No label defined (Q3976804)

[For the entire collection see Zbl 0733.00027.] Many important sequences of \(d\)-dimensional estimators \(\{X^{(n)}:\;n\to\infty\}\) have (in some open convex neighborhood \(U\) of the center \(x^*=\lim X^{(n)}\)) densities of the form \((\varepsilon=1/n)\) \[ \begin{aligned} p_ \varepsilon(x)dx &= (2\pi\varepsilon)^{-d/2}\exp(- \varepsilon^{-1} K(\varepsilon,x)-o(\varepsilon^ m))dx_ 1\dots dx_ d, \\ \varepsilon^{-1}K(\varepsilon,x) &= \varepsilon^{-1} K(x)+K_ 0(x)+ \varepsilon K_ 1(x)+\dots+\varepsilon^ m K_ m(x), \end{aligned} \] \(K(\cdot)\) is a rate function, attaining its minimum 0 in \(x^*\) with \(K''(\cdot)\) positive definite. The \(K_ j(\cdot)\) are smooth. (Saddlepoint approximations lead to such situations; \(M\)- estimators have this asymptotic form in regular cases.) A theory of asymptotic cumulant generating functions is developed with special regard to smoothness. 1) There is a uniform asymptotic expansion \[ \hbox{ln} \mathbb{E} \exp(\varepsilon^{-1} \vartheta X_ \varepsilon)=\varepsilon^{-1} \psi(\vartheta)+\psi_ 0(\vartheta)+\dots+ \varepsilon^ m\psi_ m(\vartheta)+o(\varepsilon^ m) \] when \(K_ j(\dots)\) is \((2m-2j)\)- smooth. The \(\psi_ j(\cdot)\) are then \((2m-2j)\)-smooth. \(\psi(\cdot)\), the Legendre transform of \(K(\cdot)\), is a \((2m+2)\)-smooth rate function just as \(K(\cdot)\) was assumed to be. 2) For given \(\psi(\cdot),\psi_ 0(\cdot),\dots,\psi_ m(\cdot)\) satisfying the smoothness conditions there exist unique \(K(\cdot)\), \(K_ 0(\cdot)\), \(\dots\), \(K_ m(\cdot)\) leading to these \(\psi(\cdot)\).