Skip to content

Tensor Network

A graph-structured factorization of a high-dimensional tensor into smaller tensors, with edges denoting contracted indices, open edges denoting free indices, and bond dimensions controlling capacity.

Version
v1 · 2026-09-28 · History
Domain-specific #
12487
Domain group
Natural Sciences
Origin domain
Physics
Subdomains
Quantum Many Body Physics, Tensor Factorizations → Physics
Aliases
Tensor network representation

Core Idea

A tensor network replaces one high-dimensional array or multilinear map with smaller tensors connected by an index graph. Joined legs identify indices that are summed, while open legs survive as indices of the represented result.

The representation is useful only when its factorization is compact and its contractions are feasible. Bond dimensions control how much information crosses graph cuts, while topology and contraction order govern intermediate cost; tensor-network form alone does not guarantee efficiency.

Structural Signature

Sig role-phrases:

  • Factor tensors — Encode local multilinear factors rather than one explicit global array. It is required content. Counterfactual: Without component tensors there is no network factorization.
  • Connectivity graph — Specifies which tensor indices are paired and summed. It is defining topology. Counterfactual: Changing connectivity generally changes the represented contraction.
  • Contracted indices — Transmit and sum shared degrees of freedom between factors. It is composition rule. Counterfactual: Uncontracted joined dimensions do not compose the factors.
  • Open indices — Determine the free indices and type of the resulting object. It is output interface. Counterfactual: Closing or adding an open leg changes the output object.
  • Bond dimensions — Bound information capacity across internal cuts. It is capacity control. Counterfactual: Unbounded or rapidly growing bonds remove compression advantages.
  • Contraction order — Determines intermediate sizes and computational cost. It is execution plan. Counterfactual: A poor order can make a compact representation infeasible to evaluate.

What It Is Not

  • It is not any node-and-edge diagram.
  • It is not restricted to quantum wavefunctions.
  • A low-rank representation is not automatically cheap to contract for every topology.
  • Bond dimension is not the physical dimension of an open index.
  • Closest near-miss. A factor graph also connects local factors, but it represents probabilistic factorization conventions rather than necessarily a tensor contraction over vector-space indices.

Scope of Application

  • Many-body physics. Represents states, operators, and partition functions.
  • Numerical multilinear algebra. Compresses high-dimensional arrays and maps.
  • Quantum circuits. Represents circuits and supports classical contraction simulations.
  • Probabilistic and machine learning models. Expresses structured multilinear computations when index semantics are explicit.

Clarity

State every tensor's index set and dimensions, the graph pairing, open-index order, bond dimensions, approximation error, and contraction strategy. Separate representational compactness from computational tractability.

Manages Complexity

The network exposes local factors, information bottlenecks, and evaluation dependencies that an explicit tensor hides, while keeping the approximation and cost controls visible.

Abstract Reasoning

  1. Define the target multilinear object and free indices.
  2. Choose component tensors and a connectivity graph.
  3. Assign compatible dimensions to every joined leg.
  4. Select bond dimensions and optimize tensors or derive them analytically.
  5. Choose a contraction order and quantify cost and approximation error.

Knowledge Transfer

Graphical contraction reasoning transfers across physics, statistics, and numerical analysis only when node tensors, edge-index spaces, output legs, and approximation semantics are preserved.

Examples

Canonical

A matrix product state factors an N-site wavefunction into a chain of three-index tensors; neighboring virtual indices are contracted and physical indices remain open.

Mapped back: factors → site tensors; graph → chain; contracted → virtual bonds; open → physical legs; capacity → bond dimension.

Applied / In Practice

Drawing a neural network as nodes and arrows does not make it a tensor network unless arrows specify compatible tensor indices and contraction reconstructs a defined tensorial map.

Mapped back: graph → present; tensor-index semantics → absent; verdict → not sufficient.

Structural Tensions

T1 — Compression versus Expressive Capacity. Smaller bonds reduce storage while excluding correlations that require larger ranks.

Diagnostic: How does approximation error change as bond dimensions grow?

T2 — Expressive Topology versus Contraction Tractability. Loops and higher-dimensional geometries can capture structure while making exact contraction expensive.

Diagnostic: Does the chosen graph admit a feasible contraction order?

Structural–Framed Character

Tensor Network is strongly structural as a graph-index factorization, with application-specific tensor meanings.

Structural Core vs. Domain Accent

The skeleton is factorization plus index contraction. Quantum physics adds entanglement, variational states, and operators; other domains supply their own tensor meanings and error criteria.

This entry is a kind of Decomposition.

  • Approved root. No reviewed parent entails graph-structured tensor contraction.

  • Related — tensor contraction, low-rank factorization, and graphical model. They supply the operation, compression family, and a neighboring graph-based representation.

Relationships to Other Abstractions

Local relationship map for Tensor NetworkParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.Tensor NetworkDOMAINPrime abstraction: Decomposition — is a kind ofDecompositionPRIME

Current abstraction Tensor Network Domain-specific

Parents (1) — more general patterns this builds on

  • Tensor Network is a kind of Decomposition Prime

    A Tensor Network is Decomposition of a high-dimensional tensor into smaller tensors linked by contracted indices.

Hierarchy path (1) — routes to 1 parentless root

Neighborhood in Abstraction Space

Tensor Network sits in a crowded region of the domain-specific corpus (33rd percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Coordinate Systems & Spatial Measures (29 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With

  • Neural network. Tell: Edges usually carry activations, not paired tensor indices.
  • Factor graph. Tell: Uses probabilistic factor semantics and variable nodes.
  • Tensor decomposition. Tell: May lack an explicit general contraction graph.
  • Quantum circuit. Tell: A circuit can be encoded as a tensor network but has additional temporal and operational semantics.

References

  • Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/Tensor_network (revision 1371059351).
  • Preserved source candidate: https://digitalcommons.uri.edu/phys_facpubs/241
  • Preserved source candidate: https://hal.science/hal-03282962

The frozen Wikipedia revision is discovery provenance. The retained source set was reviewed for identity, formal or operational relation, and scope. The encyclopedia's structural synthesis is bounded to those claims; a thin authority surface is recorded as a nonblocking source-strengthening repair rather than concealed.