3
$\begingroup$

Let us say have a sequence of $n$ 2-$D$ random variables $X_i=(\varepsilon_i/\sqrt{n},i\varepsilon_{i}\sqrt{6}/n^{3/2})$, where $\varepsilon_{i}$ are independent random variables such that $\mathbb{P}(\varepsilon_i =\pm 1)=1/2$. Denote by $S_n$ their sum and take a $2$-dimensional Gaussian random variable $Z$ with the same covariance matrix as $S_n$. Any standard Berry-Esseen theorem (say by Gotze or Bhachattarya or Bentkus) gives us that for any convex set $C$ we have \begin{equation} |\mathbb{P}(S_n\in C)-\mathbb{P}(Z\in C)|\leq c\gamma, \end{equation} where gamma is the sum of third absolute moments of $X_i$, which in this case behaves like $n^{-1/2}$.

Question: is there a standard way to pass from the distance between distribution functions to expectations of Lipschitz functions? That is, suppose $f$ is Lipschitz, can we still bound

\begin{equation} |\mathbb{E}f(S_n)-\mathbb{E}f(Z)| \end{equation} in therms of $\gamma$? If so, does the same bound of magnitude $\gamma$ still apply?

$\endgroup$

1 Answer 1

4
$\begingroup$

I'm fairly confident the answer is yes, but unfortunately I can't specify an exact reference just now.

If you just want to see that one can get a similar bound (but with worse dependence on $\gamma$), then it's not too hard to prove multidimensional Berry--Esseen theorems via the "Lindeberg replacement method"; see, e.g., Exercise 11.46(a) (cf. Corollary 11.59) in http://get.analysisofbooleanfunctions.org

Probably to get the correct dependence on $\gamma$ in the Lipschitz case, the best strategy would be a proof via Stein's method. Unfortunately, I didn't see the exact result you want in, say, the Chen-Goldstein-Shao text, and unfortunately this Valiant-Valiant paper's Theorem 2 (http://eccc.hpi-web.de/report/2010/179/) is not quantitatively optimal (probably).

But I feel pretty confident the result you want should be known. Perhaps try searching for 'multidimensional Stein's method'.

(If you find an appropriate theorem, please post here!)

$\endgroup$
0

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.