Contingency table probability estimation - a projection pursuit approach (Q1297867)

Let us consider a \(d\)-dimensional ordered contingency table \(\{n_{I}\}\) with \(K=\prod_{j=1}^{d}K_{j}\) cells, where \(I=(i_{1},\ldots,i_{d})\). Let \[ P=\{p(I)\;|\;I=(i_{1},\ldots,i_{d}),\;1\leq i_{j}\leq k_{j}, 1\leq j\leq d,\;\sum_{I}p(I)=1,\;p(I)\geq 0\} \] be the probabilities of a \(K\)-cell multinomial. The projection pursuit estimator of \(p(I)\) is defined by \[ \hat p^{(k)}(I)=\hat p^{(0)}(I)\prod_{j=1}^{k} g_{\theta_{j}}^{(j)}(I)=\hat p^{(k-1)}(I)g_{\theta_{k}}^{(k)}(I), \] where \(\hat p^{(k)}(I)\) is the \(k\)-th estimate of \(p(I)\), and the augmenting factor \(g_{\theta_{j}}^{(j)}(I)\) depends on a unit vector \(\theta_{j}\in R^{d}\). The authors define the projection of given cell probability estimates along an arbitrary direction \(\theta\). Then they show how to find the optimal direction \(\theta_{k}\) and the augmenting factor \(g_{\theta_{k}}^{(k)}(I)\). Compared to the kernel estimator the projection pursuit estimator suffers less from ``the curse of dimensionality'' and works better for tables with discontinuities.