What is error mitigation and why does it make quantum runs more expensive?
Status as of: August 2026.
Every quantum computer you can rent today is noisy. Error mitigation is the family of statistical tricks that squeezes a less-wrong answer out of a noisy machine — without adding a single qubit. It is not error correction: nothing is fixed in hardware, no logical qubit exists anywhere in the loop. You run more circuits, you post-process, and you pay for every extra shot. This post defines the term precisely, prices it with vendor numbers, and states the structural limit that is now a theorem, not an opinion.
What exactly is error mitigation?
Error mitigation is post-processing. You run a noisy circuit — often many deliberately modified variants of it — and combine the outputs classically so that the noise-induced bias in an expectation value partially cancels (Cai et al., Rev. Mod. Phys. 95, 045005, 2023). Three things it does NOT do:
It does not use extra qubits. It does not create a logical qubit. It does not fix the machine — the same errors happen on every shot; you are averaging around them, not removing them.
The contrast is error correction: encoding one logical qubit into many physical ones so errors are detected and undone during the computation. Correction is the structural fix, and it is not at scale — the best published demonstration is Google's Willow chip turning 101 physical qubits into exactly one logical qubit with a 0.143% error rate per cycle (Nature 638, 920–926, 2024; context in our qubit explainer). Mitigation exists because correction isn't here: it is the bridge tactic for the noisy machines you can actually rent in 2026.
How do the main techniques work?
Three families dominate practice. Each buys accuracy with shots.
| Technique | Mechanism | What it costs | Source |
|---|---|---|---|
| Debiasing (symmetrization) | Run ~25 equivalent circuit variants (different qubit mappings / gate decompositions), aggregate | Minimum 2,500 shots per task on-demand via Braket (500 direct with reservation) | Braket docs · IonQ docs, accessed Aug 2026 |
| Zero-noise extrapolation (ZNE) | Re-run the circuit at ≥3 amplified noise levels, extrapolate the observable to zero noise | ≥3× circuit executions plus growing variance; extrapolation bias not guaranteed to vanish | Giurgica-Tiron et al. 2020 · Temme et al., PRL 119, 180509 (2017) |
| Probabilistic error cancellation (PEC) | Learn a noise model, sample from its quasi-probability inverse | Sampling overhead γ² per noisy layer — grows exponentially with circuit size × error rate | Temme et al. 2017 · van den Berg et al., Nat. Phys. 2023 |
| Any mitigation (bound) | — | Worst case: superpolynomial sample count already at shallow depths | Quek et al., Nat. Phys. 20, 1648 (2024) |
The flagship use is real: IBM's 127-qubit "utility" experiment (Kim et al., Nature 618, 500, 2023) leaned on ZNE to produce expectation values the raw hardware could not. It is also the experiment that classical tensor networks later reproduced — see why a weak classical baseline ruins a quantum benchmark. Mitigation made the machine's numbers usable; it did not make them unbeatable.
Why does it make runs more expensive?
Because every technique converts accuracy into samples, and samples are the unit you pay for. Concrete arithmetic, vendor prices, accessed August 2026:
On IonQ Forte via AWS Braket, one shot costs $0.08 and a task costs $0.30 (Braket pricing; breakdown in what a real run costs). Turning debiasing on sets a floor of 2,500 shots per on-demand task (Braket docs). So the entry price of mitigated quality on that machine is 2,500 × $0.08 + $0.30 ≈ $200 per task — before your algorithm has asked for a single extra iteration. A variational loop of 100 mitigated evaluations starts at ~$20,000 on list prices.
That floor is the cheap part. ZNE multiplies executions by the number of noise levels and inflates variance. PEC is the honest extreme: its overhead is a factor γ² per noisy circuit layer, which compounds — the total sample count grows exponentially with circuit size at fixed error rate (Temme, Bravyi & Gambetta, PRL 2017; hardware protocol in van den Berg et al., 2023).
Is the cost blow-up an engineering problem or a law?
As of 2024 it is a theorem. Quek, Stilck França, Khatri, Meyer & Eisert (Nature Physics 20, 1648–1658, 2024) proved that for any error-mitigation protocol under local depolarizing noise, the worst-case number of samples needed to recover noiseless expectation values grows superpolynomially already at circuit depths comparable to current experiments — "scrambling due to noise can kick in at exponentially smaller depths than previously thought." The bound covers non-unital noise (like T1 decay on superconducting chips) too. Better engineering shifts the constants; it does not repeal the scaling.
The market is behaving as if it knows. Two reads from the frontier, both documented in earlier posts: IBM's flagship 2025 quantum-chemistry result (77 qubits, iron–sulfur clusters) abandoned the mitigation-heavy variational loop entirely, citing prohibitive runtimes, and moved to sample-based diagonalization (Robledo-Moreno et al., Science Advances 2025; analysis in our VQE post). And IonQ's headline benchmark #AQ builds error mitigation inside the reported number — its #AQ 64 announcement attributes the jump partly to "dramatic improvements in error mitigation" (IonQ blog, 13 Oct 2025; how to read vendor metrics in QV vs #AQ). Neither is an accusation — both are rational responses to the same cost curve.
For a buyer, the practical consequence is simple: any quote for "production-quality" answers from a 2026 QPU silently includes the mitigation multiplier. Ask for it explicitly — the framework is in how to evaluate a quantum pilot without burning budget.
What we know / what we don't know
We know: mitigation is post-processing on noisy machines — no logical qubit is involved (Cai et al., RMP 2023). We know its entry price on rentable hardware: 2,500-shot floor with debiasing on-demand → ≈$200/task on Forte (Braket, Aug 2026). We know PEC's overhead compounds exponentially with circuit size (Temme et al. 2017), and that a worst-case superpolynomial blow-up applies to every mitigation protocol at shallow depth (Quek et al. 2024). We know the frontier is routing around the cost: IBM's pivot away from variational loops; mitigation folded inside vendor metrics.
We don't know: the end-to-end cost multiplier of mitigation on a real industrial workload, compared on the same instance with and without it — nobody publishes that number, and it is the one a buyer actually needs. Whether newer tricks (classical shadows, measurement grouping, ML-assisted mitigation) bend the average case enough to matter — the proven bounds are worst-case. Where the crossover sits between paying the mitigation multiplier and waiting for logical qubits — it depends on logical-qubit pricing that does not exist yet. And whether Willow-class correction scales to the thousands of logical qubits that would make this whole post historical.
Our own data: none in this class. Rosetta's sealed runs are noiseless simulations or single-QPU executions without a mitigation protocol of our own; we have no measurements of mitigation overhead and claim none.
Sources
- Cai et al., Quantum error mitigation, Rev. Mod. Phys. 95, 045005 (2023)
- Quek et al., Exponentially tighter bounds on limitations of quantum error mitigation, Nature Physics 20, 1648–1658 (2024)
- Temme, Bravyi & Gambetta, PRL 119, 180509 (2017)
- van den Berg et al., Nature Physics 19 (2023)
- AWS Braket docs — error mitigation on IonQ (accessed Aug 2026)
- IonQ docs — debiasing (accessed Aug 2026)
- AWS Braket pricing (accessed Aug 2026)
- Kim et al., Nature 618, 500 (2023)
- Robledo-Moreno et al., Science Advances (2025)
- Google Quantum AI, Willow — Nature 638, 920–926 (2024)
- Giurgica-Tiron et al., Digital ZNE (2020)
- IonQ blog — #AQ 64 (13 Oct 2025)
Rosetta Q publishes verdicts with reproducible raw data. This is educational content, not a product claim.