scientific article; zbMATH DE number 3909440

From MaRDI portal
Publication:3685759





Let X be the normalized sum of a random sample \(X_ 1,...,X_ n\) of size n from a distribution \(F: X=(X_ 1+...+X_ n- n\mu)/\sqrt{n}\sigma\). If n is sufficiently large, then by the central limit theorem the distribution of X is approximated by the standard normal distribution \(\Phi\). However, the approximation is poor when the distribution F is asymmetric. The Cornish-Fisher expansion provides us with a tool for improving the approximation. Let \(\alpha_ k\) be the k-th order moment of the normalized variable \((X_ 1-\mu)/\sigma\). The Cornish-Fisher transform \(y=y_ 1(x)=x+\alpha_ 3(1-x^ 2)/6\sqrt{n}\) gives an approximation to \(y=\) \(\Phi^{-1}F_ n(x)\), which means that the random variable \(Y=y_ 1(X)\) is approximately normally distributed. We can expect the approximation is better than the one obtained by the central limit theorem. However, as the behavior of \(y_ 1(x)\) is very wild for large \(| x|\), we can not still make \(\Delta =\sup_{x}| \Pr \{X\leq x\}-\Phi (y_ 1(x))|\) sufficiently small. In order to control the wildness of the behavior of \(y_ 1(x)\), we have to reduce the correction term \(\alpha_ 3/6\sqrt{n}\cdot (1-x^ 2)\) for large \(| x|\). To this end, we introduce a transform of x defined by \[ \xi_ 1(x)=\frac{\alpha_ 3}{6\sqrt{n}}+\frac{6\sqrt{n}}{\alpha_ 3}\{(1+\frac{\alpha_ 3x}{2\sqrt{n}})^{1/3}-1\}, \] and prove the inequality \[ \sup_{x}| \Pr \{X\leq x\}-\Phi (\xi_ 1(x))| \leq 0.95 \alpha^ 2_ 3/n+R_ n, \] where \(R_ n\) is the upper bound for the error of the Edgeworth expansion, which is known to be of order 1/n. We can further improve the approximation by using \(\alpha_ 4\) as well as \(\alpha_ 3\). Numerical computations show that the approximations substantially improve the Cornish-Fisher transforms.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3685759)