Gaussian limit fields for the integrated periodogram (Q2564698)

Let \(\{X_i, i=1, \dots,n\}\) be a sample from a moving average process; define the periodogram by \[ I_{n,[nx]} (\lambda): ={1\over n} \left|\sum^{[nx]}_{t=1} X_te^{-i\lambda t} \right|^2, \] where \(x\in[0,1]\), \(\lambda \in[-\pi, \pi]\). For a smooth function \(f\) the integrated periodogram is \(\int^\lambda_{-\pi} I_{n,[nx]} (y)f(y)dy\). This paper obtains a functional central limit theorem (FCLT) for the integrated periodogram of a moving average sequence. This is applied to obtain limiting distributions for statistics based on the integrated periodogram, in particular a statistic for changepoint detection. FCLTs are standard for empirical processes of i.i.d. variables, and the limiting process of the sequential empirical process is a Kiefer process, a two-parameter Gaussian process. The Gaussian limit field obtained by the authors is based on a suitably scaled version of the Kiefer process, \[ K(x,\lambda): =2\sum^\infty_{t=1} Y_t(x) {\sin(\lambda t) \over t}, \] where the \(Y_i\) are i.i.d. Wiener processes. Theorem 3.1 establishes the FCLT for integrated periodograms of i.i.d. \(X_i\) (of mean zero and \(EX^4= \sigma^4\), or symmetric stable of index \(p\in(1,2))\) and \(f\equiv 1\). As an example, Corollary 3.2 (a) implies that \[ n^{1/2} \int^\lambda_{-\pi} \left(I_{n,[nx]} (y)-{[nx] \over n} \sigma^2 \right)dy \to_d(\lambda+\pi) Y_0(x)+ K(x,\lambda). \] Here \(Y_0\) is a Wiener process independent of \(K\). Section 4 extends these results to general linear processes \(X\) (Theorem 4.3) with finite fourth moment. A key step is Lemma 4.1 where the periodogram based on \(X\) is related to the periodogram based on the underlying i.i.d. innovations. Section 6 applies the results of the paper to the problem of detecting changepoints in stock market indices (DAX and NIKKEI).