Tensor Algebra and Index Notation

Tensor Algebra and Index Notation is the systematic calculus of multilinear objects expressed through indexed components, where covariant and contravariant indices transform oppositely under change of basis. Key result: Tensors are defined intrinsically as multilinear maps, but index notation with the Einstein summation convention provides a coordinate-free-equivalent computational machinery in which contraction, raising/lowering, and symmetrization become algebraic operations. Applies in: differential geometry, general relativity, continuum mechanics, quantum field theory, machine learning. Who/when: Ricci-Curbastro and Levi-Civita (1900) systematized the "absolute differential calculus"; Einstein popularized index notation (1916).

A tensor of type $(p, q)$ over a vector space $V$ is a multilinear map taking $p$ covectors and $q$ vectors to a scalar; equivalently, an element of the tensor product $p V \otimes \dots \otimes V \otimes q V^{*} \otimes \dots \otimes V^{*}$ . Tensor algebra matters because it furnishes the only consistent language for quantities that transform multilinearly under change of coordinates — stress, curvature, electromagnetic field, metric — and guarantees that physical laws written as tensor equations hold in every frame. Index notation encodes the abstract tensor product into upper (contravariant) and lower (covariant) slots, with the Einstein convention automating contraction. A surprising fact: the entire apparatus of raising, lowering, and the Kronecker delta acting as identity is forced once you accept that components must transform so that the underlying geometric object is invariant — the notation is not arbitrary bookkeeping but a near-rigid consequence of multilinearity.

The Core Insight

A tensor is an invariant object; its components are not. The genius of index notation is that it tracks how components must transform by attaching to each index a transformation slot. An expression is a valid tensor equation if and only if every free index matches in position (up/down) and name on both sides, and every repeated index appears exactly once up and once down. This "index balance" is a syntactic guarantee of geometric meaning — a typechecker for physics.

Formal Statement

Tensor of type $(p,q)$

Let $V$ be an $n$ -dimensional vector space over $R$ with dual $V^{*}$ . A tensor of type $(p, q)$ (contravariant rank $p$ , covariant rank $q$ ) is a multilinear map

T : p V^{*} \times \dots \times V^{*} \times q V \times \dots \times V ⟶ R .

⟳Multilinearity means linearity in each argument separately; the space of all such maps is canonically isomorphic to

V^{\otimes p} \otimes (V^{*})^{\otimes q}

Given a basis ${e_{i}}$ of $V$ and dual basis ${e^{j}}$ of $V^{*}$ with $e^{j} (e_{i}) = δ_{i}^{j}$ , the components are

T^{i_{1} \dots i_{p}}_{j_{1} \dots j_{q}} = T (e^{i_{1}}, \dots, e^{i_{p}}, e_{j_{1}}, \dots, e_{j_{q}}) .

⟳Evaluate the multilinear map on basis covectors (upper slots) and basis vectors (lower slots); multilinearity recovers

T

on arbitrary arguments.

Einstein Summation Convention

Whenever an index appears once as a superscript and once as a subscript in a single term, summation over its range $1, \dots, n$ is implied:

v^{i} e_{i} \equiv i = 1 \sum n v^{i} e_{i}, A^{i}_{j} v^{j} \equiv j = 1 \sum n A^{i}_{j} v^{j} .

⟳This is purely notational; it suppresses

\sum

for contractions. A repeated index is "dummy" (summed); an unrepeated index is "free" and labels a component.

Transformation Law

Under a change of basis $e_{i^{'}} = A^{i}_{i^{'}} e_{i}$ with inverse Jacobian $\tilde{A}^{i^{'}}_{i}$ (so $A^{i}_{i^{'}} \tilde{A}^{i^{'}}_{k} = δ_{k}^{i}$ ), the components of a $(p, q)$ -tensor transform as

T^{i_{1}^{'} \dots i_{p}^{'}}_{j_{1}^{'} \dots j_{q}^{'}} = \tilde{A}^{i_{1}^{'}}_{i_{1}} \dots \tilde{A}^{i_{p}^{'}}_{i_{p}} A^{j_{1}}_{j_{1}^{'}} \dots A^{j_{q}}_{j_{q}^{'}} T^{i_{1} \dots i_{p}}_{j_{1} \dots j_{q}} .

⟳Substitute the basis change into [2] and use multilinearity; upper indices pull the inverse Jacobian (contravariant), lower indices pull the Jacobian (covariant).

Conversely, any indexed array obeying [4] defines a unique tensor.

Quotient Theorem

Let $X^{i_{1} \dots i_{p}}_{j_{1} \dots j_{q}}$ be an array. If for every tensor $A$ the contraction $X^{\dots}_{\dots} A^{\dots}_{\dots}$ (summed over a complementary set of indices) yields a tensor, then $X$ is itself a tensor.

(X^{ij} A_{j} = B^{i} a tensor \forall A_{j} tensor) \Rightarrow X^{ij} is a tensor .

⟳Write the transformation of

B^{i}

and

A_{j}

, equate, and exploit arbitrariness of

A_{j}

to peel off its transformation matrix, leaving the tensor law for

X

Why It Holds / Derivation

Index Balance Forces Invariance

We show that an equation $L^{i}_{j} = R^{i}_{j}$ with matched free indices is frame-independent.

Step 1. Suppose both

L^{i}_{j}

and

R^{i}_{j}

are components of

(1, 1)

-tensors

L, R

. By [4],

L^{i^{'}}_{j^{'}} = \tilde{A}^{i^{'}}_{i} A^{j}_{j^{'}} L^{i}_{j}, R^{i^{'}}_{j^{'}} = \tilde{A}^{i^{'}}_{i} A^{j}_{j^{'}} R^{i}_{j} .

⟳Apply the

(1, 1)

specialization of [4] to each side independently.

Step 2. If

L^{i}_{j} = R^{i}_{j}

in one frame, subtract:

(L^{i}_{j} - R^{i}_{j}) = 0

. Multiply by

\tilde{A}^{i^{'}}_{i} A^{j}_{j^{'}}

L^{i^{'}}_{j^{'}} - R^{i^{'}}_{j^{'}} = \tilde{A}^{i^{'}}_{i} A^{j}_{j^{'}} (L^{i}_{j} - R^{i}_{j}) = 0.

⟳Linearity of the transformation in the components; a zero array maps to a zero array.

Step 3. Hence

L^{i^{'}}_{j^{'}} = R^{i^{'}}_{j^{'}}

in every frame. The equation is invariant.

■

Contraction is tensorial. Set

j = i

T^{i}_{j}

and sum:

T^{i}_{i} = \tilde{A}^{i^{'}}_{i} A^{j}_{i^{'}} T^{' ?} ⟹ trace transforms as a scalar .

The key identity

\tilde{A}^{i^{'}}_{i} A^{i}_{j^{'}} = δ_{j^{'}}^{i^{'}}

collapses the two transformation matrices, leaving a

(0, 0)

-tensor (scalar). This is why contraction lowers type by

(1, 1)

and yields an invariant.

⟳Contracting one upper with one lower index pairs a Jacobian with its inverse, which annihilate to a Kronecker delta.

Gregorio Ricci-Curbastro | 1853–1925

Ricci-Curbastro developed the absolute differential calculus between 1884 and 1900, the systematic theory of how differential expressions could be made invariant under arbitrary coordinate changes. With his student Tullio Levi-Civita, he published Méthodes de calcul différentiel absolu et leurs applications (1900), introducing the covariant/contravariant index machinery and what would become the Ricci tensor. The work languished until 1912–1915, when Einstein, guided by his friend Marcel Grossmann, recognized that this calculus was precisely the mathematics needed for a generally covariant theory of gravitation. Ricci's notation — upper and lower indices — survives essentially unchanged in modern physics.

Building Intuition

Think of a vector $v$ as an arrow that exists independently of any grid you draw. When you switch from a coarse grid to a fine grid, the basis vectors shrink but the arrow is unchanged, so its components must grow to compensate — they transform contra (against) the basis. Hence "contravariant," upper index. A covector (gradient, a stack of level surfaces) does the opposite: refine the grid and its components shrink with the basis — "covariant," lower index. The Einstein convention's "one up, one down" rule is the algebraic statement that an upper-index growth and a lower-index shrinkage cancel, producing a number everyone agrees on. Raising and lowering indices with the metric $g_{ij}$ is just the geometric act of converting "arrow" descriptions to "stack" descriptions using the inner product to measure projection.

Covariant vs. contravariant index placement and the contraction that produces an invariant scalar

Worked Examples

Contracting the Inertia Tensor with Angular Velocity

Given: A rigid body has inertia tensor (in kg·m², a

(0, 2)

-symmetric tensor; here Cartesian so up/down coincide)

I_{ij} = 201030104, ω^{j} = (1, 2, 1) rad/s .

Find: Angular momentum

L_{i} = I_{ij} ω^{j}

and rotational kinetic energy

E = \frac{1}{2} I_{ij} ω^{i} ω^{j}

. Step 1 — contract once (lower one free index $i$ remains):

L_{i} = I_{ij} ω^{j} .

⟳Sum over

j = 1, 2, 3

for each fixed

i

; one free index ⇒ result is a covector.

L_{1} = 2 (1) + 0 (2) + 1 (1) = 3

L_{2} = 0 (1) + 3 (2) + 0 (1) = 6

L_{3} = 1 (1) + 0 (2) + 4 (1) = 5

Step 2 — contract again with $ω^{i}$ (no free index ⇒ scalar):

2 E = ω^{i} L_{i} = 1 (3) + 2 (6) + 1 (5) = 20.

⟳Full contraction of a

(0, 2)

-tensor with two copies of a vector gives an invariant scalar.

Answer:

L_{i} = (3, 6, 5) kg\cdotm^{2} / s, E = 10 J

Verifying a Tensor Equation by Index Balance

Given: Claim — in Cartesian coordinates the Maxwell stress combination

T^{ij} = E^{i} E^{j} - \frac{1}{2} δ^{ij} E^{k} E_{k}

is symmetric and its divergence-related contraction is well posed. Step 1 — check index balance.

T^{ij}

has two free upper indices on both sides:

$E^{i} E^{j}$ : free $i, j$ . ✓
$δ^{ij} E^{k} E_{k}$ : $δ^{ij}$ supplies free $i, j$ ; $E^{k} E_{k}$ has $k$ up and down (dummy, summed). ✓

Both terms carry exactly free $(i, j)$ — the equation is well formed.

T^{ij} = E^{i} E^{j} - \frac{1}{2} δ^{ij} E^{k} E_{k} .

⟳Every free index appears in the same position on every additive term; dummy index

k

is balanced up/down.

Step 2 — symmetry. Swap

i \leftrightarrow j

E^{j} E^{i} - \frac{1}{2} δ^{j i} E^{k} E_{k} = T^{ij}

since

E^{i} E^{j}

and

δ^{ij}

are symmetric.

T^{ij} = T^{j i} .

⟳Product of components commutes; Kronecker delta is symmetric.

Step 3 — trace.

T^{i}_{i} = E^{i} E_{i} - \frac{1}{2} δ^{i}_{i} E^{k} E_{k}

. With

δ^{i}_{i} = 3

in 3D:

T^{i}_{i} = E^{2} - \frac{3}{2} E^{2} = - \frac{1}{2} E^{2} .

⟳

δ^{i}_{i}

sums the diagonal of the identity = dimension

n = 3

Answer:

T^{ij} is symmetric, tr T = - \frac{1}{2} E^{2} .

Common Misconceptions

A repeated index summed twice up or twice down is NOT a tensor contraction. Writing

A^{ii}

(both up) or

g^{ij} g^{ij}

(no balance) violates the convention. In Cartesian coordinates where the metric is the identity, such expressions accidentally give the right number, masking the error — but they fail to transform correctly under general coordinate changes. Always pair one upper with one lower index. Equally, a free index must appear in the same position on both sides:

v^{i} = w_{i}

is meaningless (a vector cannot equal a covector); you need

v^{i} = g^{ij} w_{j}

Consider in 2D the array $Γ^{i}_{j k}$ of Christoffel symbols. It carries three indices and looks like a $(1, 2)$ -tensor, but it is not a tensor. Under coordinate change it picks up an inhomogeneous second-derivative term:

Γ^{' i^{'}}_{j^{'} k^{'}} = \tilde{A}^{i^{'}}_{i} A^{j}_{j^{'}} A^{k}_{k^{'}} Γ^{i}_{j k} + \tilde{A}^{i^{'}}_{i} \frac{\partial ^{2} x ^{i}}{\partial x ^{j^{'}} \partial x ^{k^{'}}} .

⟳Differentiate the basis transformation; the extra term comes from the derivative acting on the position-dependent Jacobian.

The presence of the second term means

Γ

violates [4]. This is precisely why partial derivatives of tensors are not tensors, motivating the covariant derivative, whose Christoffel correction cancels the inhomogeneous piece. A correct array passing index balance can still fail to be a tensor.

Connections

Tensor algebra is the static, point-wise skeleton on which Differential Geometry builds the dynamics: smoothly varying tensors become tensor fields, and the failure of $\partial_{k}$ to be tensorial (the Christoffel Symbols counterexample) leads to the Covariant Derivative and Curvature Tensor. The metric tensor $g_{ij}$ supplies the inner product enabling Raising and Lowering Indices, linking $V$ and $V^{*}$ — this is the index-notation face of the Musical Isomorphism. Symmetrization and antisymmetrization of indices generate the Exterior Algebra of differential forms, where antisymmetric $(0, q)$ -tensors become $q$ -forms. The abstract multilinear-map definition connects to the universal property in Multilinear Algebra and the Tensor Product of vector spaces.

Real-World Applications

General Relativity — Einstein's field equations $G_{μν} = 8 π G T_{μν}$ equate the Einstein curvature tensor to the stress-energy tensor; index balance guarantees general covariance across all reference frames.
Continuum Mechanics — The Cauchy stress tensor $σ_{ij}$ relates traction to surface normal via $t_{i} = σ_{ij} n^{j}$ ; the strain tensor $ε_{ij}$ and stiffness $C_{ij k l}$ encode Hooke's law $σ_{ij} = C_{ij k l} ε_{k l}$ .
Electromagnetism — The field-strength tensor $F_{μν} = \partial_{μ} A_{ν} - \partial_{ν} A_{μ}$ unifies $E$ and $B$ ; Maxwell's equations become $\partial_{μ} F^{μν} = μ_{0} J^{ν}$ .
Crystallography & Material Science — Piezoelectric, dielectric, and elastic tensors classify anisotropic response; Neumann's principle constrains tensor components by crystal symmetry.
Machine Learning — Deep learning frameworks (TensorFlow, PyTorch) compute with multi-dimensional arrays; Einstein summation (einsum) is the direct computational descendant of index notation for batched contractions and attention mechanisms.
Fluid Dynamics — The Reynolds stress tensor and rate-of-strain tensor $S_{ij} = \frac{1}{2} (\partial_{i} u_{j} + \partial_{j} u_{i})$ govern turbulence modeling.
Quantum Information — Tensor networks (MPS, PEPS) represent many-body wavefunctions; contraction of high-rank tensors underlies DMRG and tensor-train decompositions.

Historical Development

1827 — Carl Friedrich Gauss studies intrinsic surface geometry; the first metric quadratic form $d s^{2} = E d u^{2} + 2 F d u d v + G d v^{2}$ foreshadows the metric tensor.
1854 — Bernhard Riemann's Habilitation lecture introduces $n$ -dimensional manifolds and the curvature concept, demanding a multilinear formalism.
1869 — Elwin Bruno Christoffel introduces the symbols bearing his name while studying when two quadratic differential forms are equivalent.
1884–1894 — Gregorio Ricci-Curbastro develops the absolute differential calculus, formalizing covariant and contravariant transformation.
1900 — Ricci and Tullio Levi-Civita publish their definitive memoir systematizing tensor calculus and index notation.
1912–1915 — Marcel Grossmann teaches Einstein the Ricci calculus; Einstein adopts it for general relativity.
1916 — Einstein introduces the summation convention in his foundational GR paper, eliminating explicit summation signs.
1917 — Levi-Civita and Hermann Weyl develop parallel transport and the connection, clarifying the non-tensorial nature of $Γ$ .
1923 — Élie Cartan reformulates tensor calculus via differential forms and moving frames, complementing index methods.

Subtleties & Limits

Index notation's power obscures genuine conceptual subtleties. (1) Coordinate vs. abstract indices: Penrose's abstract index notation uses indices as type markers without committing to a basis, reconciling the computational clarity of indices with coordinate-freedom — but the precise category-theoretic semantics of mixing abstract and coordinate indices remains a pedagogical minefield. (2) Density and weight: Tensor densities (e.g., $- g$ , the Levi-Civita symbol $ε_{ij k}$ ) transform with Jacobian determinant powers and are not tensors; tracking weights consistently is error-prone and lacks universal notational standards. (3) Higher categories: Generalizing tensor contraction to graphical/string-diagram calculus and to symmetric monoidal categories raises open questions about canonical forms and the complexity of optimal tensor-network contraction — the latter is NP-hard, a live problem in computational physics and machine learning. (4) Infinite dimensions: In infinite-dimensional or operator-algebraic settings, $V^{**} \neq = V$ and the clean raising/lowering duality breaks, so index notation requires careful reinterpretation.