1: \begin{abstract}

2: In optimization the duality gap between the primal and the dual problems is a

3: measure of the suboptimality of any primal-dual point. In classical mechanics

4: the equations of motion of a system can be derived from the Hamiltonian

5: function, which is a quantity that describes the total energy of the system.  In

6: this paper we consider a convex optimization problem consisting of the sum of

7: two convex functions, sometimes referred to as a composite objective, and we

8: identify the duality gap to be the `energy' of the system.  In the Hamiltonian

9: formalism the energy is conserved, so we add a contractive term to the standard

10: equations of motion so that this energy decreases linearly (\ie, geometrically)

11: with time.  This yields a continuous-time ordinary differential equation (ODE)

12: in the primal and dual variables which converges to zero duality gap, \ie,

13: optimality.  This ODE has several useful properties: it induces a natural

14: operator splitting; at convergence it yields both the primal and dual solutions;

15: and it is invariant to affine transformation despite only using first order

16: information.  We provide several discretizations of this ODE, some of which are

17: new algorithms and others correspond to known techniques, such as the

18: alternating direction method of multipliers (ADMM).  We conclude with some

19: numerical examples that show the promise of our approach. We give

20: an example where our technique can solve a convex quadratic minimization problem

21: orders of magnitude faster than several commonly-used gradient methods, including

22: conjugate gradient, when the conditioning of the problem is poor.  Our framework

23: provides new insights into previously known algorithms in the literature as well

24: as providing a technique to generate new primal-dual algorithms.

25: \end{abstract}

26: