0%

Optimization on Manifold

Manifold Optimization

Riemannian optimization problem:

where denotes a Riemannian manifold with the Riemannian metric , is a nonempty, compact, geodesically convex set, and is geodesically convex and geodesically -smooth. Here -convex function may be non-convex in the usual Euclidean space but convex along the manifold, and thus can be solved by a global optimization solver.

Riemannian Manifold

A Riemannian Manifold is a real smooth manifold equipped with a Riemannian metric . Let denote the inner product of ; and the norm of is defined as , where the metric induces an inner product structure in each tangent space associated with every .

A geodesic is a constant speed curve that is locally distance minimizing. Let and , then an exponential map maps to on , such that there is a geodesic with and . If there is a unique geodesic between any two points in , the exponential map has inverse , i.e., , and the geodesic is the unique shortest path with , where is the geodesic distance between .

Parallel transport maps a vector to , and preserves inner products and norm, that is, , where .

For any and any geodesic with and for such that , then is geodesically convex (G-convex), and an equivalent definition is formulated as follows;

where is the Riemannian gradient of at . A function is called geodesically -strongly convex (-strongly G-convex) if for any , the following inequality holds

A differential function is geodesically -smooth (G-L-smooth) if its gradient is -Lipschitz, i.e.,

Global Convergence Rate Analysis

provided the global complexity analysis of first-order algorithms for -convex optimization, and designed the following Riemannian gradient descent rule:

where is the iterate index, is an exponential map at , is a step-size or learning rate, and is the Riemannian gradient of at .

References

[1] @inproceedings{liu2017accelerated, title={Accelerated First-order Methods for Geodesically Convex Optimization on Riemannian Manifolds.}, author={Liu, Yuanyuan and Shang, Fanhua and Cheng, James and Cheng, Hong and Jiao, Licheng} }

[2]