{"id":"78c069a6-9dad-49cb-baa3-cd34460d5480","arxiv_id":"2508.16029","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new quantum optimal control algorithm that follows the geodesic on SU(2^n) converges to high-fidelity multi-qubit gates in far fewer iterations than GRAPE, including 5- and 6-qubit quantum Fourier transforms on Rydberg atom arrays.","lead":"Researchers develop a new algorithm, GEOPE, for designing quantum logic gates under hardware constraints. It follows the geodesic path on the space of unitary evolutions, and numerical tests show it converges much faster than the standard GRAPE method for multi-qubit gates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"5-qubit GRAPE comparison reuses 3-qubit hyperparameters; because second-order GRAPE is hyperparameter-sensitive (Figs. F1/F2), the 'beyond GRAPE' scaling claim may rest on an unequal comparison.","rationale":"The reader's CONDITIONAL verdict is sound. I focus on the 5-qubit comparison because it is the pivotal evidence for the scaling part of the central claim, and the paper explicitly admits the hyperparameter reuse. The 3-qubit comparisons are well-supported and fair, so this does not justify changing the verdict; it sharpens what would settle the concern. The reader's weakest_assumption (Gram-Schmidt escape) is a real algorithmic heuristic concern, but the hyperparameter asymmetry is more directly load-bearing for the comparative claim, since it could make GRAPE's failure an artifact of the evaluation rather than a genuine limitation. I therefore partially agree with the reader: the same issue is noted in their rationale, but they selected a different weakest assumption.","tokens_in":21779,"tokens_out":9312,"duration_ms":103921,"concrete_test":"Run GRAPE-Adam, GRAPE-NR, and GRAPE-RFO on the same 5-qubit QFT (L=120, Rydberg graph Fig. 3(c), ε=1e-9, 100 random seeds) with hyperparameters tuned directly on this problem, e.g. Bayesian optimization over λ for Adam, δ for NR, κ for RFO using the same cumulative-infidelity objective (Eq. E1). Report cumulative success probability versus iterations and wall-clock time. If any GRAPE variant reaches high success within ~300 iterations (or within GEOPE's wall-clock time), the claim that GRAPE cannot access the 5-qubit QFT is falsified; if all still fail, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GEOPE converges 'significantly faster than both first-order and second-order GRAPE' and finds gates 'not accessible to GRAPE' depends critically on Fig. 5. For that 5-qubit QFT, the paper explicitly states: 'we reused the hyperparameters found for 3-qubit gates instead' rather than performing a 5-qubit hyperparameter search, citing the cost of Hessian computation. This is a serious asymmetry because App. E tuned hyperparameters only on 3-qubit targets (Toffoli, CCZ, 3-QFT; Table E1), and Figs. F1/F2 show the Newton-Raphson and RFO variants are highly sensitive to their single hyperparameter (δ or κ): different values change whether a solution is found within 200 iterations. There is no justification that the 3-qubit-optimal δ/κ transfer to a 120-layer, 5-qubit control landscape with a different interaction graph and far more parameters. Thus, the 5-qubit failure of GRAPE could be an artifact of poorly chosen hyperparameters rather than a fundamental limitation. If tuned GRAPE were competitive on the 5-qubit QFT, the paper's strongest scaling conclusion and the 'unprecedented' claim would be undercut. The 3-qubit comparisons are fair and do provide genuine support for the core GEOPE method, so the right remedy is additional evidence, not rejection.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GEOPE, a quantum optimal control algorithm that replaces the fidelity-gradient ascent of GRAPE with an update direction obtained by least-squares projection of the geodesic direction (the principal-branch logarithm from the current unitary to the target) onto the span of the Jacobians of the constrained Hamiltonian parameters. A golden-section line search sets the step, and a Gram-Schmidt random step (Eq. B17) is used when the line search cannot improve fidelity. The algorithm is demonstrated on Rydberg-atom Hamiltonians for 3-qubit Toffoli/CCZ gates with 12 and 20 piecewise steps, a 5-qubit QFT with L=120, and a 6-qubit QFT with L=400. The central claims are that GEOPE converges significantly faster than both first-order (Adam) and second-order (Newton-Raphson, RFO) GRAPE and that it can find 5- and 6-qubit QFT gates that the paper's GRAPE implementations could not.","tokens_in":22136,"tokens_out":2950,"duration_ms":37763,"significance":"If the scaling claims hold, the paper would contribute a genuinely new geometric principle to quantum optimal control: instead of locally maximizing fidelity, each update follows the known geodesic to the target as closely as the constrained control landscape allows. The convex least-squares subproblem is clearly specified, the algorithm is simple to implement, and the authors provide code, which strengthens reproducibility. The 3-qubit comparisons use Bayesian-tuned hyperparameters for all methods and show large iteration-count advantages for GEOPE; this part is credible and useful. However, the paper's strongest conclusions—the 'significant' speedup over GRAPE and the 'unprecedented' 5- and 6-qubit QFT results—rest on the 5-qubit comparison of Fig. 5, which the authors admit reuses 3-qubit hyperparameters, and on a 6-qubit demonstration with no GRAPE comparison at all. Given the demonstrated sensitivity of the second-order GRAPE methods to their single hyperparameters (Figs. F1 and F2), the beyond-3-qubit claims are not yet established at the level of the paper's conclusions.","major_comments":[{"comment":"The 5-qubit QFT comparison is the load-bearing evidence for the 'beyond GRAPE' scaling claim, but the paper explicitly states that the 3-qubit hyperparameters were reused for the 5-qubit GRAPE runs. App. E tunes hyperparameters only on 3-qubit targets, and Figs. F1 and F2 show that the Newton-Raphson and RFO variants are highly sensitive to their δ or κ values: for the same gate and L, different hyperparameter choices change whether a solution is found within 200 iterations. There is no argument that δ or κ optimized for 12/20-layer 3-qubit problems transfers to a 120-layer 5-qubit landscape with a different interaction graph and far more parameters. Thus Fig. 5 does not establish that GRAPE cannot find the 5-qubit QFT with appropriate tuning; it may only establish that the reused hyperparameters were poor. The authors should either perform a 5-qubit hyperparameter search (at least for a","section":"§III, Fig. 5; App. E, Table E1; Figs. F1/F2"},{"comment":"The 6-qubit QFT result is reported as 'well beyond the capabilities of our GRAPE implementation,' but no GRAPE data, runtime, success probability, or infidelity-vs-iteration curve is shown for this case. The only quantitative detail is L=400 and that parameter values are in the repository. Since the 6-qubit claim is part of the conclusion ('unprecedented 5- and 6-qubit Quantum Fourier Transform gates'), the absence of any comparison or even a GEOPE success statistic makes the claim unverifiable from the manuscript. Provide at least the number of trials, the success rate, the final infidelity, and, if possible, a GRAPE baseline with documented hyperparameters and a wall-clock comparison.","section":"§III, 6-qubit paragraph; Ref. [45]"},{"comment":"The Gram-Schmidt escape is a load-bearing heuristic: whenever the projected geodesic direction cannot improve fidelity, the algorithm steps in a random direction orthogonal to γ, with step size ηGS=1.2ηmax. The paper states only that this 'minimises the chance that the algorithm steps back into the same minimum.' No analysis, convergence guarantee, or ablation is provided. Since the 5- and 6-qubit successes depend on escaping local minima reliably, this unverified heuristic underlies the main numerical claims. The authors should at least report the frequency with which the escape branch is taken for the reported gates and test sensitivity to ηGS and to the random seed of the escape; ideally, compare against an alternative restart strategy.","section":"App. B I, Eq. (B17); Algorithm 1"},{"comment":"The paper's speed comparisons are reported in algorithmic iterations, but App. D states that GEOPE has complexity O(KLN^4) whereas GRAPE has O(KLN^3), with K=O(n^2) and N=2^n. A factor-N-per-iteration difference is substantial for n=5–6, yet the conclusion claims GEOPE 'converges significantly faster.' Iteration count alone does not establish practical speedup; the 5-qubit text notes second-order GRAPE took hours while GEOPE took minutes, but this is anecdotal and confounded by the reused hyperparameters. Please report wall-clock times or iteration-normalized runtimes for all methods on the same hardware, and discuss whether the O(N) per-iteration overhead is offset by the observed iteration savings in the regimes advertised.","section":"App. D; §III, Fig. 4"}],"minor_comments":[{"comment":"Typo: 'illustrtated' should be 'illustrated.'","section":"Abstract"},{"comment":"The appendix defines Φ as 'Matrix constructed from the L restricted Lie algebra vectors θl'; this should be 'ϕl' to match the main text and avoid confusion with unrestricted vectors θ.","section":"App. A, notation for Φ"},{"comment":"The sum 'j∀Gj∈H' is notationally awkward and should be written as a set summation over basis elements in H; also the index of δϕ(m)_{l,j} should be made consistent with j labeling the restricted basis element.","section":"Eq. (B15)"},{"comment":"The loops 'for l ∈ (1, . . . , N2 − 1)' and 'j ← P_{N2-1}' use 'N2' where the text elsewhere writes N=2^n; the intended N^2−1 should be spelled out to avoid ambiguity.","section":"Algorithm 1"},{"comment":"References [32] and [39] are the same work (arXiv preprint and published version). Citing both is acceptable, but the main text should avoid implying they are two distinct prior methods; consider citing only the published version once the preprint is updated.","section":"Refs. [32] and [39]"}],"recommendation":"major_revision","confidential_remarks":"The 3-qubit core of the paper is sound and the geometric formulation is a refreshing contribution. My recommendation is driven by the gap between the cautious 3-qubit evidence and the sweeping 5/6-qubit claims in the abstract and conclusion. The fix—additional tuned GRAPE baselines and quantified GEOPE success statistics for 5 and 6 qubits—is within the scope of a revision and does not require rejecting the core idea. I would also encourage the editor to check the novelty disclosure relative to the patent application declared in the acknowledgments, though this is not a technical flaw."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: read this one. The L>1 piecewise geodesic pulse engineering algorithm is a real addition to quantum optimal control — the Jacobian-projection objective (Eq. 6), the line search, and the Gram-Schmidt escape are new relative to the earlier L=1 work, and the geometric rationale for why GRAPE wanders is clearly worked out, including a concrete 2-qubit example. The 3-qubit numerical section is the paper's solid core: all methods get Bayesian-tuned hyperparameters, GEOPE reaches unit cumulative success in far fewer iterations, and the robustness plots (F1/F2) show GEOPE degrades gracefully where Newton-Raphson and RFO are touchy. Code and pulse parameters are on GitHub, and the complexity discussion is honest about the O(KLN^4) vs O(KLN^3) trade-off.\n\nThe soft spots are exactly where the reader's report and the stress-test put them. The 5-qubit QFT figure compares GEOPE against GRAPE with hyperparameters inherited from 3-qubit problems, which the paper admits. Given Figs. F1/F2 show NR and RFO are hypersensitive to their single parameter — different choices flip whether a solution appears within 200 iterations — this is a serious asymmetry, not a stylistic detail. The conclusion's strongest sentence, 'beyond what is within reach for GRAPE,' is under-supported by that figure. The 6-qubit QFT result is asserted, with final parameters only in the repo and no convergence curve or fidelity history in the text. A referee should ask for either tuned 5-qubit GRAPE runs or a softened claim, and for a 6-qubit convergence plot. The Gram-Schmidt escape is a heuristic and unanalyzed; the paper says it only 'minimises the chance' of returning to the same minimum. That is a legitimate limitation to flag, but not a fatal one — the 3-qubit success doesn't depend on it and the method stands even if escape reliability varies.\n\nThe circularity concern does not land. The update derives from the known geodesic to the target and the Jacobian of the control map; there is no target fitting in the construction, and the repeated self-citation is to a distinct L=1 algorithm, so that is fine.\n\nBottom line: send this to serious peer review. The core method is novel, the 3-qubit evidence is strong, and the scaling claims are fixable with additional experiments rather than wrong. A competent referee could push the authors to either tune GRAPE at scale or tone down the 'not accessible to GRAPE' wording, and the paper would be better for it.","headline":"Genuinely new algorithm with a solid 3-qubit win over GRAPE; the 5- and 6-qubit claims need the missing comparisons before the headline 'beyond GRAPE' can be taken at face value.","tokens_in":22633,"tokens_out":2543,"would_cite":true,"duration_ms":27752,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.-a","02.40.Ky"],"model":"deepseek-v4-flash","headline":"This paper introduces GEOPE, a geodesic-guided quantum optimal control algorithm that solves each update as a convex least-squares projection of the constrained Hamiltonian's available directions onto the shortest path to the target gate.","keywords":["quantum optimal control","geodesic pulse engineering","GRAPE","Rydberg atom arrays","multi-qubit gates","Riemannian manifold","quantum Fourier transform","piecewise constant pulses"],"falsifier":"Run the five-qubit QFT benchmark (L = 120) with the Gram-Schmidt escape branch disabled: if GEOPE's success probability collapses, its advantage rests on that unverified heuristic rather than on geodesic alignment. Conversely, run the same benchmark with random escape directions but no geodesic alignment: if success is comparable, the geodesic projection itself is not doing the claimed work.","tokens_in":21652,"feed_emoji":"⚛️","tokens_out":5922,"duration_ms":66570,"temperature":0.7,"pith_summary":"This paper introduces GEOPE, an optimal-control algorithm that sets the parameters of a piecewise-constant pulse sequence by aiming each update along the shortest path—the geodesic—on the manifold of unitary matrices between the current evolution and the target gate. The update direction is found by projecting the geodesic direction onto the directions the hardware Hamiltonian can actually generate, which turns into a convex least-squares problem, followed by a line search along that direction. The authors report that GEOPE reaches high-fidelity solutions in at least ten times fewer iterations than three GRAPE variants for Toffoli, CCZ, and three-qubit QFT gates on Rydberg atom arrays, and that it can find five- and six-qubit QFT gates that their GRAPE implementations could not solve in reasonable time. If true, it makes practical optimal control of larger multi-qubit gates more feasible under realistic hardware constraints, without relying on second derivatives.","feed_headline":"GEOPE finds six-qubit QFT pulses where GRAPE stalls","feed_subtitle":"Geodesic updates on the unitary manifold converge in ten times fewer steps than GRAPE and reach 5- and 6-qubit gates.","key_machinery":"The central object is the geodesic on the Riemannian manifold SU(2^n) from the current gate U_G(Φ) to the target V, generated by Γ = -i log(U_G^† V), with the logarithm taken on its principal branch so the path is the shortest one. GEOPE's update is the minimizer of a convex least-squares problem that matches the Jacobian-generated tangent directions of the restricted Hamiltonian to this geodesic direction; when the line search cannot improve the fidelity, a Gram-Schmidt procedure steps in a direction orthogonal to the geodesic to exit the local minimum. This replaces the non-convex fidelity maximization of GRAPE with a convex projection plus line search.","core_discovery":"The central claim is that constrained quantum optimal control can be solved more efficiently by following geodesics on SU(2^n) rather than by ascending the fidelity landscape. At each algorithmic step, the current unitary and the target define a unique shortest geodesic; GEOPE computes its tangent generator, Γ = -i log(U_G^† V), using the principal branch of the matrix logarithm, and solves a linear least-squares problem to express that tangent direction as a combination of the Jacobians of the allowable control parameters as closely as the hardware restrictions permit. A golden-section line search then chooses the step size. The authors present numerical evidence that this rule converges to","pith_inferences":["Editorial inference: the geodesic projection likely acts as a preconditioner that keeps updates aligned with the global target rather than the local fidelity gradient, so the advantage over GRAPE may grow as the number of qubits or the hardware restrictions increase—this could be tested by scaling benchmarks across different interaction graphs.","Editorial inference: the unproven Gram-Schmidt escape is the least-controlled part of the loop; a deterministic second-order correction in the orthogonal space might replace it and make the algorithm's success less reliant on random restarts.","Editorial inference: the convex formulation invites combining GEOPE with constrained least-squares solvers to enforce pulse amplitude or bandwidth limits directly, which could be verified by adding box constraints to the update problem.","Editorial inference: comparing GEOPE against gradient-free methods such as CRAB on the same Rydberg benchmarks would clarify whether the speedup comes specifically from geodesic alignment or from the convex projection step alone."],"forward_implications":["Gradient-free geometric steering can outperform both first- and second-order GRAPE without computing a Hessian, so the practical bottleneck shifts from convergence rate to the cost of Jacobian evaluation.","Five- and six-qubit quantum Fourier transform gates become numerically accessible under Rydberg atom array constraints, well beyond what the paper's GRAPE implementations reached.","Because GEOPE only needs the set of accessible Hamiltonian terms, the same algorithm can be applied to ion traps, superconducting qubits, or semiconductor quantum dots by changing the restriction set.","The loss function can be extended to penalize pulse-to-pulse jumps and total evolution time, pointing toward smooth, experimentally friendlier pulses.","The same geodesic-update idea could be rephrased on Hilbert space or homogeneous spaces for state preparation, as the paper itself suggests."],"supporting_citations":[{"why":"Original GRAPE algorithm, the baseline first-order gradient-ascent method that GEOPE is compared against.","marker":"[7]"},{"why":"Accelerated Newton-Raphson GRAPE methods that serve as the second-order baselines.","marker":"[1]"},{"why":"Prior geodesic algorithm for a single time-independent Hamiltonian; GEOPE generalizes it to piecewise pulse sequences.","marker":"[39]"},{"why":"Block-matrix identity that lets GEOPE compute Jacobians of each piecewise step efficiently.","marker":"[43]"},{"why":"Adam optimizer, the adaptive first-order GRAPE variant used for comparison.","marker":"[42]"},{"why":"Modified Newton-Raphson and rational-function optimization GRAPE variants used as comparison baselines.","marker":"[44]"},{"why":"Rydberg atom array Hamiltonian model used for all numerical demonstrations.","marker":"[40]"},{"why":"Power-law r^-6 van der Waals interactions that set the coupling strengths in the Rydberg interaction graphs.","marker":"[41]"},{"why":"Bayesian optimization used to tune each method's hyperparameters for a fair comparison.","marker":"[46]"}],"fun_headline_variants":["Geodesic pulse engineering beats GRAPE on multi-qubit gates","New algorithm follows geodesics to speed up quantum gate design","Geodesic paths in SU(2^n) accelerate quantum optimal control","GEOPE: quantum gate design that outpaces GRAPE up to six qubits","From GRAPE to GEOPE: faster convergence for quantum gates"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The speedup depends on the Gram-Schmidt escape step reliably pulling the search out of local minima when the best geodesic-aligned update cannot improve the fidelity, and this heuristic is neither analyzed nor proven.","fun_headline_variants_meta":{"raw":{"variants":["Geodesic pulse engineering beats GRAPE on multi-qubit gates","New algorithm follows geodesics to speed up quantum gate design","Geodesic paths in SU(2^n) accelerate quantum optimal control","GEOPE: quantum gate design that outpaces GRAPE up to six qubits","From GRAPE to GEOPE: faster convergence for quantum gates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1228,"prompt_tokens":686,"completion_tokens":542,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":450}},"tokens_in":430,"tokens_out":542,"duration_ms":5759,"temperature":1.0,"reasoning_tokens":450,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:34:36.994775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the five-qubit QFT benchmark (L = 120) with the Gram-Schmidt escape branch disabled: if GEOPE's success probability collapses, its advantage rests on that unverified heuristic rather than on geodesic alignment. Conversely, run the same benchmark with random escape directions but no geodesic alignment: if success is comparable, the geodesic projection itself is not doing the claimed work.","supporting_citations":[{"cited_title":"Khaneja, T","cited_arxiv_id":null,"evidence_quote":"Original GRAPE algorithm, the baseline first-order gradient-ascent method that GEOPE is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Accelerated Newton-Raphson GRAPE methods that serve as the second-order baselines."},{"cited_title":"Lewis, R","cited_arxiv_id":null,"evidence_quote":"Prior geodesic algorithm for a single time-independent Hamiltonian; GEOPE generalizes it to piecewise pulse sequences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Block-matrix identity that lets GEOPE compute Jacobians of each piecewise step efficiently."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Modified Newton-Raphson and rational-function optimization GRAPE variants used as comparison baselines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Power-law r^-6 van der Waals interactions that set the coupling strengths in the Rydberg interaction graphs."}],"review_version":1}