Sample-efficient quantum error mitigation via classical learning surrogates
A paper published in Nature Quantum Intelligence describes a method that uses classical learning surrogates to reduce the number of circuit executions needed for quantum error mitigation. The approach trains a classical model on noisy quantum circuit outputs and uses it to estimate error-mitigated expectation values more efficiently.
Why it matters
Conventional quantum error mitigation techniques such as zero-noise extrapolation and probabilistic error cancellation incur large sampling overheads that scale poorly with circuit size and noise strength, making them impractical beyond small demonstrations. This work targets the main bottleneck—sample complexity—by shifting part of the estimation burden to a classical model. If the surrogate captures noise structure from limited data, it could lower the cost of obtaining reliable observables from near-term hardware, widening the range of problems where error mitigation is practical.
AI analysis — not reported by the source
What this could make possible
0–2 years
- Plausible
The method could be integrated as an optional error mitigation layer in quantum software stacks within two years, reducing shot counts by an order of magnitude for routine VQE and QAOA experiments.
If the surrogate requires only a fraction of the circuit runs that standard extrapolation or probabilistic cancellation needs, software platforms have strong incentive to adopt it. The integration path is ordinary engineering: a pre-trained model per device or per circuit family, combined with existing noise characterization tools.
2–5 years
- Speculative
Classical surrogates trained on noisy quantum data could be extended to predict full output distributions for small molecular simulations, reducing quantum runs and making hybrid workflows practical for a broader set of chemistry problems.
This requires going beyond single expectation values to multiple observables or probability distributions, and validating that the surrogate does not hallucinate correlations. If that extension works, it would shift more computational load to classical post-processing, lowering quantum resource demands.
5+ years
- Speculative
The approach could evolve into a general framework where classical models are trained once on a noisy quantum processor and then used as emulators for larger systems that have not been run, effectively pre-computing parts of quantum algorithms.
This depends on proving that surrogates can extrapolate across system sizes or noise regimes without access to exact classical simulation—a very hard problem. If achieved, it would change how near-term quantum devices are used, prioritizing training data generation over direct computation.
What would have to be true
- The classical surrogate must generalize beyond the specific noise instances used in training; otherwise it merely memorizes and offers no sample advantage.
- The method must show a net reduction in total quantum runtime after accounting for the training runs, not just a lower number of mitigation-specific samples.
- The advantage must persist under realistic hardware noise, including drift and crosstalk, which could degrade the trained surrogate.
- Benchmarks against established methods like zero-noise extrapolation and probabilistic error cancellation must be reported on problem sizes where classical simulation is not trivial.
Who’s positioned
- IBM — IBM's Qiskit Runtime already offers error mitigation primitives; a sample-efficient method could be integrated to reduce costs for users of its superconducting processors.
- Google — Google has active research in error mitigation for quantum simulation and could apply the surrogate approach to its Sycamore-class devices to extend the reach of its experiments.
- Q-CTRL — Q-CTRL builds software for error suppression and mitigation, and a new sample-efficient technique could be productized into its performance-management layers.
- Boehringer Ingelheim — As a customer of quantum chemistry services, this pharmaceutical company would benefit from lower sample requirements making hybrid VQE-type workflows more cost-effective.
What could change this
- The classical surrogate may only work in regimes where a classical simulator could have computed the result directly, undermining any quantum advantage claim.
- Training overhead could negate the sample efficiency gains, especially if the surrogate must be retrained per circuit or per noise drift.
- The method may not scale to deep circuits or high-weight observables if the surrogate's capacity cannot capture the noise correlations.
- Noise instabilities in near-term hardware could make trained surrogates unreliable over time, requiring frequent recalibration.