2
$\begingroup$

I have a doubt about the central limit theorem.

Let $(X_n)$ be an IID sequence, each $X_n$ distributed as $X$ where

$\mathbb{E}[X]=0$, $\sigma^2 = Var(X) < \infty$.

Define $S_n = X_1 + \dots + X_n$, and set

$G_n = \frac{S_n}{\sigma\sqrt{n}}$

Now, the theorem tells us that

$\lim_{n \to \infty}\mathbb{P}(G_n \le x) = \Phi(x)$ where $\Phi(x)$ is the distribution function of the standard normal distribution.

Now is it FORMALLY correct to say, for example, that

$\lim_{n \to \infty}\mathbb{P}(G_n \le 1+\frac{1}{n}) = \Phi(\lim_{n \to \infty}(1+\frac{1}{n})) = \Phi(1) \simeq 0.8413$?

Thanks to who will solve my (perhaps trivial) doubt.

$\endgroup$
  • $\begingroup$ No: $x$ is a number, not a function of $n$. $\endgroup$ – Lord Shark the Unknown Sep 1 '17 at 20:15
  • $\begingroup$ In this case, how can I estimate that probability? $\endgroup$ – LJG Sep 1 '17 at 20:17
  • $\begingroup$ The CLT is merely telling you how to compute $\lim_{n \to \infty} P(G_n \leq x)$ for fixed $x$ not depending on $n$. You might hope by some interchange of limits result or similar that you can replace $x$ by a sequence $x_n$ converging to $x$ but that's not the same result. $\endgroup$ – Ian Sep 1 '17 at 20:18
  • $\begingroup$ @LordSharktheUnknown Do you mean "not a function of $n$" instead - $x$ certainly is a function of $x$ ;) $\endgroup$ – B. Mehta Sep 1 '17 at 20:19
  • $\begingroup$ LJG: You accepted instantly the answer below. As a consequence your question basically disappeared from the list of those to answer, which made that you are left with the unique answer below. Too bad... $\endgroup$ – Did Sep 3 '17 at 11:10
1
$\begingroup$

No, it is not formally correct because $1+1/n$ is a sequence of numbers which depends on $n$, whereas the CLT as stated only holds for fixed real numbers $x$.

However, it is still true that $\lim_{n \to \infty} P(G_n \leq 1+1/n) = \Phi(1)$. In order to prove this, one only needs to use the fact that the functions $F_n(x):=P(G_n \leq x)$ are increasing, together with continuity of the limit $\Phi$. Indeed, for $N \leq n$, one has that $F_n(1+1/n) \leq F_n(1+1/N)$, and thus for any fixed $N$ $$\limsup_{n\to \infty}F_n(1+1/n) \leq \limsup_{n \to \infty} F_n(1+1/N) = \Phi(1+1/N)$$ Next, we may let $N \to \infty$ on both sides, and since the LHS does not depend on $N$, one sees by continuity of $\Phi$ that $$\limsup_{n \to \infty}F_n(1+1/n) \leq \lim_{N\to \infty} \Phi(1+1/N) = \Phi(1)$$ On the other hand, one clearly has that $F_n(1+1/n) \geq F_n(1)$ fo every $n$, and therefore we see that $$\liminf_{N \to \infty} F_n(1+1/n) \geq \liminf_{n \to \infty} F_n(1) = \Phi(1)$$ which (together with the preceding expression) gives the desired result.

In fact, because of this question, it is actually true that one has uniform convergence of the cdf's $F_n$ to $\Phi$, which means that $F_n(x_n) \to \Phi(x)$ whenever $x_n \to x$ (and the proof uses the same ideas).

$\endgroup$
  • $\begingroup$ Thank you for the reply. Could you give me some advice to estimate in a formal way that probability? $\endgroup$ – LJG Sep 1 '17 at 20:33
  • $\begingroup$ We did not study Berry–Esseen theorem. Anyway, for the theorem you talked about before, can I write safely that $\lim_{n \to \infty}\mathbb{P}(G_n \le 1+\frac{1}{n}) = \Phi(1)$? $\endgroup$ – LJG Sep 1 '17 at 20:50
  • $\begingroup$ @LJG Yes, see the linked question $\endgroup$ – Shalop Sep 1 '17 at 20:51
  • $\begingroup$ Excellent. Thank you very much. $\endgroup$ – LJG Sep 1 '17 at 20:53
  • 1
    $\begingroup$ @Shalop I've seen the changes and thank you for the further clarification. $\endgroup$ – LJG Sep 3 '17 at 7:17

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.