2
$\begingroup$

I am looking for a method to decorrelate several variables, so that their covariance matrix is diagonal, while keeping the original mean for each of them.

I found this old article which seemed pretty much what I was looking for: http://blogs.sas.com/content/iml/2012/02/08/use-the-cholesky-transformation-to-correlate-and-uncorrelate-variables.html

As I understood it:

  1. I should pick a chosen covariance matrix Cov (I chose the correlation matrix to keep the mean and standard deviation, but I am not sure it is needed)
  2. I should compute its $LL^{t}$ Cholesky decomposition matrix L(Cov)
  3. I should multiply my random variable matrix A by L(Cov) to obtain the product $B = A * L(Cov)$.
  4. The Covariance Matrix of B, Cov(B), should be equal to Cov while the mean of the modified random variables B should remain the same as the mean of the random variables of A. According to my tests, it is not.

In particular, if I used $Cov=Cov(A)^{-1}$ as input, I should obtain $L(Cov)=L(Cov(A))^{-1}$ and $Cov(B)$ should be decorrelated thus diagonal. According to my tests, it is not.

At the same time, I am confused since multiplying A by $L(I)$ with I the identity matrix used as the initial value of Cov cannot cause $Cov(B)$ to become the identity matrix if it was not initially.

I tried both with randomly distributed and normally distributed samples, I was not able to reproduce results similar to the article's and lost the means.

Did I do something wrong in my use of Cholesky decomposition to decorrelate variables ? Or if it is not supposed to work, what simple method would you advice me ?

$\endgroup$
  • $\begingroup$ Please just ask one question in a post. $\endgroup$ – Hoseyn Heydari Nov 12 '15 at 19:27
  • $\begingroup$ Thank you, I tried to edit according to your remark. $\endgroup$ – Guillaume A. Nov 13 '15 at 12:19
  • $\begingroup$ I know I could use Gram-Schmidt orthogonalisation, but it has some properties I try to avoid: the process is not symmetric between v1 and v2, and more generally, vectors are treated differently depending on the order they are added to the basis $\endgroup$ – Guillaume A. Nov 13 '15 at 15:31
2
$\begingroup$

It seems the article I quoted describes the "Whitening transformation" shown here: https://en.wikipedia.org/wiki/Whitening_transformation. Thus, I should center and reduce my variables before multiplying A by $L(Cov(A))^{−1}$. I don't get yet why this step is needed but I'll try this way.

After that, I may multiply by the original standard deviation and add the mean to recover decorrelated variables.

I'll update wether it works or not.

$\endgroup$
  • $\begingroup$ The random variables had to be centered and reduced indeed. $\endgroup$ – Guillaume A. Nov 17 '15 at 10:06
  • $\begingroup$ The process works perfectly. The covariance matrix of the decorrelated data $A'=L(Cov(A_{CenteredReduced}))^{-1}*A$ is identity. By multiplying by standard deviation and adding mean, I found what I wanted. $\endgroup$ – Guillaume A. Nov 17 '15 at 10:13

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.