Error Mitigation on Noisy Hardware
Techniques that reduce the effect of noise on near-term quantum simulations without the full overhead of error correction.
Mitigation versus correction
Quantum error correction detects and fixes errors in real time using redundant encoding, but it demands many physical qubits per logical qubit, beyond near-term hardware. Error mitigation instead accepts noisy results and post-processes many runs to estimate what the noiseless answer would have been. It reduces bias in expectation values without extra qubits, at the cost of more measurements.
Zero-noise extrapolation
Zero-noise extrapolation deliberately increases the noise (by stretching gate times or inserting identity pairs), measures the observable at several noise levels, and extrapolates back to the zero-noise limit. It works when the observable depends smoothly on the noise strength, and it is one of the most widely used mitigation techniques.
Other methods
- Probabilistic error cancellation: invert a characterized noise model by sampling, unbiased but variance-costly.
- Readout-error mitigation: correct measurement errors with a calibration matrix.
- Symmetry verification: discard runs that violate a conserved quantity such as particle number.
- Dynamical decoupling: insert pulse sequences to suppress idle-qubit decoherence.
The fundamental cost
Mitigation is not free. The sampling overhead to cancel noise grows exponentially with circuit depth and error rate, because the variance of the corrected estimator blows up. This means mitigation extends the reach of noisy devices modestly but cannot substitute for error correction at large scale, a limit made precise by recent theoretical analyses.
Where it helps
For shallow simulation circuits, small VQE chemistry instances, short-time spin dynamics, error mitigation can turn otherwise unusable noisy results into meaningful expectation values, and it is standard practice in near-term experiments. It is a bridge technology: valuable now for pushing pre-fault-tolerant devices to their limit, but not the path to the large, long-horizon simulations that require genuine error correction. Understanding its exponential overhead keeps expectations calibrated about what noisy hardware can deliver.