I have been trying to prove the following:

Consider an i.i.d. sample of random variables $\left\{ X_n \right\}$. Their distribution $F$ is said to satisfy the rapidly varying tail condition, if $\forall a > 1$ the following holds: $$ \lim_{x \to \infty} \frac{1 - F(ax)}{1 - F(x)} = 0 $$ Show that if $\displaystyle \frac{\max_{i=1,\dots,n} X_i}{b_n} \overset{\mathbb{P}}{\to} 1$ (convergence in probability), for a sequence $b_n \to \infty$, then $F$ must satisfy the rapidly varying tail condition.

I consider the definition of convergence in probability, $$ \lim_{n \to \infty} \left( \left|\frac{\max_{i=1, \dots, n}(X_1, \dots, X_n}{n} - 1 \right| > \varepsilon \right) = 0 $$ and write (for the positive part):

\begin{align*} \mathbb{P}\left(\max_{i=1,\dots,n} X_i > (1 + \varepsilon) b_n\right) &= \mathbb{P}\left[ \left( \bigcap_1^n X_i \leq (1 + \varepsilon) b_n \right)^c \right] \\ &= \mathbb{P}\left( \bigcup_1^n X_i > (1 + \varepsilon) b_n \right) = n \mathbb{P}(X_1 > (1 + \varepsilon) b_n) \\ \end{align*} For the negative part, I know that \begin{align*} \mathbb{P}\left(\max_{i=1,\dots,n} X_i < (1 - \varepsilon) b_n\right) &= \mathbb{P}\left[ \bigcap_1^n X_i < (1 - \varepsilon) b_n \right] \\ &= 1 - n\mathbb{P}(X_i \geq (1 - \varepsilon)b_n) \end{align*} Taking limits in the positive and negative parts and equating them as they both have to be equal to 0 obtain $$ \lim_{n \to \infty} n \mathbb{P}(X_1 > (1 + \varepsilon) b_n) = 0, \quad \lim_{n \to \infty} n \mathbb{P}(X_1 \geq (1 - \varepsilon) b_n) = 1 $$ However, at this point I'm stuck. I can obviously deduce that $$ \lim_{n \to \infty} \frac{1 - F(az_n)}{1 - F(z_n)} = 0, \quad a := \frac{1}{1-\varepsilon} > 1, \; z_n = (1 - \varepsilon) b_n $$ however this does not exactly imply the rapidly varying tail condition, which must hold for arbitrary $x \to \infty$. Any ideas?


We need to prove that for any $x_k\to\infty$, $$ \frac{\overline F(ax_k)}{\overline F(x_k)} \to 0 \text{ as } k\to\infty. $$ Here $\overline F(x)=1-F(x)=\mathbb P(X_1 >x)$ is the right tail of distribution of $X_1$ and $a=\frac{1+\varepsilon}{1-\varepsilon}>1$.

We have that for $z_n=(1-\varepsilon)b_n\to\infty$ as $n\to\infty$ $$n\overline F(z_n)\to 1\ \text{ and }\ n\overline F(a z_n)\to 0. $$ We can assume without loss of generality that $b_n$ (and also $z_n$) is a monotone increasing sequence.

Then for any sufficiently large $k$, there exists $n_k$ s.t. $z_{n_k}\leq x_k < z_{n_k+1}$. Then we can use that $\overline F$ is decreasing and estimate $$ \frac{\overline F(ax_k)}{\overline F(x_k)} \leq \frac{\overline F(az_{n_k})}{\overline F(z_{n_k+1})}=\frac{\overline F(az_{n_k})}{\overline F(z_{n_k})}\cdot \frac{\overline F(z_{n_k})}{\overline F(z_{n_k+1})}$$ $$= \underbrace{\frac{\overline F(az_{n_k})}{\overline F(z_{n_k})}}_{\to 0}\cdot \underbrace{\frac{n_k\overline F(z_{n_k})}{(n_k+1)\overline F(z_{n_k+1})}}_{\to 1}\cdot \underbrace{\frac{n_k+1}{n_k}}_{\to 1}\to 0 $$

| cite | improve this answer | |

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.