1: \begin{abstract}

2: Optimization on manifolds is crucial in machine learning applications such as principal component analysis and low-rank matrix completion. 

3: However, federated solutions to machine learning problems with manifold constraints remain limited due to the complex geometric properties of manifolds and the challenges associated with designing and analyzing Federated Learning (FL) algorithms with nonconvex constraints.

4: %

5: In this paper, we consider nonconvex FL over compact smooth submanifolds with heterogeneous data. We propose a computationally and communication-efficient algorithm that utilizes stochastic Riemannian gradients and a manifold projection operator, uses local updates to improve communication efficiency, and avoids client drift. Theoretically, we prove sub-linear convergence to a neighborhood of a first-order optimal solution using novel analysis techniques that jointly leverage the structures of manifolds and the properties of loss functions.  Numerical experiments demonstrate that our algorithm requires significantly less computations and communications than existing methods. 

6: \end{abstract}

7: