---
title: Power Law Explanation
url: https://www.emergentmind.com/topics/power-law
type: topic
---

# Power Law Explanation

A **power law** is a mathematical relationship in which a quantity varies as a power of another quantity. In its generic form,

$$
y=Cx^\kappa,
$$

where $C$ is a scale coefficient and $\kappa$ is an exponent. For probability distributions, power-law behavior commonly refers to a density or mass function of the form

$$
p(x)\propto x^{-\alpha},
$$

or to a survival function with a corresponding exponent. Power laws are associated with scale invariance, heavy tails, rank–size regularities, and nonlinear scaling. They occur in aggregate relations, fluctuation–mean relationships, network degree distributions, spatial interaction, financial returns, cascades, and dynamically generated intensity distributions. The term does not designate a single statistical object or mechanism: a variance–mean law, a Pareto tail, a rank–frequency relation, and a network degree bound are mathematically distinct uses of power-law scaling.

## 1. Mathematical forms and statistical distinctions

For a continuous power law above a lower cutoff $x_{\min}>0$, a normalized density is

$$
p(x)=(\alpha-1)x_{\min}^{\alpha-1}x^{-\alpha},
\qquad x\ge x_{\min},\quad \alpha>1.
$$

Its survival function is

$$
P(X\ge x)=\left(\frac{x}{x_{\min}}\right)^{1-\alpha}.
$$

Consequently, the density exponent and survival-function exponent differ by one. For discrete observations, such as degrees, word counts, or event sizes, the corresponding mass function is

$$
P(X=x)=\frac{x^{-\alpha}}{\zeta(\alpha,x_{\min})},
$$

where $\zeta(\alpha,x_{\min})$ is the generalized zeta function [1407.3492]. A discrete power law should not be fitted with a continuous density merely because both may appear linear on logarithmic axes.

The exponent controls tail decay. A larger $\alpha$ produces a steeper decline and fewer extreme observations; a smaller $\alpha$ produces a heavier tail. For the continuous model, the $m$-th moment exists only if $\alpha>m+1$. Thus, for $1<\alpha\le2$, the mean is infinite; for $2<\alpha\le3$, the mean exists but the variance is infinite [1407.3492].

Power-law terminology must also distinguish probability distributions from other scaling relations. Taylor’s power law is a variance–mean relationship,

$$
V=aM^b,
$$

where $M$ is a mean density, $V$ is its variance, and $b$ is the Taylor exponent [1505.02033]. A rank–size law relates an ordered value to its rank. A network degree law concerns the frequency of vertices with a given degree. A relation such as website hits versus unique hosts,

$$
S=CH^\gamma,
$$

describes scaling between two aggregate variables rather than the probability of observing a large value [1006.3425]. These forms may all be called power laws, but their statistical interpretation and estimation procedures differ.

Power laws are often represented on log–log axes. Taking logarithms of $y=Cx^\kappa$ gives

$$
\log y=\log C+\kappa\log x.
$$

A straight line may therefore be consistent with power-law behavior. It is not, however, sufficient evidence for a power law. Log-normal, exponential, truncated, and other heavy-tailed distributions can appear approximately linear over restricted ranges [1810.07868].

## 2. Cutoffs, truncation, and generalized power laws

An untruncated power law has no finite upper scale, but empirical systems commonly possess lower and upper limits. The lower limit may arise from detection thresholds, catalog completeness, measurement resolution, or the definition of the observable. The upper limit may reflect finite system size, finite resources, capacity constraints, or a finite observation period [1810.07868].

A truncated power law on $a\le x\le b$ has density

$$
f_{\mathrm{tpl}}(x)
=
\frac{\beta-1}{a\left(1-c^{\beta-1}\right)}
\left(\frac{a}{x}\right)^\beta,
\qquad c=\frac{a}{b}.
$$

Its survival function is not a pure power because the upper cutoff contributes an additional term. Consequently, curvature in a cumulative log–log plot does not by itself rule out a truncated power law. A double power law can instead contain two regimes, with exponent $\beta_1$ below a crossover $\theta$ and exponent $\beta_2$ above it. Global earthquake data, for example, support a regime with exponent approximately $1.655$ for moderate events and approximately $2$ above a crossover near moment magnitude $7.6$ [1810.01891].

The Generalized Power Law (GPL) replaces the constant Pareto exponent by a positive function $g(x/\sigma)$:

$$
S(x)=\left(\frac{x}{\sigma}\right)^{-g(x/\sigma)}.
$$

The highlighted GPL specification is

$$
g(z)=\alpha\left(\frac{\log z}{1+\log z}\right)^\beta,
\qquad \alpha>0,\quad \beta>-1.
$$

It allows non-Pareto behavior over the body while retaining a Pareto tail because $g(z)\to\alpha$ as $z\to\infty$. Hence,

$$
S(x)\sim \left(\frac{x}{\sigma}\right)^{-\alpha}.
$$

The ordinary Pareto distribution is recovered when $\beta=0$ [1603.01442]. In Spanish municipal-debt data, the Pareto model was supported only for selected upper tails, whereas the GPL was designed to describe the full range. The GPL had the largest reported BIC in each year, although its bootstrap goodness-of-fit test rejected it in four of seven annual samples.

A related Bayesian framework combines a truncated Zipf-polylog distribution for the body with an integer generalized Pareto distribution for the tail. The threshold is treated as uncertain rather than fixed, and a discrete model indicator compares a power-law bulk with a polylogarithmic alternative. This framework distinguishes a true Zipf boundary, $\theta=1$, from $\theta<1$, since the former has Fréchet-type heavy tails while the latter has Gumbel-type asymptotics [2008.03073].

## 3. Estimation, goodness of fit, and model comparison

The principal methodological difficulty is that the lower cutoff, upper cutoff, exponent, and distributional form are often unknown simultaneously. Estimating an exponent by ordinary least-squares regression on a log–log plot can be biased by binning, unequal variances, finite-size effects, correlated cumulative observations, and an inappropriate treatment of normalization [1407.3492].

For a fixed lower cutoff, the continuous maximum-likelihood estimator is

$$
\widehat{\alpha}
=
1+
n\left[
\sum_{i=1}^n
\log\left(\frac{x_i}{x_{\min}}\right)
\right]^{-1}.
$$

The cutoff can be selected by searching candidate values, estimating the exponent for each, and choosing the value minimizing the Kolmogorov–Smirnov distance between the empirical and fitted tail distributions. The retained sample size must be reported because a large cutoff can improve apparent fit by discarding most observations. Bootstrap resampling should re-estimate both $x_{\min}$ and $\alpha$, since their uncertainty and dependence affect inference [1407.3492].

A fitted power law can often be obtained for finite data even when it is not a plausible generating model. A simulation-based goodness-of-fit test therefore compares the observed Kolmogorov–Smirnov statistic with statistics obtained from synthetic data generated under the fitted model. A large $p$-value means that the observed discrepancy is compatible with sampling variation; it does not prove that the data were generated by a power law.

Alternative distributions should be fitted to the same observations and, where appropriate, the same cutoff. Likelihood-ratio and Vuong comparisons can assess whether a power law is better supported than a log-normal or exponential model, but a nonsignificant comparison means only that the data do not reliably distinguish the candidates. It does not establish a power law [1407.3492].

For bounded power laws, lower and upper cutoffs can also be estimated through order statistics. The expected sample minimum approaches the lower bound $a$ as sample size increases, and the expected sample maximum approaches the upper bound $b$. Averaging minima and maxima over groups of equal size and fitting their dependence on sample size provides an $O(N)$ procedure, rather than the $O(N^3)$ cost of testing all lower–upper cutoff pairs [2303.13456]. The method can extrapolate bounds beyond the smallest and largest observed values, but upper-bound estimation may converge slowly and depends on independence, sample size, and model specification.

## 4. Mechanisms generating power laws

Power laws can emerge from different mechanisms, and no single explanation is universal. In a maximum-entropy model of community sizes, if the total joining cost is linear in size, $w_k\propto k$, the resulting distribution is exponential. If the marginal joining cost decreases approximately as $1/k$, then the accumulated cost is logarithmic,

$$
w_k\sim\mu^\circ\log k,
$$

and maximum entropy gives

$$
p_k\sim k^{-\mu^\circ}.
$$

Equal sharing of entry costs among a joining particle and existing community members produces $\mu_k=\mu^\circ/(1+k)$ and therefore a power-law community-size distribution. Unequal sharing with parameter $s$ yields an asymptotic exponent $\mu^\circ/s$. The model interpolates from exponential behavior at $s=0$ to increasingly heavy-tailed behavior as sharing increases [1112.5906].

In speculative financial markets, buying in response to expected price increases and selling in response to expected decreases can produce a random-coefficient autoregressive process,

$$
r_t=a_t r_{t-1}+e_t.
$$

When $E[\log a]<0$, $\Pr(a>1)>0$, and the relevant moment conditions hold, Kesten’s theorem gives a stationary power-law tail whose exponent $p$ solves

$$
E(a^p)=1.
$$

The mechanism generates a power law but does not force $p=3$; the exponent depends on the distribution of the random feedback coefficient. The model explains heavy-tailed returns but not the persistent dependence of absolute returns associated with volatility clustering [1612.08705].

Nonlinear self-excited Hawkes processes provide another mechanism. Events modify a tension variable, and the intensity is a nonlinear function $\lambda=g(\nu)$. If $g$ grows faster than quadratically, marks are two-sided with non-positive mean, and mark tails decay rapidly, the stationary intensity density has power-law behavior. For exponential activation,

$$
g(\nu)=\lambda_0e^{\beta\nu},
$$

zero-mean marks give

$$
P_{\mathrm{ss}}(\lambda)\propto \lambda^{-2}.
$$

Negative mean marks introduce an additional exponential suppression in the transformed intensity. The mechanism combines self-excitation, memory, multiplicative amplification, and endogenous inhibition [2102.00242].

A cascade master equation produces power laws when cascade growth is accompanied by termination at a rate inversely proportional to cascade size. The discrete continuation probability

$$
P_{i,i+1}=\exp\left(-\frac{k}{i+1}\right)
$$

has a survival product whose logarithm is a harmonic sum, yielding $\pi(x)\sim x^{-k}$. In the continuum approximation,

$$
\frac{\partial p}{\partial t}
=
-a\frac{\partial p}{\partial x}
-\frac{b}{x}p.
$$

Its solution has a finite-time cutoff $x_{\max}(t)=at$ and a power-law regime with exponent $b/a$ below that cutoff [2210.09046].

## 5. Power laws in networks and spatial systems

In network science, a power-law degree distribution is often written

$$
P(\deg=k)\propto k^{-\alpha}.
$$

The deterministic power-law-bounded (PLB) framework does not require exact equality or a random graph model. Instead, it bounds the number of vertices in dyadic degree intervals. This yields

$$
d_{\ge k}
=
O\!\left(
n(t+1)^{\alpha-1}(k+t)^{1-\alpha}
\right),
$$

where $d_{\ge k}$ is the number of vertices with degree at least $k$ [1507.02426]. PLB is therefore an upper-bound condition, not proof that a graph was generated by a power-law mechanism.

For $1<\alpha<2$, simplicity of the graph imposes a second high-degree regime. There are insufficient low-degree vertices to support indefinitely many high-degree vertices through distinct edges, producing a double power-law upper bound with effective exponent

$$
\alpha'=3\alpha-\alpha^2-1.
$$

Additional PLB-neighborhood conditions constrain degree correlations along edges. They permit deterministic algorithmic analyses without assuming random edges. Under such conditions, triangle counting improves substantially, and maximum clique becomes polynomial-time solvable for $\alpha>3$. This does not contradict NP-hardness on broader classes of power-law graphs because the neighborhood restriction is stronger than a degree-distribution condition alone.

Sparsified deep neural networks provide an applied example. After pruning low-magnitude weights and retraining, node degrees in multilayer perceptrons and convolutional networks were modeled layer by layer by discrete truncated power laws. Truncation is expected because finite networks impose maximum possible degrees. During continual learning, an internal preferential-attachment model predicts that new connections are more likely between already well-connected units. The experiments support this tendency particularly when the new task can reuse existing representations [1805.01891].

Geographical gravity models use inverse power-law distance decay,

$$
I_{ij}=G\frac{P_iP_j}{r_{ij}^{b}},
$$

in contrast to exponential decay, $I_{ij}=GP_iP_j e^{-r_{ij}/r_0}$. The inverse power law implies weak but nonzero long-range interaction, whereas the exponential model emphasizes localization. An entropy-maximizing derivation produces a power law if transport cost is assumed logarithmic in distance,

$$
C_{ij}=c_1+c_0\log r_{ij},
$$

rather than linear. For Chinese interregional railway freight, the integrated power-law model had higher $R^2$, lower regression standard error, and a larger $F$ statistic than the exponential model. The fitted gravity exponent was approximately $2.1056$, interpreted by the author as an effective fractal dimension [1503.02915].

## 6. Empirical scaling, interpretation, and limitations

Power laws occur in aggregate empirical relationships that are not tail distributions. In website ratings, daily unique hosts $H$ and hits $S$ decline with rating rank $r$ according to

$$
H=C_Hr^{-\alpha},
\qquad
S=C_Sr^{-\beta}.
$$

Eliminating rank gives

$$
S=CH^\gamma,
\qquad
\gamma=\frac{\beta}{\alpha}.
$$

The common ratio $S/H$ is constant only when $\gamma=1$. For Games websites, the reported fit was $S=3.1053H^{1.1104}$, implying slightly more-than-proportional growth of hits with hosts; for Business and Finances, the displayed relationship was effectively linear, $S=2.1009H$ [1006.3425].

Taylor’s power law concerns fluctuation scaling rather than a heavy-tailed frequency distribution. In ecology, $V=aM^b$ has been reported across thousands of biological species and in non-biological quantities. Its historical significance derives from comparative empirical synthesis rather than absolute priority for the equation. Proposed explanations include aggregation and behavior, demographic and environmental stochasticity, spatial dynamics, skewed sampling distributions, and Tweedie variance functions. The exponent and coefficient can depend on species, environment, sampling scale, and estimation method. In the $N$-stars network evolution model, conditional second moments satisfy

$$
M_{w_1}\sim C E_{w_1}^{\,2},
$$

under the stated parameter and finite-moment conditions, yielding Taylor exponent $2$. This is distinct from the model’s marginal degree-distribution power law [1505.02033; 2001.03023].

Financial studies illustrate both the usefulness and fragility of empirical power-law claims. High-frequency market data have been reported to contain relative-return regimes with event-frequency slopes near $-4$, while volume-related variables often show slopes near $-2$. The reported behavior depends on instrument, year, sampling interval, binning, liquidity, and finite scaling range, and the analysis relies primarily on binned plots and descriptive slopes [2010.01199]. A separate speculative-feedback model derives power-law returns through a random-coefficient process rather than through visual fitting [1612.08705].

In geoscience, statistically supported behavior is heterogeneous. Impact fireballs and Californian earthquakes support untruncated power laws; global earthquakes support a double power law; ocean precipitation clusters and tropical cyclones support truncated power-law regimes; karst sinkholes are better described by truncated log-normal distributions; and wildfires may support either a power-law tail or a broader truncated log-normal depending on region and fitting range [1810.01891]. These findings show why hazard extrapolation must account for cutoffs, crossover regimes, uncertainty, and alternative distributions.

Several misconceptions therefore require qualification. A straight log–log segment is not proof of scale-free generation. A fitted exponent is not necessarily universal, causal, or stable across scales. A large goodness-of-fit $p$-value does not establish that a power law is true. A power-law degree distribution does not establish preferential attachment. A tail exponent should not be compared across density, survival, rank, and cumulative representations without accounting for their different definitions. Finally, a power law may describe only a body, a tail, a fluctuation–mean relation, a rank ordering, or a dynamic transient.

Power-law analysis is consequently best understood as model-based scaling inference. Reliable claims require specification of the observable, discrete or continuous support, density or cumulative convention, valid scaling interval, cutoff structure, estimation method, uncertainty procedure, goodness-of-fit assessment, and comparison with plausible alternatives. Under those conditions, power laws provide compact descriptions of scale-dependent variability across diverse systems, while their mechanisms remain specific to the dynamics, constraints, and measurement structure of each application.

Source: https://www.emergentmind.com/topics/power-law