Nonconvex Federated Learning on Compact Smooth Submanifolds With Heterogeneous Data