Fig. 10: Number of parameters for the exact wave function of a 4 × 4 system compared to the recurrent neural network (RNN) ansatz.
From: Neural network approach to quasiparticle dispersions in doped antiferromagnets
a We compare the number of parameters of the exact wave function using \(U{(1)}_{\hat{N}}\times U{(1)}_{{\hat{S}}_{z}}\) symmetry for 0≤Nh≤4 holes (blue) to the Hilbert space dimension 316 that we want to learn with the RNN ansatz. The effective number of parameters of the RNN ansatz, i.e. the number of parameters that is kept when enforcing the U(1) symmetry, with hidden dimension 30≤hd≤100 is denoted by the gray markers. b Hilbert space dimension for a local dimension of 2 (Heisenberg model), 3 (t − J model) and 4 (Fermi-Hubbard model).
