Cluster sets for partial sums and partial sum processes

Abstract

Let $X,X_1,X_2,\dots$ be i.i.d. mean zero random vectors with values in a separable Banach space $B$, $S_n=X_1+\cdots +X_n$ for $n\ge 1$, and assume $\{c_n:n\ge 1\}$ is a suitably regular sequence of constants. Furthermore, let $S_{(n)}(t),0\le t\le 1$ be the corresponding linearly interpolated partial sum processes. We study the cluster sets $A=C(\{S_n/c_n\})$ and $\mathcal{A}=C(\{S_{(n)}(\cdot )/c_n\})$. In particular, $A$ and $\mathcal{A}$ are shown to be nonrandom, and we derive criteria when elements in $B$ and continuous functions $f:[0,1]\to B$ belong to $A$ and $\mathcal{A}$, respectively. When $B={\mathbb{R}}^d$ we refine our clustering criteria to show both $A$ and $\mathcal{A}$ are compact, symmetric, and star-like, and also obtain both upper and lower bound sets for $\mathcal{A}$. When the coordinates of $X$ in ${\mathbb{R}}^d$ are independent random variables, we are able to represent $\mathcal{A}$ in terms of $A$ and the classical Strassen set $\mathcal{K}$, and, except for degenerate cases, show $\mathcal{A}$ is strictly larger than the lower bound set whenever $d\ge 2$. In addition, we show that for any compact, symmetric, star-like subset $A$ of ${\mathbb{R}}^d$, there exists an $X$ such that the corresponding functional cluster set $\mathcal{A}$ is always the lower bound subset. If $d=2$, then additional refinements identify $\mathcal{A}$ as a subset of $\{(x_1{g}_1,x_2{g}_2):(x_1,x_2)\in A,g_1,g_2\in \mathcal{K}\}$, which is the functional cluster set obtained when the coordinates are assumed to be independent.

Published in at http://dx.doi.org/10.1214/12-AOP827 the Annals of Probability (http://www.imstat.org/aop/) by the Institute of Mathematical Statistics (http://www.imstat.org)

References in corpus (1)

• Some results on two-sided LIL behavior