Renormalization Group

Definition

The renormalization group (RG) is a recipe for changing the scale at which you describe a physical system, and for tracking how the description changes as you do. One RG step has two moves:

1. Coarse-grain. Integrate out (average over) the short-wavelength, short-distance fluctuations — the fine detail you can no longer resolve at your new, blurrier scale.
2. Rescale. Shrink all lengths by the same factor b so the coarse-grained system has the same lattice spacing — the same "graininess" — as the original. Rescale the fields too, so the typical fluctuation size is restored.

After both moves you have a system described by the same kind of theory — same fields, same symmetries — but with different coupling constants. The map from old couplings to new couplings is the RG transformation. Iterating it traces out a trajectory in the space of all possible couplings: the RG flow.

Why a "group"? Doing a scale change by b₁ then by b₂ is the same as one change by b₁b₂ — the transformations compose. But coarse-graining destroys information and cannot be undone, so there is no inverse. It is really a semigroup; "renormalization group" is a historical misnomer that stuck.

The block-spin picture

Leo Kadanoff's 1966 mental image is the cleanest entry point. Take the Ising model: a lattice of spins, each ±1, with a coupling K = J/k_BT between neighbors that prefers them aligned. Group the spins into blocks — say 3×3 cells. Replace each block by a single "block spin" whose value is the majority vote of its nine members. You now have a coarser lattice with one third the spacing.

Rescale lengths by b = 3 and you are back to a lattice that looks just like the original — but the effective coupling between block spins, K′, differs from K. The function K′ = R(K) is the renormalization-group recurrence. Ask where it sends you:

• Start at high temperature (small K, disordered): blocks vote randomly, correlations shrink, and K′ < K. Flow drives K → 0, the trivial high-temperature fixed point — a completely disordered paramagnet.
• Start at low temperature (large K, ordered): blocks almost all agree, and K′ > K. Flow drives K → ∞, the zero-temperature fixed point — a perfectly ordered ferromagnet.
• Start at exactly the critical temperature T_c: the system looks the same at every magnification — fractal patches of up and down spins at all sizes. Coarse-graining changes nothing: K′ = K = K^*. This is the nontrivial critical fixed point.

The critical point sits on the knife-edge between the two trivial basins. That edge is the critical surface, and the flow along it heads to the critical fixed point.

Fixed points = scale invariance

A fixed point K^* satisfies R(K^*) = K^*: coarse-graining leaves it unchanged. Physically, the system looks identical at every length scale — it is scale-invariant (and, for the local theories relevant here, conformally invariant). That is no coincidence at a continuous phase transition: the correlation length ξ diverges there, so there is no characteristic length left, and "no characteristic length" is precisely scale invariance.

Fixed points come in flavors:

• Trivial (stable) fixed points — the K = 0 and K = ∞ ends above. They attract whole phases. Every disordered system flows to the same featureless high-T fixed point; that is why all paramagnets look alike on large scales.
• Critical (unstable) fixed points — like the Ising K^*. They are attractive within the critical surface but repulsive in the one direction that leaves it. You have to tune a knob (temperature) to land on them.
• The Gaussian fixed point — the free, non-interacting theory. Stable for d > 4, unstable for d < 4, which is the whole story behind the upper critical dimension.

Relevant vs irrelevant operators

The deepest payoff comes from linearizing the flow near a fixed point. Write the couplings as K^* plus a small perturbation, and the RG map becomes a matrix. Diagonalize it. Each eigen-direction is a scaling operator O_i with an eigenvalue Λ_i = b^y_i under a rescaling by b:

Class	Exponent yi	Behavior under coarse-graining	Meaning	Ising example
Relevant	yi > 0	Grows — pushes you away from the fixed point	A knob you must tune to stay critical	Reduced temperature t (yt = 1/ν ≈ 1.587 in 3D)
Relevant	yi > 0	Grows	Second tuning knob	Magnetic field h (yh ≈ 2.482 in 3D)
Irrelevant	yi < 0	Shrinks — dies out at large scales	Microscopic detail that washes away	Lattice anisotropy, next-neighbor coupling
Marginal	yi = 0	Neither, at leading order	Needs higher order; often gives logarithms	φ⁴ coupling at d = 4 exactly

This single classification carries enormous weight. The number of relevant operators equals the number of parameters you must fine-tune to sit exactly at the transition. Ordinary criticality has exactly two relevant operators — temperature and ordering field — so the critical point is a single point in the (T, h) plane. A tricritical point has three relevant operators; you tune three knobs and it becomes a point in a 3D parameter space. Counting relevant operators is how RG predicts the dimensionality of phase diagrams.

Universality — the surprise

Here is the punchline that won the Nobel. Two systems with utterly different microscopic physics — the liquid-gas critical point of carbon dioxide and the Curie point of a chunk of iron — produce numerically identical critical exponents. The reason: irrelevant operators are exactly the directions that flow to zero. Everything that distinguishes CO₂ from iron is encoded in irrelevant couplings, and those vanish under coarse-graining. What survives is set entirely by the fixed point, which depends only on:

• the spatial dimension d,
• the number of order-parameter components n (1 for a uniaxial magnet, 2 for an XY superfluid, 3 for an isotropic ferromagnet),
• the symmetry of the order parameter and the range of interactions.

Systems sharing these features belong to the same universality class and have the same exponents. The 3D Ising class (d = 3, n = 1) alone covers water's critical point, uniaxial ferromagnets, binary alloys, and the demixing of certain liquid mixtures — all with β ≈ 0.326, ν ≈ 0.630, γ ≈ 1.237. That a beaker of water and a bar magnet obey the same exponents to three decimal places is the kind of result that looks like a coincidence until RG explains why it cannot be anything else.

Worked example — the 2D Ising recurrence

Decimation on a 1D chain is exactly solvable and shows the machinery with real numbers. Sum out every other spin of an Ising chain with coupling K; the remaining spins acquire a new coupling

K' = (1/2) · ln( cosh(2K) )

Iterate from K = 0.5:

K0 = 0.5000
K1 = 0.5·ln(cosh 1.0000) = 0.2326
K2 = 0.5·ln(cosh 0.4653) = 0.05343
K3 = 0.5·ln(cosh 0.1069) = 0.002854
K4 ≈ 0.00000814  →  K → 0

Every starting K flows to K^* = 0. The only fixed points are the trivial pair K = 0 (disorder) and K = ∞ (order), with no critical fixed point in between — the exact statement that the 1D Ising model has no finite-temperature phase transition. RG reproduces that famous result in four lines of arithmetic.

Go to 2D and a nontrivial fixed point appears at K^* ≈ 0.4407 (Onsager's exact K_c). Linearizing the block-spin recurrence near it gives the thermal eigenvalue, and from y_t you read off the correlation-length exponent

ν = 1 / y_t     (2D Ising exact: ν = 1, y_t = 1)

Crude majority-rule block spins land near ν ≈ 1 even with a tiny calculation; the point is that a single number, y_t, from the linearized flow gives a measurable exponent.

The epsilon expansion — making 3D calculable

The masterstroke. The interacting Wilson-Fisher fixed point of φ⁴ scalar field theory exists only below the upper critical dimension d = 4. At exactly d = 4 it merges with the Gaussian fixed point. Wilson and Fisher (1972) set

d = 4 − ε

and treated ε as a small parameter, computing critical exponents as a power series in ε. For the Ising (n = 1) class:

ν = 1/2 + ε/12 + (7/162)·ε² + ...
η = ε²/54 + ...

The audacious move is to plug in ε = 1 — that is, the physical 3D case — into a series derived for small ε:

Order in ε	ν (Ising 3D, ε = 1)	Comment
0th (Gaussian)	0.500	Mean-field value
1st	0.583	One loop — already moving the right way
2nd	0.626	Two loops
Borel-resummed (5+ loops)	0.6300 (8)	Matches conformal-bootstrap / experiment
Best known (bootstrap)	0.62997	Modern benchmark

A divergent asymptotic series, evaluated at a parameter that is nowhere near small, resummed, and it nails the answer to four digits. This is why the ε-expansion is one of the celebrated triumphs of theoretical physics: it converted "compute critical exponents in 3D" — apparently hopeless — into a controlled perturbative calculation.

Connection to quantum field theory

The same machine runs in particle physics, where it is usually called "renormalization" with the group structure encoded in the beta function. A field theory is just an effective description valid below some cutoff Λ; lowering Λ is coarse-graining. The beta function β(g) = dg/dln μ says how a coupling runs with energy scale μ:

Theory	Sign of β	RG behavior of the coupling	Consequence
QCD (strong force)	β < 0	Coupling → 0 at high energy	Asymptotic freedom; quarks free at short distance
QED (electromagnetism)	β > 0	Coupling grows at high energy	Landau pole; effective charge rises with energy
φ⁴ at d = 4	β > 0 (marginal)	Triviality — coupling → 0 in the continuum	Why d = 4 is the upper critical dimension
2D non-linear sigma model	β < 0	Asymptotic freedom in 2D	Mass gap, dynamical scale generation
Conformal field theory	β = 0	Coupling fixed at all scales	A genuine RG fixed point — exact scale invariance

Gross, Politzer, and Wilczek won the 2004 Nobel for the negative QCD beta function — asymptotic freedom — which is nothing but an RG flow toward a free (Gaussian) fixed point at high energy. The Wilsonian view dissolves the old worry that field theories were "sick" because of ultraviolet divergences: every theory is effective, divergences just mean you are extrapolating below the scale where new physics enters, and renormalizability is the statement that only finitely many couplings are relevant.

Where RG shows up

• Critical phenomena. Magnets, superfluids, liquid-gas transitions, polymers, percolation — every continuous phase transition is classified by its fixed point. See phase changes.
• Particle physics. Running couplings, asymptotic freedom in QCD, grand-unification gauge-coupling unification near 10¹⁶ GeV, the hierarchy problem framed as fine-tuning a relevant operator (the Higgs mass).
• Condensed matter. The Kondo problem (Wilson's numerical RG), the Kosterlitz-Thouless transition, quantum criticality, the theory of Fermi liquids as a stable fixed point with the BCS instability as a relevant perturbation.
• Turbulence and dynamics. RG methods for the Navier-Stokes scaling, the Kardar-Parisi-Zhang equation, and reaction-diffusion fronts.
• Numerics. The density-matrix renormalization group (DMRG) is the gold-standard algorithm for 1D quantum chains; tensor-network coarse-graining generalizes it.
• Machine learning. Deep networks are often analogized to RG flows — successive layers coarse-grain features — and the analogy has motivated concrete information-theoretic studies of representation learning.

Common pitfalls and misconceptions

• "RG is just dimensional analysis." No. Naive dimensional analysis gives the Gaussian (mean-field) scaling. RG corrects it with anomalous dimensions generated by interactions — that is the entire difference between mean-field exponents and the true ones.
• "Coarse-graining loses physics." It loses short-distance physics, which is exactly the irrelevant detail. The long-distance, universal physics is what survives — losing the irrelevant is the point, not a bug.
• "The fixed point is a special material." It is a theory, a point in coupling space, not a substance. Many materials flow to it; none sits at it microscopically.
• "Relevant means important and irrelevant means ignorable everywhere." The names are technical. An irrelevant operator can still matter for non-universal quantities (the actual T_c, amplitudes) and for finite-size corrections; it is only the leading scaling behavior it does not affect.
• "Setting ε = 1 in a small-ε series is illegal." It is an asymptotic, divergent series, so it requires Borel resummation — but done carefully it is one of the most accurate tools in physics. The danger is trusting low orders blindly.
• "RG is reversible because it is a group." It is a semigroup. Coarse-graining throws away information and has no inverse; you cannot reconstruct the fine lattice from the blocks.

Why it works — the structural argument

Strip away the technique and the logic is almost embarrassingly clean. Near a continuous transition the correlation length ξ is the only long length scale, and it diverges as ξ ∼ |t|^−ν. A divergent ξ means scale invariance, which means a fixed point. The fixed point has a finite-dimensional unstable manifold — the relevant operators — whose dimension equals the number of tuning knobs. The eigenvalues y_i of the linearized flow are not free; they are properties of the fixed point, and every measurable critical exponent is an algebraic function of just two of them (y_t and y_h):

ν   = 1 / y_t
β   = (d − y_h) / y_t
γ   = (2·y_h − d) / y_t
δ   = y_h / (d − y_h)
α   = 2 − d / y_t
η   = d + 2 − 2·y_h

These relations enforce the classic scaling laws automatically — α + 2β + γ = 2 (Rushbrooke), γ = ν(2 − η) (Fisher), and the hyperscaling relation 2 − α = dν. Before RG, those identities were observed empirically and proven only as inequalities; RG explains why they hold as exact equalities. Two numbers from a single fixed point, plus the spatial dimension, fix the entire critical behavior of an enormous range of physical systems. That compression — from the bewildering specificity of materials to two universal eigenvalues — is the renormalization group's real claim to greatness.