4
$\begingroup$

Suppose that $X_i$ are independent identically distributed with finite variance and $S_n=X_1+\cdots+X_n$. One can use the Central Limit Theorem to estimate (a) $P(S_n \leq b)$ and (b) $P(a<S_n \leq b)$.

The Berry-Esseen theorem estimates the maximum possible error for the first case (a). The error is not greater than $C\frac{\rho}{\sigma^3\sqrt{n}}$.

Using the fact that $P(a<S_n \leq b)=P(S_n \leq b)-P(S_n \leq a)$ I may obtain a trivial bound for the error in the case (b): $2C\frac{\rho}{\sigma^3\sqrt{n}}$.

Is it the best possible bound for the error in the second case (b)? Or there is something better?

$\endgroup$
4
$\begingroup$

One cannot hope for anything smaller than $C\rho/(\sigma^3\sqrt{n})$, otherwise a strictly better bound than Berry-Esseen's bound would hold in case (a) (take the limit of (b) when $a\to-\infty$).

So the trivial bound is off by at most the prefactor $2C$ instead of $C$. In the cases I know, this has zero consequence.

$\endgroup$

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.