The Epsilon Expansion, RG Flows Near Four Dimensions, Critical Exponents, and Conformal Field Theory

Intuition Beginner

In 11.07.03 you saw the block-spin RG: zoom out, rewrite the Hamiltonian, and track how the couplings change. The fixed point of this process determines the critical behaviour. But block spins are hard to compute with -- the coarse-graining map generates an ever-growing set of coupling constants, and nobody has solved the recursion relations exactly in three dimensions.

Wilson and Fisher had an idea of startling elegance: compute the exponents not at $d=3$ but at $d=4-\epsilon$ and treat $\epsilon$ as a small parameter. At $d=4$ the mean-field exponents are correct -- the Gaussian fixed point is stable and the quartic interaction washes out under coarse-graining. As you lower the dimension toward $d=3$, the quartic interaction becomes relevant and a new fixed point -- the Wilson-Fisher fixed point -- splits off from the Gaussian one at a coupling strength proportional to $\epsilon$. Expanding everything in powers of $\epsilon$ and then setting $\epsilon =1$ yields predictions for $d=3$.

This sounds reckless. The expansion parameter is $\epsilon =4-d=1$, which is not small at all. But the resulting critical exponents are accurate to within a few percent of the experimental values. At five-loop order, the epsilon expansion gives $\nu =0.6294$ for the 3D Ising model, compared to the high-precision numerical value $\nu =0.629971$. An expansion around four dimensions that works in three is one of the great success stories of theoretical physics.

The key outputs are the critical exponents. For the 3D Ising model, the epsilon expansion predicts:

• $\nu \approx 0.627$ (one loop) -- experiment gives $0.630$
• $\eta \approx 0.019$ (one loop, starting at ${\epsilon }^2$) -- experiment gives $0.036$
• $\gamma \approx 1.244$ (one loop) -- experiment gives $1.237$

These numbers define whole universality classes: the Ising class ($n=1$), the XY class ($n=2$), and the Heisenberg class ($n=3$). Systems as different as the liquid-gas critical point, the superfluid transition in helium-4, and the ferromagnetic transition in iron belong to these three classes respectively, and within each class the critical exponents are identical.

At the critical point itself, the system is scale-invariant and conformally invariant: it looks the same at every length scale and under every smooth deformation that preserves angles. This is the domain of conformal field theory (CFT), which provides an exact description of the critical point. In two dimensions, CFT is so powerful that you can solve it exactly: the 2D Ising critical point is the minimal model $M(4,3)$ with central charge $c=1/2$, and all its critical exponents are known in closed form.

Visual Beginner

The RG flow near $d=4$ has a distinctive structure. At $d=4$ exactly, the quartic coupling $u$ is marginal -- it neither grows nor shrinks under the RG. At $d=4-\epsilon$ with $\epsilon >0$, two things happen:

1. The Gaussian fixed point at $u=0$ becomes unstable: $u$ is now relevant and grows under coarse-graining.
2. A new fixed point appears at $u^{\ast }\propto \epsilon$, which is stable in the $u$-direction. This is the Wilson-Fisher fixed point.

As $\epsilon$ increases from 0, the Wilson-Fisher fixed point moves further from the origin. At $\epsilon =1$ ($d=3$), it is well separated from the Gaussian fixed point and governs the critical behaviour.

The universality classes are distinguished by the symmetry of the order parameter:

Universality class	Order parameter symmetry	$n$	Physical systems
Ising	$Z_2$ (sign flip)	1	Liquid-gas, uniaxial magnets
XY	$O(2)$ (rotations in plane)	2	Superfluid He-4, thin films
Heisenberg	$O(3)$ (rotations in space)	3	Isotropic ferromagnets

Each class has a different Wilson-Fisher fixed point (labelled by $n$) and different critical exponents.

Worked example Beginner

The one-loop correlation-length exponent for the 3D Ising model. The Wilson-Fisher fixed point for the Ising model ($n=1$) in $d=4-\epsilon$ dimensions has a thermal eigenvalue:

$$y_t=2-\frac{n+2}{n+8}\epsilon +O({\epsilon }^2)=2-\frac{3}{9}\epsilon +O({\epsilon }^2)=2-\frac{\epsilon }{3}+O({\epsilon }^2).$$

The correlation-length exponent is $\nu =1/y_t$. Expand:

$$\nu =\frac{1}{2-\epsilon /3+O({\epsilon }^2)}=\frac{1}{2}\cdot \frac{1}{1-\epsilon /6+O({\epsilon }^2)}\approx \frac{1}{2}\left(1+\frac{\epsilon }{6}\right)=\frac{1}{2}+\frac{\epsilon }{12}.$$

At $\epsilon =1$ (i.e., $d=3$): $\nu \approx 1/2+1/12=7/12\approx 0.583$.

The experimental value for the 3D Ising model is $\nu \approx 0.630$. The one-loop result is off by about 7%, which is impressive for the first term in an expansion around four dimensions applied to three dimensions. Higher-order corrections (two-loop through five-loop) systematically improve the estimate:

Order	$\nu$ ($\epsilon =1$)	Error
Mean-field (${\epsilon }^0$)	0.500	21%
One-loop (${\epsilon }^1$)	0.583	7.5%
Two-loop (${\epsilon }^2$)	0.615	2.4%
Three-loop (${\epsilon }^3$)	0.626	0.6%
Five-loop + Borel	0.6294	0.1%

Check your understanding Beginner

Formal definition Intermediate+

The momentum-shell RG for ${\varphi }^4$ theory

The epsilon expansion is most naturally derived using the momentum-shell RG, which implements Wilson's coarse-graining in Fourier space rather than in real space.

Start from the Landau-Ginzburg Hamiltonian for an $n$-component field $\mathbf{\varphi }(\mathbf{x})$:

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

where $r\propto T-T_c$ is the temperature tuning parameter and $u>0$ is the quartic coupling that stabilises the free energy. The field $\mathbf{\varphi }$ has a momentum cutoff $|\mathbf{q}|\le \Lambda$.

The momentum-shell RG proceeds in three steps:

Step 1: Split. Decompose the field into slow modes ${\mathbf{\varphi }}_{<}$ with $|\mathbf{q}|<\Lambda /b$ and fast modes ${\mathbf{\varphi }}_{>}$ with $\Lambda /b<|\mathbf{q}|<\Lambda$, where $b>1$ is the scale factor.

Step 2: Integrate out. Sum over the fast modes:

$$e^{-H^{\prime }[{\mathbf{\varphi }}_{<}]}=\int \mathcal{D}{\mathbf{\varphi }}_{>}{e}^{-H[{\mathbf{\varphi }}_{<}+{\mathbf{\varphi }}_{>}]}.$$

This integration is performed perturbatively in $u$ using Feynman diagrams (the same diagrammatic technology used in quantum field theory). Each power of $u$ brings one loop to the diagrams.

Step 3: Rescale. Restore the cutoff by rescaling lengths $\mathbf{x}\to \mathbf{x}/b$ and fields $\mathbf{\varphi }\to \zeta \mathbf{\varphi }$ with $\zeta =b^{(d-2+\eta )/2}$, where $\eta$ is the anomalous dimension (zero at the Gaussian fixed point, nonzero at the Wilson-Fisher fixed point).

One-loop beta functions

The result of the momentum-shell integration to one-loop order (first order in $u$) is a set of recursion relations for the couplings. Writing the RG transformation as a continuous flow (with $b=e^{\ell }$ and $\ell$ infinitesimal) gives the beta functions:

$$\frac{dr}{d\ell }=2r+\frac{n+2}{6}{K}_{d}u\frac{{\Lambda }^d}{{\Lambda }^2+r},$$ $$\frac{du}{d\ell }=(4-d)u-\frac{n+8}{6}{K}_d{u}^2\frac{{\Lambda }^d}{({\Lambda }^2+r{)}^2},$$

where $K_d=S_d/(2\pi {)}^d$ is a geometric factor and $S_d=2{\pi }^{d/2}/\Gamma (d/2)$ is the surface area of the unit sphere in $d$ dimensions. Evaluating at $r=0$ and $\Lambda =1$ (in units of the cutoff):

$$\frac{dr}{d\ell }=2r+\frac{n+2}{6}{K}_{d}u,$$ $$\frac{du}{d\ell }=\epsilon u-\frac{n+8}{6}{K}_d{u}^2,$$

where $\epsilon =4-d$.

The Wilson-Fisher fixed point

Setting $du/d\ell =0$ gives two fixed points:

1. Gaussian: $u^{\ast }=0$ (the non-interacting theory).
2. Wilson-Fisher: $u^{\ast }=\frac{6\epsilon }{(n+8)K_4}+O({\epsilon }^2)$, where $K_4=1/(8{\pi }^2)$.

At the Wilson-Fisher fixed point, the thermal eigenvalue (the eigenvalue of the linearised RG in the $r$-direction) is:

$$y_t=2-\frac{n+2}{n+8}\epsilon +O({\epsilon }^2).$$

The eigenvalue in the $u$-direction is:

$$y_u=-\epsilon +O({\epsilon }^2),$$

confirming that the quartic coupling is irrelevant at the Wilson-Fisher fixed point for $\epsilon >0$ -- the fixed point is stable in the $u$-direction.

One-loop critical exponents

The critical exponents follow from $y_t$ and the anomalous dimension $\eta$. To one-loop order:

$$\nu =\frac{1}{y_t}=\frac{1}{2}+\frac{n+2}{4(n+8)}\epsilon +O({\epsilon }^2),$$ $$\eta =O({\epsilon }^2)\text{(the one-loop contribution vanishes)},$$ $$\gamma =(2-\eta )\nu =1+\frac{n+2}{2(n+8)}\epsilon +O({\epsilon }^2),$$ $$\alpha =2-d\nu =\frac{4-n}{2(n+8)}\epsilon +O({\epsilon }^2).$$

The remaining exponents follow from scaling relations:

$$\beta =\frac{\nu }{2}(d-2+\eta ),\delta =\frac{d+2-\eta }{d-2+\eta }.$$

Universality classes

The Wilson-Fisher fixed point depends on the number of order-parameter components $n$. This produces distinct universality classes:

Class	$n$	Symmetry	$\nu$ at $O(\epsilon )$	$\nu$ (3D numerical)
Ising	1	$Z_2$	$1/2+\epsilon /12$	0.630
XY	2	$O(2)$	$1/2+\epsilon /10$	0.671
Heisenberg	3	$O(3)$	$1/2+5\epsilon /44$	0.711
Self-avoiding walk	0	$O(n)\to 0$	$1/2+\epsilon /16$	0.588
Spherical ($n\to \infty$)	$\infty$	$O(\infty )$	$1/2+\epsilon /8$	1.000 (exact)

Key derivation: One-loop beta function for the quartic coupling Intermediate+

Proposition. The one-loop RG beta function for the quartic coupling $u$ in the $O(n)$ Landau-Ginzburg model in $d=4-\epsilon$ dimensions is $du/d\ell =\epsilon u-(n+8)K_4{u}^2/6+O(u^3)$, with a Wilson-Fisher fixed point at $u^{\ast }=6\epsilon /[(n+8)K_4]+O({\epsilon }^2)$.

Proof. The momentum-shell RG integrates out the fast modes ${\mathbf{\varphi }}_{>}$ with momenta $\Lambda /b<|\mathbf{q}|<\Lambda$. The one-loop correction to $u$ comes from the diagram with two quartic vertices connected by two internal lines (the "bubble" diagram).

The quartic vertex $u/4!$ connects four external legs. When two legs from each of two vertices are contracted to form internal propagators, the result is a correction to the remaining four-point vertex. The combinatorial factor for the $O(n)$ model is:

• The number of ways to choose which two of the four legs at each vertex are contracted: there are $\left(\genfrac{}{}{0ex}{}{4}{2}\right)=6$ pairings per vertex.
• The $O(n)$ index structure: contracting two ${\varphi }_i$ fields from each vertex gives a factor ${\delta }_{ij}{\delta }_{ij}=n$ from the index sum, plus a factor ${\delta }_{ik}{\delta }_{jk}$ from the external legs, yielding a total of $n+8$ independent tensor structures.

The momentum integral for the bubble diagram in $d$ dimensions is:

$${\int }_{\Lambda /b}^{\Lambda }\frac{d^{d}q}{(2\pi {)}^d}\frac{1}{(q^2+r{)}^2}\approx K_d\frac{{\Lambda }^d}{({\Lambda }^2+r{)}^2}\ell ,$$

where $\ell =\ln b$ is the RG "time" and $K_d=1/(8{\pi }^2)$ at $d=4$.

Assembling the combinatorial factor and the momentum integral:

$$\delta u=-\frac{n+8}{6}{K}_4{u}^2\ell .$$

Adding the engineering dimension $[u]=4-d=\epsilon$ (the quartic coupling has positive mass dimension $\epsilon$ in $d<4$):

$$\frac{du}{d\ell }=\epsilon u-\frac{n+8}{6}{K}_4{u}^2+O(u^3).$$

Fixed points satisfy $du/d\ell =0$: $u^{\ast }=0$ (Gaussian) or $u^{\ast }=6\epsilon /[(n+8)K_4]$ (Wilson-Fisher). $\square$

Bridge. This beta function is the engine of the epsilon expansion. It shows that the quartic coupling, which is marginal at $d=4$, develops a Wilson-Fisher fixed point for any $\epsilon >0$. The $n$-dependence in the denominator $(n+8)$ determines how the fixed-point location and the critical exponents vary across universality classes. The scaling hypothesis of 11.06.05 is given a microscopic foundation: the exponents predicted by scaling relations are now computed explicitly from the RG flow.

Exercises Intermediate+

Lean formalization Intermediate+

The epsilon expansion is a formal power series in $\epsilon =4-d$ whose coefficients are rational functions of $n$. Formalising it requires: (1) a type of coupling constants $(r,u)$ with a continuous RG flow given by the beta functions ${\beta }_r$ and ${\beta }_u$; (2) a fixed-point condition $\beta (u^{\ast })=0$ solvable perturbatively in $\epsilon$; (3) the eigenvalues of the linearised RG at the fixed point, expressed as power series in $\epsilon$; (4) the critical exponents as functions of these eigenvalues; (5) the anomalous dimension $\eta$ as a two-loop correction to the field rescaling. The algebraic manipulation of formal power series is within Mathlib's scope (via PowerSeries), but the physical content -- the Feynman-diagrammatic derivation of the beta function, the identification of the fixed point with a physical phase transition, and the interpretation of the exponents as measurable quantities -- requires domain-specific construction. The conformal field theory connection adds the Virasoro algebra and its representation theory, which is a substantial formalisation project in its own right.

Advanced results Master

Two-loop corrections

The one-loop beta function captures the leading dependence on $\epsilon$ but misses important corrections. At two-loop order, the beta function for $u$ acquires an $O(u^3)$ term:

$$\frac{du}{d\ell }=\epsilon u-\frac{n+8}{6}{K}_4{u}^2+\frac{(n+8{)}^2}{36}\left(\frac{6}{n+8}+\frac{3}{n+8}\right)K_4^2{u}^3+O(u^4),$$

and the anomalous dimension $\eta$ receives its first nonzero contribution:

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

For $n=1$ (Ising): $\eta ={\epsilon }^2/54$. At $\epsilon =1$: $\eta \approx 0.0185$, compared to the numerical value $\eta \approx 0.036$.

The thermal eigenvalue at the Wilson-Fisher fixed point to two-loop order is:

$$y_t=2-\frac{n+2}{n+8}\epsilon -\frac{(n+2)(n+3)}{2(n+8{)}^3}{\epsilon }^2+O({\epsilon }^3),$$

giving:

$$\nu =\frac{1}{2}+\frac{n+2}{4(n+8)}\epsilon +\frac{(n+2)(n^2+23n+60)}{8(n+8{)}^3}{\epsilon }^2+O({\epsilon }^3).$$

For $n=1$ at $\epsilon =1$: $\nu \approx 0.500+0.083+0.032=0.615$. The numerical value is 0.630.

Borel resummation

The epsilon expansion is an asymptotic series: the coefficients $a_k$ in $\nu ={\sum }_{k=0}^{\infty }{a}_k{\epsilon }^k$ grow factorially, $|a_k|\sim k!$, so the partial sums diverge for any fixed $\epsilon >0$. Nevertheless, the series contains accurate information that can be extracted by Borel resummation.

The Borel transform is:

$$B(\epsilon t)=\sum _{k=0}^{\infty }\frac{a_k}{k!}(\epsilon t{)}^k,$$

which has a finite radius of convergence (the factorial growth of $a_k$ is cancelled by $1/k!$). The Borel sum is:

$${\nu }_{\text{Borel}}(\epsilon )={\int }_0^{\infty }{e}^{-t}B(\epsilon t)dt.$$

If $B$ has no singularities on the positive real axis, this integral converges and provides a well-defined value for $\nu$. The five-loop epsilon expansion, combined with Borel resummation and conformal mapping (which maps the singularities of $B$ away from the integration contour), gives:

$${\nu }_{\text{Ising}}^{d=3}=0.6294\pm 0.0003,{\eta }_{\text{Ising}}^{d=3}=0.0335\pm 0.0025,$$

in excellent agreement with high-precision Monte Carlo simulations and the conformal bootstrap.

Conformal field theory: the continuum description of critical points

At a second-order phase transition, the correlation length diverges and the system becomes scale-invariant. In fact, scale invariance at a critical point is enhanced to conformal invariance: the system is invariant under all angle-preserving transformations, not just dilations. The critical point is therefore described by a conformal field theory (CFT).

A CFT in $d$ dimensions is characterised by:

1. Primary operators ${\mathcal{O}}_i$ with definite scaling dimensions ${\Delta }_i$. Under a scale transformation $\mathbf{x}\to \lambda \mathbf{x}$:

$${\mathcal{O}}_i(\lambda \mathbf{x})={\lambda }^{-{\Delta }_i}{\mathcal{O}}_i(\mathbf{x}).$$

2. The operator product expansion (OPE), which encodes the algebraic structure of the theory. As two operators approach each other:

$${\mathcal{O}}_i(\mathbf{x}){\mathcal{O}}_j(0)=\sum _k{C}_{ijk}|\mathbf{x}{|}^{{\Delta }_k-{\Delta }_i-{\Delta }_j}{\mathcal{O}}_k(0)+\text{descendants},$$

where $C_{ijk}$ are the OPE coefficients (structure constants) and "descendants" are derivatives of primary operators.

3. Conformal symmetry, generated by translations ($P_{\mu }$), rotations ($M_{\mu \nu }$), dilations ($D$), and special conformal transformations ($K_{\mu }$). In $d$ dimensions, the conformal group is $SO(d+1,1)$.

The critical exponents of a statistical mechanical system are related to the scaling dimensions of CFT operators by:

$${\Delta }_{\varphi }=\frac{d-2+\eta }{2},{\Delta }_{ϵ}=d-\frac{1}{\nu },$$

where $\varphi$ is the order-parameter field (the spin) and $ϵ={\varphi }^2$ is the energy-density operator.

Two-dimensional CFT: the Virasoro algebra and minimal models

In two dimensions, conformal invariance is vastly more powerful than in higher dimensions. The conformal group becomes infinite-dimensional (generated by all holomorphic and antiholomorphic transformations), and the algebra of conformal transformations is the Virasoro algebra:

$$[L_m,L_n]=(m-n)L_{m+n}+\frac{c}{12}(m^3-m){\delta }_{m+n,0},$$

where $c$ is the central charge, a number that characterises the CFT. The central charge plays a role analogous to the number of degrees of freedom: it counts the "effective number of massless fields" at the critical point.

Minimal models are a special class of 2D CFTs with central charge:

$$c=1-\frac{6(p-q{)}^2}{pq},p,q\in {\mathbb{Z}}_{>0},p,q\ge 2,\gcd (p,q)=1.$$

These theories have a finite number of primary operators and are exactly solvable. Their scaling dimensions are:

$${\Delta }_{r,s}=\frac{(pr-qs{)}^2-(p-q{)}^2}{4pq},1\le r\le q-1,1\le s\le p-1.$$

The 2D Ising model as a minimal model

The critical point of the 2D Ising model is identified with the minimal model $M(4,3)$, which has:

• Central charge $c=1-6(4-3{)}^2/(4\cdot 3)=1-6/12=1-1/2=1/2$.
• Three primary operators:
  • Identity $1$ with $\Delta =0$.
  • Spin $\sigma$ with ${\Delta }_{\sigma }=(2\cdot 1-3\cdot 1{)}^2-(4-3{)}^2/(4\cdot 4\cdot 3)=(1/16)$... More precisely: ${\Delta }_{1,2}=[(4\cdot 1-3\cdot 2{)}^2-1]/(48)=[4-1]/48=3/48=1/16$. This gives $\eta =2{\Delta }_{\sigma }=1/4$, so the spin-spin correlation function decays as $⟨\sigma (0)\sigma (r)⟩\sim r^{-1/4}$. The exact result from Onsager's solution is $\eta =1/4$, confirming the identification.
  • Energy $ϵ$ with ${\Delta }_{ϵ}={\Delta }_{2,1}=[(4\cdot 2-3\cdot 1{)}^2-1]/48=[25-1]/48=1/2$. This gives $1/\nu =d-{\Delta }_{ϵ}=2-1/2=3/2$, so $\nu =2/3$, agreeing with the exact result.

The OPE coefficients are also determined by the minimal model structure. For instance:

$$\sigma \times \sigma =1+ϵ,\sigma \times ϵ=\sigma ,ϵ\times ϵ=1.$$

This is the operator algebra of the 2D Ising CFT: two spins combine to give either the identity or the energy density; a spin and an energy density combine to give a spin; and two energy densities combine to give the identity.

The c-theorem and the RG

Zamolodchikov (1986) proved the c-theorem: in any unitary 2D quantum field theory, there exists a function $C(g)$ of the coupling constants $g$ that (a) equals the central charge $c$ at an RG fixed point, (b) decreases monotonically along RG flows ($dC/d\ell \le 0$), and (c) is stationary only at fixed points.

The $c$-theorem provides a notion of "irreversibility" for RG flows: the effective number of degrees of freedom decreases as you coarse-grain, analogous to entropy production. The Gaussian fixed point in 2D has $c=1$ (the free boson), and the Ising fixed point has $c=1/2$, so $c$ decreases along the flow from the Gaussian to the Ising fixed point.

In four dimensions, the analogous result is the a-theorem (proved by Komargodski and Schwimmer in 2011), where the "a-anomaly" plays the role of $c$.

Connections between CFT and the epsilon expansion

The epsilon expansion and CFT are two complementary approaches to the same physics:

1. The epsilon expansion computes critical exponents perturbatively as power series in $\epsilon$. It works in any dimension but gives only asymptotic series.

2. CFT provides the exact description of the critical point as a fixed point of the RG. In 2D, CFT gives exact results (closed-form expressions for all critical exponents and correlation functions). In higher dimensions, CFT is less tractable analytically but the conformal bootstrap (a numerical method that uses crossing symmetry of four-point functions) has produced the most precise critical exponents to date.

The connection is precise: the RG eigenvalues at the Wilson-Fisher fixed point are related to the scaling dimensions of operators in the corresponding CFT by $y_i=d-{\Delta }_i$. The epsilon expansion computes the anomalous parts of ${\Delta }_i$ (the deviations from their Gaussian values) as power series in $\epsilon$.

Higher-loop results and the state of the art

The epsilon expansion has been carried out to extraordinary order:

• Five-loop order for the $O(n)$ model: the beta function and anomalous dimensions are known through $O(u^6)$, giving critical exponents as polynomials in $\epsilon$ up to ${\epsilon }^5$. After Borel resummation at $\epsilon =1$, these give critical exponents accurate to four decimal places.

• Conformal bootstrap: a completely independent numerical method that uses the crossing symmetry and unitarity of CFT four-point functions to bound the scaling dimensions. For the 3D Ising model, the bootstrap gives $\nu =0.629971(4)$ and $\eta =0.036298(2)$, which are the most precise determinations available.

• Agreement: the five-loop epsilon expansion and the conformal bootstrap agree to within their respective error bars, providing a stringent cross-check on both methods.

Full proof set Master

Proposition. The anomalous dimension $\eta$ vanishes at one-loop order in the epsilon expansion for the $O(n)$ Landau-Ginzburg model, and first appears at two loops: $\eta =(n+2){\epsilon }^2/[2(n+8{)}^2]+O({\epsilon }^3)$.

Proof. The anomalous dimension $\eta$ is defined through the field rescaling $\varphi \to \zeta \varphi$ with $\zeta =b^{(d-2+\eta )/2}$. Equivalently, $\eta$ is related to the wavefunction renormalisation $Z_{\varphi }$ by $\eta =d\ln Z_{\varphi }/d\ell$.

At one-loop order, the only diagrams that contribute are the one-loop corrections to the two-point function (the "tadpole" and the "self-energy" diagrams). The tadpole diagram shifts the mass $r$ but does not depend on the external momentum $q$, so it does not contribute to $Z_{\varphi }$. The self-energy diagram at one loop involves a single quartic vertex and a single internal propagator; by momentum conservation, the internal momentum equals the external momentum, but the resulting integral is independent of $q$ after angular averaging in $d$ dimensions (the angular average of $q_{\mu }{q}_{\nu }$ is $q^2{\delta }_{\mu \nu }/d$, which renormalises the mass but not the kinetic term).

At two-loop order, the sunset diagram (two quartic vertices connected by three internal lines) provides the first momentum-dependent correction to the two-point function. Evaluating this diagram requires a two-loop momentum integral:

$${\Sigma }^{(2)}(q)\propto u^{\ast 2}\int \frac{d^d{k}_1}{(2\pi {)}^d}\frac{d^d{k}_2}{(2\pi {)}^d}\frac{k_1\cdot k_2}{(k_1^2)(k_2^2)((q-k_1-k_2{)}^2)}.$$

The $q$-dependent part of this integral gives $\partial {\Sigma }^{(2)}/\partial (q^2){|}_{q=0}\propto u^{\ast 2}\propto {\epsilon }^2$, hence $\eta \propto {\epsilon }^2$.

The exact coefficient is computed by evaluating the two-loop integral with dimensional regularisation. The result for the $O(n)$ model is:

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

For $n=1$: $\eta =3/(2\cdot 81){\epsilon }^2={\epsilon }^2/54$. $\square$

Proposition (c-theorem, Zamolodchikov 1986). In any unitary, translation-invariant, and Poincare-invariant 2D quantum field theory with a real, discrete spectrum, there exists a function $C(g)$ of the coupling constants $g$ that satisfies: (a) $C(g)\ge 0$; (b) $C$ is stationary at fixed points, where $C=c$ (the central charge of the corresponding CFT); (c) $C$ decreases monotonically under RG flow: $\mu dC/d\mu \le 0$ for all $\mu$.

Sketch of proof. Consider the two-point functions of the stress-tensor components $T(z)$ and $\bar{T}(\bar{z})$ in a generic 2D QFT. By Poincare invariance, the most general form of the two-point function is:

$$⟨T(z)T(0)⟩=\frac{c(z\bar{z})}{2z^4},$$

where $c(z\bar{z})$ is a function of the single invariant $z\bar{z}=r^2$. Define $C(r)=c(r^2)$.

At a fixed point (scale-invariant theory), $C$ must be constant (no scale to depend on), and comparison with the CFT result $⟨T(z)T(0)⟩=c/(2z^4)$ shows $C=c$.

The positivity of $C$ follows from the unitarity of the theory (reflection positivity of the Euclidean correlator). The monotonicity $dC/dr\le 0$ follows from the conservation law $\bar{\partial }T+\partial \Theta =0$ (where $\Theta$ is the trace of the stress tensor, which vanishes at a fixed point). Zamolodchikov shows that:

$$\mu \frac{dC}{d\mu }=-\frac{3}{2}\int d^{2}x|x{|}^2⟨\Theta (x)\Theta (0)⟩\le 0,$$

where the inequality follows from unitarity (the spectral representation of the correlator ensures non-negative spectral density). $\square$

Connections Master

• 11.06.05 The scaling hypothesis of Widom and Kadanoff postulated scaling relations among the critical exponents. The epsilon expansion derives these relations from first principles and computes the exponents explicitly.
• 11.07.03 The block-spin RG of 11.07.03 provided the conceptual framework (fixed points, relevant and irrelevant operators, universality). The momentum-shell RG and epsilon expansion of this unit provide the quantitative computational machinery.
• 11.07.04 The Kosterlitz-Thouless transition in 2D is a special RG flow driven by topological defects; it contrasts with the Wilson-Fisher fixed point, which is driven by the quartic interaction. The KT transition has $\eta =1/4$ at the critical point, which coincides with the 2D Ising value but arises from a different mechanism.
• 11.06.04 The Ginzburg criterion states that mean-field theory breaks down for $d<4$. The epsilon expansion confirms this: the Gaussian fixed point becomes unstable at $d=4$, and the Wilson-Fisher fixed point takes over for $d<4$.
• 11.06.03 Landau theory gives the mean-field critical exponents. These are recovered exactly at the Gaussian fixed point ($u^{\ast }=0$) in $d>4$. The Wilson-Fisher fixed point ($u^{\ast }>0$) provides the corrected exponents for $d<4$.

Historical and philosophical context Master

Wilson and Fisher's 1972 paper "Critical exponents in 3.99 dimensions" (Phys. Rev. Lett. 28, 240) is one of the most consequential single papers in statistical mechanics. The idea -- treat spatial dimension as a continuous variable and expand around the upper critical dimension -- was both audacious and technically natural. Wilson had already formulated the momentum-shell RG and computed the one-loop beta function. Fisher had independently explored perturbative methods near $d=4$. Their collaboration produced a systematic method that could be pushed to arbitrary order.

The epsilon expansion was developed through five loops by a sequence of authors. Wilson and Fisher computed one loop (1972). Brezin, Le Guillou, Zinn-Justin, and Nickel pushed the calculation to two, three, and eventually five loops (1977--1995). Each additional loop required evaluating increasingly complex Feynman diagrams; the five-loop calculation involved hundreds of diagrams and was a major computational achievement.

The discovery that the series is asymptotic (zero radius of convergence) could have been a fatal blow. Instead, it motivated the development of Borel resummation techniques specifically adapted to the epsilon expansion. The Borel method, combined with conformal mapping (to account for the singularity structure of the Borel transform), extracts remarkably accurate values from the divergent partial sums. Lipatov (1977) showed that the factorial growth of the coefficients is determined by instanton configurations (saddle points of the Euclidean action), giving the epsilon expansion a deep connection to non-perturbative physics.

The conformal field theory revolution came in 1984 with Belavin, Polyakov, and Zamolodchikov's paper "Infinite conformal symmetry in two-dimensional quantum field theory" (Nucl. Phys. B 241, 333). BPZ showed that the infinite-dimensional conformal symmetry in 2D constrains the theory so severely that it can be solved exactly. Their classification of minimal models and the exact computation of critical exponents for the 2D Ising model ($c=1/2$, $\nu =2/3$, $\eta =1/4$) provided a benchmark that the epsilon expansion, despite its power, could never match in 2D.

Zamolodchikov's c-theorem (1986) provided the first rigorous connection between CFT and the RG: the central charge $c$ is a monotonic function of the RG scale, decreasing along RG flows. This is a rigorous statement of the physical intuition that coarse-graining "loses information" -- the effective number of massless degrees of freedom decreases. The higher-dimensional analogue (the a-theorem in 4D) was proved by Komargodski and Schwimmer in 2011, nearly 25 years later.

The conformal bootstrap, revived by Rattazzi, Rychkov, Tonni, and Vichi in 2008, is the modern successor to both the epsilon expansion and CFT. It uses the crossing symmetry and unitarity of four-point correlation functions to derive rigorous bounds on scaling dimensions. For the 3D Ising model, the bootstrap has produced critical exponents to six decimal places ($\nu =0.629971(4)$, $\eta =0.036298(2)$), far exceeding the precision of the epsilon expansion and rivalling the best Monte Carlo results.

The philosophical content is rich. The epsilon expansion embodies a strategy that physicists use repeatedly: solve a problem in a regime where it is easy (here, $d=4$ where mean-field theory works), and then perturb away from that regime using a parameter (here, $\epsilon =4-d$) that measures the deviation. This strategy works when there is a continuous family of solutions connecting the easy regime to the hard one -- and when the expansion, even if formally divergent, contains enough information to be resummed. The success of this strategy in critical phenomena is a testament to the analytic structure of the RG flow: the Wilson-Fisher fixed point is an analytic function of $d$ near $d=4$, and its properties (though not the individual terms of the series) are well-behaved at $d=3$.

CFT provides a complementary philosophy: instead of computing approximate answers in three dimensions, solve the problem exactly at the critical point in two dimensions, where the infinite-dimensional conformal symmetry makes the theory exactly solvable. This "exact solution at a special dimension" strategy is also a recurring pattern in physics -- it underlies the exact solutions of the 1D and 2D Ising models, the Bethe ansatz for integrable systems, and the AdS/CFT correspondence in string theory.

Bibliography -- Epsilon expansion and perturbative RG Master

@article{wilson-fisher1972,
  author = {Wilson, Kenneth G. and Fisher, Michael E.},
  title = {Critical exponents in 3.99 dimensions},
  journal = {Phys. Rev. Lett.},
  volume = {28},
  pages = {240--243},
  year = {1972}
}

@article{wilson-kogut1974,
  author = {Wilson, Kenneth G. and Kogut, John},
  title = {The renormalization group and the $\varepsilon$ expansion},
  journal = {Phys. Reports},
  volume = {12C},
  pages = {75--199},
  year = {1974}
}

@article{brezin1976,
  author = {Br\'{e}zin, Edouard and Le Guillou, Jean Claude and Zinn-Justin, Jean},
  title = {Discussion of the $\varepsilon$-expansion for the $n$-vector model},
  journal = {Phys. Rev. D},
  volume = {15},
  pages = {1544--1556},
  year = {1977}
}

@article{fisher1974,
  author = {Fisher, Michael E.},
  title = {The renormalization group in the theory of critical behavior},
  journal = {Rev. Mod. Phys.},
  volume = {46},
  pages = {597--616},
  year = {1974}
}

@article{lipatov1977,
  author = {Lipatov, L. N.},
  title = {Divergence of the perturbation-theory series and the quasi-classical theory},
  journal = {Sov. Phys. JETP},
  volume = {45},
  pages = {216--223},
  year = {1977}
}

@book{kardar2007,
  author = {Kardar, Mehran},
  title = {Statistical Physics of Fields},
  publisher = {Cambridge University Press},
  year = {2007}
}

@book{goldenfeld1992,
  author = {Goldenfeld, Nigel},
  title = {Lectures on Phase Transitions and the Renormalization Group},
  publisher = {Addison-Wesley},
  year = {1992}
}

Bibliography -- Conformal field theory Master

@article{belavin1984,
  author = {Belavin, A. A. and Polyakov, A. M. and Zamolodchikov, A. B.},
  title = {Infinite conformal symmetry in two-dimensional quantum field theory},
  journal = {Nucl. Phys. B},
  volume = {241},
  pages = {333--380},
  year = {1984}
}

@article{zamolodchikov1986,
  author = {Zamolodchikov, A. B.},
  title = {Irreversibility of the flux of the renormalization group in a {2D} field theory},
  journal = {JETP Lett.},
  volume = {43},
  pages = {730--732},
  year = {1986}
}

@book{difrancesco1997,
  author = {Di Francesco, Philippe and Mathieu, Pierre and S\'{e}n\'{e}chal, David},
  title = {Conformal Field Theory},
  publisher = {Springer},
  year = {1997}
}

@article{rattazzi2008,
  author = {Rattazzi, Riccardo and Rychkov, Slava and Tonni, Erik and Vichi, Alessandro},
  title = {Bounding scalar operator dimensions in 4{D} field theory},
  journal = {JHEP},
  volume = {0812},
  pages = {031},
  year = {2008}
}

@article{komargodski2011,
  author = {Komargodski, Zohar and Schwimmer, Adam},
  title = {On renormalization group flows in four dimensions},
  journal = {JHEP},
  volume = {1112},
  pages = {099},
  year = {2011}
}