Geodesic Convexity and Optimization

on YouTube ↗01h 12m

" summary

Geodesic convexity explains why some nonconvex problems repeatedly return the same solution and gives concrete tools: on the PSD manifold with Fisher–Rao metric the geodesic has a closed form (noncommutative X^{1−t}Y^t), and many matrix quantities (condition number, generalized eigenvalue, log tr X^p, certain log det expressions) are geodesically convex. A metric‑learning objective trace(AS)+trace(A^{-1}D) becomes solvable in closed form (Riccati midpoint on the geodesic) and runs thousands of times faster with similar accuracy. Theory replaces Euclidean law of cosines by a curvature‑dependent inequality (factor x/tanh x) to get first‑order rates and highlights open problems for acceleration.

AI-written summary · about

" chapters (22)

1. 0:00Introduction
2. 2:34Context and motivation
3. 4:45Abstract concepts
4. 7:05Convexity in metric spaces
5. 9:25Applications and spaces
6. 11:18Riemannian manifolds
7. 13:20Formal definitions
8. 17:25Positive definite matrices
9. 20:15Riemannian metrics
10. 25:27Applications in research
11. 26:54Linear metric learning
12. 30:38Metric learning analysis
13. 36:23Theoretical applications
14. 39:23Principal component analysis
15. 42:24Optimization theory
16. 44:15Complexity theory
17. 47:11Subgradient method
18. 52:34Riemannian analysis
19. 55:15Law of cosines
20. 59:20Future work
21. 1:01:52Acceleration methods
22. 1:05:18Q&A and conclusion

" related talks

• 11mGeometric Methods for Machine Learning and OptimizationMelanie Weber · Meet the Fellows Welcome Event
• 46mCan Non-Convex Optimization be Robust?Rong Ge · Robust and High-Dimensional Statistics (Simons Institute workshop, 2018)
• 53mAn Interior-Point Method for Convex Optimization over Non-symmetric ConesDávid Papp · Hyperbolic Polynomials and Hyperbolic Programming
• 49mMetric Representations: Algorithms and GeometryAnna Gilbert · Algorithmic Advances for Statistical Inference with Combinatorial Structure
• 01h 01mDuality 1Michel Goemans · Bridging Continuous and Discrete Optimization Boot Camp 2017
• 43mComputational and Statistical Tradeoffs via Convex RelaxationVenkat Chandrasekaran · Workshop on Succinct Data Representations and Applications (Simons Institute Big Data 2013)
• 01h 18mOptimization IBen Recht · Big Data Boot Camp 2013
• 31mRandomized Riemannian Preconditioning for Quadratically Constrained ProblemsHaim Avron · Randomized Numerical Linear Algebra and Applications 2018