Open problems for the reduced-cost corner-transfer contraction of infinite PEPS environments: truncation-error control, criticality limits, symmetry and update-scheme integration, and single-layer benchmarking

Lan and Evenbly (arXiv:2306.08212, 2023) reduced the dominant cost of contracting double-layer corner-transfer environments for infinite projected entangled-pair states (iPEPS) from $O(\chi^3 D^6)$ to $O(\chi^3 D^3)$ through two rank-reducing approximations, the reduced-environment approximation and the sequential-order decomposition, built from isometries constructed out of single-layer halves of the double-layer tensor. The approximations are heuristic: no error analysis controls the two-stage isometry truncation (in particular the choice of the intermediate dimension $\chi'$ ), their formal limits near critical points or in models with stronger local entanglement are unestablished, integration with abelian/non-abelian symmetries and with variational and other update schemes is undemonstrated, and no systematic comparison with single-layer corner-transfer approaches exists. The objective is to turn the reduced-cost contraction into a validated, error-controlled tool.

Background

An infinite projected entangled-pair state (iPEPS) is a variational tensor-network ansatz for the ground state of a two-dimensional quantum lattice model directly in the thermodynamic limit: each lattice site carries a tensor whose virtual legs have bond dimension $D$ , and contracting neighboring tensors with shared legs builds the full wavefunction. Evaluating expectation values requires contracting the double-layer network $\langle\psi|\psi\rangle$ , in which bra and ket tensors are combined, so the object absorbed into the environment has local dimension $D^2$ . The standard contraction engine is the corner-transfer-matrix renormalization group (CTMRG), introduced by Nishino and Okunishi (JPSJ 65, 891 (1996)): the infinite network is iteratively absorbed into four corner matrices and four edge tensors, which are truncated to an environment (boundary) dimension $\chi$ . For double-layer environments the dominant CTMRG cost scales as $O(\chi^3 D^6)$ , and this is the practical bottleneck that keeps $D$ small in iPEPS studies.

Lan and Evenbly (arXiv:2306.08212, 2023) introduced two approximations that lower this cost to $O(\chi^3 D^3)$ . Both construct isometries (rectangular matrices with orthonormal columns, used as projectors onto a lower-dimensional subspace) out of one layer (one 'half') of the double-layer tensor and use them to project the double-layer environment onto an effective bond dimension of order $D$ instead of $D^2$ : the reduced-environment approximation applies this projection to the environment tensors, while the sequential-order decomposition applies a pair of isometries $v^{(1)}$ , $v^{(2)}$ in sequence, the first projecting onto an intermediate dimension $\chi'$ before the second brings the environment to dimension $\chi$ (the paper fixes $\chi'=\chi$ 'for the sake of concreteness'). On square-lattice spin-1/2 antiferromagnetic Heisenberg benchmarks up to $D=14$ and on the ferromagnetic transverse-field Ising model (Ising spins with nearest-neighbor $S^z S^z$ coupling and a tunable transverse field $h$ acting on $S^x$ , which drives a quantum phase transition), both approximations give results comparable to standard CTMRG at the same $\chi$ ; a single consistency check near the Ising critical field at $D=3$ , $\chi=30$ showed all three algorithms agreeing exactly.

What is missing is any control on the truncation error. The rank-reducing isometries are constructed heuristically, no bound or systematic error analysis relates the approximation quality to $\chi'$ , $\chi$ , and $D$ , and the behavior near critical points (where the correlation length diverges and the environment spectrum becomes dense) or in models with stronger local entanglement (requiring a larger bond dimension $D$ for a fixed accuracy than the source's Heisenberg benchmark) is uncharacterized. Three applicability gaps compound this: standard iPEPS codes exploit global symmetries by block-decomposing tensors (abelian cases such as particle-number conservation and non-abelian cases such as SU(2), whose irreducible multiplets are degenerate), but it is unknown how the layer-built isometries interact with such symmetry blocks; the approximations were tested only within imaginary-time evolution (a ground-state search in which the state is evolved in imaginary time so that components above the ground state decay), while variational optimization (direct minimization of the energy over the tensor elements) and other update schemes that use CTM may respond differently to an approximate environment; and the alternative strategy of contracting the bra and ket layers as two separate single-layer networks (so that the absorbed environment tensors have local dimension $D$ rather than $D^2$ ), as in Haghshenas, Gong, and Sheng (PRB 99, 174423 (2019)), was compared with the reduced-cost approach only qualitatively — the source notes that the two have the same leading cost, that its own method keeps the unit cell (the minimal block of distinct site tensors whose repetition tiles the infinite lattice) at the standard double-layer size whereas the single-layer approach expands it fourfold, and that the single-layer approach generally requires a larger $\chi$ for comparable accuracy — but never by systematic benchmarking.

Problem Statement

The objective is to turn Lan and Evenbly's reduced-cost corner-transfer-matrix contraction of double-layer iPEPS environments — the reduced-environment and sequential-order approximations — into a validated, error-controlled tool, by resolving five linked questions:

(1) Truncation-error analysis. Characterize the error introduced by the two-stage isometry truncation, in particular its dependence on the intermediate dimension $\chi'$ relative to the environment dimension $\chi$ and the PEPS bond dimension $D$ , and either derive computable (a priori or a posteriori) error bounds or establish a justified practical rule for choosing $\chi'$ , including for strongly correlated systems.

(2) Formal limits. Establish the limits of validity of the two approximations near critical points with very large correlation lengths and in models with stronger local entanglement (requiring a larger bond dimension $D$ for a fixed accuracy than the source's Heisenberg benchmark), going beyond the source's single transverse-field Ising consistency point at $D=3$ , $\chi=30$ .

(3) Symmetry integration. Implement the approximations with abelian and non-abelian symmetry-blocked tensors, quantify the additional speed-ups and memory savings, and determine how the layer-built isometries interact with symmetry-block structure.

(4) Update-scheme integration. Demonstrate the practical performance when the approximations are combined with variational optimization and with other iPEPS update schemes that use CTM, beyond the imaginary-time evolution tested in the source.

(5) Single-layer comparison. Carry out the systematic benchmarking against single-layer CTM approaches with different unit-cell requirements, quantifying the trade-offs among accuracy, boundary dimension, and unit-cell size that the source discussed only qualitatively.

The five components share one scientific objective: determining whether the $O(\chi^3 D^3)$ contraction can be trusted and used generally, with components (1)-(2) as the analytic core (what the truncation costs in accuracy and where it fails) and (3)-(5) as the applicability extensions (whether it composes with the standard machinery of the field). A complete answer resolves all five; progress on a subset is partial progress toward the single objective and must state which components remain.

References

Wangwei Lan and Glen Evenbly, 'Reduced Contraction Costs of Corner-Transfer Methods for PEPS', arXiv:2306.08212 (2023), DOI: 10.48550/arXiv.2306.08212, https://arxiv.org/abs/2306.08212
Tomotoshi Nishino and Kouichi Okunishi, 'Corner Transfer Matrix Renormalization Group Method', Journal of the Physical Society of Japan 65, 891 (1996), DOI: 10.1143/JPSJ.65.891, https://doi.org/10.1143/JPSJ.65.891
R. Haghshenas, Shou-Shu Gong, and D. N. Sheng, 'Single-layer tensor network study of the Heisenberg model with chiral interactions on a kagome lattice', Physical Review B 99, 174423 (2019), DOI: 10.1103/PhysRevB.99.174423, https://doi.org/10.1103/PhysRevB.99.174423
Jan Naumann, Erik L. Weerda, Jens Eisert, Matteo Rizzi, and Philipp Schmoll, 'Variationally optimizing infinite projected entangled-pair states at large bond dimensions: A split corner transfer matrix renormalization group approach', Physical Review B 111, 235116 (2025), DOI: 10.1103/PhysRevB.111.235116, arXiv:2502.10298, https://doi.org/10.1103/PhysRevB.111.235116
Yining Zhang, Qi Yang, and Philippe Corboz, 'Accelerating two-dimensional tensor network contractions using QR decompositions', Physical Review B 113, L201106 (2026), DOI: 10.1103/ptls-kr9z, arXiv:2505.00494, https://doi.org/10.1103/ptls-kr9z
Addison D. S. Richards and Erik S. Sørensen, 'Ace-TN: GPU-Accelerated Corner-Transfer-Matrix Renormalization of Infinite Projected Entangled-Pair States', arXiv:2503.13900 (2025), https://arxiv.org/abs/2503.13900