Alternating proximal algorithms for linearly constrained variational inequalities: application to domain decomposition for PDE's (Q642609): Difference between revisions

The authors study alternating proximal algorithms of the type \[ x_{k+1} = \text{argmin} \left \{\gamma_{k+1} f(\xi) + \frac{1}{2}\|A\xi -By_k\|_Z^2 + \frac{\alpha}{2} \|\xi-x_k\|_X^2;\quad \xi \in X \right \} \] \[ y_{k+1} = \text{argmin} \left \{\gamma_{k+1} g(\eta) + \frac{1}{2}\|Ax_{k+1} -B\eta\|_Z^2 + \frac{\nu}{2} \|\eta-y_k\|_Y^2;\quad \eta \in Y \right \} \] for the linearly constrained convex structured minimization problem \[ \min [f(x) + g(y): \; Ax=By]. \] \(X,Y,Z\) are real Hilbert spaces, \(f:X\rightarrow {\mathbb R} \cup \{ +\infty\}\), \(g:Y\rightarrow {\mathbb R} \cup \{ +\infty\}\) are closed convex functions and \(A:X \rightarrow Z, \;B:Y \rightarrow Z\) are linear continuous operators, \(\alpha\) and \(\nu\) are positive parameters. It is shown, when the iterates \(x_k\) and \(y_k\) weakly converge to the solution of the minimization problem, then \((\gamma_k\)) tends slowly to zero. The size of the subproblems solved at each iteration step is reduced due to the structured form of the objective function. Applications to domain decomposition problems for partial differential equations (PDE's) and extension of the method to maximal monotone operators are given.