Convergence to a Gaussian limit as the normalization exponent tends to 1/2 (Q806155): Difference between revisions

Consider, for example, the quadratic form \(Y_ n=\sum^{N- 1}_{j=0}\sum^{N-1}_{k=0}a_{j-k}X_ jX_ k\), where both the coefficients \((a_ k)\) and the covariances of the moving averages \(X_ 0,X_ 1,..\). are regularly varying. Suppose the \(X'_ k\)'s have spectral density \(f(x)=| x|^{-\alpha}L_ 1(x)\), \(\alpha <1\), and the \((a_ k)\) are the Fourier coefficients of a function \(g(x)=| x|^{-\beta}L_ 2(x)\), \(\beta <1\), where \(L_ 1\) and \(L_ 2\) are slowly varying at zero. It is the sum \(\alpha +\beta\) that determines the asymptotic behavior of the quadratic form \(Y_ N\) (if \(N\to \infty)\). The limiting random variable Y(\(\alpha\),\(\beta\)) can be represented as a double Wiener-Itô integral. If \(\alpha +\beta <1/2,\) then the limit is Gaussian; in the case \(\alpha +\beta >1/2\) convergence to the Gaussian fails. In Theorem 1.1 it is shown, that the behavior of Y(\(\alpha\),\(\beta\)) near the boundary \(\alpha +\beta =1/2\) is Gaussian: \[ \lim_{\alpha \to \alpha_ 0;\beta \to \beta_ 0}(\alpha +\beta - 1/2)^{1/2}Y(\alpha,\beta)=^{d}N(0,\sigma^ 2_{\alpha_ 0,\beta_ 0}),\quad \alpha_ 0+\beta_ 0=1/2. \]