For numerical differentiation, dimensionality can be a blessing! (Q4240585)

A way of modelling observational data is that in terms of the following equations \[ y_j=f({\mathbf j}h)+ \varepsilon_j,\;h=1/n,\;{\mathbf j}= (j_1, \dots, j_d)\in\{0,1, \dots,n\}^d, \tag{1} \] where \(f\in C^p(\mathbb{R}^d)\) is a function which represents some class of unknown trend, and \(\varepsilon_j\) are the measurements errors which are independent and have identically distributed normal variables with expected value 0 and variance \(\sigma^2\). Let \(D\) denote a differentiation operator of the following type \(D=\sum_{|{\mathbf p} |\leq q} c_{\mathbf p}({\mathbf x}) {\partial^{|{\mathbf p}|} \over \partial^{\mathbf p} {\mathbf x}}\). This paper deals with a general averaging-differentiation method to evaluate the \(q\)-th order derivative \(Df({\mathbf x})\), where the function \(f:\Omega\subset \mathbb{R}^d\to\mathbb{R}\) is assumed to have sufficient smoothness and the \(c_{\mathbf p}\) are constants. The information on \(f\) to be used is that given by (1). Estimates of errors in terms of the Hessian of \(f\) are given in the paper. By considering the combination of both the numerical differentiation operator and the averaging operator it is shown how to construct stable numerical differentiation rules, whose implementation is presented in a section. The end of the article is devoted to applications and computational examples.