Fast convergence of inertial dynamics and algorithms with asymptotic vanishing viscosity (Q2413084): Difference between revisions

In this paper, the authors consider fast convergence properties of the trajectories of the second order differential equation of the form \[ x''(t)+\frac{\alpha}{t}x'(t)+\nabla\Phi(x(t))=g(t),\tag{1} \] where \({\mathcal H}\) is a Hilbert space, \(\nabla\Phi\) is the gradient of the convex continously differentiable function \(\Phi: {\mathcal H} \to \mathbb R\), \(\alpha\) is a positive parameter and \(g: [0,+\infty) \to {\mathcal H}\) is a small perturbation term. At the beginning of their studies, the authors consider the homogeneous counterpart of (1), that is the case when \(g\equiv 0\) and firstly establish the minimizing property in the case when \(\alpha >0\) and argmnin \(\Phi\) is possibly empty. Next, assuming that argmnin \(\Phi\neq \emptyset\), it is proved that a solution of (1) converges weakly to a point belonging to argmnin \(\Phi\). In the subsequent step the authors examine the asymptotic behavior of the acceleration \(x''\) in the case when \(\nabla \Phi\) is locally Lipschitz continuous and argmnin\(\Phi\) is nonempty. Next section concerns the strong convergence of solutions to (1) under some geometrical or topological assumtions on \(\Phi\). The cases when the interior of \(\Phi\) is nonemty or \(\Phi\) is even (or strongly convex) are under consideration. In the subsequent step the authors analyze the asymptotic behavior of solutions to nonhomogeneous equation (1). In particular, it is proved that if \(\alpha \geq 3\), argmnin \(\Phi \neq \emptyset\) and the perturbation \(g\) satisfies some integrability assumption, then \[ \Phi(x(t))-\min\limits_{\mathcal H}\Phi={\mathcal O}\left(\frac{1}{t^2}\right), \] where \(x\) is a solution to (1). Moreover, in the case \(\alpha >3\), \(x(t)\) converges weakly, as \(t\to +\infty\), to a point in argmnin \(\Phi\). Finally, in the last section, following the proofs of the convergence in the case of the continuous dynamics, the authors prove similar results for the associated Nesterov-type algorithms. The authors illustrate their considerations by suitable examples.