Generalized Lanczos method for systematic optimization of tensor network states*

Project supported by the National Natural Science Foundation of China (Grant Nos. 11190024 and 11474331).

Huang Rui-Zhen1, 2, Liao Hai-Jun1, Liu Zhi-Yuan2, 3, Xie Hai-Dong1, 2, Xie Zhi-Yuan4, Zhao Hui-Hai5, 6, 7, Chen Jing1, 2, Xiang Tao1, 2, 8, †
       

(color online) Entanglement entropy for the ground state of the Heisenberg spin chain. (a) Converged entanglement entropy, S(∞) = S(n → ∞), as a function of L with D = 8 (squares) and 12 (circles), compared with the corresponding results obtained using a single MPS, S(n = 0) (up and down triangles), and those obtained with DMRG by keeping 1024 states (solid line). The horizontal axis is logarithmic. (b) Difference between S(n) and S(∞) as a function of n, which shows the convergence speed of S(n), for D = 8. The inset shows the size-dependence of the restarted Lanczos number n0 at which the entanglement entropy arrives within 1% of its converged value, i.e., 1 − S(n0)/S(∞) ≤ 1%.