Scaling Hypothesis: Widom Scaling, Kadanoff Block Spins, and Data Collapse

Intuition Beginner

Near a critical point, something remarkable happens: the behaviour of a system depends on only one thing — how far you are from the critical temperature $T_c$. Everything — magnetisation, heat capacity, susceptibility, correlation length — follows power laws in $|T-T_c|$. And these power laws are not independent. If you measure one critical exponent, you can predict the others.

This is the scaling hypothesis, and it is one of the most powerful ideas in statistical physics. It says that near $T_c$, there is only one relevant length scale in the problem: the correlation length $\xi$, the typical distance over which fluctuations are correlated. Every other length in the system — the lattice spacing, the system size — is irrelevant as long as $\xi$ is much larger than both. All thermodynamic quantities are governed by $\xi$, and since $\xi$ diverges as a power of $|T-T_c|$, every observable inherits a power law with an exponent tied to the divergence of $\xi$.

The consequence is dramatic. Instead of six independent critical exponents ($\alpha$, $\beta$, $\gamma$, $\delta$, $\nu$, $\eta$), only two are independent. The remaining four are determined by scaling relations — algebraic identities among the exponents. For example:

$$\alpha +2\beta +\gamma =2,\gamma =\beta (\delta -1).$$

If you measure $\beta =0.326$ and $\gamma =1.237$ for the 3D Ising model (as numerical simulations do), the scaling relations predict $\delta =1+\gamma /\beta =1+1.237/0.326\approx 4.79$ and $\alpha =2-2\beta -\gamma =2-0.652-1.237\approx 0.111$. Both match the numerical results.

The scaling hypothesis makes a prediction that can be tested directly in experiment: data collapse. If you measure the magnetisation $m$ of a ferromagnet as a function of external field $h$ at several temperatures near $T_c$, the raw curves look different at each temperature. But if you rescale the axes — plot $m/|t{|}^{\beta }$ against $h/|t{|}^{\beta \delta }$, where $t=(T-T_c)/T_c$ is the reduced temperature — all the curves collapse onto a single universal function. This collapse is strong evidence that scaling is correct.

The physical picture behind scaling was developed by Kadanoff in 1966. Imagine a lattice of spins near $T_c$. Because the correlation length is very large, spins in a block of size $b$ (much smaller than $\xi$) are strongly correlated — they behave almost as a single unit. So you can replace each block of $b\times b\times \cdots \times b$ spins by a single "block spin." The new lattice has spacing $b$ times larger, but the same physics — provided you also rescale the temperature and the magnetic field. This block-spin construction is the intuitive foundation for the scaling hypothesis and, eventually, for the renormalization group.

Visual Beginner

The scaling hypothesis predicts that thermodynamic data measured at different temperatures near $T_c$ should collapse onto a single curve when plotted in rescaled variables.

Consider the magnetic equation of state $m=m(t,h)$ near $T_c$. Raw data at temperatures $T_1,T_2,T_3$ (all close to $T_c$) give three distinct curves of $m$ vs $h$. After rescaling: $m/|t{|}^{\beta }$ vs $h/|t{|}^{\beta \delta }$, the three curves collapse onto one master curve $f_{\pm }(x)$.

Worked example Beginner

The critical exponents for the 3D Ising model are (numerical estimates): $\beta \approx 0.326$, $\gamma \approx 1.237$.

Scaling relation check — Widom: $\gamma =\beta (\delta -1)$ predicts $\delta =1+\gamma /\beta =1+1.237/0.326\approx 4.79$.

Scaling relation check — Rushbrooke: $\alpha +2\beta +\gamma =2$ predicts $\alpha =2-2(0.326)-1.237=2-0.652-1.237=0.111$.

The numerically measured values are $\delta \approx 4.790$ and $\alpha \approx 0.110$. The agreement is excellent — the scaling relations are satisfied to the precision of the numerical data.

Check your understanding Beginner

Formal definition Intermediate+

The Widom scaling ansatz

The scaling hypothesis, introduced by Widom (1965), asserts that the singular part of the free energy near a critical point is a generalised homogeneous function of the reduced temperature $t=(T-T_c)/T_c$ and the ordering field $h$. That is, there exist exponents $y_t$ and $y_h$ such that:

$$f_s(t,h)=b^{-d}{f}_s(b^{y_t}t,b^{y_h}h)$$

for any rescaling factor $b>0$, where $d$ is the spatial dimension and $f_s$ is the singular part of the free energy density. This is the Widom scaling ansatz.

The physical content is that the free energy has no scale of its own near $T_c$ — it inherits its scale entirely from the correlation length $\xi \sim |t{|}^{-\nu }$, and rescaling lengths by a factor $b$ rescales $t$ and $h$ by the appropriate powers.

Choosing $b=|t{|}^{-1/y_t}$ (so that $b^{y_t}t=\pm 1$) gives:

$$f_s(t,h)=|t{|}^{d/y_t}{f}_s\left(\pm 1,\frac{h}{|t{|}^{y_h/y_t}}\right)=|t{|}^{2-\alpha }{\Phi }_{\pm }\left(\frac{h}{|t{|}^{\Delta }}\right),$$

where $\Delta =y_h/y_t$ is the gap exponent and ${\Phi }_{\pm }$ is the scaling function for $t>0$ and $t<0$ respectively. The identification $d/y_t=2-\alpha$ follows from the definition of the specific-heat exponent.

Scaling relations from the free energy

Differentiating $f_s$ gives the critical exponents:

Order parameter ($h=0$, $t<0$): $m=-(\partial f_s/\partial h{)}_{h=0}\sim |t{|}^{d/y_t-y_h/y_t}=|t{|}^{\beta }$, so:

$$\beta =\frac{d-y_h}{y_t}.$$

Susceptibility ($h=0$): $\chi =-({\partial }^2{f}_s/\partial h^2{)}_{h=0}\sim |t{|}^{d/y_t-2y_h/y_t}=|t{|}^{-\gamma }$, so:

$$\gamma =\frac{2y_h-d}{y_t}.$$

Critical isotherm ($t=0$): choose $b=h^{-1/y_h}$ in the scaling ansatz: $f_s(0,h)=h^{d/y_h}{f}_s(0,1)$, so $m=-\partial f_s/\partial h\sim h^{d/y_h-1}=h^{1/\delta }$, giving:

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

Specific heat ($h=0$): $C=-T{\partial }^2{f}_s/\partial t^2\sim |t{|}^{d/y_t-2}=|t{|}^{-\alpha }$, giving:

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

From these four expressions in terms of two unknowns ($y_t$, $y_h$, and $d$), two independent relations among the exponents follow immediately:

$$\alpha +2\beta +\gamma =2\text{(Rushbrooke)},$$ $$\gamma =\beta (\delta -1)\text{(Widom)}.$$

These were originally derived as inequalities (Rushbrooke 1963, Griffiths 1965) from thermodynamic stability. The scaling hypothesis promotes them to equalities.

The Kadanoff block-spin construction

Kadanoff (1966) provided a physical picture for scaling. Consider a $d$-dimensional lattice with lattice spacing $a$, near a critical point where $\xi \gg a$. Define block spins by grouping the spins in a block of linear size $b\cdot a$ ($b^d$ original spins per block) and assigning each block a single effective spin $S_I$ (by majority rule or some other coarse-graining rule).

The rescaled lattice has spacing $a^{\prime }=b\cdot a$. The claim is that the Hamiltonian of the block spins has the same form as the original Hamiltonian, but with rescaled couplings:

$$t^{\prime }=b^{y_t}t,h^{\prime }=b^{y_h}h,$$

where $y_t$ and $y_h$ are the relevant eigenvalues of the rescaling transformation. The word "relevant" means that these couplings grow under rescaling — if you start at $t\ne 0$ and rescale, $t^{\prime }$ moves away from zero. (Couplings that shrink are called irrelevant; they do not affect critical behaviour.)

The correlation length transforms as ${\xi }^{\prime }=\xi /b$, because lengths are measured in units of the lattice spacing, which has increased by $b$. Since $\xi \sim |t{|}^{-\nu }$ and ${\xi }^{\prime }\sim |t^{\prime }{|}^{-\nu }$:

$$\frac{\xi }{b}=|t^{\prime }{|}^{-\nu }=|b^{y_t}t{|}^{-\nu }=b^{-y_t\nu }|t{|}^{-\nu }=b^{-y_t\nu }\xi .$$

Dividing both sides by $\xi$ gives $1/b=b^{-y_t\nu }$, so:

$$\nu =\frac{1}{y_t}.$$

This is the correlation-length scaling relation. It connects the thermal eigenvalue $y_t$ to the correlation-length exponent $\nu$.

Data collapse as experimental test

The Widom scaling ansatz for the equation of state $m=m(t,h)$ reads:

$$m(t,h)=|t{|}^{\beta }{f}_{\pm }\left(\frac{h}{|t{|}^{\Delta }}\right),\Delta =\beta \delta ,$$

where $f_{+}$ applies for $t>0$ and $f_{-}$ for $t<0$. Equivalently:

$$\frac{m}{|t{|}^{\beta }}=f_{\pm }\left(\frac{h}{|t{|}^{\beta \delta }}\right).$$

If you plot $m/|t{|}^{\beta }$ against $h/|t{|}^{\beta \delta }$ for data taken at several temperatures near $T_c$, all points should lie on a single curve $f_{\pm }$. This is data collapse, and it is the most direct experimental test of the scaling hypothesis.

The quality of the collapse depends on: (1) being close enough to $T_c$ that corrections to scaling (higher-order terms) are negligible; (2) knowing the exponents $\beta$ and $\delta$ accurately; (3) having data that span a wide enough range of $h/|t{|}^{\beta \delta }$ to reveal the shape of $f_{\pm }$. The method was used extensively in the 1960s and 1970s to determine critical exponents for fluids, magnets, and binary mixtures.

Summary of scaling relations

The six standard critical exponents are related by four scaling relations. All follow from the Widom ansatz with two independent exponents:

Relation	Formula	Type
Rushbrooke	$\alpha +2\beta +\gamma =2$	Thermodynamic
Widom	$\gamma =\beta (\delta -1)$	Thermodynamic
Fisher	$\gamma =\nu (2-\eta )$	Correlation
Josephson (hyperscaling)	$2-\alpha =d\nu$	Dimension-dependent

The first two are "thermodynamic" — they involve only exponents that can be measured from bulk thermodynamic data ($C$, $m$, $\chi$, critical isotherm). The third involves the correlation-function exponent $\eta$. The fourth involves the spatial dimension $d$ and is called hyperscaling because it depends on the dimensionality of space.

Rushbrooke and Griffiths inequalities as equalities

Before the scaling hypothesis, Rushbrooke (1963) proved the inequality $\alpha +2\beta +\gamma \ge 2$ from thermodynamic stability alone (no microscopic model needed). Griffiths (1965) proved $\alpha +\beta (1+\delta )\ge 2$. These are rigorous thermodynamic results. The scaling hypothesis predicts that these inequalities are in fact equalities — the $\ge$ becomes $=$. This prediction has been confirmed by numerical simulations and experiments to within the precision of the data.

The conceptual point is important: thermodynamics constrains the exponents but does not fully determine them. The scaling hypothesis (and, more fundamentally, the renormalization group) provides the additional physical input needed to saturate the inequalities.

Key derivation: All scaling relations from two exponents Intermediate+

Proposition. The Widom scaling ansatz $f_s(t,h)=b^{-d}{f}_s(b^{y_t}t,b^{y_h}h)$, together with the definitions of the critical exponents and the correlation-length exponent $\nu =1/y_t$, implies the four scaling relations $\alpha +2\beta +\gamma =2$, $\gamma =\beta (\delta -1)$, $\gamma =\nu (2-\eta )$, and $2-\alpha =d\nu$.

Proof. From the Widom ansatz, the four thermodynamic exponents are:

$$\alpha =2-d/y_t,\beta =(d-y_h)/y_t,\gamma =(2y_h-d)/y_t,\delta =y_h/(d-y_h).$$

Rushbrooke: $\alpha +2\beta +\gamma =(2-d/y_t)+2(d-y_h)/y_t+(2y_h-d)/y_t=2-d/y_t+2d/y_t-2y_h/y_t+2y_h/y_t-d/y_t=2$.

Widom: $\beta (\delta -1)=(d-y_h)/y_t\cdot [y_h/(d-y_h)-1]=(d-y_h)/y_t\cdot (2y_h-d)/(d-y_h)=(2y_h-d)/y_t=\gamma$.

Josephson (hyperscaling): $2-\alpha =d/y_t=d\nu$, using $\nu =1/y_t$.

Fisher: The correlation function at $T_c$ is $G(r)\sim r^{-(d-2+\eta )}$. The susceptibility is $\chi =\int d^{d}rG(r)\sim |t{|}^{-\gamma }$. At $t=0$, $G(r)\sim r^{-(d-2+\eta )}$ and $\chi$ diverges because the integral has a long-range tail. At $t\ne 0$, the correlation length cuts off the integral at $r\sim \xi$, giving $\chi \sim {\xi }^{2-\eta }\sim |t{|}^{-\nu (2-\eta )}$. Comparing with $\chi \sim |t{|}^{-\gamma }$ yields $\gamma =\nu (2-\eta )$. $\square$

Bridge. The Kadanoff block-spin picture gives the same relations from a different starting point: instead of postulating a homogeneous free energy, you postulate that coarse-graining by a factor $b$ rescales the couplings as $t^{\prime }=b^{y_t}t$, $h^{\prime }=b^{y_h}h$. The renormalization group 11.07.01 then derives the values of $y_t$ and $y_h$ from the eigenvalues of the linearised RG transformation at the fixed point, making the Kadanoff construction into a systematic calculational tool. The Wilson-Fisher fixed point 11.07.02 gives $y_t$ and $y_h$ as functions of $d$ and $n$, reproducing the correct critical exponents for all dimensions $2<d<4$.

Exercises Intermediate+

Lean formalization Intermediate+

Formalizing the scaling hypothesis requires: (1) a homogeneous-function framework with two continuous parameters $y_t$, $y_h$ and an integer dimension $d$; (2) a proof that differentiation of a generalized homogeneous function yields power-law observables with exponents expressed as rational combinations of $y_t$, $y_h$, and $d$; (3) an algebraic verification that the four scaling relations (Rushbrooke, Widom, Fisher, Josephson) hold identically given the expressions for the exponents. Parts (1) and (3) are within Mathlib's current scope (real analysis, field arithmetic). Part (2) requires a domain-specific construction for thermodynamic derivatives and their asymptotic scaling. The gap is substantial but well-scoped.

Advanced results Master

Derivation of all scaling relations from $y_t$ and $y_h$

The complete set of critical exponents, expressed in terms of the two independent scaling eigenvalues $y_t$ and $y_h$ and the dimension $d$:

$$\alpha =2-\frac{d}{y_t},\beta =\frac{d-y_h}{y_t},\gamma =\frac{2y_h-d}{y_t},\delta =\frac{y_h}{d-y_h},\nu =\frac{1}{y_t},\eta =2+d-2y_h.$$

The last expression for $\eta$ follows from the Fisher relation $\gamma =\nu (2-\eta )$ combined with the expression for $\gamma$ and $\nu$: $(2y_h-d)/y_t=(1/y_t)(2-\eta )$, giving $\eta =2+d-2y_h$.

The mean-field (Landau) exponents correspond to $y_t=2$ and $y_h=3d/2$ (for $d\ge 4$). The 2D Ising exponents ($\alpha =0$, $\beta =1/8$, $\gamma =7/4$, $\nu =1$, $\eta =1/4$) correspond to $y_t=1$ and $y_h=15/8$.

Hyperscaling and the Josephson relation

The hyperscaling relation $2-\alpha =d\nu$ is special because it is the only scaling relation that involves the spatial dimension $d$. Its physical origin is that the singular part of the free energy scales as the number of correlated volumes: $f_s\sim {\xi }^{-d}\sim |t{|}^{d\nu }$, and also $f_s\sim |t{|}^{2-\alpha }$, giving $d\nu =2-\alpha$.

Hyperscaling holds for short-range interacting systems in $d<d_c=4$ (for the Ising universality class). It has been verified numerically for $d=2$ and $d=3$ Ising models.

Breakdown of hyperscaling above $d=4$

For $d>4$, the mean-field exponents are exact: $\nu =1/2$, $\alpha =0$. Hyperscaling would predict $2-\alpha =d/2$, which gives $2=d/2$ or $d=4$. For $d>4$, $d\nu =d/2>2=2-\alpha$, so hyperscaling fails.

The physical reason is that above the upper critical dimension $d_c=4$, the dangerous irrelevant variable $\lambda$ (the quartic coupling in the Landau-Ginzburg Hamiltonian) modifies the free-energy scaling. The quartic coupling flows to zero under the RG (it is irrelevant), but it cannot be simply set to zero because the free energy is singular at $\lambda =0$. The rescaling of $f_s$ involves an extra factor from $\lambda$:

$$f_s(t,h)=b^{-d}{f}_s(b^{y_t}t,b^{y_h}h,b^{y_{\lambda }}\lambda ),$$

where $y_{\lambda }<0$ (irrelevant). Setting $\lambda$ to its fixed-point value ${\lambda }^{\ast }$ is invalid because $f_s$ depends on $\lambda$ in a singular way. The correct procedure is to evaluate $\lambda$ at its initial (microscopic) value and carry it through the scaling analysis. This produces a modified scaling form that reproduces the mean-field exponents and explains why hyperscaling breaks down.

This subtlety — the dangerous irrelevant variable — was clarified by Fisher, Ma, and Nickel (1972). It shows that scaling is not simply about discarding irrelevant variables; some irrelevant variables must be retained because the quantity of interest (here, the free energy) is non-analytic in them.

Finite-size scaling

Near a critical point in an infinite system, the correlation length $\xi \sim |t{|}^{-\nu }$ is the only relevant length scale. In a finite system of linear size $L$, $L$ provides a cutoff: $\xi$ cannot exceed $L$. The finite-size scaling hypothesis (Fisher, 1971) asserts that the free energy has the scaling form:

$$f_s(t,h,L)=L^{-d}{f}_s(L^{y_t}t,L^{y_h}h,1).$$

At the critical point ($t=0$, $h=0$): $f_s\sim L^{-d}$, and the specific heat $C\sim L^{\alpha /\nu }$, the magnetisation $m\sim L^{-\beta /\nu }$, and the susceptibility $\chi \sim L^{\gamma /\nu }$.

The finite-size scaling form provides a practical method for extracting critical exponents from numerical simulations. You simulate systems of different sizes $L$ at or near $T_c$ and measure how observables scale with $L$. The ratios $\beta /\nu$, $\gamma /\nu$, and $\nu$ itself (from the scaling of the Binder cumulant or from the shift of the pseudo-critical temperature) give the critical exponents. This method was essential for the high-precision determination of 3D Ising exponents.

The crossover between the finite-size regime ($\xi \sim L$) and the thermodynamic regime ($\xi \ll L$) occurs at $|t|\sim L^{-1/\nu }$. This defines the finite-size scaling window: for $|t|\gg L^{-1/\nu }$, finite-size effects are negligible; for $|t|\lesssim L^{-1/\nu }$, the system feels its finite extent.

Fisher's scaling law

The Fisher scaling relation $\gamma =\nu (2-\eta )$ connects the susceptibility exponent $\gamma$ to the correlation-length exponent $\nu$ and the correlation-function exponent $\eta$. Its physical origin is in the fluctuation-dissipation theorem, which relates the susceptibility to the integral of the correlation function:

$$\chi =\frac{1}{k_{B}T}\int d^{d}rG(r).$$

At $t\ne 0$, the correlation function is $G(r)\sim r^{-(d-2+\eta )}{e}^{-r/\xi }$. The integral gives $\chi \sim {\xi }^{2-\eta }$ (the power of $\xi$ from the exponential cutoff, corrected by the algebraic prefactor). Substituting $\xi \sim |t{|}^{-\nu }$ and $\chi \sim |t{|}^{-\gamma }$ yields $\gamma =\nu (2-\eta )$.

At $t=0$, the exponential cutoff is absent and $G(r)\sim r^{-(d-2+\eta )}$. The susceptibility diverges (for $d>2$, $\eta >0$ is needed for convergence), which is consistent with $\gamma >0$.

Historical development

The scaling hypothesis emerged in the mid-1960s from the convergence of several lines of investigation. Widom (1965) proposed the scaling ansatz for the equation of state based on the observation that experimental data for fluids near the critical point could be described by a single scaling function. Independently, Domb and Hunter (1965) analysed high-temperature series expansions for the Ising model and found that the exponents satisfied algebraic relations. Kadanoff (1966) provided the physical picture (block spins) that motivated the scaling hypothesis and pointed the way toward the renormalization group.

The scaling hypothesis was a major conceptual advance: it reduced the description of critical phenomena from a collection of independent exponents to a two-parameter family. But it could not predict the values of the exponents — it could only relate them. The renormalization group (Wilson, 1971) provided the missing ingredient: a systematic method for computing $y_t$ and $y_h$ from the microscopic Hamiltonian.

The passage from scaling to the renormalization group is the central conceptual arc of late-twentieth-century statistical physics. Kadanoff's block-spin construction was the physical intuition; Wilson's RG was the mathematical realisation. The scaling functions $f_{\pm }$ that Widom postulated are the renormalization-group eigenfunctions at the critical fixed point. The exponents $y_t$ and $y_h$ are the relevant eigenvalues of the RG linearisation. Scaling is the kinematics; the RG provides the dynamics.

Synthesis. The scaling hypothesis sits between Landau theory 11.06.03 and the renormalization group 11.07.01. Landau theory gives specific (mean-field) exponents; the scaling hypothesis shows that whatever the exponents are, only two are independent; the renormalization group computes those two from first principles. The Ginzburg criterion 11.06.04 tells you when Landau theory is sufficient ($d>4$); the scaling hypothesis tells you how the exponents are related in all dimensions; the RG tells you what the exponents actually are. Data collapse is the experimental signature that validates this entire framework.

Full proof set Master

Proposition. The Kadanoff block-spin rescaling, iterated $n$ times on a system with singular free energy $f_s(t,h)$, yields the Widom scaling form $f_s(t,h)=|t{|}^{2-\alpha }{\Phi }_{\pm }(h/|t{|}^{\Delta })$ with $\Delta =\beta \delta$.

Proof. Under a single block-spin transformation by factor $b$, the free energy per original site transforms as:

$$f_s(t,h)=b^{-d}{f}_s(b^{y_t}t,b^{y_h}h).$$

Iterating $n$ times:

$$f_s(t,h)=b^{-nd}{f}_s(b^{n{y}_t}t,b^{n{y}_h}h).$$

Choose $n$ so that $b^{n{y}_t}|t|=1$, i.e., $b^n=|t{|}^{-1/y_t}$. This is always possible for $t\ne 0$. Then:

$$f_s(t,h)=|t{|}^{d/y_t}{f}_s\left(\pm 1,\frac{h}{|t{|}^{y_h/y_t}}\right).$$

Define $\Delta =y_h/y_t$ and ${\Phi }_{\pm }(x)=f_s(\pm 1,x)$. Using $d/y_t=d\nu =2-\alpha$:

$$f_s(t,h)=|t{|}^{2-\alpha }{\Phi }_{\pm }\left(\frac{h}{|t{|}^{\Delta }}\right).$$

Differentiating twice with respect to $t$ (at $h=0$) recovers the specific heat: $C\sim |t{|}^{-\alpha }$. Differentiating once with respect to $h$: $m=|t{|}^{2-\alpha -\Delta }{\Phi }_{\pm }^{\prime }(h/|t{|}^{\Delta })$. At $h=0$: $m\sim |t{|}^{2-\alpha -\Delta }$. Comparing with $m\sim |t{|}^{\beta }$ gives $\beta =2-\alpha -\Delta$, so $\Delta =2-\alpha -\beta =\beta +\gamma =\beta \delta$ (using Rushbrooke and Widom relations). $\square$

Proposition. The hyperscaling relation $2-\alpha =d\nu$ holds for short-range systems in $d<d_c=4$, and breaks down for $d>d_c$ when dangerous irrelevant variables are present.

Proof. For $d<d_c$: the singular free energy scales as one correlated degree of freedom per correlation volume, $f_s\sim k_{B}T{\xi }^{-d}\sim |t{|}^{d\nu }$. Since $f_s\sim |t{|}^{2-\alpha }$ by definition of $\alpha$: $d\nu =2-\alpha$.

For $d>d_c$: the quartic coupling $u$ in the Landau-Ginzburg Hamiltonian is a dangerous irrelevant variable. Under RG rescaling by factor $b$: $u^{\prime }=b^{y_u}u$ with $y_u=4-d<0$ (irrelevant). The free energy has a singularity at $u=0$ (the Gaussian fixed point gives ${\alpha }_{\mathrm{Gauss}}=2-d/2$, but the true $\alpha =0$ for $d>4$). Retaining $u$ in the scaling form:

$$f_s(t,h,u)=b^{-d}{f}_s(b^{y_t}t,b^{y_h}h,b^{y_u}u).$$

Setting $b^{y_t}|t|=1$ and evaluating $u$ at its fixed-point value $\bar{u}$: the $u$-dependence of $f_s$ introduces an extra factor $|t{|}^{-y_u/y_t\cdot \theta }$ where $\theta$ depends on the singular dependence of $f_s$ on $u$. For the Landau-Ginzburg Hamiltonian, $\theta =d/2$ (the quartic coupling scales as $u\sim |t{|}^{(4-d)/2}$ near the Wilson-Fisher fixed point). The resulting correction modifies $d\nu$ to $d\nu -(d-4)\nu =4\nu =2$, giving $2-\alpha =2$ and $\alpha =0$ (mean-field), which violates $d\nu =2-\alpha$ for $d>4$. $\square$

Connections Master

• 11.06.02 Mean-field theory (Curie-Weiss, van der Waals) provides specific values for the critical exponents. The scaling hypothesis shows that these exponents satisfy the scaling relations (as verified in Exercise 7 of 11.06.03) — this is a consistency check on mean-field theory.
• 11.06.03 Landau theory predicts mean-field exponents. The scaling hypothesis is agnostic about the values of the exponents — it only requires that they satisfy the scaling relations. Landau theory satisfies scaling by construction (all exponents are derived from $y_t=2$, $y_h=3d/2$).
• 11.06.04 The Ginzburg criterion determines the regime of validity of mean-field theory. Within that regime, the mean-field exponents satisfy the scaling relations. Outside that regime (near $T_c$ in $d<4$), the scaling relations still hold but the exponents are different.
• 11.07.01 The renormalization group provides the microscopic foundation for the scaling hypothesis. The Kadanoff block-spin rescaling becomes the RG transformation; the exponents $y_t$, $y_h$ become the eigenvalues of the RG linearisation at the fixed point; the scaling functions are determined by the RG flow.
• 11.07.02 Block-spin renormalization of the Landau-Ginzburg Hamiltonian implements the Kadanoff construction explicitly. The RG recursion relations yield $y_t$ and $y_h$ as functions of $d$, $n$, and the coupling constants, providing the critical exponents from first principles.

Historical and philosophical context Master

The scaling hypothesis emerged from a period of intense activity in the study of critical phenomena during the 1960s. By the early 1960s, high-temperature series expansions (developed by Domb, Sykes, Fisher, and others at King's College London) had produced accurate numerical estimates of critical exponents for the 3D Ising model. These estimates did not match the mean-field (Landau) predictions. At the same time, precise experiments on fluids (by Voronel, Levelt Sengers, and others) and magnets (by Kouvel, Comly, and others) revealed that the critical exponents were universal — the same for systems with the same symmetry — but different from mean-field values.

Benjamin Widom at Cornell proposed the scaling ansatz for the equation of state in 1965. His insight was that the equation of state near the critical point should have a universal, homogeneous form: if you rescale the reduced temperature and the field by the right powers, all isotherms should collapse onto a single curve. This was partly motivated by the "law of corresponding states" in van der Waals theory, but Widom's insight was that the rescaling exponents were the critical exponents themselves, not the mean-field values.

Leo Kadanoff, then at the University of Illinois, provided the physical picture in 1966. His block-spin construction was a conceptual breakthrough: it showed that the scaling hypothesis followed from the assumption that the correlation length is the only relevant length scale near $T_c$. If you coarse-grain by a factor smaller than $\xi$, the physics is unchanged up to a rescaling of the couplings. This argument was not rigorous (it assumed that the block-spin Hamiltonian has the same form as the original Hamiltonian, which is not exactly true), but it captured the essential physics.

Kadanoff's 1966 paper ends with a remarkably prescient statement: "it is natural to think of the cell-to-cell transformation as a mathematical operation defined upon the set of all Hamiltonians." This sentence contains the seed of the renormalization group — the idea that the rescaling transformation acts on a space of Hamiltonians, and that the fixed points of this transformation determine the critical behaviour. Wilson's 1971 renormalization group was the full realisation of this idea.

The Domb-Hunter program (C. Domb and D. L. Hunter at King's College London, 1965) provided independent evidence for scaling from the analysis of series expansions. By computing the ratios of successive coefficients in high-temperature series and extrapolating, they determined critical exponents with sufficient precision to verify the scaling relations. The convergence of the experimental (Widom), theoretical (Kadanoff), and numerical (Domb-Hunter) evidence for scaling was one of the great episodes in the history of statistical physics.

The scaling hypothesis raised a deep question: what determines the two independent exponents? The renormalization group answered this question by showing that the exponents are eigenvalues of the RG transformation at the critical fixed point, and can be computed systematically (via the epsilon expansion, the high-temperature expansion, or Monte Carlo methods). The scaling hypothesis thus occupies a unique position in the logical structure of critical phenomena: it is the bridge between the phenomenology (power laws, data collapse) and the microscopic theory (the RG).

Bibliography Master

@article{widom1965,
  author = {Widom, Benjamin},
  title = {Equation of state in the critical region. {II}},
  journal = {J. Chem. Phys.},
  volume = {43},
  pages = {3898--3905},
  year = {1965}
}

@article{kadanoff1966,
  author = {Kadanoff, Leo P.},
  title = {Scaling laws for {I}sing models near {$T_c$}},
  journal = {Physics},
  volume = {2},
  pages = {263--272},
  year = {1966}
}

@article{rushbrooke1963,
  author = {Rushbrooke, George S.},
  title = {On the thermodynamics of the critical region for the {I}sing problem},
  journal = {J. Chem. Phys.},
  volume = {39},
  pages = {842--843},
  year = {1963}
}

@article{griffiths1965,
  author = {Griffiths, Robert B.},
  title = {Thermodynamic inequality near the critical point for the lattice gas and the {I}sing model},
  journal = {Phys. Rev. Lett.},
  volume = {14},
  pages = {623--624},
  year = {1965}
}

@article{dombhunter1965,
  author = {Domb, Cyril and Hunter, David L.},
  title = {On the critical behaviour of lattice gas and the {I}sing model},
  journal = {Proc. Phys. Soc.},
  volume = {86},
  pages = {1107--1116},
  year = {1965}
}

@article{fisher1971fss,
  author = {Fisher, Michael E.},
  title = {The theory of critical point singularities},
  journal = {Proc. Enrico Fermi Int. Sch. Phys.},
  volume = {51},
  pages = {1--99},
  year = {1971}
}

@article{fishermanickel1972,
  author = {Fisher, Michael E. and Ma, Shang-keng and Nickel, B. G.},
  title = {Critical exponents for long-range interactions},
  journal = {Phys. Rev. Lett.},
  volume = {29},
  pages = {917--920},
  year = {1972}
}

@article{wilson1971,
  author = {Wilson, Kenneth G.},
  title = {Renormalization group and critical phenomena. {I}. {R}enormalization group and the {K}adanoff scaling picture},
  journal = {Phys. Rev. B},
  volume = {4},
  pages = {3174--3183},
  year = {1971}
}

@book{kardar2007,
  author = {Kardar, Mehran},
  title = {Statistical Physics of Fields},
  publisher = {Cambridge University Press},
  year = {2007}
}

@book{pathria-beale2011,
  author = {Pathria, R. K. and Beale, Paul D.},
  title = {Statistical Mechanics},
  edition = {3rd},
  publisher = {Academic Press},
  year = {2011}
}

@book{goldenfeld1992,
  author = {Goldenfeld, Nigel},
  title = {Lectures on Phase Transitions and the Renormalization Group},
  publisher = {Addison-Wesley},
  year = {1992}
}