Choosing how much noise to keep in quantum generative models

When fitting examples is not enough

A quantum circuit Born machine generates bit strings by preparing and measuring a quantum state. Training adjusts the circuit so that its output distribution resembles a target distribution represented by examples. The challenge is generalization: learning the underlying distribution rather than merely reproducing the observed records.

The preprint studies models trained using maximum mean discrepancy, or MMD. This sample-based measure compares distributions through a function that scores similarities between bit strings. A small training loss alone does not establish that the model matches the unknown target distribution.

The authors first derive a bound connecting the expected generalization gap—the difference between population loss and the training-data-based estimate—to the complexity of the distributions a noisy model can reach. This result covers arbitrary noise processes and training algorithms. It does not say that every type of noise improves generalization: the bound improves only when noise reduces the relevant complexity.

Noise reduces flexibility, but introduces bias

A more explicit result requires stronger assumptions. The authors consider local Pauli errors that circuit operations spread sufficiently widely for their combined effect to be approximated by white noise. In this approximation, the output is a mixture of the ideal distribution and a uniform distribution, which assigns equal probability to every bit string.

Under these assumptions, the bound’s main complexity term shrinks with the fraction of ideal signal retained. Intuitively, noise weakens the structure that could encode accidental features of the training examples. The bound also accounts for departures from the white-noise approximation.

But a smaller generalization gap is not the same as a more accurate model. Pulling the output toward uniformity can erase genuine patterns. Even an ideal model that exactly matches the target acquires error when mixed with uniform noise. The useful amount of residual noise therefore depends on a balance between restricted flexibility and distorted output.

Spending a limited execution budget

GENIE, short for “generalization with inherent noise,” combines partial error mitigation with periodic selection of a noise level using held-out data. Its budget counts circuit executions, often called shots.

The mitigation component, capped probabilistic error reduction, preferentially reduces errors at noisier circuit locations while leaving sufficiently quiet locations unchanged. Under the paper’s model, this allocation controls how far the residual noise departs from white noise while limiting mitigation cost.

That cost matters because mitigation uses signed weights attached to measurement outcomes. Removing more noise increases statistical uncertainty, requiring additional executions for precise estimates. Below the hardware noise level, the mitigated model is accessed through weighted expectation values, not ordinary direct samples from the desired distribution.

During training, GENIE periodically freezes the circuit parameters and compares candidate noise levels. It weighs validation fit, measurement uncertainty and mitigation overhead before selecting the profile for the next training block. In the reported experiments, these comparisons consumed 2% of its execution budget.

The authors also bound cumulative selection regret under stated assumptions, including controlled parameter updates. This compares choices with the best candidate at the same current parameters, including a sampling-cost penalty—not with an optimal training schedule. A fixed validation sample leaves an uncertainty floor that more circuit executions cannot eliminate.

Results from small simulations

The synthetic experiments used 12-qubit circuits and three target distributions across 144 combinations of target, error rate, sample size and execution budget. Each condition used three seeds. GENIE was compared with no mitigation, half mitigation and full mitigation at matched execution budgets.

Its median relative loss above the best of those four methods was 0.3%, compared with 8.4% for half mitigation, 8.5% for full mitigation and 22.7% without mitigation. GENIE was within 1% of the best in 64% of conditions. These percentages describe relative MMD loss, not prediction accuracy or comparison with every possible method.

A separate 10-qubit experiment encoded Adult census records under stronger simulated noise. GENIE’s reported average absolute discrepancy across seven population shares was approximately four percentage points, versus eight to ten for fixed mitigation. Its group-size-weighted error in subgroup income rates was 16 points, versus 19–22. Income classification accuracy was approximately 68%, compared with 57–61% for the fixed strategies.

What remains unproven

All experiments were simulations on 10 or 12 qubits, not hardware demonstrations. The explicit noise-contraction result depends on sufficient scrambling, and subgroup income estimates still showed substantial errors. The study establishes neither superiority over classical models nor a practical quantum advantage. Its narrower contribution is a framework for choosing how much noise to remove when training resources are limited.