Optimal detection of a hidden target: the median rule (Q424533)

Suppose we have an observation of a sample path of the continuous process \(X_t\) started at 0 and the problem is to detect when this sample path reaches a non-negative level \(\ell\) that is not directly observable -- \(X_t\) and \(\ell\) are assumed to be independent. The problem of detecting \(\ell\) with the help of the sequential observation of \(X_t\) has two distinct formulations. The space domain formulation, where the expected distance from \(X_{\tau}\) to \(\ell\) is minimized over all stopping times \({\tau}\) of \(X_t\), and the time domain formulation, where \(\tau_{\ell}\) denotes the first entry time of \(X_t\) at the level \(\ell\) and the expected distance from \({\tau}\) to \(\tau_{\ell}\) is minimized over all stopping times \({\tau}\) of \(X_t\). The optimal stopping time will depend on the chosen distance function, and it is not clear a priori how to select the most natural distance function in this respect. The main observation of the present paper is that there exists a single stopping time of \(X_t\) that is optimal in both the space domain and the time domain with respect to the expected Euclidean distance on the real line. More precisely, assuming that \({\operatorname E}\ell<\infty\) and \({\operatorname E}\tau_{\ell}<\infty\), and denoting by \(m\) the (lowest) median of \(\ell\), the author shows that the stopping time \(\tau_m=\inf\{t\geq0|X_t=m\}\) minimises simultaneously both \({\operatorname E}|X_{\tau}-{\ell}|\) and \({\operatorname E}|\tau-\tau_{\ell}|\) over all stopping times \({\tau}\) of \(X_t\). The author shows that two-sided versions of the same problem and variational refinements of both problems lead to new classes of median/quantile rules having a dynamic character. The latter analysis is based on the method of Lagrange multipliers in optimal stopping.