The orientation is a policy

Eleven rungs of this anchor measure what a cell system costs. The area it cannot keep, the shape it trades for it, the twelve pentagons no tiling of the sphere escapes, the query that is a disc against cells that are not. Every one of those measurements takes the base polyhedron’s orientation as given.

It is not given. It is three angles, chosen by whoever designed the system, stated in the specification as a fact about the world, and scored by nobody. Every published discrete global grid has one, and the reasoning behind it is always editorial: put the vertices in the ocean, keep the awkward cells away from land.

What it cannot do #

The first thing to settle is that the orientation is not a design variable, and the argument is one line.

Rotating the solid rotates every cell rigidly. Every quantity this anchor measures a cell by — its area, its aspect, its perimeter, its distance from its neighbours — is defined by lengths on the sphere, and lengths on the sphere are invariant under rotation. So the multiset of cell qualities is exactly the same for every orientation there is: the same 384 areas at level three, in the same order, to the last bit the arithmetic carries.

That is stronger than saying the orientation is a minor effect. There is no orientation that improves a cell system, in the sense of any quantity computed from the cells alone, because such a quantity cannot see a rotation. The tiling’s area spread is 1.3263 and its worst aspect 1.2908 whichever way the solid is turned, and a comparison of two cell systems that differ only in orientation has nothing to compare.

There is a second reading of that invariance which is worth having, because it settles a question the anchor has left open. The same number of cells, in two shapes compares tilings by the distributions they produce; a cell system trades area for shape prices the trade between them. Both are comparisons of distributions, and both are therefore comparisons that the orientation cannot enter. So every result in this anchor about which scheme is better survives the orientation exactly, which is a stronger statement than any of those essays makes for itself and is worth knowing before reading them.

What it does do #

The reason anybody argues about it is that the sphere is not uniformly interesting. The cells are the same cells; the things standing on them are not.

Scored on eight stated cities, the choice is worth a great deal. At level three the spread of cell area among them runs from 1.0006 under the best orientation to 1.3865 under the worst — a swing of 39 per cent, on a tiling whose own spread is 1.3263 and never moves. The mean cell aspect the eight receive runs from 1.1265 to 1.2908.

The usual orientation — the cube with its vertices on the coordinate axes, which is what a tiling built from the obvious construction inherits — gives these eight a spread of 1.2587. That is near the bad end of the range, and it was not chosen; it is what happens when nobody chooses.

The surface is also rougher than a designer would expect, and the roughness is structural rather than numerical. A place does not gradually acquire a worse cell as the solid turns; it sits in one cell, then in another, and the two may be very different. So the score is a step function of the angles with steps a cell wide, which means an orientation cannot be tuned — a small adjustment either does nothing at all or does something large — and it means a specification quoting an orientation to four decimal places is quoting three of them for nothing.

The rule of thumb, and what it is worth #

Every designer’s account of an orientation is about vertices. The vertices carry the whole of the polyhedron’s angle deficit, they are where a Goldberg tiling’s twelve pentagons sit, and the globe on a solid shows the distortion piling up around them. So the rule is: keep the vertices away from what matters.

Measured over a ten-degree grid the correlation between distance from the nearest vertex and local cell aspect is −0.32. It has the right sign and explains a tenth of the variance. The reason is that a cube’s cells are worst at the corners of each face, which are the vertices, and also systematically different along each face’s diagonals and edges — so the field of cell quality has the symmetry of the solid rather than of eight isolated points, and a rule that mentions only the eight points is describing a small part of it.

The practical consequence is that the orientation cannot be chosen by the rule. It has to be searched, against a stated set of places, which is one loop over three angles and a lookup — and which nobody has published, because until the objective is written down there is nothing to search for.

The three angles are not equivalent #

A rotation of the sphere has three degrees of freedom, and on a solid with symmetry they are not three independent choices.

The cube’s rotation group has twenty-four elements, so the space of genuinely distinct orientations is the full rotation group divided by that — a region a twenty-fourth the size, and any sweep over all three angles visits every distinct orientation twenty-four times. That is why the sweep above fixes the third angle at zero: with the first two free it already reaches every distinct placement of the vertex set, and the third would only relabel which cell is which within a face.

The symmetry also bounds how good a placement can be. The eight vertices are spread as evenly as eight points on a sphere can be from a cube, so the largest possible distance from any point to the nearest vertex is fixed at 54.7° — the angle from a face centre to a corner. No orientation can put every place far from a vertex, because the vertices cover the sphere at that spacing whatever is done to them. The choice is entirely about which places get the good ground, and there is exactly as much good ground under every orientation.

That is the anchor’s own twelve pentagons argument arriving at the orientation. The pentagons cannot be removed and cannot be reduced in number; all that can be decided is where they go. The same holds for every vertex of every solid, and it is the reason the choice is a distributional one rather than an engineering one.

Refining does not fix it #

The obvious escape is resolution. A finer tiling has smaller cells, so surely the difference between them stops mattering.

It does not shrink. At level five, with 6,144 cells, the eight cities’ area spread runs from 1.0460 to 1.4965 — a swing of 1.43, larger than at level two. Refinement scales every cell down together and leaves the ratios exactly where they were, because those ratios are properties of the warped cube’s geometry rather than of the cell size.

So the choice is permanent. A cell system’s orientation is fixed at the moment the identifier scheme is published, because an address is an area and changing the orientation changes every address there has ever been. It is a decision made once, for the life of the system, on grounds nobody records, and it cannot be revisited by refining.

What it is worth against what it costs #

The other half of any design question is what the alternative costs, and here the answer is unusually clean: nothing.

Choosing a good orientation is a search over three angles against a stated objective, evaluated by a table lookup. The whole sweep behind the figures above is a few hundred milliseconds. There is no construction to redo, no identifier scheme to change, no property of the tiling to give up, and no trade against any other quantity — because the orientation is not trading against anything, which is the whole point of the invariance.

Compare that with the two levers this anchor has spent eleven rungs on. Changing the warp trades area against shape and every essay in the anchor is about the terms of that trade. Changing the resolution costs storage and identifier length in proportion. Changing the base solid changes the pentagon count, the face shape and every address. The orientation costs nothing at all and is worth as much as a level of refinement to the places that benefit — and it is the one lever nobody scores.

That asymmetry is the actual finding. The free parameter is the one being left to habit, and it is free precisely because it is invisible to every measurement the field makes.

What was computed, and how #

The tiling is the tangent-warped cube this anchor has used since its third rung, at levels two to five. Every cell’s area is computed by spherical polygon integration on a densified ring rather than by a planar formula, which is what a cell system trades area for shape established the anchor’s numbers on.

An orientation is applied by rotating the place by the inverse rather than the solid by the rotation, which is the same thing and needs no rebuild of the tiling — so a sweep over 300 orientations is 300 lookups rather than 300 constructions. The cell containing a place is found by nearest centre, with the centres built once per level and cached; doing it the obvious way rebuilt the tiling for every query and turned a sweep into minutes.

The eight places are cities on six continents, chosen for spread rather than for population, and the objective is deliberately the crudest one that is not degenerate. The worst aspect any of them suffers is not usable at coarse resolution: at level three it is the tiling’s global maximum of 1.2908 for every orientation tested, because with eight places and 384 cells one of them always lands in a corner cell. The spread of area among them separates the orientations at every level, so that is what is scored.

The assertions require five things separately: that the sorted list of cell areas be identical under rotation to the last bit; that a stated place’s own cell change when the solid is turned; that the best and worst orientations differ by more than fifteen per cent on the objective; that the mean aspect move too; and that distance from a vertex be a poor predictor of local cell quality — which is the negative result that makes the search necessary rather than replaceable by the rule.

Where the model stops #

Eight cities are not a population. Weighting by where people actually are would change which orientation wins and would probably widen the swing, since population is far more concentrated than eight scattered points. The measurement here is a lower bound on what the choice is worth, and the honest version needs a dataset this collection has deliberately not taken on.

The cube is the tractable case. Real systems use icosahedra, where there are twelve vertices instead of eight and the faces are triangles, so the field of cell quality has a different symmetry. Nothing in the argument depends on which solid it is — the invariance is a fact about rotations — but the size of the swing is a fact about the cube.

Nearest centre is not exactly containment. A place is assigned to the cell whose centre is closest in longitude and latitude, which agrees with true containment almost everywhere and can differ for a place near a cell boundary — and near a boundary the two candidate cells are neighbours with similar quality, so the score barely moves. A containment test against the cell’s own edges would be exact and would cost a clip per query; the difference was checked on the eight cities at every level and changed no assignment.

And an orientation is chosen once and used for decades. The objective scored here is a snapshot of where the places of interest are today, and a global grid is an addressing scheme: the identifiers it issues outlive the reasoning that placed it, so a rotation optimised for one set of cities is inherited by every later user with no record of whose interests it encoded. That is an argument for publishing the objective rather than against having one — an orientation left to habit ages exactly the same way and cannot even be re-derived afterwards.

And the search is over a grid of orientations. Twenty-four steps in the first angle and thirteen in the second is coarse, and a finer sweep finds better optima: at level three a twelve-step sweep reports a best of 1.0006 and a coarser one misses it. What is measured is therefore a lower bound on what a careful choice would achieve, which points the same way as everything else here.

The generalisation #

The rule is that a symmetry of the construction is a degree of freedom in the design, and a degree of freedom that changes no measurable property of the thing being built is exactly the one nobody scores.

The cell system’s intrinsic properties are rotation invariants, so the orientation is invisible to every quantity a specification reports. That is precisely why it goes unstated: there is nothing to state it against. What it changes is the relationship between the construction and the world, which is not a property of the construction and is the only thing anybody cares about.

This collection has met the same shape before from the other end. What survives a change of coordinates argues that only the invariants are worth reporting about a projection, because everything else is an artefact of the graticule. That argument is right and it has a boundary: a quantity that is invariant under a transformation says nothing whatever about how the object is placed, and placement is a real decision with real consequences whenever the thing being mapped is not uniform.

The habit that follows: when a specification names a constant with no derivation — an origin, an epoch, an orientation, a starting meridian — ask what it changes. If the answer is nothing about the construction, ask what it changes about the user, and expect the answer to be the whole of why it is there.

Who found it, and when #

The invariance is trivial and has never been in doubt; nobody has claimed a better orientation improves a tiling. What the literature does contain is a long series of orientation choices presented as design decisions with editorial justifications: Fuller’s icosahedral orientation of 1943 was chosen to put every vertex in water, and the modern discrete global grid specifications inherit that reasoning and state their own rotations to several decimal places without a score attached.

The reason nothing is scored is that the objective needs a set of places, and a set of places is a political object rather than a geometrical one. Choosing to optimise for eight cities on six continents, or for population, or for land area, or for the interests of whoever is paying, is a choice about whose data is well gridded — and a specification that stated its objective would have to state that too.

Which is the useful thing to say about it. The orientation of a global grid is not a technical parameter that happens to be unexamined; it is a distributional decision wearing a technical parameter’s clothes. It costs nothing to publish the objective alongside the angles, and doing so turns an aesthetic preference into a claim somebody can check.

Where the ladder goes next #

Twelve rungs have measured the cells: their areas, their shapes, their hierarchy and now their placement. What none has measured is the identifier, which is the thing a user actually handles — a number with a length, carrying a resolution it does not state, and truncating in a way that is a geometric operation rather than an arithmetic one.