Single qubit · simulated control

Calibrate a quantum gate.

Replay saved calibration runs to see how a better control pulse brings a quantum operation closer to its target. Every number here is a real evaluation from the committed held-out benchmark.

Replay of saved runs · 135 held-out runs

Replay of saved run40 / 40 evaluations

Gate infidelity · lower is better

–→–

Nominal pulse → best pulse found so far

Evaluations shown

40 / 40

Same budget for every method

Computation time

–

Measured on an Apple M5 Mac · includes optimizer overhead

Finding a better pulse

Saved traces

Bayesian optimizationGP-EI, nominal startPowellRandom searchNominal pulse

Best gate infidelity so far · logarithmic scale · values below 1e-7 drawn at the floor

Control pulse comparison

2 SEGMENTS · T = 1

Nominal pulse

x controly control

Best pulse found

x controly control

Amplitude and time are dimensionless. Controls are bounded to [−π, π]; each segment holds constant x and y controls.

Compare the methods

Method	Gate error	Trials	Reached 1e-3

What calibration does

THE METHOD

Define the operation

Choose a target gate and model a control error in a simulated qubit.

Search the control settings

Evaluate candidate pulses and use the observations to choose the next trial.

Check the whole gate

Measure similarity to the target operation, then test new error settings. Gate fidelity is not a computation success rate.

Held-out results

Method	Median final infidelity (IQR)	Reached 1e-3	Median calls-to-target	Median time / run

Robustness of a fixed pulse

GP-EI · MEDIAN OVER SEEDS

Rows: offset the pulse was calibrated at. Columns: offset it is evaluated at. The diagonal is adaptation; off-diagonal cells are transfer, and the last row is the uncalibrated nominal pulse.