The problem is to find the approximation \(g(x) = u_ 0+{\mathbf w}^ T{\mathbf u}\) of the function \(y({\mathbf x})\) in the sense of a statistical criterion. \(\mathbf w\) represents a given base of functions and the parameters \(u_ 0\), \(\mathbf u\) of approximation are defined by the following extremum condition: \(\min_{u_ 0,{\mathbf u}}{\mathbf e}^ T{\mathbf e}/(1+{\mathbf u}^ TM{\mathbf u})\) where \({\mathbf e} = (e_ 1,\dots,e_ N)^ T\), \(e_ i = y_ i - g_ i({\mathbf x}_ i)\) and \(M\) is a positive semi-definite symmetric matrix. With \(M = 0\) the above condition leads to the least squares method; if \(M\) is an identity matrix this condition leads to so called ``orthogonal regression. For an arbitrary \(M\) the solution of the above extremum condition needs in general the use of some approximation methods. The author gives a special form of \(M\) which allows to obtain an analytic solution useful to the concrete applications. It is shown, that for certain test problem the proposed method improves the approximation given by the least squares method.
This page was built for publication: A method for approximating functions
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q1802635)
