Symmetrizing and unitizing transformations for linear smoother weights (Q1861596): Difference between revisions

The author considers i.i.d. smoothers \(\hat m(X_i)=\sum_{j=1}^nw_{ij}Y_j\) as estimators of regression functions \(m(x)=E(Y_i |X_i=x)\) for i.i.d. data \((X_i,Y_i)_{i=1}^n\). It is proposed to improve any given weights \(w_{ij}\) (e.g., of Nadaraya-Watson or smoothing spline estimators) by changing them to \(w_{ij}^0\) which minimize \(\sum_{i,j=1}^n(w_{ij}-w_{ij}^0)^2\) subject to the restrictions \(w^0_{ij}=w^0_{ji}\) and \(\sum_{j=1}^n w_{ij}^0=1\), i.e., \(w_{ij}^0\) should be symmetric and unitized. An iterative algorithm is described for approximate calculation of \(w_{ij}^0\). Numerical examples are presented.