We know that if $X\sim \mathrm{Pois}(\lambda)$ and $Y\sim \mathrm{Pois}(\mu)$ are independent then $X+Y\sim \mathrm{Pois}(\lambda+\mu).$

This means that if $X_{1},\ldots,X_{n}$ are independent with $X_i \sim \mathrm{Pois}(1/n)$ (and so $E(X_i) = \operatorname{Var}(X_i)=1/n)$ then $S_n=\sum\limits_{i=1}^n X_i\sim \mathrm{Pois}(1)$ for all $n$.

If the CLT holds, it follows that $S_n\sim \mathrm{Pois}(1)$ is a normal distribution, which is not true.

What is wrong with this argument?


3 Answers 3


When you look at the definition of $S_3$, what do $X_1, X_2, X_3$ mean? They're three RV distributed $\mathrm{Pois}(1/3)$.

When you look at the definition of $S_4$, what does $X_2$ mean? IT means a RV $\mathrm{Pois}(1/4)$.

This when you look at the sequence $S_n$, the meaning of $X_i$ keeps changing. The CLT says nothing about things like that. In the CLT, you start with an infinite sequence of RVs $X_i$, and compute their partial sums, etc.

  • $\begingroup$ ...And divide their sum by n, which is not so here. $\endgroup$
    – Did
    Sep 25, 2015 at 12:41
  • $\begingroup$ Good point, Did...thanks! $\endgroup$ Sep 25, 2015 at 15:46

This is an interesting question. The central limit theorem says that the average will approach a normal distribution, meaning that for some measure of closeness to a normal distribution there is some $n$ for which the average of at least $n$ of these variables is at least that close. However, $n$ depends on $\lambda$, and this is the error that you're making by letting $\lambda$ depend on $n$.

To be explicit, CLT follows from approximating the exponent of the moment generating function with a Taylor series. The MGF of Poisson($\lambda$) is $\exp\left(\lambda(e^t-1)\right)$. The exponent can be approximated by $f(t)=\lambda\left(t+\frac{1}{2}t^2+\frac{1}{3}t^3+\cdots\right)$.

It follows from properties of MGFs that the MGF of the average of $n$ of these is $nf\left(\frac{1}{n}\right)=\lambda t+\frac{\lambda}{2n}t^2+\frac{\lambda}{3n^2}t^3+\cdots$, and the high orders of $t$ vanish much faster than the linear and squared terms.

However, if you plug in $\lambda=\frac{1}{n}$, you would get that the average approaches something not normal, and as a result you should expect the sum of the averages to be normal either.


if $X_1,\ldots,X_n$ are independent with $X_i \sim \mathrm{Pois}(1/n)$ (and so $E(X_i) = \operatorname{Var}(X_i)=1/n)$ then $S_n=\sum\limits_{i=1}^n X_i\sim \mathrm{Pois}(1)$ for all $n$.

If the CLT holds, it follows that $S_n\sim \mathrm{Pois}(1)$ is a normal distribution

No, you're misapplying the theorem.

Suppose $n=6$. Then you have $X_1,\ldots,X_6 \sim \mathrm{i.i.d. Poisson}\left(\dfrac 1 6 \right)$.

Then $X_1+\cdots+X_6 \sim \mathrm{Poisson}(1)$. And $X_1+\cdots+X_{600} \sim \mathrm{Poisson}(100) \ne \mathrm{Poisson}(1)$. And $\mathrm{Poisson}(100)$ is approximately normal with expectation $100$ and variance $100$ (so the standard deviation is $10$). But use a continuity correction here: if you want $\Pr(X_1+\cdots+X_{600}>95)$, realize that that is the same as $\Pr(X_1+\cdots+X_{600}\ge96)$ and so plug in $95.5$.

If you then change the distribution of $X_1,\ldots,X_{600}$ to $\mathrm{Poisson}\left(\dfrac 1 {600}\right)$, then their sum is indeed distributed as $\mathrm{Poisson}(1)$, which is not so close to normal. But if $X_1,\ldots,X_{6000000}\sim\mathrm{i.i.d. Poisson}\left(\dfrac 1 {600}\right)$ then again you get something that is approximately normal.

The central limit theorem says if $X_1,X_2,X_3,\ldots\sim\mathrm{i.i.d.}$ with a finite variance, then the distribution of $$ \frac{(X_1+\cdots+X_n)-\operatorname{E}(X_1+\cdots+X_n)}{\operatorname{SD}(X_1+\cdots+X_n)} $$ approaches the standard normal distribution as $n\to\infty$. It does not say that if you keep changing the distribution of each of the random variables as $n$ increases, then the same thing happens.


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.