Reinforcement learning entangling operations on spin qubits

Markus Schmitt; Mohammad Abedi

REVIEW 3 major objections 4 minor 46 references

Reinforcement learning entangling operations on spin qubits

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A reinforcement-learning agent can synthesize noise-tolerant CNOT gates for a semiconductor singlet-triplet qubit device under realistic experimental constraints, and can match a gradient-based optimizer without relying on Hamiltonian gradi

desk verdict A capable, honestly scoped numerical study of RL for CNOT synthesis on a singlet-triplet double dot, but the headline 'past error correction threshold' only holds under an experimentally impossible mid-protocol tomography assumption; in the realistic terminal-tomography setting the agent lands at NLIF ~1.8. read the letter →

arxiv 2508.14761 v1 pith:UULSW6IU submitted 2025-08-20 quant-ph

Mohammad Abedi , Markus Schmitt This is my paper

classification quant-ph

keywords reinforcementlearningquantumcontrolsinglet-tripletqubitsGaAsdoubledotCNOTgatenoise-robustpulsesynthesisprocesstomographysoftactor-critic

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that reinforcement learning can replace gradient-based optimal control for making entangling gates on a realistic semiconductor spin-qubit device. Concretely, a soft actor-critic agent is trained in a simulated gallium-arsenide singlet-triplet double dot to produce controlled-NOT (CNOT) pulses while the environment includes finite waveform rise time, hyperfine drift, slow and colored charge noise, and finite-shot process tomography. With exact unitary observation each step, the agent reaches fidelities at or above 0.999 and adjusts its pulses to different noise realizations; in the fixed-pulse noise-robust regime it reaches negative log-infidelities around 1.5-1.8, close to a gradient-based benchmark that in the comparison includes only the drift-noise contribution. If the simulator is a faithful stand-in for the device, the result matters because it shows practical control can be learned without an accurate Hamiltonian or gradients, with careful agent design - pulse-inclusive observations and noise-averaged rewards - carrying the performance.

What carries the argument

The load-bearing mechanism is the partially observable Markov decision process formulation of pulse synthesis, together with two design choices. Detuning pulses are piecewise constant but convolved with a measured impulse-response kernel, so each action has a delayed, smoothed effect; the observation includes the current shaped pulse amplitude along with the time-to-go and a unitary estimate, which restores the Markov property when the control delay approaches or exceeds one sample period. The reward is the negative log-infidelity f = -log10(1 - |Tr(U†_target U)/d|^2), either given sparsely at the final step or averaged over noise realizations, and optimization runs with the soft actor-criti

What would settle it

Run one of the learned robust CNOT protocols (for example the 20 ns, 30-action sequence) on a real GaAs singlet-triplet double dot with the same detuning calibration and measure its fidelity with around 10^5-shot process tomography. A measured fidelity well below the simulated 98% (negative log-infidelity about 1.8), or a failure of the adaptive policy to correct injected detuning fluctuations, would show that the simulator rather than the learning method produced the reported performance.

Watch

Extended reading notes

Core claim

The central claim is that a reinforcement-learning agent, using the soft actor-critic algorithm and a compact observation of (time-to-go, current pulse amplitude, current unitary estimate), can synthesize CNOT protocols for a four-dot GaAs singlet-triplet qubit system under modeled device constraints. The agent observes either the exact noisy unitary or, in the robust-pulse regime, only the pulse history, and receives a reward based on the negative log-infidelity to CNOT. With sparse final rewards and full unitary observations in a noisy environment, it finds protocols with fidelity at or above 0.999 for several (protocol length, action-count) combinations, and the same trained policy varies

Load-bearing premise

The simulated environment faithfully represents the physical device: the Hamiltonian, noise amplitudes, and impulse response are taken from earlier measurements rather than verified on hardware, so a divergence between simulation and device would invalidate the reported fidelities.

Editorial extensions

If this is right

In the adaptive regime, a single trained RL policy can act as a feedback controller: it changes its pulse sequence between noise realizations and keeps CNOT fidelity at or above 0.999, while a fixed pulse degrades.
In the robust regime, an RL agent trained on noise-averaged negative log-infidelity matches a gradient-based optimizer even when its observation is only the pulse history, showing that Hamiltonian-gradient information is not needed for benchmark-level performance.
Including the current pulse amplitude in the observation restores the Markov property under finite rise time and improves final negative log-infidelity by up to two orders of magnitude, a design rule that applies to any control system with delayed actuation.
Finite-shot process tomography limits achievable fidelity, but shot noise does not add an independent learning difficulty: performance converges to the exact-tomography result as the number of shots grows.
Higher-fidelity learned protocols also tend to use lower total pulse power, suggesting that the RL objective implicitly penalizes wasted energy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next test would be to take the noise-averaged robust pulse as the starting point for closed-loop calibration on the actual device; the paper names this as future work, and it would separate simulator-model error from policy error.
Because the robust-pulse agent learns from pulse history alone, a direct test would be to retrain the same architecture on other exchange-coupled spin-qubit platforms, such as silicon singlet-triplet or exchange-only qubits.
The observation-design result - adding the current shaped control value restores the Markov property under delayed actuation - is a general control-systems statement and could transfer to any bandwidth-limited reinforcement-learning task.
If the fidelity gap at high noise is really set by a fundamental noise bound rather than by RL optimization, then sweeping the drift-noise amplitude in the single-qubit test system should make RL and the gradient-based optimizer follow the same power law; the paper's Appendix D is consistent with this but does not prove it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a real piece of work, not a toy, and it deserves a serious referee. The authors train SAC agents to synthesize CNOT pulses on a GaAs singlet-triplet double-dot model with a measured impulse response and three noise sources, and they benchmark against qopt. They also ship code and data, which is exactly what lets you check the claims. The best parts are the design lessons: including the current pulse in the observation restores Markovianity under finite rise time, noise-averaged rewards give robust fixed pulses comparable to gradient-based optimization, and the shot-noise analysis in Sec. 4.4 is internally consistent with the tomography bound. Those are useful, transferable findings for anyone doing RL quantum control.

The soft spots are real but mostly secondary. The F ≥ 0.999 claim in Sec. 4.2 is obtained with observation ot = (T−t, εt, Ũt), i.e., the exact noisy unitary at every time step. That is not experimentally available: process tomography is destructive and requires many shots, and the protocol is only 20–50 ns. The paper's own more realistic Sec. 4.4, where the reward is based on terminal tomography with finite shots, saturates around NLIF ≈ 1.8, not 3.0. So the "past the error correction threshold" phrasing in the abstract and conclusion is not supported in any experimentally feasible setting considered in the paper. That should be fixed before publication, either by softening the claim or by showing a realistic feedback mechanism that can approach the ideal case.

Also worth flagging: the noise-averaged results report the maximum over training seeds, not the mean; the qopt comparison is honest but the cited 99.9% reference is not reproduced, with the discrepancy only partially explained; and a few figures lack error bars. None of these are load-bearing if the central argument is restated accurately.

The core positive claim—that RL can match gradient-based optimization in a realistic noisy simulator without Hamiltonian gradients—is supported within the simulator. The model is a simulator, not hardware, and the paper says so. I'd send this to peer review with a clear request to recalibrate the error-correction-threshold language and to report seed statistics rather than the max. The code and data make it checkable, and the design lessons are worth publishing even if the absolute fidelities don't transfer directly to a device.

Referee Report

3 major / 4 minor

Summary. The paper applies soft actor-critic reinforcement learning to synthesize CNOT gates in a simulated GaAs singlet-triplet double-dot device, using a Hamiltonian and noise model imported from Cerfontaine et al. It studies agent design choices: pulse-history vs unitary observations, finite rise-time Markovianity, adaptive feedback from exact noisy unitaries, noise-averaged robust pulse optimization, and simulated process tomography with shot noise. The robust-pulse results are competitive with a qopt gradient-based benchmark (NLIF≈1.5–1.8), while the adaptive setting with exact mid-protocol noisy-unitary observations reaches F≥0.999; adding tomographic measurement statistics degrades performance toward the robust limit. The paper concludes that RL is a viable Hamiltonian-gradient-free route for quantum control on this platform and provides code and data.

Significance. If the claims hold, the paper makes several useful contributions: (i) it demonstrates that including the current pulse amplitude in the observation restores Markovianity under finite rise-time distortion; (ii) it shows that noise-averaged RL can match a gradient-based optimal-control benchmark without Hamiltonian gradients; (iii) it identifies low-fluence pulse solutions as a desirable side effect; and (iv) it gives evidence that shot noise in process tomography does not confuse RL learning beyond the measurement-fidelity bound. The manuscript is unusually transparent: code and data are provided, training is multi-seed, and a qopt benchmark plus a noise-scaling cross-check in Appendix D are included. However, the headline F≥0.999 result is obtained under an observation model that is not experimentally implementable, and the reported robust-pulse comparison is based on a best-over-seeds statistic. These issues overstate the practical significance as currently written, but the underlying design-study is valuable.

major comments (3)

[Sec. 4.2 (observation model, Fig. 4/5)] The F≥0.999 adaptive result assumes the agent observes the exact noisy unitary Ũt at every control step. Mid-protocol exact process tomography is physically impossible: it is destructive, requires many repeated preparations, and takes far longer than the 20–50 ns protocol. The paper's own experimentally motivated setup in Sec. 4.4 and Fig. 7(c) reaches only NLIF≈1.8 in the large-measurement limit. The abstract/conclusion claim of 99.9% fidelity is therefore not supported in any experimentally feasible configuration in the manuscript. Please either remove or qualify the past-threshold claims, or explicitly model the cost of mid-protocol measurement and show the adaptive benefit survives.
[Sec. 4 (statistics; Fig. 6(a))] The text states 'we take the maximum over our training realisations' for the noise-averaged fidelities. If this means best over training seeds, the reported NLIF≈1.8 is an upper envelope and is not a statistically grounded estimate of achievable performance. The central claim that RL 'matched or surpassed' the qopt baseline therefore rests on a favorable seed. Please report the mean/median and spread over seeds, state the number of seeds, or justify why the maximum is the relevant design statistic.
[Sec. 4.3 (qopt comparison)] The qopt benchmark in Fig. 6(a) is described as 'only consisting of drift noise', while the RL agent is trained with all noise sources. The text also notes that slightly adjusted sample periods make qopt results comparable to Ref. [26], but those adjusted results are not shown. To substantiate the 'matching/surpassing' claim, provide a qopt run under the same noise model and optimized sampling period, or explicitly state the caveat so the comparison is on equal footing.

minor comments (4)

[Fig. 6(a) caption] The caption refers to 'a robust reward consisting of maximum negative log infidelities', but Eq. (9) defines a noise-averaged reward. Please clarify the distinction between averaging over noise realisations and taking maxima over training seeds.
[Sec. 4.4] The description of process tomography says a 'random input quantum state' is selected. Process tomography requires a fixed informationally complete set of input states and POVMs; 'random' is ambiguous and should be reworded to 'randomly selected from the tomography set'.
[Appendix C] The Gaussian-noise approximation maps a standard deviation to a number of measurements via a linear fit to the infidelity scaling. Please state the range of validity and the uncertainty of this fit, since Fig. 7(c) uses this mapping to infer large-measurement behavior.
[General] There are several typos: 'mentioned mentioned' (Sec. 4.3), 'aformentioned' (Appendix A), and 'required needed' (Conclusion).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RL control results are in-simulation optimization outcomes checked against an external qopt benchmark, with model parameters imported from independent experimental work and no fitted parameter renamed as a prediction.

full rationale

The paper's derivation chain is self-contained as an RL optimization study. The Hamiltonian, noise amplitudes, and impulse response are imported from external experimental work (Cerfontaine et al. [26]), and the benchmark is an independent implementation in the qopt library [40], not a result of the present authors. The central quantitative claims are in-simulator optimization results: the agent is trained to maximize a fidelity-based reward, and the reported NLIF values are evaluations of that same objective. This is a standard optimization evaluation, not a case where a fitted parameter is renamed as a prediction or where an output is equal to an input by construction. The paper explicitly separates regimes: Section 4.2's F >= 0.999 results use the idealized observation ot = (T-t, epsilon_t, U~t), while Section 4.4's experimentally realistic process-tomography setup converges to the noise-robust NLIF ~1.8 limit. The manuscript itself states that the realistic protocol is 'ultimately limited by the system noise' and that shot noise does not cause additional agent confusion, so the idealized threshold is not disguised as an experimentally feasible claim. The only self-citation is a background citation to Schmitt's earlier transmon RL work [17], which is not load-bearing. The lack of experimental validation is a correctness/external-validity concern, not circularity. No circular step can be exhibited from the paper's equations or citations.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities and no new physical constants. All device parameters (J0, gradients, noise amplitudes, impulse response, pulse bounds) are imported from the prior experimental modeling of Cerfontaine et al. [26]; the only in-paper fitted quantity is the Gaussian surrogate kernel of Section 4.1. RL hyperparameters follow prior literature. The load-bearing assumptions are that the Hamiltonian and noise models describe the real device and that simulator-trained protocols transfer to hardware; neither is experimentally tested here.

free parameters (4)

Gaussian surrogate kernel standard deviation = not stated numerically (fitted to the measured response)
Section 4.1 replaces the measured impulse response with a Gaussian whose standard deviation is fitted to it; the Markovianity results of Fig. 3 are computed with this surrogate.
Fast charge noise exponent alpha = 0.7 (optimistic) / 0.0 (white noise)
Section 2, Eq. 6: scenario values chosen by hand to bracket the high-frequency noise regime; they enter the noise spectrum S_e,alpha(f) used in training.
Noise amplitudes sigma_b, sigma_epsilon, S_e,alpha prefactor = 0.0105 J0; 0.0294 epsilon0; 53.8 epsilon0^2 ns
Section 2: imported from Cerfontaine et al. [26], not fitted here, but load-bearing for the realism of every reported fidelity.
RL hyperparameters (learning rate, buffer size, discount, TQC qk/qn) = Table 2 (5e-4; 100k; 0.99; 46/25)
Chosen by hand following prior RL literature (refs 36 to 38) with ablations; they condition training success but do not enter the physical model.

assumptions (5)

domain assumption The four-dot Heisenberg Hamiltonian with exchange couplings J(epsilon) and fixed gradients b12 = -b34 = J0, b23 = 7 J0 (Eq. 1) describes the device dynamics.
Section 2; imported from Cerfontaine et al. [26]. The absolute fidelity numbers inherit this model's accuracy.
standard math Trotterised evolution with time step delta-t is a faithful discrete approximation of the continuous dynamics.
Section 2: O(delta-t^2) error, 'chosen to be sufficiently small to mitigate discretisation errors'.
domain assumption The digitized impulse response from [26] represents the actual AWG-device response, including the ~1 ns delay.
Section 2, Eq. 4 and Fig. 2(b). Load-bearing for the Markovianity analysis of Section 4.1.
domain assumption Gaussian quasi-static hyperfine noise, slow charge noise, and 1/f^alpha fast charge noise capture the dominant decoherence.
Section 2, Eqs. 5 and 6; parameters imported from [26].
domain assumption Protocols trained in the simulator remain performant on the physical device (sim-to-real transfer).
Not tested in this paper; the paper positions it as future work (closed-loop fine-tuning, Section 5). This is the load-bearing premise behind the absolute fidelity claims.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Reinforcement learning entangling operations on spin qubits." pith.science (2026). https://pith.science/paper/UULSW6IU

@misc{pith2026250814761,
  author       = {Pith},
  title        = {Pith review of: Reinforcement learning entangling operations on spin qubits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UULSW6IU}},
  note         = {Machine review of arXiv:2508.14761}
}

read the original abstract

High-fidelity control of one- and two-qubit gates past the error correction threshold is an essential ingredient for scalable quantum computing. We present a reinforcement learning (RL) approach to find entangling protocols for semiconductor-based singlet-triplet qubits in a double quantum dot. Despite the presence of realistically modelled experimental constraints, such as various noise contributions and finite rise-time effects, we demonstrate that an RL agent can yield performative protocols, while avoiding the model-biases of traditional gradient-based methods. We optimise our RL approach for different regimes and tasks, including training from simulated process tomography reconstruction of unitary gates, and investigate the nuances of RL agent design.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 36 canonical work pages

[26]

High-fidelity gate set for exchange-coupled singlet-triplet qubits

Cerfontaine P, Otten R, Wolfe MA, Bethke P, Bluhm H. High-fidelity gate set for exchange-coupled singlet-triplet qubits. Physical Review B. 2020;101(15). https://doi.org/10.1103/physrevb.101.155311

work page doi:10.1103/physrevb.101.155311 2020
[1]

Control of quantum systems

Butkovskii A, Samoilenko YI. Control of quantum systems. Automation and Remote Control. 1979;40(4):485–502

work page 1979
[2]

Control of Quantum Systems 2

Butkovskii A, Samoilenko YI. Control of Quantum Systems 2. Automation and Remote Control. 1979;40(5):629–645

work page 1979
[3]

Training Schr¨ odinger’s cat: Quantum optimal control: Strategic report on current status, visions and goals for research in Europe

Glaser SJ, Boscain U, Calarco T, Koch CP, K¨ ockenberger W, Kosloff R, et al. Training Schr¨ odinger’s cat: Quantum optimal control: Strategic report on current status, visions and goals for research in Europe. The European Physical Journal D. 2015;69(12):279

work page 2015
[4]

Optimal control of coupled spin dynamics: design of NMR pulse sequences by gradient ascent algorithms

Khaneja N, Reiss T, Kehlet C, Schulte-Herbr¨ uggen T, Glaser SJ. Optimal control of coupled spin dynamics: design of NMR pulse sequences by gradient ascent algorithms. Journal of magnetic resonance. 2005;172(2):296–305

work page 2005
[5]

Second order gradient ascent pulse engineering

de Fouquieres P, Schirmer SG, Glaser SJ, Kuprov I. Second order gradient ascent pulse engineering. Journal of Magnetic Resonance. 2011;212(2):412–417

work page 2011
[6]

Global methods in optimal control theory

Krotov V. Global methods in optimal control theory. vol. 195. CRC Press; 1995. 27

work page 1995
[7]

Gradient optimization of ana- lytic controls: the route to high accuracy quantum optimal control

Machnes S, Ass´ emat E, Tannor DJ, Wilhelm FK. Gradient optimization of ana- lytic controls: the route to high accuracy quantum optimal control. arXiv preprint arXiv:150704261. 2015

work page 2015

Show all 46 references

[8]

Reinforcement learning: An introduction

Sutton RS, Barto AG. Reinforcement learning: An introduction. MIT press; 2018

2018
[9]

Kiran BR, Sobh I, Talpaert V, Mannion P, Sallab AAA, Yogamani S, et al.: Deep Reinforcement Learning for Autonomous Driving: A Survey
[10]

Smith L, Kostrikov I, Levine S.: A Walk in the Park: Learning to Walk in 20 Minutes With Model-Free Reinforcement Learning
[11]

Ouyang L, Wu J, Jiang X, Almeida D, Wainwright CL, Mishkin P, et al.: Training language models to follow instructions with human feedback
[12]

Mastering Atari, Go, chess and shogi by planning with a learned model

Schrittwieser J, Antonoglou I, Hubert T, Simonyan K, Sifre L, Schmitt S, et al. Mastering Atari, Go, chess and shogi by planning with a learned model. Nature. 2020;588(7839):604–609. https://doi.org/10.1038/s41586-020-03051-4

2020 doi
[13]

Rovibrational wave-packet manipulation using shaped midinfrared femtosecond pulses toward quantum computation: Optimization of pulse shape by a genetic algorithm

Tsubouchi M, Momose T. Rovibrational wave-packet manipulation using shaped midinfrared femtosecond pulses toward quantum computation: Optimization of pulse shape by a genetic algorithm. Physical Review A. 2008;77(5):052326

2008
[14]

Learning-Based Quantum Robust Control: Algorithm, Applications, and Experiments

Dong D, Xing X, Ma H, Chen C, Liu Z, Rabitz H. Learning-Based Quantum Robust Control: Algorithm, Applications, and Experiments. IEEE Transac- tions on Cybernetics. 2020;50(8):3581–3593. https://doi.org/10.1109/TCYB. 2019.2921424

2020
[15]

Exper- imental Deep Reinforcement Learning for Error-Robust Gate-Set Design on a Superconducting Quantum Computer

Baum Y, Amico M, Howell S, Hush M, Liuzzi M, Mundada P, et al. Exper- imental Deep Reinforcement Learning for Error-Robust Gate-Set Design on a Superconducting Quantum Computer. PRX Quantum. 2021 Nov;2:040324. https://doi.org/10.1103/PRXQuantum.2.040324

2021 doi
[16]

Universal quantum control through deep reinforcement learning

Niu MY, Boixo S, Smelyanskiy VN, Neven H. Universal quantum control through deep reinforcement learning. npj Quantum Information. 2019;5(1):33

2019
[17]

Reinforce- ment learning pulses for transmon qubit entangling gates

Nguyen HN, Motzoi F, Metcalf M, Whaley KB, Bukov M, Schmitt M. Reinforce- ment learning pulses for transmon qubit entangling gates. Machine Learning: Science and Technology. 2024;5(2):025066

2024
[18]

Model-Free Quantum Gate Design and Calibration Using Deep Reinforcement Learning

Shindi O, Yu Q, Girdhar P, Dong D. Model-Free Quantum Gate Design and Calibration Using Deep Reinforcement Learning. IEEE Transactions on Artificial Intelligence. 2024;5(1):346–357. https://doi.org/10.1109/TAI.2023.3243187

2024
[19]

Reinforce- ment learning in different phases of quantum control

Bukov M, Day AG, Sels D, Weinberg P, Polkovnikov A, Mehta P. Reinforce- ment learning in different phases of quantum control. Physical Review X. 2018;8(3):031086. 28

2018
[20]

Deep reinforcement learning for quantum gate control

An Z, Zhou D. Deep reinforcement learning for quantum gate control. Europhysics Letters. 2019;126(6):60002

2019
[21]

Model-free quan- tum control with reinforcement learning

Sivak V, Eickbusch A, Liu H, Royer B, Tsioutsios I, Devoret M. Model-free quan- tum control with reinforcement learning. Physical Review X. 2022;12(1):011059

2022
[22]

Quantum process tomography with unsupervised learning and tensor networks

Torlai G, Wood CJ, Acharya A, Carleo G, Carrasquilla J, Aolita L. Quantum process tomography with unsupervised learning and tensor networks. Nature Communications. 2023;14(1):2858

2023
[23]

Designing fast quantum gates using optimal control with a reinforcement-learning ansatz

Sarma B, Hartmann MJ. Designing fast quantum gates using optimal control with a reinforcement-learning ansatz. Physical Review Applied. 2025;23(1):014015

2025
[24]

Semiconductor spin qubits

Burkard G, Ladd TD, Pan A, Nichol JM, Petta JR. Semiconductor spin qubits. Rev Mod Phys. 2023 Jun;95:025003. https://doi.org/10.1103/RevModPhys.95. 025003

2023 doi
[25]

High-fidelity single- and two-qubit gates for two-electron spin qubits [Dissertation]

Cerfontaine P. High-fidelity single- and two-qubit gates for two-electron spin qubits [Dissertation]. R WTH Aachen University. Aachen; 2019. Ver¨ offentlicht auf dem Publikationsserver der R WTH Aachen University, Dissertation, R WTH Aachen University, 2019. Available from: ht...

2019
[27]

Storing quantum information for 30 seconds in a nanoelectronic device

Muhonen JT, Dehollain JP, Laucht A, Hudson FE, Kalra R, Sekiguchi T, et al. Storing quantum information for 30 seconds in a nanoelectronic device. Nature nanotechnology. 2014;9(12):986–991

2014
[28]

Charge Noise Spectroscopy Using Coherent Exchange Oscillations in a Singlet- Triplet Qubit

Dial OE, Shulman MD, Harvey SP, Bluhm H, Umansky V, Yacoby A. Charge Noise Spectroscopy Using Coherent Exchange Oscillations in a Singlet- Triplet Qubit. Phys Rev Lett. 2013 Apr;110:146804. https://doi.org/10.1103/ PhysRevLett.110.146804

2013
[29]

A quantum-dot spin qubit with coherence limited by charge noise and fidelity higher than 99.9%

Yoneda J, Takeda K, Otsuka T, Nakajima T, Delbecq MR, Allison G, et al. A quantum-dot spin qubit with coherence limited by charge noise and fidelity higher than 99.9%. Nature nanotechnology. 2018;13(2):102–106

2018
[30]

Fast universal quantum gate above the fault-tolerance threshold in silicon

Noiri A, Takeda K, Nakajima T, Kobayashi T, Sammak A, Scappucci G, et al. Fast universal quantum gate above the fault-tolerance threshold in silicon. Nature. 2022;601(7893):338–342

2022
[31]

Blueprint of a Scalable Spin Qubit Shuttle Device for Coherent Mid-Range Qubit Transfer in Disordered Si/SiGe/SiO 2

Langrock V, Krzywda JA, Focke N, Seidler I, Schreiber LR, Cywi´ nski L. Blueprint of a Scalable Spin Qubit Shuttle Device for Coherent Mid-Range Qubit Transfer in Disordered Si/SiGe/SiO 2. PRX Quantum. 2023 Apr;4:020305. https://doi. 29 org/10.1103/PRXQuantum.4.020305

2023 doi
[32]

Closed-loop control of a GaAs-based singlet-triplet spin qubit with 99.5% gate fidelity and low leakage

Cerfontaine P, Botzem T, Ritzmann J, Humpohl SS, Ludwig A, Schuh D, et al. Closed-loop control of a GaAs-based singlet-triplet spin qubit with 99.5% gate fidelity and low leakage. Nature communications. 2020;11(1):4144

2020
[33]

Filter-function formalism and software package to compute quantum processes of gate sequences for classical non- Markovian noise

Hangleiter T, Cerfontaine P, Bluhm H. Filter-function formalism and software package to compute quantum processes of gate sequences for classical non- Markovian noise. Physical Review Research. 2021 Oct;3(4). https://doi.org/10. 1103/physrevresearch.3.043047

2021
[34]

Exchange-based two-qubit gate for singlet-triplet qubits

Wardrop MP, Doherty AC. Exchange-based two-qubit gate for singlet-triplet qubits. Phys Rev B. 2014 Jul;90:045418. https://doi.org/10.1103/PhysRevB.90. 045418

2014 doi
[35]

Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor

Haarnoja T, Zhou A, Abbeel P, Levine S. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In: International conference on machine learning. PMLR; 2018. p. 1861–1870

2018
[36]

Haarnoja T, Zhou A, Hartikainen K, Tucker G, Ha S, Tan J, et al.: Soft Actor- Critic Algorithms and Applications
[37]

Available from: https: //arxiv.org/abs/2110.02034

Hiraoka T, Imagawa T, Hashimoto T, Onishi T, Tsuruoka Y.: Dropout Q- Functions for Doubly Efficient Reinforcement Learning. Available from: https: //arxiv.org/abs/2110.02034

arXiv
[38]

Controlling overestimation bias with truncated mixture of continuous distributional quantile critics

Kuznetsov A, Shvechikov P, Grishin A, Vetrov D. Controlling overestimation bias with truncated mixture of continuous distributional quantile critics. In: International Conference on Machine Learning. PMLR; 2020. p. 5556–5566

2020
[39]

Available from: https://arxiv.org/abs/1712.00378

Pardo F, Tavakoli A, Levdik V, Kormushev P.: Time Limits in Reinforcement Learning. Available from: https://arxiv.org/abs/1712.00378

arXiv
[40]

qopt: An Experiment-Oriented Software Package for Qubit Simulation and Quantum Optimal Control

Teske JD, Cerfontaine P, Bluhm H. qopt: An Experiment-Oriented Software Package for Qubit Simulation and Quantum Optimal Control. Physical Review Applied. 2022 Mar;17(3). https://doi.org/10.1103/physrevapplied.17.034036

2022 doi
[41]

Julia: A fresh approach to numerical computing

Bezanson J, Edelman A, Karpinski S, Shah VB. Julia: A fresh approach to numerical computing. SIAM Review. 2017;59(1):65–98. https://doi.org/10.1137/ 141000671

2017
[42]

Smooth function approximation using neural networks

Ferrari S, Stengel RF. Smooth function approximation using neural networks. IEEE Transactions on Neural Networks. 2005;16(1):24–38

2005
[43]

Adam: A method for stochastic optimization

Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980. 2014

2014
[44]

Kingma DP, Welling M.: Auto-Encoding Variational Bayes. 30
[45]

Quantum process tomography of unitary and near-unitary maps

Baldwin CH, Kalev A, Deutsch IH. Quantum process tomography of unitary and near-unitary maps. Physical Review A. 2014;90(1):012110

2014
[46]

Minimal informationally complete measurements for pure states

Flammia ST, Silberfarb A, Caves CM. Minimal informationally complete measurements for pure states. Foundations of Physics. 2005;35(12):1985–2006. 31

2005

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [26]

High-fidelity gate set for exchange-coupled singlet-triplet qubits

Cerfontaine P, Otten R, Wolfe MA, Bethke P, Bluhm H. High-fidelity gate set for exchange-coupled singlet-triplet qubits. Physical Review B. 2020;101(15). https://doi.org/10.1103/physrevb.101.155311

work page doi:10.1103/physrevb.101.155311 2020

[2] [1]

Control of quantum systems

Butkovskii A, Samoilenko YI. Control of quantum systems. Automation and Remote Control. 1979;40(4):485–502

work page 1979

[3] [2]

Control of Quantum Systems 2

Butkovskii A, Samoilenko YI. Control of Quantum Systems 2. Automation and Remote Control. 1979;40(5):629–645

work page 1979

[4] [3]

Training Schr¨ odinger’s cat: Quantum optimal control: Strategic report on current status, visions and goals for research in Europe

Glaser SJ, Boscain U, Calarco T, Koch CP, K¨ ockenberger W, Kosloff R, et al. Training Schr¨ odinger’s cat: Quantum optimal control: Strategic report on current status, visions and goals for research in Europe. The European Physical Journal D. 2015;69(12):279

work page 2015

[5] [4]

Optimal control of coupled spin dynamics: design of NMR pulse sequences by gradient ascent algorithms

Khaneja N, Reiss T, Kehlet C, Schulte-Herbr¨ uggen T, Glaser SJ. Optimal control of coupled spin dynamics: design of NMR pulse sequences by gradient ascent algorithms. Journal of magnetic resonance. 2005;172(2):296–305

work page 2005

[6] [5]

Second order gradient ascent pulse engineering

de Fouquieres P, Schirmer SG, Glaser SJ, Kuprov I. Second order gradient ascent pulse engineering. Journal of Magnetic Resonance. 2011;212(2):412–417

work page 2011

[7] [6]

Global methods in optimal control theory

Krotov V. Global methods in optimal control theory. vol. 195. CRC Press; 1995. 27

work page 1995

[8] [7]

Gradient optimization of ana- lytic controls: the route to high accuracy quantum optimal control

Machnes S, Ass´ emat E, Tannor DJ, Wilhelm FK. Gradient optimization of ana- lytic controls: the route to high accuracy quantum optimal control. arXiv preprint arXiv:150704261. 2015

work page 2015