Fisher-Geometric Diffusion in Stochastic Gradient Descent: Optimal Rates, Oracle Complexity, and Information-Theoretic Limits
arXiv:2603.02417v3 Announce Type: replace-cross
Abstract: Classical stochastic-approximation analyses treat the covariance of stochastic gradients as an exogenous modeling input. We show that under exchangeable mini-batch sampling this covariance is i…