Four-Vectors and the Metric Tensor

The relativity chapter has, so far, been about events: the postulates dethroned simultaneity, the Lorentz transformation rebuilt the coordinate grid, and Minkowski geometry organized the wreckage into an invariant interval and a light cone. This article changes gears from geography to grammar. Electrodynamics is heading somewhere specific — a formulation in which Maxwell's four equations collapse into two tensor statements and the electric and magnetic fields reveal themselves as components of a single object. To get there, every physical quantity must first be repackaged into objects with a definite transformation character: four-vectors, and the tensors built from them. The machinery has exactly two moving parts. The first is a definition with teeth: a four-vector is any quadruple of numbers that transforms between inertial frames the way the coordinates $x^\mu=(ct,x,y,z)$ do — a birth certificate, not a costume. The second is the metric tensor $\eta_{\mu\nu}=\mathrm{diag}(+1,-1,-1,-1)$ , the indefinite ruler whose single sign difference between time and space carries all of special relativity. Together they generate the entire index calculus: raising and lowering, contraction, and the manufacture of frame-invariant scalars on demand. The payoff for an electrodynamicist is concrete. Charge density and current will fuse into $J^\mu$ ; the potentials into $A^\mu$ ; the fields into the antisymmetric tensor $F^{\mu\nu}$ ; and every conservation law and wave equation of the last three chapters will be rewritten as index contractions whose frame-independence is visible at a glance. This article is deliberately a toolkit — the index gymnastics practiced here, on kinematic examples, are exactly the moves the four-current, four-potential, and field-tensor articles will perform on Maxwell's equations. Learn the grammar once; the rest of the chapter speaks it fluently.

Key takeaways

A four-vector is defined by its transformation law: $A'^\mu=\Lambda^\mu{}_\nu A^\nu$ with the same Lorentz matrix $\Lambda$ that boosts the coordinates. What a quadruple means is irrelevant; how it transforms is everything.
The Minkowski metric $\eta_{\mu\nu}=\mathrm{diag}(+1,-1,-1,-1)$ defines the invariant inner product $A\cdot B=\eta_{\mu\nu}A^\mu B^\nu=A^0B^0-\mathbf{A}\cdot\mathbf{B}$ — the same number in every inertial frame.
Lowering an index, $A_\mu=\eta_{\mu\nu}A^\nu$ , just flips the sign of the spatial components; raising undoes it. Contractions pair one upper with one lower index, never two of the same kind.
The defining property of a Lorentz transformation, $\Lambda^{\mathsf T}\eta\,\Lambda=\eta$ , is precisely what makes every such contraction a scalar.
The four-gradient $\partial_\mu=\partial/\partial x^\mu$ transforms with the inverse matrix automatically — so $\partial_\mu J^\mu$ , $\partial_\mu A^\mu$ , and $\Box=\partial_\mu\partial^\mu$ are invariant by construction, no brute-force chain rule needed.
Electrodynamics will populate this grammar with $J^\mu$ , $A^\mu$ , and $F^{\mu\nu}$ ; rank-2 tensors transform with two factors of $\Lambda$ , one per index.
Key idea: a four-vector is defined by how it transforms, and the metric turns pairs of four-vectors into numbers every inertial observer agrees on — the raw material of covariant physical law.

The core insight

A four-vector is defined by HOW it transforms, not by what it describes: any quadruple transforming like $x^\mu$ under Lorentz transformations is a four-vector, and its inner product with any other four-vector — formed with the metric — is a Lorentz scalar, identical in all frames.

Where this lives today

Every particle-physics measurement is an exercise in this article's grammar, executed a billion times per second. When the LHC's detectors record a candidate Higgs decay, the software reconstructs each daughter's four-momentum, sums them, and computes the one number worth publishing: the invariant $P^\mu P_\mu=(\sum E/c)^2-|\sum\mathbf{p}|^2$ . Because it is a scalar, physicists in the lab frame, the proton-proton center-of-momentum frame, or any boosted analysis frame extract the same resonance peak — the 125 GeV/ $c^2$ bump is a frame-proof fact precisely because it is a contraction. Detector geometry has even bent to the grammar: colliders quote particle directions in rapidity, the additive boost parameter, so that "shapes" of events are boost-covariant by construction. The metric itself has quietly become infrastructure. GPS timing algebra is Minkowski bookkeeping (plus general-relativistic corrections layered on top); pulsar-timing arrays fit arrival times with light-cone geometry across kiloparsecs; and every numerical-relativity code that predicts LIGO waveforms initializes with $\eta_{\mu\nu}$ far from the source, general relativity's $g_{\mu\nu}(x)$ reducing to our constant ruler where spacetime flattens. Within this chapter, the machinery pays its way immediately: the next two articles build the kinematic four-vectors ( $u^\mu$ , $p^\mu$ ), and then the ladder climbs through the sources and potentials to $F^{\mu\nu}$ — at which point Maxwell's equations, the ones three chapters sweated over, will be written in two strokes.

Origins · 1 minArnold Sommerfeld and how this was discovered

Arnold Sommerfeld

1868–1951 · Munich — who named the four-vector and taught relativity its index calculus

Sommerfeld is the reason this article's vocabulary exists. In two 1910 Annalen der Physik papers on the "four-dimensional vector algebra," he took Minkowski's geometric vision — its creator had died suddenly in 1909, at forty-four — and turned it into a calculus: he coined the term Vierervektor (four-vector), organized the machinery of invariant products and four-dimensional operators, and showed how electrodynamics flows through it. Einstein, who had initially dismissed the four-dimensional formalism as "superfluous learnedness," changed his mind decisively; without Sommerfeld's grammar there is no covariant electrodynamics as we write it, and arguably no tensor toolkit ready for general relativity five years later. The rest of his career compounded the legacy: his Atombau und Spektrallinien trained a generation, he supervised or mentored an absurd roster (Heisenberg, Pauli, Debye, Bethe, Pauling), was nominated for the Nobel Prize eighty-four times without winning, and his relativistic fine-structure constant $\alpha$ — born of applying this chapter's four-momentum to the hydrogen atom — still measures the strength of electromagnetism. The four-vector was named by the great teacher of twentieth-century physics, and it shows: the notation is a lesson plan.

History

1905

— Einstein's electrodynamics paper: the transformation laws exist, but as coordinate algebra, with no geometric packaging

1907

— Minkowski begins recasting relativity as four-dimensional geometry; time and space fuse into one continuum with an indefinite metric

1908

— Minkowski's Cologne address "Space and Time" makes the union public; the interval and light cone become the organizing structure

1910

— Sommerfeld coins "four-vector" and systematizes the 4D vector algebra; Bateman and Cunningham note the conformal covariance of Maxwell's equations

1913

— Minkowski's formalism, championed by Sommerfeld and Laue, becomes the standard language of Laue's first relativity textbook

1916

— Einstein's general relativity promotes the constant metric to a dynamical field

g_{\mu\nu}(x)

, with

\eta_{\mu\nu}

as its flat-spacetime limit

1956

— Synge's "Relativity: The Special Theory" completes the geometrization; four-vectors become the native language of particle kinematics

2012

— The Higgs discovery is announced as a peak in a Lorentz-invariant contraction: four-vector algebra as experimental instrument

The physical picture

Start with an ordinary arrow on a tabletop. Two surveyors with rotated grids assign it different $(x,y)$ component pairs, yet both compute the same length, because the Euclidean ruler $d\ell^2=dx^2+dy^2$ is rotation-blind. Neither component list is "the truth"; the arrow is. Special relativity forces the same discipline one dimension up and with one crucial twist: observers in relative motion assign different time and space components to the same pair of events, and the thing they agree on is not a Pythagorean length but the interval $s^2=(ct)^2-x^2-y^2-z^2$ , with its insolent minus signs. The object whose "length" that is — the displacement between two events — is the prototype four-vector, and the indefinite ruler that measures it is the metric. The minus signs are not a bookkeeping nuisance; they are the physics. A ruler that weighed time and space with the same sign would make the invariance group of spacetime an ordinary four-dimensional rotation group, and nothing could distinguish the time direction. The relative sign is what creates a light cone ( $s^2=0$ ), a causal order (Spacetime and Minkowski Geometry proved it frame-invariant), and the hyperbolic — rather than circular — character of boosts. One sign, the whole of relativity. Now the move that gives this article its power: promote the transformation law from a property to a definition. Anything — any quadruple of measured numbers, whatever apparatus produced them — that transforms under a boost exactly as $x^\mu$ does is declared a four-vector, and instantly inherits the entire geometric toolkit: an invariant norm, invariant inner products with every other four-vector, and a guaranteed slot in covariant equations. This is how relativity industrializes. We will not check, case by case, whether energy mixes with momentum or charge density with current density; we will construct quadruples certified to transform correctly ( $p^\mu$ , $J^\mu$ , $A^\mu$ ), and then any equation between such objects, with indices honestly contracted, holds in every inertial frame the moment it holds in one. The strategy of the next seven articles is exactly this, and nothing more.

Figure

One spacetime arrow, two component lists — frames disagree on the components, the metric makes them agree on the number

The math: the metric, the index machine, and why contractions are scalars

Everything begins with the prototype. An event is labeled by four coordinates gathered into the position four-vector, whose zeroth entry is time converted to a length,

x^\mu=(x^0,x^1,x^2,x^3)=(ct,\,x,\,y,\,z). \qquad (1)

where: the Greek index $\mu$ runs over $0,1,2,3$ ; Latin indices ( $i,j$ ) will always mean the spatial $1,2,3$ only. The factor of $c$ on the time slot is not decoration — it gives all four components the same dimension (length), which is what permits them to mix under boosts. A Lorentz transformation is a linear reshuffle of these labels, $x'^\mu=\Lambda^\mu{}_\nu x^\nu$ , with the summation convention in force: a repeated upper–lower index pair is summed over $0\ldots3$ automatically. For the standard boost along $x$ at speed $v=\beta c$ (the matrix derived in Lorentz Transformations),

\Lambda^\mu{}_\nu=\begin{pmatrix}\gamma & -\gamma\beta & 0 & 0\\ -\gamma\beta & \gamma & 0 & 0\\ 0&0&1&0\\ 0&0&0&1\end{pmatrix},\qquad \gamma=\frac{1}{\sqrt{1-\beta^2}}. \qquad (2)

where: the first index of $\Lambda^\mu{}_\nu$ labels the row (output component), the second the column (input). Read physically: the new time is the old time minus $\beta\times$ (old position), all stretched by $\gamma$ — simultaneity mixing, time dilation, and length contraction in one matrix. Now the definition that carries the chapter: a four-vector is any quadruple $A^\mu$ that transforms with this same matrix, $A'^\mu=\Lambda^\mu{}_\nu A^\nu$ , whenever the frame changes. Upper-index components are called contravariant. The geometry enters through the metric. The invariant interval of Spacetime and Minkowski Geometry is a quadratic form in the displacement, and we encode its coefficient pattern once and for all:

\eta_{\mu\nu}=\mathrm{diag}(+1,-1,-1,-1),\qquad s^2=\eta_{\mu\nu}\,x^\mu x^\nu=(ct)^2-x^2-y^2-z^2. \qquad (3)

where: $\eta_{\mu\nu}$ is the Minkowski metric tensor in the $(+,-,-,-)$ signature — the convention used throughout this track and its mechanics companions, and standard in particle physics and in Griffiths. (Much of the general-relativity literature uses $(-,+,+,+)$ ; only the signs of declared invariants flip, never the physics — but never mix conventions mid-calculation.) The metric is symmetric, constant, and indefinite: it is emphatically not the identity matrix, and that fact is the entire difference between spacetime and four-dimensional Euclidean space. The metric's first job is to define the inner product of any two four-vectors, and with it the norm:

A\cdot B\;\equiv\;\eta_{\mu\nu}A^\mu B^\nu\;=\;A^0B^0-\mathbf{A}\cdot\mathbf{B},\qquad A^2\equiv A\cdot A. \qquad (4)

where: $\mathbf{A}\cdot\mathbf{B}$ is the ordinary three-dimensional dot product of the spatial parts. The relative minus sign is the light cone speaking: a four-vector is called timelike ( $A^2>0$ ), null ( $A^2=0$ ), or spacelike ( $A^2<0$ ) by the sign of its norm, and the classification is frame-independent because the norm is. Derivation A below proves that frame-independence honestly. The metric's second job is bookkeeping: it converts between the two natural component types. Define the covariant (lower-index) components by

A_\mu=\eta_{\mu\nu}A^\nu\quad\Longleftrightarrow\quad A_0=A^0,\;\;A_i=-A^i. \qquad (5)

where: in this signature, lowering an index costs nothing but three sign flips on the spatial slots. The inverse metric $\eta^{\mu\nu}$ (numerically the same diagonal matrix, satisfying $\eta^{\mu\rho}\eta_{\rho\nu}=\delta^\mu{}_\nu$ ) raises indices back. With both species available, the inner product collapses to the clean pairing $A\cdot B=A^\mu B_\mu=A_\mu B^\mu$ — and the grammar rule of the entire subject emerges: legal expressions contract one upper index with one lower index, never two of a kind. An expression like $A^\mu B^\mu$ (both up) is a typo that happens to equal $A^0B^0+\mathbf{A}\cdot\mathbf{B}$ in one frame and something else in every other — it is not a scalar, and the notation is designed to make it look as wrong as it is. Why does the honest pairing produce invariants? Because of the one condition a matrix must satisfy to be a Lorentz transformation — the metric-preservation condition established in Lorentz Transformations:

\eta_{\mu\nu}\,\Lambda^\mu{}_\rho\,\Lambda^\nu{}_\sigma=\eta_{\rho\sigma}\qquad\text{(matrix form: }\Lambda^{\mathsf T}\eta\,\Lambda=\eta\text{)}. \qquad (6)

where: this is to $\eta$ what orthogonality $R^{\mathsf T}R=\mathbb{1}$ is to the Euclidean metric: the statement that $\Lambda$ preserves the ruler. Its consequences ripple everywhere: taking determinants gives $\det\Lambda=\pm1$ ; its $(0,0)$ component forces $|\Lambda^0{}_0|\ge1$ ; and — the workhorse — it makes every properly contracted expression frame-invariant, as we now show.

Derivation ACovariant components transform inversely and contractions are scalars

The goal is the load-bearing theorem of the index calculus: if $A^\mu$ and $B^\mu$ are four-vectors, the contraction $A^\mu B_\mu$ is a Lorentz scalar — the same number in every inertial frame — and the mechanism is that lower-index components transform with the inverse transformation, so the two matrices cancel in the pairing. The strategy is direct computation from the definitions, using the metric-preservation condition (6) exactly once, at the pivotal step; watching where it enters shows precisely why "transforms like $x^\mu$ " plus "contract upper with lower" is a scalar-manufacturing machine with no further conditions needed.

Step 1. Transform the covariant components from their definition. In the primed frame, $A'_\mu=\eta_{\mu\nu}A'^\nu=\eta_{\mu\nu}\Lambda^\nu{}_\rho A^\rho$ . Raise the dummy index on $A^\rho$ with the metric pair $A^\rho=\eta^{\rho\sigma}A_\sigma$ :

A'_\mu=\left(\eta_{\mu\nu}\,\Lambda^\nu{}_\rho\,\eta^{\rho\sigma}\right)A_\sigma\;\equiv\;\Lambda_\mu{}^\sigma A_\sigma.

Why the sandwich matters: the object $\Lambda_\mu{}^\sigma=\eta_{\mu\nu}\Lambda^\nu{}_\rho\eta^{\rho\sigma}$ is the transformation matrix with both indices shifted by the metric — and the next step shows it is not a new matrix at all, but the inverse-transpose of $\Lambda$ in disguise.

Step 2. Identify the sandwich as the inverse. Multiply the metric-preservation condition (6), $\eta_{\mu\nu}\Lambda^\mu{}_\rho\Lambda^\nu{}_\sigma=\eta_{\rho\sigma}$ , by $\eta^{\sigma\tau}$ and rearrange:

\left(\eta_{\sigma\nu}\,\Lambda^\nu{}_\rho\,\eta^{\rho\tau}\right)\Lambda^\sigma{}_\kappa \;=\;\delta^\tau{}_\kappa \quad\Longrightarrow\quad \Lambda_\sigma{}^\tau=\left(\Lambda^{-1}\right)^\tau{}_\sigma.

Why this is the punchline in miniature: covariant components do not need a separately postulated transformation law — the metric forces them to transform with $\Lambda^{-1}$ . In matrix language, $A'_{\text{(row)}}=A_{\text{(row)}}\Lambda^{-1}$ : rows transform inversely to columns.

Step 3. Contract and watch the cancellation:

A'^\mu B'_\mu=\Lambda^\mu{}_\nu A^\nu\left(\Lambda^{-1}\right)^\sigma{}_\mu B_\sigma=\left(\Lambda^{-1}\right)^\sigma{}_\mu\Lambda^\mu{}_\nu\,A^\nu B_\sigma=\delta^\sigma{}_\nu A^\nu B_\sigma=A^\nu B_\nu.

The transformation matrices annihilate each other pairwise. One upper index (a factor of $\Lambda$ ) paired with one lower index (a factor of $\Lambda^{-1}$ ) is the only pairing that self-destructs this way — which is the entire content of the up-down contraction rule.

Step 4. Read off the corollaries. Setting $B=A$ : every norm $A^\mu A_\mu$ is invariant — for the displacement this recovers the invariant interval, for the four-momentum it will define the rest mass, and for the four-potential contracted with the four-current it will build an invariant interaction. The theorem also runs in reverse (the quotient rule): if $A^\mu B_\mu$ is a scalar for every four-vector $B$ , then $A^\mu$ must be a four-vector — the certificate we will use to induct $J^\mu$ and $A^\mu$ into the club without re-deriving boost algebra each time. As a concrete check: for the boost (2) at $\beta=0.6$ ( $\gamma=1.25$ ), take $A^\mu=(4,5,0,0)$ and $B^\mu=(2,1,0,0)$ . Boosting gives $A'^\mu=(1.25,3.25,0,0)$ and $B'^\mu=(1.75,-0.25,0,0)$ ; the contractions are $A\cdot B=4\cdot2-5\cdot1=3$ and $A'\cdot B'=1.25\cdot1.75-3.25\cdot(-0.25)=2.1875+0.8125=3$ — equal in both frames, as the theorem requires.

The second pillar of the toolkit is a four-vector nobody has to build, because calculus supplies it. Define the four-gradient

\partial_\mu\equiv\frac{\partial}{\partial x^\mu}=\left(\frac{1}{c}\frac{\partial}{\partial t},\,\nabla\right),\qquad \partial^\mu=\eta^{\mu\nu}\partial_\nu=\left(\frac{1}{c}\frac{\partial}{\partial t},\,-\nabla\right). \qquad (7)

where: the lower index on $\partial_\mu$ is not a convention but a discovery — differentiating with respect to an upper-index coordinate produces an object that transforms covariantly, automatically. This one fact is why the conservation laws and wave equations of electrodynamics will covariantize almost without effort.

Go deeper · Derivation B · 2 minThe four-gradient is born covariant and the d'Alembertian is a scalar operator

Derivation BThe four-gradient is born covariant and the d'Alembertian is a scalar operator

The goal is to show that $\partial_\mu$ transforms with $\Lambda^{-1}$ — i.e., it is a legitimate lower-index four-vector operator — using nothing but the chain rule, and then to harvest the two consequences electrodynamics lives on: $\partial_\mu(\text{four-vector})^\mu$ is a scalar, and $\Box=\partial_\mu\partial^\mu$ is an invariant operator. The strategy is short and honest; the point of spelling it out is that Lorentz Transformations obtained the invariance of the wave operator by a brute-force chain-rule slog through $\partial^2/\partial t^2$ and $\partial^2/\partial x^2$ — here the index machinery delivers the same result in three lines, which is a fair advertisement for the whole formalism.

Step 1. Chain rule on a frame change. For the linear transformation $x'^\mu=\Lambda^\mu{}_\nu x^\nu$ , any differentiable function obeys

\frac{\partial}{\partial x'^\mu}=\frac{\partial x^\nu}{\partial x'^\mu}\frac{\partial}{\partial x^\nu}=\left(\Lambda^{-1}\right)^\nu{}_\mu\frac{\partial}{\partial x^\nu},

since inverting the linear map gives $x^\nu=(\Lambda^{-1})^\nu{}_\mu x'^\mu$ exactly. The derivative operator picks up one factor of the inverse matrix — the defining behavior of a covariant index (Derivation A, Step 2). No metric was needed; the gradient is covariant by birth.

Step 2. Contract with any four-vector field $V^\mu(x)$ :

\partial'_\mu V'^\mu=\left(\Lambda^{-1}\right)^\nu{}_\mu\Lambda^\mu{}_\rho\,\partial_\nu V^\rho=\delta^\nu{}_\rho\,\partial_\nu V^\rho=\partial_\nu V^\nu.

The four-divergence of a four-vector field is a Lorentz scalar. Why this line is worth a chapter: when $V^\mu$ becomes the four-current, $\partial_\mu J^\mu=0$ will be charge conservation as a one-frame-implies-all-frames statement; when it becomes the four-potential, $\partial_\mu A^\mu=0$ will be the Lorenz condition, invariantly imposed — both delivered in this chapter, on this line's authority.

Step 3. Contract the gradient with itself:

\Box\equiv\partial_\mu\partial^\mu=\frac{1}{c^2}\frac{\partial^2}{\partial t^2}-\nabla^2,

a scalar operator by the same cancellation: $\Box'=\Box$ acting on any scalar or on each component of a tensor equation. This is the d'Alembertian of the wave chapter, and its invariance — the fact Lorentz Transformations extracted from Maxwell's equations by demanding $\Box'=\Box$ under linear maps — is here a two-line corollary of index grammar. The wave equation $\Box A^\mu=\mu_0J^\mu$ , waiting four articles ahead, will be manifestly frame-proof for exactly this reason. As a sign audit in $(+,-,-,-)$ : $\partial^\mu=\eta^{\mu\nu}\partial_\nu$ flips only the spatial slots, so $\partial_\mu\partial^\mu=(1/c^2)\partial_t^2-\nabla^2$ , reproducing the wave operator with propagation speed $c$ ; the chain-rule matrix $\partial x^\nu/\partial x'^\mu$ for the boost (2) is (2) with $\beta\to-\beta$ , which is indeed $\Lambda^{-1}$ (confirming $\gamma^2(1-\beta^2)=1$ ).

One more storey completes the building: objects with several indices. A rank-2 tensor is a sixteen-component object transforming with one $\Lambda$ per upper index and one $\Lambda^{-1}$ per lower index,

T'^{\mu\nu}=\Lambda^\mu{}_\rho\,\Lambda^\nu{}_\sigma\,T^{\rho\sigma}\qquad\text{(matrix form: }T'=\Lambda\,T\,\Lambda^{\mathsf T}\text{)}. \qquad (8)

where: the transposition in the matrix form is just the second index contracting with the second $\Lambda$ 's column slot. Symmetry properties ( $T^{\mu\nu}=\pm T^{\nu\mu}$ ) survive transformation, so an antisymmetric tensor stays antisymmetric in every frame — which is why the field tensor $F^{\mu\nu}$ , antisymmetric with exactly six independent components, can hold three components of $\mathbf{E}$ and three of $\mathbf{B}$ frame-proof. The metric itself is the trivial case: by (6), $\eta$ "transforms" into itself — the one tensor all observers share, which is what makes it usable as a universal ruler. The roll-call, then, for the road ahead: $x^\mu$ and the four-velocity and four-momentum (next two articles) carry particle kinematics; $J^\mu=(c\rho,\mathbf{J})$ carries the sources; $A^\mu=(V/c,\mathbf{A})$ the potentials; $F^{\mu\nu}$ the fields; and scalars manufactured from them — $p^\mu p_\mu$ , $\partial_\mu J^\mu$ , $k_\mu x^\mu$ , $F_{\mu\nu}F^{\mu\nu}$ — will carry the invariant physics. Every one of those objects earns its index by the birth-certificate test, and every equation of covariant electrodynamics is a statement that two objects of the same index type are equal.

Figure

The index machine — the metric lowers, the inverse metric raises, and one up-down pair contracts to a frame-proof scalar

Limiting cases

case	result	meaning
$\beta\to0$	$\Lambda\to\mathbb{1}+O(\beta)$ , $A'^0\approx A^0$ , $\mathbf{A}'\approx\mathbf{A}-\beta A^0\hat{\mathbf{x}}$	four-vectors decouple into a Galilean scalar plus 3-vector; index machinery becomes invisible, not wrong
pure spatial rotation	$\Lambda=\begin{pmatrix}1&0\\0&R\end{pmatrix}$	$A^0$ is a rotational scalar, $\mathbf{A}$ an ordinary 3-vector: the calculus contains all of vector analysis
null four-vector	$A^\mu A_\mu=0$ with $A^\mu\ne0$	possible only because $\eta$ is indefinite; photons ( $k^\mu$ , $p^\mu$ ) live here
all signs $+$ (hypothetical)	$\eta\to\delta$ , invariance group $SO(4)$	no light cone, no causality classes — the minus signs are the physics
$(-,+,+,+)$ signature	all norms flip sign; timelike means $A^2<0$	dialect, not physics; this track and its mechanics companions lock $(+,-,-,-)$
interval between infinitesimally close events	$ds^2=\eta_{\mu\nu}dx^\mu dx^\nu$	the differential form that defines proper time in the next article

Common trapThe inner product is A⁰B⁰ MINUS A·B — and index position is information, not typography

Two errors account for most wrong answers in this subject. First, the sign: the Minkowski inner product subtracts the spatial dot product, $A\cdot B=A^0B^0-\mathbf{A}\cdot\mathbf{B}$ in this signature. Write the Euclidean $+$ out of reflex and every "invariant" you compute will silently change from frame to frame — the error usually surfaces two pages later as a rest mass that depends on the observer. Second, index altitude: $A_\mu$ and $A^\mu$ differ by real minus signs on the spatial slots, so sloppiness like " $A_\mu B_\mu$ " is not casual notation, it is a different (and frame-dependent) number. The discipline that prevents both: every repeated index must appear exactly once up and once down, and free indices must match in position on both sides of any equation. Treat a violation the way you would treat adding meters to seconds — as a type error, before ever checking the arithmetic.

Worked examples

1 · One boost, one invariant — the full index workout (confidence builder)

Setup. In frame $S$ , two events are separated by $\Delta x^\mu=(4,5,0,0)$ (units: light-seconds, i.e. $c\Delta t=4$ , $\Delta x=5$ ). Frame $S'$ moves at $\beta=0.6$ along $x$ . Find the primed components, both versions of the covariant components, and the interval in both frames.

Step 1. The boost numbers: $\gamma=(1-0.36)^{-1/2}=1.25$ , $\gamma\beta=0.75$ .

Step 2. Transform with (2): $\Delta x'^0=\gamma(\Delta x^0-\beta\,\Delta x^1)=1.25(4-3)=\mathbf{1.25}$ and $\Delta x'^1=\gamma(\Delta x^1-\beta\,\Delta x^0)=1.25(5-2.4)=\mathbf{3.25}$ . The transverse slots ride along unchanged.

Step 3. Lower indices in each frame by flipping spatial signs, per (5): $\Delta x_\mu=(4,-5,0,0)$ and $\Delta x'_\mu=(1.25,-3.25,0,0)$ .

Step 4. Contract, per (4): $\Delta x^\mu\Delta x_\mu=16-25=\mathbf{-9}$ and $\Delta x'^\mu\Delta x'_\mu=1.5625-10.5625=\mathbf{-9}$ . Identical, as Derivation A guarantees. The negative sign classifies the separation as spacelike: no signal connects these events (invariant statement), an observer exists for whom they are simultaneous ( $\beta_{\text{sim}}=\Delta x^0/\Delta x^1=0.8$ would do it), and their invariant proper distance is $\sqrt{9}=3$ light-seconds.

The lesson: the full pipeline — transform, lower, contract — takes four lines, and the invariant at the end is the only piece of the calculation all observers share.

2 · A photon's four-vector under a boost — index gymnastics meet the Doppler factor (exam level)

Setup. A plane wave (wavelength $\lambda=500$ nm, green) propagates along $+x$ in frame $S$ . Package its phase into four-vector grammar; then find what an observer receding at $\beta=0.6$ measures.

Step 1. The phase of the E8 plane wave, $\varphi=kx-\omega t$ , must be a scalar — all observers agree whether a given event sits on a crest (a wave counts its own crests; no frame can renegotiate an integer). Write it as a contraction and read off the wave four-vector:

\varphi=-k_\mu x^\mu\quad\text{with}\quad k^\mu=\left(\frac{\omega}{c},\,\mathbf{k}\right)=\frac{2\pi}{\lambda}(1,1,0,0),

where the last equality uses $\omega=ck$ for vacuum light: $k=2\pi/\lambda=1.257\times10^7$ m⁻¹, $\omega=3.77\times10^{15}$ rad/s. By the quotient rule of Derivation A (scalar contraction with arbitrary $x^\mu$ ), $k^\mu$ is certified a four-vector.

Step 2. Check its norm before boosting: $k^\mu k_\mu=(\omega/c)^2-k^2=0$ . Null — the four-vector signature of anything traveling at $c$ , and an invariant, so it will be null for every observer without further computation.

Step 3. Boost with (2) at $\beta=0.6$ (receding observer):

k'^0=\gamma\left(k^0-\beta k^1\right)=1.25\,(1-0.6)\,\frac{2\pi}{\lambda}=0.5\,\frac{2\pi}{\lambda}\quad\Longrightarrow\quad \omega'=0.5\,\omega.

Likewise $k'^1=\gamma(k^1-\beta k^0)=0.5\,(2\pi/\lambda)$ : still null ( $\omega'=ck'$ ), as promised. The received wavelength is $\lambda'=2\pi c/\omega'=\mathbf{1000}$ nm — the green photon arrives as near-infrared, redshifted by the factor $\gamma(1-\beta)=\sqrt{(1-\beta)/(1+\beta)}=0.5$ .

Step 4. Take stock. We have derived a piece of the relativistic Doppler effect as an index drill — one matrix multiplication on a certified four-vector. The full physics (arbitrary angles, aberration of starlight, the transverse Doppler shift that has no classical counterpart) belongs to Relativistic Doppler Effect and Aberration later in this chapter; the point here is the method: package, certify, boost, read off. Nothing about waves was used beyond "phase is a scalar."

The lesson: once a quantity is packaged as a four-vector, questions that look like new physics (what frequency does a moving observer see?) become one line of linear algebra.

Going further

References. Griffiths, Introduction to Electrodynamics, 4th ed., §12.1 · Jackson, Classical Electrodynamics, 3rd ed., §11.6 · Zangwill, Modern Electrodynamics, Ch. 22 · Purcell & Morin, Electricity and Magnetism, 3rd ed., Ch. 5 The mostly-minus signature $(+,-,-,-)$ gives $A^\mu B_\mu=A^0B^0-\mathbf{A}\cdot\mathbf{B}$ . Key self-consistency checks built into the formalism: the boost-matrix identity $\gamma^2(1-\beta^2)=1$ ; index-lowering sign flips with $\eta=\mathrm{diag}(+1,-1,-1,-1)$ ; the mass-shell relation $E^2/c^2-|\mathbf{p}|^2=m^2c^2$ from $p^\mu=(E/c,\mathbf{p})$ ; and $\det\Lambda=\pm1$ from $\det(\Lambda^{\mathsf T}\eta\Lambda)=\det\eta$ . No empirical constants are quoted beyond definitions. This article takes the metric-preservation condition (6) as input and derives the inverse-transpose theorem for covariant components (Derivation A) and the born-covariant four-gradient with the two-line d'Alembertian invariance (Derivation B). Worked numbers: boost $\beta=0.6$ gives $\Delta x'^\mu=(1.25,3.25,0,0)$ , interval $-9$ in both frames, cross-contraction $A\cdot B=3$ in both frames; photon drill: $\lambda=500\to1000$ nm via $\gamma(1-\beta)=0.5$ , null norm preserved.

Builds on: Postulates of Special Relativity · Lorentz Transformations · Spacetime and Minkowski Geometry Leads to: Proper Time and Four-Velocity · Relativistic Energy and Four-Momentum · The Four-Current Density · The Four-Potential · The Electromagnetic Field Tensor

Is a four-vector just any list of four quantities with the same units?

No — and this is the single most important idea of the article. Units are necessary (components that mix under boosts must be commensurable) but nowhere near sufficient. The quadruple $(E/c,\mathbf{p})$ is a four-vector; the quadruple (temperature, heat flux $/c$ ) is not, even though it has four slots, because it does not transform with $\Lambda$ . The test is always the transformation law, verified either directly (the object is built from certified four-vectors and scalars, like $p^\mu=m\,dx^\mu/d\tau$ ) or via the quotient rule (its contraction with arbitrary four-vectors is a scalar). The payoff of passing the test is enormous — instant invariants, instant covariant equations — which is why each new object in this chapter ( $J^\mu$ , $A^\mu$ ) will be explicitly walked through its certification rather than waved in on units alone.

Why do we need two kinds of components for the same vector — isn't that redundant?

It is exactly as redundant as having both a vector and the row-vector that measures it — which is to say, the redundancy is the mechanism. In Euclidean space the distinction exists but hides, because the metric is the identity and $A_i=A^i$ numerically. Minkowski's indefinite metric makes the two species differ (by spatial sign flips), and the payoff of keeping both is that frame dependence cancels automatically in any one-up-one-down pairing: contravariant components carry a factor $\Lambda$ , covariant ones carry $\Lambda^{-1}$ , and the contraction is a scalar with no further conditions. Working with only upper indices is possible but means writing $\eta_{\mu\nu}$ explicitly in every product and tracking signs by hand — the two-species notation is sign management, industrialized.

Does the choice of signature +,−,−,− versus −,+,+,+ change any physics?

None whatsoever — it changes the sign of declared invariants and some intermediate bookkeeping. In $(-,+,+,+)$ the interval of a timelike separation is negative, the four-velocity norm is $-c^2$ , and the mass-shell relation reads $p^\mu p_\mu=-m^2c^2$ . Every observable (frequencies, forces, cross-sections) comes out identical. The danger is purely operational: importing a formula from a general-relativity text (mostly $-,+,+,+$ ) into a particle-physics calculation (mostly $+,-,-,-$ ) without converting is a genuine and common error. This track and its mechanics companions lock $(+,-,-,-)$ throughout — the base convention of Griffiths ch. 12 — and every cross-referenced result has been checked to carry it.

Why does the derivative operator get a lower index when the coordinate has an upper one?

Because differentiation inverts the transformation, mechanically. If new coordinates are stretched relative to old ones, then a function's rate of change per unit new coordinate is squeezed by the inverse factor — the chain rule $\partial/\partial x'^\mu=(\partial x^\nu/\partial x'^\mu)\,\partial/\partial x^\nu$ says nothing else. Since $\partial x^\nu/\partial x'^\mu$ is exactly $(\Lambda^{-1})^\nu{}_\mu$ for a linear map, the gradient transforms the way lower-index objects do, by birth. This is a small miracle worth savoring: nature hands us, for free, exactly the object needed to contract against all the upper-index physical four-vectors — which is why conservation laws ( $\partial_\mu J^\mu=0$ ) and gauge conditions ( $\partial_\mu A^\mu=0$ ) will covariantize in one line each later in this chapter.

What exactly is the metric tensor doing physically — is it a field, a convention, or a law?

In special relativity it is the geometry itself, stated as a machine: feed it two displacement vectors and it returns their invariant interval, encoding everything the postulates established — the light cone, causal structure, time dilation. It is not a convention (its indefinite signature is forced by the invariance of $c$ ; only the overall sign is ours to choose) and not a dynamical field (it is the same constant matrix at every event, which is what "flat spacetime" means). The deepest way to see its status: $\eta$ is the unique rank-2 tensor invariant under all Lorentz transformations — the one ruler every inertial observer shares, which is precisely what qualifies it to convert between component species and manufacture scalars. In general relativity this last property is promoted: $g_{\mu\nu}(x)$ becomes position-dependent and dynamical, gravity appears as its curvature, and our $\eta$ survives as the local value any freely falling observer measures.

How does the antisymmetric field tensor manage to hold both E and B — and why couldn't they each be four-vectors?

Count and transform. An antisymmetric rank-2 tensor in four dimensions has exactly $4\cdot3/2=6$ independent components — precisely three plus three — and under the spatial-rotation subgroup those six split into two 3-vectors. Under boosts, however, the two triples mix into each other (with one $\Lambda$ per index, per equation (8)) — and that mixing is not a formal curiosity but observed physics: Transformation of Electric and Magnetic Fields derives, and Magnetism as a Relativistic Effect dramatizes, the fact that one frame's $\mathbf{E}$ is another frame's $\mathbf{E}$ -and- $\mathbf{B}$ cocktail. That behavior is fatal to any attempt to package $\mathbf{E}$ alone as (the spatial part of) a four-vector: a four-vector's spatial part mixes with its own time component under boosts, not with a different 3-vector, and there is no scalar on hand to fill either time slot. The unique container with the right size and the right mixing is the antisymmetric tensor — built explicitly in The Electromagnetic Field Tensor.

Are there useful four-component objects that are NOT four-vectors?

Yes, and knowing them prevents subtle errors. The classic in this chapter: $(\rho,\mathbf{J})/$ anything is a four-vector only because charge is invariant — the analogous package for mass density fails at relativistic speeds (mass density picks up two $\gamma$ factors, energy being a component of a four-vector itself, and ends up part of a rank-2 tensor instead — the stress-energy object built later in this chapter). Coordinate velocity $(c,d\mathbf{x}/dt)$ fails because $dt$ in the denominator is not a scalar — the fix (divide by proper time instead) is the next article's whole business. And the electric field cannot be the spatial part of any four-vector, as the previous FAQ argued. The pattern in all three failures is the same: a denominator or a density that secretly transforms. The cure is always to rebuild the quantity from certified scalars ( $d\tau$ , $\rho_0$ , $m$ ) times certified four-vectors — the construction principle the following articles apply verbatim.