Bilayer Tensor Network Analysis

• Bilayer tensor networks are structured with two layers that capture both intra- and interlayer quantum correlations, facilitating practical simulations.
• They utilize ansätze like iPEPS, PESS, and LTRG++ to reduce contraction complexity while accurately representing quantum states.
• These networks have been applied to study layered superconductors, quantum spin liquids, and frustrated magnets, validating rigorous algebraic and computational benchmarks.

A bilayer tensor network refers to a class of tensor network states or contraction schemes structured around two physical or “virtual” layers. This architecture is central to modeling quantum lattice systems with two coupled planes (as in bilayer quantum magnets and superconductors), simulating the contraction of expectation values (where “bra” and “ket” layers are involved), or optimizing algorithms for evaluating thermal and ground-state properties. Bilayer tensor networks encompass a variety of formalisms, including bilayer PEPS (Projected Entangled Pair States), bilayer PESS (Projected Entangled Simplex States), bilayer LTRG++ (Linearized Tensor Renormalization Group), and associated contraction strategies, each tailored to leveraging the layered topology for improved representational power or computational efficiency.

1. Foundational Bilayer Tensor Network Structures

2×N Grid PEPS: Algebraic-Geometric Foundations

A canonical bilayer tensor network is the 2×N Grid PEPS, where the underlying graph is a rectangle with two rows and N columns. Each vertex (site) in the grid has four entanglement edges (horizontal and vertical) and a single physical index. The local tensor at each vertex is of the form: $$T_v \in A_v \otimes B_v \otimes C_v \otimes D_v \otimes E_v$$ where $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$ (bond spaces of dimension $D$), and $B_v \cong \mathbb{C}^k$ (physical index). The global state maps the set of local tensors, contracted along all edges, into the full physical Hilbert space $W \cong (\mathbb{C}^k)^{\otimes 2N}$.

Algebraic-geometric methods are employed to characterize the resulting manifold of physical states, including dimension counts, Zariski closure considerations (especially with periodic boundary conditions), and border-rank bounds obtained via matrix flattenings. For the 2×N PEPS, the “expected” dimension (from parameter count) far exceeds the actual dimension due to contraction-induced degeneracies. Explicit constructions yield border rank

$\underline{R}(T_{\rm gen}) \geq 2^N$

for generic tensors at $D=k=2$, realizing exponentially-entangled states with modest bond dimension (Sarin, 2018).

2. Bilayer Tensor Network Ansätze for Layered Materials

iPEPS and Bilayer Architectures

For layered quantum systems (e.g. two coupled 2D lattices), bilayer iPEPS efficiently captures intra- and interlayer correlations. The ansatz places a 5-index tensor on each site in both layers, distinguishing four in-plane virtual legs (dimension $D_{xy}$) and an interlayer leg (dimension $D_\perp$): $$A^{[\ell]}_{s, u, d, l, r, \alpha}, \quad \ell = 1,2$$ with $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$0 the physical index and $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$1 the interlayer bond. The contraction of these networks approximates the full 3D norm tensor while retaining tractable complexity by separating intralayer environments (using 2D CTM) and reducing interlayer coupling to manageable effective 1D MPOs of bond dimension $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$2 (Vlaar et al., 2022).

Bilayer PESS in Frustrated Magnets

In complex geometries such as bilayer Kagome lattices, the bilayer PESS formalism represents each vertical pair as a “supersite” and uses simplex tensors ($A_v, C_v, D_v, E_v \cong \mathbb{C}^D$3) to encode triangle entanglement, while physical tensors $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$4 reflect the full two-spin Hilbert space. This captures both in-plane simplex correlations and vertical entanglement, facilitating accurate simulation of quantum spin liquids and other strongly correlated phases (Kshetrimayum et al., 2019).

3. Bilayer Contraction Schemes and Algorithmic Advances

Double-Layer Contractions and Optimizations

Expectation values in tensor network states (PEPS, PESS) often require contracting a double-layer network. The naïve contraction merges ket and bra layers, leading to local bond dimensions $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$5 and prohibitive costs ($A_v, C_v, D_v, E_v \cong \mathbb{C}^D$6 time, $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$7 memory). Xie et al. introduced an optimized mapping where the two layers are shifted and nested into a single effective layer with swap gates wherever virtual legs cross, so all local bonds remain at dimension $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$8. This “NTN” mapping reduces cost to $A_v, C_v, D_v, E_v \cong \mathbb{C}^D$9 time and $D$0 storage, effectively doubling accessible bond dimension in practical computations. The NTN scheme is lattice-agnostic, applies to any PEPS/PESS, and can exploit global symmetries for further gains (Xie et al., 2017).

Bilayer LTRG++ for Thermal Networks

LTRG++ generalizes the one-dimensional linearized thermal renormalization group to a bilayer scheme for contracting 1+1D thermal tensor networks. The central idea is to construct MPOs for $D$1 and its Hermitian conjugate, then contract their product. This halves the number of contraction sweeps, exploits canonical forms for cost-efficient contraction ($D$2 per bond vs. $D$3), and improves accuracy by two orders of magnitude over single-layer LTRG for fixed bond dimension $D$4. In the infinite system limit, LTRG++ is algebraically equivalent to transfer-matrix DMRG (TMRG), providing a unified tensor network language for finite-temperature calculations (Dong et al., 2016).

4. Physical Applications: Bilayer Quantum Materials

Superconductivity in Bilayer Models

Bilayer tensor networks underpin the analysis of unconventional superconductivity in layered compounds. For example, the bilayer $D$5–$D$6–$D$7 model for La$D$8Ni$D$9O$B_v \cong \mathbb{C}^k$0 (pressurized nickelate superconductor) demonstrates that pure magnetic rung (interlayer) exchange ($B_v \cong \mathbb{C}^k$1) stabilizes a robust $B_v \cong \mathbb{C}^k$2-wave, rung-singlet SC order in the absence of single-particle interlayer tunneling. Tensor-network simulations (DMRG, iPEPS, T-TTRG) capture algebraic decay of pairing correlations (Luttinger exponent $B_v \cong \mathbb{C}^k$3), a strong finite order parameter ($B_v \cong \mathbb{C}^k$4), and a superconducting scale $B_v \cong \mathbb{C}^k$5. The mechanism is purely magnetic and parallels recent cold-atom experiments on mixD bilayer Fermi lattices (Qu et al., 2023).

Quantum Spin Liquids in Bilayer Kagome

Tensor-network studies of Ca$B_v \cong \mathbb{C}^k$6Cr$B_v \cong \mathbb{C}^k$7O$B_v \cong \mathbb{C}^k$8 utilize the bilayer PESS formalism to confirm a gapless quantum spin-liquid ground state, reproducing experimental susceptibilities, gapless heat capacity, and the absence of magnetic ordering. The bilayer PESS ansatz, encoding both vertical (interlayer) and in-plane entanglement, quantitatively matches neutron scattering and thermodynamic measurements and demonstrates the capacity of bilayer tensor networks to model real quantum materials (Kshetrimayum et al., 2019).

5. Scaling, Complexity, and Benchmarking

The complexity of bilayer tensor networks is determined by the bond dimensions ($B_v \cong \mathbb{C}^k$9), environment cutoff ($W \cong (\mathbb{C}^k)^{\otimes 2N}$0), and the chosen contraction/optimization algorithm.

• In iPEPS bilayer contraction, CTM sweeps per layer have cost $W \cong (\mathbb{C}^k)^{\otimes 2N}$1 (intralayer), assembly and eigenvector calculations for the 1D bilayer MPO scale as $W \cong (\mathbb{C}^k)^{\otimes 2N}$2.
• Benchmarks for bilayer Heisenberg models indicate that with $W \cong (\mathbb{C}^k)^{\otimes 2N}$3, $W \cong (\mathbb{C}^k)^{\otimes 2N}$4, $W \cong (\mathbb{C}^k)^{\otimes 2N}$5, and $W \cong (\mathbb{C}^k)^{\otimes 2N}$6, the relative error in energy is $W \cong (\mathbb{C}^k)^{\otimes 2N}$7, and $W \cong (\mathbb{C}^k)^{\otimes 2N}$8 for staggered magnetization, using only a fraction (1/5 to 1/20) of the computational cost compared to full 3D contraction (Vlaar et al., 2022).
• In bilayer LTRG++ for 1D systems, computational cost remains $W \cong (\mathbb{C}^k)^{\otimes 2N}$9, but accuracy and efficiency are dramatically improved, and the scaling matches TMRG in the infinite limit (Dong et al., 2016).
• NTN contraction for expectation values is 3–5 orders of magnitude cheaper than double-layer RTN contraction for attainable bond dimensions (Xie et al., 2017).

6. Open Problems and Future Directions

Several central algebraic and computational questions remain open:

• For general $\underline{R}(T_{\rm gen}) \geq 2^N$0 PEPS (with $\underline{R}(T_{\rm gen}) \geq 2^N$1), the injectivity/maximal rank of flattenings and the explicit algebraic description of the image variety are unresolved, with current results restricted to the bilayer ($\underline{R}(T_{\rm gen}) \geq 2^N$2) case (Sarin, 2018).
• The Zariski closure of the bilayer PEPS variety (especially under periodic boundary conditions) remains uncharacterized and involves deep connections to the geometry of secant varieties, reminiscent of matrix multiplication complexities.
• In contraction schemes, whether it is possible to construct explicit bilayer tensors that saturate the quantum max-flow/min-cut bound for all bipartitions, and whether higher-order algebraic flattenings can improve the known lower bounds on border rank, is unknown.
• Algorithmically, while NTN, bilayer LTRG++, and bilayer-LCTM have dramatically improved attainable bond dimensions, hardware limitations on environment size ($\underline{R}(T_{\rm gen}) \geq 2^N$3) and bond-dimension growth continue to set practical upper bounds. Exploiting global symmetries (e.g., U(1), SU(2)), further algorithmic parallelization, and intelligent truncation strategies are active research directions.

7. Impact and Perspectives in Quantum Many-Body Theory

Bilayer tensor networks have established themselves as a vital framework in both analytical and computational quantum many-body research. They realize states with exponential entanglement scaling with modest bond dimension, enable computationally tractable contraction and optimization for systems previously out of reach, and provide accurate descriptions of complex quantum phases—ranging from quantum spin liquids to unconventional superconductors. The systematic study of their algebraic properties bridges quantum information, condensed matter, and algebraic geometry, offering ongoing insights into the expressive power and computational complexity of tensor network state spaces (Sarin, 2018, Xie et al., 2017, Dong et al., 2016, Vlaar et al., 2022, Qu et al., 2023, Kshetrimayum et al., 2019).