Quantum error correction gets the headlines, but it is not what most quantum computers actually use today. Building a single fault-tolerant logical qubit takes hundreds or thousands of physical qubits, and that overhead is mostly out of reach for current machines. So engineers reach for a different, scrappier toolkit. It is called quantum error mitigation, and it is the reason anyone can extract a usable number from a noisy processor at all.
Correction versus mitigation
The two ideas are easy to confuse. Error correction detects and fixes mistakes while a computation runs, preserving the quantum state so a long algorithm can finish cleanly. Mitigation does no such thing. It accepts that every run is contaminated by noise, then uses statistics and extra measurements to estimate what the answer would have been on a perfect machine. Correction protects the computation. Mitigation cleans up the output afterward, or shapes the inputs in advance.
That distinction matters because mitigation does not require more qubits. It trades hardware overhead for runtime overhead. You pay by running your circuit many more times, sometimes by orders of magnitude, and by doing classical post-processing on the results. For machines with dozens or low hundreds of qubits and no spare capacity for full correction, that is a bargain.
The core tricks
A handful of techniques do most of the work, and they share a common logic: deliberately learn how noise distorts results, then back the distortion out.
- Zero-noise extrapolation. Run the same circuit at several noise levels by intentionally stretching gate operations or inserting redundant pairs that cancel out. Each version is noisier than the last. Plot the answers against noise strength, then extrapolate the curve back to the point of zero noise. The extrapolated value approximates the clean result.
- Probabilistic error cancellation. Characterize the noise on each gate precisely, then represent the ideal operation as a weighted combination of noisy ones, some with negative weights. Sampling from that combination and combining the outcomes statistically cancels the error. It is powerful but expensive, demanding very accurate noise models and a flood of samples.
- Measurement error mitigation. Readout is one of the noisiest steps in a quantum computer. By measuring known calibration states, you can build a matrix describing how often a 0 is misread as a 1 and vice versa, then invert it to correct the final distribution.
- Dynamical decoupling. Insert carefully timed pulses on idle qubits so that environmental noise averages out, a borrowed trick from nuclear magnetic resonance.
The catch: it does not scale forever
Mitigation has a hard ceiling, and researchers have studied it carefully. As circuits grow deeper and wider, the noise grows too, and the number of samples needed to maintain accuracy explodes. For probabilistic error cancellation, that sampling cost rises exponentially with the amount of noise in the circuit. At some point you would need more runs than the lifetime of the universe allows. Mitigation buys you a window of useful computation just beyond what raw hardware can manage, but it cannot carry you to large fault-tolerant algorithms. Only error correction can do that.
This is why the field treats mitigation as a bridge rather than a destination. It keeps current processors scientifically relevant while the qubit counts and coherence times needed for correction slowly arrive.
Why it still matters
The most cited demonstration of mitigation at scale came from a chemistry-style physics simulation on a superconducting processor with over a hundred qubits, where mitigation produced results that tracked careful classical benchmarks even in regimes that pushed classical methods hard. The point was not that the machine solved something impossible, but that mitigation could rescue answers from circuits deep enough to be interesting. That is the practical frontier right now.
For anyone evaluating a quantum result today, mitigation should change how the claim is read. A headline number often reflects heavy post-processing and thousands of repetitions, not a single clean run. That does not make it fake, but it does make context essential. Ask how many shots were used, which mitigation methods were applied, and how the noise was characterized. The honest answer reveals how close, or how far, the machine really is to doing the job unaided.
Error mitigation is the quiet workhorse of the present era. It will not deliver the fault-tolerant future, but it is keeping the science moving until the hardware catches up.
