Past two, the list does not stop

Seventeen dollars prices the features an orbifold can have — a handle two, a cross-cap one, a mirror boundary one, a cone point of order n almost one — and enumerates the ways of spending exactly two. There are seventeen, and they are the wallpaper groups, arriving from an accounting identity that has never heard of a lattice.

Before the lattice has a say runs the same sum below two, where the total is 2 − 2/N and the answer is a finite group of order N acting on a sphere.

This essay runs it above two, where the answer changes character completely: the surface is the hyperbolic plane, and the list is infinite. That much is the sentence the theorem is usually left at. It hides the two facts worth having.

The excess is quantised, and the smallest one is 1/84 #

Nothing costs 2.001. Nothing costs 2.0001. The cheapest thing past two costs 2 + 1/84, and there is nothing at all between it and two.

That is not an artefact of a search bound. Every cost is a sum of the coins, and every coin is a fraction with a small denominator: a cone of order n costs (n − 1)/n and a corner costs (n − 1)/2n. A total is therefore a rational number with a denominator dividing a product of small integers, and the least such number strictly greater than two is 169/84.

The signature that costs it is *732 — a mirror boundary with corners of orders 7, 3 and 2, which is the triangle group drawn above. The cheapest signature with no mirrors is 732 at 2 + 1/42, and the two numbers 84 and 42 are the ones that turn up in the theorem this essay is really about.

Why the smallest excess is Hurwitz’s number #

The excess is not a bookkeeping quantity. By the Gauss–Bonnet theorem the area of an orbifold is 2π times its excess over two, so the signature costing 2 + 1/84 names the smallest hyperbolic orbifold there is: a triangle of area 2π/84 — or π/42 for the orientation-preserving group.

Now take a compact surface of genus g with g ≥ 2. Its area, at curvature −1, is 4π(g − 1). A group of automorphisms acting on it has a quotient orbifold, and the order of the group is the surface’s area divided by the orbifold’s. The largest possible order is therefore the surface’s area divided by the smallest orbifold area:

$$|G| \;\le\; \frac{4\pi(g-1)}{2\pi/42} \;=\; 84(g-1).$$

That is Hurwitz’s bound, published in 1893, and the 84 in it is the reciprocal of the excess computed above. A theorem about the automorphisms of Riemann surfaces and the classification of wallpaper patterns are the same accounting, run at different totals.

The bound is attained — Klein’s quartic curve, of genus 3, has exactly 168 automorphisms — and the groups that attain it are called Hurwitz groups. Every one of them is a quotient of the (2, 3, 7) triangle group, which is the group in the first figure.

Three geometries, one inequality #

The triangle groups are the cleanest place to see the trichotomy, because one inequality decides everything.

Three mirrors meeting at angles π/p, π/q and π/r generate a group whose fundamental domain is that triangle, and the sign of 1/p + 1/q + 1/r − 1 says which surface the triangle lives on. Positive means the angles sum to more than π, which happens on a sphere; zero means exactly π, which is the flat plane; negative means less, which is the hyperbolic plane.

Exactly three triples give zero: (2, 3, 6), (2, 4, 4) and (3, 3, 3). They are p6m, p4m and p3m1 — the three plane groups generated by reflections in a triangle, and the only flat triangle groups there are. Everything else on the list is either a finite group or a hyperbolic one, and the flat case is the boundary between them rather than a region of its own.

That is worth setting beside the seventeen. Of the seventeen wallpaper groups only three are triangle groups; the other fourteen have signatures that are not of the form *pqr. But every one of the seventeen costs exactly two, and the knife edge is the same knife edge.

The count that is not a count #

Above two the enumeration is not a classification any more, and the natural way to say so is to count. Allow an excess of a twelfth and see how many signatures there are; allow a sixth and see how many more.

Those numbers do not exist, and this essay quoted three of them. A search for signatures has to stop somewhere, and where it stops is a bound on the orders of the cone points and corners — twelve, or fourteen, or thirty. Move that bound and the answer moves with it: below an excess of a twelfth the search finds twenty-seven signatures if it stops at order twelve, twenty-nine if it stops at fourteen, thirty-three at eighteen and thirty-nine at twenty-four. There is no value it is approaching.

The reason is a single family, and setting it out makes the whole shape of the list clear.

Take the signature *n32: one mirror boundary, and three corners, of orders n, three and two. Its cost is 1 for the boundary, (n − 1)/2n for the first corner, 1/3 for the second and 1/4 for the third, and those add to (2 + 1/12) − 1/2n. At n = 7 that is 2 + 1/84, the cheapest signature there is. At n = 8 it is 2 + 1/48. At n = 100 it is a whisker under 2 + 1/12, and it is still under it at n = 10,000.

So there are infinitely many signatures costing less than 2 + 1/12, and any count of them is a count of how far the search was allowed to run. The same is true of every bound at or above a twelfth, for the same reason.

Below a twelfth the list is genuinely finite, and the bound the search needs can be computed rather than guessed. The family runs out where (2 + 1/12) − 1/2n stops being under 2 + ε, which is at n = 1/(2(1/12 − ε)). At ε = 1/20 that is fifteen, and a search allowed orders to fifteen or beyond returns fourteen signatures every time. At ε = 1/84 — the bound that settles the gap this essay opens with — it is exactly seven, which is why so short a search can prove that nothing at all lies between two and 2 + 1/84.

And the row between the two behaviours is the one worth staring at. At ε = 1/13, which is below a twelfth, the list is finite and has seventy-one members. A search stopping at order twenty-six reports thirty-nine of them and has no way of knowing it is short, because thirty-nine is what it found and finding is all a search does. The required bound there is seventy-eight, and the only reason that number is available is that the family causing the trouble is known by name.

The flat case has seventeen and there is nothing to tabulate, which is the whole difference. At exactly two the coins have to add up precisely, and the number of ways of doing that is small — small enough that the same widening test settles it in one line. At more than two the constraint is an inequality rather than an equality, and inequalities have room.

This is the sense in which the plane’s classification is a small miracle. It is not that the plane admits few patterns; it is that a knife-edge condition admits few solutions, and the plane is where the condition is an equality. One notch off the knife edge, at a twelfth, and the list is already infinite.

How the drawing is made, and what it is not #

The tilings are generated rather than drawn, by the same rule every pattern on this site follows: start with one fundamental triangle and reflect it in its own sides until nothing new appears within the drawing.

What differs is the model. Reflection in a straight line becomes inversion in a circle orthogonal to the disc’s boundary, which is what a hyperbolic mirror is in the Poincaré model, and the triangle’s vertices are placed at the hyperbolic distances the law of cosines gives, mapped into the disc by tanh(d/2). The geodesics are arcs, drawn as sampled polylines rather than as SVG arcs — a circle through two points orthogonal to the boundary has its centre outside the disc and a sweep that is easy to take the wrong way round, and sampling has no such failure mode.

Two honest limits sit on every one of these pictures, and they are why this essay’s claims are all about signatures rather than about drawings.

Nothing here is decidable in the way a plane pattern is. The vertices come from cosh and tanh, the geodesic centres are solved numerically, and there is no integer arithmetic anywhere. The round trip this site runs on every wallpaper figure — generate a pattern, forget the group, rediscover it from the point set — has no counterpart here, because the point set is not a lattice orbit and there is no finite detector to run.

The edge of the drawing is an artefact. A hyperbolic tiling has infinitely many tiles inside a bounded disc, crowding towards the boundary; the picture stops at a chosen depth. The growth is slow — about 1.2 tiles per step rather than 3 × 2ᵈ — because the reflections have order two and the triangle group has short relations, so a drawing that expected exponential growth would look as though the generator had failed.

What is exact, and it is the whole argument #

The arithmetic of costs is exact rational arithmetic, and every claim this essay makes is a claim about it.

The gap is exact. 169/84 is the least sum of coins exceeding two, and the enumeration that finds it searches every combination of features within stated bounds — with the bounds widened and the answer required not to change, which is the check any bounded search on this site owes.

The trichotomy is exact. The sign of 1/p + 1/q + 1/r − 1 is the sign of an integer, so no triangle group is ever ambiguously placed.

The correspondence between signatures and groups is not verified here, and that is the honest gap. For the seventeen, this site derives each group’s signature from its own operations and requires the two lists to match — a bijection checked in both directions. Above two there is no comparable computation, because there is no finite operation set to derive a signature from. The hyperbolic signatures are the output of an accounting identity, and the statement that each names a group is a theorem quoted rather than a fact computed.

What a hyperbolic pattern looks like to a crystallographer #

There is a reason this rung sits on the wallpaper ladder rather than off the end of it, and it is not only that the arithmetic is shared.

A hyperbolic group has no lattice, and the restriction does not apply to it. Cone points of order 5, 7, 11 and 13 are ordinary in the list above — *732 has one of order seven — and none of them is available to a plane pattern. The crystallographic restriction is a statement about integer matrices, and there are no integer matrices here because there is no lattice for them to act on. So the hyperbolic list is where the forbidden orders live, and looking at it is the clearest way of seeing that the restriction is a fact about periodicity and not about symmetry.

And the flat case is not simply the boundary; it is where the two constraints coincide. Below two, the cost decides the answer and there is no lattice; above two, the same. At exactly two the orbifold is flat, a lattice appears, and the permitted cone orders drop to 2, 3, 4 and 6 — which the accounting produces on its own, since the seventeen signatures use no other orders. Two independent arguments hand back the same five numbers: the trace of an integer matrix, and a sum of fractions adding to two.

Why the search for the cheapest is finite #

The claim that nothing costs between two and 169/84 is an assertion about an infinite list, and it is worth saying why a finite computation settles it.

The excess is a sum of rationals whose denominators are the orders of the cone points and corners, and the key observation is that a feature made cheaper by raising its order is bounded below by the price of leaving it out. A cone point of order n costs (n − 1)/n, which climbs towards one and never reaches it; a corner of order n costs half that. So adding a feature never lowers the total, and raising an order never lowers it either — the cost is monotone in both directions.

That bounds the number of features: each one costs at least a quarter, so a signature costing under 2 + ε carries at most a handful of them, and there are finitely many ways of arranging a handful.

It does not bound the orders, and the natural next sentence is false. An order high enough does not overshoot ε on its own, because a feature’s price is capped: a cone costs less than one however large its order, and a corner less than a half. So raising an order does raise the cost, but only towards a ceiling — and whether it ever crosses ε depends on where that ceiling is. For *n32 the ceiling is 2 + 1/12, so the family crosses any ε below a twelfth and no ε at or above it. That is the whole content of the section above, arriving from the arithmetic rather than from the table.

For the 1/84 claim the ceiling is crossed almost at once: the search needs cone and corner orders no higher than seven, and widening it past seven changes nothing. An enumeration whose answer stops moving when its bounds are widened is one whose bounds were not doing the work, and here it is possible to say in advance exactly where the moving stops.

The underlying number is older than the orbifold statement. The least positive value of 1 − (1/p + 1/q + 1/r) over integers is 1/42, attained at (2, 3, 7) and at nothing else, which is a small exercise in bounding one variable at a time. A mirror boundary halves every price, so the orbifold *732 costs half of that above two — 1/84 — and the same triple runs the other two geometries with the sign of the same expression deciding which.

What attains the bound, and what does not #

Hurwitz’s inequality is sharp, and the groups attaining it are sparse enough to be listed rather than described.

Genus 3. Klein’s quartic curve has 168 automorphisms, which is exactly 84(3 − 1), and the group is PSL(2, 7) — the same group that appears as the symmetries of the Fano plane, and one of the smallest simple groups that is not of prime order.

Genus 7. The next Hurwitz group is PSL(2, 8) of order 504, and 504 is 84(7 − 1).

Genus 2 attains nothing. The bound permits 84, and the largest automorphism group of a genus-two surface has order 48. So the inequality is not attained at every genus, and the genera at which it is attained form an irregular set with no simple description — which is the usual state of affairs once a clean bound meets the arithmetic of actual groups.

The area reading is the same statement with the units changed, and it is worth carrying because it is what the picture shows. A hyperbolic triangle with angles π/p, π/q and π/r has area π(1 − 1/p − 1/q − 1/r), so the minimal excess is a minimal area: the (2, 3, 7) triangle, at π/42, is the smallest fundamental domain any group acting on the hyperbolic plane can have. Carl Ludwig Siegel proved that in 1945, and it says why the tiling above looks the way it does — the triangles are as small as hyperbolic triangles are permitted to be, which is why so many of them crowd into the disc.

And it is the exact analogue of the statement the flat case makes, which is worth setting beside the seventeen. There, every group has a fundamental domain of the same area as every other once the lattice is fixed, and the classification is finite. Here the areas are quantised, bounded below, and unbounded above, and the list is infinite. Same coins, same total, and a different kind of answer on each side of two.

Who found it, and when #

The pieces arrived from three directions and were joined late.

Hurwitz proved his bound in 1893, working on Riemann surfaces, with no orbifolds and no wallpaper groups anywhere in the argument. Poincaré and Klein had constructed hyperbolic tilings in the 1880s, and Klein’s quartic — the genus-3 surface with 168 automorphisms — was published in 1878, fifteen years before the bound it attains.

Thurston introduced orbifolds in their modern form in the 1970s, in the course of the geometrisation programme, and it is the orbifold that makes all three statements one statement: the Euler characteristic of a quotient is the characteristic of the surface divided by the order of the group, and every count above is that identity with the pieces named differently.

Conway’s magic theorem, which is the version this site uses, is the same content stated so that it can be done by hand — the coins, the total, the seventeen. His notation was published in the 1990s and the whole argument appears in The Symmetries of Things in 2008. What it buys is that the sphere, the plane and the hyperbolic plane are one calculation with one number changed, which is the only reason this essay is a rung of the wallpaper ladder rather than a different subject.

What the seventeen have that the infinite list does not #

One more comparison is worth making, because it says what is special about the flat case beyond the arithmetic.

The seventeen are realised by patterns anybody can draw. Each has a lattice, a repeating cell, and a finite description: a group of operations modulo translations, and a motif. Every wallpaper figure on this site is generated from that description and handed back to a detector that rediscovers the group from the drawing.

The hyperbolic groups have none of that. There is no lattice, so there is no cell and no finite quotient to enumerate operations in; the group is infinite modulo nothing. A picture of one is a truncation of an infinite object, and the group cannot be recovered from a bounded patch by any finite procedure of the kind the plane admits.

So the difference between two dollars and two dollars and a penny is not a difference of degree. At exactly two the classification becomes decidable, and that is the property this whole site is built on. The accounting identity says the flat case is a knife edge; the machinery says the knife edge is where the arithmetic can be checked.

It is worth restating what the three geometries have in common, because the coins are the only thing they share. Below two the answer is a finite group on a sphere and the restriction never enters, since a sphere has no lattice to constrain. At two the answer is the seventeen, where the restriction is exactly what makes the list finite, and where two of the groups fold the plane into a surface with nothing marked on it. Above two the restriction is silent again, for the same reason as below: an orbifold signature is arithmetic about cone points and boundaries, and only the flat case has a lattice for those cone points to be compatible with.

Where the ladder goes next #

This rung closes the trichotomy: the same accounting has now been run below two, at two and above two, and the three answers are a finite list of families, a finite list of seventeen, and an infinite list.

Two directions lead off it, and only one of them belongs to this site.

The one that does not is the hyperbolic classification itself — Fuchsian groups, their signatures, and the moduli of the surfaces they uniformise. It is a large and beautiful subject and nothing in it is about crystals.

The one that does is what the trichotomy says about the plane’s list being closed. The seventeen are complete because the equality is a knife edge; the two hundred and thirty are complete for a related reason, and Bieberbach’s theorems make it precise — in every dimension there are finitely many crystallographic groups, and the finiteness is exactly the statement that the flat case has no room. That argument is one rung up from here and one dimension along, and it is the last structural thing the classification has to say for itself.