That step is cleanly derived (Apps. A–B) and is a genuine extension of their earlier linear PEPR work. The projection onto sine modes is the same template as before; the infinite-order piece is new.\n\nWhat they do well: multi-run log-averaged curves, matched batch/simplex size against CRAB/Nelder-Mead on two-qubit QFT, and a working three-qubit demo. Iteration and wall-time advantages look real on the reported metric. Circularity is low—the update is not fitted to the QFT target.\n\nSoft spots, in proportion. First, Appendix C’s “full 2-design” is the tensor product of six local stabilizer states. That is not a projective 2-design on the global Hilbert space; the second-moment operators do not match. So Figs. 3–6 and the CRAB objective are product-state averages, not standard Haar/average gate fidelity. Near a target unitary the error weighting differs, and the 10⁻⁸ floors should be read as “high fidelity on product inputs,” not the usual gate number. This is a real characterization error, not a fatal crack in the optimizer itself—the update rule still makes sense, and product batches often drive usable gates in practice. Second, “hyperparameter-free” is oversold: n_m, n_B, and t_f remain free; only the learning rate is gone. Third, the baseline suite is narrow (no GRAPE-class gradient method), and there is no code or integrator detail. The method is also Pauli-specific by construction.\n\nWho it is for: people already doing analog pulse engineering or CRAB-style work who want a rate-free alternative on few-qubit Pauli controls. Not a field reorganizer. It deserves a serious referee—math is honest, numerics are multi-run and comparable, limitations are mostly scope and labeling. I would engage, push them to fix the 2-design language and add a gradient baseline, and not desk-reject.","headline":"Solid methods extension of PEPR: the Pauli resummation is real and useful; the “2-design” evaluation is mislabeled product-state averaging, which softens the fidelity claims but does not kill the algorithm.","tokens_in":18822,"tokens_out":592,"would_cite":true,"duration_ms":21313,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"PEPRino finds high-fidelity multi-qubit control pulses without gradients or learning rates by resumming the fidelity response to infinite order from only the first two susceptibilities.","keywords":["quantum optimal control","pulse engineering","response theory","Quantum Fourier Transform","CRAB","multi-qubit gates","hyperparameter-free optimization","PEPRino"],"falsifier":"Replace the Pauli control operators with generic non-Pauli Hermitian generators and measure whether the two-term formula still matches the true fidelity change under a finite kick; if the predicted optimal kick no longer improves fidelity, the resummation claim fails.","tokens_in":18576,"feed_emoji":"⚛️","tokens_out":984,"duration_ms":36860,"temperature":0.7,"pith_summary":"Optimal control of quantum gates usually demands careful tuning of learning rates and other hyperparameters, and the search grows harder as the number of qubits and parameters increases. This paper introduces PEPRino, a method that constructs the local fidelity landscape under a time-local control kick by summing the infinite series of nonlinear responses into a closed formula that needs only the first- and second-order susceptibilities. The collapse of the series follows from the algebra of Pauli operators, so the algorithm can choose the best kick size with no free step-size parameter and then project that kick onto a sine-mode pulse basis. On two- and three-qubit Quantum Fourier Transforms the method reaches high fidelity; on two qubits it does so in fewer iterations and less wall-clock time than the standard CRAB/Nelder-Mead optimizer. The result is a practical, hyperparameter-free route to high-fidelity quantum operations that already scales to three qubits.","feed_headline":"Quantum pulses optimized without gradients or learning rates","feed_subtitle":"Resumming fidelity response to infinite order finds high-fidelity QFTs faster than CRAB on two and three qubits","key_machinery":"The infinite-order fidelity landscape ΔF_PEPRino = (1/2) χ^{(1)} sin(2ε) + (1/4) χ^{(2)} (1 − cos(2ε)), obtained by using the nested-commutator identities of Pauli operators that reduce every higher-order susceptibility to a multiple of χ^{(1)} or χ^{(2)}; the maximizing ε* is then projected onto the sine-mode control functions to give the hyperparameter-free update.","core_discovery":"For multi-qubit systems whose control operators are Pauli matrices or tensor products of Paulis, the change in gate fidelity under a time-local perturbation can be resummed to all nonlinear orders as ΔF = (1/2) χ^{(1)} sin(2ε) + (1/4) χ^{(2)} (1 − cos(2ε)). Maximizing this expression for the kick strength ε and projecting the optimal kick onto a finite sine basis yields parameter updates that require neither gradients nor learning rates. The resulting optimizer produces high-fidelity implementations of the Quantum Fourier Transform on two and three qubits and converges faster than CRAB with Nelder-Mead on the two-qubit case.","pith_inferences":["The same low-dimensional commutator collapse may extend to any control Lie algebra whose adjoint representation stays two-dimensional, suggesting a route beyond pure qubit Paulis.","Hybrid schemes that seed a gradient-based optimizer with PEPRino’s infinite-order step could escape flat regions of the control landscape more reliably.","The wall-time gap versus CRAB is expected to widen with qubit number, because CRAB’s simplex size grows with every added mode while PEPRino’s per-step cost is set mainly by batch size."],"forward_implications":["High-fidelity QFT control pulses for two and three qubits can be obtained without any learning-rate schedule or gradient evaluation.","Averaging susceptibilities over modest batches of initial states is already sufficient to drive global gate optimization.","The same response-projection update applies equally to state preparation and to full unitary gate synthesis.","Because each step needs only two response functions, the method remains computationally lighter than simplex methods whose cost scales with the full parameter dimension."],"fun_headline_variants":["PEPRino resumes fidelity response to infinite order for QFT pulses","All-order response projection yields gradient-free multi-qubit controls","Infinite-order fidelity resummation beats CRAB on two-qubit QFT","Pauli-control kicks optimized by closed-form ΔF without hyperparameters","Resummed χ(1),χ(2) navigates pulse landscape for high-fidelity QFT"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The closed-form landscape holds only when every control operator is built from Pauli matrices, so that nested commutators keep alternating between just two operators; if the controls are more general, the infinite series cannot be reduced to the first two terms.","fun_headline_variants_meta":{"raw":{"variants":["PEPRino resumes fidelity response to infinite order for QFT pulses","All-order response projection yields gradient-free multi-qubit controls","Infinite-order fidelity resummation beats CRAB on two-qubit QFT","Pauli-control kicks optimized by closed-form ΔF without hyperparameters","Resummed χ(1),χ(2) navigates pulse landscape for high-fidelity QFT"]},"model":"grok-4.5","effort":"low","cost_usd":0.004068,"raw_usage":{"total_tokens":1275,"prompt_tokens":841,"num_sources_used":0,"completion_tokens":87,"cost_in_usd_ticks":40684000,"prompt_tokens_details":{"text_tokens":841,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":347,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":841,"tokens_out":87,"duration_ms":7365,"temperature":1.0,"reasoning_tokens":347,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T06:43:26.748493+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replace the Pauli control operators with generic non-Pauli Hermitian generators and measure whether the two-term formula still matches the true fidelity change under a finite kick; if the predicted optimal kick no longer improves fidelity, the resummation claim fails.","supporting_citations":[],"review_version":1}