Why the average cell has six sides

Cells in a sheet — an epithelium, a soap froth, a leaf surface, the cells of a seed head — average six sides. The standard explanation is that hexagons are the most efficient way to tile a plane, which is a true statement about a different question.

The mean is six because a tiling cannot average anything else.

The counting argument #

Euler’s formula for a connected planar graph says V − E + F = 2. In a tiling where three cells meet at each vertex — which is what soap films, and most epithelia, and a generic Voronoi diagram all do — the counting is short.

Each vertex has three edges meeting it, and each edge has two ends, so 3V = 2E. Each face has some number of sides, and each edge is shared by two faces, so the sum of all the side counts is 2E.

Substituting into Euler’s formula and letting the tiling get large, the mean number of sides per face goes to exactly six.

Nothing in that argument mentions area, perimeter, efficiency or preference. It is a counting identity, and any large tiling with three-way vertices satisfies it whatever the cells are doing.

What this rules out #

The efficiency argument says that of all the ways to divide a plane into equal-area cells, hexagons have the least total perimeter. That is true — it is the honeycomb theorem, proved by Hales in 1999 — and it is a statement about the optimal tiling.

The counting argument is a statement about every tiling. A sheet of cells that were not optimising anything at all, arranged at random, still averages six sides.

So the observation “epithelial cells average six sides” does not distinguish the two stories. It is consistent with cells minimising perimeter and equally consistent with cells doing nothing of the kind, because the mean is forced either way.

That is the kind of observation this site is most interested in: one that appears to support an explanation and is actually compatible with its negation.

What would distinguish them #

The spread, not the mean.

A tiling of cells that really were minimising perimeter under equal-area constraints would be nearly all hexagons, with a narrow distribution. A tiling produced by an indifferent process has a characteristic width — a substantial fraction of fives and sevens, some fours and eights.

Measured on a golden-angle head, six-sided cells are a clear majority and there is a real tail either side. The distribution’s width is the measurement that carries information, and it is the one that usually goes unreported in favour of the mean that cannot vary.

There is also a well-known empirical regularity in real tissue — Lewis’s law, that a cell’s area rises roughly linearly with its number of sides — which is a statement about the joint distribution and says considerably more than either the mean or the spread alone.

Why a patch cannot be all hexagons #

There is a sharper version of the same counting, and it explains the exceptions.

On a sphere — a closed surface rather than a plane — the same argument with Euler’s characteristic 2 instead of 0 forces a deficit: a closed three-way tiling must contain twelve pentagons, however many hexagons it has. That is why a football has twelve black patches whatever its size, why a fullerene has twelve pentagons, and why a spherical virus capsid does too.

On a flat patch the boundary plays a similar role, and the interior cannot be uniformly hexagonal while the boundary closes. The non-hexagons are not errors in the tiling; they are what the topology requires.

The general shape #

An observation that seems to confirm an efficiency story turns out to be forced by counting, and the thing that would actually test the story — the spread — is the thing nobody reports.

The same pattern appears throughout this subject. The spiral counts are quoted as a property of a flower when they are a property of an annulus. The golden angle is quoted as optimal packing when the packing measurement does not single it out. In each case a number that carries no information has been mistaken for evidence, and the informative quantity beside it has gone unmeasured.

Where the three-way condition comes from #

The counting argument assumes that three cells meet at each vertex, and that assumption is doing real work, so it is worth asking why it holds.

For soap films it is a theorem. Plateau’s laws say that three films meet along an edge at 120°, and four edges meet at a vertex at the tetrahedral angle — anything else is unstable and rearranges. A four-way junction in a froth is a transient that resolves within moments.

For epithelia it holds for a related reason. A cell junction where four cells meet is mechanically unstable under surface tension in the same way, and real tissue resolves such vertices by a T1 transition — the four-way vertex splits into two three-way vertices with a short edge between them, and which pair of cells ends up adjacent flips. That process is one of the main ways an epithelium rearranges, and it is directly observable in developing tissue.

For a Voronoi diagram of points in general position it holds automatically: three cells meet where three points are equidistant, and four would require four points on a common circle, which is a coincidence.

So the assumption is not a modelling convenience. In each of the three cases it follows from something — stability, mechanics, genericity — and the six follows from it.

The sphere, and the twelve pentagons #

The same counting on a closed surface gives a sharper and more famous result.

On a sphere, Euler’s formula reads V − E + F = 2 rather than 0, and running the same substitution through gives a deficit that cannot be avoided: a closed tiling with three-way vertices must contain, in the hexagon-and-pentagon case, exactly twelve pentagons, however many hexagons it has.

Twelve, always. A football has twelve black patches. Buckminsterfullerene has twelve pentagonal faces among its twenty hexagons, and a larger fullerene with hundreds of atoms still has exactly twelve. An icosahedral virus capsid has twelve pentameric vertices. A geodesic dome has twelve.

None of those is a design choice. The number is topological, it is the same twelve in every case, and it is the clearest possible demonstration that the shape of a tiling is constrained by counting before anything physical enters.

Lewis’s law, and what it would take to test the efficiency story #

If the mean is uninformative, what would distinguish an optimising tissue from an indifferent one?

The spread, first. A perimeter-minimising tiling under an equal-area constraint would be nearly all hexagons; an indifferent one has a characteristic width. Measuring the width and comparing with the width a random Voronoi tiling produces is a real test, and it is rarely reported.

Lewis’s law, second, and it is the more informative. Lewis observed in the 1920s that a cell’s area rises roughly linearly with its number of sides — a seven-sided cell is bigger than a five-sided one, by a predictable amount. That is a statement about the joint distribution of area and topology, and it constrains mechanism far more tightly than either marginal.

Aboav–Weaire, third: cells with many sides tend to be surrounded by cells with few. Another joint statement, and another one that a mean cannot capture.

This site computes the mean and the spread and stops there, because the joint statistics need more care than a build-time figure can give them. Saying so is the same discipline the shell morphospace uses, where the geometry is drawn and the census is not claimed.

What the counting rules out generally #

The pattern worth carrying is that a quantity forced by topology cannot be evidence for a mechanism.

The mean side count is six in a perimeter-minimising tiling, in a random Voronoi tiling, in a soap froth, and in an epithelium that is doing nothing in particular. Observing six therefore distinguishes none of them, and an argument that runs “cells average six sides, hexagons are efficient, therefore cells are optimising” has used an observation that would have come out the same way had the conclusion been false.

That shape recurs. The spiral counts are quoted as a property of a flower when they are a property of an annulus. The golden angle is quoted as optimal packing when no packing measurement at fixed size picks it out. In each case a number that carries no information has been read as evidence, and the informative quantity sitting beside it has gone unmeasured.

What this site measures and what it does not #

Concretely, on a generated head of seven hundred primordia.

The mean is 5.93 sides over five hundred and seventy-eight bounded cells, which is six within the sampling and is forced.

The distribution has six-sided cells as a clear majority with a real tail either side — fives and sevens in quantity, fours and eights present.

The exclusions matter and are stated: cells on the convex hull are unbounded, have no area and no well-defined side count, and are dropped. About a fifth of the cells go that way on a head of this size, and a statistic that quietly kept them would be measuring the boundary.

What is not measured is the joint distribution — area against side count, or neighbour count against neighbour’s neighbour count — which is where the informative content is. That is the gap this essay names rather than fills.

Why a Voronoi tiling is the right comparison #

One methodological note, since the cells here are Voronoi cells of a point set rather than a simulation of tissue.

A Voronoi tiling assigns each point the region closer to it than to any other. For a set of cells that grew outward from initiation sites at equal rates, that is exactly the right model — the boundary between two cells is where they met, and they met halfway.

Real tissue departs from it in known ways: cells have different growth rates, surface tension straightens the boundaries, and division and rearrangement move things afterwards. But the departure is a correction to a sensible baseline rather than a reason the baseline is wrong, and the six that the counting forces survives every one of the corrections, because it survives any three-way tiling at all.

That is the useful property of a topological result. It does not care what the cells are made of.

The argument in one line, and the argument it replaces #

Six is the mean because counting edges two ways in a three-way tiling forces it.

The argument it replaces is that hexagons tile the plane with the least perimeter, which is true, was proved by Hales in 1999 after being conjectured for two millennia, and is about a different question — the optimal tiling rather than every tiling.

Both are worth knowing. Only one of them explains the observation, and it is the boring one.

Why the boring answer is the useful one #

Because it makes a sharper prediction.

The efficiency story predicts that cells should be hexagonal, and is therefore embarrassed by the fives and sevens, which have to be explained away as noise or as imperfection.

The counting story predicts the mean exactly and predicts that the exceptions are necessary — a finite patch with three-way vertices cannot be all hexagons and still close up, and a closed surface must contain exactly twelve pentagons. The exceptions stop being defects and become a consequence.

That is generally the sign of the better explanation: it accounts for what the other one has to apologise for.

A note on what “average” means here #

One technical point, since the result is about a mean and means can hide things.

The six is an asymptotic mean — it is what the average tends to as the tiling grows, and a small patch will not give exactly six because the boundary is a significant fraction of it. On the head measured here, five hundred and seventy-eight bounded cells give 5.93, and the shortfall is the boundary’s influence rather than a departure from the theorem.

It is also a mean over cells, not over vertices or edges, and the three countings give different-looking statements about the same tiling. A reader comparing two published figures should check which is being averaged.

And it assumes the tiling is simply connected — no holes. A tissue with a hole in it, or a tiling on a torus, gives a different Euler characteristic and therefore a different mean, which is the same reason a sphere forces twelve pentagons while a plane forces none.

The same counting in three subjects #

The argument in this essay is topology, so it applies wherever a three-way tiling does, and the three cases are worth naming together because they are usually met separately.

Soap froth. Plateau’s laws force three-way junctions, so the mean bubble has six sides in two dimensions. A froth coarsens over time — small bubbles vanish, large ones grow — and the mean stays at six throughout, because the topology does not care about the coarsening.

Epithelia. Cell junctions resolve four-way vertices into pairs of three-way ones, so the mean cell has six sides. This holds across tissues, species and developmental stages, which is exactly what one expects of a constraint that comes from counting rather than from biology.

Any Voronoi tiling of generic points. Three cells meet where three points are equidistant, so the mean is six. That includes the seed heads this site draws, and it includes point sets with no structure at all.

The same number in three places with no shared cause is the signature of a topological result. Reading it as a shared mechanism — as convergent evolution toward efficiency, say — would be the error this essay is about, three times over.

What to ask when a number turns up everywhere #

The habit this essay is really about is worth stating separately from the tiling.

When the same value appears across unrelated systems, there are two explanations and they call for different work. Either a shared mechanism is producing it, in which case the systems should share something else too and the prediction is testable. Or a constraint is producing it — counting, conservation, topology — in which case nothing about the systems matters and looking for a shared cause is wasted effort.

Six sides is the second kind. So, in its way, is the mean spacing in a packed head, which is fixed by area and not by the angle. Fibonacci counts are the first kind — they come from the divergence angle, and a head on another branch gives a different sequence.

Telling the two apart before reaching for an explanation is most of the work, and the test is cheap: ask what would have to change for the number to be different. If the answer is “nothing available”, it is a constraint.

There is a corollary worth stating, because it decides what a measurement should report. A constraint carries no information about the specimen, so the statistic to report is the one the constraint does not fix — and for a tiling that is the spread of the side counts rather than their mean, which separates a lattice from a random scatter by a factor of eighty on heads of nine hundred organs whose means agree to one per cent, and by more on larger heads.

That is a general instruction rather than a remark about tilings. Wherever a number turns out to be forced, the next question is which quantity the forcing leaves free, and that quantity is where the measurement was all along. It is usually one step further into the same distribution, and it usually costs nothing extra to compute, because the data it needs has already been collected to produce the number that turned out to say nothing.