Benchmarking Error Mitigation: Artefactual Improvements in Zero-Noise Extrapolation

Points clés

• ▸When noise amplification pushes past usable signal, Richardson extrapolation collapses into a fixed rescaling of one noisy measurement.
• ▸The resulting 'improvement' is independent of the noise amplification it supposedly corrects for; hardware runs overshot the ideal value by up to 21%.
• ▸A 'garbage-folding' negative control produced larger apparent improvements than genuine folding — so improvement size alone cannot validate a result.

En langage simple

Error mitigation works by running a circuit at several deliberately worsened noise levels and extrapolating back to what a noiseless machine would have said. This paper shows the trick can break silently. Once the amplified noise drowns the signal — which happens quickly on real hardware for any non-trivial circuit — the extrapolation stops tracking physics and becomes a fixed multiplier applied to one noisy number. It still returns a value, that value still looks better than the raw measurement, and nothing flags a problem. The authors' sharpest demonstration is a negative control: they fed the procedure deliberately meaningless circuit modifications and got *larger* apparent improvements than the legitimate method. The lesson generalizes well beyond quantum computing — if your correction cannot fail your own sanity check, it is not evidence.

Pourquoi c'est important

Zero-noise extrapolation underpins many near-term results, including headline quantum utility demonstrations, and is often applied implicitly inside tooling. If it can manufacture improvements that look right and are not, then a slice of published error-mitigated results needs re-examination — and every benchmark needs a negative control.