Asymptotics for some proximal-like method involving inertia and memory aspects (Q632230): Difference between revisions

The convergence behavior of the inertial proximal algorithm \[ x_{n+1}-x_n -\alpha_n(x_n-x_{n-1}) + \beta_n\partial \Phi(x_{n+1}) \ni 0, \] is analyzed. \(\Phi\) is a closed convex function in a Hilbert space, \(\partial \Phi\) denotes the subdifferential of \(\phi\) and the sequences \((\alpha_n), \;(\beta_n)\) are nonnegative. The algorithm corresponds to the discretization of the following differential inclusion (continuous gradient system with memory term) \[ \ddot{x}(t) + \gamma(t) \dot{x}(t) + \partial \Phi(x(t)) \ni 0, \quad t > 0, \] with \(\gamma\) as time-dependent damping factor. The convergence of the algorithm is analyzed by the decaying behavior of the discrete energy \[ E_n = \frac{1}{2\beta_{n-1}} |x_n-x_{n-1}|^2 + \Phi(x_n). \] The algorithm converges if there is unique minimum. In case of multiple minima, the conditions for non-convergence are given. The conditions for the convergence is still an open question, as the authors remark.