A constant that does not care which map

The period doublings come faster and faster, and the obvious thing to do with a sequence of intervals that shrink is to measure how fast.

The ratios are $4.74$, $4.64$, $4.65$. They are heading somewhere, and where they are heading is

$$\delta = 4.669201609102990\ldots$$

which is Feigenbaum’s constant. That much is a measurement. The surprise is what happens when the same measurement is made on a different map.

The same number from a different rule #

Take $x \mapsto r\sin(\pi x)$ instead. It has nothing in common with the logistic map beyond being smooth and having one hump: a different formula, a different shape, doublings at completely different parameters.

Its doublings happen at $0.7200$, $0.8333$, $0.8586$, $0.8641$ — no relation whatever to $3.0$, $3.449$, $3.544$, $3.564$. Divide consecutive gaps and the ratios are $4.48$, $4.60$, $4.65$: converging to the same $4.669$.

Do it for the tent map’s smooth relatives, for a cubic with one hump, for any of the maps a physicist might write down for a dripping tap. Same constant.

So $\delta$ does not belong to the logistic map. It belongs to the cascade, and the cascade is a thing that happens to many maps.

Why that is strange #

It is worth being precise about what is odd here, because “a constant appears in two places” is not by itself remarkable.

The parameters at which the doublings happen are completely map-dependent. There is no formula for them, they have no closed form, and two maps share none of them. What is shared is a ratio of differences of those parameters, in the limit — a quantity built entirely out of numbers that differ between the maps.

Compare it to $\pi$. That constant appears in many places because there is a circle hidden in each of them — even in a needle dropped on floorboards, where the circle is the needle’s own turning — so the appearances have a common cause a reader can point at. Here there is no shared object. The maps have different graphs, different fixed points, different attractors, different everything, and one number survives.

What the ratio is made of #

Before the explanation, it is worth being clear about which quantity converges, because three plausible candidates do not.

The doubling parameters themselves converge — to $r_\infty$ — and their limit is entirely map-dependent. The gaps between them converge, to zero, and how fast is what is at issue. What converges to something universal is the ratio of consecutive gaps:

$$\delta = \lim_{n \to \infty} \frac{r_n - r_{n-1}}{r_{n+1} - r_n}.$$

Each term of that sequence is built from three parameters that no two maps share, and the sequence converges to a number every map shares. Nothing about the individual terms hints at it: the first ratio for the logistic map is $4.74$ and for the sine map $4.48$, and those are as far apart as the maps are.

The distinction matters because it is the reason the constant was hard to believe. A shared limit of quantities that are not shared is not the kind of coincidence a numerical accident produces, and it is also not the kind of thing an argument about any one map could explain.

The mechanism: doing the same thing to a smaller picture #

The explanation Feigenbaum found in 1975 — on a hand calculator, at Los Alamos, having noticed the ratio while the numbers were still coming out — is that the cascade is generated by a process that repeats.

Look at a period-doubled window of the diagram and it contains a smaller copy of the whole diagram. That is not loose talk: applying the map twice gives a new map, and near the relevant part of the interval that new map is a rescaled version of a map of the same kind. The doubling cascade of $f$ is therefore the doubling cascade of $f \circ f$ shifted along by one, and the operation “compose with yourself and rescale” takes the family to itself.

That operation is called renormalisation. It has a fixed point: a specific function $g$ with $g(x) = -\alpha\,g(g(-x/\alpha))$, the same for every map in the family. And the slope test applies to it exactly as it applies to any other map with a fixed point — except that here the “points” are functions.

The renormalisation operator’s fixed point has one unstable direction, and its expansion rate along that direction is $\delta = 4.669$. That is the whole answer. The constant is the eigenvalue of a linearised operator at a fixed point in a space of functions, and it is universal because the fixed point is: every smooth one-humped map is drawn towards the same $g$ under repeated renormalisation, so every one of them approaches the cascade at the same rate.

The second constant #

There is a companion number that gets less attention and comes from the same place.

The doubling cascade shrinks in the parameter direction by $\delta$. It also shrinks in the value direction — the vertical splitting of the branches — by

$$\alpha = 2.502907875\ldots$$

which is the rescaling factor in the fixed-point equation above. Both are properties of $g$, both are universal, and together they say the whole diagram is self-similar with two different scale factors, one horizontal and one vertical.

That is what makes the bifurcation diagram a fractal in a precise rather than a decorative sense: it has a stated scaling law with measured constants, not merely a resemblance between its parts.

A fixed point in a space of functions #

The move that makes renormalisation work is worth naming separately, because it is the one that takes some getting used to.

Everywhere else in this field the objects being iterated are numbers, and a fixed point is a number the map leaves alone. Here the object being iterated is a function: renormalisation takes a map and returns another map, and its fixed point is a function that renormalisation leaves alone.

Everything else transfers unchanged. There is a space of objects — functions rather than points on an interval. There is a map on that space. It has a fixed point. Near the fixed point the map is approximately linear, and what happens under iteration is decided by the eigenvalues of that linearisation, exactly as a fixed point in the plane is decided by the eigenvalues of a Jacobian.

The fixed function $g$ has exactly one eigenvalue bigger than one, and it is $4.669$. Every other direction contracts, which is why every smooth one-humped map is pulled onto the same behaviour: the differences between the maps lie in the contracting directions and are wiped out, and the one direction that survives is shared.

That is also the honest answer to why the constant is universal, and it is not a statement about maps at all. It is a statement about an operator, and the maps are its inputs.

Measured before it was understood #

The order of events here is the interesting part of the history, and it is the reverse of the usual one.

Feigenbaum computed the ratio in 1975 and recognised that it was converging to something. He did not know why. The renormalisation argument came afterwards, as an explanation for a number that had already been measured — and the argument was not made rigorous for another seven years, by Lanford in 1982, in one of the first computer-assisted proofs after the four-colour theorem.

His paper was rejected twice before it appeared in 1978. The referees’ objection was reasonable on its face: a numerical coincidence between two maps is not a theorem, and the paper’s central claim was a constant with no derivation.

That is worth holding beside the four-colour proof nobody can read. Both results are true, both were established with machine assistance, and both took years to be accepted for the same reason — the mathematics community had no settled way to referee a claim whose support was a computation. Lanford’s proof needs interval arithmetic to bound the errors, and checking it means checking a program.

What the picture cannot show #

Every figure here shows finitely many doublings, and the constant is a limit.

The convergence is slow and geometric: each ratio is closer to $\delta$ than the last by roughly a factor of $\delta$ itself, so five doublings give about two correct digits and that is what the first figure reports. Getting six digits needs doublings at parameters that agree to twelve decimal places, and the period detector would have to distinguish a cycle of length sixty-four from one of length one hundred and twenty-eight in the presence of rounding. That is the same wall the sieve’s staircase runs into from the other side: more computation produces more data and no more understanding, and the understanding has to come from an argument.

So a picture of a cascade cannot show a universal constant. It can show a sequence of ratios that is not obviously going anywhere else, which is exactly what Feigenbaum had and what everybody had to be persuaded by.

There is a subtler limitation. Universality is a statement about a class of maps — smooth, one hump, quadratic maximum — and no drawing shows a class. Change the maximum from quadratic to quartic and $\delta$ changes to $7.28$; make the map non-smooth and the cascade may not happen at all. The theorem’s hypotheses are exactly the part a picture cannot carry, and they are not technicalities.

Where else the same idea went #

Renormalisation was not invented here. It came from statistical physics, where it had been developed in the 1960s to explain a different universality — the fact that water boiling, a magnet losing its magnetism and a binary alloy separating all have the same exponents near their critical points, despite having nothing physically in common.

The mechanism is the same in both settings: an operation that coarse-grains the system, a fixed point of that operation, and exponents that are eigenvalues at the fixed point. Feigenbaum’s contribution was to see that a one-dimensional map has the same structure, and the reason he saw it is that he had spent the previous years on critical phenomena.

The transferable claim is worth stating plainly. When many different systems share a number, look for an operation they all sit inside rather than an object they all contain. That is what distinguishes this from $\pi$ appearing in a needle-dropping experiment, where there genuinely is a circle.

Why the hypotheses are the theorem #

Universality classes are defined by what is shared, and the sharing here is narrower than the phrase “any map with a hump” suggests. Three conditions do the work, and each has a counterexample on the other side of it.

Smooth. A map with a corner at its maximum has a different cascade. The tent family doubles once and goes straight to chaos, with no infinite cascade at all, because the mechanism that produces the next doubling needs a curved maximum to rescale.

One hump. A map with two humps has a richer bifurcation structure and is not in this class. It has its own universality, with its own constants.

Quadratic maximum. This is the condition nobody expects to matter and it is the sharpest. A map whose peak looks like $1 - |x|^4$ rather than $1 - x^2$ has $\delta = 7.285$. The order of the maximum is the parameter that indexes the classes, and everything else about the map is irrelevant.

So “universal” does not mean “always”. It means constant across a class whose membership is decided by one local feature, and the value changes when that feature does. That is a much stronger and more useful claim than an unqualified one would be, and it is the part of the result that a picture of a cascade cannot express.

The experiment #

A universality claim about mathematics is testable in a laboratory, and it was tested.

Libchaber, in 1979–80, built a cell of liquid helium a few millimetres across and heated it from below. As the temperature difference rose, the convection rolls in it began oscillating; the oscillation doubled its period, then doubled again, and four doublings were measurable before noise swallowed the rest. The ratio of the temperature intervals came out at about $4.4$.

Four doublings against a slowly converging limit, in a physical system with a hundred million degrees of freedom, agreeing with a constant derived from a map of one variable. That result did more than any argument to make the subject respectable, and it is the strongest evidence that the universality class is real rather than an artefact of the equations people happen to write down.

The reason it works is worth one more sentence, because otherwise it reads as magic. A convecting fluid has enormous numbers of degrees of freedom, and almost all of them are strongly damped: perturbations along them die out fast. What survives is a small number of slow modes, and near the onset of oscillation the whole system’s long-run behaviour is governed by one of them. That mode is a map of one variable with a smooth maximum — so the helium is in the class, not by analogy but by membership.

That is the general shape of every successful application of this field: not “the system is chaotic” but “the system’s slow dynamics reduce to a map in a known class, and the class has constants”.

Where the ladder goes next #

This is the second rung of the period-doubling ladder: the cascade is the first, and the constant that measures it is the second. What lies past the accumulation point is the other half of the subject — orbits that never repeat at all — and the property that makes them matter is sensitivity to the starting point.

The other direction is the one this essay has been leaning on throughout. A number that turns up in unrelated systems because they share a process rather than an object is a pattern with several instances in this collection: the bell curve arrives whatever the underlying distribution, and the golden ratio turns up wherever something reproduces its own shape. Feigenbaum’s constant is the third, and it is the one that had to be measured before anybody believed it.