5.9.2 Tensor Product Basis Pairing Pattern

The Tensor Product Basis Pairing Pattern describes how basis elements from two vector spaces combine to form a basis for their tensor product space.

Tensor Product Basis Pairing Pattern is the combinatorial rule governing which basis vectors from the different factor spaces are joined together to form each basis tensor of a tensor product, together with the enumeration conventions used to order the resulting collection of paired basis tensors into a single indexed list or array.

The Pairing Rule

Given factor spaces V1, V2, …, Vn with chosen bases of sizes d1, d2, …, dn, the pairing pattern joins every possible combination of one basis vector per factor — nothing more and nothing less — to form a basis tensor of the product space:

{e_{k_{1}}^{(1)} \otimes \dots \otimes e_{k_{n}}^{(n)} ∣ 1 \leq k_{i} \leq d_{i} for each i}

The pairing pattern is exhaustive: every index tuple (k1, …, kn) drawn from the Cartesian product {1,…,d1} × {1,…,d2} × ⋯ × {1,…,dn} contributes exactly one basis tensor, and no two distinct tuples are ever identified with the same basis tensor, since the induced tensors are linearly independent.

Combinatorial Structure of the Pairing

Because the pairing pattern is exactly the Cartesian product of the individual index sets, its combinatorial size and structure follow directly from elementary counting.

Cardinality

The total number of paired basis tensors equals

∣ {1, \dots, d_{1}} \times \dots \times {1, \dots, d_{n}} ∣ = d_{1} d_{2} \dots d_{n}

which matches the dimension of the tensor product space, confirming that the pairing pattern produces exactly enough basis tensors, with no redundancy and no gaps.

No Partial or Selective Pairing

The pairing pattern admits no partial combinations — there is no meaningful basis tensor that pairs a basis vector from V1 with only some, but not all, of the remaining factors; every basis tensor in the induced basis involves exactly one contribution from every single factor space, which is a direct reflection of the tensor product's definition as an n-way, not a partial, construction.

Enumeration and Ordering Conventions

Although the pairing pattern determines which combinations occur, it does not by itself impose an order on them; a separate enumeration convention is needed to lay the paired basis tensors out as a single indexed list, which matters whenever the tensor is to be stored as a flat array or vectorized.

Lexicographic Ordering

The most common convention orders index tuples lexicographically, varying the last index fastest (row-major, matching common array storage conventions in many programming languages) or the first index fastest (column-major), analogous to the corresponding conventions for matrix storage.

Consistency Requirement

Whatever ordering convention is adopted must be applied consistently across every operation that relies on the flattened index — including vectorization, computation of the associated Kronecker product structure under change of basis, and matricization (flattening) of the tensor along a chosen grouping of factors — since two different orderings of the same pairing pattern produce coordinate vectors that differ by a fixed permutation, not by an underlying mathematical difference.

Pairing Pattern and Matricization

Grouping the factors into two blocks before applying the pairing pattern produces the row and column index sets used when a tensor is matricized (flattened) into a two-dimensional array.

Row and Column Pairing

If the factors are split into a first group (contributing the row index) and a second group (contributing the column index), the pairing pattern restricted to each group produces two separate index sets, and the tensor's matricization pairs every row index with every column index in turn — the same exhaustive Cartesian pairing principle applied one level higher, now between combined blocks rather than individual factors.

Relevance to Rank-One Recognition

Because matricization inherits its row and column index sets from the same underlying pairing pattern, the rank-one tests used to recognize decomposable tensors — checking that a flattening has rank at most one — are directly interpretable in terms of the pairing pattern: a rank-one flattening means the entrywise values across the paired row and column indices follow a single multiplicative pattern, rather than requiring several independent terms to reproduce.

Illustrative Diagram

The lines connecting the two rows of basis vectors trace a sample of the exhaustive Cartesian pairing pattern joining a three-element basis to a four-element basis, producing twelve paired basis tensors in total.