Detection of changes in linear sequences (Q1370546)

Assume that \(\{X_i,1\leq i\leq n\}\) satisfy the linear model \(X_i= \mu_i+e_i\), where the errors \(\{e_i,1\leq i\leq n\}\) are given by the linear process \(e_i=\sum_{0\leq j<\infty}a_j \varepsilon_{i-j}\), \(\{\varepsilon_i,\;-\infty <i< \infty\}\) are independent, identically distributed random variables with mean zero, positive variance \(\sigma^2=E\varepsilon^2_i\), and \(E|\varepsilon_i |^\nu <\infty\) for some \(\nu>2\). Assume also that the random variables \(\varepsilon_i\) have a smooth density function \(f\) that satisfies the condition \(\sup_{-\infty<s <\infty}|s|^{-1} \int^\infty_{-\infty} |f(t+s)- f(t)|dt<\infty\), while the coefficients \(a_j\) of the linear process \(e_i\) are such that \(a_j=O (j^{-\beta})\), as \(j\to\infty\), with some \(\beta>3/2\), and putting \(g(z)= \sum_{0\leq k<\infty} a_kz^k\), \(z\in\mathbb{C}\), assume that \(g(z)\neq 0\) for all \(|z|\leq 1\). Then \(Ee_i=0\) and \(\text{var}(e_i)= \sigma^2\sum_{0\leq k<\infty} a^2_k<\infty\). Given these conditions, one wishes to test the null hypothesis \[ H_0:\mu_1=\mu_2 =\cdots =\mu_n \] versus the alternative \[ H_A:\text{ there is an integer }k^*,\;1\leq k^*<n,\text{ such that } \mu_1=\cdots =\mu_{k^*}\neq \mu_{k^*+1}= \cdots= \mu_n. \] The tests for \(H_0\) versus \(H_A\) are to be based on large values of weighted sup-functionals of the sequence of stochastic processes \[ Z^0_n(t): =Z_m(t)-tZ_n(1), \quad 0\leq t\leq 1, \] where \[ Z_n(t)= \begin{cases} n^{-1/2} \sum_{1\leq i\leq (n+1)t}X_i, \quad & \text{if }0 \leq t<1,\\ n^{-1/2} \sum_{1\leq i\leq n} X_i,\quad & \text{if }t=1. \end{cases} \tag{1} \] For example (cf. Theorem 1.2), it is proved, assuming the above summarized conditions and that \(H_0\) holds true, that as \(n\to\infty\), one has \[ \sup_{0<t<1} \bigl|Z_n^0(t) \bigr|/q(t) @>{\mathcal D}>> \tau\sup_{0< t<1} \bigl|B(t)\bigr |/q(t)\quad \text{if and only if}\quad \lim\sup_{n\to \infty} \sup_{0<t<1} |B(t)|/q(t) <\infty, \] where \(\tau^2: =\sigma^2 (\sum_{0\leq j< \infty} a_j)^2>0\), \(q:(0,1)\to R_+\) is non-decreasing near zero and non-increasing near one, \(\inf_{\varepsilon\leq t\leq 1-\varepsilon} q(t)>0\) for all \(0<\varepsilon <1/2\), and \(\{B(t),\;0\leq t\leq 1\}\) is a Brownian bridge. It is of interest to note that the just quoted result, as well as the weighted weak convergence conclusion of Theorem 1.1 under the given conditions, coincide with the corresponding optimal ones when the \(Z_n(\cdot)\) process (1) is based on independent, identically distributed random variables having only two moments [cf., the reviewer, Stud. Sci. Math. Hung. 31, No. 1-3, 323-353 (1996; Zbl 0845.60032); ibid. 33, No. 1-3, 305-320 (1997; Zbl 0880.60029)]. Roughly speaking the same holds true for the Darling-Erdős type result [cf., \textit{D. A. Darling} and \textit{P. Erdős}, Duke Math. J. 23, 143-155 (1956; Zbl 0070.13806)] of Theorem 1.3 as well. The key to establishing the present results is the conclusion that under the given conditions the linear process \(\{e_i=\sum_{0\leq j<\infty}a_j\varepsilon_{i=j},\;-\infty<i <\infty\}\) is a stationary, strongly mixing sequence. Consequently, Theorems 1.1 -- 1.2 follow from invariance principles for tied down partial sums of such sequences that are established in the appendix of the paper. The proofs of Theorems 1.1 -- 1.3 are detailed in Section 2.