In class we gave two definitions of heavy-tailed distributions:

($\overline{F}(x) = P(X>x)$ denotes the tail of the distribution)

  1. $\liminf\limits_{x\to\infty}\frac{\overline{F}(x)}{e^{-\lambda x}}>0$ , for all $\lambda > 0$

  2. $\limsup\limits_{x\to\infty}\frac{\overline{F}(x)}{e^{-\lambda x}}=\infty$ , for all $\lambda > 0$

However, I am not sure if those two are equivalent. I don't know how to prove it, nor can I find a counterexample. I thought that maybe $X\sim Exp(\mu)$ would not satisfy this as somewhat of an edge case, but it isn't a counterexample, because of the "for all $\lambda>0$" part.

Can you prove the equivalence of those definitions, or disprove it?

  • $\begingroup$ Indeed the "for all $\lambda$" part is crucial. To go from 1. to 2., note that 1. for $\lambda$ implies $$\lim_{x\to\infty}\frac{\overline{F}(x)}{e^{-\mu x}}=\infty$$ for every $\mu<\lambda$ hence 2. holds. To get a feeling of wether the other implication holds, did you try to build some counterexamples? $\endgroup$
    – Did
    Feb 3 '17 at 6:30
  • $\begingroup$ Any complementary CDF $\bar F$ such that $\bar F(x)=e^{-\sqrt{x}}$ for some unbounded values of $x$ and $\bar F(x)=e^{-x^2}$ for some unbounded values of $x$ would disprove the implication $2.\implies 1.$ And it seems possible to build such functions $\bar F$, which leads to the question: Are you sure about the equivalence? $\endgroup$
    – Did
    Feb 3 '17 at 6:38
  • $\begingroup$ @Did Thanks for the comments! In the first one, did you mean $\mu>\lambda$ so we have $e^{\lambda x}\bar{F}(x)e^{(\mu-\lambda) x}$ which goes to $\infty$ as the first part is larger than zero? For the second comment, no, I'm not sure that they are equivalent, that's why I asked here. I don't really understand what you meant with those two complementary cdfs. To combine them somehow into one or what? Also, I should note that we are considering only positive random variables, though I don't think it makes much difference. $\endgroup$
    – Blaza
    Feb 3 '17 at 10:20
  • $\begingroup$ Re 1.: Yes, $\mu>\lambda$, sorry about the typo. Re 2.: No, the idea is to have a unique complementary CDF $\bar F$ such that $\bar F(x_n)=e^{-\sqrt{x_n}}$ and $\bar F(y_n)=e^{-y_n^2}$ with $x_n\to\infty$ and $y_n\to\infty$. $\endgroup$
    – Did
    Feb 3 '17 at 10:34
  • 1
    $\begingroup$ A specific example is to consider $$x_n=2^{4^n}\qquad a_n=e^{-\sqrt{x_n}}\qquad \frac1c=\sum_{n=0}^\infty a_n$$ and to assume that $$P(X=x_n)=ca_n$$ for every $n\geqslant0$. Then, note that $$x_{n+1}=x_n^4$$ and that $$\bar F(x_n)=\bar F(x_{n+1}-1)=c\sum_{k=n+1}^\infty a_k\sim ca_{n+1}=ce^{-\sqrt{x_{n+1}}}=ce^{-x_n^2}$$ hence $$\bar F(x_n)\sim ce^{-x_n^2}\qquad\bar F(x_n-1)\sim ce^{-\sqrt{x_n}}\sim ce^{-\sqrt{x_n-1}}$$ as desired. $\endgroup$
    – Did
    Feb 3 '17 at 11:22

Following the comments from Did.

The statements aren't equivalent.

One direction holds, i.e. the first statement implies the second, as $$\liminf\limits_{x\to\infty}\frac{\overline{F}(x)}{e^{-\lambda x}}>0\implies(\forall \mu>\lambda)\lim\limits_{x\to\infty}\frac{\overline{F}(x)}{e^{-\mu x}}=\lim\limits_{x\to\infty}\frac{\overline{F}(x)}{e^{-\lambda x}}e^{(\mu-\lambda) x}=\infty.$$ The other direction isn't true. A counterexample follows.

We'll define $$x_n=2^{4^n},\quad a_n=e^{-\sqrt{x_n}},\quad \frac1c=\sum\limits_{n=0}^{\infty}a_n,$$

and assume that $$P(X=x_n)=ca_n,\quad \forall n\in\mathbb{N}_0.$$

It follows that $x_{n+1}=x_n^4$ and $$\overline{F}(x_n)=\overline{F}(x_{n+1}-1)=c\sum\limits_{k=n+1}^{\infty}a_k \, \sim\,ca_{n+1}=ce^{-\sqrt{x_{n+1}}}=ce^{-x_n^2}.$$ Hence $$\overline{F}(x_n)\sim ce^{-x_n^2},\quad\overline{F}(x_n-1)\sim ce^{-\sqrt{x_{n}}}\sim ce^{-\sqrt{x_{n}-1}}.$$ Thus we have $$(\forall\lambda>0)\limsup\limits_{x\to\infty}\frac{\overline{F}(x)}{e^{-\lambda x}}\geq \limsup\limits_{n\to\infty}\frac{\overline{F}(x_n-1)}{e^{-\lambda (x_n-1)}}=\limsup\limits_{n\to\infty}\frac{ce^{-\sqrt{x_{n}-1}}}{e^{-\lambda (x_n-1)}}=\infty,$$ but $$\liminf\limits_{x\to\infty}\frac{\overline{F}(x)}{e^{-\lambda x}}\leq \liminf\limits_{n\to\infty}\frac{\overline{F}(x_n)}{e^{-\lambda x_n}}=\liminf\limits_{n\to\infty}\frac{ce^{-x_n^2}}{e^{-\lambda x_n}}=0.$$


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.