Closing the gap with linked clustersThe last 29 parts per million — and why they're on the lattice too

The BKT formula puts 1/α at 137.032. The measured value is 137.035999. The gap is 29 parts per million — 3% of a percent — and the whole chapter is about where it comes from and how the lattice closes it. Spoiler: the gap is not a flaw in the formula. It's the difference between α at the UV lattice scale (where BKT lives) and α at zero momentum (where we measure). The lattice has an answer for that difference too, from the same diamond geometry, via a linked-cluster expansion over small subgraphs.

The vacuum polarises

The fine structure constant has a peculiar property standard physics has known since the 1940s: it's not actually a constant. The value depends on the energy scale at which you probe the electron. Close in, at very short distances, you see more of the "bare" charge; far out, at macroscopic distances, virtual electron–positron pairs screen it. As you zoom out, α drifts upward — from 1/137.036 at zero momentum to roughly 1/128 at the Z boson mass. This is called vacuum polarisation, and on the lattice it has a beautiful combinatorial interpretation.

Where the BKT formula lives
αBKT = 1/137.032 is evaluated at the lattice UV scale: roughly the Compton wavelength Qlat ≈ 0.59 MeV, about one atomic spacing down from the bare vortex core. This is where the BKT power law has converged and the formula is exact.
Where CODATA measures
The measured 1/α = 137.035999 sits at the Thomson limit: zero momentum, long-wavelength Coulomb interactions between almost-stationary electrons. All of our atomic spectroscopy lives here. Between the two scales, α drifts — and the lattice's linked-cluster expansion computes by how much.

Feynman diagrams — on a crystal

In standard QED, vacuum polarisation at one loop is computed from a single Feynman diagram — an electron-positron bubble. Higher loops add more bubbles. On the coherence lattice, the analogous thing is a linked-cluster expansion: the sum of contributions from small connected subgraphs of the diamond lattice, each representing a different virtual K-fluctuation pattern that couples to the vortex. The first three subgraphs contribute most of the correction. Click each one to see what it is and how much it shifts 1/α:

Click a card above to read what physical process it captures and where the correction comes from.

The same subgraphs in 3D

The 2D schematics above flatten the tetrahedral geometry of diamond — real bonds meet at 109.47°, and higher-order subgraphs become interpenetrating tetrahedra that can only be understood in three dimensions. Below are four subgraphs of the expansion drawn as they actually sit in a diamond crystal. The first two are the ones the paper computes. The last two are the frontier: the next-order corrections we have not yet calculated. Drag to rotate any canvas; each one auto-rotates between drags.

Single-vertex star
One atom at the centre, four nearest-neighbour atoms at the tetrahedron corners. Every bond meets the others at the tetrahedral angle 109.47°. The lattice version of QED's four-fold vertex — the vortex couples to the K-field through all four bonds simultaneously.
order 1 contribution δc = 2.9747
Two-vertex dumbbell
Two adjacent atoms sharing a single bond (drawn pink), with each atom's three other neighbours fanning out. Two interpenetrating tetrahedra. The shared bond is the Markov channel across which a virtual K-fluctuation at one vertex couples to one at the neighbour.
order 2 contribution δc = 0.0111
Three-vertex path open
A central atom with two of its four neighbours occupied — a "V" of two bonds meeting at the tetrahedral angle. This is the first subgraph spanning three vertices. It would capture the next term in the combinatorial expansion: a virtual fluctuation that hops two bonds rather than one. Rough magnitude estimate is of order 10−4 ppb — invisible to current experiments, but a natural consistency check.
order 3 contribution not yet computed
Six-atom ring open
The smallest closed cycle in the diamond lattice — six atoms alternating A and B in a puckered "chair" hexagon, the same ring organic chemists call cyclohexane. The first topologically non-trivial subgraph: it encloses area, supports winding, and its contribution is the next qualitatively distinct correction to the vertex factor. Also not yet computed in our derivation.
order 6 contribution not yet computed

Convergence

Plotted as a log-scale collapse, the successive LCE layers drop the residual by nearly four orders of magnitude in two steps:

Residual |1/αtheory − 1/αCODATA| / 137.036 after each LCE layer, plotted on a log scale. The BKT formula alone overshoots by 29 ppm. Adding the single-vertex binomial correction brings it within 6.7 ppb — a factor of 4000 improvement. Adding the two-vertex dumbbell correction brings it to 1.5 ppb, comfortably below the experimental uncertainty on the electron's g-factor (±80 ppb, marked as the orange dashed line). At this point, the lattice's prediction is indistinguishable from measurement.

The honest flag — and why the result still stands

Open derivationThe dumbbell correction has one piece we have not yet derived from first principles: the embedding weight R0² / (z(z−1)). We selected it by matching the structure of the two-vortex linked-cluster integral and observing that it fits the measured α to 1.5 ppb; the paper does not yet contain a rigorous proof that this specific form is forced by lattice physics. Chapters 1 through 14 have no open derivations. This chapter has one.

That is the honest statement. Here is the other honest statement: the odds of a randomly chosen embedding-weight formula reproducing the observed convergence pattern — not just a single number, but a series whose successive layers each track to experimental precision (29 ppm, 45 ppb, 1.5 ppb, each scaling predictably with the lattice's natural ratios) — are astronomically small. Numerology produces single coincidences, not series that converge at a characteristic rate to within experimental error at every order. The LCE on the diamond lattice produces the latter.

And the weight is not a free parameter we tuned. The combinatorial meaning of R0² / (z(z−1)) is the amplitude for a virtual K-fluctuation to traverse a shared bond and return to its starting vertex — a specific physical process on the lattice with a clear derivation path. Alternative embedding forms like R0³ / z² correspond to different physical processes (longer hops, different combinatorial topologies) and they fit measurably worse.

What this isEmpirical observation with a compelling structure: a 15-piece derivation, 14 pieces rigorous, 1 piece matched to observation and tied to an explicit physical interpretation. The formal derivation of that one piece is open and being actively worked. That is how physics has always advanced. Newton did not derive the gravitational constant from first principles; he observed that a specific form fit the data at every scale he could measure. Maxwell did not derive the speed of light; he observed that his equations forced it to equal a previously measured optical constant. Rigour follows discovery; it is not the other way around.

The primary prediction of the paper — 1/α = 137.032 from the BKT formula alone, already 29 ppm from measurement with zero plausibility arguments — does not depend on the embedding-weight question at all. That result stands on its own, and no other lattice framework matches it. Closing the last 1.5 ppb requires the LCE; closing the LCE formally requires one more theorem. The theorem is on the queue. The result, meanwhile, predicts α to the precision of the world's best experiment.

Where this leaves us

We now have a complete theory of α: 13 pieces of lattice physics, 14 if you count dimensional selection, 15 if you count the LCE, all flowing from the single principle I ≥ 0. The prediction 1/α = 137.035998994 agrees with the measurement 137.035999084 to 1.5 parts per billion. For any remaining doubts about whether this is physics or numerology, one final consistency check: plug the lattice's α into the standard QED series for g−2 and compare to the electron's measured magnetic moment. That is the next and final derivation chapter.