Inference on the change point under a high dimensional sparse mean shift (Q2219223)

The authors of the paper consider the time series model satisfying the following equation \[ X_t = \begin{cases} \theta_1+\varepsilon_t \ \text{ if }\ t=1,\ldots,\tau;\\ \theta_2+\varepsilon_t\ \text{ if }\ t=\tau+1,\ldots,T. \end{cases} \] Here the random vectors \(X_t=\left(X_{t1}, X_{t2},\ldots,X_{tp}\right)^\intercal\in\mathbb{R}^p\) are supposed to be observed for all \(t\in\{1,\dots, T\}\), while the random vectors \(\left(\varepsilon_{t1},\varepsilon_{t2},\ldots,\varepsilon_{tp}\right)^\intercal\in\mathbb{R}^p\) are supposed to be unobserved with zero means and sub-Gaussian or sub-exponential distributions. The model dimension \(p\) is allowed to be fixed or diverging much faster than the sampling period \(T\). A plug is studied in the least squares estimator for the change point parameter \(\tau\). The sufficient conditions are obtained under which this estimator possesses sufficient adaptivity against plug in estimates of mean parameters \(\theta_1,\theta_2\) in order to yield an optimal rate of convergence. Feasible algorithms for implementation of the proposed methodology are provided. Theoretical results of the paper are supported with Monte Carlo simulations.