Let $$\boldsymbol{x} = \begin{bmatrix} x\\ y\\ \end{bmatrix}, \boldsymbol{\mu} = \begin{bmatrix} \mu_x\\ \mu_y\\ \end{bmatrix}, \boldsymbol{\Sigma}=\begin{bmatrix} \sigma_x & \rho\sigma_x\sigma_y\\\ \rho\sigma_x\sigma_y & \sigma_y\\ \end{bmatrix}$$ where $\rho$ is the Pearson correlation between $x$ and $y$. Suppose I have the following bivariate normal distribution

$$\boldsymbol{x} \sim N(\boldsymbol{\mu},\boldsymbol{\Sigma})$$


$$R= \begin{bmatrix} \cos(\theta) & \sin(\theta)\\ \sin(\theta) & -\cos(\theta)\\ \end{bmatrix}$$

I want to decompose the covariance matrix $\boldsymbol{\Sigma}$ such that the resulting decomposition uses rotational matrices. According to this article


I can decompose the covariance matrix via eigendecomposition

$$\Sigma = VLV^{-1}$$ and more specifically, I can write it as

$$\Sigma = RSSR^{-1}$$

where $S=\sqrt{L}$ and $R=V$. Unfortunately, I'm unfamiliar with eigendecomposition so I'm having trouble finding how to do this. The example in the article is for an uncorrelated case.

Can someone show me how to eigendecompose the covariance matrix into the $RSSR^{-1}$ form?

I tried for instance to figure out what $L$ is by doing

$$R^{-1}\Sigma R = L$$

But if I plug that into Wolfram Alpha (to check if I'm going in the right direction), I get something horribly ugly, see here

  • 1
    $\begingroup$ Any constructive feedback would be great for how to improve the question. I've spent a lot of time to narrow the problem I am facing, at each step trying to do my own work. But this step is one I just don't understand. $\endgroup$ – Stan Shunpike Jan 19 '20 at 8:30

The covariance matrix is PSD and symmetric, so a standard eigen decomposition results in $$\Sigma=VDV^T$$ where $V=V^T$. $V$ is orthonormal matrix, and for it to be rotation matrix, you also need $|V|=1$ where you can arrange it by suitable choice of signs for the eigenvectors.

Note: $\Sigma$ matrix above should contain $\sigma_x^2,\sigma_y^2$ and your example rotation matrix has $|R|=-1$, you can use the one in wikipedia.


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.