This series claims that four mathematical structures — topology, dynamics, closure, computation — recur across physical systems. There is an older and far better-established account of where emergent properties come from, and it has one structure, not four: broken symmetryA system whose governing law is symmetric under some transformation, but whose actual state is not. The law does not pick a direction; the system does.. Philip Anderson's More is Different1 is the canonical statement of it, and for fifty years it has been the default answer.
So the honest question for this framework is not whether broken symmetry is important. It is: where does broken symmetry sit relative to the four skeletons? Is it a fifth structure we have missed, is it a rival account of the same ground, or is it something inside the structure we already drew? A sweep of roughly forty documented symmetry breakings across the twelve multi-skeleton systemsThe twelve systems in the matrix that carry two or more skeletons: ecosystem, machine learning, protein folding, active nematics, tokamak plasma, brain, FQHE, quasicrystal, cardiac tissue, immune system, development, and the BZ reaction. in the matrix returned an answer, and then returned a second, less comfortable one that this page also has to report.
What follows is five findings. The first two support the framework. The third is a serious problem for it. The fourth names two systems that defeat broken symmetry outright. The fifth is a rival explanation of the framework's own headline number, generated from the framework's own data, which we have not been able to rule out.
1. What broken symmetry is
A magnet is the textbook case. The physics of a lump of iron does not prefer any direction in space — rotate the whole experiment and every law reads the same. Cool it below the Curie temperature anyway and the spins line up along some direction. The law kept its symmetry; the state threw it away. That gap is the whole phenomenon.
Anderson's argument was that this gap is where new physics enters. Once a direction is chosen, the system acquires properties — rigidity, a definite orientation, a restoring force — that no amount of analysis of a single atom would predict. More is different because the collective state has less symmetry than the parts.
Two pieces of vocabulary carry the rest of this page.
- G and H.
Gis the symmetry group of the disordered state — everything the law allows.His what survives the choice: the subgroup that still leaves the ordered state alone. The set of equally-good ordered states is the quotientG/H, called the order-parameter manifoldThe space of all the different ordered states the system could equally well have fallen into. For a magnet in three dimensions it is the sphere of directions the magnetisation could point; every point on it is an equally valid ground state.. For a three-dimensional magnet,Gis all rotations,His the rotations about the chosen axis, andG/His a sphere. - Goldstone modes. Moving along
G/Hcosts no energy — every point on it is an equally good ground state. So a slow, long-wavelength twist from one ground state to a neighbouring one is almost free, and appears as a gapless excitation.2 Spin waves in a magnet, phonons in a crystal, and — as section 5 gets to — phasonsThe Goldstone mode of a quasicrystal: a rearrangement of the tiling that costs no energy in the ideal case, because it corresponds to sliding the cut through the higher-dimensional lattice. in a quasicrystal are all the same fact in different clothes.
The two directions out of a ground state
The potential is symmetric under every rotation. The ball cannot sit at the top, so it picks a point on the brim — and the moment it does, the symmetry is gone from the state but not from the potential. Watch the energy readout in each of the two directions it can then move.
Press a button. The ball is sitting in the brim, in one arbitrarily chosen ground state.
Everything in the rest of this page is a consequence of one thing: G/H is a space, and spaces
have shape.
2. Where it shows up in these systems
The sweep went system by system through the twelve multi-skeleton entries in the matrix and collected documented symmetry breakings from the published literature on each. It found about forty. The table lists them by system, with where each one sits in the four-skeleton picture.
Granularity caveat — read this before the table
“About forty” is a soft number and it should be treated as one. Whether the cardiac spiral core and the chirality of that spiral are one breaking or two is a judgement call, and a different reviewer would reasonably return 32 or 50. The count is soft. The pattern is not — the distribution across skeletons does not change under any regrouping we tried, because no regrouping moves an instance from one skeleton to another.
| System | Documented symmetry breakings | Sits on |
|---|---|---|
| Development | Anterior–posterior axis; dorsal–ventral axis; left–right chirality; the segmentation clock (continuous time-translation → discrete); Turing-type digit patterning; planar cell polarity; the gastrulation site; cell-fate choice among equivalent cells | S2 S1×S2 |
| Cardiac | Planar wave → spiral (the core breaks translation and rotation); spiral chirality; alternans (period doubling breaks discrete time-translation); discordant alternans (nodal lines break the spatial symmetry of the alternans phase) | S2 S1×S2 |
| BZ reaction | Homogeneous → oscillatory (Hopf); Turing patterning; spiral core and its chirality; pacemaker nucleation breaking translation | S2 S1×S2 |
| Active nematics | Isotropic → nematic (rotational); spontaneous flow transition; ±½ disclination unbinding into active turbulence; nematic → polar ordering where it occurs | S2 S1×S2 |
| Tokamak plasma | L–H transition and the pedestal (radial translation); magnetic islands from broken axisymmetry; the sawtooth crash (helical symmetry); zonal-flow generation from symmetric turbulence | S2 S1×S2 |
| Brain | Orientation pinwheels in V1 (rotational symmetry of the feature map); ocular dominance columns (translation); perceptual bistability resolving to one percept; travelling cortical waves | S2 S1×S2 |
| Quasicrystal | Liquid → quasicrystal, breaking translation and rotation; selection of an icosahedral orientation; the phason degree of freedom left behind by the extra broken translations | S2 S1×S2 |
| Protein folding | Coil → globule; the handedness of the α-helix; the glassy, ergodicity-breaking part of the folding funnel | S2 |
| Machine learning | Symmetry breaking of the loss landscape at the interpolation threshold; neural collapse of the feature geometry; mode collapse in generative training | S2 |
| Ecosystem | Vegetation banding (Turing-type, breaking translation); competitive exclusion among near-equivalent competitors | S2 |
| Immune | Germinal-centre competition resolving to a dominant clone; the self/non-self discrimination threshold | S2 |
| FQHE | Only one, and it is a different phase from the topologically ordered one: the nematic and stripe phases in higher Landau levels break rotational symmetry. The fractional plateaus themselves break nothing — see section 6. | S2 |
| Magnets (comparator, one skeleton) |
Paramagnet → ferromagnet. One breaking. The purest instance in the entire set, and section 7 is about why that matters. | S2 |
3. F1 — broken symmetry is an edge, not a fifth skeleton
Read the right-hand column. Every instance in the sweep sits inside S2 (Dynamics / Criticality) or on the S1×S2 edge. None sits on S3 (Closure). None sits on S4 (Computation).
That is a clean result and it has a clean reading: broken symmetry is not a fifth vertex of the tetrahedron that we forgot to draw. It is what the S1×S2 edge is made of. A symmetry breaks (that is the dynamics), and the order parameter it leaves behind has a shape, and the defects in that shape are the topology. One phenomenon, sitting across two skeletons.
The near-misses, named honestly
S3 and S4 are not empty because nobody looked. There are three candidates, and all three fail for the same reason.
- Machine learning: the permutation symmetry of hidden units in a layer. Real, and really broken by training.
- Brain: the permutation symmetry of neurons within an assembly.
- Immune: the permutation symmetry of clones in a naive repertoire, broken by clonal selection.
Each of these is a symmetry of the substrate — of which physical unit plays which role. None is a symmetry of the closure relation or of the computation being performed. Relabel every hidden unit and the network computes the identical function; relabel every clone and the repertoire covers the identical antigen space. A symmetry whose breaking changes nothing about the structure in question is not a symmetry breaking of that structure.
We are stating this as a caveat rather than burying it, because it is the honest version and it is more interesting than a clean sweep. If someone finds a genuine symmetry of a closure relation — not of its substrate — that breaks, F1 is in trouble and we would like to know.
4. F2 — where topology is shallow, it is the defect classification of a broken symmetry
This is not our result. It is David Mermin's, from 1979, and it is one of the cleanest theorems in condensed matter physics.3
The claim: defects in an ordered medium are classified by the
homotopy groupsHomotopy groups count the holes in a space by asking which loops (or spheres) inside it can be shrunk to a point and which cannot. π₁ counts loops; π₂ counts spheres.
of the order-parameter manifold G/H. Point defects in two dimensions and line defects in three are
classified by π₁(G/H); point defects in three dimensions by π₂(G/H). Not
approximately, not heuristically — the homotopy class of the loop is a conserved label, and a defect
carrying a non-trivial one cannot be removed by any continuous rearrangement.
The mechanism is worth understanding directly, because it turns an abstract group into a picture. Draw a
loop in the material around the defect. At every point of that loop the order parameter has some value, so
the loop traces out a path in G/H. If that path can be shrunk to a point inside G/H, the defect can
be combed away. If it cannot, the defect is stuck — and no local surgery, no annealing, no amount of
patience will remove it. Topological protection is exactly this and nothing more.
The Mermin map
Left: the director field of a nematic around a ±½ disclination, with a
loop you can move. Right: the order-parameter manifold RP², drawn as a hemisphere seen
from above — its rim is the equator, and points on opposite sides of the rim are the same
point, because a director has no arrow. As the marker walks the loop, its image traces a path on
the right. Move the loop off the core and watch the path change character.
Press Walk the loop. Nothing has been traced yet.
| System | Defect | Classified by |
|---|---|---|
| Active nematics | ±½ disclinations | π₁(RP²) = Z₂ |
| Cardiac tissue | Phase singularity at a spiral core | π₁(S¹) = Z4 |
| BZ reaction | Spiral core | π₁(S¹) = Z |
| Tokamak plasma | Magnetic islands | Field-line topology change from broken axisymmetry |
| Brain (V1) | Orientation pinwheels | Defects of a broken rotational symmetry in the feature map5 |
| Development | Planar-cell-polarity vortices, Turing defects | π₁ of the polarity manifold |
What this costs the framework. In every row above, the topology present in the system is downstream of a symmetry breaking. It is not an independent structure that happens to co-occur with criticality; it is the defect classification of the ordering. That makes “topology recurs across domains” a weaker claim than it looks. Six of the systems in the matrix carry topology at depth 1 for what is, on this reading, a single reason: they ordered, and ordering leaves defects.
The weakest row, flagged rather than hidden
The V1 pinwheel case is not a defect in physical space. It is a defect in a feature map — a chart of which cortical location prefers which stimulus orientation. The mathematics of the defect is the same, but the order parameter is a measured tuning property of neurons, not a physical field with a Hamiltonian and a free energy. Whether Mermin's classification applies with full force to a map that develops under activity-dependent plasticity is genuinely open. Treat that row as the weakest one in the table.
5. F3 — the topology column holds at least four incompatible invariants
This is the most important thing on this page and the least comfortable, so we will state it flatly before softening anything.
The systems in the matrix scored as “topological” are not all topological in the same sense. There are at least four different classifying invariants involved. They do not classify the same objects. They do not agree on when two systems are equivalent. And they are not depths, aspects, or special cases of one another.
| Invariant | Systems | What it classifies |
|---|---|---|
Mermin homotopy defectsπn(G/H) |
Nematics, cardiac, BZ, tokamak, brain, development | Defects of a broken symmetry |
| Topological order (modular tensor category) |
FQHE, anyons6 | Ground-state degeneracy on a surface, and anyon statistics. No symmetry is broken |
| Knot and link invariants of a 1-D embedding |
Protein folding8 | Entanglement of a chain in 3-space |
| Cut-and-project from a higher-dimensional lattice |
Quasicrystals9 | Aperiodic order. More symmetry broken than a crystal, not less |
Put the four side by side and the incompatibility is immediate. A Mermin defect requires a broken symmetry; topological order requires that no symmetry be broken and no local order parameter exist. A knot invariant classifies a curve's embedding and is indifferent to whether anything ordered at all. Cut-and-project classifies a tiling by the geometry of a slice through a lattice in a higher dimension, and the quasicrystal is the one case here where the ordered state has less symmetry than a crystal does, not more.
Ask any of them the question “are these two systems topologically the same?” and they answer about different objects. There is no common category in which the answers can be compared.
The fourth row is the one that most people have not seen worked, and it is worth a minute because it also closes the loop back to section 1. A quasicrystal is not a badly-behaved crystal. It is a periodic crystal in a higher dimension, sliced at an irrational angle and seen edge-on.
Cut and project, and the phason
A perfectly periodic square lattice. A strip through it at an irrational slope
(1/φ). Project every lattice point inside the strip onto the line and you get a
quasiperiodic sequence of two spacings — long and short, in the ratio φ, never repeating.
Now slide the strip sideways. Points swap in and out, the sequence rearranges, and
nothing costs anything: that is a phason, the Goldstone mode of the extra broken translations.
Press Slide the strip. The sequence below is the current slice.
Notice what that invariant is about: not a defect, not a ground-state count, not an embedding — a window in a space the material does not live in. It answers a question the other three do not ask.
Four invariants, four different questions
Each panel animates its own classification. Read the question under each one: they are not the same question asked at four depths. That is F3.
1 · Mermin homotopy
2 · Topological order
3 · Knot invariants
4 · Cut and project
Press Play all four. Then read the four questions and ask what would have to be true for them to be one question.
An open problem for this framework, in its own voice
The series currently encodes the first-versus-second split — Mermin defects at depth 1, topological order at depth 2 — as a depth: a score of 1 against a score of 2. That encoding says these are the same kind of thing, more or less of it. They are not. It is a difference of kind, recorded as a difference of degree.
So either somebody shows that these four are aspects of one classification — which would be a real result and we would very much like to see it — or “topology” is not one skeleton, and the S1 column in the matrix is at least two columns wearing one label. We do not know which. We are not going to pretend the scoring settles it.
6. F4 — two systems defeat broken symmetry outright, from opposite sides
Section 3 found broken symmetry everywhere it looked. Two systems in the matrix are where it fails, and they fail in opposite directions. Both already have pages in the wider series; neither is currently flagged as special, and both should be.
FQHE and topological order — order with no symmetry to break
Landau's theory says: to have a phase transition, break a symmetry; to tell two phases apart, find a local order parameter. The fractional quantum Hall states7 refuse both. States at different filling fractions have identical symmetry, and no local measurement anywhere in the bulk distinguishes them. There is nothing for Landau's tool to grip.
What does distinguish them is global: the number of degenerate ground states depends on the genusThe number of handles on a surface. A sphere has genus 0, a doughnut genus 1, a two-holed doughnut genus 2. Topological order shows up as a ground-state count that depends on this number. of the surface the system lives on, and the excitations braid with fractional statistics. This is the founding statement of the topological-order field.6 Order without symmetry breaking, and it took a new invariant to see it.
Glass — a broken symmetry that is an artifact of the calculation
The glass transition is the mirror image. The theory that describes it best speaks of replica symmetry breakingA technique in which one analyses n identical copies (“replicas”) of a disordered system and takes n to zero. In the spin-glass solution the symmetry among the replicas is broken — but the replicas are a calculational device, not physical copies.10, and that phrase has misled a great many readers. The broken symmetry is the permutation symmetry of the replicas — of copies introduced to make an average tractable. It is a feature of the calculation.
Physically, a glass has no broken spatial symmetry and no order parameter. A snapshot of a glass is statistically indistinguishable from a snapshot of the liquid it froze out of. What happens is ergodicity breakingThe system stops exploring its full configuration space on any accessible timescale. It is trapped in one region — not because a symmetry was broken, but because the barriers became insurmountable.: the system stops exploring, without ever having chosen a direction.
Where Landau's tool has nothing to grip
Left: a magnet. Put a probe anywhere and it reads the order parameter — the local magnetisation — and that single local number tells you which phase you are in. Right: a topologically ordered state on a torus. There are several degenerate ground states, and no local probe anywhere distinguishes them. Only a measurement all the way around a cycle does.
Press Move the probe. Watch the two local readouts, then press Measure around the cycle.
Why these two matter more than the other fourteen
Anyone claiming to succeed Landau theory has to explain the two cases Landau theory cannot handle: order without a broken symmetry (FQHE), and rigidity without order (glass). These are the two most informative systems in the matrix for testing any framework, including this one, and they are currently scored as ordinary rows.
A correction to the FQHE row, and why it was made this way
The matrix used to score FQHE as carrying closure (S3) at depth 1, on the strength of the description
“self-generated anyons”. It no longer does: that cell is now 0. Two reasons, and the second is
the one that actually decided it.
First, the conservative reading of closure. Closure in this series means a self-sustaining
network that produces its own components from a
food setThe set of molecules available freely from the environment. A reflexively autocatalytic and food-generated (RAF) set is one where every reaction is catalysed from within the set and every component can be built up from the food set.
— the RAF condition.11 The conservative reading holds that a
fusion algebra closing on itself is not yet a food-generated production network.
⛔ Corrected 2026-08-10: this sentence previously read “Anyons are excitations
created by externally perturbing the medium” — which is false. Anyonic
quasiparticles are collective excitations generated by the two-dimensional electron gas itself;
the substrate produces them, they are not injected from outside (Nayak et al.,
Rev. Mod. Phys. 80, 1083 (2008), doi:10.1103/RevModPhys.80.1083). Physical generation
by the substrate is Kauffman’s generation criterion instantiated in matter, so any remaining case
for the conservative value must rest on the food-set and catalysis conditions, not on where anyons come from.
The steel-man
for the old value is real and should not be buried: Fibonacci fusion
τ × τ = 1 + τ genuinely does regenerate τ from τ, which is closure in
the fusion-category sense, and “anyon fusion as a RAF” remains an open candidate connection in this
series — it is why the cell was ever a 1. But the fusion algebra closing on itself is a statement about
an algebra, not about a network of physical components producing each other from a food set.
Second, and decisively: the predictor was frozen before the analysis
Two copies of this matrix disagreed on that cell. One of them had been committed to version control and frozen as the predictor for an analysis before that analysis was run; the other had not. Editing a frozen predictor after seeing results is post-hoc tuning, and it is a drift signal no matter how defensible the new value looks. So the frozen copy wins by default and the unfrozen copy is the one that gets corrected — which is the direction the change went here, and it would have gone that way even if the argument above had come out the other side. That is the entire reason for freezing a predictor, in one concrete instance.
7. F5 — a rival explanation of stack height, offered against ourselves
The matrix scores each system by stack heightHow many of the nine emergent properties a system carries: four skeletons at two depths each, plus physical realisation. The framework's headline number.: how many of the nine emergent properties it carries. Development tops the table at 7. The framework's reading of that is that Development carries more distinct mathematical structures than anything else in the set.
Count symmetry breakings instead and a competing account appears, built entirely out of the framework's own data:
| System | Symmetry breakings in the sweep | Stack height |
|---|---|---|
| Development | ~8 | 7 |
| Cardiac | ~4 | 6 |
| BZ reaction | ~4 | 6 |
| Immune | ~2 | 5 |
| Brain | ~4 | 5 |
| Magnets | 1 | 1 |
Magnets are the purest broken-symmetry system in the entire set and they have a stack height of 1. That single fact is what makes the rival account interesting rather than trivial. It is not that broken symmetry predicts stack height by volume — a system with one enormous, perfectly clean breaking sits at the bottom of the table.
What separates the high-stack systems is not that more distinct structures are present. It is that breaking recurs — at different scales, sequentially, each one acting on the state the last one left behind. Development breaks a symmetry, then breaks another symmetry of the result, then another, eight times over, at scales from the whole embryo down to a pair of adjacent cells. A magnet breaks one symmetry, magnificently, once, and stops.
The idea that what happens where two symmetry-breaking transitions meet is its own object of study is not new — it is the classical theory of bicritical pointsA point in a phase diagram where two distinct ordered phases and the disordered phase all meet, so that two different symmetry breakings become degenerate. The behaviour there belongs to neither transition alone.16, and it is a reasonable place to look for what a cascade actually is.
A cascade, and what it might be measuring
Development, one breaking at a time. Each step acts on the state the last one left behind, and each kills a named group. Watch the surviving symmetry shrink — and then compare the chart: instances of breaking against stack height, across the systems in the matrix.
Press Break the next symmetry. The embryo starts spherically symmetric, and every step from here is irreversible.
We are not resolving this
On this reading, “stack height” may be measuring the depth of a symmetry-breaking cascade and reporting it as a count of independent structures. If that is right, the four skeletons would be four ways of describing what a cascade leaves behind, rather than four independent things — a considerably weaker claim than the one this series makes.
We cannot currently distinguish the two accounts with the data in the matrix, because every high-stack system in it is also a deep-cascade system. That is a confound, not a finding, and the next section says what would break it. A page that cannot name the strongest objection to itself is not worth reading; this is ours.
8. What would settle it
Two of the five findings above are open. Here is what evidence would decide each, so that this page can be checked rather than believed.
For F3 — is “topology” one skeleton or several?
- Settled in favour of one skeleton by a construction that recovers Mermin defect classes, modular tensor categories, knot invariants and cut-and-project windows as special cases of a single classification, with the specialisation map written down. Not an analogy — a map.
- Settled against by finding two systems that the matrix scores identically on S1 but which no invariant relates: same score, no shared classifying structure, and no prediction that transfers from one to the other. One clean pair would do it.
- Sharpened either way by asking what the S1 score is supposed to predict. If a depth-2 S1 score predicts something a depth-1 score does not — some measurable, in some system — then the depth encoding is doing work and the incompatibility is survivable. If nobody can name that measurable, the column is bookkeeping.
For F5 — cascade depth, or independent structures?
- The decisive case is a system with a high stack and a shallow cascade — many distinct skeleton-properties, but few sequential symmetry breakings. The framework predicts such systems exist. The rival account predicts they do not. Neither of us has one, and the matrix contains no example. Finding one settles it in the framework's favour.
- The reverse case — a deep cascade with a low stack — would settle it the other way. Magnets are not it: one breaking, one stack.
- Failing either, the honest report is that the two accounts are observationally equivalent on the present data, and the stack-height number should be quoted with that caveat attached.
And a note on the S1×S2 edge grading
Findings F1 and F2 together are the reason the matrix now grades the S1×S2 edge as established and grades self-organised criticality and the edge of chaos as phenomena. “Established” there means one specific thing: a classification theorem exists. Mermin 1979 is such a theorem and it is forty-seven years old. Self-organised criticality12 and the edge of chaos13 are real, named, repeatedly observed — and have no classifying invariant behind them; in the edge-of-chaos case the λ parameter that was supposed to be the invariant was deflated in 199314 and nothing replaced it. The S1×S4 edge keeps its grade for the same reason the S1×S2 edge gains one: Kitaev's construction15 is a theorem, not an observation. That is a re-grading of evidence class, not of importance, and it is the reading this series takes, not a claim about community consensus.
Recap
- F1
- ~40 documented symmetry breakings, all on S2 or the S1×S2 edge, none on S3 or S4. Broken symmetry is what the S1×S2 edge is made of, not a fifth skeleton. The S3/S4 near-misses are substrate symmetries.
- F2
- Where topology is shallow it is Mermin's defect classification of a broken symmetry — downstream of the ordering, not independent of it. This weakens “topology recurs”.
- F3
- Four incompatible invariants share the S1 column. Recorded as a difference of degree; it is a difference of kind. Open problem.
- F4
- FQHE has order with no broken symmetry. Glass has rigidity with no order. Both defeat Landau theory, from opposite sides.
- F5
- Stack height may be measuring cascade depth. Magnets: one breaking, stack 1. Unresolved, and confounded in the present data.