Best Khintchine type inequalities for sums of independent, rotationally invariant random vectors (Q5936412)

From MaRDI portal
scientific article; zbMATH DE number 1613322
Language Label Description Also known as
English
Best Khintchine type inequalities for sums of independent, rotationally invariant random vectors
scientific article; zbMATH DE number 1613322

    Statements

    Best Khintchine type inequalities for sums of independent, rotationally invariant random vectors (English)
    0 references
    0 references
    0 references
    0 references
    15 November 2002
    0 references
    Let \((\Omega,P)\) be a probability space. Let \(X:(\Omega, P)\rightarrow {\mathbb{R}}^n\) be a rotation invariant random vector, i.e. \[ P(X\in A)= P(X\in R^{-1} (A)) \] for all Borel sets \(A\subseteq \mathbb{R}^n\) and all orthogonal transformations \(R\in O(n)\). If \(X\) takes its values in the Euclidean sphere \(S^{n-1}\subseteq \mathbb{R}^n\), \(X\) is uniformly distributed on \(S^{n-1}\). If \(\xi, \eta :(\Omega, P)\rightarrow \mathbb{R}\) are real random variables in \(L_1(\Omega, P)\), we say that \(\xi\) dominates \(\eta\), denoted by \(\eta \prec\xi\), if \[ \mathbb{E}\eta=\mathbb{E}\xi \quad \text{and} \quad\int_s^\infty P(\eta >t) dt \leq\int_s^\infty P(\xi >t) dt \] for all \(s\in \mathbb{R}^n\). The basic result in the paper is: Proposition 1. Let \((X_i)\), \((Y_i)\) be two sequences of independent, rotationally invariant random vectors in \(\mathbb{R}^n\) such that \(\|X_i\|^2\prec\|Y_i\|^2\) for \(i=1,\dots,k\). Then \[ \mathbb{E}\Biggl\|\sum_{i=1}^{k}X_i \Biggr\|^p\leq\text{ (resp. \(\geq\)) }\mathbb{E}\Biggl\|\sum_{i=1}^{k}Y_i\Biggr\|^p \] for \(p\geq 3\) or \(p, n\geq 2\) (resp. for \(1\leq p\leq 2\) and \(n\geq 3\)). If one of the sequences is normal, the \(p\)th moment is easily calculated. In such case the authors obtain: Theorem 2. Let \((X_i)\) be an i.i.d sequence of rotationally invariant random vectors in \(\mathbb{R}^n\). Let \(Z\) be a rotationally invariant normal random vector in \(\mathbb{R}^n\) such that \(\|X_1\|^2\prec\|Z\|^2\). Then \[ \Biggl(\mathbb{E}\Biggl\|\sum_{i=1}^{k}\alpha_i X_i\Biggr\|^p\Biggr)^{1/p}\leq \text{ (resp. \(\geq\)) } (\mathbb{E}\|Z\|^p)^{1/p} \Biggl(\sum_{i=1}^{k}|\alpha_i|^2\Biggr)^{1/2} \] for \(p\geq 3\) or \(p, n\geq 2\) (resp. for \(1\leq p\leq 2\) and \(n\geq 3\)). The constant \(\left(\mathbb{E}\|Z\|^p\right)^{1/p}\) is the best possible. In the case when \(X_i\)'s are uniformly distributed on the unit sphere in \(\mathbb{R}^{n}\), the best constant is \[ (\mathbb{E}\|Z\|^p)^{1/p}=\sqrt{\frac 2n} \Biggl(\Gamma\biggl(\frac{p+n}{2}\biggr)\left /\Gamma\biggl(\frac n2\biggl)\Biggr)^{1/p}.\right. \] With this constant, the Khinchin type inequality in this case also holds for \( 1\leq p\leq 2\), \(n=2\).
    0 references
    0 references
    Khinchin type inequalities
    0 references
    random vector
    0 references
    0 references