Alternating and Gaussian fermionic Isometric Tensor Network States
Yantao Wu, Zhehao Dai, Sajant Anand, Sheng-Hsuan Lin, Qi Yang, Lei Wang, Frank Pollmann, Michael P. Zaletel · 2025
Read Alternating and Gaussian fermionic Isometric Tensor Network States by Yantao Wu, Zhehao Dai, Sajant Anand, Sheng-Hsuan Lin, Qi Yang, Lei Wang, Frank Pollmann, Michael P. Zaletel (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-10695.txt.
Introduction
Yantao Wu and co-authors improve a class of tensor network states used to simulate two-dimensional quantum matter. Their 2025 paper works with isometric tensor networks, which are efficient but carry a restriction whose cost was unclear. The team's variant flips the direction of the network's arrows from column to column, hence the name alternating.
They also build a Gaussian fermionic version of the tool, which allows exact checks on free-particle models. Several lines of evidence favor the alternating design. Entanglement flows more efficiently along the arrows, and the matching quantum circuit is deeper and more expressive. Ground state energies and scaling beat the original design across test models, including an interacting magnet model. The authors argue the simple flip is a real upgrade, not a cosmetic one.
Catalog record
| CREATOR | Yantao Wu, Zhehao Dai, Sajant Anand, Sheng-Hsuan Lin, Qi Yang, Lei Wang, Frank Pollmann, Michael P. Zaletel |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 162 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-10695
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Alternating and Gaussian fermionic Isometric Tensor Network States Yantao Wu, Zhehao Dai, Sajant Anand, Sheng-Hsuan Lin, Qi Yang, Lei Wang, Frank Pollmann, Michael P. Zaletel Abstract Isometric tensor networks in two dimensions enable efficient and accurate study of quantum many-body states, yet the effect of the isometric restriction on the represented quantum states is not fully understood. We address this question in two main contributions. First, we introduce an improved variant of isometric tensor network states (isoTNS) in two dimensions, where the isometric arrows on the columns of the network alternate between pointing upward and downward, hence the name alternating isometric tensor network states . Second, we introduce a numerical tool – isometric Gaussian fermionic TNS (isoGfTNS) – that incorporates isometric constraints into the framework of Gaussian fermionic tensor network states. We demonstrate in numerous ways that alternating isoTNS represent many-body ground states of two-dimensional quantum systems significantly better than the original isoTNS. First, we show that the entanglement in an isoTNS is mediated along the isometric arrows and that alternating isoTNS mediate entanglement more efficiently than conventional isoTNS. Second, alternating isoTNS correspond to a deeper, thus more representative, sequential circuit construction of depth 𝒪 ( L x ⋅ L y ) \mathcal{O}(L_{x}\cdot L_{y}) compared to the original isoTNS of depth 𝒪 ( L x + L y ) \mathcal{O}(L_{x}+L_{y}) . Third, using the Gaussian framework and gradient-based energy minimization, we provide numerical evidences of better bond-dimension scaling and variational energy of alternating isoGfTNS for ground states of various free fermionic models, including the Fermi surface and the band insulator. Finally, benchmarking on the transverse field Ising model, we demonstrate that alternating isoTNS provides substantially improved performance and stability relative to the original isoTNS for ground state search algorithm in interacting systems. I Introduction Recently, two-dimensional (2D) tensor network states (TNS) [ 1 , 2 , 3 , 4 ] have become a valuable tool for studying strongly interacting quantum phases of matter [ 5 , 6 , 7 , 8 ] . Despite the usefulness of 2D TNS, several outstanding research challenges remain that limit their applicability, namely reducing the complexity of contracting the tensor network to compute physical observables and improving the convergence speed in optimizing the variational tensors. For one-dimensional (1D) TNS, i.e., matrix product states (MPS), the isometric form is the key to the stability and efficiency of ground state MPS algorithms such as the density matrix renormalization group (DMRG) [ 9 , 10 ] and time evolution MPS algorithms based on the time-evolving block-decimation [ 11 ] or the time-dependent variational principle [ 12 ] . Thus, it is natural to consider the generalization of isometric form to 2D. The most straightforward version was proposed in [ 13 ] shown in Fig. 1 , known as isometric tensor network states (isoTNS). Similar versions are also proposed in [ 14 , 15 ] . In an isoTNS, each bond in the 2D network is assigned an arrow, and each tensor, when viewed as a map A from the outgoing legs to the incoming legs, is required to be an isometric linear map, A^{\dagger}A=\mathbbm{I} . In particular, all physical legs are incoming. This allows both the efficient exact contraction of the norm of the 2D TNS and evaluation of expectation values of local operators on the orthogonal hypersurface [ 13 ] . In 1D, every MPS can be exactly brought into the isometric form with the same maximal bond dimension through gauge transformations. In other words, the manifolds of generic MPS of bond dimension \chi and MPS in isometric form of bond dimension \chi are the same. In 2D, the isoTNS is a more restricted ansatz than general TNS with the same bond dimension, and the generic and isometric manifolds of dimension \chi are believed to be inequivalent. It is an open question on how the representation power of TNS is affected by imposing the isometric constraint. Quantum computation provides a separate motivation for studying the representability of isoTNS. Any 1D MPS of length L can be prepared on a quantum computer by a local sequential circuit of depth L [ 16 ] . This is not true for 2D TNS unless the network has an isometric form. Thus, it is prudent to understand the impact isometric constraints can have on the resulting sequential circuits, especially as the representability of sequential quantum circuits has recently attracted significant interests [ 17 , 18 , 19 ] . Moreover, there actually exists freedom in choosing the isometric direction of the tensors while maintaining the algorithmic advantages and sequential-circuit representations. For examples, see Fig. 1 and Fig. 1 where we show the networks of primary interest in this work. This network design choice has not been explored in previous isoTNS works [ 13 , 20 , 21 , 22 , 23 ] . However, this turns out not to be a gauge degree of freedom and can significantly affect the representability of the TNS ansatz and sequential circuit. To understand the effect of different isometric forms, we provide insight from two perspectives – entanglement structure and the corresponding sequential quantum circuits – and support this with numerical evidence of representing ground states of both non-interacting fermionic and interacting spin systems. In this paper, we show that in an isoTNS, entanglement is, in a sense made precise later, mediated along the isometric arrows. An isoTNS is capable of representing almost any short-ranged correlated state, but it can be inefficient in capturing entanglement perpendicular to its isometric directions. This severely limits the application of isoTNS to some of the most important physical systems like Fermi liquids. Here, we propose a novel isometric ansatz, alternating isoTNS (alt-isoTNS), shown in Fig. 1 , which greatly improves the representability of isoTNS while keeping all algorithmic advantages over general TNS. To differentiate from alt-isoTNS, hereafter, we call the original isoTNS [ 13 ] , shown in Fig. 1 , the uniform isoTNS (uni-isoTNS). \captionlistentry \captionlistentry \captionlistentry Figure 1 : A 5\times 5 uniform (a) and alternating (b) isoTNS with open boundary condition. The orthogonality center is the single, upper right tensor with all incoming legs. The orthogonality hypersurface is the set of red tensors. Note for the alt-isoTNS that the isometry arrow direction of the columns alternates, while for uni-isoTNS all the column arrows point up. (c) Contour plot of the k -space occupation number of the uni- (left) and alt-isoTNS (right) at bond dimension \chi=16 , representing the Fermi surface ground state of Eq. 29 . Another independent contribution of this paper is to introduce a tool to efficiently study the representability of isoTNS — isometric Gaussian fermionic tensor network states (isoGfTNS). IsoGfTNS is a variational ansatz resulting from imposing isometric constraint on the Gaussian fermionic tensor network states (GfTNS). The general problem of 2D TNS representability can be very difficult to study due to high computational complexity in optimizing a generic, interacting 2D TNS. However, analysis can be performed by imposing the Gaussian constraint on the variational ansatz, which simplifies the calculation, and by choosing free fermion systems as a benchmark problem, where the exact solution exists [ 24 ] . Since the key quantity restricting the TNS representability is the entanglement