A crumple has no tail

A crumple is what a sheet does when it is folded at random and flattened: a pattern of straight creases meeting at vertices, arrived at without a design, with no repeating unit and no symmetry. It is a genuine crease pattern — every vertex satisfies the conditions, because the paper satisfied them by being paper — and it is the least structured object this collection can produce.

That makes it the natural place to test whether disorder costs anything. A corrugation is regular, so its constraints are uniform and a search through them might reasonably be expected to run smoothly. A crumple has no such excuse.

It costs nothing. From three folds to eight, the search takes eight, sixteen, eighteen, thirty-four, thirty-eight and seventy steps on eight, sixteen, eighteen, thirty-five, thirty-nine and seventy-one panels. That is one step per panel, less the last panel that needs no decision, and not one wrong guess anywhere in the family.

What “deepening” means here #

The ladder needs a definition, because a crumple has no obvious size parameter and deeper could mean several things.

Here it means the number of folds the sheet was put through before flattening: three, four, five, six, seven, eight. Each additional fold adds creases, adds vertices and adds panels, and it does so unevenly — three folds give eight panels and eight give seventy-one, which is not a smooth progression because a fold that happens to cross several existing creases adds far more than one that does not.

That unevenness is useful rather than a nuisance. It means the ladder’s six points sample panel counts from eight to seventy-one without anybody choosing them, and the linearity of the cost is therefore a fit through points that were not placed. A family whose sizes had been chosen would be a family whose sizes had been chosen.

Rarer, at the same price #

What does change as a crumple deepens is how many of its letterings agree with themselves.

The consistent share falls from thirty-four of forty on a shallow crumple to eleven of forty on a deep one — a factor of three across the range — for the same reason it falls on a grid: more vertices means more places a contradiction could sit, and a lettering has to avoid every one of them at once.

The search does not notice. Its cost tracks the panel count and nothing else, and the panel count is a measure of how big the pattern is rather than of how disordered it is.

So the crumple ladder repeats, on the least structured family available, what the box-pleating ladder shows on the most structured one: rarity is a statement about a set, cost is a statement about a procedure, and the two are free to move independently. Measuring them on families at opposite ends of the order-disorder axis and getting the same answer is what makes it a general observation rather than an artefact of regularity.

Why disorder is not difficulty #

The intuition that a disordered pattern should be harder comes from thinking of a search as a hunt through a space. A structured space has patterns to exploit; a shapeless one does not; so a shapeless one should take longer.

That intuition is about a search that samples. This search does not sample. It propagates: writing a letter on one crease forces letters on its neighbours through the conditions at their shared vertex, and those force more, and the forcing runs until it stops. What decides the cost is how much of the sheet each decision settles, and that has very little to do with whether the pattern is regular.

A crumple’s vertices are irregular in their angles and perfectly ordinary in their degree — four creases meeting at a point, with a strictly smallest sector, which is exactly the configuration that admits two labellings once one crease is known. So the propagation runs through a crumple as decisively as through a grid, because the local structure it depends on is the same local structure.

Regularity, in other words, is a property of the arrangement of vertices and the propagation depends on the kind of vertex. Those are different things, and the collection’s habit of calling a pattern “structured” conflates them.

There is a sharper version of that available, and it is worth stating because it predicts where the linearity would break.

The cost of a propagating search depends on how many decisions are genuinely free after the propagation has run. On a pattern whose vertices are all degree four with a strictly smallest sector, each vertex contributes one free bit and the propagation settles the rest, so the count of free decisions equals the count of vertices and the cost is linear. What would break it is a pattern with many vertices whose conditions admit several labellings even once a neighbour is known — a degree-six or degree-eight vertex, or one whose sectors tie.

Crumples do not produce those. A sheet folded at random and flattened produces creases that cross in pairs, and a crossing of two creases is a degree-four vertex. So the crumple’s disorder is entirely in where its vertices are and what angles they carry, and not at all in what kind of vertex they are — which is precisely the axis the search cares about.

What a crumple is genuinely hard about #

There are questions about a crumple that are difficult, and it is worth naming them so the flat line above is not read as saying the object is simple.

Which folded state it is in. A crumple’s letters agreeing with themselves is necessary and not sufficient, and the search over stackings — which panel lies above which — is the expensive one. On patterns of a dozen panels that search visits hundreds of thousands of nodes, and it does not scale — which is the refusal only a search can make and the expensive half of every question here.

How many folded states it has. The crumple keeps its options measured this and found the count large, which is the opposite of what a sheet crushed in a fist looks like.

What it did on the way. A crumple’s crease pattern is the record of a process, and the process is not recoverable from the record — several sequences of folds produce the same final pattern, and nothing in the pattern says which one happened. The decision a crumple has taken is how much of that record survives, and the answer is less than a reader expects.

None of those is answered by the search this essay is about, and all of them are harder than it.

The word “random” is doing less work than it looks #

There is a temptation to read a crumple as a random crease pattern, and the collection has been careful not to, for a reason this ladder makes concrete.

A crumple is not drawn from any distribution over crease patterns. It is the output of a construction — fold along a line, fold again, flatten — and every crumple in the ladder is a member of whatever family that construction produces. Which patterns are the random ones is the question this collection keeps returning to, and the answer keeps being that there is no such thing: there are only constructions, and each one produces the patterns it produces.

That matters here because the flat cost line is a property of the construction. Crumpling by repeated folding produces degree-four vertices exclusively, and every conclusion above depends on that. A different way of producing a “random” pattern — perturbing a grid’s vertices, say, or drawing lines and letting them cross three at a time — would produce a different family with different vertex degrees, and there is no reason to expect the same line.

So the honest headline is not disorder does not cost. It is this construction’s disorder does not cost, and the mechanism explains why: the construction is disordered in the coordinates and perfectly regular in the combinatorics.

The cost per panel is one over the degree #

The mechanism has a formula, and the formula predicts the sign of the deviation the essay’s proposed experiment would produce — which turns an open test into a check.

The search’s genuine decisions are one per interior vertex, since each degree-four vertex contributes one free bit and the propagation settles the rest. So the cost per panel is the ratio of vertices to panels, and that ratio is fixed by the degree.

For a pattern whose interior vertices all have degree $d$: every crease has two ends, so $dV = 2E$; and panels, vertices and creases satisfy $V + F = E + 1$. Substituting,

$$\frac{V}{F} \;\approx\; \frac{1}{\tfrac{d}{2} - 1}$$

One decision per panel at degree four. A half at degree six. A third at degree eight.

Which says the line should bend downward #

That is the opposite of the essay’s expectation, and it is what the Yoshimura already shows.

A degree-six family has half as many vertices for the same number of panels, so it has half as many free decisions, so its search should cost below one step per panel rather than above. The Yoshimura’s fifty-seven steps on sixty-five panels — 0.88, and the deviation blunted by its many boundary vertices, which carry no condition — is that prediction arriving.

So a triangulated sheet or a three-line crumple would not break the line by rising above it. It would fall below it, toward a half, and confirm the mechanism by moving in a direction nobody would have chosen as evidence for it.

The refutation the essay proposes is therefore not a refutation but a calibration, and it has a predicted value rather than merely a predicted direction. That is a better experiment than the one described, because a family that came in at one step per panel — neither one nor a half — would genuinely refute the account.

What the ladder’s own growth says #

One more thing falls out of the ladder’s numbers and it bounds how far the family can be pushed.

Panels go 8, 16, 18, 35, 39, 71 across six folds, which is a factor of 1.55 per fold in the geometric mean; creases behave the same way. A fold that duplicated the whole sheet would give two, and this gives one and a half — because a fold reflects only the material on one side of its crease and duplicates that, not everything.

So a crumple grows by half again per fold, and the ladder’s six points span a factor of nine in size. That range is what makes a straight line through them worth fitting; six points a factor of two apart would not have distinguished linear from anything else.

The comparison that would settle it properly #

What would make all of this considerably stronger is a family built deliberately to break the line, and it is worth describing because it has not been built.

Take a construction that produces vertices of degree six as freely as degree four — a triangulated sheet, or a crumple made by folding along three concurrent lines rather than two crossing ones. By the mechanism above, a degree-six vertex admits more labellings once a neighbour is known, so the propagation should settle less of the sheet and the search should have genuine choices where the degree-four families have none. If the cost then rises above one step per panel, the mechanism is confirmed against a case designed to refute it. If it does not, the mechanism is wrong and something else is producing the linearity.

Either outcome is worth having, which is what makes it the right next measurement rather than a sixth family on the same line. The Yoshimura already hints at the answer — its degree-six vertices give fifty-seven steps on sixty-five panels, below the line rather than above it — and one corrugation is not a family.

Which theorem was checked, and how #

Every lettering the search returns is written back onto the crumple and put past two instruments that did not produce it: the four conditions at every interior vertex, read by the pattern’s own reader, and a folded sheet rebuilt from the coordinates and walked for a circle in the arcs.

The claim of linear cost is asserted rather than observed: every crumple in the ladder is required to cost no more than one step per panel, and the assertion fails on the first one that does not. That is stronger than a straight line through six points, because a search could visit exactly n nodes while taking a wrong turn and recovering, and the requirement rules that out.

The rarity measurement comes from the sampler, which shares no code with the search — it draws letterings that pass every vertex condition and tests each afterwards. The two agreeing about which crumples have consistent letterings at all is what makes the pair a comparison rather than one instrument reported twice.

What the picture cannot show #

A crumple here is a simulation rather than a sheet. It is produced by folding a rectangle along random lines and flattening, which reproduces the combinatorics of a crumpled sheet and not its mechanics: real crumpling involves stretching, curved ridges, plastic deformation and a great deal of stored elastic energy, none of which is in this model. Four things that are not true sets out the gap, and the honest reading of everything above is that it is about a piecewise-flat pattern with disordered vertices rather than about crushed paper.

The ladder is also six patterns from one seed, deepening in one direction. Six is few, one seed is one, and a different random construction could in principle produce a crumple whose search behaves differently. Nothing in the mechanism suggests it would — the argument depends only on the vertices having degree four and a strictly smallest sector — but the measurement is six points.

And the flat line is flat in decisions. Each step propagates the conditions over the whole pattern, so a seventy-panel crumple’s steps cost more than an eight-panel one’s, and a wall-clock reading would rise faster than linearly while the decision count stayed exactly linear.

What a folder would recognise #

There is a hand version of this and it is the ordinary experience of unfolding a crumpled sheet.

Flatten a ball of paper and the creases are visible immediately — as ridges and valleys, already assigned, because the paper has made the assignment by being folded. Nobody looks at a crumpled sheet and wonders which way each crease goes; the sheet has decided, and it has decided consistently, because a consistent assignment is the only kind a sheet can physically hold.

That is the whole content of the search’s flat line, in the hand. The paper solves the problem by construction, in one operation, with no search at all — and the reason it can is exactly the reason the search is cheap: at every vertex the local situation determines almost everything, so a globally consistent answer assembles itself from local ones.

The letters a crumple was given is where that was first measured, and the reading here adds only that the cheapness survives at every depth. A sheet crumpled harder is not a sheet that has solved a harder problem.

What the ladder is evidence for #

Five families now sit on the same line — grids, leaves, crumples, corrugations and cut patterns — and it is worth being explicit about what that accumulation buys, because the answer is not simply “more confidence”.

Each family fails a different reason for the linearity. If the cost were low because the patterns are regular, the crumples would break it. If it were low because the patterns are small, the sixteen-division grid at two hundred and fifty-six panels would break it. If it were low because the patterns were designed by somebody who knew the search, the crumples and the cut patterns would break it, since neither is designed at all. If it were low because the vertices are sparse, the grids would break it.

None of them does, and what survives every one of those objections is the mechanism: a degree-four vertex with a strictly smallest sector admits two labellings once one crease is known, and a sheet made of those has one free decision per vertex. A collection of families is not evidence by weight of numbers — it is evidence because each member rules out a different alternative explanation, and choosing families for that reason is the only thing that makes a fifth one worth measuring.

Where the ladder goes next #

Every family measured so far — grids, leaves, crumples, corrugations, cut patterns — costs one step per panel and never backtracks, which starts to look like a fact about the search rather than about any of them. The exception is instructive: the one family whose cost is not linear is the tessellation patches, and what makes them different is not disorder but an edge, where the propagation that sweeps cleanly across everything else runs out of sheet.