1: \begin{abstract}

2: The exploration of neural network quantum states has become widespread in the search for ground states in complex quantum many-body systems. However, achieving high precision remains challenging due to the intricate sign structures and the exponential growth of the Hilbert space.

3: In this work, we propose a neural network state method that is confined to a significantly smaller symmetric subspace, by the full space summation and Markov chain Metropolis sampling approaches.

4: Using symmetries and group theory, the proposed method significantly reduces the number of parameters in neural network states and achieves better accuracy and convergence properties.

5: % Our framework requires fewer neurons and samples to attain satisfactory convergence, while also possessing the capability to simulate excitations.

6: We validate our method using the frustrated spin-$\frac{1}{2}$ $J_1$-$J_2$ antiferromagnetic Heisenberg chain and compare its performance against NetKet, the standard library of neural network states.

7: The results indicate that our symmetrized neural network states achieve a substantial improvement over the conventional neural network states method, reducing energy errors by two orders of magnitude.

8: % We further explore the performance of two network structures: a complex network with complex parameters, and two separate networks with real parameters for a complex wavefunction. We observe no discernible difference in accuracy between these structures.

9: We also compare degenerate eigenstates with different quantum numbers, highlighting the advantage of operating within a smaller variational space.

10: \end{abstract}

11: