Universality and Universality Classes

Here is a fact that should be surprising. A fluid near its liquid–gas critical point, a uniaxial ferromagnet near its Curie temperature, and a binary alloy near its order–disorder transition are utterly different at the microscopic level — different constituents, different forces, different lattices. Yet near their critical points they share the same critical exponents: $\beta\approx0.326$ , $\gamma\approx1.237$ , $\nu\approx0.630$ . This insensitivity to microscopic detail is called universality, and the collection of systems sharing one set of exponents is a universality class. The phenomenological side of this story — which systems fall into which class, and the scaling relations their exponents obey — is developed under the heading of universality classes as an empirical scaling hypothesis. This article supplies the explanation: why universality is true at all, delivered by the renormalization group. The mechanism is the geometry of RG flow. Under repeated coarse-graining and rescaling, a microscopic Hamiltonian traces a trajectory through the infinite-dimensional space of all possible couplings. Systems tuned to criticality have a diverging correlation length, and their trajectories are drawn toward a fixed point of the RG transformation. Every Hamiltonian on the critical surface flowing into the same fixed point ends up with identical long-distance physics. The microscopic differences between two such systems are directions along which the flow contracts — irrelevant scaling fields — so they vanish in the scaling limit. What survives is controlled by the fixed point alone. The data that pick out which fixed point are few and global: the spatial dimensionality $d$ , the symmetry of the order parameter (its number of components $n$ and breaking pattern), and the range of interactions. Everything else — lattice type, the precise short-range couplings, further-neighbor terms — is irrelevant in the technical RG sense and drops out of the exponents. Universality is sharp but bounded: leading exponents and certain dimensionless amplitude ratios are universal; absolute amplitudes, the nonuniversal $T_c$ , and the metric factors relating physical to scaling variables are not. Understanding exactly what is and is not universal is the core conceptual payoff of the RG view of critical phenomena.

Key takeaways

Universality is the empirical and theoretical fact that wildly different systems share identical critical exponents, amplitude ratios, and scaling functions near a continuous transition; the renormalization group explains why.
Systems whose Hamiltonians flow under coarse-graining to the same RG fixed point have the same long-distance critical behavior, regardless of microscopic detail — this is the RG statement of universality.
A universality class is fixed by only a few global data: spatial dimension $d$ , the symmetry (number of order-parameter components $n$ and its breaking pattern), and the interaction range; everything else is an irrelevant perturbation.
Irrelevant couplings die under the flow, washing out lattice structure and particular interaction strengths; only the relevant scaling fields survive to control the singularities.
The same exponents describe the liquid–gas critical point, uniaxial magnets, and binary alloys (all $d=3$ , $n=1$ Ising), while $d=3$ XY ( $n=2$ ) and Heisenberg ( $n=3$ ) form distinct classes.
Key idea: critical behavior is controlled by an RG fixed point; every system whose flow approaches that fixed point inherits identical exponents and scaling functions, because the microscopic differences between them are irrelevant scaling fields that contract to zero under coarse-graining — so a class is labeled only by dimension, order-parameter symmetry, and interaction range.

The core insight

Critical behavior is controlled by an RG fixed point; every system whose flow approaches that fixed point inherits identical exponents and scaling functions, because the microscopic differences between them are irrelevant scaling fields that contract to zero under coarse-graining. Universality classes are labeled by the fixed point, which depends only on dimensionality, order-parameter symmetry/components, and interaction range.

Where this lives today

Universality is the license behind the entire practice of computing critical exponents in idealized models. A Monte-Carlo study of the nearest-neighbor Ising model on a simple cubic lattice — a caricature of any real magnet — yields exponents that match measured values in chemically unrelated materials, precisely because the lattice model and the real material share $(d,n,\text{symmetry})$ and flow to the same fixed point. The numerical conformal bootstrap now determines the 3D Ising class exponents to five or six significant figures by treating the fixed point as a conformal field theory, and those numbers are used as reference values across condensed-matter and statistical physics. The concept extends far beyond thermal criticality. Quantum phase transitions at zero temperature fall into universality classes fixed by dimension, symmetry, and dynamical exponent; non-equilibrium steady states (directed percolation, surface growth) have their own classes; and the same logic organizes the critical behavior of systems as varied as the deconfinement transition in QCD and the onset of turbulence. Wherever the phrase "belongs to the … universality class" is used, the underlying claim is the one this article explains: two systems flow to the same RG fixed point, so their long-distance physics is identical up to nonuniversal rescalings.

The physical picture

Imagine viewing a critical system through a sequence of progressively blurrier lenses — that is coarse-graining. At criticality the correlation length is infinite, so the system looks statistically identical at every magnification: it is scale-invariant. Two different microscopic models, viewed through enough blur, become indistinguishable, because their fine-grained differences drop below resolution. That converged, magnification-independent image is the fixed point, and any two systems whose blurred images converge to it belong to the same universality class. The flow toward a fixed point has stable directions (irrelevant) and unstable ones (relevant). The critical surface is the basin within which every relevant field has been tuned to zero; on it, the flow runs purely along irrelevant directions into the fixed point. All the microscopic complexity a real material carries — every extra coupling it has — corresponds to displacements along irrelevant directions, and these shrink to nothing under iteration. Only the handful of relevant directions, temperature and field, remain, and they are shared by every member of the class. Why do only $d$ , $n$ , and symmetry matter? The relevant/irrelevant split is set by the scaling dimensions of operators, which depend on dimensionality. Which fixed point exists at all is dictated by the symmetry the Hamiltonian must respect: an $n$ -component order parameter with $O(n)$ symmetry admits a specific fixed-point structure. Change $n$ , or break the symmetry differently, and you land on a different fixed point — hence a different class. A useful slogan: universality is forgetting. The RG systematically forgets everything about a system except its symmetry, dimension, and interaction range; what survives the forgetting is exactly what is universal.

Figure

Many microscopic Hamiltonians, one fixed point: the geometry of universality

Figure

Different classes, different exponents: same-class systems agree, the 2D-Ising outlier stands apart

The core equations

The RG transformation maps the set of all couplings to a new set after coarse-graining and rescaling; a fixed point is a coupling set that maps to itself. This is the object that defines a universality class.

\mu_a' = \mathcal{R}_b(\mu_a)\,,\qquad \mu^* = \mathcal{R}_b(\mu^*) \qquad (1)

where: $\mu_a$ — the full set of couplings (a point in coupling space); $\mathcal{R}_b$ — RG transformation with rescaling factor $b>1$ (coarse-grain, then rescale); $\mu^*$ — a fixed point, invariant under $\mathcal{R}_b$ . At a fixed point the system reproduces itself under coarse-graining, so the correlation length obeys $\xi=\xi/b$ , forcing $\xi=0$ or $\xi=\infty$ ; the $\xi=\infty$ fixed points are the critical ones. Linearizing the map near the fixed point defines the scaling fields — the eigen-directions that transform multiplicatively — whose eigenvalue exponents are the universal data.

\delta\mu_a' = \sum_b M_{ab}\,\delta\mu_b\,,\qquad g_i' = b^{y_i}\,g_i \qquad (2)

where: $\delta\mu_a$ — deviation from the fixed point; $M_{ab}=(\partial\mathcal{R}_a/\partial\mu_b)|_{\mu^*}$ — linearization matrix; $g_i$ — scaling field along eigen-direction $i$ ; $y_i$ — RG eigenvalue exponent. A field is relevant ( $y_i>0$ , grows), irrelevant ( $y_i<0$ , dies), or marginal ( $y_i=0$ ); the exponents $y_i$ are properties of the fixed point, so they are shared by every system that flows there — the origin of universality. For an Ising-like transition there are exactly two relevant scaling fields, and the singular free energy is a homogeneous function of them once the irrelevant fields have flowed away.

f_s(g_t,g_h,\{g_i\}) = b^{-d}\,f_s\!\big(b^{y_t}g_t,\,b^{y_h}g_h,\,\{b^{y_i}g_i\}\big) \qquad (3)

where: $f_s$ — singular part of the free-energy density; $g_t\propto t$ — thermal scaling field; $g_h$ — magnetic (ordering-field) scaling field; $\{g_i\}$ — irrelevant scaling fields ( $y_i<0$ ). Iterating until $b^{y_t}|g_t|=1$ collapses this to a two-variable scaling form because, with $y_i<0$ , the irrelevant arguments $g_i|g_t|^{-y_i/y_t}\to0$ as $g_t\to0$ . Every static critical exponent is a function of the two relevant eigenvalues and the dimension. Only two numbers are independent.

2-\alpha=\frac{d}{y_t},\;\; \nu=\frac{1}{y_t},\;\; \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 \qquad (4)

where: $y_t,y_h$ — the two relevant eigenvalues; $d$ — dimension. Reach for these when you have the eigenvalues and want the exponents, or in reverse when you have measured exponents and want to check they come from a single fixed point with two relevant fields. Because all exponents derive from the same two eigenvalues, the scaling and hyperscaling relations hold automatically as identities within every class.

\alpha+2\beta+\gamma=2\,,\quad \gamma=\beta(\delta-1)\,,\quad \nu d=2-\alpha\,,\quad \gamma=\nu(2-\eta) \qquad (5)

where: $\alpha,\beta,\gamma,\delta,\nu,\eta$ — the six standard critical exponents. These are Rushbrooke, Widom, Josephson (hyperscaling), and Fisher; they are consequences of the eigenvalue formulas, not independent inputs. Hyperscaling $\nu d=2-\alpha$ fails above the upper critical dimension, where the dangerous irrelevant variable intervenes. Universality is layered: exponents are universal, and so are the scaling function and dimensionless amplitude ratios — but absolute amplitudes and $T_c$ are not.

\frac{\Gamma_+}{\Gamma_-}\,,\quad \frac{A_+}{A_-}\,,\quad \frac{\xi_0^+}{\xi_0^-}\quad\text{universal};\qquad \Gamma_\pm,\;T_c,\;c_t,\;c_h\quad\text{nonuniversal} \qquad (6)

where: $\Gamma_\pm$ — susceptibility amplitudes above/below $T_c$ ; $A_\pm$ — specific-heat amplitudes; $\xi_0^\pm$ — correlation-length amplitudes; $c_t,c_h$ — nonuniversal metric factors relating physical variables to scaling fields. Each amplitude factorizes into a universal fixed-point number times a system-dependent metric factor; in a $+/-$ ratio the metric factor cancels, leaving a pure fixed-point number — a more stringent test of universality than exponents alone.

Derivation AFrom a shared fixed point to identical critical exponents

The goal is to show rigorously that two distinct microscopic Hamiltonians flowing to the same fixed point must have identical exponents, with their microscopic differences surviving only as irrelevant scaling fields that decay away. Take two systems with bare couplings $\mu^{(A)}$ and $\mu^{(B)}$ , both tuned to criticality so each lies on the critical surface $S$ of a fixed point $\mu^*$ . The strategy is to decompose each Hamiltonian into scaling fields near the fixed point, show the difference between the two trajectories contracts under iteration, and observe that the exponents are computed from the linearization matrix at the common fixed point — data that belongs to $\mu^*$ alone.

Step 1. Decompose each Hamiltonian into scaling fields: $\mu=\mu^*+\sum_i g_i\,e_i$ . For systems on the critical surface, all relevant scaling fields are tuned to zero ( $g_t=g_h=0$ at exact criticality), so the only nonzero displacements are along irrelevant eigen-directions. Why: being on the critical surface means precisely that the relevant knobs have been set to their critical values, leaving only irrelevant differences between the two systems.

Step 2. Show the difference contracts: under $n$ iterations, $g_i^{(A)}(n)-g_i^{(B)}(n)=b^{n y_i}\big(g_i^{(A)}-g_i^{(B)}\big)\to0$ because every remaining direction is irrelevant, $y_i<0$ . Why: an irrelevant eigenvalue multiplies the displacement by $b^{y_i}<1$ each step, so the two trajectories approach each other geometrically and both sit arbitrarily close to $\mu^*$ after enough steps.

Step 3. Locate the exponents in fixed-point data: the linearization $M_{ab}=(\partial\mathcal{R}_a/\partial\mu_b)|_{\mu^*}$ is evaluated at the common fixed point, so its eigenvalues $b^{y_i}$ — and every exponent through $2-\alpha=d/y_t$ , $\nu=1/y_t$ , and the rest — depend only on $\mu^*$ . Why: the derivative that defines the eigenvalue exponents is taken at $\mu^*$ , not along either trajectory, so it cannot know which system arrived there.

Step 4. Conclude: systems $A$ and $B$ , however different microscopically, share $y_t$ and $y_h$ and therefore identical $\alpha,\beta,\gamma,\delta,\nu,\eta$ ; their microscopic differences survived only as irrelevant fields that decayed. Why: this is universality — the assertion that the class, not the microscopic model, determines the exponents, made into a theorem about RG flow.

Go deeper · Derivation B · 1 minUniversal scaling function and amplitude ratios

Derivation BUniversal scaling function and amplitude ratios

The goal is to derive how the scaling function is universal only up to nonuniversal metric factors, explaining precisely why amplitude ratios — but not the amplitudes themselves — are universal. Start from the homogeneous free energy with irrelevant fields already set to zero, and track how the system-dependent relation between physical variables ( $t,h$ ) and scaling fields ( $g_t,g_h$ ) enters the observables. The plan is to insert the linear metric relations, differentiate to obtain a susceptibility amplitude, and then take a $+/-$ ratio to cancel the metric factors.

Step 1. Write the homogeneous free energy $f_s(t,h)=|t|^{2-\alpha}\Psi_\pm(h/|t|^{\Delta})$ with $\Delta=\beta\delta=y_h/y_t$ , and relate physical to scaling variables by nonuniversal metric factors $g_t=c_t\,t[1+O(t)]$ , $g_h=c_h\,h[1+O(h)]$ . Why: the scaling function $\Psi_\pm$ written in $g$ -variables is universal, but rewriting it in physical $t,h$ absorbs the system-dependent constants $c_t,c_h$ , which are analytic (nonsingular) at the transition.

Step 2. Differentiate twice to get the susceptibility amplitude: $\chi=-\partial^2 f_s/\partial h^2=\Gamma_\pm|t|^{-\gamma}$ with $\Gamma_\pm=c_h^2\,c_t^{\gamma}\,\psi_\pm$ , where $\psi_\pm=\Psi_\pm''(0)$ is a universal number. Why: the second field-derivative brings down $c_h^2$ and the temperature prefactor brings $c_t^{\gamma}$ , so the amplitude $\Gamma_\pm$ is a universal fixed-point number multiplied by nonuniversal metric powers — hence $\Gamma_\pm$ itself is not universal.

Step 3. Take the above/below- $T_c$ ratio: $\Gamma_+/\Gamma_-=\psi_+/\psi_-$ , a pure fixed-point number. Why: both amplitudes carry the identical prefactor $c_h^2 c_t^{\gamma}$ , so it cancels in the ratio, leaving only the universal ratio of scaling-function values — the same logic gives universal $A_+/A_-$ and $\xi_0^+/\xi_0^-$ .

Step 4. Read the layered structure: exponents (from eigenvalues) are universal; the scaling function and amplitude ratios (from the fixed-point scaling function) are universal; absolute amplitudes, $T_c$ , and metric factors are not. Why: two systems in the same class agree exactly on the first two layers and can differ freely on the third — which is exactly what experiments find, e.g. the 3D Ising ratio $\Gamma_+/\Gamma_-\approx4.76$ shared across fluids and magnets.

Limiting cases

Universality classes are organized primarily by $(d,n)$ and symmetry; the standard classes and their controlling features (exponent values recalled from the critical-exponents and 2D-Ising owners):

Case	Result	Meaning
Ising ( $n=1$ , $d=3$ )	$\beta\approx0.326$ , $\gamma\approx1.237$ , $\nu\approx0.630$ , $\eta\approx0.036$	Liquid–gas critical point, uniaxial ferromagnets, binary alloys; discrete $\mathbb{Z}_2$ symmetry, scalar order parameter.
XY ( $n=2$ , $d=3$ )	$\alpha\approx-0.0151$ , $\nu\approx0.6717$ , $\gamma\approx1.3178$	Superfluid $^4$ He $\lambda$ -transition, easy-plane magnets; $O(2)$ symmetry; the slightly negative $\alpha$ is a precision test.
Heisenberg ( $n=3$ , $d=3$ )	$\nu\approx0.7112$ , $\gamma\approx1.396$ , $\eta\approx0.0375$	Isotropic ferromagnets; $O(3)$ symmetry, three-component vector order parameter.
Mean-field / Gaussian ( $d>4$ )	$\beta=\tfrac12$ , $\gamma=1$ , $\nu=\tfrac12$ , $\delta=3$ , $\eta=0$	Above the upper critical dimension the Gaussian fixed point governs; exponents are $n$ -independent Landau values (also exact for infinite-range models).
2D Ising (Onsager, exact)	$\beta=\tfrac18$ , $\gamma=\tfrac74$ , $\nu=1$ , $\eta=\tfrac14$ , $\alpha=0\,(\log)$	Exactly solved; the lower-dimensional Ising class, distinct from $d=3$ — demonstrates the strong $d$ -dependence of exponents at fixed symmetry.

Worked examples

Example 1 — The 2D-Ising class from its eigenvalues, and its hyperscaling consistency. This computes an exact universality class from two numbers and checks it closes.

Step 1. Take the two relevant eigenvalues of the 2D-Ising fixed point, $y_t=1$ and $y_h=15/8$ , in dimension $d=2$ .

Step 2. Compute the exponents from the eigenvalue formulas: $\nu=1/y_t=1$ ; $\alpha=2-d/y_t=2-2=0$ ; $\beta=(d-y_h)/y_t=(2-\tfrac{15}{8})=\tfrac18$ ; $\gamma=(2y_h-d)/y_t=(\tfrac{30}{8}-2)=\tfrac74$ ; $\delta=y_h/(d-y_h)=\tfrac{15/8}{1/8}=15$ ; $\eta=d+2-2y_h=4-\tfrac{30}{8}=\tfrac14$ .

Step 3. Confirm these are exactly Onsager's 2D-Ising exponents $(\alpha,\beta,\gamma,\delta,\nu,\eta)=(0,\tfrac18,\tfrac74,15,1,\tfrac14)$ , as recalled from the exact 2D-Ising solution.

Step 4. Test Rushbrooke: $\alpha+2\beta+\gamma=0+\tfrac14+\tfrac74=2$ . The relation holds with equality, confirming the set is a hyperscaling-consistent class. The 2D-Ising class is generated entirely by $y_t=1$ , $y_h=15/8$ , and satisfies $\alpha+2\beta+\gamma=2$ exactly.

Example 2 — Why $d=3$ splits into distinct classes. This shows that changing only the order-parameter symmetry moves you to a different class.

Step 1. Fix $d=3$ and compare $n=1$ (Ising) with $n=3$ (Heisenberg): both have continuous transitions but different order-parameter components.

Step 2. Read the exponents: 3D Ising has $\nu\approx0.630$ , while 3D Heisenberg has $\nu\approx0.711$ — a difference of about 13%, far outside any measurement error.

Step 3. Attribute the difference to the fixed point: the $O(n)$ symmetry enters the fixed-point equations through the combinatorial factor $n+8$ , so the eigenvalues, and hence $\nu$ , depend on $n$ .

Step 4. Conclude that dimensionality alone does not fix a class: a uniaxial magnet ( $n=1$ ) and an isotropic ferromagnet ( $n=3$ ), both in $d=3$ , sit in different universality classes with different exponents ( $\nu\approx0.630$ vs $\approx0.711$ ).

Origins · 1 minuniversality class and how this was discovered

universality class

the RG label of a critical fixed point · A universality class is the set of all systems whose Hamiltonians flow under the renormalization group to the same critical fixed point, and therefore share identical critical exponents, scaling functions, and dimensionless amplitude ratios. The concept crystallized in the 1960s from the empirical discovery that fluids, magnets, and alloys share exponents, and it acquired its explanation with Wilson's renormalization group: a class is labeled solely by the spatial dimension

d

, the order-parameter symmetry (its number of components

n

and breaking pattern), and the interaction range — every microscopic difference being an irrelevant scaling field. It is the organizing principle that lets an idealized lattice model predict the critical behavior of a real material.

History

1900

— Early measurements hint that different fluids share a "law of corresponding states" near criticality

1944

— Onsager's exact 2D Ising solution proves exponents can be nonclassical and strongly dimension-dependent

1965

— Widom's scaling hypothesis posits the singular free energy is a generalized homogeneous function

1966

— Kadanoff's block-spin picture links scaling to coarse-graining and anticipates fixed points

1971

— Wilson's renormalization group explains universality as flow to a common fixed point

1972

— Wilson and Fisher compute class-dependent exponents, showing they depend only on d and n

2002

— Pelissetto and Vicari's review tabulates high-precision exponents and amplitude ratios for the standard classes

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 14 · Kardar, Statistical Physics of Fields, Ch. 4–5 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 12 · Goldenfeld, Lectures on Phase Transitions and the Renormalization Group · Cardy, Scaling and Renormalization in Statistical Physics · Pelissetto & Vicari, Critical Phenomena and Renormalization-Group Theory, Phys. Rep. 368 (2002). The epsilon-expansion gives the $n$ -dependence of the correlation-length exponent as $\nu=\tfrac12+\tfrac{n+2}{4(n+8)}\varepsilon$ (Wilson & Fisher). The lambda-point value $\alpha\approx-0.0151$ is from the microgravity Lambda-Point Experiment (Lipa et al.). Scaling-relation consistency ( $\alpha+2\beta+\gamma=2$ , etc.) and the exponent formulas in $y_t,y_h$ follow from the eigenvalue formalism developed above.

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

What is the difference between "universality" and a "universality class"?

Universality is the phenomenon: the observation, and its RG explanation, that critical behavior is insensitive to microscopic detail. A universality class is the bookkeeping unit: the specific set of systems that share one fixed point, hence one set of exponents, scaling functions, and amplitude ratios. In RG terms, universality is the fact that flow converges to fixed points; a universality class is the basin of attraction of a particular critical fixed point, labeled by $(d,n,\text{symmetry},\text{range})$ . The phenomenology of which systems land where, and the scaling relations, is developed by the universality-classes owner; this article supplies the RG mechanism behind it.

How can systems as different as a fluid and a magnet share exponents?

Because near criticality both are described, after coarse-graining, by the same effective field theory — a single scalar order parameter with $\mathbb{Z}_2$ symmetry — and both flow to the 3D Ising fixed point. The fluid's order parameter is the density difference $\rho-\rho_c$ and the magnet's is the magnetization, but each is a one-component field with the same symmetry, so their microscopic differences (molecular forces vs exchange interactions, continuum vs lattice) are irrelevant scaling fields. What controls the exponents is the fixed point they share, not the microscopic constituents. This fluid–magnet identity was the historical empirical signature of universality.

Why do only two exponents turn out to be independent?

Because for an ordinary critical point there are exactly two relevant scaling fields — temperature and ordering field — with eigenvalues $y_t$ and $y_h$ . Every static exponent is an explicit function of these two numbers and the dimension: $\nu=1/y_t$ , $\beta=(d-y_h)/y_t$ , and so on. So once you know two exponents you know them all, and the scaling relations (Rushbrooke, Widom, Fisher, Josephson) are automatic identities rather than extra assumptions. If a transition had three relevant fields (a tricritical point, say), there would be three independent exponents instead.

Does universality mean two systems in a class have the same susceptibility?

No — that is the most common misconception. They have the same exponent $\gamma$ , so their susceptibilities diverge as the same power of $|t|$ , and the same scaling function up to rescaling of its arguments. But the absolute susceptibility amplitude $\Gamma_\pm$ is nonuniversal: it carries system-dependent metric factors relating physical variables to the scaling fields. What is universal beyond the exponent is the dimensionless ratio $\Gamma_+/\Gamma_-$ , in which the metric factors cancel. So universality constrains shapes and ratios, not absolute magnitudes.

Why does hyperscaling fail above four dimensions but the other scaling relations survive?

Hyperscaling, $\nu d=2-\alpha$ , assumes the singular free energy scales purely as $\xi^{-d}$ with no extra dependence on irrelevant couplings. Above the upper critical dimension $d=4$ the quartic coupling is irrelevant but dangerous: the free energy is singular as it goes to zero, so keeping its dependence modifies the free-energy homogeneity and breaks hyperscaling — you get $\alpha=0$ rather than the naive $2-d/2$ . The other relations (Rushbrooke, Widom, Fisher) do not invoke the dimension explicitly and continue to hold with the mean-field values. So dimensionality-dependent relations are the ones at risk above $d_c$ .

How is a marginal operator different, and what does it do to universality?

A marginal operator has eigenvalue $y_i=0$ : to linear order it neither grows nor decays, so power counting cannot decide its fate — one must go to the next order in the flow. Marginal operators, as at the upper critical dimension, produce logarithmic rather than power-law corrections, and they can make two systems share exponents only up to logarithms. They also allow lines of fixed points (continuously varying exponents) in special cases, such as the eight-vertex model. So a marginal operator is a genuine exception to the clean relevant/irrelevant dichotomy and must be handled beyond leading order.