\(E\)-optimal designs for polynomial regression (Q2366747): Difference between revisions

The authors prove a conjecture of \textit{F. Preitschopf} [``Bestimmung optimaler Versuchspläne in der polynomialen Regression''. (1989; Zbl 0719.62087)] that \(E\)-optimal designs for polynomial regression with constant variance in the interval \([-1,1]\) are supported by Chebyshev points. The proof uses a certain duality between \(E\)-optimality and \(c\)- optimality and the experience that the Chebyshev polynomial \(\sum^ d_{i=0}a_ ix^ i=a'f(x)\) plays an essential role for \(c\)- optimality. The main result is that a \(c\)-optimal design w.r.t. \(c=KK'a\) is also \(E\)- optimal w.r.t. the parameter subsystem \(K'\Theta\). This design is supported by the Chebyshev points \(s_ i=\cos((d-i)\pi/d)\); \(i=0,1,\dots,d\); with weights \(w_ i=w_{d-i}=(-1)^{d-i}u_ i/\| K'a\|\) where the \(u_ i\) are determined from \(\sum^ d_{i=0}u_ if(s_ i)=KK'a\).