I am reading Berkes et al. (2003) about the GARCH model. Could someone help me figure out the proof of one lemma in the paper?

$\mathbf{Lemma.}$ If $\{\xi_k, 0\leq\xi_k<\infty\}$ is a sequence of identically distributed random variables satisfying \begin{equation} \mathrm{E}\log^{+}|\xi_0|<\infty, \tag{1} \end{equation} then $\sum_{k=0}^{\infty}z^{k}\xi_k$ converges with probability one for any $|z|<1$. Note, $\log ^{+} x=\log x$ if $x>1$, and $0$ otherwise.

$\mathbf{Proof.}$ By the Borel-Cantelli lemma it is enough to prove that, for any $\zeta>1$, \begin{equation} \sum_{k=1}^{\infty}P\{|\xi_{k}|>\zeta^{k}\}<\infty. \tag{2} \end{equation}

The distribution of $\xi_k$ does not depend on $k$, so \begin{align} \sum_{k=1}^{\infty} P\{|\xi_{k}|>\zeta^{k}\} &=P\{\log^{+}|\xi_k|>k\log\zeta\}\nonumber \\ &=\sum_{k=1}^{\infty} P\{\log^{+}|\xi_0|>k\log\zeta\} \tag{3} \\[10pt] &\leq\mathrm{E} \log^{+} |\xi_0|/\log \zeta, \tag{4} \end{align} and thus $(1)$ implies $(2)$.


$\mathbf{Question.}$ It seems to me that $(3)$ does not imply $(4)$. It is natural to apply the Markov inequality to $(3)$ and we have $\sum_{k=0}^{\infty} P\{\log^{+}|\xi_0|>k\log\zeta\}\leq \sum_{k=1}^{\infty}\frac{\mathrm{E}\log^{+}|\xi_0|}{k\log\zeta}$. Since the harmonic sequence, $\sum_{k=1}^{\infty}\frac{1}{k}$, does not converge, we cannot get $(4)$ by using the Markov inequality. Did I miss something here?


  • $\begingroup$ You should add a reference to the paper. $\endgroup$ – Juho Kokkala Aug 5 '16 at 3:53
  • 1
    $\begingroup$ I think I figured it out by using $\mathrm{E}X=\int_{0}^{\infty} P(X>s)ds$ for any nonnegative random variable $X$. $\endgroup$ – JRBNB Aug 5 '16 at 15:23
  • $\begingroup$ Could you edit the title to highlight the actual problem? I don't know the terminology of these inequalities, but probably you do, so you could include a relevant name. Having GARCH in the title is not useful, IMHO. Also, the GARCH tag seems irrelevant as the inequality is probably not intrinsically specific to GARCH models, is it? $\endgroup$ – Richard Hardy Aug 5 '16 at 17:00
  • $\begingroup$ My original title is 'Is this proof wrong?' And the only tag I used is 'probability'. Some people help me make modifications. If you like you also can do it. $\endgroup$ – JRBNB Aug 5 '16 at 17:33

I think I asked a dumb question. The proof is right. The missing step in the proof is as follows.

\begin{align} &~~~~\sum_{k=1}^{\infty} P\{\frac{\log^{+}|\xi_0|}{\log\zeta}>k\}\\ &\leq \sum_{k=1}^{\infty} \int_{k-1}^k P\{\frac{\log^{+}|\xi_0|}{\log\zeta}>s \}ds\\ &=\mathrm{E} \log^{+} |\xi_0|/\log \zeta, \end{align}


Your argument does not invalidate their argument. Basically, they say $A<B$, where $B<\infty$, while you are saying, $A<B\times C$, where $C=\infty$. Both arguments are compatible and your argument does not invalidate theirs.

  • $\begingroup$ @JRBNB I don't know whether they used it or not. In fact, you should provide a link to the paper (specially because you are copy-pasting bits of it). My point is about the material you just posted. Your inequallity corresponds to $A< B\times C$, where $A < B$ is their inequality. $\endgroup$ – Kevin Levrone Aug 4 '16 at 23:31
  • $\begingroup$ @JRBNB I have to go now. I suggest you post your questions when you have more time to interact and show more interest in your own questions. Good luck with your reading. $\endgroup$ – Kevin Levrone Aug 4 '16 at 23:37

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.