1
$\begingroup$

Let $X_1,...,X_n$ be iid observations from $N(0,1)$. Let $\overline{X}=\dfrac{1}{n}\sum_{i=1}^n X_i$ and $S^2=\dfrac{1}{n}\sum_{i=1}^n (X_i-\overline{X})^2$. Then is it true that $\sqrt{n}\sup_x |\Phi(\dfrac{x-\overline{X}}{S})-\Phi(x)|\stackrel{p}{\to} 0$? Here $\Phi(.)$ is standard normal cdf.

Seems something like Berry Esseen bound but I think I am getting it is stochastically bounded and not really going to 0.

$\endgroup$

1 Answer 1

3
$\begingroup$

No, it is not true. First fix any $x$ and consider $\overline{X^2}=\frac1n\sum_{i=1}^nX_i^2$, $S^2=\overline{X^2}-\bigl(\overline X\bigr)^2$ and $h(s,t)=\Phi\left(\dfrac{x-s}{\sqrt{t-s^2}}\right)$. Then use multivariate Delta method to prove that $$ \sqrt{n} \left(\Phi\biggl(\dfrac{x-\overline{X}}{S}\biggr)-\Phi(x)\right)=\sqrt{n}\biggl(h\biggl(\overline X, \overline{X^2}\biggr)-\underbrace{h(\mathbb EX_1,\mathbb EX_1^2)}_{h(0,1)}\biggr)\xrightarrow{\mathcal D}\mathcal N(0,\sigma_x^2), $$ where $\sigma^2_x = \nabla h(0,1)^T\cdot \Sigma \cdot \nabla h(0,1)$ and $\Sigma$ is a covariance matrix of $(X_1,X_1^2)$.

Here $\text{Var}(X_1)=1$, $\text{Var}(X_1^2)=2$, $\text{Cov}(X_1, X_1^2)=0$, so $$ \Sigma=\begin{pmatrix}1 & 0 \cr 0 & 2 \end{pmatrix} $$ $$ \frac{\partial h}{\partial s}\bigg|_{s=0, t=1} = -\varphi(x) = -\frac{1}{\sqrt{2\pi}}e^{-x^2/2} $$ and $$ \frac{\partial h}{\partial t}\bigg|_{s=0, t=1} = -\frac12\varphi(x) = -\frac{1}{2\sqrt{2\pi}}e^{-x^2/2} $$ Please check the derivatives, although their values do not matter much. And $$ \sigma^2_x = \varphi^2(x)\cdot 1+\left(\frac{\varphi(x)}{2}\right)^2\cdot 2 = \frac32\varphi^2(x). $$ Say, for $x=0$, $$ \sqrt{n} \left(\Phi\biggl(\dfrac{x-\overline{X}}{S}\biggr)-\Phi(x)\right) \xrightarrow{\mathcal D} \mathcal N\biggl(0,\frac{3}{4\pi}\biggr) $$ For any $\varepsilon>0$ $$ \liminf_{n\to\infty}\mathbb P\left(\sqrt{n}\sup_x \biggl|\Phi\biggl(\dfrac{x-\overline{X}}{S}\biggr)-\Phi(x)\biggr|>\varepsilon\right) \geq \liminf_{n\to\infty}\mathbb P\left(\sqrt{n} \biggl|\Phi\biggl(\dfrac{0-\overline{X}}{S}\biggr)-\Phi(0)\biggr|>\varepsilon\right) $$ $$ =\mathbb P(|Y|>\varepsilon)=2\Phi\left(-\frac{2\sqrt{\pi}\varepsilon}{\sqrt3}\right)>0 $$ where $Y\sim N\bigl(0,\frac{3}{4\pi}\bigr)$.

$\endgroup$
1
  • $\begingroup$ Yup I realised that, essentially by using delta method after posting this question. Thanks anyway! $\endgroup$ Feb 1, 2020 at 5:08

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.