Optimal and Learning-Based Quantum Control

Quantum control research asks which operations can be achieved with available fields and information, and how reliably they can be implemented in a real system. Optimization produces a candidate solution to a specified objective. It does not establish that the objective captures the experiment, that the model is correct, or that the result is globally optimal.

Required background. Optimal-control computation supplies gradients and numerical checks; open-system optimal control supplies the physical setting.

The objective defines what improvement means

A typical model uses $H(t)=H_0+\sum_j u_j(t)H_j$ with bounded controls. A robust objective can average a task error over a specified uncertainty ensemble or optimize a worst-case bound. Add penalties only when they represent actual constraints, such as bandwidth, fluence, leakage, or sensitivity to drift.

For example,

$$J[u]=\mathbb E_{\theta\sim p}[1-F_\theta(u)] +\lambda\int_0^T\sum_j |u_j(t)|^2dt$$

trades average infidelity against control effort under the chosen distribution $p$. The units and normalization of $\lambda$ must make the terms comparable. A favorable average can conceal failures in a rare but operationally important region; report those tails separately.

Established: controllability theory and numerical methods solve many specified model problems. Active: robust control with incomplete models, scalable optimization, constrained hardware, and trustworthy learning-based adaptation. Koch et al. (2022) review these connections. Liang et al. (2026) identify estimation, delay, and learning issues for monitored feedback. The literature check was 2026-10-11; the later source is explicitly a preprint.

Compare methods at equal information and cost

Gradient methods, derivative-free search, Bayesian optimization, and reinforcement learning use different information. An agent trained on millions of simulator calls should not be compared with a baseline allowed only a few experimental trials without reporting that difference. Include training, calibration, model identification, and deployment cost.

The key question is often generalization. Does a pulse optimized for nominal parameters work across measured drift? Does a learned policy exploit an unphysical simulator shortcut? Does a faster method reach the same accuracy under the same stopping criterion? The AI-assisted control assessment owns data-driven tool comparisons; this page owns the underlying control contract.

Starter project: robust pulses against a bounded uncertainty

Implement a driven two-level gate with detuning and amplitude uncertainty. Compare a simple analytical pulse, a nominal numerical optimum, and an ensemble-robust pulse. Fix total duration, peak amplitude, time resolution, and any bandwidth constraint.

Verify numerical gradients against finite differences on a small instance, check the ideal limiting case, and refine propagation independently of the optimization grid. Tune on one parameter set and test on a denser held-out set, including its boundaries. Report channel or gate error, worst-case performance, control effort, runtime, and the number of model evaluations.

A laptop suffices for the initial problem. Deliver reproducible pulses, propagation checks, and a robustness surface. Stop interpreting an improvement if it vanishes with a finer propagation step or if it violates the bandwidth or amplitude constraint. This baseline is not novel by itself; an original extension must introduce a justified new uncertainty, hardware constraint, or theoretical guarantee.

Where a graduate project can become research

One direction derives certificates for a restricted robust-control problem. Another compares model-based and record-based adaptation under the same observation budget. A third studies whether low-dimensional pulse parameterizations retain performance while improving transfer to hardware.

A valuable negative result identifies a reachable-set obstruction or a control-noise tradeoff that optimization cannot remove. Distinguish inability of one search algorithm to find a pulse from proof that no allowed pulse exists. Experimental validation needs independent calibration and interleaved baselines so that ordinary drift is not mistaken for algorithmic improvement.

Entry points

Read Koch et al. for the landscape, then reconstruct one algorithm and its assumptions from the stable owner. Reassess when a controllability theorem changes the feasible set or when held-out hardware data contradict the model used to optimize the controls.

References

• C. P. Koch et al., “Quantum optimal control in quantum technologies. Strategic report on current status, visions and goals for research in Europe,” EPJ Quantum Technology 9, 19 (2022). Preprint.
• W. Liang, J. Zhang, G. Guo, and D. Dong, “Measurement-Based Feedback of Open Quantum Systems: A Control-Theoretic Review and Tutorial,” arXiv:2609.06111v1 (2026). Preprint.