A single-layer framework of variational tensor network states

We propose a single-layer tensor network framework for the variational determination of ground states in two-dimensional quantum lattice models. By combining the nested tensor network method [Phys. Rev. B 96, 045128 (2017)] with the automatic differentiation technique, our approach can reduce the computational cost by three orders of magnitude in bond dimension, and therefore enables highly efficient variational ground-state calculations. We demonstrate the capability of this framework through two quantum spin models: the antiferromagnetic Heisenberg model on a square lattice and the frustrated Shastry-Sutherland model. Even without GPU acceleration or symmetry implimention, we have achieved the bond dimension of 9 and obtained accurate ground-state energy and consistent order parameters compared to prior studies. In particular, we confirm the existence of an intermediate empty-plaquette valence bond solid ground state in the Shastry-Sutherland model. We have further discussed the convergence of the algorithm and its potential improvements. Our work provides a promising route for large-scale tensor network calculations of two-dimensional quantum systems.

💡 Research Summary

The paper introduces a novel single‑layer variational tensor‑network framework designed to efficiently determine ground‑state wavefunctions of two‑dimensional quantum lattice models. Traditional variational tensor‑network approaches rely on evaluating the energy functional E = ⟨Ψ|H|Ψ⟩/⟨Ψ|Ψ⟩, which requires contracting a double‑layer tensor network (the bra and ket layers stacked together). When the bond dimension D is increased, the effective bond dimension of the double‑layer network becomes D², leading to memory and computational costs that scale as D⁸ and D¹² respectively for the widely used corner‑transfer‑matrix renormalization group (CTMRG) contraction. These prohibitive costs become a severe bottleneck for gradient‑based optimization with automatic differentiation (AD), because AD must store all intermediate tensors to compute ∂E/∂X, where X collects all variational parameters.

To overcome this limitation, the authors adopt the nested tensor‑network (NTN) method, originally proposed to reduce the cost of expectation‑value calculations. The key insight is to treat the physical indices σ as ordinary link indices that connect the bra and ket tensors within a single layer, rather than summing them out to form a reduced double‑layer tensor. By inserting auxiliary δ‑function tensors X (at every crossing of virtual bonds) and composite tensors Y (which also carry the physical index), the double‑layer network is re‑expressed as a single‑layer network whose effective bond dimension is only D·d (where d is the physical dimension, typically 2 for spin‑½). Consequently, the CTMRG contraction now scales as D⁶ in memory and D⁹ in computation, a reduction of three to four orders of magnitude compared with the original formulation.

The authors integrate this NTN representation with modern AD frameworks (PyTorch and Zygote). The variational parameters X are flattened tensors of all local PEPS tensors. The energy E(X) is computed by contracting the single‑layer NTN using CTMRG; AD automatically builds the computational graph, enabling the gradient ∂E/∂X to be obtained without explicit back‑propagation code. With the gradient in hand, standard first‑order optimizers such as L‑BFGS or Adam are employed to update X iteratively until the energy converges. This loop—(i) initialize a random PEPS, (ii) evaluate E and its gradient via AD‑enhanced NTN‑CTMRG, (iii) perform a gradient step, (iv) repeat—constitutes the complete variational algorithm.

The method is benchmarked on two paradigmatic models on infinite 2D lattices: (1) the square‑lattice spin‑½ antiferromagnetic Heisenberg model, and (2) the frustrated Shastry‑Sutherland model. Despite using no GPU acceleration or explicit symmetry exploitation, the authors achieve a bond dimension D = 9, which is substantially larger than what previous variational PEPS studies could reach without symmetry. For the Heisenberg model they obtain an energy per site of −0.66944, in excellent agreement with high‑precision quantum Monte Carlo and previous tensor‑network results, and they reproduce the expected Néel order parameter. For the Shastry‑Sutherland model they map out the phase diagram as a function of the frustration ratio J/J′, confirming the presence of an intermediate empty‑plaquette valence‑bond‑solid (VBS) phase around J/J′ ≈ 0.7–0.8. The energy and order‑parameter convergence are smoother than in earlier studies, demonstrating the stability of the gradient‑based optimization when combined with the NTN reduction.

The paper also provides a detailed analysis of computational resources. Memory consumption drops from O(D⁸) to O(D⁶), and wall‑clock time per gradient evaluation reduces by roughly two orders of magnitude, enabling the authors to perform hundreds of optimization steps at D = 9 on a standard CPU workstation. They discuss potential extensions: incorporating global symmetries (U(1), SU(2)) directly into the tensors would further reduce the effective bond dimension; adaptive selection of the CTMRG environment dimension could improve accuracy‑cost balance; and parallel or multi‑GPU implementations could push the feasible bond dimension beyond D = 12.

In summary, the work presents a powerful and practical framework that merges the nested single‑layer tensor‑network representation with automatic differentiation. By dramatically lowering the scaling of both memory and computation, it makes high‑bond‑dimension variational PEPS calculations accessible without specialized hardware or symmetry tricks. This opens the door to large‑scale, accurate studies of strongly correlated 2D quantum systems, including frustrated magnets, spin liquids, and topologically ordered phases, where conventional tensor‑network methods have been limited by prohibitive computational costs.