Consider $$D=|D|=\frac{\sum_{i=1}^n \sum_{j=1}^n I[|{\sqrt{T}}\hat\theta_{ij,T}|>f(n)]}{g(n)}, $$ where $i \ne j$, $I$ is the indicator function, $ |{\hat\theta_{ij,T}}|<1$, $f(n)=O(\ln(n))$, $g(n)=O(n^2)$, and $f(n)/T=o(1)$.

The aim is to show that $D$ converges almost surely to $0$.

Taking the expectation, I know the following:

$$ E|D|=\frac{\sum_{i=1}^n \sum_{j=1}^n Pr[|{\sqrt{T}\hat\theta_{ij,T}}|>f(n)]}{g(n)}\, {\le}\,\,{\sup_{ij}}\,Pr[|{\sqrt{T}\hat\theta_{ij,T}}|>f(n)] \,\, {\le} \,\, {e^{-\frac{f(n)}{\rho}}} $$

where ${\rho}={\sup_{ij}}\, {\rho_{ij}}$ is positive and bounded.

From the Markov inequality: $$ Pr(|D|>{\epsilon})\,\,{\le}\,\,\frac{E|D|}{\epsilon} \,\,{\le}\,\, \frac{{e^{-\frac{f(n)}{\rho}}}}{\epsilon} $$ for some positive $\epsilon$ and hence $\lim_{n,T\to\infty} Pr(|D|>{\epsilon})=0.$

Is this result sufficient for $D$ to converge almost surely to 0? If not, how can I show that $D$ converges almost surely to 0?

  • 2
    $\begingroup$ This is just a suggestion but I recommend restricting the use of latex to the equations in your question, since mathjax takes extra time to render in the browser (especially on mobile devices). $\endgroup$ – j.c. Dec 10 '17 at 20:42
  • 1
    $\begingroup$ The formatting is indeed very strange, and most likely done by some robot (converting from another format?), maybe somebody will guess what exactly. $\endgroup$ – YCor Dec 10 '17 at 21:38
  • 1
    $\begingroup$ Thank you @j.c.. I am very new to this. I will try that next time. $\endgroup$ – user0735 Dec 10 '17 at 21:44

What you have proved is that the $D$ converges in probability as $n\rightarrow\infty$, and I do not see why $T\rightarrow\infty$ should involve in the limit in your conclusion since you neither assume decay speed of $\hat{\theta}_{i,j,T}$ nor restrict the increasing speed of $T$. So I think you missed some assumption here.

Moreover, if $I[|{\hat\theta_{ij,T}}|>f(n)]$(or the double sum itself is monotonic) is monotonic as $T,n\rightarrow\infty$, which is very likely to be true since $|{\hat\theta_{i,j,T}}|<1$. Then using Skorohod Theorem you know there is a subsequence $Z_{k_n}$ of $$Z_n:=\sum_{i=1}^n \sum_{j=1}^n I[|{\hat\theta_{ij,T}}|>f(n)]$$ which converges to $0$ almost surely $P$, and since such a sequence $Z_{k_n}$ is selected from a monotonic sequence $Z_n$, the original sequence must also converge almost surely by comparison lemma.

  • $\begingroup$ Thank you @Henry.L for your response. Sorry I forgot T, which I have now added. Without the use of monotonicity, would it make sense to show that the sum of n from 1 to infinity of the exponential term is bounded in order to get almost sure convergence? (thinking along the lines of the Borel-Cantelli lemma). $\endgroup$ – user0735 Dec 10 '17 at 21:55
  • $\begingroup$ I think you need monotonicity with $T$ taken into consideration now. I think a bit into BC lemma but do not quite see how it comes into play in this scenario... $\endgroup$ – Henry.L Dec 11 '17 at 15:34
  • $\begingroup$ I don't quite see how with an indicator function I can have monotonicity. $\endgroup$ – user0735 Dec 13 '17 at 12:23

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.