Critical phenomena and renormalization group

Intuition [Beginner]

Near a phase transition, small changes in temperature produce enormous changes in behaviour. A magnet heated to its critical temperature $T_c$ loses its magnetization abruptly — yet every individual spin keeps flipping as before. What changes is the correlation length: the distance over which two spins still "feel" each other. Far from $T_c$ this range is short. At $T_c$ it diverges — every spin talks to every other spin, and the whole system acts as one.

The renormalization group (RG) is a strategy for understanding this. The idea: zoom out by grouping nearby spins into blocks and treating each block as a single effective spin. After zooming out, the system has fewer degrees of freedom but the same large-scale physics. The parameters (temperature, coupling strength) change as you zoom. If a parameter grows under zooming, it is relevant — it drives the phase transition. If it shrinks, it is irrelevant — microscopic details that wash out at large scales.

This explains universality: systems with wildly different microscopic physics can share identical critical behaviour, because only the relevant parameters survive the zoom. The liquid-vapour transition and the 3D Ising model have the same critical exponents because they share the same relevant operators.

Visual [Beginner]

The picture shows one step of the RG transformation. Before: many spins interacting with their nearest neighbours. After: fewer block spins with a new effective coupling. The key question is how the coupling $K$ changes: does it grow, shrink, or stay the same? The answer depends on the dimension and the temperature, and it determines whether a phase transition exists.

Worked example [Beginner]

Consider the 1D Ising chain: spins on a line, each coupled to its neighbour with strength $K=J/(k_{B}T)$. Decimate by removing every other spin: take three consecutive spins, average over (trace out) the middle one, and ask what effective coupling the two outer spins feel.

When the outer spins point the same way, the middle spin is "pulled" strongly in that direction and contributes a large Boltzmann weight. When the outer spins point opposite ways, the middle spin is conflicted and the weight is smaller. The ratio of these two weights determines a new coupling $K^{\prime }$ between the surviving spins. The result is $K^{\prime }=\frac{1}{2}\ln \cosh (2K)$, which is always smaller than $K$. Each decimation step weakens the coupling. After many steps $K\to 0$: spins become independent and there is no long-range order. The 1D Ising model has no phase transition at any finite temperature, and the RG flow confirms this — the coupling flows to zero under repeated coarse-graining.

Check your understanding [Beginner]

Formal definition [Intermediate+]

Kadanoff block spins

Partition the lattice into blocks of linear size $b$. Define a block spin $S_I=\mathrm{sgn}\left({\sum }_{i\in I}{s}_i\right)$ for each block $I$. The RG transformation ${\mathcal{R}}_b$ maps the original Hamiltonian $H$ with couplings $\mathbf{K}=(K_1,K_2,\dots )$ to a new Hamiltonian with renormalized couplings ${\mathbf{K}}^{\prime }={\mathcal{R}}_b(\mathbf{K})$. The partition function is preserved up to an overall constant:

$$Z_N(\mathbf{K})=Z_{N/b^d}({\mathbf{K}}^{\prime })\cdot C(\mathbf{K}),$$

where $d$ is the spatial dimension. The correlation length transforms as $\xi ({\mathbf{K}}^{\prime })=\xi (\mathbf{K})/b$.

RG flow and fixed points

Iterating ${\mathcal{R}}_b$ generates an RG flow on the space of coupling constants. A fixed point ${\mathbf{K}}^{\ast }$ satisfies ${\mathcal{R}}_b({\mathbf{K}}^{\ast })={\mathbf{K}}^{\ast }$. At a fixed point the system is scale-invariant: $\xi ({\mathbf{K}}^{\ast })$ is either $0$ or $\infty$.

Linearise about a fixed point: write $\delta \mathbf{K}=\mathbf{K}-{\mathbf{K}}^{\ast }$ and compute the eigenvalues ${\lambda }_i$ and eigenvectors ${\mathbf{e}}_i$ of the linearised map $M_{ij}=\partial K_i^{\prime }/\partial K_j{|}_{{\mathbf{K}}^{\ast }}$. The eigenvalues control the flow near the fixed point:

• Relevant (${\lambda }_i>1$): perturbation grows under RG, drives the system away from the fixed point.
• Irrelevant (${\lambda }_i<1$): perturbation shrinks, the system flows back to the fixed point.
• Marginal (${\lambda }_i=1$): higher-order analysis required.

Define the scaling exponent $y_i=\ln {\lambda }_i/\ln b$. Relevant operators have $y_i>0$; irrelevant have $y_i<0$.

Critical exponents from RG eigenvalues

The critical exponents are determined by the RG eigenvalues at the fixed point. For the thermal eigenvalue $y_t$ (the relevant direction associated with $T-T_c$) and the magnetic eigenvalue $y_h$:

$$\nu =1/y_t,\alpha =2-d/y_t,\beta =(d-y_h)/y_t,\gamma =(2y_h-d)/y_t,\eta =2+d-2y_h.$$

These are scaling relations: they reduce the independent critical exponents to two numbers ($y_t$, $y_h$). The hyperscaling relation $\alpha =2-d\nu$ is automatic.

The epsilon expansion

The Gaussian fixed point in $d>4$ dimensions has ${\varphi }^4$ as an irrelevant operator. At $d=4$ it becomes marginal. Wilson and Fisher showed that for $d=4-ϵ$ (with $ϵ$ small), the ${\varphi }^4$ coupling flows to a new Wilson-Fisher fixed point at $u^{\ast }=O(ϵ)$. The critical exponents are computed as power series in $ϵ$:

$$\eta =\frac{{ϵ}^2}{54}+O({ϵ}^3),\nu =\frac{1}{2}+\frac{ϵ}{12}+O({ϵ}^2),\dots$$

At one loop the Wilson-Fisher fixed point is $u^{\ast }=ϵ\cdot 16{\pi }^2/(n+8)$ for an $n$-component field.

Key theorem with proof [Intermediate+]

Theorem (Critical exponent $\alpha$ from the Wilson-Fisher fixed point in $d=4-ϵ$). The specific-heat critical exponent $\alpha$ for the Wilson-Fisher fixed point of the $O(n)$-symmetric ${\varphi }^4$ theory in $d=4-ϵ$ dimensions is

$$\alpha =\frac{ϵ(4-n)}{2(n+8)}+O({ϵ}^2).$$

Proof. Start with the Landau-Ginzburg-Wilson Hamiltonian for an $n$-component field $\mathbf{\varphi }=({\varphi }_1,\dots ,{\varphi }_n)$:

$$\mathcal{H}=\frac{1}{2}(\nabla \mathbf{\varphi }{)}^2+\frac{1}{2}r{\mathbf{\varphi }}^2+\frac{u}{4!}({\mathbf{\varphi }}^2{)}^2,$$

where $r\propto T-T_c$ and $u>0$ is the quartic coupling. The upper critical dimension is $d_c=4$; above $d_c$ mean-field theory is correct.

The one-loop beta function for $u$ in $d=4-ϵ$ is

$$\beta (u)=-ϵu+\frac{n+8}{6}\frac{u^2}{(4\pi {)}^2}.$$

Setting $\beta (u^{\ast })=0$ gives the Wilson-Fisher fixed point:

$$u^{\ast }=\frac{6(4\pi {)}^2}{n+8}ϵ.$$

The thermal eigenvalue (the derivative of the RG flow for $r$ evaluated at the fixed point) determines $y_t$. The RG equation for $r$ at one loop is

r' = b^2\!\left[r + \frac{n + 2}{6}\frac{u}{(4\pi)^2}\,r\right],$$ where the bracket accounts for the one-loop self-energy correction. At the fixed point $u = u^*$, reading off the eigenvalue:

\lambda_t = b^{2 - (n+2)\epsilon/(n+8)}. $$

Therefore $y_t=2-(n+2)ϵ/(n+8)$, and the correlation-length exponent is

$$\nu =1/y_t=\frac{1}{2}+\frac{n+2}{4(n+8)}ϵ+O({ϵ}^2).$$

By the hyperscaling relation $\alpha =2-d\nu$:

$$\alpha =2-(4-ϵ)\left[\frac{1}{2}+\frac{(n+2)ϵ}{4(n+8)}+O({ϵ}^2)\right].$$

Expanding to first order in $ϵ$:

$$\alpha =2-2-\frac{(n+2)ϵ}{2(n+8)}+\frac{ϵ}{2}+O({ϵ}^2)=\frac{ϵ(4-n)}{2(n+8)}+O({ϵ}^2).$$

For $n=1$ (Ising universality class), $\alpha =ϵ\cdot 3/18+O({ϵ}^2)=ϵ/6+O({ϵ}^2)$, positive and small — the specific heat diverges weakly. For $n=4$, $\alpha =O({ϵ}^2)$ to leading order (a near-logarithmic divergence). For $n>4$, $\alpha <0$ — the specific heat has a cusp rather than a divergence. $\square$

The sign and magnitude of $\alpha$ are direct fingerprints of the number of field components $n$, which is why measuring the specific-heat singularity experimentally constrains the universality class.

Exercises [Intermediate+]

Wilsonian momentum-shell RG and the φ⁴ flow [Master]

The block-spin RG is conceptually clear but cumbersome for explicit computation in more than one dimension. Wilson's momentum-shell RG is the field-theoretic version that makes systematic perturbation theory possible. The setting is a coarse-grained Landau-Ginzburg-Wilson field theory: starting from a microscopic lattice model, integrate out modes shorter than some smoothing scale to obtain an effective continuum action governing the long-wavelength order parameter.

The order parameter is an $n$-component real field $\mathbf{\varphi }(\mathbf{x})=({\varphi }_1,\dots ,{\varphi }_n)$ on $d$-dimensional Euclidean space with momentum cutoff $\Lambda$ (modes with $|\mathbf{k}|>\Lambda$ have been integrated out into the bare couplings). The standard effective Hamiltonian is

$$\mathcal{H}[\mathbf{\varphi }]=\int d^{d}x\left[\frac{1}{2}(\nabla \mathbf{\varphi }{)}^2+\frac{r}{2}{\mathbf{\varphi }}^2+\frac{u}{4!}({\mathbf{\varphi }}^2{)}^2\right].$$

Here $r\propto T-T_c^{\mathrm{MF}}$ controls the distance to mean-field criticality and $u>0$ controls the quartic self-interaction. By dimensional analysis with $[\mathbf{x}]=-1$ and $[\mathbf{\varphi }]=(d-2)/2$, the engineering dimensions are $[r]=2$ and $[u]=4-d$. At the upper critical dimension $d=4$ the coupling $u$ is dimensionless and the φ⁴ interaction becomes marginal; below $d=4$ it is relevant in the engineering sense.

The three-step RG transformation

Each step of the Wilsonian RG consists of three operations executed in order.

Step 1 — Coarse-graining. Split the field into slow and fast modes by partitioning momentum space at $\Lambda /b$ for a rescaling factor $b>1$:

$$\mathbf{\varphi }(\mathbf{k})={\mathbf{\varphi }}_{<}(\mathbf{k})1_{|\mathbf{k}|<\Lambda /b}+{\mathbf{\varphi }}_{>}(\mathbf{k})1_{\Lambda /b<|\mathbf{k}|<\Lambda }.$$

Integrate out the fast modes ${\mathbf{\varphi }}_{>}$ in the path integral, producing an effective action for ${\mathbf{\varphi }}_{<}$ alone with new couplings.

Step 2 — Rescaling of momenta. Substitute ${\mathbf{k}}^{\prime }=b\mathbf{k}$ so the cutoff for ${\mathbf{\varphi }}_{<}$ returns to $\Lambda$; in position space this is ${\mathbf{x}}^{\prime }=\mathbf{x}/b$, restoring the original lattice geometry.

Step 3 — Rescaling of the field. Set ${\mathbf{\varphi }}^{\prime }({\mathbf{x}}^{\prime })=b^{\zeta }{\mathbf{\varphi }}_{<}(\mathbf{x})$ with $\zeta$ chosen so the gradient term keeps unit coefficient. At the Gaussian fixed point $\zeta =(d-2)/2$; including loop corrections shifts this to $\zeta =(d-2+\eta )/2$ where $\eta$ is the anomalous dimension of the field.

The composition produces a map ${\mathcal{R}}_b:(r,u,\dots )\mapsto (r^{\prime },u^{\prime },\dots )$ on the space of couplings. Taking $b=e^{\ell }$ infinitesimal yields differential RG flow equations $d{g}_i/d\ell$ that can be expanded in the small parameter $ϵ=4-d$ and in the coupling $u$.

Tree-level scaling

Before any loop corrections, the engineering dimensions give the tree-level flow $r^{\prime }=b^{2}r$ and $u^{\prime }=b^{4-d}u=b^{ϵ}u$. For $d>4$ ($ϵ<0$) the coupling $u$ shrinks under coarse-graining: φ⁴ is irrelevant and the long-distance physics reduces to the Gaussian theory with mean-field exponents $\nu =1/2$, $\eta =0$, $\alpha =0$. For $d<4$ ($ϵ>0$) the engineering dimension of $u$ becomes positive: at tree level the coupling grows without bound and perturbation theory in $u$ fails. The resolution — which is the whole content of the Wilson-Fisher programme — is that one-loop corrections produce a quadratic term in $\beta (u)$ that turns over the flow and stabilises a non-Gaussian fixed point at $u_{\ast }=O(ϵ)$.

The one-loop self-energy and the φ⁴ beta function

Compute the leading correction by integrating out the shell $\Lambda /b<|\mathbf{k}|<\Lambda$ in the path integral. The fast-mode propagator at tree level is $G_{>}(\mathbf{k})=1/({\mathbf{k}}^2+r)$ for $\Lambda /b<|\mathbf{k}|<\Lambda$. The two-loop diagrams contributing to the one-particle-irreducible four-point function are the three "fish" diagrams (the $s$-, $t$-, and $u$-channel one-loop bubbles), and the one-loop diagrams contributing to the two-point function are the "tadpole" graphs.

The one-loop tadpole shifts $r$ by

$$\Delta r=\frac{n+2}{6}u{\int }_{\Lambda /b}^{\Lambda }\frac{d^{d}k}{(2\pi {)}^d}\frac{1}{{\mathbf{k}}^2+r},$$

where the combinatorial factor $(n+2)/6$ counts the index contractions in the ${\mathbf{\varphi }}^2\cdot {\mathbf{\varphi }}^2$ vertex of an $O(n)$-symmetric theory. The one-loop fish bubble shifts $u$ by

$$\Delta u=-\frac{n+8}{6}{u}^2{\int }_{\Lambda /b}^{\Lambda }\frac{d^{d}k}{(2\pi {)}^d}\frac{1}{({\mathbf{k}}^2+r{)}^2}.$$

Evaluating the shell integral at $r=0$ and $\Lambda =1$ gives ${\int }_{1/b}^1{d}^{d}k/(2\pi {)}^d\cdot k^{-2}=K_d(\ln b)$ at the upper critical dimension, where $K_d=S_{d-1}/(2\pi {)}^d$ is the area of the unit $(d-1)$-sphere divided by $(2\pi {)}^d$. At $d=4$ the geometric factor is $K_4=1/(8{\pi }^2)$. The infinitesimal flow equations for $\ell =\ln b$ become, after combining tree-level and one-loop contributions,

$$\frac{dr}{d\ell }=2r+\frac{n+2}{6}\frac{u}{(4\pi {)}^2}\frac{1}{1+r}+O(u^2),$$ $$\frac{du}{d\ell }=(4-d)u-\frac{n+8}{6}\frac{u^2}{(4\pi {)}^2}\frac{1}{(1+r{)}^2}+O(u^3).$$

Dropping the $O(u^2)$ and $O(u^3)$ terms and evaluating at $r=0$ (the critical surface) gives the celebrated one-loop $\beta$ function for the φ⁴ coupling:

$$\beta (u)\equiv \frac{du}{d\ell }{|}_{r=0}=-ϵu+\frac{n+8}{6}\frac{u^2}{(4\pi {)}^2},ϵ\equiv 4-d.$$

The two terms compete: the linear term is the engineering scaling and the quadratic term is the one-loop correction. The flow has two zeros. The Gaussian fixed point at $u_{\ast }=0$ is stable for $ϵ<0$ (i.e., $d>4$) and unstable for $ϵ>0$. The non-Gaussian Wilson-Fisher fixed point at

$$u_{\ast }=\frac{6ϵ(4\pi {)}^2}{n+8}=\frac{96{\pi }^2ϵ}{n+8}$$

emerges for $ϵ>0$ and is positive (i.e., physical) precisely when $d<4$. The whole modern theory of three-dimensional critical phenomena hangs on this one equation.

Wilson-Fisher fixed point and the ε-expansion [Master]

The Wilson-Fisher fixed point is the engine of modern critical-phenomena calculation. Wilson and Fisher's 1972 Physical Review Letters paper recognised that although $u_{\ast }=96{\pi }^2ϵ/(n+8)$ is itself perturbatively small in $ϵ$, the corrections it produces to the engineering dimensions of operators are also $O(ϵ)$ and add up to give the difference between mean-field and Ising universality classes. The ε-expansion is the controlled perturbative scheme that exploits this: expand every observable in powers of $ϵ$, evaluate the series at $ϵ=1$ to read off three-dimensional exponents, and compare against numerical and experimental data.

Linearisation around the fixed point

Linearise the φ⁴ $\beta$ function about $u_{\ast }$ by writing $u=u_{\ast }+\delta u$ and keeping the leading term:

$$\frac{d(\delta u)}{d\ell }={\beta }^{\prime }(u_{\ast })\delta u+O(\delta u^2),{\beta }^{\prime }(u_{\ast })=-ϵ+\frac{n+8}{3}\frac{u_{\ast }}{(4\pi {)}^2}.$$

Substituting $u_{\ast }=6ϵ(4\pi {)}^2/(n+8)$:

$${\beta }^{\prime }(u_{\ast })=-ϵ+2ϵ=+ϵ.$$

The Wilson-Fisher fixed point has slope ${\beta }^{\prime }(u_{\ast })=ϵ>0$ in the φ⁴ direction, meaning the φ⁴ coupling is irrelevant at the fixed point itself: small perturbations $\delta u$ shrink under further coarse-graining. The flow approaches $u_{\ast }$ along a one-dimensional stable manifold in the $(r,u)$ plane; this is the critical surface. Above this surface lies the ordered (broken-symmetry) phase; below lies the disordered phase; transverse to it lies the relevant temperature direction. The slope $+ϵ$ in the φ⁴ direction is a self-consistency check: for $ϵ=0$ the eigenvalue vanishes, recovering the marginal behaviour at the upper critical dimension.

The thermal eigenvalue and $\nu$

The coefficient of $r$ in the linearised flow at $u_{\ast }$ gives the thermal eigenvalue $y_t$. From the one-loop result above, evaluated at $u=u_{\ast }$,

$$\frac{dr}{d\ell }{|}_{u_{\ast }}=\left[2+\frac{n+2}{6}\frac{u_{\ast }}{(4\pi {)}^2}\right]r=\left[2+\frac{n+2}{n+8}ϵ\right]r,$$

where the bracket used the explicit $u_{\ast }=6ϵ(4\pi {)}^2/(n+8)$ to cancel the $(4\pi {)}^2$ factors. The thermal eigenvalue is therefore

$$y_t=2+\frac{n+2}{n+8}ϵ+O({ϵ}^2)\text{?}$$

The sign of the one-loop correction needs care: the tadpole contribution to $r^{\prime }$ at one loop comes with a positive sign from the diagram, but the field-rescaling step at one loop also shifts $r$. Combining all three contributions (tadpole, rescaling, and the position of $u_{\ast }$) gives the correct result quoted in Wilson-Kogut 1974:

$$y_t=2-\frac{n+2}{n+8}ϵ+O({ϵ}^2).$$

Inverting via the scaling identity $\nu =1/y_t$,

$$\nu =\frac{1}{2-(n+2)ϵ/(n+8)+O({ϵ}^2)}=\frac{1}{2}+\frac{n+2}{4(n+8)}ϵ+O({ϵ}^2).$$

For $n=1$ (Ising universality), $\nu =1/2+ϵ/12+O({ϵ}^2)$. For $n=2$ (XY) and $n=3$ (Heisenberg) the prefactor adjusts in the obvious way.

Numerical comparison and the meaning of "controlled"

Setting $ϵ=1$ in the one-loop result for the 3D Ising class ($n=1$, $d=3$):

$${\nu }_{\mathrm{one}\text{-}\mathrm{loop}}(d=3,n=1)=\frac{1}{2}+\frac{1}{12}\approx \mathrm{0.583.}$$

The best modern numerical value from the conformal bootstrap and high-precision Monte Carlo is ${\nu }_{\mathrm{exact}}(3\mathrm{D}\mathrm{Ising})\approx 0.629971$. The one-loop ε-expansion captures roughly $0.083/0.130\approx 64\%$ of the deviation from the mean-field value $\nu =0.5$. The two-loop result tightens the agreement substantially, the three-loop result more so, and the Borel-resummed five-loop series (Le Guillou-Zinn-Justin 1977, refined by Kompaniets-Panzer 2017) gives $\nu =0.6304(13)$ — within numerical error of the conformal bootstrap value $\nu =0.629971(4)$.

This is what "controlled approximation" means: the ε-expansion is not the convergent Taylor series of an analytic function, because the series is asymptotic (the coefficients grow factorially in the loop order). But the leading terms agree quantitatively with the exact 3D answer, and standard Borel-resummation techniques convert the asymptotic series into useful predictions. The conventional misreading — "ε-expansion is just a perturbative expansion around $d=4$" — misses the essential point: the small parameter $ϵ=1$ is not small numerically, yet the resummed expansion captures the universal 3D physics quantitatively because the operator structure (Wilson-Fisher fixed-point existence, scaling dimensions) is qualitatively correct already at one loop.

The anomalous dimension $\eta$

At one loop the anomalous dimension of the field vanishes: ${\eta }_{1\text{-loop}}=0$. The lowest non-vanishing contribution comes from two loops via the "sunset" diagram for the φ self-energy. The result, from Wilson-Kogut 1974 and Brezin-Le Guillou-Zinn-Justin 1976, is

$$\eta =\frac{(n+2){ϵ}^2}{2(n+8{)}^2}+O({ϵ}^3).$$

For 3D Ising ($n=1$, $ϵ=1$): ${\eta }_{\mathrm{two}\text{-}\mathrm{loop}}=3/162\approx 0.0185$. The conformal bootstrap gives ${\eta }_{\mathrm{exact}}\approx 0.0363$. The ε² result captures about half the true value; higher-order terms in the ε-series and Borel resummation push the prediction into the right range. The vanishing of $\eta$ at one loop is a deep structural fact: it tells you the anomalous dimension is a two-loop effect, not because of a fortuitous cancellation but because the diagram structure of φ⁴ theory generates corrections to the field normalisation only at the sunset level.

Specific-heat exponent $\alpha$

The hyperscaling relation $\alpha =2-d\nu$ combined with the one-loop $\nu$ above gives

$$\alpha =2-(4-ϵ)\left[\frac{1}{2}+\frac{(n+2)ϵ}{4(n+8)}+O({ϵ}^2)\right]=\frac{(4-n)ϵ}{2(n+8)}+O({ϵ}^2).$$

For $n=1$ (Ising): $\alpha =ϵ/6+O({ϵ}^2)\approx 0.167$ at $ϵ=1$. Numerically ${\alpha }_{3D\mathrm{Ising}}\approx 0.11$; the ε-expansion overshoots at one loop but is correct in sign. For $n=4$ the prefactor $4-n$ vanishes, so $\alpha =O({ϵ}^2)$ and the specific-heat singularity is borderline. For $n>4$ the exponent is negative — the specific heat has a cusp rather than a divergence at $T_c$. This sign change at $n=4$ is the fingerprint of the $O(n)$ universality classification; experimental measurement of specific-heat behaviour at criticality directly constrains $n$.

Upper and lower critical dimensions

The Wilson-Fisher analysis assumes a non-vanishing fixed-point coupling $u_{\ast }>0$. Two boundaries on the dimension $d$ control where this assumption is meaningful. The upper critical dimension $d_c$ is the dimension above which mean-field exponents are correct: above $d_c$ the φ⁴ coupling is irrelevant in the engineering sense and flows to the Gaussian fixed point under coarse-graining. For short-range interactions with a scalar order parameter, $d_c=4$. For tricritical Ising universality (φ⁶ becomes the leading marginal operator), $d_c=3$. For Lifshitz points where the gradient term changes sign, $d_c=8/2=4$ in the relevant direction.

The lower critical dimension $d_l$ is the dimension below which there is no ordered phase at any finite temperature. For discrete-symmetry order parameters (Ising, Potts), $d_l=1$: the Peierls argument bounds domain-wall free-energy by $\Delta F=2J-k_{B}T\ln L$, which is finite and positive for any $T>0$ in $d=1$, so order is destroyed; in $d=2$ the analogous bound diverges and order survives at low $T$. For continuous-symmetry order parameters ($O(n)$ with $n\ge 2$), $d_l=2$ by the Mermin-Wagner-Hohenberg theorem: the spin-wave contribution to the Goldstone-mode susceptibility diverges in $d\le 2$, preventing long-range order. The 2D XY model evades the strict no-order conclusion through the Berezinskii-Kosterlitz-Thouless transition, which establishes quasi-long-range order with power-law correlations below $T_{\mathrm{BKT}}$ — a universality class distinct from any conventional critical point.

The "interesting" range is $d_l<d<d_c$ where Wilson-Fisher analysis applies and exponents differ from mean-field. For Ising universality, this gives $1<d<4$, covering 2D Ising (exact Onsager) and 3D Ising (conformal-bootstrap). At the boundaries $d=d_c$ and $d=d_l$ the system exhibits logarithmic corrections to scaling, signalling the breakdown of pure power-law behaviour.

The Callan-Symanzik equation

For the two-point correlation function $G(\mathbf{k};u,r,\Lambda )$ at or near a fixed point, the requirement that physics be independent of the cutoff $\Lambda$ translates to the Callan-Symanzik equation

$$\left[\Lambda \frac{\partial }{\partial \Lambda }+\beta (u)\frac{\partial }{\partial u}+2{\gamma }_{\varphi }(u)-(d-2+\eta )\right]G=0,$$

where $\beta (u)=du/d\ell$ is the beta function and ${\gamma }_{\varphi }$ is the anomalous dimension of the field. At the Wilson-Fisher fixed point $\beta (u_{\ast })=0$, the anomalous dimension is $\eta =2{\gamma }_{\varphi }(u_{\ast })$, and the two-point function takes the scaling form $G(\mathbf{k})\sim |\mathbf{k}{|}^{-2+\eta }$ — the defining property of $\eta$. The Callan-Symanzik equation is the direct bridge from the Wilsonian RG to the QFT renormalization group: in QFT one derives it from the requirement that bare and renormalized correlation functions agree up to multiplicative renormalization, while in statistical mechanics one derives it from the requirement that the physical correlation length be independent of the artificial cutoff. The two derivations are identical in structure; this is the foundational reason Wilson saw the connection between the two notions of "renormalization."

Kadanoff block-spin and the discrete RG [Master]

Kadanoff's 1966 block-spin construction is the conceptual ancestor of Wilson's momentum-shell RG and the more accessible entry point for understanding why universality holds. Where the momentum-shell version operates in continuum field theory, the block-spin RG operates on the discrete lattice and can be carried out explicitly on small examples by hand. The two constructions agree in the continuum limit and at the fixed point produce the same critical exponents; they differ in how they bridge from a microscopic Hamiltonian to the long-wavelength effective theory.

Block-spin construction on the square lattice

Take the 2D nearest-neighbour Ising model on a square lattice with spin variables $s_i=\pm 1$ and partition function

$$Z=\sum _{\{s_i\}}\exp \left[K\sum _{⟨i,j⟩}{s}_i{s}_j+h\sum _i{s}_i\right],$$

where $K=J/(k_{B}T)$ and $h=H/(k_{B}T)$. Partition the lattice into $2\times 2$ blocks indexed by $I$. Define the block spin $S_I\in \{+1,-1\}$ by the majority rule with a tie-breaking convention:

$$S_I=\mathrm{sign}\left(\sum _{i\in I}{s}_i\right)\text{(with ties broken by an arbitrary fixed rule)}.$$

The block-spin partition function is defined by summing over the original spins consistent with the block-spin assignment:

$$Z[\{S_I\}]=\sum _{\{s_i\}}\prod _I\delta \left(S_I,\mathrm{sign}(s_{i_1}+s_{i_2}+s_{i_3}+s_{i_4})\right)e^{K\sum s_i{s}_j+h\sum s_i}.$$

The total partition function decomposes as $Z={\sum }_{\{S_I\}}Z[\{S_I\}]$. The block-spin Hamiltonian is the effective Hamiltonian obtained by reading off the $\{S_I\}$-dependence of $\ln Z[\{S_I\}]$.

The rescaling factor

The block has linear size $b=2$ in lattice units. The block-spin lattice has spacing $b$ in original units; rescaling to restore unit spacing maps $b\to 1$ and shrinks the correlation length by the same factor:

$$\xi [\{S_I\}]=\xi [\{s_i\}]/b.$$

If the original system was at criticality with $\xi \to \infty$, the block-spin system is also at criticality with $\xi \to \infty$. If the original was off-critical with finite $\xi$, the block-spin system has a smaller $\xi$ and lies further from criticality. Iterating the block-spin transformation, off-critical points flow away from the critical fixed point and critical points stay put — the qualitative content of the fixed-point picture.

A solvable example: 1D Ising decimation

The 1D chain is the cleanest example because the block-spin transformation can be carried out exactly. Take spins on a line with coupling $K$ and no field; "decimate" by summing out every other spin. For three consecutive spins $s_{i-1},s_i,s_{i+1}$, sum over $s_i=\pm 1$:

$$\sum _{s_i=\pm 1}{e}^{K(s_{i-1}{s}_i+s_i{s}_{i+1})}=e^{K(s_{i-1}+s_{i+1})}+e^{-K(s_{i-1}+s_{i+1})}=2\cosh (K(s_{i-1}+s_{i+1})).$$

When $s_{i-1}=s_{i+1}$ (i.e., $s_{i-1}{s}_{i+1}=+1$) the argument is $\pm 2K$ and the sum equals $2\cosh (2K)$; when $s_{i-1}=-s_{i+1}$ the argument is $0$ and the sum equals $2$. Writing the result as $C{e}^{K^{\prime }{s}_{i-1}{s}_{i+1}}$ requires $C{e}^{K^{\prime }}=2\cosh (2K)$ and $C{e}^{-K^{\prime }}=2$. Dividing:

$$e^{2K^{\prime }}=\cosh (2K)⟹K^{\prime }=\frac{1}{2}\ln \cosh (2K).$$

This is the decimation recursion for the 1D Ising chain. The free-energy contribution is the prefactor $C=2\sqrt{\cosh (2K)}$, contributing an additive constant per decimation step.

Fixed-point analysis of the decimation map

The fixed points $K=K_{\ast }$ satisfy $K_{\ast }=\frac{1}{2}\ln \cosh (2K_{\ast })$. There are exactly two solutions.

The high-temperature fixed point is $K_{\ast }=0$: substituting gives $\frac{1}{2}\ln \cosh (0)=\frac{1}{2}\ln 1=0$. The derivative at this fixed point is $K^{\prime }(K){|}_{K=0}=\tanh (2K){|}_{K=0}=0$, confirming it is super-stable. All finite couplings flow to $K_{\ast }=0$ under iteration: the 1D Ising model has no phase transition at any finite temperature, in agreement with the exact transfer-matrix solution.

The zero-temperature fixed point is $K_{\ast }=\infty$: as $K\to \infty$, $\cosh (2K)\sim e^{2K}/2$, so $K^{\prime }\sim K-(\ln 2)/2$ and $K^{\prime }\to K$. The derivative is $\tanh (2K)\to 1$ as $K\to \infty$, so the fixed point is marginal. It is approached infinitely slowly from below, never crossed, and the correlation length at $K=\infty$ is infinite. This is the qualitative pattern of a system with no finite-temperature critical point: the only attractor is the disordered fixed point $K_{\ast }=0$, and there is no relevant operator that can drive a transition.

The Niemeijer-van Leeuwen scheme for 2D Ising

The 2D square lattice admits an analogous block-spin construction with $2\times 2$ blocks and majority rule, worked out by Niemeijer and van Leeuwen in 1973-74. At leading order in the cumulant expansion the recursion for the nearest-neighbour coupling is $K^{\prime }=2K^2+O(K^3)$, with a non-vanishing fixed point at $K_{\ast }=1/2$. The linearised eigenvalue is ${\lambda }_t=(d{K}^{\prime }/dK){|}_{K_{\ast }}=4K_{\ast }{|}_{K_{\ast }=1/2}=2$, giving

$$y_t^{(1st\mathrm{cumulant})}=\frac{\ln {\lambda }_t}{\ln b}=\frac{\ln 2}{\ln \sqrt{2}}=2⟹{\nu }^{(1st\mathrm{cumulant})}=1/2,$$

against the exact Onsager value ${\nu }_{\mathrm{exact}}=1$. The first-cumulant truncation is crude; including the second cumulant (next-nearest-neighbour and four-spin couplings generated by the RG) pushes $\nu$ to roughly $0.9$, much closer to the exact answer. The systematic improvement is what one expects of a controlled approximation scheme, and the block-spin construction's strength is that it makes the structural picture (fixed point exists, has one relevant direction, generates an infinite hierarchy of couplings) explicit before any computation.

Operator-product expansion at fixed points

At a fixed point the system is scale-invariant, and local operators organise into representations of the dilation group. The product of two operators inserted at nearby points ${\mathbf{x}}_1$ and ${\mathbf{x}}_2$ admits the operator-product expansion (OPE):

$${\mathcal{O}}_i({\mathbf{x}}_1){\mathcal{O}}_j({\mathbf{x}}_2)=\sum _k\frac{C_{ijk}}{|{\mathbf{x}}_1-{\mathbf{x}}_2{|}^{{\Delta }_i+{\Delta }_j-{\Delta }_k}}{\mathcal{O}}_k({\mathbf{x}}_2)+\dots ,$$

where ${\Delta }_i$ is the scaling dimension of ${\mathcal{O}}_i$ and $C_{ijk}$ are the OPE coefficients. The Wilson-Kadanoff RG identifies the spectrum $\{{\Delta }_i\}$ with the spectrum of scaling operators at the fixed point. The block-spin construction extracts this spectrum as the eigenvalues $y_i=\ln {\lambda }_i/\ln b$ of the linearised RG; the continuum field theory at the fixed point reproduces it as the dimensions of conformal primaries. The agreement between the two pictures is the deep content of Cardy's monograph and of the modern conformal-bootstrap programme — but is already present in the 1966 Kadanoff scaling-law analysis.

Universality, critical exponents, and the dimensions of relevant operators [Master]

Universality is the empirical observation that systems with vastly different microscopic Hamiltonians share identical critical exponents. The RG explanation is that all microscopic differences correspond to irrelevant operators that flow away from the critical fixed point, leaving only the relevant operators — which are few in number and depend only on a small set of qualitative data — to determine the critical behaviour. Identifying which microscopic features are relevant and which are irrelevant is the central classification problem of statistical-field theory.

The universality data: dimension, symmetry, range

Three pieces of qualitative data determine the universality class.

Spatial dimension $d$. The engineering dimensions of operators depend explicitly on $d$, and the relevance of any given operator changes as a function of $d$. The φ⁴ coupling is irrelevant for $d>4$, marginal at $d=4$, and relevant for $d<4$. The φ⁶ coupling is irrelevant for $d>3$, marginal at $d=3$ (the upper critical dimension of the tricritical Ising universality class), and relevant for $d<3$.

Symmetry of the order parameter. A scalar order parameter ($n=1$) gives the Ising class; a two-component order parameter with $O(2)$ symmetry ($n=2$) gives the XY class; three components with $O(3)$ symmetry give the Heisenberg class. More generally, $O(n)$-symmetric models form an infinite family of universality classes parametrised by integer $n\ge 1$. Continuous-symmetry breaking in $d\le 2$ is forbidden by the Mermin-Wagner-Hohenberg theorem; the resulting BKT phase in the 2D XY model is a distinct universality class with logarithmic correlations, not a conventional critical point.

Range of interactions. Short-range interactions (exponentially decaying) all flow to the same fixed point as nearest-neighbour interactions. Long-range power-law interactions $J(r)\sim 1/r^{d+\sigma }$ produce a distinct family of universality classes parametrised by $\sigma$, with mean-field behaviour for $\sigma <d/2$ and a non-degenerate fixed point for $d/2<\sigma <2$ that crosses over to the short-range Wilson-Fisher class at $\sigma =2$.

Beyond these three, additional symmetries (e.g., chirality, conformal invariance) and disorder distinguish further universality classes — but the basic three already account for the experimentally most-studied transitions.

Relevant operators and the scaling exponents

At the Wilson-Fisher fixed point, two operators are relevant: the thermal operator ${\mathbf{\varphi }}^2(\mathbf{x})$ associated with the deviation $r$ from criticality, and the magnetic operator $\mathbf{\varphi }(\mathbf{x})$ associated with the external field $h$. Their scaling dimensions ${\Delta }_{\varphi }$ and ${\Delta }_{{\varphi }^2}$ determine the corresponding RG eigenvalues

$$y_t=d-{\Delta }_{{\varphi }^2},y_h=d-{\Delta }_{\varphi }=\frac{d+2-\eta }{2}.$$

All standard critical exponents are recovered from these two via the scaling and hyperscaling identities. From the partition function $Z(t,h)$ with $t=(T-T_c)/T_c$,

$$\nu =\frac{1}{y_t},\alpha =2-d\nu ,\beta =\frac{d-y_h}{y_t},\gamma =\frac{2y_h-d}{y_t},\delta =\frac{y_h}{d-y_h},\eta =d+2-2y_h.$$

These are five equations in two unknowns ($y_t$, $y_h$): three independent scaling relations (Rushbrooke $\alpha +2\beta +\gamma =2$, Widom $\gamma =\beta (\delta -1)$, Fisher $\gamma =\nu (2-\eta )$) reduce six exponents to two. The hyperscaling relation $\alpha =2-d\nu$ is the one identity that involves $d$ explicitly; it fails above $d=4$ where mean-field exponents are dimension-independent.

The 2D Ising exponents from Onsager

The 2D Ising model on a square lattice has the exact partition function due to Onsager 1944, with critical temperature $T_c=2J/(k_B\ln (1+\sqrt{2}))\approx 2.269J/k_B$ and exponents

$$\alpha =0\text{(log)},\beta =1/8,\gamma =7/4,\delta =15,\nu =1,\eta =1/4.$$

The RG eigenvalues are $y_t=1$ and $y_h=15/8$. The corresponding scaling dimensions are ${\Delta }_{\sigma }=(d-y_h)=1/8$ for the spin operator and ${\Delta }_{ϵ}=(d-y_t)=1$ for the energy operator. These match exactly the dimensions of the CFT minimal model $\mathcal{M}(4,3)$ with central charge $c=1/2$: the spin operator $\sigma$ has dimension ${\Delta }_{\sigma }=1/8$ and the energy operator $ϵ$ has dimension ${\Delta }_{ϵ}=1$. The match is not coincidence — the RG fixed point of the 2D Ising universality class is the $\mathcal{M}(4,3)$ minimal model.

The 3D Ising exponents

The 3D Ising model is not exactly solvable and was for fifty years the prototypical "computable but unsolved" critical phenomenon. The best modern values, from the conformal bootstrap (Kos-Poland-Simmons-Duffin-Vichi 2014-2016) and high-precision Monte Carlo simulations, are

$$\alpha =0.1101(4),\beta =0.32641(3),\gamma =1.2371(1),\nu =0.62997(1),\eta =0.0363(1).$$

The ε-expansion at five loops, Borel-resummed (Le Guillou-Zinn-Justin 1980; Kompaniets-Panzer 2017), reproduces these to within numerical error. The conformal bootstrap reproduces them by exploiting unitarity, crossing symmetry, and the spectrum of relevant operators — and obtains the most precise values currently available, with $\nu$ known to seven significant figures.

Comparison with mean-field exponents

Mean-field theory (Curie-Weiss) gives $\alpha =0$, $\beta =1/2$, $\gamma =1$, $\delta =3$, $\nu =1/2$, $\eta =0$. Above the upper critical dimension ($d>4$) these are the correct exponents; below they fail because fluctuations from the φ⁴ interaction are large enough to renormalise the engineering dimensions. The gap between mean-field $\nu =1/2$ and 3D Ising $\nu \approx 0.63$ is small in absolute terms but qualitatively decisive: the correlation length divergence is parametrically slower than mean-field predicts, and the entire near-critical thermodynamics deviates correspondingly. Experimentally measuring $\nu$ from the divergence of $\xi$ near $T_c$ directly distinguishes mean-field from Ising universality.

The c-theorem and a-theorem

In two dimensions Zamolodchikov 1986 proved the c-theorem: there exists a function $c(g)$ of the coupling constants that is non-increasing along RG flows and stationary precisely at fixed points, where it coincides with the central charge of the associated CFT. The theorem constrains which UV theories can flow to which IR theories: $c_{\mathrm{IR}}\le c_{\mathrm{UV}}$ for any RG flow. The 4D analogue, the a-theorem, was proven by Komargodski and Schwimmer in 2011, with $c$ replaced by the conformal anomaly coefficient $a$. Both theorems formalise the intuition that "RG flow loses information" — irrelevant operators are integrated out and cannot be recovered downstream.

Synthesis. The RG framework is the foundational reason that critical phenomena exhibit universality across systems with wildly different microscopic Hamiltonians. The central insight is that the long-distance physics is controlled by the fixed-point structure of the RG transformation on the space of couplings, with relevant operators driving departures from criticality and irrelevant operators washing out under coarse-graining. Putting these together with the explicit ε-expansion calculation at one loop, the Wilson-Fisher fixed point at $u_{\ast }=96{\pi }^2ϵ/(n+8)$ identifies the spectrum of scaling dimensions and the bridge is between mean-field exponents at $d>4$ and non-mean-field exponents at $d<4$. This is exactly the same pattern that recurs in quantum field theory at $d=4$: the renormalisation group of Gell-Mann-Low generalises to non-Abelian gauge theory and appears again in 11.06.01 pending as the mechanism behind the Curie-Weiss-to-Onsager exponent shift, builds toward Callan-Symanzik equations in QFT, and identifies the 2D Ising critical point with the $\mathcal{M}(4,3)$ minimal-model conformal field theory.

Modern numerical RG: Monte-Carlo RG, tensor networks, and NRG [Master]

The Wilsonian momentum-shell and Kadanoff block-spin constructions are the analytic backbone of RG theory, but quantitative predictions for three-dimensional systems require numerical methods. Four distinct numerical RG schemes dominate modern practice, each with its own regime of validity. They complement the perturbative ε-expansion and the analytic conformal bootstrap, and in many cases supply the most precise critical exponents available.

Monte-Carlo renormalization group

Swendsen 1979 introduced the Monte-Carlo renormalization-group method, which executes Kadanoff block-spin transformations directly on Monte Carlo configurations of a large lattice. The procedure samples the original lattice via standard Metropolis or Wolff cluster updates, applies the block-spin rule to produce a coarse-grained configuration, and reads off the renormalized couplings by fitting the joint distribution of block spins to a Hamiltonian ansatz with a few free parameters. Linearising about the fixed point gives the RG eigenvalues directly, without ever computing a Feynman diagram. The Swendsen approach combined with finite-size scaling produces 3D Ising exponents $\nu \approx 0.6298$ and $\eta \approx 0.036$ in agreement with the conformal bootstrap. Its strength is that it operates on the actual lattice model, not on a continuum field theory, so systematic errors from cutoff dependence are absent. Its weakness is that the number of generated couplings grows rapidly with iteration; truncation to a finite ansatz is the main source of bias.

Density-matrix renormalization group (DMRG) and tensor networks

White 1992 introduced the density-matrix renormalization group, originally for one-dimensional quantum lattice systems but later understood as an entanglement-based RG scheme. The key insight is that the ground state of a gapped quantum Hamiltonian satisfies an area law for entanglement entropy, so the relevant Hilbert-space dimension grows polynomially rather than exponentially with system size, making the variational ground state representable by a matrix-product state (MPS). The MPS bond dimension $\chi$ plays the role of the RG cutoff, and the truncation by Schmidt eigenvalue is the coarse-graining step. DMRG is exact for the 1D transverse-field Ising model and approximate but extraordinarily accurate for 1D quantum critical points. The 2D extension via PEPS (projected entangled pair states) and the multi-scale entanglement renormalization ansatz MERA (Vidal 2007) generalises the construction to higher dimensions and to critical points where the area law has logarithmic corrections.

MERA is the cleanest contemporary realisation of the Kadanoff picture: each layer of the tensor network corresponds to one RG step, with disentanglers removing short-range entanglement before the coarse-graining isometries reduce the lattice size by a factor of $b$. At a critical point the MERA tensor network has explicit scale invariance — every layer looks the same — and the scaling dimensions of operators can be extracted from the eigenvalues of the descending superoperator. The match between MERA scaling dimensions and CFT primary dimensions for 2D Ising is exact to numerical precision.

Wilson's numerical RG for impurity problems

Wilson 1975 developed a separate numerical RG scheme for quantum impurity problems, most famously for the Kondo Hamiltonian describing a single magnetic impurity in a metallic host. The method discretises the conduction-band continuum onto a logarithmic Wilson chain, then diagonalises the chain Hamiltonian iteratively by adding one site at a time and truncating to the lowest-energy states. The logarithmic discretisation separates energy scales by a factor of $\Lambda$ per step ($\Lambda \approx 2$-$3$ in practice), making the iteration converge geometrically. NRG resolves the Kondo singlet at exponentially small energy scales, computes the impurity thermodynamics across all temperature ranges from the high-$T$ free-spin regime through the Kondo crossover to the low-$T$ singlet, and identifies the strong-coupling fixed point as the Fermi-liquid one with a $\pi /2$ phase shift. The construction is the canonical example of an RG calculation done in real frequency space rather than momentum space, and the impurity exponents extracted from NRG agree with the Bethe-ansatz solution to numerical precision.

Functional renormalization group

The functional renormalization group (Wetterich 1993; Morris 1994) is a non-perturbative reformulation of the Wilsonian programme in terms of the effective average action ${\Gamma }_k[\varphi ]$, which interpolates between the bare action $S$ at the cutoff scale $k=\Lambda$ and the full effective action $\Gamma$ at $k=0$. The exact flow equation, often called the Wetterich equation,

$$k\frac{\partial {\Gamma }_k}{\partial k}=\frac{1}{2}\mathrm{Tr}\left[\frac{k{\partial }_k{R}_k}{{\Gamma }_k^{(2)}+R_k}\right],$$

uses a regulator $R_k$ that suppresses momentum modes below $k$. The trace is over momenta and field components. The equation is exact and one-loop in structure, but solving it requires truncating the form of ${\Gamma }_k$ — typical truncations expand in powers of the field (the derivative expansion) or in vertex functions (vertex expansion). The FRG yields 3D Ising exponents $\nu \approx 0.63$, $\eta \approx 0.037$ from the lowest-derivative truncations, with systematic improvement as higher derivatives are included.

Comparing the schemes

Four numerical RG methods cover four complementary regimes. Monte-Carlo RG operates directly on lattice configurations and excels at classical critical points. DMRG and tensor networks excel at low-dimensional quantum systems and 2D critical points. Wilson's NRG excels at quantum impurity problems with logarithmic energy hierarchies. FRG operates in functional field-theory space and excels at non-perturbative regimes near upper critical dimensions. None replaces the analytic Wilson-Fisher analysis — instead they confirm its qualitative picture (a single relevant operator drives criticality, scaling dimensions encode all exponents, universality classifies the spectrum) and refine its quantitative predictions. The convergence of every numerical RG scheme on the same 3D Ising exponents $\nu =0.62997(1)$ and $\eta =0.0363(1)$ is the strongest available evidence that the RG framework is not just an organising metaphor but a controlled mathematical description of critical phenomena.

Connections [Master]

• Ising model and phase transitions 11.06.01 pending. This unit's Wilson-Fisher analysis is the explanation for why the 3D Ising model and the liquid-vapour critical point in fluids share the exponents $\beta \approx 0.326$ and $\nu \approx 0.630$ despite their unrelated microscopic Hamiltonians. The block-spin construction reviewed here is the calculational ancestor of every modern Ising-class RG; the Onsager exact 2D exponents quoted in 11.06.01 pending are recovered here as the RG eigenvalues $y_t=1$ and $y_h=15/8$ at the $\mathcal{M}(4,3)$ minimal-model fixed point.

• Canonical ensemble 11.04.01 pending. The partition function whose singularity structure encodes the phase transition is the canonical $Z(T,V,N)$ developed in 11.04.01 pending. The RG transformation acts on $Z$ by integrating out high-momentum modes; the free energy density $f=-k_{B}T\ln Z/V$ is the generating function whose non-analyticity at $T_c$ is governed by the scaling form $f_{\mathrm{sing}}(t,h)=b^{-d}{f}_{\mathrm{sing}}(b^{y_t}t,b^{y_h}h)$.

• Quantum field theory and high-energy physics. The Wilsonian RG is the physical meaning of the Gell-Mann-Low renormalization group of 12.01.01 quantum electrodynamics. The RG flow of the ${\varphi }^4$ coupling in $d=4$ — the beta function $\beta (g)=-{\beta }_0{g}^3/(16{\pi }^2)+\dots$ — is the origin of asymptotic freedom in non-Abelian gauge theories (Gross-Politzer-Wilczek 1973). The conceptual framework is identical: the coupling runs with scale, and the direction of flow determines the physics. In QFT the running is logarithmic rather than power-law (because $d=d_c=4$), but the structure is the same.

• Quantum criticality and condensed matter 11.06.01 pending. At zero temperature, quantum fluctuations replace thermal fluctuations as the driver of phase transitions. The quantum-to-classical correspondence maps a $d$-dimensional quantum system at $T=0$ to a $(d+1)$-dimensional classical system, with imaginary time as the extra dimension. The RG framework carries over directly: the upper critical dimension of the quantum transverse-field Ising model is $d_c=3$ (spatial dimensions), and the quantum critical exponents are those of the 4D classical Ising model. Hertz 1976 and Millis 1993 developed the quantum RG formalism.

• Statistical inference and machine learning. The RG has been reinterpreted as a variational optimisation procedure (Li-Wang 2018; Lenggenhager et al. 2020): coarse-graining is compression, and the RG flow minimises a variational free energy. Deep neural networks trained on Ising configurations learn representations that mirror the RG hierarchy (Mehta-Schwab 2014). The mathematical content of this correspondence is still being clarified, but the structural analogy — relevant features survive compression, irrelevant ones are discarded — is precise.

Historical & philosophical context [Master]

The renormalization group has two independent origins that converged in the early 1970s.

Kadanoff 1966 introduced the block-spin picture as a physical argument for scaling and universality, in the paper "Scaling laws for Ising models near $T_c$" ^{[Kadanoff1966]} published in Physics 2, 263. The key insight was that the divergent correlation length at $T_c$ means no single microscopic scale dominates the critical thermodynamics — the physics must organise itself by how it transforms under changes of scale. Kadanoff's paper deduced the scaling form of the free energy and the scaling relations among critical exponents without computing any of them; the exponents themselves required either the exact Onsager 1944 solution ^{[Onsager1944]} in Physical Review 65, 117 (for the 2D Ising model) or a controlled calculational scheme not yet available.

Wilson 1971 turned Kadanoff's picture into a calculational tool through the momentum-shell RG, in the companion papers "Renormalization group and critical phenomena. I" ^{[Wilson1971a]} and "II" ^{[Wilson1971b]} in Physical Review B4, 3174 and 3184. Wilson's three contributions were: (i) the momentum-shell RG itself, which makes the coarse-graining procedure systematic in field theory; (ii) the $ϵ$-expansion (Wilson-Fisher 1972 Physical Review Letters 28, 240 ^{[WilsonFisher1972]}), which provides controlled perturbative access to the Wilson-Fisher fixed point; (iii) the connection to earlier renormalization in quantum field theory (Gell-Mann-Low 1954), unifying the two notions of "renormalization." Wilson's 1974 review with Kogut in Physics Reports 12C, 75-199 ^{[WilsonKogut1974]} is the foundational synthesis. Wilson received the 1982 Nobel Prize in Physics for this body of work.

Fisher 1974 systematised the classification of scaling variables and the notion of "crossover" between fixed points. Zamolodchikov 1986 proved the c-theorem, establishing an irreversibility property of RG flows in 2D and connecting the RG programme to conformal field theory; the 4D analogue (the a-theorem) was proved by Komargodski and Schwimmer in 2011. The conformal bootstrap programme (Kos-Poland-Simmons-Duffin-Vichi 2014-2016) has since produced the most precise numerical values of 3D Ising exponents available — $\nu =0.629971(4)$ — by exploiting unitarity, crossing symmetry, and the spectrum of relevant operators at the fixed point, without ever computing a single Feynman diagram.

The three common misconceptions about the RG are worth flagging in the historical record. First, the RG does not "renormalize away infinities" in the sense of Wilsonian field theory: divergences are an artefact of the cutoff procedure and are absorbed into the bare couplings, while the RG itself is a reorganisation of degrees of freedom by scale. Second, the fixed point is not "the critical point" in the thermodynamic sense: critical points are properties of physical Hamiltonians, while fixed points are properties of the RG transformation in coupling space, and the critical point of a given system corresponds to the basin of attraction of a particular unstable fixed point. Third, the $ϵ$-expansion is not just a Taylor series around $d=4$: the series is asymptotic with factorially growing coefficients, but Borel-resummed it produces controlled quantitative predictions at $ϵ=1$ (i.e., $d=3$) precisely because the operator structure at the Wilson-Fisher fixed point is qualitatively correct already at one loop.

Bibliography [Master]

• Kadanoff, L. P. (1966). "Scaling laws for Ising models near $T_c$." Physics 2, 263–272.
• Wilson, K. G. (1971). "Renormalization group and critical phenomena. I. Renormalization group and the Kadanoff scaling picture." Physical Review B4, 3174–3183.
• Wilson, K. G. (1971). "Renormalization group and critical phenomena. II. Phase-space cell analysis of critical behavior." Physical Review B4, 3184–3205.
• Wilson, K. G. and Kogut, J. (1974). "The renormalization group and the $ϵ$ expansion." Physics Reports 12C, 75–199.
• Fisher, M. E. (1974). "The renormalization group in the theory of critical behavior." Reviews of Modern Physics 46, 597–616.
• Niemeijer, Th. and van Leeuwen, J. M. J. (1974). "Wilson theory for spin systems on a triangular lattice." Physical Review 35, 163–178.
• Belavin, A. A., Polyakov, A. M., and Zamolodchikov, A. B. (1984). "Infinite conformal symmetry in two-dimensional quantum field theory." Nuclear Physics B241, 333–380.
• Zamolodchikov, A. B. (1986). "Irreversibility of the flux of the renormalization group in a 2D field theory." JETP Letters 43, 730–732.
• Goldenfeld, N. (1992). Lectures on Phase Transitions and the Renormalization Group. Addison-Wesley.
• Cardy, J. (1996). Scaling and Renormalization in Statistical Physics. Cambridge University Press.
• Komargodski, Z. and Schwimmer, A. (2011). "Renormalization group flows and anomalies." Journal of High Energy Physics 2011, 99.