I am trying to work out an example I came up with in which I have to compute the variance of a squared sum of dependent random variables: \begin{align*} W:=\mathbf{V} \left( \sum_{i=1}^n X_i \right)^2. \end{align*} To motivate a bit, we know that if the $X_i$ are iid then $W \leq cn^2$ for some constant $c$, see e.g. Variance of squared sum of i.i.d random variables. I suspect that something similar is true in the case the $X_i$ are only "weakly dependent". Let's take this to mean, in the present situation, that $$ corr(X_i^2,X_j^2) \leq r^{|i-j|}, 0<r<1\\ corr(X_i,X_j) \leq r'^{|i-j|}, 0<r'<1 $$ for all $i,j$.


Assuming the above and that the $X_i$ have, say, uniformly bounded moments of order 2 and 4, $ E X_i^2 \leq k, E X_i^4 \leq k'$ for all $i$, can we work out a bound of the form $W \leq C n^2$ where $C$ only depends on $r,r',k, k'$?

Remark: I've been trying to compute this for a few hours now, writing $$ \mathbf{V} \left( \sum_{i=1}^n X_i \right)^2 = \sum_{i,j,k,l} Cov (X_iX_j, X_lX_k). $$ The problem here seems to bound the $(i,j)$ and $(k,l)$ pairwise dependencies, and if anyone knows how to do this I would appreciate the help.

Remark2: I did a similar thing for $\mathbf{V} \left( \sum_{i=1}^n X_i \right)\leq Cn$ where $C$ is roughly speaking proportional to the $l^1$ norm of $r'^{|i|}$ above, which sort of inspired this question.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.