Fast convex optimization via inertial dynamics with Hessian driven damping

Abstract

We first study the fast minimization properties of the trajectories of the second-order evolution equation $$\ddot{x}(t) + \frac{\alpha}{t} \dot{x}(t) + \beta \nabla^2 \Phi (x(t))\dot{x} (t) + \nabla \Phi (x(t)) = 0,$$ where $\Phi:\mathcal H\to\mathbb R$ is a smooth convex function acting on a real Hilbert space $\mathcal H$, and $\alpha$, $\beta$ are positive parameters. This inertial system combines an isotropic viscous damping which vanishes asymptotically, and a geometrical Hessian driven damping, which makes it naturally related to Newton's and Levenberg-Marquardt methods. For $\alpha\geq 3$, $\beta >0$, along any trajectory, fast convergence of the values $$\Phi(x(t))- \min_{\mathcal H}\Phi =\mathcal O\left(t^{-2}\right)$$ is obtained, together with rapid convergence of the gradients $\nabla\Phi(x(t))$ to zero. For $\alpha>3$, just assuming that $\Phi$ has minimizers, we show that any trajectory converges weakly to a minimizer of $\Phi$, and $ \Phi(x(t))-\min_{\mathcal H}\Phi = o(t^{-2})$. Strong convergence is established in various practical situations. For the strongly convex case, convergence can be arbitrarily fast depending on the choice of $\alpha$. More precisely, we have $\Phi(x(t))- \min_{\mathcal H}\Phi = \mathcal O(t^{-\frac{2}{3}\alpha})$. We extend the results to the case of a general proper lower-semicontinuous convex function $\Phi : \mathcal H \rightarrow \mathbb R \cup \{+\infty \}$. This is based on the fact that the inertial dynamic with Hessian driven damping can be written as a first-order system in time and space. By explicit-implicit time discretization, this opens a gate to new $-$ possibly more rapid $-$ inertial algorithms, expanding the field of FISTA methods for convex structured optimization problems.