Tensors on smooth manifolds

Intuition [Beginner]

A vector is an arrow with direction and magnitude. On a manifold, vectors live at a single point, inside the tangent space at that point. The tangent space is a flat copy of ${\mathbb{R}}^n$ attached to each point, giving you a place to put arrows. Velocity at a point, acceleration at a point -- these are tangent vectors.

But vectors are only half the story. A covector (also called a one-form) is a machine that eats a vector and produces a number. If a vector is "a thing with direction," a covector is "a question you can ask about a thing with direction." The gradient of a function is a covector: it asks "how much does this function change if I move in that direction?"

A tensor of type $(r,s)$ is a machine with $r$ slots for covectors and $s$ slots for vectors. You feed it the right ingredients and it returns a single real number. The answer depends linearly on each input -- doubling any one input doubles the result. A vector is a $(1,0)$-tensor (one covector slot). A covector is a $(0,1)$-tensor (one vector slot). A matrix, in the right context, behaves as a $(1,1)$-tensor.

The object that makes general relativity work is the metric tensor $g$, a $(0,2)$-tensor. It takes two vectors at the same point and returns a number we interpret as their inner product. On flat spacetime, $g$ is the Minkowski metric $\eta$, which gives each vector its squared length -- negative for timelike vectors, positive for spacelike ones, zero for lightlike ones. On a curved manifold, $g$ varies from point to point, encoding how the geometry bends.

Three reasons GR needs this language. The gravitational field is the metric tensor field -- a smooth assignment of a metric tensor to each point of spacetime. The matter content of the universe is encoded in the stress-energy tensor $T$, also a $(0,2)$-tensor field. The curvature of spacetime is described by the Riemann curvature tensor, a $(1,3)$-tensor field. Einstein's equation $G=8\pi T$ is a tensor equation relating two $(0,2)$-tensor fields. Without tensors you cannot write it down.

The crucial property is coordinate independence. A tensor is not its components. The components $g_{\mu \nu }$ are the numbers you get when you evaluate the metric on basis vectors of a specific coordinate system. Change the coordinates and every component changes, but the tensor itself -- the multilinear map -- does not. A tensor equation that holds in one coordinate system holds in every coordinate system.

This is not cosmetic. Coordinate systems are human choices -- labels we slap on spacetime to do calculations. Physics must not depend on labels. Tensors are the mathematical objects that transform under coordinate changes in exactly the right way to keep physical statements label-independent. This is why GR is formulated in tensor language.

A tensor field is a smooth assignment of a tensor to each point. "Smooth" means the components vary smoothly as you move from point to point. The metric tensor field $g$ is the central object of Riemannian and Lorentzian geometry -- it tells you how to measure lengths, angles, areas, volumes, and curvature.

Three misconceptions are worth flagging. A tensor is not a matrix. A matrix is a grid of numbers; a tensor is a multilinear map. In a specific basis you can represent a $(0,2)$-tensor as a matrix, but the representation changes with the basis while the tensor does not. Second, not every $(0,2)$-tensor is a metric. A metric must be symmetric and non-degenerate, and for GR it must have the right signature -- one negative and three positive eigenvalues. A random $(0,2)$-tensor need not satisfy any of these. Third, a tensor lives at a point. A tensor field lives on the whole manifold. Conflating the two is like conflating "a number" with "a function."

Visual [Beginner]

Picture a tensor as a machine with input slots. A $(0,2)$-tensor has two arrow-shaped slots -- you insert two tangent vectors and the machine outputs a single real number. The metric tensor $g$ is exactly this: a device for turning pairs of vectors into inner products.

At each point $p$ of a manifold, the tangent space $T_{p}M$ is a flat vector space. The metric tensor assigns an inner product to this space. In flat Minkowski spacetime every tangent space gets the same inner product $\eta$ with signature $(-,+,+,+)$. On a curved manifold the inner product varies from point to point -- and that variation is the curvature.

The diagram captures the operational meaning: the metric tensor is not a grid of numbers but a rule for pairing vectors. The grid (the matrix of components $g_{\mu \nu }$) is what you get when you choose a specific basis for the tangent space. A different basis gives a different grid but the same pairing rule.

Worked example [Beginner]

The Schwarzschild metric describes spacetime outside a spherical mass $M$. In coordinates $(t,r,\theta ,\phi )$ the line element is

$$d{s}^2=-(1-\frac{r_s}{r})c^{2}d{t}^2+(1-\frac{r_s}{r}{)}^{-1}d{r}^2+r^{2}d{\theta }^2+r^2{\sin }^2\theta d{\phi }^2,$$

where $r_s=2GM/c^2$ is the Schwarzschild radius. The metric tensor $g_{\mu \nu }$ has four non-zero components, all diagonal:

$$g_{tt}=-(1-\frac{r_s}{r}),g_{rr}=(1-\frac{r_s}{r}{)}^{-1},g_{\theta \theta }=r^2,g_{\phi \phi }=r^2{\sin }^2\theta .$$

This is Lorentzian for $r>r_s$: the $tt$ component is negative while $g_{rr}$, $g_{\theta \theta }$, $g_{\phi \phi }$ are all positive. One negative eigenvalue and three positive gives the signature $(-,+,+,+)$, the same sign pattern as flat Minkowski spacetime.

The inverse metric $g^{\mu \nu }$ satisfies $g^{\mu \rho }{g}_{\rho \nu }={\delta }_{\nu }^{\mu }$. Since $g_{\mu \nu }$ is diagonal, each component of the inverse is just the reciprocal:

$$g^{tt}=-(1-\frac{r_s}{r}{)}^{-1},g^{rr}=(1-\frac{r_s}{r}),g^{\theta \theta }=\frac{1}{r^2},g^{\phi \phi }=\frac{1}{r^2{\sin }^2\theta }.$$

The determinant follows from multiplying the diagonal entries:

$$\det (g)=g_{tt}\cdot g_{rr}\cdot g_{\theta \theta }\cdot g_{\phi \phi }=-r^4{\sin }^2\theta .$$

This vanishes at $r=0$ (the curvature singularity) and at $\theta =0,\pi$ (the coordinate singularity at the poles, where the spherical coordinate system degenerates).

Check your understanding [Beginner]

Formal definition [Intermediate+]

Let $M$ be an $n$-dimensional smooth manifold, $T_{p}M$ the tangent space at $p\in M$, and $T_p^{\ast }M$ the cotangent space (the dual of $T_{p}M$). A tensor of type $(r,s)$ at $p$ is a multilinear map

$$T:\underset{r}{\underbrace{T_p^{\ast }M\times \cdots \times T_p^{\ast }M}}\times \underset{s}{\underbrace{T_{p}M\times \cdots \times T_{p}M}}\to \mathbb{R}.$$

The set of all such tensors is denoted $T_s^r(T_{p}M)$. It is a vector space of dimension $n^{r+s}$.

If $\{e_{\mu }\}$ is a basis for $T_{p}M$ and $\{e^{\mu }\}$ is the dual basis (defined by $e^{\mu }(e_{\nu })={\delta }_{\nu }^{\mu }$), then the components of $T$ in this basis are

$$T^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}=T(e^{{\mu }_1},\dots ,e^{{\mu }_r},e_{{\nu }_1},\dots ,e_{{\nu }_s}).$$

The tensor product of $A\in T_{s_1}^{r_1}(T_{p}M)$ and $B\in T_{s_2}^{r_2}(T_{p}M)$ is the tensor $A\otimes B\in T_{s_1+s_2}^{r_1+r_2}(T_{p}M)$ defined by

$$(A\otimes B)({\alpha }^1,\dots ,{\alpha }^{r_1+r_2},V_1,\dots ,V_{s_1+s_2})=A({\alpha }^1,\dots ,{\alpha }^{r_1},V_1,\dots ,V_{s_1})\cdot B({\alpha }^{r_1+1},\dots ,{\alpha }^{r_1+r_2},V_{s_1+1},\dots ,V_{s_1+s_2}).$$

Under a coordinate change $x^{\mu }\to x^{{\mu }^{\prime }}$, the components of a $(r,s)$-tensor transform as

$$T^{{\mu }_1^{\prime }\cdots {\mu }_r^{\prime }}{}_{{\nu }_1^{\prime }\cdots {\nu }_s^{\prime }}=\frac{\partial x^{{\mu }_1^{\prime }}}{\partial x^{{\mu }_1}}\cdots \frac{\partial x^{{\mu }_r^{\prime }}}{\partial x^{{\mu }_r}}\frac{\partial x^{{\nu }_1}}{\partial x^{{\nu }_1^{\prime }}}\cdots \frac{\partial x^{{\nu }_s}}{\partial x^{{\nu }_s^{\prime }}}{T}^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}.$$

Each upper index picks up a factor of $\partial x^{{\mu }^{\prime }}/\partial x^{\mu }$ and each lower index a factor of $\partial x^{\nu }/\partial x^{{\nu }^{\prime }}$. This transformation law is the defining characteristic that distinguishes tensors from arbitrary arrays of functions.

A tensor field of type $(r,s)$ on $M$ is a smooth section of the tensor bundle $T_s^{r}M={⨆}_{p\in M}{T}_s^r(T_{p}M)$ -- that is, a smooth assignment $p\mapsto T(p)\in T_s^r(T_{p}M)$. Smoothness means that in any coordinate chart the components $T^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}(p)$ are smooth functions of the coordinates.

A metric tensor on $M$ is a $(0,2)$-tensor field $g$ that is (i) symmetric: $g(X,Y)=g(Y,X)$ for all vector fields $X,Y$; (ii) non-degenerate: $g(X,V)=0$ for all $V$ implies $X=0$. The metric is Lorentzian if at every point the matrix $g_{\mu \nu }$ has signature $(-,+,\dots ,+)$ -- one negative and $n-1$ positive eigenvalues. It is Riemannian if the signature is $(+,+,\dots ,+)$. GR uses Lorentzian metrics; the single negative eigenvalue corresponds to the time direction.

The metric provides a canonical isomorphism between $T_{p}M$ and $T_p^{\ast }M$ via index raising and lowering: given a vector $V^{\mu }$, the corresponding covector is $V_{\mu }=g_{\mu \nu }{V}^{\nu }$; given a covector ${\omega }_{\mu }$, the corresponding vector is ${\omega }^{\mu }=g^{\mu \nu }{\omega }_{\nu }$, where $g^{\mu \nu }$ is the inverse matrix satisfying $g^{\mu \rho }{g}_{\rho \nu }={\delta }_{\nu }^{\mu }$.

Key theorem with proof [Intermediate+]

Theorem (Coordinate independence of tensor equations). Let $T$ be a tensor field of type $(r,s)$ on a smooth manifold $M$. If $T=0$ (i.e., $T^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}=0$ for all index values) in one coordinate system, then $T=0$ in every coordinate system.

Proof. Let $x^{\mu }$ be coordinates in which $T^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}=0$, and let $x^{{\mu }^{\prime }}$ be an arbitrary new coordinate system. By the tensor transformation law,

$$T^{{\mu }_1^{\prime }\cdots {\mu }_r^{\prime }}{}_{{\nu }_1^{\prime }\cdots {\nu }_s^{\prime }}=\frac{\partial x^{{\mu }_1^{\prime }}}{\partial x^{{\mu }_1}}\cdots \frac{\partial x^{{\mu }_r^{\prime }}}{\partial x^{{\mu }_r}}\frac{\partial x^{{\nu }_1}}{\partial x^{{\nu }_1^{\prime }}}\cdots \frac{\partial x^{{\nu }_s}}{\partial x^{{\nu }_s^{\prime }}}\underset{=0}{\underbrace{T^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}}}.$$

Each component $T^{{\mu }_1^{\prime }\cdots {\mu }_r^{\prime }}{}_{{\nu }_1^{\prime }\cdots {\nu }_s^{\prime }}$ in the new coordinates is a linear combination (with Jacobian-factor coefficients) of the old components, which are all zero. A linear combination of zeros is zero. Hence $T=0$ in the new coordinates. Since the new coordinate system was arbitrary, $T=0$ in all coordinate systems. $■$

The converse observation is equally important: not every array of functions that vanishes in one coordinate system vanishes in all. The Christoffel symbols ${\Gamma }_{\mu \nu }^{\rho }$ are a standard counterexample. They involve first derivatives of the metric and appear in the geodesic equation, but they do not satisfy the tensor transformation law. In flat space, one can choose Cartesian coordinates in which all ${\Gamma }_{\mu \nu }^{\rho }=0$, yet the same connection has non-zero components in polar coordinates. This non-tensorial behaviour is precisely what makes Christoffel symbols useful -- they measure the failure of a coordinate system to be "straight," which is coordinate-dependent information.

The theorem has an immediate physical corollary: any equation of the form $T=S$, where $T$ and $S$ are tensor fields of the same type, is coordinate-independent if and only if $T-S$ is a tensor field. Einstein's field equations $G_{\mu \nu }=\frac{8\pi G}{c^4}{T}_{\mu \nu }$ have this form -- both sides are $(0,2)$-tensor fields, so the equation is a coordinate-independent statement about the relationship between geometry (the Einstein tensor $G_{\mu \nu }$) and matter (the stress-energy tensor $T_{\mu \nu }$).

The tensoriality of $G_{\mu \nu }$ and $T_{\mu \nu }$ is what guarantees that the field equations make the same physical prediction regardless of the coordinate system used to express them. A non-tensorial equation would not have this property and would not be an acceptable physical law in the framework of GR.

This result also clarifies why the search for physical laws in GR reduces to constructing appropriate tensor fields. If a candidate physical quantity fails to transform as a tensor, it cannot appear as a standalone term in a physical equation -- its value would depend on the observer's coordinate choice, violating the principle of general covariance. The stress-energy tensor $T_{\mu \nu }$, the Einstein tensor $G_{\mu \nu }$, and the Riemann curvature tensor $R^{\rho }{}_{\sigma \mu \nu }$ all pass this test. The Christoffel symbols do not; they appear inside covariant derivatives that produce tensorial results, but never standalone.

Bridge. The coordinate-independence theorem builds toward 13.04.01 pending, where Einstein's field equations $G_{\mu \nu }=(8\pi G/c^4)T_{\mu \nu }$ take the form of a tensor equality -- both sides being $(0,2)$-tensor fields -- and appears again in 13.03.01 pending, where the Riemann curvature tensor is verified to be a genuine tensor despite being constructed from the non-tensorial Christoffel symbols. The central insight is that tensoriality is the mathematical content of the physical principle of general covariance: a tensor equation makes the same prediction in every coordinate system. This is exactly the bridge between the coordinate transformation laws developed in this unit and the invariant geometric quantities of differential geometry, and it generalises the flat-space tensor analysis of 10.05.01 pending to arbitrary curved manifolds.

Exercises [Intermediate+]

Lean formalization [Intermediate+]

Mathlib contains the algebraic and differential-geometric prerequisites for tensor calculus but does not yet assemble them into the physics-specific structures GR requires.

The relevant layers are: LinearAlgebra.TensorProduct and LinearAlgebra.TensorAlgebra for the algebraic tensor product of modules; Mathlib.Geometry.Manifold.Basic, ContMDiffSection, and the tangent/cotangent bundle constructions for smooth manifold theory; and Mathlib.LinearAlgebra.QuadraticForm for indefinite quadratic forms at the linear-algebra level. Together these provide a basis for defining tensor fields as smooth sections of tensor bundles over manifolds.

What is missing is a PseudoRiemannianMetric structure on a smooth manifold that carries signature information (distinguishing Lorentzian from Riemannian), the Levi-Civita connection derived from such a metric, and the associated curvature tensors (Riemann, Ricci, Einstein). The tensor transformation law itself is straightforward to formalise given the existing TensorProduct machinery; the physics gap is the Lorentzian-signature structure and its geometric consequences. The lean_status: none field reflects this gap. Tyler's review attests intermediate-tier correctness.

Advanced results [Master]

Beyond the foundational definitions, three structural developments of the tensor formalism are central to both the mathematics of Lorentzian geometry and the physics of GR.

The Lie derivative

The Lie derivative measures the rate of change of a tensor field when dragged along the flow of a vector field $X$. For a function $f$, ${\mathcal{L}}_{X}f=X(f)=X^{\mu }{\partial }_{\mu }f$. For a vector field $Y$, the Lie derivative is the Lie bracket: ${\mathcal{L}}_{X}Y=[X,Y]$, with components $({\mathcal{L}}_{X}Y{)}^{\mu }=X^{\nu }{\partial }_{\nu }{Y}^{\mu }-Y^{\nu }{\partial }_{\nu }{X}^{\mu }$. For a general $(r,s)$-tensor $T$, the Lie derivative is

$$({\mathcal{L}}_{X}T{)}^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}=X^{\rho }{\partial }_{\rho }{T}^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}-\sum _{i=1}^r{T}^{{\mu }_1\cdots \rho \cdots {\mu }_r}{}_{{\nu }_1\cdots {\nu }_s}{\partial }_{\rho }{X}^{{\mu }_i}+\sum _{j=1}^s{T}^{{\mu }_1\cdots {\mu }_r}{}_{{\nu }_1\cdots \rho \cdots {\nu }_s}{\partial }_{{\nu }_j}{X}^{\rho }.$$

Each upper index contributes a correction term with $\partial X$, and each lower index contributes one with the opposite sign. The Lie derivative of a tensor field is again a tensor field of the same type -- it does not require a connection.

A vector field $X$ satisfying ${\mathcal{L}}_{X}g=0$ is a Killing vector field. The equation ${\mathcal{L}}_{X}g=0$ expands to the Killing equation ${\nabla }_{\mu }{X}_{\nu }+{\nabla }_{\nu }{X}_{\mu }=0$ when a Levi-Civita connection is available. Killing vectors generate continuous isometries of the metric. The Schwarzschild metric admits four independent Killing vectors (one timelike, ${\partial }_t$, and three associated with spherical symmetry), reflecting its static and spherically symmetric character. The number of linearly independent Killing vectors is bounded above by $n(n+1)/2$ on an $n$-dimensional manifold; this maximum is attained on maximally symmetric spaces (Minkowski, de Sitter, anti-de Sitter).

Proposition (Cartan's trick for the Lie derivative). On a manifold with a metric $g$ and its Levi-Civita connection $\nabla$, the Lie derivative of the metric along a vector field $X$ satisfies ${\mathcal{L}}_X{g}_{\mu \nu }={\nabla }_{\mu }{X}_{\nu }+{\nabla }_{\nu }{X}_{\mu }$. Consequently, $X$ is a Killing vector field if and only if ${\nabla }_{\mu }{X}_{\nu }+{\nabla }_{\nu }{X}_{\mu }=0$.

The proof uses the fact that the Levi-Civita connection is torsion-free and metric-compatible. The torsion-free condition ${\nabla }_{\mu }{X}_{\nu }-{\nabla }_{\nu }{X}_{\mu }={\partial }_{\mu }{X}_{\nu }-{\partial }_{\nu }{X}_{\mu }$ lets one replace partial derivatives in the coordinate formula for ${\mathcal{L}}_{X}g$ with covariant derivatives (the Christoffel-symbol terms cancel pairwise by symmetry of $g$). Metric compatibility ${\nabla }_{\rho }{g}_{\mu \nu }=0$ then gives the clean result. The Killing equation ${\nabla }_{\mu }{X}_{\nu }+{\nabla }_{\nu }{X}_{\mu }=0$ follows immediately. In $n$ dimensions, the Killing equation imposes $n(n+1)/2$ independent PDE constraints on the $n$ components of $X$, which constrains the dimension of the solution space to at most $n(n+1)/2$.

Exterior calculus and differential forms

A differential $k$-form at a point $p$ is a totally antisymmetric $(0,k)$-tensor: $\omega \in {\Lambda }^k{T}_p^{\ast }M$. The space ${\Lambda }^k{T}_p^{\ast }M$ has dimension $\left(\genfrac{}{}{0ex}{}{n}{k}\right)$. The wedge product $\wedge :{\Lambda }^k\times {\Lambda }^{\ell }\to {\Lambda }^{k+\ell }$ is the antisymmetrised tensor product:

$$(\alpha \wedge \beta )(V_1,\dots ,V_{k+\ell })=\frac{1}{k!\ell !}\sum _{\sigma \in S_{k+\ell }}\text{sgn}(\sigma )\alpha (V_{\sigma (1)},\dots ,V_{\sigma (k)})\beta (V_{\sigma (k+1)},\dots ,V_{\sigma (k+\ell )}).$$

The exterior derivative $d:{\Lambda }^k\to {\Lambda }^{k+1}$ is defined by $d\omega =(k+1){\partial }_{[{\mu }_0}{\omega }_{{\mu }_1\cdots {\mu }_k]}d{x}^{{\mu }_0}\wedge \cdots \wedge d{x}^{{\mu }_k}$, where square brackets denote antisymmetrisation. The exterior derivative satisfies $d^2=0$ and is coordinate-independent.

Differential forms carry a cohomology theory: the de Rham cohomology groups $H_{\text{dR}}^k(M)=\ker (d_k)/\text{im}(d_{k-1})$. On a compact oriented $n$-manifold without boundary, $H_{\text{dR}}^n(M)\cong \mathbb{R}$ and Poincare duality gives $H_{\text{dR}}^k(M)\cong H_{\text{dR}}^{n-k}(M)$. The Euler characteristic satisfies $\chi (M)={\sum }_k(-1{)}^k\dim H_{\text{dR}}^k(M)$.

Stokes' theorem unifies the classical integral theorems. If $M$ is an oriented $n$-manifold with boundary $\partial M$ and $\omega$ is a compactly supported $(n-1)$-form, then

$${\int }_{M}d\omega ={\int }_{\partial M}\omega .$$

This single statement contains the fundamental theorem of calculus ($n=1$), Green's theorem ($n=2$), Gauss's divergence theorem ($n=3$), and the general Stokes theorem in any dimension. In GR, Stokes' theorem underlies the derivation of conservation laws from the divergence-free property of the stress-energy tensor.

The Hodge star on Lorentzian manifolds

On an $n$-dimensional oriented (pseudo-)Riemannian manifold with metric $g$ and volume form $ϵ=\sqrt{|g|}d{x}^0\wedge \cdots \wedge d{x}^{n-1}$, the Hodge star operator $\star :{\Lambda }^k\to {\Lambda }^{n-k}$ is defined by

$$\alpha \wedge \star \beta =g(\alpha ,\beta )ϵ$$

for all $k$-forms $\alpha ,\beta$, where $g(\alpha ,\beta )$ is the induced inner product on $k$-forms. On a Lorentzian manifold with signature $(-,+,+,+)$ in four dimensions, the Hodge star satisfies $\star (\star \alpha )=(-1{)}^{k+1}\alpha$ for a $k$-form -- note the extra $(-1)$ compared to the Riemannian case, arising from the indefinite signature.

The codifferential $\delta :{\Lambda }^k\to {\Lambda }^{k-1}$ is defined as $\delta =(-1{)}^{nk+n+1}\star d\star$ (with sign conventions varying across the literature). It is the adjoint of $d$ with respect to the $L^2$ inner product on forms. The Laplace-de Rham operator (also called the Hodge Laplacian) is $\Delta =d\delta +\delta d$. On functions ($0$-forms) in a Lorentzian setting, this reduces to the d'Alembertian $\square =g^{\mu \nu }{\nabla }_{\mu }{\nabla }_{\nu }$, the wave operator that governs the propagation of fields in curved spacetime.

The Hodge star is central to the formulation of Maxwell's equations in curved spacetime: the source-free Maxwell equations become $dF=0$ and $d\star F=0$, where $F$ is the electromagnetic $2$-form. This is one of the cleanest illustrations of how tensor calculus (specifically exterior calculus) captures physics in a coordinate-independent way. The source-full equations add a current $3$-form $J$: $d\star F=4\pi \star J$, and charge conservation follows from $d^2=0$ applied to both sides: $d\star J=0$.

On an even-dimensional manifold, the Hodge star on middle-degree forms connects to topological invariants. The signature of a $4k$-dimensional manifold can be expressed in terms of the index of the operator $d+\delta$ restricted to even-degree forms (the Hirzebruch signature theorem). This bridge between differential-geometric operators and topological invariants is a recurring pattern in modern geometry and mathematical physics. In four dimensions the Hodge star on $2$-forms satisfies ${\star }^2=-1$ on self-dual forms and ${\star }^2=+1$ on anti-self-dual forms; this splitting underlies the ASD (anti-self-dual) Yang-Mills equations and the Donaldson invariants of smooth $4$-manifolds.

Tensor symmetries and representation theory

The algebraic structure of tensor spaces admits a rich symmetry theory. A tensor of type $(0,k)$ can be decomposed into its symmetric and antisymmetric parts, and more generally into irreducible representations of the symmetric group $S_k$ acting by index permutation. The Young symmetriser formalism makes this precise: each Young diagram with $k$ boxes corresponds to a specific symmetry type, and the space of $(0,k)$-tensors decomposes as a direct sum of subspaces indexed by these diagrams.

For GR the most important symmetries are: total antisymmetry (differential forms, $k$-forms with $\left(\genfrac{}{}{0ex}{}{n}{k}\right)$ independent components); total symmetry (the metric $g_{\mu \nu }$, the Ricci tensor $R_{\mu \nu }$); and the specific algebraic symmetries of the Riemann tensor ($R_{\rho \sigma \mu \nu }=-R_{\sigma \rho \mu \nu }=-R_{\rho \sigma \nu \mu }$, $R_{\rho \sigma \mu \nu }=R_{\mu \nu \rho \sigma }$, plus the first Bianchi identity $R_{[\rho \sigma \mu ]\nu }=0$). These reduce the $n^4=256$ nominal components of $R_{\rho \sigma \mu \nu }$ in four dimensions to $20$ independent ones. The Weyl tensor, the trace-free part of the Riemann tensor, carries $10$ independent components and encodes the conformally invariant portion of the curvature.

Pullbacks and pushforwards of tensor fields

Given a smooth map $\varphi :M\to N$ between manifolds, the pushforward ${\varphi }_{\ast }:T_{p}M\to T_{\varphi (p)}N$ maps tangent vectors forward. For $v\in T_{p}M$ and $f\in C^{\infty }(N)$, $({\varphi }_{\ast }v)(f)=v(f\circ \varphi )$. In coordinates, $({\varphi }_{\ast }v{)}^{\alpha }=(\partial y^{\alpha }/\partial x^{\mu })v^{\mu }$ where $y^{\alpha }$ are coordinates on $N$. The pushforward is the natural way to transport directional information from one manifold to another along a smooth map.

The pullback ${\varphi }^{\ast }:T_{\varphi (p)}^{\ast }N\to T_p^{\ast }M$ goes in the opposite direction for covectors: $({\varphi }^{\ast }\omega )(v)=\omega ({\varphi }_{\ast }v)$. More generally, the pullback extends to $(0,s)$-tensor fields: for a $(0,s)$-tensor $S$ on $N$, the pullback ${\varphi }^{\ast }S$ is the $(0,s)$-tensor on $M$ defined by

$$({\varphi }^{\ast }S{)}_p(v_1,\dots ,v_s)=S_{\varphi (p)}({\varphi }_{\ast }{v}_1,\dots ,{\varphi }_{\ast }{v}_s).$$

In components: $({\varphi }^{\ast }S{)}_{{\mu }_1\cdots {\mu }_s}=(\partial y^{{\alpha }_1}/\partial x^{{\mu }_1})\cdots (\partial y^{{\alpha }_s}/\partial x^{{\mu }_s})S_{{\alpha }_1\cdots {\alpha }_s}$. The pullback of a symmetric $(0,2)$-tensor is symmetric, and the pullback of a positive-definite metric is positive-definite (when $\varphi$ is an immersion). However, the pullback of a Lorentzian metric need not be Lorentzian -- the inclusion of the $x$-axis into Minkowski space pulls $\eta$ back to the Riemannian metric $d{x}^2$.

A diffeomorphism $\varphi :M\to M$ satisfying ${\varphi }^{\ast }g=g$ is an isometry. The collection of all isometries forms a Lie group, the isometry group of $(M,g)$. For Minkowski spacetime this group is the Poincare group $\text{ISO}(1,3)={\mathbb{R}}^4⋊O(1,3)$; for the Schwarzschild solution it is $\mathbb{R}\times SO(3)$ (time translation and spherical symmetry). Isometries are the symmetry transformations of GR, and Killing vector fields are their infinitesimal generators.

The metric and causal structure

On a Lorentzian manifold $(M,g)$ with signature $(-,+,+,+)$, the metric partitions the non-zero tangent vectors at each point into three causal classes. A vector $v\in T_{p}M$ is timelike if $g(v,v)<0$, **spacelike** if $g(v,v)>0$, and null (or lightlike) if $g(v,v)=0$ with $v\ne 0$.

The null cone at $p$ is the set of null vectors in $T_{p}M$, forming a double cone that separates timelike vectors (inside) from spacelike vectors (outside). This cone is the relativistic light cone -- the set of all directions light can travel through $p$.

Proposition (Conformal invariance of the null cone). Let $g$ and $\hat{g}={\Omega }^{2}g$ be two conformally related Lorentzian metrics on $M$ with $\Omega >0$ smooth. A non-zero vector $v$ is null with respect to $g$ if and only if it is null with respect to $\hat{g}$.

The proof is immediate: $\hat{g}(v,v)={\Omega }^{2}g(v,v)$, and ${\Omega }^2>0$ implies $\hat{g}(v,v)=0$ if and only if $g(v,v)=0$. The conformal invariance of the null cone underlies Penrose's conformal compactification: by rescaling the metric, the asymptotic structure of spacetime (future null infinity ${\mathcal{I}}^{+}$, past null infinity ${\mathcal{I}}^{-}$) becomes accessible as a finite boundary. The causal structure -- which events can influence which -- depends only on the conformal class of $g$, not on the metric itself.

A causal curve is a piecewise-smooth curve whose tangent vector is everywhere timelike or null. If $p$ and $q$ are connected by a future-directed causal curve, we write $p⪯q$ and say "$p$ is in the causal past of $q$." The relation $⪯$ is a partial order only when the spacetime is causal (contains no closed causal curves). Spacetimes violating this condition, such as the Godel metric, permit causal-pathology paradoxes.

Theorem (Global hyperbolicity). On a globally hyperbolic spacetime, the causality relation $⪯$ is a closed partial order, and the sets $J^{+}(p)=\{q:p⪯q\}$ (causal future) and $J^{-}(p)=\{q:q⪯p\}$ (causal past) are closed. Moreover, $M$ is diffeomorphic to $\mathbb{R}\times \Sigma$ where $\Sigma$ is a Cauchy surface and the metric takes the form $g=-{\beta }^{2}d{t}^2+h_t$ with $\beta >0$ and $h_t$ a Riemannian metric on $\Sigma$ depending smoothly on $t$.

Global hyperbolicity is the standard causality condition imposed on physical spacetimes. It guarantees well-posedness of the initial-value problem for hyperbolic PDEs (the wave equation $\square \varphi =0$, Maxwell's equations, and Einstein's equations themselves). The causal structure determined by $g$ is a topological invariant of the Lorentzian manifold that has no analogue in Riemannian geometry. In GR the causal structure replaces the fixed Newtonian notion of absolute time: the metric tensor, through its null-cone geometry, decides what "before" and "after" mean at each point.

Synthesis. The tensor formalism is the foundational reason that general relativity can express physical laws in a coordinate-independent manner. The Lie derivative generalises directional differentiation to arbitrary tensor fields without requiring a connection, the pullback-pushforward calculus identifies tensorial quantities on one manifold with those on another via smooth maps, and the causal structure determined by the Lorentzian metric governs which events can influence which. Putting these together, the metric tensor $g_{\mu \nu }$ encodes distances, angles, curvature, causal structure, and gravitational dynamics in a single $(0,2)$-tensor field. The bridge is between the abstract multilinear algebra of this unit and the concrete gravitational physics of 13.04.01 pending, where the Einstein tensor -- built from the metric and its derivatives -- equates to the stress-energy content of matter. This pattern recurs throughout differential geometry: a single tensor field, equipped with the right algebraic symmetries, generates the entire geometric and physical content of the theory.

Full proof set [Master]

Proposition (Pullback preserves tensor type and symmetries). Let $\varphi :M\to N$ be a smooth map and let $S$ be a $(0,s)$-tensor field on $N$. Then $\phi^ S$isa$(0, s)$-tensorfieldon$M$.If$S$issymmetric,then$\phi^* S$issymmetric.If$\phi$isadiffeomorphismand$S$isnon-degenerate,then$\phi^* S$ is non-degenerate.*

Proof. For the $C^{\infty }(M)$-multilinearity claim, let $f\in C^{\infty }(M)$ and compute at $p\in M$:

$$({\varphi }^{\ast }S{)}_p(f\cdot v_1{|}_p,v_2{|}_p,\dots ,v_s{|}_p)=S_{\varphi (p)}({\varphi }_{\ast }(f\cdot v_1{|}_p),{\varphi }_{\ast }{v}_2{|}_p,\dots ,{\varphi }_{\ast }{v}_s{|}_p).$$

Since ${\varphi }_{\ast }(f\cdot v_1{|}_p)=f(p)\cdot {\varphi }_{\ast }(v_1{|}_p)$ (the pushforward is $\mathbb{R}$-linear at each point), the right-hand side equals

$$f(p)\cdot S_{\varphi (p)}({\varphi }_{\ast }{v}_1{|}_p,{\varphi }_{\ast }{v}_2{|}_p,\dots )=f(p)\cdot ({\varphi }^{\ast }S{)}_p(v_1{|}_p,v_2{|}_p,\dots ).$$

Linearity in each argument follows identically, confirming that ${\varphi }^{\ast }S$ is $C^{\infty }(M)$-multilinear and hence a $(0,s)$-tensor field.

For symmetry, let $S$ be symmetric and compute:

$$({\varphi }^{\ast }S{)}_p(v_i,v_j,\dots )=S_{\varphi (p)}({\varphi }_{\ast }{v}_i,{\varphi }_{\ast }{v}_j,\dots )=S_{\varphi (p)}({\varphi }_{\ast }{v}_j,{\varphi }_{\ast }{v}_i,\dots )=({\varphi }^{\ast }S{)}_p(v_j,v_i,\dots ).$$

For non-degeneracy under a diffeomorphism: suppose $({\varphi }^{\ast }S{)}_p(v,w)=0$ for all $w\in T_{p}M$. Then $S_{\varphi (p)}({\varphi }_{\ast }v,{\varphi }_{\ast }w)=0$ for all $w$. Since $\varphi$ is a diffeomorphism, ${\varphi }_{\ast }:T_{p}M\to T_{\varphi (p)}N$ is a linear isomorphism, so ${\varphi }_{\ast }w$ ranges over all of $T_{\varphi (p)}N$ as $w$ ranges over $T_{p}M$. Hence $S_{\varphi (p)}({\varphi }_{\ast }v,u)=0$ for all $u\in T_{\varphi (p)}N$. Non-degeneracy of $S$ forces ${\varphi }_{\ast }v=0$, and injectivity of ${\varphi }_{\ast }$ gives $v=0$. $\square$

Proposition (Isometries form a Lie group). The set $\text{Isom}(M,g)$ of all isometries of a Lorentzian manifold $(M,g)$ forms a finite-dimensional Lie group under composition. Its Lie algebra is the space of Killing vector fields.

Proof (sketch). An isometry $\varphi$ satisfies ${\varphi }^{\ast }g=g$. By the previous proposition, ${\varphi }^{\ast }g$ is a symmetric non-degenerate $(0,2)$-tensor field; the equation ${\varphi }^{\ast }g=g$ imposes $n(n+1)/2$ constraints on $\varphi$ at each point. Myers and Steenrod (1939) proved that the isometry group of a Riemannian manifold is a Lie group; Kobayashi (1961) extended this to the pseudo-Riemannian case. The key step is showing that an isometry is determined by its value and first derivative at a single point (the Myers-Steenrod rigidity lemma), which bounds the dimension of $\text{Isom}(M,g)$ above by $n+n(n+1)/2=n(n+3)/2$. The Lie-algebra elements are the vector fields $X$ satisfying ${\mathcal{L}}_{X}g=0$, which is exactly the Killing equation. $\square$

Connections [Master]

• Differential geometry foundations 03.01.01 provides the smooth-manifold setting in which tensor fields are defined. The tangent and cotangent bundles, transition functions, and smooth structures developed there are the substrate on which this unit builds the tensor-algebraic apparatus.

• Special relativity 10.05.01 pending uses flat-space tensors throughout -- the Minkowski metric ${\eta }_{\mu \nu }$, the four-momentum $p^{\mu }$, the electromagnetic field tensor $F^{\mu \nu }$. This unit generalises those objects from the fixed flat geometry of SR to arbitrary curved manifolds. The tensor transformation law is the same; what changes is that the metric is no longer constant.

• Einstein field equations [13.04.01, pending] are the downstream consumer of the tensor formalism. The Einstein tensor $G_{\mu \nu }$ and the stress-energy tensor $T_{\mu \nu }$ are both $(0,2)$-tensor fields, and the field equation $G_{\mu \nu }=(8\pi G/c^4)T_{\mu \nu }$ is a tensor equation whose coordinate independence is guaranteed by the theorem proved in this unit.

• Schwarzschild solution 13.05.01 pending is the concrete worked example: the Schwarzschild metric tensor $g_{\mu \nu }$ is a $(0,2)$-tensor field with Lorentzian signature on a specific manifold. The inverse metric, determinant, and signature analysis in this unit's worked example are the first computational steps in understanding that solution.

• Tensor symmetries and the Riemann tensor [13.03.NN, pending] extend the tensor formalism to curvature. The Riemann tensor $R^{\rho }{}_{\sigma \mu \nu }$ is a $(1,3)$-tensor field constructed from the metric and its first and second derivatives. Its algebraic symmetries (antisymmetry in the last two indices, the first Bianchi identity, interchange symmetry) are structural properties of a specific tensor type, not general features of all tensors.

• Differential forms 03.04.02 and exterior derivative 03.04.04 are the antisymmetric corner of the tensor world. Differential forms are totally antisymmetric $(0,k)$-tensors, and the exterior derivative is a natural differential operator on them. The Hodge star, discussed in Advanced results, connects the antisymmetric and metric structures.

Historical & philosophical context [Master]

Gregorio Ricci-Curbastro and his student Tullio Levi-Civita published "Methodes de calcul differentiel absolu et leurs applications" in Mathematische Annalen 54 (1900), pages 125-201, establishing what they called absolute differential calculus -- the systematic calculus of tensors on manifolds, including the covariant derivative, the metric, and the transformation laws. Ricci-Curbastro had been developing these ideas since the 1880s, building on Christoffel's 1869 work on quadratic differential forms and Lipschitz's extensions of Riemann's geometry. The 1900 paper with Levi-Civita was the comprehensive synthesis.

Einstein encountered tensor calculus through Marcel Grossmann, his ETH classmate and colleague, around 1912. Grossmann recognised that Ricci-Curbastro and Levi-Civita's formalism was precisely the mathematical language Einstein needed for a theory of gravity in which the metric of spacetime becomes a dynamical field. The collaboration between Einstein and Grossmann produced the "Entwurf" theory of 1913 -- a precursor to GR that had the right geometric picture but the wrong field equations. By November 1915 Einstein had the correct field equations, and the Ricci-Curbastro-Levi-Civita calculus became the standard language of gravitational physics.

The philosophical point is structural. Before the tensor calculus, the only language for writing physical laws in curvilinear coordinates was the older component-by-component approach, in which every equation came with a forest of correction terms (the Christoffel symbols) that made coordinate independence hard to verify. Tensor notation absorbs these correction terms into the transformation law itself: if an object is a tensor, coordinate independence is automatic. This is not a notational convenience -- it is a conceptual clarification. The tensor formalism makes precise the distinction between quantities that depend on the choice of coordinates (which are physically meaningless) and quantities that do not (which carry physical content).

Weyl's 1918 Raum, Zeit, Materie was the first textbook to present tensor calculus as the natural language of GR, followed by Eddington's 1923 The Mathematical Theory of Relativity. The modern index-free notation (using abstract index conventions and bundle-theoretic language) was consolidated by the differential-geometry community in the mid-twentieth century, particularly through the influence of Kobayashi and Nomizu's Foundations of Differential Geometry (1963, 1969). Penrose's abstract index notation, introduced in 1968, combines the computational convenience of indices with the coordinate independence of the invariant approach.

Bibliography [Master]

Primary literature:

• Ricci-Curbastro, G. & Levi-Civita, T., "Methodes de calcul differentiel absolu et leurs applications", Math. Ann. 54 (1900), 125-201. [Originating paper for absolute differential calculus.]
• Einstein, A., "Die Feldgleichungen der Gravitation", Sitzungsberichte Preuss. Akad. Wiss. (1915), 844-847. [The field equations in their final form.]
• Einstein, A. & Grossmann, M., "Entwurf einer verallgemeinerten Relativitatstheorie und einer Theorie der Gravitation", Zeit. Math. Phys. 62 (1913), 225-261. [The precursor collaboration introducing tensor methods to gravitational physics.]
• Christoffel, E. B., "Ueber die Transformation der homogenen Differentialausdrucke zweiten Grades", J. Reine Angew. Math. 70 (1869), 46-70. [Predecessor work on quadratic forms and the Christoffel symbols.]

Textbooks and monographs:

• Weyl, H., Raum, Zeit, Materie (Springer, 1918; 7th ed. 1983). [First textbook presentation of tensor calculus in GR.]
• Eddington, A. S., The Mathematical Theory of Relativity (Cambridge, 1923; 2nd ed. 1924). [Early systematic treatment.]
• Kobayashi, S. & Nomizu, K., Foundations of Differential Geometry, Vols. I-II (Wiley, 1963, 1969). [Consolidated the modern bundle-theoretic approach.]
• Misner, C. W., Thorne, K. S. & Wheeler, J. A., Gravitation (Freeman, 1973). [Comprehensive GR reference with extensive tensor calculus.]
• Wald, R. M., General Relativity (University of Chicago Press, 1984). [Modern apex anchor for master-tier GR.]
• Carroll, S. M., Spacetime and Geometry: An Introduction to General Relativity (Addison-Wesley, 2004). [Accessible intermediate-tier treatment of tensors and manifolds.]
• Schutz, B. F., A First Course in General Relativity, 2nd ed. (Cambridge, 2009). [Standard introductory treatment of tensor calculus in GR.]
• Hartle, J. B., Gravity: An Introduction to Einstein's General Relativity (Addison-Wesley, 2003). [Physics-first approach with tensors introduced through physical motivation.]

Mathematical foundations:

• Lee, J. M., Introduction to Smooth Manifolds, 2nd ed. (Springer, 2013). [Standard reference for smooth-manifold theory and tensor fields.]
• Lee, J. M., Riemannian Manifolds: An Introduction to Curvature (Springer, 1997). [Curvature and metric geometry at the intermediate level.]
• Nakahara, M., Geometry, Topology and Physics, 2nd ed. (Taylor & Francis, 2003). [Bridge between mathematics and physics, covering tensor calculus, differential forms, and fibre bundles.]
• Penrose, R. & Rindler, W., Spinors and Space-Time, Vol. 1 (Cambridge, 1984). [Introduces the abstract index notation used throughout modern GR.]