Central limit theorems for quadratic forms in random variables having long-range dependence (Q1071370)

Let f(x) and g(x) be integrable real symmetric functions on [-\(\pi\),\(\pi\) ] that are bounded on subintervals that exclude the origin. Let \(X_ 1,X_ 2,..\). be a mean zero stationary Gaussian sequence with spectral density f(x), and let \(...,-a_ 1,a_ 0,a_ 1,...\) be the Fourier coefficients of g(x). It is shown that the distribution of the normalized quadratic form \[ Z_ N=N^{- 1/2}\{\sum^{N}_{i=1}\sum^{N}_{j=1}a_{i-j}X_ iX_ j- E\sum^{N}_{i=1}\sum^{N}_{j=1}a_{i-j}X_ iX_ j\} \] converges to a normal distribution if there exist constants \(\alpha <1\) and \(\beta <1\) with \(\alpha +\beta <1/2\) such that for each \(\delta >0\), \(f(x)=O(| x|^{-\alpha -\delta})\) and \(g(x)=O(| x|^{- \beta -\delta})\) as \(x\to 0.\) Of particular interest are the cases where \(f(x)\sim x^{-\alpha}L_ 1(x)\) and \(g(x)\sim x^{-\beta}L_ 2(x)\) as \(x\to 0\) with \(L_ 1\) and \(L_ 2\) slowly varying. The exponents \(\alpha\) and \(\beta\) are allowed to be positive, zero or negative. The sequence \(\{X_ j\}\) is said to exhibit a long-range dependence when \(\alpha >0\). When \(\alpha <0\), the covariances \(r_ k=EX_ jX_{j+k}\) satisfy \(\sum^{+\infty}_{k=- \infty}r_ k=0.\) Suppose \(f(x)\sim x^{-\alpha}L_ 1(x)\) and \(g(x)\sim x^{-\beta}L_ 2(x)\) as \(x\to 0\). \textit{M. Rosenblatt} [Proc. 4th Berkeley Sympos. math. Statist. Probab. 2, 431-443 (1961; Zbl 0105.118)] showed that in the special case \(1/2<\alpha <1\) and \(a_{i-j}=\delta_{ij}\), the quadratic form \(\sum^{N}_{i=1}\sum^{N}_{j=1}a_{i-j}X_ iX_ j\), adequately normalized, converges to a non-normal distribution. The assumption \(a_{i-j}=\delta_{ij}\) implies g(x) constant and thus \(\beta =0\). The result of the paper shows that the normalized quadratic form \(Z_ N\) converges to a normal distribution when \(1/2<\alpha <1\) and \(\beta <1/2-\alpha <0\). If \(\alpha\leq 1/2\), it is even possible to choose \(\beta >0\) as long as \(\beta <\min (1/2-\alpha\), 1). Results of this type can be used in the study of the asymptotic behavior of maximum likelihood estimators related to the sequence \(\{X_ j\}\); see the authors' paper ''Large sample properties of parameter estimates for strongly dependent stationary time series.'' Ann. Stat. 14, 517-532 (1986). Examples of sequences \(\{X_ j\}\) satisfying \(f(x)\sim x^{- \alpha}L_ 1(x)\) that are of special interest include fractional Gaussian noise and fractional ARMA.