There are two ways to deal with the errors that plague a quantum computer. One is to fix them outright, spreading a single unit of information across many physical qubits so mistakes can be detected and reversed. That approach, error correction, is powerful and expensive, and it is not ready for the machines companies run today. The other way is cheaper and stranger: let the errors happen, measure how badly they corrupt the answer, and mathematically walk the result back toward what a perfect machine would have produced. This second approach is called error mitigation, and it is quietly doing much of the heavy lifting in the current era of quantum computing.
Correction versus mitigation
The distinction matters. Error correction protects a computation in real time and can, in principle, run an arbitrarily long circuit reliably. But it demands a crushing overhead, often hundreds or thousands of physical qubits per protected logical qubit. Mitigation buys no such guarantee. It cannot make an individual run of a circuit correct. What it can do is improve the average you extract after running that circuit many times. Since quantum programs already rely on repeated sampling, this fits the workflow neatly. You are not trying to save any single shot. You are trying to recover the right expectation value from a cloud of noisy ones.
Zero-noise extrapolation
The most widely used trick is zero-noise extrapolation. The idea sounds almost too simple. If you cannot turn the noise off, turn it up on purpose. Run your circuit at its native error rate, then run it again with the noise deliberately amplified, say by stretching gate pulses or by inserting pairs of gates that cancel out logically but still accumulate real errors. Do this at several noise levels and you get a curve showing how your answer degrades as noise climbs. Then you extrapolate that curve backward, past the point you can actually reach, to estimate what the result would be at zero noise. It is a bit like measuring how fast paint fades in strong, medium, and weak sunlight, then projecting the original color.
The method has obvious limits. Extrapolation assumes the noise behaves in a predictable, well-shaped way, and pushing too far past the measured points can produce nonsense. But when the underlying error model cooperates, zero-noise extrapolation can sharpen a result dramatically at a modest cost in extra circuit runs.
Cancelling errors on paper
A more aggressive technique is probabilistic error cancellation. Here the machine's noise is carefully characterized in advance, so you know roughly how each gate distorts the ideal operation. You then build a recipe that mixes together several slightly altered versions of your circuit, some of which are weighted with negative signs. Averaged together with those signs, the distortions cancel and the ideal answer emerges. The catch is severe: the sampling cost grows exponentially with the amount of noise you are trying to cancel, so a noisier machine forces you to run vastly more shots. It works, but it does not scale forever.
Other tricks round out the toolkit. Readout mitigation corrects for the fact that measuring a qubit sometimes reports the wrong value, using a calibration of how often each misread happens. Dynamical decoupling slips harmless pulses into idle moments to shield qubits from slow drift. Symmetry verification throws away runs that violate a physical law the answer must obey, such as a conserved particle count.
Why it matters now
Error mitigation is not a permanent solution, and nobody in the field pretends otherwise. The sampling overhead climbs steeply as circuits grow, and at some point the number of repetitions becomes impractical. That ceiling is exactly why full error correction remains the long-term goal. But mitigation is what makes the machines of the present useful rather than merely impressive. It is the reason researchers can run chemistry and materials simulations on noisy hardware and get numbers worth comparing against classical methods.
The likely future is a blend. As early error correction arrives, mitigation techniques will layer on top, cleaning up whatever residual noise the correction leaves behind. For now, these statistical rewinds are the bridge across the gap between the hardware we have and the hardware we want. They do not make a quantum computer honest. They make it honest enough, on average, to trust.
