Inertial Primal-Dual Dynamics Methods Featuring Implicit Hessian-Driven Damping for Convex Optimization Problems in Continuous and Discrete Time

What the math gives to ML

The paper develops inertial primal-dual dynamics for strongly convex objectives under linear equality constraints, with damping generated implicitly through gradients evaluated at an extrapolated future point. The transferable asset is not merely acceleration, but the use of a resolvent-like implicit damping step whose energy analysis controls both parameter velocity and constraint violation. A practical neural-network adaptation is an optimizer that combines momentum, dual residual correction, and a Hessian-preconditioned implicit velocity solve, with diagonal or low-rank Hessian approximations making the method feasible on large models.