Suppose we have a sequence $x_i, \ldots, x_n$ of correlated random variables with pairwise covariances $\Sigma_{ij}$, (the $(i,j)$-entry of the covariance matrix $\Sigma$). Consider $\hat{x}=\sum_{j}c_jx_j$, where $c_j=\frac{\sum_i K_{ij}}{\sum_{i,m} K_{im}}$ (here $K_{ij}$ denotes the $(i,j)$-entry of the inverse of the covariance matrix $K = \Sigma^{-1}$). If each $x_j$ has mean $\mu$, then one can easily see that $\mathbb{E}[\hat{x}]=\mu$. However, I'm having trouble proving that $Var[\hat{x}]=Cov[\hat{x},\hat{x}]=(\sum_{i.j}K_{ij})^{-1}$. Can anyone shed some light on how to obtain this value for the variance? My instinct was to use $Var[X]=E(X^2)-[E(X)]^2$, but this only creates more trouble when we get to the $[E(X)]^2$ term. Any help is much appreciated.


Substituting in the formula for the variance the definition of the variable $\hat{x}$ and the fact that $E[\hat{x}]=\mu$:


We know that $$E(x_jx_k)=\text{Cov}[x_j,x_k]+E[x_j]E[x_k]=\Sigma_{jk}+\mu^2$$

and now we would like to try and compute the sum in the first expression. First we notice that $\sum_{i}c_i=1$. Then we substitute the 2nd equation into the sum and decompose:


so the mean of the variable cancels and we are left with the evaluation of sum which can be performed as follows:

$$\sum_{jk}{c_jc_k\Sigma_{jk}}=\frac{1}{(\sum_{im}K_{im})^2}\sum_{i_1i_2}\sum_{jk}K_{i_1j}\Sigma_{jk}K^{T}_{ki_2}\\=\frac{1}{(\sum_{im}K_{im})^2}\sum_{i_1i_2}(K\Sigma K^{T})_{i_1i_2}\\=\frac{1}{(\sum_{im}K_{im})^2}\sum_{i_1i_2}(K^{T})_{i_1i_2}\\=\frac{1}{(\sum_{im}K_{im})^2}\sum_{i_1i_2}K_{i_1i_2}=(\sum_{ij}K_{ij})^{-1}\\$$

and finally we conclude that:



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.