# Zipf's law

Zipf's law (/zɪf/) is an empirical law formulated using mathematical statistics that refers to the fact that many types of data studied in the physical and social sciences can be approximated with a Zipfian distribution, one of a family of related discrete power law probability distributions. Zipf distribution is related to the zeta distribution, but is not identical.

Parameters Probability mass function Zipf PMF for N = 10 on a log–log scale. The horizontal axis is the index k . (Note that the function is only defined at integer values of k. The connecting lines do not indicate continuity.) Cumulative distribution function Zipf CDF for N = 10. The horizontal axis is the index k . (Note that the function is only defined at integer values of k. The connecting lines do not indicate continuity.) $s\geq 0\,$ (real)$N\in \{1,2,3\ldots \}$ (integer) $k\in \{1,2,\ldots ,N\}$ ${\frac {1/k^{s}}{H_{N,s}}}$ where HN,s is the Nth generalized harmonic number ${\frac {H_{k,s}}{H_{N,s}}}$ ${\frac {H_{N,s-1}}{H_{N,s}}}$ $1\,$ ${\frac {H_{N,s-2}}{H_{N,s}}}-{\frac {H_{N,s-1}^{2}}{H_{N,s}^{2}}}$ ${\frac {s}{H_{N,s}}}\sum \limits _{k=1}^{N}{\frac {\ln(k)}{k^{s}}}+\ln(H_{N,s})$ ${\frac {1}{H_{N,s}}}\sum \limits _{n=1}^{N}{\frac {e^{nt}}{n^{s}}}$ ${\frac {1}{H_{N,s}}}\sum \limits _{n=1}^{N}{\frac {e^{int}}{n^{s}}}$ For example, Zipf's law states that given some corpus of natural language utterances, the frequency of any word is inversely proportional to its rank in the frequency table. Thus the most frequent word will occur approximately twice as often as the second most frequent word, three times as often as the third most frequent word, etc.: the rank-frequency distribution is an inverse relation. For example, in the Brown Corpus of American English text, the word the is the most frequently occurring word, and by itself accounts for nearly 7% of all word occurrences (69,971 out of slightly over 1 million). True to Zipf's Law, the second-place word of accounts for slightly over 3.5% of words (36,411 occurrences), followed by and (28,852). Only 135 vocabulary items are needed to account for half the Brown Corpus.

The law is named after the American linguist George Kingsley Zipf (1902–1950), who popularized it and sought to explain it (Zipf 1935, 1949), though he did not claim to have originated it. The French stenographer Jean-Baptiste Estoup (1868–1950) appears to have noticed the regularity before Zipf.[not verified in body] It was also noted in 1913 by German physicist Felix Auerbach (1856–1933).

## Other datasets

The same relationship occurs in many other rankings unrelated to language, such as the population ranks of cities in various countries, corporation sizes, income rankings, ranks of number of people watching the same TV channel, and so on. The appearance of the distribution in rankings of cities by population was first noticed by Felix Auerbach in 1913. Empirically, a data set can be tested to see whether Zipf's law applies by checking the goodness of fit of an empirical distribution to the hypothesized power law distribution with a Kolmogorov–Smirnov test, and then comparing the (log) likelihood ratio of the power law distribution to alternative distributions like an exponential distribution or lognormal distribution. When Zipf's law is checked for cities, a better fit has been found with exponent s = 1.07; i.e. the $n^{th}$  largest settlement is ${\frac {1}{n^{1.07}}}$  the size of the largest settlement.

## Theoretical review

Zipf's law is most easily observed by plotting the data on a log-log graph, with the axes being log (rank order) and log (frequency). For example, the word "the" (as described above) would appear at x = log(1), y = log(69971). It is also possible to plot reciprocal rank against frequency or reciprocal frequency or interword interval against rank. The data conform to Zipf's law to the extent that the plot is linear.

Formally, let:

• N be the number of elements;
• k be their rank;
• s be the value of the exponent characterizing the distribution.

Zipf's law then predicts that out of a population of N elements, the normalized frequency of elements of rank k, f(k;s,N), is:

$f(k;s,N)={\frac {1/k^{s}}{\sum \limits _{n=1}^{N}(1/n^{s})}}$

Zipf's law holds if the number of elements with a given frequency is a random variable with power law distribution $p(f)=\alpha f^{-1-1/s}.$ 

It has been claimed that this representation of Zipf's law is more suitable for statistical testing, and in this way it has been analyzed in more than 30,000 English texts. The goodness-of-fit tests yield that only about 15% of the texts are statistically compatible with this form of Zipf's law. Slight variations in the definition of Zipf's law can increase this percentage up to close to 50%.

In the example of the frequency of words in the English language, N is the number of words in the English language and, if we use the classic version of Zipf's law, the exponent s is 1. f(ks,N) will then be the fraction of the time the kth most common word occurs.

The law may also be written:

$f(k;s,N)={\frac {1}{k^{s}H_{N,s}}}$

where HN,s is the Nth generalized harmonic number.

The simplest case of Zipf's law is a "1f function." Given a set of Zipfian distributed frequencies, sorted from most common to least common, the second most common frequency will occur ½ as often as the first. The third most common frequency will occur ⅓ as often as the first. The fourth most common frequency will occur ¼ as often as the first. The nth most common frequency will occur 1n as often as the first. However, this cannot hold exactly, because items must occur an integer number of times; there cannot be 2.5 occurrences of a word. Nevertheless, over fairly wide ranges, and to a fairly good approximation, many natural phenomena obey Zipf's law.

In human languages, word frequencies have a very heavy-tailed distribution, and can therefore be modeled reasonably well by a Zipf distribution with an s close to 1.

As long as the exponent s exceeds 1, it is possible for such a law to hold with infinitely many words, since if s > 1 then

$\zeta (s)=\sum _{n=1}^{\infty }{\frac {1}{n^{s}}}<\infty .\!$

where ζ is Riemann's zeta function.

## Statistical explanation

A plot of the rank versus frequency for the first 10 million words in 30 Wikipedias (dumps from October 2015) in a log-log scale.

Although Zipf’s Law holds for all languages, even non-natural ones like Esperanto, the reason is still not well understood. However, it may be partially explained by the statistical analysis of randomly generated texts. Wentian Li has shown that in a document in which each character has been chosen randomly from a uniform distribution of all letters (plus a space character), the "words" with different lengths follow the macro-trend of the Zipf's law (the more probable words are the shortest with equal probability). Vitold Belevitch in a paper, On the Statistical Laws of Linguistic Distribution offered a mathematical derivation. He took a large class of well-behaved statistical distributions (not only the normal distribution) and expressed them in terms of rank. He then expanded each expression into a Taylor series. In every case Belevitch obtained the remarkable result that a first-order truncation of the series resulted in Zipf's law. Further, a second-order truncation of the Taylor series resulted in Mandelbrot's law.

The principle of least effort is another possible explanation: Zipf himself proposed that neither speakers nor hearers using a given language want to work any harder than necessary to reach understanding, and the process that results in approximately equal distribution of effort leads to the observed Zipf distribution.

Similarly, preferential attachment (intuitively, "the rich get richer" or "success breeds success") that results in the Yule–Simon distribution has been shown to fit word frequency versus rank in language and population versus city rank better than Zipf's law. It was originally derived to explain population versus rank in species by Yule, and applied to cities by Simon.

## Related laws

A plot of word frequency in Wikipedia (November 27, 2006). The plot is in log-log coordinates. x  is rank of a word in the frequency table; y  is the total number of the word’s occurrences. Most popular words are "the", "of" and "and", as expected. Zipf's law corresponds to the middle linear portion of the curve, roughly following the green (1/x)  line, while the early part is closer to the magenta (1/x0.5) line while the later part is closer to the cyan (1/(k + x)2.0) line. These lines correspond to three distinct parameterizations of the Zipf–Mandelbrot distribution, overall a broken power law with three segments: a head, middle, and tail.

Zipf's law in fact refers more generally to frequency distributions of "rank data," in which the relative frequency of the nth-ranked item is given by the Zeta distribution, 1/(nsζ(s)), where the parameter s > 1 indexes the members of this family of probability distributions. Indeed, Zipf's law is sometimes synonymous with "zeta distribution," since probability distributions are sometimes called "laws". This distribution is sometimes called the Zipfian distribution.

A generalization of Zipf's law is the Zipf–Mandelbrot law, proposed by Benoît Mandelbrot, whose frequencies are:

$f(k;N,q,s)={\frac {[{\text{constant}}]}{(k+q)^{s}}}.\,$

The "constant" is the reciprocal of the Hurwitz zeta function evaluated at s. In practice, as easily observable in distribution plots for large corpora, the observed distribution can be modelled more accurately as a sum of separate distributions for different subsets or subtypes of words that follow different parameterizations of the Zipf–Mandelbrot distribution, in particular the closed class of functional words exhibit s lower than 1, while open-ended vocabulary growth with document size and corpus size require s greater than 1 for convergence of the Generalized Harmonic Series.

Zipfian distributions can be obtained from Pareto distributions by an exchange of variables.

The Zipf distribution is sometimes called the discrete Pareto distribution because it is analogous to the continuous Pareto distribution in the same way that the discrete uniform distribution is analogous to the continuous uniform distribution.

The tail frequencies of the Yule–Simon distribution are approximately

$f(k;\rho )\approx {\frac {[{\text{constant}}]}{k^{\rho +1}}}$

for any choice of ρ > 0.

In the parabolic fractal distribution, the logarithm of the frequency is a quadratic polynomial of the logarithm of the rank. This can markedly improve the fit over a simple power-law relationship. Like fractal dimension, it is possible to calculate Zipf dimension, which is a useful parameter in the analysis of texts.

It has been argued that Benford's law is a special bounded case of Zipf's law, with the connection between these two laws being explained by their both originating from scale invariant functional relations from statistical physics and critical phenomena. The ratios of probabilities in Benford's law are not constant. The leading digits of data satisfying Zipf's law with s = 1 satisfy Benford's law.

$n$  Benford's law: $P(n)=$
$\log _{10}(n+1)-\log _{10}(n)$
${\frac {\log(P(n)/P(n-1))}{\log(n/(n-1))}}$
1 0.30103000
2 0.17609126 −0.7735840
3 0.12493874 −0.8463832
4 0.09691001 −0.8830605
5 0.07918125 −0.9054412
6 0.06694679 −0.9205788
7 0.05799195 −0.9315169
8 0.05115252 −0.9397966
9 0.04575749 −0.9462848

## Applications

In information theory, a symbol (event, signal) of probability $p$  contains $\log _{2}(1/p)$  bits of information. Hence, Zipf law for natural numbers: $\Pr(x)\approx 1/x$  is equivalent with number $x$  containing $\log _{2}(x)$  bits of information. To add information from a symbol of probability $p$  into information already stored in a natural number $x$ , we should go to $x'$  such that $\log _{2}(x')\approx \log _{2}(x)+\log _{2}(1/p)$ , or equivalently $x'\approx x/p$ . For instance, in standard binary system we would have $x'=2x+s$ , what is optimal for $\Pr(s=0)=\Pr(s=1)=1/2$  probability distribution. Using $x'\approx x/p$  rule for a general probability distribution is the base of Asymmetric Numeral Systems family of entropy coding methods used in data compression, which state distribution is also governed by Zipf law.

Zipf's law also has been used for extraction of parallel fragments of texts out of comparable corpora.