Speaker
Description
Variational quantum algorithms operate in a challenging statistical regime: objective functions and gradients must be inferred from a finite number of noisy quantum measurements. In this talk I present uncertainty-aware and bias-aware statistical learning methods for optimizing parametrized quantum circuits, with a focus on the Variational Quantum Eigensolver.
This framework strengthens the classical learning component of VQE through physics-informed Gaussian-process kernel learning, from which confidence-region methods, adaptive shot allocation, and the Bayesian parameter shift rule stem. A complementary bias analysis characterises when sequential minimal optimization energy estimates become unreliable and shows how this phenomenon can be exploited to speed up optimization without sacrificing statistical accuracy.
The methods are evaluated using exact classical simulations of spin-chain Hamiltonians, enabling quantitative comparisons between optimizers under a fixed cumulative shot budget. The results illustrate how explicit modeling of uncertainty and estimator bias can improve the efficiency and interpretability of variational quantum optimization.