Readout Error

Readout error occurs when the classical measurement apparatus incorrectly identifies a qubit’s state, reporting $|0\rangle$ when the qubit is in $|1\rangle$ or vice versa. On current quantum processors, readout errors are typically 0.5% to 5% per qubit, making them one of the largest individual error sources. Unlike gate errors, which compound through the circuit, readout errors affect only the final measurement step, but they corrupt every shot of every circuit.

Readout error model

Readout errors are modeled by a confusion matrix (also called an assignment matrix) for each qubit:

$M = \begin{pmatrix} P(0|0) & P(0|1) \\ P(1|0) & P(1|1) \end{pmatrix}$

where $P(m|s)$ is the probability of measuring outcome $m$ given the true state $s$. For example:

$M = \begin{pmatrix} 0.98 & 0.03 \\ 0.02 & 0.97 \end{pmatrix}$

means the qubit in $|0\rangle$ is correctly read as 0 with 98% probability, and the qubit in $|1\rangle$ is correctly read as 1 with 97% probability. The asymmetry ($P(1|0) \neq P(0|1)$) is typical: $|1\rangle$ states are more likely to be misread because excited states can decay to $|0\rangle$ during the measurement process.

For $n$ qubits, the full readout error model is a $2^n \times 2^n$ confusion matrix, though it is often approximated as a tensor product of single-qubit matrices (assuming independent readout errors).

Physical causes

Superconducting qubits: Readout uses a dispersive measurement where a microwave probe pulse acquires different phase shifts depending on the qubit state. The returned signals for $|0\rangle$ and $|1\rangle$ form two partially overlapping distributions in the IQ plane (in-phase and quadrature components). Overlap between these distributions causes classification errors. $T_1$ decay during the measurement pulse (typically 0.5 to 2 microseconds) causes $|1\rangle$ to relax to $|0\rangle$ before readout completes.

Trapped ions: Readout uses state-dependent fluorescence. An ion in $|1\rangle$ (bright state) emits photons when illuminated; an ion in $|0\rangle$ (dark state) does not. Errors come from insufficient photon counts, off-resonant scattering, and photon detection noise. Trapped ion readout fidelities typically exceed 99.5%.

Mitigation techniques

Readout error mitigation (M3, matrix-free measurement mitigation): Characterize the confusion matrix by preparing known states and observing the measurement statistics. Then apply the inverse (or pseudo-inverse) of the confusion matrix to measured probability distributions. This is purely classical post-processing.

# Qiskit example of readout error mitigation
from qiskit_ibm_runtime import SamplerV2 as Sampler

# The Qiskit Runtime Sampler includes built-in
# readout error mitigation via the resilience_level option
sampler = Sampler(backend, options={"resilience_level": 1})

Twirling: Randomly flip qubits before measurement and classically undo the flips, symmetrizing the confusion matrix to make mitigation more robust.

Repeated readout: Measure each qubit multiple times and take a majority vote. This requires mid-circuit measurement capability and adds latency.

Why it matters for learners

Readout error is often the first type of error encountered when running circuits on real hardware. It is conceptually simple (a classical confusion between 0 and 1), relatively easy to characterize and mitigate, and provides a good entry point for understanding the broader topic of quantum error mitigation. Separating readout errors from gate errors in your mental model helps when diagnosing why a circuit produces unexpected results.

See also

• Measurement
• Quantum Error Mitigation
• Noise Model
• Fidelity
• Shot