AI RESEARCH

Fisher-Geometric Diffusion in Stochastic Gradient Descent: Optimal Rates, Oracle Complexity, and Information-Theoretic Limits

arXiv CS.LG

ArXi:2603.02417v2 Announce Type: replace-cross 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 identified by the sampling mechanism itself: to leading order it is the projected covariance of per-sample gradients.