Manifold optimization

previous arrow
next arrow
Slider

Opti­mization on manifolds: an overview

Let f:\mathcal{M} \rightarrow \mathbb{R} be a smooth real-valued function on a Riemannian manifold \mathcal{M}. We consider

\displaystyle{\min_{w \in \mathcal{M}} f(w)},

where w \in \mathcal{M} is a given model variable. This problem has many applications; for example, in principal component analysis (PCA) and the subspace tracking problem, which is the set of r-dimensional linear subspaces in \mathbb{R}^d. The low-rank matrix completion problem and tensor completion problem are promising applications concerning the manifold of fixed-rank matrices/tensors. The linear regression problem is also defined on the manifold of fixed-rank matrices. The independent component analysis (ICA) is an example defined on the oblique manifold.

Manifold optimization conceptually translates a constrained optimization problem into an unconstrained optimization problem on a nonlinear search space (manifold). Building upon this point of view, this problem can be solved by Riemannian deterministic algorithms (the Riemannian steepest descent, conjugate gradient descent, quasi-Newton method, Newton’s method, end trust-region (TR) algorithm, etc.) or Riemannian stochastic algorithms.

Riemannian steepest descent.

Publication