Numerical differentiation for two-dimensional scattered data (Q2573438): Difference between revisions

The authors present a method, based on Tikhonov regularization and a Green's function, for estimating the gradient of a 2-variable function \(f\) given by ``mildly scattered'' and slightly noisy data. That is, the data comprises values \(f(x_i,y_i) + \delta_i\) where the points \((x_i,y_i)\) are fairly evenly distributed over the domain of definition \(\Omega\), in a sense they make precise. They show that they compute the exact gradient of an approximating function \(\psi\), and as the maximum spacing of the points and the maximum \(| \delta_i| \) go to zero, the gradient of \(\psi\) converges in \(L_2\) to that of \(f\). Two numerical examples, one with data on a grid on a disc and one with a randomly perturbed grid on a rectangle, show that the method is quite effective. The main work of the algorithm depends on the domain and the points, not the data values. Therefore it will be quite efficient with multiple sets of data over the same points. The authors say the method applies equally in higher dimensions.