Two things called folding

Two of the most-discussed folding problems in science share a word and very little else. A protein folds, in the sense that a chain of amino acids arrives at a particular three-dimensional arrangement. A sheet folds, in the sense that a flat piece of material arrives at a particular layered arrangement.

The vocabulary encourages a transfer of intuition that does not survive examination. This essay computes the one comparison that can be made honestly and uses it to say precisely where the analogy breaks.

The counts, and why they are not the story #

Start with what is genuinely comparable. Both problems have a configuration space that grows exponentially with the size of the object, and both are therefore hopeless to search.

A crease pattern with k creases has two choices per crease — mountain or valley — so two-to-the-k assignments. At ten creases that is 1,024.

A chain with k rotatable joints, each with some small number of favoured states, has that number raised to the k. At three states per joint and ten joints it is 59,049, and at the twenty-odd states per residue that a real protein backbone has it is a number with no useful name.

The observation that this makes exhaustive search impossible is old and correct, and it is the least interesting thing either problem has to say. Exponential spaces are ordinary; what distinguishes problems is what can be said about the space without searching it.

What a local condition is #

Here is the distinction the essay turns on, and it is worth defining carefully.

A condition is local if it can be evaluated by looking at a bounded neighbourhood of one place in the object, with no information about anywhere else.

The flat-folding conditions are local in exactly that sense. Developability, Kawasaki, Maekawa and big-little-big are all statements about the sectors and creases meeting at a single vertex. Evaluating them requires knowing nothing about the rest of the sheet — not its size, not its other vertices, not which vertices the pattern has anywhere else.

Because they are local, they can be applied to every vertex of a pattern of any size in time proportional to the number of vertices. A cheap test that removes most of a space is worth vastly more than a clever search of the whole of it, and this is the structural advantage the folding of sheets has.

Why a chain has no such test #

The chain problem has nothing playing this role, and the reason is geometric rather than a matter of nobody having found one.

The property a folded chain has to satisfy is that residues far apart along the chain end up close together in space. Whether two residues are close depends on every torsion angle between them — turn one joint in the middle and everything downstream of it swings. So no bounded neighbourhood of the chain determines anything about which contacts are made.

That is the definition of non-local, and it is not an artefact of any particular model. It is a consequence of the object being a chain: position accumulates along it, and accumulation is exactly what defeats locality.

A sheet does not have this problem because its constraints are about angles at a point, which do not accumulate. Two vertices at opposite ends of a pattern constrain each other only through the intervening geometry, and the local conditions never need to know.

It is worth testing that claim against the obvious objection, which is that a folded sheet also brings distant parts of itself together and therefore also has non-local structure. It does — and that is precisely the part the local conditions cannot see. Which layer ends up on top of which is a global question about the whole sheet, the local theorems say nothing about it, and it is where the difficulty of the sheet problem lives.

So the sheet has both kinds of structure: local constraints about angles, which are cheap and powerful, and global constraints about layers, which are neither. The chain has only the second kind. Reading the comparison that way makes it less a claim about which problem is harder and more a claim about what each one has available, which is the more useful framing and the one that survives.

The asymmetry is also why the two fields’ hard results look so different. The sheet’s hardness result is about layer ordering, arrived at after the local questions were settled; there is nothing corresponding to the local questions in the chain case for a hardness result to be arrived at after.

What the filter is worth, measured #

The site has computed the filter’s strength in several places and it is large.

At a degree-four vertex with unequal sectors, four of the sixteen assignments fold — a removal of seventy-five percent, at one vertex. Applied to a pattern with a hundred vertices, if the vertices were independent, that would leave a fraction of the space too small to write down.

They are not independent, of course, and the surviving set is larger than the naive product suggests. The census of how many assignments fold is the honest version of that arithmetic, and the headline is unchanged: flat-foldable patterns are a vanishingly small and highly structured subset of all patterns.

What locality buys, as an exponent #

The filter’s strength was given above as a percentage at one vertex, and the useful form is what it does to the exponent of the whole pattern, because that is the quantity a chain has no version of.

Take an $n$ by $n$ grid of quadrilaterals. It has $2n(n-1)$ creases and $(n-1)^2$ interior vertices. Without any conditions the assignments number $2^{2n(n-1)}$ — an exponential in the area.

Each vertex removes a factor. Where the sectors are equal, Maekawa alone applies and the factor is two, leaving $2^{n^2-1}$, which is still an exponential in the area. Where a sector is strictly smallest, the smallest-sector lemma applies as well and the factor is four, leaving

$$2^{\,2n(n-1) - 2(n-1)^2} = 2^{\,2(n-1)}$$

which is an exponential in the perimeter.

So the local conditions convert a bulk count into a boundary count, and that is the whole of what locality is worth here. A sheet a hundred panels on a side has $2^{19800}$ assignments and, if its vertices are generic, about $2^{198}$ admissible ones. Both numbers are hopeless to enumerate and they are not the same kind of hopeless: the first grows with the paper and the second with its edge.

Which is exactly what the smallest-sector lemma does #

That also isolates which condition is responsible, and it is not the famous one.

Maekawa’s factor of two removes one of the two doublings per crease and leaves the exponent proportional to the area. The second factor of two — the one the smallest-sector lemma supplies, and only where a sector is strictly smallest — is what collapses the exponent to the boundary. A map, whose vertices are four right angles, never gets it: its admissible letterings number two to the power of its panel count, which is the bulk figure, and nine enumerated map sizes confirm that exactly.

The difference between a bulk exponential and a boundary one is a single lemma about which sector is smallest, and it is available only because the lemma can be checked at one point with a protractor.

Nothing of this shape exists for a chain. There is no quantity at a residue whose value removes a factor from the count of the whole chain, because the count of the whole chain is not a product over residues in any way a bounded neighbourhood can see. The sheet’s exponent can be attacked locally; the chain’s cannot be attacked at all.

The part where the analogy briefly works #

There is one respect in which the two problems genuinely rhyme, and it is worth granting because it makes the disanalogy sharper.

Both have the property that local conditions do not determine the global answer. Every vertex of a crease pattern can satisfy all four theorems and the sheet can still fail to fold, because the layers have to fit together globally and nothing local knows about layers. That gap is one of the first things this site established, and it is why the global problem is NP-hard while the local test runs in linear time.

The chain has the same shape of gap and starts from a worse position: it has no local test to fall short, so the entire difficulty is global from the beginning.

So the accurate statement of the relationship is: the sheet problem is the chain problem’s easier cousin, made easier by exactly one thing, and that thing is locality.

The same distinction, at the routing scale #

The previous rung supplies a third case, and it lands on the sheet’s side of the line for a reason worth extracting.

Routing a scaffold through a shape is an exponential search problem with a cheap necessary condition attached: count the two lattice colours and compare. That condition is local in the relevant sense — it is a property of the cells, computed one at a time, with no reference to any route.

So DNA origami, despite being a chain problem in every physical respect, has the structure of the sheet problem: a cheap filter that refuses, an expensive oracle that decides, and a gap between them. The reason is that its difficulty was moved into a combinatorial object — a lattice — where locality is available again.

That suggests the right way to read the whole comparison. Locality is not a property of the physical system; it is a property of the representation. A problem becomes tractable-ish when somebody finds a representation in which a cheap local condition says something, and folding sheets is a subject that arrived with such a representation already in it. Chains have not been so lucky, which is a fact about the state of the art rather than a theorem.

Where the word came from, and what it costs #

The borrowing of the word is not anybody’s mistake and it does have a cost worth naming.

“Folding” for a protein predates any of the mathematics on this site and describes a real phenomenon — a chain arriving at a compact structure — with no reference to sheets. “DNA origami” is a deliberate metaphor, coined for a technique whose relationship to paper folding is that both produce a designed shape from a single starting object.

The cost is that arguments transfer where they should not. Results about flat-foldability say nothing about chains, because they are theorems about angles around a point and a chain has no such angles. Results about protein folding say nothing about sheets, for the mirror reason. Both directions of transfer are made regularly and neither is supported by anything.

The specific transfers worth watching for are the ones about difficulty, because difficulty claims are the most quotable and the least portable. “Folding is NP-hard” is a true statement about deciding whether a crease pattern folds flat, and it is not a statement about proteins. “Folding cannot be searched exhaustively” is true of both and is a statement about exponential growth rather than about folding. A reader who collects these without their antecedents ends up with a general impression that folding is hard, which is worth nothing and feels like knowledge.

This site has a standing habit that heads it off, which is naming which theorem a figure is about. Flat-foldability, rigid-foldability and developability are three different properties with three different tests, and the discipline of saying which one is meant is what makes it obvious when a claim has drifted to a fourth thing entirely.

What is actually shared #

Something real does survive the comparison, and it is a matter of method rather than of results.

Both fields discovered that the useful question is not “what is the folded state” but “what constrains the space of possible folded states”. Both found that the constraints available are cheap and incomplete. And both spend most of their effort on the gap between what the constraints permit and what actually occurs.

That is a genuine kinship and it is a kinship of epistemic situation rather than of geometry. It is also the reason the two literatures are worth reading in each other’s company, provided nobody carries a theorem across.

A third shared feature is worth adding because it is the one this site is best placed to see. Both fields have a version of the problem where the answer is known to exist and getting to it is the difficulty — a native structure that a protein reaches every time, a folded state a pattern demonstrably has — and in both cases the existence of the answer is what makes the search question interesting rather than what makes it easy. The site has a whole essay on a state no motion reaches, which is the sheet version of exactly that gap, and reading it beside a discussion of folding kinetics is more productive than reading any theorem across.

What the picture cannot show #

The hero figure plots two counts on one axis and the comparison is honest only in a narrow sense.

The two counts are counting different things. Two-to-the-k assignments is a count of labelled patterns, most of which do not fold; the chain’s count is of geometric configurations, most of which are physically absurd because they self-intersect. Neither number is a count of anything a physicist would call a state, and putting them on one axis is a comparison of growth rates rather than of quantities.

The figure is therefore about the shapes of two curves and about the filter marked on one of them, and the numbers themselves should not be read across. That is an unusual thing for a figure on this site to admit, and the alternative — drawing one curve and describing the other in prose — would have made the comparison less checkable rather than more honest.

The idealisation, named #

The chain’s states are treated as a fixed small number per joint, which is a caricature of a torsional energy landscape with broad minima and barriers between them. Real conformational counting is a much more careful business.

The sheet’s assignments are treated as two per crease, which ignores that a crease can also be flat — unfolded — and that a real pattern has a continuum of fold angles between the flat state and the closed one. This site’s own rigid-folding machinery works in that continuum and the assignment count is the flat-state shadow of it.

Both simplifications are in the direction of making the counts smaller than the truth, and neither changes the locality argument at all, which is the point of the essay.

Where this ladder goes next #

The molecular-folding ladder ends here with a negative result, which is an unusual thing to build a rung on and the right shape for this one.

The field has two rungs left and both look back at it. The same corrugation appears in four unrelated places — an argument about geometry recurring where lineage does not — and it is worth reading against this essay, because it is the case where a shared shape does mean something and the reasons it does are precisely the reasons the folding analogy fails.

Then the field’s own account of what it does not know, which is where the rule every figure here was built to gets stated as a table rather than as a preamble.

The last thing to carry out of this rung is a question rather than a conclusion. If locality is a property of the representation rather than of the system, then the interesting research question for any hard folding problem is not how to search its space better but whether some other description of the same object admits a cheap local test. That is what the lattice abstraction did for scaffold routing, and it is what the vertex conditions did for sheets long before anybody framed them that way. Whether a chain has such a description is not something this site can answer, and it is a better question than which of the two problems is harder.