Modelling long-memory time series with finite or infinite variance: a general approach (Q2742773)

The conditions on some filter with a certain transfer function to obtain by linear filtering a second-order white noise are investigated. The authors study the family of generalized fractional filters, that is the family of filters whose transfer functions are sums of a finite number of functions \(F\) satisfying the assumption that there exists \(a>0\) such that \(F(z)\) is analytic in the open disk \(|z |<a.\) If \(a<\infty\), \(F(z)\) has a finite number of distinct singularities on the circle \(|z |= a (a e^{i\lambda_j})\) and for every \(j \in \{ 1, \ldots, m\}\) there exists \(d_j (-d_j \notin N)\) such that \(F(z) = \bigl( 1- ze^{-i\lambda_j}a^{-1}\bigr)^ {-d_j} h_j(z)\), where \(h_j\) is analytic in a neighborhood of \(a e^{i\lambda_j}\) and \(h_j (a e^{i\lambda_j}) \neq 0.\)NEWLINENEWLINENEWLINELet \(F(z)= \sum_{n=0}^\infty a_n z^n\) be the series expansion of \(F.\) The authors describe the asymptotic behaviour of the coefficients \(a_n.\) They use their results for obtaining, for example, the following assertion.NEWLINENEWLINENEWLINELet \(\{\varepsilon_n\}\) be a strong \(S\alpha S\) white noise with parameter \((0, 0, \alpha)\), where \(\alpha \in [0, 2]\) (\(E e^{it\varepsilon_i} = \exp\{ - |t|^{\alpha}\}\)) and let \(F\) be a GF filter with \(a\geq 1\), \(d = \max d_j < 1 - \alpha^{-1}.\) Then the series \(X_n = \sum_{j=0}^\infty a_j \varepsilon_{n-j}\) converge almost surely and the induced \((X_n)\) is an \(S\alpha S\) process with marginal parameter \(\bigl( 0, 0, \sum |a_n|^{\alpha}\bigr).\)NEWLINENEWLINENEWLINEOther conditions for the existence of the induced stationary \(S\alpha S\) processes are given. The asymptotic dependence structure of these processes via the codifference and the covariance sequences, respectively, is described.