Rate of convergence in bootstrap approximations (Q1109447)

X\({}_ 1,X_ 2,..\). are iid random variables with zero mean and variance 1. Let C denote the collection \((X_ 1,...,X_ n)\), and let \((X_ 1^*,...,X_ n^*)\) be a collection drawn at random from C, by sampling with replacement. Define \[ \bar X=n^{- 1}\sum^{n}_{j=1}X_ j,\quad \bar X^*=n^{- 1}\sum^{n}_{j=1}X_ j^*,\quad {\hat \sigma}^ 2=n^{- 1}\sum^{n}_{j=1}(X_ j-\bar X)^ 2,\quad S=n^{1/2} \bar X, \] \[ S^*=n^{1/2}(\bar X^*-\bar X)/{\hat \sigma},\quad F_ n(x)=P(S\leq x),\quad and\quad F_ n^*(x)=P(S^*\leq x| C). \] Let G denote the standard normal distribution function. For n moderate to large, \(F_ n\) can be approximated by G. The bootstrap approximation states that \(F_ n\) should be close to \(F_ n^*\). Roughly speaking, it is shown that \(F_ n\) is closer to \(F_ n^*\) than to G when \(X_ 1\) has a finite third moment. More precisely, suppose \(P(| X_ 1|)>x)=x^{-a}K(x)\), where \(2\leq a\leq 3\) and K is slowly varying at infinity. Also assume that \(P(X_ 1>x)/P(| X_ 1| >x)\) converges as x approaches infinity. When \(2\leq a<3\), the normal approximation and the bootstrap approximation are asymptotically equivalent. When \(a=3\), there are cases where the bootstrap is asymptotically superior, and cases where the opposite is true. Thus the case \(a=3\), corresponding to the third moment being either ``just finite'' or ``just infinite'', forms the boundary between circumstances where the bootstrap performs better than the normal approximation and circumstances where it does not.