Question about the strong law of large numbers (to build understanding)

202 Views Asked by At

I'm learning about the LLN and CLT for the first time and I'm having some trouble. I've read through other posts but I have a quirky (likely dumb) question about denominators...

The Strong LLN says for $X_1,...,X_n$ iid with common mean $\mu$,

$$P \left( \lim_{n \rightarrow \infty} \frac{X_1 + ... + X_n }{n} = \mu \right) = 1$$

My question 1: Instead of dividing by $n$, is it true that if I divide by a function of $n$, say $f(n) = Cn^k$ where $k \ge 1$ then the limit will converge to some constant always? (not necessarily $\mu$)

$$P \left( \lim_{n \rightarrow \infty} \frac{X_1 + ... + X_n }{f(n)} = \text{some constant} \right) = 1$$

I believe this is true because $Var\left(\frac{S_n}{f(n)}\right) = \frac{n \sigma^2}{(f(n))^2}$ and as long as $f(n)$ grows as fast as $n$ then the variance will go to zero as $n$ goes to infinity.

My question 2: If instead of dividing by $n$ in the Strong LLN, if I divide by $ \sigma\sqrt{n}$ then is it true that

$$P \left( \lim_{n \rightarrow \infty} \frac{X_1 + ... + X_n }{\sigma \sqrt{n}} = \text{some constant} \right) \neq 1$$

and it won't "converge" to a constant because the RV $\frac{S_n}{\sigma \sqrt{n}}$ is a Normal RV with variance $1$.

My question 3: It seems to me what you divide $S_n$ by is very important and can lead to very different results like the LLN or CLT... is that a somewhat correct way of thinking?

Thank you!!

1

There are 1 best solutions below

4
On BEST ANSWER

Regarding 1 and 2: If you divide by $Cn^k$ then the limit will be $\mu/C$ (if $k=1$) or $0$ (if $k>1$). When $k<1$ the result will be $\infty$ with the same sign as $\mu$ unless $\mu=0$, in which case the situation is more complicated.

Regarding 3, yes: essentially you have different results when you measure deviation of the sample mean from the population mean in different scales relative to $n$. Under "typical" circumstances (when $X_i$ has a MGF existing in a neighborhood of $0$):

  • $O(1)$ scale fluctuations asymptotically vanish (this is LLN)
  • $O(n^{-1/2})$ scale fluctuations persist with probability strictly between $0$ and $1$ (this is CLT)
  • $O(1)$ scale fluctuations decay exponentially fast in probability (this is Cramer's theorem, a type of large deviation principle).