fmudesign: singular correlation matrix

[runhol]

The latest fmudesign uses transformation of correlated variables via Cholesky factorization. To perform the Cholesky
factorization the matrix has to be symmetric positive definite (non-singular), Could this be generalized to the case
of singular correlation matrices (symmetric positive semidefinite) by using f.ex. SVD instead of Cholesky?

[dafeda]

I think this is technically possible, but not sure it's something that we want to do.
One way the correlation matrix becomes singular is if two or more variables are 100% correlated.
When do we want to model using 100% correlated variables?
Great if @tralsos can chip in here.

[tralsos]

I think the use case for 100% correlated variables is there. E.g. you have a case where want to test what is the oil water contact in two segments is the same, compared to that they vary independently. Or you have some prediction parameters that are correlated, but you want to test what happens if they are 100% correlated.

[azizovilgar7]

I have a similar issue using it, as we have some cases similar to what Trine mentioned.

[tralsos]

Input from Gyrid Johnsen:

Correlation 1 is (often) used for FWL and GOC uncertainties.
I have encountered this issue in two different projects lately (and luckily I knew of the change on fmudesign)

1. In one solved by setting a constant dummy number (-999) in design input and copying the relevant columns in the final design matrix
2. In the other by setting correlation 0.99999 and having no decimals in the output.
3. Can also be solved with @risk or by scripting.
   Solution 1) requires manual steps – solution 2 is ok if you don’t need decimals - solution 3) is not transparent.
   So it would help a lot of projects if this feature is “fixed”.