Scaling Hypothesis and Scaling Relations

Above the level of mean-field (Landau) theory, the critical exponents that govern continuous phase transitions are not independent fitting parameters: experiments and exact solutions reveal tight algebraic relations among them. The scaling hypothesis, advanced by Widom, Kadanoff, Fisher and others in the mid-1960s, is the minimal assumption that explains these relations. It asserts that the part of the free energy responsible for critical singularities is a generalized homogeneous function of its thermodynamic arguments, controlled by a small number of scaling exponents. Physically the hypothesis is motivated by the divergence of the correlation length ξ. As $T\to T_c$ , ξ grows without bound and becomes the only relevant length scale; microscopic lengths drop out. Kadanoff's block-spin picture makes this concrete: coarse-graining the system into blocks of side $b$ produces a rescaled system with $t\to b^{y_t}t$ and $h\to b^{y_h}h$ , while the singular free-energy density transforms with the factor $b^{-d}$ . Iterating and demanding self-similarity at $T_c$ yields the homogeneity law for $f_s$ . From that single functional equation, mechanical differentiations with judicious choices of $b$ generate every standard power law — specific heat (α), spontaneous magnetization (β), susceptibility (γ), critical isotherm (δ), correlation length (ν), correlation decay (η). Because all six are built from only two free-energy exponents (and $d$ ), they cannot be independent; eliminating $y_t,y_h$ produces the scaling relations. Rushbrooke and Widom involve only thermodynamic exponents and hold in any dimension; the hyperscaling relation involving $d$ additionally assumes that the singular free-energy density scales as one inverse correlation volume, $f_s \sim \xi^{-d}$ . The deeper justification — and the actual computation of $y_t,y_h$ — comes from the renormalization group, where the scaling exponents are the eigenvalues of the linearized RG flow at a fixed point. The scaling hypothesis sits between phenomenology and the RG: it is the structural statement the RG proves.

Key takeaways

Near a critical point the singular free-energy density is a generalized homogeneous function of the reduced temperature $t=(T-T_c)/T_c$ and ordering field $h$ : $f_s(t,h)=b^{-d}f_s(b^{y_t}t,\,b^{y_h}h)$ for any rescaling $b$ .
Differentiating this one relation and choosing $b$ to fix an argument reproduces all thermodynamic power laws and expresses α, β, γ, δ, ν, η in terms of just two numbers $y_t,y_h$ .
The scaling relations — Rushbrooke $α+2β+γ=2$ , Widom $γ=β(δ−1)$ , Fisher $γ=ν(2−η)$ , Josephson $2−α=dν$ — are exact equalities that slash the data needed to characterize a class.
For short-range Ising and $O(n)$ $\phi^4$ universality classes, conventional hyperscaling fails above the upper critical dimension $d_c=4$ , where bulk critical exponents take their mean-field values and dangerous irrelevant variables modify the naive scaling picture.
Scaling implies data collapse: the equation of state becomes a single-variable function of scaling combinations such as $m/|t|^β$ vs $h/|t|^{βδ}$ — a stringent test.
Key idea: assume $f_s(t,h)=b^{-d}f_s(b^{y_t}t,b^{y_h}h)$ for arbitrary $b$ ; eliminating $b$ collapses all critical behavior onto two exponents, so every scaling relation follows from one homogeneity law.

The core insight

Assume the singular free-energy density obeys $f_s(t,h)=b^{-d}f_s(b^{y_t}t,b^{y_h}h)$ for arbitrary $b$ . Eliminating $b$ collapses all critical behavior onto two exponents $y_t,y_h$ , so every critical exponent — and hence every relation among them — follows from one homogeneity law.

Where this lives today

Finite-size scaling — the lattice-of-finite-size cousin of the scaling hypothesis, $f_s(t,h,L)=b^{-d}f_s(b^{y_t}t,b^{y_h}h,L/b)$ — is the workhorse of modern computational critical-phenomena studies. Setting $b=L$ gives $\chi=L^{\gamma/\nu}\tilde\chi(tL^{1/\nu})$ , and collapsing Monte Carlo data measured at several sizes $L$ extracts $T_c$ , ν and $\gamma/\nu$ to high precision. The same data-collapse logic underlies conformal-bootstrap determinations of 3D Ising exponents and the analysis of quantum critical points in cold-atom and quantum-critical systems, where a common scaling variable for a tuning parameter $r$ is $r/T^{1/(\nu z)}$ .

The physical picture

Near $T_c$ , the correlation length $\xi$ becomes much larger than microscopic lengths, and the system appears approximately scale invariant on length scales between the microscopic scale and $\xi$ . Scale invariance of the physics, not just of geometry, forces the free energy to be homogeneous — its shape repeats under simultaneous rescaling of length and of the couplings $t,h$ . The two figures below carry the whole idea: Kadanoff blocking that produces the homogeneity law, and the data-collapse it predicts.

Figure

Kadanoff blocking: b×b blocks of spins become effective coarse-grained degrees of freedom, while the scaling fields transform as t → bʸᵗ t and h → bʸʰ h.

Because the blocking can be iterated for any factor and the result is smooth, $b$ may be treated as a continuous free parameter. Choosing $b=|t|^{-1/y_t}$ removes the $t$ -dependence and exposes how every observable scales with $|t|$ ; choosing $b=h^{-1/y_h}$ does the same for the field. This freedom is exactly why two numbers suffice. A key testable consequence is data collapse: measuring the magnetization at many temperatures and fields and plotting $m/|t|^\beta$ against $h/|t|^{\beta\delta}$ collapses the data onto two scaling branches, corresponding to $t>0$ and $t<0$ , up to nonuniversal metric factors. Deviations quantify corrections to scaling.

Figure

Data collapse: raw M(t,h) curves at several temperatures on one side of T_c (left) collapse onto a single scaling branch M/|t|^β vs. h/|t|^{βδ} (right); including temperatures on both sides of T_c gives the two branches 𝓕₊ and 𝓕₋

The core equations

The whole edifice rests on one homogeneity law. It is a hypothesis about the singular part $f_s$ alone (the analytic background subtracted); everything else is calculus.

f_s(t,h)=b^{-d}\,f_s\!\left(b^{y_t}\,t,\ b^{y_h}\,h\right)\qquad\forall\,b>0 \qquad (1)

where: $f_s$ — singular part of the free-energy density; $t=(T-T_c)/T_c$ — reduced temperature; $h$ — ordering field conjugate to the order parameter; $b$ — arbitrary length-rescaling (block) factor; $d$ — spatial dimension; $y_t,y_h$ — thermal and magnetic RG scaling exponents (both positive, i.e. relevant). Physical reading: the free energy of the blocked system, per unit volume, is the same function with the couplings pushed further from (or nearer to) criticality. Anchoring the thermal argument by choosing $b=|t|^{-1/y_t}$ turns (1) into the Widom scaling form — a prefactor times a universal function of one scaling variable.

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

where: $\Phi_\pm$ — universal scaling functions for $t>0$ and $t<0$ ; $\Delta\equiv y_h/y_t$ is the gap exponent; the scaling relations derived below give $\Delta=\beta\delta=\beta+\gamma$ ; α — specific-heat exponent. Physical reading: all the temperature dependence lives in the prefactor $|t|^{2-\alpha}$ ; the field enters only through the single combination $h/|t|^{\Delta}$ . Each thermodynamic exponent is a derivative of $f_s$ evaluated at $h=0$ . This is where the two eigenvalues fan out into six exponents.

C_{\mathrm{sing}} \propto \left.\frac{\partial^2 f_s}{\partial t^2}\right|_{h=0} \sim |t|^{-\alpha},\quad m\sim\frac{\partial f_s}{\partial h}\Big|_{0}\sim|t|^{\beta},\quad \chi\sim\frac{\partial^2 f_s}{\partial h^2}\Big|_{0}\sim|t|^{-\gamma} \qquad (3)

where: $C$ — specific heat at $h=0$ ; $m$ — order parameter / spontaneous magnetization; χ — susceptibility $\partial m/\partial h$ . Physical reading: differentiate the scaling form, count powers of $|t|$ , and read off the exponent. Carrying out those derivatives on (2) gives the order-parameter exponents purely in terms of $y_t,y_h,d$ — the master result of the article.

\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 (4)

where: δ — critical-isotherm exponent ( $m\sim h^{1/\delta}$ at $t=0$ ); ν — correlation-length exponent ( $\xi\sim|t|^{-\nu}$ ); η — anomalous (Fisher) dimension. Physical reading: within the scaling and hyperscaling framework, six exponents are determined by two independent scaling dimensions and therefore satisfy four independent relations.

Derivation cards

Derivation AFrom the homogeneity law to the complete set of exponents

A single functional equation for $f_s$ , differentiated and with $b$ chosen freely, yields all six exponents in terms of $y_t,y_h,d$ and hence every scaling relation. The trick throughout is that $b$ is arbitrary, so we set it to whatever fixes one argument, then read the leading power of $|t|$ (or $h$ ) in each thermodynamic derivative. This card is the engine that turns one hypothesis into the entire exponent table.

Step 1. Fix the thermal argument by setting $b=|t|^{-1/y_t}$ in (1), giving $f_s(t,h)=|t|^{d/y_t}\,\Phi_\pm(h|t|^{-y_h/y_t})$ . Define $2-\alpha\equiv d/y_t$ ; since the correlation length rescales as $\xi\sim b\sim|t|^{-1/y_t}$ , we also read $\nu=1/y_t$ . Why this choice: it removes the arbitrary scale and leaves the temperature dependence as a clean prefactor times a function of one variable.

Step 2. Take two $t$ -derivatives at $h=0$ (using $\Phi_\pm(0)$ finite): $C\sim|t|^{d/y_t-2}=|t|^{-\alpha}$ , confirming $\alpha=2-d/y_t$ .

Step 3. One and two $h$ -derivatives bring down factors $|t|^{-y_h/y_t}$ : $m\sim|t|^{(d-y_h)/y_t}$ so $\beta=(d-y_h)/y_t$ , and $\chi\sim|t|^{(d-2y_h)/y_t}$ so $\gamma=(2y_h-d)/y_t$ .

Step 4. On the critical isotherm ( $t=0$ ) set $b=h^{-1/y_h}$ : $f_s\sim h^{d/y_h}$ so $m\sim h^{d/y_h-1}$ , giving $1/\delta=(d-y_h)/y_h$ , i.e. $\delta=y_h/(d-y_h)$ . Using the correlation-function result $\chi\sim\xi^{2-\eta}$ derived below, together with $\chi\sim|t|^{-\gamma}$ , $\xi\sim|t|^{-\nu}$ , $\gamma=(2y_h-d)/y_t$ , and $\nu=1/y_t$ , gives $\eta=d+2-2y_h$ .

Step 5. Substituting the boxed expressions and simplifying yields $\alpha+2\beta+\gamma=2$ (the $y_h$ terms cancel), $\gamma=\beta(\delta-1)$ , $\gamma=\nu(2-\eta)$ , and $2-\alpha=d\nu$ . Only the last contains $d$ , so only it depends on hyperscaling.

Go deeper · Derivation B · 1 minThe Fisher relation: susceptibility as an integral of correlations

Derivation BThe Fisher relation: susceptibility as an integral of correlations

The one relation that links thermodynamics to spatial correlations — Fisher's $\gamma=\nu(2-\eta)$ — comes from the fluctuation–dissipation identity that the susceptibility is the volume integral of the two-point correlation function, cut off at the correlation length. This card shows how the exponent η of the spatial decay controls the thermodynamic exponent γ, tying the correlation-function language to the free-energy language.

Step 1. By fluctuation–dissipation, $\chi\sim\int^{\xi}\!\mathrm{d}^d r\,G(r)$ with the critical correlation function $G(r)\sim r^{-(d-2+\eta)}$ for $r\ll\xi$ .

Step 2. The integrand times the volume element scales as $r^{d-1}\cdot r^{-(d-2+\eta)}=r^{1-\eta}$ ; for $\eta<2$ , the integral is dominated by its upper cutoff $r\sim\xi$ , giving $\chi\sim\xi^{2-\eta}$ . Why the upper cutoff dominates: the exponent $2-\eta>0$ means the integral grows with its ceiling, so the largest correlated scale ξ sets the answer.

Step 3. Insert $\xi\sim|t|^{-\nu}$ : $\chi\sim|t|^{-\nu(2-\eta)}$ . Comparing with $\chi\sim|t|^{-\gamma}$ gives $\gamma=\nu(2-\eta)$ .

Limiting cases

Case	Result	Meaning
Mean-field / Landau ( $d\ge d_c=4$ )	$\alpha=0,\ \beta=\tfrac12,\ \gamma=1,\ \delta=3,\ \nu=\tfrac12,\ \eta=0$	Rushbrooke and Widom hold ( $0+1+1=2$ , $1=\tfrac12\cdot2$ ); inserting the mean-field exponents into $2-\alpha=d\nu$ gives $2=d/2$ , satisfied only at $d=4$ ; above $d_c=4$ , conventional hyperscaling fails.
2D Ising (Onsager exact)	$\alpha=0\,(\log),\ \beta=\tfrac18,\ \gamma=\tfrac74,\ \delta=15,\ \nu=1,\ \eta=\tfrac14$	Satisfies all four exactly: $0+\tfrac14+\tfrac74=2$ ; $\tfrac74=\tfrac18(15-1)$ ; $\tfrac74=1\cdot(2-\tfrac14)$ ; $2-0=2\cdot1$ .
3D Ising (best estimates)	$\alpha\approx0.110,\ \beta\approx0.326,\ \gamma\approx1.237,\ \nu\approx0.630,\ \eta\approx0.036$	Hyperscaling $2-\alpha=3\nu$ : $1.890\approx1.890$ ; Rushbrooke $0.110+0.652+1.237=1.999$ . Relations hold within errors.
Spherical / $O(n\to\infty)$ , $2<d<4$	$\nu=\tfrac{1}{d-2},\ \eta=0,\ \gamma=\tfrac{2}{d-2}$	Exponents are $d$ -dependent yet obey hyperscaling exactly — a clean nontrivial check of Fisher and Josephson away from mean field.

Worked examples

Example 1 — Predict two exponents from two measurements. Suppose an experiment on a 3D uniaxial magnet returns $\beta=0.326$ and $\delta=4.79$ . Scaling lets you predict γ and α without further measurement.

Step 1. Widom gives $\gamma=\beta(\delta-1)=0.326\times(4.79-1)=0.326\times3.79=\mathbf{1.236}$ .

Step 2. Rushbrooke gives $\alpha=2-2\beta-\gamma=2-0.652-1.236=\mathbf{0.112}$ .

Step 3. Hyperscaling then cross-checks the correlation-length exponent: $\nu=(2-\alpha)/d=(2-0.112)/3=1.888/3=\mathbf{0.629}$ . All three match the accepted 3D-Ising values ( $\gamma\approx1.237$ , $\alpha\approx0.110$ , $\nu\approx0.630$ ) — a two-measurement prediction of the whole set.

Example 2 — Exponents from RG eigenvalues (2D Ising). The RG gives the eigenvalues $y_t=1/\nu=1$ and $y_h=(d+2-\eta)/2$ with $d=2,\eta=1/4$ , so $y_h=(2+2-1/4)/2=15/8$ . Recover γ, ν, η from (4).

Step 1. $\gamma=(2y_h-d)/y_t=(2\cdot\tfrac{15}{8}-2)/1=(\tfrac{30}{8}-2)=\tfrac{14}{8}=\mathbf{7/4}$ .

Step 2. $\nu=1/y_t=\mathbf{1}$ and $\eta=d+2-2y_h=2+2-\tfrac{30}{8}=4-\tfrac{15}{4}=\mathbf{1/4}$ .

Step 3. Check Fisher: $\nu(2-\eta)=1\cdot(2-\tfrac14)=\tfrac74=\gamma$ . Consistent — the eigenvalue route and the relation route agree exactly, as they must.

Origins · 1 minscaling hypothesis and how this was discovered

scaling hypothesis

formulated mid-1960s (Widom 1965, Kadanoff 1966, Fisher 1967) · The scaling hypothesis crystallized from three nearly simultaneous strands: Benjamin Widom's 1965 homogeneity ansatz for the equation of state, Leo Kadanoff's 1966 block-spin picture giving it a coarse-graining rationale, and Michael Fisher's synthesis of the exponent relations. It is the phenomenological statement that Kenneth Wilson's renormalization group would later derive from first principles, earning Wilson the 1982 Nobel Prize.

History

1944

— Onsager solves the zero-field 2D Ising model exactly, determining the critical temperature and logarithmic specific-heat singularity and demonstrating behavior beyond mean-field theory

1963

— Rushbrooke states the inequality α+2β+γ≥2 (later an equality under scaling)

1965

— Widom proposes the homogeneous equation of state, α+2β+γ=2 and γ=β(δ−1)

1966

— Kadanoff introduces block spins, giving scaling its coarse-graining picture

1967

— Fisher formalizes the scaling relations and the anomalous dimension η

1971

— Wilson's renormalization group derives scaling and computes the exponents

1982

— Wilson receives the Nobel Prize for the renormalization group

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 12 (Phase Transitions: Critical Phenomena) · Kardar, Statistical Physics of Fields, Ch. 4 (The Scaling Hypothesis) and Ch. 5 (RG) · Stanley, Introduction to Phase Transitions and Critical Phenomena, Ch. 11 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 12 · Goldenfeld, Lectures on Phase Transitions and the Renormalization Group, Ch. 9–10 · Widom, J. Chem. Phys. 43, 3898 (1965); Kadanoff, Physics 2, 263 (1966); Fisher, Rev. Mod. Phys. 46, 597 (1974). All four scaling relations can be verified by substituting the $(y_t,y_h,d)$ expressions from (4) and confirming the cancellations. Numerically: the 2D Ising exact values $\alpha=0,\beta=1/8,\gamma=7/4,\delta=15,\nu=1,\eta=1/4$ and the 3D Ising estimates $\alpha\approx0.110,\beta\approx0.326,\gamma\approx1.237,\nu\approx0.630,\eta\approx0.036$ (Pelissetto–Vicari, Phys. Rep. 368, 549 (2002)) each satisfy all four relations. Mean-field values $\alpha=0,\beta=1/2,\gamma=1,\delta=3,\nu=1/2,\eta=0$ satisfy Rushbrooke and Widom but satisfy hyperscaling only at $d=4$ .

Builds on: Critical Exponents · Continuous Phase Transitions · Correlation Length and Its Divergence · Landau Theory of Phase Transitions. Leads to: Universality and Universality Classes · Correlation Function · Order Parameter.

Is the scaling hypothesis derived or assumed?

Within thermodynamics it is an assumption — you postulate that $f_s$ is a generalized homogeneous function. Its justification comes from the renormalization group, where the homogeneity law is a consequence of the RG flow equation for the free energy and the exponents $y_t,y_h$ are eigenvalues of the flow linearized about a fixed point. So historically it was a phenomenological leap (Widom, Kadanoff) later proven by Wilson's RG. The article treats it as the structural statement the RG establishes.

Why does the free energy pick up a factor b⁻ᵈ under blocking?

Free energy is extensive: the total singular free energy of the system is invariant under an exact coarse-graining of the partition function, but the number of degrees of freedom drops by $b^d$ when you group $b^d$ sites into one block. Dividing the invariant total by the number of blocks gives a free-energy density that is $b^{-d}$ times the original. That volume factor is what makes hyperscaling — the relation involving $d$ — possible.

What is the difference between Rushbrooke and hyperscaling?

Rushbrooke ( $\alpha+2\beta+\gamma=2$ ) follows purely from differentiating the Widom scaling form; the dimension $d$ cancels, so it holds in every dimension, above and below $d_c$ . Hyperscaling ( $2-\alpha=d\nu$ ) needs the extra physical input that the singular free energy density is one unit of $k_BT$ per correlation volume $\xi^d$ . That assumption fails above $d_c=4$ , so hyperscaling fails there while Rushbrooke keeps holding. Testing whether hyperscaling holds is a way to locate $d_c$ .

How is δ different from the other exponents?

δ is the only exponent defined on the critical isotherm ( $t=0$ ) rather than along the $h=0$ axis. It measures how the order parameter responds to the field exactly at $T_c$ : $m\sim h^{1/\delta}$ . In the scaling form you access it by choosing $b=h^{-1/y_h}$ instead of $b=|t|^{-1/y_t}$ , which is why $\delta=y_h/(d-y_h)$ mixes $y_h$ and $d$ differently from β and γ. Consistency of the two choices is guaranteed by the single homogeneity law — that internal consistency is one of the hypothesis's strongest checks.

What does "data collapse" actually test?

It tests the scaling function $\Phi_\pm$ , not just the exponents. Two systems in the same universality class must collapse onto the same master curve (up to two non-universal scale factors) — a far more stringent statement than merely sharing exponents. In practice you plot $m/|t|^\beta$ versus $h/|t|^{\beta\delta}$ ; if the exponents are right but the curves still do not overlap, you are seeing corrections to scaling or a wrong background subtraction. Good collapse across many temperatures is the practical certificate that a scaling description applies.

Do the scaling relations hold as inequalities or equalities?

Historically Rushbrooke and Griffiths first appeared as thermodynamic inequalities (e.g. $\alpha+2\beta+\gamma\ge2$ ), provable from convexity of the free energy without any scaling assumption. The scaling hypothesis upgrades them to equalities. Experimentally the near-saturation of the inequalities was itself the first strong evidence for scaling. If a measured set of exponents violated an inequality, something would be badly wrong; if it satisfies the equality, scaling is confirmed to that precision.

Can I use scaling if my system is not a magnet?

Yes — scaling is about symmetry and dimensionality, not chemistry. The liquid–gas critical point (order parameter = density difference, field = chemical-potential offset), binary-alloy ordering, and the superfluid $\lambda$ -transition all obey the same homogeneity law with the appropriate $t$ and $h$ . Only the class (set by $d$ , order-parameter symmetry, interaction range) changes the exponent values; the structure — six exponents from two, four relations — is universal.