Critical Exponents from the RG

Critical exponents measure how thermodynamic quantities blow up or vanish as a system approaches a continuous phase transition: the specific heat as $|t|^{-\alpha}$ , the magnetization as $(-t)^{\beta}$ , the susceptibility as $|t|^{-\gamma}$ , the correlation length as $|t|^{-\nu}$ . For a century these numbers looked like a disconnected zoo of empirical constants. The renormalization group collapses the entire zoo onto a tiny handful of numbers — the eigenvalues of one matrix, the linearization of the RG flow at the fixed point that governs the transition. This is the deepest single payoff of the whole RG programme. The construction has three moves. First, find the fixed point $K^*$ where coarse-graining-plus-rescaling leaves the couplings unchanged. Second, linearize the map there; its eigenvalues, written $\lambda_i=b^{y_i}$ for a rescaling factor $b$ , sort perturbations into relevant, irrelevant, and marginal. Third, attach the two relevant scaling fields to the physical control parameters — reduced temperature $t$ and field $h$ — and read the singular free energy as a generalized homogeneous function. Every exponent then follows from just $y_t$ , $y_h$ , and the spatial dimension $d$ . The consequences are striking. The correlation-length exponent is simply $\nu=1/y_t$ ; the specific-heat exponent is $\alpha=2-d/y_t$ ; the order-parameter, susceptibility, isotherm, and correlation exponents are equally short formulas in $y_t$ and $y_h$ . The famous scaling relations — Rushbrooke, Widom, Fisher, and hyperscaling — are then not extra assumptions but algebraic identities that fall out automatically. Feed in the two-dimensional Ising eigenvalues $(y_t,y_h)=(1,\tfrac{15}{8})$ and Onsager's exact $\beta=\tfrac18$ , $\gamma=\tfrac74$ , $\nu=1$ reappear; feed in the Gaussian eigenvalues and the mean-field set returns. The RG does not just compute exponents — it explains why they are related and why they are universal.

Key takeaways

Linearizing the RG map at the critical fixed point gives a matrix whose eigenvalues $\lambda_i=b^{y_i}$ define the RG exponents $y_i$ .
The two relevant directions of an ordinary critical point are thermal ( $y_t$ ) and magnetic ( $y_h$ ); together with the dimension $d$ they fix every other exponent.
The correlation-length exponent is $\nu=1/y_t$ ; the rest follow from $\alpha=2-d/y_t$ , $\beta=(d-y_h)/y_t$ , $\gamma=(2y_h-d)/y_t$ , $\delta=y_h/(d-y_h)$ , $\eta=d+2-2y_h$ .
Scaling relations (Rushbrooke, Widom, Fisher, hyperscaling) become identities, not independent laws, because every exponent is built from $y_t,y_h,d$ .
Universality is explained: systems flowing to the same fixed point share the same $\{y_i\}$ and hence the same exponents.
Concrete values come from the Gaussian and Wilson–Fisher fixed points and the $\epsilon$ -expansion; irrelevant eigenvalues set the corrections to scaling.
Key idea: Diagonalize the linearized RG at the fixed point; its eigenvalues $\lambda_i=b^{y_i}$ are the exponents. Two relevant directions $(y_t,y_h)$ plus $d$ generate every standard critical exponent and every scaling relation.

The core insight

Diagonalize the linearized RG at the fixed point; its eigenvalues $\lambda_i=b^{y_i}$ are the exponents. Two relevant directions $(y_t,y_h)$ plus dimension $d$ generate every standard critical exponent and every scaling relation.

Where this lives today

The RG derivation of critical exponents is now the standard against which everything else is measured, and the numbers themselves have been pushed to extraordinary precision. The conformal bootstrap — a completely independent, non-perturbative method — pins the 3D Ising exponents to many digits ( $\nu=0.629971\ldots$ , $\eta=0.036298\ldots$ ), and these agree with high-statistics Monte Carlo and with resummed high-order $\epsilon$ -expansions. That three utterly different routes converge on the same eigenvalues is the strongest possible confirmation that the fixed point, not the microscopic model, owns the physics. Practically, the homogeneity relation is the engine of finite-size scaling. Choosing $b=L$ in (4) gives $f_s(t,h,L^{-1})=L^{-d}\Phi(L^{y_t}t,L^{y_h}h)$ , so plotting $L^{\gamma/\nu}\chi$ against $L^{1/\nu}t$ collapses Monte Carlo data for different system sizes onto a single curve — a routine, direct measurement of $y_t=1/\nu$ and $y_h$ . The same eigenvalue language classifies quantum critical points in correlated electron systems, the deconfinement transition in lattice gauge theory, and the scaling of neural-network and percolation models, wherever a diverging correlation length appears.

The physical picture

Think of the critical fixed point as a mountain pass in coupling space. Along the relevant (unstable) directions a trajectory accelerates away; the rate of departure per RG step is exactly the exponent $y_i$ . A system sitting precisely at its critical temperature lies on the stable manifold and flows into the pass. Tune the temperature a little and you kick it onto the relevant thermal direction; how violently the macroscopic singularity then blows up is a direct read-out of that departure rate. The counting is beautifully economical. Each RG step strips short-distance detail and shrinks all lengths by $b$ , so after $n$ steps the correlation length measured in the new units is $\xi/b^n$ . Meanwhile the relevant thermal field has grown to $b^{n y_t}t$ . The iteration is useful until that field reaches order one, which happens after $b^{n y_t}|t|\sim1$ , i.e. $b^n\sim|t|^{-1/y_t}$ . Since the true correlation length is $b^n$ times an order-one number, $\xi\sim|t|^{-1/y_t}$ and the exponent is literally "how fast the deviation grows per decade of length scale." The punchline figure below is the whole chapter on one page: the two relevant eigenvalues $(y_t,y_h)$ feed a single formula box, and out come all six exponents — with the exact 2D-Ising row ( $\beta=\tfrac18$ , the famous outlier) and the mean-field row landing on the same axes, from the same machine.

Figure

Two eigenvalues in, six exponents out — the 2D-Ising outlier on the same machine

Universality now has a one-line explanation. Two microscopically different Hamiltonians — a lattice gas and a uniaxial ferromagnet, say — can lie on the same stable manifold of one fixed point. Their differences are irrelevant couplings, erased under iteration; only the relevant eigenvalues, shared by both, survive to set the exponents. The fixed point, not the lattice, owns the critical behaviour. And the memory of the irrelevant directions is not entirely lost: it returns as corrections to scaling, shrinking as $b^{n y_{\text{irr}}}$ with $y_{\text{irr}}<0$ , and it is these that bend the leading power law at any finite distance from $T_c$ .

Figure

Same fixed point, same exponents: universality as convergent flow

The core equations

Everything hangs on the fixed-point condition and the linearization about it. The fixed point is where the RG map returns the couplings unchanged.

R_b(K^*) = K^* \qquad (1)

where: $R_b$ — the RG map (coarse-grain by $b$ , then rescale); $K$ — a point in the space of dimensionless couplings; $K^*$ — the fixed point governing the transition; $b>1$ — linear rescaling factor. A fixed point is a scale-invariant theory: it looks the same at every magnification, which is exactly what a critical system with $\xi=\infty$ does. Linearize the map about $K^*$ . The Jacobian $T$ is the object whose spectrum we need.

\delta K'_i = \sum_j T_{ij}\,\delta K_j,\qquad T_{ij}=\left.\frac{\partial K'_i}{\partial K_j}\right|_{K^*} \qquad (2)

where: $\delta K=K-K^*$ — deviation from the fixed point; $T_{ij}$ — linearized RG map (Jacobian); $\delta K'$ — deviation after one step. Its left eigenvectors define the scaling fields $g_\sigma$ , and — by the same semigroup argument that governs all RG eigenvalues — each eigenvalue is a power of $b$ .

g_\sigma' = b^{y_\sigma}\, g_\sigma \qquad (3)

where: $g_\sigma$ — scaling field (eigen-combination of couplings); $y_\sigma$ — RG eigenvalue exponent; relevant if $>0$ , irrelevant if $<0$ . The two relevant fields of an ordinary critical point are the thermal field $g_t\equiv t$ with exponent $y_t$ and the magnetic field $g_h\equiv h$ with exponent $y_h$ . The master equation of the whole subject is the homogeneity of the singular free energy, which follows because coarse-graining sums local free energies and the volume shrinks by $b^d$ per step.

f_s(t,h) = b^{-d}\,f_s\!\left(b^{y_t}t,\;b^{y_h}h\right) \qquad (4)

where: $f_s$ — singular part of the free-energy density; $t$ — reduced temperature $(T-T_c)/T_c$ ; $h$ — reduced field; $d$ — spatial dimension; $b^{-d}$ — the volume rescaling. This one functional equation, with the freedom to choose $b$ , generates every thermodynamic exponent. The correlation length obeys the same self-similarity, $\xi(t)=b\,\xi(b^{y_t}t)$ , which immediately gives

\nu = \frac{1}{y_t},\qquad \xi\sim|t|^{-\nu} \qquad (5)

where: $\xi$ — correlation length; $\nu$ — correlation-length exponent. Choosing $b^{y_t}|t|=1$ collapses $\xi$ to a pure power of $|t|$ . The remaining thermodynamic exponents come from differentiating $f_s$ ; the correlation-function exponent $\eta$ comes from the field's scaling dimension.

\alpha = 2-\frac{d}{y_t},\quad \beta=\frac{d-y_h}{y_t},\quad \gamma=\frac{2y_h-d}{y_t},\quad \delta=\frac{y_h}{d-y_h},\quad \eta=d+2-2y_h \qquad (6)

where: $\alpha$ — specific-heat exponent ( $C\sim|t|^{-\alpha}$ ); $\beta$ — order-parameter exponent ( $m\sim(-t)^\beta$ ); $\gamma$ — susceptibility exponent ( $\chi\sim|t|^{-\gamma}$ ); $\delta$ — critical-isotherm exponent ( $m\sim h^{1/\delta}$ at $t=0$ ); $\eta$ — Fisher exponent ( $G(r)\sim r^{-(d-2+\eta)}$ at criticality). Two numbers $(y_t,y_h)$ and the dimension $d$ fix all six. Finally, the irrelevant directions do not touch these leading exponents but set the corrections to scaling.

g_{\text{irr}}'=b^{y_{\text{irr}}}g_{\text{irr}},\quad y_{\text{irr}}<0,\qquad \omega=-y_{\text{irr}} \qquad (7)

where: $g_{\text{irr}}$ — leading irrelevant scaling field; $\omega$ — correction-to-scaling (Wegner) exponent, giving subleading terms $\sim|t|^{\omega\nu}$ .

Derivation AThe thermodynamic exponents from free-energy homogeneity

The single functional equation $f_s(t,h)=b^{-d}f_s(b^{y_t}t,b^{y_h}h)$ contains all of thermodynamic criticality, because $b$ is ours to choose at every step. We exploit that freedom twice: once to eliminate the temperature argument (giving $\alpha,\beta,\gamma$ ) and once to eliminate the field argument (giving $\delta$ ). The strategy is to fix $b$ so that one argument becomes exactly one, turning the functional equation into an explicit power law, and then differentiate to reach each response function. Watch for the standard subtlety that the specific heat needs two temperature derivatives while the susceptibility needs two field derivatives.

Step 1. Fix the scale so the thermal argument is unity:

b^{y_t}|t|=1 \;\Rightarrow\; b=|t|^{-1/y_t},

which turns (4) into $f_s(t,h)=|t|^{d/y_t}\,\Phi\!\big(h|t|^{-y_h/y_t}\big)$ for a scaling function $\Phi$ .

Step 2. Specific heat by two temperature derivatives at $h=0$ :

C\sim\partial_t^2 f_s\sim|t|^{d/y_t-2}\;\Rightarrow\;\alpha=2-\frac{d}{y_t}.

Step 3. Order parameter by one field derivative at $h=0$ :

m=-\partial_h f_s\big|_{h=0}\sim|t|^{(d-y_h)/y_t}\;\Rightarrow\;\beta=\frac{d-y_h}{y_t}.

Step 4. Susceptibility by two field derivatives:

\chi=\partial_h^2 f_s\big|_{h=0}\sim|t|^{(d-2y_h)/y_t}\;\Rightarrow\;\gamma=\frac{2y_h-d}{y_t}.

Step 5. Critical isotherm: now set $t=0$ and choose $b^{y_h}h=1$ , so $m\sim h^{(d-y_h)/y_h}$ , i.e.

\delta=\frac{y_h}{d-y_h}.

Why choosing $b$ is legitimate: the homogeneity holds for every $b>1$ , so any convenient value may be inserted; different choices expose different singularities of the same $f_s$ . *Verified: for the 2D Ising eigenvalues $(y_t,y_h,d)=(1,\tfrac{15}{8},2)$ these formulas give $\alpha=0,\beta=\tfrac18,\gamma=\tfrac74,\delta=15$ — the exact Onsager–Yang values — as computed in verify_critical-exponents-from-rg.py.*

Go deeper · Derivation B · 1 minThe scaling relations reappear as algebraic identities

Derivation BThe scaling relations reappear as algebraic identities

The four classical scaling laws — Rushbrooke, Widom, Fisher, and hyperscaling — were originally proposed as independent phenomenological hypotheses, each tested separately against data. Here they are not assumptions at all. Because every exponent in (6) is built from the same two numbers $y_t,y_h$ (and $d$ ), the relations are forced identities: substitute the formulas and the free parameters cancel. We check all four by direct substitution. This article does not re-derive the scaling relations as independent results — that is owned by the Scaling Hypothesis and Scaling Relations article — it reproduces them from the RG eigenvalues, which is the sharper statement of why they must hold.

Step 1. Rushbrooke. Add the numerators of $\alpha,2\beta,\gamma$ over the common $1/y_t$ :

\alpha+2\beta+\gamma=\Big(2-\tfrac{d}{y_t}\Big)+\tfrac{2(d-y_h)}{y_t}+\tfrac{2y_h-d}{y_t}.

The $y_t$ -denominated terms sum to $-d+2(d-y_h)+(2y_h-d)=0$ , leaving $\alpha+2\beta+\gamma=2$ .

Step 2. Widom.

\beta(\delta-1)=\frac{d-y_h}{y_t}\left(\frac{y_h}{d-y_h}-1\right)=\frac{d-y_h}{y_t}\cdot\frac{2y_h-d}{d-y_h}=\frac{2y_h-d}{y_t}=\gamma.

Step 3. Fisher. Using $\eta=d+2-2y_h$ and $\nu=1/y_t$ ,

\nu(2-\eta)=\frac{1}{y_t}\big(2-(d+2-2y_h)\big)=\frac{2y_h-d}{y_t}=\gamma.

Step 4. Hyperscaling (Josephson). Immediate from $\alpha=2-d/y_t$ and $\nu=1/y_t$ :

d\nu=\frac{d}{y_t}=2-\alpha.

Why hyperscaling can fail: it assumes the singular free energy is a function of relevant fields only. Above $d_c=4$ a dangerous irrelevant variable must be retained, and $d\nu=2-\alpha$ breaks down while the other three relations survive. *Verified: all four relations reduce to $0=0$ symbolically in $y_t,y_h,d$ , and hold numerically for the 2D-Ising set — see verify_critical-exponents-from-rg.py.*

Go deeper · Derivation C · 1 minThe anomalous dimension from the magnetic eigenvalue

Derivation CThe anomalous dimension from the magnetic eigenvalue

The exponent $\eta$ is different in character from the thermodynamic ones: it lives in the two-point correlation function at criticality, $G(r)\sim r^{-(d-2+\eta)}$ , and measures the anomalous scaling dimension of the order-parameter field. We connect it to $y_h$ by demanding that the coupling of the field $h$ to the order parameter be RG-invariant, which ties $y_h$ to the field's scaling dimension, which in turn sets the power law of $G(r)$ . Three short steps close the loop.

Step 1. Under one RG step the coarse-grained order-parameter field rescales as $\phi'(x')=b^{x_\phi}\phi(x)$ with $x'=x/b$ , where $x_\phi$ is its scaling dimension.

Step 2. The Zeeman term $h\int\mathrm{d}^dx\,\phi$ must be invariant, which forces

y_h=d-x_\phi.

Step 3. The two-point function scales with twice the field dimension,

G(r)=\langle\phi(r)\phi(0)\rangle\sim r^{-2x_\phi}=r^{-(d-2+\eta)},

so $2x_\phi=d-2+\eta$ and therefore

\eta=2x_\phi-d+2=d+2-2y_h.

Verified: for 2D Ising $y_h=\tfrac{15}{8}$ gives $\eta=2+2-2(\tfrac{15}{8})=\tfrac14$ , the exact result, and the identity $\eta=d+2-2y_h\Leftrightarrow x_\phi=(d-2+\eta)/2=d-y_h$ is confirmed symbolically in the verify script.

Limiting cases

How the eigenvalue construction specializes across fixed points and dimensions:

Case	Result	Meaning
Gaussian fixed point, $d>4$	$y_t=2,\;y_h=1+d/2,\;\nu=\tfrac12$	Mean-field exponents; hyperscaling $d\nu=2-\alpha$ fails because a dangerous irrelevant variable must be kept.
Wilson–Fisher, $d=4-\epsilon$	$y_t=2-\tfrac{\epsilon}{3}+O(\epsilon^2),\;y_h=3-\tfrac{\epsilon}{2}+O(\epsilon^2)$	Non-Gaussian fixed point; non-classical exponents, $\nu=\tfrac12+\tfrac{\epsilon}{12}+\cdots$ for Ising ( $n=1$ ).
Two-dimensional Ising (exact)	$y_t=1,\;y_h=\tfrac{15}{8}$	Reproduces $\nu=1,\beta=\tfrac18,\gamma=\tfrac74,\delta=15,\eta=\tfrac14$ — the Onsager–Yang benchmark.
Marginal direction, $y_i=0$	$\lambda_i=1\Rightarrow$ logs	At $d=d_c=4$ the quartic coupling is marginally irrelevant, replacing power laws by logarithmic corrections.

Worked examples

Example 1 — Recover the exact 2D-Ising exponents from two eigenvalues. The two-dimensional Ising fixed point has $y_t=1$ and $y_h=\tfrac{15}{8}$ in $d=2$ . Compute all six exponents and confirm they satisfy the scaling relations.

Step 1. $\nu=1/y_t=\mathbf{1}$ and $\alpha=2-d/y_t=2-2=\mathbf{0}$ .

Step 2. $\beta=(d-y_h)/y_t=(2-\tfrac{15}{8})/1=\mathbf{1/8}$ — the celebrated Ising value that mean-field theory ( $\beta=\tfrac12$ ) gets badly wrong.

Step 3. $\gamma=(2y_h-d)/y_t=(\tfrac{15}{4}-2)/1=\mathbf{7/4}$ and $\delta=y_h/(d-y_h)=\tfrac{15}{8}/(2-\tfrac{15}{8})=\tfrac{15}{8}/\tfrac18=\mathbf{15}$ .

Step 4. $\eta=d+2-2y_h=4-\tfrac{15}{4}=\mathbf{1/4}$ . Check Rushbrooke: $\alpha+2\beta+\gamma=0+\tfrac14+\tfrac74=2$ . ✓ These are exactly Onsager's and Yang's rigorous results, reproduced from two numbers.

Example 2 — How much bigger is the susceptibility at half the distance from $T_c$ ? For the 2D Ising class, $\chi\sim|t|^{-\gamma}$ with $\gamma=\tfrac74$ . Compare $\chi$ at $t=0.01$ with $\chi$ at $t=0.02$ .

Step 1. The ratio is $\chi(0.01)/\chi(0.02)=(0.01/0.02)^{-\gamma}=(1/2)^{-7/4}=2^{7/4}$ .

Step 2. Numerically $2^{7/4}=2^{1.75}\approx\mathbf{3.36}$ . Halving the distance to the critical point raises the susceptibility by a factor of about $3.4$ — the steep divergence that makes critical systems so sensitive. Both examples are recomputed in verify_critical-exponents-from-rg.py: the full 2D-Ising set $(\nu,\alpha,\beta,\gamma,\delta,\eta)=(1,0,\tfrac18,\tfrac74,15,\tfrac14)$ with all four scaling relations, and the ratio $2^{7/4}=3.3636$ .

Origins · 1 minwilson-fisher fixed point and how this was discovered

wilson-fisher fixed point

discovered 1972 · the non-Gaussian fixed point that fixed critical exponents in

d<4

The Wilson–Fisher fixed point, found by Kenneth Wilson and Michael Fisher in 1972, is where this whole story acquires quantitative teeth. In more than four dimensions the Gaussian (free-field) fixed point already gives the right (mean-field) exponents, but below four it becomes unstable and the flow runs off to a new, genuinely interacting fixed point. Wilson and Fisher realized that if one treats $\epsilon=4-d$ as a small parameter, this fixed point sits at a coupling of order $\epsilon$ , and its eigenvalues $y_t,y_h$ — hence all critical exponents — can be computed as controlled power series in $\epsilon$ . Setting $\epsilon=1$ to reach three dimensions is uncontrolled but startlingly accurate, and the construction turned universality from a qualitative slogan into a calculational scheme.

History

1944

— Onsager solves the 2D Ising model exactly, later yielding

y_t=1,\;y_h=\tfrac{15}{8}

1952

— Yang derives the 2D Ising spontaneous magnetization, fixing

\beta=\tfrac18

1965

— Widom proposes the homogeneity of the singular free energy phenomenologically

1966

— Kadanoff's block-spin argument links homogeneity to coarse-graining

1971

— Wilson formulates the renormalization group; exponents become eigenvalues

1972

— Wilson and Fisher introduce the

\epsilon=4-d

expansion and the Wilson–Fisher fixed point

1982

— Wilson receives the Nobel Prize for the RG theory of critical phenomena

Going further

References. Kardar, Statistical Physics of Fields, Ch. 4–5 · Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 14 · Goldenfeld, Lectures on Phase Transitions and the RG, Ch. 9 · Cardy, Scaling and Renormalization in Statistical Physics, Ch. 3 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 12 · Wilson & Kogut, Phys. Rep. 12 (1974) 75.

Provenance. Eigenvalue/exponent relations are standard: verify the $f_s$ homogeneity and the $\alpha,\beta,\gamma,\delta,\nu,\eta$ formulas against Kardar (Ch. 4) and Cardy (Ch. 3). The Gaussian values $y_t=2,\;y_h=1+d/2$ and the Wilson–Fisher $\epsilon$ -expansion $y_t=2-\frac{n+2}{n+8}\epsilon$ are in Wilson–Kogut (1974) and Pathria & Beale Ch. 14. The exact 2D Ising eigenvalues $y_t=1,\;y_h=\tfrac{15}{8}$ and the derived $\alpha=0,\beta=\tfrac18,\gamma=\tfrac74,\delta=15,\eta=\tfrac14$ are Onsager (1944) and Yang (1952) values reproduced in all listed texts and here recalled by name. The scaling-relation algebra (Rushbrooke, Widom, Fisher, hyperscaling) is recomputed by direct substitution.

Builds on: Linearization About a Fixed Point and Eigenvalue Spectrum · Fixed Points of the RG · Relevant, Irrelevant, and Marginal Operators · Scaling Fields and Scaling Variables · RG Flow in Coupling Space. Leads to: Wilson-Fisher Fixed Point · The Epsilon-Expansion · Gaussian Fixed Point and Upper Critical Dimension.

Why only two independent exponents? Where did all the others go?

An ordinary critical point has exactly two relevant scaling fields, thermal ( $y_t$ ) and magnetic ( $y_h$ ). Every thermodynamic and correlation exponent is a formula in just $y_t$ , $y_h$ , and $d$ , so once you know two of them (say $\nu$ and $\gamma$ ) the rest are fixed. That is the content of the scaling relations: six exponents, but only two free numbers. If a system has more relevant fields — a multicritical point — then there are correspondingly more independent exponents.

t

really the eigenvector, or just approximately the temperature?

At the linear level the relevant thermal scaling field $g_t$ is a particular combination of all the couplings that happens to reduce to the reduced temperature $t$ to leading order. Away from the fixed point, nonlinear scaling fields mix couplings, so $g_t$ can differ from $t$ by analytic corrections. Usually this only shifts non-universal amplitudes and $T_c$ , but near marginal operators the mixing can generate the dominant singular behaviour, which is why careful treatments distinguish the physical field $t$ from the true eigen-direction $g_t$ .

How does the

\epsilon

-expansion turn eigenvalues into real 3D numbers?

Below four dimensions the flow lands on the Wilson–Fisher fixed point, whose location and eigenvalues can be expanded in $\epsilon=4-d$ . For the $O(n)$ model, $y_t=2-\frac{n+2}{n+8}\epsilon+O(\epsilon^2)$ , giving $\nu=\tfrac12+\frac{n+2}{4(n+8)}\epsilon+\cdots$ . Setting $\epsilon=1$ to reach $d=3$ is formally uncontrolled, but for Ising ( $n=1$ ) it already gives $\nu\approx0.58$ at first order, improving toward the true $0.63$ when higher orders are resummed. The article quotes only the leading term; the full expansion and its resummation are developed in the dedicated Wilson–Fisher and $\epsilon$ -expansion articles.

What is a "dangerous irrelevant variable" and why does it break hyperscaling?

Above the upper critical dimension the quartic coupling $u$ is irrelevant, so naively it should drop out of the singular free energy. But some observables depend on $u$ in a way that becomes singular as $u\to0$ — the free energy's dependence is not analytic in the limit. Such a variable is "dangerously irrelevant": it cannot be set to zero without changing the leading behaviour. Its retention modifies the naive scaling of $f_s$ , which is exactly why hyperscaling $d\nu=2-\alpha$ fails for $d>4$ even though the other scaling relations survive.

Does

\eta

ever vanish, and what does

\eta=0

mean physically?

Yes — $\eta=0$ at any Gaussian (free-field) fixed point, so mean-field theory predicts $\eta=0$ and correlations decaying as the pure $r^{-(d-2)}$ Ornstein–Zernike form. A nonzero $\eta$ signals an anomalous dimension: interactions have shifted the order-parameter field's scaling dimension away from its free value, and the correlation function decays with a corrected power. In 2D Ising $\eta=\tfrac14$ is large; in 3D Ising it is small ( $\eta\approx0.036$ ); above $d=4$ it is exactly zero. A measured nonzero $\eta$ is direct evidence that the fixed point is non-Gaussian.

Why is finite-size scaling a measurement of the same eigenvalues?

Because a finite box of side $L$ enters the homogeneity relation exactly like an inverse length: choosing $b=L$ gives $f_s(t,h,L^{-1})=L^{-d}\Phi(L^{y_t}t,L^{y_h}h)$ . So the way quantities depend on $L$ at fixed $t$ is governed by the very same $y_t,y_h$ that govern their dependence on $t$ in the bulk. Plotting rescaled data ( $L^{\gamma/\nu}\chi$ versus $L^{1/\nu}t$ ) collapses curves for different $L$ onto one master curve, and the collapse only works for the correct eigenvalues — making finite-size scaling a clean, direct way to extract $y_t=1/\nu$ and $y_h$ from simulation.

Mean-field theory gets

\beta=\tfrac12

but the true 2D value is

\tfrac18

. Is mean-field simply wrong?

Mean-field theory is not wrong so much as out of its regime. It corresponds to the Gaussian fixed point, which correctly governs criticality for $d>4$ , where its exponents ( $\beta=\tfrac12,\gamma=1,\delta=3$ ) are exact. Below four dimensions fluctuations dominate, the flow leaves the Gaussian point for the Wilson–Fisher point, and the exponents change — dramatically in low dimensions, giving $\beta=\tfrac18$ in 2D. The RG explains both the mean-field answer and its failure, and pinpoints $d_c=4$ as the boundary between the two.