The Minkowski Metric

Euclid's geometry is carried by one formula: $\mathrm{d}\ell^2 = \mathrm{d}x^2 + \mathrm{d}y^2 + \mathrm{d}z^2$ . Every theorem about lengths, angles, circles and straight lines follows from that sum of squares, and the rigid motions of space are exactly the transformations leaving it alone. Minkowski's insight, delivered in Cologne in September 1908, was that spacetime has a formula of its own — the same formula, with one sign changed.

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

One minus sign, and everything is different. A sum of squares is positive definite: no nonzero vector has zero length, and the level sets are spheres. A difference of squares is indefinite: nonzero vectors can have zero norm — those are the light rays — and the level sets are hyperbolae that never cross their asymptotes. The rigid motions are no longer rotations but boosts, and $\cos^2 + \sin^2 = 1$ becomes $\cosh^2 - \sinh^2 = 1$ . And the indefiniteness has a consequence nobody expects. The defining condition of a Lorentz transformation, $\Lambda^{\mathsf{T}}\eta\,\Lambda = \eta$ , forces — by itself, with no physics added — $\det\Lambda = \pm 1$ and $(\Lambda^{0}{}_{0})^2 \geq 1$ : two signs that cannot vary continuously. The Lorentz group is therefore not one connected thing but four disconnected pieces, and every experiment ever performed lives in exactly one of them. Parity and time reversal are the other three. That is what a metric signature buys you.

Key takeaways

The Minkowski metric $\eta_{\mu\nu}$ is the constant, symmetric, non-degenerate tensor whose contraction with a coordinate separation gives the invariant interval, $\mathrm{d}s^2 = \eta_{\mu\nu}\mathrm{d}x^{\mu}\mathrm{d}x^{\nu}$ . It is the geometry of flat spacetime, and its signature is what makes that geometry Lorentzian rather than Euclidean.
This chapter's locked convention is $(+,-,-,-)$ , timelike-positive: $\eta = \mathrm{diag}(+1,-1,-1,-1)$ , so $\mathrm{d}s^2 > 0$ is timelike, $c^2\mathrm{d}\tau^2 = +\mathrm{d}s^2$ , $L_0 = \sqrt{-\Delta s^2}$ , and the four-velocity has $U\cdot U = +c^2$ .
$\eta$ is its own inverse, so raising and lowering an index simply flips the sign of the three spatial components — hence the invariant scalar product $A\cdot B = \eta_{\mu\nu}A^{\mu}B^{\nu} = A^{\mu}B_{\mu} = A^0B^0 - \vec{A}\cdot\vec{B}$ .
A Lorentz transformation is defined as a linear map preserving $\eta$ : $\Lambda^{\mathsf{T}}\eta\,\Lambda = \eta$ . This is exactly the statement that the interval is preserved for every separation, and it is where every relativistic transformation law comes from.
Key idea: from $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ alone, taking determinants gives $\det\Lambda = \pm 1$ , and the $00$ component gives $(\Lambda^{0}{}_{0})^2 - \sum_i (\Lambda^{i}{}_{0})^2 = 1$ , so $|\Lambda^{0}{}_{0}| \geq 1$ and the band $-1 < \Lambda^{0}{}_{0} < 1$ is EMPTY. Two discrete sign invariants that no continuous path can flip — so the Lorentz group has exactly FOUR connected components, and physics lives in the proper orthochronous one. The metric's indefiniteness is the entire reason.

The core insight

A metric is not a coordinate system; it is the machine that turns coordinate differences into physical numbers. Give me $\eta$ and I can compute every invariant in special relativity — proper times, proper lengths, invariant masses, the phase of a wave — without knowing who is measuring. Take $\eta$ away and the four coordinates are meaningless labels. The signature carries the physics: one plus because there is one time direction, three minuses because there are three of space, and the mixture because a light ray must cover a nonzero distance in zero interval.

Where this lives today

The metric is the piece of special relativity that never gets superseded — it gets promoted. In general relativity the constant array $\eta_{\mu\nu}$ becomes a field, $g_{\mu\nu}(x)$ , and that field is gravity: Einstein's equations are equations for the metric, and every observable — light bending, orbital precession, a black-hole ringdown — is a statement about arc length computed with $g$ . The weak-field expansion $g_{\mu\nu} = \eta_{\mu\nu} + h_{\mu\nu}$ makes it explicit, and $h_{\mu\nu}$ propagating at $c$ is a gravitational wave: LIGO measures $h \sim 10^{-21}$ , a deviation from Minkowski's metric twenty-one decimal places out. In quantum field theory $\eta$ is everywhere and invisible: it is what makes $p^2 = m^2c^2$ Lorentz-invariant, and what forces the relative sign in the Dirac and Klein–Gordon operators. And the four-component structure derived above is no curiosity — it is why parity ( $P$ ) and time reversal ( $T$ ) are discrete questions, askable separately of every interaction, and why their answers surprised everyone: the weak interaction violates $P$ maximally (Wu, 1957) and violates $CP$ (Christenson, Cronin, Fitch and Turlay, 1964). Neither result would be well-posed if the Lorentz group were connected. The signature is a target too: rotating optical resonators, clock comparisons and astrophysical photon-dispersion bounds constrain departures from $\eta$ below $10^{-18}$ , and have found nothing.

The physical picture

Rotate the axes of a Euclidean plane and a fixed vector's components change but its length does not: the invariant is $x^2 + y^2$ , the level sets are circles, and the transformations carry an angle $\theta$ with $\cos^2\theta + \sin^2\theta = 1$ . Boost the axes of the $(ct, x)$ plane and the same story runs with one sign changed: the invariant is $c^2t^2 - x^2$ , the level sets are hyperbolae, and the transformations carry a rapidity $\varphi$ with $\cosh^2\varphi - \sinh^2\varphi = 1$ . The boost matrix is a hyperbolic rotation:

\Lambda(\varphi) = \begin{pmatrix} \cosh\varphi & \sinh\varphi \\ \sinh\varphi & \cosh\varphi \end{pmatrix}, \qquad \cosh\varphi = \gamma, \quad \sinh\varphi = \gamma\beta.

Every entry of that matrix is fixed by the demand that it preserve $\eta$ , so the pair $(\gamma, \gamma\beta)$ is constrained to lie on the unit hyperbola $\gamma^2 - (\gamma\beta)^2 = 1$ — a real constraint on real objects. At $\beta = 0.6$ : $\gamma = 1.25$ , $\gamma\beta = 0.75$ , and $1.25^2 - 0.75^2 = 1$ exactly. For the cosmic-ray muons Frisch and Smith timed down Mount Washington at $\beta = 0.995$ : $\gamma = 10.0125$ , $\gamma\beta = 9.9624$ , and $10.0125^2 - 9.9624^2 = 1.0000$ . Every particle in the universe sits somewhere on that curve, and the metric is what put it there. The difference between a circle and a hyperbola is the whole difference between Newton and Einstein. A circle closes; a hyperbola runs off to infinity, asymptotic to the light cone but never reaching it — which is exactly why $c$ is unattainable, and why no boost can turn a timelike direction into a spacelike one.

Interactive Simulation

The Inner Product in a Rotated Frame — boost the frame and watch the Minkowski inner product stay invariant while the individual components change.

Figure

PUNCHLINE — one sign, two geometries. Left: Euclid's invariant x² + y² = 1 is a CIRCLE, its isometries are rotations, and cos²θ + sin²θ = 1. Right: Minkowski's invariant (ct)² − x² = 1 is a HYPERBOLA, its isometries are boosts, and cosh²φ − sinh²φ = 1. Every physical velocity sits on that hyperbola at the point (γ, γβ). Marked: β = 0.6 ⇒ (1.25, 0.75), and 1.25² − 0.75² = 1 exactly; and the Frisch–Smith muon at β = 0.995 ⇒ (10.0125, 9.9624), with 10.0125² − 9.9624² = 1.0000. The circle closes; the hyperbola never reaches its asymptote — which is why c is unattainable

The core equations

\mathrm{d}s^2 = \eta_{\mu\nu}\,\mathrm{d}x^{\mu}\mathrm{d}x^{\nu} = c^2\mathrm{d}t^2 - \mathrm{d}x^2 - \mathrm{d}y^2 - \mathrm{d}z^2, \qquad \eta_{\mu\nu} = \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & -1 & 0 & 0 \\ 0 & 0 & -1 & 0 \\ 0 & 0 & 0 & -1 \end{pmatrix} \qquad (1)

where: $x^{\mu} = (ct, x, y, z)$ , $\mu = 0,1,2,3$ ; $\eta_{\mu\nu}$ — the Minkowski metric, a constant symmetric non-degenerate $(0,2)$ tensor of signature $(1,3)$ . Physical reading: this is the chapter's locked convention, $(+,-,-,-)$ , timelike-positive — stated here once and used everywhere: $\mathrm{d}s^2 > 0$ is timelike, $c^2\mathrm{d}\tau^2 = +\mathrm{d}s^2$ , $L_0 = \sqrt{-\Delta s^2}$ , $U\cdot U = +c^2$ . The other convention gets an exact dictionary in the warning card below. $\eta$ is its own inverse, the single most useful fact about it:

\eta^{\mu\alpha}\eta_{\alpha\nu} = \delta^{\mu}{}_{\nu}, \qquad \eta^{\mu\nu} = \mathrm{diag}(+1,-1,-1,-1) = \eta_{\mu\nu} \ \ \text{(same entries)} \qquad (2)

where: $\eta^{\mu\nu}$ — the inverse metric, used to raise indices. Physical reading: $\eta^2 = I$ , so raising and lowering are the same operation done twice, and each simply flips the sign of the three spatial components. This is not true in a curved spacetime, where the inverse metric is a genuinely different array — it is a gift of flatness, and it makes index gymnastics cheap.

A_{\mu} = \eta_{\mu\nu}A^{\nu} = (A^0, -A^1, -A^2, -A^3), \qquad A\cdot B = \eta_{\mu\nu}A^{\mu}B^{\nu} = A^{\mu}B_{\mu} = A^0B^0 - \vec{A}\cdot\vec{B} \qquad (3)

where: $A^{\mu}$ — a contravariant four-vector; $A_{\mu}$ — its covariant partner. Physical reading: this is the only way to build a Lorentz scalar from two four-vectors, and every invariant in relativity is one of these contractions. Note that $A\cdot A$ can be positive, negative or zero for a nonzero $A$ — the indefiniteness again, and the reason a photon's four-momentum is nonzero with zero norm. The defining condition of a Lorentz transformation is that it leaves $\eta$ numerically unchanged:

\eta_{\alpha\beta}\,\Lambda^{\alpha}{}_{\mu}\,\Lambda^{\beta}{}_{\nu} = \eta_{\mu\nu}, \qquad\text{i.e.}\qquad \Lambda^{\mathsf{T}}\eta\,\Lambda = \eta \qquad (4)

where: $\Lambda$ — a $4\times4$ real matrix. Physical reading: this is the group. Every boost, every rotation, and nothing else satisfies it; it is exactly equivalent to "the interval is preserved for every separation"; and this article's owned result is squeezed out of it with no further input. (The parameter count is worked out in Lorentz Invariance and Invariants.) Finally, the metric hands you the clock:

c^2\,\mathrm{d}\tau^2 = +\,\mathrm{d}s^2 \qquad\Longrightarrow\qquad \mathrm{d}\tau = \frac{\mathrm{d}s}{c} = \mathrm{d}t\sqrt{1 - v^2/c^2} \qquad (5)

where: $\tau$ — proper time. Physical reading: the plus sign is the convention doing its job. A timelike $\mathrm{d}s^2$ is positive here, so its root is real and $\mathrm{d}\tau$ is real and positive. In the other convention the same equation reads $c^2\mathrm{d}\tau^2 = -\mathrm{d}s^2$ , and an author who mixes the two produces an imaginary clock.

Deriving the structure

Derivation AThe metric is the interval, and its algebra follows in four lines

The interval is already established as the invariant of a boost; the job here is to see it as the contraction of a tensor and to extract the algebra that makes four-vector calculations possible. Every subsequent invariant in relativity — proper time, invariant mass, the phase of a wave, the field invariants — is this same contraction with different vectors plugged in, so this one card does work for the rest of the subject.

Step 1. Write the interval as a quadratic form: $\mathrm{d}s^2 = \sum_{\mu\nu}\eta_{\mu\nu}\mathrm{d}x^{\mu}\mathrm{d}x^{\nu}$ with $x^0 = ct$ and $\eta = \mathrm{diag}(+1,-1,-1,-1)$ . Only the symmetric part of a matrix contributes to a quadratic form, so $\eta$ may be taken symmetric with no loss. Why: not a definition dressed up as a derivation. The interval was fixed by physics; $\eta$ is the array that reproduces it, so its entries are data, not choices — apart from an overall sign, which is the convention.

Step 2. Invert it. $\eta$ is diagonal with entries $\pm1$ , so $\eta^2 = I$ : $\eta$ is its own inverse, the raised-index array $\eta^{\mu\nu}$ has the same entries as $\eta_{\mu\nu}$ , and $\eta^{\mu\alpha}\eta_{\alpha\nu} = \delta^{\mu}{}_{\nu}$ . Why: this is why index gymnastics is trivial in flat spacetime. Lowering an index on $A^{\mu} = (A^0, \vec{A})$ gives $A_{\mu} = (A^0, -\vec{A})$ — a sign flip on the spatial part, nothing more.

Step 3. Build the scalar product: $A\cdot B \equiv \eta_{\mu\nu}A^{\mu}B^{\nu} = A^{\mu}B_{\mu} = A^0B^0 - \vec{A}\cdot\vec{B}$ — symmetric, bilinear, and (next card) Lorentz invariant. But indefinite: $A\cdot A = (A^0)^2 - |\vec{A}|^2$ is a difference of squares, so a nonzero four-vector can have positive, negative or zero norm. Why: this is where the analogy with Euclid breaks, and it breaks productively. The null vectors — nonzero, zero-norm — are the light rays and the photon four-momenta. No positive-definite metric has any.

Step 4. Read off the clock. For a timelike displacement take the co-moving frame, $\mathrm{d}x^{\mu} = (c\,\mathrm{d}\tau, \vec{0})$ , so $\mathrm{d}s^2 = \eta_{00}(c\,\mathrm{d}\tau)^2 = +c^2\mathrm{d}\tau^2$ and $\mathrm{d}\tau = \mathrm{d}s/c$ , a real positive number. Divide the four-displacement by $\mathrm{d}\tau$ to get $U^{\mu} = \gamma(c, \vec{v})$ , and take its norm: $U\cdot U = \gamma^2(c^2 - v^2) = c^2\gamma^2(1-\beta^2) = \boldsymbol{+c^2}$ . Why: note the sign. In our convention $U\cdot U = +c^2$ ; in $(-,+,+,+)$ it is $-c^2$ . Both are right in their own house, and an article that quotes one while deriving in the other has produced a particle moving at an imaginary speed — the single most common casualty of a mixed signature. *Verified in verify_aq-sr-minkowski-metric.py: $\eta$ is symmetric with $\det\eta = -1$ and $\eta^2 = I$ ; lowering an index flips exactly the three spatial signs; $A\cdot B$ reduces to $A^0B^0 - \vec{A}\cdot\vec{B}$ ; the nonzero vector $(1,1,0,0)$ has zero norm; and $U\cdot U$ simplifies to $+c^2$ and demonstrably not $-c^2$ .*

Go deeper · Derivation B · 2 minLorentz transformations are DEFINED by the equation that preserves the metric

Derivation BLorentz transformations are DEFINED by the equation that preserves the metric

In an introductory course the boost is derived from the postulates and the metric noticed afterwards. Structurally the logic runs the other way, and it is worth inverting once and for all: a Lorentz transformation is a linear map preserving $\eta$ . That one matrix equation contains the boosts, the rotations, the invariance of every scalar product, and — as the next card shows — the topology of the group.

Step 1. Demand the interval be preserved for every separation. If $x'^{\mu} = \Lambda^{\mu}{}_{\nu}x^{\nu}$ , then $\Delta s'^2 = \Delta x^{\mathsf{T}}(\Lambda^{\mathsf{T}}\eta\Lambda)\Delta x$ ; requiring this to equal $\Delta x^{\mathsf{T}}\eta\,\Delta x$ for every $\Delta x$ means the two symmetric matrices are equal: \boxed{\ \Lambda^{\mathsf{T}}\eta\,\Lambda = \eta\ } Why: "for every $\Delta x$ " is what upgrades a scalar identity into a matrix identity — a symmetric bilinear form vanishes identically only if its matrix does. Equation (4) is therefore not stronger than interval-invariance; it is exactly equivalent to it.

Step 2. Check that the boost obeys it. With $\Lambda(\varphi)$ the hyperbolic rotation above, $\Lambda^{\mathsf{T}}\eta\Lambda$ has $00$ entry $\cosh^2\varphi - \sinh^2\varphi = 1$ , $11$ entry $\sinh^2\varphi - \cosh^2\varphi = -1$ , and off-diagonal entries $\cosh\varphi\sinh\varphi - \sinh\varphi\cosh\varphi = 0$ . That is $\eta$ . Why: the hyperbolic identity $\cosh^2 - \sinh^2 = 1$ is doing precisely the job that $\cos^2 + \sin^2 = 1$ does for a rotation preserving $\delta_{ij}$ . Same structure, one sign.

Step 3. Spatial rotations obey it too — they leave $c\,\mathrm{d}t$ alone and preserve $\mathrm{d}x^2 + \mathrm{d}y^2 + \mathrm{d}z^2$ , hence $\mathrm{d}s^2$ . So equation (4) collects boosts and rotations into one group, the Lorentz group, with no further assumption. Why: the unification is not cosmetic. It is why a boost in $x$ followed by a boost in $y$ is not a pure boost but a boost composed with a rotation, and that the two live in one group is what makes it possible.

Step 4. Harvest the invariance of every scalar product: $A'\cdot B' = (\Lambda A)^{\mathsf{T}}\eta(\Lambda B) = A^{\mathsf{T}}(\Lambda^{\mathsf{T}}\eta\Lambda)B = A^{\mathsf{T}}\eta B = A\cdot B$ . Why: one line, and every Lorentz scalar in physics is proved invariant at once — the interval, the proper time, $U\cdot U = c^2$ , the invariant mass, the phase of a plane wave. That economy is why the metric formulation replaced the coordinate one within a decade. *Verified in verify_aq-sr-minkowski-metric.py: the change in the interval under an arbitrary linear map $L$ is exactly $\Delta x^{\mathsf{T}}(L^{\mathsf{T}}\eta L - \eta)\Delta x$ ; the boost and the rotation both satisfy $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ and both preserve $\mathrm{d}s^2$ ; a matrix that fails the condition demonstrably changes the interval.*

The owned result: the Lorentz group has four pieces

Now squeeze equation (4). One matrix equation, no physics attached — and it contains a fact about the topology of spacetime's symmetry group.

Go deeper · Derivation C · 3 minFrom the metric condition alone, the Lorentz group falls into four disconnected components

Derivation CFrom the metric condition alone, the Lorentz group falls into four disconnected components

Take $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ and add nothing. Two numbers extracted from it — the determinant and the time-time entry — turn out to be confined to sets that are disconnected, and a continuous path inside the group cannot jump a gap. So the group is not one connected blob but four separate ones, with the identity in only one. This is precisely where the metric's indefiniteness bites: for a positive-definite metric the same argument yields only the determinant, and the rotation group has two components rather than four.

Step 1. Take determinants. From $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ , $\det(\Lambda^{\mathsf{T}})\det(\eta)\det(\Lambda) = \det(\eta)$ . Since $\det(\Lambda^{\mathsf{T}}) = \det(\Lambda)$ and $\det\eta = -1 \neq 0$ , we may divide: (\det\Lambda)^2 = 1 \qquad\Longrightarrow\qquad \det\Lambda = +1 \ \ \text{or} \ \ -1. Why: no intermediate value is permitted. $\det$ is a continuous function on the group, and it takes values in a two-point set — so it is constant along any continuous path. Already the group cannot be connected.

Step 2. Now read the $00$ component of equation (4). Written out, $(\Lambda^{\mathsf{T}}\eta\Lambda)_{00} = \sum_{\alpha\beta}\Lambda^{\alpha}{}_{0}\,\eta_{\alpha\beta}\,\Lambda^{\beta}{}_{0} = (\Lambda^{0}{}_{0})^2 - \sum_{i=1}^{3}(\Lambda^{i}{}_{0})^2$ , and this must equal $\eta_{00} = +1$ . Hence (\Lambda^{0}{}_{0})^2 = 1 + \sum_{i=1}^{3}(\Lambda^{i}{}_{0})^2 \ \geq\ 1. Why: here is the indefiniteness, in one equation. The spatial terms enter with a minus sign, so moving them to the other side makes $(\Lambda^{0}{}_{0})^2$ at least one — never less. For a positive-definite metric the same manipulation would bound the entry from above, and nothing discrete would follow.

Step 3. Draw the consequence. $|\Lambda^{0}{}_{0}| \geq 1$ means $\Lambda^{0}{}_{0} \geq +1$ or $\Lambda^{0}{}_{0} \leq -1$ : the open band $-1 < \Lambda^{0}{}_{0} < 1$ is empty. So $\mathrm{sign}(\Lambda^{0}{}_{0})$ is a second discrete invariant, and it too cannot change along a continuous path — to flip it a matrix would have to cross the forbidden band. A $\Lambda$ with $\Lambda^{0}{}_{0} \geq 1$ maps future-pointing timelike vectors to future-pointing ones and is orthochronous; one with $\Lambda^{0}{}_{0} \leq -1$ reverses time. Why: the arrow of time is not smuggled in by hand. It is a topological feature of the symmetry group, forced by the sign pattern of the metric.

Step 4. Count. Two independent signs, $\det\Lambda = \pm1$ and $\mathrm{sign}(\Lambda^{0}{}_{0}) = \pm1$ , give four classes, and all four are realised by explicit matrices: \begin{array}{lccl} I & \det = +1 & \Lambda^0{}_0 = +1 & \textbf{proper orthochronous } L^{\uparrow}_{+} \\ P = \mathrm{diag}(1,-1,-1,-1) & \det = -1 & \Lambda^0{}_0 = +1 & L^{\uparrow}_{-} \\ T = \mathrm{diag}(-1,1,1,1) & \det = -1 & \Lambda^0{}_0 = -1 & L^{\downarrow}_{-} \\ PT = -I & \det = +1 & \Lambda^0{}_0 = -1 & L^{\downarrow}_{+} \end{array} Each satisfies $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ , and each sits in a different box. Why: parity and time reversal are not extra symmetries bolted onto relativity. They are the other three components of the Lorentz group, and the metric produced them from nothing but its signature.

Step 5. Say what is proved and what is standard. The two invariants prove the group has at least four pieces, no two joinable by a continuous path. That each piece is itself connected — so there are exactly four — is the standard completion (every proper orthochronous $\Lambda$ factors as a boost times a rotation, and both families reach the identity continuously). Multiplying any element of $L^{\uparrow}_{+}$ by $P$ , $T$ or $PT$ lands in exactly one of the other three, so the four components are the four cosets of $L^{\uparrow}_{+}$ . Physics lives in $L^{\uparrow}_{+}$ . Why: this is why "Lorentz invariance" in practice means invariance under the identity component, and why $P$ and $T$ are separate, discrete questions — questions to which nature answers no. Parity is violated by the weak interaction, and $CP$ (hence $T$ ) with it. Neither violation was even well-posed until the metric's signature had split the group into pieces you could interrogate one at a time. *Verified in verify_aq-sr-minkowski-metric.py: $\det(L^{\mathsf{T}}\eta L) = (\det L)^2\det\eta$ symbolically, so $(\det\Lambda)^2 = 1$ ; the $00$ component of $\Lambda^{\mathsf{T}}\eta\Lambda - \eta$ simplifies exactly to $(\Lambda^0{}_0)^2 - \sum_i(\Lambda^i{}_0)^2 - 1$ ; $(\Lambda^0{}_0)^2 \geq 1$ over 5000 random spatial columns; and all four of $I$ , $P$ , $T$ , $PT$ satisfy the metric condition, realise all four sign combinations, and keep their component when multiplied by any boost.*

Figure

The four connected components of the Lorentz group, forced by Λᵀ η Λ = η alone. The determinant is ±1 (never between); the time-time entry satisfies (Λ⁰₀)² = 1 + Σᵢ(Λⁱ₀)² ≥ 1, so the grey band −1 ＜ Λ⁰₀ ＜ 1 is EMPTY. Two discrete signs, four boxes, and no continuous path from one box to another. Physics lives in the proper orthochronous box, the one containing the identity, every boost and every rotation

Limiting cases

Case	Result	Meaning
$\eta \to \delta$ (all $+$ )	Euclidean geometry	positive definite: no null vectors, level sets are spheres, isometries are rotations, and the group has only two components
$A\cdot A = 0$ , $A \neq 0$	null vector	the signature's fingerprint — impossible for any positive-definite metric; these are the photons

Common trapThe signature dictionary — read this before you trust a sign from any other book

About half the literature uses the opposite convention. Taylor–Wheeler and MTW use $(-,+,+,+)$ ; Rindler and most particle-physics texts, and this chapter, use $(+,-,-,-)$ . No physical result depends on the choice — but every intermediate sign does, and mixing two sources mid-derivation is the surest way to manufacture an imaginary clock. The exact dictionary:

quantity	this chapter, $(+,-,-,-)$	the other, $(-,+,+,+)$
metric	$\eta = \mathrm{diag}(+1,-1,-1,-1)$	$\eta = \mathrm{diag}(-1,+1,+1,+1)$
interval	$\Delta s^2 = c^2\Delta t^2 - \Delta r^2$	$\Delta s^2 = -c^2\Delta t^2 + \Delta r^2$
timelike	$\Delta s^2 > 0$	$\Delta s^2 < 0$
spacelike	$\Delta s^2 < 0$	$\Delta s^2 > 0$
proper time	$c\,\Delta\tau = +\sqrt{\Delta s^2}$	$c\,\Delta\tau = \sqrt{-\Delta s^2}$
proper length	$L_0 = \sqrt{-\Delta s^2}$	$L_0 = +\sqrt{\Delta s^2}$
scalar product	$A\cdot B = A^0B^0 - \vec{A}\cdot\vec{B}$	$A\cdot B = -A^0B^0 + \vec{A}\cdot\vec{B}$
four-velocity	$U\cdot U = +c^2$	$U\cdot U = -c^2$

The translation rule is a single global sign: $\Delta s^2 \to -\Delta s^2$ , and every scalar product with it. What does not change: which pairs of events are timelike, the numerical value of any proper time or length, the transformation matrices, the condition $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ (which holds with either $\eta$ ), and the four-component structure above. The physics is convention-blind. Your algebra is not.

Worked examples

1 · Where the Frisch–Smith muon sits on Minkowski's unit hyperbola

The metric condition $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ forces the boost's first column to obey $(\Lambda^{0}{}_{0})^2 - (\Lambda^{1}{}_{0})^2 = 1$ . For a pure boost that column is $(\gamma, \gamma\beta)$ . Check it against a real measured particle: the cosmic-ray muons that Frisch and Smith timed at $\beta = 0.995$ .

Step 1 — identify. For a boost along $x$ , $\Lambda^{0}{}_{0} = \gamma$ and $\Lambda^{1}{}_{0} = \gamma\beta$ . The metric constraint of Step 2 above therefore reads $\gamma^2 - (\gamma\beta)^2 = 1$ — which is $\gamma^2(1-\beta^2) = 1$ , the definition of $\gamma$ . The Lorentz factor is not an extra postulate; it is the $00$ component of the metric condition.

Step 2 — apply. At $\beta = 0.995$ , $\gamma = 1/\sqrt{1 - 0.995^2} = 1/\sqrt{0.009975} = \mathbf{10.0125}$ and $\gamma\beta = \mathbf{9.9624}$ . The muon's boost sits at the point $(10.0125,\ 9.9624)$ in the $(\Lambda^{0}{}_{0}, \Lambda^{1}{}_{0})$ plane.

Step 3 — the boundary number. Is it on the hyperbola? (10.0125)^2 - (9.9624)^2 = 100.2502 - 99.2502 = \mathbf{1.0000}. Exactly one, to every digit the numbers carry. Every particle ever accelerated, in every laboratory, sits on this one curve — and the curve is nothing but the $00$ entry of $\eta$ .

Step 4 — run it backwards. Ask for $\Lambda^{0}{}_{0} = 10$ exactly. The constraint gives $\Lambda^{1}{}_{0} = \sqrt{100 - 1} = 9.9499$ , so $\beta = \mathbf{0.99499}$ : a tenfold time dilation costs you $0.995c$ — exactly the speed the muons were travelling, and exactly why the Mount Washington experiment worked. Note the sanity check the metric hands you free: $|\Lambda^{0}{}_{0}| \geq 1$ always, so $\gamma < 1$ is not merely unphysical — it is forbidden by the matrix equation. No such Lorentz transformation exists.

2 · Translating one calculation between the two conventions

A particle moves at $\beta = 0.6$ . Compute $U\cdot U$ and the interval to an event $(c\Delta t, \Delta x) = (5, 3)$ light-seconds, in both signatures, and check that the physics agrees.

Step 1 — our convention, $(+,-,-,-)$ . With $\gamma = 1.25$ and $\gamma\beta = 0.75$ , the four-velocity is $U^{\mu} = (1.25c,\ 0.75c,\ 0,\ 0)$ . Then U\cdot U = \eta_{\mu\nu}U^{\mu}U^{\nu} = (1.25c)^2 - (0.75c)^2 = 1.5625c^2 - 0.5625c^2 = \mathbf{+c^2}. And the interval: $\Delta s^2 = 5^2 - 3^2 = \mathbf{+16\ \mathrm{ls}^2}$ — positive, so timelike, and the proper time is $c\Delta\tau = +\sqrt{16} = 4$ , i.e. $\Delta\tau = 4$ seconds.

Step 2 — the other convention, $(-,+,+,+)$ . Same particle, same events, same components. Now U\cdot U = -(1.25c)^2 + (0.75c)^2 = \mathbf{-c^2}, \qquad \Delta s^2 = -5^2 + 3^2 = \mathbf{-16\ \mathrm{ls}^2}. Every sign has flipped, and "timelike" now means $\Delta s^2 < 0$ . The proper time is $c\Delta\tau = \sqrt{-\Delta s^2} = 4$ .

Step 3 — the boundary number. Both conventions give $\Delta\tau = \mathbf{4\ seconds}$ — not approximately, identically. The clock does not know which book you are reading. What changed is which square root you take: $+\sqrt{\Delta s^2}$ in ours, $\sqrt{-\Delta s^2}$ in theirs. Get that pairing wrong and you end up with $\Delta\tau = 4i$ seconds.

Step 4 — what is convention-free. The same boost matrix satisfies $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ with either $\eta$ ; its determinant is $+1$ in both; $\Lambda^{0}{}_{0} = 1.25 \geq 1$ in both. The group, the four components, the physics: identical. Only the bookkeeping moved.

Physics of the portrait

Origins · 2 minMinkowski metric and how this was discovered

Minkowski metric

1908–present · Cologne · Göttingen · every textbook since

The object is $\eta_{\mu\nu}$ : an invariant quadratic form of indefinite signature, whose isometries are exactly the Lorentz transformations. It arrived, essentially whole, in a single lecture. Hermann Minkowski (1864–1909) was a mathematician of the first rank long before he touched relativity — he founded the geometry of numbers — and he taught at the Eidgenössische Polytechnikum in Zürich from 1896 to 1902, where one of his students was Albert Einstein. From 1907, at Göttingen, he set about recasting Einstein's kinematics and Lorentz's electrodynamics in four-dimensional form, and on 21 September 1908 he delivered Raum und Zeit to the 80th Assembly of German Natural Scientists and Physicians in Cologne. Its opening sentence is the most quoted in the subject: "Henceforth space by itself, and time by itself, are doomed to fade away into mere shadows, and only a kind of union of the two will preserve an independent reality." What he supplied to back the rhetoric was this metric. Einstein, who had thought the reformulation superfluous mathematics, was using it to build general relativity within four years. Minkowski saw none of that: he died of appendicitis in Göttingen on 12 January 1909, less than four months after the Cologne lecture, aged 44. The lecture was published posthumously, and the geometry has carried his name ever since.

History

1905

— Einstein's Zur Elektrodynamik bewegter Körper derives the Lorentz transformation from two postulates, without identifying the quadratic form it preserves.

1908

— Minkowski delivers Raum und Zeit in Cologne on 21 September, introducing the four-dimensional continuum, the spacetime diagram and the invariant interval, and recasting relativity as geometry.

1909

— Minkowski dies of appendicitis in Göttingen on 12 January, aged 44; Raum und Zeit is published posthumously.

1911

— Max von Laue publishes Das Relativitätsprinzip (Vieweg, Braunschweig), an early monograph written in Minkowski's four-dimensional language, which helps make the geometric formalism standard.

1916

— Einstein's general relativity replaces the constant η with a dynamical field g, of which η is the flat-spacetime case.

1956

— Lee and Yang propose that parity, one of the discrete Lorentz components, may be violated by the weak interaction; Wu and collaborators confirm it in 1957.

Going further

Texts. Rindler, Introduction to Special Relativity, 2nd ed., §7–8 · Misner, Thorne & Wheeler, Gravitation, ch. 1–2 and Taylor & Wheeler, Spacetime Physics, 2nd ed., ch. 1–2 (both in $(-,+,+,+)$ — read the dictionary above first) · Weinberg, Gravitation and Cosmology, ch. 2. Primary source: Minkowski, Raum und Zeit (Cologne, 21 September 1908), translated in The Principle of Relativity (Dover).

Provenance. The base article was already in the chapter's locked signature $(+,-,-,-)$ , and every step has nonetheless been re-derived rather than transcribed; no numeric or sign error was found in it. The metric's role as this chapter's signature anchor is discharged explicitly: the convention is stated in equation (1), the exact $(-,+,+,+)$ dictionary is given in the warning card, and the convention-independence of the physics is demonstrated in worked example 2. Recomputed in verify_aq-sr-minkowski-metric.py: $\eta^2 = I$ , $\det\eta = -1$ ; $U\cdot U = +c^2$ (demonstrably not $-c^2$ ); $\Lambda^{\mathsf{T}}\eta\Lambda = \eta$ for a boost and a rotation; $(\det\Lambda)^2 = 1$ ; the $00$ component reduces exactly to $(\Lambda^0{}_0)^2 - \sum_i(\Lambda^i{}_0)^2 = 1$ ; all four of $I, P, T, PT$ satisfy the metric condition and realise all four sign combinations; the muon anchors $\gamma = 10.0125$ , $\gamma\beta = 9.9624$ , $\gamma^2 - (\gamma\beta)^2 = 1.0000$ ; and $\beta = 0.6 \Rightarrow \gamma = 1.25$ , $\gamma\beta = 0.75$ , $\varphi = \ln 2$ .

Builds on. The Spacetime Interval · Events and Worldlines · Lorentz Invariance and Invariants. Leads to. Four-Vectors · Timelike, Spacelike and Null Separations · The Light Cone · Hyperbolic Geometry of Boosts.

Does any physical result depend on the choice of metric signature?

None whatever, and it is worth being emphatic, because the question arises the moment a student opens a second textbook. Proper times, proper lengths, invariant masses, cross-sections, decay rates — every measurable number is identical in $(+,-,-,-)$ and $(-,+,+,+)$ . What differs is the sign of $\Delta s^2$ , the sign of every scalar product, and which square root extracts a physical quantity. The translation is one global sign, $\Delta s^2 \to -\Delta s^2$ . The danger is never in using one convention; it is in silently using both, and the classic symptom is an imaginary proper time. State your signature at the top of the page and do not move.

Why does the metric have to be indefinite? Could spacetime not just be four-dimensional Euclidean space?

Because light. A positive-definite metric has no null vectors: $A\cdot A = 0$ forces $A = 0$ . But a light ray connects two distinct events at zero interval — that is what the constancy of $c$ means geometrically — so the metric must admit nonzero vectors of zero norm, and the only way is a mixture of signs. Everything else follows: level sets become hyperbolae rather than spheres, isometries become boosts rather than rotations, the light cone appears as the zero-set, and the symmetry group falls into four pieces. Four-dimensional Euclidean space is a perfectly good geometry; it is not the one we live in.

What exactly makes the four components disconnected, rather than just four labelled subsets?

Two subsets are disconnected when no continuous path inside the group joins them. Both $\det\Lambda$ and $\Lambda^0{}_0$ are continuous functions of the matrix entries. If a path led from $\det\Lambda = +1$ to $-1$ , the determinant would have to take intermediate values on the way — but $(\det\Lambda)^2 = 1$ forbids every value except $\pm1$ , so no such path exists. Likewise for $\Lambda^0{}_0$ : going from $\geq +1$ to $\leq -1$ means crossing the open band $(-1,1)$ , and $(\Lambda^0{}_0)^2 = 1 + \sum_i(\Lambda^i{}_0)^2 \geq 1$ says that band contains no Lorentz matrix at all. Two forbidden crossings, four boxes, nothing continuous joining them.

If parity and time reversal are Lorentz transformations, why does "Lorentz invariance" not include them?

Because in practice "Lorentz invariance" is shorthand for invariance under the identity component $L^{\uparrow}_{+}$ — the boosts and rotations, continuously connected to doing nothing at all. That component is what the relativity principle is about: no experiment can detect uniform motion, because uniform motion is a continuous deformation of rest. $P$ and $T$ are different animals — discrete operations, not deformations of the identity — and whether nature respects them is a separate empirical question, with a separate answer. That answer, famously, is no: parity is maximally violated by the weak interaction, and $CP$ (hence, by the $CPT$ theorem, $T$ ) with it. Special relativity survives untouched, because none of it touches $L^{\uparrow}_{+}$ .

Why can I be sloppy about index position in Euclidean space but not here?

Because in Euclidean space the metric is $\delta_{ij}$ , so lowering an index changes nothing: $A^1 = A_1$ . Here $\eta = \mathrm{diag}(+1,-1,-1,-1)$ , so lowering flips the sign of the spatial components: $A_1 = -A^1$ . The consequences are immediate. $A^{\mu}B_{\mu} = A^0B^0 - \vec{A}\cdot\vec{B}$ is a Lorentz scalar; $\sum_{\mu}A^{\mu}B^{\mu}$ , both indices up, is $A^0B^0 + \vec{A}\cdot\vec{B}$ and is not invariant — a different number in every frame, and one of the commonest errors in a first course. The summation convention is designed to make the mistake visible: a repeated index appears once up and once down, and if it does not, what you have written is not a scalar.

Is the Minkowski metric a tensor, and what does that even mean if its components never change?

It is a tensor, and its components never changing is a special feature, not an exception. A $(0,2)$ tensor transforms as $\eta'_{\mu\nu} = \Lambda^{\alpha}{}_{\mu}\Lambda^{\beta}{}_{\nu}\,\eta_{\alpha\beta}$ — and equation (4) says exactly that this equals $\eta_{\mu\nu}$ . So $\eta$ is transforming like a tensor; it simply happens to be invariant under the transformations we are allowed, just as $\delta_{ij}$ is invariant under rotations. That numerical invariance is not a property of tensors in general; it is the defining property of the Lorentz group, turned around. In general relativity the allowed transformations are no longer Lorentz ones, the components do change, and $g_{\mu\nu}(x)$ becomes a dynamical field. The tensor was always a tensor. Only the group shrank.