Split invariance principles for stationary processes (Q653310)

The optimal Wiener approximation of partial sums \(S_n\) of i.i.d. random variables with \(EX_1=0\), var \(X_1=1\) and \(E|X|^p<\infty\) for some \(p>2\), is the statement that after suitable enlarging of the probability space, there exists a Wiener process \(\{W(t), t\geq0\}\) such that \[ S_n=W(n)+o(n^{\frac{1}{p}}) \text{ a.s.} \] Similar results are obtained for a class of weakly dependent stationary sequences \(\{X_n,n\geq1\}\) satisfying \(E|X_1|^p<\infty\), \(2<p\leq4\). (See [\textit{W. Liu} and \textit{Z. Lin}, Stochastic Processes Appl. 119, No. 1, 249--280 (2009; Zbl 1157.60311)] and [\textit{W. B. Wu}, Ann. Probab. 35, No. 6, 2294--2320 (2007; Zbl 1166.60307)]). The proofs do not work for \(p>4\) and no existing method for dependent approximation yields a.s. rate better than \(o(n^{\frac{1}{4}})\). In this paper, the authors prove that when including in the approximation the second Wiener component \(W_2(t)\), whose scaling factor is smaller than that of \(W_1(t)\), it is possible to obtain more exact error terms. Let \(\{Y_k, k\in\mathbb{Z}\}\) be a random process, let \(p\geq1\) and \(\delta(m)\rightarrow0\), The processes considered here are weakly \(M\)-dependent in \(L^p\) with rate function \(\delta(\cdot)\), i.e., they are such that: {\parindent=8mm \begin{itemize}\item[(A)]For any \(k\in\mathbb{Z}\), \(m\in\mathbb{N}\) one can find a random variable \(Y_k^{(m)}\) with finite \(p\)th moment such that \[ \|Y_k-Y_k^{(m)}\|_p\leq\delta(m). \] \item[(B)] For any disjoint intervals \(I_1, \ldots, I_n\) of integers and any positive integers \(m_1,\ldots,m_r\) the vectors \(\{Y_j^{(m_1)},j\in I_1\},\ldots,\{Y_j^{(m_2),j\in I_r}\}\) are independent provided the distance \(d(I_k,I_l)>\max\{m_k,m_l\}\) for \(1\leq k<l\leq r\). \end{itemize}} Hence, sequences satisfying conditions (A) and (B) are approximable in the \(L^p\) sense by \(m\)-dependent processes of any fixed order \(m\geq1\) with termwise approximation error \(\delta(m)\). The main results are as follows. \hangindent 12pt Theorem 1. Let \(p>1\), \(\eta>0\) and let \(\{Y_k,k\in \mathbb{Z}\}\) be a centered stationary sequence, weakly \(M\)-dependent in \(L^p\) with rate function \(\delta (m)\leq m^{-A}\), where \(A>\frac{p-2}{2 \eta}\left(1-\frac{1+\eta}{p}\right)\vee 1\); \(\frac{1+\eta}{2}<\frac{1}{2}\). Then the series \(\sigma^2=\sum\limits_{k\in \mathbb{Z}}EY_oY_k\) is absolutely convergent, and \(\{Y_k, k\in\mathbb{Z}\}\) can be redefined on a new probability space together with two Wiener processes \(\{W_1(t), t\geq0\}\) and \(\{W_2(t),t\geq0\}\) so that \[ \sum^n_{k=1} Y_k=W_1(s_n^2) + W_2(t_n^2)+O(n^{\frac{1+\eta}{p}}) \text{ a.s.} \] where \(\{S_n\}\) and \(\{t_n\}\) are nondecreasing sequences with \(s_n^2\sim \sigma^2_n\), \(t_n^2\sim c_n^{\gamma}\) for some \(0<\gamma<1\), \(c>0\). \hangindent 12pt Theorem 2 contains a similar statement for a weakly \(M\)-dependent process \(\{Y_k, k\in\mathbb{Z}\}\) with \(p>2\) and for the rate function \[ \delta(m)\ll \exp\{-\rho m\}, \quad \varrho>0. \] The convergence rate here is \(O(n^{\frac{1}{p}}\log^2n)\). Examples of weakly \(M\)-dependent processes are discussed in Section 3. \(m\)-dependent processes, NED processes, linear processes, nonlinear time processes, augmented GARCH sequences, ergodic sums and strong approximation problems, with applications of Theorems 1 and 2, are considered for them. The purpose of Sections 4 and 5 is to show how to deal with a large class of limit theorems of weakly dependent sequences via the approximation results of Section 2. In Section 4 the order of magnitude of \[ \max\limits_{1\leq k\leq n-a_n}\max\limits_{1\leq l\leq a_n}\left|\sum\limits_{j=k+1}^{k+l}Y_j\right| \] is considered. Section 5 deals with change-point tests with an epidemic alternative. The main theorems are proved in Section 6.