Tensors and Index Notation in Flat Spacetime

Write a physical law twice — once in a train station, once on the moving train — and if it is a genuine law the two versions must say the same thing. Special relativity makes that demand precise: the laws of physics take the identical form for every inertial observer, and the coordinates two such observers assign to the same event are related by a Lorentz transformation. The problem is bookkeeping. A quantity like velocity or field strength is a list of numbers, and every one of those numbers is different for the two observers; checking by hand that an equation survives the change of frame is hopeless once the objects have more than a few components. Index notation is the solution. Instead of writing components and tracking how each one transforms, we attach little labels — indices — that encode the transformation rule in their very placement, upper or lower. A tensor is then defined not by what it is made of but by how its components move under $\Lambda$ , and a correctly balanced tensor equation is guaranteed to hold in every inertial frame the instant it holds in one. The engine underneath is the Minkowski metric $\eta_{\mu\nu}=\mathrm{diag}(-1,+1,+1,+1)$ , the object Lorentz transformations are defined to leave invariant. It supplies the inner product that distinguishes spacetime from bare $\mathbb{R}^4$ , lets you lower an upper index and raise a lower one, and turns the contraction of a matched up-down pair into a Lorentz scalar — the frame-independent kernel of any tensor. The spacetime interval, proper time, the rest mass, and the field invariants of electromagnetism are all such contractions. In flat spacetime, in inertial coordinates, the metric is a table of constants, so ordinary partial derivatives already behave as tensors: no connection is needed. That single gift is what lets all of special relativity be written in a few compact lines, and it is the grammar that, upgraded with covariant derivatives, becomes general relativity.

Key takeaways

A tensor of type $(p,q)$ is defined by its transformation law: $p$ upper indices each carry a factor of $\Lambda$ , $q$ lower indices each carry a factor of $\Lambda^{-1}$ — the index structure is the definition, not the number of components.
Upper (contravariant) and lower (covariant) indices transform inversely; the Einstein summation convention silently sums any index that appears once up and once down.
The Minkowski metric $\eta_{\mu\nu}=\mathrm{diag}(-1,+1,+1,+1)$ and its inverse $\eta^{\mu\nu}$ lower and raise indices, trading a vector for its covector partner.
Contracting a matched upper-lower pair yields a Lorentz scalar; a valid tensor equation has identical free-index structure on both sides, each summed index once up once down.
The Kronecker delta $\delta^\mu{}_\nu$ , the symmetric and antisymmetric parts, and the Levi-Civita symbol $\epsilon_{\mu\nu\rho\sigma}$ are the building blocks; in flat inertial coordinates $\partial_\mu$ transforms tensorially with no connection.
Key idea: attach indices that encode transformation behaviour — an equation between tensors of the same type, with matching free indices, is automatically true in every inertial frame. Raise and lower with $\eta$ , sum one up against one down, and physics becomes manifestly Lorentz covariant.

The core insight

Attach indices that encode transformation behaviour: an equation between tensors of the same type, with matching free indices, is automatically true in all inertial frames. Raise and lower with $\eta$ , sum over one up and one down, and physics becomes manifestly Lorentz covariant.

Where this lives today

Index notation is not a historical relic; it is the everyday working language of every field that touches relativity. In lattice quantum chromodynamics and in the amplitude calculations behind Large Hadron Collider predictions, physical results are extracted as full contractions of tensors over spacetime and internal indices, and the same balance-of-free-indices grammar catches errors in expressions with dozens of terms before a single number is computed. In numerical relativity, the codes that model colliding black holes and produce the gravitational-waveform templates matched against LIGO and Virgo data are, at their core, evolution equations for the metric tensor $g_{\mu\nu}$ written index by index, with the flat $\eta_{\mu\nu}$ of this article as the asymptotic boundary condition far from the source. Modern differential-geometry software and tensor-algebra packages automate raising, lowering, and contraction directly from the transformation law of equation (1). And the conceptual bridge is exact: promote the constant metric $\eta_{\mu\nu}$ to a position-dependent $g_{\mu\nu}(x)$ and the partial derivative $\partial_\mu$ to a covariant derivative $\nabla_\mu$ carrying Christoffel symbols, and every piece of index grammar developed here — the summation convention, the symmetric-antisymmetric split, the invariance of contractions — carries over unchanged into the Einstein field equations.

The physical picture

Think of an upper index as naming a thing that lives in spacetime — a displacement, a velocity, an arrow you could in principle draw — and a lower index as naming a measuring instrument — a gradient, a stack of level surfaces, a device that eats an arrow and reports a number. A contravariant vector $V^\mu$ is a thing; a covariant covector $W_\mu$ is an instrument. When the instrument acts on the thing — one index up, one down, summed — the output is a single number, and that number is the same for every observer. That operation is the contraction, and it is the only legal way to collapse two indices in a tensor equation. The transformation rules are engineered for exactly this. A boost drags $V^\mu$ around by the matrix $\Lambda$ and drags $W_\mu$ around by the inverse matrix $\Lambda^{-1}$ ; when you pair them the two motions cancel, $\Lambda\,\Lambda^{-1}=\delta$ , and the number survives untouched. Summing two upper indices, or two lower ones, would leave dangling factors of $\Lambda$ that do not cancel — the result would depend on who is looking, so it cannot be a physical statement. The metric $\eta$ is the universal adaptor that converts any thing into the matching instrument, so anything can be paired with anything.

Figure

The grammar of indices: a contravariant vector V^μ (an arrow) paired with a covariant one-form W_μ (a family of level surfaces) contracts to a single frame-invariant number V^μ W_μ. Under a boost the Λ that transforms V^μ is exactly undone by the Λ⁻¹ that transforms W_μ, so V'^μ W'_μ = V^μ W_μ

Index gymnastics is grammar, not physics. A correctly typed equation — the same free indices on both sides, each summed index appearing once up and once down — is guaranteed to be a sensible tensor statement, whatever it turns out to mean physically. A mismatched equation is like a type error in a compiler: it cannot possibly be right, and you can see this without knowing any dynamics, just by counting indices. The payoff is that the sign of a vector's own squared norm, the trace of a mixed tensor, and the value of any full contraction are permanent labels no observer can alter. The second figure makes this concrete with real numbers: a single timelike four-vector, its invariant norm, and the boost that carries it to its own rest frame.

Figure

Numeric invariance under a boost: the timelike four-vector V^μ = (5, 3, 0, 0) has squared norm V·V = η_μν V^μ V^ν = −16 in every inertial frame. A boost of β = 0.6 rotates it to V'^μ = (4, 0, 0, 0) in its rest frame, where −4² + 0² = −16 once more. Both points sit on the invariant hyperbola (ct)² − x² = 16

The core equations

The definition of a tensor is a transformation law, and everything else is a consequence of it. A tensor of type $(p,q)$ is any object whose components in a new inertial frame are obtained from the old ones by one factor of $\Lambda$ for each upper index and one factor of $\Lambda^{-1}$ for each lower index. The index structure you write down is a promise about how the object behaves under a change of observer.

T'^{\mu_1\cdots\mu_p}{}_{\nu_1\cdots\nu_q}=\Lambda^{\mu_1}{}_{\alpha_1}\cdots\Lambda^{\mu_p}{}_{\alpha_p}\,(\Lambda^{-1})^{\beta_1}{}_{\nu_1}\cdots(\Lambda^{-1})^{\beta_q}{}_{\nu_q}\,T^{\alpha_1\cdots\alpha_p}{}_{\beta_1\cdots\beta_q} \qquad (1)

where: $\mu,\nu,\alpha,\beta$ — spacetime indices running over $0,1,2,3$ ; $p$ — number of upper (contravariant) indices; $q$ — number of lower (covariant) indices; $\Lambda^\mu{}_\nu$ — the Lorentz transformation matrix; $(\Lambda^{-1})^\nu{}_\mu$ — its inverse. Physical reading: a scalar is $(0,0)$ , a four-vector $(1,0)$ , a covector $(0,1)$ ; the metric and the electromagnetic field strength are $(0,2)$ . The type is the object's genome. The Einstein summation convention is the piece of shorthand that makes index notation usable: whenever an index appears exactly once as an upper and once as a lower on the same side of an equation, it is silently summed over its four values. An index that is not repeated is a free index; it must appear, in the same position, on both sides of every valid equation.

A^\mu B_\mu \equiv \sum_{\mu=0}^{3} A^\mu B_\mu,\qquad \text{free index: once per side;}\ \ \text{dummy index: once up, once down} \qquad (2)

where: $A^\mu B_\mu$ — a contraction, hence a scalar; the repeated $\mu$ is a dummy (summed) index and can be renamed at will; a free index labels a genuine component. Physical reading: the balance of free indices is a conservation law for grammar — an equation with a free $\mu$ on the left and a free $\nu$ on the right is simply ill-formed. The Minkowski metric is the object that makes spacetime geometric rather than a bare list of coordinates, and it is defined to be invariant under Lorentz transformations. Its inverse $\eta^{\mu\nu}$ is numerically identical in inertial frames, and the two are related by the contraction that defines any inverse.

\eta_{\mu\nu}=\mathrm{diag}(-1,+1,+1,+1),\qquad \eta^{\mu\alpha}\eta_{\alpha\nu}=\delta^\mu{}_\nu \qquad (3)

where: $\eta_{\mu\nu}$ — components of the flat metric in inertial coordinates; the single $-1$ in the time slot is the whole of relativistic geometry; $\eta^{\mu\nu}$ — inverse metric; $\delta^\mu{}_\nu$ — the Kronecker delta, equal to $1$ when $\mu=\nu$ and $0$ otherwise, the unique $(1,1)$ tensor with the same components in every frame. The metric's job in daily calculation is to raise and lower indices — to convert a contravariant vector into its covariant partner and back. This is how a "thing" is turned into the matching "instrument" so that it can be contracted with another thing.

V_\mu=\eta_{\mu\nu}V^\nu,\qquad V^\mu=\eta^{\mu\nu}V_\nu \quad\Rightarrow\quad V_0=-V^0,\ \ V_i=+V^i \qquad (4)

where: $V^\mu$ — contravariant components; $V_\mu$ — covariant components. Physical reading: in the $(-,+,+,+)$ signature, lowering an index flips the sign of the time component only and leaves the three spatial components numerically alone — the single most common source of dropped minus signs. Contracting a vector with its own lowered partner gives the Lorentz-invariant inner product, the direct generalization of the Euclidean dot product with one crucial sign flip. The sign of a vector's squared norm is a frame-independent label.

V\cdot W=\eta_{\mu\nu}V^\mu W^\nu=V_\mu W^\mu=-V^0W^0+\vec V\cdot\vec W \qquad (5)

where: $V\cdot W$ — scalar product of two four-vectors, invariant under every Lorentz transformation; $\vec V\cdot\vec W$ — the ordinary three-dimensional dot product of the spatial parts. Physical reading: $V\cdot V<0$ is timelike, $=0$ null, $>0$ spacelike, and no boost can move a vector between these classes. Any two-index tensor splits uniquely into a symmetric and an antisymmetric part, a decomposition that respects the transformation law and so is meaningful in every frame. The symmetric metric and the antisymmetric field strength are the two archetypes.

T_{(\mu\nu)}=\tfrac12(T_{\mu\nu}+T_{\nu\mu}),\qquad T_{[\mu\nu]}=\tfrac12(T_{\mu\nu}-T_{\nu\mu}),\qquad T_{\mu\nu}=T_{(\mu\nu)}+T_{[\mu\nu]} \qquad (6)

where: round brackets $(\ )$ denote symmetrization, square brackets $[\ ]$ antisymmetrization. Physical reading: in four dimensions the symmetric part carries $10$ independent components and the antisymmetric part $6$ , adding to the full $16$ of a general rank-2 tensor. Finally, in flat spacetime with inertial coordinates the partial derivative is itself a legitimate tensor: it carries a lower index and needs no correction terms. This is the simplification that lets special relativity be written without the machinery of connections.

\partial_\mu\equiv\frac{\partial}{\partial x^\mu},\qquad \partial^\mu=\eta^{\mu\nu}\partial_\nu,\qquad \epsilon_{\mu\nu\rho\sigma}:\ \epsilon_{0123}=+1,\ \text{totally antisymmetric} \qquad (7)

where: $x^\mu=(ct,x,y,z)$ — inertial coordinates; $\partial_\mu$ — the four-gradient, a covector; $\epsilon_{\mu\nu\rho\sigma}$ — the Levi-Civita permutation symbol, which changes sign under any index swap and vanishes whenever two indices repeat. Physical reading: $\partial_\mu V^\mu$ is a four-divergence and hence a Lorentz scalar, which is why conservation laws take the form $\partial_\mu J^\mu=0$ .

Derivation AWhy the spacetime interval is the same in every frame

The single fact that anchors the entire formalism is that the contraction $\eta_{\mu\nu}\,\mathrm{d}x^\mu\mathrm{d}x^\nu$ returns the same number to every inertial observer. If that one contraction is invariant, then so is the norm of every four-vector, the classification into timelike and spacelike, and any equation built by contracting matched indices. The goal here is to prove the invariance directly from the coordinate transformation law and the single defining property of a Lorentz matrix, exhibiting the interval as the contraction of a $(0,2)$ tensor with two displacements. Nothing beyond linear substitution is required, and the result is the definition of the Lorentz group itself.

Step 1. Transform the coordinate displacement with the Lorentz matrix, $\mathrm{d}x'^\mu=\Lambda^\mu{}_\alpha\,\mathrm{d}x^\alpha$ , which is exact because inertial frames are related linearly and $\Lambda$ is a constant matrix. Why: this is just the statement that the differentials transform as a $(1,0)$ tensor, the very thing an upper index promises.

Step 2. Substitute both differentials into the primed interval, giving $\mathrm{d}s'^2=\eta_{\mu\nu}\Lambda^\mu{}_\alpha\Lambda^\nu{}_\beta\,\mathrm{d}x^\alpha\,\mathrm{d}x^\beta$ , so the whole frame-dependence is now packed into the pair of $\Lambda$ factors sandwiching the metric. Why: isolating the transformation matrices is what lets the next step collapse them in one move.

Step 3. Impose the defining condition of a Lorentz transformation, $\eta_{\mu\nu}\Lambda^\mu{}_\alpha\Lambda^\nu{}_\beta=\eta_{\alpha\beta}$ , which in matrix form is $\Lambda^{\mathsf T}\eta\,\Lambda=\eta$ and is precisely the requirement that $\Lambda$ preserve the metric. Why: this is the four-dimensional analogue of an orthogonal matrix preserving Euclidean length; a general linear map would not leave the two $\Lambda$ 's collapsing to a bare $\eta$ .

Step 4. Read off the result, $\mathrm{d}s'^2=\eta_{\alpha\beta}\,\mathrm{d}x^\alpha\,\mathrm{d}x^\beta=\mathrm{d}s^2$ : the interval is invariant. Why: the contraction of the $(0,2)$ metric with two displacement vectors is a genuine scalar, which is why proper time and proper length are quantities all observers agree on.

Go deeper · Derivation B · 1 minWhy a matched contraction lowers the rank of a tensor

Derivation BWhy a matched contraction lowers the rank of a tensor

The reason contraction is the only legal way to sum indices is that a matched upper-lower sum lands you back on a genuine tensor — one of lower rank — whereas an unmatched sum does not. The cleanest case is the trace of a mixed $(1,1)$ tensor, which the argument below shows is a Lorentz scalar. The goal is to track the transformation factors explicitly and watch the $\Lambda$ and $\Lambda^{-1}$ annihilate into a Kronecker delta, leaving a number that no observer can change. The same cancellation, applied to one pair of indices inside a larger tensor, is what reduces a $(p,q)$ tensor to a $(p-1,q-1)$ tensor.

Step 1. Write the transformation of a mixed tensor, $T'^\mu{}_\nu=\Lambda^\mu{}_\alpha\,(\Lambda^{-1})^\beta{}_\nu\,T^\alpha{}_\beta$ , one $\Lambda$ for the upper index and one $\Lambda^{-1}$ for the lower. Why: this is just equation (1) specialized to $p=q=1$ , the smallest tensor that is neither a vector nor a scalar.

Step 2. Form the trace by setting the free indices equal and summing, $T'^\mu{}_\mu=\Lambda^\mu{}_\alpha\,(\Lambda^{-1})^\beta{}_\mu\,T^\alpha{}_\beta$ , so that $\mu$ has changed from a free label into a dummy summed index. Why: contraction is exactly this act of identifying an upper with a lower slot, and it is legal only because the two carry inverse transformation factors.

Step 3. Collapse the matrices with the group property $(\Lambda^{-1})^\beta{}_\mu\Lambda^\mu{}_\alpha=\delta^\beta{}_\alpha$ , which holds by the very definition $\Lambda^{-1}\Lambda=\mathbb{1}$ . Why: the delta is the fingerprint of a successful cancellation; had both indices been upper, two $\Lambda$ 's would remain and no delta would appear.

Step 4. Conclude $T'^\mu{}_\mu=\delta^\beta{}_\alpha T^\alpha{}_\beta=T^\alpha{}_\alpha$ : the trace is invariant, so a $(1,1)$ tensor contracts to a $(0,0)$ scalar. Why: physically, the trace of the stress-energy tensor $T^\mu{}_\mu$ is a coordinate-free number that vanishes for the electromagnetic field, a statement true in every frame at once.

Go deeper · Derivation C · 1 minWhy the four-gradient carries a lower index

Derivation CWhy the four-gradient carries a lower index

It is not obvious a priori that differentiation should produce a tensor at all, and in curved coordinates it does not. In flat spacetime with inertial coordinates, however, the partial derivative $\partial_\mu$ transforms cleanly as a covector, and the reason is the chain rule: because coordinates transform with $\Lambda$ , derivatives with respect to them pick up the inverse $\Lambda^{-1}$ , which is precisely the lower-index rule. The goal is to make this explicit, so that the four-divergence $\partial_\mu V^\mu$ can be recognized as a scalar and conservation laws can be written in manifestly invariant form.

Step 1. Apply the multivariable chain rule to the derivative in the new frame, $\dfrac{\partial}{\partial x'^\mu}=\dfrac{\partial x^\nu}{\partial x'^\mu}\dfrac{\partial}{\partial x^\nu}$ , which simply expresses one frame's derivative in terms of the other's. Why: the Jacobian $\partial x^\nu/\partial x'^\mu$ is the object that decides how the derivative transforms, so everything hinges on computing it.

Step 2. Invert the coordinate transformation, $x^\nu=(\Lambda^{-1})^\nu{}_\alpha x'^\alpha$ , and differentiate to get the constant Jacobian $\dfrac{\partial x^\nu}{\partial x'^\mu}=(\Lambda^{-1})^\nu{}_\mu$ . Why: the Jacobian is constant precisely because a Lorentz transformation is linear, which is what keeps $\partial_\mu$ tensorial without correction terms.

Step 3. Assemble the two pieces to find $\partial'_\mu=(\Lambda^{-1})^\nu{}_\mu\,\partial_\nu$ , exactly the lower-index transformation law, so $\partial_\mu$ is a genuine $(0,1)$ tensor. Why: it follows immediately that $\partial_\mu V^\mu$ contracts an up against a down and is therefore a Lorentz scalar — the four-divergence — which is why $\partial_\mu J^\mu=0$ and $\partial_\mu T^{\mu\nu}=0$ are frame-independent conservation laws.

Limiting cases

Index notation reduces to familiar structures whenever a piece of the machinery switches off.

Case	Result	Meaning
Spatial sub-block	$\eta_{ij}=\delta_{ij},\ i,j\in\{1,2,3\}$	The purely spatial part of the metric is the ordinary Euclidean metric, so three-vector dot and cross products reappear and lowering a spatial index changes nothing numerically.
Scalar, rank zero	$\phi'=\phi$	A $(0,0)$ tensor has no indices and is frame-invariant outright — proper time, rest mass, and electric charge are examples.
Identity tensor	$\delta^\mu{}_\nu=\Lambda^\mu{}_\alpha(\Lambda^{-1})^\alpha{}_\nu$	The Kronecker delta is the unique mixed tensor with identical components in every frame; raising one index turns it into $\eta^{\mu\nu}$ .
Antisymmetric rank two	$T_{[\mu\nu]}$ : $6$ components	In four dimensions an antisymmetric pair has exactly six independent entries — the three electric and three magnetic components of $F_{\mu\nu}$ .
Non-relativistic limit	$v/c\to0:\ \eta_{\mu\nu}V^\mu V^\nu\to -c^2(\mathrm{d}t)^2$	For slow motion the time slot dominates and contractions reproduce Galilean invariants up to the overall factor $-c^2$ .

Common trapV^μ and V_μ are just two names for the same numbers

They are not. In the $(-,+,+,+)$ signature, lowering an index with $\eta$ flips the sign of the time component, so $V_0=-V^0$ while the spatial components are unchanged — the contravariant and covariant lists genuinely differ. Two neighbouring traps ride along with this one. First, never sum two upper or two lower indices: only a matched up-down pair transforms with $\Lambda\,\Lambda^{-1}=\delta$ and yields an invariant, whereas $A^\mu B^\mu$ leaves dangling factors of $\Lambda$ and is meaningless. Second, never confuse a free index with a dummy: a free index must appear once, in the same up-or-down position, on both sides of the equation, while a dummy is summed and can be renamed freely — writing a free $\mu$ on one side and a free $\nu$ on the other is a grammatical error before it is ever a physical one. Check index balance the way a compiler checks types: a mismatch cannot be right.

Worked examples

1 · Lowering, contracting, and boosting a four-vector (fully numeric)

Take the four-vector $V^\mu=(5,3,0,0)$ in units with $c=1$ , in the metric $\eta_{\mu\nu}=\mathrm{diag}(-1,+1,+1,+1)$ . Find its covariant components, its invariant norm, and verify by an explicit boost that the norm does not change.

Step 1. Lower the index with the metric, $V_\mu=\eta_{\mu\nu}V^\nu$ . Only the time slot flips sign, so $V_\mu=(-5,\,3,\,0,\,0)$ . Notice at once that the covariant list is not the contravariant list — the leading sign has changed, exactly the trap of the warning above.

Step 2. Contract to get the invariant norm, $V\cdot V=V_\mu V^\mu=(-5)(5)+(3)(3)+0+0=-25+9=\mathbf{-16}$ . The norm is negative, so $V^\mu$ is timelike; there exists an inertial frame in which its spatial part vanishes and it points purely along the time axis.

Step 3. Find that frame. A boost along $x$ with speed $\beta$ has $\gamma=(1-\beta^2)^{-1/2}$ and acts as $V'^0=\gamma(V^0-\beta V^1)$ , $V'^1=\gamma(V^1-\beta V^0)$ . The spatial component vanishes when $\beta=V^1/V^0=3/5=\mathbf{0.6}$ , giving $\gamma=(1-0.36)^{-1/2}=1/0.8=\mathbf{1.25}$ .

Step 4. Apply the boost: $V'^0=1.25(5-0.6\times3)=1.25\times3.2=\mathbf{4}$ and $V'^1=1.25(3-0.6\times5)=1.25\times0=\mathbf{0}$ , so $V'^\mu=(4,0,0,0)$ . Its norm is $-4^2+0^2=\mathbf{-16}$ , identical to Step 2. The four numbers changed completely — from $(5,3,0,0)$ to $(4,0,0,0)$ — but the contraction is frozen at $-16$ , and its square root, $4$ , is the invariant "rest" magnitude of the vector. Both component lists sit on the single hyperbola $(ct)^2-x^2=16$ of the second figure.

2 · Splitting a rank-two tensor and counting components (fully numeric)

Any two-index tensor decomposes into a symmetric and an antisymmetric part. Verify the decomposition on a concrete pair of components, then count how many independent numbers each part holds in four dimensions.

Step 1. Take a rank-2 tensor whose $01$ and $10$ entries are $T_{01}=6$ and $T_{10}=2$ . Symmetrize: $T_{(01)}=\tfrac12(T_{01}+T_{10})=\tfrac12(6+2)=\mathbf{4}$ . Antisymmetrize: $T_{[01]}=\tfrac12(T_{01}-T_{10})=\tfrac12(6-2)=\mathbf{2}$ .

Step 2. Check that the parts rebuild the original. Adding, $T_{(01)}+T_{[01]}=4+2=6=T_{01}$ ; and because the symmetric part is unchanged while the antisymmetric part flips sign on swapping indices, $T_{(10)}+T_{[10]}=4-2=2=T_{10}$ . The split is exact and unique — every rank-2 tensor is the sum of its symmetric and antisymmetric parts.

Step 3. Count components in four dimensions. A general rank-2 tensor has $4\times4=\mathbf{16}$ entries. The symmetric part is fixed by its diagonal and upper triangle, $4\times5/2=\mathbf{10}$ independent numbers; the antisymmetric part has zero diagonal and $4\times3/2=\mathbf{6}$ independent numbers. The tally $10+6=16$ balances exactly.

Step 4. Read the physics off the count. The symmetric $10$ -component class is where the Minkowski metric $\eta_{\mu\nu}$ and the stress-energy tensor $T^{\mu\nu}$ live; the antisymmetric $6$ -component class is exactly the electromagnetic field strength $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$ , whose six entries are the three electric and three magnetic field components. Counting indices has told us why the electromagnetic field has six components and not four or sixteen.

3 · The trace of the Kronecker delta

A one-line contraction that every relativist uses without thinking: what is $\delta^\mu{}_\mu$ ?

Step 1. The Kronecker delta $\delta^\mu{}_\nu$ equals $1$ when its indices agree and $0$ otherwise. Setting the upper equal to the lower and summing gives $\delta^\mu{}_\mu=\delta^0{}_0+\delta^1{}_1+\delta^2{}_2+\delta^3{}_3=1+1+1+1=\mathbf{4}$ .

Step 2. The number $4$ is the dimension of spacetime, and it is a Lorentz scalar by the trace argument of the second derivation. This is why factors of $4$ (or of the dimension $d$ in $d$ -dimensional generalizations) appear throughout relativity — in the trace of the identity, in $\eta^{\mu\nu}\eta_{\mu\nu}=4$ , and in the trace-reversal that turns the Ricci tensor into the Einstein tensor. A contraction has counted the dimensions of the world.

Physics of the portrait

The whole apparatus of upper and lower indices was not built for relativity at all. It was invented as pure mathematics — a general calculus of quantities that keep their meaning under any change of coordinates — a generation before Einstein needed it, by an Italian mathematician whose work was so unfashionable that it nearly vanished.

Origins · 1 minGregorio Ricci-Curbastro and how this was discovered

Gregorio Ricci-Curbastro

1853–1925 · the invention of tensor calculus and index notation

Gregorio Ricci-Curbastro spent his career at the University of Padua developing what he called the absolute differential calculus — a systematic algebra of quantities, expressed through their indexed components, that transform in a definite way when coordinates change. Building on Riemann's geometry and on Christoffel's 1869 study of quadratic differential forms, Ricci worked out through the 1880s and 1890s the rules of raising, lowering, and contracting indices, and the covariant derivative, that we still use unaltered. For years the work was regarded as an obscure formalism of interest to almost no one. The turning point was the long memoir Méthodes de calcul différentiel absolu et leurs applications, which Ricci published in 1900 in the Mathematische Annalen jointly with his former student Tullio Levi-Civita; it codified the entire calculus and, almost as an afterthought, listed physical applications. When Einstein, struggling between 1912 and 1915 to express gravity as the curvature of spacetime, was pointed by his friend Marcel Grossmann toward exactly this "absolute differential calculus", he found the mathematical language he had been missing — and in adopting it he renamed its central objects tensors, from the Latin for "that which stretches", after the elastic-stress tensors that were among its first uses. The index notation of this article, and the summation convention Einstein added in 1916 to tame it, are Ricci's calculus applied to the flat special-relativistic case: the training ground before the curved metric $g_{\mu\nu}$ replaces the constant $\eta_{\mu\nu}$ .

History

1869

— Christoffel writes quadratic differential forms in terms of the symbols now bearing his name

1892

— Ricci-Curbastro founds the absolute differential calculus, the systematic algebra of tensor components

1900

— Ricci-Curbastro and Levi-Civita publish the memoir that codifies tensor calculus for physics

1908

— Minkowski recasts special relativity as geometry on a four-dimensional spacetime

1916

— Einstein's general relativity puts tensors to work and introduces the summation convention

Going further

Texts. Misner, Thorne & Wheeler, Gravitation (1973), ch. 1–3 (the geometry of tensors) · Wald, General Relativity (1984), ch. 2–3 and appendix · Carroll, Spacetime and Geometry (2004), ch. 1 · Schutz, A First Course in General Relativity (2009), ch. 2–3 · Hartle, Gravity (2003), ch. 5. All equations and conventions use the $(-,+,+,+)$ metric. Key results in this article include: the Lorentz-group condition $\Lambda^{\mathsf T}\eta\,\Lambda=\eta$ ; the raising and lowering rules with the sign flip $V_0=-V^0$ ; the covariant field strength $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$ ; the metric inverse identity $\eta^{\mu\alpha}\eta_{\alpha\nu}=\delta^\mu{}_\nu$ ; the cancellation $\Lambda^\mu{}_\alpha(\Lambda^{-1})^\beta{}_\mu=\delta^\beta{}_\alpha$ ; the numeric example $V^\mu=(5,3,0,0)$ giving norm $-16$ , boost $\beta=0.6$ , $\gamma=1.25$ , $V'^\mu=(4,0,0,0)$ ; the component count $16=10+6$ with decomposition $T_{01}=6,\,T_{10}=2\Rightarrow T_{(01)}=4,\,T_{[01]}=2$ ; and $\delta^\mu{}_\mu=4$ . No empirical numbers are asserted beyond the convention-dependent metric entries.

Builds on: Minkowski Spacetime · Minkowski Metric and Signature · Lorentz Transformations · Four-Vectors. Leads to: Proper Time · Spacetime Interval · Causal Structure and Light Cones.

Ask the tutor: hand me any indexed expression and I will tell you whether it is a legal tensor equation, or lower, raise, and contract a four-vector for you step by step.

What is the actual difference between an upper and a lower index?

They mark opposite transformation behaviour. An upper (contravariant) index transforms with the Lorentz matrix $\Lambda$ ; a lower (covariant) index transforms with its inverse $\Lambda^{-1}$ . Geometrically, an upper index labels a vector — an arrow, a displacement, a velocity — while a lower index labels a covector or one-form, an object that eats a vector and returns a number, like a gradient. The metric $\eta$ converts between the two by flipping the sign of the time component in the $(-,+,+,+)$ signature. The reason the distinction matters is contraction: only a matched upper-lower pair transforms invariantly, so keeping the vertical position of every index straight is what guarantees your equations mean the same thing in every frame.

What is the Einstein summation convention, and why is it safe to leave the sum implicit?

The convention says: whenever an index appears exactly twice on one side of an equation, once as an upper and once as a lower, sum it over all four values $0,1,2,3$ without writing the summation sign. It is safe because such a repeated index is always a dummy — it is summed away and does not survive into the result — so it can even be renamed at will, $A^\mu B_\mu=A^\nu B_\nu$ . The convention is more than shorthand: by insisting the repeated index be one up and one down, it makes the only legal contraction the frame-invariant one, so the notation quietly refuses to let you write a non-covariant sum. An index that appears only once is free, and it must match on both sides.

Why does the metric have that minus sign, and does the sign choice matter?

The minus sign is the whole of relativistic geometry. It is what separates the time direction from the three spatial ones, giving the invariant interval $\mathrm{d}s^2=-c^2\mathrm{d}t^2+\mathrm{d}x^2+\cdots$ a fundamentally different character from a Euclidean distance and allowing non-zero vectors to have zero length (the null vectors along the light cone). The overall sign is a convention: this article uses the mostly-plus signature $(-,+,+,+)$ favoured by MTW, Wald, and Carroll, in which timelike vectors have negative norm; much of particle physics uses mostly-minus $(+,-,-,-)$ , in which every norm flips sign. No measurable quantity depends on the choice — proper time, energy, and the mass shell come out identical — but you must state your signature and never mix the two.

How do I check quickly whether an equation is a valid tensor equation?

Count indices. First, every free index — one that is not summed — must appear exactly once on each side of the equation, in the same position, upper or lower. If the left side has a free upper $\mu$ and the right side has a free lower $\mu$ , or a free $\nu$ instead, the equation is ill-formed. Second, every summed index must appear exactly twice on a side, once up and once down; two ups or two downs signal an error. This is a purely mechanical check you can run without knowing any physics, and it catches most algebra mistakes instantly — it is the relativist's equivalent of a compiler's type check. If the index structure balances, the equation is at least a legitimate tensor statement, true in every frame if true in one.

Why can partial derivatives be treated as tensors here but not in general relativity?

Because in flat spacetime with inertial (Cartesian) coordinates the metric components are constants. The four-gradient $\partial_\mu$ transforms as a covector purely by the chain rule, since the Jacobian of a Lorentz transformation is a constant matrix, and there are no correction terms. In curved spacetime, or even in flat spacetime written in curvilinear coordinates like spherical polars, the metric varies from point to point, and differentiating a tensor's components no longer yields a tensor — the basis vectors themselves change, and one must add Christoffel-symbol terms to compensate. That upgraded derivative is the covariant derivative $\nabla_\mu$ . Flat-space index notation is the special case where the connection vanishes, which is exactly why it is the right place to learn the grammar first.

Why is the electromagnetic field a tensor with six components rather than two three-vectors?

Because the electric and magnetic fields mix into one another under a boost, and only a single antisymmetric rank-2 tensor captures that mixing correctly. The field strength $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$ is antisymmetric, so in four dimensions it has exactly $4\times3/2=6$ independent components — precisely the three components of $\mathbf{E}$ and the three of $\mathbf{B}$ . Stacking $\mathbf{E}$ and $\mathbf{B}$ as two separate three-vectors would be wrong: neither is a four-vector, and a moving observer sees a magnetic field where a stationary one sees only an electric field. Writing them as the single object $F_{\mu\nu}$ makes Maxwell's equations, $\partial_\mu F^{\mu\nu}=\mu_0 J^\nu$ and $\partial_{[\mu}F_{\nu\rho]}=0$ , manifestly the same in every inertial frame.