On the removal of ill conditioning effects in the computation of optimal controls

First, it is shown how ill conditioning effects may arise when discrete optimal control problems with linear dynamics, and convex cost and constraints are solved by ''primal'' methods such as linear or quadratic programming, or certain gradient methods. A dual method is then presented, which exploits to the utmost the dynamical structure of the optimal control problem. This method has been found to perform very well and does not suffer from ill-conditioning effects.