This paper proposes a practical way to calibrate complicated control pulses for two-qubit gates by focusing on only a few directions that actually affect gate error. The authors show that, although an optimal control pulse can have many adjustable pieces, the gate’s fidelity is sensitive to only a small, low-rank set of pulse distortions. They use this observation to speed up closed‑loop experimental calibration and demonstrate it on a neutral‑atom entangling gate.

The key idea is to compute the Hessian of the gate fidelity. The Hessian is a matrix that tells you how the fidelity changes for small changes to each control parameter. Diagonalizing that matrix shows a small number of eigenvectors with nonzero eigenvalues. Those eigenvectors define a principal subspace of pulse changes that matter to leading order; changes orthogonal to that space do not change fidelity at quadratic order. For the Rydberg-mediated controlled‑Z (CZ) gate studied here, the authors find only five nonzero Hessian eigenvalues. They trace that small number to two independent leakage channels and a nonlinear phase error that can occur in the target unitary.

In the experiment the team applied this “low‑rank Hessian optimization” to an amplitude‑robust CZ gate on metastable‑state 171Yb nuclear‑spin qubits. They verified the predicted sensitive directions and then optimized only within that small subspace using closed‑loop feedback from the experiment. The optimized two‑qubit gate reaches a raw fidelity of 0.9959(2). After postselecting on no detected atom loss, the fidelity rises to 0.99902(7). They also report that the optimized gate performance is essentially unchanged when the laser power is varied by up to 20 percent.

The method matters because high‑fidelity, fast entangling gates are essential for scaling quantum processors and for lowering the cost of error correction. Scanning or tuning every parameter in a high‑dimensional pulse is slow and often stalls. By finding the small set of directions that actually affect fidelity and only searching there, calibration becomes much faster and avoids the slow convergence or fidelity floors seen in other strategies. The authors also show in simulations that optimizing the Hessian eigenvectors converges rapidly compared with several other commonly used approaches.