Quantum error mitigation helps extract more useful information from a noisy quantum circuit without fully correcting every error. This guide explains the main techniques, their tradeoffs, and a repeatable benchmarking workflow for deciding whether mitigation produced a meaningful improvement.
Overview
Today’s quantum processors produce imperfect results because operations, qubit interactions, measurement, and environmental effects introduce noise. A circuit that is correct in an ideal simulator can therefore return biased expectation values or an inaccurate distribution when executed on hardware. Quantum error mitigation does not create a noise-free device. Instead, it uses circuit design, calibration data, repeated measurements, or statistical post-processing to estimate what the result might have been with less noise.
Mitigation is different from quantum error correction. Error correction uses additional physical qubits and fault-tolerant procedures to protect logical information. Mitigation generally works with the available noisy circuit and attempts to reduce the effect of errors on a particular observable or output. Its value depends on the circuit, backend, noise pattern, number of shots, and quality of the assumptions used.
Before choosing a technique, define the quantity you want to estimate. An expectation value for a variational algorithm may benefit from a different method than a full bitstring distribution. For a broader introduction to circuit size and execution cost, see the quantum circuit resource estimation guide. If your experiment uses VQE or QAOA, also establish the ansatz, objective function, and classical optimizer before adding mitigation; otherwise, improvements can be confused with optimization noise.
Core techniques at a glance
| Technique | Required assumptions or data | Circuit or sampling overhead | Noise sensitivity | Implementation difficulty | Good fit |
|---|---|---|---|---|---|
| Measurement error mitigation | Calibration of readout behavior, often from prepared basis states | Calibration circuits and post-processing; may grow with qubit count | Most useful for readout bias; less effective for errors during gates | Low to moderate | Shallow circuits, population estimates, and final measurement correction |
| Zero-noise extrapolation | A way to run related circuits at several effective noise levels and a model for extrapolation | Multiple circuit variants and additional shots | Sensitive to the noise-scaling method and extrapolation model | Moderate | Expectation values from short or medium-depth circuits |
| Probabilistic error cancellation | A noise model or experimentally estimated error channels | Can require substantial sampling because signed estimates increase variance | Highly sensitive to model mismatch and noise strength | High | Experiments where a detailed noise model and sampling budget are available |
| Symmetry verification | A known conserved quantity or valid subspace | Extra measurements, post-selection, or discarded shots | Depends on whether errors move results outside the known symmetry sector | Moderate | Hamiltonian simulations and algorithms with clear physical constraints |
These categories can be combined, but combination is not automatically better. For example, readout mitigation may be a sensible first step before testing zero-noise extrapolation, while adding several high-overhead methods can make statistical uncertainty difficult to interpret.
Checklist by scenario
For a new noisy quantum circuit
- Define the target. Record the observable, probability, energy estimate, or classification metric that matters. Avoid treating a visually different histogram as proof of improvement.
- Build an ideal reference. Run the same logical circuit in an ideal simulator when the system size permits. If exact simulation is too costly, use a smaller instance, a classically tractable special case, or a trusted reference calculation.
- Run an unmitigated hardware baseline. Keep the transpiled circuit, backend, shot count, qubit mapping, measurement layout, and execution conditions. The baseline must be reproducible.
- Inspect the circuit. Record depth, two-qubit gate count, measured qubits, and any inserted routing operations. A mitigation result is hard to assess when the compared circuits are not logically equivalent.
- Choose one primary method. Start with measurement error mitigation for a readout-dominated problem, symmetry verification when a valid subspace is known, or zero-noise extrapolation when the goal is an expectation value and several noise-scaled executions are practical.
- Repeat with matched sampling. Give the unmitigated and mitigated estimates comparable shot budgets where possible. If the mitigated result uses substantially more samples, report that cost.
- Quantify uncertainty. Use repeated batches, bootstrap intervals, analytical approximations, or another appropriate method. A closer point estimate is not persuasive if its uncertainty is much larger.
For readout error mitigation
Use calibration circuits that prepare known computational-basis states and measure how often each state is reported correctly. Check whether the calibration is feasible for the number of qubits involved; a full calibration can become expensive as the system grows. Consider a reduced or structured model only when its assumptions are documented and tested.
Apply the calibration consistently to the data and preserve the raw counts. Mitigated probabilities can be negative or fail to sum exactly to one when an inversion or approximate correction is used. That behavior is a warning to inspect the method and uncertainty, not an invitation to clip values silently.
For zero-noise extrapolation
Execute logically equivalent circuits at multiple effective noise levels. Noise scaling can involve gate folding or another framework-supported transformation, but the transformation should preserve the intended ideal computation. Fit the resulting observable to a simple extrapolation model and avoid using more model flexibility than the number of noise levels supports.
Compare several choices of noise levels and fit ranges where practical. If the extrapolated value changes substantially when one point is removed, the result is fragile. Report the measured values as well as the extrapolated estimate.
For symmetry verification
Write down the symmetry before running the experiment. Examples include a conserved parity or a Hamiltonian subspace, but the relevant constraint depends on the algorithm and problem. Measure the information needed to identify valid outcomes, then compare post-selected and unfiltered estimates. Discarding data can reduce bias while increasing variance, so track the retained-shot fraction.
For VQE workflows, mitigation should be evaluated alongside energy error, optimizer stability, and convergence behavior. The VQE implementation workflow provides a useful baseline for separating these effects.
What to double-check
- Logical equivalence: Confirm that transpilation, qubit mapping, parameter binding, and measurement interpretation match across all compared runs.
- Calibration age: Readout and noise characteristics can change. Record when calibration data was collected relative to the experiment.
- Shot allocation: Mitigation may require extra circuits or more samples. Compare total execution cost, not only the final error.
- Observable range: Check whether post-processing produces physically implausible values and explain how those values were handled.
- Statistical significance: Use uncertainty intervals or repeated trials before claiming that mitigation improved accuracy.
- Model dependence: Probabilistic error cancellation and extrapolation can fail when the assumed noise model or scaling behavior is inaccurate.
- Framework behavior: Verify the current SDK and provider documentation for supported transformations, calibration formats, backend restrictions, and result conventions. APIs and mitigation modules change over time.
For reproducibility, store source code, package versions, circuit representations, transpilation settings, calibration data, raw counts, processed results, and random seeds where applicable. The reproducible quantum computing experiments workflow can serve as a project checklist.
Common mistakes
Calling every post-processing step mitigation. Renormalizing a histogram or smoothing a curve does not establish that physical error was reduced. Explain the mechanism and its assumptions.
Comparing only with the raw hardware result. A mitigated estimate should also be compared with an ideal or trusted reference whenever possible. Otherwise, a change can be measured without knowing whether it moved in the right direction.
Using mitigation to hide an unsuitable circuit. Excessive depth, unnecessary two-qubit gates, and poor qubit connectivity can dominate the error budget. Optimize the circuit first; mitigation is not a substitute for resource reduction. The quantum gates reference can help review gate choices and syntax across frameworks.
Overfitting extrapolation. A high-degree fit can pass through a few noisy points while producing an unstable zero-noise estimate. Prefer a simple model and show sensitivity to reasonable alternatives.
Ignoring discarded shots. Post-selection may improve a conditional estimate while leaving too little data for a reliable conclusion. Always report how many shots remain.
When to revisit
Revisit this workflow whenever you change the backend, qubit layout, circuit depth, SDK version, transpilation settings, measurement scheme, or target observable. Also revisit it before a new seasonal planning or benchmarking cycle, especially if earlier calibration data is being reused. Provider interfaces and supported mitigation options can change after software updates, so confirm the current API rather than copying an older code example unchanged.
A practical return checklist is:
- Re-establish the ideal or trusted reference.
- Capture a fresh unmitigated baseline and its uncertainty.
- Measure the circuit’s current depth, gate mix, and shot budget.
- Choose the least complex mitigation method that matches the dominant error and target observable.
- Run matched comparisons, preserve raw data, and quantify uncertainty.
- Record accuracy improvement together with added circuit executions, shots, discarded samples, and runtime.
- Keep the conclusion narrow: state where mitigation helped, where it was unstable, and whether the extra cost was justified.
Used this way, quantum error mitigation becomes a controlled benchmarking tool rather than a last-minute correction. It helps developers understand noisy quantum circuits, compare hardware runs fairly, and decide when a result is strong enough to support the next experiment.