The claim that survives

This site spends most of its length taking claims apart, which risks leaving an impression that the subject is a pile of errors. It is not. One of the three famous assertions about it is correct, and it is correct in a form sharper and more interesting than the version usually repeated.

The quantity #

How well a number x is approximated by fractions is measured by how small q·|qx − p| can be made, taking the best p for each denominator q. A number with good approximations gets that quantity very small; a number that resists them keeps it large.

Hurwitz proved in 1891 that for every irrational number the quantity can be pushed below 1/√5 = 0.4472, and that the golden ratio is the number for which it cannot be pushed any lower. It is the extreme case, and the bound is attained by it and by nothing else that is not equivalent to it.

Measured here, the golden angle scores 0.4377 — the sampling floor approaching the theoretical value — against 0.3306 for the best of fifteen hundred other angles across a seventy-five degree window.

That is not a narrow win. It is a different regime.

Why the sweep cannot find it #

A detail worth noticing, because it is the same difficulty the subject has in miniature.

The sweep steps in twentieths of a degree and therefore consists entirely of rational angles. Every one of them, being rational, has a denominator at which the approximation is exact and the score collapses — so the sweep can never land on the answer, and the best it reports is whichever grid point happens to be furthest from a low-denominator fraction.

The golden angle had to be evaluated exactly and inserted. Anything else would have been measuring the grid.

That is a small methodological point with a large moral: the property being measured is invisible to any finite sampling of the parameter, which is exactly why the packing measurements at fixed head size could not find it either.

Why this is the right form of the claim #

Three reasons.

It is exact. Hurwitz’s theorem is a theorem. There is no head size at which it stops applying, no criterion under which it gives a different answer, and no dependence on how the measurement is made.

It explains the phyllotaxis. The parastichy numbers are the denominators of the good approximations, so an angle with no good approximations at any denominator never develops the rows that a good approximation produces. That is what keeps the gaps bounded, and it is the mechanism the packing story gestures at without stating.

It explains the extremeness. Every irrational angle eventually fills the disc; the golden angle is the one that never has a bad phase on the way. That is a statement about all scales at once, which is why no measurement at one scale can show it.

What it does not claim #

It does not claim a plant computes it, or benefits from the last decimal place, or would be measurably worse at 137.3°.

The biological version is much weaker and is enough: a meristem that places each primordium away from its neighbours is pushed away from angles whose multiples pile up, and the golden angle is where that pressure has nowhere left to push. The theorem says why the attractor is where it is. It does not say that a plant needs to be there.

Confusing those two is how “the most irrational number” acquired its air of mysticism. The arithmetic is remarkable; the plant is just avoiding its own neighbours.

The three claims, together #

The nautilus is a golden spiral. Out by a factor of 2.14, and the one free choice in the measurement cannot account for it.

Sunflower spirals are always Fibonacci. True of a branch, not of plants — the Lucas branch gives 47 and 76, and the same model reaches it.

137.5° is optimal. Right, once the criterion is named. Not as packing on a whole head, where the spacing criteria are decided by the first handful of organs and area evenness prefers rational angles, but as resistance to rational approximation, where it is the extreme case and there is a theorem — whose geometric form is the closest pair a noble head keeps away from its centre.

One in three is a better record than this subject’s reputation suggests, and the one that survives is the one worth telling.

Continued fractions, and why they are the right tool #

The quantity is defined by a search over all fractions, which sounds expensive. It is not, because the best approximations to a number are its continued fraction convergents and nothing else — a theorem, and the reason continued fractions are the natural language for this question.

Writing a number as

$$x = a_0 + \cfrac{1}{a_1 + \cfrac{1}{a_2 + \cfrac{1}{a_3 + \dotsb}}}$$

the partial quotients aᵢ control everything. A large partial quotient means the next convergent is a very good approximation: truncating just before a large aᵢ leaves a tiny remainder. A number with a large quotient somewhere is well approximated by a fraction with a modest denominator.

So the number hardest to approximate is the one whose partial quotients are all as small as possible, and the smallest a partial quotient can be is 1. The number with every quotient equal to 1 is

$$\phi = 1 + \cfrac{1}{1 + \cfrac{1}{1 + \dotsb}}$$

which is the golden ratio, and its convergents are ratios of consecutive Fibonacci numbers. There is nothing left to tune. That is why the extreme case is φ and not some carefully chosen alternative — it is the boundary of the parameterisation rather than a point within it.

What “noble” means #

Numbers whose continued fraction tails are all 1s are called noble, and they form the equivalence class attaining Hurwitz’s bound. φ is one, and so is anything of the form (aφ + b)/(cφ + d) with ad − bc = ±1, which includes (1+√5)/2 shifted and scaled in many ways.

That has a consequence for this site’s subject that is easy to miss: the golden angle is not the only maximally irrational divergence angle. The Lucas angle, which the model reaches on its second branch, is also noble, and its spiral counts are Lucas numbers for the same arithmetic reason that the golden angle’s are Fibonacci.

So the correct statement is not “φ is uniquely good”. It is “the noble numbers are the maximally resistant class, φ is its simplest member, and the dynamics reaches φ from most starting conditions and the others from some”. Every part of that is checkable and the last part is the part that involves biology.

Why the sweep could not find it #

An implementation detail worth stating, because it is the kind of thing that silently makes a result meaningless.

The comparison sweeps fifteen hundred angles across a window and evaluates the quantity at each. Every one of those grid points is a rational number — a floating-point value is a dyadic rational — and a rational number’s approximation quantity is zero, since it is exactly equal to a fraction.

The measurement therefore does not evaluate the quantity at any irrational at all. It evaluates a truncated version — the smallest q·|qx − p| over denominators up to some limit — which for a rational with a large denominator behaves like an irrational until the search reaches it.

That works, but it means the golden angle could never appear as a grid point, and a sweep that just happened to include a value near it would report a number driven by that value’s own denominator rather than by proximity to φ. So the golden angle is evaluated separately, from its exact definition, and inserted into the plot.

The figure shows it as a distinct mark for that reason, and the distinction is not cosmetic. It is the difference between a comparison and a coincidence.

The shape of a claim worth keeping #

Three things separate this claim from the two that fail.

It has a stated criterion. “Optimal” without a criterion is not a claim, and the packing measurements show what happens when four plausible criteria are supplied to the same question: the two about cells are won by rational angles, and the two about distance rank every noble angle alike.

It has a theorem. Hurwitz’s bound is not an empirical observation about sampled angles; it is a proof that nothing does better, and the measurement here is a demonstration rather than evidence.

And it has a mechanism that reaches it. A property being optimal is not an explanation of why an organism has it. The dynamical model produces the angle from a local rule with no reference to arithmetic, which closes the gap between “this value is special” and “this value occurs”.

Claims about pattern in living things usually have none of the three. This one has all three, which is why it is the one that survives.

What the three claims have in common #

Looking across them, the difference between the one that survives and the two that fail is not that the first is about a truer fact. It is structural.

The nautilus claim has no criterion attached: “the shell is a golden spiral” does not say in what sense, and when a sense is supplied — the growth factor per turn — it is wrong by a factor of two.

The Fibonacci claim has no scope attached: “sunflower spirals are Fibonacci numbers” does not say which sunflowers, at which radius, or what happens otherwise, and every one of those turns out to matter.

This claim has both. The criterion is resistance to rational approximation. The scope is all irrationals, with a theorem covering them. And the number that satisfies it is reached by a mechanism that is not told about it.

The lesson generalises past this subject. A claim about pattern in nature is worth something in proportion to how much of it can be wrong, and the two failing claims are so loosely stated that there was never anything for the evidence to contradict.

The convergents, and how fast they are bad #

One more measurement makes the extremity concrete.

The convergents of φ are 1/1, 2/1, 3/2, 5/3, 8/5, 13/8, 21/13 — consecutive Fibonacci ratios. The error of the qth convergent is about 1/(√5 q²), and the constant √5 is the largest possible: for any other irrational, infinitely many convergents beat it.

Compare with a number having a large partial quotient. The continued fraction of π is [3; 7, 15, 1, 292, …], and the 292 means the convergent just before it — 355/113 — is extraordinarily good, matching π to seven digits with a three-digit denominator. π is in that sense easy to approximate; φ has no such moment anywhere in its expansion, because there is no partial quotient bigger than 1 to produce one.

Translated into lattices: a divergence angle at 355/113 of a turn would produce a pattern that is essentially 113 rays for any head with fewer than about 12,000 primordia. The gap essay is that effect measured. The golden angle’s worst case is the mildest available, at every scale.

What it would take to overturn this one #

A claim worth keeping should come with the conditions under which it would fail, and this one has them.

It would fail if the criterion were shown to be the wrong one — if some other property of a divergence angle turned out to drive the biology and the golden angle were merely near-optimal for it. That is a live possibility rather than a formality, and the way to settle it is to state the competing criterion and measure both.

It would fail if the dynamics did not reach the value. The arithmetic result stands alone as mathematics, but its relevance to plants depends on a mechanism arriving at it, and a model that landed somewhere else would leave the theorem true and beside the point.

It would fail as an explanation if the branch structure were richer than measured — if plants turned out to be spread across many noble angles rather than concentrated on two branches, the maximality of φ would stop distinguishing anything.

None of those has happened. Listing them is what separates a claim from a slogan, and the two claims this site rejects have no such list because there is nothing specific enough in them to fail.

Why the arithmetic version is also the more interesting one #

There is a version of this essay that would treat the surviving claim as a consolation prize — two famous assertions demolished, one left standing. That gets it backwards.

The packing version, in the form that survives measurement, is the same fact read off a head: away from the centre, the closest pair every noble head keeps is √(2/√5), Hurwitz’s constant under a square root. Read only as geometry — this arrangement fits more in — it would have been pleasant and not surprising, since something has to be best.

The arithmetic version is a fact about numbers being approximable, which has nothing to do with plants, discs, or space at all. Hurwitz proved it in 1891 about the real line. That the extreme case of a theorem in Diophantine approximation should be the value a greedy local rule about repulsion converges on, and that the value should then be visible in the number of spirals on a sunflower, is a genuinely odd chain, and every link in it is checkable.

Popular accounts reach for wonder and land on the golden ratio being everywhere, which is not true. The thing that is true is stranger and takes three paragraphs, and this site is an argument that the three paragraphs are worth it.

The other noble angles, and what would find them #

A prediction this site can state and not test, which is worth recording as such.

If the branches are noble numbers and the basins order their frequencies, then beyond the Fibonacci and Lucas branches there should be a third, at a noble angle with a smaller basin, producing a sequence with no common name — the one anomalous series that the botanical literature reports occasionally and files as error.

Finding it would need angles recovered from a large number of specimens, with radii recorded, and a histogram. The prediction is that the histogram has thin peaks in specific places rather than a spread around 137.5°, and the places are computable in advance.

That is what a claim with a mechanism buys: not just an explanation of the common case, but a statement about what the rare cases should be.

The same fact from the other end #

Expansion found a second place where this angle turns out to be unreachable by rationals, and it arrived from the opposite direction.

The sweep in this essay could never land on the golden angle because every grid point is rational and every rational collapses the lattice — the value had to be evaluated exactly and inserted by hand, which reads as a limitation of the method.

On a cylinder, the rationals are not an artefact of a grid. The branch points of the lattice tree — the configurations where three spiral families are equally short and the lattice is exactly equilateral — sit at divergences of 5/14, 15/38, 37/98, 99/258, 257/674 of a turn. Every one is rational, with denominator 2(m² + mn + n²), and the sequence converges on 137.50776° without ever reaching it.

So the geometry’s own distinguished points are rational, and the golden angle is their accumulation point. It is at no fork.

Which sharpens the claim this essay makes rather than softening it. The angle is not where the geometry puts a special configuration; it is the limit of where the geometry puts them, and its distinction remains the arithmetic one — that it resists rational approximation better than any other number, which is why a lattice built on it never develops the rows a near-rational one does.

And an eighth criterion that does not #

The disorder of the tissue invites a repair of the folklore that would at last give the golden angle a geometric criterion: if rationals are ordered, the most badly approximable angle should be the most disordered. It is testable on the Farey interval between 8/21 and 5/13, where the golden angle is the noble number.