1: \begin{abstract}

2:         We study the long time behavior of an underdamped mean-field Langevin (MFL) equation,

3: 	and provide a general convergence as well as an exponential convergence rate result under different conditions.

4: 	The results on the MFL equation can be applied to study the convergence of the Hamiltonian gradient descent algorithm for the overparametrized optimization.

5: 	We then provide a numerical example of the algorithm to train a generative adversarial networks (GAN).

6: \end{abstract}

7: