Triad Representation for the Anisotropic Tensor Renormalization Group in Four Dimensions

Yuto Sugimoto, Shoichi Sasaki · 2025

Read Triad Representation for the Anisotropic Tensor Renormalization Group in Four Dimensions by Yuto Sugimoto, Shoichi Sasaki (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-2507-21909.txt.

Introduction

Yuto Sugimoto and Shoichi Sasaki published this physics computing study in 2025. Tensor renormalization group methods simulate lattice systems by repeatedly compressing huge networks of numbers. They promise a way around the sign problem that blocks standard simulations of dense quark matter. An anisotropic variant speeds up four-dimensional calculations, but costs stay heavy. The authors propose a further variant built on a triad network representation.

Their method, triad-ATRG, grows more slowly in cost as the network's bond size increases. It does so while losing little accuracy in free energy and other physical quantities. The team also wrote parallel versions of both the older and new methods for multiple graphics processors. Those versions ran far faster than processor-based code on a four-dimensional test system. The work aims to make large lattice studies practical for more research groups.

Catalog record

CREATORYuto Sugimoto, Shoichi Sasaki
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-21909

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Triad Representation for the Anisotropic Tensor Renormalization Group in Four Dimensions Yuto Sugimoto, Shoichi Sasaki Abstract The development of tensor renormalization group (TRG) algorithm in higher dimensions is an important and urgent task, as the TRG is expected to provide a way to overcome the sign problem in lattice QCD calculations at finite density. One possible approach that enables faster computations in four-dimensional lattice theories is the anisotropic tensor renormalization group (ATRG). However, the computational cost remains substantial and requires significant computational resources. In this paper, we propose a novel algorithm, called the Triad-ATRG, which is based on the ATRG and other improved TRG variants with triad network representation. This method achieves lower scaling with respect to the bond dimension, while minimizing the loss of accuracy in the free energy and other physical quantities. We also present parallel implementations of both the ATRG and Triad-ATRG on multiple GPUs, which significantly improve performance compared to CPU-based calculations for the four-dimensional system. 1 Introduction The tensor renormalization group (TRG) levinTensorRenormalizationGroup2007 is a powerful numerical method for computing the partition function of quantum many-body systems. The core idea behind the TRG is to apply a sequence of low-rank approximations using singular value decomposition (SVD). The accuracy of the approximation is controlled by the bond dimension D , which determines the number of singular values retained. This procedure reduces computational complexity while preserving the essential physical information of the system. One of the most important advantages of the TRG framework is that its computational cost is proportional to O(\log{V}) with respect to the system volume V , making it easy to take the thermodynamic limit. Furthermore, the absence of the sign problem is expected to achieve applications in a wide range of theories, including finite-density QCD. At present, many applications of the TRG method and its variants have been reported Shimizu:2012wfa ; Shimizu:2012zza ; Yu:2013sbi ; shimizuGrassmannTensorRenormalization2014 ; Shimizu:2014fsa ; Unmuth-Yockey:2014afa ; Zou:2014rha ; wangPhaseTransitionsFerromagnetic2014 ; takedaGrassmannTensorRenormalization2015 ; Yang:2015rra ; nakamotoComputationCorrelationFunctions2016 ; zhaoTensorNetworkAlgorithm2016 ; kawauchiTensorRenormalizationGroup2016 ; Bal:2017mht ; Sakai:2017jwp ; sakaiApplicationTensorNetwork2018 ; Kadoh:2018hqq ; Shimizu:2017onf ; Kadoh:2018tis ; Kuramashi:2018mmi ; yoshimuraCalculationFermionicGreen2018 ; Akiyama:2019chk ; Bazavov:2019qih ; akiyamaPhaseTransitionFourdimensional2019 ; Kadoh:2019ube ; akiyamaTensorRenormalizationGroup2020 ; Delcamp:2020hzo ; Iino:2019rxt ; PhysRevE.101.012124 ; akiyamaTensorRenormalizationGroup2021 ; Akiyama:2020soe ; Akiyama:2021zhf ; Bloch:2021mjw ; Fukuma:2021cni ; hirasawaTensorRenormalizationGroup2021a ; Milde:2021vln ; Nakayama:2021iyp ; kuwaharaTensorRenormalizationGroup2022 ; Akiyama:2021glo ; Akiyama:2022pse ; Akiyama:2022eip ; Jha:2022pgy ; Luo:2022eje ; Akiyama:2023hvt ; Akiyama:2024qgv ; Samlodia:2024kyi ; akiyamaTensorRenormalizationGroup2024a ; Akiyama:2024qer ; Pai:2024tip ; kannoGrassmannTensorRenormalization2024 ; luoCriticalEndpointsThreedimensional2025a ; tsengBondWeightedTensor2025 . From an algorithmic perspective, the original TRG was designed for two-dimensional classical systems, and its extension to higher dimensions is not straightforward. The higher-order tensor renormalization group (HOTRG) xieCoarsegrainingRenormalizationHigherorder2012 was developed to extend the TRG to higher dimensions by implementing coarse-graining transformations along a selected axis at each step. Recently, the HOTRG has been successfully applied to four-dimensional systems akiyamaPhaseTransitionFourdimensional2019 ; Akiyama:2019chk ; Milde:2021vln . These achievements are of particular importance, as our target theory, QCD, is defined in (3\!+\!1) -dimensional spacetime. However, due to its relatively simple algorithm structure, the HOTRG suffers from high computational cost in higher dimensions, since its computational cost is proportional to O(D^{4d-1}) in d dimensions. As a result, increasing the bond dimension in three and four-dimensional systems poses a significant computational challenge. In order to address this situation, several variants of the HOTRG have been proposed to reduce the computational cost in higher dimensions. Among these methods, the anisotropic TRG (ATRG) adachiAnisotropicTensorRenormalization2020 has been demonstrated to achieve a substantial cost reduction, scaling as O(D^{2d+1}) . This lower cost enables the study of four-dimensional systems with substantially larger bond dimensions, and in fact, the ATRG has been successfully applied to four-dimensional systems with/without the sign problem Akiyama:2019chk ; akiyamaTensorRenormalizationGroup2020 ; Akiyama:2020soe ; Akiyama:2021zhf ; Akiyama:2022eip ; Akiyama:2023hvt . In addition to the ATRG, there exist other algorithms that have been developed for the purpose of reducing computational costs kadohRenormalizationGroupTriad2019 ; Milde:2021vln ; nakayamaApplicationProjectiveTruncation2024 ; nakayamaRandomizedHigherorderTensor2023 . It is noteworthy that the Triad-TRG kadohRenormalizationGroupTriad2019 , the Triad-MDTRG nakayamaApplicationProjectiveTruncation2024 ; nakayamaRandomizedHigherorderTensor2023 , achieve a lower scaling than the HOTRG by decomposing all tensors into three-legs. The Triad-TRG and a series of MDTRG variants also adopt approximate contraction schemes based on the randomized singular value decomposition (RSVD) doi:10.1137/090771806 , which was originally introduced in the two-dimensional TRG moritaTensorRenormalizationGroup2018 . In particular, the MDTRG incorporates several strategies so as to suppress additional systematic errors introduced by the tensor decomposition. As a result, the MDTRG achieves an accuracy comparable to that of the HOTRG when simulating the three-dimensional Ising model. However, the application of the MDTRG to four-dimensional systems remains unreported. In this study, we propose a novel approach, the Triad-ATRG, which combines the ATRG and the triad representation kadohRenormalizationGroupTriad2019 ; nakayamaRandomizedHigherorderTensor2023 to enable practical computations with larger bond dimensions for four-dimensional systems. We also present parallel implementations of both the ATRG and Triad-ATRG on multiple GPUs, which significantly improve performance compared to CPU-based calculations. This paper is organized as follows. In Sec. 2 , we review the basics of both the HOTRG and ATRG methods, as well as the triad representation. Then, we introduce the Triad-ATRG method, which is more efficient than the original ATRG method for the case of the four-dimensional system. The idea of combining the Triad-ATRG and the MDTRG technique to achieve a higher accuracy calculation is also described. We also discuss the GPU parallelization for the ATRG and Triad-ATRG methods. In Sec. 3 , we present the numerical results of applying the Triad-ATRG method to the four-dimensional Ising model. We then demonstrate that the Triad-ATRG method achieves a comparable accuracy to the ATRG method, but with a significantly reduced computational cost. In Sec. 4 , we summarize our work and discuss future prospects. 2 Triad-ATRG method 2.1 Overview of the ATRG, MDTRG, and triad representation In this section, we briefly review the anisotropic tensor renormalization group (ATRG) method adachiAnisotropicTensorRenormalization2020 . The ATRG method is based on the higher-order tensor renormalization group (HOTRG) method xieCoarsegrainingRenormalizationHigherorder2012 , with some additional approximation techniques. Let us consider a target tensor network that is connected on a d -dimensional square lattice on the volume V=L^{d} . We assume that this tensor network is homogeneous, implying that all tensors located at each lattice site are equivalent. This condition is satisfied when we consider a partition function of a quantum many-body system, which has a local translational invariance. In such case, the partition function of the system is expressed as Z=\mathrm{tTr}\prod_{n}T_{i_{1}(n)i_{2}(n)\dots i_{d}(n)j_{1}(n)j_{2}(n)\cdots j_{d}(n)}. (1) where n denotes a lattice site. In the above equation, the symbol T represents a local tensor; i_{\mu}(n) corresponds to the bond connecting site n to n+\hat{\mu} , while j_{\mu}(n) is a shorthand for the opposite bond connecting site n to n-\hat{\mu} , i.e. , j_{\mu}(n)\equiv i_{\mu}(n-\hat{\mu}) . The tensor trace \mathrm{tTr} indicates the summation over all internal indices under the constraint i_{\mu}(n)=j_{\mu}(n+L\hat{\mu}) for \mu=1,2,\dots,d under periodic boundary conditions. A fundamental technique common to all TRG methods is the approximation of the target tensor network by updating the local tensors through SVD. In the HOTRG case, the approximation is performed along a selected axis by contracting the unit-cell tensor, denoted by \Gamma , which is composed of two adjacent local tensors. The unit-cell tensor along \hat{1} -axis is

Read the complete text