1: \begin{abstract}

2:   This paper studies the projected saddle-point dynamics associated to

3:   a convex-concave function, which we term saddle function.  The

4:   dynamics consists of gradient descent of the saddle function in

5:   variables corresponding to convexity and (projected) gradient ascent

6:   in variables corresponding to concavity.  \changes{We examine the

7:     role that the local and/or global nature of the

8:     convexity-concavity properties of the saddle function plays in

9:     guaranteeing convergence and robustness of the dynamics. } Under

10:   the assumption that the saddle function is twice continuously

11:   differentiable, we provide a novel characterization of the

12:   omega-limit set of the trajectories of this dynamics in terms of the

13:   diagonal blocks of the Hessian. Using this characterization, we

14:   establish global asymptotic convergence of the dynamics under local

15:   strong convexity-concavity of the saddle function. When strong

16:   convexity-concavity holds globally, we establish three results.

17:   First, we identify a Lyapunov function \changes{(that decreases strictly along

18:   the trajectory)} for the projected

19:   saddle-point dynamics when the saddle function corresponds to the

20:   Lagrangian of a general constrained convex optimization

21:   problem. Second, \changes{for the particular case} when the saddle

22:   function is the Lagrangian of an \changes{equality-constrained}

23:   optimization problem, we show input-to-state stability of the

24:   saddle-point dynamics by providing an ISS Lyapunov function. Third,

25:   we \changes{use the latter result to} design an opportunistic

26:   state-triggered implementation of the dynamics. Various examples

27:   illustrate our results.

28: \end{abstract}

29: