Let us consider a distribution $\mathcal{L}$ on a finite set of integers (actually I'm even happy with any RV on the set $\{-1,0,1\}$) (with probability not to extremes : roughly around 1/4 for -1, 1 and 1/2 for 0).

I'm looking at the sum of such variables, so let set $Z_n=\sum_{i=1}^n X_i$ for a family of i.i.d variables $X$ drawn under $\mathcal{L}$ and I want to estimate the probability for $Z_n$ to be lower than 0. I would like a lower bound of the shape $ F_n(x) \geq \Phi(x)$ or some exponential lower bound (the upper bound is easy to get from sub-gaussian variables or Hoeffding bounds).

Hence, I tried to approximate the cumulative distribution $F_n(x)$ of the (conveniently normalised, recentered) $Z_n$ by the normal CDF $\Phi(x)$ and evaluate this on $x=\frac{\sqrt{n}E(X)}{\sigma}$, where $\sigma$ is the standard deviation of the $X_i$.

My problem is that I don't want an actual Berry-Esseen like inequality, but a one-sided one (plus the uniform decay in $K/\sqrt{n}$ is too weak for my purpose since I ultimately want to look at $F_n(x)^\alpha$ for very large $\alpha$). Going further than Berry-Esseen I looked at Edgeworth expansions, but since I'm on the tails of the distributions, I can't really figure out what are the signs of the remainders in the expression...


Your Answer

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

Browse other questions tagged or ask your own question.