Autonomous Hamiltonian certification and changepoint detection

Modern quantum devices require high-precision Hamiltonian dynamics, but environmental noise can cause calibrated Hamiltonian parameters to drift over time, necessitating expensive recalibration. Detecting when recalibration is needed is challenging, especially since the very gates required for sophisticated verification protocols may themselves be miscalibrated. While cloud quantum computing services implement heuristic routines for triggering recalibration, the fundamental limits of optimal recalibration are not yet known. We develop efficient Hamiltonian certification and changepoint detection protocols in the autonomous setting, where we cannot rely on an external noiseless device and use only single-qubit gates and measurements, making the protocols robust to the calibration issues for multi-qubit operations they aim to detect. For unknown $n$-qubit Hamiltonians $H$ and $H_0$ with operator norm bounded by $M$, our certification protocol distinguishes whether $|H-H_0|F\geqε$ or $|H-H_0|F\leq O(ε/\sqrt{n})$ with sample complexity $O(nM^2\ln(1/δ)/ε^2)$ and total evolution time $O(nM\ln(1/δ)/ε^2)$. We achieve this by evolving random stabilizer product states and performing adaptive single-qubit measurements based on a classically simulable hypothesis state. Extending this to continuous monitoring, we develop an online changepoint detection algorithm using the CUSUM procedure that achieves a detection delay time bound of $O(nM\ln(M\mathbb{E}\infty[T])/ε^2)$, matching the known asymptotically optimal scaling with respect to false alarm run time $\mathbb{E}\infty[T]$. Our approach enables quantum devices to autonomously monitor their own calibration status without requiring ancillary systems, entangling operations, or a trusted reference device, offering a practical solution for robust quantum computing with contemporary noisy devices.

💡 Research Summary

The paper addresses a pressing practical problem in quantum computing: how a quantum device can autonomously detect when its own Hamiltonian parameters have drifted enough to require recalibration, without relying on any external, perfectly calibrated hardware or entangled ancillae. Existing certification and learning methods typically assume access to a trusted reference system, multi‑qubit entangling gates, or sophisticated control that themselves would be unreliable if the device is already mis‑calibrated. To break this circular dependency, the authors propose “autonomous” protocols that use only single‑qubit state preparation and measurement—operations that are generally the most robust on current devices.

The first contribution is an efficient Hamiltonian certification protocol. The device is supplied with random stabilizer product states (tensor products of single‑qubit Pauli eigenstates). Each state is evolved under the unknown device Hamiltonian H for a carefully chosen short time t, then each qubit is measured in a basis that is adaptively selected based on a classical simulation of the ideal Hamiltonian H₀. This adaptive measurement maximizes the information gain about the fidelity between the evolved unknown state and the simulated reference state. By analyzing the relationship between fidelity loss and the Frobenius norm of the Hamiltonian difference, the authors show that with evolution time t = Θ(ε/(nM)) the fidelity deviation scales as t²‖H−H₀‖_F². Applying Chernoff‑Hoeffding bounds yields a sample complexity N = O(n M² log(1/δ)/ε²) and a total evolution time T = O(n M log(1/δ)/ε²). Importantly, both scale linearly with the number of qubits n, and the protocol requires no entangling gates, ancillae, or knowledge of the exact time‑evolution operator e^{−iH₀t}.

The second contribution extends the certification routine to continuous monitoring. The authors model a time‑varying Hamiltonian sequence {H_i} that at some unknown changepoint ν switches from being ε‑close to H₀ to being ε‑far. They feed the single‑sample outcome of the certification step into a classical CUSUM (cumulative sum) statistic, which accumulates log‑likelihood ratios between the “pre‑change” and “post‑change” hypotheses. When the CUSUM statistic exceeds a threshold h, a changepoint is declared and the most likely changepoint location is estimated as the argmax of the statistic. By choosing h≈log(E_∞