Put a line together with a circle and you get π: the diameter, laid along the circumference, fits three times and a little more. Put a vertical line together with a horizontal one and you get √2, the diagonal of the square they make. Nobody measured these into existence and nobody asks where they came from — they are what those shapes do when they play together.
This page asks whether the fine structure constant belongs to the same family. Put spheres together with a wave — a wave that has to close on itself, and spheres of phase influence whose radii are fixed by nothing but the distance between two particles — add a touch of relativity and quantum mechanics, and the geometry leaves exactly one dimensionless number free. It comes out at 1/137.
Put a quantum amplitude around a circumference and it cannot be arbitrary: after one full trip it must arrive at the phase it started with. That selects a discrete family of modes: whole numbers of wavelengths, and half-numbers too, which come back inverted and close on a second turn. Drag the mode number away from those values and the wave fails to join up — the gap in rose is the phase it never closes. standard
The ground mode is the one with a single wavelength around the loop, λ = 2πR, which is the same statement as one unit of angular momentum:
Everything that follows is built on this single closed wave. The next panel asks what its amplitude has to be.
The amplitude comes from radial velocity uncertainty Δṙ, which quantum mechanics forbids from being exactly zero. That uncertainty makes the travelled path longer than the circle it rides on. Relativity pulls the other way: at orbital speed v lengths contract by √(1−(v/c)²).
The two effects cancel exactly at one value. When Δṙ/v and v/c are the same number, the contracted wavy path has precisely the length 2πR of the plain circle — and a dot running each of them keeps step forever. That number is 0.0072974. consequence
The two path lengths are related by a right triangle: the wavy path is the hypotenuse, the circumference one leg, the radial excursion the other. Writing α for the ratio Δṙ/v, the circumferential constant that survives is not π but
and the amplitude is A = απR/2. Nothing about electromagnetism has been assumed — α is so far only a name shared by two ratios of speeds that had no reason to be equal.
This is what justifies promoting π. The contracted path is not merely shorter: it is exactly the wave belonging to a circle of circumference 2π′R. The particle still covers 2πR of path in one period, but the circumference it winds around has become 2π′R. So π′ is not an extra assumption bolted on — it is what the lengthening and the contraction leave behind once they have both acted.
This is the meeting point the rest of the page stands on. Bohr's quantization of angular momentum and Einstein's kinematics come from different theories, and here they meet in one right triangle and leave a single invariant behind. Everything that follows carries that invariant from the plane into space.
The flat picture has one direction transverse to the motion: radially, in the plane. Space has two. Rotate the circle out of the page and the second one shows itself — a wave running up out of the orbital plane. Nothing in a two-body configuration prefers either, so isotropy splits the excursion equally between them and each carries A/√2. Only the in-plane one moves the phase surface; the other runs tangent to it and leaves the radius alone. consequence
Schrödinger's cat is alive and dead in equal measure: (|alive⟩ + |dead⟩)/√2. Each branch carries amplitude 1/√2 because probabilities are squares of amplitudes, and the two squares, ½ and ½, have to add up to one.
The excursion here has the same structure. It is an equal combination of two orthogonal states, radial and binormal, and each carries A/√2 so that the squares add back to A². As with the cat, only one branch matters for what happens next: the radial one moves the surface, and the binormal one runs along it. So the boundary sees A/√2, not A.
The analogy covers the bookkeeping: equal weights, with squares that sum to the whole. It makes no claim that the boundary is in a measured superposition.
A second and separate freedom remains: the orbital plane can sit at any angle to the line joining the two particles whose interaction the next panel builds. Averaging the in-plane component over that angle with uniform measure gives a factor 2/π, and the two together fix the boundary displacement.
Drop a needle at random onto a floor ruled with parallel lines. How often it crosses a line depends only on the average length of its shadow across the lines, and averaging over a half-turn of orientations gives 2/π. That is why π turns up in a problem with no circle in it.
The tilt average is the same integral. The in-plane component plays the needle, the interaction axis is the normal to the rulings, and (1/π)∫sin θ dθ = 2/π is the mean shadow. Choose a needle of length R and rulings spaced by the sphere diameter 2R, and the crossing probability is exactly 1/π. That number returns in the corrections.
The correspondence holds for rotation about a single axis, which is the one degree of freedom the construction assumes. A fully isotropic orientation in three dimensions would carry the weight sin θ dθ/2 and a different average.
Both surfaces move inward by A′, so the separation closes by twice that, and in units of the radius the shift is
Give each of two particles a sphere of phase influence with radius equal to their separation — the only choice that introduces no second length. The sphere volume counts the modes available to one particle; the lens where the spheres overlap counts the modes they share. The condition posited is that the ratio of the two equals the same invariant that selected closed modes in the plane.
Sharp classical spheres give 16/5 = 3.2, too few shared modes. Quantum boundaries are not sharp: each surface is displaced inward, the lens grows, and the ratio falls. Approach with the coarse control, then land with the vernier — within a few vernier steps it settles exactly onto closure. postulate
The inward displacement of each surface works out to √2αR/2, so the separation becomes s = 1 − √2α. The factor √2 is the one established in the panel above.
With the displacement written in terms of α, the closure condition becomes an equation in α alone. Across the whole physically open range it has a single root. The value of the fine structure constant was never put in; α entered only as the unknown. consequence
In 1928 Dirac combined quantum mechanics with special relativity, the same pairing this page uses, and found that the electron's magnetic moment comes out fixed. The gyromagnetic ratio is g = 2, with nothing to adjust. It was a triumph, and it was off by about a part in a thousand. Measurements in 1947 put g slightly above 2, and in 1948 Schwinger accounted for the difference: ae = (g−2)/2 = α/2π ≈ 0.00116.
Feynman recalled watching someone throw a plate in the Cornell cafeteria. He noticed that its wobble and its spin ran at a ratio of 2:1, a factor of two that comes purely from the kinematics of a spinning rigid body. He worked out the equations for fun, and later said that this play led him back to the problems that became his quantum electrodynamics.
Dirac's g = 2 is a factor of two of the same kind. It is fixed by how spin and motion are tied together, not by any number put in by hand.
The plate is classical mechanics. It shows that a clean factor of two can come from rotation alone, not how the Dirac equation produces it.
The miss shows up as precession. In a magnetic field, the electron's momentum turns at the cyclotron rate while its spin precesses at a rate proportional to g. If g were exactly 2, the two would turn in lockstep forever. They do not: the spin creeps ahead by a fraction ae per turn, and Penning-trap experiments measure that slow drift. The most precise of them (Fan et al. 2023) supplies the value this page is compared against.
The spin runs ahead because of the vacuum. The electron continually emits and reabsorbs virtual photons, and these fluctuations shift its coupling to the field. This is Schwinger's vertex correction. The same quantum vacuum, confined between two uncharged metal plates, pushes them together; Casimir predicted that effect in 1948, the same year. The two are separate calculations, and the g−2 correction is not a Casimir force. What they share is the fact that empty space has structure, and that structure leaves a mark on clean first answers.
The geometric construction follows the same pattern. With rigid spheres, the leading order lands at 137.190, which is 0.112% from measurement, about the same size as Dirac's miss. That match in size is expected rather than remarkable: it is the size any first correction in α/π ≈ 0.0023 has. In both cases the remainder is assigned to fluctuations around a rigid first answer. The next two sections show how the geometry accounts for its part in a thousand.
The leading order treats the phase surface as a fixed wall. It is not: virtual quanta sample the lens and cross it, and both processes change the effective separation. Every correction turns out to be a moment of one Bernoulli variable whose parameter is 1/π — the same normalisation already used in the tilt average. Its variance drives the one-loop compression, its thinned mean drives every traversal. posited mechanism scaling exact coefficients
What the traversal orders weigh is not how many quanta are absorbed, but how much is not already represented in the leading-order average. A mode stays represented only if it was in that average and survived the crossing, so the surviving weight is q = p/e = 0.1171 and the correction carries the remaining 0.8829, halved for interchange. Absorption enters only through the second of the two gates, acting on the thin baseline stream.
The chain alternates in sign but is a single negative feedback of gain α/π = 1/430, not a sequence of separate pushes: it resums to 1/(1+α/π) and its net effect is compressive.
The traversal chain is a geometric series and sums in closed form, so the final number carries no truncation. The same lens also returns the leading anomalous magnetic moment, α/2π, from factors already fixed above. check, not derivation
Between two parallel plates, only the vacuum modes that fit the gap survive, while every mode is allowed outside. The imbalance presses the plates together.
The second-order term uses the same logic. Under static counting, there are π − 1 modes outside the lens for every mode inside it, and that excess acts as a pressure that compresses the lens. This is the largest correction, and it is the one that takes the agreement from a part in a thousand to half a part per million.
This is a counting argument, not a stress computation. The vacuum stress on a single conducting sphere is known to be outward (Boyer 1968), and the direct calculation on the lens boundary remains open as an independent test.
Each needle drop is a yes-or-no event: the needle crosses a line or it does not, with probability 1/π at the spacing above. That makes it a Bernoulli trial, the same variable from which every correction is built.
The variance of that trial is p(1−p) = (π−1)/π² = 0.217, which is exactly the second-order coefficient. The needle that fixed the leading order is reused here rather than replaced by something new.
Montmort, 1708: n guests check their hats, and the hats come back at random. As n grows, the chance that no guest gets their own hat tends to 1/e. The number of matches is Poisson with mean one, and 1/e is its zero count.
The transmission factor is the same quantity. One lens lies between the spheres, so there is on average one chance of absorption per crossing, and the chance of crossing untouched is e⁻¹. A mode stays represented only if it is a crossing needle and passes through untouched: q = 1/(eπ).
The same 1/e also appears in the secretary problem, where the best stopping rule succeeds with probability 1/e. There it comes from an optimisation rather than a Poisson count, so the hat-check problem is the exact counterpart.
Closure fixes where the lens has to sit, and that barely moves between orders. What changes is how the displacement gets paid for. With both corrections off, the leading term covers it alone at α⁻¹ = 137.190. Switch on the Casimir compression and it takes 1.16×10⁻⁵ out of the budget — the lens is squeezed inside the outline closure demands. Scroll α⁻¹ down until it seats back in. The traversal chain does the same thing on a far smaller scale; turn it on alone to see it at its own magnification.
| Term included | Solution for α⁻¹ | Relative error |
|---|
Teal digits agree with the measured 137.035999166; the first amber digit is where the series and the measurement part company.
The two atomic-recoil determinations disagree with each other at about 5.5σ, so no single measurement can settle the question today. The construction still commits to a direction: improved recoil work should converge near 137.0359991768 rather than toward either present central value.
| Determination | α⁻¹ |
|---|
Read back from the end, the construction draws on a short list of objects that physics and probability already know well. Each one is an average, a superposition or a count, and each has been checked independently many times over.
| Element | Role here | Familiar form |
|---|---|---|
| λ = 2πR | phase closure of the ground mode | Bohr's quantization, L = ℏ |
| π√(1−α²) | wave lengthening meets Lorentz contraction | Pythagoras and special relativity |
| 1/√2 | equal split between two transverse states | Schrödinger's cat |
| 2/π, 1/π | average over one rotation; baseline membership | Buffon's needle |
| (π−1)/π² | variance of membership; second-order coefficient | Bernoulli trial |
| π − 1 | exterior mode excess compressing the lens | Casimir effect |
| e⁻¹ | untouched crossing of one lens | Montmort's hat-check problem |
| α/2π | lens cap fraction × tilt average × α | Schwinger's term — a check, not an input |
This is why the list of ingredients stays short. Familiarity says the rules are natural; it does not say nature uses them. The leading order is a single postulate and lands within 1.1×10⁻³ unaided. The corrections add six modelling decisions, each the first option the geometry suggests and none chosen by comparison with the measured value. What decides the matter lies outside this page: two computations open to standard methods (the hard-wall two-loop vacuum polarization and the vacuum stress on the lens boundary) and the next generation of measurements of α.