PaperlayerView exact revision on arXiv

Source-linked AI summary

The Tensor Networks Anthology: Simulation techniques for many-body quantum lattice systems

Pietro Silvi, Ferdinand Tschirsich, Matthias Gerster, Johannes Jünemann, Daniel Jaschke, Matteo Rizzi, Simone Montangero

arXiv:1710.03733v2quant-ph

TL;DR

The anthology addresses the computational challenge of simulating many-body quantum systems by presenting tensor-network methods and numerical tools. It describes data structures, symmetry-aware algorithms, and loop-free-network methods for low-energy states, presenting a toolkit for many low-entanglement, finite-size one-dimensional simulations.

  • Problem

    The configuration space of quantum many-body problems grows exponentially with system size, restricting exact diagonalization to small systems.

    Source: §1 · p. 4

  • Method

    The anthology presents tensor-network simulation methods alongside their numerical data structures, operations, symmetry techniques, and ground-state optimization strategies.

    Source: §1 · pp. 4–5, §6 · p. 103

  • Results

    The resulting toolkit can address many low-entanglement scenarios for one-dimensional systems, including common finite-size simulations.

    Source: §6 · p. 103

  • Takeaways & Limitations

    The anthology serves as both a programmer’s manual and a review of tensor-network methods, and can support a technical introductory course.

    Source: §1 · p. 5

    In symmetric single-tensor updates, local optimizations preserve link representations, so sector quantum numbers and degeneracies cannot be handled variationally.

    Source: §5.1.3 · p. 85

Abstract

from arXiv · show

We present a compendium of numerical simulation techniques, based on tensor network methods, aiming to address problems of many-body quantum mechanics on a classical computer. The core setting of this anthology are lattice problems in low spatial dimension at finite size, a physical scenario where tensor network methods, both Density Matrix Renormalization Group and beyond, have long proven to be winning strategies. Here we explore in detail the numerical frameworks and methods employed to deal with low-dimension physical setups, from a computational physics perspective. We focus on symmetries and closed-system simulations in arbitrary boundary conditions, while discussing the numerical data structures and linear algebra manipulation routines involved, which form the core libraries of any tensor network code. At a higher level, we put the spotlight on loop-free network geometries, discussing their advantages, and presenting in detail algorithms to simulate low-energy equilibrium states. Accompanied by discussions of data structures, numerical techniques and performance, this anthology serves as a programmer's companion, as well as a self-contained introduction and review of the basic and selected advanced concepts in tensor networks, including examples of their applications.

1 Introduction

This section presents tensor networks as entanglement-controlled representations for many-body problems beyond exact diagonalization and introduces them as a computationally focused framework. It covers gauge transformations that preserve the represented state and runtime gauge adaptation in DMRG/MPS to support useful numerical manipulations and improve speed and precision.

1 Introduction

  • The anthology combines a review of tensor-network simulation methods with a programmer’s manual, highlighting a symmetric-compatible single-tensor ground-state update with subspace expansion.It focuses on numerical data structures and algebraic operations rather than an extended review of physical results or a general pedagogical introduction.Source: §1 · pp. 4–5

1.1 Structure of the Anthology

  • The anthology progresses from tensors and linear algebra through symmetries and loop-free networks to generalized DMRG and ground-state search, with symmetry material optional.Its modular hierarchy is intended to map onto a tensor-network programming library, whose ground-state component can be replaced by tasks such as real-time dynamics.Source: §1.1 · p. 5, Figure 1 · p. 6

1.2 Short History of Tensor Networks for Quantum Many-Body Problems

  • DMRG was recast as optimization over Matrix Product States, whose ability to capture low-entanglement one-dimensional ground states under area laws explains its success.The MPS reformulation enabled extensions including out-of-equilibrium evolution, thermodynamic-limit and finite-temperature methods, and access to spectral properties.Source: §1.2 · pp. 5, 7

1.3 Tensor Network States in a Nutshell

  • Tensor networks represent a many-body wavefunction by contracting smaller tensors, with bounded local connectivity and auxiliary dimensions enabling computational efficiency when target states have low entanglement.The ansatz directly controls entanglement, and its bounds depend on network geometry; real-space representations can exploit area laws.Source: §1.3 · p. 8

1.4 Many-Body Statistics

  • Tensor-network ansätze flexibly accommodate different local degrees of freedom and quantum statistics, including bosons and fermions; short-range parity-preserving 1D models suit Jordan–Wigner mappings, while direct exchange-statistics rules are useful in 2D.Source: §1.4 · pp. 8–9

1.5 Typically Addressed Problems

  • Typical 1D applications include low-energy state search, unitary dynamics, thermal-state simulation, and open-system evolution, with dynamics most effective when entanglement growth is small.The anthology treats finite systems, using extrapolation for thermodynamic-limit properties rather than directly addressing infinite homogeneous networks.Source: §1.5 · pp. 9–10, §1.6 · p. 10

1.6 Accessible Information in Tensor Networks

  • Efficiently contractible networks support calculations of norms, local observables, correlations, operator expectation values, and entanglement properties, while contractibility is not general for networks such as PEPS.Loop-free networks have efficient contractibility when link dimensions are bounded; suitable gauge choices can further reduce norm-calculation costs.Source: §1.6 · pp. 10–11

1.7 The Tensor Network Gauge

  • Tensor-network gauge transformations locally change auxiliary-link representations without changing the represented physical state or its physical quantities.The anthology introduces this gauge freedom for use throughout its later discussion.Source: §1.7 · p. 11
  • A matrix and its inverse can be inserted across a single auxiliary link without changing the network geometry or the quantum state it represents.The transformation leaves invariant the composite tensor formed by contracting the neighboring tensors across that link.Source: §1.7 · p. 11, Figure 3 · p. 12
  • Combining single-link gauge transformations preserves the many-body state and can be performed efficiently; runtime gauge adaptation in DMRG/MPS can improve speed and precision.The section previews reviews of linear-algebra implementations, symmetric-tensor counterparts, and the advantages of gauge conditions.Source: §1.7 · p. 12

2 Tensors and Tensor Operations

This section introduces tensors as the data structures underlying tensor networks and the operations used to build and execute tensor-network algorithms. It covers tensor contractions, including efficient contraction strategies, and decompositions into linked factors that reconstruct the input and can yield smaller factors or useful isometries.

2.1 Tensors

  • A tensor is a finite-dimensional complex array whose links index its elements; zero-, one-, and two-link tensors are scalars, vectors, and matrices, respectively.Links also provide a graphical representation of tensors as network edges, while tensors are represented as nodes.Source: §2.1 · p. 13

2.2 Basic Tensor Operations

  • Tensor operations divide into memory-and-shape manipulations, generally requiring no FLOPs, and linear-algebra operations, which require them.Because basic tensor operations account for much of algorithm runtime, implementations should optimize them and can use libraries such as BLAS or LAPACK.Source: §2.2 · p. 15

2.2.1 Tensor Manipulations

  • Permutation, fusion, splitting, and index remapping reshape tensor links, enabling tensors to be reorganized or converted to matrices without necessarily moving data.Permutation is linear in the number of elements, while fusion and splitting can take O(1); splitting may require a subsequent permutation, and compression or inflation changes link dimensions.Source: §2.2.1 · pp. 15, 17, 18

2.2.2 Tensor Linear Algebra Operations

  • Contractions generalize matrix multiplication: permuting and fusing links casts them as matrix products, after which remaining links can be split and reordered.This enables optimized multiplication and parallelization, with permutation overhead to consider; partial traces sum over paired indices, and tensor norms contract a tensor with its complex conjugate.Source: §2.2.2 · pp. 18, 21
  • Contractions often dominate computational cost, and network geometries with known optimal contraction strategies can address this challenge.The search for efficient contraction strategies is generally difficult.Source: §2.2.2 · p. 21
  • Tensor decomposition replaces one tensor with linked factors whose contraction restores it within a chosen precision, often yielding smaller factors or useful isometries.A simple decomposition reverses a contraction; specialized decompositions aim to provide favorable factor sizes or isometries for larger networks.Source: §2.2.2 · p. 22
  • After bipartitioning and fusing tensor links, QR yields Q and upper-trapezoidal R, with Q a left isometry and the intermediate dimension set by the smaller fused dimension.The fused links are split again to restore the outer network geometry.Source: §2.2.2 · p. 22
  • SVD singular values reveal entanglement and enable link compression: removing zero or negligible values reduces dimension, while discarded values can bound truncation error and the factors remain isometric.Discarding negligible singular values approximates the tensor and can improve tensor-network algorithm performance; approximate factorization methods can control stochastic and truncation errors.Source: §2.2.2 · p. 25
  • Square-matrix diagonalization uses exact or iterative methods, while exponentiation uses Taylor/Padé expansion or EXPOKIT; combining tensor operations supports local-operator exponentiation in TEBD.The tensor-network approach can form an exact exponential through a full tensor and matrix, or represent a Taylor expansion using smaller tensors; numerical stability conditions apply.Source: §2.2.2 · pp. 24–25, Figure 11 · p. 24

3 Abelian Symmetric Tensors and Operations

Symmetries can reduce quantum many-body problems to simpler decoupled sectors, benefiting both computation and targeted variational studies. The section reviews Abelian symmetries and how their information is incorporated into tensor data structures, networks, and operations.

3 Abelian Symmetric Tensors and Operations

  • Symmetries block-diagonalize the Hamiltonian into simpler problems by separating states according to conserved quantities.In quantum systems, the conserved quantities correspond to operators commuting with the Hamiltonian.Source: §3 · p. 26
  • Symmetry-aware numerical methods can speed up computation and directly target states with a chosen quantum number.This canonical-ensemble approach is more direct and robust than selecting an average particle number by tuning chemical potential across multiple sectors.Source: §3 · p. 26
  • The section develops Abelian symmetry tools, from quantum groups and tensor encodings to symmetry-organized data structures, networks, and operations.It also provides examples of Abelian groups encountered in typical problems.Source: §3 · p. 26

3.1 Quantum Symmetries in a Nutshell: Abelian Groups

Quantum symmetries are Hamiltonian-commuting group representations that organize energy eigenstates into irreducible-representation sectors. The section summarizes Abelian representations and common groups, emphasizing their simpler tensor-network fusion rules and options for reducing non-Abelian symmetries.

3.1 Quantum Symmetries in a Nutshell: Abelian Groups

  • A symmetry is a unitary group representation commuting with the Hamiltonian, so energy eigenstates can be chosen within irreducible-representation subspaces.A sector or charge labels the irrep; within-sector states share an energy independent of their secondary quantum number by Schur’s lemma.Source: §3.1 · p. 26
  • Abelian symmetries have one-dimensional irreps and simple fusion rules, while non-Abelian tensor networks require tracking Clebsch-Gordan coefficients.The anthology focuses on Abelian groups; a non-Abelian model can instead use an Abelian subgroup, with unpredicted energy degeneracies expected.Source: §3.1 · pp. 26–27
  • Irrep inversion maps a phase to its negative, distributes over fusion, and the identical irrep is the self-adjoint zero label.Abelian Hamiltonians are block-diagonal by irrep sector without additional constraints.Source: §3.1 · p. 27
  • Z_n symmetries describe cyclic rotations, including Z2 parity and the Ising model, while U(1) covers integer conservation such as particle number.The spinless Bose-Hubbard model is another U(1) example; U(1) fusion adds integer labels.Source: §3.1 · p. 27
  • The integer group Z describes conservation of an interval-valued quantity, with discrete translations on an infinite one-dimensional lattice as an example.Its irrep labels lie in [0, 2π), and fusion adds labels modulo 2π.Source: §3.1 · pp. 27–28
  • Independent Abelian symmetries combine by direct product, with composite irrep labels formed from tuples that fuse independently.The XYZ model for an even number of sites has Z2 × Z2 symmetry, also called D2.Source: §3.1 · p. 28
  • Common non-Abelian symmetries can be reduced to Abelian subgroups, such as D_n to Z_n or Fermi-Hubbard SO(4) to U(1) × U(1).The latter keeps particle-number conservation for each spin species.Source: §3.1 · p. 28

3.2 Encodable Symmetries in Tensor Networks

Tensor networks can efficiently encode pointwise symmetries, whose transformations act independently on each degree of freedom. The section distinguishes global and lattice-gauge cases and sets aside other symmetry classes that lack a general treatment here.

3.2 Encodable Symmetries in Tensor Networks

  • Pointwise symmetries act independently at each site through local representations, making them particularly suitable for tensor-network implementation.The local representation may vary explicitly by site.Source: §3.2 · p. 28
  • Global pointwise symmetries act across every site, commonly expressing conservation of total particle number or total spin magnetization.Their representations are usually homogeneous across sites.Source: §3.2 · pp. 28–29
  • Lattice gauge symmetries have restricted supports, often neighboring pairs or plaquettes, and their extensive number can make them computationally useful.Other symmetries, including translation, lattice rotation/reflection, and chiral symmetry, lack a general treatment in this anthology.Source: §3.2 · p. 29
  • The tensor-network construction targets many-body states with well-defined quantum numbers using invariant tensors and symmetric links.The discussion primarily follows Ref., which specializes the construction of Ref. to Abelian groups.Source: §3.2 · p. 29

3.3 Constructing Symmetric Tensors

The section organizes symmetric tensors by assigning group representations to links and requiring tensors to be invariant under their linked representations. This yields a block structure and data-handling scheme for symmetry-aware tensor networks.

3.3.1 Abelian Symmetric Links

  • A symmetric link index is the pair of a quantum number and a degeneracy label, with indices of the same quantum number grouped into a degeneracy space.The associated group representation acts diagonally, with each quantum number's irrep repeated according to its degeneracy.Source: §3.3.1 · pp. 29–30

3.3.2 Abelian Symmetric Tensors

  • A tensor is invariant when the linked representations leave it unchanged; incoming links use inverted representations, so transformations cancel on internal network links.Internal links run out of one tensor and into another, giving W(g)W†(g) = 1.Source: §3.3.2 · pp. 30–31

3.3.3 Inner Structure of Abelian Symmetric Tensors

  • Only sets of link quantum numbers that fuse to the identical irrep contribute, and each matching set has a degeneracy tensor over its degeneracy spaces.A structural tensor selects the matching sets, while the degeneracy tensors hold their entries.Source: §3.3.3 · p. 31
  • The maximal number of nonzero tensor elements is limited to those in degeneracy tensors, often far fewer than in a nonsymmetric counterpart.Only matching quantum-number sets contribute degeneracy tensors.Source: §3.3.3 · pp. 31–33
  • A selector link with a single non-degenerate quantum number turns an invariant tensor into a covariant tensor targeting that sector.The selector representation determines the physical state's sector.Source: §3.3.3 · p. 33, §3.3.4 · p. 35

3.3.4 Symmetric Links and Tensor Data Handling

  • A symmetric tensor object stores its links, link directions, present quantum-number matches, and a degeneracy tensor for each present match.Matches whose degeneracy tensor vanishes need not be listed.Source: §3.3.4 · p. 33

3.4 Symmetric Tensor Networks

For a symmetric Hamiltonian, distinct symmetry sectors are dynamically decoupled, so a simulation can target one sector. The section constructs invariant or covariant tensor-network states from symmetric tensors, using a selector link for the covariant case.

3.4 Symmetric Tensor Networks

  • A symmetric Hamiltonian dynamically decouples different symmetry sectors, allowing a tensor-network simulation to restrict to one selected sector.This restriction is stated to hold without loss of generality.Source: §3.4 · p. 35
  • If every tensor is invariant, the resulting tensor-network state is invariant; conversely, an arbitrary invariant state can be decomposed this way, though link dimensions may grow for networks with loops.The construction is stated for invariant states, with possible dimension growth when the network is not loop-free.Source: §3.4 · p. 35
  • A covariant state belongs to one selected irrep and can be represented by invariant tensors with a single selector link fixed to that irrep.The selector link has dimension one and can be attached to any one tensor in the network.Source: §3.4 · pp. 35–36

3.5 Symmetric Tensor Operations

This section develops operations on symmetric tensors and links that preserve quantum-number and degeneracy structure, covering subtensor manipulation, fusion and splitting, contraction, and decomposition. It describes block-wise linear algebra using degeneracy blocks and discusses how computational savings depend on the operation and degeneracy distribution.

3.5.1 Link and Tensor Manipulations

  • Symmetric tensors can be initialized from user elements, zeros, random values, or an identity matrix, with matches and degeneracy tensors set according to the chosen construction.User elements require fusion-rule checks; zero initialization has no present matches or allocated degeneracy tensors, and the identity uses opposite link directions with matching quantum numbers and identity blocks.Source: §3.5.1 · p. 36
  • Link inversion maps quantum numbers to their inverses, whereas tensor Hermitian conjugation flips link directions without inverting representations and complex-conjugates degeneracy tensors.Inverting a link within a tensor also flips its direction and updates the match while preserving the fusion rule; degeneracy tensors remain unchanged.Source: §3.5.1 · p. 38
  • Subtensor access and readout address elements or degeneracy blocks by quantum numbers and indices, while readout extracts block-wise ordinary subtensors from retained matches.Assignment handles shared matches by ordinary subtensor assignment, zeros matches present only in the destination, and adds matches present only in the source after initializing their degeneracy tensors to zero.Source: §3.5.1 · p. 40
  • Symmetric link fusion maps tuples of quantum numbers and degeneracy indices bijectively to fused indices, combining degeneracies from original links and colliding quantum-number combinations.Concatenating partial degeneracies lets consecutive links be fused or split by moving contiguous memory sections without reordering degeneracy-tensor elements; the resulting operations cost at most linear in the number of symmetric tensor elements.Source: §3.5.1 · p. 41
  • A fuse-node stores the mapping between original and fused links; specialized storage avoids inefficient tensor contraction, and node permutation rearranges mapping data rather than fused tensor links.Contracting a fuse-node creates the fused tensor, while its direction-inverted adjoint splits it; both operations reduce to ordinary fusion or splitting of degeneracy tensors.Source: §3.5.1 · p. 42
  • Fusion combines symmetric links by quantum number and degeneracy offsets, while splitting restores the original links and may leave zero degeneracy tensors for cleanup.A split can use a fuse-node created elsewhere; immediately fusing and splitting the same links restores the tensor, apart from possible zero subtensors.Source: Figure 21 · p. 45, §3.5.1 · pp. 45–46
  • Fused symmetric tensors form block-diagonal matrices, but absent sector matches may require explicit identity blocks for matrix operations such as exponentiation.Link clean-up removes redundant indices by intersecting a link's representation with the fusions of the other links at both connected tensors.Source: §3.5.1 · pp. 45–47

3.5.2 Symmetric Tensor Linear Algebra Operations

  • Symmetric contraction requires shared links to match in quantum numbers and degeneracy dimensions with opposite directions, then contracts matching degeneracy tensors and combines repeated output matches.The specialized procedure avoids introducing blocks for matches absent from the input tensors, unlike a full fuse-multiply-split approach.Source: §3.5.2 · pp. 47–48
  • For peaked degeneracies, contraction cost is governed by the largest block contraction; the maximum speedup over a full nonsymmetric contraction scales with the product of the sector counts on uncontracted links.Avoiding contractions of absent degeneracy tensors can provide a larger practical gain than optimized matrix multiplication on larger blocks.Source: §3.5.2 · pp. 48–49
  • Symmetric SVD fuses a tensor into a block matrix, decomposes each present block independently, connects the factors through an intermediate symmetric link, and restores the original links.The decomposition yields symmetric factors and, for SVD, singular values; truncation may require computing values across all blocks before selecting which to discard.Source: §3.5.2 · p. 49, Figure 23 · p. 50
  • When truncating SVD results, quasi-degenerate singular values across sectors should be kept or discarded together because they may form a multiplet of an underlying non-Abelian symmetry.Source: §3.5.2 · pp. 49, 51
  • For QR and singular-value decompositions, independent block processing can yield quadratic gains in active-sector count and enables natural parallelization.The advantage comes from replacing one large decomposition with operations on multiple smaller blocks.Source: §3.5.2 · p. 51

3.5.3 Computational Cost Estimation

  • Symmetric operations are often cheaper than ordinary ones when quantum-number bookkeeping adds negligible overhead, because symmetry restricts allowed blocks and nonlinear operations benefit most from smaller blocks.The cost depends on the dimensions and number of degeneracy tensors involved.Source: §3.5.3 · pp. 51–52
  • Flat degeneracy distributions provide the greatest savings: they minimize operation costs, and symmetric-tensor memory use is reduced by a factor of the maximum number of sectors.In this distribution, many small degeneracy tensors contribute to the reduction.Source: §3.5.3 · p. 53
  • For narrowly peaked distributions, symmetric-operation costs approach those of ordinary operations, while benefits generally grow as link indices spread across more quantum numbers.A dominant degeneracy dimension can eliminate the advantage of exploiting symmetry.Source: §3.5.3 · pp. 52–53

4 Loop-Free Tensor Networks

This section explains how loop-free tensor-network structure and gauge freedom support simpler contractions, isometric and Schmidt-decomposition forms, and local compression and gauge operations. It also covers encoding symmetry without increasing bond dimensions and preparing symmetric ansatz states, including a general randomized construction.

4 Loop-Free Tensor Networks

  • Smart gauging imposes contraction rules that can eliminate many explicit contractions, reducing the numerical work for loop-free networks.Figure 25 illustrates this simplification for calculating a local observable.Source: §4 · p. 55, Figure 25 · p. 56

4.1 General Aspects of Loop-Free Tensor Networks

  • Each virtual link defines a bipartition whose Schmidt rank is bounded by its bond dimension; discarding the smallest Schmidt values is optimal for single-link compression.Simultaneous compression on multiple bonds is not fidelity-optimal, and loop-free algorithm complexity is polynomial in the maximal bond dimension D.Source: §4.1 · pp. 57–58

4.2 Gauges in Loop-Free Tensor Networks

  • Center gauges install tensor isometries that simplify contractions, local error estimates, operator application, measurements, and normalization during variational optimization.The canonical gauge also interprets the ansatz as simultaneous, independent, truncated Schmidt decompositions across network-induced bipartitions.Source: §4.2 · p. 58
  • Loop-free TNs can be transformed into symmetric tensors without increasing bond dimensions, while retaining unitary and canonical gauges.The symmetry gauge remains protected under algorithms formulated with symmetric tensor operations.Source: §4 · p. 55, §4.2 · p. 58

4.2.1 Gauge Freedom

  • Link-local gauge transformations connect equivalent loop-free representations and can reduce every bond to its Schmidt-rank minimum.The gauge matrices change the tensors without modifying the encoded state.Source: §4.2.1 · p. 59

4.2.2 Unitary Gauge

  • QR decompositions install the unitary gauge from outer nodes inward, producing isometries and potentially reducing bond dimensions; moving its center costs O(dist(c1, c2)).The procedure preserves the physical state.Source: §4.2.2 · pp. 60, 62

4.2.3 Canonical Gauge

  • SVDs convert unitary to canonical gauge, yielding simultaneous Schmidt decompositions and minimal bond dimensions on every link.Stored Schmidt values enable local re-isometrization, but inverting small values can cause numerical instability.Source: §4.2.3 · pp. 62, 64
  • For local unitary evolution, the canonical gauge avoids costly reinstatement because unitary operations preserve the Schmidt decomposition except on the bonds they act on.The same can hold approximately for near-unitaries or identities, including virtual-link compression by truncated SVDs that locally restore the gauge.Source: §4.2.3 · p. 66

4.2.4 Symmetry Gauge

  • The symmetry gauge is theoretically installable by changing the physical basis, contracting and symmetrically decomposing the state, then canonicalizing; local basis changes preserve bond dimensions.The resulting symmetric network can attain the same minimal bond dimension as a non-symmetric one.Source: §4.2.4 · p. 66
  • In loopy networks, the canonical gauge is unavailable and loop constraints can increase bond dimensions, so symmetric-tensor speedups may come with a bond-dimension trade-off.The symmetric tensor ansatz can still be beneficial for loopy geometries.Source: §4.2.4 · p. 66

4.2.5 Real Gauge

  • The real gauge makes tensor elements real and, when installable for both state and operators, can speed computation and halve the variational space in diagonalization with time-reversal symmetry.A local basis transformation can often combine real and symmetry gauges when time-reversal and global pointwise Abelian symmetries are present.Source: §4.2.5 · p. 67

4.3 Preparation of Symmetric TN Ansatz States

  • Symmetric-state initialization must choose compatible representations for every virtual link; product states, expected particle distributions, and problem-specific degeneracy profiles offer different starting strategies.More allowed representation matches waste less information, while the most suitable degeneracy distribution depends on the system and expected state.Source: §4.3 · pp. 67–68

4.3.1 Randomized Symmetric Ansatz States

  • A randomized construction for arbitrary symmetry groups builds and cleans up virtual links, trims them to a maximal bond dimension, then assigns random tensors to every allowed match.The procedure guarantees a non-trivial network in which all degeneracy tensors can contribute to the state.Source: §4.3.1 · p. 68

5 Ground State Search with Loop-Free Tensor Networks

The section extends DMRG ground-state search to tensor networks without cycles, where gauge freedom supports efficient variational optimization. Local optimization becomes a simple eigenvalue problem, and the approach uses contraction, SVD/QR decomposition, and simple eigenvalue solving.

5 Ground State Search with Loop-Free Tensor Networks

  • Loop-free TNs let local variational energy-minimization problems be formulated as simple eigenvalue problems, unlike loopy TNs that often require approximate approaches.The section presents this as a consequence of the absence of cycles.Source: §5 · p. 71
  • The ground-state search scheme extends DMRG to any tensor-network architecture without cycles.It is described as a robust scheme for achieving Hamiltonian ground states.Source: §5 · p. 71
  • Loop-free geometry also offers gauge freedom, while loopy networks can require more expensive contractions and make symmetry handling less straightforward.Periodic-boundary DMRG is cited as requiring specially engineered methods to address these issues.Source: §5 · p. 71
  • The approach reduces linear-algebra demands to contraction, SVD/QR decomposition, and simple eigenvalue problems while exploiting adaptive TN gauges.These are identified as the three basic procedures for the selected loop-free setup.Source: §5 · p. 71

5.1 Ground State Search Algorithm for the Adaptive-Gauge Loop-Free TN

This section describes variational optimization of loop-free tensor networks through iterative local problems, including symmetry-aware operations where applicable and extensions to excited-state targeting. It also covers variational compression at fixed bond dimension and two state-orthogonalization approaches that balance fidelity and orthogonality.

5. Ground State Search Algorithm for the Adaptive-Gauge Loop-Free TN

  • The algorithm repeatedly optimizes a local tensor with the remaining network fixed, sweeping the optimization center until the state approximates the ground state.This decomposes the full diagonalization into smaller problems; the fixed tensors form an environment used to construct a reduced effective Hamiltonian.Source: §5.1 · p. 71, Figure 34 · p. 72

5.1.1 High-Level Operations

  • Efficient interactions keep the total dimension of TPO links crossing each network bipartition non-scaling or weakly scaling with system size.Representations trade off the number of lightweight interactions against fewer or larger TPO links; a global TPO can be less efficient for long-range inhomogeneous interactions.Source: §5.1.1 · p. 74
  • The method renormalizes Hamiltonian terms toward a unitary-gauge center and combines them into an effective Hamiltonian for optimizing its tensor.With local-support interactions, isometry can make renormalization yield the identity when an interaction acts trivially on the other links, saving computation.Source: §5.1.1 · pp. 74–76
  • Projector penalties can target excited states when a known ground state is penalized, but their global support prevents summing projector terms or exploiting isometry during renormalization.The penalty is εp|Ψp⟩⟨Ψp|, with εp much larger than the energy gap; the resulting renormalized projectors are generally not identities.Source: §5.1.1 · p. 78

5.1.2 Ground State Search Algorithm in Detail

  • The detailed algorithm requires a Hamiltonian-compatible, acyclic network with N physical links and virtual-link dimensions no greater than D.The geometry should also support the required entanglement distribution; acyclicity means there is a unique path between every pair of nodes.Source: §5.1.2 · p. 81
  • Each sweep minimizes the effective Hamiltonian at a center, moves the gauge and updates terms along the path, and tests convergence with a bond-dimension-dependent energy threshold.The threshold should avoid unnecessary sweeps beyond the accuracy limit imposed by finite bond dimension; its dependence on bond dimension is fitted empirically for the model.Source: §5.1.2 · pp. 84–85

5.1.3 Optimizer – Strategies

  • Single-tensor updates are inexpensive but can constrain symmetry sectors, while double-tensor updates restore variational freedom at higher cost and subspace expansion bridges their cost and convergence tradeoff.Subspace expansion pads a link, alternately optimizes its two adjacent tensors, then restores the original bond dimension by SVD.Source: §5.1.3 · pp. 85–87

5.1.4 Optimizer – Execution

  • Because central gauging reduces each local problem to a standard eigenvalue problem, sparse effective Hamiltonians are efficiently handled by iterative Lanczos or Arnoldi methods using only matrix-vector actions.These methods target only the ground eigenstate, unlike direct diagonalization, which computes the full spectrum.Source: §5.1.4 · p. 89

5.1.5 Low Excited States

  • The search targets excited states by adding energy penalties for overlap with lower eigenstates, so each target requires that all lower states have already been determined.Penalty energies must be large enough to exceed the energy gap to the target state; each projector adds an interaction term to the effective Hamiltonian.Source: §5.1.5 · p. 90

5.1.6 Variational Compression, Addition and Orthogonalization

  • The ground-state algorithm can compress a loop-free network by minimizing with a single-projector Hamiltonian, but severe compression may miss the optimum, making preliminary local truncations useful for initialization.When only one link's bond dimension changes, Schmidt-value truncation is already an efficient optimal strategy.Source: §5.1.6 · pp. 90–91
  • Variational ground-state optimization approximates a sum of same-geometry tensor networks with the closest state achievable at fixed maximal bond dimension D.Constructing the exact sum first gives bond dimension D′ and then requires compression; drastic compression from D′ is less efficient unless block-diagonal contraction data structures are used.Source: §5.1.6 · p. 91
  • Subtracting the shared overlap and variationally adding states yields the highest-fidelity approximation to an orthogonal state at fixed D, but generally not exact orthogonality.The procedure extends to more than two states, including modified Gram–Schmidt.Source: §5.1.6 · p. 92
  • Direct variational compression enforces orthogonality by adding positive-energy projectors onto states to orthogonalize against; as εp → ∞, orthonormality becomes a hard constraint at the expense of fidelity.The optimizer’s eigenproblem then reduces to orthogonalization against the effective projectors.Source: §5.1.6 · p. 92

5.2 A Practical Application: Binary Tree Tensor Networks

The binary tree tensor network (bTTN) treats open and periodic one-dimensional systems on equal footing and offers moderate bond-dimension scaling. Its practical implementation combines central gauging, eigensolvers, update schemes, and efficient observable evaluation.

5.2 A Practical Application: Binary Tree Tensor Networks

  • A bTTN treats open- and periodic-boundary one-dimensional systems on equal footing, using three-link tensors for moderate bond-dimension scaling.Its hierarchical structure also makes two arbitrary sites readily reachable.Source: §5.2 · p. 92
  • For D < 500, QR gauging takes roughly two thirds the time of SVD while sharing its O(D^3.8) scaling.The recommendation is to use QR for gauging and reserve SVD for compression.Source: §5.2 · p. 92, Figure 48 · p. 93
  • ARPACK is the most suitable of the benchmarked eigensolvers for a bTTN ground-state algorithm.Using the previous center tensor as an iterative-solver initial guess can reduce operator applications, especially in final sweeps.Source: §5.2 · pp. 92, 94, Table 1 · p. 94, Figure 49 · p. 95
  • With N sites, a bTTN has Q = N − 2 tensors; single-/double-tensor updates use O(ND^3)/O(ND^4) memory and O(ND^4)/O(ND^6) runtime, independent of boundary conditions.The dominant contractions scale as O(D^4) and O(D^5), respectively; the double-tensor update also has an O(D^6) SVD.Source: §5.2 · pp. 94, 96
  • A center-gauged bTTN evaluates local observables with three tensors and two-point observables with O(log^2 N) contractions.The hierarchical structure also underlies the network’s stated predisposition to describe critical systems with algebraically decaying correlations.Source: §5.2 · p. 96

5.3 Binary Tree Tensor Networks – Examples

The examples apply bTTN ground-state algorithms to a critical transverse-field Ising model and a rotating-barrier Bose-Hubbard ring. They compare symmetry handling and update strategies in convergence and computational cost.

Spin-1/2 Ising Model

  • In the Ising benchmark, exploiting Z2 symmetry improves convergence, especially at N = 1024, and practical runtime and memory savings are close to the predicted factor of 1/4 and one-half.The symmetry-based single-tensor update fixes the even- and odd-sector degeneracy dimensions to D/2.Source: §5.3 · pp. 98, 100, Figure 52 · p. 99
  • Both update schemes converge to optimal link representations in the Ising benchmark, with the double-tensor update requiring slightly fewer sweeps.The comparison uses N = 128 and D = 40.Source: §5.3 · p. 98, Figure 53 · p. 100

Bose-Hubbard Ring With a Rotating Barrier

  • For the Bose-Hubbard ring, both schemes converge to optimal link representations, but double-tensor updates cost more in runtime and memory.The model has U(1) boson-number symmetry; its sector degeneracies are selected variationally because they cannot be assigned a priori.Source: §5.3 · pp. 100–101, Figure 55 · p. 102
  • For the rotating-barrier ring, U(1) double-tensor updates run roughly 4 times faster than unsymmetric single-tensor updates, while subspace expansion is roughly 3 times faster at D = 50.The subspace-expansion approach also has favorable bond-dimension scaling.Source: §5.3 · p. 101, Figure 55 · p. 102

6 Conclusions

The anthology assembles tensor-network data structures, numerical routines, symmetry handling, and loop-free network methods into a toolkit for finite-size, low-dimensional many-body simulations. It also identifies more computationally challenging two-dimensional systems as an important open direction.

6 Conclusions

  • The anthology presents data structures and efficient algebraic routines, explains how Abelian symmetries reduce effective variational spaces, and reviews loop-free network geometry and ground- and low-excited-state optimization strategies.Its numerical comparisons address the precision, performance, advantages, and scaling of those strategies.Source: §6 · p. 103

6.1 Outlook and Perspectives

  • Two-dimensional tensor-network simulation remains challenging because designing a geometry that contracts efficiently while capturing the 2D area law is difficult.The conclusion also notes loop-free networks’ demonstrated use for some two-dimensional topological models and their potential for out-of-equilibrium, finite-temperature, and open-system simulations.Source: §6.1 · p. 103

A Anti-Unitary Symmetries and Time-Reversal

The section extends symmetric tensor networks to anti-unitary symmetries, including time reversal, which can sometimes reduce Hamiltonian calculations from complex to real linear algebra. For suitable local transformations, this simplification can be encoded by restricting tensor elements to be real.

A Anti-Unitary Symmetries and Time-Reversal

  • Anti-unitary symmetries, including time reversal, can in some cases reduce Hamiltonian problems from complex to real linear algebra.The section extends the symmetric-tensor-network formalism to include anti-unitary group representations.Source: A Anti-Unitary Symmetries and… · p. 104
  • When T^2 = +1, the Hamiltonian is real in an invariant basis, which can be constructed from |S⟩ + T|S⟩ superpositions without diagonalizing H.This case is noted for integer spin and some spin models without explicit spin degrees of freedom.Source: A Anti-Unitary Symmetries and… · p. 104
  • For locally transformable lattice models, a generalized gauge transformation can encode certain anti-unitary symmetries by restricting tensor elements to be real, a TN “real gauge.”The simplest case is a Hamiltonian already real in |S⟩ and invariant under T = K.Source: A Anti-Unitary Symmetries and… · p. 104

B Graphical Notation Legend

Figure 56 serves as the legend for the manuscript’s graphical notation. It identifies tensor-network components and selected geometric properties.

B Graphical Notation Legend

Loading 1710.03733v2…