Let $(X_i)_{i\in\mathbb{N}}$ be a sequence of $L^2$ random variables with expected value $m$ for all $n$. Let $S_n=\sum_{i=1}^n X_i$ and $|\mathrm{Cov}(X_i,X_j)|\leq\epsilon_{|i-j|}$ for finite, non-negative constants $\epsilon_k$. Show that:

(1) If $\lim_{n\to\infty} \epsilon_n=0$ then $S_n/n\to m$ in $L^2$ and probability

(2) If $\sum_{k=1}^\infty \epsilon_k<\infty$, then $\mathrm{Var}(S_n/n)$ is of order $O(1/n)$ and $S_n/n\to m$ almost surely

(1) First of, I found a similar looking question here, but we don't have, that the constants are bounded by $1$, so I don't know if the Chebyshev-inequality approach works here.

(2) We have $\mathrm{Var}(S_n/n)=\dfrac{1}{n^2}Var(S_n)=\dfrac{1}{n^2}\sum_{i\ne j}\mathrm{Cov}(X_i,X_j)=\dfrac{1}{n^2}(\sum_{k=1}^n\mathrm{Cov}(X_i,X_i)+\sum_{i=1}^{n-1}\sum_{j=i}^n \mathrm{Cov}(X_i,X_j))\leq\dfrac{1}{n^2}(n\cdot\epsilon_0+\sum_{i=1}^{n-1}\sum_{j=i}^n \mathrm{Cov}(X_i,X_j))$.
This is where I'm stuck.


We can assume that $m=0$. Then take $\delta\gt 0$ write for a fixed $C$: $$\frac 1{n^2} \mathbb E[S_n^2]\leqslant \frac 1{n^2}\sum_{i,j=1}^n\varepsilon_{|i-j|} [|i-j|\leqslant C]+\frac 1{n^2}\sum_{i,j=1}^n\varepsilon_{|i-j|} [|i-j|\gt C].$$ If we choose it such that $\varepsilon_k \leqslant \delta$ if $k\geqslant C$, then we have $$\frac 1{n^2} \mathbb E[S_n^2]\leqslant \frac{2Cn}{n^2}\sup_k \varepsilon_k+\delta,$$ hence the sequence $\left(\mathbb E[S_n^2]/n^2\right)_{n\geqslant 1}$ converges to $0$.

Once we showed the good asymptotic of $\mathbb E[S_n^2]$, an application of the Borel-Cantelli's lemma yields $2^{-n}S_{2^n}\to 0$ almost surely. Indeed, define for a fixed $j$ the event $A_n:=\{|S_{2^n}|\gt 2^n/j\}$ for a fixed $\delta$. Then the series $\sum_n\mathbb P(A_n)$ is convergent, and we get $\mathbb P\left(\limsup_{n\to\infty} A_n \right)=0$. As a consequence, on a set $\Omega_j$ of probability one, we have $|S_{2^n}(\omega)| 2^n\lt1/j $ for $n\geqslant N(j,\omega)$.

In order to deduce the result for the whole sequence, note that $$\mu\left\{\max_{2^n\leqslant k\lt 2^{n+1}}\frac{|S_k|}{2^n}>\delta\right\} \leqslant 2^n\max_{2^n\leqslant k\lt 2^{n+1}}\mu\left\{\frac{|S_k|}{2^n}>\delta\right\}\\ \leqslant 2^n\max_{2^n\leqslant k\lt 2^{n+1}}\frac 1{\delta^2}\mathbb E\left[\frac{S_k^2}{2^{2n}}\right].$$

| cite | improve this answer | |
  • $\begingroup$ If I understand the first part correctly, we need $e_k\to 0$ to be able to choose such a $\delta$. I can't really follow you for the second part, could you expand on it? Edit: By that I mean the application of Borell-Cantelli. $\endgroup$ – blst Jan 6 '15 at 17:36
  • $\begingroup$ @blst I have edited. $\endgroup$ – Davide Giraudo Jan 6 '15 at 22:29
  • $\begingroup$ How do we know that $\sum_n\mathbb P(A_n)$ is convergent? $\endgroup$ – blst Jan 7 '15 at 13:13
  • $\begingroup$ Use Chebychev's inequality to bound $\mathbb P(A_n)$. $\endgroup$ – Davide Giraudo Jan 7 '15 at 13:53
  • $\begingroup$ @DavideGiraudo Regarding the part for the whole sequence: Shouldn't we prove that $\mu\left\{\max_{2^n\leqslant k\lt 2^{n+1}}{|S_k|}/{2^n}>\delta\right\}$ is less than or equal to a summable sequence (in $n$)? It seems to me that, since $\mathbb E[S_k^2]=O(k)$ and $2^n \leq k<2^{n+1}$, the last step is bounded by $B/\delta^2$ for some $B>0$, hence it is bounded by a constant (because $\delta$ is fixed). Am I missing something? $\endgroup$ – Monte Carlo Aug 10 '15 at 10:49

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.