Tensor Products of Vector Spaces

Multiplication is the one operation a vector space does not come with. You can add two vectors and scale them, but there is no built-in way to "multiply" a vector in $V$ by a vector in $W$ . The tensor product $V\otimes W$ supplies exactly this missing product, in the only way that respects linearity in each factor separately. Its defining feature is not a formula for its elements but a job description: $V\otimes W$ is the space through which every bilinear map out of $V\times W$ factors as a single linear map. That one clause — the universal property — fixes the space uniquely up to a canonical isomorphism and hands you, for free, the trace, the induced inner product, contraction, and the operator $A\otimes B$ , with no consistency condition left to check. Pick bases and the abstraction becomes concrete: $\{e_i\otimes f_j\}$ is a basis, so the dimensions multiply, $\dim(V\otimes W)=mn$ — in sharp contrast to the direct sum $V\oplus W$ , where they add. The elementary tensors $v\otimes w$ are outer products, rank-one matrices; they span the space but do not fill it, and the least number needed to build a general element is its rank. That single fact — that not every tensor is a product — is the entire mathematical content of quantum entanglement. This article builds the construction from its universal property, computes with the Kronecker product, and follows the thread to the Schmidt decomposition and the composite-system Hilbert space.

Key takeaways

The tensor product $V\otimes W$ turns bilinear maps out of $V\times W$ into linear maps out of $V\otimes W$ : the universal solution to the bilinearity problem, unique up to a canonical isomorphism.
If $\dim V=m$ and $\dim W=n$ then $\dim(V\otimes W)=mn$ , with basis $\{e_i\otimes f_j\}$ ; the elementary tensors $v\otimes w$ span the space but are not all of it — a generic element is a sum of several.
On linear maps the tensor product gives $A\otimes B$ , whose matrix is the Kronecker product, with $(A\otimes B)(v\otimes w)=(Av)\otimes(Bw)$ , eigenvalues $\lambda_i\mu_j$ , $\operatorname{tr}(A\otimes B)=\operatorname{tr}A\,\operatorname{tr}B$ , and $\det(A\otimes B)=(\det A)^n(\det B)^m$ .
Tensor products of Hilbert spaces carry the induced inner product $\langle v_1\otimes w_1,\,v_2\otimes w_2\rangle=\langle v_1,v_2\rangle\langle w_1,w_2\rangle$ — the rigorous foundation for composite quantum systems.
Any $|\psi\rangle\in V\otimes W$ has a Schmidt decomposition into at most $\min(m,n)$ orthonormal product terms; Schmidt rank $1$ means separable, rank $>1$ means entangled.
Key idea: Define the tensor product by what it does, not by its elements: $V\otimes W$ is the unique (up to canonical isomorphism) space through which every bilinear map factors as a linear map. Bases then give $\dim(V\otimes W)=mn$ , and operators act factorwise via the Kronecker product.

Where this lives today

The tensor product became the working substrate of quantum information. A register of $N$ qubits is not $N$ separate two-state systems but a single vector in $(\mathbb{C}^2)^{\otimes N}$ of dimension $2^{N}$ — the exponential that makes quantum computers powerful and classically hard to simulate is literally the "multiply, don't add" of tensor dimensions. Entanglement, defined here as Schmidt rank greater than one, is the resource that teleportation, superdense coding, and Bell-inequality tests spend; the entanglement entropy of Worked Example 2 is now measured directly in cold-atom and superconducting-qubit platforms. On the classical side that same exponential is the enemy, and the response is tensor-network methods: matrix product states and the density-matrix renormalization group compress a many-body wavefunction by keeping only its dominant Schmidt terms, reaching spin chains and quantum chemistry brute-force storage never could. Beyond physics, low-rank tensor decompositions — the higher-order cousins of the SVD above — compress neural-network weights and recommender models. Every one applies the two facts this article proves: dimensions multiply, and not every tensor is a product.

The physical picture

The fastest way to hold the tensor product in your head is the outer product. Take a column vector $v\in\mathbb{F}^m$ and a row vector $w^{\mathsf T}$ with $w\in\mathbb{F}^n$ ; the product $vw^{\mathsf T}$ is an $m\times n$ array with $(i,j)$ entry $v_i w_j$ . This is the elementary tensor $v\otimes w$ , and it is special: every row is a multiple of $w^{\mathsf T}$ and every column a multiple of $v$ , so it has rank one. The whole space of $m\times n$ matrices is $\mathbb{F}^m\otimes\mathbb{F}^n$ , and the rank-one matrices are exactly the simple tensors inside it — with $mn$ independent entries, the dimension count the first figure makes concrete.

Figure

The elementary tensor v⊗w is the outer product vwᵀ — a rank-1 matrix. With dim V = 3 and dim W = 2 the whole space V⊗W is the 6-dimensional space of all 3×2 matrices, so dim(V⊗W) = mn = 3·2 = 6

The universal property is the right way to remember everything else. To build a map out of $V\otimes W$ you never define it on the messy general element $\sum_k v_k\otimes w_k$ and check that different representations agree; you specify a bilinear map out of $V\times W$ , and the tensor product converts it, automatically and uniquely, into one linear map. Every named construction — $A\otimes B$ , contraction, the trace, the inner product — is built this way. On operators the payoff is spectral: because $A$ and $B$ act factorwise, the joint eigenvectors are the products $v_i\otimes w_j$ with eigenvalues $\lambda_i\mu_j$ . The second figure is that grid, with real numbers attached — the source of the multiplicative trace and determinant.

Figure

Spectra multiply: the eigenvalues of A⊗B are the products λᵢμⱼ. For factor eigenvalues 2, 3 and 1, 5 the four products 2, 10, 3, 15 sum to tr = 30 = trA·trB = 5·6 and multiply to det = 900 = (detA)²(detB)² = 6²·5²

The dimension multiplies for the same reason the eigenvalues do: labelling a basis element needs one index from $V$ and one from $W$ , independently — $m$ times $n$ . The direct sum is the opposite bookkeeping, a vector in one factor or the other, so dimensions add. That contrast, multiply versus add, is the whole difference between a composite system and two systems side by side. Entanglement is then nothing more exotic than the fact that a $2\times 2$ matrix need not have rank one: the Bell array $\bigl(\begin{smallmatrix}1&0\\0&1\end{smallmatrix}\bigr)$ is rank two, hence not a single outer product, hence not separable.

The core equations

Everything begins with a single map. The canonical bilinear map sends a pair $(v,w)$ to the elementary tensor $v\otimes w$ , linear in each slot when the other is fixed; its images generate the whole space.

\beta:V\times W\to V\otimes W,\qquad \beta(v,w)=v\otimes w \qquad (1)

where: $V,W$ — vector spaces over a field $\mathbb{F}$ ; $v\otimes w$ — the elementary (simple, decomposable) tensor, image of $(v,w)$ ; $\beta$ — the universal bilinear map, linear separately in $v$ and $w$ . Reading: $\beta$ is not linear on the product $V\times W$ (that would double-count scalars); it is bilinear, and the tensor product is the machine that repairs this. The property that defines $V\otimes W$ — and pins it down uniquely — is that $\beta$ is the most general bilinear map there is: any other factors through it, once, as a linear map.

\forall\,\text{bilinear }B:V\times W\to U\;\;\exists!\,\text{linear }\tilde B:V\otimes W\to U,\quad B=\tilde B\circ\beta \qquad (2)

where: $U$ — any target space; $B$ — an arbitrary bilinear map to be "straightened"; $\tilde B$ — the unique linear map with $\tilde B(v\otimes w)=B(v,w)$ . Reading: to define a linear map out of $V\otimes W$ it is necessary and sufficient to give a bilinear map out of $V\times W$ — the tool that makes every later construction well defined. The elementary tensors obey exactly the bilinearity relations forced by equation (1) — the only relations, from which everything true in $V\otimes W$ follows plus linearity.

(v_1+v_2)\otimes w=v_1\otimes w+v_2\otimes w,\quad v\otimes(w_1+w_2)=v\otimes w_1+v\otimes w_2,\quad (cv)\otimes w=c(v\otimes w)=v\otimes(cw) \qquad (3)

where: $c$ — any scalar in $\mathbb{F}$ . Reading: the scalar slides freely between the factors, so $v\otimes w$ fixes $(v,w)$ only up to $(cv,\,c^{-1}w)$ — a first hint that simple tensors form a curved cone, not a subspace. Choose a basis in each factor and the products of basis vectors give a basis of the whole space — where the object acquires coordinates and its dimension becomes a product.

\{\,e_i\otimes f_j\,\}_{1\le i\le m,\,1\le j\le n}\ \text{is a basis of }V\otimes W,\qquad \dim(V\otimes W)=mn \qquad (4)

where: $\{e_i\}$ — basis of $V$ , $\dim V=m$ ; $\{f_j\}$ — basis of $W$ , $\dim W=n$ . Reading: two independent index ranges multiply, so $V\otimes W$ is combinatorially far larger than either factor and than the direct sum $V\oplus W$ of dimension $m+n$ . A general tensor is a combination of the basis products, and its coefficients form a matrix — reading that matrix as a matrix is the key to rank and entanglement.

t=\sum_{i=1}^{m}\sum_{j=1}^{n} t^{ij}\,e_i\otimes f_j,\qquad t^{ij}\in\mathbb{F} \qquad (5)

where: $t^{ij}$ — the $mn$ coordinates, arranged as an $m\times n$ array; $t$ — a generic tensor, simple ( $=v\otimes w$ ) iff this array has rank $\le 1$ . Reading: the tensor rank of $t$ equals the matrix rank of $(t^{ij})$ — rank $1$ separable, higher entangled. Linear maps on the factors induce a single map on the product, defined by its action on simple tensors — legitimate by the universal property (2).

(A\otimes B)(v\otimes w)=(Av)\otimes(Bw),\qquad [A\otimes B]=\begin{pmatrix}a_{11}B & \cdots & a_{1m}B\\ \vdots & & \vdots\\ a_{m1}B & \cdots & a_{mm}B\end{pmatrix} \qquad (6)

where: $A:V\to V'$ , $B:W\to W'$ — linear maps on the two factors; $a_{ij}$ — entries of $A$ , each multiplying the whole block $B$ . Reading: the block matrix is the Kronecker product, the coordinate form of $A\otimes B$ in the product basis. Because the joint eigenvectors are products, the spectrum multiplies and the scalar invariants follow at once.

\operatorname{spec}(A\otimes B)=\{\lambda_i\mu_j\},\qquad \operatorname{tr}(A\otimes B)=\operatorname{tr}A\,\operatorname{tr}B,\qquad \det(A\otimes B)=(\det A)^{n}(\det B)^{m} \qquad (7)

where: $\lambda_i,\mu_j$ — eigenvalues of $A$ and $B$ ; $m=\dim V$ (size of $A$ ), $n=\dim W$ (size of $B$ ). Reading: the exponents are asymmetric — each factor's determinant is raised to the other factor's dimension, since each eigenvalue of $A$ recurs $n$ times in the product spectrum. When the factors are inner-product spaces, exactly one inner product on $V\otimes W$ makes the basis products orthonormal — the product of the factor inner products on simple tensors.

\langle v_1\otimes w_1,\;v_2\otimes w_2\rangle=\langle v_1,v_2\rangle_V\,\langle w_1,w_2\rangle_W \qquad (8)

where: $\langle\cdot,\cdot\rangle_V,\langle\cdot,\cdot\rangle_W$ — the inner products on $V$ and $W$ . Reading: this makes the tensor product of Hilbert spaces a Hilbert space — the postulated state space of a composite quantum system. Finally, every vector in $V\otimes W$ has a canonical two-sided orthonormal expansion — the Schmidt decomposition — whose length measures entanglement.

|\psi\rangle=\sum_{k=1}^{r}\sqrt{p_k}\;|u_k\rangle\otimes|\phi_k\rangle,\qquad r\le\min(m,n),\ \ \sum_k p_k=1 \qquad (9)

where: $\{u_k\}\subset V$ , $\{\phi_k\}\subset W$ — orthonormal sets; $p_k>0$ — squared Schmidt coefficients (reduced-density eigenvalues); $r$ — the Schmidt rank. Reading: $r=1$ is separable, $r>1$ entangled, and the spread of the $p_k$ measures how entangled.

Derivation AThe universal property pins the tensor product down to one space

The goal is to justify the phrase "unique up to a canonical isomorphism" — the claim that lets us speak of the space $V\otimes W$ even though a dozen concrete constructions exist. We show that any two spaces satisfying the universal property (2) are joined by a unique isomorphism respecting their bilinear maps, so no genuinely different tensor product exists and every later formula is construction-independent. The argument is a pure diagram chase that never touches elements, which is exactly the point.

Step 1. Suppose two spaces $T_1$ and $T_2$ , with bilinear maps $\beta_1:V\times W\to T_1$ and $\beta_2:V\times W\to T_2$ , both satisfy the universal property. Apply the universality of $T_1$ to the bilinear map $\beta_2$ : there is a unique linear $\phi:T_1\to T_2$ with $\phi\circ\beta_1=\beta_2$ . Why: the property says any bilinear map — in particular $\beta_2$ — factors uniquely through $\beta_1$ , which manufactures the comparison map $\phi$ .

Step 2. Symmetrically, universality of $T_2$ applied to $\beta_1$ gives a unique linear $\psi:T_2\to T_1$ with $\psi\circ\beta_2=\beta_1$ . Why: the two spaces play interchangeable roles.

Step 3. Compose: $(\psi\circ\phi)\circ\beta_1=\psi\circ\beta_2=\beta_1$ . But $\mathrm{id}_{T_1}$ also satisfies $\mathrm{id}_{T_1}\circ\beta_1=\beta_1$ , and the universal property makes the factoring map unique, so $\psi\circ\phi=\mathrm{id}_{T_1}$ ; likewise $\phi\circ\psi=\mathrm{id}_{T_2}$ . Why: uniqueness is the entire lever — two maps that do the same job must be equal.

Step 4. Hence $\phi$ is an isomorphism with inverse $\psi$ , intertwining $\beta_1$ and $\beta_2$ . The tensor product is therefore unique up to this canonical isomorphism, and any invariant — dimension, trace, inner product — is the same in every model. Why: canonical means the isomorphism is forced by the structure, not chosen, so there is genuinely one tensor product.

Go deeper · Derivation B · 1 minEigenvalues of a Kronecker product are products of the factor eigenvalues

Derivation BEigenvalues of a Kronecker product are products of the factor eigenvalues

The aim is to prove the spectral statement in equation (7) and read off the trace and determinant identities as corollaries, since those do much of the practical work here. We assume $A$ and $B$ are diagonalizable for transparency; the general case follows by triangularizing both factors (Schur form), whose Kronecker product is still triangular with the product array on its diagonal. The mechanism is the single line $(A\otimes B)(v\otimes w)=(Av)\otimes(Bw)$ applied to eigenvectors.

Step 1. Let $\{v_i\}$ be an eigenbasis of $V$ with $Av_i=\lambda_i v_i$ , and $\{w_j\}$ an eigenbasis of $W$ with $Bw_j=\mu_j w_j$ . Then $\{v_i\otimes w_j\}$ is a basis of $V\otimes W$ by equation (4). Why: products of two bases form a basis, so it suffices to know the operator on these $mn$ vectors.

Step 2. Apply the operator's defining action: $(A\otimes B)(v_i\otimes w_j)=(Av_i)\otimes(Bw_j)=(\lambda_i v_i)\otimes(\mu_j w_j)$ . Why: $A\otimes B$ acts factorwise on simple tensors, so eigenvectors of the factors give eigenvector candidates for the product.

Step 3. Pull the scalars out through bilinearity (3): $(\lambda_i v_i)\otimes(\mu_j w_j)=\lambda_i\mu_j\,(v_i\otimes w_j)$ , so $v_i\otimes w_j$ is an eigenvector with eigenvalue $\lambda_i\mu_j$ . Why: scalars slide freely across $\otimes$ , merging two eigenvector equations into one.

Step 4. There are $mn=\dim(V\otimes W)$ such eigenpairs, a full eigenbasis, so the listed products are the entire spectrum with multiplicity. Summing gives $\operatorname{tr}(A\otimes B)=\sum_{i,j}\lambda_i\mu_j=\bigl(\sum_i\lambda_i\bigr)\bigl(\sum_j\mu_j\bigr)=\operatorname{tr}A\,\operatorname{tr}B$ , and multiplying gives $\det(A\otimes B)=\prod_{i,j}\lambda_i\mu_j=(\det A)^{n}(\det B)^{m}$ . Why: once the spectrum is the full product grid, the invariants are its sum and product, and each factor eigenvalue recurs the other factor's dimension times — the source of the exponents $n$ and $m$ .

Go deeper · Derivation C · 1 minSchmidt decomposition from the singular value decomposition

Derivation CSchmidt decomposition from the singular value decomposition

The goal is to show that every bipartite vector $|\psi\rangle\in V\otimes W$ can be written as a single orthonormal sum of product terms, the number of them equal to the matrix rank of its coefficient array — the sharp, computable criterion for entanglement. The engine is the SVD of that array, so the Schmidt decomposition is really the SVD in disguise. We work in finite-dimensional Hilbert spaces with orthonormal bases, and identify the Schmidt coefficients with the reduced-density-matrix eigenvalues at the end.

Step 1. Expand in an orthonormal product basis: $|\psi\rangle=\sum_{i,j}C_{ij}\,e_i\otimes f_j$ , collecting the coefficients into an $m\times n$ complex matrix $C=(C_{ij})$ . Why: the entire state is encoded in one matrix, and its entanglement is a property of that matrix.

Step 2. Take the SVD $C=U\Sigma V^{\dagger}$ , with $U$ ( $m\times m$ ) and $V$ ( $n\times n$ ) unitary and $\Sigma=\operatorname{diag}(\sigma_1,\dots,\sigma_r,0,\dots)$ , $\sigma_k>0$ , $r=\operatorname{rank}C\le\min(m,n)$ ; in components $C_{ij}=\sum_k U_{ik}\,\sigma_k\,\overline{V_{jk}}$ . Why: the SVD exists for every matrix and diagonalizes it between two orthonormal frames — exactly the two-sided orthonormality the Schmidt form demands.

Step 3. Define $|u_k\rangle=\sum_i U_{ik}\,e_i$ and $|\phi_k\rangle=\sum_j \overline{V_{jk}}\,f_j$ . Unitarity of $U,V$ makes $\{u_k\}$ and $\{\phi_k\}$ orthonormal, and substituting regroups the double sum into $|\psi\rangle=\sum_{k=1}^{r}\sigma_k\,|u_k\rangle\otimes|\phi_k\rangle$ . Why: the unitary columns are the rotated bases in which the state is diagonal, turning $mn$ coefficients into $r$ .

Step 4. Interpret. If $r=1$ then $C$ has rank one and $|\psi\rangle=v\otimes w$ is separable; if $r>1$ the state is entangled. Normalizing so $\sum_k\sigma_k^2=1$ and setting $p_k=\sigma_k^2$ , the reduced density operator $\rho_V=\operatorname{tr}_W|\psi\rangle\langle\psi|=\sum_k p_k\,|u_k\rangle\langle u_k|$ has the $p_k$ as eigenvalues. Why: the squared singular values are the populations seen by one subsystem alone, so the Schmidt rank and coefficients are the entanglement content — the numbers feeding the entropy $S=-\sum_k p_k\log p_k$ .

Limiting cases

The tensor product degenerates gracefully — each special case is a familiar object in disguise.

Case	Result	Meaning
One factor one-dimensional	$\mathbb{F}\otimes W\cong W$	Scalars tensor trivially, $c\otimes w\mapsto cw$ ; $\mathbb{F}$ is the unit of $\otimes$ , just as $1$ is the unit of multiplication.
Both coordinate spaces	$\mathbb{F}^m\otimes\mathbb{F}^n\cong\mathbb{F}^{m\times n}$	Order-two tensors are matrices; $v\otimes w$ is the outer product $vw^{\mathsf T}$ , a rank-one matrix.
Versus the direct sum	$\dim(V\oplus W)=m+n$ vs $\dim(V\otimes W)=mn$	Direct sum stacks independent subsystems (add); tensor product couples them into composite states (multiply).
Dual times space	$V^*\otimes W\cong\mathrm{Hom}(V,W)$	$\varphi\otimes w$ acts as $v\mapsto\varphi(v)\,w$ , a rank-one operator; sums give every linear map.
Repeated tensoring	$V^{\otimes k},\ \dim=m^k$	Higher powers carry the symmetric and antisymmetric subspaces — the home of bosonic and fermionic Fock spaces.

Common trapEvery element of V⊗W is a product v⊗w

This is the most common and most consequential error in the subject, and it is false. The elementary tensors only span $V\otimes W$ ; the generic element is a sum $\sum_k v_k\otimes w_k$ that cannot be collapsed to one product. The coefficient matrix $(t^{ij})$ of a simple tensor has rank $\le 1$ , and most matrices have higher rank — the identity $\bigl(\begin{smallmatrix}1&0\\0&1\end{smallmatrix}\bigr)$ does not factor as an outer product. In quantum language this is exactly the gap between separable and entangled states, so treating a generic tensor as if it factored quietly assumes away all of entanglement. Two smaller traps ride along: $v\otimes w=0$ iff $v=0$ or $w=0$ , and $v\otimes w$ fixes $(v,w)$ only up to $(cv,\,c^{-1}w)$ .

Worked examples

1 · A Kronecker product, end to end (confidence builder)

Let $A=\bigl(\begin{smallmatrix}2&1\\0&3\end{smallmatrix}\bigr)$ and $B=\bigl(\begin{smallmatrix}1&0\\2&5\end{smallmatrix}\bigr)$ . Build $A\otimes B$ explicitly and verify its trace, determinant, and spectrum against the factorwise identities.

Step 1. Read off the factor invariants. $A$ is upper triangular with eigenvalues $\{2,3\}$ , so $\operatorname{tr}A=5$ , $\det A=6$ ; $B$ is lower triangular with eigenvalues $\{1,5\}$ , so $\operatorname{tr}B=6$ , $\det B=5$ . These are the only inputs the identities need.

Step 2. Assemble the Kronecker product (equation 6): each entry $a_{ij}$ multiplies the whole block $B$ , giving a $4\times 4$ matrix.

A\otimes B=\begin{pmatrix}2B & 1B\\ 0B & 3B\end{pmatrix}=\begin{pmatrix}2&0&1&0\\4&10&2&5\\0&0&3&0\\0&0&6&15\end{pmatrix}

Step 3. Trace: sum the diagonal, $2+10+3+15=\mathbf{30}$ , and the identity predicts $\operatorname{tr}A\,\operatorname{tr}B=5\cdot 6=\mathbf{30}$ — agreement.

Step 4. Spectrum: the matrix is block upper triangular (lower-left block $0B=0$ ), so its eigenvalues are those of the diagonal blocks. Now $2B$ has eigenvalues $\{2,10\}$ and $3B$ has $\{3,15\}$ , so $\operatorname{spec}(A\otimes B)=\{2,10,3,15\}=\{\lambda_i\mu_j\}$ — exactly the product grid of the second figure.

Step 5. Determinant as the product of the spectrum: $2\cdot 10\cdot 3\cdot 15=\mathbf{900}$ . With $m=n=2$ the identity predicts $(\det A)^2(\det B)^2=6^2\cdot 5^2=\mathbf{900}$ — agreement. Result: $\operatorname{tr}(A\otimes B)=30$ , $\det(A\otimes B)=900$ , spectrum $\{2,10,3,15\}$ , all matching the factorwise formulas.

2 · Separable or entangled? Reading rank off the coefficient matrix

Two two-qubit states (bases $\{|0\rangle,|1\rangle\}$ ). State (a): $|\psi_a\rangle\propto|00\rangle+|01\rangle+|10\rangle+|11\rangle$ . State (b): $|\psi_b\rangle=0.6\,|00\rangle+0.8\,|11\rangle$ . Decide which is entangled, and quantify it.

Step 1. Write each coefficient matrix $C$ , rows indexed by the first qubit and columns by the second (equation 5): $C_a=\bigl(\begin{smallmatrix}1&1\\1&1\end{smallmatrix}\bigr)$ and $C_b=\bigl(\begin{smallmatrix}0.6&0\\0&0.8\end{smallmatrix}\bigr)$ .

Step 2. Rank of $C_a$ : the two rows are identical, so $\operatorname{rank}C_a=1$ — a simple tensor, so state (a) is separable, $|\psi_a\rangle\propto(|0\rangle+|1\rangle)\otimes(|0\rangle+|1\rangle)$ . Schmidt rank $1$ , zero entanglement.

Step 3. Rank of $C_b$ : diagonal with nonzero entries $0.6,0.8$ , so $\operatorname{rank}C_b=2$ — not a single outer product, so state (b) is entangled. Its singular values sit on the diagonal, $\sigma_1=0.8,\ \sigma_2=0.6$ , giving populations $p_1=0.64,\ p_2=0.36$ ( $0.6^2+0.8^2=1$ , so $|\psi_b\rangle$ is normalized).

Step 4. Quantify with the reduced density matrix $\rho_A=\operatorname{diag}(0.36,0.64)$ and its von Neumann entropy $S=-\sum_k p_k\log_2 p_k=0.531+0.412=\mathbf{0.94\ \text{bits}}$ , short of the maximal $1$ bit a balanced Bell state ( $p_1=p_2=0.5$ ) gives. Result: (a) is separable (Schmidt rank 1); (b) is entangled, Schmidt rank 2, entanglement entropy $\approx 0.94$ bits.

Physics of the portrait

The tensor product had to be invented three times — as an algebra of directed magnitudes, a calculus of coordinates, and finally a universal object — so its portrait is less a person than a two-century relay.

Origins · 1 mintensor product and how this was discovered

tensor product

the universal bilinear construction V⊗W · multilinear algebra and quantum mechanics

The tensor product is one of the oldest ideas in linear algebra and one of the last to be given a clean definition. Hermann Grassmann built the first algebra of "extensive quantities" in his 1844 Ausdehnungslehre — but the book was so far ahead of its notation that it was ignored for a generation. The physical need surfaced in crystal physics and continuum mechanics, where quantities like stress carried two directional indices at once; Woldemar Voigt coined the word "tensor" in 1898, and Ricci-Curbastro and Levi-Civita systematized the coordinate calculus in 1900 — the tool Einstein seized on for general relativity. The abstract definition came later: von Neumann's 1932 Mathematical Foundations of Quantum Mechanics declared a composite system's state space to be the tensor product of the component Hilbert spaces, and Hassler Whitney's 1938 paper on tensor products of abelian groups isolated the universal property as the definition. That one-line characterization — the space through which every bilinear map factors — is why entanglement, contraction, and the metric tensor are facets of one object.

History

1844

— Grassmann's Ausdehnungslehre founds the algebra of extensive (directed) quantities

1898

— Voigt coins the term "tensor" for stress-like quantities in crystal physics

1900

— Ricci and Levi-Civita systematize the absolute differential calculus of tensors

1932

— von Neumann formulates a composite quantum system as a tensor-product Hilbert space

1938

— Whitney defines the tensor product by its universal property

Going further

Texts. Stone & Goldbart, Mathematics for Physics (CUP, 2009), chapter on multilinear algebra · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 4 · Hassani, Mathematical Physics, 2nd ed., Ch. 26 (tensor algebra) · Byron & Fuller, Mathematics of Classical and Quantum Physics, Ch. 3 · Nakahara, Geometry, Topology and Physics, 2nd ed., §2.6 (tensors) · Roman, Advanced Linear Algebra, 3rd ed., Ch. 14 (Tensor Products) · Nielsen & Chuang, Quantum Computation and Quantum Information, §2.1.7 and §2.5 (Schmidt decomposition). The key results in this article are: the dimension formula $\dim(V\otimes W)=mn$ with product basis $\{e_i\otimes f_j\}$ ; the Kronecker spectrum $\{\lambda_i\mu_j\}$ with identities $\operatorname{tr}(A\otimes B)=\operatorname{tr}A\,\operatorname{tr}B$ and $\det(A\otimes B)=(\det A)^n(\det B)^m$ ; the $4\times 4$ Kronecker product of Example 1 with trace $30$ , determinant $900$ , and spectrum $\{2,10,3,15\}$ ; the entropy of Example 2, $-0.36\log_2 0.36-0.64\log_2 0.64\approx 0.94$ bits; and the uniqueness of the tensor product via the standard universal-property argument. No nonstandard numbers are asserted.

Prerequisites. Vector Spaces and Subspaces · Linear Independence, Basis, and Dimension · Linear Operators and Matrix Representations · Inner-Product Spaces and Norms. Leads to. Adjoint and Hermitian Operators · Eigenvalues, Eigenvectors, and the Characteristic Equation · Determinants and Their Properties.

Ask the tutor: Ask me why the identity matrix is entangled but a product of coin flips is not — the same rank test, two very different stories.

Why not just define v⊗w as the outer product vwᵀ and be done?

For coordinate spaces $\mathbb{F}^m$ and $\mathbb{F}^n$ that model is perfectly correct — $v\otimes w=vw^{\mathsf T}$ and $V\otimes W$ is the space of $m\times n$ matrices. But it is basis-dependent and does not generalize: abstract spaces have no canonical "row" and "column", infinite-dimensional Hilbert spaces have no finite matrices, and repeated products $V\otimes W\otimes X$ get awkward fast. The universal property sidesteps all of this — it defines $V\otimes W$ by the maps into and out of it, works in any setting, and by uniqueness agrees with the outer-product picture wherever that applies. You get the concrete model as a special case, plus a definition that survives leaving it.

What is the difference between V⊗W and V⊕W in one sentence?

The direct sum $V\oplus W$ places the two spaces side by side, so an element is a pair $(v,w)$ and dimensions add ( $m+n$ ); the tensor product couples them, so an element is a sum of products and dimensions multiply ( $mn$ ). Physically, the direct sum is two systems you merely describe together, while the tensor product is one composite system whose parts can be correlated — entangled. The mnemonic is "or" versus "and": a basis vector of the sum is $e_i$ or $f_j$ (pick a side), of the product $e_i$ and $f_j$ together (one from each). That distinction is why $N$ qubits need $2^N$ amplitudes.

Why do the determinant exponents come out asymmetric, (det A)ⁿ (det B)ᵐ?

Because each eigenvalue of a factor is reused as many times as the other factor is large. The spectrum of $A\otimes B$ is the full $m\times n$ grid of products $\lambda_i\mu_j$ . A fixed $\lambda_i$ appears once per column, i.e. $n$ times; a fixed $\mu_j$ once per row, i.e. $m$ times. The determinant is the product over the whole grid, so $\prod_{i,j}\lambda_i\mu_j=\bigl(\prod_i\lambda_i\bigr)^{n}\bigl(\prod_j\mu_j\bigr)^{m}=(\det A)^{n}(\det B)^{m}$ . The trace, a sum over the same grid, factors symmetrically as $(\sum\lambda_i)(\sum\mu_j)$ , which is why only the determinant carries the swapped exponents. If you misremember which is which, rederive it from a $2\times 2$ example in ten seconds.

Is entanglement a physics phenomenon or a fact about tensor products?

Both, and that is the point. Mathematically, "entangled" just means "not a simple tensor" — the coefficient matrix has rank greater than one; nothing quantum in that, and the identity matrix has been rank two since long before quantum mechanics. What quantum theory adds is the postulate that a composite system's state space is the tensor product of the component Hilbert spaces (equation 8), making these non-product vectors physically realizable — plus the measurement rules that let their correlations violate Bell inequalities, which no classical joint distribution can. The linear algebra supplies the possibility of entanglement; the physics supplies its reality.

How is the Schmidt decomposition different from just diagonalizing something?

It is a two-sided diagonalization, which is why it exists for every state. Ordinary diagonalization rotates one basis and fails for non-normal or non-square objects. A bipartite state's coefficient matrix $C$ is generally rectangular with no eigenvalues of its own, but the SVD rotates the $V$ -basis and $W$ -basis independently to make $C$ diagonal with nonnegative entries — the Schmidt coefficients. Since the SVD exists for any matrix, the Schmidt form exists for any state. Its length (Schmidt rank) and coefficients (reduced-density eigenvalues) are basis-independent invariants, which makes them legitimate measures of entanglement.

Where does the tensor product show up outside quantum mechanics?

Almost everywhere a quantity carries more than one directional index. The metric $g_{\mu\nu}$ and stress tensor $\sigma_{ij}$ live in $V^*\otimes V^*$ , the moment-of-inertia tensor in $V\otimes V$ ; index notation $T^{ij}$ is the expansion $\sum T^{ij}\,e_i\otimes f_j$ of equation (5), and contraction is the pairing $V^*\otimes V\to\mathbb{F}$ . Coupling two angular momenta decomposes a tensor product of $SU(2)$ representations, $V_{j_1}\otimes V_{j_2}=\bigoplus_{J=|j_1-j_2|}^{j_1+j_2}V_J$ , with the Clebsch–Gordan coefficients as the change of basis. And $\mathrm{Hom}(V,V)\cong V^*\otimes V$ writes every operator as a sum of rank-one pieces. Same construction, many costumes.