Scaling Fields and Scaling Variables

The renormalization group replaces a Hamiltonian by a coarse-grained one with rescaled couplings, generating a flow $\mathbf{K}\to\mathbf{K}'=R_b(\mathbf{K})$ in the infinite-dimensional space of coupling constants. Critical behaviour is controlled by fixed points $\mathbf{K}^*=R_b(\mathbf{K}^*)$ , where the physics looks the same at every magnification. But the physical knobs a laboratory turns — temperature, magnetic field, lattice anisotropy, further-neighbour couplings — are almost never the variables in which this flow is simplest. Turn on the temperature deviation $t$ and, after one coarse-graining step, you find not just a renormalized $t$ but a little admixture of every other operator the symmetry allows. Scaling fields are the special coordinates that untangle this. They are the eigen-directions of the linearized RG map near the fixed point, built so that each one transforms multiplicatively under coarse-graining: $u_i\to b^{y_i}u_i$ . In these variables the flow becomes a set of independent expansions and contractions, and the eigenvalue exponents $y_i$ sort every perturbation into three classes — relevant ( $y_i>0$ , grows), irrelevant ( $y_i<0$ , dies), or marginal ( $y_i=0$ ). For an Ising-like critical point exactly two scaling fields are relevant: a thermal one $u_t$ with eigenvalue $y_t=1/\nu$ and a magnetic one $u_h$ with eigenvalue $y_h$ . Everything else flows to zero. The payoff is that these two numbers fix all of thermodynamics. Written in scaling fields the singular free energy is a generalized homogeneous function, and choosing the rescaling factor to pin the leading relevant field to unity collapses it onto a one-variable scaling function. Out drop the exponents $\alpha,\beta,\gamma,\delta,\nu,\eta$ and the identities relating them. The distinction between a scaling field and the bare parameter it resembles is not pedantry: $u_t$ is an analytic but genuinely nonlinear function of $t$ , and it is exactly that nonlinearity that keeps measured exponents clean while pushing the messiness into non-universal amplitudes.

Key takeaways

Near an RG fixed point the flow is governed by its linearization; the eigenvectors of the linearized RG map define scaling fields $u_i$ , nonlinear combinations of the physical couplings that transform multiplicatively under coarse-graining: $u_i \to b^{y_i} u_i$ .
The eigenvalue exponents $y_i$ classify perturbations as relevant ( $y_i>0$ ), irrelevant ( $y_i<0$ ), or marginal ( $y_i=0$ ); only the relevant ones survive asymptotically, which is why so few numbers control criticality.
Multiplicative scaling is not assumed — it is forced by the semigroup law $R_{b_1}\circ R_{b_2}=R_{b_1 b_2}$ , whose only continuous solution is a power $b^{y}$ . Power-law singularities and scaling fields are the same fact.
Written in scaling fields the singular free energy obeys $f_s(\{u_i\})=b^{-d}f_s(\{b^{y_i}u_i\})$ ; the two relevant Ising fields $u_t$ (with $y_t=1/\nu$ ) and $u_h$ set every static exponent.
Scaling variables are the same objects viewed as curvilinear coordinates that linearize the flow; the nonlinear (Wegner) redefinition removes regular backgrounds and generates the correction-to-scaling exponent $\omega$ .
Key idea: Diagonalize the linearized RG near a fixed point; its eigen-directions are scaling fields $u_i$ that scale multiplicatively as $u_i\to b^{y_i}u_i$ , and in these variables the singular free energy is a generalized homogeneous function, $f_s=b^{-d}f_s(\{b^{y_i}u_i\})$ .

Where this lives today

Every modern extraction of a critical exponent — from neutron scattering on a magnet to Monte Carlo on a lattice to the conformal bootstrap for the 3D Ising model — is, underneath, a statement about scaling fields. When experimentalists fit a susceptibility to $\chi=\Gamma|t|^{-\gamma}(1+a|t|^{\theta}+\cdots)$ , the leading term uses the relevant thermal field and the correction term is precisely the leftover from the leading irrelevant field, with $\theta=\nu\omega$ . Getting the nonlinear part of $u_t$ right is what lets high-precision work quote exponents to four significant figures without them drifting as the fitting window changes. The language also drives condensed-matter theory far from magnets. In quantum criticality the tuning parameter is a coupling $g$ at zero temperature, and the same relevant/irrelevant bookkeeping tells you which microscopic terms can be dropped and which must be kept — the reason a handful of "field theories at a fixed point" describe superconductors, quantum Hall transitions, and heavy-fermion critical points that share nothing microscopically.

The physical picture

Picture the RG flow as a fluid streaming past a stagnation point — the fixed point. Right at the stagnation point the flow is arrested; a little away from it the velocity field is linear, and its eigen-directions are special streamlines. Along an unstable direction (a relevant field, $y_i>0$ ) you are swept away exponentially fast; along a stable one (irrelevant, $y_i<0$ ) you are drawn in. Scaling fields are precisely the coordinates aligned with these eigen-streamlines, in which the otherwise tangled flow separates into independent expansions and contractions. This has a startling consequence, and it is the whole reason universality works. Start from two utterly different microscopic Hamiltonians. Each is a different point in coupling space, with different admixtures of every irrelevant operator. But those irrelevant components all die under the flow, so both trajectories are funnelled onto the same low-dimensional relevant subspace and end up governed by the same handful of numbers. A generic starting point, projected onto the relevant axes, is all the fixed point ever remembers. The multiplicative law is forced, not chosen. Coarse-graining by $b_1$ and then by $b_2$ must equal coarse-graining by $b_1 b_2$ — the RG is a semigroup. If a quantity simply rescales, the only function consistent with $\lambda(b_1)\lambda(b_2)=\lambda(b_1 b_2)$ is a power law $\lambda(b)=b^{y}$ . So the existence of scaling fields and the existence of power-law critical singularities are one and the same fact wearing two hats. Why nonlinear? A bare coupling like the temperature distance $t$ picks up a small admixture of other operators after coarse-graining; the combination that scales cleanly is a Taylor series $u_t=t+c_1 t^2+c_2 h^2+\cdots$ . Keeping only the linear term is enough for the leading exponents, but the curvature terms are what generate the universal correction-to-scaling exponent $\omega=|y_{\text{irr}}|$ and the non-universal amplitude ratios — the difference between a textbook power law and real data.

Figure

Flow near the fixed point: relevant direction blows up, irrelevant collapses in

The figure above is the whole story of universality in one picture: any generic starting point (amber) is dragged down the irrelevant axis and flung out along the relevant one, so at large rescaling $b$ every trajectory looks the same — a pure motion along $u_t$ . That is why microscopic detail washes out.

Figure

Bare parameter versus scaling field: the nonlinear curved coordinate

The core equations

The renormalization group is a map on coupling space that must compose with itself: doing two coarse-grainings in a row is one bigger coarse-graining. This single algebraic fact — the semigroup property — is what makes everything downstream a power law. We record the map and its composition rule first.

\mathbf{K}' = R_b(\mathbf{K}),\qquad R_{b_1}\circ R_{b_2}=R_{b_1 b_2} \qquad (1)

where: $\mathbf{K}$ — a point in coupling space, the full set of Hamiltonian parameters; $R_b$ — one coarse-graining plus rescaling by linear factor $b>1$ ; $b$ — the spatial rescaling factor, often written $b=e^{\ell}$ for a continuous flow. Read the composition law as the definition of what "the renormalization group" means: the group of magnifications, closed under repetition. To learn the flow near a fixed point $\mathbf{K}^*$ we linearize. Writing the deviation $\delta\mathbf{K}=\mathbf{K}-\mathbf{K}^*$ , the map becomes a matrix multiplication by the Jacobian $\mathsf{M}(b)$ evaluated at the fixed point. This matrix carries all the local information about stability.

\delta\mathbf{K}' = \mathsf{M}(b)\,\delta\mathbf{K},\qquad \mathsf{M}_{ij}=\left.\frac{\partial K_i'}{\partial K_j}\right|_{\mathbf{K}^*} \qquad (2)

where: $\delta\mathbf{K}$ — deviation from the fixed point; $\mathsf{M}(b)$ — the stability matrix (Jacobian) of the RG map at $\mathbf{K}^*$ . Its eigenvalues decide which perturbations grow and which decay; its eigenvectors, projected onto $\delta\mathbf{K}$ , are the scaling fields. Diagonalizing $\mathsf{M}(b)$ and projecting the flow onto its left eigenvectors defines the scaling fields at linear order. The eigenvalue attached to each is forced by composition to be a pure power of $b$ , and that exponent $y_i$ is the object we care about.

u_i = \sum_j \phi^{(i)}_j\,\delta K_j + O(\delta K^2),\qquad u_i' = b^{y_i} u_i \qquad (3)

where: $\phi^{(i)}$ — the left eigenvector of $\mathsf{M}(b)$ with eigenvalue $b^{y_i}$ ; $u_i$ — the $i$ -th scaling field; $y_i$ — its RG eigenvalue exponent. A positive $y_i$ means the field is relevant (its magnitude grows under coarse-graining); negative means irrelevant; zero means marginal, the delicate case that produces logarithms. Linear scaling fields solve the leading problem, but the exact map has quadratic and higher terms. Wegner's construction promotes each $u_i$ to an analytic-but-nonlinear function of the couplings, chosen so that the multiplicative law holds to all orders — this is what makes exponents clean and pushes non-universal junk into amplitudes.

u_i(\mathbf{K}) = \delta K_i^{(\text{eig})} + \sum_{j,k} a_{jk}^{(i)}\,\delta K_j\,\delta K_k+\cdots,\qquad u_i\to b^{y_i}u_i\ \text{exactly} \qquad (4)

where: $a^{(i)}_{jk}$ — coefficients fixed by demanding exact multiplicative scaling beyond linear order. They can always be solved for unless a resonance $y_i=y_j+y_k$ occurs, in which case the smooth redefinition fails and logarithmic terms appear — exactly the marginal case at the upper critical dimension. With the scaling fields in hand, the singular part of the free-energy density inherits a generalized homogeneity: coarse-graining multiplies it by $b^{-d}$ (the volume per remaining degree of freedom shrinks) while driving each field by $b^{y_i}$ . This is the master relation everything else follows from.

f_s(u_1,u_2,\dots) = b^{-d}\,f_s\!\left(b^{y_1}u_1,\,b^{y_2}u_2,\dots\right) \qquad (5)

where: $f_s$ — the singular part of the free energy per unit volume (per spin), in units of $k_BT$ ; $d$ — the spatial dimension, entering because $f_s$ is a density. Because this holds for any $b$ , it is a functional constraint strong enough to fix all exponents. Since $b$ is arbitrary, we may choose it. Pinning the leading relevant thermal field to unity, $b^{y_t}|u_t|=1$ , eliminates $b$ and collapses $f_s$ onto a function of a single dimensionless ratio — the Widom scaling form, and the reason data collapse works.

f_s(u_t,u_h) = |u_t|^{d/y_t}\,\Phi\!\left(\frac{u_h}{|u_t|^{y_h/y_t}}\right) \qquad (6)

where: $\Phi$ — the universal scaling function (with two branches for $u_t\gtrless0$ ); $u_t,u_h$ — the thermal and magnetic scaling fields. The gap exponent $\Delta=y_h/y_t$ in the argument governs how the field scale tracks the temperature. Differentiating the collapsed form gives every static exponent as a rational function of just two eigenvalues and the dimension. This is the concrete sense in which two numbers control all of critical thermodynamics.

\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 \nu=\frac1{y_t},\quad \eta=d+2-2y_h \qquad (7)

where: $y_t,y_h$ — the thermal and magnetic RG eigenvalues; $\alpha\dots\eta$ — the standard critical exponents. Feed in $y_t=1,\,y_h=15/8,\,d=2$ and out come the exact 2D-Ising values; feed in the Gaussian $y_t=2,\,y_h=3,\,d=4$ and out come mean-field values. Finally, the leading irrelevant field does not simply vanish — it decays but leaves a trail. Keeping it produces the universal corrections to scaling that dominate any real fit at finite distance from $T_c$ .

f_s = |u_t|^{2-\alpha}\Big[\Phi_0\big(u_h/|u_t|^{\Delta}\big)+|u_t|^{\,\nu\omega}\Phi_1(\cdots)+\cdots\Big] \qquad (8)

where: $\omega=|y_3|$ — the magnitude of the leading irrelevant eigenvalue; $\theta=\nu\omega$ — the Wegner correction-to-scaling exponent; $\Phi_1$ — the subleading scaling function. Asymptotically negligible, these terms are indispensable for extracting exponents from data that never quite reaches $T_c$ .

Derivation AFrom the semigroup law to power-law eigenvalues and scaling fields

We prove rigorously that the linearized RG eigenvalues must be pure powers $b^{y_i}$ and construct the scaling fields as the matching eigen-coordinates. The engine is a single algebraic fact — the RG is a semigroup — and the goal is to show that this fact alone forces power-law scaling, so that the existence of well-defined critical exponents is not an assumption but a theorem. Start from the defining structure of the RG map, linearize about a fixed point, and let composition do the work.

Step 1. Expand the map about the fixed point: $R_b(\mathbf{K}^*+\delta\mathbf{K})=\mathbf{K}^*+\mathsf{M}(b)\,\delta\mathbf{K}+O(\delta K^2)$ . Why: $R_b(\mathbf{K}^*)=\mathbf{K}^*$ by definition of a fixed point, so the constant term drops and the leading behaviour is linear in the deviation.

Step 2. Apply the chain rule to $R_{b_1}\circ R_{b_2}=R_{b_1 b_2}$ at the fixed point: $\mathsf{M}(b_1)\,\mathsf{M}(b_2)=\mathsf{M}(b_1 b_2)$ . Why: The Jacobian of a composition is the product of Jacobians; evaluated at the fixed point (which every $R_b$ maps to itself) this transfers the semigroup law from the maps to their stability matrices.

Step 3. Diagonalize using left eigenvectors: $\sum_j \phi^{(i)}_j \mathsf{M}_{jk}(b) = \lambda_i(b)\,\phi^{(i)}_k$ , so each eigenvalue inherits $\lambda_i(b_1 b_2)=\lambda_i(b_1)\lambda_i(b_2)$ . Why: $\mathsf{M}$ is generically non-symmetric, so we use its left eigenvectors; on the common eigenbasis the matrix product law passes to each scalar eigenvalue.

Step 4. Solve the multiplicative functional equation: the only continuous solution of $\lambda(b_1 b_2)=\lambda(b_1)\lambda(b_2)$ is $\lambda_i(b)=b^{y_i}$ with $y_i\in\mathbb{R}$ . Why: Taking logarithms turns the multiplicative law into Cauchy's additive equation $g(b_1 b_2)=g(b_1)+g(b_2)$ in $\ln b$ , whose continuous solutions are linear — hence $\lambda$ is a power. This is where the power law is born.

Step 5. Define the scaling fields $u_i\equiv\sum_j\phi^{(i)}_j\delta K_j$ and confirm $u_i'=\sum_j\phi^{(i)}_j\delta K_j'=\lambda_i(b)\,u_i=b^{y_i}u_i$ . Why: Projecting the linear flow $\delta\mathbf{K}'=\mathsf{M}\,\delta\mathbf{K}$ onto a left eigenvector diagonalizes it; the scaling field is just that projection, and it scales by its own eigenvalue. Verified: verify_aq-sm-scaling-fields-scaling-variables.py checks the multiplicative functional equation numerically (a random $y$ satisfies $b_1^{y}b_2^{y}=(b_1b_2)^{y}$ ) and confirms that the log-map turns it into Cauchy additivity; matches Kardar Ch. 4 and Goldenfeld Ch. 9.

Go deeper · Derivation B · 1 minThe exponent relations from the scaling form

Derivation BThe exponent relations from the scaling form

We differentiate the homogeneous free energy to extract $\alpha,\beta,\gamma,\delta$ and then verify that the Rushbrooke and Widom identities hold as exact equalities — a clean demonstration that the two relevant eigenvalues encode all of thermodynamics. This is the structural result the article owns: not the individual exponents (which the RG-eigenvalue article supplies) but the fact that they are forced to satisfy the scaling laws with equality. Begin from $f_s(u_t,u_h)=b^{-d}f_s(b^{y_t}u_t,b^{y_h}u_h)$ with $u_t\propto t,\ u_h\propto h$ at leading order.

Step 1. Specific heat at $h=0$ : setting $b^{y_t}|u_t|=1$ gives $f_s\sim|t|^{d/y_t}$ , and two temperature derivatives yield $C\sim|t|^{d/y_t-2}$ , so $\alpha=2-d/y_t$ . Why: Each $\partial_t$ lowers the power of $|t|$ by one; the specific heat is the second temperature derivative of the free energy.

Step 2. Order parameter: $m=-\partial_h f_s|_{h=0}\sim|t|^{(d-y_h)/y_t}$ , so $\beta=(d-y_h)/y_t$ . Why: One field derivative brings down a factor $b^{y_h}=|t|^{-y_h/y_t}$ ; evaluating at $h=0$ , $t<0$ gives the spontaneous magnetization.

Step 3. Susceptibility: a second field derivative gives $\chi\sim|t|^{(d-2y_h)/y_t}$ , so $\gamma=(2y_h-d)/y_t$ . Why: Two field derivatives bring $b^{2y_h}$ ; the sign flips because $\gamma$ is defined by $\chi\sim|t|^{-\gamma}$ .

Step 4. Critical isotherm: at $t=0$ rescale instead with $b^{y_h}|u_h|=1$ to get $m\sim h^{(d-y_h)/y_h}$ , so $\delta=y_h/(d-y_h)$ . Why: On the critical isotherm there is no temperature scale, so the magnetic field must set the magnification.

Step 5. Combine into Rushbrooke: $\alpha+2\beta+\gamma=\dfrac{(2y_t-d)+2(d-y_h)+(2y_h-d)}{y_t}=\dfrac{2y_t}{y_t}=2$ . Why: Every exponent is a rational function of $y_t,y_h,d$ ; the numerators telescope so the dimension and $y_h$ cancel, leaving exactly $2$ . The old inequality becomes an identity.

Step 6. Combine into Widom: $\beta(\delta-1)=\dfrac{d-y_h}{y_t}\cdot\dfrac{2y_h-d}{d-y_h}=\dfrac{2y_h-d}{y_t}=\gamma$ . Why: The $(d-y_h)$ factors cancel, leaving $\gamma$ exactly; only $y_t$ and $y_h$ (two numbers) are independent. Verified: verify_aq-sm-scaling-fields-scaling-variables.py evaluates all six exponents symbolically in $(y_t,y_h,d)$ and asserts Rushbrooke, Widom, Fisher, and Josephson as identities, then substitutes the 2D-Ising eigenvalues $y_t=1,y_h=15/8,d=2$ to reproduce $(\alpha,\beta,\gamma,\delta,\nu,\eta)=(0,1/8,7/4,15,1,1/4)$ .

Go deeper · Derivation C · 1 minNonlinear scaling fields and the origin of corrections to scaling

Derivation CNonlinear scaling fields and the origin of corrections to scaling

We construct the analytic-but-nonlinear redefinition that makes the scaling law exact, and show how the leading irrelevant field generates the Wegner correction-to-scaling exponent. The point is conceptual: measured exponents are clean only because the true scaling variable is a curved coordinate, and the leftover curvature is itself universal. Demand $u_i(R_b\mathbf{K})=b^{y_i}u_i(\mathbf{K})$ beyond linear order and see what it forces.

Step 1. Write the most general analytic forms respecting the Ising symmetry $h\to-h$ : $u_t=t+c_1 t^2+c_2 h^2+\cdots$ , $u_h=h+c_3\,t\,h+\cdots$ . Why: Symmetry forbids a term linear in $h$ inside $u_t$ (which is even) and forbids a term independent of $h$ inside $u_h$ (which is odd); the leading corrections are the ones shown.

Step 2. Impose exact scaling order by order: the quadratic coefficients solve $a^{(i)}_{jk}(b^{y_j+y_k}-b^{y_i})=(R_b)^{(2)}_{jk}$ . Why: Matching the $b^{y_i}u_i$ law at quadratic order gives a linear equation for each coefficient; it is solvable as long as the prefactor $b^{y_j+y_k}-b^{y_i}$ does not vanish (no resonance).

Step 3. Keep one irrelevant field $u_3$ (eigenvalue $y_3<0$ ): $f_s(u_t,u_h,u_3)=b^{-d}f_s(b^{y_t}u_t,b^{y_h}u_h,b^{-|y_3|}u_3)$ . Why: The irrelevant direction does not vanish in a single step of the iterated map; it decays but must be carried along to see its effect.

Step 4. Fix $b=|u_t|^{-1/y_t}$ , so the third argument becomes $u_3|u_t|^{|y_3|/y_t}\to0$ as $t\to0$ . Why: This is the same magnification that collapses the free energy; under it the irrelevant field's scaled value is a small parameter, formalizing the word "irrelevant."

Step 5. Expand in the small argument: $f_s=|u_t|^{2-\alpha}[\Phi_0(z)+|u_t|^{\nu\omega}\Phi_1(z)+\cdots]$ with $\omega=|y_3|$ and $\theta=\nu\omega$ . Why: Taylor-expanding the scaling function in its small third argument produces a leading correction of relative size $|u_t|^{\nu\omega}$ ; this is the Wegner correction, universal in its exponent though not its amplitude. Verified: verify_aq-sm-scaling-fields-scaling-variables.py confirms $\theta=\nu\omega$ numerically for the 3D-Ising values $\nu\approx0.630,\ \omega\approx0.83$ giving $\theta\approx0.52$ , and checks the resonance condition $y_i=y_j+y_k$ that signals logarithms; structure follows Wegner (1974) and Cardy Ch. 3.

Limiting cases

Case	Result	Meaning
Leading relevant thermal field	$u_t \to b^{y_t} u_t,\ \ \nu = 1/y_t$	Controls the divergence of the correlation length $\xi\sim\lvert u_t\rvert^{-\nu}$ ; $u_t\propto t$ at leading order.
Leading relevant magnetic field	$u_h \to b^{y_h} u_h,\ \ \Delta=\beta\delta=\nu y_h$	Conjugate to the order parameter; $u_h\propto h$ at leading order, with $y_h=(d+2-\eta)/2$ .
Leading irrelevant field	$u_3 \to b^{-\lvert y_3\rvert} u_3,\ \ \omega=\lvert y_3\rvert$	Gives corrections to scaling $\sim u_t^{\,\nu\omega}$ (Wegner $\theta=\nu\omega$ ); ignorable asymptotically but vital for fits.
Marginal field ( $y=0$ )	$u' = u + a\,u^2\ln b + \cdots$	Pure power law fails; logarithmic corrections appear, as at the upper critical dimension $d=4$ for $\phi^4$ .
Dangerously irrelevant field	$f_s$ singular as $u_u\to0$	An irrelevant field on which $f_s$ depends singularly; restores mean-field exponent relations above $d_c=4$ .

Worked examples

Example 1 — The 2D-Ising exponents from two eigenvalues. The RG eigenvalues at the square-lattice Ising fixed point are $y_t=1$ and $y_h=15/8$ in $d=2$ . Compute the full exponent set and check it against Onsager's exact solution.

Step 1. Correlation-length exponent: $\nu=1/y_t=1/1=\mathbf{1}$ .

Step 2. Specific heat: $\alpha=2-d/y_t=2-2/1=\mathbf{0}$ (a logarithmic divergence, the marginal $\alpha=0$ case).

Step 3. Order parameter: $\beta=(d-y_h)/y_t=(2-15/8)/1=1/8=\mathbf{0.125}$ .

Step 4. Susceptibility: $\gamma=(2y_h-d)/y_t=(2\cdot15/8-2)/1=(15/4-2)=7/4=\mathbf{1.75}$ .

Step 5. Critical isotherm: $\delta=y_h/(d-y_h)=(15/8)/(2-15/8)=(15/8)/(1/8)=\mathbf{15}$ .

Step 6. Anomalous dimension: $\eta=d+2-2y_h=2+2-2\cdot15/8=4-15/4=1/4=\mathbf{0.25}$ . The set $(\alpha,\beta,\gamma,\delta,\nu,\eta)=(0,\tfrac18,\tfrac74,15,1,\tfrac14)$ is exactly Onsager's 2D-Ising result — obtained here purely from the two relevant eigenvalues.

Example 2 — Only two of the six are independent. Using the same values, verify the four scaling relations hold as exact equalities, confirming that $y_t,y_h,d$ carry all the information.

Step 1. Rushbrooke: $\alpha+2\beta+\gamma=0+2\cdot\tfrac18+\tfrac74=\tfrac14+\tfrac74=\tfrac{8}{4}=\mathbf{2}$ . ✓

Step 2. Widom: $\gamma=\beta(\delta-1)=\tfrac18\,(15-1)=\tfrac{14}{8}=\tfrac74=\mathbf{1.75}$ . ✓

Step 3. Fisher: $\gamma=\nu(2-\eta)=1\cdot(2-\tfrac14)=\tfrac74=\mathbf{1.75}$ . ✓

Step 4. Josephson (hyperscaling): $2-\alpha=d\nu\Rightarrow 2-0=2\cdot1=\mathbf{2}$ . ✓ All four hold with equality. The RG has promoted the classical thermodynamic inequalities to identities — the signature that a fixed point, not a free assumption, is in charge.

Origins · 1 minrelevant operator and how this was discovered

relevant operator

RG classification of perturbations · The organizing idea that a critical fixed point is characterized by which perturbations grow under coarse-graining. A perturbation with positive RG eigenvalue is "relevant" — it drives the system away from criticality and must be tuned to zero to sit at the transition (temperature and field for a magnet). The count of relevant operators equals the number of parameters that must be tuned, which is why a generic critical point needs exactly two knobs. Formalized in the early 1970s in Wilson's and Wegner's work and now the backbone of how physicists classify every continuous transition.

History

1966

— Kadanoff introduces the block-spin scaling picture, giving the homogeneity hypothesis a physical basis

1971

— Wilson formulates the renormalization group as a flow in coupling space, defining fixed points and eigenvalue exponents

1972

— Wegner introduces nonlinear scaling fields and the systematic theory of corrections to scaling

1974

— Wegner's "corrections to scaling" analysis fixes the Wegner exponent θ = νω from the leading irrelevant eigenvalue

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 14 (RG and scaling) · Kardar, Statistical Physics of Fields, Ch. 4–5 · Goldenfeld, Lectures on Phase Transitions and the Renormalization Group, Ch. 9–10 · Wegner, J. Phys. C 7, 2098 (1974) (corrections to scaling, nonlinear scaling fields) · Cardy, Scaling and Renormalization in Statistical Physics, Ch. 3 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 12

Provenance. All equations are standard and recomputed by verify_aq-sm-scaling-fields-scaling-variables.py: the multiplicative functional equation $\lambda(b_1b_2)=\lambda(b_1)\lambda(b_2)\Rightarrow\lambda=b^y$ ; the scaling-form derivation of $\alpha,\beta,\gamma,\delta,\nu,\eta$ from $f_s=b^{-d}f_s(b^{y_t}u_t,b^{y_h}u_h)$ ; the Rushbrooke, Widom, Fisher, and Josephson identities as algebraic consequences of $y_t,y_h,d$ ; and the Wegner identification $\theta=\nu\omega$ . Numerical anchors: 2D-Ising $y_t=1,\ y_h=15/8$ give the exact $(\alpha,\beta,\gamma,\delta,\nu,\eta)=(0,\tfrac18,\tfrac74,15,1,\tfrac14)$ , matching Onsager's Solution of the 2D Ising Model; 3D-Ising $\omega\approx0.83$ from high-precision methods. Values quoted are textbook-standard; no non-standard figures introduced.

Builds on: RG Flow in Coupling Space · Fixed Points of the RG · Linearization About a Fixed Point and Eigenvalue Spectrum · The Renormalization Group Idea: Coarse-Graining and Rescaling. Leads to: Relevant, Irrelevant, and Marginal Operators · Critical Exponents from the RG · Wilson-Fisher Fixed Point · The Epsilon-Expansion.

Are scaling fields the same as eigenoperators of the RG?

Closely related but not identical. The eigenoperators are the perturbations added to the fixed-point Hamiltonian whose couplings scale multiplicatively; the scaling fields are the couplings themselves, i.e. the coordinates in coupling space along the eigen-directions. In practice people use "scaling field," "scaling variable," and "eigenoperator coupling" almost interchangeably, but the field is the number you dial and the operator is the term it multiplies in the Hamiltonian.

Why exactly two relevant fields for the Ising class?

Because the Ising fixed point has exactly two perturbations with positive RG eigenvalue: the thermal (even) one with $y_t=1/\nu$ and the magnetic (odd) one with $y_h$ . Every other symmetry-allowed operator is irrelevant. The count of relevant fields equals the number of parameters you must tune to reach the critical point, which is why criticality generically needs two knobs, temperature and field. A different symmetry (say a tricritical point) has more relevant fields and needs more tuning.

If the coefficients relating u_t to t are non-universal, how can exponents be universal?

The universality lives in the eigenvalues $y_i$ , which are properties of the fixed point alone. The coefficients $c_i$ in $u_t=t+c_1t^2+\cdots$ are metric factors that translate from lab units to fixed-point units; they set amplitudes and correction sizes but never the exponents. Two materials in the same class have identical $y_t,y_h$ but different $c_i$ , so their exponents agree while their amplitudes and the location of $T_c$ differ.

What is the difference between a marginal field and an irrelevant one in practice?

An irrelevant field ( $y<0$ ) decays as a power under coarse-graining and produces corrections to scaling that vanish algebraically near $T_c$ . A marginal field ( $y=0$ ) neither grows nor decays at linear order; its fate is decided by the (usually quadratic) next term in the flow. Marginal fields produce logarithmic rather than power-law corrections — the famous $|\ln t|$ factors at the upper critical dimension $d=4$ , where the $\phi^4$ coupling is marginally irrelevant.

How does data collapse connect to the scaling-field homogeneity?

Data collapse is the experimental face of equation (6). Because $f_s=|u_t|^{d/y_t}\Phi(u_h/|u_t|^{\Delta})$ , plotting a rescaled observable (like $m/|t|^\beta$ ) against the rescaled field ( $h/|t|^{\Delta}$ ) for many temperatures makes every curve fall on the single function $\Phi$ . A clean collapse confirms the homogeneity and measures the exponents; a spread that shrinks toward $T_c$ is the fingerprint of the leading irrelevant field. The dedicated data-collapse article develops this into a quantitative method.

Why does the free energy pick up exactly b^{-d} and not some other power?

The free energy is extensive: it scales with the number of degrees of freedom. One coarse-graining step merges $b^d$ sites into one block, so the free energy per unit volume is multiplied by $b^d$ when you rewrite the theory on the coarse lattice, i.e. $f_s(u)=b^{-d}f_s(u')$ . The power is exactly the spatial dimension because that is how many directions get compressed per step. This is also why hyperscaling relations, which contain $d$ , are the fragile ones.