Detection of multiple changes in a sequence of dependent variables (Q120317)

The author considers a vector \(\mathbf{Y} = (Y_i,\;1\leq i \leq n)\) of \(\mathbb{R}^p\)-valued random variables marginally distributed with \(P_{i,\theta (i/n)}\) where \(\theta (\tau)\) is a mapping from \((0,1]\) into a compact parameter space. The function \(\theta (\tau)\) is assumed to be piecewise constant in the \(K\) intervals defined by \(0<\tau _1 <\ldots <\tau _{K-1} < \tau _K =1\), with corresponding values \(\theta _1,\ldots ,\theta _K\). The problem is to estimate \(\mathbf {\tau } = (\tau _1,\dots ,\tau _{K-1})\) and \(\mathbf {\theta } = (\theta _1,\dots ,\theta _K)\) from the observations \(\mathbf{Y}\). This problem is solved by formulating assumptions under which for all \(1\leq j \leq K\) there exist consistent minimum contrast estimators \(\hat {\theta }_{n,j}(\mathbf{Y}_j)\) of \(\theta _j\) for \(\mathbf{Y}_j = (Y_i \:\;n\tau _{j-1} < i < n\tau _j)\), where the contrast functions are of the form \(W_n (\mathbf{Y}_j, \theta _j)\). The new contrast functions \[ J_n (\mathbf {\tau }, \mathbf {\theta })= W_n (\mathbf{Y}_1,\theta _1) + \ldots + W_n (\text\textbf{Y}_K, \theta _K) \] under the stated assumptions are shown to lead to consistent minimum contrast estimators \(\hat {\mathbf \tau }_n\), \(\hat {\mathbf \theta }_n\). The rate of consistency is evaluated, which is to some extent invariant w.r.t.\ the dependence structure of the observations \(\mathbf{Y}\).