Convergence analysis for the Cauchy problem of Laplace's equation by a regularized method of fundamental solutions (Q609547)

A Cauchy problem for the Laplace equation in a neighborhood of a circle \( \Gamma = \{ x \in \mathbb{R}^2 \, | \, | x | = \varrho \} \) is considered. Here \( \varrho > 0 \), and \( | \cdot | \) denotes the Euclidean distance in \( \mathbb{R}^2 \). The outer problem under consideration is \( \Delta u = 0 \) in \( \Omega_1 = \{ x \, | \, \varrho < | x | < r_1 \} \), and the inner problem is \( \Delta u = 0 \) in \( \Omega_2 = \{ x \, | \, r_2 < | x | < \varrho \} \), with \( 0 < r_2 < \varrho < r_1 \). In both cases the boundary conditions are \( u|_\Gamma = f, \;\frac{\partial u}{\partial \nu}|_\Gamma = g \), where \( \frac{\partial u}{\partial \nu} \) denotes the outer normal derivative of \( u \) on \( \Gamma \). It is supposed that perturbed data \( f^\delta, g^\delta \) are available with \( | f^\delta(x) - f(x) | \leq \delta, \;| g^\delta(x) - g(x) | \leq \delta \) for \( x \in \Gamma \). The authors consider a Tikhonov regularized collocation method which is based on the method of fundamental solutions. Here the approximate solution \( u_{\alpha,N}^\delta \) is of the form \( u_{\alpha,N}^\delta(x) = q_\alpha^\delta + \sum_{s=1}^2 \sum_{j=0}^{N-1} (\lambda_\alpha^\delta)_{s,j} G(x,y_{s,j}) \), where \( \alpha \) denotes the regularization parameter, and the coefficients \( \{(\lambda_\alpha^\delta)_{s,j}\} \) and \( q_\alpha^\delta \) are specified below. In addition, \( G(x,y) = - \tfrac{1}{2\pi} \log | x -y | \) denotes the fundamental solution of the Laplace equation, and the outer source points \( \{y_{1,j}\} \) and the inner source points \( \{y_{2,j}\} \) are uniformly distributed on circles \( \{ x \in \mathbb{R}^2 \, | \, | x | = R_1\} \) and \( \{ x \in \mathbb{R}^2 \, | \, | x | = R_2\} \), respectively. Here \( R_1, R_2 \) are real numbers such that \( 0 < R_2 < \varrho < r_1 < R_1 \) for the outer problem, and \( 0 < R_2 < r_2 < \varrho < R_1 \) is required for the inner problem. The collocation coefficients \( q_\alpha^\delta \) and \( \{(\lambda_\alpha^\delta)_{s,j}\} \) are determined by using Tikhonov's regularization method applied to the following perturbed system of \( 2N + 1 \) linear equations for the \( 2N+1 \) unknowns \( q \) and \( \{\lambda_{s,j}\} \): \( \sum_{s=1}^{2} \sum_{j=0}^{N-1} \lambda_{s,j} = 0 \) (an invariance condition) and the collocation equations \( u_N(x_i) = f^\delta(x_i), \;\frac{\partial u_N}{\partial \nu}(x_i) = g^\delta(x_i) \) for \( i = 0,1,\ldots, N-1 \). Here \( u_N(x) = q + \sum_{s=1}^2 \sum_{j=0}^{N-1} \lambda_{s,j} G(x,y_{s,j}) \), and \( \{x_{i}\} \) are uniformly distributed grid points on \( \Gamma \). It is shown that if the functions \( f \) and \( g \) have complex Fourier series expansions with coefficients that decay exponentially, and if in addition the Tikhonov parameter satisfies \( \alpha = c \delta \) for some constant factor \( c > 0 \), and if \( N(\delta) = c_2 | \log \sqrt{\delta} | \) for some \( c_2 > 0 \) specified in the paper, then for some \( 0 < a < 1 \) and each \( r \) with \( a\varrho < r < \varrho/a \) there holds \( \sup_{| x | = r} | u_{\alpha,N(\delta)}^\delta(x) - u(x) | = \mathcal{O}(\sqrt{\delta} | \log \sqrt{\delta} | + \gamma^{N(\delta)}) \) as \( \delta \to 0 \), with some \( 0 < \gamma = \gamma(r) < 1 \). Finally, results of some numerical experiments are presented.