The Signal
Boosting used fewer circuit runs than PEC in simulated quantum learning
A new preprint compares boosting noisy quantum classifiers with probabilistic error cancellation, finding lower circuit-execution costs for boosting in many simulated cases. The study tests four-qubit circuits in binary-classification simulations, not on quantum hardware.

A task-specific comparison
A boosting approach to noisy quantum classification used fewer circuit executions than probabilistic error cancellation in 39 of 56 simulated comparisons, according to a new arXiv preprint. The method combines imperfect quantum classifiers rather than first trying to reconstruct each classifier’s ideal, noiseless output.
Submitted on November 19, 2026, the preprint is titled “Beyond Error Mitigation: Using Boosting for Noisy Quantum Machine Learning.” Anastasiia Lukianenko, Yash J. Patel, Nurbolat Kenbayev and Vedran Dunjko wrote the paper.
The comparison focuses on supervised binary classification. In that task, correct prediction depends on the sign and margin of the classifier’s output rather than on accurate reconstruction of every noiseless expectation value. The preprint argues that recovering every ideal expectation value may be more than the learning task requires.
What the methods were asked to do
The researchers compared AdaBoostR and Real AdaBoost with probabilistic error cancellation, or PEC. PEC estimates ideal circuit expectation values from noisy executions using a sampling procedure; the paper’s resource calculation applies its sampling overhead to the measured cost of a reference learner. Boosting instead trains noisy quantum circuits as weak classifiers and combines their predictions.
Both approaches were assessed against common learning targets. The study counted the circuit executions required to reach a specified empirical margin-loss target, and separately compared resources for reaching a validation-accuracy target. The authors also evaluated test accuracy, which was not used to set those targets.
The numerical study used a teacher–student setup: teacher circuits generated labels, and separate student circuits were trained to predict them. The simulations used four-qubit circuits at depths from one to five, with gate-local depolarizing noise. The full-grid comparison used projected PEC costs, while the researchers simulated PEC sampling directly in selected settings.
Results and limits
The preprint reports that boosting required fewer circuit executions in 39 of 56 comparisons across algorithms, depths, noise strengths and learning targets. For the classification target, its cost advantage generally increased with physical noise, but the trend across circuit depths was not consistent. These are simulation results, not measurements from a quantum processor.
Test performance also favored boosting in the study’s aggregate comparisons. The median test-accuracy difference was positive in all 36 comparisons reported for that analysis, although some individual teacher tasks favored the reference. For classification, all 22 method–depth–noise cells meeting the paper’s reporting threshold showed a positive median difference; the median across those cells was 2.15 percentage points.
The authors say the results depend on a simulator-assisted training protocol and a particular resource-counting convention. They identify hardware testing, more realistic noise models and tasks with an independently established quantum-over-classical advantage as areas for further work.
Sources
- Original source · Preprint Beyond Error Mitigation: Using Boosting for Noisy Quantum Machine Learning arXiv:2611.25752