A noisy quantum processor does not usually return the expectation that the ideal circuit predicts. Error mitigation tries to estimate that ideal expectation from experiments on the noisy device. Here is one elementary method: deliberately run related experiments at different noise levels, then extrapolate their measured expectations back to zero noise. The method does not repair an individual quantum state, and its conclusion depends on how the noise is scaled.
What is being estimated?
Let be the ideal output of a circuit and let be a Hermitian observable. The target is the real number . Suppose a reproducible family of experiments indexed by a small noise strength has expectation , with . This equality is an assumption about the whole experiment, including preparation and readout: an error that remains when is not removed by extrapolating this parameter.
Definition 1.7.1
(Two-point zero-noise extrapolation)
Fix a noise multiplier and a noise strength such that both and are accessible. The two-point extrapolated value is
In practice and are replaced by measured sample means, so the same weighted formula is a random estimator.
Concept lensAlgebraic
Extrapolation as interpolation evaluated at zero
The formula is a Lagrange interpolation. The unique degree-one polynomial through and is , and evaluating at gives weights on and on . The estimator is therefore exact whenever the response is an affine function of , and the weights sum to one so that a constant response is reproduced.
For fixed , assume that, near zero,
Proposition 1.7.2
(Bias of the two-point estimate)
If with for small , and is fixed, then
Proof
(Canceling the linear term)
Substitute the expansion at and at . The constant terms give , which divided by is . The linear terms give . The quadratic terms give , which divided by is . The remainder is , bounded in absolute value by , which is for fixed .
Proof X-ray
Where the cancellation comes from
Two hypotheses carry the argument. Smoothness of to second order lets the response be treated as a polynomial with a controlled remainder; without it there is no linear term to cancel. The weights and are chosen so that they sum to one, which preserves , and so that , which kills the slope. The curvature is untouched, because the weight on the quadratic term is ; it survives as the residual . The remainder estimate uses that is fixed as : if grew with the constant in would not be uniform. A curved response therefore does not meet the extrapolating line at the true intercept.
This calculation justifies cancellation of first-order bias, not an assertion that all noise disappears.
A calculable qubit example
Take ideal input , no intended gates, and measure . The ideal expectation is . Suppose each of two successive noise locations applies the qubit depolarizing channel
for a density matrix . Since , one application multiplies the expectation by . Thus .
Worked example 1.7.3
(Doubling both noise locations at )
Assume a second experiment scales both locations to while leaving everything else unchanged; this requires . It has , so
At the noisy expectation is , the doubled-noise expectation is , and the extrapolated value is . This agrees with , since here and . The linear error is gone, but the residual bias is . This model is deliberately simple: increasing gate repetitions in a real circuit does not automatically replace every noise channel by while preserving the ideal computation. That physical scaling must be justified or calibrated separately.
Common mistake
(Repeating a noise location does not double )
Applying twice in a row multiplies the expectation by , so , not . At the effective strength is , not . Repeating both locations twice gives , whereas . The two-point formula must then use the true multiplier : , which equals . Using with the repeated-gate data gives , a larger error.
Worked example 1.7.4
(An affine response is extrapolated exactly)
Suppose a single noise location with one depolarizing channel acts on , so . Then , and for any
The bias vanishes exactly, as the interpolation view predicts. The edge case shows that the residual error of the method is entirely a statement about curvature of the response, not about the choice of .
Sampling cost and interpretation
Assume independent shots in the two experiments, each measuring with outcomes .
Proposition 1.7.5
(Variance of the shot-based estimator)
Let be the sample means of independent outcomes with true means and . Then has expectation and variance
Proof
(Variances of independent sample means)
A outcome has , so . The sample mean of independent copies has variance . The two experiments are independent, so the variance of is , which is the stated expression. Linearity gives the expectation.
At a fixed total shot budget, mitigation can reduce systematic bias while increasing statistical uncertainty. Negative weights can also produce an estimate outside even though each measured sample mean lies inside that interval; the extrapolated number is not itself the outcome of a measurement.
Worked example 1.7.6
(Shot cost at )
With and , the unmitigated sample mean from shots has variance . For the extrapolated estimator with ,
For the standard deviations are about (extrapolated) and (unmitigated), a factor of about . The bias reduction from to is bought with this larger statistical error.
Counterexample mode
Fixed readout errors survive
Suppose a true outcome is reported as with probability , and a true is reported as with probability , independently of . Then the reported expectation is
Indeed the true outcome probabilities are ; applying the two misreporting probabilities gives this affine response. Write and , where . The extrapolated value is , which generally differs from . Scaling the modeled circuit noise does not remove fixed readout errors. A useful mitigation claim must say which errors were varied, which were held fixed, and how uncertainty was assessed.
Practice with solutions
Quick check
(For and noise multiplier , calculate the two-point extrapolation and its bias.)
Apply the defining formula and cancel the linear term.
Solution · Answer
, with bias . Check: the linear terms give and the quadratic terms give , which halves to .
Quick check
(In the qubit example, take . Calculate both expectations and the extrapolation. Is the residual zero?)
Use and .
Solution · Answer
, , and ; its bias is , so the residual is not zero. It agrees with .
Quick check
(Why does choosing extremely close to one make finite-shot estimates delicate?)
Look at the weights and .
Solution · Answer
The magnitudes of both weights contain , amplifying sampling fluctuations; independent variances are multiplied by the squares of the weights. This does not rule out a particular optimum away from one, which depends on the noise response and shot budget.