{"id":"b99ceead-25b6-4168-a7fc-28abf7cd253b","arxiv_id":"2608.03656","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Lie-group version of iLQR synthesizes multi-qubit gates with machine-precision infidelity in one iteration when all Hamiltonian terms are available, and converges faster early under 2-local constraints.","lead":"Researchers adapted a control-theory algorithm, iLQR, to the curved space SU(N) and tested it on quantum gate design. In idealized settings it hit machine-precision fidelities in one iteration; under realistic drive restrictions it converged faster early but more often got stuck in local minima.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Two-local convergence comparison rests on the untested first-order BCH truncation of Eq. (10)/(A14), not just the dexp order tested in Sec. IV B4.","rationale":"The reader's weakest assumption is indeed the first-order BCH linearization, and I agree that this is the most load-bearing mathematical approximation in the paper. However, I would phrase the concern more narrowly: the numerical convergence curves are likely computed by propagating the exact Schrödinger dynamics in the forward pass, so the BCH truncation affects the quality of the descent direction and feedback gains rather than the fidelity evaluation itself. That still matters for the central claim because a poor surrogate dynamics can create artificial early progress or artificial local minima. The paper explicitly tests dexp truncation order but never tests the first-order BCH truncation, despite Sec. V showing that the five-qubit failures are not cured by more time steps or higher dexp order. The full-algebra one-iteration result is not a validating test because it is essentially an exact linearization at the initial identity trajectory with complete control. I therefore keep the reader's CONDITIONAL verdict: the concern is plausible and testable, but it is not established as a fatal error. If the proposed finite-difference Jacobian rerun reproduces the reported curves, the concern is resolved and an ACCEPT would be justified; if not, the numerical comparison would need major reinterpretation. Secondary issues (no released code, hyperparameter selection on the benchmark gates, and the triviality of the full-basis single-iteration result) reinforce the conditional status but are not the main load-bearing concern.","tokens_in":25141,"tokens_out":15624,"duration_ms":158338,"concrete_test":"Re-run the Sec. IV B2/B3 optimizations (at least CCZ, Toffoli, WX4, WZ4) with the backward-pass Jacobians f_eta,k and f_u,k replaced by finite-difference derivatives of the exact Lie-algebra equation of motion Eq. (9) about each nominal trajectory, keeping hyperparameters, seeds, and line search fixed. If the median iteration-by-iteration infidelity curves and final distributions match Figs. 2, 3, 5, and 6, the BCH truncation is validated; if they diverge materially, the reported early-iteration advantage is an artifact of the approximate dynamics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical comparison in the 2-local setting (Figs. 2, 3, 5, 6) uses iLQR updates whose descent directions and feedback gains are computed from the first-order BCH linearization Eq. (10) / Eq. (A14), where the error O(||(eta_k, sigma delta u_k)||^2) is discarded. This is a different and more fundamental approximation than the dexp series truncation that Sec. IV B4 tests. If the neglected BCH terms are comparable to the retained terms at the update sizes produced in early iterations, then the reported 'faster early convergence' of the Lie-group variant could reflect an optimistic surrogate dynamics rather than a genuine geometric advantage. The paper's own Discussion (Sec. V) states that neither increasing the number of time steps nor increasing the dexp order improves the five-qubit results, which is consistent with first-order BCH error being a limiting factor. The full-algebra single-iteration result does not test this: with zero initial controls and the full basis, the linearization is exact at the initial point, so it cannot validate the truncation away from that point.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper adapts the iterative linear quadratic regulator (iLQR) to the special unitary group SU(N) for multi-qubit gate synthesis. It derives Lie-algebraic equations of motion for the error state via the adjoint representation and a first-order Baker-Campbell-Hausdorff linearization, defines a geodesic final cost, extends the formulation to smoothed controls and 2-local drive restrictions, and provides a worst-case time-complexity analysis. Numerically, it compares the Lie-group iLQR against a Euclidean iLQR on two- to five-qubit benchmark gates. The central claims are that, with full Lie-algebra basis access, the Lie-group formulation reaches infidelities below 10^-15 in a single optimization iteration, and that under 2-local restrictions it converges faster in early iterations but exhibits greater sensitivity to initialization and a stronger tendency toward local minima. The paper also reports per-iteration wall-clock runtimes and tests the effect of the dexp series truncation order.","tokens_in":25438,"tokens_out":6901,"duration_ms":71209,"significance":"If the numerical claims hold, the paper makes a useful contribution to quantum optimal control: it provides a concrete SU(N) adaptation of an established Lie-group trajectory optimization framework, with analytic derivative expressions, a complexity analysis, and a systematic comparison against the Euclidean iLQR on physically motivated multi-qubit gates. The derivation in Section III and Appendix A is plausible and follows standard Lie-group calculus, and the paper honestly acknowledges limitations in the Discussion, including the unsatisfactory five-qubit constrained results and the fact that increasing time steps or dexp order does not cure them. The dexp-order experiment in Section IV.B4 is a good example of probing a secondary approximation. However, the evidence for the headline early-convergence advantage is currently provisional: the first-order BCH truncation is not directly validated, the hyperparameter selection is performed on the same benchmark problems used for the final comparison, and the code and exact hyperparameter values are not available for independent audit.","major_comments":[{"comment":"The 2-local numerical comparison rests on the first-order BCH linearization in Eq. (10) and Eq. (A14), which drops O(||(eta, sigma delta u)||^2) terms. Section IV.B4 tests only the dexp series truncation order, not the validity of this more fundamental first-order truncation at the control and error magnitudes produced by the optimizer. The full-algebra single-iteration demonstration with zero initial controls cannot validate the approximation away from the identity. Please add a direct check: compare the linearized eta_{k+1} from Eq. (10) with the exact eta_{k+1} from Eq. (9) along representative optimizer trajectories for the 2-local cases, and report the sizes of the neglected terms. Without such a check, the faster early decrease of the Lie-group curves in Figs. 2, 3, 5, and 6 could reflect the surrogate linearization rather than a genuine geometric advantage of the true SU(N) dynamics. The Discussion's observation in Section V that more time steps and higher dexp orders do not help the five-qubit gates is consistent with this concern, but it is not itself a measurement of the truncation error.","section":"Section III.B, Eq. (10) / Appendix A3 / Section IV.B4"},{"comment":"The hyperparameters for the 2-local comparison are selected by Bayesian optimization using cost functions evaluated on the same benchmark gates and the same N_s=25 seed protocol that later produce the reported medians. The manuscript does not state whether the 25 evaluation seeds are disjoint from the seeds used by the Bayesian optimizer. This creates a risk of in-sample hyperparameter selection that can bias the comparison between the Lie-group and Euclidean variants. Please either evaluate on a holdout seed set or explicitly state and justify the overlap, and report the chosen hyperparameter values for each method and gate. Without this, the claim that the Lie-group formulation consistently converges faster in early iterations is not yet fully supported.","section":"Section IV.B1-IV.B5, Eqs. (20)-(21)"},{"comment":"The central evidence of the paper is numerical, but the code is not available with the submission and the exact hyperparameter configurations (initial amplitude scale u_init, running cost R_u, final cost Q_f, and the Bayesian-optimization settings) are not reported. Reference [29] points to a generic organization page rather than a specific archived repository. To make the comparison auditable, please include a table of the best hyperparameters for every reported experiment, all random seed metadata, and make the code available as a review artifact or a persistent DOI in the revision.","section":"Section IV.B1 and Ref. [29]"},{"comment":"The single-iteration full-algebra results in Table I are reported without any seed variation, confidence intervals, or statement of how many independent runs were performed. Since the claim is about reaching machine-precision infidelities in one iteration, a few repetitions with different seeds (at least for the multi-qubit gates) are needed to rule out numerical accidents and to make the comparison with the two-iteration GEOPE result meaningful. If the observed behavior is deterministic, state so explicitly.","section":"Section V / Table I"}],"minor_comments":[{"comment":"The phrase \"single-time-step setting\" is ambiguous: it should be clarified whether this means a single optimization iteration, a discretization with T=1, or both, especially because Eq. (6) and the single-qubit example in Section IV.A1 use T=80.","section":"Section IV.A2"},{"comment":"The Fredkin gate expansion lists the term |101><110| twice and appears to omit |110><101|; please check the formula.","section":"Appendix C, Eq. (C2)"},{"comment":"The text states that the implementation retains terms up to third order in the dexp series, while Fig. 4 uses orders {0,1,3,7}. Please clarify how these order labels are counted so that the implementation and the experiment are consistent.","section":"Section IV.B4"},{"comment":"The symbol T denotes both the number of time steps and the final-state notation, while the target unitary is U_g; please rename one of these to avoid confusion in the definitions of the Bayesian optimization scores.","section":"Eqs. (20)-(21)"},{"comment":"The sentence \"the dexp series truncation is a secondary approximation that catches some higher-order terms but not all\" is imprecise; it would be helpful to state exactly which terms are retained in the implemented Jacobians and which are dropped.","section":"Section III.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of a quantum-control journal, and the derivation appears sound. My main concern is that the headline numerical comparison is not yet reproducible and may be affected by in-sample hyperparameter selection and by the untested first-order BCH truncation. All requested checks are feasible in a revision, so I would not reject the paper. I would also encourage the editor to require a code depository with a persistent identifier, since the paper's contribution is primarily numerical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The punchline: this is a solid, incremental paper. It takes the Lie-group iLQR/DDP framework that exists in robotics and applies it to SU(N) gate synthesis, with SU(N)-specific derivative formulas and a systematic comparison against Euclidean iLQR on two- to five-qubit gates. The one-iteration full-algebra results at 10^-15 are credible—they follow naturally from using the geodesic distance as the terminal cost—and the 2-local results honestly report early convergence advantages plus a tendency to local minima and degradation at five qubits. The complexity analysis is useful and shows equal worst-case O(T N^6) per iteration with a smaller state dimension in the Lie-group version.\n\nWhat the paper does well: the derivation of Eq. (10)/(A14) is careful, the BCH and dexp manipulations check out, and the numerical study is broader than most QOC papers. The authors also test dexp truncation order and two different Bayesian optimization score functions, which is more than many method papers do. The discussion of limitations is candid, including the statement that increasing T or dexp order does not fix the five-qubit results.\n\nNow the soft spots, in proportion. The stress-test note I was sent worries that the first-order BCH linearization in Eq. (10) makes the convergence curves a surrogate dynamics. I think that is a misreading. In iLQR the forward pass propagates the true dynamics—here Eq. (9) (or the original unitary evolution)—and the linearized Eq. (10) is used only to compute the descent direction and feedback gains in the backward pass. So the reported infidelities are from actual Schrödinger evolution. The BCH truncation can degrade the control law, but it does not make the plotted trajectories unreal. The paper could have stated this more explicitly, but the algorithm in Sec. III A already says g_new_{k+1}=f(g_new_k,u_new_k).\n\nThe real weaknesses are reproducibility and hyperparameter selection. The code is not released (only promised upon acceptance), and the 2-local results use hyperparameters selected by Bayesian optimization on the same benchmark gates. That makes the quantitative claims somewhat self-selected, and there are no held-out gates. Table I also omits the actual hyperparameter values and variance, though for the full-algebra setting the runs appear deterministic. These are fixable in revision. The contribution is not a breakthrough; its significance is confined to quantum optimal control, where it could be a useful warm-start tool.\n\nWho this is for: researchers in pulse-level gate synthesis, especially those already using iLQR or GEOPE. It deserves a serious referee. I would not desk reject it. My recommendation: send it to peer review, ask for the code and exact hyperparameters as a condition of acceptance, and have one referee verify the forward-pass claim about true dynamics.","headline":"A clean, incremental adaptation of Lie-group iLQR to SU(N) gate synthesis; the central convergence claims survive scrutiny, and the main weaknesses are reproducibility and benchmark-specific tuning, not the BCH truncation worry.","tokens_in":25923,"tokens_out":5291,"would_cite":true,"duration_ms":49489,"reading_group":"maybe","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 claims that Lie-group iLQR on SU(N) reaches infidelities below 10^-15 in one optimization iteration in the full-drive setting and converges faster early under 2-local drives, at the cost of more sensitivity to initialization.","keywords":["iterative linear quadratic regulator","quantum gate synthesis","Lie-group control","SU(N)","pulse optimization","trajectory optimization","2-local Hamiltonian","quantum optimal control"],"falsifier":"Take an optimized control sequence from a five-qubit WZ4 run and compare, step by step, the linearized state update of Eq. (10) with the exact discrete update eta_{k+1} = -i log(exp(-\\xi_k dt) exp(i eta_k) exp((\\xi_k + \\delta \\xi_k) dt)) from Eq. (9): if the accumulated normed difference grows to the same order as the reported final infidelity, the single-iteration convergence claim is an artifact of the linearization rather than a property of the true dynamics. A simpler variant is to propagate the optimized pulses under the exact Schr\\\"odinger equation and compare the resulting fidelity to the value reported by the optimizer.","tokens_in":24955,"feed_emoji":"⚛️","tokens_out":6123,"duration_ms":51797,"temperature":0.7,"pith_summary":"This paper adapts the iterative linear quadratic regulator (iLQR), a trajectory-optimization workhorse, from flat Euclidean coordinates to the curved manifold SU(N) that quantum gates naturally evolve on. It claims that when every Lie-algebra basis element is available as a drive Hamiltonian, the Lie-group formulation reaches gate infidelities below $10^{-15}$ in a single optimization iteration for two- through five-qubit gates, while standard Euclidean iLQR needs many iterations. Under physically realistic 2-local drive restrictions, the Lie-group version converges faster in early optimization iterations but is more sensitive to initialization and more prone to local minima. The practical stake is pulse synthesis: a geometry-aware optimizer that warms up quickly could shorten the pulse-design pipeline for quantum hardware without changing worst-case complexity.","feed_headline":"One iteration lands multi-qubit gates at 10^-15 infidelity","feed_subtitle":"Geometry-aware iLQR on SU(N) beats flat-coordinate iLQR in the full-drive setting and is faster early under 2-local constraints.","key_machinery":"The load-bearing object is the Lie-algebraic error state eta_k^\\wedge = -i log(U_k^\\dagger \\tilde U_k), which measures the geodesic gap between the current unitary and the reference unitary on SU(N). Around this error state the paper linearizes the equation of motion as eta_{k+1} \\approx Ad_{\\exp(-i(H_0+H_d u_k)\\$\\sigma$ dt)}(\\eta_k) + dt\\, H_d\\, dexp_{-i(H_0+H_d u_k)\\$\\sigma$ dt}(\\$\\sigma$ \\delta u_k), using the adjoint representation of the group and the differential of the matrix exponential, with the dexp series kept to third order. This linearization turns the Bellman backward pass into a standard quadratic regulator solve while keeping the state on the manifold, and it is what lets controls be initialized at zero and still reach machine-precision infidelity in a single step when the full algebra is accessible.","core_discovery":"The central claim is that treating the quantum state as an element of SU(N) rather than as a flat vector changes the optimization landscape enough to matter. The authors build the trajectory on Lie-algebra elements eta_k = -i log(U_k^\\dagger \\tilde U_k), the geodesic displacement from the reference to the candidate unitary, and linearize the discrete Schr\\\"odinger update through the adjoint representation Ad and the differential-of-exponential map dexp. In the idealized full-algebra setting, this makes the optimizer converge to infidelities below $10^{-15}$ in one iteration on CNOT, CCZ, Fredkin, Toffoli, and the X/Z weighted-parity gates, whereas Euclidean iLQR requires 10-22 iterations; when only 1- and 2-local drives are allowed, the Lie-group median converges faster at early iterations for three- and four-qubit gates but plateaus at worse final fidelities for five-qubit gates and exhibits broader run-to-run spread.","pith_inferences":["One untested extension is to initialize 2-local restricted runs with the single-iteration, full-algebra solution; because the full-algebra solve is near-instant, this could sidestep the sensitivity to initialization that the paper observes.","A second-order DDP variant on SU(N) would likely reduce the local-minima tendency, since the paper identifies the first-order linearization as the source of its approximation error and DDP is the natural next rung.","The dexp-order scan suggests an adaptive policy: choose the truncation order per gate based on the size of higher-order ad-terms encountered along the trajectory, rather than fixing order three for everything."],"forward_implications":["If correct, the Lie-group iLQR can serve as a one-shot pulse synthesizer for full-algebra benchmarks, producing machine-precision gates from zero-initialized controls.","For 2-local constrained gates, the faster early convergence means useful moderate-fidelity pulses can be obtained with fewer iterations for three- and four-qubit gates, though not for the five-qubit parity gates tested.","Because the worst-case complexity is O(T N^6) per iteration for both formulations, the Lie-group version's smaller state dimension (N^2-1 versus 2N^2) gives it a wall-clock edge for three- and four-qubit gates, a lead that reverses at five qubits.","The paper's suggestion that Lie-group solutions warm-start the Euclidean variant would combine fast early progress with the Euclidean method's better late-stage five-qubit behavior."],"supporting_citations":[{"why":"Supplies the Lie-group DDP/iLQR derivation, including the linearized dynamics that this paper adapts to SU(N).","marker":"[18]"},{"why":"Provides the Lie-algebraic geodesic cost function that the paper uses as its final cost.","marker":"[19]"},{"why":"Defines the Euclidean iLQR formulation for quantum control, the baseline against which convergence is compared.","marker":"[5]"},{"why":"Introduces the geodesic pulse engineering method and benchmark setting (full algebra, zero initialization) used for comparison.","marker":"[15]"},{"why":"Provides the time-dependent GEOPE method and the O(T K N^4) complexity contrast.","marker":"[16]"},{"why":"Supplies the dexp series expansion used in the linearized dynamics.","marker":"[21]"}],"fun_headline_variants":["One-shot SU(N) iLQR hits 10^-15 gate fidelity","Lie-group iLQR converges in 1 step vs 20 for gates","SU(N) geometry cuts gate optimization iterations","Full-drive gates: single iLQR iteration to 10^-15","2-local constrained gates: faster start, worse finish"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the first-order Baker-Campbell-Hausdorff linearization of the Lie-group dynamics, which drops terms of order ||(eta_k, \\$\\sigma$ \\delta u_k)||^2 and truncates the dexp series at third order, accurately describes the real Schr\\\"odinger evolution at the control amplitudes the optimizer produces; if that approximation breaks down, the reported convergence curves do not describe the true quantum dynamics.","fun_headline_variants_meta":{"raw":{"variants":["One-shot SU(N) iLQR hits 10^-15 gate fidelity","Lie-group iLQR converges in 1 step vs 20 for gates","SU(N) geometry cuts gate optimization iterations","Full-drive gates: single iLQR iteration to 10^-15","2-local constrained gates: faster start, worse finish"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1303,"prompt_tokens":928,"completion_tokens":375,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":286}},"tokens_in":544,"tokens_out":375,"duration_ms":3964,"temperature":1.0,"reasoning_tokens":286,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:48:28.742475+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an optimized control sequence from a five-qubit WZ4 run and compare, step by step, the linearized state update of Eq. (10) with the exact discrete update eta_{k+1} = -i log(exp(-\\xi_k dt) exp(i eta_k) exp((\\xi_k + \\delta \\xi_k) dt)) from Eq. (9): if the accumulated normed difference grows to the same order as the reported final infidelity, the single-iteration convergence claim is an artifact of the linearization rather than a property of the true dynamics. A simpler variant is to propagate the optimized pulses under the exact Schr\\\"odinger equation and compare the resulting fidelity to the value reported by the optimizer.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Lie-group DDP/iLQR derivation, including the linearized dynamics that this paper adapts to SU(N)."},{"cited_title":"For the three- qubit WX 2 and WZ 2 gates, the faster convergence rate of the median run is most pronounced","cited_arxiv_id":null,"evidence_quote":"Defines the Euclidean iLQR formulation for quantum control, the baseline against which convergence is compared."},{"cited_title":"Trowbridge, A","cited_arxiv_id":null,"evidence_quote":"Introduces the geodesic pulse engineering method and benchmark setting (full algebra, zero initialization) used for comparison."},{"cited_title":"Heimann, F","cited_arxiv_id":null,"evidence_quote":"Provides the time-dependent GEOPE method and the O(T K N^4) complexity contrast."},{"cited_title":"Bhattacharyya, P","cited_arxiv_id":null,"evidence_quote":"Supplies the dexp series expansion used in the linearized dynamics."}],"review_version":1}