One crease decides the sheet

Paper that folds itself ends on a difficulty rather than a solution. A self-folding sheet has to supply the fold and then choose what to fold into, and the second half is where these things fail: at a degree-four vertex there are two branches out of the flat state, they are equally downhill, and no amount of torque prefers one. The obvious conclusion is that every vertex needs its own bias — a pre-crease, a differential shrink, an asymmetry built in — and that a sheet with a hundred vertices needs a hundred decisions.

That conclusion is wrong, and the measurement that shows it is one the flat-folding half of this site has already made and got the opposite answer from.

The experiment, as the flat half ran it #

How little the conditions decide fixed one crease of a pattern, propagated Kawasaki, Maekawa and the big-little-big lemma to a fixed point, and counted what followed. The numbers were discouraging and they have not improved: on a Miura of thirty-eight creases, one. On a waterbomb patch of seventy-six, one. On a Yoshimura of eighty-six, one. On a square twist tessellation of eighty-four, three.

The reason is not subtle. The conditions are conditions on letters, and a vertex admits several labellings that satisfy all of them — eight at a degree-four vertex with a tie, four without. Fixing one crease removes at most half of one vertex’s options, and half of a set of four is still two, so the intersection with the neighbour removes nothing and the wave stops at the first vertex it reaches.

There is a temptation to read that as a defect of the propagation and to reach for a search instead. It is not. A search finds an assignment, which is a choice among many rather than a consequence of the pattern; the propagation reports what is entailed, and what is entailed is almost nothing. The distinction matters here because the rigid side is about to report something entailed, and the two have to be measured the same way for the comparison to mean anything.

The experiment, as the rigid half runs it #

The rigid question fixes a fold angle rather than a letter, and that changes the arithmetic completely.

A degree-four developable vertex is a spherical four-bar. Fix the fold angle at one of its creases and the other three are determined, up to a choice between two configurations — not several labellings but two configurations, each of which is a complete set of four numbers. Carry either to the neighbouring vertex along their shared crease, and the neighbour’s other three angles are determined in the same way.

So the propagation is a search over branches, two per vertex, and what makes it strong is not the branching factor but the checking: a branch taken at one vertex is tested against every crease that vertex shares with a vertex already solved, and the contradictions arrive immediately.

On the general quadrilateral mesh built for the previous rung — four interior vertices, twelve creases, no straight crease family, no two vertices alike — fixing one fold angle determines all twelve creases and leaves exactly one consistent set of angles out of the sixteen branch chains available. Nothing else survives.

On a four-by-four Miura the propagation determines all twenty-four creases, and on a five-by-five all forty. There the count of consistent sets is four and eight rather than one, and the reason is the Miura’s own symmetry: it has mirror configurations that are genuinely different foldings of the same pattern at the same drive. Every one of them is complete — the ambiguity is a small discrete set of whole answers, not a large continuum of partial ones.

Two things are worth separating in that result. The propagation is complete — every crease gets a number — and it is almost unique, which is a stronger and more surprising property. Completeness follows from the vertex solve: as long as the walk can reach every vertex through shared creases, every crease is eventually determined. Uniqueness does not follow from anything local: it is the consistency checks killing branches, and on the irregular mesh they kill fifteen of the sixteen chains.

What a self-folding sheet actually needs #

The consequence is the practical one. A sheet that folds itself needs a bias, and the bias has to break the symmetry between the two branches at a vertex. What this measurement says is that one such bias is enough for the sheet, because the fold angle it establishes is carried to every neighbour by the closure and nothing else survives being carried.

That is a different engineering problem from the one the difficulty implies. One vertex has to be told which way to go — a pre-crease, a hinge with a stop, an actuator with a preferred direction, a corner held while the rest is released — and the geometry does the remainder. A hundred-vertex sheet does not need a hundred decisions; it needs one decision and a mesh whose loops close.

The condition attached to that is not small, and it is the previous rung’s subject: the loops have to close. On a mesh whose loops do not, the propagation does not merely become ambiguous — it fails outright. Fixing one fold angle on a mesh five per cent away from the Miura in the slide family determines one crease and then finds no consistent continuation at all, out of every branch chain there is.

That is a useful failure mode. A propagation that returns nothing is a mesh that does not fold, reported before anything is built.

Which theorem was checked, and how #

The flat propagation is forcedFrom, unchanged since the essay that introduced it: it propagates the local conditions to a fixed point and never branches, so what it reports is what those conditions entail, which is a smaller and more stable quantity than what a search can find. Using a search would flatter the flat side and blur the comparison.

The rigid propagation is a depth-first walk over branch choices, with a vertex admitted only when it shares a crease with one already solved, and every shared crease checked to within 10⁻⁷ radians. Two things are excluded deliberately. Configurations in which some crease has a fold angle of zero are rejected, because a configuration that leaves a crease unfolded is a folded state of a different pattern — the one with that crease rubbed out — and it satisfies every closure without being an answer to the question. And the walk is capped, so a mesh that admits many solutions reports many rather than running forever.

The control is the Miura, whose one degree of freedom is established elsewhere by two other routes. If the propagation had failed to determine a Miura completely, the propagation would be wrong rather than the Miura interesting.

There is a second practical reading, and it is about diagnosis rather than actuation. Because the propagation is a search that either completes or dies, running it is a test: hand it a mesh and a drive, and it answers whether that mesh has a folded configuration at that drive, together with the configuration if there is one. That is a cheaper question than solving the whole system at once, and it is the question a designer actually asks — not “is this mesh rigid-foldable in principle” but “does this mesh have somewhere to go from here”.

It also gives the failure a location. When the walk dies it dies at a particular vertex, having been handed a fold angle its own closure cannot accommodate, and the vertex it dies at is the one whose neighbourhood the designer has to change. A residual computed over a whole sheet says only that something is wrong.

Where the model stops #

The comparison is between two different questions, and the essay’s claim is only about the contrast, not about one being better. Flat-folding asks which letters can be written; rigid folding asks what angles a sheet passes through. A pattern can propagate perfectly under the second and have no letters at all under the first — a rigid folding that never reaches a flat state has no assignment to speak of.

The rigid propagation also assumes rigid panels, which real material is not. A self-folding sheet made of a laminate with a shape-memory layer bends its panels a little, and a little bending is exactly what buys a mesh whose loops do not quite close the ability to move anyway. So the total refusal above is a refusal for an idealised sheet, and the practical version of it is a stiffness rather than an impossibility — which is structural rather than geometric and is not developed here.

Finally, the meshes are small: four interior vertices in the general case, nine and sixteen in the Miuras. Propagation over a large irregular mesh needs the many-face solve that the previous rung names as owed.

What the picture cannot show #

The propagation figure is two bars, which is an honest summary and a poor picture of what happens. What actually happens in the rigid case is that fifteen of the sixteen branch chains die at the second or third vertex, and the survivor completes; the shape of that — a search that collapses immediately — is not visible in a ratio.

Nor is the bias. Nothing in any figure shows what physically makes one vertex choose, because that is a matter of material rather than of geometry and this site’s licence stops at the geometry.

The generalisation #

The contrast is not really between two kinds of folding. It is between a rule that constrains a discrete quantity and a rule that constrains a continuous one, and it is a general fact about propagation that the second travels and the first does not.

A letter is one of two values. A condition on letters removes candidates from a shortlist, and a shortlist with two survivors is as good as no information to a neighbour that had two candidates of its own. A fold angle is a real number. A condition on fold angles is an equation, an equation with one unknown has isolated solutions, and isolated solutions are information a neighbour can use.

That is why the two halves of this subject feel so different to work in. The flat-folding side is combinatorial, its constraints are weak individually, and its difficulty is that the search space is enormous. The rigid side is algebraic, its constraints are strong individually, and its difficulty is that they are over-determined — the sheet is either exactly right or it does not move at all. The same pattern, drawn once, supports both, and neither difficulty is visible from the other.

Who found it, and when #

That a developable quadrilateral mesh which folds has one degree of freedom is Tomohiro Tachi’s, and it is the fact this essay is reading from the other end: one degree of freedom means one number, and one number means one decision. Self-folding as an engineering programme is a great deal younger — the shape-memory and light-activated sheets of the last two decades — and the branch problem is well known in that literature as the reason self-folding demonstrations use patterns with strong pre-creases everywhere.

What is new here is the comparison, and it is available only because both propagations are written down here and can be run on the same pattern. The flat number and the rigid number have never been placed side by side, because they are usually computed by different people for different reasons.

One more thing is worth recording because it is the sort of number that does not survive being remembered. The propagation on the irregular mesh visited sixteen partial states before finishing, and on the five-by-five Miura a hundred and eighty-seven. Those are not search costs worth optimising; they are evidence about how quickly the contradictions arrive. A propagation that had to explore thousands of chains before finding its survivor would be a different phenomenon, and would not support the claim that one bias suffices — because a sheet does not backtrack.

How fast the branches die #

The propagation’s strength is the checking rather than the branching, and the arithmetic of that is worth setting out, because it is what turns almost unique from an impression into a measurement.

A walk that chooses between two configurations at every interior vertex has $2^v$ branch chains available to it. The irregular mesh has four interior vertices, so sixteen chains, of which one survives — a sixteenth. The four-by-four Miura has nine, so five hundred and twelve chains, of which four survive: eight in a thousand. The five-by-five has sixteen, so sixty-five thousand five hundred and thirty-six chains, of which eight survive, which is one part in eight thousand.

The survivors grow and their share collapses. One, four, eight, against sixteen, five hundred and twelve, sixty-five thousand. A larger sheet is a larger space of ways to be wrong and very nearly the same small set of ways to be right, which is the same statement as one degree of freedom arrived at by counting rather than by kinematics.

The visited counts say the search never sees most of that space. Sixteen partial states on the irregular mesh, which is every chain, because four vertices is too few for pruning to matter. A hundred and eighty-seven on the five-by-five, against sixty-five thousand available — three parts in a thousand. The contradictions arrive within two or three vertices of the seed and the rest of the tree is never entered, which is the property the essay’s engineering claim rests on. A sheet does not backtrack, and a search that had to would be describing something a sheet cannot do.

One bias, and the assumption inside it #

The claim that one decision suffices carries an assumption that is invisible on every mesh above and is worth naming, because it is a design rule rather than a caveat.

The propagation travels along shared creases. A vertex is admitted when it shares a crease with a vertex already solved, so what the walk covers is the connected component of the shared-crease graph containing the seed. On a Miura, on the irregular mesh, and on every pattern in this collection that graph is connected, and the assumption never shows.

It need not be. A sheet with two creased regions separated by a band of uncreased paper has a shared-crease graph in two pieces, and a fold angle fixed in one of them says nothing whatever about the other: the flat band between them can be traversed by the paper at any angle the two regions like. Such a sheet needs one bias per component, and no amount of geometry will carry a decision across paper that has no creases in it.

So the finding is properly stated with its condition attached. A self-folding sheet needs one decision per connected component of its shared-crease graph, and the reason a hundred-vertex Miura needs exactly one is that a Miura is one component. That is a cheap thing to check on a design and a very expensive thing to discover after building it, since a two-component sheet given one bias does not fail — it folds half-way, correctly, and stops. That is the same shape of trouble as a condition that holds everywhere it is asked and not over the sheet, arriving from the actuation end: everything local is right and the object is not what was wanted.

Where the ladder goes next #

The bias has to be somewhere, and where is not indifferent: a vertex at the edge of the sheet has fewer neighbours to convince and a vertex in the middle has more, so the propagation from different seeds is not equally fast even when it is equally complete. That is measurable and is not measured here.

The other direction is what a panel with real depth does to all of this, where the two branches stop being symmetric for a reason that has nothing to do with the pattern.