On Gaussian approximation of Hilbert space valued discrete time martingales (Q1897889)

Let \(H\) be a separable Hilbert space with the inner product \((\cdot,\cdot)\) and with the norm \(|\cdot|\). We denote the space of \(H\)-valued random elements defined on a fixed probability space \((\Omega, {\mathcal F}, {\mathbf P})\) by \(L_0(H)\). We use \({\mathcal L}(X)\) to denote the distribution of \(X\in L_0(H)\). For each integer \(n\), let \(X_{n1}, \dots, X_{nk_n}\in L_0 (H)\) be a square integrable martingale difference sequence with respect to \(\sigma\)-fields \({\mathcal F}_{n0} \subset {\mathcal F}_{n1} \subset \cdots\subset {\mathcal F}_{nk_n}\), that is, suppose that \(X_{ni}\) is \({\mathcal F}_{n,i-1}\)-measurable, \({\mathbf E}_iX_{ni}=0\) a.s. and \({\mathbf E} |X_{ni} |^2 <\infty\) for all \(n\) and \(i=1,2, \dots, k_n\), where \({\mathbf E}_i\) denotes the conditional expectation with respect to \({\mathcal F}_{n,i-1}\). The conditional covariance operator \(Q_{ni}\): \(H\to H\) of \(X_{ni}\) given \({\mathcal F}_{n,i-1}\) is defined by \(Q_{ni}x= {\mathbf E}_i (X_{ni},x) X_{ni}\) for \(x\in H\), \(i=1, \dots, k_n\). Let \(Y\in L_0(H)\) be a zero mean Gaussian random element with the covariance operator \(Q\). Let \(q: H\to\mathbb{R}\) be a continuous seminorm on \(H\). Define \[ \Delta_{q,n}= \sup_{r>0} \biggl|{\mathbf P} \bigl\{q(S_n) <r\bigr\}-{\mathbf P} \bigl\{q(Y) <r\bigr\} \biggr|, \] where \(S_n= \sum^{k_n}_{k=1} X_{nk}\). Conditions under which the Hilbert space martingale CLT holds can be found in [\textit{H. Walk}, Z. Wahrscheinlichkeitstheorie Verw. Geb. 39, 135-150 (1977; Zbl 0342.62060) and \textit{A. Jakubowski}, Probab. Math. Stat. 9, No. 1, 95-114 (1988; Zbl 0669.60010)]. Using various sets of assumptions, many authors have derived bounds on the rate at which \(\Delta_{q,n}\) converges to zero [see, e.g., \textit{W. S. Rhee} and \textit{M. Talagrand}, J. Multivariate Anal. 20, 303-320 (1986; Zbl 0606.60010), the author, Lith. Math. J. 31, No. 3, 345-355 (1991) and Litov. Mat. Sb. 31, No. 3, 497-512 (1991; Zbl 0777.60027) and \textit{P. Gudynas}, ibid. 31, No. 1, 34-42 (1991) resp. ibid. 31, No. 1, 50-61 (1991; Zbl 0728.60011)]. In praticular, from Rhee and Talagrand (loc. cit.) the estimate \(\Delta_{q,n}= O(n^{-1/6})\) follows provided the following conditions are fulfilled: (A) \(q\) is three times Frechet differentiable on the set \(\{q(x) =1\}\) and \(\sup\{|q^{(k)} (x)|: q(x)=1\} \leq C<\infty\), for \(k=1,2,3\); (B) \(\sum^{k_n}_{k=1} Q_{ni}= Q\) a.s.; (C) the random variable \(q(Y)\) has a bounded density; (D) \(|X_{ni} |\leq Cn^{-1/2}\) a.s. for all \(i=1, \dots, k_n\). \dots Theorem 1. Let \(q\) be a seminorm on \(H\) such that \(q\in {\mathcal C}_{2,1} (H)\). Assume that conditions (B) and (C) are satisfied. Then for every \(s>2\) there exists a constant \(C=C(s,Q)\) such that \(\Delta_{q,n} \leq CL^{\alpha (s)}_{s,n}\). If, in addition, \((\text{D}')\) \(|X_{ni} |\leq M_n\) a.s. for all \(i=1, \dots, k_n\), then \(\Delta_{q,n} \leq CM_n^{1/3}\). Note that this bound is sharp even for independent \(X_{n1}, \dots, X_{nk_n}\).