Tensor Algebra and the Tensor Product

Physics keeps writing down objects with more than one index — the stress tensor $\sigma^{ij}$ , the field strength $F_{\mu\nu}$ , the metric $g_{\mu\nu}$ , the Riemann curvature $R^{\rho}{}_{\sigma\mu\nu}$ — and then declaring, almost as an afterthought, how their components transform under a change of coordinates. The tensor product reverses this order of business. Given a single vector space $V$ and its dual $V^*$ , it manufactures every one of these multi-index objects cleanly, without ever choosing coordinates, so that the index gymnastics and transformation laws come out as theorems rather than definitions. Fix a basis afterward and the indices reappear; they were bookkeeping all along. The engine that makes this work is a single universal property. A bilinear map $B:V\times W\to U$ is linear in each slot but not in the pair $(v,w)$ , which puts it just out of reach of ordinary linear algebra. The tensor product $V\otimes W$ is a new vector space with a bilinear map $\otimes$ , engineered so that every bilinear $B$ becomes an ordinary linear map $\tilde B:V\otimes W\to U$ — multilinearity converted, once and for all, into linearity. Because this property is stated by an existence-and-uniqueness clause, it pins $V\otimes W$ down up to a unique isomorphism: any two honest constructions are canonically the same. Concretely, with bases one realizes $V\otimes W$ as the span of formal products $e_i\otimes f_j$ subject only to bilinearity, and its dimension is the product of the factor dimensions — the fact that most sharply separates $\otimes$ from the direct sum $\oplus$ . Iterating gives the tensor algebra $T(V)$ , the free associative algebra whose quotients are the symmetric algebra, the exterior algebra of differential forms, and the Clifford algebra of spinors. Throughout, $V$ is finite-dimensional over a field $\mathbb{F}$ (think $\mathbb{R}$ or $\mathbb{C}$ ) with $n=\dim V$ ; curvature, connections, and forms are left to neighbouring articles.

Key takeaways

The tensor product $V\otimes W$ is the universal space that linearizes bilinearity: every bilinear map out of $V\times W$ factors uniquely through it as a linear map.
If $\{e_i\}$ and $\{f_j\}$ are bases of $V$ and $W$ , then $\{e_i\otimes f_j\}$ is a basis of $V\otimes W$ , so $\dim(V\otimes W)=\dim V\cdot\dim W$ — a product, not the sum $m+n$ of the direct sum.
A type- $(p,q)$ tensor is an element of $V^{\otimes p}\otimes(V^*)^{\otimes q}$ : a multilinear object with $p$ contravariant (upper) and $q$ covariant (lower) slots.
The identifications $\mathrm{Hom}(V,W)\cong W\otimes V^*$ and $\mathrm{End}(V)\cong V\otimes V^*$ make $\otimes$ the natural home of linear maps, contraction, and the trace.
The tensor algebra $T(V)=\bigoplus_{k\ge0}V^{\otimes k}$ is the free associative algebra on $V$ , parent of the symmetric, exterior, and Clifford algebras.
Key idea: The tensor product converts multilinearity into linearity — $V\otimes W$ is the unique space through which every bilinear map factors — so indices, transformation laws, contraction, and the metric are all bookkeeping for elements of iterated tensor products of $V$ and $V^*$ .

The core insight

The tensor product converts multilinearity into linearity: $V\otimes W$ is the unique space through which every bilinear map factors. Everything about tensors — indices, transformation laws, contraction, the metric — is bookkeeping for elements of iterated tensor products of $V$ and $V^*$ .

Where this lives today

The tensor product is the working algebra of modern quantum physics and a frontier of computation. In quantum information the state space of a joint system is $\mathcal H_A\otimes\mathcal H_B$ , and "most tensors are not simple" is the precise statement that entanglement is generic — the resource that quantum computers, teleportation, and Bell tests all exploit. That same structure powers large-scale simulation: tensor-network methods such as matrix product states (MPS/DMRG) and PEPS represent an exponentially large many-body wavefunction as a web of small contracted tensors, turning impossible ground-state calculations into tractable ones — and the language crossed into machine learning, where every deep-learning framework is literally a "tensor" library doing these contractions on GPUs. On the geometric side, the same $V\otimes V^*$ bookkeeping that names the metric runs LIGO's numerical-relativity codes, evolving the ten components of $g_{\mu\nu}$ to predict the waveforms of merging black holes. And separability — deciding whether a mixed state factors or is entangled — is computationally hard in general, keeping the simple-versus-general distinction at the center of quantum-complexity research.

The physical picture

Think of a simple tensor $v\otimes w$ as a stored recipe: fed a bilinear measuring device it returns a number, but it is kept as an honest vector so recipes can be added and scaled. The tensor product is the smallest space holding all such recipes, and its defining trick is that the awkward two-argument linearity of a bilinear map becomes ordinary one-argument linearity inside it. The cleanest model is matrices: when $\dim V=m,\dim W=n$ , the space $V\otimes W$ is exactly the $m\times n$ matrices, a simple tensor $v\otimes w$ is the rank-one outer product $v\,w^{\mathsf T}$ , and a general tensor is any matrix — a sum of rank-one pieces. "Most matrices have rank greater than one" is precisely "most tensors are not simple."

Figure

The tensor product multiplies dimensions. Bases of V (dimension 2) and W (dimension 3) give 2 × 3 = 6 basis tensors eᵢ ⊗ fⱼ, whereas the direct sum V ⊕ W has only 2 + 3 = 5 dimensions. The outer product v ⊗ w with v = (2, 3)ᵀ and w = (1, 0, 4) is the rank-one 2 × 3 matrix shown

Indices are just coordinates on these spaces. An upper index labels a basis of $V$ (contravariant, like a vector's components), a lower index labels a basis of $V^*$ (covariant, like a functional); the placement records which factor a slot belongs to, so contraction can pair one of each. This is why the same construction, iterated over $V$ and $V^*$ , delivers vectors, covectors, operators, metrics, and curvature as special cases distinguished only by their counts of upper and lower slots. The second figure shows the most physical instance: the split of $V\otimes V$ into symmetric and antisymmetric parts, which in four-dimensional spacetime separates the ten metric components from the six of the electromagnetic field.

Figure

In four-dimensional spacetime V ⊗ V has 4 × 4 = 16 components. The swap τ splits them into a symmetric part Sym²V of dimension 10 — the metric gμν, diagonal plus upper triangle — and an antisymmetric part Λ²V of dimension 6 — the field strength Fμν, strict upper triangle with zero diagonal. The counts add: 10 + 6 = 16

The core equations

Everything begins with a single map. The tensor product supplies a target space $V\otimes W$ and a rule that turns an ordered pair of vectors into one vector living inside it; every identity below describes how this map behaves.

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

where: $V,\,W$ — vector spaces over a field $\mathbb{F}$ ; $V\otimes W$ — their tensor product, a new vector space; $v\otimes w$ — a simple (decomposable) tensor, the image of the pair. Reading it: pairs go in, single honest vectors come out, ready to be added and scaled. The map $\otimes$ is not linear in the pair but linear in each slot separately. These bilinearity relations are not theorems; they are the defining relations built into the construction, and they let a scalar slide freely between the factors.

(\alpha v_1+\beta v_2)\otimes w=\alpha(v_1\otimes w)+\beta(v_2\otimes w),\qquad v\otimes(\alpha w_1+\beta w_2)=\alpha(v\otimes w_1)+\beta(v\otimes w_2) \qquad (2)

where: $\alpha,\beta$ — scalars in $\mathbb{F}$ ; $v_i\in V,\,w_j\in W$ . Reading it: $(2v)\otimes w=v\otimes(2w)=2(v\otimes w)$ , so $\otimes$ is not a commutative product — scalars cross the $\otimes$ sign, the vectors never do. The property that names $V\otimes W$ uniquely is universality: every bilinear map out of $V\times W$ is secretly a linear map out of $V\otimes W$ , in exactly one way.

\forall\,B:V\times W\to U\ \text{bilinear}\quad \exists!\ \tilde B:V\otimes W\to U\ \text{linear},\qquad \tilde B(v\otimes w)=B(v,w) \qquad (3)

where: $U$ — any target vector space; $B$ — a bilinear map; $\tilde B$ — the unique linear factorization; $\exists!$ — "there exists a unique." Reading it: specifying a bilinear map is the same labour as specifying a linear map on the tensor product, so proofs reduce to checking simple tensors. Choosing a basis of each factor and forming all one-from-each products gives a basis of the tensor product, so its dimension is the product of the two.

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

where: $\{e_i\}$ — basis of $V$ , $m=\dim V$ ; $\{f_j\}$ — basis of $W$ , $n=\dim W$ ; $mn$ — product, not the sum $m+n$ . Reading it: this is the algebraic difference between $\otimes$ and $\oplus$ , and why a rank-2 tensor in $n$ dimensions has $n^2$ components. Iterating over $V$ and $V^*$ produces the general tensor. A type- $(p,q)$ tensor lives in $p$ copies of $V$ and $q$ of $V^*$ , and expanding it on the product basis exposes the sea of upper and lower indices.

T=T^{i_1\cdots i_p}{}_{j_1\cdots j_q}\;e_{i_1}\otimes\cdots\otimes e_{i_p}\otimes \varepsilon^{j_1}\otimes\cdots\otimes \varepsilon^{j_q} \qquad (5)

where: $T^{\cdots}{}_{\cdots}$ — components, summed over repeated indices (Einstein convention); $e_i$ — basis of $V$ in the contravariant (upper) slots; $\varepsilon^j$ — dual basis with $\varepsilon^j(e_i)=\delta^j_i$ in the covariant (lower) slots. Reading it: the index picture is just components on this basis, and it fixes the transformation law — upper indices transform by the Jacobian $\partial x'^i/\partial x^k$ , lower by its inverse, so a contracted pair is invariant. The most useful identification makes linear maps into tensors: a map $V\to W$ is the same datum as an element of $W\otimes V^*$ — a covector reads the input, a vector supplies the output.

\mathrm{Hom}(V,W)\;\cong\;W\otimes V^*,\qquad (w\otimes\varphi)(v)=\varphi(v)\,w \qquad (6)

where: $\varphi\in V^*$ — a covector measuring the input; $w\in W$ — the output direction; $\cong$ — canonical isomorphism in finite dimensions. Reading it: with $W=V$ this is $\mathrm{End}(V)\cong V\otimes V^*$ , so an operator carries one upper and one lower index, $A^i{}_j$ — the index structure of a matrix explained. Contraction is dual to building tensors up: it feeds a vector slot into a covector slot, collapsing both to a scalar. On an operator's single pair of slots it is exactly the trace, basis-independent by construction.

C:V\otimes V^*\to\mathbb{F},\qquad C(v\otimes\varphi)=\varphi(v),\qquad C(A)=A^i{}_i=\operatorname{tr}A \qquad (7)

where: $C$ — the contraction map, a linear functional; $\varphi(v)$ — the covector's value on the vector; $\mathbb{F}$ — the ground field. Reading it: on a higher tensor one pairs one upper with one lower slot, lowering the type by $(1,1)$ , e.g. $T^{ij}{}_{jk}\mapsto$ a $(1,1)$ tensor. Stacking all tensor powers of $V$ into one graded object gives the tensor algebra, whose multiplication is $\otimes$ itself. It is "free": it imposes no relations beyond linearity and associativity.

T(V)=\bigoplus_{k=0}^{\infty}V^{\otimes k},\qquad V^{\otimes 0}=\mathbb{F},\ V^{\otimes 1}=V,\qquad V^{\otimes p}\otimes V^{\otimes q}=V^{\otimes(p+q)} \qquad (8)

where: $T(V)$ — the full (contravariant) tensor algebra; $V^{\otimes k}$ — the degree- $k$ piece; $\oplus$ — graded direct sum, multiplication raising degree additively. Reading it: quotient by $v\otimes w-w\otimes v$ for the symmetric algebra, by $v\otimes w+w\otimes v$ for the exterior algebra, by $v\otimes v-\langle v,v\rangle$ for the Clifford algebra.

Derivation ABilinear maps factor uniquely through the tensor product

The goal is to prove that $V\otimes W$ deserves its name by verifying the universal property in full: that for every target $U$ the bilinear maps $V\times W\to U$ correspond bijectively to the linear maps $V\otimes W\to U$ , realized by an explicit, unique factorization $\tilde B$ . This is the single fact from which every canonical isomorphism in the article descends. Let $B:V\times W\to U$ be bilinear; the claim is $\operatorname{Bilin}(V\times W,U)\cong\operatorname{Hom}(V\otimes W,U)$ .

Step 1. Define the factorization on a general tensor by summing $B$ over its simple pieces, $\tilde B\big(\sum_a v_a\otimes w_a\big)=\sum_a B(v_a,w_a)$ ; every element is such a finite sum since simple tensors span. Why: setting $\tilde B(v\otimes w)=B(v,w)$ is forced, and additivity dictates the value on sums.

Step 2. Check well-definedness. Two expressions for the same tensor differ by the defining bilinearity relations of $\otimes$ , and a bilinear $B$ annihilates exactly those relations, so the sum is independent of the representation. Why: this is the crux — well-definedness fails for an arbitrary function on pairs, and matching bilinearity to bilinearity rescues it.

Step 3. Confirm linearity and uniqueness together. Additive and homogeneous on the spanning simple tensors, $\tilde B$ is linear on all of $V\otimes W$ ; and any linear $L$ with $L(v\otimes w)=B(v,w)$ agrees with $\tilde B$ on that spanning set, hence everywhere. Why: the " $\exists!$ " upgrades existence into a characterization, forbidding two inequivalent tensor products.

Step 4. Close the bijection. Each linear $L:V\otimes W\to U$ conversely yields a bilinear $(v,w)\mapsto L(v\otimes w)$ , and $B\mapsto\tilde B$ and $L\mapsto L\circ\otimes$ are mutually inverse. Why: a bijection both ways makes "bilinear map" and "linear map on the tensor product" interchangeable — the whole point. This matches Roman and Stone & Goldbart.

Go deeper · Derivation B · 1 minThe dimension of the tensor product is the product of the dimensions

Derivation BThe dimension of the tensor product is the product of the dimensions

The goal is to prove rigorously that $\{e_i\otimes f_j\}$ is a basis of $V\otimes W$ , hence $\dim(V\otimes W)=mn$ — both spanning and independence — using the universal property to manufacture the functionals that detect independence. This result most sharply separates $\otimes$ from $\oplus$ and underlies every component count in the article. Let $\{e_i\}_{i=1}^m$ and $\{f_j\}_{j=1}^n$ be bases of $V$ and $W$ .

Step 1. Show spanning by pushing bilinearity through, $v\otimes w=\big(\sum_i v^i e_i\big)\otimes\big(\sum_j w^j f_j\big)=\sum_{i,j}v^iw^j\,(e_i\otimes f_j)$ ; since simple tensors span, so do the $e_i\otimes f_j$ . Why: bilinearity distributes the product over the basis expansions, reducing all simple tensors to the $mn$ basis products.

Step 2. Build detecting functionals. For each $(k,l)$ the map $\beta_{kl}(v,w)=v^k w^l$ is bilinear, so by universality it descends to a linear $\tilde\beta_{kl}:V\otimes W\to\mathbb{F}$ . Why: independence cannot be tested on formal sums; the universal property supplies honest functionals that read out one coordinate slot each.

Step 3. Evaluate on the candidate basis, $\tilde\beta_{kl}(e_i\otimes f_j)=\delta^k_i\,\delta^l_j$ . If $\sum_{i,j}c^{ij}(e_i\otimes f_j)=0$ , applying $\tilde\beta_{kl}$ gives $c^{kl}=0$ for every $(k,l)$ . Why: a family on which a dual family acts as $\delta^k_i\delta^l_j$ is automatically independent — the biorthogonality argument.

Step 4. Count. Spanning plus independence make $\{e_i\otimes f_j\}$ a basis of $mn$ elements, so $\dim(V\otimes W)=mn$ , not $m+n$ . Why: the matrix model $V\otimes W\cong\mathbb{F}^{m\times n}$ has $mn$ entries, agreeing with Arfken–Weber–Harris and Riley–Hobson–Bence.

Go deeper · Derivation C · 1 minEvery rank-two tensor splits into symmetric and antisymmetric parts

Derivation CEvery rank-two tensor splits into symmetric and antisymmetric parts

The goal is to prove that when $2$ is invertible in $\mathbb{F}$ , the space $V\otimes V$ decomposes as a direct sum $\mathrm{Sym}^2V\oplus\Lambda^2V$ realized by complementary projectors built from the swap operator, with dimensions $\tfrac{n(n+1)}2$ and $\tfrac{n(n-1)}2$ adding to $n^2$ . This is the algebraic reason the metric and the field strength, both rank-2 covariant tensors, are so different. Define the swap $\tau(u\otimes v)=v\otimes u$ on $V\otimes V$ .

Step 1. Form the projectors $P_{\pm}=\tfrac12(\mathrm{id}\pm\tau)$ . Since $\tau^2=\mathrm{id}$ , one has $P_+^2=P_+$ , $P_-^2=P_-$ , $P_+P_-=0$ , and $P_++P_-=\mathrm{id}$ . Why: an involution splits a space into its $\pm1$ eigenspaces, and $P_\pm$ are the eigenprojectors — the factor $\tfrac12$ is why the characteristic must differ from $2$ .

Step 2. Decompose $T=P_+T+P_-T$ : in components $T^{(ij)}=\tfrac12(T^{ij}+T^{ji})$ is symmetric and $T^{[ij]}=\tfrac12(T^{ij}-T^{ji})$ antisymmetric, and the split is unique. Why: $P_+T$ and $P_-T$ lie in eigenspaces for distinct eigenvalues, meeting only at zero.

Step 3. Count: $\dim\mathrm{Sym}^2V=\binom{n+1}{2}=\tfrac{n(n+1)}2$ (unordered pairs with repetition), $\dim\Lambda^2V=\binom{n}{2}=\tfrac{n(n-1)}2$ (strict pairs). Why: a symmetric array is fixed by its diagonal plus upper triangle, an antisymmetric one by its strict upper triangle with zero diagonal — the two triangles shaded in the figure.

Step 4. Verify completeness, $\tfrac{n(n+1)}2+\tfrac{n(n-1)}2=\tfrac{n^2+n+n^2-n}2=n^2=\dim(V\otimes V)$ . Why: for $n=4$ this is $10+6=16$ — the metric's ten components and the field strength's six — exactly the relativity bookkeeping.

Limiting cases

The tensor product specializes and degenerates in ways that double as sanity checks on the general theory.

Case	Result	Meaning
One factor is the ground field	$V\otimes_{\mathbb{F}}\mathbb{F}\;\cong\;V$	Tensoring with the field does nothing, $v\otimes\lambda\mapsto\lambda v$ : $\mathbb{F}$ is the unit object for $\otimes$ , as $1$ is for multiplication.
Simple versus general tensors	$\dim(V\otimes W)=mn$ , simple tensors a measure-zero cone	Rank-one matrices are vanishingly rare; a generic tensor is a genuine sum of simple pieces, so entanglement is the rule, not the exception.
Dual and Hom identifications	$\mathrm{Hom}(V,W)\cong W\otimes V^,\ (V\otimes W)^\cong V^\otimes W^$	Canonical in finite dimensions; in infinite dimensions only an injection, until $\otimes$ is completed to a topological tensor product.
Symmetric–antisymmetric split (rank 2)	$V\otimes V=\mathrm{Sym}^2V\oplus\Lambda^2V$	Valid when $\operatorname{char}\mathbb{F}\neq2$ ; $\tfrac{n(n+1)}2+\tfrac{n(n-1)}2=n^2$ . The parts never mix, so the adjectives are basis-free.
Higher rank, $k\ge3$	$V^{\otimes k}$ has $n^k$ components; Sym and $\Lambda$ no longer exhaust it	The swap group is now $S_k$ , and mixed-symmetry pieces (Young diagrams) appear alongside the fully symmetric and antisymmetric parts.

Common trapThe tensor product is not the direct sum, and ⊗ is not an ordinary product

The single most common error is to conflate $V\otimes W$ with $V\times W$ or $V\oplus W$ . The direct sum has dimension $m+n$ and its elements are pairs $(v,w)$ that add slot by slot; the tensor product has dimension $mn$ and most of its elements are sums of simple tensors $v\otimes w$ that cannot be collapsed to a single product — the "entangled" ones. A second trap is treating $\otimes$ as a commutative multiplication: it is bilinear, so $(2v)\otimes w=v\otimes(2w)=2(v\otimes w)$ , but $v\otimes w\ne w\otimes v$ in general. Finally, do not read a tensor's transformation law as its definition — the law is a consequence of which slots the object occupies, not an axiom imposed by hand.

Worked examples

1 · Splitting a rank-2 tensor and counting components (fully worked)

Take $V$ with $\dim V=2$ and the rank-2 tensor with components $T^{ij}=\begin{pmatrix}1&2\\4&3\end{pmatrix}$ . Decompose it into symmetric and antisymmetric parts, confirm the component counts, and then read off the spacetime ( $n=4$ ) counts that separate the metric from the field strength.

Step 1. Form the symmetric part $T^{(ij)}=\tfrac12(T^{ij}+T^{ji})$ . With transpose $T^{ji}=\begin{pmatrix}1&4\\2&3\end{pmatrix}$ , this is $\tfrac12\begin{pmatrix}2&6\\6&6\end{pmatrix}=\begin{pmatrix}1&3\\3&3\end{pmatrix}$ , unchanged under swapping indices. The symmetric part is $\begin{pmatrix}1&3\\3&3\end{pmatrix}$ .

Step 2. Form the antisymmetric part $T^{[ij]}=\tfrac12(T^{ij}-T^{ji})=\tfrac12\begin{pmatrix}0&-2\\2&0\end{pmatrix}=\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ . Its diagonal is forced to zero and the off-diagonal entries are negatives, so it carries one independent number. The antisymmetric part is $\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ .

Step 3. Check reconstruction and counts. The parts sum to $\begin{pmatrix}1&3\\3&3\end{pmatrix}+\begin{pmatrix}0&-1\\1&0\end{pmatrix}=\begin{pmatrix}1&2\\4&3\end{pmatrix}=T^{ij}$ . For $n=2$ the dimensions are $\tfrac{2\cdot3}2=3$ symmetric and $\tfrac{2\cdot1}2=1$ antisymmetric, $3+1=4=2^2$ . The split is $4=3+1$ for $n=2$ .

Step 4. Scale to spacetime. For $n=4$ the same formulas give $\dim\mathrm{Sym}^2V=\tfrac{4\cdot5}2=10$ and $\dim\Lambda^2V=\tfrac{4\cdot3}2=6$ , with $10+6=16=4^2$ . The symmetric ten are the independent components of a metric $g_{\mu\nu}$ ; the antisymmetric six are the components of $F_{\mu\nu}$ — three electric and three magnetic. In four dimensions $16=10+6$ : metric versus field strength.

2 · The trace as a contraction, and its basis independence (fully worked)

Regard the operator $A=\begin{pmatrix}2&1&0\\0&3&1\\1&0&4\end{pmatrix}$ as a $(1,1)$ tensor $A^i{}_j\in V\otimes V^*$ with $\dim V=3$ . Contract its two slots, and confirm that the result survives a change of basis.

Step 1. Contraction pairs the single upper slot with the single lower slot, $C(A)=A^i{}_i=A^1{}_1+A^2{}_2+A^3{}_3$ . Reading the diagonal, $C(A)=2+3+4$ . The contraction is $\operatorname{tr}A=9$ .

Step 2. See why this is coordinate-free. A $(1,1)$ tensor transforms by similarity $A\mapsto SAS^{-1}$ , and $\operatorname{tr}(SAS^{-1})=\operatorname{tr}(S^{-1}SA)=\operatorname{tr}A$ by cyclicity, so $9$ belongs to the operator, not the basis. The upper index transforms by the Jacobian and the lower by its inverse; contracting cancels the two exactly. The contracted scalar $9$ is basis-independent.

Step 3. Connect it to the tensor picture. Writing $A=\sum_{i,j}A^i{}_j\,e_i\otimes\varepsilon^j$ and applying $C(e_i\otimes\varepsilon^j)=\varepsilon^j(e_i)=\delta^j_i$ collapses the sum to $\sum_i A^i{}_i$ — the same $9$ , seen as evaluating each dual-basis covector on its partner. Contraction $=$ trace $=$ the canonical $V\otimes V^*\to\mathbb{F}$ pairing.

Physics of the portrait

The tensor product arrived decades before physics knew it needed it. Its inventor was not a physicist but a Prussian schoolteacher whose masterwork was ignored for a generation, only to become the silent scaffolding of relativity, quantum mechanics, and gauge theory.

Origins · 1 minHermann Grassmann and how this was discovered

Hermann Grassmann

1809–1877 · Stettin, Prussia · founder of multilinear and exterior algebra

Hermann Günther Grassmann was a gymnasium teacher in Stettin who never held a university chair, yet in his 1844 Ausdehnungslehre (theory of extension) he built, essentially from nothing, the first systematic calculus of multilinear quantities — the ancestor of the tensor product, the exterior product, and linear algebra as we now teach it. He treated vectors as abstract entities to be added and multiplied, defined an antisymmetric product that is the modern wedge $\wedge$ , and grasped dimension, independence, and subspaces years before they became standard. The work was so far ahead of its notation-poor time that mathematicians found it impenetrable; discouraged, Grassmann turned to linguistics, producing a celebrated translation of the Rig-Veda and a law of Indo-European phonology (Grassmann's law) that made him more famous among philologists in his lifetime. Vindication came late: Hankel, Clifford, and Gibbs recognized in the 1870s–80s that his extension algebra contained the vector and exterior calculus physics was groping toward, and Cartan's differential forms are Grassmann's antisymmetric products made into a calculus. When Ricci and Levi-Civita assembled the tensor calculus Einstein would use for gravity, and when quantum theory placed composite systems in tensor products of Hilbert spaces, both built on the foundation Grassmann laid alone, in a schoolroom, for an audience not yet ready.

History

1844

— Grassmann's Ausdehnungslehre founds multilinear and exterior algebra

1884

— Ricci-Curbastro begins the absolute differential (tensor) calculus

1900

— Ricci and Levi-Civita publish the tensor calculus Einstein will adopt

1916

— Einstein casts general relativity entirely in tensor language

1938

— Whitney characterizes the tensor product by its universal factorization property

Going further

Texts. Roman, Advanced Linear Algebra, 3rd ed., Ch. 14 — the cleanest universal-property treatment · Stone & Goldbart, Mathematics for Physics, Ch. 10 · Nakahara, Geometry, Topology and Physics, 2nd ed., §2.5 · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 4 · Hassani, Mathematical Physics, Ch. 25–26 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, tensor chapter. Key numerical checks: (1) $\dim(V\otimes W)=mn$ , and a $(2,1)$ tensor at $n=4$ has $n^3=64$ components; (2) the rank-2 split satisfies $\tfrac{n(n+1)}2+\tfrac{n(n-1)}2=n^2$ , giving $10+6=16$ at $n=4$ ; (3) the matrix $\begin{pmatrix}1&2\\4&3\end{pmatrix}$ decomposes as $\begin{pmatrix}1&3\\3&3\end{pmatrix}+\begin{pmatrix}0&-1\\1&0\end{pmatrix}$ ; (4) the trace $A^i{}_i=2+3+4=9$ is invariant under $A\mapsto SAS^{-1}$ .

Prerequisites. Vector spaces and linear maps · Dual Spaces and Linear Functionals · Bilinear and multilinear maps. Leads to. Covariant and Contravariant Tensors · The Metric Tensor · Tensor Contraction and the Trace · Differential Forms · Manifolds and Tangent Spaces.

Ask the tutor: Ask me to walk through why a specific physical object — the metric, $F_{\mu\nu}$ , a spin-coupled state, or the stress tensor — lands in exactly its tensor-product space, and we will count its slots and components together.

What is the difference between V ⊗ W and V ⊕ W?

They are built from the same two spaces but answer different questions. The direct sum $V\oplus W$ collects a vector from each space independently: its elements are pairs $(v,w)$ , its dimension is $m+n$ , and it fits when two systems merely coexist. The tensor product $V\otimes W$ collects interactions: its elements are sums of products $v\otimes w$ , its dimension is $mn$ , and it fits when a quantity depends bilinearly on one input from each space — a bilinear form, a joint quantum state, a matrix. Quick tell: if the natural map into your space is linear in the pair, you want $\oplus$ ; if linear in each argument separately, you want $\otimes$ .

Why is v ⊗ w called a "simple" tensor, and what does that have to do with entanglement?

A simple (or decomposable) tensor can be written as a single product $v\otimes w$ ; as a matrix it has rank one. Most tensors are not simple — they are sums like $e_1\otimes e_1+e_2\otimes e_2$ , the identity matrix, rank two, provably not any single $v\otimes w$ . In quantum mechanics the joint state of two systems lives in $\mathcal H_A\otimes\mathcal H_B$ ; a simple tensor $|\psi\rangle\otimes|\phi\rangle$ is a separable (product) state, while a non-simple tensor is an entangled state whose parts have no independent description. So "most tensors are not simple" is, word for word, "most joint states are entangled" — the backbone of quantum information.

What does the universal property actually buy me in practice?

It converts hard constructions into one-line checks. To define a linear map out of $V\otimes W$ you never worry about the messy relations among simple tensors: you just write a bilinear map on $V\times W$ , and the universal property hands you a unique linear map for free. That is exactly how the trace, the determinant, contraction, and every canonical identification in this article are built. It also settles uniqueness: because the property pins $V\otimes W$ down up to a single canonical isomorphism, you may use any construction — quotient of a free space, matrices, bilinear forms — and know they all agree.

Why do tensors carry upper and lower indices?

Because a general tensor is assembled from two kinds of slot, and the index position records which. An upper (contravariant) index labels a factor of $V$ and transforms by the Jacobian, like a vector's components; a lower (covariant) index labels a factor of $V^*$ and transforms by the inverse Jacobian, like a functional. The identification $\mathrm{End}(V)\cong V\otimes V^*$ is why an operator has one of each, $A^i{}_j$ : it reads a vector (lower slot) and returns one (upper slot). The point of the opposite laws is that contracting an upper against a lower index cancels the two Jacobians, leaving an invariant — which is why physical laws carry balanced indices.

Is the tensor product commutative or associative?

There are canonical isomorphisms $V\otimes W\cong W\otimes V$ (the swap) and $(U\otimes V)\otimes W\cong U\otimes(V\otimes W)$ , so up to these natural identifications $\otimes$ is both commutative and associative — which is what lets us write $V^{\otimes k}$ without parentheses. But equality is too strong: $v\otimes w$ and $w\otimes v$ are different elements in different (isomorphic) spaces, and treating them as literally equal is exactly the mistake that collapses $\Lambda^2V$ , where antisymmetry $v\otimes w=-w\otimes v$ is the whole content. So: commutative and associative up to canonical isomorphism, never as a naive equality.

Does any of this survive in infinite dimensions, as for Hilbert spaces?

Partly, and the caveats matter. The algebraic tensor product is defined in any dimension and the universal property holds verbatim. What breaks are the finite-dimensional identifications: $\mathrm{Hom}(V,W)\cong W\otimes V^*$ becomes only an inclusion (the algebraic product captures just the finite-rank operators), and $(V\otimes W)^*\cong V^*\otimes W^*$ weakens to an injection. Quantum mechanics therefore uses the completed (Hilbert) tensor product $\mathcal H_A\,\hat\otimes\,\mathcal H_B$ , the completion in the natural norm, so that infinite superpositions of product states — the generic entangled states — are included. Same algebra, plus a completion step.