Distributions and heavy tails

Finite-size scaling and data collapse

In a finite sample, a heavy-tailed distribution rarely runs as a pure power law all the way to infinity. System boundaries, grid size, or observation windows bend the tail downward at a size-dependent cutoff. Finite-size scaling aligns distributions recorded across multiple system sizes onto a single master curve, letting you estimate both the power-law exponent and the cutoff fractal dimension while checking whether one scaling ansatz actually holds.

Updated · 7 min read

On this page
  1. What the finite-size scaling ansatz assumes
  2. How data collapse minimizes overlap variance in log-coordinates
  3. Reading collapse diagnostics and spotting boundary traps
  4. Why an attractive collapse is not proof of universality
  5. Running FSS Collapse in Autoplot
  6. Frequently asked questions

What the finite-size scaling ansatz assumes

When avalanches, fracture bursts, or critical clusters grow inside a finite domain of linear size L, large events cannot exceed a characteristic cutoff scale sc∼LDx. Following Salman, Ahadi, and Truskinovsky (2026), Eq. 43 writes the size distribution at size L as

P(s;L)=s−xF(s/LDx)

Here s is the positive event observable (such as stress drop ΔΣ or dissipated energy ΔW), x is the power-law decay exponent (τ for stress drops, ε for energy drops; denoted τ in Autoplot's search UI), Dx is the fractal cutoff dimension controlling how fast the upper cutoff grows with system size L, and F(u) is the dimensionless scaling function describing both the critical power-law plateau and the finite-size rollover.

Before running a multi-size collapse, establish on individual sizes that a heavy tail actually exists and where lower non-scaling structure ends. See log binning for width-normalized density estimation across orders of magnitude, and how to choose xmin for a power law for separating small-scale discretization or threshold effects from the scaling regime.

How data collapse minimizes overlap variance in log-coordinates

Data collapse turns visual curve alignment into an explicit variance minimization over log-rescaled densities (Salman, Ahadi, and Truskinovsky, 2026, Eqs. 44–50):

  1. Bin raw samples on logarithmic grids: for each dataset with linear system size Lk, restrict positive observations to [xmin,k,xmax,k], build log-spaced bin edges, evaluate geometric bin centers, and normalize bin counts by nused·Δsi to obtain probability densities P(si;Lk).
  2. Rescale into collapsed log10 coordinates: for a candidate pair (x,Dx), map each non-empty bin center si and density P(si;Lk) to uk,i=log10(si/LkDx) and vk,i=log10(P(si;Lk)LkDxx)
  3. Find the strict common overlap [umin,umax]: set umin=maxk[miniuk,i] and umax=mink[maxiuk,i], linearly interpolate each curve onto a uniform grid {cj} spanning [umin,umax] (for example 80 points), and retain the valid grid subset V where all NL curves are defined inside their measured bounds.
  4. Minimize cross-size variance Q(x,Dx): compute the average variance across all NL curves at each valid grid point j∈V: Q(x,Dx)=1|V|∑j∈V1NL−1∑k=1NL(vˆk(cj)−v¯(cj))2 where v¯(cj)=1NL∑k=1NLvˆk(cj) is the cross-size mean at cj.
  5. Select the best (x,Dx) pair on the search grid: evaluate Q(x,Dx) across the (x,Dx) search grid (optionally narrowing onto a local fine grid around the best coarse point) to find the parameters that minimize variance.
  6. Estimate operational 1-σ errors from the Q profile: project Q(x,Dx) onto 1D minimum-variance profiles along x and Dx, fit a local quadratic q(δ)≈Qmin+aδ2 around the minimum grid point, and take the half-width where the fitted variance doubles to 2Qmin: σ=Qmin/a
Autoplot FSS Collapse layer card named Avalanche FSS in X&Y Plot, showing Data 1 through Data 4 with Add Data, closed Collapse Search (tau 0.8–1.8, D 1.8–3.4), Mark & stroke (Solid), Label & legend, fitted tau 1.3260, D 2.6614, quality 5.132e-04, series 4, Estimated errors tau +/- 0.0275 and D +/- 0.132, per-size counts L=32 to L=256, and Master curve analytic form note
Autoplot's FSS Collapse layer card stacks size-indexed datasets, search ranges, fitted exponents, estimated errors, and per-size bin counts in one inspector.

Reading collapse diagnostics and spotting boundary traps

A clean numerical minimum does not automatically mean the data obeys finite-size scaling. Always read the summary table, search bounds, and overlap together:

Diagnostic signalWhat it tells youAction to take
Optimum hits a search limit (τ min/max or D min/max)True minimum lies outside your search box or Q slopes monotonically without a local valleyWiden (τ,D) search bounds and confirm a true interior minimum appears
No stable overlap found or very short [umin,umax]Curves barely overlap in s/LD under min overlap or aggressive per-size cutoffsRelax unnecessary xmin/xmax cutoffs or widen L spacing gradually rather than dropping intermediate sizes
Crisp collapse for NL=2 sizes onlyTwo decaying curves can align under many (x,Dx) tradeoffs across short intervalsAdd at least one or two intermediate or larger system sizes so both slope and finite-size rollover lock down
τ agrees across regimes while D jumps sharplyBulk power-law scaling persists while event geometry changesInspect spatial morphology or cutoff structure separately; exponent similarity alone does not imply identical geometry

In crystal plasticity simulations by Salman, Ahadi, and Truskinovsky (2026), preyield and postyield dissipated energy bursts illustrate that fourth row directly: energy exponents stay close (ε=0.9195 preyield vs ε=1.0207 postyield across ~5–6 decades) even as the cutoff fractal dimension jumps from nearly point-like localized events (Dε=0.3640) to system-spanning branched bands (Dε=0.9836), while stress drops shift in both exponent (τ=0.9070 to τ=1.2476) and dimension (Dτ=0.3987 to Dτ=1.4379).

Autoplot X&Y Plot displaying four finite-size scaling collapsed distributions for L=32, L=64, L=128, and L=256 in blue, red, green, and purple overlaid with a smooth Avalanche FSS master curve on log10(s / L^D) versus log10(P(s) · L^{Dτ}) axes
Four synthetic truncated power-law datasets (L = 32, 64, 128, 256; 60,000 samples each) collapsed in Autoplot to τ = 1.3260 ± 0.0275 and D = 2.6614 ± 0.132 (quality 5.132e-04), overlaid with the fitted analytic master curve.

Why an attractive collapse is not proof of universality

Visual curve collapse is a consistency check, not a hypothesis test or proof of universality. Three guardrails keep a finite-size scaling figure scientifically defensible:

  1. Read quality as an overlap variance, not a p-value: quality (Qmin) is the mean variance across interpolated log densities, not a chi-square statistic or statistical goodness-of-fit test, and depends on default bins, min overlap, interp pts, and cutoffs.
  2. Test sensitivity to cutoffs and bin count: vary default bins (for example 20 vs 30 vs 40) and re-check small-s discretization cutoffs (raw log10 xmin) to verify fitted (τ,D) remain stable.
  3. Check single-size fits alongside the collapse: confirm that individual system sizes support a heavy tail rather than curving continuously everywhere using power law vs lognormal tails or a binned vs CCDF check in how to fit a power law to data.
In Autoplot

Running FSS Collapse in Autoplot

Autoplot's FSS Collapse layer (Run Analysis → FSS Collapse in an X&Y Plot figure, Plus) implements the variance-minimization collapse of Salman, Ahadi, and Truskinovsky (2026) locally on raw positive sample columns without requiring you to pre-bin files outside the app:

  • Size-indexed inputs (Data): use + Add Data to add two or more positive numeric columns (>= 5 points and >= 3 non-empty bins each), assign each row its linear system size L, and optionally enter per-row raw log10 xmin or raw log10 xmax cutoffs and default bins (default 30, clamped >= 3). Cutoffs are converted from base-10 logarithms (10value) before binning; the figure itself renders collapsed coordinates.
  • Coarse-to-fine search (Collapse Search): set bounds and resolution (τ min/max/steps, D min/max/steps), toggle Fine search around best grid point (fine steps), adjust interp pts (default 80) and min overlap (default 0.5 decades), and click the orange gear button to compute or rerun.
  • Outputs and warnings (Mark & stroke, card summary, Figure Inspector → Results): Autoplot plots every collapsed L series alongside an optional master curve (fitted analytic form using D and τ, falling back to the cross-size mean curve if the weighted analytic fit fails), prints scalar τ, D, quality (Qmin), quadratic profile Estimated errors, and per-L used counts/bins, and surfaces boundary (τ or D on search limits) or overlap (No stable overlap found) warnings. Autoplot does not plot the 2D Q(τ,D) surface or 1D profile curves directly.

See features and pricing.

Frequently asked questions

Frequently asked questions
Should I pass raw event values or pre-binned histogram heights into FSS Collapse?Pass raw positive event columns (one variable per system size L), not already binned histogram counts or densities. Autoplot applies each row's raw log10 xmin and raw log10 xmax cutoffs first and constructs width-normalized logarithmic probability densities internally before searching (τ,D).
What does Autoplot's quality metric mean, and can I compare it across papers?quality is the mean cross-size variance Qmin of interpolated log10(P·LDτ) values over the shared overlap grid V. Smaller Q indicates tighter overlap for fixed binning, overlap settings, and sample size NL, not an absolute p-value—only compare Q across runs or parameter settings that share the same binning, cutoffs, and overlap window.
Why did FSS Collapse warn that best τ or D is on the search boundary?If optimal τ equals τ min/τ max or optimal D equals D min/D max, the true variance minimum likely lies outside your configured interval—or Q decreases monotonically across the search box without an interior valley. Widen the search bounds and rerun so the minimum is strictly bracketed inside the grid.
How does Autoplot compute Estimated errors for τ and D?Autoplot projects the grid quality surface Q(τ,D) onto 1D minimum profiles around the best grid point, fits a local quadratic polynomial q(δ)≈Qmin+aδ2, and prints σ=Qmin/a (the half-width where the fitted variance doubles to 2Qmin). Because Autoplot outputs these scalar widths without plotting the Q(τ,D) landscape itself, check stability across nearby binning and cutoff choices as well.
What happens if one system size covers a much narrower x-range than the others?Because [umin,umax] takes the strict intersection across every active L series, a single poorly populated or heavily truncated curve shrinks the valid overlap domain V. If umax−umin drops below min overlap, Autoplot emits "No stable overlap found; relax cutoffs or minimum overlap." Inspect per-L used/total counts and adjust cutoffs or exclude incomplete trial runs intentionally.

Sources

Finite-size scaling equations, the variance-minimization quality score Q(x,Dx), operational 1-σ profile error estimation, and crystal-plasticity empirical contrasts are adapted from O. U. Salman, A. Ahadi, and L. Truskinovsky, "Self-induced marginality in plastically deformed crystals," *Physical Review Materials* 10, 103611 (2026), DOI `10.1103/3rwf-d3yx` (published under CC BY 4.0). Autoplot UI labels, logarithmic binning rules, quality, and profile_sigma error calculation come directly from Autoplot's Python engine (compute_fss_collapse in math_logic.py) and FSSCollapseCardEditor, verified on 15 November 2026.

Try it on your own data.

Autoplot's FSS Collapse card bins raw size-indexed columns on logarithmic scales, searches (τ, D) on a coarse-to-fine grid, plots collapsed curves with a master fit, and flags boundary or overlap failures. Part of Plus (free for 30 days).

Compare power law vs lognormal →
Download free for Mac

Apple silicon · macOS 15 or later · Free plan · Plus free for 30 days

Send to my Mac

Autoplot runs on a Mac. Send the link to yourself, or open it in the Mac App Store.