Hello,

In the context of ridge regression I need to compute:

Code: Select all

(X*X^T + lambda*Identity)

where lambda is a scalar. In order to compute the "lambda * Identity" I did:

Code: Select all

MatrixXd reg_I = lambda * MatrixXd::Identity(m, n);

but this turns out to be slow (and unnecessary) i.e. multiplying by a bunch of zeros. I then tried:

Code: Select all

DiagonalMatrixXd reg_I(m, n);
reg_I.identity();
reg_I *= lambda;

but apparently I can't combine DiagonalMatrix and MatrixXd in the main expression? I get a compilation error for this:

Code: Select all

(X*X^T + reg_I)

Then the question is how can I do this elegantly and efficiently in Eigen and by elegantly I mean simply not compromising on the code readability and end up writing a loop over the diagonal elements of X*X^T and adding lambda.

TIA,
Best regards,
Giovanni

First compute X2 = X*X^T and then do:

Code: Select all

X2.diagonal().array() -= lambda;


This is what X*X^T - reg_I should do for you in the future.

hi ggael,

Always excellent support, Thank you

I will miss having the visible ridge formula but the performance improvement really does justify it.

Cheers,
Giovanni