The quantity nobody argues about

Relativity is usually introduced as a list of things that stop being fixed. A rod’s length depends on who measures it. The interval between two ticks depends on who counts. Whether two things happen at the same moment depends on who is asked. Presented that way the theory reads as a dismantling, and its most important content — that something specific is put in place of what was removed — arrives late if at all.

The quantity that survives is

$$s^2 = c^2t^2 - x^2,$$

and its status is not a convenience. Any two observers in uniform relative motion disagree about $t$ and about $x$, in a way fixed by their relative speed, and agree about that combination to whatever precision either can measure.

Why the diagram is unusable without it #

There is a practical reason for meeting the interval early, and it is one that a first course usually skips.

The temptation is to assume that a unit is a unit — that a centimetre along the tilted time axis is the same tick as a centimetre along the vertical one. It is not, and assuming it produces exactly the confusion in which each observer appears to find the other’s clock running slow, which reads as a contradiction. The hyperbolae fix the scale: a unit of proper time is wherever the $s=1$ curve crosses, and on the page that is $\gamma$ times further out.

Once the scale is right, the apparent symmetry stops being paradoxical and becomes a statement about projections. Each observer projects the other’s worldline onto their own time axis, and each projection is shorter than the thing projected — exactly as two people walking on diverging paths each see the other recede, with no contradiction and no privileged walker.

Push the same construction to 0.8 c, where $\gamma$ is 1.667, and the tilted axes close much further onto the light line — and the useful thing about watching them move is that nothing happens abruptly anywhere. Everything in the picture is continuous in $\beta$: the axes approach the 45° line and never reach it, which is the geometric form of the statement that no boost reaches the speed of light. There is no last frame before the wall, because there is no wall on the diagram, only an asymptote.

Three kinds of separation, and only three #

The sign of the interval divides every pair of events into three classes, and the division does not depend on who is doing the dividing.

That trichotomy is what makes relativity a theory with a causal structure rather than a free-for-all. The events whose time order is negotiable are exactly the events that cannot affect one another; the events that can affect one another have an order every frame agrees about. Nothing has to be added to the theory to protect cause from effect — the protection is the sign of a quadratic form.

The disagreement itself is easy to draw: two events simultaneous in one frame are separated in time in another, by an amount that depends on how far apart they are. What the interval adds to that picture is the boundary. The reordering is possible only for spacelike pairs, so the set of events whose “now” is negotiable is precisely the set outside the light cone — and it is the same set for everybody, which is the only reason the negotiation is harmless.

The null case deserves a moment. Two events joined by a light ray have zero interval however far apart they are — the emission of a photon at a distant star and its absorption in an eye are separated by nothing at all in this metric. It is the clearest sign that $s^2$ is not a distance in any ordinary sense: distances vanish only between coincident points, and this one vanishes along an entire cone.

Proper time is a length #

The most useful thing the interval does is give a worldline a length, and that length is what a clock carried along it reads.

The minus sign is doing all of that work, and it is worth naming as the single structural difference between this geometry and Euclid’s. In a Euclidean plane the straight line between two points is the shortest path and a detour costs extra length. Here a detour saves length, because the spatial part is subtracted, and the more the path deviates the less time it accumulates. A twin who leaves and returns has taken a detour through spacetime, and the arithmetic of who is younger is the arithmetic of which path is longer — nothing more.

The usual telling of that story reaches for the acceleration at the turnaround, and the acceleration is real and is what breaks the symmetry between the two twins. But it is not what produces the number. The number is a path length, computed above with no acceleration anywhere in it, and the turnaround matters only because it is what makes one path bent.

Minkowski, and a remark that aged well #

The interval was not in Einstein’s 1905 paper. What that paper has is the transformation, derived from two postulates, with the consequences worked out one at a time — the contraction, the dilation, the failure of simultaneity — and no geometry at all.

Minkowski, who had taught Einstein mathematics at Zurich and had thought him a lazy student, supplied the geometry in 1908, in a lecture whose opening is the most quoted sentence 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 had noticed is that the whole of the 1905 paper is the statement that a particular quadratic form is invariant, and that the transformations preserving it form a group in the same way rotations do.

Einstein’s initial reaction was that it was superfluous erudition. He changed his mind within four years, and for a specific reason: the general theory could not have been written without it. Generalising “the interval is invariant” to “the interval is invariant with a position-dependent metric” is a sentence that can be written down; generalising a list of consequences about rods and clocks is not.

That sequence — a physical result, then a geometric reformulation that looks like decoration, then a generalisation that is only available in the geometric language — is worth noticing because it recurs. The reformulation is doing work even when it adds no new prediction, because it changes what the next question can be.

The transformation, seen as a rotation #

Once the interval is the invariant, the Lorentz transformation is not a strange rule about lengths and times; it is whatever transformation preserves it, in the same way that a rotation is whatever transformation preserves $x^2+y^2$.

Rotations are parameterised by an angle and preserve a circle; Lorentz transformations are parameterised by a rapidity and preserve a hyperbola, with $\cosh$ and $\sinh$ where the cosine and sine were. $\gamma$ is the $\cosh$ of the rapidity, and the fact that it runs away to infinity rather than coming back round is the same fact as the hyperbola being unbounded where a circle is not.

That reading resolves a question that otherwise looks like a coincidence: why velocities do not simply add. Angles add under successive rotations, and it is rapidities that add under successive boosts — the velocity is the hyperbolic tangent of the rapidity, and the tangent of a sum is not the sum of the tangents. The velocity-addition formula is the hyperbolic-tangent addition identity, and the speed of light is unreachable because the tangent approaches one asymptotically however large its argument.

The composition rule follows from that immediately. Adding 0.9 c to 0.9 c gives 0.994 c rather than 1.8, which looks arbitrary in velocities and is an ordinary addition in the parameter: 1.472 plus 1.472 is 2.944, and the hyperbolic tangent of that is 0.994. A formula that has to be memorised in one coordinate is a sum in the other, and the invariant hyperbola is what says which coordinate that is.

The apparatus the whole thing can be derived from is a light pulse bouncing between two mirrors, seen from a frame in which the clock is moving: the path is longer, the speed is the same, so the tick is longer, by $\gamma$. What the interval adds is that this is not a peculiarity of light clocks. The interval between two ticks is a property of the pair of events, so every clock carried alongside must agree with it, whatever mechanism it runs on — a caesium standard, a decaying muon, a heartbeat.

The muon, which is the measurement #

The interval is not an interpretive convenience; it is measured, routinely, in a way that makes the geometry hard to avoid.

Muons are produced in the upper atmosphere by cosmic rays, at around 15 km, and decay with a half-life of 1.56 microseconds. Even at the speed of light a muon covers 470 metres in a half-life, so the number arriving at sea level ought to be smaller than the number produced by a factor of about $2^{-32}$ — one in four thousand million. The measured factor is nearer one in ten.

Both frames account for it and neither needs the other. In the ground’s frame the muon’s clock runs slow by γ, which for a typical 3 GeV muon is about 30, so it lives thirty times longer and covers thirty times the distance. In the muon’s own frame it lives exactly 1.56 microseconds and the atmosphere is contracted to 500 metres, which it crosses easily. The two descriptions disagree about every quantity in them and agree about the only thing that can be checked: how many muons reach the detector.

That agreement is the interval doing its work. The number of decays a muon undergoes between two events is a count of ticks of a clock it carries, which is a proper time, which is the length of its worldline — and a length is not something two observers can disagree about. The experiment was done in 1940 by Rossi and Hall with a detector on a mountain and one at sea level, and it remains the cleanest demonstration that the geometry is the physics rather than a way of talking about it.

The paradox the trichotomy dissolves #

The classification into timelike, null and spacelike is worth exercising on the puzzle it settles most cleanly, because the resolution is not a trick and does not require any new physics.

A pole ten metres long is carried at a speed for which $\gamma = 2$ toward a barn five metres long with a door at each end. In the barn’s frame the pole is contracted to five metres, so there is a moment when it is entirely inside, and both doors can be shut at once with the pole enclosed. In the pole’s frame the barn is contracted to two and a half metres and the pole is ten, so the pole cannot possibly be inside at any moment whatever.

Both descriptions are correct, and the apparent contradiction is entirely in the phrase at once.

Shutting the two doors is two events, and in the barn’s frame they are simultaneous and five metres apart. Their interval is $0 - 25$, which is negative: the pair is spacelike separated. So there is no frame- independent fact about their order, and in the pole’s frame they are not simultaneous at all. The far door shuts and reopens before the pole’s front end reaches it, and the near door shuts after the pole’s back end has passed. The pole is never enclosed, and each door still shuts at a moment when no pole is in it.

Every question with an answer gets the same answer in both frames, and it is worth naming which questions those are. Did the far door strike the pole? That is a question about whether two events coincide, and coincidence is invariant — if two things are at the same place at the same time in one frame, they are in every frame. Was the pole damaged? Same. Did the doors both shut? Yes, in both frames. Were they shut at the same time? That question has no frame-independent answer, and it is the only one the paradox rests on.

The general form is the reason the trichotomy matters. A pair of events whose order is negotiable is a pair that cannot influence each other, so no chain of cause and effect can be reordered and no contradiction can be constructed out of the disagreement. Every relativistic paradox is built by treating a spacelike pair as though its order were a fact, and every resolution consists of pointing that out.

The clocks that have to be corrected #

Proper time is measured continuously, on a working system that would fail within minutes if the geometry were wrong.

A satellite navigation system works by trilateration on travel times: each satellite broadcasts its position and the time it broadcast, and a receiver solves for where and when it must be for four such signals to be consistent. The whole method is a comparison of clocks, and the clocks are moving relative to the ground and sitting higher in a gravitational field.

The kinematic part is the one this essay computes. A satellite at twenty thousand kilometres orbits at about 3.9 kilometres a second, which gives $\gamma - 1 \approx 8.3\times10^{-11}$, so its clock records less proper time than a ground clock by about seven microseconds a day. That is the interval along a worldline, and it is the muon experiment again with a caesium clock in place of a muon.

The gravitational part is larger and belongs to the general theory rather than to this essay: a clock higher in a potential well runs fast, and at that altitude the effect is about forty-six microseconds a day. The two have opposite signs, and the net is that a satellite clock gains around thirty-eight microseconds against a ground clock every day.

Thirty-eight microseconds is eleven kilometres of light travel. An uncorrected system would accumulate that much error in ranging every day, which is not a degradation but a total failure — the position solution would be useless within a couple of minutes and absurd within an hour.

The correction is not applied in software after the fact. The satellites’ clocks are deliberately built to run at a slightly different frequency from their nominal one before launch, so that once in orbit they tick at the rate a ground observer wants. That offset is a permanent, designed-in acknowledgement that proper time is path-dependent.

Two smaller corrections are applied continuously and are worth naming because they are the same geometry at a finer grain. Each satellite’s orbit is slightly elliptical, so its speed and altitude vary through an orbit and the two relativistic terms do not stay constant; a periodic correction of up to some tens of nanoseconds is computed per satellite. And because the Earth rotates while a signal is in flight, the receiver’s own worldline has moved between transmission and reception, which contributes over a hundred nanoseconds for a signal crossing the disc. None of these is an error term added to a Newtonian calculation. They are what it means to compute a proper time along a real path.

What it costs, and where it stops #

The interval is not positive definite, so it is not a distance. Ordinary geometric intuition — the triangle inequality, the idea that a small separation means a small distance — fails, and fails in ways that are easy to import without noticing. The reversed triangle inequality for timelike paths is the twin result; the vanishing interval along the cone is the null result; and any argument that treats $s^2$ as a Pythagorean sum will get both backwards.

It is a flat-space statement. In general relativity the interval survives as a local quantity with a position-dependent metric, and there is in general no way to define an interval between distant events at all. The invariant becomes $\mathrm{d}s^2 = g_{\mu\nu}\mathrm{d}x^\mu\mathrm{d}x^\nu$ and the global structure that made the flat case so simple is gone: the hyperbolae in the first figure have no counterpart in a curved spacetime.

It says nothing about what is happening. Two events at a given interval may be a photon’s emission and absorption, a particle’s decay, or nothing at all. The interval is a property of the pair of points and carries no dynamics — which is exactly why it is invariant, and exactly why a great deal more machinery is needed before anything can be predicted.

Where the same invariant turns up next #

The habit of asking what a transformation preserves, rather than what it changes, is the most portable thing in this subject, and the interval is the first instance rather than the only one.

Pair a particle’s energy with its momentum in the same way and the invariant is $E^2-p^2c^2$, which is the rest mass squared — the same hyperbolae with different labels on the axes. That is what mass and energy being one quantity amounts to: the mass is not one of the two components but the invariant length of the pair, so a particle’s mass is the same in every frame while its energy and momentum are not. A photon’s version of the same statement is that the invariant is zero, which is exactly the null case above.

Pair a wave’s frequency with its wavenumber and the invariant is the phase, which is a count of crests and cannot be frame-dependent — a crest is either at a detector or it is not. That is the honest derivation of the relativistic Doppler effect, and it explains why a transverse shift exists at all: the frequency and wavenumber transform together, so a source moving across the line of sight still shifts even with no component of velocity along it.

Pair the electric and magnetic fields and the invariants are $E^2-c^2B^2$ and $\mathbf{E}\cdot\mathbf{B}$. A configuration that is purely magnetic in one frame has a negative first invariant, and no frame can make it purely electric — which is the frame-independent content of the statement that a magnetic field is electrostatics seen sideways, and the limit on how far that reading can be pushed.

What the picture cannot show #

Every diagram on this page has one space dimension. That is not merely a simplification; it hides the structure most people expect a light cone to have. With two spatial dimensions the cone is a cone; with three it is a three-dimensional surface in four-dimensional spacetime, and no drawing exists. What is lost with the extra dimensions is the transverse behaviour — that a boost leaves transverse lengths alone, and that light travels in every direction rather than along two lines at 45°.

The hyperbolae are also drawn only in the future quadrant. The full set of events at interval $s$ from the origin is two branches, future and past, and in more than one spatial dimension it is a hyperboloid — the surface a particle of fixed mass occupies in momentum space, which is where the same geometry reappears with energy and momentum in place of time and position.

Energy against speed is the same hyperbola drawn in different variables. Energy and momentum transform between frames exactly as time and position do, and their invariant combination $E^2 - p^2c^2$ is the rest mass — so a particle’s mass is the length of its energy–momentum vector in precisely the sense that a proper time is the length of a worldline. The null case survives the translation too: a photon’s vector has zero length, which is the whole content of the statement that it has no mass.

And the scale is a choice. Drawing light at 45° amounts to measuring time in metres, or distance in seconds; a diagram in seconds and metres would have light at a slope of $3\times10^8$ and would be unusable. That convention is so universal that it is easy to forget it was made, and it is the reason the interval looks symmetric between its two terms when in ordinary units it is not.

The ladder from here #

Later rungs on this anchor: the Lorentz transformation derived as the linear map preserving the interval, with rapidity as the parameter; the invariant interval in momentum space, where $E^2-p^2c^2$ is the rest mass and the same hyperbolae reappear; four-vectors as the objects the transformation acts on, and invariants as their inner products; proper time as an integral along an arbitrary worldline, and the accelerated case; and the metric in curved spacetime, where the interval survives locally and everything global is lost.

The neighbouring ladders are the spacetime diagram, which this puts a scale on, simultaneity, which is the disagreement the invariant bounds, and the twin paradox, which is a statement about path length in this metric and nothing else.