REVIEW 3 major objections 4 minor 18 references
Initial tensor construction for the tensor renormalization group
T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read This paper claims that any partition function written as a product of Boltzmann factors can be converted exactly—without expansions or decompositions—into a locally connected tensor network for the tensor renormalization group, and that a…
desk verdict An exact and clean delta-shift recipe for TRG initial tensors, with an overreaching claim about removing initial-tensor dependence on thin evidence. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is Algorithm 1: given a Boltzmann factor $K$ depending on several site variables, insert a delta function $\delta_{\sigma,a}$ to duplicate one of its indices, then shift that delta by one lattice spacing and fuse it into a neighboring tensor, repeating along a directed tree that connects all sites of $K$ to a chosen origin. Each arrow of the tree converts a shared index, which would appear in more than two tensors, into two new indices that each appear on exactly two neighboring tensors, producing a locally connected tensor network. The tree-choice optimization is the rectilinear Steiner tree problem, and the number of arrows determines the bond dimension of the initial tensor. For the accuracy claim, the load-bearing object is the 'squeezer' $P$, constructed from both left and right SVD isometries and singular values, which replaces the single-sided isometry used in HOTRG truncation.
What would settle it
Contract the tensor network produced by Algorithm 1 for a small lattice, such as a $4\times4$ Ising model with periodic boundary conditions, and compare with the original Boltzmann sum; exact equality is required by construction, so any discrepancy falsifies the method. For the squeezer claim, run boundary-HOTRG with the delta-initial tensors on a three-dimensional Ising or $\mathbb{Z}_2$ gauge model and check whether convergence in bond dimension matches the expansion-based initial tensors, which the paper does not test.
Extended reading notes
Core claim
The central claim is that the initial tensor construction does not assume any specific properties of the Boltzmann weights. Given a Boltzmann factor $K_{\sigma_{\hat r}\ldots}$ depending on several site variables, the algorithm inserts a delta function $\delta_{\sigma,a}$ that duplicates one index, shifts that delta to a neighboring site, and fuses it into the tensor there; repeating this along a directed tree that connects all sites of $K$ to a chosen origin yields a new tensor whose indices are each shared by exactly two neighboring tensors. This is an exact rewriting of the partition function, so no information is lost before truncation. The choice of tree is the rectilinear Steiner tree problem, and the number of arrows sets the final tensor's bond dimension. The paper further claims that the accuracy of TRG truncations can depend strongly on the initial tensor's symmetry: HOTRG with the delta-constructed asymmetric tensor is less accurate than with the Taylor-constructed symmetric one at fixed bond dimension, but when the coarse-graining step uses squeezers built from both left and right singular-value-decomposition isometries, the two initial tensors give the same accuracy. Replacing isometries by squeezers is stated to remove the dependence on the form of the initial tensors.
Load-bearing premise
The claim that squeezers remove the initial-tensor dependence for arbitrary TRG methods is transferred from one 2D Ising benchmark and from a construction summarized from earlier work, so the general improvement rests on that transfer holding beyond the tested case.
Editorial extensions
If this is right
- Any partition function written as a product of Boltzmann factors, including non-local interactions such as those in eq. (9), can be converted exactly into a TRG-ready tensor network without Taylor, character, or orthogonal-function expansions.
- The size of the initial tensor is controlled by the number of arrows in the Steiner tree, so the numerical cost of a TRG calculation can be estimated before coarse-graining begins.
- The same construction applies to inhomogeneous systems and to observables with impurity tensors, not just translation-invariant partition functions.
- Within the tested 2D Ising benchmark, boundary-type squeezers make HOTRG's free-energy error independent of whether the initial tensor came from the delta construction or from a Taylor expansion, removing a source of algorithm bias.
- For lattice gauge theories, the paper reports that the construction already works for the $\mathbb{Z}_2$ gauge model in a companion work, keeping open the path to sign-problem-free gauge calculations.
Reading between the lines
- Editorial extension: because the construction is exact, the only approximation in a TRG run is the truncation, so the delta-initial tensors could serve as a representation-independent benchmark for comparing different TRG algorithms on the same model.
- Editorial extension: the paper leaves the Steiner tree step to be solved by hand for small interaction graphs; automating a minimal-tree search for larger Boltzmann weights would be a direct practical follow-up.
- Editorial extension: the squeezer claim is demonstrated on one 2D Ising benchmark; applying boundary-type squeezers to non-symmetric initial tensors in three-dimensional or gauge models would test whether the 'removes dependence' statement holds beyond the tested case.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a systematic method for constructing locally connected tensor networks as initial inputs for tensor renormalization group (TRG) methods. Starting from a partition function written as a product of Boltzmann factors, the method introduces auxiliary indices through delta functions and shifts those deltas along a directed tree, so that each index appears on exactly two neighboring tensors. The construction is illustrated on the two-dimensional Ising model, a nonlocal four-spin interaction, and a three-dimensional model. The paper also compares HOTRG and boundary-HOTRG for two different initial tensor choices in the 2D Ising model and argues that boundary-HOTRG's squeezers remove the dependence of TRG accuracy on the initial tensor form.
Significance. The exact initial-tensor construction in Sections 2 and 3 is a genuinely useful contribution: it is parameter-free, involves no expansions or decompositions, and is applicable to arbitrary Boltzmann weights. The worked identities in Eqs. (4)-(8) and (10) are internally consistent, and the use of the Onsager result as an external benchmark for the Ising free energy is appropriate. The Steiner-tree formulation in Section 4 correctly identifies a practical algorithmic cost. However, the improvement claim in Section 5 is supported by only a single numerical example, so the significance of the 'removing initial-tensor dependence' part is presently conditional. If the claim is either strengthened by additional tests or appropriately qualified, the paper would be a solid proceedings contribution.
major comments (3)
- [§5, Fig. 5] The abstract and conclusions state that replacing isometries by squeezers 'removes the dependence on the form of the initial tensors', but the evidence is a single benchmark: the 2D Ising free energy at the critical temperature with V=2^20. No off-critical temperature, second model, different boundary condition, or error estimate is provided. Since this is the load-bearing evidence for the improvement claim, either add further tests or qualify the claim to the tested case.
- [§5, squeezer construction] The b-HOTRG squeezers P_1 and P_2 are not derived in this manuscript but are imported from Refs. [8,15], and the statement that the construction applies 'in any TRG method' is supported only by references. Because the removal of initial-tensor dependence is a central contribution, the paper should either explain the mechanism by which the squeezers guarantee independence for arbitrary TRG algorithms or explicitly defer the full claim to Ref. [8] with a precise statement of what the present paper demonstrates.
- [§3, after Algorithm 1] The notation after Algorithm 1 is confusing: the text says the new tensor K'(x) depends only on 'a_x and a_{x+|v|}', but the preceding definition and the two-dimensional examples use indices a_v and a_{v+μ}. This should be corrected to a_v and a_{v+μ}, or the convention should be defined precisely, so that the reader can follow the index bookkeeping in Eqs. (7) and (10).
minor comments (4)
- [Eq. (10)] The delta-function structure in Eq. (10) is not obvious from the short description of the three index shifts; a brief derivation or a labelled intermediate step would help the reader verify the mapping of the a, b, and c indices.
- [§5, Fig. 5] The lower two b-HOTRG curves appear essentially coincident in the printed figure. A short data table or a zoomed inset would make the claimed 'same accuracy' quantitative and would show the D-dependence more clearly.
- [§5, first paragraph] The sentence 'All TRG methods introduce a truncation, which minimizes a cost function' is too broad; some contraction schemes are exact for their respective tensor-network classes. Suggest softening to 'The TRG methods considered here'.
- [§6] The conclusion states that the method applies to general observables and inhomogeneous systems, but this is asserted rather than demonstrated. A one-sentence description of how impurity tensors or spatially varying Boltzmann factors enter Algorithm 1 would make the scope claim substantiated.
Circularity Check
No circularity in the central construction: Algorithm 1 is an exact index relabelling, and the squeezer-based improvement is an independently stated numerical claim rather than a fitted input or self-referential reduction.
full rationale
The claimed derivation chain in Sections 2 and 3 is self-contained and exact. Equation (4) inserts a delta function; equations (5)-(8) shift it by one site and show the partition function is unchanged, reducing to the original Ising Boltzmann factor when the delta is summed. Equation (10) performs the same relabelling for a non-local four-site interaction, and Algorithm 1 is the direct generalization. No Boltzmann weight property is expanded or decomposed, and no parameter is fitted to the exact solution used in Fig. 5. The later b-HOTRG discussion imports the squeezer construction from the authors' previous work [8] and from [15], but the definitions are quoted in the text and the numerical comparison is presented as a benchmark, not as a prediction derived from those definitions. The statement that 'Replacing isometries by squeezers removes the dependence on the form of the initial tensors' is supported by the Fig. 5 benchmark together with citations; this is an evidentiary and generality limitation, not a circular step in which the conclusion is equivalent to the input. I therefore find no reduction of a claimed result to its own fit or definition.
Assumptions & free parameters
assumptions (4)
- domain assumption The partition function is given as a product of Boltzmann factors depending on a finite set of lattice sites and is exactly represented by summing over all physical degrees of freedom.
- standard math Inserting delta functions and relabeling summed indices under periodic boundary conditions preserves the value of the partition function.
- domain assumption For any Boltzmann weight, a directed nearest-neighbor tree connecting all involved sites to an origin exists and is chosen to minimize arrows.
- domain assumption The boundary-HOTRG squeezer construction of [14,15] and [8] correctly defines a truncation that preserves the partition function to controlled accuracy.
Cite this review
Pith. "Pith review of Initial tensor construction for the tensor renormalization group." pith.science (2026). https://pith.science/paper/LTVUXW3P
@misc{pith2026250111810,
author = {Pith},
title = {Pith review of: Initial tensor construction for the tensor renormalization group},
year = {2026},
howpublished = {\url{https://pith.science/paper/LTVUXW3P}},
note = {Machine review of arXiv:2501.11810}
}
read the original abstract
We propose a method to construct the initial tensor representation of partition functions and observables for the tensor renormalization group (TRG). The TRG is a numerical calculation technique that utilizes a tensor network representations of physical quantities to investigate physical properties without encountering the sign problem. To apply the TRG, it is essential to construct a locally connected tensor network suitable for recursive coarse-graining. We present a systematic approach for translating a general tensor representation of the partition function to this form. Furthermore, we show the dependence of TRG algorithms on the choice of the initial tensor network representation and propose an improvement of TRG algorithms in this respect
Figures
Figures from the paper (3 more)
Reference graph
Works this paper leans on
-
[8]
K. Nakayama and M. Schneider,Initial tensor construction and dependence of the tensor renormalization group on initial tensors, Phys. Rev. D110 (2024) 094501 [2407.14226]
- [15]
-
[1]
Y. Meurice, R. Sakai and J. Unmuth-Yockey,Tensor lattice field theory for renormalization and quantum computing, Rev. Mod. Phys.94 (2022) 025005 [2010.06539]
arXiv 2022
-
[2]
Bañuls,Tensor Network Algorithms: A Route Map, Ann
M.C. Bañuls,Tensor Network Algorithms: A Route Map, Ann. Rev. Condensed Matter Phys. 14(2023) 173 [2205.10345]
arXiv 2023
-
[3]
M. Levin and C.P. Nave,Tensor renormalization group approach to two-dimensional classical lattice models,Phys. Rev. Lett.99 (2007) 120601 [cond-mat/0611687]
arXiv 2007
- [4]
-
[5]
K. Nakayama, L. Funcke, K. Jansen, Y.-J. Kao and S. Kühn,Phase structure of the CP(1) modelin the presenceof atopological𝜃-term,Phys.Rev.D 105(2022)054507 [2107.14220]
arXiv 2022
-
[6]
Y. Kuramashi and Y. Yoshimura,Three-dimensional finite temperature Z2 gauge theory with tensor network scheme, JHEP 08(2019) 023 [1808.08025]
arXiv 2019
Show all 18 references
-
[7]
Yosprakob and K
A. Yosprakob and K. Okunishi,Tensor renormalization group study of the three-dimensional SU(2) and SU(3) gauge theories with the reduced tensor network formulation, 2406.16763
-
[9]
K.H. Pai, S. Akiyama and S. Todo,Grassmann tensor renormalization group approach to (1 + 1)-dimensional two-color lattice QCD at finite density, 2410.09485
-
[10]
Y. Liu, Y. Meurice, M.P. Qin, J. Unmuth-Yockey, T. Xiang, Z.Y. Xie et al.,Exact blocking formulas for spin and gauge models, Phys. Rev. D88(2013) 056005 [1307.6543]
2013 arXiv
-
[11]
Zhao, Z.Y
H.H. Zhao, Z.Y. Xie, Q.N. Chen, Z.C. Wei, J.W. Cai and T. Xiang,Renormalization of tensor-network states, Phys. Rev. B81 (2010) 174411 [1002.1405]
2010 arXiv
-
[12]
Z.Y. Xie, J. Chen, M.P. Qin, J.W. Zhu, L.P. Yang and T. Xiang,Coarse-graining renormalization by higher-order singular value decomposition,Phys. Rev. B86(2012) 045139 [1201.1144]
2012 arXiv
-
[13]
Sherwani,Algorithms for VLSI Physical Design Automation, Springer New York (1999), 10.1007/b116436
N.A. Sherwani,Algorithms for VLSI Physical Design Automation, Springer New York (1999), 10.1007/b116436
1999 doi
-
[14]
S. Iino, S. Morita and N. Kawashima,Boundary tensor renormalization group,Phys. Rev. B 100 (2019) 035449 [1905.02351]
2019 arXiv
-
[16]
Kaufman,Crystal statistics
B. Kaufman,Crystal statistics. ii. partition function evaluated by spinor analysis, Phys. Rev. 76(1949) 1232. 9 Initial tensor construction for the tensor renormalization group Katsumasa Nakayama
1949
-
[17]
Kadoh and K
D. Kadoh and K. Nakayama,Renormalization group on a triad network, 1912.02414
1912 arXiv
-
[18]
Nakayama,Randomized higher-order tensor renormalization group, 2307.14191
K. Nakayama,Randomized higher-order tensor renormalization group, 2307.14191. 10
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.