On the rate of convergence in the martingale central limit theorem (Q1952437)

Let \(\underline{X}_n=(X_1,\ldots,X_n)\) be a square integrable martingale difference sequence of random variables, i.e. \(E(X_i|X_1,\ldots,X_{i-1})=0\) and \(E(X_i^2)<\infty\) for all \(i=1,\ldots,n\), so that \(s^2(\underline{X}_n)=\sum_{i=1}^nE(X_i^2)\) and \(V^2(\underline{X}_n)=s^{-2}(\underline{X}_n)\sum_{i=1}^n E(X_i^2|X_1,\ldots,X_{i-1})\) are well-defined. The martingale CLT says that if \(V^2(\underline{X}_n)\rightarrow1\) in probability as \(n\to\infty\) and a Lindeberg condition is satisfied, then \(s^{-1}(\underline{X}_n)\sum_{i=1}^nX_i\) converges in distribution to a standard normal random variable. Typically, bounds on the rate of convergence in this CLT involve the \(L_p\)-norm \(\|V^2(\underline{X}_n)-1\|_p\) for some \(p\in[1,\infty]\) in the form \(\|V^2(\underline{X}_n)-1\|_p^{p/(2p+1)}\). In the present paper it is shown that the exponent \(p/(2p+1)\) is optimal. This closes a notable gap in the theory of bounds on the rate of convergence in the martingale CLT.