Benchmarking the readout of a superconducting qubit for repeated measurements

Readout of superconducting qubits faces a trade-off between measurement speed and unwanted back-action on the qubit caused by the readout drive, such as $T_1$ degradation and leakage out of the computational subspace. The readout is typically benchmarked by integrating the readout signal and choosing a binary threshold to extract the "readout fidelity". We show that readout fidelity may significantly overlook readout-induced leakage errors. Such errors are detrimental for applications that rely on continuously repeated measurements, e.g., quantum error correction. We introduce a method to measure the readout-induced leakage rate by repeatedly executing a composite operation - a readout preceded by a randomized qubit-flip. We apply this technique to characterize the readout of a superconducting qubit, optimized for fidelity across four different readout durations. Our technique highlights the importance of an independent leakage characterization by showing that the leakage rates vary from $0.12\%$ to $7.76\%$ across these readouts even though the fidelity exceeds $99.5\%$ in all four cases.