Asymptotics of partial sums of linear processes with changing memory parameter (Q2393664): Difference between revisions

Let \(\{\zeta_s,s\in \mathbb{Z}\}\) be a standardized stationary and ergodic martingale difference. Let \(d(\tau)\) \((\tau\in(0,1])\) be a function and \(\{\psi_j(d)\}\) be a parameter class. As models of linear processes with nonconstant long memory or those of locally stationary long-memory, we usually consider \[ X_t=\sum^\infty_{j=0}\psi_j(d_t)\zeta_{t-j}\;(t=1, \dotsc,n) \] and classes of invertible time-varying fractional integrated processes of type \[ X^a_t=\sum^\infty_{j=0}a_j(t)\zeta_{t-j}. \] In this paper, the authors focus on the study of asymptotics of partial sums of ``truncated'' versions of the above mensioned processes \[ S_n(\tau)= \sum^{[n\tau]}_{t=1}X_t=\sum^{[n\tau]}_{t=1}\sum^{t-1}_{j=0} \psi_j(d_t)\zeta_{t-j}. \] Among others, the authors prove that, under some set of conditions on \(\{d_t\}\), \(d_+,C_1\) and \(\psi_j(d)\) \[ \frac {\log n}{n^{d+\frac 12}}S_n{\overset {D[0,1]}\longrightarrow}C_1 J_{d+}(\tau)\quad\mathrm{and}\quad\frac{\overline X_n}{\sqrt{\overline X_n}}{\overset\mathrm{law}\longrightarrow}N(0,1)\;(n\to\infty) \] hold, where \(J_{d_+}\) is a type II fractional Brownian motion defined by \[ J_{d_+}(\tau)=\int^\tau_0(\tau-x)^{d_+}W(dx) \] and \(\{W(dx):x \in[0,1]\}\) is a standard Gaussian white noise with zero mean and variance \(dx\) and that, under another set of conditions, \[ \left(\frac {\log^{\frac 2\gamma}n}{n}\right)^{d_{\max}+\frac 12}S_n\left( \tau_{\max}+\frac{\tau}{\log^{\frac 1\gamma}n}\right){\overset{\mathcal D}{(R)}\longrightarrow}U(\tau)\;(n\to\infty) \] holds, where the limit process \(U(\tau)\) is defined as a stochastic integral with respect to a Gaussian white noise \(\eta(du)\) on the line with zero mean and variance \(du\). Analogous asymptotics for time-varying fractionally integrated processes are also obtained.