Speaker
Description
Isometric tensor networks provide globally optimal low-rank approximations without environment computation. However, this property is guaranteed only at the orthogonality center, making node-level parallelization difficult. In contrast, node-level parallelization has been developed based on Vidal canonical form [1,2], though the canonical conditions and the truncation optimality are gradually lost during variational optimization and time evolution. We propose a two-layer network architecture which combines the advantages of both representations. One layer strictly maintains an isometric structure, while the other use Vidal canonical form. This hybrid design aims to balance approximation accuracy and computational parallelism.
[1] G. Vidal, Phys. Rev. Lett. 91, 147902 (2003).
[2] R. -Y. Sun, T. Shirakawa, and S. Yunoki, Phys. Rev. B 110, 085149 (2024).