Quantum error mitigation tries to improve estimates produced by noisy quantum hardware without fully encoding the computation inside a fault-tolerant error-correcting code. It trades extra circuit executions, calibration information and modelling assumptions for reduced bias in selected observables.
The central object is often not the noisy output state itself but an expectation value μ=Tr(Oρideal) that we want to estimate. A noisy device returns μ(λ), where λ represents an effective noise scale. Error-mitigation methods combine several noisy estimates or reweight noisy circuit samples so the final estimator better approximates μ(0).
This guide develops zero-noise extrapolation, Richardson cancellation, probabilistic error cancellation, quasiprobability sampling cost, symmetry verification, measurement-error mitigation, virtual-distillation ideas and bias–variance trade-offs. Guide 25 owns fault tolerance; this guide deliberately does not claim mitigation is a substitute for scalable quantum error correction.
Characterise noise → choose a mitigation model → run extra circuits or reweight samples → reduce bias → pay increased variance and calibration cost → validate against a trusted reference.
1. Why noisy expectation values are biased
Suppose the ideal circuit prepares ρ and we want μ=Tr(Oρ).
A noisy implementation applies channel 𝒩, giving
μ_noisy=Tr[O𝒩(ρ)].
The sampling error of estimating μnoisy can be reduced by more shots, but the systematic difference μnoisy−μ remains.
Error mitigation targets that systematic bias while accepting that the estimator variance often increases.
2. Bias and variance are different errors
For estimator μ̂, mean-squared error decomposes as
MSE=Bias(μ̂)²+Var(μ̂)
when estimating one scalar target.
A mitigation protocol can decrease bias and still worsen MSE if its sampling variance explodes.
The right comparison is therefore total estimation cost at a target precision, not just whether the corrected central value moved closer to a benchmark once.
3. Zero-noise extrapolation
Suppose one can execute the same logical circuit at several effective noise scales λ, c₂λ, c₃λ,… while keeping the ideal unitary unchanged.
Expand an observable expectation around λ=0:
μ(λ)=μ(0)+a₁λ+a₂λ²+a₃λ³+…
Combine measured values with coefficients chosen to cancel the leading powers of λ.
Temme, Bravyi and Gambetta proposed this Richardson-extrapolation strategy for short-depth noisy circuits. [1]
4. First-order Richardson extrapolation
Using μ(λ) and μ(cλ), choose coefficients γ₁,γ₂ satisfying
γ₁+γ₂=1;γ₁+cγ₂=0.
Solving gives
γ₁=c/(c−1), γ₂=−1/(c−1).
Thus
μ_ZNE=[c μ(λ)−μ(cλ)]/(c−1)=μ(0)+O(λ²).
5. Worked ZNE example
Suppose c=2, μ(λ)=0.72 and μ(2λ)=0.61.
Then
μ_ZNE=2(0.72)−0.61=0.83.
The extrapolated value can lie outside the interval spanned by the noisy data because it is estimating an intercept, not averaging observations.
If the observable spectrum lies in [−1,1], 0.83 is physically admissible. An extrapolated value 1.4 would signal model/statistical failure or need for constrained post-processing rather than a valid expectation.
6. How to scale noise
Real hardware rarely exposes a knob labelled λ.
Noise can be amplified using techniques such as:
- pulse stretching;
- gate folding, for example U→UU†U;
- inserting identity-equivalent gate pairs;
- controlled changes to idle duration.
The method must preserve the ideal logical operation while changing the relevant noise predictably enough for extrapolation.
7. Gate folding
Replace ideal gate U by
U(U†U)^m.
Ideally this is still U. Physically it uses more noisy gates and therefore increases exposure to noise.
Local folding applies this gate by gate; global folding folds a whole circuit segment.
If coherent errors interfere differently after folding, the effective noise may not scale linearly with circuit length. ZNE must validate the scaling model empirically.
8. Higher-order Richardson cancellation
With K noise scales ciλ, choose coefficients γi satisfying
Σ_iγ_i=1
and
Σ_iγ_i c_i^m=0, m=1,…,K−1.
The residual bias begins at O(λK) under the analytic expansion assumptions.
However, coefficients can become large and alternating, greatly amplifying statistical uncertainty.
9. Variance amplification in ZNE
If independent estimates μ̂i have variances σi², then
Var(Σ_iγ_i μ̂_i)=Σ_iγ_i²σ_i².
For first-order c=2, coefficients are {2,−1}. If the two noisy estimates have equal variance σ², the extrapolated variance is
4σ²+σ²=5σ².
Achieving the original standard error therefore needs roughly five times as many total effective shots under equal-allocation assumptions.
10. Probabilistic error cancellation
PEC begins from a calibrated model of noisy operations that are actually implementable.
Represent an ideal operation 𝒰 as a quasiprobability linear combination
𝒰=Σ_i q_i 𝒩_i
where the coefficients qi can be negative.
Sample implementable noisy operation 𝒩i with probability |qi|/γ, where
γ=Σ_i|q_i|.
Multiply the measurement result by the signed weight γ·sgn(qi). Averaging restores the ideal operation expectation if the calibrated model is exact.
11. Why negative quasiprobabilities are expensive
If all qi were ordinary probabilities summing to one, γ=1 and there would be no extra sampling penalty.
Noise inversion generally requires negative coefficients, so γ>1.
Each sample weight has magnitude γ. Variance therefore scales roughly as γ² per mitigated operation block, and for many gates the total overhead can grow exponentially with circuit noise volume.
This is the fundamental PEC trade-off: asymptotically unbiased estimation under a trusted noise model at the cost of potentially severe sampling overhead.
12. Worked PEC overhead
Suppose one effective mitigated gate decomposition has γ=1.08 and a circuit contains 50 independent layers with the same overhead factor.
The total absolute-weight factor is approximately
Γ=1.08^50≈46.9.
The variance overhead scale is Γ²≈2200.
A modest 8% quasiprobability overhead per layer therefore becomes extremely expensive over depth. This is why PEC is most plausible for shallow circuits or very low physical error rates.
13. PEC depends on noise-model accuracy
PEC is unbiased only relative to the learned model.
If calibration misses coherent drift, crosstalk, non-Markovianity or context dependence, the quasiprobability inverse cancels the wrong channel.
Increasing shots cannot remove this model bias.
Noise learning and mitigation should therefore be interleaved with drift checks and held-out benchmark circuits.
14. Symmetry verification
Suppose the ideal state lies in a known eigenspace of symmetry S:
S|ψ⟩=s|ψ⟩.
If noise produces outcomes outside that symmetry sector, those outcomes can be rejected or algebraically projected away.
For a ±1 symmetry, the projector is
P_s=(I+sS)/2.
The symmetry-verified expectation of observable O commuting with Ps can be written
⟨O⟩_sv=Tr(OP_sρP_s)/Tr(P_sρ).
15. Symmetry verification does not detect all errors
An error E that commutes with the symmetry or maps states within the same eigenspace can remain invisible.
Example: if parity is fixed, a two-qubit bit flip may preserve parity even though it changes the logical state.
Symmetry verification removes only components violating the checked constraints. It is a targeted filter, not a complete error-correcting code.
16. Postselection cost
If pacc=Tr(Pρ) is the probability an experimental shot passes the symmetry test, only that fraction contributes after postselection.
To retain M accepted samples on average, one needs approximately M/pacc raw shots.
If pacc=0.4, the raw-shot overhead is 2.5 even before considering changed estimator variance inside the accepted subset.
17. Measurement-error mitigation
Readout error is classical after the final quantum state reaches the detector basis in many architectures.
Let p be the true ideal measurement-outcome probability vector and q the observed vector. A calibration matrix A gives
q=Ap.
If A is known and well conditioned, estimate
p≈A^{-1}q.
Direct inversion can produce negative probabilities under finite sampling or amplify noise when A is ill conditioned. Constrained estimators or model-structured inversions are often preferable.
18. Worked one-qubit readout correction
Suppose the readout confusion matrix is
A=[[0.96,0.08],[0.04,0.92]].
Columns correspond to prepared true states 0 and 1.
If the observed probabilities are q=(0.56,0.44), then detA=0.88 and
A^{-1}=(1/0.88)[[0.92,−0.08],[−0.04,0.96]].
Multiplication gives p≈(0.5455,0.4545).
The correction is modest because the calibration matrix is close to identity and well conditioned.
19. Tensor-product readout models
For n qubits, a full confusion matrix has 2n×2n entries and becomes impossible to calibrate densely at large n.
A scalable approximation assumes independent local readout:
A≈A_1⊗A_2⊗…⊗A_n.
Correlated measurement-error models add selected pair or neighbourhood corrections.
Model complexity should match evidence: an over-simple tensor product misses crosstalk; a full matrix overfits and is exponentially costly.
20. Virtual distillation
Suppose a noisy state has spectral decomposition
ρ=Σ_i p_i|i⟩⟨i|
with desired component p₀ largest.
Then
ρ^M/Tr(ρ^M)
amplifies the relative weight of the dominant eigenvector because ratios become (pi/p₀)M.
Virtual-distillation protocols estimate observables of this nonlinear effective state using multiple copies and permutation measurements rather than physically preparing ρM/TrρM.
21. Purity amplification has a limit
If the dominant eigenvector of noisy ρ is not the ideal state, virtual distillation amplifies the wrong vector.
Coherent systematic errors can rotate the leading eigenvector away from the target while leaving it nearly pure.
Purifying a biased state is not the same as correcting its coherent orientation.
22. Clifford data regression
One can train a classical correction model using circuits close to the target that are classically simulable, for example circuits in which selected non-Clifford gates are replaced by Clifford operations.
Compare noisy hardware results with exact classical values on the training set, fit a regression, then apply it to the target circuit.
This can capture device-specific bias without an explicit microscopic noise model, but validity depends on whether the training circuits span the same error behaviour as the target.
23. Combining mitigation methods
Readout mitigation, symmetry verification and ZNE can be layered.
But corrections do not necessarily commute mathematically. Postselection changes the sample distribution; readout inversion can alter symmetry estimates; ZNE coefficients can amplify every upstream statistical error.
A combined pipeline should therefore be tested end to end on benchmark circuits rather than assuming independent gains multiply cleanly.
24. Error mitigation is observable specific
Many mitigation protocols estimate a chosen observable accurately without reconstructing a globally accurate density matrix.
A corrected energy estimate does not imply every correlation function, output bitstring probability or fidelity is equally corrected.
This is a strength—one need not solve full tomography—but it limits what can be inferred from a successful mitigated scalar.
25. Scaling limits
Mitigation overhead generally worsens with circuit depth, error rate and system size.
Fundamental analyses connect mitigation cost to distinguishability/resource measures and show that suppressing errors without redundancy requires increased sampling or stronger prior knowledge. [4]
There is no free asymptotic replacement for fault tolerance hidden inside classical post-processing.
26. Common misconception: mitigation makes the noisy state equal to the ideal state
Many methods only produce a corrected estimator for selected observables. The physical hardware state can remain noisy.
27. Common misconception: ZNE always improves accuracy
If noise scaling is nonlinear, statistics are weak or high-order coefficients become large, extrapolation can increase total error.
28. Common misconception: PEC is model free
PEC explicitly depends on a calibrated representation of ideal operations as quasiprobability combinations of implementable noisy operations. Model error becomes mitigation bias.
29. Worked synthesis problem
A variational circuit estimates energy E using Pauli expectations. At nominal noise λ, E(λ)=−1.42. After twofold gate folding, E(2λ)=−1.30.
Step 1: First-order ZNE. For c=2, E_0≈2E(λ)−E(2λ)=−2.84+1.30=−1.54.
Step 2: Physical check. If a known rigorous lower bound on the Hamiltonian ground energy is −1.50, the extrapolated −1.54 is impossible and cannot be accepted literally.
Step 3: Diagnose. Possible causes include statistical uncertainty, nonlinear noise scaling or coherent folding artefacts.
Step 4: Improve. Increase shots, add a third noise scale, compare alternative folds and inspect whether the fitted extrapolation respects the known energy bound.
Step 5: Conclusion. A mitigation result should be judged against independent physical constraints, not trusted because the corrected number moved farther from the noisy value.
30. Practice set
- What quantity does error mitigation often target?
- Write the MSE bias–variance decomposition.
- State the first-order ZNE formula for scale factor c.
- Why can ZNE coefficients increase variance?
- Name two ways to scale noise experimentally.
- What is a PEC quasiprobability decomposition?
- Define the PEC overhead γ.
- Why can PEC overhead grow rapidly with depth?
- Write the ±1 symmetry projector.
- Why can symmetry verification miss some errors?
- What linear model describes readout mitigation?
- Why is mitigation not a substitute for fault-tolerant quantum error correction?
Answers
- An ideal expectation value or other selected observable statistic.
MSE=Bias²+Variance.[cμ(λ)−μ(cλ)]/(c−1).- Large alternating coefficients multiply shot noise from each noisy estimate.
- Examples: pulse stretching, gate folding, added idle time and identity-equivalent gate insertion.
- An ideal operation written as a signed linear combination of noisy implementable operations.
γ=Σ_i|q_i|.- Layer overheads multiply, so estimator variance can scale exponentially with total noise volume.
P_s=(I+sS)/2.- Errors preserving the checked symmetry sector remain invisible.
q=Ap, followed by a stable inverse or constrained estimate of p.- Mitigation does not suppress physical error propagation indefinitely; its sampling/model overhead worsens with circuit scale.
Sources and further study
[1] Kristan Temme, Sergey Bravyi and Jay M. Gambetta, Error Mitigation for Short-Depth Quantum Circuits, Physical Review Letters 119, 180509 (2017). Introduces zero-noise extrapolation and quasiprobability error cancellation for short-depth circuits.
[2] Abhinav Kandala and colleagues, Error mitigation extends the computational reach of a noisy quantum processor, Nature 567, 491–495 (2019). Experimental error-mitigation demonstrations for variational quantum circuits.
[3] Sergey Bravyi, Sarah Sheldon, Abhinav Kandala, David C. McKay and Jay M. Gambetta, Mitigating measurement errors in multiqubit experiments, Physical Review A 103, 042605 (2021). Measurement-error mitigation with structured calibration models.
[4] Ryuji Takagi, Suguru Endo, Shintaro Minagawa and Mile Gu, Fundamental limits of quantum error mitigation, Physical Review Research 3, 033178 (2021). Resource-theoretic analysis of mitigation overhead and limitations.
Continue through Quantum Mathematics
Guide 53: Quantum Optimal Control, Lie-Algebraic Controllability, GRAPE and Pulse Engineering optimises hardware control before execution. Guide 54: Continuous Quantum Measurement, Stochastic Master Equations, Quantum Filtering and Feedback estimates conditional states from live measurement records. Guide 56: Quantum Contextuality, Kochen–Specker Theorem, Compatibility Graphs and Contextual Fractions studies a foundational quantum resource beyond ordinary noise correction.