arXiv:2609.38713
When Error Mitigation Makes Things Worse: Budget-Aware Evaluation, Extrapolation Failure, and the Calibration Trust Boundary
Guilin Zhang, Kai Zhao, Xiquan Cui, Henry Heng, Xu Chu, Aletta Johanna Blanken
Abstract
Error mitigation is expected to turn noisy quantum measurements into more useful estimates. We show that it can instead amplify error. Under control-offset miscalibration, the evaluated unconstrained zero-noise extrapolation (ZNE) estimators are worse than no mitigation on 38-63% of simulated instances and produce extreme tail errors. A small IBM Heron study finds worsening on 80-88% of 18 instances across two depths, with mean error 3.0-4.1 times the raw error. In simulation, the median changes little, so median-only monitoring misses the failures. We then compare methods at the same online shot budget per evaluation and report offline training cost separately. After that cost is amortized, a budget-conditioned neural corrector occupies the low-budget end of the simulated accuracy-cost frontier and falls back to the raw estimate when an ensemble disagrees; its hardware transfer remains unsuccessful. Finally, we treat provider-reported calibration as an input that is not bound to the execution-time device state. A white-box projected-gradient stress test on this metadata increases the corrector's error 8.1 times within the feature ranges used for training and evades the disagreement monitor on half of the worsened cases. Together, the results connect budget-aware evaluation, tail risk, and calibration integrity in near-term quantum learning pipelines.
AI Summary
The paper examines when error mitigation worsens expectation estimates, comparing methods at fixed online shot budgets and testing calibration metadata as an attack surface. Simulations use hardware-efficient circuits with mean magnetization, stochastic, correlated, and coherent-offset noise, alongside a small IBM Heron experiment. Under coherent miscalibration, evaluated unconstrained zero-noise extrapolation (ZNE) estimators worsen 38–63% of simulated instances. Medians remain near raw errors while means develop large tails; one absolute error reaches 371 for an observable bounded by $[-1,1]$. On hardware, Richardson ZNE worsens 80% and 88% of instances at two depths, with mean errors 3.0 and 4.1 times raw errors across 18 instances. Physical-range clipping would cap absolute error at 2, but clipped or constrained estimators are not evaluated. A shot-budget-conditioned residual neural corrector, trained on simulated noisy–ideal pairs, reduces mean error by 72–81% under coherent miscalibration across equal-budget comparisons. CDR overtakes it at the largest budget; neural training cost is accounted for separately and requires amortization. Hardware transfer fails, yielding mean-error reductions of $-3\%$ and $-94\%$. A white-box projected-gradient attack perturbing calibration features within their training ranges raises neural error from 0.0082 to 0.0671; ensemble disagreement detects half the worsened cases. Stated limitations include simulation-dominated evidence, fixed circuit families and diagonal observables, separately trained models, best-validation-seed selection, heavy-tail statistical uncertainty, and one small hardware batch. Attack budgets are not calibrated to physical drift, and random-perturbation controls are absent.
Error found: The claim that direct prediction is worse than raw execution by 57–927% at every size and noise regime contradicts the scaling table, which includes only 7%, 3%, and rounded 0% worsening.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- Accepted for a 5-minute short oral presentation at the NeurIPS 2026 Workshop on Secure and Trustworthy Quantum Machine Learning (SaTQuML). Revised public author version