Negative curvature obstructs acceleration for geodesically convex optimization, even with exact first-order oracles.
Christopher CriscitielloNicolas BoumalPublished in: CoRR (2021)
Keyphrases
- convex optimization
- interior point methods
- low rank
- first order logic
- total variation
- primal dual
- convex relaxation
- higher order
- multiscale
- basis pursuit
- active learning
- convex formulation
- scale space
- norm minimization
- semidefinite program
- machine learning
- natural images
- markov random field
- lower bound
- computer vision