A graph has no numbers at all

Every field the survey has looked at splits its quantities into sizes and shapes, with counts in a third class. This one is entirely the third class, and the reason is that it has no parameters.

There is nothing to scale #

It is worth being sure that is not a trick of how the field chooses to describe things. A mechanism plainly has lengths, and a chain is a mechanism’s description with the lengths removed — so is the field simply declining to look at half its object?

No, and the reason is that the questions the field asks genuinely do not involve the lengths. How many distinct chains have eight links and one freedom; whether two drawings are the same chain; which links can be grounded to give distinct mechanisms; how a chain decomposes into Assur groups. Every one of those has the same answer for every set of lengths, so the lengths are not being ignored, they are irrelevant.

Ask the scaling probe for a chain’s exponent and it returns nothing, correctly — every quantity it computes is unchanged under a scaling and also unchanged under any perturbation whatever, because there is no continuous parameter to perturb.

The field’s whole output is integers, and the integers are exact.

One four-link chain. Two six-link chains. Sixteen at eight links. Two hundred and thirty at ten. Those numbers are not measurements of anything and could not be improved by a better instrument.

A chain is which links are joined to which. It has no lengths, so a scaling has nothing to act on; it has no angles either, and no positions.

The field’s whole output is integers, and the integers are exact.

A spectrum is numbers and is not a measurement #

One apparent counter-example is worth clearing, because the field does compute lists of real numbers.

A graph has an adjacency spectrum: the eigenvalues of its adjacency matrix. Those are real numbers, they are not integers, and the site computes them as part of telling two chains apart quickly.

They are not measurements and they have no units. They are functions of a finite combinatorial object, computed exactly, and the same chain always gives the same list to machine precision. Perturbing them is meaningless because there is nothing to perturb.

So a real number is not automatically a measurable quantity, and has an exponent under a scaling is a sharper test than is not an integer. A graph spectrum reads zero under the scaling probe and zero under the control, exactly like a mobility, and lands in the same class despite not being an integer.

A quantity’s class is decided by how it responds to perturbations of the machine, not by whether it happens to be a whole number. That is worth having explicitly, since three of the four classes here have so far been introduced with integers as their examples.

Where the evidence comes from instead #

A field with no measurements still has to be right, and the shape of its evidence is different from every other field’s.

A canonical form. Two chains are the same chain when one is a relabelling of the other, so the question are these two chains one is answered by computing a labelling-independent form of each and comparing. That is exact, it is decidable, and it either works or is provably wrong.

An enumeration. The count of chains at a given link count is obtained by generating every candidate and deduplicating. That is exhaustive rather than sampled, so the answer is a proof rather than an estimate.

And a refusal. The site’s own enumeration is checked by feeding it two chains that are the same drawn differently and requiring them to be identified, and two that differ in one edge and requiring them to be separated.

None of that is a measurement. It is the discipline of a combinatorial argument, and its failure modes are a wrong canonical form or an incomplete generation — both of which produce a wrong integer with no residual, no error bar and no signal.

The other fields’ counts are not like this #

Putting this field’s counts beside the site’s others sharpens what is unusual about them.

A mobility is an integer computed from a mechanism’s parameters. It is locally constant, it has a boundary somewhere in parameter space, and the distance to that boundary is a measurable margin.

A Grashof class is the same: an integer with a continuum underneath it.

A chain count has no parameters at all. There is no continuum, no boundary and no margin. It is not locally constant, it is constant, and the only thing that could change it is a different question.

That puts the field in a fourth position rather than in the count essay’s third. A count derived from parameters and a count with no parameters behave differently, and the difference is exactly whether a margin exists to be reported.

It also decides what a mistake looks like. A count derived from parameters can be wrong because the parameters were near a boundary; a chain count can only be wrong because the enumeration was.

Errors of a kind measurement never sees #

That is worth dwelling on because it is the field’s real hazard and it is unlike anything else on the site.

Every other field’s errors show up as a number being slightly wrong: a residual above the noise, a routes-disagreement, a sensitivity that does not match a finite difference. This field’s errors are integers being wrong by one, and an integer wrong by one looks exactly like an integer.

The only defence is a second route, and the field’s second routes are the same ones every combinatorial argument uses: count a different way, check against a published table, verify a known small case exhaustively.

That is a weaker kind of evidence than a residual at 10⁻¹⁶ and it is the best available, and it is worth saying so plainly rather than letting the field’s exactness suggest more certainty than it has. A number that is exact is not thereby a number that is right, and the two are easily conflated when there is no error bar to draw attention to the difference.

That is worth dwelling on because it is the field’s real hazard and it is unlike anything else on the site.

The site’s own instance is instructive: an enumeration reaching 8,494 labelled graphs at eight links to produce sixteen distinct chains, and the cost being dominated by the group it is quotienting by rather than by the size of the answer. A generation that missed a case would produce fifteen, and nothing about fifteen would look wrong.

That is a weaker kind of evidence than a residual at 10⁻¹⁶ and it is the best available, and it is worth saying so plainly rather than letting the field’s exactness suggest more certainty than it has.

Neither of those contains a number an instrument could read. The next one does contain numbers, which makes it the useful test of the claim rather than another illustration of it: a graph’s spectrum is a list of real values, it is exquisitely sensitive to the structure, and it is still not a measurement of anything, because it is computed from the adjacency and carries no units.

What a measurement of a mechanism recovers about its chain #

The answer is: all of it, trivially, and the triviality is the point.

Look at a mechanism and count its links and joints. That is the chain, completely, and no instrument is needed. Any observation that distinguishes the mechanism from a different one at all determines its chain, because the chain is what the mechanism visibly is.

So this field is at the extreme end of a spectrum the whole survey has been walking along. At one end are the bodies field’s clearances, every one a length, none recoverable from angles. In the middle are the linkage fields, part shape and part size. At this end is a field whose entire content is recoverable from a glance.

A count is free, as the count essay puts it, and a field made entirely of counts is a field a measurement cannot fail at.

The one thing an observation could get wrong #

Saying a chain is recoverable from a glance overstates it slightly and the exception is worth naming.

What a glance gives is the links and the joints as they appear. What can differ from that is whether two apparently separate links are one part, whether an apparent joint is a joint or a fixed connection, and whether a joint that looks like a pin is a pin.

Every one of those is a modelling decision rather than a measurement, and every one changes the chain. Three parallel bars drawn as five links and six pins is a different chain from the same assembly with two of the bars taken as one part, and the mobility formula gives different answers for the two.

So the chain is recoverable from an observation once the modelling is settled, and settling the modelling is not an observation. That is the third limit of the whole survey arriving in the field with no parameters — a model nobody thought of, here in the form of a part somebody counted as two.

A field with no measurement error still has modelling error, and in this field modelling error is the only kind there is.

And a margin does not exist #

The count essay’s recommendation — report a count with the margin of the continuous quantity underneath it — does not apply here, and the exception is worth stating because it is the only one.

A mobility has a margin: the rank decision’s gap. A Grashof class has a margin: the distance to s + l = p + q. A restraint count has a margin: how far a contact can move.

A chain’s count has none. There is no continuous quantity underneath it, no boundary to be near, and no perturbation that would change it. Sixteen eight-link chains is sixteen, and the only way for it to be wrong is for the enumeration to have been wrong.

So the field’s counts are in a fourth position again: counts with no margin, because there is no continuum for a margin to live in. They are the most certain numbers on the site and the least checkable, and both follow from the same absence.

Sixteen, and how it is known #

The field’s most quotable number deserves a paragraph on where it comes from, since it is the archetype of what this essay is about.

Sixteen eight-link chains with one degree of freedom. The route is: generate every labelled graph with the right degree sequence, reduce each to a canonical form, and count the distinct forms. At eight links that is 8,494 labelled graphs producing sixteen distinct chains.

The ratio of 8,494 to 16 is the field’s own gotcha: an enumeration’s cost is usually the group it is quotienting by rather than the size of its answer. Five hundred labelled graphs per distinct chain is the symmetry group of the relabelling being divided out one candidate at a time.

At ten links the same route reaches two hundred and thirty and, on this site’s own implementation, does not return in reasonable time — which is recorded rather than worked around.

So the number sixteen is a proof and the number two hundred and thirty is quoted from elsewhere, and the two have different standings. A field whose evidence is exhaustive is a field whose evidence stops when the exhaustion becomes infeasible, which is a limit of a completely different kind from anything measurement imposes.

Where the numbers do come in #

A chain acquires numbers the moment it is given lengths, and that transition is the field’s own boundary.

Choosing the chain before the lengths is the field’s central practical argument: the search space of mechanisms factors into a discrete choice and a continuous one, and the discrete choice is small enough to enumerate while the continuous one is not.

So this field hands its answer to another field, and the handover is where the sizes and shapes appear. A chain plus lengths is a mechanism; a mechanism has parameters; the parameters have exponents; and everything the survey says elsewhere applies from that point on.

The factorisation is what makes the whole subject tractable, and it is a factorisation into a part with no numbers and a part that is all numbers.

What the tolerance field would say about a chain #

A useful thought experiment, because it shows the boundary from the other side.

The tolerance field takes a mechanism’s parameters, gives them ranges and asks how far the output moves. Hand it a chain and it has nothing to work with: no parameter, no range, no output that varies.

The right answer is that a chain has no tolerance, and it is worth saying because a designer might reasonably ask how sensitive is this to the topology and the question has no meaning. A chain is not sensitive to anything; a different chain is a different machine.

What is sensitive is the boundary between chains, and the boundary lives in the mechanism’s parameters rather than in its graph. A four-bar whose coupler length goes to zero becomes a three-link chain; a mechanism whose two pins coincide becomes a different graph. Those transitions are continuous in the parameters and discrete in the graph, and where they happen is a question about the parameters.

A chain’s robustness is a property of the mechanism that realises it, which is the general statement of why a count with no continuum still has a boundary somewhere — the boundary is one level down, in the field the chain hands its answer to.

The field where the question does not apply #

A survey that asks one question of every field will eventually meet the field where the question has no meaning, and finding it is informative rather than a waste of a chapter.

There is nothing here to scale. A chain is which links are joined to which, it has no continuous parameters at all, and so the scaling probe returns nothing, the tolerance field has nothing to perturb, and the whole apparatus of identification has no quantity to be uncertain about. A reader who has followed the survey through eight fields and expects a ninth row of exponents finds an empty one, and the emptiness is the result.

What that marks is a boundary inside the subject rather than a gap in the work. On one side are twenty-four fields whose objects have parameters, share a structure, and answer to the same three questions — what scales, what a measurement recovers, what a tolerance does. On the other is one field whose objects are combinatorial, where every answer is an integer, and where the only failure available is an enumeration that missed a case.

The line between them is exactly where a chain hands over to a mechanism. A chain plus lengths is a machine; a machine has parameters; the parameters have exponents; and everything the survey says elsewhere applies from that point on and not one step before it.

That handover explains why the subject factors the way it does, and the factoring is not arbitrary. Choosing the chain is a discrete search over a finite set with no numbers in it, and choosing the lengths is a continuous optimisation with every number in it. They are different kinds of problem, they want different tools, and doing them in the wrong order — picking lengths and hoping a chain emerges — is a mistake this site has an essay about.

The evidence here is of a different kind too, and it is weaker. A residual at 10⁻¹⁶ says a construction closed. Nothing in this field produces one. An enumeration reaching 8,494 labelled graphs at eight links to yield sixteen distinct chains is defended by counting a second way, checking a published table, and verifying small cases exhaustively — and a generation that missed a case would return fifteen, with nothing about fifteen looking wrong. There is no margin, no conditioning, no perturbation that reveals the error, because there is nothing continuous for it to live in.

So the honest statement is that this field’s exactness is not the same kind of certainty as the rest of the site’s, and it is worth saying plainly rather than letting the integers imply more than they carry. An exact answer and a well-evidenced one are different properties, and this is the one field where a result can easily have the first without the second.