1
$\begingroup$

I could not derive the weak law of large numbers from the central limit theorem for i.i.d. random variables with $0 < \operatorname{Var}(X) < \infty$.

The central limit theorem gives $$\frac{\sum X_i - n\cdot E(X_1)}{\sqrt{\operatorname{Var}(X_i)n})} \to N(0,1)$$

Now I want to show

$$\lim P\left(\middle|\frac{\sum X_i - E(X_1)}{n}\middle| \geq \epsilon\right) = 0 \quad\forall \epsilon>0$$

I know that the cdf converges pointwise. It should be really easy as it is intuitively clear that it should hold, but I can't find the correct way to write and pin it down.

$\endgroup$
  • 1
    $\begingroup$ One can prove the wlln using Chebyshev's inequality, and one doesn't need any nontrivial information about the shape of the distribution. $\endgroup$ – Michael Greinecker Nov 16 '13 at 17:36
  • $\begingroup$ Yes. But it is asked for how to derive it from the much stronger central limit theorem.. $\endgroup$ – user109265 Nov 16 '13 at 17:44
  • $\begingroup$ I understand the question. What I don't understand is why the CLT should be of much use here. $\endgroup$ – Michael Greinecker Nov 16 '13 at 17:48
3
$\begingroup$

For every positive $t$, consider the events $A_n^t=[|S_n-nE[X]|\geqslant t\sqrt{n\mathrm{var}(X)}]$ and $B_n^t=[|S_n-nE[X]|\geqslant nt]$. Then:

  1. The CLT says that $P[A_n^t]\to P[|Z|\geqslant t]$ when $n\to\infty$, for every positive $t$, where $Z$ is standard normal.
  2. The WLLN says that $P[B_n^t]\to0$ when $n\to\infty$, for every positive $t$.
  3. For every positive $t$ and $s$, there exists some finite $N$ such that $B_n^t\subseteq A_n^s$ for every $n$ large enough (can you show this?).

With these elementary facts in mind, assume that 1. holds, that is, that the CLT is valid and consider some positive $t$ and $\varepsilon$. Then:

  • There exists some finite $s$ such that $P[|Z|\geqslant s]\leqslant\varepsilon$.
  • By 1., this implies that $P[A_n^s]\leqslant2\varepsilon$ for every $n$ large enough, say $n\geqslant N_1$.
  • By 3., $P[B_n^t]\leqslant P[A_n^s]$ for every $n$ large enough, say $n\geqslant N_2$.

Thus, for every $n\geqslant\max(N_1,N_2)$, $P[B_n^t]\leqslant2\varepsilon$. Since $\varepsilon$ is arbitrary, this is the WLLN.

$\endgroup$

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.