Scaling Form of the Free Energy from the RG

For a century the scaling hypothesis was a beautiful guess. Widom and Kadanoff observed that near a critical point the singular part of the free energy behaves as a generalized homogeneous function of temperature and field, and from that single assumption every relation among the critical exponents fell out — Rushbrooke, Widom, Fisher, Josephson. But the hypothesis explained nothing about why the free energy should be homogeneous, and it produced no values for the exponents. It was a pattern with no mechanism. The renormalization group supplies the missing mechanism, and this article's job is to show exactly how. The homogeneity is not postulated; it is a direct consequence of how the free energy is assembled under coarse-graining. Each RG step integrates out short-wavelength fluctuations, writing their contribution as a smooth background plus a renormalized free energy of the surviving coarse system. Because the partition function is unchanged while $b^d$ degrees of freedom are merged into one, the free energy per unit volume picks up a factor $b^{-d}$ , and the couplings flow. Written in the scaling fields $t$ and $h$ — which near the fixed point transform multiplicatively as $t\to b^{y_t}t$ , $h\to b^{y_h}h$ — this becomes the master relation $f_s(t,h)=b^{-d}f_s(b^{y_t}t,b^{y_h}h)$ . Once you have that functional equation, every static exponent follows by differentiation and one clever choice of the arbitrary rescaling $b$ . Setting $b=|t|^{-1/y_t}$ magnifies until the temperature field is order one, and the leading singularities of specific heat, magnetization, susceptibility, and critical isotherm read off directly. Crucially the resulting $\alpha,\beta,\gamma,\delta$ are all built from the same two eigenvalues, so the scaling relations are not separate facts to be checked — they are algebraic identities. The construction even reveals why hyperscaling, the relation containing the dimension $d$ , is the fragile one: it rests on the free energy scaling like an inverse correlation volume, an assumption that fails above the upper critical dimension.

Key takeaways

Under one RG step with rescaling factor $b$ , the singular free energy per spin obeys $f_s(t,h)=b^{-d}f_s(b^{y_t}t,b^{y_h}h)$ — a homogeneity relation forced by extensivity and the linearized flow, not assumed.
The two RG eigenvalues $y_t$ (thermal) and $y_h$ (magnetic) are the only inputs; every static critical exponent is a rational function of $y_t,y_h,d$ .
Choosing $b=|t|^{-1/y_t}$ collapses the free energy onto a single scaling function of $h/|t|^{y_h/y_t}$ — the precise statement of data collapse.
All scaling relations — Rushbrooke $\alpha+2\beta+\gamma=2$ , Widom $\beta\delta=\beta+\gamma$ , Josephson $2-\alpha=d\nu$ , Fisher $\gamma=\nu(2-\eta)$ — become algebraic identities.
Hyperscaling contains $d$ explicitly because it assumes $f_s\sim\xi^{-d}$ ; it fails above the upper critical dimension where a dangerously irrelevant variable modifies the matching.
Key idea: Coarse-graining multiplies the free-energy density by $b^d$ while driving the scaling fields by $b^{y_t},b^{y_h}$ ; the resulting homogeneity of $f_s$ turns the old scaling hypothesis into a theorem and encodes all static critical behaviour in just $y_t$ , $y_h$ , and $d$ .

Where this lives today

The scaling form is the workhorse behind every quantitative test of the RG. When a Monte Carlo study of a three-dimensional magnet reports $\nu=0.6300(4)$ , the analysis is a scaling-form fit: rescale the free energy or its derivatives by the predicted powers, demand collapse, and read the exponents. The same machinery underlies universal amplitude ratios — quantities like $A_+/A_-$ (specific-heat amplitudes above and below $T_c$ ) that, unlike the amplitudes themselves, are fixed entirely by the scaling function $\Phi_\pm$ and therefore provide tests of universality even sharper than the exponents. The construction reaches well beyond thermal criticality. In quantum phase transitions the free-energy density is replaced by a ground-state energy density and temperature by a coupling $g$ , but the homogeneity relation and its consequences carry over verbatim, with a dynamic exponent $z$ inserted to account for time scaling differently from space. This is why one scaling form organizes the critical behaviour of systems as different as superfluid helium films, quantum Hall plateau transitions, and the onset of antiferromagnetism in heavy-fermion metals.

The physical picture

Think of the free energy as built layer by layer. Each RG step integrates out the shortest-wavelength fluctuations and records their contribution as a smooth, analytic piece plus a renormalized free energy of what remains. Iterate, and the singular behaviour is pushed entirely into how the coarse free energy depends on the flowing couplings. Near a fixed point those couplings move geometrically — by fixed powers of $b$ — so the free energy inherits a power-law, homogeneous structure. The homogeneity is thus a bookkeeping fact about how contributions accumulate, not a mysterious symmetry. The freedom to choose $b$ is a magnifying glass you control. The physics is invariant under the RG, so you may zoom until the system is no longer near criticality — until $b^{y_t}|t|=1$ . At that magnification the correlation length has shrunk to about one lattice spacing and the free energy is a smooth function of the rescaled field. The prefactor you accumulated on the way, $b^{-d}=|t|^{d/y_t}$ , is the leading singularity. You have converted a hard question about a divergence into an easy question about a smooth function at a chosen magnification. Data collapse is the experimental face of this equation, and it is worth taking literally: plot $f_s/|t|^{d/y_t}$ (or $m/|t|^\beta$ ) against $h/|t|^{y_h/y_t}$ for many temperatures, and all the curves land on one universal function $\Phi_\pm$ . The collapse working is direct visual proof that only two exponents govern the whole critical region. When it fails near $T_c$ , the failure is informative — it is the leading irrelevant field announcing itself. Hyperscaling deserves special attention because it is the relation that breaks. It says the free energy of a single correlation volume is of order $k_BT$ : $f_s\,\xi^{d}\sim 1$ . This is a geometric statement — it counts independent fluctuating blocks — which is why it carries the spatial dimension $d$ and why it collapses when fluctuations stop dominating above $d_c=4$ . There a quartic coupling that is technically irrelevant refuses to be set to zero (it is "dangerously irrelevant"), and the naive counting of blocks fails.

Figure

Zooming with b: the free-energy prefactor becomes the singularity

The picture above is the mechanism in three frames: coarse-grain until the correlation length is one cell, and the whole temperature singularity has moved into the accumulated prefactor $|t|^{d/y_t}$ . That is why two eigenvalues suffice.

Figure

Scaling relations printed as the output: two eigenvalues fix all six exponents

The core equations

The master relation is the whole article in one line: coarse-grain once, and the singular free-energy density transforms with a volume factor $b^{-d}$ while its scaling-field arguments are driven outward by their eigenvalues. Everything downstream is differentiation and a smart choice of $b$ .

f_s(t,h)=b^{-d}\,f_s\!\left(b^{y_t}t,\,b^{y_h}h\right) \qquad (1)

where: $f_s$ — singular part of the free energy per unit volume (per spin), in units of $k_BT$ ; $b$ — linear rescaling factor of one RG step ( $b>1$ ); $d$ — spatial dimension; $t$ — reduced-temperature scaling field, $t\propto(T-T_c)/T_c$ ; $h$ — reduced-field scaling field; $y_t,y_h$ — thermal and magnetic RG eigenvalues at the fixed point. Read it as: the free energy at criticality looks the same at every magnification, up to a known prefactor. Because equation (1) holds for any $b$ , we exploit the freedom by pinning the leading relevant field. Setting $b^{y_t}|t|=1$ eliminates $b$ and collapses $f_s$ onto a function of one scaling combination — the origin of every data-collapse plot.

f_s(t,h)=|t|^{d/y_t}\,\Phi_{\pm}\!\left(\frac{h}{|t|^{y_h/y_t}}\right) \qquad (2)

where: $\Phi_\pm$ — universal scaling function, with a $+$ branch for $t>0$ (disordered) and a $-$ branch for $t<0$ (ordered); $y_h/y_t=\Delta=\beta\delta$ — the gap exponent controlling the field–temperature trade-off. The prefactor $|t|^{d/y_t}$ carries the temperature singularity; the function carries the shape. Two of the exponents are not thermodynamic derivatives at all but properties of correlations, and they too come from the eigenvalues. The correlation length follows from the trivial geometric rescaling of length; the anomalous dimension from the scaling dimension of the order-parameter field.

\nu=\frac{1}{y_t},\qquad \eta=d+2-2y_h \qquad (3)

where: $\nu$ — correlation-length exponent, $\xi\sim|t|^{-\nu}$ ; $\eta$ — anomalous dimension of the order-parameter field. Here $\nu=1/y_t$ because length rescales by $b$ while $t$ rescales by $b^{y_t}$ ; $\eta$ encodes the deviation of the two-point function from its Gaussian $r^{-(d-2)}$ decay. Differentiating the collapsed free energy delivers the four thermodynamic exponents. Each is a rational function of $y_t,y_h,d$ , which is the entire reason the scaling relations exist.

\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} \qquad (4)

where: $\alpha$ — specific heat $C\sim|t|^{-\alpha}$ ; $\beta$ — order parameter $m\sim(-t)^{\beta}$ ; $\gamma$ — susceptibility $\chi\sim|t|^{-\gamma}$ ; $\delta$ — critical isotherm $m\sim h^{1/\delta}$ . Two numbers, four exponents — hence two constraints, which are the scaling relations below. The first constraint, Rushbrooke, is now an equality. In the old thermodynamics it was an inequality proved from convexity; the RG makes the slack vanish.

\alpha+2\beta+\gamma=2 \qquad (5)

where: an identity the moment $\alpha,\beta,\gamma$ are written in terms of $y_t,y_h,d$ — the numerators telescope. This is the sharpest signal that a single fixed point, not independent thermodynamics, is in control. The second constraint, Widom, ties the critical isotherm to the coexistence and susceptibility exponents through the gap exponent.

\gamma=\beta(\delta-1),\qquad \beta\delta=\beta+\gamma=\Delta \qquad (6)

where: $\Delta=y_h/y_t$ — the gap exponent, the power of $|t|$ needed to keep $h$ in the scaling combination. It sets the field scale $|t|^{\Delta}$ separating linear response from critical-isotherm behaviour. The third and fourth relations contain the dimension explicitly. Josephson (hyperscaling) and Fisher close the set — but hyperscaling's dependence on $d$ is the seed of its later breakdown.

2-\alpha=d\nu,\qquad \gamma=\nu(2-\eta) \qquad (7)

where: hyperscaling $2-\alpha=d\nu$ holds only for $d\le d_c=4$ ; Fisher $\gamma=\nu(2-\eta)$ contains no $d$ and survives above $d_c$ . The presence or absence of $d$ predicts exactly which relations break. Finally, an irrelevant coupling $u$ (eigenvalue $y_u<0$ ) can be added. Generically it produces harmless corrections to scaling; the special "dangerous" case, where $f_s$ depends on it singularly, is what rescues mean-field exponents above four dimensions.

f_s(t,h,u)=b^{-d}f_s(b^{y_t}t,\,b^{y_h}h,\,b^{y_u}u),\quad y_u<0 \qquad (8)

where: $u$ — least-irrelevant coupling, eigenvalue $y_u<0$ . For generic $u$ this gives corrections $\sim|t|^{-y_u/y_t}$ ; a dangerously irrelevant $u$ enters $\Phi$ singularly and instead modifies the leading exponents for $d>d_c$ .

Derivation AHyperscaling and the specific-heat exponent

We derive Josephson's hyperscaling relation $2-\alpha=d\nu$ directly from the master scaling form, exhibiting exactly where the dimension $d$ enters and hence why this relation — unlike Rushbrooke or Widom — can break down above four dimensions. This is the structural point the article owns: the RG does not just reproduce the scaling relations, it tells you which are robust and which are fragile, by tracking whether $d$ appears. Set $h=0$ , magnify until the temperature field is order unity, and read off the leading singularity.

Step 1. Specialize the master form to zero field: $f_s(t,0)=b^{-d}f_s(b^{y_t}t,0)$ . Why: On the coexistence axis $h=0$ there is only the thermal field to worry about, so the two-argument relation collapses to one argument.

Step 2. Choose the magnification $b=|t|^{-1/y_t}$ , so $b^{y_t}|t|=1$ . Why: This zooms out until the renormalized temperature is $O(1)$ ; the right-hand free energy becomes a pure constant $\Phi_\pm(0)$ , isolating the singular prefactor.

Step 3. Read off the singular free energy: $f_s(t,0)=|t|^{d/y_t}\,\Phi_\pm(0)$ . Why: The entire temperature dependence now sits in the accumulated prefactor $b^{-d}=|t|^{d/y_t}$ ; the dimension $d$ enters here, as the power of the volume factor.

Step 4. Differentiate twice: $C=-T\,\partial^2 f/\partial T^2\sim|t|^{d/y_t-2}$ , so comparing with $C\sim|t|^{-\alpha}$ gives $\alpha=2-d/y_t$ . Why: Each temperature derivative lowers the exponent by one; the specific heat is the second derivative of the free energy with respect to temperature.

Step 5. Insert $\nu=1/y_t$ : $2-\alpha=d/y_t=d\nu$ . Why: The correlation-length exponent is the inverse thermal eigenvalue, so $d/y_t=d\nu$ . The explicit $d$ is inherited from the volume prefactor in Step 3 — that is the geometric origin of hyperscaling and of its breakdown above $d_c$ . Verified: verify_aq-sm-scaling-free-energy-rg.py checks $2-\alpha=d\nu$ symbolically in $(y_t,d)$ and confirms 2D-Ising ( $y_t=1,d=2$ ) gives $\alpha=0$ , $d\nu=2$ , consistent with Onsager's $\nu=1,\alpha=0$ .

Go deeper · Derivation B · 1 minWidom relation from the field derivative

Derivation BWidom relation from the field derivative

We obtain Widom's relation $\beta\delta=\beta+\gamma$ purely algebraically from the two eigenvalues, showing it is an identity with no residual freedom — the gap exponent $\Delta=y_h/y_t$ appears symmetrically on both sides. Because this relation contains no $d$ , it will survive above the upper critical dimension where hyperscaling fails, a contrast the article uses to teach which relations are robust. Differentiate the scaling form with respect to $h$ in complementary limits.

Step 1. Magnetization from the scaling form: $m=-\partial f_s/\partial h=|t|^{(d-y_h)/y_t}\,\Phi'_\pm(h/|t|^{y_h/y_t})$ . Why: One field derivative reduces the prefactor exponent by $y_h/y_t$ and differentiates the scaling function; the chain rule supplies the inner factor $|t|^{-y_h/y_t}$ .

Step 2. Spontaneous magnetization at $h\to0,\ t<0$ : $m\sim(-t)^{(d-y_h)/y_t}$ , so $\beta=(d-y_h)/y_t$ . Why: In the ordered phase $\Phi'_-(0)$ is a nonzero constant, so the magnetization scales as the prefactor alone.

Step 3. Susceptibility: a second field derivative gives $\chi\sim|t|^{(d-2y_h)/y_t}$ , so $\gamma=(2y_h-d)/y_t$ . Why: Each field derivative reduces the exponent by another $y_h/y_t$ ; the sign convention $\chi\sim|t|^{-\gamma}$ flips it.

Step 4. Critical isotherm from $f_s(0,h)\sim|h|^{d/y_h}$ : $m\sim h^{(d-y_h)/y_h}=h^{1/\delta}$ , so $\delta=y_h/(d-y_h)$ . Why: At $t=0$ the magnification must be set by the field, $b^{y_h}|h|=1$ , giving a pure power of $h$ .

Step 5. Combine: $\beta\delta=\dfrac{d-y_h}{y_t}\cdot\dfrac{y_h}{d-y_h}=\dfrac{y_h}{y_t}=\dfrac{(d-y_h)+(2y_h-d)}{y_t}=\beta+\gamma$ . Why: The $(d-y_h)$ factors cancel and the numerator regroups into $\beta+\gamma$ ; the gap exponent $\Delta=y_h/y_t$ is what both sides equal. No $d$ survives — Widom is dimension-independent. Verified: verify_aq-sm-scaling-free-energy-rg.py asserts $\beta\delta=\beta+\gamma$ symbolically and checks the 2D-Ising numbers $\beta=1/8,\delta=15,\gamma=7/4$ give $\beta\delta=15/8=\beta+\gamma$ .

Go deeper · Derivation C · 1 minBreakdown of hyperscaling from a dangerously irrelevant variable

Derivation CBreakdown of hyperscaling from a dangerously irrelevant variable

We show how the irrelevant quartic coupling above four dimensions refuses to be set to zero and thereby restores mean-field exponents while breaking Josephson's relation — the concrete meaning of "dangerously irrelevant." This is the payoff of tracking where $d$ enters: the geometric argument $f_s\sim\xi^{-d}$ is exactly the one that fails. Work at the Gaussian fixed point where $y_t=2$ and the quartic coupling has eigenvalue $y_u=4-d$ .

Step 1. Include $u$ in the zero-field scaling form: $f_s(t,0,u)=b^{-d}f_s(b^{y_t}t,0,b^{y_u}u)$ with $y_t=2$ , $y_u=4-d$ . Why: At the Gaussian fixed point these are the exact eigenvalues; for $d>4$ , $y_u<0$ so $u$ flows toward zero and looks irrelevant.

Step 2. Note the naive matching would fail: if $\Phi(\cdot,u)$ had a smooth $u\to0$ limit, hyperscaling would give $\alpha=2-d\nu=2-d/2$ , a $d$ -dependent value contradicting mean field. Why: Mean-field theory gives $\alpha=0$ for all $d>4$ ; a $d$ -dependent $\alpha$ would be inconsistent, signalling that $u$ cannot simply be dropped.

Step 3. Recognize the singular dependence: $f_s\sim|t|^{d/y_t}\,\Psi(u\,|t|^{-y_u/y_t})$ with $\Psi(x)\to x^{p}$ as $x\to0$ . Why: The tree-level $\phi^4$ free energy scales as a power of $u$ , so the scaling function diverges as $u\to0$ — the coupling is "dangerous" and must be retained.

Step 4. Fix $p=d/2-2$ from the tree-level free energy: $f_s\sim|t|^{d/y_t+p\,y_u/y_t}u^{p}$ , and the temperature power becomes $|t|^{2}$ . Why: Demanding the $\phi^4$ theory's mean-field free energy $f\sim t^2/u$ fixes the singular power; the exponents combine so the net temperature power is exactly $2$ , i.e. $\alpha=0$ .

Step 5. Conclude: because $u$ enters $f_s$ singularly, the geometric argument $f_s\sim\xi^{-d}$ is invalid for $d>d_c=4$ ; hyperscaling $2-\alpha=d\nu$ fails while Widom and Fisher (which lack $d$ ) survive. Why: The block-counting picture behind hyperscaling assumes only $\xi$ sets the scale; the dangerous $u$ introduces a second scale, breaking exactly the $d$ -dependent relation and leaving the dimension-free ones intact. Verified: verify_aq-sm-scaling-free-energy-rg.py confirms that for $d>4$ mean-field values ( $\alpha=0,\beta=1/2,\gamma=1,\delta=3,\nu=1/2$ ) violate $2-\alpha=d\nu$ (since $2\ne d/2$ for $d>4$ ) yet satisfy Widom and Fisher, and that at $d=4$ hyperscaling is marginally restored ( $2=d/2$ ).

Limiting cases

Case	Result	Meaning
Zero field, $h=0$	$f_s(t,0)\sim\lvert t\rvert^{d/y_t}\Phi_\pm(0)$	Choosing $b=\lvert t\rvert^{-1/y_t}$ gives the specific-heat singularity with $2-\alpha=d/y_t=d\nu$ (hyperscaling).
Critical isotherm, $t=0$	$f_s(0,h)\sim\lvert h\rvert^{d/y_h}$	Set $b=\lvert h\rvert^{-1/y_h}$ ; differentiating gives $m\sim\lvert h\rvert^{1/\delta}$ , so $\delta=y_h/(d-y_h)$ .
Mean-field, $d>d_c=4$	$y_t=2,\ y_h=3,\ d_c=4$	Gaussian eigenvalues; a dangerously irrelevant $u$ pins exponents to mean-field and breaks hyperscaling for $d>4$ .
2D-Ising fixed point	$y_t=1,\ y_h=\tfrac{15}{8},\ d=2$	Yields $\nu=1,\alpha=0,\beta=\tfrac18,\gamma=\tfrac74,\delta=15$ — Onsager's exact exponents from two eigenvalues.

Worked examples

Example 1 — The full 2D-Ising exponent set from two eigenvalues. The square-lattice Ising fixed point has $y_t=1$ , $y_h=15/8$ in $d=2$ . Show these two numbers reproduce Onsager's exact exponents and satisfy hyperscaling.

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

Step 2. $\beta=(d-y_h)/y_t=(2-15/8)/1=(16/8-15/8)=1/8=\mathbf{0.125}$ .

Step 3. $\gamma=(2y_h-d)/y_t=(30/8-2)/1=(30/8-16/8)=14/8=7/4=\mathbf{1.75}$ .

Step 4. $\delta=y_h/(d-y_h)=(15/8)/(1/8)=\mathbf{15}$ .

Step 5. Josephson check: $2-\alpha=2$ and $d\nu=2\cdot1=2$ . They agree, so hyperscaling holds in $d=2$ , exactly as it must below $d_c=4$ . The complete set $(\alpha,\beta,\gamma,\delta,\nu)=(0,\tfrac18,\tfrac74,15,1)$ is Onsager's — derived here from the master scaling form alone.

Example 2 — Reading exponents off the $\epsilon$ -expansion eigenvalues. At the Wilson-Fisher fixed point in $d=4-\epsilon$ the eigenvalues are $y_t=2-\epsilon/3+O(\epsilon^2)$ and $y_h=3-\epsilon/2+O(\epsilon^2)$ for $n=1$ . Compute $\nu$ to first order in $\epsilon$ and evaluate at $\epsilon=1$ .

Step 1. $\nu=1/y_t=\dfrac{1}{2-\epsilon/3}=\dfrac12\cdot\dfrac{1}{1-\epsilon/6}$ .

Step 2. Expand for small $\epsilon$ : $\nu=\tfrac12(1+\tfrac{\epsilon}{6}+\cdots)=\tfrac12+\dfrac{\epsilon}{12}+O(\epsilon^2)$ .

Step 3. Confirm against the standard one-loop result $\nu=\tfrac12+\dfrac{(n+2)}{4(n+8)}\epsilon$ ; at $n=1$ the coefficient is $3/36=1/12$ . ✓

Step 4. Set $\epsilon=1$ (i.e. $d=3$ ): $\nu=\tfrac12+\tfrac{1}{12}=7/12\approx\mathbf{0.583}$ . This one-loop value undershoots the true 3D-Ising $\nu\approx0.630$ — the honest truncation gap the $\epsilon$ -expansion article discusses. The scaling form is what turned the two computed eigenvalues into a measurable exponent.

Origins · 1 minscaling dimension and how this was discovered

scaling dimension

The RG scaling dimension of an operator · The number that says how a field or coupling stretches under a change of length scale. In the RG a scaling operator with dimension

x

contributes to the free energy through an eigenvalue

y=d-x

, so relevant operators have small scaling dimension and irrelevant ones large. The scaling dimension of the order-parameter field,

x_\phi=(d-2+\eta)/2

, fixes the anomalous dimension

\eta

and the decay of the two-point function

G(r)\sim r^{-2x_\phi}

. It is the bridge between the abstract RG eigenvalues and the concrete exponents an experiment measures, and the central quantity that conformal field theory later computes exactly in two dimensions.

History

1965

— Widom proposes the homogeneity hypothesis for the singular free energy, deriving the exponent relations

1966

— Kadanoff's block-spin argument gives the homogeneity a physical picture and predicts data collapse

1971

— Wilson's renormalization group derives the homogeneity from coarse-graining and computes the eigenvalues

1972

— Wilson and Fisher compute the eigenvalues in the epsilon-expansion, turning the scaling form into real numbers

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 14 (Phase Transitions: Renormalization Group) · 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, Ch. 9 · Cardy, Scaling and Renormalization in Statistical Physics, Ch. 3

Provenance. All equations are standard and recomputed by verify_aq-sm-scaling-free-energy-rg.py: the master form $f_s=b^{-d}f_s(b^{y_t}t,b^{y_h}h)$ ; the exponent formulas $\alpha=2-d/y_t$ , $\beta=(d-y_h)/y_t$ , $\gamma=(2y_h-d)/y_t$ , $\delta=y_h/(d-y_h)$ , $\nu=1/y_t$ , $\eta=d+2-2y_h$ ; the Rushbrooke, Widom, Josephson, and Fisher identities; the 2D-Ising values $y_t=1,y_h=15/8$ giving Onsager's $\beta=1/8,\gamma=7/4,\delta=15,\nu=1,\alpha=0$ ; and the mean-field/hyperscaling-breakdown check for $d>4$ . The $\epsilon$ -expansion eigenvalues $y_t=2-\epsilon/3$ , $y_h=3-\epsilon/2$ are the standard Wilson-Fisher one-loop results. This article recalls the phenomenological Scaling Hypothesis and Scaling Relations by name as the target it derives, and cross-checks against Onsager's Solution of the 2D Ising Model.

Builds on: The Renormalization Group Idea: Coarse-Graining and Rescaling · Linearization About a Fixed Point and Eigenvalue Spectrum · Scaling Fields and Scaling Variables · Relevant, Irrelevant, and Marginal Operators. Leads to: Critical Exponents from the RG · Gaussian Fixed Point and Upper Critical Dimension · Wilson-Fisher Fixed Point · RG Flow in Coupling Space.

How is this different from the old Widom-Kadanoff scaling hypothesis?

The old scaling hypothesis postulated that $f_s$ is a generalized homogeneous function and then derived the exponent relations. It could not say why the function is homogeneous or what the exponents are. The RG derives the homogeneity from coarse-graining — the $b^{-d}$ prefactor and the multiplicative flow of the scaling fields — and computes the eigenvalues $y_t,y_h$ at a fixed point (exactly in 2D, perturbatively via the $\epsilon$ -expansion). So this article turns a hypothesis into a theorem and gives it predictive content.

Why do only two eigenvalues appear when there are infinitely many couplings?

Because at an Ising-like fixed point only two couplings are relevant — the thermal and magnetic scaling fields. Every other coupling is irrelevant and flows to zero under coarse-graining, contributing at most corrections to scaling. So although coupling space is infinite-dimensional, the asymptotic critical behaviour lives on a two-dimensional relevant subspace, and $f_s$ depends on just $y_t,y_h,d$ .

What exactly is the scaling function Phi, and is it universal?

$\Phi_\pm$ is the free energy evaluated at the reference magnification, as a function of the single scaling variable $h/|t|^\Delta$ . Its exponents and its overall shape are universal — the same for every system in the universality class — but it carries two non-universal metric factors that set the scales of $t$ and $h$ . Rescaling by those two amplitudes brings the scaling functions of different materials into exact coincidence, which is why data from a magnet and a fluid collapse onto the same curve.

Is hyperscaling ever restored above four dimensions?

Not in the ordinary sense. For $d>4$ the exponents stick at mean-field values, and $2-\alpha=d\nu$ reads $2=d/2$ , which is false for $d>4$ . The resolution is that a modified "hyperscaling" holds if you use the effective thermal length inflated by the dangerous irrelevant variable, but the naive relation with the bare $d$ is genuinely broken. At exactly $d=4$ the two sides coincide ( $2=d/2$ ) up to logarithmic corrections.

How does finite system size fit into this scaling form?

The inverse system size $L^{-1}$ is just another scaling field, with the trivial eigenvalue $y_L=1$ (length rescales geometrically). Adding it, $f_s(t,h,L^{-1})=b^{-d}f_s(b^{y_t}t,b^{y_h}h,bL^{-1})$ , and choosing $b=L$ gives the finite-size scaling form. This is the foundation of how simulations extract $T_c$ and the exponents from modest lattices; the finite-size-scaling article develops it in full.

Why is the singular part, not the full free energy, homogeneous?

Each coarse-graining step produces a smooth analytic contribution from the integrated-out short-wavelength modes, plus the renormalized free energy of the coarse system. The analytic pieces accumulate into a regular background that is not homogeneous — it is a perfectly ordinary Taylor series in $t$ and $h$ . Only the leftover singular part inherits the multiplicative flow of the couplings and hence the homogeneity. Subtracting the background correctly is essential before fitting exponents or amplitudes.