Rates of convergence for multivariate normal approximation with applications to dense graphs and doubly indexed permutation statistics (Q888478): Difference between revisions

Based on quotations taken from the article: The authors provide a new general theorem for multivariate normal approximation on convex sets. The theorem is formulated in terms of a multivariate extension of Stein couplings. Let \(W\) and \(Z\) be \(d\)-dimensional random vectors, \(d\geq 1\), where \(Z\) has standard \(d\)-dimensional Gaussian distribution. This paper is concerned with bounding the quantity \[ d_c({\mathcal L}(W), {\mathcal L}(Z)):=\sup_{A\in {\mathcal A}}|\operatorname{P}(W\in W)-\operatorname{P}(Z\in A)|, \] where \({\mathcal A}\) denotes the collection of all convex sets in \(\mathbb R^d\). A triple of square integrable \(d\)-dimensional random vectors \((W,W',G)\) is called a \(d\)-dimensional Stein coupling if \[ \operatorname{E}\{G'F(W')-G'F(W)\}=\operatorname{E}\{W'F(W)\} \] for all \(F:\mathbb R^d\to \mathbb R^d\) for which the expectations exist. Throughout the article, \(|x|\) denotes the Euclidean norm of \(x\in \mathbb R^d\), and \({\mathbf I}_d\) denotes the \(d\)-dimensional identity matrix. To shorten the formulas, write \(\operatorname{E}^W(\cdot)\) to denote conditional expectation \(\operatorname{E}(\cdot|~W)\). With this, the main result of the article reads as follows: let \((W,W',G)\) be a \(d\)-dimensional Stein coupling. Assume that \(\mathrm{Cov}(W)={\mathbf I}_d\). With \(D=W'-W\), suppose that there are positive constants \(\alpha\) and \(\beta\) such that \[ |G|\leq \alpha,\quad |D|\leq \beta. \] Then, there is a universal constant \(C\) such that \[ d_c({\mathcal L}(W), {\mathcal L}(Z))\leq \] \[ C\Bigl[d^{7/4}\alpha \operatorname{E}|D|^2 + d^{1/4}\beta+ d^{7/4}\alpha^{1/2} B^{1/2}_1+ d^{3/8}B_2+ d^{1/8} B^{1/2}_3\Bigr], \] where \(Z\) is a \(d\)-dimensional standard Gaussian random vector and \[ B_1:=\sqrt{\mathrm{Var}\operatorname{E}^W|D|^2},\quad B_2:=\sqrt{\sum_{i,j=1}^d \mathrm{Var}\operatorname{E}^W(G_iD_j)}, \] \[ B_3:=\sqrt{\sum_{i,j,k=1}^d \mathrm{Var}\operatorname{E}^W(G_iD_jD_k)}. \] The authors apply the results to a homogeneity test in dense random graphs and to prove multivariate asymptotic normality for certain doubly indexed permutation statistics.