Is it obtained simply by truncating a Gompertz variable?

No; the stated construction transforms it.

Does fitting this family prove that an AI score is calibrated?

No; calibration requires comparing predictions with outcomes.

AI Glossary / Statistics and probability / Advanced

Unit Gompertz distribution

Definition

The unit Gompertz distribution is a continuous probability distribution on the open interval from zero to one, constructed by an exponential transformation of a positive Gompertz random variable. It provides a possible model for bounded continuous measurements such as proportions or normalized indices. It is distinct from an ordinary Gompertz lifetime distribution and from a binomial model for success counts. Its suitability must be assessed for the data, and exact endpoint values require additional treatment.

What it means

The word unit describes the interval of possible values, not a single parameter or a probability of one. A transformation maps an unbounded positive variable into a bounded variable while retaining a tractable cumulative distribution function. This creates a family of shapes that can be compared with other bounded-data models. It does not mean every percentage should follow this distribution. In particular, observed proportions based on small denominators are discrete measurements with sampling variability that a simple continuous model may not capture.

How it works

Use positive parameters alpha and beta and define F(x) = exp[-alpha(x^(-beta) minus 1)] for zero less than x less than one. Differentiation gives density alpha times beta times x^(-beta minus 1) times F(x). The quantile at probability p is [1 minus log(p)/alpha]^(-1/beta). One construction starts with a Gompertz variable Y having survival exp[-alpha(exp(beta y) minus 1)] and sets X = exp(-Y). Because this transformation decreases, the CDF of X uses the survival probability of Y. Parameters can be estimated by likelihood, followed by diagnostic checks and comparison with suitable alternatives.

Three practical examples

Example 1: An analyst considers the family for continuous quality indices strictly between zero and one. Example 2: A simulation uses its quantile function to generate bounded scenario values with a chosen shape. Example 3: A research team compares it with beta and Kumaraswamy models for a normalized measurement. These are candidate uses, not evidence that the family outperforms alternatives generally. Dependence, endpoint masses and the measurement process remain separate modeling issues.

Think of it this way

Think of taking positive distances along a line and compressing them into a ruler that runs from zero to one. Large distances move close to zero, while distances near zero move close to one. The compression changes the probability pattern in a precise way. It creates a bounded model, but does not establish that the ruler describes every real bounded measurement.

Worked example

Set alpha = 1 and beta = 1. Then F(x) = exp[-(1/x minus 1)]. At x = 0.5, the probability that X is at most 0.5 is exp(-1), approximately 0.367879. The median solves F(x) = 0.5, giving x = 1/(1 minus log(0.5)), approximately 0.590616. Substituting that value back into the CDF returns 0.5. The density at x = 0.5 is x^(-2) times F(x), equal to 4/e, approximately 1.4715. A density above one is valid because probabilities are areas under the density, not its height at a point. These values illustrate the stated parameterization and are not fitted results from an observed dataset.

Illustrative case study

In this fictional case study, Laurel Quality models a continuous normalized score that never equals zero or one in its initial sample. Analysts compare unit Gompertz and other bounded families using fit diagnostics and predictive assessment. They also check whether future data can contain exact endpoints and whether the score's normalization rule changes over time. When reviewing sources, they notice a later correction to the original distribution paper and use the corrected account rather than copying early claims about density or hazard shapes.

Common misconceptions and mistakes

Common mistakes include confusing the unit Gompertz family with the ordinary Gompertz distribution, treating its parameters as a mean and variance, and using a density height as an event probability. Another mistake is applying a continuous interior-support model to exact zeros or ones without explanation. The original publication contains properties corrected in subsequent work, so broad claims about universal log-concavity or hazard shapes should not be copied uncritically. Formula verification and source checking matter even for a named distribution.

Business use

Businesses may consider this family when a genuinely continuous bounded measurement needs a parametric description. Compare relevant tail probabilities and quantiles, not only a single goodness-of-fit score. Keep the normalization convention and endpoint handling explicit. If scores are generated by an AI system, calibration, subgroup behavior and drift are not established by fitting their marginal distribution. A well-fitting curve can summarize observed scores while still saying little about whether those scores represent reliable probabilities or useful decisions.

Check your understanding

What is F(0.5) when both parameters equal one?

Answer: Approximately 0.367879.

What is the median?

Answer: Approximately 0.590616.

Why does the transformation use the original survival function?

Answer: Exponentiating the negative of Y reverses the ordering.

What should be checked when consulting the original paper?

Answer: Later corrections to some of its stated properties.

Related terms

Cumulative distribution function
Related terms include Gompertz distribution
bounded distribution
beta distribution
quantile function and maximum likelihood
but they imply different probability laws

Learn next

Learn next how decreasing transformations affect cumulative probabilities and how Jacobians produce transformed densities. Then compare bounded families using residuals and predictive checks. Investigate zero-or-one inflation when endpoint observations are meaningful rather than changing them arbitrarily to fit an interior-only model.

Sources

Sources: Mazucheli, Menezes and Dey, Unit-Gompertz Distribution with Applications, https://rivista-statistica.unibo.it/article/view/8497; Anis and De, An Expository Note on Unit-Gompertz Distribution with Applications, https://rivista-statistica.unibo.it/article/view/11135. The latter corrects properties in the original paper. The CDF, median and density calculations are original and independently checked; the quality-score case is fictional.

Previous: Unit clause Next: Unit preference