Numerical differentiation of 2D functions from noisy data. (Q1416390)

The authors deal with the problem of approximating the gradient of a bivariate function \(f\) when a set of noisy data of \(f\) are given. The noise has expected value \(0\) and an unknown covariance matrix. The proposed method consists of two steps. First, the data are smoothed by an orthogonal multiresolution analysis of \(L^2(\mathbb{R}^2)\). The orthonormal scaling function is the tensor product \(\phi(x,y)= \varphi(x)\varphi(y)\) where \(\varphi\) is a scaling function of a coiflet. Then, using the smoothed data and usual difference operators, the authors construct a set of values approximating the gradient. From this a gradient estimator is obtained. The asymptotic properties of this gradient estimator are studied and convergence results are provided. An extension to scattered data is presented too. Numerical examples illustrate this method.