Two-qubit gate fidelity

Fidelity is reported as a number like 99.4% (equivalently, an error rate of 6×10⁻³). It quantifies, on average, how faithfully the physical entangling gate (CZ, ECR, or a Mølmer–Sørensen gate) implements its ideal unitary. Because almost every useful circuit is dominated by entangling operations, this figure — together with how many entangling gates your compiled circuit needs — largely determines whether a result is signal or noise.

Common methods differ materially. Randomized benchmarking (RB) averages over random Clifford sequences and is insensitive to state-preparation and measurement errors. Interleaved RB isolates one specific gate. Cross-entropy benchmarking (XEB) derives fidelity from sampling statistics of random circuits. Layered/holistic benchmarks (such as IBM's error-per-layered-gate, EPLG) measure gates operating together in a realistic layer — usually yielding worse (more honest, for applications) numbers than isolated-pair RB.

Two fidelity figures are not directly comparable when: (1) the aggregation differs — a device-median is not a best-pair, and vendors sometimes publish the best pair; (2) the method differs — RB, XEB, and layered benchmarks answer different questions; (3) the gates differ — a CZ and an MS gate have different durations and error mechanisms; (4) the date differs — calibration drifts daily. Always check the record behind the number: every fidelity figure in the QPU137 catalog states its method and aggregation where the vendor disclosed them, and says so where they did not.