I am on a project and I had to study about signal detection, Bayes decision theory etc. The basic paper I am reading, gives an equation for LLR, referring to these equations:

$H_0 : Y\sim N(\mu_0,\Sigma_0)$

$H_1 : Y\sim N(\mu_1,\Sigma_1)$

The log-likelihood ratio is given then by:

$$\log(L(y))=\frac{1}{2}\left[\log\frac{| \Sigma_0 |}{| \Sigma_1 |} + y^T(\Sigma_0^{-1} - \Sigma_1^{-1})y+2 (\mu_1^T\Sigma_1^{-1} -\mu_0^T \Sigma_0^{-1} )y+\mu_0^T \Sigma_0^{-1}\mu_0-\mu_1^T\Sigma_1^{-1}\mu_1\right]$$

But as I wanted to reproduce the final equation, I ended in this: $$\log(L(y))=\frac{1}{2}\left[\log\frac{| \Sigma_0 |}{| \Sigma_1 |}+y^{T}(\Sigma _0^{-1}-\Sigma_1^{-1})y+(\mu_1^T\Sigma_1^{-1}-\mu_0^T \Sigma_0^{-1}) y + \mu_0^T \Sigma_0^{-1} \mu_0-\mu_1^T\Sigma_1^{-1}\mu_1+y^T(\Sigma_1^{-1}\mu_1-\Sigma_0^{-1} \mu_0)\right]$$

The above are equal if only:

$$y^T(\Sigma_1^{-1}\mu_1-\Sigma_0^{-1}\mu_0) = (\mu_1^T\Sigma_1^{-1}-\mu_0^T\Sigma_0^{-1})y$$

which I find obviously wrong.

Can someone help me with this?

  • $\begingroup$ The equality holds if the covariance matrices are symmetric, which is true. $\endgroup$ – Chester Jul 13 '15 at 23:49
  • $\begingroup$ Ok, I really missed that! Thank you! $\endgroup$ – user7127 Jul 14 '15 at 0:11

$$ y^T(\Sigma_1^{-1}\mu_1-\Sigma_0^{-1}\mu_0) = (\mu_1^T\Sigma_1^{-1}-\mu_0^T\Sigma_0^{-1})y $$ Either of these is the transpose of the other. But these are $1\times1$ matrices. Every $1\times1$ matrix of real numbers is its own transpose.


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.