Asymptotic representation of the process of the likelihood ratio in the case of a discontinuous density (Q1920703)

Let \(X_1,\dots,X_n\) be a real sample consisting of independent random variables (rvs) with common density \(f(x,\theta)\) with respect to the Lebesgue measure. Assume that the density is continuous in \(x\) everywhere except at a point \(x(\theta)\) which depends on an unknown parameter \(\theta\in \Theta\subset \mathbb{R}\) and at which the density has a discontinuity of the first kind: \[ 0\neq p(\theta)=f(x(\theta)+0,\theta)\neq f(x(\theta)-0,\theta)=q(\theta)\neq 0. \] We do not consider the case of an arbitrary but finite number of jumps, merely to obtain simpler conditions and formulas. Assuming the true value of the parameter \(\theta\) to be fixed, let us write \(p\), \(q\), and \(x'\) in place of \(p(\theta)\), \(q(\theta)\), and \(x'(\theta)\) (the prime denoting the derivative with respect to \(\theta)\). In the study of the asymptotic behavior of the maximum likelihood estimators, an important role is played by the process that equals the logarithm of the likelihood ratio for the suitably-normalized argument \(Y_n(u)=\sum L(X_i,\theta+u/n)\). Here the symbol \(\sum\) stands for the summation over the index \(i\) from 1 to \(n\). It was proven by \textit{I. A. Ibragimov} and \textit{R. Z. Khas'minskij} [Asymptotic theory of estimation. (1979); see the review of the English translation (1981; Zbl 0467.62026)] that, under appropriate assumptions, the distribution of the process \(Y_n(u)\) converges in some sense to the distribution of the process \[ Y(u)=u(p-q)x'+(\nu_+(px'u)-\nu_-(-qx'u))\text{ln}(q/p) \] as \(n\to \infty\), where \(\nu_\pm(u)\) are independent standard Poisson processes extended by zero for \(u<0\). The aim of the present article is to obtain the asymptotic representation \[ Y_n(u)=Y(u)+u\eta_n n^{-1/2}+O(\ln^2n/n),\;|u|\leq c\ln n, \] on a suitable probability space, which is true with probability \(\geq 1-O(\text{ln}^2n/n)\), where the sequence \(\{\eta_n\}\) of rvs is independent of \(Y\) and converges in distribution to the normal law with zero mean.