Asymptotically Optimal Solutions in the Change-Point Problem (Q4954385)

Let \(X = (x_1,x_2, \ldots, x_{\theta-1}, x_{\theta}, \ldots)\) be a given sequence of independent observations, the first \(\theta - 1\) of them having distribution \(G_1\) and the others having \(G_2 \neq G_1\), \(\theta\) being unknown. Let \(P_t\) denote the distribution on the space of sequences \(X\) when \(\theta = t\) and \(E_t\) be the expectation. Section 1 of the paper shows that, for the problem of testing the hypotheses \(H_1 = \{\theta = j \}\), the maximum likelihood criterion (estimator) is asymptotically optimal within the class of ''asymptotically homogeneous'' tests (as \(\theta \to \infty\)). In Section 2, an estimate of the ''average likelihood'' is proved to be optimal for \(\theta\) in the mean-square sense. In both sections, limit distributions of estimators are found which are asymptotically Bayes in a wide class of prior distributions. Sections 3 and 4 treat (sequential) procedures of finding stopping times \(\tau\) (for \(X\)) adequate for soonest alarm after the change point \(\theta\). Under the restrictions \(\lim_{\theta \to \infty} P_{\theta}(\tau < \theta) \leq \varepsilon\) for the class of ''asymptotically homogeneous'' stopping times the problem considered is to find \(\tau\) rendering \(\varphi_t(\tau) = E_t(\tau - t |\tau \geq t)\) minimal. The first order optimality is established for the stopping time obtained similarly to the cumulative sums method [see, e.g., \textit{E.S. Page}, Biometrika 41, 100-115 (1954; Zbl 0056.38002); \textit{G. Lorden}, Ann. Math. Statistics 42, 1897-1908 (1971; Zbl 0255.62067)] and to the method introduced by \textit{M. Pollak} [Ann. Stat. 15, 749-779 (1987; Zbl 0632.62080)]. Corresponding asymptotic representations for \(\varphi_t(\tau)\) as \(t \to \infty\) are obtained.