I was wondering if there is a general procedure of solving for the primary variables of a linear or quadratic or, in general, a convex program after already having solved the dual program.

The problem I am working on is a least squares problem with l1-norm regularization terms and these regularization terms can be changed into constraints on a quadratic program for the dual variables. I can solve this for the dual variables. One of these ends up being the residual between the data and the model. So I can easy arrive at the model. However the model is a sum of various terms and I would like to know them individually, not just their sum (the model).

The only way I know how to get them is to do a second minimization problem of minimizing the l1-norms with the least squares term now held constant (so negligible). This is even more involved than the first problem but still reducible to a linear program and so is solvable.

However I suspect this is not the simplest route and that there should be a way of writing the primary variable solution in terms of the dual variable solutions.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.