{"id":"c1fd1baa-01d0-4a00-bcd8-088e5b375357","arxiv_id":"2412.15816","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A continuous-embedding gradient optimizer designs single-flux-quantum pulse sequences that achieve 99.9%+ simulated two-qubit gate fidelities in a tunable-coupler architecture.","lead":"Researchers simulated high-fidelity two-qubit gates for superconducting transmon circuits driven by single flux quantum pulses, reporting fidelities above 99.9%. The work could make cryogenic digital control of quantum processors more practical by reducing wiring and memory overhead.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No post-binarization fidelity check is reported in Secs. IV-V; the headline fidelities may be for the continuous relaxation, not for a physically realizable binary SFQ pulse train.","rationale":"The paper's core contribution is a gradient-based method for optimizing discrete SFQ pulses through a continuous embedding, so the bridge from the relaxed continuous parameters back to physical binary pulses is the critical assumption. The reader flagged this exact gap, and I agree it is the most load-bearing. The 5-level truncation, delta-function kick model, and leakage treatment are also unvalidated approximations, but those affect the accuracy of the simulation model; the binarization gap can invalidate the central physical claim even if the model is otherwise exact. The paper is plausible and interesting, and the requested post-discretization fidelity check is straightforward, so the CONDITIONAL verdict from the reader remains appropriate. This stress-test does not change that verdict.","tokens_in":9857,"tokens_out":12133,"duration_ms":113046,"concrete_test":"Re-run the simulation for the best reported cases (e.g., 40 GHz, T=80 ns for CZ/CNOT; hold time 17 ns for fSim) after rounding each optimized θ_sfq to the nearest bit (0 or 1) at every clock slot and each coupler on/off time to the nearest clock tick, using the identical Hamiltonian, logical basis, and average-fidelity definition; report the post-discretization fidelities together with the leakage population per logical basis state. If the infidelity increases by more than ~0.1 percentage point or the fidelity drops below the claimed 0.999/0.9999 thresholds, the headline results are not representative of an experimentally realizable SFQ pulse sequence and the abstract/conclusion claims need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. IV A (Eqs. (8)-(10)) relaxes the discrete SFQ amplitudes to continuous values θ_sfq in [0,1] and adds penalty terms to push them toward 0/1 and to align coupler times with the SFQ clock, but the paper never states that the optimized parameters are exactly binary after optimization, nor that the final sequence is re-evaluated after rounding. The penalty is soft: with γ starting at 1e-5 and increasing by 1.1 every 20 updates for 150 cycles, the final weight is large but finite, and the logarithmic barrier in Eq. (10) actually keeps the optimizer away from exact 0/1 during the run (it only decays, not vanishes). Unless a separate thresholding step and re-simulation are performed, the fidelities in Figs. 2, 3, and 5 and the abstract (99.99% fSim, >99.9% CZ/CNOT) may describe continuous drive amplitudes that an SFQ source cannot produce, since each SFQ pulse has a quantized area Φ0. Fig. 4 shows binary-looking kick patterns, but no thresholding rule or post-discretization fidelity numbers are given. This gap directly undermines the physical realizability of the reported two-qubit gate fidelities.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents two approaches for realizing two-qubit gates (fSim, CZ, CNOT) in a tunable-coupler transmon architecture controlled by single-flux-quantum (SFQ) pulses. The first approach relaxes the discrete SFQ pulse amplitudes to continuous variables and uses gradient-based optimization with penalties to drive them toward binary values. The second approach uses an exact decomposition of CZ/CNOT into two fSim gates and single-qubit rotations, using a previously reported SFQ-based single-qubit control method. The authors report simulated average gate fidelities of about 99.99% for fSim and above 99.9% for CZ and CNOT, and show that the decomposition method requires significantly less memory to store the pulse sequences.","tokens_in":10096,"tokens_out":10849,"duration_ms":90820,"significance":"If the reported fidelities are achieved for physically realizable binary pulse sequences, the paper demonstrates a scalable, microwave-free route to high-fidelity two-qubit gates, which would be an important step for SFQ-based quantum control. The use of auto-differentiation for continuous embedding of discrete pulses is a practical technique, and the analytical decomposition provides a compact, memory-efficient representation. However, the manuscript does not presently demonstrate that the optimized continuous pulse parameters are converted to binary SFQ sequences without significant fidelity loss, which is essential for the physical relevance of the headline numbers. The paper is therefore interesting but requires a key verification before the claims can be accepted.","major_comments":[{"comment":"The optimization relaxes discrete SFQ amplitudes to continuous values in [0,1] and adds the penalty P and the logarithmic barrier Φ. Because the barrier in Eq. (10) keeps the amplitudes strictly in the interior of the interval during the run, and the penalty weight γ is finite at the end of the schedule, the optimized parameters are not guaranteed to be exactly binary. The manuscript never states that the final parameters are thresholded to 0 or 1, nor that the resulting binary pulse train is re-simulated. Since each SFQ pulse has a quantized area, the fidelities reported for gradient-optimized CZ and CNOT gates in Fig. 5 and in the abstract likely describe continuous drive amplitudes that an SFQ source cannot produce. Please provide the fidelity of the thresholded/rounded binary sequence, or explicitly demonstrate that the optimizer converges to exact binary values within a specified tolerance.","section":"Sec. IV A, Eqs. (8)-(10); Sec. V, Figs. 4-5"},{"comment":"The memory comparison between the optimized and decomposition-based sequences (3200 vs. less than 200 bits per qubit) is presented without the encoding details. The optimized sequence is a raw per-slot binary pulse train, while the decomposition exploits the encoding of Ref. [21] and the repeated structure of the sequence. Please specify the encoding rules and the exact bit-count calculation so that the claimed memory reduction can be reproduced.","section":"Sec. V, final paragraph"}],"minor_comments":[{"comment":"The phrase 'average a gate fidelity' should be 'average gate fidelity'.","section":"Abstract"},{"comment":"Ref. [21] is a preprint by overlapping authors; the text should note that it is not yet peer-reviewed, if that is the case.","section":"Sec. I"},{"comment":"The schedule 'after every sequence of 20 parameter updates' is ambiguous; please specify whether this is 20 gradient steps or 20 L-BFGS-B iterations.","section":"Sec. IV A"},{"comment":"The y-axis of Fig. 5 is logarithmic with a range that makes exact infidelities for the best points hard to read; consider adding numerical values in the text.","section":"Sec. V, Fig. 5"},{"comment":"The sentence 'The four kick plots in the top and middle panels correspond to the four possible SFQ-clock slots during a qubit period' could be clarified with a concrete example of how the slots are indexed.","section":"Sec. V, Fig. 4 caption"},{"comment":"The truncation to n = ±50 and five levels is stated, but there is no convergence check with respect to the truncation; a brief statement that the results are converged would be helpful.","section":"Sec. III A"}],"recommendation":"major_revision","confidential_remarks":"The main technical concern is the absence of a post-discretization fidelity check for the continuous relaxation. If the authors can provide the fidelity of the rounded binary sequences and it remains above the stated thresholds, the paper would be suitable for publication. The analytical decomposition and the memory-efficient encoding are interesting, but the memory claim needs more detail. Also, Ref. [21] is a closely related preprint by overlapping authors; the editorial office should ensure that its status is clear. The manuscript would benefit from a data/code availability statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful contribution here is the optimization method: relaxing discrete SFQ pulse presences to continuous variables and using JAX autodiff with L-BFGS-B to hit 99.99% fSim and >99.9% CZ/CNOT in simulation. That is a real improvement over the genetic-algorithm approach in Refs. [9,12], and the analytical fSim-pair decomposition with memory reduction is a nice practical add-on.\n\nThe Hamiltonian model and simulation basis are standard and carefully built: charge basis, diagonalization, five-level truncation, Löwdin orthogonalization for the degenerate logical states, Suzuki-Trotter. The calibration of idling fluxes is sensible. The paper is honest that this is simulation, and it does not oversell experimental readiness.\n\nThe main soft spot is the one the stress-test flags: Sec. IV A relaxes the binary pulse presence to continuous values and adds penalties, but there is no statement that the final optimized parameters are exactly 0/1, and no re-simulation after rounding. The penalty schedule is soft—gamma ends around 12, mu decays to roughly 1e-6—so nothing forces exact binary values. Fig. 4 shows binary-looking kicks, but that is a plot, not a check. The reported fidelities may be for continuous pulses, which an SFQ source cannot produce. This is not a fatal flaw; the fix is straightforward: round, re-simulate, report. But as written, the headline number for a physical pulse train is unverified. A second, smaller gap: no code or data, so the optimization details are not reproducible from the text alone. The 5-level truncation and delta-function kick model are standard in the subfield, so I would not call them flaws, though an error budget would strengthen the paper.\n\nThe citation pattern is fine; the self-citation to Ref. [21] is to the single-qubit sequences they actually use.\n\nThis paper is for people working on SFQ control and cryogenic digital control architectures. It deserves a serious referee: the method is novel and the results are plausible, but the authors need to close the discretization gap and, ideally, release code. Send it to review, with a request for a post-rounding fidelity report and code.","headline":"Gradient-based continuous embedding for SFQ two-qubit gates is a genuine step forward, but the missing post-discretization fidelity check leaves the headline numbers unverified against physically realizable pulses.","tokens_in":10651,"tokens_out":2264,"would_cite":true,"duration_ms":19126,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.Lx","85.25.-j"],"model":"deepseek-v4-flash","headline":"A continuous-embedding optimizer turns discrete single-flux-quantum pulses into two-qubit gates with simulated fidelities near 99.99 percent.","keywords":["single-flux-quantum control","tunable-coupler architecture","transmon qubits","two-qubit gates","gradient-based optimal control","continuous embedding","fSim gate","CZ/CNOT gates"],"falsifier":"Take one optimized CZ sequence from Section V, round every relaxed pulse amplitude to the nearest binary value (0 or 1), keep the coupler times on the SFQ clock, and re-simulate the resulting pulse train under the same Hamiltonian; if the average fidelity drops below the reported 0.999 (or below 0.9999 for the fSim gate), the claim that these discrete pulse trains realize the reported fidelities is not yet established.","tokens_in":9667,"feed_emoji":"⚛️","tokens_out":15121,"duration_ms":119426,"temperature":0.7,"pith_summary":"The paper proposes a way to program two-qubit gates on a chip with two transmon qubits (a common type of superconducting qubit) coupled through a tunable coupler, using single-flux-quantum (SFQ) pulses, short digital voltage spikes each carrying one flux quantum, instead of microwave control. The central claim is that the discrete, bit-like nature of these pulses can be handled by a continuous embedding: relax each pulse to a continuous amplitude, let gradient descent optimize the gate fidelity, and add penalties that pull the amplitudes back to physical 0-or-1 values. In simulation this yields fSim gates (an entangling $XX+YY$ rotation with a $ZZ$ phase) with average fidelity around 99.99% and CZ and CNOT gates above 99.9%, at durations under 80 ns. The paper also gives a semi-analytical construction of CZ and CNOT from two fSim layers plus single-qubit rotations, cutting the memory needed to store pulse schedules from thousands of bits per qubit to under two hundred. The reported fidelities are simulated, not measured, and the payoff would be a digital, cryogenic-friendly control route to scalable transmon processors.","feed_headline":"Simulated SFQ pulses hit 99.99 percent for a two-qubit gate","feed_subtitle":"A continuous-embedding optimizer designs digital SFQ pulse trains for CZ, CNOT, and fSim gates at the 99.9 percent level.","key_machinery":"Two mechanisms carry the argument. The first is the continuous embedding of the discrete SFQ control: each pulse amplitude $\\theta^i_{\\rm sfq}\\in[0,1]$ is a relaxed presence/absence bit, and the coupler on/off times are real-valued. The cost function (8) adds a binarizing penalty $P$ and a log-barrier smoothing term $\\Phi$, with hyperparameters annealed on a schedule, so gradient descent can navigate the smooth landscape while the final solution is pushed toward physical 0/1 pulses and clock-aligned times. The second is the fSim-pair decomposition identity for CZ, hence CNOT: with $\\Gamma(\\theta,\\phi)=e^{-i\\theta(XX+YY)/2}e^{-i\\phi ZZ/4}$, the circuit $R_x(\\xi)\\Gamma(\\theta,\\phi)R_x(2\\alpha)\\Gamma(-\\theta,\\phi)R_x(\\xi)R_x(\\eta)R_x(-\\eta)$ realizes CZ up to equal single-qubit $Z$ rotations whenever Eq. (14) holds, giving a continuous family of decompositions parameterized by the fSim angles; the authors pick the member near a $\\sqrt{i\\mathrm{SWAP}}$ gate (hold time 17 ns, total duration 23.4 ns) as the shortest practical native gate. The simulation itself uses a five-level-per-transmon charge basis, symmetric orthogonalization of the degenerate logical states at the idling point, and fourth-order product-formula time evolution, all written with auto-differentiation so gradients of the roughly 6400 parameters cost about two forward passes.","core_discovery":"On its own terms, the paper's discovery is that the binary combinatorial problem of choosing where to place SFQ kicks, and when to move the coupler, can be solved as a smooth optimization problem. The variational parameters are, for every clock tick and every qubit, a binary amplitude indicating the presence or absence of an SFQ pulse, together with the start and end times of the coupler excursions. Relaxing the amplitudes to the interval $[0,1]$ and the coupler times to real values turns the cost $C(\\vec\\theta)=1-F(\\vec\\theta)+P(\\vec\\theta)+\\Phi(\\vec\\theta)$ into a differentiable function, with $P$ penalizing non-binary amplitudes and misaligned clock times and $\\Phi$ a logarithmic barrier that keeps the search away from boundaries. Minimizing this cost by second-order gradient descent reproduces, in simulation, fSim gates with average fidelity on the order of 0.9999 and CZ and CNOT gates above 0.999, at SFQ clock frequencies of 20 and 40 GHz and durations of 70-80 ns. The semi-analytical alternative exploits the identity that a CZ gate can be decomposed into two fSim gates, with parameters satisfying Eq. (14), plus single-qubit rotations; choosing an fSim hold time near 17 ns that corresponds to an $XX+YY$ angle close to $\\pi/4$ reaches fidelities close to 0.999 while storing each qubit's pulse sequence in under 200 bits.","pith_inferences":["The paper reports simulation-only fidelities; a direct test is to round the optimized continuous amplitudes to binary kicks and re-simulate the resulting pulse train, since the paper does not state whether the discretized sequences were re-evaluated.","The continuous-embedding recipe is generic enough that the same penalty-and-barrier schedule could be applied to other discrete-control problems, such as flux-latching gates or DAC-step control, where combinatorial searches are currently the default.","The decomposition family in Eq. (14) leaves freedom in choosing the fSim angles; that freedom could be searched over to find implementations that are more robust to parameter drift or that minimize leakage, rather than only the shortest one selected here.","If the discretized schedules hold up, the next likely bottleneck is error accumulation across many gates and idling intervals, since the idling point still carries residual conditional phases not characterized in the paper."],"forward_implications":["SFQ control becomes a credible digital alternative to microwave control for two-qubit gates in tunable-coupler transmon processors, with simulated fidelities comparable to microwave-based gates.","Because the gate set is generated by streams of identical flux-quantum pulses rather than shaped microwave waveforms, the controller can be moved close to the cryogenic chip, easing cabling and heat-load constraints.","The analytical fSim-pair decomposition reduces the memory required to store a CZ or CNOT schedule from 3200 bits per qubit to under 200 bits, simplifying the classical-to-quantum interface.","Higher SFQ clock frequencies and longer gate durations both lower infidelity, so faster future SFQ controllers can trade speed against gate quality.","With both qubits idling at the same target frequency, single-qubit gates are obtained without qubit-specific optimization, which simplifies scaling to more qubits."],"supporting_citations":[{"why":"Supplies the tunable-coupler circuit and its Hamiltonian, which define the system simulated throughout the paper.","marker":"[15]"},{"why":"Establishes the model of SFQ pulses as short voltage spikes whose area is one flux quantum, the control drive used in the optimizations.","marker":"[6]"},{"why":"Provides the earlier genetic-algorithm approach to SFQ control that the gradient-based method is intended to improve on.","marker":"[9]"},{"why":"Reports earlier SFQ-based two-qubit gate construction, motivating the search for more memory-efficient and structured pulse sequences.","marker":"[12]"},{"why":"Gives the decomposition of a CZ gate into two fSim layers and single-qubit gates that underlies the analytical construction.","marker":"[16]"},{"why":"Supplies the single-qubit SFQ pulse sequences and the compact binary encoding used to estimate the memory savings.","marker":"[21]"},{"why":"Provides the penalty and smoothing treatment for binary variables that the continuous-embedding cost function adopts.","marker":"[25]"},{"why":"Supplies the symmetric orthogonalization procedure used to define the logical basis for the degenerate qubit states.","marker":"[24]"},{"why":"Implements the auto-differentiation that makes gradient evaluation across the roughly 6400 pulse parameters computationally feasible.","marker":"[26]"}],"fun_headline_variants":["Gradient descent tunes SFQ pulses to 99.99% fidelity","Digital SFQ control hits 99.9% gate fidelity","Binary SFQ pulses made differentiable for high-fidelity gates","Smooth embedding optimizes SFQ two-qubit pulses","Smooth optimization designs digital SFQ pulse trains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper does not state whether the fidelities in Section V are computed before or after the relaxed continuous pulse parameters of Section IV A are discretized to binary SFQ kicks; if the discrete pulse train is not re-simulated, the headline numbers may describe pulses a real controller cannot produce.","fun_headline_variants_meta":{"raw":{"variants":["Gradient descent tunes SFQ pulses to 99.99% fidelity","Digital SFQ control hits 99.9% gate fidelity","Binary SFQ pulses made differentiable for high-fidelity gates","Smooth embedding optimizes SFQ two-qubit pulses","Smooth optimization designs digital SFQ pulse trains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000857,"raw_usage":{"total_tokens":3758,"prompt_tokens":1017,"completion_tokens":2741,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":2658}},"tokens_in":633,"tokens_out":2741,"duration_ms":17218,"temperature":1.0,"reasoning_tokens":2658,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:03:59.476477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one optimized CZ sequence from Section V, round every relaxed pulse amplitude to the nearest binary value (0 or 1), keep the coupler times on the SFQ clock, and re-simulate the resulting pulse train under the same Hamiltonian; if the average fidelity drops below the reported 0.999 (or below 0.9999 for the fSim gate), the claim that these discrete pulse trains realize the reported fidelities is not yet established.","supporting_citations":[{"cited_title":"McDermott and M","cited_arxiv_id":null,"evidence_quote":"Establishes the model of SFQ pulses as short voltage spikes whose area is one flux quantum, the control drive used in the optimizations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the earlier genetic-algorithm approach to SFQ control that the gradient-based method is intended to improve on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reports earlier SFQ-based two-qubit gate construction, motivating the search for more memory-efficient and structured pulse sequences."},{"cited_title":"Arute, K","cited_arxiv_id":null,"evidence_quote":"Gives the decomposition of a CZ gate into two fSim layers and single-qubit gates that underlies the analytical construction."},{"cited_title":"Compact Pulse Schedules for High-Fidelity Single-Flux Quantum Qubit Control","cited_arxiv_id":"2309.04606","evidence_quote":"Supplies the single-qubit SFQ pulse sequences and the compact binary encoding used to estimate the memory savings."},{"cited_title":"Murray and K.-M","cited_arxiv_id":null,"evidence_quote":"Provides the penalty and smoothing treatment for binary variables that the continuous-embedding cost function adopts."},{"cited_title":"Mayer, On L¨ owdin’s method of symmetric orthogonal- ization, Int","cited_arxiv_id":null,"evidence_quote":"Supplies the symmetric orthogonalization procedure used to define the logical basis for the degenerate qubit states."},{"cited_title":"Bradbury, R","cited_arxiv_id":null,"evidence_quote":"Implements the auto-differentiation that makes gradient evaluation across the roughly 6400 pulse parameters computationally feasible."}],"review_version":1}