Necessary and sufficient conditions for limit theorems for quadratic variations of Gaussian sequences (Q2417012)

In this paper, the author develops a general and elementary approach to handle the asymptotic behavior of quadratic variations of Gaussian processes. The object of study is \( \| Y^n \|^2:= \sum_{k=1}^n (Y^{(n)}_k)^2\) , where \(Y^n=(Y^{(n)} _1, \dots, Y^{(n)}_n) \), \(n\ge 1\), is a sequence of centered Gaussian vectors. The limit in probability of the sequence \(\| Y^n \|^2\), if it exists, is called the quadratic variation of the sequence \(Y^n\). First, if the energy, defined as \(\lim_{n\rightarrow \infty} E ( \| Y^n\|^2)\) exists, then the quadratic variation exists if and only if the sequence \(Y^n\) has a \(2\)-planar variation, defined as the limit of \(\|\Gamma^{(n)}\|^2_2\), where \(\Gamma^{(n)}\) is the covariance matrix of \(Y^n\) and \(\| \Gamma^{(n)} \|_2\) denotes its Euclidean norm. Clearly, the convergence is also in \(L^p\) for any \(p\ge 1\). Using the concentration inequality for Gaussian measures and the Borell-Cantelli lemma, the author obtains the following two sufficient conditions for the almost sure convergence to zero of the centered quadratic variations: \(\sup_{n\ge 1} E(\| Y^n\|^2)<\infty\) and \(\| \Gamma^{(n)} \|_2 = o( 1/ \log n)\). The third type of results are concerned with the Gaussian fluctuations of the normalized and centered quadratic variations \[ Z_n= \frac { \| Y^n \|^2 - E(\| Y^n \|^2)} { \sqrt{\mathrm{Var} (\| Y^n \|^2 - E(\| Y^n \|^2))}}. \] Applying the Fourth-Moment theorem by the reviewer and \textit{G. Peccati} [Ann. Probab. 33, No. 1, 177--193 (2005; Zbl 1097.60007)], necessary and sufficient conditions for the convergence in distribution of \(Z_n\) to the law \(N(0,1)\) are given. In particular, a simple condition in the case of a finite non-zero energy, is \(\lambda^*(n):=\max_{1\le k\le n} | \lambda_k^n| =o( n^{-1/2})\), where \(\lambda^n_k, 1\le k\le n\), are the eigenvalues of \(\Gamma^{(n)}\). Furthermore, in this context one can deduce a Berry-Esseen bound of the form \(C \sqrt{n} \lambda^*(n)\). These general results are applied in Section 3 to quadratic variations of Gaussian processes. That is, if \(X= (X_t)_{t\in [0,T]}\) is a given centered Gaussian process, the sequence \(Y^n\) is, in this case, \[ Y^{(n)}_k= \frac { X_{t_k}- X_{t_{k-1}} } { \sqrt{\phi(t_k-t_{k-1})}}, \qquad k=1,\dots, N(n), \] where \(\pi_n= \{ 0=t^n_0 < t^n_1 < \cdots <t^n_{N(n)} =T\} \) is a sequence of partitions of the interval \([0,T]\) and \(\phi\) is a given function. The case of second order quadratic variations is also considered. In the last part of the paper, the previous results are applied to a list of basic examples of Gaussian stochastic processes, where known important results are recovered. The examples include Brownian motion, fractional Brownian motion, sub-fractional Brownian motion and bifractional Brownian motion. The study is focused on first-order quadratic variations for which criteria for almost sure convergence are obtained, together with Berry-Essen bounds for the convergence of the distribution function of the normalized and centered quadratic variations to the normal distribution function.