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.
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.
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. Inclusion test: Identify component tensors, their index dimensions, the graph pairing contracted indices, remaining open indices, and the contraction or approximation regime. Exclusion test: Exclude a generic graph with scalar labels, an explicit dense tensor without factorization, and a diagram whose edges do not denote index identifications. Nearest boundary: A factor graph also connects local factors, but it represents probabilistic factorization conventions rather than necessarily a tensor contraction over vector-space indices. Exit condition: The object ceases to be this tensor network when the index-pairing graph or component tensors no longer reconstruct the claimed multilinear object. Common misclassifications: 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. Nearest named distinctions: Neural network: Edges usually carry activations, not paired tensor indices. Factor graph: Uses probabilistic factor semantics and variable nodes. Tensor decomposition: May lack an explicit general contraction graph. Quantum circuit: A circuit can be encoded as a tensor network but has additional temporal and operational semantics.
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¶
- Define the target multilinear object and free indices.
- Choose component tensors and a connectivity graph.
- Assign compatible dimensions to every joined leg.
- Select bond dimensions and optimize tensors or derive them analytically.
- 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.
Relationships to Other Abstractions¶
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
- Tensor Network → Decomposition
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
- Symbolic Cholesky Decomposition — 0.90
- Matrix Multiplication — 0.90
- Whitney Sum — 0.88
- Artificial Neural Network — 0.88
- Algebraic Surface — 0.88
Computed from structural-signature embeddings · 2026-10-08