Deviation inequalities and moderate deviations for estimators of parameters in bifurcating autoregressive models (Q405492)

A general \(p\)th-order asymmetric bifurcating autoregressive process with unknown parameters \[ \theta= \begin{pmatrix} a_0 & a_1 &\cdots & a_p\\ b_0 & b_1 &\cdots &b_p\end{pmatrix}^t \] can be written in the abbreviated vector form \[ \mathbb{X}_{2n}= A\mathbb{X}_n+ \eta_{2n},\qquad \mathbb{X}_{2n+1}= B\mathbb{X}_n+ \eta_{2n+1}, \] for all \(n\geq 2^{p-1}\). Here, \(\mathbb{X}_n= (X_n, X_{[n/2]},\dots, X_{[n/2^{p-1}]})^t\) is the regression vector, \(\{(\varepsilon_{2k}, \varepsilon_{2k+1})\}\) is a noise process, \(\eta_{2n}=(a_0+ \varepsilon_{2n})e_1\) and \(\eta_{2n+1}= (b_0+ \varepsilon_{2n+1})e_1\) with \(e_1= (1,0,\dots, 0)^t\in\mathbb{R}^p\), and \[ A= \begin{pmatrix} a_1 & a_2 &\cdots & a_p\\ 1 & 0 &\cdots & 0\\ 0 &\cdot &\cdot &\cdot\\ 0 &\cdot & 1 &\cdot\end{pmatrix}\quad\text{and}\quad B= \begin{pmatrix} b_1 & b_2 & \cdots & b_p\\ 1 & 0 &\cdots & 0\\ 0 &\cdot &\cdot &\cdot\\ 0 &\cdot & 1 &\cdot\end{pmatrix}. \] For all \(n\geq 1\), denote the \(n\)th generation by \(\mathbb{G}_n= \{2^n, 2^n+1,\dots, 2^{n+1}-1\}\) and put \(\mathbb{F}_n= \bigcup^n_{k=0} \mathbb{G}_k\). Let \[ S_n= \sum_{k\in\mathbb{T}_{n,p-1}} \begin{pmatrix} 1 & \mathbb{X}^t_k\\ \mathbb{X}_k & \mathbb{X}_k\mathbb{X}^t_k\end{pmatrix}\quad\text{and}\quad \Sigma_n= I_2\otimes S_n, \] where \(\otimes\) stands for the matrix Kronecker product. Let \(\widehat\theta_n\) (\(n\geq p\)) be the least-squares estimator of \(\theta\). Then, we can write \[ \widehat\theta_n- \theta= \Sigma^{-1}_{n-1} \sum_{k\in\mathbb{T}_{n-1,p-1}} (\varepsilon_{2k}, \varepsilon_{2k}\mathbb{X}_k, \varepsilon_{2k+1}, \varepsilon_{2k+ 1}\mathbb{X}_k)^t. \] In this paper, dividing driven noises into two cases (the i.i.d.\ case and the conditionally independent case), the authors prove, among others, that deviation inequalities of the type \[ \operatorname{P}(\|\widehat\theta_n- \theta\|> \delta)\leq Ce^{C_n(\delta)} \] hold and that, under suitable conditions, \((\sqrt{|\mathbb{T}_{n-1}|}(\widehat\theta_n- \theta)/v_{|\mathbb{T}_{n-1}|})_{n\geq 1}\) \((v_n\uparrow\infty)\) satisfies the moderate deviation principle. The estimation of the parameter in the noise process is also considered.