The chapter in which Nishino and Okunishi introduced the corner transfer matrix renormalization group (CTMRG) for two-dimensional (2D) classical lattice models closed by formulating its three-dimensional (3D) extension: divide the cubic system into eight corner transfer tensors (CTTs), build the density submatrix (DSM) as their product, and renormalize the CTTs from its dominant eigenvectors. The construction was implemented by the same authors in 1998, but the resulting algorithm is quantitatively inaccurate: the spectra of the density submatrices for one- and two-dimensional cuts of the 3D lattice decay so slowly that truncation is ineffective even away from criticality, and the approach was abandoned in favor of variational tensor-product-state approximations, while the field adopted tree-type coarse-graining schemes (the higher-order and anisotropic tensor renormalization groups, HOTRG and ATRG) that lack corner structure and controlled entanglement handling in 3D. The authors' 2022 retrospective still lists both efficient control of short-range entanglements in higher dimensions and the nature of the 3D corner-double-line tensor as open, the latter identified as a possible cause of the failure. The problem is to settle the viability of corner-transfer, density-matrix-based renormalization in three dimensions: either realize it as a practical, quantitatively accurate renormalization group for 3D classical lattice models with truncation errors controlled by the number of retained states, or identify and characterize the structural obstruction that defeats it.

Background

Classical lattice models as transfer matrices. A classical lattice model in dd spatial dimensions — for example the Ising model with spins si=±1s_i=\pm 1 and Boltzmann weights exp⁡(βJsisj)\exp(\beta J s_i s_j) — has a partition function ZZ that can be written as the trace of a product of transfer matrices, objects whose repeated application builds up the lattice layer by layer. Exact evaluation is generally impossible: the row-to-row transfer matrix of an L×LL\times L square lattice already has dimension qLq^L for qq-state spins. Numerical real-space renormalization group (RG) methods attack this by iteratively replacing many local degrees of freedom with fewer effective ones while keeping the dominant information.

Density-matrix truncation in one and two dimensions. The density matrix renormalization group (DMRG) of S. R. White (Phys. Rev. Lett. 69, 2863 (1992)) made such truncation precise for one-dimensional (1D) quantum chains: the basis to keep is chosen from the dominant eigenvectors of the reduced density matrix of the half-chain. For 2D classical models, Baxter's corner transfer matrix (CTM) — the quarter-plane contribution to the half-plane transfer matrix, introduced for the eight-vertex model (R. J. Baxter, J. Stat. Phys. 15, 485 (1976)) — factorizes the half-plane transfer matrix into four corner matrices, so that the reduced density submatrix (DSM) ρ\rho of the boundary spins is a product ρ=ABCD\rho=ABCD of four CTMs. Nishino and Okunishi's corner transfer matrix renormalization group (CTMRG) (J. Phys. Soc. Jpn. 65, 891 (1996)) truncates the corner basis to the mm dominant eigenvectors of ρ\rho and iterates, giving a highly efficient real-space RG for 2D classical models; its accuracy rests on the fast decay of the DSM spectrum in 2D. Its sibling, the product wave function renormalization group (PWFRG), feeds the same corner construction back into 1D quantum chains as a rapid infinite-system DMRG.

The three-dimensional construction. For a 3D classical model on the cubic lattice, the natural generalization replaces the square system by a cube divided into eight corner transfer tensors (CTTs), supplemented by six additional tensors carrying the plane- and edge-structure indices of the cubic lattice, and constructs the DSM as the product of the CTTs; diagonalizing this DSM and keeping the mm dominant eigenvectors renormalizes the CTTs. In their 1997 chapter, Nishino and Okunishi formulated this construction and stated the obstruction: unlike the renormalized CTMs in 2D, the renormalized CTT is not diagonal, and the numerical cost was prohibitive even for small mm. The construction was implemented for the cubic-lattice Ising model in 1998 and failed quantitatively: the DSM spectra for one- and two-dimensional cuts of the 3D lattice decay so slowly that any feasible mm discards essential weight, even at temperatures moderately away from criticality.

Why this is the hard frontier. The modern understanding of why truncation schemes fail, and how to repair them, comes from tensor-network coarse graining. The tensor renormalization group (TRG) of Levin and Nave (Phys. Rev. Lett. 99, 120601 (2007)), its higher-order variant HOTRG (Phys. Rev. B 86, 045139 (2012)) and its cheaper anisotropic variant ATRG (Phys. Rev. B 102, 054432 (2020)) — the current standard for 3D classical models — truncate by local singular value decompositions and generate tree-type networks whose fixed points contain corner-double-line (CDL) tensors, i.e. purely short-range entanglement loops that survive coarse graining and mask the critical fixed point. In 2D, tensor network renormalization (TNR) (G. Evenbly and G. Vidal, Phys. Rev. Lett. 115, 180405 (2015)) removes this short-range entanglement with disentanglers and restores correct critical fixed points, including logarithmic corrections; no equally efficient scheme exists in 3D. In their 2022 retrospective, Okunishi, Nishino, and Ueda state that efficient control of short-range entanglements in higher dimensions is not established and that the nature of the 3D version of a CDL tensor — a possible cause of the 3D corner-transfer-tensor RG failure — remains to be clarified. Very recent work (2024–2026) builds entanglement-filtered 3D real-space RG schemes — schemes that explicitly discard short-range entanglement at each coarse-graining step — with partially controlled errors, and a quantitative extrapolation methodology at fixed retained dimension mm ('finite-mm scaling', Phys. Rev. E 101, 062111 (2020)) is available in 2D; but no corner-transfer, density-matrix-based RG for 3D classical models works today. A 2025 variational route reformulates the 3D classical problem as a leading-eigenvalue problem for its layer-to-layer transfer matrix and contracts its environments with a split CTMRG built from two-dimensional corner matrices, reaching Monte-Carlo-comparable accuracy for the 3D Ising model; it optimizes a variational tensor-product state rather than iterating a three-dimensional corner-transfer renormalization group, and it explicitly attributes the failure of the direct 3D CTMRG generalization to the slow decay of the spectra of density matrices constructed from corner-transfer tensors.

Problem Statement

Determine whether the corner-transfer, density-matrix-based renormalization group can be extended to three-dimensional classical lattice models as a practical and quantitatively accurate method, in the sense formulated by Nishino and Okunishi. The question is discharged by either of two linked outcomes, which together settle the viability of the approach.

  1. Constructive goal. A real-space renormalization group for 3D classical lattice models in which the truncation basis at each step is chosen from the dominant eigenvectors of a density submatrix built from corner transfer tensors — the natural 3D generalization of the 2D CTMRG, such as the eight-corner-transfer-tensor construction with its supplemental plane- and edge-structure tensors — implemented so that the truncation error is demonstrably controlled by the number mm of retained states and by the number of renormalization steps, and benchmarked on standard 3D models — at least the cubic-lattice Ising model (critical temperature, critical exponents ν\nu of the correlation length and η\eta of the spin field, and free energy) — with quantitative accuracy defined as errors in these quantities that decrease with mm and with the number of renormalization steps and approach their established benchmark values.

  2. Diagnostic goal. A characterization of the obstruction that defeats the three-dimensional construction: an identification of the structural reason — such as the 3D analogue of the corner-double-line tensor, or equivalently the entanglement-structure of 3D corner environments — why the spectra of the corner density submatrices for one- and two-dimensional cuts of the 3D lattice decay so slowly, in a form that quantitatively explains the documented failure of the 1998 algorithm for the cubic-lattice Ising model and locates the obstruction in the corner-transfer, density-matrix construction itself rather than in the details of that implementation.

Either outcome resolves the question. Schemes that only apply tree-type coarse graining — local singular-value truncation without a corner-based or environment-based density submatrix, as in standard TRG/HOTRG/ATRG — do not answer it, since the distinguishing feature of this method is precisely the density-matrix (environment-based) truncation that underlies the success of DMRG and CTMRG in one and two dimensions. Likewise, variational tensor-product-state optimizations for 3D models that contract only two-dimensional corner environments — after compressing the 3D network to an effective layered problem — do not answer it, since they bypass rather than realize the three-dimensional corner-transfer construction whose viability is at issue.

References

  1. T. Nishino and K. Okunishi, 'Density matrix and renormalization for classical lattice models', in Strongly Correlated Magnetic and Superconducting Systems, Lecture Notes in Physics (Springer, Berlin, 1997), pp. 167-183. DOI: 10.1007/bfb0104638, https://doi.org/10.1007/bfb0104638; preprint arXiv:cond-mat/9610107, https://arxiv.org/abs/cond-mat/9610107
  2. S. R. White, 'Density matrix formulation for quantum renormalization groups', Phys. Rev. Lett. 69, 2863 (1992). DOI: 10.1103/PhysRevLett.69.2863, https://doi.org/10.1103/PhysRevLett.69.2863
  3. R. J. Baxter, 'Corner transfer matrices of the eight-vertex model. I. Low-temperature expansions and conjectured properties', J. Stat. Phys. 15, 485 (1976). DOI: 10.1007/BF01020802, https://doi.org/10.1007/BF01020802
  4. T. Nishino and K. Okunishi, 'Corner Transfer Matrix Renormalization Group Method', J. Phys. Soc. Jpn. 65, 891 (1996). DOI: 10.1143/jpsj.65.891, https://doi.org/10.1143/jpsj.65.891
  5. T. Nishino and K. Okunishi, 'A Density Matrix Algorithm for 3D Classical Models', J. Phys. Soc. Jpn. 67, 3066 (1998). DOI: 10.1143/jpsj.67.3066, https://doi.org/10.1143/jpsj.67.3066
  6. K. Okunishi and T. Nishino, 'Kramers-Wannier Approximation for the 3D Ising Model', Prog. Theor. Phys. 103, 541 (2000). DOI: 10.1143/ptp.103.541, https://doi.org/10.1143/ptp.103.541
  7. T. Nishino, K. Okunishi, Y. Hieida, N. Maeshima, and Y. Akutsu, 'Self-consistent tensor product variational approximation for 3D classical models', Nucl. Phys. B 575, 504 (2000). DOI: 10.1016/S0550-3213(00)00133-4, https://doi.org/10.1016/S0550-3213(00)00133-4; arXiv:cond-mat/0001083, https://arxiv.org/abs/cond-mat/0001083
  8. N. Maeshima, Y. Hieida, Y. Akutsu, T. Nishino, and K. Okunishi, 'Vertical density matrix algorithm: A higher-dimensional numerical renormalization scheme based on the tensor product state ansatz', Phys. Rev. E 64, 016705 (2001). DOI: 10.1103/PhysRevE.64.016705, https://doi.org/10.1103/PhysRevE.64.016705; arXiv:cond-mat/0101360, https://arxiv.org/abs/cond-mat/0101360
  9. A. Gendiar and T. Nishino, 'Latent heat calculation of the three-dimensional q = 3, 4, and 5 Potts models by the tensor product variational approach', Phys. Rev. E 65, 046702 (2002). DOI: 10.1103/PhysRevE.65.046702, https://doi.org/10.1103/PhysRevE.65.046702
  10. M. Levin and C. P. Nave, 'Tensor Renormalization Group Approach to Two-Dimensional Classical Lattice Models', Phys. Rev. Lett. 99, 120601 (2007). DOI: 10.1103/PhysRevLett.99.120601, https://doi.org/10.1103/PhysRevLett.99.120601
  11. R. Orús and G. Vidal, 'Simulation of two-dimensional quantum systems on an infinite lattice revisited: Corner transfer matrix for tensor contraction', Phys. Rev. B 80, 094403 (2009). DOI: 10.1103/PhysRevB.80.094403, https://doi.org/10.1103/PhysRevB.80.094403; arXiv:0905.3225, https://arxiv.org/abs/0905.3225
  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' (HOTRG), Phys. Rev. B 86, 045139 (2012). DOI: 10.1103/PhysRevB.86.045139, https://doi.org/10.1103/PhysRevB.86.045139
  13. G. Evenbly and G. Vidal, 'Tensor Network Renormalization', Phys. Rev. Lett. 115, 180405 (2015). DOI: 10.1103/PhysRevLett.115.180405, https://doi.org/10.1103/PhysRevLett.115.180405
  14. H. W. J. Blöte, W. Guo, and M. P. Nightingale, 'Scaling in the vicinity of the four-state Potts fixed point', J. Phys. A 50, 324001 (2017). DOI: 10.1088/1751-8121/aa7b53, https://doi.org/10.1088/1751-8121/aa7b53; arXiv:1710.06088, https://arxiv.org/abs/1710.06088
  15. D. Adachi, T. Okubo, and S. Todo, 'Anisotropic tensor renormalization group', Phys. Rev. B 102, 054432 (2020). DOI: 10.1103/PhysRevB.102.054432, https://doi.org/10.1103/PhysRevB.102.054432
  16. H. Ueda, K. Okunishi, K. Harada, R. Krčmár, A. Gendiar, S. Yunoki, and T. Nishino, 'Finite-m scaling analysis of Berezinskii-Kosterlitz-Thouless phase transitions and entanglement spectrum for the six-state clock model', Phys. Rev. E 101, 062111 (2020). DOI: 10.1103/PhysRevE.101.062111, https://doi.org/10.1103/PhysRevE.101.062111; arXiv:2001.10176, https://arxiv.org/abs/2001.10176
  17. K. Okunishi, T. Nishino, and H. Ueda, 'Developments in the Tensor Network — from Statistical Mechanics to Quantum Entanglement', J. Phys. Soc. Jpn. 91, 062001 (2022). DOI: 10.7566/JPSJ.91.062001, https://doi.org/10.7566/JPSJ.91.062001; arXiv:2111.12223, https://arxiv.org/abs/2111.12223
  18. X. F. Liu, Y. F. Fu, W. Q. Yu, J. F. Yu, and Z. Y. Xie, 'Variational Corner Transfer Matrix Renormalization Group Method for Classical Statistical Models', Chin. Phys. Lett. 39, 067502 (2022). DOI: 10.1088/0256-307X/39/6/067502, https://doi.org/10.1088/0256-307X/39/6/067502; arXiv:2203.17098, https://arxiv.org/abs/2203.17098
  19. S. Nyckees, A. Rufino, F. Mila, and J. Colbois, 'Critical line of the triangular Ising antiferromagnet in a field from a C3-symmetric corner transfer matrix algorithm', Phys. Rev. E 108, 064132 (2023). DOI: 10.1103/PhysRevE.108.064132, https://doi.org/10.1103/PhysRevE.108.064132; arXiv:2306.09046, https://arxiv.org/abs/2306.09046
  20. X. Lyu and N. Kawashima, 'Three-dimensional real-space renormalization group with well-controlled approximations', Phys. Rev. E 111, 054140 (2025). DOI: 10.1103/PhysRevE.111.054140, https://doi.org/10.1103/PhysRevE.111.054140; arXiv:2412.13758, https://arxiv.org/abs/2412.13758
  21. X. Lyu and N. Kawashima, 'Essential Difference Between 2D and 3D From the Perspective of real-space Renormalization Group', J. Stat. Phys. 193, 32 (2026). DOI: 10.1007/s10955-026-03587-1, https://doi.org/10.1007/s10955-026-03587-1; arXiv:2311.05891, https://arxiv.org/abs/2311.05891
  22. X. Lyu and N. Kawashima, 'Lattice-reflection symmetry in tensor-network renormalization group with entanglement filtering in two and three dimensions', Phys. Rev. B 113, 144109 (2026). DOI: 10.1103/z7sk-7sq8, https://doi.org/10.1103/z7sk-7sq8; arXiv:2510.19428, https://arxiv.org/abs/2510.19428
  23. Z.-H. Wang and L.-P. Yang, 'The Klein bottle ratio of two-dimensional ferromagnetic Potts models', arXiv:2604.00870 (2026), https://arxiv.org/abs/2604.00870
  24. X.-Z. Xu, T.-Y. Lin, and G.-M. Zhang, 'Efficient optimization of variational tensor-network approach to three-dimensional statistical systems', Phys. Rev. B 112, 134403 (2025). DOI: 10.1103/7dbj-w7cn, https://doi.org/10.1103/7dbj-w7cn; arXiv:2506.19339, https://arxiv.org/abs/2506.19339