1: \begin{abstract}

2: Neural networks has recently attracted much interest as useful representations

3: of quantum many body ground states, which might help address the infamous

4: sign problem. Most attention was directed at their representability

5: properties, while possible limitations on finding the desired optimal

6: state have not been suitably explored. By leveraging well-established

7: results applicable in the context of infinite width, specifically

8: regarding the renowned neural tangent kernel and conjugate kernel,

9: a comprehensive analysis of the convergence and initialization characteristics

10: of the method is conducted. We reveal the dependence of these characteristics

11: on the interplay among these kernels, the Hamiltonian, and the basis

12: used for its representation. We introduce and motivate novel performance

13: metrics and explore the condition for their optimization. By leveraging

14: these findings, we elucidate a substantial dependence of the effectiveness

15: of this approach on the selected basis, demonstrating that so-called

16: ``stoquastic\textquotedblright{} Hamiltonians are more amenable to

17: solution through neural networks than those suffering from a sign

18: problem.

19: \end{abstract}

20: