On polynomial-based projection indices for exploratory projection pursuit (Q2638692)

For p-variate data Y and a (projecting) vector \(\theta \in R^ p\), \textit{J. H. Friedman} [J. Am. Stat. Assoc. 82, 249-266 (1987; Zbl 0664.62060)] proposed an index of interestingness of \(\theta\). Y as follows: Projected data \(\theta\). Y are transformed by standard normal distribution to \(U_{\theta}=\ell.\Phi (\theta.Y)-1\) which has uniform distribution on (-1,1) iff Y has a p-variate normal distribution. As the index I, an integrated squared difference of the densities of \(U_{\theta}\) and \(\ell /2\) is then taken. Difficulties with such an index appear for distributions having tails heavier than the normal distributions. Therefore the author has suggested a different index J, namely the directly (i.e. without transformation) taken integral of the squared difference of the densities of \(\theta\). Y and the normal density. The present author shows that if we denote by \(\theta_ 1\) a theoretical most interesting direction (under the assumption that data are distributed according to F and the index I is considered) and by \({\hat \theta}{}_ 1\) its estimate (which is derived in the paper via Fourier coefficients of the estimate of the density \(f(x)=dF/dx)\) then there exists a continuous-path, zero-mean Gaussian process and a continuous function c(\(\theta\)) such that if \(\theta^*\) is a point maximizing \(\xi^ 2(\theta)c(\theta)\), then \[ n^{1/2}({\hat \theta}_ 1- \theta_ 1)\to \xi (\theta^*)\times (\theta^*)\theta^*\text{ in } distribution. \] An analogous result is derived for the index J. The estimates \(\hat I\) and \(\hat J\) of I and J, respectively, are also discussed.