Uniformity of double saddlepoint conditional probability approximations (Q1383912): Difference between revisions

Let \({\mathbf T}= (t^1,\dots, t^d)\) be a random vector with a cumulative moment function \(K(\tau)\). Let a vector with the subscript \(-1\) denote the vector with first entry removed. Suppose that the multivariate saddlepoint \(\widehat{\theta}\) solves \(K^j(\widehat{\tau})= t^j\) for all \(j\in \{1,\dots, d\}\) and the reduced multivariate saddlepoint \(\widetilde{\tau}\) solves \(K^j(\widetilde{\tau})= t^j\) for all \(j\in \{2,\dots, d\}\), subject to \(\widetilde{\tau}_1=0\), \(\rho^{*2}=| K''(\widehat{\tau})/ K_{-1}'' (\widetilde{\tau})|\), \(\breve z_1= \widehat{\tau}_1 \rho^*\). Put \[ \widehat{w}_1= \sqrt{ 2[K(\widetilde{\tau})- \widetilde{\tau}{\mathbf t}+ \widehat{\tau}{\mathbf t}-K(\widehat{\tau})]}. \] Then, the double saddlepoint approximation is given by \[ F^\dag(t^1\mid{\mathbf t}_{-1})= \Phi(\sqrt{n} \widehat{w}_1)+ \frac{\varphi(\sqrt{n} \widehat{w}_1)} {\sqrt{n}} \Biggl[ \frac{1}{\widehat{w}_1}- \frac{1}{\breve z_1} \Biggr]. \] The author shows that under some conditions \[ F(t^1\mid{\mathbf t}_{-1})=F^\dag(t^1\mid{\mathbf t}_{-1})+\varphi(\sqrt{n} \widehat{w}_1)E_1({\mathbf t})/ (n\sqrt{n}) \] where \(E_1({\mathbf t})\) is bounded, and that the analogous result holds for a random row vector \({\mathbf T}\) confined to a lattice of form \({\mathbf a}+ {\mathbf Z}^d\). Some applications are also considered.