Testing the Goodness of Fit of Parametric Regression Models with Random Toeplitz Forms (Q4455914)

A regression model \(Y_{i,n}=g(t_{i,n})+\varepsilon(t_{i,n})\) is considered, where \(t_{i,n}\in[0,1]\) are fixed random points, \(g\) and Var\((\varepsilon(t))=\sigma^2(t)\) are unknown, smooth enough \(\varepsilon(t_{i,n})\) are independent and of zero mean. Estimates of \[ M^2=\min_{v\in U}\int_0^1(v(t)-g(t))^2\nu(dt) \] are considered, where \(U\) is a subset of \(L_2([0,1],\nu)\). Quadratic forms \(Q_n\) of \(Y_n=(Y_{1,n},\dots,Y_{n,n})\) with Toeplitz band matrices are considered as estimators for \(M^2\) in linear models generated by \(U= \text{span}[f_1,\dots,f_d]\). For the nonlinear models \(U=\{f_\vartheta,\vartheta\in\Theta\}\), \(\Theta\subseteq R^k\), it is proposed to use \[ \widehat{M}^2_n=\inf_{\vartheta\in\Theta} Q_n(Y_n-f_{\vartheta,n}). \] Tests for the hypothesis \(g\in U\) (i.e., \(M^2=0\)) are proposed based on \(\widehat M^2_n\). Results of simulations are presented.