Error Mitigation vs Error Correction

Two outputs with different contracts#

Define a mitigated observable estimate versus an encoded logical operation and list what each output certifies.

Mitigation infers a better estimate from noisy runs: calibration, extrapolation to zero noise, post-processing, cleverly designed sampling. The qubits stay noisy; the answer gets cleaner. Correction encodes information into a larger system, extracts syndrome data, and actively protects or restores the logical state. The representation itself changes.

Quantum error mitigation estimates observables under noise using additional assumptions and sampling rather than encoding a protected logical state. [error-mitigation-review]

Worked comparison: keep the resource currencies separate#

A mitigation record names the observable, circuit family, noise-scaling method, calibration data, shot allocation, estimator, bias model, variance or interval, and unmitigated baseline. A correction record names code, logical state or operation, physical error model, syndrome rounds, decoder, physical and logical failure denominators, and overhead. Their common word “error” does not make their outputs commensurate.

Derive the extrapolate and its variance#

For the linear two-scale fixture, write $E(\lambda)=E_0+a\lambda+O(\lambda^2)$. Evaluating at λ and 3λ and cancelling the first-order term gives $\widehat E_0=(3E(\lambda)-E(3\lambda))/2$. The synthetic values 0.74 and 0.62 yield 0.80. The cancellation is only as credible as linearity between the chosen scales and the claim that scaling preserves the target circuit semantics.

If the two estimates are independent with variances $v_1$ and $v_3$, the extrapolated variance is $(9v_1+v_3)/4$. Even equal input variances become 2.5 times larger. Allocate shots using those coefficients or report the avoidable variance penalty. Reusing correlated calibration data changes the covariance term and must be modeled rather than silently assuming independence.

A third scale tests curvature. Fit the declared linear and quadratic models, inspect residuals, and treat model disagreement as mitigation uncertainty. Choosing scales after observing which pair lands nearest the desired answer is analysis leakage. The raw noisy estimates and the unmitigated result remain visible beside the extrapolate.

Zero-noise extrapolation spends variance#

Use a small synthetic fit to show bias reduction, variance amplification, shots, and extrapolation assumptions.

The boundary matters because the two have different ceilings. Mitigation can improve near-term experiments whose circuits are shallow enough to survive at all. Fault-tolerant computation — arbitrarily long, arbitrarily reliable — requires correction plus reliable logical operations. No sampling trick crosses that gap.

Quantum error correction encodes information and uses syndrome information to diagnose correctable errors without measuring logical data directly. [nielsen-chuang][gottesman-stabilizer]

Trace what correction actually supplies#

Correction begins by encoding one logical degree of freedom across several physical systems. Each cycle couples data to ancillas, measures checks, decodes a syndrome history, and updates a physical correction or Pauli frame. Logical error per cycle is estimated from failures of the encoded task. It is neither a gate-error percentage nor an estimator bias, and its denominator must name the logical memory or operation tested.

More distance raises data/check qubits, extracted syndrome bits, decoder work, and often logical-operation time. Decoder latency is seconds; code distance is dimensionless; physical qubits and syndrome cycles are counts. These quantities jointly constrain a design but cannot be summed into an “overhead score.”

Correction spends qubits, cycles, and latency#

Trace encode-syndrome-decode-frame-update and keep physical and logical errors separate.

Mitigation pays in shots. Its estimates are still sampled, so for a Bernoulli estimator $\operatorname{SE}(\widehat p)\approx\sqrt{p(1-p)/S}$, and methods like zero-noise extrapolation tend to raise variance as they remove bias — the total shot cost can grow steeply with circuit size.

Fault-tolerant computation requires reliable logical operations and control beyond one error-detection or correction demonstration. [fault-tolerant-roads]

One table, incompatible units#

Compare shots, estimator variance, physical qubits, code distance, logical error per cycle, and decoder time without summing them.

Correction pays in qubits. The overhead estimate from Chapter 43 applies: $Q_{\mathrm{physical}}\approx Q_{\mathrm{logical}}q_{\mathrm{per\ logical}}$ , with the multiplier set by the code, the distance, and the physical error rate. Mitigation changes how noisy estimates are interpreted; correction changes how information is represented and protected. Budget accordingly.

The Shor and Steane codes are canonical examples of encoded correction and do not share mitigation's sampling-resource contract. [shor-code-1995][steane-code-1996]

Reject boundary-crossing claims#

“Mitigation reduced the error rate” is incomplete. It may mean one expectation moved closer to a trusted value; report absolute estimator bias and uncertainty. “The code corrected an error” may mean one ideal injected fault was reversed; report the correctable set and extraction assumptions. “Fault tolerant” requires error propagation control and improving logical behavior under increased protection, not merely an encoded circuit.

Postselection illustrates the risk. Discarding runs that trigger a check can improve a conditional observable while lowering acceptance probability and changing the sampled population. The evidence record needs accepted shots, attempted shots, selection rule, and any induced bias. It does not demonstrate preservation of an arbitrary unknown state across the discarded cases.

Claims that cross the boundary without evidence#

Reject concrete statements that equate a cleaner estimate, error detection, correction, and fault tolerance.

A strong answer: mitigation can reduce bias in estimates from noisy runs, but it does not encode arbitrary logical information for long computations. Error correction uses redundancy and syndrome extraction to protect logical states, and fault tolerance adds reliable logical gates, repeated syndrome cycles, decoding, and control — at substantial resource overhead. A mitigated demo may be valuable evidence. It becomes evidence of protected logical computation only when logical error behavior itself is demonstrated.

Near-term utility claims must keep mitigation evidence distinct from fault-tolerance roadmaps. [nisq-preskill][error-mitigation-review]

Let the workload decide#

Choose mitigation when the deliverable is a bounded set of observables, the noisy circuits remain informative, a trusted scaling or calibration model is testable, and the shot budget is acceptable. Reverse the choice when variance or model sensitivity overwhelms the claimed bias reduction. Choose correction when the deliverable requires protected state lifetime, repeated logical operations, or depth beyond the physical error budget; reverse an architecture choice when logical suppression fails to improve with added protection.

A project may use both: correction supplies a logical experiment and mitigation adjusts a residual observable. The report must preserve both layers. Mitigating logical data does not retroactively establish that the code achieved its logical error target, and correction does not validate the mitigation model.

The comparison table should therefore carry evidence labels. A mitigated estimate is “model-derived from noisy samples”; its uncertainty includes sampling and scaling assumptions. A corrected syndrome trace is “ideal executable fixture” unless noisy extraction was modeled. A logical-error measurement is “empirical under named code and protocol.” Rows with different labels may inform one engineering decision, but they are not interchangeable observations.

Bias, variance, and shot cost also need distinct columns. An extrapolate can move closer to the known synthetic target while its confidence interval widens; that is a bias–variance trade, not unqualified improvement. Correction has a different failure curve: added protection is useful only when measured logical error decreases under the same denominator and operating protocol.

Reversal criteria make the decision falsifiable. Abandon the chosen mitigation when held-out scales show curvature or required shots exceed the workload budget. Reject the correction configuration when repeated syndrome extraction raises logical failure, decoder deadlines are missed, or the physical error model lies outside the code’s demonstrated suppression regime.

Choose by workload and claim#

Match shallow observable estimation or long protected computation to the appropriate evidence path and reversal condition.

The second trap: "more physical qubits solve it." More qubits without lower error rates, syndrome extraction, and logical operations build a larger noisy system, not a more reliable one. Scale is an ingredient, not the recipe.

Quantum error mitigation estimates observables under noise using additional assumptions and sampling rather than encoding a protected logical state. [error-mitigation-review]