0
$\begingroup$

I have a model of an information transmission system Y = XH + N, where X is a diagonal matrix with the transmitted "symbols" (known), H is a column vector which distorts the transmitted symbols and N is a "noise (column) vector" which is sampled from a zero mean Gaussian distribution of known variance. An estimator of the distortion H would be: $\hat{H} = (X^H X)^{-1} X^H Y$. Now, if we apply this estimator to the model, we obtain

$(X^H X)^{-1} X^HX H + (X^H X)^{-1} X^HN = H + (X^H X)^{-1} X^HN$. This implies $(X^H X)^{-1} X^HN$ is the error in the estimator. I am interested to get the mse of this estimator. I was thinking of proceeding as follows: $C_{\tilde{N}} = E[((X^H X)^{-1} X^HN)((X^H X)^{-1} X^HN)^H]$

$ = (X^H X)^{-1} X^HE[NN^H]\ X (X^H X)^{-1}$

$ = (X^H X)^{-1} X^H (\sigma^2 I) \ X (X^H X)^{-1}$

$ = (X^H X)^{-1} X^H (\sigma^2 I) \ X (X^H X)^{-1} = \sigma^2 (X^H X)^{-1}$

We now have, I believe, the covariance matrix of the estimator noise (since X is a matrix here). Can I obtain the MSE of the estimator from this? And, in general, if we have a covariance matrix of the form $\sigma^2 P$, where P is guarenteed to be positive definite (as a covariance matrix should be) but no other information is known (P is not diagonal), how could I calculate the MSE?

$\endgroup$
  • $\begingroup$ Any citation (reference books?) to such a result would be appreciated. $\endgroup$ – Analon92 Mar 25 at 3:05
0
$\begingroup$

So it turns out this was rather simple, when I thought about it. $MSE = E[|H-\hat{H}|^2] = E[(H-\hat{H})^H (H-\hat{H})]$

$=E[((X^H X)^{-1} X^HN)^H (X^H X)^{-1} X^HN]$

which is a real number.

What I had calculated was $E[(H-\hat{H}) (H-\hat{H})^H]$, which is a rank 1 matrix. To get the MSE, we just need to sum the diagonal elements of this correlation matrix, which can be derived using basic linear algebra.

$\endgroup$

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.