Researchers bypass complexity limits in quantum learning

Researchers at Nanjing University of Information Science and Technology and the Chinese Academy of Sciences have developed a new method for learning continuous quantum dynamical trajectories, a challenge that has limited understanding of complex quantum systems. Published on August 24, 2026, in Quantum Science and Technology, their work introduces physics-informed kernel ridge regression, which doubles the information gained from each quantum simulation.

The team acknowledges a current obstacle to wider adoption: the data supporting the study is not yet publicly available due to a lack of a suitable data repository in the field. This new framework promises to compress complex quantum dynamics into classical predictive models, bridging quantum simulation and machine learning.

Physics-Informed Kernels Bypass Query Complexity in Quantum Dynamics

A new approach to learning quantum dynamics reduces the number of required quantum computer queries by encoding physical laws directly into machine learning algorithms. Conventional models treat quantum observables as generic time series, demanding an impractical number of samples to accurately capture rapid quantum fluctuations. The researchers first defined a theoretical lower bound; any learning protocol measuring independent copies without quantum memory requires at least Ω(T/ε^2) oracle queries, revealing a quadratic penalty that makes dense sampling impossible.

To overcome this, they introduced physics-informed kernel ridge regression, or PI-KRR, which integrates the Heisenberg equation as a differentiable constraint within the learning process. This framework, when combined with classical shadow tomography, simultaneously reconstructs trajectories for multiple local observables, with measurement overhead scaling at only O( log M).

The team also demonstrated robustness against errors common in near-term intermediate-scale quantum (NISQ) devices; isolated measurement outliers are suppressed as 1/sqrt(m) with a training size of m, and systematic Hamiltonian miscalibrations introduce only linearly bounded prediction errors. Numerical experiments using transverse-field Ising models showed PI-KRR resolves sharp features, such as light-cone fronts, with up to two orders of magnitude lower mean absolute error than standard kernel methods. These results suggest that PI-KRR offers a significant improvement in accuracy and efficiency for learning quantum dynamics.