Variationally optimizing infinite projected entangled-pair states at large bond dimensions: A split corner transfer matrix renormalization group approach
Jan Naumann, Erik L. Weerda, Jens Eisert, Matteo Rizzi, Philipp Schmoll · 2025
Read Variationally optimizing infinite projected entangled-pair states at large bond dimensions: A split corner transfer matrix renormalization group approach by Jan Naumann, Erik L. Weerda, Jens Eisert, Matteo Rizzi, Philipp Schmoll (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-2502-10298.txt.
Introduction
Simulating quantum systems on a two-dimensional lattice is powerful but costly, because the math objects involved grow very fast. Jan Naumann, Erik Lennart Weerda, Jens Eisert, Matteo Rizzi, and Philipp Schmoll introduced a faster algorithm in 2025 for a leading method called PEPS. The bottleneck is a step that merges two layers of the lattice, which eats time and memory at large sizes.
Their split approach keeps the two layers separate and adds new environment tensors to hold the needed information. Benchmarks on quantum lattice models showed large speedups in finding the lowest-energy state, with accuracy preserved. The authors offer the method for large-scale simulations of quantum materials.
Catalog record
| CREATOR | Jan Naumann, Erik L. Weerda, Jens Eisert, Matteo Rizzi, Philipp Schmoll |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 284 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-10298
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Variationally optimizing infinite projected entangled-pair states at large bond dimensions: A split corner transfer matrix renormalization group approach Jan Naumann, Erik L. Weerda, Jens Eisert, Matteo Rizzi, Philipp Schmoll Abstract Projected entangled-pair states (PEPS) have become a powerful tool for studying quantum many-body systems in the condensed matter and quantum materials context, particularly with advances in variational energy optimization methods. A key challenge within this framework is the computational cost associated with the contraction of the two-dimensional lattice, crucial for calculating state vector norms and expectation values. The conventional approach, using the corner transfer matrix renormalization group (CTMRG), involves combining two tensor network layers, resulting in significant time and memory demands. In this work, we introduce an alternative “split-CTMRG” algorithm, which maintains separate PEPS layers and leverages new environment tensors, reducing computational complexity while preserving accuracy. Benchmarks on quantum lattice models demonstrate substantial speedups for variational energy optimization, rendering this method valuable for large-scale PEPS simulations. I Introduction One of the main challenges in condensed matter physics is developing tools to predict the properties of interacting quantum matter classically. A versatile and important family of classical simulation methods is provided by tensor networks [ 1 , 2 , 3 , 4 ] – a set of techniques that has originated from the early development of the density matrix renormalization group approach [ 5 , 6 ] and that has matured into a large body of powerful methods. Of those methods, two-dimensional tensor networks in the form of infinite projected entangled-pair states (PEPS) [ 7 ] have substantially matured in recent years due to the development of variational energy optimization techniques, paving the way for them to become one of the most important tools in the study of quantum many-body systems and two-dimensional quantum materials. While the gradient of the energy expectation value used in the optimization was constructed manually by summing up a large number of relevant tensor network diagrams in the pioneering works [ 8 , 9 ] , the more modern approach performs this task in an automated way – typically by utilizing automatic differentiation (AD) [ 10 , 11 , 12 ] . These developments have boosted both the accuracy as well as the applicability of PEPS in the study of intricate frustrated quantum magnetism [ 13 , 14 , 15 , 16 , 17 , 18 ] , topologically ordered phases [ 19 , 20 , 21 ] , or itinerant electron systems [ 22 ] . However, it has already been pointed out that the high computational cost of the variational optimization can be a limitation for variational PEPS studies of particularly complex models [ 23 ] . One of the key steps in any numerical PEPS simulation is the contraction of the two-dimensional tensor network, required to compute the norm of the state vector as well as expectation values. Various methods, such as boundary MPS [ 24 , 25 , 26 ] and tensor coarse-graining techniques [ 27 , 28 , 29 , 30 ] have been developed to address this challenge. However, the corner transfer matrix renormalization group (CTMRG) algorithm [ 31 , 32 , 33 , 34 , 35 , 36 , 37 ] has emerged as the most widely used approach, particularly when combined with automatic differentiation and energy optimization. Despite that, it is precisely this CTMRG part that is still the primary bottleneck in the calculations, both in time and memory. This is because (i) it implements a computationally expensive iterative renormalization procedure to find the fixed-point tensors representing the (approximate) contraction of the infinite lattice, and (ii) it has to be performed in each evaluation of the energy, which happens at least once – but likely multiple times – during every optimization step. In the original formulation of CTMRG, the norm of the PEPS state vector is computed by combining the two TN layers for dual vector \bra{\psi} and state vectors \ket{\psi} into an effective double-layer network. While this results in an empirically stable and robust algorithm, it leads to a high computational cost due to the effective squaring of the PEPS bond dimension. Here, we instead propose an alternative split-CTMRG scheme that gains an advantage in terms of computational complexity by keeping the two layers of the PEPS separate, and defining individual fixed-point environment tensors for them. In the past, several proposals were made to improve the efficiency of the CTMRG scheme. Refs. [ 38 , 39 ] have proposed a substantially more computationally efficient contraction algorithm by flattening the double-layer tensor network for the norm into a single-layer structure, at the cost of expanding the elementary unit cell. More recently, Ref. [ 40 ] presented a variant of the CTMRG algorithm, in which the PEPS tensors and their conjugates are sequentially absorbed into conventional CTMRG environment tensors. Both of these approaches were proposed and tested only in the context of ground state search with imaginary-time evolution, i.e., the simple or full update [ 24 , 41 ] . However, when moving to variational PEPS optimization we are confronted with specific requirements. This is because the variational PEPS framework not only has the most to gain from reducing the computational cost of computing environment tensors but also imposes the strictest accuracy requirements on expectation value calculations. These strict requirements are due to the fact that variational PEPS optimization relies on accurate calculations of the energy gradient. Without an accurate energy approximation, the optimization will be ineffective, especially near the targeted minimum. In order to achieve the optimal accuracy, the projectors used in the CTMRG scheme are of crucial importance. In this work, we therefore employ the philosophy of the best established projectors from the conventional CTMRG [ 34 , 35 , 36 ] to define the new projectors for the split-CTMRG algorithm. We show that this yields very precise expectation values, while substantially reducing the cost of the algorithm compared to the conventional CTMRG, allowing us to push the limits and applicability of the infinite PEPS ansatz. This work is structured as follows: In Sec. II , we briefly outline the conventional CTMRG algorithm and introduce the new split-CTMRG version, highlighting important differences and sources of efficiency improvements. In Sec. III , we then provide numerical benchmarks on a quantum lattice model, demonstrating state-of-the-art accuracy and showcasing improved results due to significantly reduced computational costs. Finally, we conclude this work in Sec. IV . II Method Before presenting our alternative version of the CTMRG algorithm in Sec. II.2 , we briefly revise the context of infinite projected entangled-pair states and the general task of computing their norms in Sec. II.1 . II.1 Infinite PEPS and the CTMRG Infinite projected entangled-pair states (PEPS) are an ansatz class of variational many-body state vectors of quantum lattice models \ket{\psi}=\sum_{\{s_{i}\}}C^{\{s_{i}\}}\ket{\{s_{i}\}}, (1) in which the the coefficient tensor C^{\{s_{i}\}} is presumed to have a substructure of a tensor network with two dimensional planar connectivity. On a square lattice, this tensor network consists of rank-five tensors, which are periodically repeated \displaystyle\begin{split}\includegraphics[scale={1.0}]{figures/infinitePEPS_1.pdf}\end{split}. (2) While we show here a fully translational invariant network, it is also possible to define a non-trivial unit cell of tensors that is periodically repeated. The open legs of the tensors in