Computing excited states with isometric tensor networks in two-dimensions Journal: Arxiv
Alec Dektor, Runze Chi, Roel Van Beeumen, Chao Yang · 2025
Read Computing excited states with isometric tensor networks in two-dimensions Journal: Arxiv by Alec Dektor, Runze Chi, Roel Van Beeumen, Chao Yang (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-2510-20063.txt.
Introduction
Alec Dektor and three co-authors introduce a subspace iteration method for finding low-lying excited states of quantum many-body systems on two-dimensional lattices. Their approach rests on a new block isometric projected entangled pair state form, which extends the block matrix product state idea from one-dimensional chains to two dimensions.
The new form allows exact block orthogonalization, controlled local truncation through singular value decompositions, and efficient measurement of observables. The authors test the inexact iteration on the two-dimensional transverse-field Ising and Heisenberg models and compare with existing pair state methods. Their results show that block isometric tensor networks offer a scalable way to study excitations beyond one dimension.
Catalog record
| CREATOR | Alec Dektor, Runze Chi, Roel Van Beeumen, Chao Yang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-20063
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Computing excited states with isometric tensor networks in two-dimensions Journal: Arxiv Alec Dektor, Runze Chi, Roel Van Beeumen, Chao Yang Abstract We present a new subspace iteration method for computing low-lying eigenpairs (excited states) of high-dimensional quantum many-body Hamiltonians with nearest neighbor interactions on two-dimensional lattices. The method is based on a new block isometric projected entangled pair state (block-isoPEPS) ansatz that generalizes the block matrix product state (MPS) framework, widely used for Hamiltonians defined on one-dimensional chains, to two-dimensions. The proposed block-isoPEPS ansatz offers several attractive features for PEPS-based algorithms, including exact block orthogonalization, controlled local truncation via singular value decompositions, and efficient evaluation of observables. We demonstrate the proposed inexact subspace iteration for block-isoPEPS by computing excitations of the two-dimensional transverse-field Ising and Heisenberg models and compare our results with existing PEPS methods. Our results demonstrate that block isometric tensor networks provide a scalable framework for studying excitations in quantum many-body systems beyond one dimension. 1 Introduction We consider the problem of computing a few of the algebraically smallest eigenpairs, i.e., ground and excited states, of Hamiltonians with pairwise interactions H=\sum_{(j_{1},j_{2})\in S}H_{j_{1}}H_{j_{2}},\qquad H_{j}=I^{\otimes(j-1)}\otimes M_{j}\otimes I^{\otimes(L-j)} (1) where S\subset\{1,\ldots,L\}\times\{1,\ldots,L\} specifies the interacting pairs on L local d -dimensional vector spaces (referred to as sites), each M_{j}\in\mathbb{C}^{d\times d} is a Hermitian matrix, and I denotes the d\times d identity matrix. Hamiltonians of the form ( 1 ) appear in model quantum many-body systems such as spin chains and lattices with nearest-neighbor interactions. The number of entries in the matrix ( 1 ) grows exponentially fast with the parameter L making classical methods for finding eigenpairs too expensive for problems of interest. Tensor networks are a powerful framework for computing eigenvectors of ( 1 ) using a reasonable number of degrees of freedom determined by an auxiliary parameter called the bond-dimension. For instance, when ( 1 ) represents a Hamiltonian describing a quantum spin chain with nearest-neighbor interactions, i.e., S contains only pairs of the form (j,j+1) , and has separated eigenvalues, the eigenvectors corresponding to the algebraically smallest eigenvalues admit accurate matrix product state (MPS) approximations with small bond-dimension [ 1 ] . Several MPS algorithms have been developed to compute eigenvectors [ 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 ] of ( 1 ). These algorithms enable practitioners to routinely compute accurate eigenvector approximations of (quasi) one-dimensional quantum many-body Hamiltonians ( 1 ) with L>100 . However, due to the one-dimensional connectivity of the MPS tensor network, these methods are often not suitable for quantum many-body systems with nearest-neighbor interactions in two-dimensional geometries such as rectangular lattices. In this case the set S of interacting pairs contains (j_{1},j_{2}) with j_{1}\ll j_{2} which induces long-range interactions in the MPS, leading to prohibitively large bond-dimensions in MPS eigenvector approximations. These limitations have motivated the development of algorithms based on tensor networks with additional connectivity designed to efficiently capture the entanglement structure of quantum many-body systems defined on two- and higher-dimensional geometries. Projected entangled pair states (PEPS) are a class of tensor networks with connectivity designed to efficiently capture entanglement in two-dimensional geometries [ 11 ] . Their connectivity makes PEPS significantly more demanding than MPS in terms of algorithmic design and computational resources. For example, exact contraction (required for computing norms, inner products, and observables) of PEPS networks has computational cost scaling exponentially with the 2D system size, whereas the corresponding contraction for MPS scales linearly with 1D system size (and polynomially with the bond-dimension). To mitigate the exponential scaling of PEPS contraction, several approximate schemes have been developed that trade accuracy for reduced cost [ 12 ] . While these methods avoid exponential scaling, they remain computationally expensive, often scaling as high powers of the PEPS bond-dimension \chi , e.g., \mathcal{O}(\chi^{10}) . An alternative strategy for obtaining computationally efficient PEPS algorithms is to impose isometric constraints on the tensors in the PEPS network, so that contracting large regions of the network is trivial. Imposing sufficient isometric constraints yields an isometric PEPS (isoPEPS) [ 13 ] with an orthogonality center similar to canonical forms of MPS (which are central to many MPS algorithms). However unlike MPS, the orthogonality center cannot be moved around the PEPS without either increasing bond-dimension or introducing error. As such, the set of isoPEPS with fixed bond dimension \chi is a strict subset of all PEPS with bond-dimension \chi . While the variational power of isoPEPS is still not fully understood, recent work has identified classes of physical states isoPEPS can represent exactly or efficiently [ 14 ] and has established complexity-theoretic limits on their contraction and sampling [ 15 ] . The computational cost of many isoPEPS algorithms scale as \mathcal{O}(\chi^{7}) [ 16 ] , offering a favorable alternative to the \mathcal{O}(\chi^{10}) scaling of many generic PEPS algorithms. Several isoPEPS algorithms have been recently developed for computing a single algebraically smallest eigenpair, i.e., the ground-state, and real-time evolution of nearest neighbor two-dimensional systems [ 13 , 17 , 18 , 19 , 20 , 21 ] . However, methods for computing several of the algebraically smallest eigenpairs, i.e., excited states, with isoPEPS remain unexplored. Figure 1: Tensor network diagrams for representing three tensors. The block MPS network, shown here with 3 sites, provides a compact representation of states with limited entanglement in 1D geometries. In this paper we propose the block (isometric) PEPS ansatz, shown here with 9 sites, as a compact representation of several states with limited entanglement in 2D geometries. In this paper, we introduce a new block-isoPEPS tensor network ansatz shown in fig. 1 (right) for computing several eigenpairs of nearest neighbor two-dimensional quantum many-body systems. Our approach generalizes the state averaged MPS (or block tensor train) framework [ 22 , 23 , 24 , 25 ] fig. 1 (left), a well-established tensor network ansatz for computing several eigenpairs in one-dimensional systems, to two dimensions. The block-isoPEPS ansatz offers several desirable properties for handling nearest-neighbor Hamiltonians in two-dimensions. It retains the properties of isoPEPS, providing direct control over truncation errors through singular values and enabling efficient computation of local observables. Additionally, the isometric constraints enable exact and efficient orthogonalization of a block of tensor network states. To move the orthogonality center and block index around the tensor network we present the block Moses Move, an extension of the Moses Move introduced in [ 13 ] . Using the block Moses Move we propose a new isoPEPS subspace iteration method for computing several eigenpairs with computational complexity scaling as \mathcal{O}(\chi^{7}) . Our approach differs from existing PEPS methods [ 26 , 27 , 28 , 29 ] , which do not impose orthogonality constraints during the optimization. One common strategy is to construct the excitation ansatz in the tangent space of an optimized ground state and then solve the effective Hamiltonian problem within this non-orthogonal subspace to extract excitation energies. Due to the restricted form of the ansatz, such approaches can only capture local excitations above the ground state. Alternatively, one can avoid these special excitation ansatz by directly optimizing the functional defined by the inverse norm matrix multiplied by the Hamiltonian matrix [ 29 ] within the variational space. We benchmark the proposed