Quantum generative models need not strip all noise: GENIE picks a noise level from the shot budget and stays near the best in most simulated conditions
Synopsis
When a quantum circuit Born machine, a generative model that draws samples directly from measurement probabilities of a quantum state, is trained under a finite shot budget, hardware noise both contracts the generalization gap and pulls the output toward uniform, so the best noise level can be intermediate and unknown in advance. GENIE treats residual noise as a tunable regularizer: capped probabilistic error reduction realizes a chosen noise level, and a controller selects that level during training from held-out data within a fixed budget of circuit executions. In simulations on three synthetic targets and the Adult census data, it stays close to the best of four noise-control strategies, whereas every fixed noise level falls short in many settings.
Interpretation
When training a quantum circuit Born machine, retaining part of the hardware noise can shrink the generalization gap between training loss and loss against the true distribution, without removing all noise. Earlier quantum generative-model work largely treated noise as an error to suppress, and judged error mitigation by how much bias it removes at a given sampling cost. For any noise process, the generalization gap is bounded by the complexity of the reachable distribution class; under scrambled local Pauli noise that complexity contracts by the retained-signal factor, but the same noise also biases the model toward the uniform distribution.
GENIE turns the noise level into a hyperparameter chosen during training, using held-out data to decide how much noise to keep within a fixed budget of circuit executions. The hardware noise level is fixed by the device, and lowering it requires error mitigation whose sampling cost grows exponentially with the amount of noise removed. Capped probabilistic error reduction mitigates only the noisiest locations and keeps the remaining noise close to white noise; the controller periodically spends a small part of the budget probing candidate levels and selects the level for the next training block from validation performance and remaining budget.
In simulated comparisons at matched circuit-execution budgets, adaptive level selection is steadier than any fixed noise level. Fixed levels reverse rank across budgets and error rates: full mitigation is worst at the smallest budget, and training without mitigation stalls on blob and wfc as the budget grows. On three synthetic datasets, GENIE stays close to the best of the four methods in most conditions, with a median regret far below the three fixed levels; on the Adult data at strong noise it gives the most accurate within-group income rates and income predictions of the four methods.
Perspective
The conclusions rest on simulated circuits: both the synthetic and the census experiments use simulated circuits with stochastic Pauli noise, and no real hardware was used. The source proposes testing the effective depolarizing approximation and noise control on larger circuits and hardware, and deriving contraction bounds for other noise channels.
The effective depolarizing approximation requires enough circuit scrambling, so the precision of the residual-noise estimate depends on circuit architecture and depth; the controller compares candidate levels at the current parameters rather than planning a schedule for the whole run, and its reliability depends on the held-out sample size; the source reports that in the census experiment the held-out sample was already sufficient to select useful noise levels.
