Provable and scalable quantum Gaussian processes for quantum learning

Quantum machine learning has faced serious limitations in scalability, interpretability, and suitability for quantum data. To address this, researchers from Los Alamos National Laboratory and other institutions introduce quantum Gaussian processes (QGPs), a Bayesian framework that treats unknown quantum transformations as priors over Gaussian processes. Under specific structural and symmetry conditions, unitary quantum stochastic processes naturally become Gaussian processes, allowing standard ML techniques like regression, classification, and Bayesian optimization to be applied directly to quantum systems. The key innovation is injecting a physics-informed inductive bias via the quantum kernel of the process, making the model both interpretable and data-efficient.

The authors prove that matchgate (free-fermionic) evolutions produce the first family of provable and scalable QGPs where the unknown unitary acts on all qubits non-trivially. They demonstrate accurate long-range extrapolation on quantum dynamics, learning phase diagrams of many-body systems, and sample-efficient Bayesian optimization for quantum sensing tasks. The results suggest QGPs offer a simpler, more structured path forward for quantum learning, sidestepping the complexity of existing quantum neural networks while maintaining provable guarantees.