Rescaling Eigenvectors of Correlation Matrix

97 Views Asked by At

I am looking to scale a set of conditional eigenvalues calculated for a normalized dataset back to the original basis of the data.

Please bear with me as linear algebra is not my forte! Please feel free to correct me in places where I have gone wrong.

  • a normalized set of data, $Z$, where;

    $Z = \frac{A - E(X)}{\sigma(X)}$

    the expectation values and standard deviations, $E(X)$ and $\sigma(X)$ are known while $A$ is an unknown dataset. I am using the normalized dataset since I am comparing variables that are not the same unit.

  • the associated covariance matrix, $\Sigma(Z)$, where:

    $\Sigma_{norm}(Z) = \frac{1}{n-1}Z^{T}Z$

  • the conditional covariance, $\Sigma_{norm}(X |Y)$, where:

    $Z = \left( \begin{array}{c} X \\ \hline Y \end{array} \right) $

    $\Sigma_{norm}(Z) = \left( \begin{array}{c|c} J & L^{T} \\ \hline L & K \end{array} \right) $

    $\Sigma_{norm}(X |Y) = J - L^{T}K^{-1}L$,

  • The 'conditional eigenvalues' of the conditional matrix

    $\Sigma_{norm}(X |Y) = U_{norm}^{T}D_{norm}U_{norm}$

In the normalized basis, the radii of the hyperellipse are defined by:

$r_{i} = \sqrt{\lambda_{i}}U_{i}$

In my case this gives a 3D Ellipse as follows for a certain Z:

Eigenvector Ellipse

I wish to obtain the same ellipse, scaled to the basis of $A$.

My attempt at a solution:

I notice that calculating the covariance matrix for Z is the same as calculating the correlation matrix. The covariance matrix and correlation matrix are related according to:

$\Sigma_{Norm}(X) = (\text{diag}(\Sigma(X))^{\frac{-1}{2}}\Sigma(X)(\text{diag}(\Sigma(X)))^{\frac{-1}{2}}$

Since $(\text{diag}(\Sigma))^{\frac{-1}{2}}$ is symmetric, I can obtain the unnormalized covariance matrix by using:

$\Sigma(X) = (\text{diag}(\Sigma(X)))^{\frac{1}{2}}\Sigma_{norm}(X)(\text{diag}(\Sigma(X)))^{\frac{1}{2}}$

Ideally then, I need to calculate $\text{diag}(\Sigma(X|Y))^{\frac{1}{2}}$. Is there a way to do this that does not require the calculation of the covariance matrix?

Can anyone offer any advice?

Kindest regards!