Finite-Size Scaling and the RG

A bulk critical point is defined by a diverging correlation length $\xi$ and the power-law singularities that ride on it. But any real or simulated system is finite, and a finite sum of Boltzmann factors over a finite configuration space is a perfectly analytic function of temperature and field — it has no singular points at all. There can be no genuine divergence in a finite box. And yet experiments on thin films and, above all, Monte Carlo simulations show sharp, growing, systematically shifting peaks that look every bit like the onset of a transition. Finite-size scaling is the quantitative theory of how the infinite-volume singularities are rounded and shifted by the finite size $L$ , and the renormalization group is what makes it universal. The key conceptual move, due to Fisher and to Fisher and Barber, is that the correlation length cannot exceed the system size. When the bulk $\xi(t)$ would diverge, in a finite box it saturates at a value of order $L$ , so the dimensionless ratio $L/\xi$ replaces $\xi$ itself as the controlling variable. Within the RG this is formalized by adding $L^{-1}$ to the list of scaling fields. Under a coarse-graining step by factor $b$ , a fixed physical box of $L/a$ cells becomes $L/(ba)$ cells: $L\to L/b$ , i.e. $L^{-1}\to bL^{-1}$ . The inverse size therefore scales with the trivial geometric eigenvalue $y_L=1$ — the same rescaling that defines a length. Combining $L^{-1}$ (eigenvalue $1$ ) with the thermal field $t$ (eigenvalue $y_t=1/\nu$ ) and the field $h$ (eigenvalue $y_h$ ) in the RG homogeneity relation produces the finite-size scaling ansatz. Choosing $b=L$ removes the explicit size dependence and leaves universal functions of $L^{1/\nu}t$ and $L^{y_h}h$ . Every observable's finite-size behaviour then follows by differentiation: the peak of the susceptibility grows as $L^{\gamma/\nu}$ , the order parameter at $T_c$ falls as $L^{-\beta/\nu}$ , the pseudo-critical temperature shifts as $L^{-1/\nu}$ , and the Binder cumulant becomes size-independent so its curves cross at $T_c$ . This is not a nuisance to be eliminated — it is the single most powerful practical tool for measuring exponents and locating critical points from finite data.

Key takeaways

In a finite system of size $L$ true singularities are impossible: the partition function is a finite sum of analytic terms, so peaks are rounded rather than divergent.
The RG cures the puzzle by treating $L^{-1}$ as a scaling variable with eigenvalue $y_L=1$ : coarse-graining shrinks $L$ just as it grows $t$ and $h$ .
The central result is the FSS form: every singular quantity equals a power of $L$ times a universal function of $L/\xi\sim L\,t^{\nu}$ , so all critical physics is governed by the ratio of system size to correlation length.
Peaks shift from bulk $T_c$ by $T_c(L)-T_c\sim L^{-1/\nu}$ and round over a width $\sim L^{-1/\nu}$ ; the susceptibility peak grows as $L^{\gamma/\nu}$ , the specific heat as $L^{\alpha/\nu}$ .
At $T_c$ the order parameter scales as $\langle|m|\rangle\sim L^{-\beta/\nu}$ and the Binder cumulant becomes size-independent — the sharpest way to locate $T_c$ and measure $\nu$ .
Above the upper critical dimension $d>4$ hyperscaling fails and FSS is modified: the relevant length is $L^{d/2}$ , the Gaussian finite-size scaling regime.
Key idea: Add $L^{-1}$ to the RG scaling fields with eigenvalue $y_L=1$ ; the homogeneity of the singular free energy then forces every observable to be a power of $L$ times a universal function of $L^{1/\nu}t$ , so singularities are rounded over $\Delta t\sim L^{-1/\nu}$ , shifted by $\sim L^{-1/\nu}$ , with peaks growing as $L^{\gamma/\nu}$ , $L^{\alpha/\nu}$ .

The core insight

Add the inverse system size $L^{-1}$ to the RG scaling fields with eigenvalue $y_L=1$ ; the homogeneity of the singular free energy then forces every observable to be a power of $L$ times a universal function of the single combination $L^{1/\nu}t$ (equivalently $L/\xi$ ). Singularities are rounded over a window $\Delta t\sim L^{-1/\nu}$ and shifted by $\sim L^{-1/\nu}$ , with peak amplitudes growing as $L^{\gamma/\nu}$ , $L^{\alpha/\nu}$ .

Where this lives today

Finite-size scaling is the workhorse of modern computational statistical mechanics. A study of the three-dimensional Ising model runs lattices of side $L=16,32,64,128,\dots$ , plots the Binder cumulant to fix $T_c$ from the exponent-free crossing, then collapses $L^{-\gamma/\nu}\chi$ and $L^{\beta/\nu}\langle|m|\rangle$ against $tL^{1/\nu}$ to extract $\nu=0.6300(4)$ and the exponent ratios. The slope of the cumulant at the crossing, growing as $L^{1/\nu}$ , gives an independent handle on $\nu$ . This is how the best classical estimates of critical exponents — competitive with the conformal bootstrap — are obtained, entirely from finite systems. The reach extends into real materials through the critical Casimir effect. A fluid confined between plates a distance $L$ apart, tuned near its bulk critical point, feels a universal force whose amplitude is fixed by the universality class and the boundary conditions — the thermodynamic analog of the electromagnetic Casimir force, but tunable with temperature. Measured in binary liquid mixtures and near the superfluid transition of helium films, these forces are a direct experimental confirmation of finite-size scaling functions, not just exponents, closing the loop between the RG homogeneity relation and a force you can measure with a colloid.

The physical picture

A finite box is a low-pass filter on correlations. The bulk correlation length wants to grow without bound as $T\to T_c$ , but it physically cannot exceed the box: $\xi$ saturates at $\sim L$ . So the right question is never "how big is $\xi$ ?" but "is $\xi$ bigger or smaller than $L$ ?" — and the only dimensionless answer is the ratio $L/\xi$ , the master variable of finite-size scaling. Everything else is bookkeeping on that ratio. Coarse-graining literally eats the system. Each RG step with factor $b$ replaces blocks of $b^d$ degrees of freedom by one, so a box of $N=(L/a)^d$ sites becomes $(L/ba)^d$ sites: $L$ shrinks to $L/b$ . Hence $L^{-1}$ grows by $b$ under the RG — it behaves exactly like a relevant field with eigenvalue $1$ , marching the finite system away from the critical fixed point at the same geometric rate that distances rescale. There is nothing mysterious about $y_L=1$ ; it is just the dimension of an inverse length. Finite-size scaling exhibits an especially strong form of universality. Not only the exponents but the entire scaling functions of $L^{1/\nu}t$ are universal — once you fix the universality class, the spatial dimension, the aspect ratio of the box, and the boundary conditions (periodic, free, antiperiodic). This is why data collapse works: rescale the axes by the right powers of $L$ and curves from $L=8,16,32,64$ fall on top of one another. The scaling exponents are universal, but the scaling functions depend on the geometry and boundary conditions. For example, periodic and free boundaries belong to the same universality class yet have different scaling functions. Think of $T_c(L)$ , the temperature where the finite susceptibility peaks, as a pseudo-critical point. It is a smooth, well-defined feature of an analytic function, drifting toward the true $T_c$ as $L^{-1/\nu}$ . The peak grows as $L^{\gamma/\nu}$ and narrows as $L^{-1/\nu}$ . Finite-size scaling describes how thermodynamic singularities are rounded and shifted in finite systems. The resulting finite-size dependence provides a practical route to measuring critical exponents.

Figure

The correlation length saturates at the box: ξ cannot exceed L

The figure shows the mechanism: the bulk correlation length (dashed) would blow up at $t=0$ , but in a finite box it flattens off at $\xi\sim L$ — a bigger box saturates at a higher plateau. That saturation is why a finite system rounds the transition instead of showing a true divergence.

Figure

Susceptibility peak: height grows as L^(γ/ν), position shifts as L^(−1/ν)

The core equations

The starting point is the RG homogeneity relation for the singular free-energy density, extended to include the inverse system size $L^{-1}$ as an additional scaling field with RG eigenvalue $y_L=1$ . The standard finite-size scaling relations follow from this equation.

f_s(t,h,L^{-1}) = b^{-d}\, f_s\!\big(b^{y_t}t,\; b^{y_h}h,\; b\,L^{-1}\big) \qquad (1)

where: $f_s$ — singular free energy per unit volume; $t$ — reduced temperature $(T-T_c)/T_c$ ; $h$ — reduced field; $L^{-1}$ — inverse linear size, a relevant scaling field; $b$ — RG factor ( $b>1$ ); $y_t=1/\nu$ , $y_h$ — thermal and magnetic eigenvalues; $d$ — spatial dimension. The third argument scales as $b\,L^{-1}$ because an inverse length has RG dimension $1$ . Because $b$ is free, choose $b=L$ . The explicit size dependence disappears, leaving a universal function of two scaling variables — the finite-size scaling ansatz that every collapse plot verifies.

f_s(t,h,L^{-1}) = L^{-d}\,\Psi\!\big(L^{1/\nu}t,\; L^{y_h}h\big) \qquad (2)

where: $\Psi$ — universal scaling function (fixed by class, dimension, geometry, and boundary conditions); $L^{1/\nu}t\sim(L/\xi)^{1/\nu}$ — the thermal scaling variable; $L^{y_h}h$ — the field scaling variable. Setting $b=L$ makes $b\,L^{-1}=1$ disappear as an argument. The physical origin of the whole subject is that the correlation length saturates: at criticality the only available length is $L$ , so $\xi\propto L$ rather than diverging.

\xi(t,L) = L\,\mathcal{X}\!\big(L^{1/\nu}t\big),\qquad \mathcal{X}(0)=O(1) \qquad (3)

where: $\mathcal{X}$ — universal function with a finite value at zero argument; at $t=0$ the bulk $\xi\to\infty$ is cut off at $\xi\propto L$ . This is the concrete statement that a box of side $L$ cannot support correlations longer than itself. Differentiating the ansatz gives each observable's finite-size law. The susceptibility, two field derivatives of $f_s$ , grows as a power of $L$ set by the ratio $\gamma/\nu$ .

\chi(t,L) = L^{\gamma/\nu}\,\Phi_\chi\!\big(L^{1/\nu}t\big),\qquad \frac{\gamma}{\nu}=2y_h-d \qquad (4)

where: $\gamma=\nu(2y_h-d)$ — susceptibility exponent; $\Phi_\chi$ — universal scaling function, peaked at a finite argument. At $T_c$ ( $t=0$ ) this reads $\chi\sim L^{\gamma/\nu}$ , the peak-height law. The order parameter, one field derivative, decays as a power of $L$ at criticality — a finite box always has a nonzero but shrinking magnetization.

\langle|m|\rangle(t,L) = L^{-\beta/\nu}\,\Phi_m\!\big(L^{1/\nu}t\big),\qquad \frac{\beta}{\nu}=d-y_h \qquad (5)

where: $\beta=\nu(d-y_h)$ — order-parameter exponent; $\langle|m|\rangle$ — mean absolute magnetization, nonzero in a finite system even for $t>0$ because of finite-size symmetry restoration. The negative power says the ordered moment is squeezed as the box grows at fixed $T_c$ . For finite systems, pseudocritical temperatures approach the thermodynamic critical temperature according to:

T_c(L) - T_c(\infty) \;\propto\; L^{-1/\nu} \qquad (6)

where: $T_c(L)$ — the temperature where a chosen observable peaks (amplitude is observable-dependent and non-universal); $1/\nu=y_t$ — the shift exponent. Extrapolating $T_c(L)$ to $L\to\infty$ recovers the true critical temperature. Finally, the Binder cumulant is special: it is dimensionless, so no power of $L$ multiplies it. Its curves for different sizes therefore cross at $T_c$ , separating the location of the transition from the measurement of exponents.

U_L \equiv 1 - \frac{\langle m^4\rangle}{3\,\langle m^2\rangle^{2}} = \mathcal{U}\!\big(L^{1/\nu}t\big) \qquad (7)

where: $U_L$ — fourth-order Binder cumulant; $\langle m^k\rangle$ — $k$ -th moment of the order parameter; $\mathcal{U}$ — universal function. At $t=0$ , $U_L=\mathcal{U}(0)=U^\ast$ for every $L$ , so all curves pass through one crossing point — the standard high-precision $T_c$ locator.

Derivation AFrom RG homogeneity to all finite-size scaling laws

We start from the single homogeneity relation with $L^{-1}$ as a relevant field and derive the susceptibility, specific heat, order parameter, peak shift, and rounding width in one differentiation exercise. The owned structural result is that a single geometric fact — $y_L=1$ — plus the two bulk eigenvalues generates the entire finite-size phenomenology, so every FSS exponent is a bulk exponent divided by $\nu$ , and the whole machinery of extracting critical data from finite lattices reduces to reading powers of $L$ . All the laws flow from equation (1).

Step 1. Write the homogeneity relation with $y_t=1/\nu$ , $y_L=1$ : $f_s(t,h,L^{-1})=b^{-d}f_s(b^{1/\nu}t,b^{y_h}h,bL^{-1})$ , valid near the fixed point with irrelevant fields set to zero. Why: This is leading-order scaling; corrections from irrelevant operators are separate and subleading.

Step 2. Choose the rescaling factor $b=L$ to absorb the explicit system-size dependence: $f_s=L^{-d}\mathcal{F}(L^{1/\nu}t,L^{y_h}h)$ . Why: The size scale absorbs the arbitrary $b$ , leaving a universal function of the two physical scaling variables.

Step 3. Susceptibility: $\chi=\partial^2 f_s/\partial h^2|_{h=0}=L^{2y_h-d}\Phi_\chi(L^{1/\nu}t)=L^{\gamma/\nu}\Phi_\chi$ . Why: Two field derivatives bring $L^{2y_h}$ ; combined with the $L^{-d}$ prefactor and $\gamma=\nu(2y_h-d)$ this is $L^{\gamma/\nu}$ .

Step 4. Specific heat: $C=-T\partial^2 f_s/\partial t^2=L^{2/\nu-d}\Phi_C(L^{1/\nu}t)=L^{\alpha/\nu}\Phi_C$ ; if $\alpha<0$ the peak saturates to a constant. Why: Two temperature derivatives bring $L^{2/\nu}$ ; with $\alpha=2-\nu d$ this is $L^{\alpha/\nu}$ . A negative $\alpha$ means the exponent is negative and the peak does not grow.

Step 5. Order parameter: $m=\partial f_s/\partial h|_{h=0}=L^{y_h-d}\Phi_m=L^{-\beta/\nu}\Phi_m$ , using $\beta=\nu(d-y_h)$ ; at $t=0$ , $\langle|m|\rangle\sim L^{-\beta/\nu}$ . Why: One field derivative brings $L^{y_h}$ ; the net power is $y_h-d=-\beta/\nu$ , a decay.

Step 6. Pseudocritical temperature shift: $\partial_t X=0\Rightarrow L^{1/\nu}t^\ast=c_X\Rightarrow t^\ast\propto L^{-1/\nu}$ . Why: The maximum of the universal function sits at a fixed value $c_X$ of its argument; the constant $c_X$ (hence the shift amplitude) is observable-dependent.

Step 7. Rounding: the crossover from bulk to finite-size behavior occurs when $L^{1/\nu}|t|=\mathcal O(1)$ , implying a transition width $\Delta t\sim L^{-1/\nu}$ . Why: The scaling function changes character when its argument is $O(1)$ ; that fixes a temperature window narrowing as $L^{-1/\nu}$ . The exponent relations $\gamma/\nu=2y_h-d$ , $\alpha/\nu=2/\nu-d$ , $\beta/\nu=d-y_h$ hold symbolically; inserting the 2D-Ising values $\nu=1$ , $y_h=15/8$ gives $\gamma/\nu=7/4$ , $\beta/\nu=1/8$ , $\alpha/\nu=0$ .

Go deeper · Derivation B · 1 minWhy the Binder cumulant crosses at T_c

Derivation BWhy the Binder cumulant crosses at T_c

We show that the fourth-order Binder cumulant becomes asymptotically independent of system size at criticality because the powers of $L$ cancel between the numerator and denominator. Consequently, Binder-cumulant curves for different system sizes intersect near $T_c$ , providing a practical estimator of the critical temperature. This is the practical crown jewel of the article: a quantity that separates "where is the transition" from "what are the exponents," which is exactly what a careful measurement needs. Start from the moment scaling and form the ratio.

Step 1. Each moment scales as $\langle m^k\rangle\sim L^{-k\beta/\nu}g_k(L^{1/\nu}t)$ . Why: The order parameter carries $L^{-\beta/\nu}$ , so its $k$ -th moment carries the $k$ -th power of that prefactor, times a universal function.

Step 2. Form the cumulant ratio: $\dfrac{\langle m^4\rangle}{\langle m^2\rangle^2}=\dfrac{g_4}{g_2^2}\cdot\dfrac{L^{-4\beta/\nu}}{(L^{-2\beta/\nu})^2}=\dfrac{g_4}{g_2^2}$ . Why: The prefactors are $L^{-4\beta/\nu}$ on top and $L^{-4\beta/\nu}$ on the bottom; they cancel exactly, leaving a ratio of universal functions of $L^{1/\nu}t$ alone.

Step 3. Evaluate the Binder cumulant at the critical point $(t=0)$ : $U_L(0)=1-\tfrac13\,g_4(0)/g_2(0)^2=U^\ast$ , a universal number independent of $L$ . Why: At zero argument the universal functions take fixed values, so $U^\ast$ is the same for every size — all curves $U_L(T)$ pass through it at $T_c$ .

Step 4. Scaling of the Binder-cumulant slope: $\mathrm{d}U_L/\mathrm{d}T|_{T_c}\propto L^{1/\nu}$ . Why: Differentiating $\mathcal{U}(L^{1/\nu}t)$ in $T$ brings down $L^{1/\nu}$ ; the steepening of the crossing with size directly measures $\nu$ . The powers of $L$ cancel exactly in $\langle m^4\rangle/\langle m^2\rangle^2$ , and the crossing slope scales as $L^{1/\nu}$ — in the 2D-Ising model $1/\nu=1$ .

Limiting cases

Case	Result	Meaning
$L\gg\xi$ (far from $T_c$ )	$f_s\sim\xi^{-d}[\text{analytic in }L^{-1}]$	The box dwarfs $\xi$ ; finite-size corrections are exponentially small $\sim e^{-L/\xi}$ . Bulk thermodynamics applies.
$L\ll\xi$ (critical region)	$\chi\sim L^{\gamma/\nu},\ \langle\lvert m\rvert\rangle\sim L^{-\beta/\nu}$	The box, not $\xi$ , sets the scale; every singular quantity is a pure power of $L$ at $T_c$ .
Crossover, $L\sim\xi$	$X=L^{x/\nu}\Phi(L^{1/\nu}t)$	The universal scaling function has its maximum at a fixed value of $L^{1/\nu}t$ , implying the finite-size shift $T_c(L)-T_c\propto L^{-1/\nu}$ .
$d>4$ (above $d_c$ )	$\chi(T_c,L)\sim L^{d/2},\ T_c(L)-T_c\sim L^{-d/2}$	Hyperscaling fails; the dangerous irrelevant variable makes the effective length $L^{d/2}$ : Gaussian FSS.
Marginal $d=d_c=4$	$X\sim L^{x/\nu}(\ln L)^{\hat{x}}$	At the upper critical dimension multiplicative logarithmic corrections appear ( $\phi^4$ /Ising in $d=4$ ).

Common trapShift is not rounding, and the peak is not at the bulk T_c

A susceptibility maximum in a finite system generally occurs at a pseudocritical temperature rather than at the thermodynamic critical temperature. The pseudo-critical temperature $T_c(L)$ where a given observable peaks differs from $T_c(\infty)$ by $\sim L^{-1/\nu}$ , with an amplitude that is non-universal and observable-dependent — the specific-heat peak and the susceptibility peak generally shift to different places. Extrapolate $T_c(L)$ to $L\to\infty$ , or use the $L$ -independent Binder crossing. And beware $d\ge4$ : the textbook replacement $b=L$ gives the wrong powers because hyperscaling fails — use Gaussian FSS with $L^{d/2}$ .

Worked examples

Example 1 — The 2D-Ising doubling test. At the square-lattice Ising fixed point $\nu=1$ and $\gamma=7/4$ . Predict how the susceptibility peak and the order parameter at $T_c$ change when the lattice is doubled.

Step 1. Peak exponent: $\gamma/\nu=(7/4)/1=7/4=\mathbf{1.75}$ , so $\chi(T_c,L)\propto L^{7/4}$ .

Step 2. Doubling ratio: $\dfrac{\chi_{\max}(2L)}{\chi_{\max}(L)}=2^{\gamma/\nu}=2^{7/4}$ .

Step 3. Compute $2^{7/4}=2\cdot2^{3/4}=2\cdot1.6818=\mathbf{3.364}$ . Doubling from $L=32$ to $L=64$ raises the peak by a factor $\approx3.36$ .

Step 4. Order parameter: $\beta=1/8$ , so $\beta/\nu=1/8$ and $\langle|m|\rangle(T_c)\propto L^{-1/8}$ ; doubling multiplies it by $2^{-1/8}=1/1.0905=\mathbf{0.917}$ , an $8.3\%$ drop per doubling.

Step 5. Peak shift: $T_c(L)-T_c\propto L^{-1/\nu}=L^{-1}$ , so doubling the lattice halves the distance of the peak from the true $T_c=2/\ln(1+\sqrt2)\approx\mathbf{2.269}\,J/k_B$ . Reading the log-log slope of $\chi_{\max}$ versus $L$ across several doublings returns $\gamma/\nu=7/4$ directly.

Example 2 — Gaussian finite-size scaling in five dimensions. For $d=5>d_c=4$ the mean-field exponents $\nu=1/2$ , $\gamma=1$ hold. Show that naive FSS gives the wrong power and that the correct one is $L^{d/2}$ .

Step 1. Naive prediction: $\chi(T_c,L)\overset{?}{\sim}L^{\gamma/\nu}=L^{1/(1/2)}=L^{2}$ .

Step 2. But the quartic coupling $u$ is dangerously irrelevant for $d>4$ : the free energy is singular as $u\to0$ , so $u$ cannot be set to zero and the naive $b=L$ matching is invalid.

Step 3. Restoring the singular $u$ -dependence inflates the effective thermal length to $\ell\sim L^{d/2}$ , giving $\chi(T_c,L)\sim L^{d/2}=L^{5/2}=\mathbf{L^{2.5}}$ , not $L^2$ .

Step 4. Consistency at $d=4$ : the two pictures merge, since $d/2=2$ equals $\gamma/\nu=2$ — the naive and Gaussian powers coincide at the upper critical dimension (up to logarithms). For $d=5$ they differ by $L^{1/2}$ , a factor that has been confirmed in simulations of the five-dimensional Ising model.

Step 5. The shift likewise becomes $T_c(L)-T_c\sim L^{-d/2}=L^{-5/2}$ rather than the naive $L^{-1/\nu}=L^{-2}$ — a concrete, measurable diagnostic that hyperscaling has failed.

Origins · 1 minfinite-size scaling and how this was discovered

finite-size scaling

the RG theory of finite systems · The framework that predicts how the singular behaviour of an infinite system is rounded and shifted in a finite box, treating the inverse system size as a relevant scaling variable of RG eigenvalue one. Introduced by Michael Fisher and developed with Michael Barber in the early 1970s, it converted the "problem" of finite simulations into their greatest strength: by collapsing observables from several sizes onto universal functions of

L^{1/\nu}t

, one measures

\nu

\gamma

\beta

and locates

T_c

from lattices of a few thousand sites. It is the backbone of computational critical phenomena and, through the critical Casimir effect, of the thermodynamics of confined fluids and thin films.

History

1971

— Fisher formulates the finite-size scaling idea at the Enrico Fermi summer school

1972

— Fisher and Barber publish the finite-size scaling theory (Phys. Rev. Lett. 28, 1516)

1976

— Nightingale develops phenomenological finite-size scaling using transfer matrices on strips

1981

— Binder introduces the fourth-order cumulant crossing as a precise locator of the critical temperature

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd/4th ed., chapters on phase transitions and finite-size scaling · Kardar, Statistical Physics of Fields, RG and scaling chapters · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, scaling and RG chapters · Cardy, Scaling and Renormalization in Statistical Physics (finite-size scaling) · Barber, in Phase Transitions and Critical Phenomena, vol. 8 (ed. Domb & Lebowitz), "Finite-size scaling" · Fisher & Barber, Phys. Rev. Lett. 28, 1516 (1972) · Privman (ed.), Finite Size Scaling and Numerical Simulation of Statistical Systems Key results collected here: the eigenvalue $y_L=1$ follows from geometric rescaling $L^{-1}\to bL^{-1}$ ; the master collapse uses $b=L$ ; the exponent relations are $\gamma/\nu=2y_h-d$ , $\alpha/\nu=2/\nu-d$ , $\beta/\nu=d-y_h$ , with 2D-Ising values $\nu=1$ , $y_h=15/8$ giving $\gamma/\nu=7/4$ , $\beta/\nu=1/8$ , $\alpha/\nu=0$ ; the Binder-cumulant $L$ powers cancel exactly; the doubling ratio is $2^{7/4}\approx3.364$ ; Gaussian FSS gives $\chi\sim L^{d/2}$ and shift $\sim L^{-d/2}$ for $d>4$ , with the two pictures merging at $d=4$ where $d/2=\gamma/\nu=2$ ; and $T_c=2/\ln(1+\sqrt2)\approx2.269\,J/k_B$ is Onsager-exact.

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

If a finite system is analytic, in what sense does it "have" a transition?

It doesn't, strictly — the transition is a property of the $L\to\infty$ limit. What a finite system has is a smooth, well-defined pseudo-critical feature: a rounded susceptibility peak at $T_c(L)$ , a magnetization that is small but nonzero above $T_c$ , and so on. Finite-size scaling is precisely the theory of how these analytic features sharpen and shift into the true singularity as $L$ grows. The "transition" you see in a simulation is the shadow the bulk singularity casts on a finite box.

Why is the eigenvalue of L exactly 1 and not something class-dependent?

Because $L^{-1}$ is just an inverse length, and lengths rescale geometrically under the RG regardless of the universality class. One coarse-graining step by $b$ measures all lengths in units enlarged by $b$ , so a fixed box of size $L$ spans $L/b$ rescaled units: $L^{-1}\to bL^{-1}$ , i.e. $y_L=1$ . This is why $y_L$ never carries class-dependent information — all the class dependence lives in $y_t$ and $y_h$ . (Above $d_c=4$ the effective thermal length is inflated by the dangerous irrelevant variable, which is a separate effect.)

How is finite-size scaling different from data collapse?

Finite-size scaling is the RG theory that predicts the collapse form $X=L^{x/\nu}\Phi(L^{1/\nu}t)$ ; data collapse is the method of plotting rescaled data to test it and measure exponents. They are two faces of the same relation. In practice one uses FSS to know which powers to rescale by, then performs the collapse to extract the numbers. The data-collapse article develops the plotting method; this article derives the scaling laws it relies on.

Do boundary conditions really change the answer?

They change the scaling functions and universal amplitudes, though not the exponents. A periodic box, a box with free surfaces, and one with antiperiodic boundaries share $\nu,\gamma,\beta$ but have different $\Phi_\chi$ , different Binder crossing values $U^\ast$ , and different critical Casimir amplitudes. This is why careful FSS work always states the boundary conditions and aspect ratio: comparing a periodic-box $U^\ast$ to a free-surface one is comparing different universal numbers, even within the same universality class.

Why does the specific-heat peak sometimes not grow with L?

Because $\alpha$ can be negative. The specific heat scales as $L^{\alpha/\nu}$ , so when $\alpha<0$ the exponent is negative and the "peak" approaches a finite constant (plus a shrinking bump) as $L\to\infty$ rather than diverging. The 3D Ising model has $\alpha\approx0.11>0$ so its specific-heat peak grows weakly, but many systems (like the 3D XY model, $\alpha\approx-0.015$ ) have $\alpha<0$ and show a saturating cusp. This is a common source of confusion when the specific heat looks "flat" in a simulation.

Can finite-size scaling distinguish first-order from continuous transitions?

Yes, and cleanly. At a first-order transition the latent-heat delta function becomes a smooth peak in the specific heat of height $\sim L^{d}$ (a volume law), centered at a shift $\sim L^{-d}$ , driven by tunneling between coexisting phases. This is a different scaling from the continuous case ( $L^{\alpha/\nu}$ with $\alpha/\nu<d$ ), so measuring how the peak height grows with $L$ is a decisive numerical criterion: volume-law growth signals first order, a smaller power signals a continuous transition.