A curve has no panels

Almost everything in this subject gets easier by subdivision. A curved surface is approximated by facets and the error falls with the square of the panel size. A circle is approximated by a polygon and nobody objects. A machine that can only make straight creases can make a great many of them, close together, and the usual expectation is that a curved crease is just a straight crease problem with a large number in it.

It is not, and the reason is a conservation law rather than a numerical difficulty.

The number under each panel is the same number. That is the finding, and everything below is either what it means or how it was checked.

The object all of it is about is not exotic. It is a line on a sheet of paper that happens to bend.

What a rigid folding actually is #

The word “panel” is doing all the work, so it needs to be exact. A rigid folding is a sheet divided into finitely many regions, each of which stays perfectly flat throughout the motion, joined along finitely many straight lines about which they hinge. Nothing bends anywhere except at a hinge, and a hinge is a line.

That definition is not a modelling convenience. It is what makes rigid origami buildable: a flat region can be a sheet of aluminium or a solar panel, and a hinge can be a hinge.

The classic object on the rigid side is a straight-crease vertex, whose panels move as one linked assembly.

The distinction between paper and panels is normally a matter of degree: paper bends a little, panels do not, and the gap between the two shows up as tolerance and thickness. For a curved crease it is not a matter of degree at all.

The one quantity that refuses to improve #

Take a curved crease and replace it by k straight segments. The sheet either side of it becomes 2k flat regions, which is a rigid folding of something. The question is what it is a rigid folding of.

Three quantities can be tracked as k rises, and they behave in three different ways.

The greatest distance between the chords and the curve falls as k⁻¹·⁹⁹⁸, which is the sagitta of a chord and is why a sixteen-segment drawing already looks right. The turning at each joint falls as k⁻¹·⁰⁰⁰, which is what dividing a fixed amount between k joints does. Both exponents are least-squares fits to the measured values, not constants written into the figure.

The third series is flat. At four segments the joints of this crease collect 128.34° of turning between them; at 128 segments they collect 128.34°, and the difference between the two is under a millionth of a millionth of a degree, which is the arithmetic and not the geometry.

Why the total cannot fall #

The reason is short enough to state in a sentence and is worth stating carefully, because it is the whole argument.

The total turning of a polyline is the angle from its first direction to its last, accumulated along the way. The approximation starts at one end of the crease with the crease’s own tangent and finishes at the other with the crease’s own tangent, because the chords have their endpoints on the curve. So the total turning of any approximation is the angle between those two tangents — which is a property of the curve, decided before any approximation was chosen, and identical for every choice.

Refining does not remove kink. It divides it.

That is a conservation law in the ordinary sense: a fixed quantity being redistributed among a changing number of places, with the sum invariant. It is also why the flat series in the figure is exactly flat rather than nearly flat. A quantity that fell slowly could be argued about; this one does not fall.

There is a second reading of the same number, and it connects this essay to the rung below it. The total turning is the integral of the crease’s curvature along its length — the curvature it has as a line drawn on the flat sheet. Folding cannot change that curvature, because it is measured within the surface and folding is an isometry. So the obstruction here is exactly the quantity that makes a curved crease worth having: the thing that gives the fold its shape is the thing panels cannot represent, and they are not two properties that happen to conflict but one property looked at twice.

The rulings are not free either #

There is a second and independent reason a panel cannot exist here, and it does not mention approximation at all.

A curved crease forces the surface either side of it to be ruled: through each point of the crease runs a straight line lying in the surface, and the surface is the union of those lines. For a fold of the kind these figures draw, the rulings leave the crease at a fixed angle on both sides.

Tangent lines to a circle are not parallel. Two neighbouring rulings therefore meet, at a finite distance, and the piece of sheet between them narrows to nothing where they do. It is a wedge, not a strip. A panel is a flat piece with area, and this has area only for as long as the rulings are ignored.

Take the rulings seriously and the sheet either side of a curved crease is a family of straight lines with no two of them parallel, which is a developable surface and is precisely not a finite collection of flat pieces. Both halves of the argument arrive at the same place from opposite directions: subdividing the crease keeps the kink, and honouring the rulings keeps the width from existing.

Which theorem was checked, and how #

The claim is an invariance, and an invariance is easy to fake by computing it once and printing it several times. So the check is built the other way round.

Every angle is measured off the chords that were built. The directions come out of an arctangent on the segment endpoints, the joint angles are differences of those directions, and the total is their sum. Nothing reads back the curve’s own turning and compares it with itself; the constant emerges from the construction, and only then is it compared with the curve.

The ends are counted. The turning from the crease’s tangent on to the first chord, and off the last chord back on to the tangent, are part of the total. Omitting them is the standard way this measurement goes quietly wrong: the interior joints alone give a total that appears to creep upward with the segment count, which reads as slow convergence to something and is an artefact of having dropped two terms.

The falling exponents are fitted, not asserted. The generator refuses to draw if the per-joint turning does not fall as the count to the power −1 within two parts in a hundred, or if the distance from the curve does not fall as the power −2 within six. Those are the two ways subdivision is supposed to help, and requiring them is what makes the third series informative rather than merely flat.

The range is wider than the drawing. Whatever segment counts a figure shows, the check runs from the smallest of them to eight times the largest, doubling. A constant that held over four counts and drifted over ten would be a numerical coincidence.

What the pictures cannot show #

The hero figure’s last panel is the honest admission. At sixteen segments the approximation is visually indistinguishable from the smooth crease, and there is no way to draw the difference, because the difference is not in the ink. It is in a number underneath, and a reader who trusted the drawing over the number would conclude the opposite of what is true.

The figures also cannot say whether any of these panelled patterns folds. They are drawn flat, as crease patterns, and whether a given arrangement of panels has a rigid motion at all is a separate question decided by conditions at the vertices — which for a curved crease live only at the finitely many points where creases cross, and where the curvature does not appear at all.

The idealisations are the usual three and one that is specific. The sheet has no thickness. A hinge is an exact line. A panel is perfectly flat and stays perfectly flat. And “panel” is being used in its strict sense throughout: a region that never bends anywhere, at any stage of the motion. Relax that by any amount and the argument stops applying immediately — which is exactly what real hardware does, and is the next section.

What is built instead #

None of this says a curved fold cannot be made. It says a curved fold cannot be made out of panels, and the objects that exist are made out of something else.

Paper does the job because paper bends. Every point of that surface away from a crease is bent in one direction, which costs a sheet nothing, and the object is exactly the family that produced the best things in the subject decades before anybody could compute one. What it is not is a mechanism with hinges.

Hardware wanting the same shapes has three ways out and each abandons something. It can accept panels that bend, which means the flat-panel model is being used as an approximation and the residual has to be absorbed somewhere. It can make the crease a strip of compliant material rather than a line, which is the same move a real crease radius forces anyway. Or it can subdivide and accept the kink, which is the honest reading of the figures above: the kink at each joint does fall as one over k, so at enough segments each individual joint is small enough for a hinge to absorb, even though the total never budges.

That last trade is the practical content of the whole essay. Subdividing does not make the problem go away, and it does make each remaining piece of the problem small — which is a different and much weaker statement, and the one an engineer can act on.

The crease that turns both ways #

The invariance argument as stated reads the turning with its sign — the angle from the first chord’s direction to the last, accumulated along the way — and on the creases drawn here that is the whole story, because a circular arc turns one way throughout. A crease with an inflection in it turns one way and then the other, and there the signed total can be zero while every joint still kinks.

So the signed total is the wrong invariant in general, and the right one is the total turning taken without regard to sign: the sum of the absolute kink at every joint. On a crease that never changes its direction of turning the two coincide, which is why the measurements above are exactly flat rather than nearly so. On an S-shaped crease they part company, and only the second is the obstruction.

The second behaves slightly differently under refinement and the difference is worth knowing. A signed total is exactly constant, because it telescopes to the difference between two tangent directions. An absolute total does not telescope; it rises toward the integral of the crease’s curvature magnitude along its length, approaching it from below as the segments shorten. So refining a serpentine crease makes the obstruction slightly larger rather than leaving it alone.

Either way it never goes to zero, and that is the only thing the argument needs. A crease with any curvature anywhere has a positive integral of curvature magnitude, every inscribed polyline collects some fraction of it, and no amount of subdivision removes what subdivision only redistributes.

What the segmentation actually costs #

The practical reading — subdivide until each joint is small enough for a hinge to absorb — has a price that can be written down, and writing it down shows that the obvious lever does not work.

A hinge or a compliant strip can take some angular kink, call it its capacity. The total to be distributed is fixed by the crease, so the number of segments needed is that total divided by the capacity. Nothing about the choice of segmentation changes it: taking fewer segments makes each joint too sharp, taking more wastes nothing but does not reduce the count that was required.

Then count what those segments cost in paper. The chords themselves are not the expense — they sum to about the crease’s own length however many there are, so the paper their radii consume does not grow with the count. What grows is everything else: each joint needs its own ruling creases so that the panels either side can stay flat, and those run across the sheet rather than along the crease. Their number is the segment count, so the paper they consume rises in proportion to it.

Which gives the shape of the bill. The paper a segmented curved crease consumes is set by the crease’s total turning divided by the hinge’s capacity, and a designer has two levers on it, neither of them the segmentation. One is the shape being asked for, since the total turning is a property of the drawn curve and is fixed before anything is built. The other is the material, since the capacity is a property of the hinge. Choosing how finely to subdivide is not a lever at all — it is a consequence of the other two, and the cost was decided when the curve was drawn.

Who noticed it, and when #

Curved-crease folding is old as a practice and recent as a theory. The Bauhaus preliminary course had students fold concentric circles into a saddle in the 1920s, and nobody involved was computing anything. David Huffman — better known for a coding algorithm — spent much of the 1970s and 1980s making curved-crease sculptures and working out the geometry that governs them, including the ruling behaviour the figures above assert.

The statement that a curved crease admits no rigid folding is a twenty-first-century one, and it belongs to the literature that grew up around rigid origami as an engineering discipline rather than around curved creases as an art. It is the sort of result that only becomes worth writing down once somebody has a machine that wants to make one: as long as curved folds were made by hand in paper, the question of whether the panels exist does not arise, because there are no panels and nobody was looking for any.

Erik Demaine, Martin Demaine, Duks Koschitz and Tomohiro Tachi assembled the modern account of curved-crease folding in the 2010s, recovering and extending Huffman’s work; the ruling and tangent-circle geometry drawn here follows that treatment. The invariance argument itself is elementary once stated, which is normally a sign that the difficulty was in noticing that the question mattered.

Where the ladder goes next #

Below this rung the ladder is about what a curved crease is: that the paper either side is forced into a shape nobody creased, that the sculptors got there first, that one number controls everything the fold adds, and that the conditions live at the crossings. This rung is the negative result those four make available, and it is the first one on the anchor that says what cannot be done rather than what can.

Above it sits the question this essay deliberately does not touch. Given a target surface, find the creases that produce it — the inverse problem, which is open in general, and which is where curved folding stops being a subject with theorems and becomes a subject with examples. The panel result narrows it usefully: whatever the answer turns out to be, it is not a list of flat pieces.

Sideways, the interesting neighbour is the machine, which can only make what it can specify. A crease pattern of straight lines is a finite list of coordinates and a curved crease is a function, and the difference between those two kinds of specification is a much larger part of why the subject looks the way it does than the geometry alone would suggest.