Timezone: »

 
Sufficient Conditions for Non-asymptotic Convergence of Riemannian Optimization Methods
Vishwak Srinivasan · Ashia Wilson

@
in OPT 2022: Optimization for Machine Learning »
Event URL: https://openreview.net/forum?id=1jRxg6oUaG »

Motivated by energy based analyses for descent methods in the Euclidean setting, we investigate a generalisation of such energy based analyses for descent methods over Riemannian manifolds. In doing so, we find that it is possible to derive curvature-free guarantees for such descent methods, improving on work by Zhang and Sra [2016]. This analysis allows us to study acceleration of Riemannian gradient descent in the geodesically-convex setting, and improve on an existing result by Alimisis et al 2021]. Finally, extending the analysis by Ahn and Sra [2020], we attempt to provide some sufficient conditions for the acceleration of Riemannian descent methods in the strongly geodesically convex setting.

Author Information

Vishwak Srinivasan (Massachusetts Institute of Technology)
Ashia Wilson (MIT)

More from the Same Authors