The lognormal distribution is a continuous probability distribution for a positive random variable whose natural logarithm follows a normal distribution. Equivalently, exponentiating a normally distributed variable produces a lognormally distributed variable. Its values are strictly positive, and its density is typically asymmetric, with a concentration of relatively small observations and a long right-hand tail.
Lognormal models arise naturally when quantities change through the multiplication of many random factors rather than through the addition of random increments. They are used for particle sizes, material properties, environmental concentrations, incomes, and financial prices, although their suitability depends on the generating process and the observed range of values. Their mathematical appeal comes from a useful combination: calculations involving logarithms reduce to normal-distribution calculations, while the original-scale distribution accommodates substantial skewness and variation across orders of magnitude.
The distribution also illustrates several important distinctions in probability and statistical modeling. Its arithmetic mean differs from its median and geometric mean; all its positive moments exist, but its moment-generating function is infinite for positive arguments; and its moments do not uniquely identify its distribution. Products and ratios of jointly lognormal variables remain lognormal, whereas sums generally do not.
Definition, density, and parameterizations
A variable has a lognormal distribution with parameters and when
The notation is common, although some authors put , rather than , in the second parameter position. Stating explicitly that and are the mean and standard deviation of prevents this ambiguity. Neither is, in general, the mean or standard deviation of .
The probability density function is
and is zero for . The factor comes from the derivative of the logarithmic transformation. It is essential: merely drawing a normal-shaped curve against a logarithmic horizontal axis does not give the density per unit of the original variable. A histogram using equal-width logarithmic bins represents a different density scale from one using equal-width original-scale bins.
Writing for the standard normal cumulative distribution function, the lognormal cumulative distribution is
Consequently, its quantile function is
These expressions make probability calculations straightforward even though the density itself is not symmetric. For example, an interval containing the central 95% of the population is approximately
Its endpoints are equally far from multiplicatively, not additively.
An alternative parameterization uses as a scale parameter and as a shape parameter. Multiplying changes the units or overall scale without changing relative dispersion. The population geometric mean is , and the geometric standard deviation is . Thus, the interval from to contains approximately 68.27% of the population.
A three-parameter version introduces a location threshold :
Its support is , and , rather than , is normal. This shifted model differs materially from the ordinary two-parameter distribution, particularly in estimation near the lower endpoint. Changing the base of the logarithm only rescales the log-scale parameters; it does not define a different family.
Strictly, a logarithm is applied to a dimensionless ratio. For a dimensional quantity, the definition can be written , where is a fixed reference unit. Changing measurement units then shifts but leaves unchanged.
Moments, shape, and tail behavior
The moments of a lognormal variable follow directly from the exponential moments of a normal variable. For every real ,
Thus, both positive and negative power moments are finite. In particular, the expected value and variance are
The median is , while the mode is . For every nondegenerate lognormal distribution, mode, median, and mean occur in that increasing order.
The coefficient of variation, defined as standard deviation divided by mean, is
It depends only on , explaining why this parameter controls relative dispersion. Skewness and excess kurtosis are respectively
and
Both grow rapidly with . For example, if and , the median is 1, the mean is approximately 1.649, and the mode is approximately 0.368. Describing such a population with a single “typical value” therefore requires specifying which measure is intended.
The right tail is heavier than an exponential tail but lighter than every fixed inverse power. More precisely, if , normal-tail asymptotics give
as . It follows that for every , whereas for every . The lognormal is therefore heavy-tailed under the definition based on the absence of finite positive exponential moments, but it is not a power-law distribution.
Although every power moment exists, is infinite whenever . The formal series constructed from the moments consequently does not supply a moment-generating function around zero. The Laplace transform exists for , but does not generally have an elementary expression.
More unusually, the lognormal distribution is moment-indeterminate. To see the mechanism, let denote its density and set
These are nonnegative densities, and integration against gives the same result for every nonnegative integer . The oscillatory contribution vanishes after exponential tilting of the underlying normal density. Distinct distributions can therefore share every ordinary moment of a lognormal variable. Knowledge of all integer moments alone is insufficient to establish lognormality.
Multiplicative origins and historical development
The central mechanism behind lognormal behavior is multiplication. Suppose
Taking logarithms gives
If the logarithms of the factors are independent, or satisfy suitable weak-dependence conditions, and meet the relevant finite-variance assumptions, the central limit theorem can make their standardized sum approximately normal. The product is then approximately lognormal over the range where that normal approximation is accurate.
This is a multiplicative counterpart of the familiar additive explanation of normal variation. Errors that add on an absolute scale tend to motivate normal models; errors that act proportionally tend to motivate lognormal models. A sequence of percentage increases and decreases acts multiplicatively on the level of a quantity. Likewise, variation in a product of positive physical factors becomes additive after taking logarithms. The approximation is not automatic: strong dependence, a dominant factor, or sufficiently irregular log-factor tails can prevent normal convergence.
A continuous-time version is geometric Brownian motion. If
where is a Wiener process, the solution is
For each fixed , is lognormal with log-scale variance . The correction arises from stochastic calculus; omitting it confuses the drift of the level with the drift of its logarithm. This stochastic process is not stationary: its marginal parameters change with time, and values at different times are dependent.
Historical development preceded this modern process formulation. In 1879, Francis Galton discussed the geometric mean and multiplicative variation, and Donald McAlister gave a mathematical treatment of a distribution based on normally distributed logarithms. Their work helped establish a framework for variation measured by ratios rather than differences. In 1931, Robert Gibrat developed proportional-growth ideas in economic analysis, associating the size of an economic unit with random growth rates that did not depend on its current size.
John Aitchison and J. A. C. Brown’s 1957 monograph, The Lognormal Distribution, consolidated mathematical properties and applications. Subsequent work extended its role in reliability, environmental measurement, communications, and finance. These developments also clarified that a plausible multiplicative story does not uniquely identify a distribution.
In particular, proportional growth over a fixed time horizon can produce a lognormal cross-section under suitable assumptions, but entry, exit, size constraints, and heterogeneous ages can alter the resulting population substantially. Pure geometric Brownian motion does not by itself produce a stationary distribution of sizes. Models incorporating stopping times or additional economic mechanisms can yield power-law tails instead. The connection between multiplication and lognormality is consequently a mechanism with explicit conditions, not a universal law of growth.
Transformations, dependence, and multivariate distributions
The lognormal family is closed under several multiplicative transformations. If , then for and nonzero real ,
In particular, the reciprocal has parameters and . These properties follow immediately by taking logarithms. They explain why changing physical units, forming reciprocal rates, or raising positive quantities to powers often preserves lognormal structure.
Products and ratios require a statement about joint dependence. If have jointly normal logarithms, then
is lognormal whenever the corresponding linear combination of logarithms has positive variance. Separate lognormal marginal distributions are not enough: their logarithms must have the appropriate joint normal structure. Independence of the lognormal variables is one sufficient condition, but it is not necessary.
The multivariate lognormal distribution is defined by
using the multivariate normal distribution. If is positive definite, the joint density on the positive orthant is
The product of reciprocal coordinates is the absolute Jacobian determinant of the logarithmic transformation.
Dependence on the original scale differs from dependence on the logarithmic scale. If , then the covariance is
and the correlation coefficient is
Consequently, a desired original-scale correlation cannot generally be inserted directly into a normal covariance matrix. Even a valid correlation matrix on the original scale may be incompatible with specified lognormal marginal parameters.
The restriction is especially visible for negative dependence. With equal log-scale standard deviations , the most negative possible correlation under a joint lognormal construction is , attained when the two logarithms have correlation . Large marginal dispersion therefore limits how negatively correlated the original variables can be.
Conditioning also remains tractable. The logarithms of one subset, conditional on fixed values of another subset, have a normal conditional distribution, so the corresponding conditional distribution on the original scale is lognormal. This supports prediction and simulation of dependent positive quantities. For simulation, a factorization such as Cholesky decomposition generates correlated normal logarithms, which are then exponentiated. By contrast, sums and differences are not generally lognormal; a difference can even take negative values.
Parameter estimation and statistical inference
For independent positive observations , estimation of the two-parameter model reduces to normal inference for . The log-likelihood is
The Jacobian term does not involve the parameters, so it does not change their maximizing values. The maximum-likelihood estimators are
The usual unbiased estimate of the log-scale variance instead uses denominator :
This distinction matters in small samples. Moreover, absence of bias for an estimator of or does not imply absence of bias after nonlinear transformation. The maximum-likelihood estimator of the population mean is
but it is not generally unbiased. The ordinary sample mean of the original observations is unbiased for the population mean, although its variability can be substantial when the right tail is pronounced.
Exact normal-theory inference is available for log-scale parameters. With independent observations and ,
has a Student distribution with degrees of freedom. Exponentiating the endpoints of a confidence interval for gives an interval for the population median . It does not give an interval for the arithmetic mean. Similarly,
has a chi-squared distribution, providing exact intervals for log-scale variance.
A prediction interval for one future observation has a different uncertainty factor:
The extra 1 under the square root accounts for variation in the future observation itself. Intervals for the arithmetic mean involve both log-scale parameters nonlinearly; likelihood methods, asymptotic approximations, or bootstrap methods are commonly used.
Moment estimation offers another route. Given an original-scale mean and variance , the matching parameters satisfy
Replacing and with sample quantities yields method-of-moments estimates, but large observations can dominate them.
Censoring and truncation require modified likelihoods. An observation known only to lie below a detection limit contributes a cumulative probability, not a density at an arbitrarily substituted value. Estimating an unknown shift parameter in the three-parameter model introduces additional complications because the support depends on that parameter. The simple transformed-normal formulas apply directly only to the unshifted, uncensored setting.
Lognormal regression and interpretation on different scales
A standard lognormal regression model writes
with errors independent across observations conditional on the predictors. Under these assumptions, fitting linear regression by ordinary least squares to the logarithms also produces maximum-likelihood estimates of the regression coefficients. The model specifies the entire conditional distribution of a positive response, not merely a convenient transformation for plotting.
The conditional median and mean are different:
Exponentiating a fitted log response therefore estimates the conditional median under the model, not the conditional arithmetic mean. This distinction follows from Jensen’s inequality: exponentiation is convex, so the exponential of an expectation is generally smaller than the expectation of the exponential.
For a predictor entering linearly without interactions, a one-unit increase changes the conditional median by a factor , holding other predictors fixed. Under constant , it changes the conditional mean by the same factor. The exact percentage change is ; the frequently used interpretation is an approximation valid for small coefficients. If both response and a predictor are logged, the coefficient is an elasticity: it describes proportional response changes associated with proportional predictor changes within the specified model.
Constant log-scale variance implies a particular original-scale variance:
Thus, original-scale standard deviation grows in proportion to the conditional mean. If log-scale variance instead depends on predictors, say , the mean correction must also depend on them. A single global retransformation factor then fails to represent the model.
Normality is not required merely to fit a regression to log responses, but it is required for the exact lognormal interpretation. For a more general model
the mean is
A residual-based smearing factor estimates the second term without imposing normal errors when the residual distribution is appropriately stable across predictor values. Estimation becomes more complicated when that distribution changes with the predictors.
Lognormal regression must also be distinguished from a generalized linear model using a logarithmic link. A gamma regression with log link directly models ; ordinary regression of models a transformed response. Although both may describe positive, right-skewed data, their likelihoods and interpretation differ.
Zeros cannot occur in a pure lognormal response. Adding an arbitrary constant before taking logarithms changes the statistical model and the meaning of its coefficients. Populations containing genuine zeros and positive continuous outcomes can instead be represented by separate components for occurrence and positive magnitude. Associations estimated by any of these models do not, without further assumptions, establish causal effects.
Sums, approximations, and numerical computation
Unlike products, sums of lognormal variables do not generally remain within the lognormal family. If
its density can be expressed through convolutions in the independent case, but no general elementary formula results. This problem occurs in aggregate losses, total pollutant exposure, accumulated costs, and interference from multiple radio transmitters. The distinction matters because an aggregate may look lognormal in its central range while having noticeably different tail probabilities.
A widely used approximation, associated with Fenton’s 1960 work and commonly called the Fenton–Wilkinson approximation, matches the first two moments of to a lognormal distribution. If
the approximating parameters are
For dependent components,
Ignoring positive dependence can therefore substantially understate aggregate variation. Matching moments provides a convenient approximation, not a theorem that the sum is lognormal, and it does not guarantee accurate extreme quantiles.
Two different limiting regimes clarify sum behavior. Because lognormal variables have finite variance, sums of independent, identically distributed lognormals satisfy the ordinary central limit theorem as the number of terms tends to infinity. After centering and scaling, their distribution approaches a normal distribution. When is large, however, finite-sample convergence can be slow because rare, large observations account for much of the mean and variance.
For a fixed number of independent, identically distributed lognormal variables, the extreme right tail behaves differently:
This is an instance of the subexponential property: an exceptionally large sum is asymptotically dominated by one exceptionally large summand. It does not contradict the central limit theorem, because one statement concerns a fixed sum at increasingly extreme thresholds, whereas the other concerns an increasing number of terms near the centered bulk.
Simulation is straightforward using
A Monte Carlo method can estimate aggregate probabilities or nonlinear expectations, although ordinary sampling becomes inefficient for very rare events. Numerical integration and specialized tail approximations provide alternatives.
Numerical stability requires care when parameters or thresholds are extreme. Direct exponentiation can overflow in floating-point arithmetic, and computing a survival probability as can lose precision when is close to 1. Log-survival calculations avoid this subtraction. Likewise, sums of exponentials can be evaluated on the logarithmic scale using
which prevents unnecessary overflow while preserving the aggregate’s logarithm.
Applications, model assessment, and boundaries of validity
In aerosol science and particle technology, lognormal distributions describe many particle-size populations over specified ranges. Particle diameters often span orders of magnitude, making geometric mean diameter and geometric standard deviation useful summaries. The weighting convention is crucial: a distribution counted by particle number differs from one weighted by surface area or mass. If diameter is lognormal and weighting is proportional to , the weighted log-diameter distribution remains normal with mean shifted from to and unchanged variance. For constant-density spheres, mass weighting corresponds to .
Environmental measurements such as contaminant concentrations can be strongly right-skewed because they combine variation in sources, dilution, transport, and sampling conditions. A lognormal model can represent positive concentrations in a relatively homogeneous setting. Nevertheless, mixed sources, seasonal changes, measurements below detection limits, and spatial dependence can produce distributions that are not adequately represented by one lognormal population. Concentrations measured across sites are also not automatically independent observations.
In reliability and duration analysis, the lognormal distribution models lifetimes whose logarithms are approximately normal. Its hazard function,
begins near zero, rises, and eventually declines toward zero. It can therefore represent a population whose conditional failure rate increases and subsequently decreases. It does not impose the constant hazard of an exponential distribution, nor the globally monotone hazard found in some other lifetime models. This shape is a model property, not a general assertion about physical aging.
In finance, geometric Brownian motion makes future prices lognormal and continuously compounded returns normal over a fixed horizon. This is central to the classical Black–Scholes option-pricing framework introduced in 1973. The price distribution and return distribution must not be confused: a lognormal price does not imply a lognormal log return. Empirical return series commonly show changing volatility and departures from normality, so the classical model is an idealization rather than an exact description of market observations.
Income and size distributions provide another important boundary. A lognormal model can describe a substantial central range while failing in the upper tail, where a Pareto-type power law may fit differently. On a finite observed range, both can appear plausible. A straight-looking portion of a log-log plot is insufficient to establish power-law behavior, and visible right skewness alone is insufficient to establish lognormality.
Assessment typically begins with the logarithms: their histogram, normal quantile plot, dependence structure, and variation across subgroups reveal different aspects of fit. Formal goodness-of-fit tests must account for parameters estimated from the same data; null distributions derived for fully specified models cannot always be reused unchanged. Comparisons with gamma, Weibull, mixture, or other positive-data models depend on the intended target—central prediction, arithmetic means, or extreme exceedances.
Pooling distinct populations can conceal multimodality or generate apparent heavy tails. Conversely, small samples may contain too little tail information to discriminate between models with similar central behavior. A fitted lognormal distribution therefore specifies an entire extrapolating model, including probabilities far beyond the observed range; agreement in the bulk does not by itself validate those extreme-tail probabilities.