1: \begin{abstract}

2: There has been great interest in using tools from

3: dynamical systems and numerical analysis of differential equations

4: to understand and construct new optimization methods.

5: In particular, recently a new paradigm has emerged that applies ideas from   mechanics and geometric integration to obtain accelerated optimization methods on Euclidean spaces. This has

6: important consequences given that

7: accelerated methods are the workhorses behind many machine

8: learning applications.

9: In this paper we

10: build upon these advances and propose a

11: framework for dissipative and constrained Hamiltonian systems that

12: is suitable for solving optimization problems on arbitrary smooth manifolds.

13: Importantly, this allows us to leverage

14: the well-established theory of symplectic integration

15: to derive ``rate-matching'' dissipative integrators. This

16: brings  a new perspective to optimization on manifolds whereby

17: convergence guarantees follow by construction from classical arguments in symplectic geometry and backward error analysis.

18: Moreover, we construct two dissipative generalizations of leapfrog  that are straightforward to implement:

19: one for Lie groups and homogeneous spaces, that relies on the tractable geodesic flow or a retraction thereof,

20: and the other  for

21: constrained submanifolds  that is

22: based on a dissipative generalization of the famous RATTLE integrator.

23: \end{abstract}

24: