1: \begin{abstract}

2: We consider a continuous-time optimization method

3: based on a dynamical system, where 

4: a massive particle starting at rest moves in the 

5: conservative force field generated by the objective 

6: function, without any kind of friction.

7: We formulate a restart criterion 

8: based on the mean dissipation of the kinetic

9: energy, and we prove a global convergence

10: result for strongly-convex functions. 

11: Using the Symplectic Euler discretization

12: scheme, we obtain an iterative optimization 

13: algorithm. 

14: We  have considered a discrete mean dissipation

15: restart scheme, but we have also

16:  introduced a new restart procedure based on

17: ensuring at each iteration a decrease of

18: the objective function greater than the one

19: achieved by a step of the classical gradient method.

20: For the discrete conservative algorithm,

21: this last restart criterion is capable of

22: guaranteeing a qualitative

23:  convergence result.

24: We apply the same restart scheme to the

25: Nesterov Accelerated Gradient (NAG-C), and we

26: use this restarted NAG-C as benchmark in the

27: numerical experiments.

28: In the smooth convex problems considered,

29: our method shows a faster convergence rate than

30: the restarted NAG-C.

31: We propose an extension of our discrete conservative

32: algorithm to composite optimization: in

33: the numerical tests involving non-strongly convex

34: functions with $\ell^1$-regularization,

35:  it has better

36: performances than the well known efficient

37: Fast Iterative Shrinkage-Thresholding Algorithm,

38:  accelerated with an adaptive restart scheme.

39: \keywords{Convex optimization,

40: accelerated first-order optimization, 

41: restart strategies, conservative dynamical model.}

42: % \PACS{PACS code1 \and PACS code2 \and more}

43: % \subclass{MSC code1 \and MSC code2 \and more}

44: 

45: \end{abstract}

46: