Quantum error mitigation can make results from noisy quantum hardware more useful without requiring a full fault-tolerant error-correction system. This guide explains the main techniques, their assumptions and tradeoffs, and a reproducible experiment template for comparing raw, mitigated and ideal results.
Overview
Quantum hardware produces errors through imperfect gates, measurement, qubit relaxation, dephasing, crosstalk and other device-specific effects. A simulator can model some of these effects, but a hardware run adds calibration drift, queue-time changes and connectivity constraints. The result is that a circuit’s measured expectation value may differ from the value produced by an ideal simulation.
Error mitigation aims to estimate a better result from noisy executions. It does not remove errors from the quantum state in the same way that quantum error correction seeks to do. Instead, it uses additional measurements, circuit transformations, calibration data or problem-specific constraints to reduce the effect of noise on a chosen observable.
That distinction matters. Mitigation may improve an energy estimate, correlation, success probability or expectation value, but it can also increase sampling cost, amplify statistical uncertainty or produce an estimate outside a physically plausible range. Treat each result as an estimate with an uncertainty and a documented method—not as a guaranteed correction.
Technique comparison
| Technique | What it targets | Additional work | Best fit | Main limitation |
|---|---|---|---|---|
| Readout error mitigation | Measurement and assignment errors | Calibration circuits and post-processing | Shallow circuits where measurement error is a major contributor | It does not correct errors that occurred during state preparation or gates |
| Zero-noise extrapolation | Errors accumulated during circuit execution | Run noise-scaled versions of the circuit and extrapolate toward lower noise | Expectation-value experiments with a controllable noise-scaling method | Extrapolation can magnify statistical noise and depends on a useful noise model |
| Probabilistic error cancellation | Known or estimated noise processes | Characterization, circuit sampling and weighted post-processing | Small experiments where accuracy is more important than execution efficiency | Sampling overhead can grow quickly, especially for deeper circuits |
| Symmetry checks and post-selection | Results that violate known problem symmetries | Measure or infer a conserved quantity and reject inconsistent outcomes | Problems with a reliable parity, particle-number or other conserved quantity | Discarding data can increase variance and may bias results if assumptions are wrong |
These methods can be combined, but combining them is not automatically better. For example, readout mitigation may be applied before analyzing a zero-noise extrapolation series, yet the calibration and uncertainty implications should be tested rather than assumed. For a broader decision framework, see Quantum Error Mitigation Techniques: A Practical Decision Guide for Developers.
Checklist by scenario
Scenario 1: Measurement results look wrong, but the circuit is shallow
- Run the circuit with a known input state, such as an all-zero basis state, and record the observed bitstrings.
- Repeat with additional basis states if your mitigation workflow supports a calibration matrix or assignment model.
- Check whether the calibration was collected on the same qubits, measurement basis and approximate time window as the experiment.
- Compare raw and mitigated distributions, not only one final expectation value.
- Inspect whether mitigation creates negative quasi-probabilities or unusually large corrections.
Readout error mitigation is a reasonable first test when the circuit has few gates and the main discrepancy appears at measurement. It is not a substitute for investigating gate errors, state-preparation errors or an incorrect observable definition.
Scenario 2: You are estimating an expectation value on hardware
- Define the observable precisely, including its basis decomposition if it contains multiple Pauli terms.
- Choose a noise-scaling method that changes the effective noise while preserving the intended ideal operation as closely as possible.
- Run the original circuit and each scaled circuit with the same shot budget where practical.
- Repeat the series sufficiently to estimate run-to-run variation, especially if the device or queue conditions may change.
- Fit a simple extrapolation model first. Compare its result with the raw value and with an ideal or high-fidelity simulator reference.
- Report the fit, scaling factors, shot counts and uncertainty instead of publishing only the extrapolated number.
This is the core pattern for a zero-noise extrapolation tutorial: deliberately create several noisy estimates, model how the observable changes with noise, and estimate its value near the zero-noise limit. The extrapolated value is model-dependent, so use it as a diagnostic estimate rather than proof that the hardware produced an ideal state.
Scenario 3: A known symmetry can validate or filter outcomes
- Write down the symmetry and why the algorithm should preserve it under ideal execution.
- Determine whether the symmetry can be measured directly, inferred from the bitstring, or evaluated through additional circuits.
- Measure the fraction of shots that pass the check before applying post-selection.
- Compare estimates with and without rejected shots.
- Test the same procedure on an ideal simulation and a noisy simulation to identify whether the filter removes errors or simply removes a large part of the sample.
Symmetry checks are especially useful in chemistry and optimization workflows when the expected particle number, parity or another invariant is known. They are unsafe when the assumed symmetry is not actually preserved by the ansatz, Hamiltonian, constraints or implementation.
Scenario 4: You need a reproducible comparison across simulator and hardware
- Freeze the circuit structure, qubit mapping, transpilation or compilation settings and observable definition.
- Run an ideal simulator to establish a reference value.
- Run a noisy simulator with an explicitly documented noise model, if available.
- Run the same logical experiment on hardware using a fixed shot budget and record backend, device settings and execution date.
- Apply one mitigation technique at a time: raw, readout-mitigated, extrapolated or post-selected.
- Compare absolute error against the ideal reference, confidence intervals, execution count and rejected-shot rate.
Simulator choice affects interpretation. A statevector simulator is useful for an ideal reference, while a noisy or tensor-network simulator may support different circuit sizes and noise assumptions. The practical differences are discussed in Quantum Simulators Compared.
What to double-check
Use an explicit experiment record
For every run, save the circuit or circuit hash, number of qubits, observable, basis rotations, shots, qubit layout, compilation settings, mitigation method, calibration data and software versions. SDK interfaces and provider-specific mitigation integrations change over time, so record the exact environment rather than relying on a notebook title.
Separate bias from variance
Mitigation often attempts to reduce bias—the difference between the noisy estimate and the ideal value—but may increase variance. A result that is closer to the ideal value in one run may be less reliable overall. Use repeated trials or resampling where appropriate, and include uncertainty bars or confidence intervals.
Check the comparison baseline
An ideal simulator reference is valuable, but it may not scale to the same circuit size as hardware. If you change the circuit through transpilation, noise scaling or basis decomposition, make sure the ideal comparison uses the corresponding logical operation. When an exact ideal reference is unavailable, compare methods against smaller circuits or analytically known cases.
Review sampling requirements
Readout calibration consumes executions. Zero-noise extrapolation requires multiple circuit variants. Probabilistic error cancellation can require a substantial number of samples, and post-selection discards shots. Before choosing a technique, estimate the total execution count and decide whether the expected improvement justifies the cost.
Common mistakes
- Calling mitigation error correction: mitigation improves an estimate under assumptions; it does not create a fault-tolerant logical qubit.
- Using an ideal simulator as the only test: include a noisy simulator or small hardware experiment to test whether the method behaves as expected.
- Changing several variables at once: changing shots, circuit depth, qubit mapping and mitigation method makes the result difficult to attribute.
- Extrapolating from too few or poorly separated noise levels: a fit can look precise while being unstable outside the observed range.
- Ignoring calibration drift: readout data collected at a different time or on different qubits may not describe the experiment.
- Reporting only the best number: show raw, mitigated and ideal results, along with the method and uncertainty.
- Applying post-selection without checking acceptance: a low acceptance rate can make the estimate expensive and statistically fragile.
- Assuming an SDK example is version-independent: verify current API behavior, backend support and required primitives before adapting code from a tutorial.
When to revisit
Revisit your error-mitigation workflow whenever the circuit, backend or software environment changes. A new qubit layout, transpiler configuration, shot budget, ansatz depth or observable can change which error source dominates. Recalibrate readout mitigation when the relevant device conditions change, and rerun baseline tests after moving from simulation to hardware.
As a practical maintenance checklist, review the workflow before a new project phase, after an SDK or provider integration update, and whenever a previously stable result shifts unexpectedly. Recheck the assumptions behind symmetry filters and noise scaling at the same time. Keep a small regression suite containing an ideal case, a noisy simulation and one hardware-compatible circuit so that changes can be detected quickly.
For a complete experiment, record these three outputs side by side:
- Ideal: the simulator or analytical reference.
- Raw: the unmitigated hardware or noisy-simulator estimate.
- Mitigated: the estimate after one clearly identified technique.
Then compare accuracy, uncertainty, shot cost and stability across repeated runs. That simple discipline makes quantum error mitigation a measurable engineering workflow rather than a last-minute post-processing step. If the experiment is part of a variational algorithm, connect the mitigation results to optimizer behavior and total cost; the end-to-end VQE workflow provides a useful companion process for that transition.