Endpoint estimation for observations with normal measurement errors (Q2418000)

Let $X_1,\ldots,X_n$ be IID random variables with a continuous distribution function $F_X$. The authors consider the problem of estimating the (assumed finite) endpoint $\theta=\sup\{x : F_X(x)<1\}$ when each $X_i$ is contaminated with IID Gaussian measurement error, with each Gaussian random variable having mean zero and unknown variance $\sigma^2$. That is, the observed data are $Y_1,\ldots,Y_n$, where $Y_i=X_i+\epsilon_i$ and the $\epsilon_i$ are IID Gaussian with unknown variance. \par Under the assumption of the existence of a second-order expansion of $1-F_X(x)$ for $x$ near $\theta$ (which places $X$ in the domain of attraction of the Weibull distribution), the authors propose an estimator $\hat{\sigma}$ for $\sigma$, and the estimator $\hat{\theta}=Y_{n,n}-\hat{\sigma}\sqrt{2\log n}$ for $\theta$, where $Y_{n,n}$ is the maximum of the $Y_i$. Under some assumptions, asymptotic normality of the estimators $\hat{\sigma}$ and $\hat{\theta}$ is established. \par The authors also use simulations to investigate the finite sample behaviour of their proposed endpoint estimator, and the paper concludes with an application to long jump data.