High-dimensional change-point detection under sparse alternatives (Q2313279): Difference between revisions

This paper deals with the detection of a change-point in the \(d\)-dimensional Gaussian model \[ X_i = \theta + \Delta \theta_{\tau} \1_{i > \tau} + \xi_i, \qquad 1 \leq i \leq n, \] where \(\theta \in \mathbb R^d\) is a baseline parameter, \(\Delta \theta_{\tau}\) is a jump vector, \(\tau \in \left\{1,\dots,n-1\right\}\) is the change-point location, and \(\xi_i \sim \mathcal N \left(0, I_d\right)\) are i.i.d. standard Gaussian vectors. The standing assumption is that the number of non-zero components of \(\Delta \theta_{\tau}\) is \(p\), and that neither \(p\), nor \(\tau\) or \(\text{supp} \left(\Delta \theta_{\tau}\right)\) are known. The authors study the behavior of the corresponding testing problem as \(d\), \(n\) and \(p\) tend to \(\infty\), distinguishing between high sparsity (\(p \asymp d^{1-\beta}\) with \(\beta \in \left(1/2,1\right)\)) and low sparsity (\(p \asymp d^{1-\beta}\) with \(\beta \in \left(0,1/2\right)\)), by deriving minimax separation rates. The upper bounds are derived using a \(\chi^2\)-based test, which is also investigated in simulations and on real data.