Domain decomposition iterative procedures for solving scalar waves in the frequency domain (Q1392765)

The subject of this paper is a numerical solution of the scalar complex boundary value problem \[ -\Delta u-p(x)^2u+iq(x)^2u=f(x), \quad x\in\Omega\subset {\mathbb R}^d, \] with the boundary condition \[ {\partial u\over{\partial \nu}}+i\alpha(x)u=0,\quad x\in \Gamma=\partial \Omega, \] where \(\alpha=\alpha_r-i\alpha_i\), \(\alpha_r>0\), \(\alpha_i\geq 0\), \(0<p_0\leq p(x)\leq p_1<\infty\), \(0<q_0\leq q(x)\leq q_1<\infty\). The author applies the finite element technique with bilinear elements on squares, and a non-overlapping domain decomposition \(\Omega=\cup_{j=1}^M\Omega_j\), with interfaces \(\Gamma_{jk}=\Omega_j\cap\Omega_k\) and partial borders \(\Gamma_j= \Gamma\cap\Omega_j\). On the interfaces \(\Gamma_{jk}\) Robin boundary conditions of the form \({{\partial u_j}\over{\partial\nu_j}}+i\beta u_j= -{{\partial u_k}\over{\partial\nu_k}}+i\beta u_k\) are used. In the finite-dimensional version the normal derivatives in the Robin conditions are replaced by forward and backward divided differences. The algebraic version of this problem, a system of linear equations with complex symmetric matrix, is difficult. All classical iterative processes applied to it converge slowly, or do not converge at all; moreover, the required accuracy of the numerical solution results in the necessity to operate with a large volume of data. In such a situation the use of a domain decomposition method is quite natural. The convergence of the finite element solution (without domain decomposition) to the weak solution of the original problem is a classical result. The author proves that if \(q(x)\) is strictly positive and if a certain inequality is satisfied by the complex parameter \(\beta\) of the Robin interface conditions, the usual iterative algorithm applied to the decomposed finite element problem converges to the finite element solution. Also some conditions on subdomain sizes and on \(\beta\) are given, which minimize the spectral radius of the iteration matrix of the method. Moreover, the author suggests to apply some special accelerating iterative procedure and a method of automatic choice of the efficient parameter \(\beta\) in order to make iterative process faster. The paper contains numerical results.