1: \begin{abstract}

2: This paper proposes a general framework of

3: Riemannian adaptive optimization methods.

4: The framework encapsulates several stochastic optimization algorithms

5: on Riemannian manifolds and

6: incorporates the mini-batch

7: strategy that is often used in deep learning. 

8: Within this framework, we also propose AMSGrad

9: on embedded submanifolds of Euclidean space.

10: Moreover, we give convergence analyses valid for both

11: a constant and a diminishing step size.

12: Our analyses also reveal the relationship between the convergence rate

13: and mini-batch size.

14: In numerical experiments,

15: we applied the proposed algorithm to principal

16: component analysis and the low-rank matrix completion problem,

17: which can be considered to be Riemannian optimization problems.

18: Python implementations of the methods used

19: in the numerical experiments are available at

20: \url{https://github.com/iiduka-researches/202408-adaptive}.

21: \end{abstract}

22: