Expressibility and Trainability Analysis of Hardware-Efficient Ansatz Variants in Variational Quantum Eigensolver with a Linear Mixing Model Maudina Rohmah, Teguh Budi Prayitno, Yanoar P. Sarwono
Badan Riset dan Inovasi Nasional
Abstract
This study evaluates nine hardware-efficient ansatz (HEA) variations in an H2 VQE system, measuring expressibility via KL divergence and barren plateaus via gradient variance. Results show that high-expressibility ansatzes like RyRx often suffer from vanishing gradients. Conversely, complex mixing structures like HRy offer superior stability. These findings support a linear mixing model: excessive expressibility can hinder optimization by homogenizing state space and reducing critical directional information.
