[Scikit-learn-general] Fwd: Use of SVD in LDA

[Derek Seiple]

I am trying to write a c++ port of the LDA.fit and LDA.predict_proba
functions.

What I don't understand is why the training can use the Singular Value
Decomposition. The end result is that the decision_function uses the
"scalings" values (among other things) obtained from the training to
calculate the probabilities.

Can someone please help explain from a conceptual point of view what the
"scalings" are and how they are obtained/why they are used? Some help
understanding why the SVD can be used in the training would be helpful as
well.

Thank you,

Derek