4
$\begingroup$

The celebrated Berry-Esseen inequality tells us that the rate of convergence in the univariate CLT is of magnitude $\frac{1}{\sqrt{n}}$ for sums $S_n=X_1+\cdots+X_n$ of independent random variables $X_i$ with bounded third moments. Consider the following example: let $\varepsilon_{i}$ stand for independent Rademacher random variables, that is, we have $\mathbb{P}(\varepsilon_{i}=\pm 1)=1/2$. I am interested in the random vector $V_n=(R_n,R'_n)$, where $R_n=\varepsilon_{1}+\cdots+\varepsilon_{n}$ and $R'_{n}$ is an independent copy of $R_n$. It is easy to verify that for fixed $r\geq \sqrt{2}$ we have $\mathbb{P}(||V_n||\leq r)\approx \frac{c_{r}}{n}$ for some $c_{r}>0$. I would like to estimate the proximity between the law of $V_n$ and a correspoding 2-dimensional Gaussian distribution. For an estimate of the proximity that would make sense one would hope for a bound of magnitude at most $\frac{c}{n}$. Are such results known?

$\endgroup$
4
$\begingroup$

It follows from the main result in the paper

MR1309710 (95k:60015) Nagaev, S. V.; Chebotarëv, V. I. On the Edgeworth expansion in a Hilbert space. (Russian) Limit theorems for random processes and their applications, 170--203, 304, Trudy Inst. Mat., 20, Izdat. Ross. Akad. Nauk Sib. Otd. Inst. Mat., Novosibirsk, 1993

that you have the Edgeworth expansion of the form $$P(\|V_n\|^2\le r)=P(\|G\|^2\le r)+\sum_{1\le m\le(k-2)/2}n^{-m} Q_{2m}(r)+\Delta_{n,k}(r),$$ where $G$ is a mean-zero Gaussian vector with the same covariance matrix as that of $V_n$, $k\ge4$, $Q_{2m}$ are certain functions, and $\Delta_{n,k}(r)$ is a remainder. So, the main asymptotic term in the "error" $P(\|V_n\|^2\le r)-P(\|G\|^2\le r)$ of the normal approximation is $Q_2(r)/n$, indeed on the order of $1/n$.

$\endgroup$
  • $\begingroup$ Thank you so much! Exactly the thing I was looking for! $\endgroup$ – TOM Mar 8 at 13:34

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.