scientific article; zbMATH DE number 2172779
A nonlinear regression model \(y=\eta (\theta)+\varepsilon\), \(\theta \in \Theta\), \(\varepsilon \sim N(o,\sigma ^2I)\), is called intrinsically linear if by some reparametrization \(\beta =\beta (\theta)\) it can be transformed to a linear model, or equivalently, if its expectation surface \(\mathcal E = {\eta (\theta); \theta \in \Theta }\), is planar. In the general case \(\mathcal E\) is curved, and the geometrical idea of the paper is to approximate the model by a model \[ y= A\beta (\theta)+\alpha , \quad \theta \in \Theta , \qquad \varepsilon \sim N\bigl (0,\sigma ^2I\bigr), \] so as to minimize the distance of the expectation surfaces of the new and the original models \[ \operatorname {min}_{A,\alpha ,\beta (.)} \int _\Theta \,\Bigl \| \eta (\theta)-\bigl [A\,\beta (\theta) + \alpha \bigr ]\Bigr \| ^2\pi (\theta)\,d\theta. \] Here \(\pi (\cdot)\) is the density of the prior guess of the position of \(\theta \). The main result is that the dimension and the form of \(A\) is obtained by applying the method of principal components to the matrix \[ \text{Var}_\pi (\eta)=\int _\Theta \bigl [\eta (\theta)-E_\pi (\eta)\bigr ]\bigl [\eta (\theta) -E_\pi (\eta)\bigr ]^T \pi (\theta)\,d\theta, \] which itself can be approximated by Monte-Carlo simulations. Then \(\alpha \) and \(\beta (\theta)\) are obtained straightforwardly in terms of \(A\). Six examples illustrating this and other methods are presented.
This page was built for publication:
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4680165)