Testing and estimating change-points in the covariance matrix of a high-dimensional time series (Q2306269)

Consider the vector time series \(Y_i(a)=(Y^{(1)}_i(a),\dots,Y^{(d)}_i(a))^T\) with coordinates \[ Y^{(\nu)}_i(a)=\sum_{j=0}^\infty a^{(\nu)}_j \epsilon_{i-j}, \] where \(\nu=1,\dots,d\); \(i=1,\dots,n\). The coefficients \(\{a^{(\nu)}_j\}\) satisfy some decay assumption, the zero mean errors \(\{\epsilon_j\}\) satisfy restrictions on the second and third moments, \(d\) stands for a dimension which may be very high. This framework describes a large class of multivariate linear processes, including vector autoregressive moving average models, and spiked covariance models. Change point occurs at time point \(\tau=[n\theta]\), \(0<\theta<1\), if \(Y_i=Y_i(b)\) at \(i \leq \tau\) and \(Y_i=Y_i(c)\) at \(i > \tau\) with different coefficients \(\{b^{(\nu)}_j\}\) and \(\{c^{(\nu)}_j\}\). Detecting procedure does not estimate coefficients \(\{b^{(\nu)}_j\}\) and \(\{c^{(\nu)}_j\}\) but it is assumed that \(v^T \mathrm{Var}(Y(b))w \ne v^T \mathrm{Var}(Y(c))w\) for some vectors \(v, w\) which can be determined by principal component analysis. Therefore, a possible change point is detected by the CUSUM-type statistic \[ \max_{1 \leq k < n} n^{-1/2} v^T(S_k-(k/n) S_n)w, \] where \(S_k=\sum_{i \leq k} Y_i Y_i^T\). Asymptotic estimates (as \(n \to \infty\)) are proved. Simulation results are given.