Impossible in two derivatives, possible in one

The argument this collection is built on is four lines long. Gaussian curvature is intrinsic, by the Theorema Egregium. A sphere has curvature 1/R² and a plane has zero. An isometry preserves curvature. Therefore no isometry exists, and no map is faithful.

Every step of that is correct. What is worth noticing is what the first step computes: curvature is built from second derivatives, on both of the routes this site uses — from the second fundamental form on one and from derivatives of the metric on the other. So the theorem is a statement about maps that have two derivatives, and it is silent about maps that have fewer.

Take one away and the theorem is not merely silent. It is false.

The mechanism, in one dimension #

Take a segment and shrink it by a factor λ < 1. It is now too short. Add a corrugation across its own direction:

$$f(t) = \left(\lambda t,\; \frac{a}{N}\sin(Nt)\right), \qquad t \in [0, 2\pi].$$

Its speed is √(λ² + a²cos²(Nt)), which does not depend on N at all. So the length the corrugation restores is set entirely by the amplitude-to-slope ratio, and the number of wiggles is free.

That is the whole trick, and it is worth saying twice because everything follows from it. a buys length. N buys closeness. They are independent.

Solve for the a that makes the length exactly right — for λ = 0.9 it is 0.6241 — and every member of the family f_N is an exact isometry of the original segment onto its image, at every N. Meanwhile the image stays within a/N of the shortened straight line, so as N grows the family converges uniformly to a map that is emphatically not an isometry: it is the map that shrinks everything by a tenth.

What converges and what does not #

The measurements over six wiggle counts:

• the length is exact at every one, to 10⁻⁹, which is the quadrature’s own precision;
• the deviation from the shortened straight line falls exactly as 1/N, from 0.312 at N = 2 to 0.0098 at N = 64;
• the tangent’s departure is 34.74° at N = 2 and 34.74° at N = 64;
• the curvature rises from 1.25 to 39.94, in proportion to N.

The third of those is why one corrugation is not enough. Uniform convergence — the first and second lines — is convergence in C⁰, and a C⁰ limit of isometries need not be an isometry, which is exactly what the family demonstrates. To have a C¹ limit the derivative must converge too, and here it does not: the tangent oscillates through ±34.74° however fine the corrugation.

That is not a failure of the construction. It is the reason the construction is iterated.

The iterated version, where the derivative does converge #

Recover only a fraction θ of the missing length at each stage. Then the amplitude needed at stage k falls geometrically, and so does the angle each stage tilts the tangent through — which means the derivative does converge, because the sum of the tilts converges.

The wiggle count is free at every stage, so it can be chosen as large as the closeness demands. Over eight stages, taking back half the deficit each time and multiplying the wiggle count by six:

Three columns, three fates. The deficit goes to zero: the limit is an isometry. The tilts go to zero and their sum converges: the limit has a continuous derivative. The curvature runs to infinity: the limit has no second derivative at all.

Which is exactly the gap the impossibility theorem lives in. There is no contradiction, because the theorem’s hypothesis is not met — and the theorem’s hypothesis turns out to be doing far more work than the way it is usually stated suggests.

What the sphere would have to make up #

The construction above is one-dimensional and the theorem is two-dimensional, so it is worth being explicit about what the real object has to do.

A spherical cap of angular radius 30° has a boundary 4.51 per cent shorter than flat paper of the same radius. Stated as an angle rather than a length that is a cone deficit of 16.23°, and stated as a curvature it is a total Gaussian curvature of 0.8418 — three routes to one number, and the same number the angular excess of a triangle reports and the total curvature of a region integrates.

The C¹ flattening does not remove that deficit. It hides it in corrugations too fine to see and too rough to differentiate twice, which is the honest description of what Nash and Kuiper’s maps are: isometric, continuous in their first derivatives, and possessing no surface normal that varies continuously.

What “no second derivative” means for a surface #

It is worth pinning down what the limit object actually lacks, because “not twice differentiable” sounds like a technicality about smoothness classes and it is not.

A surface’s second derivatives are what give it a shape in the ordinary sense: a normal direction that varies continuously, a curvature at each point, a tangent plane that the surface hugs to second order. The limit of the iterated corrugation has a tangent plane at every point — that is what C¹ means — and the tangent plane’s orientation is continuous. What it does not have is any control on how fast that orientation turns: the curvature at every stage is larger than at the last, without limit.

So the object is smooth enough to have a direction everywhere and rough enough to have no curvature anywhere. That is a strange combination and it is exactly the combination the theorem’s hypothesis excludes.

For a reader who wants a physical picture: it is a sheet of paper that has been crumpled at every scale simultaneously, so that any patch of it, however small, contains crumples. The lengths measured along it are the sphere’s; the directions are continuous; and no part of it is flat enough to be a page.

The refusal: the smooth case is untouched #

The paired assertion is what makes the result a measurement rather than a rhetorical trick. With two derivatives available, the obstruction is intact and this site’s own machinery says so: the unit sphere’s curvature computed from its metric alone comes out at 0.9999967, which is the differencing step’s own error, and the cap deficit is positive at every angular radius.

So the pair is:

• with two derivatives, no isometric flattening exists, and the curvature computed intrinsically proves it;
• with one, an isometric flattening exists, and a construction produces one.

Both halves are needed. Without the second the theorem looks like a fact about geometry; with it, the theorem is revealed as a fact about smoothness, and the sentence a sphere cannot be flattened acquires a qualifier it has never carried in a cartography text.

What was computed, and how #

The speed integral is evaluated by composite Simpson on the closed form at twenty thousand panels, and separately by summing chord lengths along 8,192 sampled points. The two agree, which is the check that the amplitude solved for is the amplitude that makes the length right rather than the one that makes one of the two integrators say so.

The amplitude itself is found by bisection on the mean speed, to eighty iterations, so the length ratio comes out at 1.000000000 rather than at something near it. That precision is not decoration: the whole claim is that the corrugated curve is exactly isometric at every N, and a construction that was isometric to three figures would be demonstrating a different and much weaker thing.

The iterated sequence takes each stage’s amplitude from the same bisection, applied to a curve whose speed is already what the previous stages left. The wiggle count is multiplied by a fixed factor per stage, which is a choice: the theorem allows any growth fast enough, and the numbers in the table depend on it. What does not depend on it is the pattern of the three columns, which is the result.

The cap deficit is a closed form and is quoted as three equivalent quantities — a length ratio, a cone angle and a total curvature — because each of them is the natural currency of a different part of this collection and they must agree.

Why the amplitude has to be so large #

One number in the corrugation table deserves its own paragraph, because it is the quantitative content of “corrugations are not free”.

Recovering a tenth of the length requires tilting the tangent through 34.74 degrees. That is not a small perturbation of a straight line: at every point of the corrugated curve, the direction of travel is a third of a right angle away from the direction the shortened segment runs in.

The relation is not linear. The length gained by a corrugation of tangent tilt θ goes like the mean of sec over the oscillation, so recovering a hundredth of the length costs about 11.5 degrees and recovering a half costs about 66. Small length deficits are cheap in amplitude and large ones are not, which is why the iterated construction takes the deficit down in geometric steps: each stage acts on a curve that is already nearly the right length, so each stage’s tilt is small.

That is also why the tilts sum to something finite. The tilt at stage k is roughly proportional to the square root of the deficit remaining, and a deficit halving each stage gives tilts falling by a factor of √2 — a convergent series, and the convergence of the derivative is precisely the convergence of that series.

Where the model stops #

The construction is one-dimensional. A curve corrugated across its own direction is the mechanism of Nash’s proof reduced to the case where every quantity is an exact integral. The real construction iterates corrugations in a sequence of directions on a surface, and keeping track of which lengths are already right while a new direction is corrugated is the substance of the proof. Nothing here reproduces that.

Nothing here is a map anybody could draw. The limit of the iterated construction is a fractal-like object with no tangent plane varying continuously and no scale below which it looks flat. A printed map of it would be a grey smear at any resolution.

The wiggle counts are absurd by construction. By the eighth stage the count is 6.7 million, and continuing gives numbers with no physical meaning. That is what “arbitrarily fine” costs when it is written down as a table.

C¹ is not the boundary anybody has found exactly. Nash and Kuiper’s result holds for C¹; the Theorema Egregium needs C². What happens between them — for maps with a derivative satisfying a Hölder condition — is a genuine research question, and the threshold exponent is known only within a range.

The generalisation #

The lesson is about how a theorem’s hypotheses should be read, and it generalises well past cartography.

“A sphere cannot be flattened” is a statement everybody in this subject knows and nobody qualifies. The qualified version — a sphere cannot be flattened by a map with two continuous derivatives — sounds like pedantry until the unqualified version turns out to be false, at which point the qualifier is the entire content.

For map-making the practical consequence is nil, and it is worth saying so plainly rather than pretending otherwise. Nobody wants a corrugated map. Every projection in use is analytic or piecewise analytic, and lives firmly inside the theorem’s hypotheses.

What changes is the shape of the impossibility. It is not that the sphere and the plane are incompatible in some absolute sense; it is that they are incompatible at second order, and the incompatibility is precisely the curvature that second order measures. That is a more precise statement of what the collection’s central fact is, and a more honest one — because the trade-off every projection makes is a trade-off between quantities that are also second-order, and the fact that they cannot all be satisfied has the same source.

Who found it, and when #

Gauss published the Theorema Egregium in 1827, in the Disquisitiones generales circa superficies curvas, and the word egregium is his: remarkable, standing out from the herd.

John Nash published his C¹ isometric embedding theorem in 1954, and Nicolaas Kuiper extended it to the sharp dimension in 1955. The result was received as a paradox — it says a unit sphere can be isometrically embedded inside a ball of any radius, which sounds impossible and is not — and Gromov’s later convex integration recast the technique as a general method rather than a trick.

The pictures came very much later. The Hévéa project in Grenoble computed and rendered a C¹ isometric embedding of a flat torus in 2012, and the images are the first time anybody could see what such a surface looks like: a fractal-textured object whose normal exists nowhere in the classical sense.

Cartography’s own literature has never needed the result, which is why it is absent from it, and why the qualifier is missing from the sentence everybody repeats.

Why the printer restores the theorem #

The obvious question a cartographer asks is whether any of this is usable, and the answer is no — but the interesting part is which step forbids it, because it is not the one it appears to be.

It is not that the surface is ugly. A corrugated sheet with true distances on it would be an extraordinary object and somebody would find a use for it.

It is that the construction has no finest scale, and every real medium does. The corrugations are nested without limit: any patch, however small, contains crumples, and the isometry is achieved only in the limit of that infinite nesting. A printed sheet has a finest ink width, a screen has a pixel, a stored geometry has a coordinate precision. Below that, the corrugations cannot be represented and are simply not drawn.

And a corrugation that is not drawn is a corrugation that has been smoothed. Truncating the nesting at any finite stage leaves a surface that is perfectly ordinary — twice differentiable, with a curvature everywhere, and therefore subject to the theorem again. Its distances are not the sphere’s. The 4.51 per cent deficit at 30° that the infinite construction hides in ever-finer folds reappears in whatever is left, because the folds that were carrying it are the ones the printer dropped.

So the resolution of the medium is what converts the possible back into the impossible, and it does so at any resolution whatever. There is no printer fine enough: the construction needs infinitely many scales and every printer has finitely many, so the gap is not a matter of degree.

That is the honest cartographic status of the result. It is a true statement about the smoothness hypothesis in Gauss’s theorem, it repairs a sentence that is repeated without its qualifier, and it offers nothing to anybody making a map — not because the object is impractical but because the object cannot be an artefact at all. Every drawing of it is a drawing of something else.

Which is worth holding beside what the essay’s own figures do. The corrugation pictures are exactly such truncations: they show two or three stages, they are twice differentiable, and their distances are wrong. They are illustrations of a process rather than pictures of its limit, and the limit has no picture.

Where the ladder goes next #

Eight rungs have established that curvature is the obstruction, that it is intrinsic, that it is the whole of the obstruction locally, and now that it is an obstruction at second order rather than absolutely.

What the ladder has never done is ask about the obstruction’s global half. Two surfaces with the same curvature at every point are locally isometric and need not be globally so — a plane and a cylinder are the standing example, with curvature zero apiece and no isometry between them that covers both. The obstruction there is topological rather than metric, and it is the other half of what stops a map being faithful.