sharpness

This paper addresses the open question of maximum step size for gradient descent convergence on non-L-smooth objectives, introducing adaptive methods that operate at the edge of stability and can minimize sharpness globally.