I'm interested in estimates of the normal approximation for binomial distributions, i.e. in estimates of

$$\sup_{x\in\mathbb R}\left|P\left(\frac{B(p,n)-np}{\sqrt{npq}} \le x\right) - \Phi(x)\right|$$

From the Berry-Essen theorem I can deduce

$$\sup_{x\in\mathbb R}\left|P\left(\frac{B(p,n)-np}{\sqrt{npq}} \le x\right) - \Phi(x)\right| \le \frac{C(p^2+q^2)}{\sqrt{npq}}$$

with $C \le 0.4748$.

My question: Are there better estimates for the normal approximation of the binomial distribution?

The Berry-Esseen theorem is quite general because it can be applied to each sum of i.i.d random variables. So I guess there are better estimates for the special case of the binomial distribution...

  • $\begingroup$ Have you found a tighter bound since the OP? Could you share your findings, please? Thank you $\endgroup$
    – Confounded
    Aug 12, 2021 at 12:51
  • $\begingroup$ This doesn't seem to be a very useful formula for the bound since it can return values much greater than $1$ which is nonsensical as the difference between any two CDFs can't be larger than 1. Is the formula correct? $\endgroup$
    – Confounded
    Aug 12, 2021 at 14:01
  • $\begingroup$ @Confounded It's a long time since I worked in this field. However I do not recall to have a better method than the Berry-Essen theorem (for the constant I might have found a better estimation). mathoverflow.net/questions/220030/… is related to this as well $\endgroup$ Aug 22, 2021 at 21:29

1 Answer 1


I'm not sure if you can actually expect too much of an improvement here. In some sense the binomial distribution is actually a (near) worst case for Berry-Esseen.

By Chebyshev's inequality, a positive proportion of the mass of the binomial distribution is located between $np-\sqrt{npq}$ and $np+\sqrt{npq}$. Together with the pigeonhole principle, this implies that a single value in this interval is taken on with probability proportional to $\frac{1}{\sqrt{npq}}$ (there are much more precise results than this; see for example Steven Dunbar's notes on Local Limit Theorems), meaning that the CDF of the binomial distribution has jumps of this size.

On the other hand, the normal distribution is continuous. So on one side or the other of the jump, the approximation error must be at least $\frac{C'}{\sqrt{npq}}$, which matches the Berry-Esseen bound up to a constant factor ($p^2+q^2$ is always between $1/2$ and $1$).


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.