Data Collapse and Scaling Plots

Data collapse is the daily bread of computational critical phenomena. A modern study of the three-dimensional Ising model simulates lattices of side $L=16,32,64,128,\dots$, locates $T_c$ from the Binder-cumulant crossing (which needs no exponents), and then collapses $L^{\beta/\nu}M$ and $L^{-\gamma/\nu}\chi$ against $tL^{1/\nu}$, tuning the exponent ratios to optimize the overlap. This delivers $\nu=0.6300(4)$ and $\eta=0.036(1)$ to a precision that rivals the conformal bootstrap, all from finite lattices no real experiment could match. The method also reaches into the laboratory and beyond equilibrium. Coexistence curves for argon, nitrogen, and carbon dioxide collapse onto a single master curve with $\beta\approx0.326$ — the 3D-Ising value — the modern, exponent-resolved version of the law of corresponding states, proving that liquid–gas and uniaxial-magnet criticality share one universality class. The same logic collapses percolation cluster-size data, dynamic relaxation curves (extracting the dynamic exponent $z$), and zero-temperature quantum-critical data in the coupling $g$, so that a single diagnostic certifies scaling and measures exponents across an enormous range of physics.

Imagine many magnetization curves $M(h)$ taken at different temperatures near $T_c$. Each is a smeared step, and no two look alike. The scaling hypothesis says these curves are secretly the same object: stretch the field axis by $|t|^{-\beta\delta}$ and squash the magnetization axis by $|t|^{-\beta}$, and every curve lies exactly on top of one master curve. The temperature has been scaled away. What remains is a single universal shape — and the fact that the trick works is the physics. Finite-size scaling has the same spirit but trades $|t|$ for $L$. At exactly $T_c$ the only length left in the problem is the box size $L$, so RG analysis forces $M\sim L^{-\beta/\nu}$. Off-critical data then collapse when plotted as $L^{\beta/\nu}M$ against the ratio $L/\xi\sim tL^{1/\nu}$, because that ratio is the only dimensionless way the finite system can "know" whether it is critical. Simulations that could never reach the thermodynamic limit read the exponents straight off this collapse. Think of the master curve as a fingerprint of the fixed point. Two utterly different systems — a uniaxial ferromagnet and a liquid–gas critical point — share the same fingerprint if they flow to the same fixed point. The collapse plot is how you read the fingerprint: the axes are scrubbed of every material-specific scale except two non-universal amplitudes. That is the deep claim, and it is falsifiable: if two systems in the same predicted class refuse to collapse onto one curve, the classification is wrong. A word of caution built into the method: collapse is only asymptotic. Include data too far from $T_c$, or on too-small lattices, and corrections-to-scaling terms $\propto|t|^{\omega\nu}$ or $L^{-\omega}$ become visible as a systematic spread that shrinks as you approach criticality. This is not noise — it is the leading irrelevant operator, and a careful worker either restricts the fitting window or adds the correction term explicitly.

The two panels above are the claim of the RG made visible: three temperature curves that share nothing on the left become indistinguishable points on one curve on the right, once you rescale by the correct $\beta$ and $\delta$. The collapse is the thesis, and the outliers are the data points that fall exactly onto the universal curve.

Everything starts from the RG homogeneity of the singular free energy: coarse-grain once, the density picks up $b^{-d}$, and the scaling fields are driven by their eigenvalues. Data collapse is nothing more than this relation with $b$ eliminated in favour of a physical scale.

where: $f_s$ — singular part of the free-energy density; $b$ — RG rescaling factor ($>1$); $d$ — spatial dimension; $y_t,y_h$ — thermal and magnetic RG eigenvalues; $t,h$ — reduced temperature $(T-T_c)/T_c$ and reduced field. The relation holds for every $b$, which is the freedom we spend to make a collapse. Choosing $b=|t|^{-1/y_t}$ removes the temperature from the first slot and leaves a function of a single scaling variable — the mathematical form the collapse plot verifies.

where: $\Phi$ — the universal scaling function, one branch for $t>0$ and one for $t<0$. The prefactor sets the vertical rescaling; the argument sets the horizontal one. The observable an experiment actually plots is the magnetization, obtained by one field derivative. Its collapse form names the two exponents that scale the axes.

where: $\mathcal{M}_\pm$ — universal scaling function for $t\gtrless0$; $\beta=(d-y_h)/y_t$ scales the vertical axis; $\beta\delta=y_h/y_t$ scales the horizontal axis. Plot $M|t|^{-\beta}$ against $h|t|^{-\beta\delta}$ and all isotherms collapse onto $\mathcal{M}_\pm$. At criticality the correlation length would diverge, but in a finite box it is cut off. The controlling variable becomes the ratio $L/\xi$, and $L$ enters as a scaling field with the trivial eigenvalue of an inverse length.

where: $\xi$ — correlation length; $\nu$ — correlation-length exponent; $y_L=1$ — RG eigenvalue of $L^{-1}$, fixed by geometric rescaling ($L^{-1}\to bL^{-1}$). This is what lets $L$ replace $|t|$ as the scaling variable at $T_c$. Adding $L^{-1}$ to the scaling fields and choosing $b=L$ gives the finite-size collapse form for the order parameter — the workhorse of computational critical phenomena.

where: $L$ — linear system size; $\widetilde{\mathcal{M}}$ — finite-size scaling function; $tL^{1/\nu}\sim(L/\xi)^{1/\nu}$ — the only relevant scaling variable. Plot $L^{\beta/\nu}M$ against $tL^{1/\nu}$ and data from all sizes collapse. The susceptibility and the Binder cumulant collapse the same way, and the cumulant has a special virtue: no power of $L$ multiplies it, so its curves cross at $T_c$ independently of any exponent.

where: $\chi$ — magnetic susceptibility, $\gamma/\nu=2y_h-d$; $U_4=1-\langle m^4\rangle/(3\langle m^2\rangle^2)$ — the dimensionless fourth-order Binder cumulant. Because $U_4$ has no $L$ prefactor, curves for different $L$ intersect at $T_c$, separating the location of $T_c$ from the measurement of exponents. Finally, real data never quite collapse: the leading irrelevant field adds a size- or temperature-dependent correction that must be modelled or fit away.

where: $\omega$ — leading correction-to-scaling exponent; $a$ — non-universal amplitude. The bracketed factor is the systematic residual that shrinks as $t\to0$ or $L\to\infty$; it is the fingerprint of the leading irrelevant operator, not experimental error.

Example 1 — A temperature-independent collapse in mean field. Take the mean-field equation of state, which gives $M\sim|t|^{\beta}$ at $h=0$ with $\beta=1/2$ and $M\sim h^{1/\delta}$ at $t=0$ with $\delta=3$. Show explicitly that $M/|t|^{\beta}$ is a temperature-independent function of $h/|t|^{\beta\delta}$.

Step 1. The scaling variable is $x=h/|t|^{\beta\delta}$ with $\beta\delta=\tfrac12\cdot3=3/2$, so $x=h/|t|^{3/2}$.

Step 2. Build a concrete mean-field-style master curve $\mathcal{M}(x)=x/(1+|x|^{4/3})^{1/2}$, which is linear at small $x$ (linear response) and $\to x^{1/3}$ at large $x$ (critical isotherm, $1/\delta=1/3$).

Step 3. Then $M(t,h)=|t|^{1/2}\mathcal{M}(h/|t|^{3/2})$. Evaluate at $h=0.02$ for $|t|=0.04$ and $|t|=0.01$: for $|t|=0.04$: $x=0.02/0.04^{1.5}=0.02/0.008=2.5$, and $M=0.04^{0.5}\mathcal{M}(2.5)$; for $|t|=0.01$: $x=0.02/0.01^{1.5}=0.02/0.001=20$, and $M=0.01^{0.5}\mathcal{M}(20)$.

Step 4. The rescaled magnetization $M/|t|^{1/2}=\mathcal{M}(x)$ depends only on $x$, not on $t$ separately. Two different temperatures with the same $x$ give the same rescaled value — the definition of a collapse.

Step 5. Large-argument check: $\mathcal{M}(x)\to x^{1/3}$, so $M\to|t|^{1/2}(h|t|^{-3/2})^{1/3}=h^{1/3}$ with the $|t|$ dependence cancelling — the critical isotherm $\delta=\mathbf{3}$ emerges from the tail, as it must.

Example 2 — The finite-size doubling test for 2D-Ising. At the 2D-Ising fixed point $\nu=1$, $\gamma=7/4$, so $\chi_{\max}\propto L^{\gamma/\nu}=L^{7/4}$. Predict the peak-height ratio when the lattice is doubled.

Step 1. $\gamma/\nu=(7/4)/1=7/4=\mathbf{1.75}$.

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

Step 3. Compute $2^{7/4}=2^{1.75}=2\cdot2^{0.75}=2\cdot1.6818=\mathbf{3.364}$.

Step 4. So going from $L=32$ to $L=64$ should raise the susceptibility peak by a factor $\approx3.36$; measuring that ratio across several doublings and reading the log-log slope returns $\gamma/\nu=7/4$ directly. Simultaneously $M_{\max}$ falls by $2^{\beta/\nu}=2^{1/8}\approx\mathbf{1.09}$ per doubling, giving $\beta/\nu=1/8$.

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 12–14 · 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. 9–10 · Cardy, Scaling and Renormalization in Statistical Physics, Ch. 3 · Landau & Lifshitz, Statistical Physics Part 1, Ch. XIV

Provenance. All formulas are standard and recomputed by verify_aq-sm-data-collapse-scaling-plots.py: the exponent relations $\alpha=2-d/y_t$, $\beta=(d-y_h)/y_t$, $\gamma=(2y_h-d)/y_t$, $\delta=y_h/(d-y_h)$, $\nu=1/y_t$; the mean-field collapse demonstration with $\beta=1/2,\delta=3$; the 2D-Ising ratios $\gamma/\nu=7/4$ (so $\chi_{\max}$ doubling ratio $2^{7/4}\approx3.364$) and $\beta/\nu=1/8$; and $T_c=2/\ln(1+\sqrt2)\approx2.269\,J/k_B$ from Onsager. The 3D-Ising $\beta\approx0.326$ is a high-precision Monte-Carlo/conformal-bootstrap value. This article cross-references the RG Scaling Form of the Free Energy and Finite-Size Scaling and the RG by name, and anchors on Onsager's Solution of the 2D Ising Model.

Builds on: The Renormalization Group Idea: Coarse-Graining and Rescaling · Scaling Fields and Scaling Variables · Critical Exponents from the RG · Fixed Points of the RG. Leads to: Wilson-Fisher Fixed Point · Gaussian Fixed Point and Upper Critical Dimension · Relevant, Irrelevant, and Marginal Operators.