{"id":"5507c9b3-1a7d-4055-b1ed-b7beccc34763","arxiv_id":"2411.19785","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"The paper reports neural-network-optimized, angle-continuous pulse families for native C1P and C2P Rydberg phase gates, with simulated infidelities of 3.4e-4 and 1.45e-3.","lead":"Researchers trained neural networks to find laser pulse families that implement two- and three-qubit phase gates on neutral atom hardware. The resulting gates are fast, smooth, use global addressing, and reach simulated infidelities near 1e-3 to 1e-4.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (15) prints the Rydberg decay term with a +i sign, which under the evolution convention of Eq. (16) produces gain, not decay; the reported infidelities therefore cannot be reproduced as written.","rationale":"The reader identified the uniform-equidistant interaction assumption of Eq. (12) as the weakest point, which is a real robustness concern but is standard for a first theoretical pulse-family proposal and does not invalidate the simulated central claim. A more fundamental internal issue is the sign of the decay term in Eq. (15). Because the paper's fidelity numbers are obtained by simulating the full Hamiltonian with decay, the sign of Hdecay directly determines whether those numbers describe loss. Written as +i Γ/2, the term is anti-Hermitian in the wrong direction: under Eq. (16), it amplifies Rydberg amplitude. This makes the reported quantitative claim unreproducible from the manuscript alone. The most plausible resolution is that the code uses the conventional -i sign and Eq. (15) contains a typo; in that case the underlying optimization results could still be valid. The appropriate response is to keep the reader's CONDITIONAL verdict and add an explicit condition: verify the decay sign in the code or by recomputation from the released NN weights, and correct Eq. (15) if needed. The geometric-uniformity issue remains a secondary, separate limitation that should be stated clearly in the paper but is not the central blocker.","tokens_in":16784,"tokens_out":8895,"duration_ms":86290,"concrete_test":"Locate the implementation of Hdecay in the training or evaluation code; if code is not available, recompute the mean infidelities from the released NN weights under the two candidate Hamiltonians H1q ± i Γ/2 |r><r| using Eq. (16). Check that the conventional '-' sign reproduces the reported values 3.4e-4 (C1P) and 1.45e-3 (C2P), while the '+' sign does not. If the '-' sign reproduces the numbers, the paper needs only a sign correction; if the '+' sign reproduces them, the central claim is not supported as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The numerical headline—mean infidelities of 3.4e-4 for C1P and 1.45e-3 for C2P, both claimed to include Rydberg decay—rests on the non-Hermitian term in Eq. (15), Hdecay = +i Γ/2 |r><r|, added to H1q and evolved via Eq. (16), U = T exp(-i∫H dt). With this sign, a Rydberg amplitude acquires e^{+Γt/2} and the Rydberg population grows instead of decaying. Physical decay requires Hdecay = -i Γ/2 |r><r|. If the training or evaluation code follows the printed sign, the quoted 'decay-included' fidelities are not decay losses; if the code uses the conventional '-' sign, Eq. (15) is a typo. Either way, the manuscript is internally inconsistent at exactly the point that supports the central quantitative claim. The reader's concern about the uniform geometry assumption of Eq. (12) is a legitimate robustness limitation, but the sign inconsistency is more load-bearing because it attacks the simulated fidelity numbers themselves, not merely their transfer to real hardware. The definition in Eq. (B3), (1-F)_r = F - Fdecay, only has the intended positive sign if Fdecay < F, which the '+' sign does not guarantee.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses the design of parameterized multi-qubit gates for neutral-atom platforms. The authors train chained neural networks to output detuning pulses for controlled-phase gates C1P and C2P as continuous (piecewise-continuous) functions of the phase angle φ, with fixed Rabi frequency and global addressing. They simulate the dynamics under the full Hamiltonian with finite Rydberg blockade (B = 21.1) and include a non-Hermitian term intended to model Rydberg decay, reporting mean infidelities of 3.4 × 10^-4 for C1P and 1.45 × 10^-3 for C2P. They also compare pulse durations with known time-optimal results and provide the trained network weights as data. The central claim is that these are the first families of CkP pulses for Rydberg platforms, with fidelities suitable for current hardware.","tokens_in":17129,"tokens_out":7227,"duration_ms":59943,"significance":"If the numbers are correct, the contribution is a practical recipe for native parameterized multi-qubit gates, which is relevant for variational algorithms and circuit compilation. The numerical protocol is careful: C1P uses full-Hamiltonian simulation with finite blockade, C2P uses a two-stage refinement from infinite-blockade to finite-blockade optimization, and the reported fidelities are obtained by forward integration of the stated Hamiltonian. The release of trained network weights is a concrete reproducibility asset. However, the quantitative claims are undermined by a sign error in the decay Hamiltonian, and the reported fidelities depend on an idealized symmetric geometry. With the sign corrected and the numbers re-evaluated, the work would be a solid contribution.","major_comments":[{"comment":"The non-Hermitian term is written as H_decay = +i Γ/2 |r><r|. With the evolution convention of Eq. (16), U = T exp(-i ∫ H dt), the imaginary term produces e^{+Γt/2} growth in the Rydberg amplitude, i.e., gain rather than decay. Consequently, the quoted mean infidelities 'including Rydberg decay' (3.4e-4 for C1P and 1.45e-3 for C2P) cannot be reproduced from the printed equations. If the simulation code uses the conventional -i Γ/2, Eq. (15) is a typo; if it uses the printed sign, the fidelity numbers do not represent decay losses. Either way, the manuscript must be corrected and the reported fidelities recomputed or re-confirmed. The definition in Eq. (B3), (1-F)_r = F - F_decay, also assumes F_decay < F, which is not guaranteed with the '+' sign.","section":"Eq. (15), Sec. II A"},{"comment":"The text states 'The C2P gate is described by the matrix U = 13 - 2e^{iφ}|111><111| – equivalent to the definition in Eq. (5).' This is not equivalent: this matrix gives eigenvalue 1 - 2e^{iφ} on |111>, whereas Eq. (5) requires e^{iφ}. The correct expression is 1_3 + (e^{iφ} - 1)|111><111|. While the actual numerical optimization uses Eq. (5), the printed statement is a mathematical error that should be fixed.","section":"Sec. II A, after Eq. (14)"},{"comment":"The high-fidelity numbers are obtained under the assumption that all atom pairs have identical interaction strength V and that a single global detuning pulse is applied. This idealized symmetric geometry is not tested against variations in interatomic distances, laser intensity inhomogeneity, or finite temperature. Since the paper concludes that the gates have 'immediate benefits for current neutral atom hardware,' the authors should either provide robustness analysis under realistic inhomogeneities or explicitly state that the reported fidelities apply only to the ideal equidistant, globally addressed configuration.","section":"Eqs. (8)-(12) and Sec. III"}],"minor_comments":[{"comment":"V as defined via -C6/|Ri-Rj|^6 is negative for positive C6, so B = V/Ωmax would be negative. Please define B = |V|/Ωmax or clarify the sign convention.","section":"Eq. (12), Sec. II A"},{"comment":"The reference for the additional data [56] lacks a URL or DOI; for reproducibility, the data should be accessible via a persistent link.","section":"Data availability"},{"comment":"The claim that the C1P pulse is 'time-optimal' is not proven. The agreement of Tφ with a known arcsinh fit is suggestive, but a proof or a precise comparison to a time-optimal control solution would be needed.","section":"Introduction and Fig. 4"},{"comment":"The journal name 'A VS Quantum Sci.' should be 'AVS Quantum Sci.'.","section":"Reference [10]"},{"comment":"Because the evolution is non-Hermitian when decay is included, the projected operator P U_out P is not unitary; the use of the Hilbert-Schmidt distance as a fidelity measure should be justified or the normalization made explicit.","section":"Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The manuscript addresses a timely problem and the training methodology is sound. The main concern is the sign of the decay term in Eq. (15), which, if taken literally, invalidates the headline fidelity numbers. I recommend major revision with a request to correct the sign and re-evaluate all decay-inclusive fidelities. The geometry assumption is a common idealization in this field, but the authors should state its scope explicitly. The C2P unitary expression after Eq. (14) is also erroneous and must be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine step forward for Rydberg pulse families: it adapts the chained neural-network method of Sauvage and Mintert to C1P and C2P gates, uses piecewise domain training sensibly, and the two-stage finite-blockade refinement for C2P is a thoughtful fix for a real failure mode. The C1P time curve reproduces known time-optimal results, and the fidelities come from forward simulation against the full Hamiltonian, not from circular fitting. If the decay sign is corrected, the method is credible and likely useful for neutral-atom NISQ work.\n\nThe soft spot is load-bearing. Equation (15) prints Hdecay = +iΓ/2 |r⟩⟨r|, and evolution uses U = T exp(−i∫H dt). That combination gives Rydberg population growth, not decay. Either the code follows the printed plus sign, in which case the reported “decay-included” fidelities are not decay losses at all, or the code uses the conventional minus sign, in which case Eq. (15) is a typo at exactly the point that supports the central quantitative claim. The paper must state which sign the code actually uses and correct the equation before any reviewer can trust the numbers. This is not a minor formatting issue; it is an internal contradiction with the paper’s own equations.\n\nThe uniform-geometry assumption of Eq. (12) is a legitimate robustness limitation, but secondary: real arrays have intensity inhomogeneity and unequal distances, and the paper does not simulate those. The reader’s concern about that is fair, but the sign error is more urgent. Also, the training code is only available on request, and the published artifact is just the weights, so full reproducibility is not yet there.\n\nWho gets value from this paper: quantum control researchers and neutral-atom experimentalists interested in parametrized multiqubit gates. It deserves a serious referee, but only conditionally: the authors must correct Eq. (15), confirm which sign the simulator uses, and ideally release the training code. I would not cite this version until that is resolved. My recommendation is to send it to peer review with a clear request to fix the sign and re-verify the headline fidelities.","headline":"Solid pulse-family optimization for Rydberg controlled-phase gates, but the printed decay Hamiltonian has a sign error that turns decay into gain; the headline fidelities are unreproducible as written.","tokens_in":17662,"tokens_out":2706,"would_cite":false,"duration_ms":25395,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-network optimizer finds piecewise-continuous detuning pulses that implement $\\mathrm{U_{C_1P}}(\\phi)$ and $\\mathrm{U_{C_2P}}(\\phi)$ on neutral atoms at simulated mean infidelities of $3.4\\times10^{-4}$ and $1.45\\times10^{-3}$, with…","keywords":["parametrized quantum gates","Rydberg blockade","neutral atom quantum computing","neural network control","quantum optimal control","C1P gate","C2P gate","pulse families"],"falsifier":"Run the same trained pulses in a simulation or experiment with atoms placed at unequal separations, giving pairwise interaction strengths differing by about 10%, or with a 2% per-site Rabi-frequency spread, and compute the mean gate fidelity over $\\phi\\in(0,\\pi]$. If the mean infidelity rises above roughly $10^{-2}$, the central claim of hardware-applicable pulse families is falsified for real arrays.","tokens_in":16578,"feed_emoji":"⚛️","tokens_out":7582,"duration_ms":59047,"temperature":0.7,"pith_summary":"The paper aims to show that a single trained neural network can output, for any requested phase angle $\\phi$ in $(0,\\pi]$, a laser-detuning pulse that implements a controlled-phase gate on neutral atoms, first with one control qubit ($\\mathrm{C_1P}$) and then with two ($\\mathrm{C_2P}$). Because the pulse family is produced in one training pass rather than re-optimized per angle, it could remove the per-iteration overhead that gradient-ascent control would impose in variational and compiled circuits. For $\\mathrm{C_1P}$ the simulated mean infidelity is $3.4\\times10^{-4}$, and for $\\mathrm{C_2P}$, after re-optimization from an infinite-blockade starting point, $1.45\\times10^{-3}$ including Rydberg decay and finite blockade. The pulses use a fixed Rabi frequency, global addressing, smooth detuning, and durations near or at the time-optimal values, which is what makes them plausible on current hardware.","feed_headline":"Trained networks deliver smooth phase-gate pulses for every angle","feed_subtitle":"Simulated mean infidelities of 3.4e-4 for C1P and 1.45e-3 for C2P, using one global laser.","key_machinery":"The argument rests on the blockade-regime decomposition of the $N$-qubit Hilbert space into decoupled two-level systems. For $\\mathrm{C_1P}$ the nontrivial transition is $|11\\rangle$ to the bright state $|b_2\\rangle=(|1r\\rangle+|r1\\rangle)/\\sqrt{2}$ with Rabi frequency $\\sqrt{2}\\Omega_{\\max}$; $\\mathrm{C_2P}$ adds $|111\\rangle$ to $|b_3\\rangle=(|11r\\rangle+|1r1\\rangle+|r11\\rangle)/\\sqrt{3}$ at $\\sqrt{3}\\Omega_{\\max}$. The optimization uses these effective dynamics, and optionally the full finite-blockade Hamiltonian, and treats the Schr\\\"odinger equation as a neural ODE so gradients can be backpropagated through the solver into the chained networks $N_T(\\phi)$ and $N_C(\\phi,t)$; the trained networks output $T_\\phi$ and $\\Delta(\\phi,t)$ directly for any $\\phi$.","core_discovery":"The central claim is that parametrized multiqubit phase gates $\\mathrm{U_{C_1P}}(\\phi)$ and $\\mathrm{U_{C_2P}}(\\phi)$ are natively realizable on Rydberg-neutral-atom hardware by piecewise-continuous families of detuning pulses $\\Delta(\\phi,t)$, with the Rabi frequency held fixed at $\\Omega_{\\max}$ and the same pulse applied to all atoms. Optimizing two chained neural networks, one that outputs pulse duration $T_\\phi$ and one that outputs the detuning waveform, against the gate-infidelity cost $J$ converges to controls whose simulated mean infidelities, including Rydberg-state decay and finite blockade at $B=21.1$, are $3.4\\times10^{-4}$ for $\\mathrm{C_1P}$ and $1.45\\times10^{-3}$ for $\\mathrm{C_2P}$. The $\\mathrm{C_1P}$ pulse is time-optimal, and the $\\mathrm{C_2P}$ pulse at $\\phi=\\pi$ is within 2.6% of the known time-optimal $\\mathrm{C_2Z}$ duration. The paper states that this is the first $\\mathrm{C_kP}$ pulse family on the Rydberg platform.","pith_inferences":["Beyond the paper: the same chained-neural-network ansatz should produce pulse families for other parametrized unitaries whose dynamics decompose into a few bright-state manifolds, such as fan-out gates, as long as the effective Hamiltonian stays low-dimensional.","Beyond the paper: because only detuning is modulated, the protocol's robustness to laser-intensity inhomogeneity may be the next bottleneck; a testable extension is to include per-atom Rabi variations in the simulation and quantify how infidelity grows.","Beyond the paper: in variational algorithms one could query the trained network on the fly for the current $\\phi$, eliminating the repeated classical re-optimization step entirely, which is the practical payoff the authors point toward.","Beyond the paper: hardware validation on arrays with non-equidistant spacing is the natural next step, since the symmetric-geometry assumption is the main place the reported infidelities could degrade."],"forward_implications":["A $\\mathrm{C_1P}$ gate can be run with a single global laser, no single-site addressing, and a duration that tracks the time-optimal curve $f(\\phi)=a\\,\\mathrm{arcsinh}(b\\phi)$.","The $\\mathrm{C_2P}$ gate runs at about $16.87/\\Omega_{\\max}$ for $\\phi=\\pi$, only roughly 2.6% above the time-optimal $\\mathrm{C_2Z}$ gate, while still using only global controls.","Including these native gates in compiled circuits replaces decompositions that are about 2.2 times ($\\mathrm{C_1P}$) and 4.6 times ($\\mathrm{C_2P}$) longer in two-qubit-gate time, reducing error accumulation on near-term devices.","The trained network weights serve as initial guesses for further optimization, for instance toward robustness to laser-intensity noise.","The method is in principle extendable to more than two control qubits, limited mainly by Hilbert-space dimension and training difficulty."],"supporting_citations":[{"why":"Defines the variational family-control method, chained networks outputting pulses for all angles, that this paper adapts to neutral atoms.","marker":"[23]"},{"why":"Provides the time-optimal CkZ pulse durations and bright-state analysis that set the time bounds and comparison baselines.","marker":"[13]"},{"why":"Derives the two-level bright-state decomposition and time-optimal C1P pulse law that the C1P results reproduce.","marker":"[34]"},{"why":"Demonstrates high-fidelity Rydberg entangling gates experimentally and supplies the experimental context for the pulse families.","marker":"[15]"},{"why":"Gives the strontium-88 finite blockade strength $B=21.1$ and the error-budget framework used in the simulations.","marker":"[40]"},{"why":"Supplies the Rydberg-state lifetime and decay modeling used for the reported infidelity numbers.","marker":"[41]"},{"why":"Introduces neural ODEs, the mechanism by which gradients are backpropagated through the Schr\\\"odinger solver.","marker":"[51]"},{"why":"Motivates why CkP gates reduce circuit depth and compilation overhead on neutral-atom hardware.","marker":"[12]"}],"fun_headline_variants":["Neural nets design smooth pulses for native multiqubit gates","AI-crafted pulses enable low-error C1P and C2P gates on neutral atoms","Smooth global pulses: neural networks find efficient Rydberg gates","One global laser trains smooth pulses for multiqubit gates","Optimized pulses for multiqubit phase gates on neutral atoms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported fidelities assume every atom pair has the same interaction strength and every atom sees the same laser intensity and detuning; real arrays have unequal spacings and intensity inhomogeneities that are not included in the simulation.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets design smooth pulses for native multiqubit gates","AI-crafted pulses enable low-error C1P and C2P gates on neutral atoms","Smooth global pulses: neural networks find efficient Rydberg gates","One global laser trains smooth pulses for multiqubit gates","Optimized pulses for multiqubit phase gates on neutral atoms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001048,"raw_usage":{"total_tokens":4453,"prompt_tokens":1045,"completion_tokens":3408,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":3316}},"tokens_in":661,"tokens_out":3408,"duration_ms":20329,"temperature":1.0,"reasoning_tokens":3316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:48:39.074626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same trained pulses in a simulation or experiment with atoms placed at unequal separations, giving pairwise interaction strengths differing by about 10%, or with a 2% per-site Rabi-frequency spread, and compute the mean gate fidelity over $\\phi\\in(0,\\pi]$. If the mean infidelity rises above roughly $10^{-2}$, the central claim of hardware-applicable pulse families is falsified for real arrays.","supporting_citations":[{"cited_title":"Bluvstein, S","cited_arxiv_id":null,"evidence_quote":"Provides the time-optimal CkZ pulse durations and bright-state analysis that set the time bounds and comparison baselines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates high-fidelity Rydberg entangling gates experimentally and supplies the experimental context for the pulse families."},{"cited_title":"Kalinowski, N","cited_arxiv_id":null,"evidence_quote":"Gives the strontium-88 finite blockade strength $B=21.1$ and the error-budget framework used in the simulations."},{"cited_title":"Dlaska, K","cited_arxiv_id":null,"evidence_quote":"Supplies the Rydberg-state lifetime and decay modeling used for the reported infidelity numbers."},{"cited_title":"Saffman, I","cited_arxiv_id":null,"evidence_quote":"Introduces neural ODEs, the mechanism by which gradients are backpropagated through the Schr\\\"odinger solver."}],"review_version":1}