12
$\begingroup$

The proof of the Berry-Esseen theorem found in Terence Tao's notes (https://terrytao.wordpress.com/2010/01/05/254a-notes-2-the-central-limit-theorem/, Theorem 37) starts by "smoothing" the cumulative distribution cutoff function, by convolving with a function whose fourier transform is compactly supported.

First of all, what is the motivation for doing this?

Secondly, why not smooth the function by convolving with a compactly supported test function instead?

Any insights into the proof of Berry-Esseen would be appreciated.

$\endgroup$
6
+50
$\begingroup$

First of all, it is worth to mention that we really want to work with characteristic functions (Fourier transforms), since adding independent random variables corresponds to a plain multiplication (as opposed to convolutions).

Smoothing the cdf $F$ by convolution with a nice function $H$ (which is equivalent to adding an independent random variable with a smooth cdf) gives the following:

  1. The resulting cdf $G = F*H$ has bounded density. This allows to write for any smooth cdf $\widetilde G$ with the same mean $$ G(x) - \widetilde G(x) = \frac{i}{2\pi} \int_{\mathbb{R}}\frac{\psi(t) - \widetilde\psi(t)}{t} e^{-itx} dt,\tag{1} $$ where $\psi$ and $\widetilde\psi$ are the characteristic functions of $G$ and $\widetilde G$.

  2. Taking $\widetilde G = \Phi * H$, where $\Phi$ is the standard Gaussian cdf, we get from $(1)$ $$ \big(F - \Phi\big)*H(x) = \frac{i}{2\pi} \int_{\mathbb{R}}\frac{\varphi(t) - e^{-t^2/2}}{t}\omega(t) e^{-itx} dt,\tag{2} $$ where $\varphi$ and $\omega$ are characteristic functions of $F$ and $H$. Since $\omega$ has finite support, the integral in $(2)$ is over a finite interval. This allows to write Taylor expansion for the characteristic function $\varphi$ with error terms admitting nice bounds in this interval.

  3. There are now two errors: smoothing error (replacing $F$ by $G$) and normal approximation error, controlled by $(2)$. Adjusting the "window" $T$ (the width of support of $\omega$) increases one error and decrease another, so we can optimize the error.

If we took compactly supported $H$, not $\omega$, this could still increase smoothness of $F$. However, the analysis of characteristic functions (and, as I have explained, we want to analyze them rather than cdfs) would not be simplified, as explained in 2-3 above.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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