Tensor Renormalization Group (TRG)

Updated 2 December 2025

Tensor Renormalization Group (TRG) is a method that coarse-grains high-rank tensor networks to compute partition functions and correlation functions in lattice models.
It uses sequential SVD-based truncations to compress local tensor spaces, balancing computational cost with controlled accuracy through the bond dimension parameter.
Advancements such as HOTRG, boundary TRG, and hybrid schemes enhance robustness and scalability, enabling precise evaluations even in high entanglement or disordered systems.

The Tensor Renormalization Group (TRG) is a class of real-space renormalization group methods for evaluating the partition functions and correlators of classical and quantum lattice models by systematically coarse-graining high-rank tensor networks. Introduced by Levin and Nave (2007), TRG establishes a scalable, variationally controlled alternative to conventional Monte Carlo and momentum-space RG, capable of achieving high-precision thermodynamics, correlation functions, and critical properties for a diverse range of spin, dimer, gauge, and quantum systems, often free of the sign problem. Over the past decade, TRG has evolved through a series of algorithmic variants, including higher-order, hybrid, and boundary-aware refinements, enabling systematic improvement of computational efficiency, robustness, and fidelity under conditions of high entanglement and in the presence of inhomogeneity, boundaries, and disorder.

1. Tensor Network Representation of Lattice Models

Lattice partition functions in classical statistical mechanics and quantum many-body systems can be mapped to the contraction of local tensors on a graph reflecting the lattice geometry. For a classical two-dimensional model with local Hilbert space of dimension $d$ , the partition function may be written as a contraction of rank-4 (square lattice) or rank-3/rank-6 (triangular/honeycomb/dimer models) tensors, where each index represents a local state or bond configuration (Güven et al., 2010, Roychowdhury et al., 2014). In quantum systems and gauge theories, character expansions or Trotter-Suzuki decompositions convert interactions and time evolution into tensor network form, often with complex or Grassmann-valued components (Meurice et al., 2013, Campos et al., 2021, Yosprakob, 24 Jan 2025).

Mapping the original (potentially non-locally connected) partition function to a locally connected tensor network suitable for TRG usually requires algorithmic transformation, such as the delta-function insertion and minimal spanning tree procedure, which reduces multi-site Boltzmann factors to a network of rank-2D tensors with uniquely paired indices on each edge (Nakayama et al., 2024, Nakayama et al., 21 Jan 2025). The initial tensor construction, including its symmetries, can significantly influence downstream TRG truncation performance and must be carefully optimized for robust, model-agnostic coarse-graining.

2. Core Algorithm: Sequential Coarse-Graining and Truncation

The essential TRG step groups blocks of tensors and applies local singular value decompositions (SVD) or higher-order SVD (HOSVD) to compress composite local spaces and truncate to a maximum bond dimension $\chi$ (Güven et al., 2010, Meurice et al., 2013, Yosprakob, 24 Jan 2025). For the Levin–Nave TRG on the square lattice, two neighboring tensors are viewed as a matrix, SVD is performed, and only the top $\chi$ singular values and associated states are kept. After recontracting, a new coarse-grained tensor network is obtained, with the effective lattice spacing doubled in each direction.

In higher-order and general lattices (HOTRG, HOSRG), tensor blocking proceeds along one or more directions alternately, and truncation occurs after suitable reshaping and environment-aware analysis of the local tensor contractions (Chen et al., 2019). The cost per iteration for basic square-lattice TRG is $O(\chi^6)$ , with HOTRG raising this to $O(\chi^7)$ . Hybrid, projective, and environment-optimized variants reduce this cost, sometimes to $O(\chi^5)$ , while boundary-TRG methods maintain accuracy even in asymmetric or inhomogeneous initial tensor settings (Nakamura et al., 2018, Nakayama et al., 2024).

3. Convergence, Accuracy, and Computational Scaling

The accuracy of TRG is systematically controllable via the truncation parameter $\chi$ , which limits entanglement retained at each RG step. For gapped, short correlation-length systems (e.g., triangular-lattice dimer models), modest $\chi$ yields sub-per-mille precision in free energies and correlators even for large lattices (Roychowdhury et al., 2014). At criticality, where the entanglement entropy grows logarithmically with system size, much larger $\chi$ is required; errors in free energy and observables typically decrease as a power law, $|ΔF|\sim\chi^{-b}$ with $b\sim3$ –$5$ depending on algorithm and model (Chen et al., 2019, Adachi et al., 2020).

Algorithmic scaling for the most common variants is summarized below:

Method	Cost per RG step	Memory	Bond-dimension scaling
TRG (Levin–Nave)	$O(\chi^6)$	$O(\chi^4)$	Polynomial (critical)
HOTRG / HOSRG	$O(\chi^7)$	$O(\chi^4)$	Higher, but robust
Boundary-TRG (BTRG, etc)	$O(\chi^7)$	$O(\chi^4)$	Symmetry robust
Projective/TRG-PT	$O(\chi^5\,n_{\rm itr})$	$O(\chi^4)$	Fast, variational
Bond-weighted TRG	$O(\chi^5)$	$O(\chi^3)$	Enhanced fixed-point
Core-tensor RG (CTRG)	$O(\chi^4)$	$O(\chi^3)$	Linear shrinkage
Stochastic Hybrid	$O(V\chi^6)$	Model-dependent	Statistical error

Environment-aware (SRG, ∂TRG, variational boundary) and projective-truncation methods can improve accuracy at fixed resource cost by globally optimizing truncations based on backward gradients or explicitly constructed (possibly variational) environments (Chen et al., 2019, Song et al., 14 Aug 2025).

4. Extensions: Boundaries, Disorder, Fermions, and Real-Time Evolution

Boundary TRG formulations introduce explicit boundary tensors and track boundary fixed-points, enabling calculation of surface and interface observables, surface free energies, and direct access to boundary scaling spectra and conformal towers under criticality (Iino et al., 2019). These formulations are essential for extracting boundary conformal field theory data and characterizing systems with open, free, or fixed boundary conditions.

Disordered and frustrated systems (spin glasses, random-bond Ising models) require topological-invariant extensions (pTRG), which preserve internal gauge invariance and allow consistent estimation of physical quantities even under strong frustration or bond-dilution (Wang et al., 2013). Coarse-graining remains controlled, and local observables can be computed via backward environment iteration.

For fermionic or gauge systems, TRG is adapted to handle Grassmann or non-Abelian variables via appropriate expansion (e.g., character or auxiliary field), analytical contraction of matrix indices, and removal of non-local entanglement structures (e.g., via the armillary sphere construction in lattice QCD, which eliminates non-physical loop entanglement) (Yosprakob, 24 Jan 2025).

In quantum systems, TRG can contract higher-dimensional tensor networks representing the (imaginary-time) evolution of 1D chains, enabling calculation of entanglement entropy, central charges, and extraction of effective Hamiltonians and real-time evolution operators (using a formal analytic continuation). Modern formulations achieve precise estimates for entanglement scaling and central charge (e.g., $c=0.49997(8)$ for the 1D Ising chain at $D=96$ ) (Hayazaki et al., 2 Sep 2025, Meurice et al., 2016, Hite et al., 2024).

5. Algorithmic Innovations and Hybrid Schemes

Recent development in TRG includes projective-truncation (Nakamura et al., 2018) (lowering asymptotic complexity to $O(\chi^5)$ via variational projectors), hybrid stochastic approaches (replacing deterministic truncation with statistically unbiased estimators using noise vectors, eliminating truncation bias at the cost of controllable statistical variance) (Ohki et al., 2021), and differentiable TRG (∂TRG), which employs automatic differentiation for deep-sweep optimization of tensor environments, yielding error reductions by orders of magnitude (Chen et al., 2019).

Bond-weighted TRG (BTRG) introduces explicit weight tensors to absorb powers of singular values, stabilizing coarse-graining flows at criticality with invariant singular-value spectra and robust fixed-point tensors (Adachi et al., 2020). Variational boundary-based renormalization constructs optimal projectors from a globally computed environment, closely matching the global entanglement structure and further improving efficiency and robustness, with generalizations to 3D and higher (Song et al., 14 Aug 2025).

Linear-shrinking core-tensor RG (CTRG) updates only rows and columns adjacent to a core tensor, resulting in per-iteration costs scaling as $O(\chi^4)$ and linearly shrinking system size, while retaining comparable accuracy to standard TRG for moderate to large $\chi$ (Lan et al., 2019).

6. Physical Observables, Criticality, and Applications

TRG methods compute not only free energies but also local magnetizations, spin-spin and dimer-dimer correlation functions, multi-site entropies, critical exponents, and phase diagrams. Extraction of phase boundaries is achieved via finite-size scaling of long-distance correlators, and boundary TRG allows access to surface-order parameters and conformal spectra (Güven et al., 2010, Iino et al., 2019, Wang et al., 2013). Correlation functions are calculated using impurity tensor insertions; for critical models with algebraic correlations, accurate treatment requires large $\chi$ or the use of disentangling extensions (e.g., TNR, Loop-TNR).

TRG's deterministic nature allows it to bypass sign problems that invalidate Monte Carlo, enabling study of quantum systems at finite density, with chemical potential, or in the presence of external fields, as demonstrated for O(N), principal chiral, and non-Abelian lattice models (Meurice et al., 2013, Campos et al., 2021, Yosprakob, 24 Jan 2025).

Applications include quantum dimer models (Roychowdhury et al., 2014), Lorentzian/Euclidean quantum gravity via Regge calculus (Ito et al., 2022), interacting quantum fields (Campos et al., 2021), and as benchmarking/validation tools for emerging quantum simulation and NISQ devices, especially for real-time evolution in non-integrable models (Hite et al., 2024).

7. Limitations, Open Problems, and Rigorous Results

While classical TRG is variationally controlled and systematically improvable, it is limited by the polynomial cost scaling in $\chi$ and the inability of standard truncation to capture infinite-range entanglement at criticality—exponential convergence in $\chi$ is only achieved by tensor network renormalization (TNR) or by adding explicit disentanglers (Roychowdhury et al., 2014, Lan et al., 2019).

Recent work has established computer-assisted, rigorous control over the high-temperature RG fixed-point flow for generic 2D tensor networks, introducing master functions and bounding box ("hat-tensor") techniques. For explicit initial deviation magnitudes $\|\widehat{b}\|\lesssim 0.02$ in 63 orthogonal sectors, one can prove flow to the high-T fixed point for the 2D Ising and XY models, yielding explicit bounds on analyticity domains in $(\beta)$ (Ebel et al., 3 Jun 2025).

A remaining challenge is the efficient and accurate simulation of models with extensive disorder, frustration, and long-range interactions, as well as full treatment of non-Abelian symmetry in higher dimensions with analytically tractable truncation schemes and numerical stability (Wang et al., 2013, Yosprakob, 24 Jan 2025). Hybrid and boundary-optimized formulations, deep variational training, and further integration with quantum computational methods represent active frontiers.

References