Truncated Power Law in Heavy-Tailed Models

Updated 10 July 2026

Truncated power laws are distributions that follow a power-law behavior over a range but include explicit cutoff mechanisms—such as exponential decay, bounded support, or threshold alteration—to reflect finite constraints.
They are used in fields like astrophysics, geoscience, medical text analysis, and network science to capture deviations from pure scale invariance due to real-world limitations.
Estimation methods, including maximum likelihood and goodness-of-fit tests, are employed to differentiate truncated power laws from pure power laws, optimizing model selection based on empirical and structural validations.

A truncated power law is a power-law-like distribution whose scale-free regime is curtailed by an explicit cutoff, a finite support, or a threshold-induced deformation. In the literature surveyed here, the term covers several non-identical but closely related constructions: a power law with exponential cut-off, $f(x)\propto x^{-\alpha}e^{-\lambda x}$ ; a power law restricted to a bounded interval $[a,b]$ ; finite-time or finite-size laws with a hard upper cutoff; and thresholded forms such as $N(x)\,dx \propto (x+x_0)^{-a}dx$ . Across these formulations, the common idea is that empirical systems often display approximate scaling over a substantial range but depart from a pure power law because arbitrarily large events are suppressed by finite time horizons, finite system size, observational limits, or other structural constraints (Quiroz et al., 2020, Corral et al., 2012, Roman et al., 2022).

1. Definitions and mathematical variants

The most common empirical form is the power law with exponential cut-off,

$f(x) \propto x^{-\alpha} e^{-\lambda x},$

with $x\ge x_{\min}$ , scaling exponent $\alpha$ , and cut-off parameter $\lambda$ . In text analysis of medical discharge reports and in interbank-network tails, this is the explicit truncated power law fitted to data (Quiroz et al., 2020, Vandermarliere et al., 2014).

A distinct but standard statistical construction is the bounded-support truncated power law on $a\le x\le b$ ,

$f(x) = \frac{(\alpha-1)a^{\alpha-1}}{1 - \left( \frac{a}{b} \right)^{\alpha-1}} \cdot \frac{1}{x^\alpha},$

with complementary cumulative distribution

$S(x) = \frac{(a/x)^{\alpha-1} - (a/b)^{\alpha-1}}{1 - (a/b)^{\alpha-1}} \qquad (\alpha \neq 1).$

This formulation is central in geoscience methodology for fitting and testing truncated power-law regimes over objectively selected intervals (Corral et al., 2012, Corral et al., 2018).

A third variant is the finite-time cascade law derived from a master equation,

$[a,b]$ 0

with solution

$[a,b]$ 1

and $[a,b]$ 2 for $[a,b]$ 3. Here the truncation is a hard cutoff at $[a,b]$ 4, not an exponential taper (Roman et al., 2022).

A related but not identical family is the thresholded power law,

$[a,b]$ 5

also described as generalized Pareto type II or Lomax. In this case $[a,b]$ 6 encodes physical or instrumental thresholds, while $[a,b]$ 7 represents background contamination (Aschwanden, 2015).

Form	Representative expression	Truncation mechanism
Exponential cut-off	$[a,b]$ 8	Soft tail suppression
Bounded support	$[a,b]$ 9 on $N(x)\,dx \propto (x+x_0)^{-a}dx$ 0	Explicit lower and upper bounds
Finite-time hard cutoff	$N(x)\,dx \propto (x+x_0)^{-a}dx$ 1 for $N(x)\,dx \propto (x+x_0)^{-a}dx$ 2	Maximum size set by observation time
Thresholded power law	$N(x)\,dx \propto (x+x_0)^{-a}dx$ 3	Lower-threshold deformation and finite upper range

These forms are often grouped under the same label because each preserves a power-law regime over part of the support while introducing a systematic departure from pure scale invariance. The precise mathematical choice matters for inference and for physical interpretation.

2. Mechanisms that generate truncation

Finite time is one direct mechanism. In the master-equation approach for cascades, the cutoff emerges because a cascade can only grow if enough time is available; the maximal observable size therefore grows linearly with observation time, $N(x)\,dx \propto (x+x_0)^{-a}dx$ 4 (Roman et al., 2022). This gives a theoretical route from Markov dynamics to a truncated power law, rather than treating the cutoff as merely phenomenological.

Finite system size and finite extent provide a second mechanism. In astrophysical instability statistics, upper truncation at the largest events is attributed to finite system size, while the low-end distortion can reflect physical instability thresholds or incomplete sampling (Aschwanden, 2015). In self-consistent stellar-dynamical models, truncated power-law spheres are defined directly in radius by

$N(x)\,dx \propto (x+x_0)^{-a}dx$ 5

so that the system has finite spatial extent by construction (Baes et al., 2023).

Finite correlation time is a third mechanism. In anomalous diffusion, tempered fractional Gaussian noise replaces pure power-law correlations with exponentially or power-law tempered kernels. For exponential tempering,

$N(x)\,dx \propto (x+x_0)^{-a}dx$ 6

and the truncation of correlations produces crossover from anomalous diffusion to normal diffusion at long times (Molina-Garcia et al., 2018).

Finite sample size and finite network size also induce cutoffs. In Bayesian graph models with BFRY node weights,

$N(x)\,dx \propto (x+x_0)^{-a}dx$ 7

the truncation at $N(x)\,dx \propto (x+x_0)^{-a}dx$ 8 ensures all moments exist for finite $N(x)\,dx \propto (x+x_0)^{-a}dx$ 9 and models the degree cutoffs observed in finite real-world graphs (Lee et al., 2017).

Taken together, these results support a broad interpretation: truncation is rarely an arbitrary cosmetic modification. It often encodes physically meaningful finiteness in time, space, resources, memory, or observation.

3. Estimation, goodness-of-fit, and model discrimination

The dominant inferential strategy is maximum likelihood estimation. In medical text analysis, all candidate distributions—power law, lognormal, exponential, stretched exponential, and truncated power law—were fit by MLE, with $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 0 and $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 1 estimated following Clauset et al. methods for the power-law family (Quiroz et al., 2020). In the bounded-support framework, one fixes $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 2 and $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 3, maximizes the log-likelihood for $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 4, and evaluates the empirical complementary cumulative distribution against the fitted model (Corral et al., 2012).

Goodness-of-fit is typically assessed by the Kolmogorov-Smirnov statistic plus Monte Carlo or bootstrap calibration. A systematic procedure for truncated and non-truncated power laws sweeps candidate intervals $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 5, fits $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 6 by MLE, computes the KS distance, simulates synthetic samples from the fitted model, refits each sample, and estimates a $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 7-value as the fraction of simulated KS distances exceeding the empirical one (Corral et al., 2012). An improved version of this protocol was later used to determine whether a power-law tail is improved by a truncated log-normal and to choose the range maximizing the accepted log-range $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 8 (Corral et al., 2018).

Likelihood-ratio testing is the standard tool for discriminating between heavy-tailed alternatives. In the medical-report study, negative LR with $f(x) \propto x^{-\alpha} e^{-\lambda x},$ 9 favored the alternative over the pure power law; for the full corpus the truncated power law yielded LR $x\ge x_{\min}$ 0, $x\ge x_{\min}$ 1, and the lognormal LR $x\ge x_{\min}$ 2, $x\ge x_{\min}$ 3 (Quiroz et al., 2020). Interbank-network analysis likewise used MLE, KS-based tail selection, and likelihood-ratio comparisons across candidate distributions (Vandermarliere et al., 2014).

Several papers address limitations of standard workflows. One critique is that some Clauset-style procedures can fail to recognize true simulated power-law tails and do not work well when extended to upper-truncated power laws, motivating the alternative KS-plus-Monte-Carlo protocol (Corral et al., 2012). Another issue is that standard MLEs generally require $x\ge x_{\min}$ 4 and $x\ge x_{\min}$ 5 to be specified or guessed. A later least-squares tool was proposed to estimate $x\ge x_{\min}$ 6, $x\ge x_{\min}$ 7, and $x\ge x_{\min}$ 8 simultaneously by fitting expected order-statistic bins of a truncated power law (Pezzuto et al., 2023).

Model comparison near the boundary between lognormal and power law can also be numerically unstable. A reparameterization of the truncated lognormal,

$x\ge x_{\min}$ 9

was introduced so that $\alpha$ 0 yields an exact power law and $\alpha$ 1 becomes the power-law exponent; this avoids the divergence of the standard $\alpha$ 2 parameterization near the power-law limit (Pueyo, 2014).

A different line of methodology avoids fixing $\alpha$ 3 altogether. Non-parametric power-law surrogates preserve the likelihood function for all $\alpha$ 4 by redistributing observed prime factors while respecting lower or upper cutoffs and optional constraints such as Markov transitions (Moore et al., 2022). This extends inference from a single fitted exponent to the entire family of admissible power laws.

4. Empirical evidence across disciplines

Empirical studies consistently show that truncated power laws are often competitive with, and sometimes superior to, pure power laws, but they are not universally dominant.

In medical discharge reports, 20,000 reports from MIMIC-III were lowercased, tokenized, and analyzed by frequency. The pure power law was ruled out by a goodness-of-fit test with $\alpha$ 5, whereas the truncated power law was very strongly favored over the pure power law with LR $\alpha$ 6, $\alpha$ 7. The lognormal was also favored, and CCDF plots showed that the pure power law fit the extreme tail but not the head, whereas the truncated power law and lognormal tracked the full range more closely (Quiroz et al., 2020).

In the Russian interbank network, all studied distributions were heavy tailed, the fat tail typically contained 20% of the data, and the tail was mostly described well by a truncated power law. Yet the stretched exponential and log-normal fit the full range of the data best. The conclusions were robust across aggregation windows and across growth, maturity, normal, and crisis periods, with the caveat that the statistical distinction among tail models was often weak (Vandermarliere et al., 2014).

In geoscience, the objective fitting protocol found that tropical cyclones and rain precipitation clusters over oceans exhibit truncated power-law regimes, whereas karst sinkholes and wildfires are better described by truncated log-normals. Global earthquakes were found to follow a double power law rather than a single truncated power law (Corral et al., 2018). This domain therefore illustrates both the utility and the limits of the truncated-power-law hypothesis.

Earthquake seismic moments provide a more specific caution. When the Gutenberg-Richter power law was compared against the tapered Gutenberg-Richter and truncated gamma distributions on the global CMT catalog, the truncated gamma achieved the highest likelihood and significantly outperformed the pure power law, including after the 2004 Sumatra-Andaman earthquake (Serra et al., 2015). Similarly, in stellar-cluster mass-function fitting, a doubly truncated gamma outperformed both the lognormal and a four power law in the tested datasets (Zaninetti, 2014). These cases indicate that exponential-tail alternatives other than the truncated power law can be statistically preferable.

In network science, Bayesian simple-graph models based on truncated BFRY variables yield asymptotic degree probabilities

$\alpha$ 8

with truncation parameter $\alpha$ 9 setting a finite-size cutoff in observable degrees (Lee et al., 2017). Here the truncated power-law idea appears not only as a fitted marginal but as a prior-generating mechanism.

5. Consequences for dynamics, observables, and asymptotics

Truncation changes more than tail probabilities; it can alter the qualitative behavior of aggregate observables.

For sums of heavy-tailed random vectors, the effect depends on the growth rate of the truncation threshold $\lambda$ 0. In the soft truncation regime,

$\lambda$ 1

much of the heavy-tailed large-deviation structure is retained. In the hard truncation regime,

$\lambda$ 2

the same is lost and the sums behave analogously to light-tailed variables on the relevant large-deviation scales (Chakrabarty, 2011). This is a precise statement that truncation can either preserve or suppress heavy-tailedness, depending on how the cutoff scales with sample size.

In stochastic transport, truncating power-law correlations produces diffusion crossovers. Exponential tempering yields short-time anomalous diffusion, $\lambda$ 3, and long-time normal diffusion, $\lambda$ 4; power-law tempering can instead generate crossover from one anomalous regime to another, depending on the tempering exponent $\lambda$ 5 (Molina-Garcia et al., 2018).

In strong gravitational lensing, a truncated power-law surface-density model,

$\lambda$ 6

was used to derive deflection, shear, and magnification for finite-extent lenses. Mock observations showed that when images form outside the truncation radius, the interior slope can still be recovered for elliptical lenses. This directly contradicts the commonly stated idea that lensing is insensitive to the mass distribution interior to the image annulus (O'Riordan et al., 2020).

In stellar dynamics, truncation can be imposed at the density level rather than by energy truncation. For truncated power-law spheres with tangential Cuddeford anisotropy, self-consistency requires

$\lambda$ 7

and the paper shows that this condition is both necessary and sufficient for the family under study (Baes et al., 2023). In this context, truncation enables finite-extent models with closed expressions for potential and velocity-dispersion profiles.

A recurring misconception is to treat all departures from pure power laws as interchangeable. The surveyed literature instead distinguishes several structurally different families: exponential cutoffs, hard support truncation, thresholded power laws, tapered Gutenberg-Richter laws, truncated gamma laws, truncated lognormals, and rank-size laws with cutoffs such as generalized Lavalette forms (Aschwanden, 2015, Serra et al., 2015, Ausloos, 2014). Similar empirical behavior on log-log plots does not imply identical mechanisms or identical inferential consequences.

A second misconception is that apparent scaling automatically indicates a genuine scale-free process. Averaging an exponential distribution can produce asymptotic forms $\lambda$ 8 in one dimension and $\lambda$ 9 in two dimensions. These are described as spurious power laws because the exponent reflects Euclidean dimension and the law can be reduced back to an underlying exponential process (Chen, 2013). This suggests that some observed truncations or apparent scaling ranges should be tested against simpler latent-scale alternatives before being interpreted as evidence of complexity.

A third misconception is that pure power laws are the natural null model and truncation is merely a correction. In several data sets the pure power law is either ruled out or not preferred: medical discharge reports favored the truncated power law and lognormal over the pure power law (Quiroz et al., 2020); interbank tails were often best fit by truncated power laws but with substantial ambiguity among alternatives (Vandermarliere et al., 2014); global earthquake moments were better fit by a truncated gamma (Serra et al., 2015); and sinkholes and wildfires were better described by truncated log-normals in the analyzed geoscience datasets (Corral et al., 2018).

The broader implication is methodological rather than doctrinal. A truncated power law is best viewed as one member of a family of heavy-tailed models that encode finite-size, finite-time, or threshold effects in distinct ways. Its usefulness is greatest when the cutoff can be tied to a mechanism and when model selection is performed against plausible alternatives rather than assumed a priori.