What universality is worth

Thirteen rungs ago this field was handed a polynomial and asked for a mechanism. It has one. This rung says what that is worth.

What was established #

The identity. A polynomial in the coordinates of a two-link arm’s tip is a constant plus a finite sum of cosines of whole-number combinations of the arm’s two angles, checked against a direct evaluation at twenty-seven hundred random angle pairs across nine curves, worst disagreement $1.8\times10^{-14}$.

The gadgets. Four constructions — a rigid offset, a rhombus taken as a bisector, a rhombus held on a line, a parallelogram — meeting their closed forms to $2.7\times10^{-13}$ across their sweeps. From them: negation, doubling, addition, and every whole-number multiple.

The machine. Nine curves compiled, assembled, driven, and each one’s tracing point on its own curve to $10^{-11}$ or better over its working arc, checked by evaluating a polynomial the mechanism’s constraints never mention.

The cost. Five bars for a line, four hundred and thirteen for a quintic, growing as the fourth power of the degree, three quarters of the largest machine spent carrying angles rather than computing them.

The failures. Sixteen assemblies of a twenty-bar machine, four of which draw a different curve at the same closure residual; braces that remove them at one redundant equation each; and singular configurations that no bracing removes and that end every machine’s arc.

What the theorem said, and what it did not #

The classical statement is that every plane algebraic curve is traced, in a neighbourhood, by a point of some linkage. It is true, and everything above is consistent with it.

What it does not say is anything about size. A statement of the form there exists is silent about how large the thing is, and the silence is not an oversight — a bound would be a different theorem, and the bounds that have been proved since are enormous.

What it does not say is how much of the curve. In a neighbourhood is doing a great deal of work in that sentence, and this field measured what the neighbourhood is on nine specific machines. The answers run from two radians of driving angle down to a tenth.

What it does not say is which component. A linkage’s configuration space has components, the construction fixes only the bar lengths, and Kempe’s original argument attended to one component and not the others — which is the gap that later work closed.

None of that makes the theorem less true. It makes it a theorem about existence, which is what it was always advertised as, and the point of building one of the objects is that existence proofs do not tell anybody what the objects are like.

The gap between a proof and an artefact, in one table #

The three columns most worth putting together are the ones the theorem does not have.

A line: two terms, five bars with a slide or twelve all-revolute, two radians of arc. A machine somebody could make, drawing an exact straight line, and beaten in practice by four bars that are wrong by nine per cent.

A lemniscate: five terms, fifty bars, $1.3$ radians. Still drawable on paper. Not something anybody would fabricate.

A general quintic: eighteen terms, four hundred and thirteen bars, $0.17$ radians — and braced, so that it does not have assemblies drawing other curves, one thousand and ninety-six unknowns rather than four hundred and eighty.

The progression is the whole story. The construction stops being a mechanism somewhere between a conic and a cubic, and everything past that point is a proof written in a notation that happens to have bars in it.

That is not a criticism of the construction. It is the answer to a question the theorem does not ask, and it took building nine of them to find out where the line falls.

The three numbers worth carrying #

If this ladder leaves three figures behind, they should be these.

Four hundred and thirteen bars for a general quintic. That is what universality costs at the smallest degree at which anybody would want it. It is not a machine; it is an argument in the shape of a machine.

Three quarters of it is transport. The size is not caused by the algebra being hard. It is caused by a linkage having no way to refer to a quantity computed elsewhere, so that every reference is built out of matter, one parallelogram per unit of distance.

A tenth of a radian. The largest machine draws about ten degrees of its curve before a gadget reaches a configuration where two of its placements merge and it comes out computing something else.

Why the construction is nevertheless a different kind of answer #

Against all that, it would be easy to file this field as a curiosity, and there is one respect in which that would be wrong.

Every other answer this site gives to find me a mechanism that does this is a search. Burmester’s construction takes three positions and returns a finite set. An optimiser takes a sampled path and returns whatever it found from wherever it started. A census takes a chain and returns the linkages that could realise it. In each case something has to be looked for, the looking can fail, and whether it succeeded is a question with its own machinery — defect tests, root counts, objective functions.

Compilation is not a search. There is no starting guess, nothing to fail, and nothing to check afterwards except that the result is what the procedure says it is. Handed a polynomial, it produces a machine in under a second, and the machine draws the curve.

That is a different relationship between a demand and a mechanism, and it is the reason the field is not a footnote to synthesis. Synthesis asks which mechanism is closest. Compilation asks which mechanism is this, and the answer is a construction rather than a candidate.

What an existence proof is for #

There is a way of hearing all of this as a complaint about the theorem, and it would be the wrong reading, so it is worth setting out what a result of this shape is actually doing.

Before it, the honest position on can a linkage draw this curve was: nobody knows, and every case that has been settled was settled by somebody clever. Peaucellier’s cell took the nineteenth century decades to find and was reported as news. The straight line had been an open question in the subject for a hundred years, and a great many people had produced approximations while assuming that an exact answer might not exist.

After it, the question is closed for every algebraic curve at once, and closed in the direction that stops people looking for impossibility arguments. That is worth a great deal, and it is worth exactly as much whether the linkage has ten bars or ten million.

An existence proof retires a question. What it does not do is answer the question anybody actually had, which was usually what is the mechanism, and the gap between those two is where a construction like this one lives. Building one is how the gap gets measured.

The measurement that could not have been guessed #

Of everything in the four hundred and thirteen, one number was not predictable from the outside, and it is the one worth having built the thing for.

That most of a compiled machine is transport is not something anybody would infer from reading the construction. The construction reads as a sequence of clever angle gadgets, and its arithmetic is where the interest and the ingenuity are; the parallelograms are the boring part, mentioned in passing. Counting them says the boring part is three quarters of it, and the fraction is still rising.

That reverses where a reader’s attention should go. Anybody trying to make a smaller universal construction would naturally attack the arithmetic — better multipliers, fewer reflectors, cleverer sharing — and would be optimising a quarter of the machine whose other three quarters grow twice as fast. The place to attack is transport, and transport is not a feature of this construction but a property of what a bar can say.

A measurement changed which half of the object is interesting. That is the ordinary justification for building something that already had a proof.

Where the two meet #

They meet at exactly one place, and it is worth stating because it is the boundary this field’s slate was drawn against.

A four-bar’s coupler point traces a curve, and that curve has an implicit equation of degree six — a fact the neighbouring rung measures rather than quotes. So an ordinary four-bar is, read backwards, a compiled machine for a particular sextic: four bars where the compiler would use hundreds.

The comparison says what the compiler is missing. It handles every curve the same way, so it cannot notice that a particular sextic is a four-bar’s own curve and needs four bars. A universality construction is a procedure that never recognises anything, and its cost is the price of that.

Which means the interesting question is not can a linkage draw this curve — it can — but what is the smallest linkage that draws it, and nothing in this field answers that. It is a genuinely open problem, the compiler’s answer is an upper bound, and on the one case where a better answer is known the upper bound is a hundred times too big.

The three instruments, in retrospect #

A field that ends by counting what it has should count its instruments too, because the choice of them was the decision that made the rest possible.

The closure residual was inherited. Every field on this site reads it, it is the norm of the constraint equations at a converged position, and it says whether the bars are the lengths they were made. Here it says that on every branch, right and wrong, and is therefore the field’s least informative number.

The curve residual is the specification, checked, and it is the one that carries the claim. Its virtue is complete independence: the machine’s constraint list does not mention the polynomial, so evaluating the polynomial at the machine’s own tracing point is a question from outside the mechanism entirely.

The departure is the finest, and it was built for this field. It compares each gadget’s actual output angle against the angle its specification asks for, given the machine’s own two arm angles, and it names which part of a four-hundred-bar machine has gone wrong and at which driving angle.

Building the third turned up a false alarm worth carrying: a reflector’s far vertex crosses the pivot and comes out on the opposite ray with the gadget working perfectly, so the first version reported a departure of $\pi$ on half the catalogue. Every machine it flagged was drawing its curve to $10^{-13}$. An instrument that is wrong about a working machine costs more than no instrument, because there is no prior reason to suspect the instrument rather than the machine.

What this field did not take #

Three things it would have been natural to take and did not, recorded here so that they read as open rather than as done.

The configuration-space direction. The stronger modern results are not about curves but about spaces: every compact real algebraic set is, up to diffeomorphism, the configuration space of some linkage. That is a bigger statement in a different direction, it connects to the space of configurations the serial field already draws, and nothing here goes near it.

The spurious curve’s own equation. Past the end of its arc a machine draws a definite curve, and which one is answerable: it is the vanishing set of the polynomial with one term’s frequency altered. Nothing here computes it.

A predicted arc. Every gadget’s singularities are known in closed form, so a machine’s working arc could be predicted from the term angles and then confirmed by driving. This field measures it and does not predict it, which leaves the site’s habit of two independent routes unsatisfied on that one quantity.

The shape of the result, stated once #

A construction that works for everything works for the easy cases the same way it works for the hard ones, which is to say badly. That sentence is the whole of what building this taught, and it is not specific to linkages.

The circle is the cleanest demonstration. Handed $x^2 + y^2 = r^2$, the compiler produces eleven bars, and eleven bars is a two-link arm with a fixed elbow — a rigid body on a pivot, which is a crank, which is one bar. The compiler did not recognise a circle; it multiplied out a polynomial, found one surviving frequency pair, and built the apparatus that one term calls for. That the apparatus turned out to be a crank in disguise is a fact about the answer rather than a step in the method.

A general procedure has no way to be clever about a special case, because being clever about special cases is exactly what makes a procedure not general. The eleven bars are the price of not having to think, paid on a case where thinking would have cost nothing.

What the field cost, in the site’s own terms #

One last accounting, which belongs in the field that carries it.

This field is one library, thirteen figure generators and a set of gate assertions, and every number in every essay comes off an object that was built and driven. Nine machines were compiled, from five bars to four hundred and thirteen; every one of them was positioned by the same Newton–Raphson solver every other field on this site uses, on the same analytic Jacobian, judged against the same closure tolerance.

That matters for a specific reason. A field about universality could very easily have been a field about algebra with pictures attached — the identity is exact, the gadget relations are exact, and the whole thing could be drawn from closed forms without a solve anywhere. It would have looked identical.

What the solve buys is the second half of the field: the branches, the singularities, the arcs. None of those exists in the algebra. They exist in the mechanism, and they are found by asking a mechanism to move and watching what it does — which is the site’s founding habit and is why nothing here is drawn that was not solved.

What a reader should leave with #

Universality is real. A curve given as an equation can be turned into a linkage by a mechanical procedure, and the linkage draws it exactly — not nearly, not at sampled points, but to the floor of the arithmetic at every position it has.

The linkage is also too large to build, works over an arc measured in degrees, has assemblies that draw other curves, and is beaten by four bars on the one case anybody has ever needed.

Both halves are the result. The first is the theorem and the second is what a reader gets by building one, and the reason for the thirteen rungs between them is that neither half means much without the other.

There is a version of this essay that ends by calling the construction beautiful and useless, and it would be half right. The useless half is measured above and is not in dispute. The other half is that a procedure turning an equation into a machine, with no insight anywhere in it, is a genuinely different relationship between a demand and a mechanism from anything else in this subject — and the fact that the machines it produces are absurd does not make the relationship less real.

A construction is worth what it establishes, not what it produces. What this one establishes is that the class of curves a linkage can draw exactly is not a special class at all: it is all of them, and it was never a question about mechanisms but about how much apparatus anybody was willing to build.