Minimal time robust control for two superconducting qubits Authors: Niril George, Joseph L Allen, Robert Kosut, Eran Ginossar Published: 2025 ABSTRACT AbstractHigh-fidelity quantum gates are crucial for achieving fault-tolerant quantum computing; however, decoherence significantly reduces gate fidelities during long operation times. Although optimal control techniques can theoretically minimize these operation times, they often neglect realistic uncertainties in system parameters. In this work, we demonstrate that by using robust optimal control strategies, the cross-resonance gate in superconducting systems can be operated within 64 ns, achieving fidelities of while maintaining robustness against up to 10% uncertainty in a single parameter. Alternatively, by extending the control time to 71 ns, we achieve fidelities of with robustness against up to 3% uncertainty. Our results identify the minimal control times attainable with experimentally feasible pulses and system parameters, as well as the maximum allowable static parameter error for high-fidelity operations. Furthermore, we demonstrate simultaneous robustness against both static and time-dependent errors by generating 100 ns control pulses () that maintain robustness against 10% static parameter error and time-dependent parameter fluctuations two orders of magnitude stronger than typical experimental noise. These findings demonstrate a viable open-loop strategy for implementing fast, high-fidelity quantum gates in the presence of realistic system uncertainties that would otherwise degrade conventional control pulses.Subject terms: Engineering, Physics Introduction Superconducting qubits are among the most promising candidates for quantum computing architectures1–4. A high-precision implementation of desired unitary operations is essential for effective quantum computation5. Errors in quantum gates primarily stem from incoherent processes, characterized by the and time constants, and from unitary errors due to imperfect control protocols and implementations6. Reducing gate times can mitigate incoherent errors by making their effects negligible. However, a fundamental speed limit exists for gate operations7–14. While these speed limits are typically derived analytically for simple two-level systems, more complex multi-level systems require numerical methods to determine minimal control times, which are inherently longer than the theoretical speed limits 15. Optimal control techniques have been widely employed to improve the fidelity of desired quantum operations by optimizing system parameters and pulse shapes16–24. These techniques can be extended to mitigate the effects of noise, control errors, and system uncertainties, forming what is known as robust optimal control (ROC). Sampling-based ROC, for example, involves the search for design variables that maximize either the average or the worst-case gate fidelity across a set of uncertain parameters25–27. More advanced techniques employ stochastic sampling where new uncertainty samples are generated at each iteration, thereby improving generalization and scalability to higher-dimensional errors28,29. Other strategies achieve robustness analytically, including composite pulse sequences that cancel amplitude and detuning errors30,31, dynamically corrected gates that embed the target operation in engineered pulse sequences to suppress decoherence and control imperfections32–35, and DRAG (derivative removal by adiabatic gate) and its variants, which use derivative-quadrature pulse shaping to mitigate leakage and phase errors36–39. Machine learning has also been used to improve robustness, from simulation-based training40–42 to on-hardware training of control pulses43. More advanced frameworks seek to optimize for both robustness and speed, shifting the focus to time-optimal robust control44–48; however, extending these results to multi-qubit gates in superconducting circuits remains an open challenge. The cross-resonance gate is one of the most effective superconducting quantum gates for generating entanglement49–51. It is an all-microwave gate implemented on fixed-frequency qubits, leveraging the cross-coupling between them. This gate is advantageous for superconducting transmon qubits, which offer long coherence times52, reduced charge noise53, and high single-qubit gate fidelities54, owing to the use of fixed-frequency transmons that can be engineered for optimal performance. The minimal overhead of the gate, requiring only microwave control, makes it particularly suitable for scaling up quantum computers. Reported gate times for the cross-resonance gate have been relatively slow compared to those for flux-tunable gates. Published benchmarks include at gate times of 160 ns from Sheldon et al.55, rising to with error-mitigation techniques at 484 ns56. System-wide characterizations of large-scale IBM processors report median two-qubit fidelities of with gate times 57, while in-situ deep reinforcement–learning optimization has improved the fidelities to with similar gate times on the same hardware43. Cross-resonance–enabled two-qubit gates have also demonstrated in 180 ns58. While fidelities at this level are sufficient for fault-tolerant implementations within the surface-code architecture5,59,60, there remains room for improvement in both fidelity and operation time. A time-optimal control approach was applied to the cross-resonance gate15, revealing that significantly shorter gate times are possible: it numerically determined an unconstrained quantum speed limit of ns and identified experimentally feasible pulses achieving fidelity in 70 ns with a bandwidth-constrained single-carrier drive. Robust optimal control has been applied to similar cQED gates, demonstrating multi-level model fidelities of in 199 ns robust to static parameter uncertainty61. More recently, reinforcement learning approaches have generated pulses (249 ns) that can achieve with robustness to hardware drifts of up to in simulation62, while Baum et al.43 experimentally demonstrated that their deep reinforcement learning optimized gates maintained fidelities against calibration drifts for over 25 days. In this work, we employ the robust optimal control methodology from 25 to ascertain the minimal gate time necessary for achieving high-fidelity operations while ensuring robustness against uncertainties in the system’s intrinsic parameters. Our model encompasses the four lowest energy states of the transmon, effectively representing two directly coupled multi-level qubits with direct driving mechanisms. We investigate varying levels of static uncertainty in a single system parameter, spanning from to , which are within the bounds of experimental relevance. Our findings reveal that control pulses with durations of 64 ns maintain fidelities exceeding 0.99 across all considered uncertainty levels. Additionally, pulses of 71 ns achieve fidelities above 0.999 for uncertainties up to . Notably, securing a worst-case fidelity of within multi-level systems for gate times under 100 ns constrains the maximum permissible uncertainty to . This highlights an inherent trade-off between uncertainty and performance, underscoring the importance of fabricating hardware with well-characterized uncertainties, rather than merely augmenting control strategies in existing systems. We further show that our framework exhibits robustness to time-dependent errors by generating control pulses under 100 ns achieving that maintain level of robustness against static parameter error while simultaneously establishing robustness against time-dependent parameter fluctuations. We confirmed this robustness not only in realistic noise simulations but also in strong noise simulations where parameter fluctuations are amplified by two orders of magnitude. Our results enable the application of robust optimal control techniques in future two-qubit gate experiments, demonstrating that pulse-shaping methods can effectively incorporate uncertainties to attain high fidelities with minimal gate durations. The remainder of this paper is organized as follows: In “Cross-resonance gate for multi-level qubits”, we define the theoretical model for the cross-resonance gate used in this study. Section “Robust optimal control” then describes the robust optimal control methods and the algorithm employed to achieve the reported results. Next, in “Non-robust time optimization”, we determine the minimal time for the cross-resonance gate using experimentally feasible parameterizations without considering robustness. Following this, “Robust time optimization” presents the results of the search for a time-optimal cross-resonance gate robust to static parameter uncertainty. In “Robust control against time dependent errors”, we demonstrate that the robust cross-resonance gate operation remains effective under extreme conditions by establishing simultaneous robustness against static and time-dependent errors. Finally, “Conclusion” presents our conclusions. Cross-resonance gate for multi-level qubits In this work, we analyze the cross-resonance gate between two transmons coupled by a bus resonator, with microwave drives directly applied to each transmon, which replicates the experimental setup reported by Sheldon et al. 55. When the resonator’s fundamental frequency is significantly higher than the transition frequencies of the individual transmons, the Hamiltonian can be projected onto the zero-excitation subspace of the resonator. By additionally applying the rotating wave approximation, we obtained the following lab-frame effective Hamiltonian: Here, and represent the annihilation and creation operators of transmon j, is the frequency of transmon j, is its anharmonicity, J is the coupling strength between the two transmons, and is the pulse envelope (control) for drive k on transmon j with carrier frequency and phase . We assumed that the Hamiltonian for the transmon is of the form of a Duffing oscillator19, which is valid if the anharmonicity is in the transmon regime. To perform a cross-resonance gate, the “control” qubit is driven by a direct microwave drive at the frequency of the “target” qubit50,63. Due to the cross-coupling between the qubits, this generates an entangling operation. To see this, consider a Hamiltonian of two directly coupled qubits with a single microwave drive on the first qubit, with control only in the x-quadrature: This Hamiltonian indicates that entanglement can be generated by the coupling between the qubits by simply waiting an appropriate amount of time. However, as J is generally small, this time is very long. Instead, because the coupling between the qubits is weak, such that , we can diagonalize the first line of the Hamiltonian to effectively decouple the qubits. Using this transformation, the drive term can be transformed, so the full Hamiltonian becomes with and being the decoupled, shifted qubit frequencies. Therefore, by choosing the drive frequency , a two-qubit entangling operation using can be performed. For simplicity, in Eq. 2, only single-quadrature control was chosen. Generally, in experiments, there is full control of both quadratures for the drives. Additionally, for direct drives on the qubits, there are two microwave control lines, one for each qubit, as is the case in ref. 55 . In this case, in the drive frame, the Hamiltonian becomes with being the detuning of transmon j from the drive frequency. For the simulations, we transformed the Hamiltonian into the dressed basis and truncated the matrix to include only the first four energy levels of both transmons. Higher energy levels were excluded from our analysis, as our simulations (see Sec. B and Fig. 4) showed that energy levels beyond the third had minimal occupation when the drive power for the cross-resonance drive was set to match the parameters in55. Furthermore, we precisely matched the system parameters to those used in the referenced study for a fair comparison. Specifically, the frequencies were set at GHz and GHz, and both transmons exhibited an anharmonicity of MHz. The bus resonator frequency was maintained at 6.31 GHz, resulting in an estimated coupling of MHz between the two qubits. Population dynamics of robust cross-resonance gates. This figure shows the time evolution of populations (in the computational subspace and the third and fourth levels) during the 71 ns cross-resonance gate obtained from SCP optimizations. The subplots illustrate the dynamics for gates optimized against varying levels of static J coupling uncertainty (): the left panel corresponds to a 0% target (standard), the middle to , and the right to . Robust optimal control Many experimental procedures have refined the use of standard pulse shapes by calibrating the pulses to minimize errors due to drifts in the control fields or the inevitable filtering they undergo. However, as these approaches reach the limits of what is achievable with standard pulses and current hardware, there is a limit on the achievable gate time, as a standard pulse requires a certain duration to perform the desired gate. Quantum optimal control provides a method for creating shaped, intelligent pulses that can allow for shorter gate times. However, these are often limited in their effectiveness, as the optimal control simulation requires a predefined model with specific system parameters. It is acknowledged that parameter estimation in multi-qubit superconducting systems requires improvement64,65. Previously, it has been shown that if parameter uncertainty is not included in the optimization, the fidelity over the uncertain range significantly decreases away from the simulated parameter value25,61. Measuring system parameters is resource-intensive, and as the scale of quantum computers increases, this becomes a significant issue. As systems scale up, this leads to greater uncertainty in the system parameters and therefore likely gives rise to optimal control solutions that, when applied experimentally, achieve significantly lower fidelity than simulated. Robust optimal control seeks to ensure that this drop-off in fidelity is reduced considerably by including robustness in the search. This allows optimal-control-shaped pulses to be less limited by the constraints of the simulated model and to provide experimentally feasible results. Robustness aids in scaling up quantum computers by ensuring less variability in the performance of quantum gates across a multi-qubit processor and making the implementation more reliable overall. The work in this paper used numerical optimization techniques to find optimal pulse shapes to perform the gate of interest in minimal time. In particular, the interest was in achieving fast gate times by finding pulses that are robust to uncertainties. By decreasing the times, incoherent errors, such as decoherence, can be effectively removed. In this case, the focus was on the coherent errors resulting from, for example, parameter uncertainty. Sampling-based robust optimization takes a sample of points, , from some range, , of the uncertain parameters (or errors)25. The fidelity is calculated from the time evolution of the Hamiltonian for each of the sampled parameter values, resulting in a range of fidelities, . Using this range of fidelities, the aim is then to either maximize the average fidelity of the range or to maximize the worst-case fidelity in the uncertain range25. In general, for quantum computing, maximizing the average fidelity over the uncertain range is not enough. If the average fidelity is maximized, there may exist a value in the range of uncertainty that performs far below the desired threshold. In this case, if the real system parameter has a value at this sub-optimal point, then the actual performance of the gate would also be sub-optimal. Indeed, in this situation, slow parameter drifts may often encounter adverse parameter values and thus degrade the fidelity over the duration of a quantum algorithm. If, instead, the worst-case fidelity is maximized for the uncertain range, it is guaranteed that all values in the range perform at least as well as this worst-case fidelity, ensuring that slow parameter drifts do not adversely affect the fidelity during operation. Therefore, maximizing the worst-case fidelity is the ideal target for quantum gates. The numerical optimization algorithm used for this work was based on the Sequential Convex Programming (SCP) algorithm25. This is a gradient-based, local search optimizer, similar in concept to GRAPE66. The SCP algorithm produces piecewise-constant amplitudes over N uniform time intervals, , where T is the final time of the system evolution. The Hamiltonian at time is then given by: where is the drift (i.e., not controlled)) Hamiltonian, j is the index of the control, is the constant amplitude at time , and is the time-independent control Hamiltonian associated with control j. Using the piecewise-constant approximation, the unitary evolution of the system U(T) is given by , with . In the piecewise-constant approximation, as the number of time intervals grows, so does the dimension of the control space. The optimization goal is to maximize a cost function of the form: with being the projector into the computational subspace, the size of the computational subspace, U(T) the time evolution of the full system at time T, and W the desired unitary operator. This desired unitary operator was chosen as it is related to the CNOT operation via: The optimizer is initialized by choosing an initial control, , which is the set of all complex-valued amplitudes at each time step. This initial control is within the feasible set E, defined by user-defined constraints on the control. The trust region, , effectively tells the optimizer the largest change that can be made to each individual time-step pulse while respecting the local linearization for the convex optimization update. Specifically, the trust region is adjusted after each optimization step: if the actual fidelity increases, the region becomes larger, whereas if the actual fidelity does not increase, the region is reduced to increase the likelihood of finding a change that enhances the fidelity. The uncertain set, , is created to define the size of the static uncertainty in the parameters of interest. From this uncertain set, a fixed sample of uncertain values, , is taken. At each iteration of the optimizer, SCP seeks to find an optimal solution for the increment, , by maximizing the linearized fidelities, , subject to and . Upon the return of the optimal that satisfies all the constraints, the new fidelities are calculated and compared with the fidelities from the previous iteration. If , the control is updated as , the trust region is increased, and the optimization step begins again. If, however, , the trust region is instead decreased, and the optimization step is repeated with the original . The key to the SCP algorithm is that the step to find the increment, , is a convex optimization. In addition, the algorithm compares the worst-case fidelity across the uncertain range at each iteration to maximize it. In the GRAPE algorithm, the method for increasing the fidelity is to update the control as , where k denotes the relevant piecewise-constant amplitude in the control and is a predefined increment. As the control is directly changed by the inclusion of the gradient, the only way to include robustness is to calculate the average fidelity across the uncertain range, e.g., , and then use this for the gradient calculation and the update step. The update step then guarantees that the average fidelity over the uncertain range is maximized, not the worst-case fidelity. Other works have discussed sampling-based optimization for robust control by maximizing the average fidelity67. Alongside static parameter errors, superconducting qubits are also subject to time-dependent fluctuations in both the drift and control Hamiltonians. These arise from low-frequency drifts and 1/f noise in the device parameters 68–72, as well as broadband noise and imperfections in the classical control electronics 73–79, all of which can significantly impact gate fidelity. Consequently, to ensure high-fidelity operation, the robust optimal control framework must exhibit simultaneous robustness against both static and time-dependent errors. This robustness is achieved by incorporating an ensemble of time-dependent errors trajectories, sampled from the underlying error statistics, directly into the cost function: Here L independent trajectories are generated for each static sample , and they modulate the relevant Hamiltonian parameters in time. The robust objective is then taken as , so that the optimisation targets the worst-case fidelity over the static error set, with each static point evaluated in terms of its average performance under time-dependent errors. The finite set of static samples is selected once at the beginning of the optimisation and kept fixed throughout, whereas the time-dependent errors trajectories are regenerated at every SCP iteration. This resampling prevents overfitting to any particular error realisation and drives the search towards pulses whose performance reflects the underlying error statistics. By choosing L to be sufficiently large, the variance of this stochastic estimate is kept small enough for stable trust-region updates, yielding control pulses that are simultaneously robust to static parameter errors and to realistic time-dependent errors. Non-robust time optimization Before investigating time optimization for robust optimal control, it is useful to examine time minimization without robustness. As in almost all quantum control problems, solutions to time-optimization formulations are local. There are a variety of methods for finding local minimal time solutions. In simpler systems, a quantum speed limit can be derived using analytical methods. These speed limits are unique to the system of interest and the desired operation. However, for the multi-level systems in question, numerical methods are required to find the minimal operation time, for example, by including the gate time in the cost function to penalize long gate times 80, or by plotting fidelity as a function of time 15. We adopted the latter strategy, running multiple optimizations from random initializations over a range of pulse durations to identify the shortest times required for high-fidelity operations. Implementing the fidelity mapping required careful consideration of practical experimental constraints. While state-of-the-art AWGs have sampling rates of , typical experimental setups are operated with AWGs that have sampling rates of . Accordingly, we formulate the optimization over latent variables c(t) and obtain the physical drive by Gaussian spectral shaping, described mathematically as . Here, is the final pulse envelope applied to the hardware, c(t) denotes the underlying control waveform optimized by the algorithm, and are the Fourier transform and its inverse, respectively, is a Gaussian low-pass filter in angular frequency (enforcing a prescribed bandwidth), and the product represents the filtered Fourier spectrum of the optimized control. This procedure acts as a built-in low-pass that constrains the pulse bandwidth, mitigates high-frequency leakage to unwanted transitions, and ensures sufficient smoothness for accurate reproduction by the hardware. In our work, the filtering was set to have a minimum effective time step of (larger for longer pulses), which is compatible with standard AWG capabilities. With these parameters defined, optimization was performed for pulse durations ranging from to in increments. For each duration, the SCP algorithm was executed with 25 independent random initial guesses, terminating upon reaching the maximum iteration limit or achieving convergence (see Table 1 for all optimization parameters). The results of these standard (non-robust) optimizations are plotted in Fig. 1, where the black line shows the average gate infidelity () as a function of pulse duration. These optimizations yield average fidelities of in just , in , and in only , representing a twofold reduction in gate time for similar fidelity levels compared to current experimental demonstrations43,55–58 and is notably consistent with the time reported in previous time-optimal (non-robust) studies of similar systems15. Table 1 Optimization hyperparameters. Core algorithm parameters AlgorithmSequential convex programmingRandom initializations25Initial trust region0.001Trust region increase factor1.15Trust region decrease factor0.5 Stopping conditions Max iterations30000Fidelity convergence over 10 iterationsTrust region convergence over 10 iterations Time grid Non-robust optimization1–100 (1 ns steps)Static robust optimization1–100 (1 ns steps)Simultaneous robust optimization100ns Sampling parameters Non-robust optimization1 Error free hamiltonianStatic robust optimization3 static hamiltonians (, 0, )Simultaneous robust optimization3 (100 Dynamic Noisy + 1 Static Noise-Free) HamiltoniansThis table details the core parameters for the Sequential Convex Programming (SCP) algorithm, the stopping conditions for convergence, and the specific time grid and sampling strategies used for the non-robust (“Non-robust time optimization”), static robust (“Robust time optimization”), and simultaneous robust (“Robust control against time dependent errors”) optimization scenarios. The term Dynamic indicates that the stochastic error realizations are refreshed at each iteration. Optimization hyperparameters. This table details the core parameters for the Sequential Convex Programming (SCP) algorithm, the stopping conditions for convergence, and the specific time grid and sampling strategies used for the non-robust (“Non-robust time optimization”), static robust (“Robust time optimization”), and simultaneous robust (“Robust control against time dependent errors”) optimization scenarios. The term Dynamic indicates that the stochastic error realizations are refreshed at each iteration. Robust cross-resonance gates performance. The top panel plots the cross-resonance gate infidelity () achieved by the SCP optimization algorithm as a function of the target gate time ( to ), where each curve corresponds to an optimization targeting robustness against a specific maximum static uncertainty in the J coupling (), ranging from 0% (standard optimization, black line) up to (colored lines, see legend). For each gate time and uncertainty target, the optimization was run multiple times with different random initial pulses; the plotted infidelity at each T is the average over random starts of the worst-case fidelity across the three sampled couplings. Background shading indicates target fidelity regions: (red, infidelity ) and (light blue, infidelity ). The minimum pulse duration for fidelity (optimized for robustness) is . For fidelity (optimized for robustness), the minimum duration is . The bottom panels show the infidelity () of two representative pre-optimized pulses when subjected to a range of errors in the J coupling: the left panel shows the performance of a pulse optimized for a gate time targeting robustness in , achieving the desired fidelity within this uncertainty range, while the right panel shows a pulse optimized for targeting uncertainty, achieving within its range. The solid line indicates the infidelity of the best optimized pulse simulated at the corresponding coupling error, while the shaded area represents the infidelity range across an ensemble of optimized pulses (from the other starting points) tested at that error. The vertical dotted lines mark the ± error range for which each pulse was specifically optimized. Robust time optimization To provide an example of our robust optimal control framework and its effect on the minimal control time, a simplified model was considered where only a single static uncertainty was included in the system parameters. Typically, most system parameters in a two-qubit coupled system are easily measured using standard qubit spectroscopy with relatively low uncertainty. However, the qubit-qubit coupling cannot be directly measured using such methods. Instead, it is deduced via fitting to measurements 81. For example, one method involves quantum state preparation of a control qubit (which introduces some error), followed by spectroscopic measurements on the target qubit, and then modeling of the coupling based on electromagnetic simulations of the modes. This leads to the qubit-qubit coupling J being the most uncertain system parameter, and it was therefore the focus of this work. To incorporate robustness into the time optimization, we adapted the optimization procedure outlined in “Non-robust time optimization” with the modified objective to maximize the worst-case gate fidelity across a specified uncertainty range in the selected Hamiltonian parameters. As before, the SCP algorithm was executed for pulse durations from to (utilizing 25 independent initializations for each duration, as detailed in Table 1), for varying sizes of uncertainty in the coupling strength J, ranging from no uncertainty up to error, in increments of . For each uncertainty range, we sampled three values (which were kept fixed throughout the optimization): the maximum positive error, zero error, and the maximum negative error. Keeping the zero-error Hamiltonian is essential to ensure that we do not end up with a solution that has reduced fidelity at the most probable point. It is worth mentioning that when the error range is small, sampling the zero-error point may not be necessary 82. However, in a larger error range, it may be beneficial to sample more points 26. The results of the robust optimization in Fig. 1 show that pulses optimized for robustness against a uncertainty in the coupling strength J can achieve a worst-case fidelity of in just . Furthermore, by constraining the uncertainty to , worst-case fidelities above are attainable in (an example of this time-optimal robust pulse is shown in Fig. 3). Although this is slower than the non-robust case, it remains twice as fast as current experimental demonstrations43,55–58. These results also represent an improvement over previously reported robust control strategies for similar systems. For instance, we achieved a worst-case with robustness faster than reinforcement learning approaches that achieved average fidelities () against static drift in 62. While at , we achieved a much larger uncertainty robustness of . Example of a Robust Cross-Resonance Control Pulse. This pulse was obtained from the SCP optimization, achieving a fidelity of with a duration of 71 ns while maintaining robustness against static uncertainty in the J coupling. The bottom panels of Fig. 1 validate the effectiveness of the generated pulses by testing two example pulses against variations in the J coupling. The left panel confirms that the pulse (optimized for with robustness) successfully maintains its target fidelity across the entire error range. Similarly, the right panel shows that the pulse (designed for with robustness) achieves its fidelity goal within the specified window. At these short gate times, decoherence effects are negligible, as transmon coherence times are orders of magnitude longer—typically tens of microseconds 55 and up to for state-of-the-art devices 83. We observed that there is a trade-off between gate operation time and static uncertainty. This can be understood as arising from an equivalence between the time constraint and the robustness requirement, as illustrated by the following simplified argument. Given a general Hamiltonian with a control , the equation of motion is: When we reduce the time duration of the control task by rescaling time with , this is equivalent to a perturbation of the Hamiltonian, because: where and . In a situation where fidelity is locally a convex function of the two constraints, the fidelity may be maintained by compensating for the change in the time constraint with a change in uncertainty. Related ideas were discussed regarding the trade-off between fidelity and time-optimal control 84. The results also indicate that the fidelity values do not increase linearly with gate operation time and that there appears to be a maximum reachable fidelity value, whereby an increase in gate operation time will not yield higher fidelity. For uncertainties greater than , the best achievable worst-case fidelities are no greater than , within the constraints of the simulated system. This suggests that the optimal level of uncertainty for this system is or less; if this level of measurement uncertainty for the J coupling can be reached, then high-fidelity, short, robust pulses can be achieved. It should be noted that to achieve robust and high-fidelity results, the transfer function of the AWGs, transmission lines, connectors, and other electronics should be included in the optimization conducted for a specific experiment. This will counteract a possible reduction in fidelity caused by transfer function distortions 61. Robust control against time dependent errors To demonstrate simultaneous robustness against static errors and time-dependent errors, we subject the Hamiltonian parameters of the cross-resonance gate to stochastic variations. These variations simulate the effect of environmental noise and control imperfections under both realistic conditions and strong noise regimes, where fluctuations are amplified by two orders of magnitude. To systematically implement these fluctuations, we model each fluctuations, , as a stochastic process constructed from a sum of N Fourier components85: The spectral character of the perturbation is determined by the distributions for the amplitudes and frequencies , allowing us to simulate common noise types. For each noise realization, the phases are drawn uniformly from , and the amplitudes are initially drawn from a zero-mean Gaussian distribution. To generate (flicker) noise, these amplitudes are then scaled by , with frequencies sampled from a logarithmic distribution over the frequency band . For white noise, the amplitudes are used as-is (unscaled), and the frequencies are sampled from a uniform distribution over . This process models the physical effects of fluctuations by modulating the ideal system parameters via: where we define as the time-dependent physical fluctuation. The amplitudes of the underlying stochastic process are scaled so that the root mean square (RMS) of the physical fluctuation matches the target values specified in Table 2. Table 2Noise environment parameters.Parameter fluctuationRealistic environmentStrong noise environmentQubit frequency () Hz HzCoupling ()Control amplitude ()Control phase () rad radCrosstalk variation ()The table lists the target strength for each simulated noise source, quantified as the root mean square (RMS) of the physical fluctuation, . Two scenarios are presented: a Realistic Environment with strengths derived from experimental literature, and a Strong Noise Environment that amplifies these fluctuations by two orders of magnitude to serve as a stress test. Noise environment parameters. The table lists the target strength for each simulated noise source, quantified as the root mean square (RMS) of the physical fluctuation, . Two scenarios are presented: a Realistic Environment with strengths derived from experimental literature, and a Strong Noise Environment that amplifies these fluctuations by two orders of magnitude to serve as a stress test. We applied this stochastic model to the primary sources of noise found in fixed-frequency transmons, which can be broadly categorized into fluctuations affecting the drift and control Hamiltonians. For the drift Hamiltonian, the primary fluctuations are caused by charge noise, variations in the critical current, and microscopic two-level defects. These sources predominantly manifest as fluctuations in the qubit’s frequency () on the order of and in the coupling strength () of approximately 68–70,86. Physically, these fluctuations have a 1/f spectral density at low frequencies, which we modeled using a 1/f spectrum over the range and a white noise profile over 71,72. For the control Hamiltonian, we considered two main noise sources. The first is pulse imperfections, which manifest as amplitude and phase noise. Amplitude noise appears as fluctuations in the quadrature amplitudes, and , with typical variations of 73,74, while phase noise appears as a time-dependent mixing of these quadratures, corresponding to fluctuations of about per unit amplitude75. Similar to the drift parameters, their spectral character is 1/f at low frequencies, transitioning to white noise at higher frequencies. This was modeled with a 1/f spectrum over and a white noise component over 76–78. The second control noise source is classical crosstalk, where control signals for one qubit inadvertently affect another. This typically arises from how control drives are integrated into the circuit, with the magnitude of the crosstalk depending on the coupling interaction between the drive and neighboring qubits 50. To model classical crosstalk, the control term (third term) of Eqn. 4 was expanded to include terms that represent the classical coupling of one qubit’s control drive to the other qubit: The time-dependent crosstalk strength from the drive on qubit j to qubit i, , was modeled as a stochastic process with a mean magnitude of and short-timescale variations on the order of 79, following a 1/f spectrum over and a white noise component over . We performed the robust optimization using the objective defined in Eq. (8), employing static samples corresponding to the nominal coupling and the error bounds (). For the dynamic average, we utilized independent fluctuations generated via Eq. (11) at each iteration. We found that this ensemble size, combined with the inclusion of the fluctuation-free trajectory, kept the variance of the objective sufficiently low to ensure stable trust-region updates. The specific strengths of the fluctuations for the scenarios tested are detailed in Table 2. Our key finding is that we can generate sub-100 ns control pulses achieving fidelities that are simultaneously robust to static parameter errors (up to ) and to time-dependent errors of all parameters (up to two orders of magnitude larger than typical experimental noise). Figure 2 illustrates this by validating the performance of representative robust pulses against variations in the coupling J across 100 independent fluctuations. This confirms that the framework can handle significant time-dependent errors without compromising static robustness. Performance of cross-resonance gates optimized for simultaneous robustness (). This figure shows the gate infidelity, , of two representative pulses explicitly optimized to ensure simultaneous robustness against static coupling errors and time-dependent parameter fluctuations. The left panel shows the performance of a pulse optimized for static error in alongside realistic experimental noise, achieving the desired fidelity. The right panel shows a pulse optimized for the same static error but targeting robustness against a strong-noise environment, where parameter fluctuations are two orders of magnitude larger. (Both noise environments are defined in Table 2.) The solid line shows the infidelity of these robust pulses in a noise-free environment, while the shaded area represents the infidelity range across 100 unique, time dependent error realizations. The vertical dotted lines mark the ± static error range for which each pulse was specifically optimized. Conclusion In this work, we demonstrated that our robust optimal control framework can generate cross-resonance operations achieving fidelities of in 64 ns with robustness against static uncertainties of up to , and fidelities of in 71 ns with robustness against uncertainties of up to . This result, achieved in , is nearly as fast as the non-robust time-optimal limit of 15, yet it simultaneously incorporates robustness to uncertainty. This combination of speed and robustness significantly surpasses previous robust strategies, which required to achieve similar fidelities and robustness levels61,62. This also represents a twofold improvement in gate speed compared to other experimental demonstrations of the cross-resonance gate 43,55–58, all while maintaining robustness against the highly uncertain qubit-qubit coupling. Furthermore, we demonstrated that this robustness extends to time-dependent errors, validating the method’s effectiveness under both realistic conditions and strong-noise regimes up to two orders of magnitude beyond experimental norms. These results are highly promising for fault-tolerant quantum computing. Fidelities of are sufficient for surface-code quantum error correction59,87–89, a leading error-correction strategy for superconducting devices90–93. Moreover, by accepting a tighter uncertainty tolerance to reach , the encoding overhead for many error-correcting codes can be dramatically reduced89,94,95. We have further shown a fundamental trade-off in control: for any given fidelity target, there is a maximum level of uncertainty a device can tolerate, especially at short gate times. This principle is particularly relevant for noisy intermediate-scale quantum (NISQ) technologies96,97, where fidelities can be relaxed to gain speed. Our results demonstrate that gate times of 51 ns can be realized for the full range of uncertainties, provided that a lower fidelity of approximately is tolerated. This performance meets the target for many NISQ applications, and the robustness of the results ensures less variability across multi-qubit processors, which is ideal for scaling up quantum computers. We highlight that the framework presented here are not limited to the cross-resonance gate and the methodology can be extended to any optimal control goal, such as other quantum gates. In addition, it is not limited to superconducting qubits; the methods outlined can be applied to any quantum system with any desired control objective. While analytical explanations for pulse robustness exist, they are limited to a few idealized cases (e.g., the well-known DRAG protocol36) and do not account for practical constraints such as uncertainties. Numerous studies by Rabitz and colleagues have shown that the control landscape topology becomes increasingly fractured and complex under such constraints98. This highlights the need for further dedicated research to elucidate the control landscape and pulse dynamics in these challenging scenarios. Here, we have demonstrated the use of complex-shaped pulses that achieve very high-fidelity gates, marking a significant advancement toward practical applications in quantum computing. Appendix A: Example of a robust cross-resonance control pulse Figure 3 illustrates the control pulse for the shortest cross-resonance gate produced by the SCP optimization algorithm that achieves a fidelity of while maintaining robustness against static uncertainty in the J coupling. Appendix B: Population dynamics See Fig. 4. Author contributions E.G. conceived and supervised the project. N.G., J.L.A., and R.K. contributed to the development of the robust optimal control framework. N.G. performed the numerical simulations and data analysis. N.G. and J.L.A wrote the main manuscript text. All authors reviewed the manuscript. Funding E.G. acknowledges financial support from the EPSRC grant EP/T001062/1, and the EPSRC strategic equipment grant EP/L02263X/1. For R.K this material is based upon work supported by, or in part by, the U. S. Army Research Laboratory and the U. S. Army Research Office under contract/grant number W911NF2310307. Data availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.