14.23.5 Tensor Map Product Error Pattern Boundary

Tensor Map Product Error Pattern Boundary is the catalog of recurring mistaken beliefs about the tensor product of maps that mark the edge between a correct working understanding of the construction and a set of superficially plausible but false generalizations drawn from composition, direct sums, or ordinary scalar arithmetic.

Confusing Tensor Product With Composition

The Error

A common error treats $f\otimes g$ as though it were another way of writing the composite $f\circ g$, particularly when $f$ and $g$ happen to act on the same space, leading to statements like $(f\otimes g)\left(v\right)=f\left(g\right(v\left)\right)$.

Why It Is Wrong

The tensor product map is defined on $V\otimes W$, an entirely different space from $V$ or $W$ alone, and its output on a simple tensor is $f\left(v\right)\otimes g\left(w\right)$, not a single vector obtained by feeding the output of one map into the other; composition, by contrast, requires the codomain of one map to equal the domain of the other and produces an ordinary vector in that shared codomain.

Confusing Tensor Product With Direct Sum

The Error

A related error assumes the matrix of $f\otimes g$ is block-diagonal, with $F$ and $G$ placed side by side as independent blocks, matching the structure of a direct sum of operators.

Why It Is Wrong

The tensor product matrix is a full block matrix in which every block is a scaled copy of $G$, with no block equal to zero unless the corresponding entry of $F$ is zero, in sharp contrast with the direct sum matrix $F\oplus G$, which genuinely is block-diagonal with off-diagonal blocks identically zero; mistaking one structure for the other leads to dimension errors, since $F\otimes G$ has size $pq\times nm$ while $F\oplus G$ has size $(p+q)\times (n+m)$.

Assuming Rank Adds Instead of Multiplies

The Error

Because rank behaves additively for a direct sum, $\text{rank}(f\oplus g)=\text{rank}\left(f\right)+\text{rank}\left(g\right)$, it is tempting to assume the same additive law holds for the tensor product.

Why It Is Wrong

Rank is multiplicative, not additive, for the tensor product of maps: $\text{rank}(f\otimes g)=\text{rank}\left(f\right)\cdot \text{rank}\left(g\right)$, a direct consequence of the image factorization $\text{im}(f\otimes g)=\text{im}\left(f\right)\otimes \text{im}\left(g\right)$; applying the additive formula instead produces a rank estimate that is wrong except in degenerate cases where one of the two ranks equals one.

Distributing a Sum of Maps Incorrectly

The Error

Expanding $(f_1+f_2)\otimes (g_1+g_2)$, it is common to write only two terms, $f_1\otimes g_1+f_2\otimes g_2$, by analogy with a misremembered distributive law.

Why It Is Wrong

Bilinearity in both slots simultaneously forces all four cross terms to appear, $f_1\otimes g_1+f_1\otimes g_2+f_2\otimes g_1+f_2\otimes g_2$, and dropping the mixed terms silently discards genuine contributions to the sum, an error most easily caught by checking dimension or rank consistency after expansion.

Assuming a Fixed Ordering Convention Without Stating It

The Error

A candidate matrix is compared against a reference Kronecker product without first confirming that both use the same lexicographic ordering of the combined basis, leading to a mismatch that looks like a genuine computational error but is actually only a permutation of rows and columns.

Why It Is Wrong to Treat as a Genuine Discrepancy

As established in the discussion of the ordering convention, two matrices differing only by the choice of which index varies fastest represent the same underlying map and are related by conjugation with a fixed permutation matrix; treating this difference as an error in the underlying construction, rather than as a bookkeeping mismatch, misdiagnoses the actual source of any disagreement.

Assuming Diagonalizability Transfers Without Qualification

The Error

Assuming that because eigenvalues of $T\otimes S$ are always products of eigenvalues of $T$ and $S$, the operator $T\otimes S$ must always be diagonalizable whenever this eigenvalue relationship is exhibited.

Why It Is Wrong

Diagonalizability of $T\otimes S$ requires diagonalizability of both $T$ and $S$ individually, as established at the operator representation boundary; the mere existence of eigenvalues and eigenvectors satisfying the product relation does not, by itself, guarantee that these eigenvectors span the whole space, which is the additional condition diagonalizability actually requires.