Hamiltonian Descent Methods

Chris J. Maddison, Daniel Paulin, Yee Whye Teh +2

We propose a family of optimization methods that achieve linear convergence using first-order gradient information and constant step sizes on a class of convex functions much large…