In which throwing away detail, on purpose and repeatedly, turns out to be the whole idea.
In about 1977 a post-graduate mentor was asked by a nine-year-old what the cutting-edge fields in advanced mathematics were. He answered: renormalization group theory — and it will probably win a Nobel prize. Kenneth Wilson collected it in 1982. The method collected a second one in 2004. This page is what he was pointing at.
Water boils. A magnet loses its magnetism at the Curie point. A random network of pipes suddenly percolates. A forest fire either dies out or spreads forever. These systems share no ingredients: one is molecules with dipole interactions, one is electron spins, one is pure graph combinatorics.
Near their transitions, they share numbers.
Measure how the order parameter grows below the transition and you get a power law with an exponent β. Measure how the susceptibility diverges and you get γ. Measure how far correlations reach and you get ν. Wildly different microscopic systems return the same values, often to several decimal places. Liquid–gas and uniaxial magnets: same numbers. This is universalityDifferent microscopic systems exhibiting identical critical exponents and scaling functions. Not a coincidence and not an analogy — a consequence of what survives repeated coarse-graining., and before 1971 nobody could explain it.
The key object is the correlation lengthThe distance over which fluctuations remain statistically related. Below it, the system looks correlated; beyond it, independent. Written ξ (xi). ξ: the distance over which one part of the system still “knows about” another. Away from a transition it is a few lattice spacings, and the system looks like independent patches. As you tune towards the critical point it diverges:
ξ ∼ |t|−ν, t = (T − Tc) / TcAt t = 0 there is no characteristic length left. Fluctuations exist at every scale simultaneously, nested in one another, and a photograph of the system zoomed by any factor is statistically the same photograph. The system is scale invariant.
Everything below requires ξ → ∞. A first-order transition — ice to water, most condensation, and quite possibly many transitions that look critical from a distance — has a finite correlation length, latent heat, and coexisting phases. There is no scale invariance, so there is no fixed point, no universality and no exponents. Ask whether a transition is continuous before reaching for any of this.
Leo Kadanoff, 1966. If the system looks the same at every scale, then describe it at a coarser scale and see what happens to the description.
Take a lattice of Ising spins. Group them into blocks of side b. Replace each block by a single spin, by majority rule. Now shrink all lengths by b so the new lattice has the same spacing as the old one. You have a system of the same type — spins on a lattice — but with different interaction strengths.
Let K = (K1, K2, …) be the full set of coupling constants describing a system — nearest-neighbour coupling, next-nearest, four-spin terms, everything the symmetry allows. A coarse-graining by factor b defines a map
K′ = Rb(K)The map is not invertible — information was discarded — so Rb forms a semigroup, not a group: Rb ∘ Rb′ = Rbb′. The name “renormalization group” is, strictly, a misnomer that stuck.
Iterating Rb traces a trajectory through the space of all possible couplings. The interesting points are the ones that do not move:
A system sitting at K* is unchanged by coarse-graining: it looks identical at every scale. Since each step divides the correlation length by b, a fixed point must have ξ* = ξ*/b, so ξ* is either 0 or ∞. The ξ = 0 fixed points are the boring ones — perfect order and perfect disorder. The ξ = ∞ fixed point is the critical point.
Near the fixed point write K = K* + δK and linearize:
δK′ = M δK, Mij = ∂K′i/∂Kj at K*Diagonalize M. Its eigenvalues are conventionally written λi = byi, and the eigenvectors are the scaling directions. Three cases, and everything hangs on them:
| Case | Name | Behaviour under coarse-graining | Consequence |
|---|---|---|---|
| yi > 0 | relevant | grows; the flow runs away from the fixed point | must be tuned to zero to reach criticality — one knob each |
| yi < 0 | irrelevant | shrinks; the flow forgets it | this is universality — the microscopic details live here |
| yi = 0 | marginal | undecided at linear order | needs higher order; can give logarithms |
For a system with one thermal relevant direction yt and one magnetic one yh, in dimension d, all the familiar exponents are arithmetic on those two numbers:
| Exponent | In terms of eigenvalues | What it measures |
|---|---|---|
| ν | 1 / yt | divergence of the correlation length |
| α | 2 − d/yt | specific-heat singularity |
| β | (d − yh) / yt | growth of the order parameter |
| γ | (2yh − d) / yt | divergence of susceptibility |
| δ | yh / (d − yh) | response at the critical isotherm |
| η | d + 2 − 2yh | anomalous dimension of the correlator |
Two independent numbers generate six exponents, which is why they obey identities — Rushbrooke α + 2β + γ = 2, Widom γ = β(δ−1), Fisher γ = ν(2−η), and hyperscaling 2 − α = dν. These were empirical puzzles before the RG and are corollaries after it. That is the single most convincing thing the theory did.
The first three relations follow from the two-eigenvalue structure alone. 2 − α = dν additionally assumes the free energy scales purely as ξ−d, and it fails above the upper critical dimension, in the presence of dangerous irrelevant variables, and in random-field systems. A measured hyperscaling violation is therefore informative, not necessarily an error — but it does rule out the simple picture.
This one you can do by hand in ten minutes, and it is honest: it produces a real answer and the answer is “no phase transition”.
Take a chain with coupling K = J/kBT. Sum out every second spin — decimationA real-space RG in which a subset of degrees of freedom is summed over exactly, leaving an effective model for the rest. Exact in 1D; approximate in higher dimensions., with b = 2. For the middle spin σ2 between neighbours σ1, σ3:
∑σ2=±1 eK(σ1σ2 + σ2σ3) = 2 cosh(K(σ1+σ3))Demand that this equal eg + K′σ1σ3. Taking the two cases σ1σ3 = ±1 and dividing:
Fixed points: K* = 0 (infinite temperature) and K* = ∞ (zero temperature). Nothing in between.
Check the stability. For small K, cosh(2K) ≈ 1 + 2K² so K′ ≈ K² < K — the flow moves towards zero. For large K, cosh(2K) ≈ e2K/2 so K′ ≈ K − ½ln2 — it decreases by a constant each step and eventually falls to zero too.
The 1966 picture was physically compelling and computationally stuck. Wilson’s move (1971) was to coarse-grain in momentum space instead: integrate out field modes in a shell between Λ/b and Λ, then rescale. Applied to the φ4 theory this gives flow equations for the mass and quartic coupling that can actually be solved — perturbatively, in an unexpected small parameter.
Wilson and Fisher, 1972: treat the dimension itself as continuous and expand around four. Write d = 4 − ε.
For the one-component (Ising) case the results begin
ν = ½ + ε/12 + O(ε²), η = ε²/54 + O(ε³)In three dimensions ε = 1, which is not small. First order gives ν ≈ 0.583 against a true value of 0.62997 — about 7% off. The series is asymptotic, not convergent; adding terms naively eventually makes it worse, and the good modern values come from resummation or from the conformal bootstrap, not from truncating this. The ε-expansion is a proof of mechanism, not a calculator.
| Universality class | β | γ | ν | Notes |
|---|---|---|---|---|
| Mean field / Landau | 1/2 | 1 | 1/2 | exact for d > 4 |
| 2D Ising (Onsager, exact) | 1/8 | 7/4 | 1 | α = 0 (logarithmic) |
| 3D Ising (bootstrap) | 0.32642 | 1.23708 | 0.62997 | the most precisely known non-trivial class5 |
| 2D percolation (exact) | 5/36 | 43/18 | 4/3 | geometric, no Hamiltonian needed |
| Directed percolation (1+1) | 0.2765 | — | ν⊥ 1.0970 | the generic absorbing-state class6 |
“The RG” names a strategy, not a single algorithm. The strategy is: define a coarse-graining, follow what it does to the description, find the fixed points. A dozen quite different machines implement it.
| Flavour | Coarse-graining is… | Best at | Origin |
|---|---|---|---|
| Real-space / block-spin | grouping sites, majority rule | intuition; exact in 1D by decimation | Kadanoff 1966 |
| Momentum-shell (Wilsonian) | integrating out a shell of high-k modes | the ε-expansion; critical exponents | Wilson 19711 |
| Field-theoretic | changing the renormalization scale μ; beta functions | particle physics; running couplings | Gell-Mann & Low 1954; Callan, Symanzik 1970 |
| Functional / exact (FRG) | a flowing effective action Γk | non-perturbative problems; no small parameter needed | Wetterich 1993 |
| Numerical RG | iterative diagonalization of a discretized bath | impurity problems — it solved the Kondo problem | Wilson 1975 |
| DMRG | keeping the highest-weight eigenstates of the reduced density matrix | 1D quantum chains, to machine precision | White 19922 |
| Tensor-network / MERA / TNR | removing short-range entanglement, then coarse-graining | quantum states; flows in wavefunction space | Vidal 2007; Evenbly & Vidal 2015 |
| Conformal bootstrap | not an RG at all | the most precise exponents we have; uses only symmetry and consistency | Rattazzi et al. 2008 |
The last row is there deliberately. The bootstrap gets better numbers than any RG by never coarse-graining anything — it imposes conformal symmetry and crossing consistency and squeezes the allowed region until one point is left. Worth knowing that the RG is the explanatory framework and no longer the most precise computational one.
In the field-theoretic version one tracks how a coupling depends on the energy scale at which you measure it:
β(g) = μ ∂g/∂μThe sign is everything. In QED β > 0: the coupling grows with energy. In QCD the non-abelian gluon self-interaction flips it — β < 0 — so the coupling shrinks at high energy and quarks behave as if free when probed hard. That is asymptotic freedom, and it is why the RG collected a second Nobel prize in 2004.3
A tutorial that only sells the method is not useful. Four boundaries, all sharp.
No diverging correlation length, no scale invariance, no fixed point. There is nothing for the RG to compute. Before applying any of this, establish that the transition is continuous.
The Landau–Ginzburg–Wilson framework begins by writing a free energy as a functional of a local order parameter. Topological orderA gapped quantum phase distinguished by no local measurement — no symmetry broken, no local order parameter. Classified instead by anyon content and modular data. is defined by not having one. Three independent sources state the consequence in print:
Tensor-network RG does flow in wavefunction space and does find fixed points — but there is no scaling theory around them, so there are no exponents. This is an open problem, precisely stated, not an oversight.
An autocatalytic set is a set-theoretic object: reflexively autocatalytic and food-generated, or not. There is no scale parameter, no coarse-graining that preserves closure in a controlled way, and hence no RG. Computation is worse: the natural invariant — Turing degree — is degenerate, since almost everything non-trivial is universal. The nearest thing to an exception is coarse-graining of cellular automata.10
This one causes real confusion, so it gets its own box.
Static exponents are fixed by dimensionality, order-parameter symmetry and interaction range. The dynamic exponent z — relating relaxation time to length, τ ∼ ξz — is fixed by something else entirely: conservation laws and mode couplings. Hohenberg and Halperin classified dynamics into separate models A, B, C… for exactly this reason.4 A non-conserved order parameter gives z ≈ 2.17; a conserved one gives z = 4 — with identical static exponents.
So two systems sharing β while differing in z is not an anomaly and not a partial failure. It is what the theory predicts. Their static classification agrees and their dynamic classification does not. Reporting that as a puzzle means the two classifications have been silently merged.
Partly, and the boundary is worth being exact about — because “different domains share numbers” is precisely the observation the RG was built to explain, and precisely the observation that is easiest to over-claim.
What the RG genuinely delivers. A mechanism — not an analogy — by which systems with unrelated microphysics end up with identical critical behaviour. The convergence is real, it is derivable, and it comes with an index: the short list of dimensionality, symmetry and range. Two systems converge if and only if those agree.
What it does not deliver. Convergence of anything else. Sharing a universality class means sharing exponents and scaling functions — it does not mean sharing a mechanism, a structure, or a vocabulary. Two systems in the 3D Ising class have nothing else in common, and that is the point of the theorem rather than a limitation of it.
So: as a candidate mechanism for domain convergence, the RG is the strongest one available, and it comes with a built-in test. If two domains are claimed to converge, the RG says to check whether their dimensionality, order-parameter symmetry and interaction range agree — and if they do not, the shared numbers need a different explanation. It also says which quantities should transfer (static exponents) and which have no business transferring (dynamic ones, absent matching conservation laws). That asymmetry is a prediction, and it is checkable.
Coda.
Around 1977, a nine-year-old on a college campus asked his mentor what was cutting-edge in advanced mathematics. The answer was renormalization group theory, with the prediction that it would win a Nobel prize.
Kenneth Wilson received the 1982 Nobel Prize in Physics “for his theory for critical phenomena in connection with phase transitions”. The prediction was right within five years. It was right a second time in 2004, when Gross, Politzer and Wilczek received the prize for asymptotic freedom — a statement about the sign of a beta function, and therefore also a renormalization-group result.
Written for Bruce Stephenson, and for Bruce A Metzger, who was right.