How is the minimum variance unbiased estimator derived for some RV $X\sim \mathcal{N}(0,\sigma^2)$, given $\mathbf{Y}=X\cdot \mathrm{ones}_{N\times 1}+\mathbf{w}$, where $\mathbf{w}$ is $N$-dimensional multivariate-normal with mean $[0, \dots, 0]^T$ and covariance matrix $\mathbf{\Sigma}$, and is independent of $X$?

In words, what is the best estimator for a signal in correlated noise, what is the distribution?

I know the answer is well-known and buried somewhere in Kay's Estimation Theory text but I cannot find it at the present. (am still looking)

If $\mathbf{\Sigma}$ were diagonal (that is, uncorrelated noise), we know that this estimator is some linear combination of the $\mathbf{Y}$ and it's Gaussian distributed about the signal with variance $\sum_{n=1}^N \frac{1}{\mathbf{\Sigma}_{n,n}}$ (cf. Maximal-ratio combining)


1 Answer 1


From Kay's Fundamentals of Statistical Signal Processing, Volume I:

$$\textbf{Example 4.4 - DC Level in Colored Noise}$$ $\text{We now extend Example 3.3 to the colored noise case. If }x[n]=A+w[n]\text{ for }n=0,1,\dots,N-1,\text{ where }w[n]\text{ is }\textit{colored}\text{ Gaussian noise with }N\times N\text{ covariance matrix }\mathbf{C},\text{ it immediately follows from (4.25) that with }\mathbf{H}=\mathbf{1}=[1\ 1\dots 1]^T, \text{the MVU estimator of the DC level is}$ \begin{array}\hat{A} &=& (\mathbf{H}^TC^{-1}\mathbf{H})^{-1}\mathbf{H}^TC^{-1}\mathbf{x} \\ &=&\displaystyle \frac{\mathbf{1}^T\mathbf{C}^{-1}\mathbf{x}}{\mathbf{1}^T\mathbf{C}^{-1}\mathbf{1}}\end{array} $\text{and the variance is, from (4.26),}$ $$\begin{array}\displaystyle \text{var}(\hat{A}) &=& (\mathbf{H}^TC^{-1}\mathbf{H})^{-1} \\ &=& \displaystyle \frac{1}{\mathbf{1}^T\mathbf{C}^{-1}\mathbf{1}}. \end{array}$$

iThe whole series is a great reference.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.