Minimax large deviations risk in change-point problems (Q1376541)

Let \(\theta\) be a one-dimensional parameter which belongs to the interval \((-T,T)\) where \(T\) is a given positive constant and consider the estimation problem of \(\theta\), in change-point models, from the observations \(X^{(n)}\) which satisfy the Itô equation \[ \dot X^{(n)}(t)= AI(t>0)+ \frac{\sigma}{\sqrt{n}} \dot W^{(n)}(t), \quad -\infty< t<\infty. \] Here \(A\) and \(\sigma\) are given positive constants and \(\dot W^{(n)}\) is a standard Gaussian white noise. For an arbitrary estimator \(\widehat{\theta}_n\) of \(\theta\) obtained from \(X^{(n)}\) and a fixed constant \(c (>0)\) define the minimax Bahadur-type risk \[ \beta_n(c)= \inf_{\widehat{\theta}_n} \sup_{|\theta|< T}n^{-1} \log_\theta^{(n)} (|\widehat{\theta}_n- \theta|\geq c). \] This risk evaluates the Cramér-type large deviations of estimators and closely relates to the optimal confidence intervals if one considers \((\widehat{\theta}_n- c,\widehat{\theta}_n+c)\) as the confidence interval for \(\theta\). The authors proves that \[ \lim_{n\to\infty} \beta_n(c)= -A^2c/(4\sigma^2) \quad\text{for any }c<T. \] In addition, a more general discrete change-point model is also analyzed.