1: \begin{abstract}

2:   This paper presents a robust, distributed algorithm to solve general

3:   linear programs. The algorithm design builds on the characterization

4:   of the solutions of the linear program as saddle points of a

5:   modified Lagrangian function.  We show that the resulting

6:   continuous-time saddle-point algorithm is provably correct but, in

7:   general, not distributed because of a global parameter associated

8:   with the nonsmooth exact penalty function employed to encode the

9:   inequality constraints of the linear program. This motivates the

10:   design of a discontinuous saddle-point dynamics that, while enjoying

11:   the same convergence guarantees, is fully distributed and scalable

12:   with the dimension of the solution vector.

13:   % In particular, this dynamics is appropriate when the global

14:   % objective is to have the aggregate of agent states converge to a

15:   % solution of a linear program. This is in contrast to other

16:   % consensus-based distributed optimization formulations in which

17:   % agents must agree on the entire solution vector. 

18:   We also characterize the robustness against disturbances and link

19:   failures of the proposed dynamics.  Specifically, we show that it is

20:   integral-input-to-state stable but not input-to-state stable. The

21:   latter fact is a consequence of a more general result, that we also

22:   establish, which states that no algorithmic solution for linear

23:   programming is input-to-state stable when uncertainty in the problem

24:   data affects the dynamics as a disturbance. Our results allow us to

25:   establish the resilience of the proposed distributed dynamics to

26:   disturbances of finite variation and recurrently disconnected

27:   communication among the agents.  Simulations in an optimal control

28:   application illustrate the results.

29: \end{abstract}

30: