Learning when quantum error mitigation is worth the cost

Choosing a remedy, not inventing one

Quantum circuits suffer from several kinds of error. Gates can disturb the state they are meant to manipulate, information can decay while qubits wait, and measurements can misidentify the result. Different remedies address different problems. Applying the same remedy throughout an experiment can therefore waste resources or miss opportunities to improve an answer.

Adaptive Policy-Guided Error Mitigation, or APGEM, puts a controller above the quantum-learning task. Rather than developing a new way to correct measurements, it learns which available intervention offers the best balance of output quality and measurement cost.

The menu contains six choices. Zero-noise extrapolation deliberately runs noisier versions of a circuit to estimate its noise-free output. The study’s probabilistic error-cancellation implementation approximately reverses an effective depolarizing channel. Measurement mitigation corrects readout confusion. Dynamical decoupling uses pulse sequences to counter coherent errors during idle periods. Circuit optimization recompiles the calculation to reduce noise exposure. The final option is to do nothing.

How the controller learns

The controller uses a contextual bandit: a decision-making algorithm that chooses an action, observes its payoff and updates its estimates. Unlike a learner given the outcomes of every alternative, it receives feedback only for the technique it selected.

Its inputs include estimated noise rates, an unmitigated probe’s performance, properties of the measured distribution, circuit statistics and optimization signals. These are transformed into 24 numerical features. The LinUCB algorithm uses those features to estimate each method’s utility while exploring uncertain alternatives during learning.

Utility is not simply accuracy. Each task’s performance is normalized to a common scale, then scored relative to a task-specific usability threshold. Measurement cost and estimator variance incur penalties. An expensive correction is worthwhile only when its quality benefit compensates for those penalties. The result therefore depends on the researchers’ chosen thresholds and weights.

What the simulations found

The evaluation covers a variational classifier, a quantum-kernel classifier, an energy-estimation task, a generative model and a reinforcement-learning routing task. Most use four-qubit, two-layer circuits. Sixteen constructed noise conditions per task produce 80 task-condition combinations.

All quantum execution takes place in the Qiskit Aer simulator. The researchers first cache repeated executions for each task, noise condition and mitigation choice. Controllers then learn and are evaluated using this response table. This supports paired comparisons, but it is not a demonstration of adaptation during a live hardware training run. For the classifier and energy task, model parameters are first optimized without noise.

The main comparison reports mean utility of 0.435 for APGEM, against 0.356 for the strongest fixed strategy, which was no mitigation. The manuscript reports recovery of 95.5% of an oracle’s utility. Here, the oracle means choosing the best available technique according to average utility estimated from repeated executions—not obtaining a perfect quantum answer. The percentage is therefore not a prediction accuracy or an error-removal rate.

Choices varied with the workload and noise. Error cancellation was prominent for energy estimation and generation, while the kernel classifier often did best without mitigation. A second contextual controller, using Thompson sampling, achieved similar utility, suggesting that conditioning choices on context mattered more than the particular selection algorithm.

Where the evidence stops

A further simulation used noise models derived from eight IBM device calibration snapshots, supplemented with controlled coherent errors. Across 32 blocks, APGEM reportedly reached 99.9% of oracle utility, with every choice within 0.01 utility of the best option. These were hardware-informed simulations, not executions on those processors. Several choices were nearly tied; on the coherent-error snapshots specifically, dynamical decoupling was statistically indistinguishable from APGEM.

A three-to-eight-qubit sweep held the reported decision measurement budget at 13,312 shots. That demonstrates behavior under the tested budget, not that a fixed budget will preserve accuracy at larger scales or overcome quantum simulation’s exponential memory requirements.

Other qualifications remain. Some context features rely on ideal-reference information, whose availability for harder practical problems is not established here. The manuscript also reports differing regret and exact-choice agreement figures across sections, which need clarification. As a preprint, the work has not established its conclusions through peer review.

The supported finding is narrower than a general solution to quantum noise: within these small simulated workloads and this cost function, learning which mitigation method to use performed better than committing to one method throughout.