{"id":"d5ca7bcb-f907-4249-ae39-02f44d0cbea8","arxiv_id":"1908.06283","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A push-pull objective, using the target plus orthogonal operators, improves convergence and fidelity of GRAPE and Krotov quantum control, demonstrated numerically and in an NMR singlet-order experiment.","lead":"This paper proposes a \"push-pull\" way to optimize quantum control pulses: the objective rewards reaching the target operation and penalizes being close to a random set of orthogonal operations. Tests on simulated qubits and on an NMR molecule show faster convergence and higher final fidelities than standard GRAPE and Krotov routines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The PP-Krotov update rule (Eq. 14) is not derived from the stated objective JPP and is self-referential in the control amplitudes; the PP-Krotov numerics and NMR demonstration are therefore not yet verified.","rationale":"The reader's weakest_assumption focuses on post-hoc selection of Lbest and alpha. That is a real generalization concern, but the more immediate internal obstacle is that the PP-Krotov algorithm, which generates a large share of the numerical evidence and the only experimental demonstration, is not shown to optimize the defined objective. Eq. 14 contains an unexplained additional term and is an implicit equation in the updated control amplitudes, making the algorithm underspecified as written. This is a correctness risk that is independent of parameter selection: even with a fixed L and alpha, a reader cannot verify that the reported PP-Krotov curves correspond to push-pull optimization of JPP. The concrete re-derivation and rerun would settle the issue. Because the suspect Eq. 14 underlies the Krotov-based claims, I move the verdict from CONDITIONAL to UNVERDICTED pending that check; the GRAPE-based results may still support the central claim, but the Krotov and experimental legs are not yet verifiable.","tokens_in":8329,"tokens_out":9984,"duration_ms":112392,"concrete_test":"Independently derive the Krotov stationarity condition for piecewise-constant controls maximizing JPP of Eq. 5, including the resource penalty. Verify whether the first-order update is exactly Eq. 11 with the combined boundary B_N = <Ut|U0:N>Ut - (alpha/L) sum_l <Vl|U0:N>Vl for gate control (and the corresponding SC boundary), or whether it contains the extra ~v term of Eq. 14. If the extra term is absent, implement the derived update and rerun the two-qubit CNOT and state-transfer benchmarks of Fig. 2e-h and the TCP LLS sequence of Fig. 4 using L = 5 and alpha = 0.2. If the mean final fidelity advantage over pull-only disappears, or if the 45 ms sequence no longer predicts average fidelity above 95%, then the PP-Krotov and NMR support for the central claim fails; if Eq. 14 reduces to the derived update after fixing the implicit definition, the concern is resolved.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim requires PP-Krotov to actually optimize the push-pull objective JPP = F - alpha Fo, but the connection is not established. For a Krotov maximization of this objective, the terminal Lagrange multiplier should be the single combined boundary B_N = d(F - alpha Fo)/dU, namely B_N = <Ut|U0:N>Ut - (alpha/L) sum_l <Vl|U0:N>Vl for gate control, and the first-order update should be Eq. 11 with this combined B_N. Instead, Eq. 14 appends an extra term (alpha delta/L) sum_l [~v_{jkl} - (1/lambda_k)<C^{(i-1)}_{jl}|A_k U^{(i)}_{0:j-1}>], and then defines ~v^{(i)}_{jkl} = (alpha eta/L)[u^{(i)}_{jk} - (1/lambda_k) sum_l Im<C^{(i)}_{jl}|A_k U^{(i)}_{0:j}>]. This extra term does not follow from the variational principle stated in the paper, and because ~v depends on the same u^{(i)}_{jk} being updated, Eq. 14 is an implicit equation; the paper does not explain how it is solved. Since PP-Krotov supplies much of the numerical evidence (Fig. 2e-h for Krotov, Fig. 3 for QFT) and the only experimental demonstration (Fig. 4, TCP LLS preparation), the central claim is not yet supported by these results unless Eq. 14 can be independently derived or the implemented algorithm is specified and shown to optimize JPP. No code or complete parameter sets are provided, so this cannot be checked from the manuscript alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'push-pull optimization of quantum controls' (PPOQC), a modification of the standard quantum optimal-control objective that adds a penalty term depending on a set of operators orthogonal to the target. The combined objective is JPP = F - alpha Fo - sum lambda_k r_k (Eq. 5), where Fo is the average fidelity to L orthogonal operators (V_l for gate control, R_l for state control). The authors show how to incorporate this objective into GRAPE, giving PP-GRAPE via modified gradients (Eq. 7), and into Krotov's method, giving PP-Krotov via a revised update rule (Eq. 14). They report numerical results for two-qubit CNOT and singlet-state transfer (Fig. 2), for quantum Fourier transforms on up to seven qubits (Fig. 3), and an NMR demonstration preparing long-lived singlet order in 2,3,6-trichlorophenol with a pulse sequence that is 30% shorter and yields 27% higher singlet order than the standard method (Fig. 4). The central claim is that the push-pull objective leads to faster convergence, better exploration of parameter space, and higher final fidelities than standard pull-only algorithms.","tokens_in":8682,"tokens_out":4394,"duration_ms":47404,"significance":"If the central claim holds, the paper offers a simple and potentially general modification to two widely used quantum-control optimization routines, with large reported improvements (in some cases two orders of magnitude in infidelity) and a concrete experimental demonstration. The strength of the manuscript lies in the transparent definition of the push-pull objective, the explicit incorporation into GRAPE and Krotov frameworks, the broad numerical evidence up to seven qubits, and the experimental verification of a PP-Krotov state-control sequence. The paper also honestly shows that the computational overhead of PP-Krotov grows linearly in L and that the advantage factor varies by task. However, the manuscript does not provide machine-checked proofs or reproducible code, and the PP-Krotov update is introduced without a derivation from the stated objective. The reported advantage factors are computed after selecting Lbest on the same test problems, and the push weight alpha is tuned on the same suite, so the out-of-sample predictive content of the numerical claims is not yet established.","major_comments":[{"comment":"The update rule in Eq. (14) is not derived from the variational principle applied to the stated objective JPP in Eq. (5). For a Krotov maximization of the combined terminal objective F - alpha Fo, the terminal costate should be the single combined boundary B_N = <Ut|U0:N>Ut - (alpha/L) sum_l <Vl|U0:N>Vl for gate control, and the first-order update should have the form of Eq. (11) with this combined boundary. Instead, Eq. (14) appends an extra term containing ~v_{jkl}, which is defined in terms of the same u^{(i)}_{jk} being updated, so the equation is implicit and the paper does not explain how it is solved or iterated. Since PP-Krotov supplies much of the numerical evidence (Figs. 2e-h and 3) and the experimental demonstration (Fig. 4), the central claim requires either a derivation of Eq. (14) from a well-defined Lagrangian/costate system or a precise statement of the actually implemented iteration together with a proof or numerical check that it optimizes (or at least monotonically improves) JPP.","section":"Push-pull Krotov (PP-Krotov), Eq. (14)"},{"comment":"The text states that the set of orthogonal operators 'can be generated randomly and efficiently in every iteration,' but the numerical experiments do not specify whether {V_l} (or {R_l}) is fixed during a given optimization run or regenerated at each iteration. If the set is regenerated during the run, the objective JPP itself changes from iteration to iteration, so the comparison with pull-only methods is not a comparison against a well-defined fixed objective. If the set is fixed, that should be stated explicitly. This ambiguity affects the interpretation of the convergence curves in Figs. 2 and 3 and should be resolved before the numerical superiority claim can be assessed.","section":"Introduction and Numerical analysis"},{"comment":"The advantage factor is defined as (1-F(L=0))/(1-F(Lbest)) with Lbest chosen as the set giving the maximum mean final fidelity on the same test problems, and the push weight is fixed at alpha=0.2 after a scan over alpha reported in Supplementary Fig. 6. These selections are made in-sample on the very problems used for the comparison, so the reported gains (up to a factor of 64) may reflect favorable selection rather than a general property of the method. To support the claim that push-pull optimization is superior, the authors should either use a predetermined rule for alpha and L (for example, a fixed small L and alpha in [0.1, 0.3]) or report hold-out/cross-validated performance on problems not used for tuning.","section":"Numerical analysis, Fig. 2(m-p), and Supplementary Fig. 6"},{"comment":"No code, data, or complete parameter sets (for example, the values of epsilon, delta, eta, kappa, penalty constants lambda_k, and the initial-guess generation protocol) are provided, and the NMR pulse sequence is shown only graphically. Because the central claim rests on numerical comparisons and the experimental demonstration depends on a specific implementation of Eq. (14), the absence of these details prevents independent verification of the results. The authors should make the code and data available or, at minimum, provide full parameter tables and a precise algorithmic pseudocode for both PP-GRAPE and PP-Krotov.","section":"Numerical analysis and Experimental demonstration"}],"minor_comments":[{"comment":"The push weight alpha is restricted to -1 <= alpha <= 1, but the text does not explain why negative values are allowed or what a negative push weight would mean dynamically; this deserves a brief comment.","section":"Eq. (5) and Introduction"},{"comment":"The notation in the Krotov section is confusing: the performance function is J but the Lagrangian is denoted L, and the stationarity condition 'partial L / partial F = 0' mixes the functional and a variable; defining the variational derivative clearly would improve readability.","section":"Krotov optimization, Eqs. (8)-(12)"},{"comment":"The caption contains the typo 'Eovlution' (should be 'Evolution'), and the figure itself is visually dense; labeling the panels and axes more explicitly would help.","section":"Supplementary Fig. 7"},{"comment":"Reference [9] is cited as 'eMagRes (2007)' without volume or article identifier, which is incomplete by journal standards.","section":"References"},{"comment":"The Bloch-sphere model is presented as a heuristic hint rather than a derivation; this is acceptable, but the text should explicitly caution that the model uses an instantaneous state and does not capture the full feedback dynamics of the optimization.","section":"Supplementary Material, 'A naive model'"},{"comment":"The comparison with the 'standard method' (Ref. [35]) reports a 27% higher singlet order and a 30% shorter sequence, but the standard pulse shape and its fidelity under the same RF inhomogeneity are not shown; a quantitative side-by-side comparison would strengthen the experimental claim.","section":"Experimental section, Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the push-pull idea is potentially useful, but the current version needs major revision before I can recommend acceptance. The most serious technical concern is the PP-Krotov update rule (Eq. 14), which is not derived from the stated objective and is implicit in the controls; this underpins both the numerical and the experimental claims. The in-sample selection of Lbest and alpha also needs to be addressed, and the lack of code/data makes verification difficult. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Push-pull is a simple idea: add a penalized average fidelity to a handful of operators orthogonal to the target, so the gradient both pulls to the target and pushes away from those states. The GRAPE variant is clean and the gradient formula (Eq. 7) is correct given the stated objective. Numerical trends are consistent—PP-GRAPE beats pull-only in mean final fidelity across all tested cases, and the two-qubit and QFT tests suggest the benefit does not vanish at 7 qubits. The NMR demonstration is a real experiment and shows a meaningful improvement in singlet order preparation. Credit where due: the trick is easy to adopt and likely helpful in practice.\n\nThe soft spots are real. First, the PP-Krotov update (Eq. 14) is not derived from the push-pull objective. A variational treatment of JPP would give a single combined boundary for the target and the orthogonal operators; instead the paper appends an extra term containing ~v, which depends on the very u being updated. That is an implicit equation, and no solution method or derivation is given. Since the QFT numerics and the NMR experiment rest on PP-Krotov, that half of the paper is unverified. Second, the advantage factors (up to 64) are computed after choosing Lbest as the set with the best mean final fidelity, and alpha is scanned in the supplement; the headline numbers are therefore in-sample, not out-of-sample. Third, no code or full parameter sets are provided, so the Krotov behavior cannot be independently checked.\n\nThe GRAPE half could stand alone as a useful note. The Krotov half needs either a proper derivation, a clear statement that it is a heuristic (with numerical evidence that it actually moves in the right direction), or released code. The authors may well have a working heuristic, but the paper currently oversells it as a variational method.\n\nFor a quantum-control researcher looking for a cheap improvement to GRAPE, this is worth reading. The larger claims need more support. I would send it to peer review—the idea is novel enough to warrant referee time—but would ask for a rewrite of the Krotov section and an out-of-sample performance analysis. If the Krotov update cannot be justified, the paper should be trimmed to the GRAPE result and the experiment.","headline":"A clean and useful trick for GRAPE, but the Krotov implementation is underived and the headline numbers are tuned in-sample; send it for review but demand a rewrite.","tokens_in":9225,"tokens_out":6086,"would_cite":false,"duration_ms":55621,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adding a penalized average fidelity to a set of orthogonal target operators speeds convergence and improves solutions in quantum control.","keywords":["quantum optimal control","push-pull optimization","GRAPE","Krotov","orthogonal operators","long-lived singlet order","NMR","quantum Fourier transform"],"falsifier":"Fix $L$ and $\\alpha$ before the runs and average the final fidelity over many randomly drawn orthogonal-operator sets; if the mean no longer beats pull-only GRAPE and Krotov on the CNOT, singlet-transfer, and QFT tasks, the central claim fails. An 8-qubit QFT with $L=1$ would test whether the advantage survives further scaling.","tokens_in":8101,"feed_emoji":"⚛️","tokens_out":6242,"duration_ms":60119,"temperature":0.7,"pith_summary":"This paper tries to establish that a quantum control objective becomes stronger when a set of operators orthogonal to the target is included alongside the target itself. The resulting push-pull objective $J_{\\mathrm{PP}} = F - \\alpha F_o - \\sum_k \\lambda_k r_k$ is reported to converge faster and to reach higher final fidelities than standard GRAPE and Krotov optimizations, with advantage factors up to 64 in numerical tests. A sympathetic reader would care because the modification is simple, works for both gate control and state control, and appears to help exactly where pull-only methods stall in local minima.","feed_headline":"Push-pull trick sharpens quantum control up to 64-fold","feed_subtitle":"Adding orthogonal operators to the fidelity target speeds up GRAPE and Krotov, and trims an NMR singlet pulse by 30 percent.","key_machinery":"The load-bearing object is the push-pull performance function $J_{\\mathrm{PP}} = F - \\alpha F_o - \\sum_k \\lambda_k r_k$: the target fidelity $F$ pulls the search toward the target, while $F_o$, the mean fidelity to $L \\le d-1$ orthogonal operators, pushes the search away from directions orthogonal to it. In GRAPE this enters by replacing the gradient with $G = g(U_t) - (\\alpha/L)\\sum_l g(V_l)$; in Krotov it enters through extra co-sequences and terminal Lagrange multipliers for each orthogonal operator. The argument is carried by this gradient modulation, which the supplement shows varying more rapidly than in pull-only runs.","core_discovery":"The central claim is that replacing the target-only objective with the push-pull performance function $J_{\\mathrm{PP}} = F - \\alpha F_o - \\sum_k \\lambda_k r_k$—where $F$ is the target fidelity and $F_o$ is the average fidelity to $L$ orthogonal operators that each have zero overlap with the target—makes quantum control optimization converge faster and to better solutions. The mechanism is a pull toward the target combined with a push away from the orthogonal operators, which the authors argue produces more thorough exploration of the parameter space. Numerically, PP-GRAPE and PP-Krotov beat their pull-only counterparts in mean final fidelity in every tested case, and an NMR experiment preparing long-lived singlet order used a 30% shorter sequence with 27% higher singlet order than the standard one.","pith_inferences":["Because the advantage factor is computed after picking $L_{\\mathrm{best}}$ from the same runs and the push weight is scanned in the supplement, the reported gains are in-sample; averaging over pre-fixed $(L,\\alpha)$ choices would give a fairer estimate.","The push term behaves like an exploration regularizer that modulates gradients during the search, so PPOQC may also help in other settings where target-only gradients stall, such as open-system control and gate compilation.","A natural testable extension is an adaptive push weight $\\alpha(i)$ that starts large to explore and decays toward zero; if such a schedule preserves the gains, sensitivity to the fixed value $\\alpha=0.2$ would no longer matter."],"forward_implications":["Adding a push term to the performance function improves mean final fidelity over pull-only GRAPE and Krotov in every two-qubit case tested, with advantage factors up to 64.","PPOQC can be added to both gradient-ascent and variational-principle routines; PP-GRAPE adds negligible computing time as $L$ grows, while PP-Krotov's time rises linearly.","The advantage persists for larger registers: a single PP-Krotov sequence implements an $n$-qubit quantum Fourier transform with $n$ up to 7.","An experimental PP-Krotov sequence prepares long-lived singlet order in NMR with a 30% shorter pulse that yields 27% higher singlet order than the standard sequence and tolerates 10% RF inhomogeneity."],"supporting_citations":[{"why":"Supplies the GRAPE gradient method and the pull-only baseline that PP-GRAPE extends by subtracting averaged push gradients.","marker":"[13]"},{"why":"Foundational Krotov variational-principle optimization that PP-Krotov modifies with extra co-sequences for orthogonal operators.","marker":"[16]"},{"why":"Krotov's global optimal control theory provides the variational background for the PP-Krotov update rules.","marker":"[33]"},{"why":"Provides the Gram-Schmidt orthogonalization used to construct operators orthogonal to the target.","marker":"[32]"},{"why":"Establishes that singlet order in a homonuclear spin pair outlives $T_1$, the physical effect the NMR experiment exploits.","marker":"[34]"},{"why":"Describes the standard long-lived-state preparation sequence used as the experimental baseline; PP-Krotov's sequence is 30% shorter and yields 27% higher singlet order.","marker":"[35]"}],"fun_headline_variants":["Push-pull trick beats pull-only quantum control","Orthogonal targets boost quantum optimization","Push-pull quantum control speeds convergence","Shorter NMR pulse with push-pull method","Push-pull optimization sharpens quantum gates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a small, randomly regenerated set of orthogonal operators, used with a fixed push weight, reliably improves optimization on every control task; the paper asserts this empirically after selecting the best operator-set size from the same runs, rather than proving it or cross-validating it.","fun_headline_variants_meta":{"raw":{"variants":["Push-pull trick beats pull-only quantum control","Orthogonal targets boost quantum optimization","Push-pull quantum control speeds convergence","Shorter NMR pulse with push-pull method","Push-pull optimization sharpens quantum gates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000129,"raw_usage":{"total_tokens":1072,"prompt_tokens":844,"completion_tokens":228,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":163}},"tokens_in":460,"tokens_out":228,"duration_ms":2961,"temperature":1.0,"reasoning_tokens":163,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:50:19.899556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix $L$ and $\\alpha$ before the runs and average the final fidelity over many randomly drawn orthogonal-operator sets; if the mean no longer beats pull-only GRAPE and Krotov on the CNOT, singlet-transfer, and QFT tasks, the central claim fails. An 8-qubit QFT with $L=1$ would test whether the advantage survives further scaling.","supporting_citations":[{"cited_title":"Doria, T","cited_arxiv_id":null,"evidence_quote":"Supplies the GRAPE gradient method and the pull-only baseline that PP-GRAPE extends by subtracting averaged push gradients."},{"cited_title":"De Fouquieres, S","cited_arxiv_id":null,"evidence_quote":"Foundational Krotov variational-principle optimization that PP-Krotov modifies with extra co-sequences for orthogonal operators."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Krotov's global optimal control theory provides the variational background for the PP-Krotov update rules."},{"cited_title":"Krotov, Global methods in optimal control theory , Vol","cited_arxiv_id":null,"evidence_quote":"Describes the standard long-lived-state preparation sequence used as the experimental baseline; PP-Krotov's sequence is 30% shorter and yields 27% higher singlet order."}],"review_version":1}