Power law · aiwiki.page

A power law is a relationship in which one quantity is proportional to a fixed power of another. Its basic form is $f(x)=Cx^k$ , where $C$ is a constant coefficient and $k$ is the exponent. Such relationships describe geometric scaling, inverse-square forces, and many empirical patterns. A power-law relationship between measured quantities is distinct from a power-law probability distribution, which describes how frequently different values occur. (reference.wolfram.com)

Mathematical form and scaling

For a positive input $x$ , a power-law function can be written

f(x)=Cx^k.

The exponent may be any real number. With $C>0$ , positive exponents describe increasing relationships, negative exponents describe decreasing relationships, and $k=0$ gives a constant. Direct substitution shows that doubling $x$ multiplies $f(x)$ by $2^k$ : a square law gives a factor of four, whereas an inverse-square law gives a factor of one-quarter. These are consequences of the power-model form. (reference.wolfram.com)

Its defining scaling property follows algebraically:

f(ax)=a^k f(x),\qquad a>0.

Changing the input scale therefore changes only the output multiplier, not the functional shape. This property is called scale invariance. For dimensional physical quantities, the coefficient supplies the units required by the relationship; scale invariance does not mean that units can be disregarded. (reference.wolfram.com)

For positive $C$ , taking logarithms gives

\log f(x)=\log C+k\log x.

Thus an exact power law appears as a straight line on a log–log plot, with slope $k$ . This transformation is an algebraic identity, but a visually straight segment in noisy observations is not sufficient evidence for a power-law distribution. (reference.wolfram.com)

Power-law probability distributions

In statistics, the term commonly denotes a distribution whose upper tail has density proportional to $x^{-\alpha}$ . A normalized continuous model is

p(x)=\frac{\alpha-1}{x_{\min}} \left(\frac{x}{x_{\min}}\right)^{-\alpha}, \qquad x\ge x_{\min}>0,\quad \alpha>1.

Here $x_{\min}$ is the lower boundary of the modeled range. This is a Pareto distribution of the first kind, with Pareto shape parameter $\alpha-1$ . Authors use different exponent conventions, so the density exponent and the Pareto shape parameter must not be assumed identical. (itl.nist.gov)

Integrating this probability density function yields the survival probability for a random variable $X$ :

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

The survival function, equal to one minus the cumulative distribution function for this continuous model, consequently has an exponent one smaller in magnitude than the density exponent. This difference follows directly from integration and matters when comparing plots. (itl.nist.gov)

A discrete counterpart assigns a probability mass function proportional to $n^{-\alpha}$ on integers above a minimum, with normalization obtained by summation. Zipf’s law instead commonly describes a power-law relationship between an item’s frequency or size and its rank. Rank-based and density-based exponents are related, but are not interchangeable. (cs.cornell.edu)

Tails and moments

Power-law distributions are examples of heavy-tailed distributions: their upper tails decay more slowly than exponentially decreasing tails. Large observations therefore remain comparatively influential. Empirical power-law behavior often applies only above a threshold, rather than throughout a distribution. (cs.cornell.edu)

For the continuous model above, direct evaluation of the defining integral gives

E[X^q]= \frac{\alpha-1}{\alpha-1-q}x_{\min}^{q}, \qquad q<\alpha-1.

At or above that boundary, the integral diverges. The mean is finite only when $\alpha>2$ , and the variance is finite only when $\alpha>3$ . These conditions are mathematical properties of an idealized distribution extending without an upper bound, not assertions that a finite observed dataset contains infinite values. (cs.cornell.edu)

Examples and generating mechanisms

A deterministic example occurs in classical mechanics. Newtonian gravitation between two point masses has magnitude

F=G\frac{m_1m_2}{r^2}.

At fixed masses, force follows a power law in separation with exponent $-2$ . The same expression applies to spherically symmetric bodies using their center-to-center separation. This relationship concerns force and distance, not a frequency distribution of forces. (openstax.org)

Statistical power laws have been investigated in word frequencies, city populations, and network connectivity. Proposed generating mechanisms include preferential attachment, in which already well-connected entities attract additional connections, and critical behavior associated with phase transitions. Different mechanisms can produce similar distributions; observing a power-law tail does not uniquely establish its cause. (cs.cornell.edu)

Empirical identification

Reliable identification requires specifying the fitted range and assessing competing explanations. Linear regression on logarithmically transformed histogram counts can produce inaccurate exponent estimates. A high apparent linearity alone does not establish that observations follow the proposed distribution. (epubs.siam.org)

A widely used framework combines maximum-likelihood estimation of the exponent and lower threshold with goodness-of-fit testing. Simulated samples allow comparison between observed discrepancies and those expected under the fitted model. Likelihood-ratio comparisons assess alternatives such as a lognormal distribution or a power law with an exponential cutoff. Failure to reject a power law indicates compatibility, not proof or superiority over every alternative. Limited samples and uncertain tail boundaries can leave several models plausible. (epubs.siam.org)