CRGD augments the objective with a penalty on negative Hessian eigenvalues to create a Lyapunov function guaranteeing convergence to second-order stationary points at user-selectable rates.
Sym- metry, saddle points, and global optimization landscape of nonconvex matrix factorization,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
math.OC 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Saddle Point Evasion via Curvature-Regularized Gradient Dynamics
CRGD augments the objective with a penalty on negative Hessian eigenvalues to create a Lyapunov function guaranteeing convergence to second-order stationary points at user-selectable rates.