The Spacetime Interval

Two observers watch the same pair of events. One says they were $3$ seconds and $5$ light-seconds apart; the other, sailing past at $0.6c$ , says $6.25$ seconds and $3.75$ light-seconds. Neither is mistaken. Time separations and space separations are not properties of events at all — they are properties of the slicing, and each observer slices spacetime his own way. So is there anything left — something the two events possess in themselves, surviving the change of slicing? There is, and it is the only such thing. Form the combination \Delta s^2 = c^2\Delta t^2 - \Delta x^2 - \Delta y^2 - \Delta z^2, and both observers get the identical number, exactly, always. Not an approximation, not a small-velocity result, not a happy accident of the algebra: it is what a Lorentz transformation is. The spacetime interval stands in the same relation to Minkowski geometry as Pythagoras' theorem does to Euclid's — with one sign flipped, and that sign changes everything. Everything changes because a minus sign makes the geometry indefinite. Euclidean distance is a sum of squares, always positive, and the long way round always takes you further. Minkowski's interval is a difference of squares, and the long way round always takes you less proper time. The straight worldline between two events is not the shortest path. It is the longest — and that single inversion is the twin paradox, the whole of it, before any twins have been introduced.

Key takeaways

The spacetime interval between two events, $\Delta s^2 = c^2\Delta t^2 - \Delta x^2 - \Delta y^2 - \Delta z^2$ , is the same number in every inertial frame — even though $\Delta t$ and $\Delta x$ are separately frame-dependent. It is the metric-induced norm of the separation four-vector, $\Delta s^2 = \eta_{\mu\nu}\Delta x^{\mu}\Delta x^{\nu}$ .
Invariance is not assumed; it is derived. The light postulate gives $\Delta s^2 = 0 \Rightarrow \Delta s'^2 = 0$ ; homogeneity forces linearity, so $\Delta s'^2 = k(|v|)\,\Delta s^2$ ; applying the boost and its inverse gives $k^2 = 1$ ; continuity at $v = 0$ selects $k = +1$ .
What the interval means depends on its sign. Timelike ( $\Delta s^2 > 0$ ): $\Delta s^2 = +c^2\Delta\tau^2$ , so $\sqrt{\Delta s^2}/c$ is the proper time a clock present at both events reads. Spacelike ( $\Delta s^2 < 0$ ): $\Delta s^2 = -L_0^2$ , so $\sqrt{-\Delta s^2}$ is the proper length. The minus sign is not optional.
Level sets of the interval are invariant hyperbolae $c^2t^2 - x^2 = \text{const}$ . Under a boost, an event slides along its hyperbola: $\Delta t$ and $\Delta x$ both change, and $\Delta s^2$ does not.
Key idea: the interval obeys a REVERSED triangle inequality. For future-directed timelike vectors, $\tau(A + B) \geq \tau(A) + \tau(B)$ — the exact opposite of Euclid — so of all worldlines joining two timelike-separated events, the straight one has the LONGEST proper time, and any detour costs you time you never get back. The twin paradox is not a paradox. It is this inequality.

The core insight

Boosting is a hyperbolic rotation: it mixes $c\Delta t$ and $\Delta x$ into each other the way an ordinary rotation mixes $x$ and $y$ . A rotation preserves $x^2 + y^2$ ; a boost preserves $c^2\Delta t^2 - \Delta x^2$ . That is the entire content of relativistic kinematics — time dilation, length contraction, the relativity of simultaneity are all just the observation that the components of a fixed vector change when you turn the axes. What does not change is its length, and in spacetime "length" means the interval.

Where this lives today

The interval is not a stepping stone to relativistic physics — it is relativistic physics, in the sense that every calculation in the subject is ultimately a statement about it. In particle physics the invariant mass of a system, $M^2c^4 = (\sum E)^2 - (\sum \vec{p}c)^2$ , is the interval of the total four-momentum: the same quadratic form, in momentum space instead of position space. Every resonance ever discovered — the $Z$ , the Higgs — was found as a bump in that invariant, precisely because it is the one number all observers and all detectors agree on. In general relativity the interval becomes the whole theory. Coordinates lose their meaning, the metric $g_{\mu\nu}(x)$ becomes a dynamical field, and Einstein's equations are equations for $\mathrm{d}s^2$ : the bending of light, the perihelion of Mercury, the ringdown of a merging black hole are all statements about arc length in a curved spacetime, and the flat interval of this article is what they reduce to locally. And it is in your pocket. An uncorrected GPS satellite clock would accumulate a $38.6\ \mu\mathrm{s}$ error per day, and $c$ times that is $11.6\ \mathrm{km}$ of position error, every day, forever. Relativity is not an exotic correction to a system that would otherwise work slightly less well. Without it the system does not work at all.

The physical picture

Draw a circle of radius $5$ and rotate the axes: every point moves, its $x$ and $y$ both change, and yet it stays on the circle, because the rotation preserves $x^2+y^2 = 25$ . That is what invariance looks like — not "nothing moves" but "everything moves along a level set". Minkowski's version replaces the circle with a hyperbola. The level set $c^2t^2 - x^2 = 25\ \mathrm{ls}^2$ is a curve running up the diagram, asymptotic to the $45^\circ$ light lines. Under a boost every event on it slides along it — the time coordinate grows, the space coordinate grows, and the invariant stays exactly $25$ . Take the event $(ct, x) = (5, 0)$ light-seconds, the second tick of a clock at rest at the origin. Boost to $\beta = 0.6$ and the same event acquires coordinates $(6.25, 3.75)$ : a bigger time and a nonzero position. Both numbers changed. $6.25^2 - 3.75^2 = 39.0625 - 14.0625 = 25$ . Nothing changed. That is time dilation, watched as a geometric fact rather than a conspiracy of rods and clocks: the moving observer assigns $6.25$ seconds to a tick the rest observer calls $5$ , because he is reading a different component of the same fixed spacetime vector. The vector never moved. The hyperbolae come in two families, and the split is the light cone. Outside it, $c^2t^2 - x^2 = -16\ \mathrm{ls}^2$ is a spacelike hyperbola opening left and right. The event $(0, 4)$ — one end of a $4$ -light-second rod, at time zero — boosts to $(3, 5)$ : it acquires a time coordinate, and its spatial coordinate stretches from $4$ to $5$ . And $3^2 - 5^2 = -16$ . That is the relativity of simultaneity and length contraction, in the same picture, doing the same thing.

Figure

PUNCHLINE — the invariant hyperbolae. A boost slides events ALONG the level sets of the interval: both coordinates change, the interval does not. Timelike hyperbola c²t² − x² = +25 ls²: the event (5, 0) — a clock's fifth tick — boosts at β = 0.6 (γ = 1.25) to (6.25, 3.75), and 6.25² − 3.75² = 25 exactly. Spacelike hyperbola c²t² − x² = −16 ls²: the event (0, 4) — the far end of a 4-ls rod — boosts to (3, 5), and 3² − 5² = −16 exactly. Both events moved. Neither interval did

The core equations

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

where: $\Delta x^{\mu} = (c\Delta t, \Delta x, \Delta y, \Delta z)$ — the separation four-vector; $\eta_{\mu\nu}$ — the Minkowski metric (developed in The Minkowski Metric). Physical reading: the interval is the norm induced by the metric — not a distance imposed on spacetime from outside, but the length the geometry itself assigns. Timelike is positive in this chapter. The opposite convention $(-,+,+,+)$ — Taylor–Wheeler, MTW — negates $\Delta s^2$ throughout, and the base article for this entry was written in it; every sign below has been rebuilt, not relabelled. In differential form, the line element:

\mathrm{d}s^2 = c^2\,\mathrm{d}t^2 - \mathrm{d}x^2 - \mathrm{d}y^2 - \mathrm{d}z^2 \qquad (2)

where: $\mathrm{d}s^2$ — the interval between neighbouring events. Physical reading: this is what makes the interval a metric rather than merely an invariant. Integrating $\mathrm{d}s$ along a curve gives its arc length, and for a timelike curve that arc length is a clock reading. The sign of $\Delta s^2$ decides which physical quantity it names:

\text{timelike } (\Delta s^2 > 0): \ \ \Delta s^2 = +c^2\Delta\tau^2; \qquad \text{spacelike } (\Delta s^2 < 0): \ \ \Delta s^2 = -L_0^2 \qquad (3)

where: $\Delta\tau$ — proper time; $L_0$ — proper length. Physical reading: $\Delta\tau = \sqrt{\Delta s^2}/c$ and $L_0 = \sqrt{-\Delta s^2}$ . Each root is taken of whichever sign the class makes positive, and swapping them is the classic way to manufacture an imaginary clock. (The S1 article The Invariant Spacetime Interval establishes the invariance kinematically; this one recasts it as the norm from which all the geometry follows.) For a worldline with arbitrary motion, the proper time is the arc length:

\Delta\tau = \frac{1}{c}\int \mathrm{d}s = \int\sqrt{1 - \frac{v(t)^2}{c^2}}\ \mathrm{d}t \qquad (4)

where: $v(t)$ — the instantaneous speed. Physical reading: every clock, however it moves, reads the arc length of its own worldline. This is the functional the owned result extremises — and the extremum is a maximum. Finally, the level sets:

c^2t^2 - x^2 = \text{const} \qquad (5)

where: the invariant hyperbolae. Physical reading: boosts move events along these curves and never off them, which is why they calibrate the tilted axes of a spacetime diagram. The light cone is their common asymptote.

Where the invariance comes from

Derivation ADeriving the invariance of the interval, and pinning k to exactly plus one

Most treatments prove the interval invariant by writing out a boost and grinding. That is a check, not a derivation: it presupposes the Lorentz transformation, which was itself derived from something. The honest route runs the other way — from the light postulate alone, plus homogeneity and isotropy, to the invariance of $\Delta s^2$ for every pair of events, not just null ones. The step everybody waves at is the last, where a constant $k$ must be shown equal to $+1$ and not $-1$ . This card does it properly.

Step 1. Start with light. A flash at the origin at $t = 0$ reaches the sphere $x^2+y^2+z^2 = c^2t^2$ , so the events on it have $\Delta s^2 = 0$ . The second postulate says the primed observer sees the same flash expand at the same speed $c$ , so for him too $\Delta s'^2 = 0$ . Therefore: $\Delta s^2 = 0 \Rightarrow \Delta s'^2 = 0$ . Why: this is all the light postulate gives you. It constrains the interval only on the light cone — a measure-zero set. Everything else must come from symmetry.

Step 2. Homogeneity of spacetime — no event is special — forces the coordinate transformation to be linear. (A nonlinear map would make the transformation law depend on where you are.) Hence $\Delta s'^2$ , being a quadratic expression in the primed coordinates, is a quadratic form in the unprimed separations. Why: linearity is what turns "the transformation preserves a cone" into "one quadratic form vanishes wherever another does", which is an algebraic statement with teeth.

Step 3. Two non-degenerate quadratic forms in four variables that vanish on the same cone must be proportional. So $\Delta s'^2 = k\,\Delta s^2$ for some constant $k$ , independent of the events. Isotropy — no direction is special — means $k$ can depend only on the magnitude of the relative velocity: $k = k(|v|)$ . Why: at this point we already know the interval is invariant up to an overall factor. All that is left is to kill the factor.

Step 4. Kill it with reciprocity. Boost from $S$ to $S'$ with velocity $v$ , then from $S'$ back to $S$ with velocity $-v$ . The composition is the identity, so $\Delta s^2 = k(|-v|)\,k(|v|)\,\Delta s^2$ , hence $k(|v|)^2 = 1$ and $k = \pm 1$ . Why: this is the only place the group structure is used, and it is enough to reduce a whole function of $v$ to a choice of sign.

Step 5. Pick the sign, and pick it for a reason. $k(|v|)$ takes values in the two-point set $\{+1,-1\}$ ; it is a continuous function of $v$ (the transformation depends continuously on the boost); and at $v = 0$ the transformation is the identity, so $k(0) = +1$ . A continuous function into a discrete set is constant on a connected domain. Therefore $k \equiv +1$ for every $v$ , and $\Delta s'^2 = \Delta s^2$ exactly, for every pair of events, timelike, null or spacelike. Why: this is the step usually skipped. Without continuity you cannot exclude $k = -1$ , which would swap timelike and spacelike between frames and destroy causality — so the harmlessness of $k$ is not a triviality, it is a theorem, and it rests on the boosts forming a connected group containing the identity. *Verified in verify_aq-sr-spacetime-interval.py: null separations map to null separations; $\Delta s'^2 - \Delta s^2$ simplifies to zero identically; the boost composed with its $-v$ inverse is the identity, so $k(v)k(-v) = 1$ ; $k^2 = 1$ has roots $\pm 1$ ; the boost at $v = 0$ is the identity, giving $k(0) = +1$ ; and $\Delta s'^2 = +1 \cdot \Delta s^2$ (and demonstrably not $-1 \cdot \Delta s^2$ ).*

Go deeper · Derivation B · 2 minWhat the interval measures, and why one root has a minus sign in front of it

Derivation BWhat the interval measures, and why one root has a minus sign in front of it

An invariant is a number; a metric is a number that measures something. This card converts $\Delta s^2$ into two physical readings and pins down the sign that the base article for this entry got wrong. The rule that emerges is mechanical, and worth memorising: go to the frame in which the separation is as simple as it can be, and read off what is left.

Step 1. Take a timelike pair, $\Delta s^2 > 0$ . Boost to the frame in which they are co-located, $\Delta x' = 0$ (that frame exists precisely because the pair is timelike). Then equation (1) collapses to $\Delta s^2 = c^2\Delta t'^2$ , and $\Delta t'$ is what a clock present at both events reads — the proper time. So \Delta\tau = \frac{\sqrt{\Delta s^2}}{c}, a positive square root of a positive number. Why: the interval of a timelike pair is not analogous to a time — in one frame it literally is the time, multiplied by $c$ . Every other frame's $\Delta t$ is larger, which is time dilation.

Step 2. Take a spacelike pair, $\Delta s^2 < 0$ . Boost to the frame in which they are simultaneous, $\Delta t' = 0$ (which exists precisely because the pair is spacelike). Then $\Delta s^2 = -\Delta x'^2$ , and $|\Delta x'|$ is the distance a ruler reads between two simultaneous events — the proper length. So L_0 = \sqrt{-\Delta s^2}. Why: the minus sign is not optional. $\Delta s^2$ is negative here, so $\sqrt{\Delta s^2}$ is imaginary. The base article for this entry defined $\Delta\sigma = \sqrt{\Delta s^2}$ for spacelike pairs — correct in its own $(-,+,+,+)$ convention and catastrophic in ours. The general rule: take the real root of whichever sign the class makes positive.

Step 3. Now generalise to a curve. Equation (2) gives $\mathrm{d}s^2 = c^2\mathrm{d}t^2 - \mathrm{d}x^2 = c^2\mathrm{d}t^2(1 - v^2/c^2)$ along a worldline moving at $v = \mathrm{d}x/\mathrm{d}t$ . Taking the root and integrating gives the arc length, $\Delta\tau = \int\sqrt{1 - v^2/c^2}\,\mathrm{d}t$ — equation (4). Why: this is what promotes the interval from an invariant to a metric. A clock is a device that measures the arc length of its own worldline, and nothing else. It has no access to $t$ ; it never did.

Step 4. Read off the level sets. Fixing $\Delta s^2 = c^2\Delta\tau^2$ traces a hyperbola $c^2t^2 - x^2 = \text{const} > 0$ , opening in the time direction: the locus of all events one proper second from the origin, according to every observer. Fixing $\Delta s^2 = -L_0^2 < 0$ traces the spacelike hyperbolae opening sideways. Their common asymptote, $\text{const} = 0$ , is the light cone. Why: these curves are the calibration marks of a spacetime diagram. The unit tick on a tilted $ct'$ axis is where that axis crosses the hyperbola $c^2t^2 - x^2 = 1$ — not where the tilted axis crosses a Euclidean circle, which is the single commonest way to draw a Minkowski diagram wrong. *Verified in verify_aq-sr-spacetime-interval.py: substituting the co-location boost gives $\Delta t'^2 = \Delta s^2/c^2$ ; substituting the simultaneity boost gives $\Delta x'^2 = -\Delta s^2$ ; and the two figure events $(5,0) \to (6.25, 3.75)$ and $(0,4) \to (3,5)$ preserve $\Delta s^2 = +25$ and $-16$ exactly.*

The owned result: the straight worldline is the longest

Euclid's triangle inequality says a detour is never shorter than the direct route. Minkowski's says the opposite, and it is not a curiosity — it is the reason a travelling twin comes home young.

Go deeper · Derivation C · 2 minThe reversed triangle inequality, and why the inertial worldline maximises proper time

Derivation CThe reversed triangle inequality, and why the inertial worldline maximises proper time

In Euclidean space, $|A + B| \leq |A| + |B|$ : the direct path is the shortest. In Minkowski space, for future-directed timelike vectors, the inequality reverses: $\tau(A+B) \geq \tau(A) + \tau(B)$ . The direct path is the longest. The whole of the twin paradox is contained in that reversal, and it is worth proving properly rather than asserting, because the sign flip comes from exactly one place — the indefiniteness of the metric — and once you have seen where, the paradox stops being paradoxical for good.

Step 1. First a lemma: the reversed Cauchy–Schwarz inequality. For future-directed timelike $A$ and $B$ , claim $A \cdot B \geq \sqrt{A\cdot A}\,\sqrt{B\cdot B}$ — note the direction of the inequality, which is backwards from the Euclidean $|A\cdot B| \leq |A||B|$ . To prove it, go to the rest frame of $A$ , where $A = (a, \vec{0})$ with $a = \sqrt{A\cdot A} > 0$ . Why: an invariant inequality may be proved in whichever frame makes it easiest, and the rest frame of $A$ costs nothing.

Step 2. In that frame write $B = (b^0, \vec{b})$ . Timelike and future-directed means $b^0 > 0$ and $(b^0)^2 - |\vec{b}|^2 = B\cdot B > 0$ , so $b^0 = \sqrt{B\cdot B + |\vec{b}|^2} \geq \sqrt{B\cdot B}$ . Then $A\cdot B = a\,b^0 \geq a\sqrt{B\cdot B} = \sqrt{A\cdot A}\sqrt{B\cdot B}$ , with equality iff $\vec{b} = 0$ , i.e. iff $B$ is parallel to $A$ . Why: here is the sign flip, in one line. The spatial part $|\vec{b}|^2$ enters the norm with a minus, so making $B$ spatially non-parallel to $A$ increases $b^0$ relative to $\sqrt{B\cdot B}$ . In Euclid it would decrease the projection. The indefinite metric turns the inequality over.

Step 3. Now the triangle. $(A+B)\cdot(A+B) = A\cdot A + 2A\cdot B + B\cdot B \geq A\cdot A + 2\sqrt{A\cdot A}\sqrt{B\cdot B} + B\cdot B = \big(\sqrt{A\cdot A} + \sqrt{B\cdot B}\big)^2$ . Both sides are positive (the sum of two future timelike vectors is future timelike), so we may take square roots: \sqrt{(A+B)^2} \ \geq\ \sqrt{A^2} + \sqrt{B^2}, \qquad\text{i.e.}\qquad \tau(A+B) \geq \tau(A) + \tau(B), with equality iff $A \parallel B$ . Why: read it as physics. $A$ and $B$ are the two legs of a bent worldline; $A + B$ is the straight worldline joining the same endpoints. The straight one accumulates more proper time, and strictly more unless the "bend" is not a bend at all.

Step 4. Go continuous. For a worldline from $P$ to $Q$ , the proper time is $\tau[\text{path}] = \int\sqrt{1 - v^2/c^2}\,\mathrm{d}t$ (equation 4). Work in the frame where $P$ and $Q$ are co-located. The integrand is $\leq 1$ , with equality only when $v = 0$ , so $\tau \leq \int \mathrm{d}t = \Delta t$ — and $\Delta t$ is exactly what the stationary (inertial) worldline accumulates. Varying properly: $\partial L/\partial v = -v/\sqrt{1-v^2} = 0$ gives $v = 0$ , and the second derivative $\partial^2 L/\partial v^2 = -(1-v^2)^{-3/2}$ is negative for every $v$ , so the integrand is strictly concave and the stationary point is a strict maximum. Why: not a minimum, not a saddle — a maximum, and the negative second variation is the proof. "Geodesic" in Lorentzian geometry means longest, not shortest, for timelike curves.

Step 5. Cash it in. Two events, $10$ years apart in the Earth frame, same place. The stay-at-home twin's worldline is straight: $\Delta s^2 = 10^2 - 0 = 100$ , so $\tau = 10$ years. The traveller goes out at $0.6c$ for $5$ years (reaching $3$ light-years) and returns: each leg has $\Delta s^2 = 5^2 - 3^2 = 16$ , so $\tau_{\text{leg}} = 4$ and $\tau_{\text{total}} = 8$ . And $10 > 8$ — the straight worldline is longer, by exactly $2$ years. Why: nothing about clocks, nothing about acceleration, nothing about who turns around — just the observation that a bent path through Minkowski spacetime is shorter than a straight one. The asymmetry the "paradox" complains about is the asymmetry between a straight line and a bent one, and it is visible on the diagram. *Verified in verify_aq-sr-spacetime-interval.py: over 20 000 random pairs of future-directed timelike vectors, $A\cdot B \geq \sqrt{A^2}\sqrt{B^2}$ and $\tau(A+B) \geq \tau(A) + \tau(B)$ without exception, with equality exactly when parallel; $\partial L/\partial v = 0$ solves to $v = 0$ and $\partial^2 L/\partial v^2 < 0$ everywhere; and the twin numbers $10 > 4 + 4$ reproduce.*

Figure

The twin paradox is the reversed triangle inequality. Both twins start and end at the same two events, 10 years apart in the Earth frame. The straight worldline accumulates Δs² = 100 ls-yr², so τ = 10 yr. The bent worldline is two legs of Δs² = 5² − 3² = 16 each, so τ = 4 + 4 = 8 yr. The detour COSTS 2 years — the exact opposite of Euclid, where a detour always costs extra distance

Limiting cases

Case	Result	Meaning
$\Delta s^2 > 0$	timelike	$\Delta\tau = \sqrt{\Delta s^2}/c$ — a real clock reading; the pair is causally connectable
$\Delta s^2 < 0$	spacelike	$L_0 = \sqrt{-\Delta s^2}$ — a real ruler reading; the pair is causally disconnected
$v \ll c$	$\Delta\tau \approx \Delta t(1 - \tfrac12\beta^2)$	the fractional time loss is $\beta^2/2$ — this single term is the whole SR part of the GPS correction
$c \to \infty$	$\Delta s^2/c^2 \to \Delta t^2$	the interval degenerates to coordinate time: absolute simultaneity returns, and the spatial part drops out entirely
bent vs straight	$\tau_{\text{bent}} < \tau_{\text{straight}}$	the reversed triangle inequality; every detour is paid for in proper time, with no exceptions

Common trapThe interval is not a distance, and the straight line is not the shortest path

Two Euclidean habits must be actively unlearned. First, $\Delta s^2$ is a difference of squares, not a sum, so it can be positive, negative or zero — "the square root of the interval" is a meaningless phrase until you know the class and the convention. Second, and more insidiously, the extremal property is inverted: between two timelike-separated events the straight worldline is the path of maximum proper time, and every deviation makes the clock read less. Carry the Euclidean intuition across and you expect the travelling twin to age more. The physical statement is: an inertial clock always beats a non-inertial one over the same pair of events, and its winning margin, $\Delta t - \Delta t/\gamma$ , is fixed by the geometry.

Worked examples

1 · What the interval does to your phone

A GPS satellite orbits with a period of $11$ h $58$ min. Compute its speed from that period, then use the interval to find how much its onboard clock loses per day relative to a clock on the ground, from special relativity alone.

Step 1 — identify. From Kepler's third law with $GM_{\oplus} = 3.986004418\times10^{14}\ \mathrm{m^3\,s^{-2}}$ and $P = 43\,080\ \mathrm{s}$ , the orbital radius is $a = (GM P^2/4\pi^2)^{1/3} = 2.656\times10^{7}\ \mathrm{m}$ — an altitude of about $20\,200\ \mathrm{km}$ . The orbital speed is $v = \sqrt{GM/a} = \mathbf{3874\ m/s}$ , so $\beta = v/c = \mathbf{1.292\times10^{-5}}$ .

Step 2 — apply the interval. Over a coordinate time $\Delta t$ the satellite covers $\Delta x = v\,\Delta t$ , so \Delta s^2 = c^2\Delta t^2 - v^2\Delta t^2 = c^2\Delta t^2\left(1 - \beta^2\right) \quad\Longrightarrow\quad \Delta\tau = \frac{\sqrt{\Delta s^2}}{c} = \Delta t\sqrt{1-\beta^2} = \frac{\Delta t}{\gamma}. Nothing here is a separate "time dilation formula": it is equation (3) applied to a timelike interval, and nothing else.

Step 3 — the boundary number. Expand: $\Delta t - \Delta\tau = \Delta t\big(1 - \sqrt{1-\beta^2}\,\big) \approx \Delta t\,\beta^2/2$ . With $\beta^2/2 = 8.35\times10^{-11}$ and $\Delta t = 86\,400\ \mathrm{s}$ (one day), \Delta t - \Delta\tau = 86\,400 \times 8.35\times10^{-11} = 7.21\times10^{-6}\ \mathrm{s} = \mathbf{7.2\ \mu s\ per\ day}. The satellite clock runs slow by $7.2$ microseconds a day.

Step 4 — why it matters. Seven microseconds of clock error is $c \times 7.2\times10^{-6} = \mathbf{2.2\ km}$ of position error, accumulating every day. (General relativity supplies a larger correction of the opposite sign — the satellite's clock runs fast by about $45.8\ \mu\mathrm{s/day}$ because it sits higher in the Earth's gravitational potential — for a net $+38.6\ \mu\mathrm{s/day}$ , which the satellites are pre-corrected for before launch. That piece is gravitational and is not derived here.) The special-relativistic $7.2\ \mu\mathrm{s}$ , though, is nothing but equation (3), computed for a spacecraft.

2 · The twins, as a length calculation

One twin stays home. The other flies out at $0.6c$ for $5$ years of Earth time, turns round, and flies back. Compute both proper times directly from the interval, with no mention of time dilation.

Step 1. The two events that both twins share are DEPART and REUNION, separated in the Earth frame by $\Delta t = 10\ \mathrm{yr}$ and $\Delta x = 0$ .

Step 2. The stay-at-home twin's worldline is the straight segment joining them. Its interval is $\Delta s^2 = (10)^2 - 0^2 = 100$ (in light-year units), so his proper time is $\tau_{\text{stay}} = \sqrt{100} = \mathbf{10\ years}$ . For an inertial observer, proper time and coordinate time coincide — which is what "at rest" means.

Step 3. The traveller's worldline is two straight segments. On the outbound leg $\Delta t = 5\ \mathrm{yr}$ and $\Delta x = 0.6 \times 5 = 3\ \mathrm{ly}$ , so $\Delta s^2 = 5^2 - 3^2 = \mathbf{16}$ and $\tau_{\text{out}} = \sqrt{16} = 4\ \mathrm{yr}$ . The return leg is the mirror image: another $4$ years. Total: $\tau_{\text{travel}} = \mathbf{8\ years}$ .

Step 4 — the boundary number and the moral. The traveller returns $\mathbf{2\ years}$ younger, computed without ever using $\gamma$ (as a check, $10/\gamma = 10/1.25 = 8$ ✓). The asymmetry has nothing to do with acceleration as such and everything to do with the shape of the two paths: one is straight, one is bent, and by the reversed triangle inequality the straight one is longer. Push to $c$ and the interval goes to zero — the ultimate detour costs all of your proper time. (The paradox and its resolutions are treated in the twin-paradox articles; the geometry is here.)

Physics of the portrait

The interval has no discoverer's name attached because it is not a discovery so much as a change of subject: the moment Minkowski proposed that $c^2t^2 - x^2$ was the thing to look at, relativity stopped being a theory about rods and clocks and became geometry.

Origins · 2 minSpacetime interval and how this was discovered

Spacetime interval

1908–present · Minkowski's Cologne lecture · the calibration hyperbola · every GPS satellite ever launched

Einstein's 1905 paper contains the Lorentz transformation but not the interval. It was Hermann Minkowski who, in the 1908 Cologne lecture Raum und Zeit, saw that the transformation preserves a quadratic form, and that this quadratic form is the geometry — that space and time were not two things being mixed but one thing being sliced. What he provided was not a new prediction but a new object, and the object turned out to be the theory. Every result in special relativity became a statement about the interval: time dilation is a timelike vector's time component growing as it is rotated; length contraction is the same for a spacelike vector's space component; the relativity of simultaneity is the fact that a hyperbolic rotation has no invariant notion of "horizontal". Einstein, who had been Minkowski's student, at first regarded the reformulation as superfluous learning; within four years he was using it to build general relativity, where the interval — now with a position-dependent metric — is the only structure left, and there is nothing to compute with except $\mathrm{d}s^2$ . It is also, quietly, infrastructure: the $7.2\ \mu\mathrm{s}$ a day that every GPS satellite's clock loses is $\sqrt{\Delta s^2}/c$ and nothing else.

History

1905

— Einstein's Zur Elektrodynamik bewegter Körper derives the Lorentz transformation from two postulates, without ever writing down the invariant it preserves.

1908

— Minkowski's Raum und Zeit, delivered in Cologne on 21 September, identifies the invariant quadratic form and recasts relativity as the geometry of a four-dimensional continuum.

1909

— Minkowski dies in Göttingen on 12 January, aged 44; the Cologne lecture is published posthumously.

1911

— Max von Laue publishes Das Relativitätsprinzip (Vieweg, Braunschweig), an early monograph written in Minkowski's four-dimensional language.

1916

— Einstein's general relativity promotes the interval to a position-dependent line element, making it the only surviving structure of the theory.

1971

— Hafele and Keating fly caesium clocks around the world in both directions and measure the accumulated proper-time differences against a ground clock.

1980

s — the GPS constellation goes operational with its satellite clocks pre-corrected for the relativistic rate offset (SR −7.2 μs/day, GR +45.8 μs/day, net +38.6 μs/day).

Going further

Texts. Rindler, Introduction to Special Relativity, 2nd ed., §7–8 · Taylor & Wheeler, Spacetime Physics, 2nd ed., ch. 1 (the interval as the central object) · Morin, Special Relativity, ch. 4 · Misner, Thorne & Wheeler, Gravitation, ch. 1–2 · Hartle, Gravity, ch. 4. The S1 treatment, The Invariant Spacetime Interval, works the invariance out kinematically from the Lorentz transformation; this article derives it from the light postulate and recasts it as the metric-induced norm.

Provenance. The base article was written entirely in the opposite signature $(-,+,+,+)$ : $\Delta s^2 = -c^2\Delta t^2 + \Delta x^2$ , $\eta = \mathrm{diag}(-1,+1,+1,+1)$ , proper time from $\Delta s^2 = -c^2\Delta\tau^2$ , proper length from $\Delta\sigma = \sqrt{+\Delta s^2}$ , hyperbolae $-c^2t^2 + x^2 = \text{const}$ . All five are false in this chapter's locked signature and have been re-derived, not relabelled. Anchors recomputed in verify_aq-sr-spacetime-interval.py: the GPS chain ( $P = 43\,080$ s, $a = 26\,561$ km, $v = 3874\ \mathrm{m/s}$ , $\beta = 1.292\times10^{-5}$ , SR loss $7.21\ \mu\mathrm{s/day}$ , matching the canonical $-7.2$ ); the twin chain ( $10$ yr vs $4+4 = 8$ yr, deficit $2$ yr); the hyperbola events $(5,0)\to(6.25,3.75)$ at $\Delta s^2 = +25$ and $(0,4)\to(3,5)$ at $\Delta s^2 = -16$ ; and the reversed triangle inequality over 20 000 random future-timelike pairs, with a strictly negative second variation.

Builds on. Events and Worldlines · Spacetime Diagrams · Lorentz Invariance and Invariants. Leads to. Timelike, Spacelike and Null Separations · The Light Cone · The Minkowski Metric · Proper Time and Proper Length · Four-Vectors.

Why is the interval a difference of squares rather than a sum?

Because a sum would be wrong. If the invariant were $c^2\Delta t^2 + \Delta x^2$ , boosts would be ordinary rotations in the $(ct,x)$ plane, and a large enough "rotation" could turn a time separation into a space separation — you could turn round in time the way you turn round in space, and causality would be gone. The minus sign is what keeps the light cone in place: the level sets are hyperbolae, which never cross their asymptotes, not circles, which sweep through every direction. The deep reason is the light postulate: a form preserving the cone $c^2t^2 = r^2$ must be proportional to $c^2t^2 - r^2$ , and that form is indefinite whether we like it or not.

If proper time is the arc length of a worldline, does an accelerating clock need a special formula?

No, and that is the point of the arc-length form. Equation (4), $\tau = \int\sqrt{1 - v(t)^2/c^2}\,\mathrm{d}t$ , holds for any motion: the integrand depends on the instantaneous speed and on nothing else — not on the acceleration, not on its history. This is the clock postulate, and it has experimental content: muons in a storage ring experience proper accelerations of order $10^{18}\,g$ , and their decay rate still tracks $\gamma$ computed from instantaneous speed alone. An accelerating clock needs no new formula; it just needs its worldline integrated.

What is the interval between two events on a light ray, and what does a photon experience?

The interval is exactly zero: that is the definition of null, and every observer agrees, even while disagreeing wildly about the $\Delta t$ and $\Delta x$ that go into it. The tempting next step — "so no time passes for a photon" — has no meaning behind it. Proper time is the arc length of a worldline parametrised by its own clock, and a null curve has zero arc length, so it cannot be parametrised that way. There is no rest frame for a photon, not because the frame is exotic but because the limit does not exist: boosting toward $c$ sends $\gamma \to \infty$ , not to any frame. The honest statement is that the interval along a light ray is zero, and "the photon's point of view" is not something the theory offers.

Doesn't the twin paradox need acceleration to resolve it?

Acceleration is how the traveller's worldline gets bent, but it is not what makes her younger — the bend is. The reversed triangle inequality compares two paths between the same two events and finds the straight one longer; it does not care how the bend was achieved. You can even remove the acceleration entirely: let a passing outbound rocket copy the Earth's clock reading, a passing inbound rocket copy the outbound one's, and hand the total to Earth. No object accelerates, and the elapsed proper time along the bent path is still $8$ years against Earth's $10$ . What breaks the symmetry is that only one worldline is straight. "Who accelerated?" is a symptom; "which path is straight?" is the cause.

Why does the derivation need continuity to rule out k = −1?

Because $k = -1$ is not algebraically absurd — it satisfies $k^2 = 1$ perfectly well — and it would be a disaster: it would send $\Delta s^2 \to -\Delta s^2$ , turning every timelike pair spacelike in the primed frame and demolishing the causal order. What excludes it is that $k$ is a continuous function of the boost velocity, takes values only in $\{+1, -1\}$ , and equals $+1$ at $v = 0$ (where the transformation is the identity). A continuous function into a two-point set is constant on a connected domain, and the boosts form a connected family containing the identity, so $k = +1$ everywhere. The argument is short but not empty: it is exactly where the connectedness of the proper orthochronous Lorentz group enters, and it is why time reversal and parity, which sit in the other components, are not boosts.

What is the interval good for once you leave kinematics?

It becomes the template for every invariant in relativistic physics. Replace the separation four-vector by the four-momentum and the same quadratic form gives the invariant mass, $m^2c^4 = E^2 - (pc)^2$ — the number every observer and every detector agrees on, and the number a resonance appears as a bump in. Replace it by the wave four-vector and you get the invariant phase of a plane wave, from which the relativistic Doppler effect follows. Replace the flat metric by a curved one and the construction is general relativity. The interval's importance is not that it happens to be conserved: it is the pattern — a metric contraction of two four-vectors — from which every Lorentz scalar in physics is built.