Weak Law of Large Numbers for Dependent Random Variables with Bounded Covariance

19k Views Asked by At

I'm currently stuck on the following problem which involves proving the weak law of large numbers for a sequence of dependent but identically distributed random variables. Here's the full statement:

  • Let $(X_n)$ be a sequence of dependent identically distributed random variables with finite variance.

  • Let $\displaystyle S_n = \sum_{i=1}^n X_i $ denote the $n^\text{th}$ partial sum of the random variables $(X_n)$.

  • Assume that Cov$(X_i,X_j) \leq c^{|i-j|}$ for $i, j \leq n$ where $|c| \leq 1$.

Is it possible to show that $\displaystyle \frac{S_n}{n} \rightarrow \mathbb{E}[X_1]$ in probability? In other words, is it true that given any $\epsilon>0$,

$$ \lim_{n\rightarrow \infty} \mathbb{P}\bigg[\Big|\frac{S_n}{n} - \mathbb{E}[X_1]\Big| > \epsilon\bigg] = 0$$

EDIT: Following some comments, it turns out that I had the right approach so I've gone ahead and answered my own question below.

1

There are 1 best solutions below

0
On BEST ANSWER

Fix $\epsilon > 0$ and $n \in \mathbb{N}$, then we can use Chebyshev's inequality to see that

$$\mathbb{P}\bigg[\Big|\frac{S_n}{n} - \mathbb{E}[X_1]\Big| > \epsilon\bigg] \leq \frac{\text{Var}\Big(\frac{S_n}{n}\Big)}{\epsilon^2}$$

where

$$\displaystyle \text{Var}\Big(\frac{S_n}{n}\Big)= \frac{\text{Var}(S_n)}{n^2} \leq \frac{\sum_{i=1}^n\sum_{j=1}^n \text{Cov}{(X_i,X_j)}}{n^2} \leq \frac{\sum_{i=1}^n\sum_{j=1}^n c^{|i-j|}}{n^2} $$

We can then explicitly calculate the double sum $\sum_{i=1}^n\sum_{j=1}^n c^{|i-j|}$ as follows:

$$\begin{align} \sum_{i=1}^n\sum_{j=1}^n c^{|i-j|} &= \sum_{i=1}^n c^{|i-i|} + 2\sum_{i=1}^n\sum_{j=1}^{i-1} c^{|i-j|} \\ &= n + 2\sum_{i=1}^n\sum_{j=1}^{i-1} c^{|i-j|} \\ &= n + 2\sum_{i=1}^n\sum_{j=1}^{i-1} c^{i-j} \\ &= n + 2\sum_{i=1}^n c^i \frac{1 - c^{-i}}{1-c^{-1}} \\ &= n + 2\sum_{i=1}^n \frac{c^i + 1}{1-c^{-1}} \\ &= n + \frac{2c}{c-1} \sum_{i=1}^n c^{i}-1 \\ &= n + \frac{2c}{c-1} \big(\frac{1-c^{n+1}}{1-c} -n \big)\\ &= n + \frac{2c}{(c-1)^2}(c^{n+1}+1) + \frac{2c}{c-1}n\\ \ \end{align}$$

Thus,

$$\lim_{n\rightarrow\infty} \mathbb{P}\bigg[\Big|\frac{S_n}{n} - \mathbb{E}[X_1]\Big| > \epsilon\bigg] = \lim_{n\rightarrow\infty} \frac{\text{Var}\Big(\frac{S_n}{n}\Big)}{\epsilon^2} \leq \lim_{n\rightarrow\infty} \frac{n + \frac{2c}{(c-1)^2}(c^{n+1}+1) + \frac{2c}{c-1}n}{n^2 \epsilon^2} = 0 $$

Seeing how our choice of $\epsilon$ was arbitrary, the statement above holds for any $\epsilon > 0 $ and shows that $\frac{S_n}{n} \rightarrow E[X_1]$ in probability, as desired.

This shows the validity of the theorem for $c<1$, but not for $c=1$. We can easily extend the demonstration to all cases in which $|\mbox{Cov}(X_i,X_j)|\le f_{|i-j|}$ where $\lim_{i\to\infty}f_i=0$. Indeed in this case it is simple to show that $$\lim_{n\to\infty}{1\over n^2}\sum_{i=1}^n\sum_{j=1}^n \text{Cov}{(X_i,X_j)}\le \lim_{n\to\infty}{1\over n^2}\sum_{i=1}^n\sum_{j=1}^n |\text{Cov}{(X_i,X_j)}|\le \lim_{n\to\infty}{1\over n^2}\sum_{i=1}^n\sum_{j=1}^nf_{|i-j|}=0$$