Abstract:We prove the claim in the title under mild conditions which are usually satisfied when trying to establish asymptotic normality. We assume strictly stationary and absolutely regular data.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is about the equivalence between the asymptotic normality of U - statistics and the convergence of Wasserstein distance. Specifically, the author proves that under certain conditions, the asymptotic normality of U - statistics (that is, $\sqrt{n}U_h$ weakly converges to $N(0, \sigma^2)$) is equivalent to the convergence under Wasserstein distance $d_2$.
### Background of the paper
U - statistics are a widely used class of statistics, including sample mean, variance, k - order moment estimators, Gini mean difference, point estimates of cumulative distribution functions, and distance covariance, etc. The general form of U - statistics is:
\[ U_h := U_h(X_1, \ldots, X_n) := \binom{n}{m}^{-1} \sum_{1 \leq i_1 < \cdots < i_m \leq n} h(X_{i_1}, \ldots, X_{i_m}), \]
where $X_1, \ldots, X_n$ are observations from a separable metric space $X$, and $h: X^m \to \mathbb{R}$ is a centrally symmetric kernel function.
### Main results
The main result of the paper is the following theorem:
**Theorem 1**: Let $(X_k)_{k \in \mathbb{N}}$ be a process taking values in a separable metric space $X$, and $h: X^m \to \mathbb{R}$ be a centrally symmetric kernel function. Assume that one of the following two conditions holds:
1. $(X_k)_{k \in \mathbb{N}}$ is independently and identically distributed, and $\|h(X_1, \ldots, X_m)\|_{L^2} < \infty$.
2. $(X_k)_{k \in \mathbb{N}}$ is strictly stationary and absolutely regular, and $\|h(X_1, \ldots, X_m)\|_{L^p} < \infty$ for some $p > 2$. Moreover, the mixing coefficient $\beta(n)=O(n^{-r})$ for some $r > \frac{mp}{p - 2}$.
Then $\sqrt{n}U_h$ converges in distribution to $N(0, \sigma^2)$ if and only if it converges to the same limit under Wasserstein distance $d_2$.
### Explanation
1. **Asymptotic normality**: This means that $\sqrt{n}U_h$ weakly converges in distribution to the normal distribution $N(0, \sigma^2)$, where $\sigma^2>0$.
2. **Wasserstein distance**: The Wasserstein distance $d_2$ is a method for measuring the difference between two probability distributions, defined as:
\[ d_2^2(X, Y) := \inf E[(X' - Y')^2], \]
where $\inf$ is taken over all random vectors $(X', Y')$ such that the distributions of $X'$ and $Y'$ are equal to the distributions of $X$ and $Y$ respectively. The convergence of Wasserstein distance is stronger than weak convergence because it also implies the convergence of certain moments.
### Proof ideas
1. **Hoeffding decomposition**: U - statistics can be expressed as the first term $\sqrt{n}U_n^{(1)}(h; \xi)$ of its Hoeffding decomposition plus the remaining terms. If it can be proved that $\sqrt{n}U_n^{(1)}(h; \xi)$ is asymptotically normal and the remaining terms converge under $o_P(1/\sqrt{n})$, then the asymptotic normality of $\sqrt{n}U_h$ can be proved.
2. **Treatment of mixed data**: For mixed data, it is necessary to assume $p$-moment ($p > 2$) instead.