Problem setup: Suppose $X_1, \ldots, X_n$ is an i.i.d. sample from $F_X$ (CDF), and $Y_1, \ldots, Y_n$ is another i.i.d. sample from $F_Y$ (also CDF). In addition, $h(z_1, \ldots, z_n)$ is a real-valued symmetric function. Let $H_X = h(X_1, \ldots, X_n)$ and $H_Y = h(Y_1, \ldots, Y_n)$. Denote $D_{KS}(X, Y)$ as the Kolmogorov-Smirnov distance between $X \sim F_X$ and $Y \sim F_Y$, that is,

$$D_{KS}(X, Y) = \sup_{z \in \mathbb R} |F_X(z) - F_Y(z)|.$$

Question: How should I use $D_{KS}(X, Y)$ to bound $D_{KS}(H_X, H_Y)$, as in finding a function $g$ such that $D_{KS}(H_X, H_Y) \leq g(D_{KS}(X, Y))$? One can make assumptions on $F_X$, $F_Y$, and $h$, but should be as minimal as possible. (for example, compact support of $F_X$, $h$ being Lipschitz, and etc.).

My intuition is that if $X$ and $Y$ are from two close distributions (in terms of KS distance), so should $H_X$ and $H_Y$, but I couldn't find a way to formally make a statement.

Any help or ideas would be much appreciated!


1 Answer 1


If you think of $h$ as a test statistic, then you should be able to get substantial differences between the distributions of $h({\mathbf X})$ and $h({\mathbf X})$ when the K-S distance between the distributions of $X$ and $Y$ is of order $n^{-1/2}$.

For example, take $X\sim N(0,1)$ and $Y\sim N(\mu,1)$ and $h({\mathbf X})= n^{-1/2}\sum_i X_i$. Then $h(X)\sim N(0,1)$ and $h(Y)\sim N(n^{1/2}\mu,1)$, which is not small if $\mu$ is of order $n^{-1/2}$ or larger.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.