1: \begin{abstract}%

2: 

3: Generative Adversarial Networks (GANs) 

4: learn an implicit generative model 

5: from data samples through a two-player game. 

6: In this paper, we study the existence of Nash equilibrium 

7: of the game which is consistent as the number of data samples grows to infinity.

8: In a realizable setting where the goal is to estimate

9: the ground-truth generator of a stationary Gaussian process, 

10: we show that the existence of consistent Nash equilibrium 

11: depends crucially on the choice of the discriminator family. 

12: The discriminator defined from 

13: second-order statistical moments 

14: can result in non-existence of Nash equilibrium, 

15: existence of consistent non-Nash equilibrium, 

16: or existence and uniqueness of consistent Nash equilibrium,

17: depending on whether symmetry properties of the generator family

18: are respected. We further study empirically

19: the local stability and global convergence of gradient descent-ascent 

20: methods towards consistent equilibrium.

21: 

22: \end{abstract}

23: