The Pareto distribution is a continuous probability distribution whose probability of exceeding a value decreases as a power of that value. In its standard, type I form, it describes a positive random variable with a minimum value and a shape parameter , such that
The parameter controls the relative frequency of very large observations: smaller values produce heavier tails and greater concentration in the largest outcomes.
Named after Vilfredo Pareto, who studied upper-income distributions in the late nineteenth century, the distribution is a basic model in statistics, econometrics, insurance, and the analysis of extremes. Its importance extends beyond exact Pareto models: many distributions have tails that approach a Pareto form even when their central portions look quite different. Its scale invariance, simple formulas, and potentially infinite moments make it both analytically useful and a warning against applying familiar finite-variance methods without checking their assumptions.
Historical origins and the meaning of Pareto behavior
Pareto developed his empirical law of income distribution during the 1890s, including in his Cours d’économie politique of 1896–1897. His analysis concerned the number of people whose incomes exceeded a specified level. Written in modern notation, the relationship is
where counts incomes above , is a constant, and determines the slope on logarithmic coordinates. Converting this count into a probability produces the survival function of a Pareto model above an appropriate lower threshold.
The historical law was therefore a statement about exceedance counts, not about the height of an income histogram. This distinction remains important. A survival function proportional to corresponds to a density proportional to . Exponents quoted in empirical studies can differ by one simply because one study describes the density while another describes the survival function. Rank-based descriptions introduce another convention: if observations are ordered from largest to smallest, size is approximately proportional to rank raised to .
The Pareto distribution is a particularly simple power law, but “power law” is a broader mathematical description rather than the name of one probability distribution. Power-law relationships occur in physical scaling, network degrees, frequencies, and rank-size patterns. Some concern discrete variables; others describe only an asymptotic relationship. A continuous Pareto distribution instead specifies a normalized probability model on a particular support, with a defined lower endpoint and parameter values.
The widely cited Pareto principle, or “80/20 rule,” is related to this history but is not the definition of the distribution. Pareto models allow many degrees of concentration, not just an 80/20 allocation. Moreover, observing that a small fraction of cases accounts for a large fraction of a total does not identify a Pareto distribution. Lognormal distributions, mixtures, and bounded distributions can also produce concentrated outcomes.
Modern usage distinguishes three claims of increasing strength: a dataset is highly unequal; its upper tail is approximately power-law; and its observations above a threshold follow an exact Pareto distribution. These claims require different evidence. A large maximum or a visually straight portion of a logarithmic plot may motivate a tail model, but neither establishes the exact distribution. In applications, the Pareto law is commonly an approximation for high incomes, large claims, or other upper-tail observations, rather than a description of the entire population.
Definition, parameters, and scale invariance
For the type I distribution, the cumulative distribution function is
Its probability density function is
Both parameters must be positive. The scale parameter has the same units as , while the shape parameter is dimensionless. Changing the unit of measurement changes , but not .
The survival function, denoted , often provides the clearest interpretation. For any multiplier ,
Thus, multiplying an already admissible threshold by ten reduces its exceedance probability by , regardless of the threshold’s absolute magnitude. For , an observation is one hundred times less likely to exceed than to exceed . For , the reduction is only tenfold.
A stronger expression of this property uses conditional probability:
Consequently, conditional on exceeding , the ratio has a Pareto distribution with minimum and the same shape parameter. Raising the threshold changes the scale but preserves the form. This is multiplicative threshold stability, not the additive memorylessness of an exponential random variable.
The quantile function is
In particular, the median is . The density decreases throughout its support, so its maximum occurs at the lower endpoint. Nevertheless, the distribution has no finite upper bound. A very large observation is not outside the model merely because it lies far above most observations.
The hazard rate, defined as density divided by survival probability, is
It decreases with . When the variable represents duration, this means that, conditional on having already lasted to , the instantaneous termination rate is lower at larger . Such an interpretation depends on the application: a decreasing mathematical hazard does not by itself establish a physical aging mechanism.
Terminology surrounding the shape parameter requires care. Some fields call the tail index; others reserve “tail index” for its reciprocal, . Reporting the survival formula alongside the parameter avoids this ambiguity.
Moments, transformations, and simulation
The Pareto distribution illustrates how a valid probability model can lack a finite mean or variance. For a real exponent ,
The integral converges precisely when , giving
Because the support is bounded away from zero, all negative-order moments exist. The limitation concerns sufficiently large positive powers and the upper tail.
The expected value is
For , the expectation is infinite. The variance is
For , the mean is finite but the variance is infinite. Higher standardized moments require still larger shape parameters: conventional skewness requires , and kurtosis requires .
These distinctions have practical consequences. With , the mean is and the variance is . With , the mean is , but no finite population variance exists. With , every individual observation is finite with probability one, yet the population mean is infinite. An infinite expectation describes the mathematical accumulation of increasingly rare, increasingly large possibilities; it does not mean that an observed value must itself be infinite.
A useful transformation is
For ,
so follows an exponential distribution with rate . In particular,
Logarithmic moments remain finite even when ordinary positive moments diverge. This transformation underlies simple estimators, diagnostic plots, and exact confidence procedures.
For simulation, if follows a uniform distribution on , then
has the Pareto law. Equivalently, an exponential draw can be transformed through . In numerical work, computing first can avoid intermediate overflow. Extremely small , or a uniform draw very close to zero, can produce values beyond a computer’s floating-point range.
The positive moment-generating function is infinite for every , since exponential growth eventually dominates polynomial tail decay. This does not prevent use of Laplace transforms with a negative exponent, or Monte Carlo methods, but it rules out arguments that assume an ordinary moment-generating function exists near zero.
Inequality, Lorenz curves, and top shares
For nonnegative quantities with a finite mean, the Lorenz curve describes the cumulative share of the total held by the bottom fraction of the population. For a Pareto distribution with ,
The scale parameter cancels. Thus, within this model, multiplying all incomes or holdings by the same constant changes their levels but not their relative inequality.
The corresponding Gini coefficient is
As increases, relative dispersion decreases and approaches zero. As decreases toward from above, approaches one. At , for example, . These formulas assume a Pareto distribution for the whole modeled population, not merely for its uppermost observations.
The fraction of the total held by the top fraction is
This formula gives a precise connection to the 80/20 rule. Requiring the top to hold gives
which implies approximately
An 80/20 split therefore corresponds to one particular shape parameter. It is not shared by all Pareto distributions, and a single observed split is insufficient to demonstrate a Pareto law.
Another useful identity concerns the average value above a threshold :
The ratio of this conditional mean to the threshold is constant. The conditional mean excess is consequently
These relationships help connect tail parameters to grouped income data and to losses above insurance deductibles.
A critical qualification arises when a Pareto model applies only above a high threshold. The resulting Lorenz and Gini formulas then describe the conditional upper-tail population, not society as a whole. Overall inequality also depends on the fraction below the threshold, their distribution, and the upper tail’s contribution to total resources.
For , the usual population Lorenz curve normalized by a finite mean is unavailable. A finite dataset still has a calculable empirical Lorenz curve and Gini coefficient, because its observed sum is finite. That fact does not supply a finite population mean or validate the preceding population formulas. Economic interpretation also depends on the measured quantity: income, wealth, and consumption have different supports, and net wealth may be zero or negative, unlike a type I Pareto variable.
Related distributions and broader tail classes
An exact Pareto distribution is one member of the much broader class of heavy-tailed distributions. In a common probabilistic definition, a nonnegative distribution is heavy-tailed if its positive exponential moments are infinite. Pareto tails satisfy this condition, but not every heavy-tailed distribution is Pareto. The lognormal distribution, for example, is heavy-tailed under this definition without having a power-law survival function.
A particularly important class consists of distributions with regularly varying tails:
where is slowly varying, meaning that
for every fixed . An exact Pareto tail has constant . More generally,
This is asymptotic scale invariance: the Pareto relationship becomes increasingly accurate at high thresholds without necessarily holding at moderate values.
The generalized Pareto distribution has survival function
on its admissible support, with scale . Its exponential limit occurs at . If is type I Pareto and , then the excess , conditional on , follows this distribution with
Positive corresponds to a power-law upper tail; negative corresponds to a finite upper endpoint.
The Lomax distribution, often called Pareto type II, is obtained by subtracting the minimum:
It has support beginning at zero rather than at a positive lower bound. Software conventions differ, so a function named “Pareto” may implement a type I model, a Lomax model, or a location-scale extension. The support and survival formula identify the convention more reliably than the name.
Other transformations are equally useful. The reciprocal has a beta distribution with parameters . A positive power is Pareto with minimum and shape . A distribution truncated at an upper endpoint instead has survival
Truncation makes all positive moments finite and removes exact scale invariance near .
Discrete rank-frequency laws, including the Zipf distribution, are related through power-law scaling but are not interchangeable with a continuous Pareto model. Their normalization, likelihoods, and interpretation depend on integer-valued support.
Extremes, sums, and limit laws
The Pareto distribution plays a central role in extreme value theory. Let be independent observations with the same Pareto distribution, and let
The characteristic size of the maximum is therefore proportional to , rather than to the logarithm of sample size as in an exponential-tail setting.
For ,
The limit is a Fréchet distribution. This result extends to regularly varying tails with suitable normalization. It explains why Pareto-type tails belong to a common domain of attraction even when their central distributions differ substantially.
Sums show a different but related sensitivity to . When , the law of large numbers implies that the sample mean converges to the finite population mean. When , the variance is also finite, and the ordinary central limit theorem gives a normal limit for the centered sample mean under square-root-of- scaling.
For , the sample mean remains consistent, but ordinary finite-variance normal theory does not apply. Centered sums, normalized on the order of , approach a non-Gaussian stable distribution. The characteristic scale of sample-mean fluctuations is then , which decreases more slowly than . At the boundary , logarithmic corrections enter the normalization.
For , sums normalized by also have a stable limit, but there is no finite mean around which to center them. A few of the largest observations can contribute a substantial fraction of the total even in large samples. The boundary again needs special treatment, including logarithmic centering in the appropriate stable-limit formulation.
A complementary result concerns rare large sums. For a fixed positive integer ,
This is the subexponential, or “one large jump,” property: an exceptionally large aggregate is asymptotically driven by one exceptionally large component. The statement assumes independent nonnegative components and fixed ; it is not an unrestricted approximation for arbitrarily growing portfolios.
These results distinguish three questions that are sometimes conflated: whether an average converges, how quickly its fluctuations decrease, and what drives extreme aggregate outcomes. A finite mean answers only the first. It does not guarantee a finite standard error under conventional formulas or make large observations negligible at realistic sample sizes.
Parameter estimation and uncertainty
Suppose are independent Pareto observations and the lower threshold is fixed and known. The likelihood function is
Differentiating its logarithm gives the maximum likelihood estimator
It depends on logarithmic ratios rather than the arithmetic mean, so it remains usable when the population mean or variance is infinite.
Let
The exponential transformation implies that has a gamma distribution with shape and rate . Therefore
The maximum likelihood estimator is upward biased in finite samples. The estimator is unbiased for , although unbiasedness alone does not determine which estimator has the best overall performance.
The same transformation supplies an exact confidence interval. Since
a confidence interval with coverage is
where is the -quantile of the chi-squared distribution with degrees of freedom. These exact formulas require a genuinely fixed lower endpoint and an exact Pareto model.
When is unknown, its maximum likelihood estimate is the sample minimum , and the corresponding shape estimate is
The support now depends on an estimated parameter. Consequently, uncertainty formulas derived for known cannot simply be reused unchanged. Selecting a tail threshold from a larger dataset introduces an additional inferential problem: the chosen threshold need not be the true lower endpoint of a population-wide Pareto distribution.
For upper-tail estimation, write the ordered observations as
The Hill estimator of the reciprocal tail index , using the largest observations, is
Under suitable regular-variation assumptions, it estimates the tail index beyond exact Pareto models. Its behavior depends strongly on : too few upper observations produce high variability, while too many may include observations outside the Pareto-like region.
With known , Bayesian inference also has a simple form. A gamma prior on , with shape and rate , gives a gamma posterior with shape and rate . The simplicity follows from the exponential likelihood for log-ratios, not from an assumption that the original data have finite moments.
Empirical assessment and model limitations
A common diagnostic plots empirical survival probabilities against observed values on logarithmic axes. An exact Pareto model implies
so the theoretical plot is a straight line with slope . Such plots are useful for locating possible scaling regions and detecting curvature, truncation, or changes in tail behavior. They do not, by themselves, provide a reliable test of a Pareto model.
Several statistical difficulties affect graphical interpretation. Empirical survival probabilities at different thresholds are dependent, because they are calculated from overlapping observations. Their uncertainty is also uneven, becoming especially large in the sparsely observed upper tail. Fitting a straight line by ordinary least squares to logarithmic histogram counts adds further sensitivity to bin boundaries, empty bins, and the transformation of count errors. The resulting slope is generally not the same as a properly estimated likelihood parameter.
A more explicit procedure fits a Pareto law above a candidate threshold, measures the discrepancy between fitted and empirical distributions, and assesses that discrepancy through simulation. A Kolmogorov–Smirnov distance is one possible discrepancy measure. When parameters or thresholds are estimated from the data, simulated datasets must undergo the same fitting procedure to calibrate a meaningful goodness-of-fit assessment. Standard critical values for a completely specified distribution do not automatically apply.
Threshold choice creates a bias–variance tradeoff. A low threshold retains many observations but may include a non-Pareto central region. A high threshold may better approximate the limiting tail while leaving too few observations for precise estimation. Parameter-stability plots and Hill plots reveal how estimates change as the threshold moves, but a visually stable segment is not conclusive evidence. Many empirical distributions approach their limiting tail form slowly.
Competing models matter. A lognormal or stretched-exponential tail can resemble a straight line over a limited logarithmic range. A truncated Pareto may fit moderate extremes while predicting far fewer exceptionally large outcomes than an unbounded model. Likelihood comparisons require attention to common support, parameter estimation, and whether the alternatives are nested. A failure to reject a Pareto model means that the available data do not show a sufficiently large discrepancy under the chosen test; it does not prove the tail is exactly Pareto.
Data collection can alter the apparent tail. Top-coded records, reporting thresholds, rounding, missing high-value observations, and changes in measurement definitions affect inference. Time dependence and clustering can reduce the information available relative to an independent sample of the same size. An exact independent-observation model should therefore be distinguished from a descriptive tail fit.
Extrapolation is particularly sensitive to the shape parameter. An estimated survival probability at involves , so uncertainty in has an increasingly large effect as the extrapolation moves farther beyond the observed threshold.
Generative mechanisms and practical applications
Several different mechanisms can generate Pareto or Pareto-like tails. One transparent construction combines exponential growth with an exponentially distributed age or stopping time. Suppose
where and is exponential with rate . Then
The resulting shape parameter is . Faster growth relative to the stopping rate produces a heavier tail. The construction is exact, but its relevance to a particular dataset depends on whether its growth and stopping assumptions are credible.
Random multiplicative recursions provide a broader mechanism. In models of the form
a stationary distribution can have a power-law tail under suitable contraction, moment, and regularity conditions. A characteristic tail equation is
This is associated with Kesten-type processes. The additive term helps prevent collapse toward zero, while occasional multiplicative expansions generate large values. The equation alone is not sufficient: existence of a stationary solution and the relevant conditions on and are essential.
In economics, Pareto tails are used for high incomes, wealth, firm sizes, and related quantities. Applications often combine a flexible distribution for the main body with a Pareto upper tail. This separation matters because entry, taxes, accounting definitions, household composition, and growth processes can affect different parts of the distribution differently. A fitted tail exponent is a descriptive parameter; it does not by itself identify the mechanism producing inequality.
In insurance, the distribution models large claim severity and payments above high deductibles. For a deductible , the expected uncapped payment is
where . If , this expectation is infinite under the unbounded model. Policy limits change the calculation by capping payments and making the relevant payment expectation finite.
For financial loss modeling, the quantile at probability level is
For a continuous Pareto loss with , expected shortfall is
The multiplier grows rapidly as approaches one, showing how losses beyond a fixed quantile can be much more sensitive to tail shape than the quantile itself.
Pareto-like models also occur in workload sizes, transmission durations, and file sizes in computing. Large jobs can dominate totals and waiting times, especially when second moments are large or infinite. These applications require identifying the variable being modeled: a heavy-tailed distribution of job sizes is not automatically a heavy-tailed model for arrivals, dependence, or system capacity. Physical limits, changing technology, and finite observation windows can also produce effective truncation that is absent from the idealized distribution.