Tensor Network / PEPS Formalism

• Tensor Network / PEPS formalism is a framework that represents high-dimensional quantum states using interconnected lower-order tensors to capture area-law entanglement.
• PEPS arranges local tensors on lattice vertices with contracted virtual bonds, enabling efficient simulation of 2D quantum many-body physics, machine learning, and gauge theories.
• Advanced contraction methods such as BMPS, CTM, and TRG, along with canonical forms, ensure computational stability and scalability in modeling complex quantum systems.

A tensor network is a graphical and algebraic framework for efficiently parameterizing and manipulating high-dimensional arrays via networks of lower-order tensors interconnected by contracted indices. The Projected Entangled Pair States (PEPS) formalism is a central two-dimensional tensor network class, generalizing Matrix Product States (MPS) to higher spatial dimensions and providing a scalable variational ansatz for quantum many-body wavefunctions, statistical models, machine learning, and gauge theories. PEPS encode area-law entanglement and locality by arranging site tensors on the vertices of a lattice, with virtual indices contracted along edges encoding correlations. The bond dimension $D$ of the virtual indices governs the entanglement capacity and accuracy of the state.

1. Mathematical Structure of PEPS

PEPS are constructed as follows. For a square lattice of $N=L\times L$ sites, assign to each site $i$ a local tensor $$A^{[i]}_{s_i; \alpha_i, \beta_i, \gamma_i, \delta_i}$$, where $s_i$ labels the physical degree of freedom (local Hilbert space of dimension $d$) and $$\alpha_i, \beta_i, \gamma_i, \delta_i = 1,\dotsc, D$$ are virtual indices associated with the four cardinal directions. The global wavefunction is

$$|\Psi\rangle = \sum_{\{s_i\}} \left( \sum_{\{\text{bond indices}\}} \prod_{i=1}^N A^{[i]}_{s_i; \alpha_i, \beta_i, \gamma_i, \delta_i} \right) |s_1 \cdots s_N\rangle,$$

where the contraction runs over all internal virtual indices according to lattice adjacency (e.g., $$\alpha_{i,\text{right}} = \alpha_{i+\hat{x},\text{left}}$$) (Orus, 2013). PEPS ansätze extend naturally to arbitrary graphs, non-square geometries, and higher dimensions, and the structure admits generalizations to fermionic, bosonic, and gauge-symmetric settings.

In diagrammatic notation, a PEPS tensor is depicted as a node with one physical and several virtual legs, with bonds contracted along edges. Open physical legs correspond to variational degrees of freedom and observables.

2. Contraction Schemes and Computational Complexity

Exact global contraction of PEPS, i.e., evaluation of amplitudes or expectation values, is generically $\#$P-hard in two or more dimensions due to exponential scaling in $N=L\times L$0 (system linear size) (Orus, 2013). The bottleneck arises from the exponential growth of boundary MPS bond dimensions or the size of contracted "environment" tensors.

Major approximate contraction algorithms include:

• Boundary-Matrix Product State (BMPS) method: Reduces 2D networks into a sequence of MPS/MPO contractions, compressing successive rows (or columns) and truncating bond dimensions via singular value decomposition (SVD). The leading cost is $N=L\times L$1, with $N=L\times L$2 the boundary MPS bond dimension (Lubasch et al., 2013, Pang et al., 2020).
• Corner Transfer Matrix (CTM) Renormalization: Approximates the environment of a patch by iteratively absorbing rows/columns into corner and edge tensors, truncating the bond/environment dimension $N=L\times L$3 (Orus, 2013).
• Tensor Renormalization Group (TRG) and variants: Coarse-grains the 2D tensor network by SVD-based splitting and truncation (Orus, 2013).
• Cluster update (CU$N=L\times L$4) schemes: Interpolate between highly local "simple update" (neglect environment, cheap but less accurate) and full boundary-MPS/CTM environments via clustering of environment rows/columns, controlling computational cost and accuracy (Lubasch et al., 2013).

Algorithmic advances include randomized SVD for network refactorization, reshape-avoiding orthogonalization, and distributed-memory implementations that enable scaling PEPS contraction to larger bond dimensions and system sizes (Pang et al., 2020).

3. Entanglement, Injectivity, and Topological Order

A key feature of PEPS is their compliance with the area law of entanglement: the entanglement entropy $N=L\times L$5 of a region scales as $N=L\times L$6, the boundary length, with maximal $N=L\times L$7 (Orus, 2013). This makes PEPS the natural ansatz class for ground states of local gapped 2D Hamiltonians.

Injectivity plays a foundational role in the classification and physical properties of PEPS:

• Injective PEPS: The map from virtual to physical degrees of freedom is injective on some finite block; such PEPS have unique gapped ground states of local parent Hamiltonians.
• G-injective and symmetry-injective PEPS: Virtual legs transform under a symmetry group $N=L\times L$8, and tensors exhibit invariance under $N=L\times L$9 actions, enabling robust construction of topologically ordered and symmetry-enriched topological (SET) phases (Garre-Rubio, 2019).

Non-injective PEPS (including those for topological phases such as toric code or chiral phases) exhibit ground-state degeneracies and stable edge modes (Fernández-González et al., 2011, Wahl, 2015).

The "fundamental theorem" for injective PEPS establishes that any two injective tensors generating the same state are related by a local virtual gauge transformation (Garre-Rubio, 2019). This underlies classification schemes and symmetry actions on the PEPS manifold.

4. Physical and Machine Learning Applications

PEPS and their generalizations have a broad range of applications:

• Quantum Many-Body Physics: Accurate ground states for frustrated spin systems (e.g., Heisenberg antiferromagnets—energy per site within $i$0 of QMC benchmarks at $i$1 (Orus, 2013)), parent Hamiltonians for exactly solvable models (toric code, AKLT, string-nets, RVB) (Schuch et al., 2012, Orus, 2013, Wahl, 2015), chiral topological phases, and lattice gauge theories (Kelman et al., 2024, Canals et al., 2024).
• Critical Systems and Fermi Surfaces: Fermionic PEPS and Gaussian PEPS accurately capture critical phenomena, including Fermi surfaces with power-law scaling of the energy error $i$2 for exponent $i$3, where $i$4 is the typical bond dimension (Mortier et al., 2020).
• Quantum Field Theories: The cPEPS framework provides the continuum limit of lattice PEPS and enables variational encoding of relativistic field theory vacua with manifest area-law entanglement and symmetries (Shachar et al., 2021).
• Machine Learning: Two-dimensional PEPS formalism provides state-of-the-art supervised learning models for images, outperforming MPS/TTN architectures and matching MLP accuracies with significantly fewer parameters, while leveraging the locality and low-entanglement structure of natural data (Cheng et al., 2020).
• Real-Time Dynamics: Time-evolution protocols using Trotter decompositions, embedded-bond and cluster variational optimizations, and powerful environment approximations extend the reach of PEPS to non-equilibrium dynamics, including critical systems and boundary CFT characterization (Ponnaganti et al., 2023).

5. Symmetries, Gauge Constraints, and Classification

Symmetries are encoded in PEPS both at the physical and virtual level:

• 0-form global and on-site symmetries can be represented as virtual actions pulled through each tensor, giving rise to symmetry-enriched phases and order parameters (Garre-Rubio, 2019).
• 1-form symmetries and higher-form anomalies are classified by algebraic constraints on the virtual legs, projective phases, and braiding relations of virtual symmetry operators, enabling construction and optimization within constrained tangent spaces and systematic classification of anyonic excitations (Tan et al., 2024).
• Gauge Symmetries and Lattice Gauge Theories: Gauge-invariant PEPS are constructed by embedding Gauss-law constraints in symmetric vertex tensors, enabling the exact representation of gauge-invariant subspaces using symmetric tensors only and providing duality correspondences to globally symmetric spin systems (Canals et al., 2024). Fermionic and Gaussian PEPS constructions further extend the ansatz to lattice QCD and more general lattice gauge models (Kelman et al., 2024).

Classification of PEPS phases is achieved via group extension and cohomology data, symmetry fractionalization parameters, and parent Hamiltonian analysis (Garre-Rubio, 2019, Zhang et al., 2021). G-injective PEPS realize string-net and quantum double models and provide a direct route to study anyon condensation, topological transitions, and SET classification.

6. Canonical Forms, Stability, and Algorithmic Aspects

Gauge freedom on the virtual indices of PEPS is intrinsic. Canonical forms—the fixing of gauges to minimize redundancy—play a crucial role in stabilizing algorithms:

• Minimal Canonical Form (MCF): A rigorous canonicalization for arbitrary-dimensional PEPS, defined by minimal Frobenius norm up to the gauge orbit closure, ensures decidability of gauge equivalence for all contraction graphs (Acuaviva et al., 2022).
• 1D Canonical Form: For MPS, canonicalization is exact (left, center, right-orthogonal). In 2D, only partial or approximate canonical forms are achievable (e.g., column/row isometries by iterative sweeps). Approximate canonicalization reduces the condition number of the environment matrix and controls the amplification of local truncation and noise errors, with O(1) amplification in ideal cases (Zhang et al., 2020).
• Error Amplification and Conditioning: In MPS/PEPS, worst-case error amplification is governed by the singular-value spectrum of environment matrices; canonicalization suppresses amplification and yields numerically stable variational and evolution algorithms (Zhang et al., 2020).

Efficient contraction and optimization schemes combine approximate environments, Monte Carlo sampling (for Gaussian PEPS in gauge theories), and distributed-memory parallelism, with randomized SVD, to scale PEPS algorithms to large system sizes and bond dimensions (Pang et al., 2020, Kelman et al., 2024).

7. Extensions, Challenges, and Outlook

Ongoing and emerging developments in PEPS research include:

• Fermionic and Chiral PEPS: Extensions to arbitrary statistics, crystalline symmetries, and topological invariants, including Chern numbers and symmetry-protected indices (Wahl, 2015, Zhang et al., 2021).
• Continuous PEPS (cPEPS): Continuum generalizations for quantum field theory, manifesting area-law entanglement and symmetry, and yielding universal convergence rates to QFT vacua (Shachar et al., 2021).
• Gauge Theories and Real-Time Dynamics: Robust simulation of non-Abelian lattice gauge theories and quantum chromodynamics, leveraging Gaussian PEPS and variational Monte Carlo methods (Kelman et al., 2024).
• Algorithmic Advances: Scalable, stable contraction and optimization routines (randomized SVD, environment approximations, parallelization), finite-depth preparation circuits using globally symmetric tensors, and product-state initialization (Pang et al., 2020, Canals et al., 2024).
• Classification and Symmetry: Systematic understanding of higher-form symmetries, symmetry fractionalization, anyon condensation, and the full mapping between PEPS tensors, cocycle data, and physical invariants (Garre-Rubio, 2019, Tan et al., 2024).

The PEPS formalism, both as a conceptual and computational framework, remains a central pillar of modern approaches to high-dimensional quantum matter, quantum information, machine learning, and lattice gauge field theory, unifying rigorous mathematically controlled descriptions with scalable practical algorithms (Orus, 2013, Lubasch et al., 2013, Garre-Rubio, 2019, Cheng et al., 2020, Kelman et al., 2024, Canals et al., 2024).