2

I am aware that in a convex optimization problem, the initial solution does not matter as the algorithm guarantees convergence to the global minimum/maximum. But what if the initial solution does not satisfy one of the constraints? I.e. it is technically not a "solution" as it violates a constraint. Will using this as the initial guess still lead to the global solution?

Thanks.

Tomas Jorovic
  • 3,983
  • 3
  • 27
  • 38
  • One technique is to 'minimize' the maximum constraint violation until the current solution becomes feasible. This is still a convex problem. – copper.hat Dec 19 '12 at 03:31

1 Answers1

1

It depends on what algorithm is being used, and even on the specific implementation of the algorithm: the violation of the constraint might create a situation which should never happen when the constraints are satisfied, and so the algorithm itself might not specify what to do in such a case.

Robert Israel
  • 448,999
  • Hi Robert, I see. Thanks for the input. I'm using a gradient based method by the way, in particular, Projected Gradient Descent. – Tomas Jorovic Dec 19 '12 at 01:50
  • I think in "projected gradient descent", starting from an infeasible point should not matter, because projections will bring it inside anyway. – user25004 Dec 19 '12 at 04:19