Gradient-estimator design overcomes trainability barriers in neural-network-based variational optimization

DOI10.48550/arXiv.2609.22342MaRDI QIDQ7353464arXivFDO

Authors Baigeng Wang, Chenan Wei, Yi-Ran Xue, Rui Wang

Publication date 16 September 2026

Published in arXiv

Tldr We derive a direct unbiased gradient estimator and an adaptive minimum-variance phase (AMVP) estimator to overcome trainability barriers in neural-network-based variational optimization, achieving compact networks with over an order of magnitude less GPU time while matching or surpassing density matrix renormalization group accuracy and chemical accuracy in challenging molecular systems.

Full work available at URL https://doi.org/10.48550/arXiv.2609.22342

arXiv classification

cond-mat.dis-nncond-mat.str-elcs.LGphysics.chem-phphysics.comp-ph

Cites work

• Monte Carlo Gradient Estimation in Machine Learning
• Optimization methods for large-scale machine learning
• Quantum Monte Carlo approaches for correlated systems
• Solving the quantum many-body problem with artificial neural networks
• The density-matrix renormalization group in the age of matrix product states
• Variance reduction techniques for gradient estimates in reinforcement learning

This page was built for publication: Gradient-estimator design overcomes trainability barriers in neural-network-based variational optimization

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7353464)