Feedback-controlled saddle gradient processes (Q1914124)

The author considers the feedback stabilization problem for nonlinear control systems of the form \[ dx/dt= \pi_Q (x- \alpha \nabla L_x (x,p))-x, \qquad dp/dt= \pi_P (p+ \alpha \nabla L_p (x,p)) -p, \] where \((x,p)\in P\times Q\), \(P\subset \mathbb R^n\), \(Q\subset \mathbb R^m\) and the equilibrium state \(x^*\), \(p^*\) is a saddle of \(L\). Here \(\pi_Q\) and \(\pi_P\) denote the projections onto \(Q\) and \(P\), respectively. As a special case the saddle function \(L\) is supposed to be a Lagrange function of a convex programming problem. For this problem several feedback controllers depending on \(x\), \(p\), \(\dot x\) and \(\dot p\) entering additively in the arguments of \(\nabla L\) are presented and proved to be stabilizing. Furthermore the efficiency of these methods in terms of the speed of convergence to the equilibrium is discussed.