I have data from a random variable $X$:


I am interested in the asymptotic distribution of the sample average $\bar{x}=\frac{\sum_{i=1}^N x_{i}}{N}$.

These draws are independent but not identically distributed. They have mean $\mu_{i}$ and variance $\sigma^2_{i}<\infty$. (sorry if notation is not strict enough)

I know the Lindeberg–Lévy Central Limit Theorem does not hold because moments are not identical. But assume the Lyapunov’s condition holds. Then,

$$ \frac{1}{s_{n}}\sum_{i=1}^N(x_{i}-\mu_{i}) \xrightarrow{d} N(0,1) $$

where $s_{n}^2 = \sum_{i=1}^N \sigma_{i}^2 $.

Is there anything I can be certain about the sample average? Some sufficient condition? Can I not infer some degree of asymptotic normality?

My data support normality of the average. I wonder if there is some theoretical basis behind that, or just mere coincidence.

  • $\begingroup$ If you have Greene, theorem D.19 gives a variant of CLT that only requires independence. $\endgroup$ – VCG Sep 12 '16 at 20:25
  • 1
    $\begingroup$ Lindeberg CLT does not require the same first or second moments. It is a sufficient condition too. In essence, you only need a sequence of RVs whose don't get too big too fast (in the precise sense given in the theorem). $\endgroup$ – user75138 Sep 12 '16 at 21:11

Note that if you have the sample average $\bar{x}=\frac{1}{n}\sum_{i=1}^nx_{i}$, then using the same "sample average" operator on the individual moments to define $\hat{\mu}=\bar{\mu}$ and $\hat{\sigma}^2=\overline{\sigma^2}$, your Lyapunov equation becomes $$\frac{\bar{x}-\hat{\mu}}{\hat{\sigma}/\sqrt{n}} \xrightarrow{d} N(0,1)$$ This is essentially the same as the i.i.d. result for the asymptotic sampling distribution of the sample mean.


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.