Details stop counting at the critical point
2026-08-03 · Self-Organization & Criticality
A lump of iron and a canister of xenon: one is a hard metal, the other a colourless gas. Different atoms, different structure, and even a different thing going on — one loses its magnetism, the other loses its liquid surface. Yet they share a number, and they agree on it to three decimal places. Nobody can say what that number has to do with iron, or with xenon.
Water at 99 °C is still water. At 100 °C it is steam. There is no half-water-half-steam substance in between. That is a completely different kind of event from "the water got hotter" — getting hotter is continuous and you can stop anywhere, whereas becoming steam is a flip, and on the far side of it you have a different thing.
The counter-intuitive part comes next. At exactly the moment of the flip, a system forgets its own biography. Whether it is iron or xenon, how heavy its atoms are, how far apart they sit, whether the lattice is square or skewed — all of it, the stuff that normally decides everything, walks off stage. What is left is a handful of numbers, and those numbers are shared across substances that have nothing whatever in common.
Five issues on this site deal with criticality, each from a different angle. Issue 10 is bifurcation (the system jumps to a different steady state), issue 17 is percolation (connectivity suddenly spans the system), issue 18 is self-organized criticality (the system climbs to the critical point on its own), issue 36 is critical slowing down (how to see it coming). This issue asks one question only: standing exactly at the critical point, which properties of the system still count? The answer is: very few — and the few are remarkably orderly.
To talk about a flip you first need something that tells you whether it has happened. That something is called the order parameter, and its definition is economical: it is exactly zero on the un-flipped side and non-zero on the flipped side.
Take a magnet. Inside a lump of iron every atom behaves a little like a compass needle with its own orientation. Heat the iron above 770 °C and those orientations scatter every which way, cancelling out; the lump has no magnetism at all — the order parameter is zero. Cool it down and the needles spontaneously line up more or less together, and the lump becomes a magnet — the order parameter is non-zero. That 770 °C has a name: the Curie point.
Or take a sealed vessel of water. At room temperature there is liquid below, vapour above, and a visible surface between them, because liquid is far denser than vapour. The order parameter here is that density gap. Heat the vessel to 374 °C at around 220 atmospheres and the gap closes to zero — the surface disappears, and what fills the vessel is neither liquid nor gas. That point is water's critical point.
The order parameter is not just any convenient metric. You cannot use temperature, because it is non-zero on both sides; you cannot use density, for the same reason. What you need is the quantity that specifically marks whether a symmetry has been broken: at high temperature all directions are equivalent, at low temperature the system is forced to pick one, and the order parameter records how far it has committed.
Why single out continuous transitions? Because there is nothing universal about a jump: how high it jumps and where it jumps are entirely that substance's own business. Whereas at a continuous transition something strange happens — the distance over which the system correlates with itself grows without bound, and that single fact is what forces everything that follows.
Next time somebody says a system is "undergoing a transition", ask for a number first: a number that should be exactly zero on the pre-transition side. If they cannot write one down, "transition" is currently a feeling. If they can, you immediately have something to monitor — and you can tell whether it jumps (first-order: no warning, so you defend in advance) or climbs continuously (in which case there are precursors worth measuring).
In 1945 the physical chemist Edward Guggenheim did something very plain. He took the liquid–gas coexistence data then available for eight substances — neon, argon, krypton, xenon, nitrogen, oxygen, carbon monoxide and methane — and plotted them all on one chart.
In absolute units the eight curves have nothing to do with each other. Neon's critical temperature is 44 K, xenon's is 290 K, more than a factor of six apart; the densities differ several-fold as well. Eight unrelated domes.
Then he did something a schoolchild could do: he divided each substance's temperature by its own critical temperature and its density by its own critical density. The eight curves promptly became one.
Collapsing is not by itself surprising: normalise any two roughly similar curves and they will end up close. What matters is the shape at the top of the dome.
The closer you get to the critical point, the smaller the liquid–gas density gap. It shrinks in a very particular way: the gap goes as (1 − T/Tc) raised to a power β — where T is the current temperature and Tc the substance’s critical temperature, so the bracket is simply "how far from the critical point we still are". That power is a critical exponent — a number describing how fast some quantity vanishes, or blows up, as the critical point is approached. By convention these are named with Greek letters: β for the order parameter, γ for the strength of fluctuations, ν for the correlation distance, and so on.
Guggenheim's fit gave β = 1/3. Later precise experiments and calculations give 0.326.
Here is the pivot of the whole issue: an ordinary magnet in three-dimensional space also has β = 0.326 near its critical point. Not "roughly" — in the narrow band close to the critical point they agree to the precision experiment and computation can reach. (Step away from it and they stop agreeing; section 05 is about exactly that.) The same number turns up again in the order–disorder transition of binary alloys, and in the moment two liquids stop mixing and separate into layers. These systems share no ingredients at all: electron spins in the magnet, molecular crowding in the fluid, which atom sits on which lattice site in the alloy.
Physicists call the set of all systems sharing one set of critical exponents a universality class. The list above forms one class, known as the three-dimensional Ising class, after the extremely crude lattice model that first pinned it down → ref · the Ising model.
Start with something you can watch with your eyes. Warm a thick-walled glass tube of carbon dioxide slowly toward its critical point and the clear liquid inside abruptly turns milky, opaque enough to block light. Take the heat away and it clears again. This is critical opalescence.
The reason it goes milky is that the tube fills with blobs of differing density, and those blobs happen to be about the same size as the wavelength of visible light, so they scatter it in every direction.
This introduces the key quantity of the issue: the correlation length, usually written ξ. It means: how far apart two places can be and still feel each other. In room-temperature water ξ is a few molecules long — shove one molecule and three neighbours react, beyond that nobody cares. At the critical point ξ grows without bound: the entire vessel is pulling on itself end to end.
The reason the blobs come in every size — from a few molecules to the whole tube — is precisely that no size is the typical one. That is the same face worn by avalanches in issue 18, which have no typical size either; the difference is that there the system climbed to criticality by itself, and here somebody tuned it there.
The operation performed in that figure is called coarse-graining: merge neighbouring units into one, then carry on treating that as the new unit. It amounts to taking a step back and throwing detail away.
Leo Kadanoff proposed this line of attack in 1966, and Kenneth Wilson turned it into a machinery that could actually compute things during the 1970s — the renormalization group, for which he received the 1982 Nobel Prize in Physics. Its central result, in plain words, is this:
Every round of coarse-graining turns the parameters describing the system (how strong the bonds are, how heavy the atoms are, whether next-nearest neighbours interact, and so on) into a new set of parameters. Do it repeatedly and the vast majority of parameters shrink toward zero — these are called "irrelevant"; only a very few grow instead, and those are "relevant". Behaviour at the critical point is fixed by the relevant few alone.
That is the answer to why details do not count: not that they are absent, but that they are systematically ground away as the magnification is stepped back. And a diverging ξ is exactly what lets you keep stepping back — there is still structure to see however far you go, so every finite-scale detail gets ground down to nothing. Far from the critical point ξ is short, two steps back and there is nothing left to look at, so you never reach the stage where details are erased. There, details rule completely.
When you inherit a system nobody can explain, measure the correlation length first: pick any observable and ask how far apart two locations can be and still correlate. If ξ is short, build the local, detailed, reductionist model — a coarse-grained one will be wrong. If ξ approaches the size of the system, stop spending on detail: no amount of microscopic precision will make it into the conclusion, and the job is to identify the two or three relevant parameters instead. Make this call before you build the model, not after it fails.
If the details are ground away, what is left? The renormalization group's answer is startlingly short — three things, and not one of them concerns what the material is made of.
One: the dimension of space. The same rules on a line, on a sheet and in a brick give entirely different results. On a line there is no transition at all: at any temperature above absolute zero, any single site can flip and cut the line into two mutually indifferent halves, and the cost of doing so is fixed — fluctuations can always afford it. So order on a line cannot survive any finite temperature. (Incidentally this is exactly the mistake Ernst Ising made in 1924: having solved the one-dimensional case, he concluded there was no transition in any dimension. Twenty years later Lars Onsager solved two dimensions exactly and showed there is.)
Two: how many components the order parameter has. That is to say: when the system flips, is the available choice binary, or a continuous circle of directions? If the little compass needles in a magnet can only point up or down, that is binary; if they can point at any angle in a plane, that is a circle. The two cases have different critical exponents — they are two different universality classes. This condition is really a question about what symmetry the system has.
Three: how far the interaction reaches. "Affects only its immediate neighbours" and "every unit directly affects every other unit" are two different worlds. The second is called mean-field behaviour, and its β is a clean 1/2. There is also a lovely result attached: once the dimension of space exceeds four, short-range systems degenerate into mean-field behaviour too — there are so many neighbours that "neighbours" and "everybody" stop differing.
Fix those three and the universality class is fixed, and with it the exponents. Every other piece of information — iron or nickel, Tc at 770 °C or 358 °C, how high the curve sits — is locked out.
The sharpest test on record comes from liquid helium. Helium-4 becomes a superfluid near 2.17 K, and that transition belongs to the three-dimensional XY class (the one whose order parameter is a full circle of directions). The experiment was flown into Earth orbit — on the ground, gravity’s pressure gradient rounds the transition itself off and you cannot get that close — and the analysis published in 2003 gives one of its critical exponents as α = −0.0127, to a few parts in a thousand. That number has nothing to do with magnets, and nothing to do with the chemistry of helium; it follows from three dimensions, a circle of directions, and short-range interaction.
Before carrying one system's behaviour over to another, clear three gates: how many units does each unit directly touch (dimension) · is the choice it faces binary or a continuous circle (symmetry) · does influence stop at the neighbours or reach everyone (range). Clear all three and the analogy can carry numbers. Fail one and the analogy stays at the level of a story — do not compute with it. The gate that fails most often is the third: human social networks are full of long-range links, which puts them in a different universality class from any neighbours-only model.
Universality is one of the most beautiful results of twentieth-century physics, and also one of the most heavily abused when carried into the social sciences. Know its edges before using it.
First, it holds only inside a narrow corridor beside the critical point. Leave the corridor and two substances part company at once. How wide the corridor is, is itself completely non-universal — it belongs to each substance. Some have a scaling window wide enough to catch easily; some need temperature control to within a ten-thousandth of a degree.
Second, Tc itself is not universal in the slightest. This is the easiest thing to forget. What is universal is the shape of the curve close to the critical point; where the critical point sits, how high the curve is, how big the fluctuations are, all belong to the material. So "these two systems are in the same universality class" does not license you to infer one system's critical point from the other's. It licenses inferences about shape.
Third, social systems probably do not meet the premises. Calling an opinion flip, a market crash or an organizational turnaround a "phase transition" reads well, but not one of the premises comes cheap: the derivation of universality wants a system at equilibrium (with a well-defined "temperature" measuring the strength of random disturbance), units that are broadly identical, symmetric coupling (I influence you as much as you influence me), and enough units that finite-size effects can be neglected. Human systems fail essentially all four: there is no temperature, they are never at equilibrium, people differ enormously, influence is one-way (some people get listened to), and an organization has tens or hundreds of members — a dozen orders of magnitude short of "infinite".
Worse, there is the third gate from the previous section: social networks are full of long-range links, and long-range links change the universality class outright, pushing the system toward mean-field behaviour. So even if a social phenomenon genuinely is a phase transition, it is most likely not in the class you reached for.
Fourth, "it looks like a phase transition" is nearly worthless as evidence. An S-shaped curve can be produced by a phase transition, and equally by a pile of quite different mechanisms — accumulating positive feedback, a spread of individual thresholds, plain autocorrelation in a time series. To claim a transition seriously, the minimum is two independently measured critical exponents plus a check that the scaling relation between them holds. A single exponent fitted to a single curve establishes nothing.
Before the word "transition" goes into a conclusion, answer four questions: ① which quantity is the order parameter, and is it really exactly zero on one side? ② do your data sit inside the scaling window or outside it? ③ have you independently measured a second exponent? ④ are there enough units that finite-size effects can be ignored? If you cannot answer all four, downgrade "phase transition" to "there appears to be a threshold". The weaker claim still guides action — find the threshold, watch it — without importing a pile of corollaries you never checked.
A very strong constraint: it compresses the space of possible macroscopic behaviours from infinitely many down to a handful of classes. Which means that once you have measured a set of exponents you can work backwards to the system's dimension, symmetry and interaction range — and those three are often much harder to observe directly than the exponents are. Universality's real use is as a probe pointing backwards, not as a predictor pointing forwards.
The intuition runs like this: the higher the dimension, the more neighbours each unit has, and the closer what it feels is to a plain average, with individual random deviations averaging out. Above four dimensions that averaging has overwhelmed the fluctuations entirely, and the system behaves exactly as if each unit felt only a mean field. It is worth noticing that the three dimensions we live in sit just below that threshold — fluctuations in three dimensions still barely matter, which is why there are distinct universality classes to speak of at all.
No, or at least it cannot be assumed to be. A critical point has a definition — an order parameter, a correlation length, measurable exponents — whereas in most uses "the edge of chaos" names no order parameter and nobody offers a measurable exponent. The test is simple: when someone says a system is "at the edge of chaos", ask which quantity is the order parameter and how the correlation length is measured. When there is no answer, the phrase is a metaphor.
Formally it makes sense — a long correlation length does mean high sensitivity. But the same property means noise propagates across the whole as well, and response times become extremely long (which is the subject of issue 36, critical slowing down). So it is not free: what you buy is sensitivity to all inputs, including the ones you did not want. If you are going to try, settle first who filters the inputs.
Broadly the same pair — order parameter and correlation length are both wanted either way. The difference is which threshold you watch. To bring a flip about, watch for the correlation length starting to grow, because that is when the system begins responding as a whole. To prevent one, watch for the order parameter leaving zero, because by then the flip has happened. The lag between those two events is the window in which intervention is possible.