Global Convergence and Error Propagation in Neural Gradient Flows: A Riemannian Optimization Framework
Published in arXiv preprint, 2026
Reformulates minimizing-movement steps of energy minimization as Riemannian gradient flow on a neural increment manifold, proving exponential local convergence and O(δ) global function-space error propagation for Gauss–Newton-preconditioned training.
Recommended citation: S. Zheng, Y. Wang, H. Yang. (2026). "Global Convergence and Error Propagation in Neural Gradient Flows: A Riemannian Optimization Framework." arXiv:2605.27779.
Download Paper