Two-Qubit Gate Fidelity

Two-qubit gates are the hardest operations to execute well in quantum computing. Single-qubit gates routinely exceed 99.9% fidelity on leading platforms; two-qubit gates are slower, more sensitive to noise, and typically lag behind by one to two orders of magnitude in error rate. Because almost every useful quantum algorithm requires many two-qubit gates, two-qubit gate fidelity is the single most important hardware metric for estimating how deep a circuit a device can reliably execute.

Why two-qubit gates are harder

Two-qubit gates require physical coupling between qubits. On superconducting processors, this coupling is mediated by a shared microwave resonator or a tunable coupler. To entangle two qubits, the interaction must be turned on for the duration of the gate, which means both qubits are simultaneously exposed to the coupling channel and to any noise it introduces. The interaction also affects neighboring qubits through residual ZZ coupling (always-on cross-talk), adding phase errors to qubits that were supposed to be idle.

The gate itself takes longer than a single-qubit gate. A typical superconducting two-qubit gate (CNOT or CZ via cross-resonance or parametric coupling) runs for 100-400 ns, compared to 10-50 ns for a single-qubit gate. More time means more decoherence: both qubits accumulate $T_1$ and $T_2$ errors during the gate. Trapped-ion two-qubit gates via the Molmer-Sorensen interaction take 10-100 microseconds, which is long enough that careful dynamical decoupling or cooling may be needed to maintain fidelity.

How two-qubit gate fidelity is measured

Interleaved randomized benchmarking (IRB) is the most common protocol. A standard randomized benchmarking experiment establishes a reference decay rate for random Clifford circuits on the qubit pair. A second experiment interleaves the target two-qubit gate between every random Clifford. The ratio of the two decay rates isolates the error contribution of the specific gate, giving an average gate infidelity that is independent of state preparation and measurement errors.

Process tomography fully characterizes the quantum process implemented by the gate, reconstructing a $4 \times 4$ process matrix (for two qubits) and comparing it to the ideal. The process fidelity is the overlap between the two matrices. Process tomography requires exponentially many measurements in qubit count, making it impractical beyond a few qubits, but it provides richer diagnostic information than IRB.

Cycle benchmarking and direct randomized benchmarking variants characterize the fidelity of simultaneously executed two-qubit gates across a full processor, capturing crosstalk effects that single-pair IRB misses.

Reported values across platforms

As of 2024-2025, leading platforms report two-qubit gate fidelities in the range of:

• Superconducting qubits (IBM, Google): 99.0-99.7% for individual pairs; simultaneous gate fidelities typically 0.3-1% lower.
• Trapped ions (Quantinuum, IonQ): 99.5-99.9%, benefiting from long coherence times and all-to-all connectivity, though absolute gate times are much longer.
• Neutral atoms (QuEra, Atom Computing): 99.0-99.5%, improving rapidly as Rydberg gate protocols mature.

These numbers are not directly comparable without knowing which benchmarking protocol was used, which qubit pairs were selected, and whether the measurement includes crosstalk from neighboring operations.

Why it matters for learners

Two-qubit gate fidelity determines how many gates a circuit can contain before the output becomes dominated by noise. A rough rule is that a circuit with $N$ two-qubit gates, each at fidelity $F$, has total fidelity approximately $F^N$. At $F = 0.99$ and $N = 100$, total fidelity is roughly 37%. At $F = 0.999$ and $N = 100$, it is 90%. This is why the difference between 99% and 99.9% is not incremental but transformative in terms of the algorithms that become feasible. The fault-tolerance threshold also depends critically on two-qubit gate fidelity: the surface code threshold is around 99% (1% error per gate), meaning hardware operating near that boundary has almost no margin for other error sources.

See also

• Gate Fidelity
• Randomized Benchmarking
• Crosstalk
• CNOT Gate
• Quantum Volume
• Quantum Error Rate