Fig. 9: The effect of U(1) and spatial symmetries.
From: Neural network approach to quasiparticle dispersions in doped antiferromagnets
Relative error for t − J (dark blue) and t − Jz (light blue) models on a 4 × 4 square lattice with one hole, t/Jz = 3 and periodic boundaries, averaged over 100 training steps with 200 samples per step (error bars denote the respective standard deviation). We use a recurrent neural networks (RNN) with implemented \(U(1)=U{(1)}_{\hat{N}}\times U{(1)}_{{\hat{S}}_{z}}\) symmetry, U(1) and C4 symmetry, implemented via the cost function and the RNN ansatz, with a hidden dimension of hd = 70. For the t − Jz model, we provide the relative energy errors as numbers in light blue.
