Gamma and Beta Distributions

Positive quantities, bounded proportions, and a shared family of probability constructions

Statistics

Probability

Distributions

Gamma and beta distributions, their parameterizations, moments, transformations, conjugate roles, and biological interpretation.

Author

Jiahong Sun

Published

September 13, 2026

Modified

September 13, 2026

1 Topic at a Glance

One-sentence summary: The gamma distribution models positive magnitudes and accumulated waiting times, while the beta distribution models uncertain probabilities or continuous proportions between zero and one.

Distribution	Support	Parameters used here	Mean	Variance
Gamma	$x>0$	shape $a>0$, rate $b>0$	$a/b$	$a/b^2$
Beta	$0<x<1$	shapes $a>0$, $b>0$	$a/(a+b)$	$ab/\{(a+b)^2(a+b+1)\}$
Prerequisites	Exponential and geometric distributions; transformations of random variables

The symbols $a$ and $b$ play different roles in the two families. For gamma distributions, software and texts also alternate between a rate $b$ and a scale $\theta=1/b$; stating the convention is essential.

2 The Gamma Distribution

A positive random variable $X$ has a gamma distribution with shape $a>0$ and rate $b>0$ when

\[ \boxed{ f_X(x)=\frac{b^a}{\Gamma(a)}x^{a-1}e^{-bx}, \qquad x>0, } \]

where the gamma function is

\[ \Gamma(a)=\int_0^\infty t^{a-1}e^{-t}\,dt. \]

The substitution $t=bx$ shows that the density integrates to one. Integration by parts gives $\Gamma(a+1)=a\Gamma(a)$, so $\Gamma(n)=(n-1)!$ for positive integers $n$.

2.1 Moments and shape

For any real $r>-a$,

\[ E(X^r)=\frac{\Gamma(a+r)}{\Gamma(a)b^r}. \]

Taking $r=1$ and $r=2$ yields

\[ \boxed{E(X)=\frac ab,\qquad \operatorname{Var}(X)=\frac{a}{b^2}.} \]

The moment-generating function is

\[ M_X(t)=\left(\frac{b}{b-t}\right)^a, \qquad t<b. \]

When $a<1$, the density decreases from an integrable singularity at zero. When $a=1$, it is exponential. When $a>1$, its mode is $(a-1)/b$. Increasing $a$ while holding the mean fixed reduces relative variation and makes the distribution less skewed.

2.2 Waiting times and additivity

If $T_1,\ldots,T_k$ are independent exponential waiting times with common rate $b$, then

\[ T_1+\cdots+T_k\sim\operatorname{Gamma}(k,b). \]

Thus the time of the $k$th event in a homogeneous Poisson process is gamma distributed. More generally, the moment-generating function proves

\[ X_i\sim\operatorname{Gamma}(a_i,b)\text{ independently} \quad\Longrightarrow\quad \sum_iX_i\sim\operatorname{Gamma}\left(\sum_i a_i,b\right). \]

The common rate is required. Adding gamma variables with different rates does not generally produce another gamma variable.

2.3 Parameterization warning

The shape–scale density is

\[ f_X(x)=\frac{x^{a-1}e^{-x/\theta}}{\Gamma(a)\theta^a}, \qquad \theta>0, \]

with mean $a\theta$ and variance $a\theta^2$. It describes exactly the same family when $\theta=1/b$. Confusing rate and scale replaces a parameter by its reciprocal and can radically change simulated values or likelihood calculations.

3 The Beta Distribution

A random variable $P$ has a beta distribution with shapes $a>0$ and $b>0$ when

\[ \boxed{ f_P(p)=\frac{1}{B(a,b)}p^{a-1}(1-p)^{b-1}, \qquad 0<p<1, } \]

where

\[ B(a,b)=\int_0^1p^{a-1}(1-p)^{b-1}\,dp =\frac{\Gamma(a)\Gamma(b)}{\Gamma(a+b)}. \]

Its principal moments are

\[ \boxed{ E(P)=\frac{a}{a+b}, \qquad \operatorname{Var}(P)= \frac{ab}{(a+b)^2(a+b+1)}. } \]

For $a,b>1$, the mode is $(a-1)/(a+b-2)$. The uniform distribution on $(0,1)$ is $\operatorname{Beta}(1,1)$. Shapes below one put more density near boundaries; $a,b>1$ produce an interior mode; and unequal shapes produce skewness.

3.1 Mean–concentration parameterization

Writing

\[ \mu=\frac{a}{a+b}, \qquad \kappa=a+b \]

gives $a=\mu\kappa$ and $b=(1-\mu)\kappa$. Then

\[ E(P)=\mu, \qquad \operatorname{Var}(P)=\frac{\mu(1-\mu)}{\kappa+1}. \]

This separates location from concentration. At a fixed mean, larger $\kappa$ concentrates probability more tightly around $\mu$. Calling $\kappa$ a sample size is only an analogy unless the model was constructed from actual commensurate observations.

4 The Beta–Gamma Construction

Let

\[ U\sim\operatorname{Gamma}(a,1), \qquad V\sim\operatorname{Gamma}(b,1), \qquad U\perp\!\!\!\perp V. \]

Define

\[ S=U+V, \qquad P=\frac{U}{U+V}. \]

The inverse transformation is $U=PS$, $V=(1-P)S$, with absolute Jacobian determinant $s$. Therefore

\[ \begin{aligned} f_{P,S}(p,s) &=\frac{p^{a-1}(1-p)^{b-1}}{B(a,b)} \frac{s^{a+b-1}e^{-s}}{\Gamma(a+b)}. \end{aligned} \]

The factorization shows that

\[ \boxed{P\sim\operatorname{Beta}(a,b),\qquad S\sim\operatorname{Gamma}(a+b,1),\qquad P\perp\!\!\!\perp S.} \]

The beta variable represents the fraction of two independent gamma contributions attributable to the first; the gamma variable represents their total. This construction generalizes to the Dirichlet distribution.

5 Conjugate Roles

5.1 Beta prior with a binomial likelihood

\[ Y\mid P=p\sim\operatorname{Binomial}(n,p), \qquad P\sim\operatorname{Beta}(a,b), \]

then multiplication of likelihood and prior gives

\[ P\mid Y=y\sim\operatorname{Beta}(a+y,b+n-y). \]

The posterior updates the two shape parameters with successes and failures. This is the conjugate calculation introduced in Bayes’ theorem and Bayesian updating. It does not make beta distributions appropriate for every proportion: structural zeros or ones, dependence among trials, misclassification, and variation across experimental units require a richer observation model.

5.2 Gamma prior with a Poisson likelihood

If conditionally independent counts satisfy

\[ Y_i\mid\Lambda=\lambda\sim\operatorname{Poisson}(t_i\lambda), \qquad \Lambda\sim\operatorname{Gamma}(a,b), \]

where $t_i$ are known exposures, then

\[ \boxed{ \Lambda\mid y \sim\operatorname{Gamma}\left(a+\sum_i y_i, b+\sum_i t_i\right). } \]

The rate parameter accumulates exposure, while the shape accumulates events. Mixing a Poisson count over a gamma-distributed rate also yields a negative-binomial marginal distribution, explaining one important source of count overdispersion.

6 Worked Biological Example

Suppose a sequencing calibration experiment observes $y=18$ correctly classified control molecules out of $n=20$ conditionally independent trials. With a $\operatorname{Beta}(2,2)$ prior for the assay’s classification probability $p$,

\[ p\mid y\sim\operatorname{Beta}(20,4), \]

whose posterior mean is

\[ E(p\mid y)=\frac{20}{24}\approx0.833. \]

The raw observed fraction is $0.9$; the posterior mean is pulled toward the prior mean $0.5$. Statistically, the posterior describes $p$ under the binomial sampling model and chosen prior. Biologically, it supports generalization only to specimens and conditions represented by the calibration design.

Molecules or reads from one library are technical units. Estimating performance across laboratories, organisms, or batches requires independently replicated biological or operational units and a hierarchical model if their probabilities differ.

7 Assumptions and Model Choice

Intended use	Crucial assumption	Common failure	Sensitivity or alternative
Gamma waiting-time model	positive duration generated by accumulated constant-rate waits	time-varying hazard, censoring, or competing events	survival model with appropriate hazard and observation process
Gamma magnitude model	continuous positive support with gamma-shaped mean–variance relation	zeros, multimodality, or heavy tails	two-part, mixture, log-normal, or robust model
Beta proportion model	a continuous latent quantity strictly inside $(0,1)$	observed boundary values or discrete numerator/denominator data	zero/one-inflated beta or explicit binomial model
Conjugate updating	stated conditional independence and common parameter	clustered trials or heterogeneous rates	hierarchical beta–binomial or gamma–Poisson model

8 Common Pitfalls

Failing to state rate versus scale: the same numeric value implies reciprocal means under the two gamma conventions.
Using a beta density for a count proportion without its denominator: $3/5$ and $300/500$ have different measurement uncertainty; model the counts when available.
Treating density at zero or one as point probability: an ordinary beta law assigns zero probability to exact boundary values.
Adding gamma variables with unequal rates as though closure still held: common rate is part of the additivity result.
Reading conjugacy as validation: convenient algebra does not check sampling, dependence, or prior assumptions.
Confusing technical observations with biological replication: deeper measurement within one specimen does not estimate between-specimen variation.

9 Connections to the Knowledge Map

Exponential and geometric distributions gives the one-event waiting-time case of the gamma family.
Poisson distribution and Poisson processes supplies counts and arrival times linked to gamma variables.
Bernoulli and binomial distributions provides the likelihood paired with a beta prior.
Bayes’ theorem and Bayesian updating explains posterior normalization and prediction.
Chi-squared, Student’s t, and F distributions shows that $\chi^2_\nu$ is $\operatorname{Gamma}(\nu/2,1/2)$ under the rate convention.
Dirichlet distribution extends normalized independent gamma variables to multiple proportions.

10 Take-home Messages

The gamma distribution models positive quantities; its rate and scale parameterizations must never be left implicit.
The beta distribution models a latent continuous probability or proportion on $(0,1)$, with shapes controlling its mean and concentration.
Normalizing two independent gamma variables produces a beta proportion independent of their gamma-distributed total.
Beta–binomial and gamma–Poisson conjugacy give transparent updates, but their sampling assumptions still require scrutiny.
Boundary values, heterogeneous units, dependence, and the biological replication structure often require extensions.

11 References

Johnson, N. L., Kotz, S., and Balakrishnan, N. Continuous Univariate Distributions, Volume 2, 2nd ed. Wiley, 1995.
Casella, G. and Berger, R. L. Statistical Inference, 2nd ed. Duxbury, 2002.
Gelman, A. et al. Bayesian Data Analysis, 3rd ed. CRC Press, 2013.
R Core Team. “The Gamma Distribution.” R Statistical Functions Documentation. https://stat.ethz.ch/R-manual/R-devel/library/stats/html/GammaDist.html.
R Core Team. “The Beta Distribution.” R Statistical Functions Documentation. https://stat.ethz.ch/R-manual/R-devel/library/stats/html/Beta.html.

--- title: "Gamma and Beta Distributions" subtitle: "Positive quantities, bounded proportions, and a shared family of probability constructions" description: "Gamma and beta distributions, their parameterizations, moments, transformations, conjugate roles, and biological interpretation." author: "Jiahong Sun" date: "2026-09-13" date-modified: "2026-09-13" categories: [Statistics, Probability, Distributions] toc: true number-sections: true draft: false --- # Topic at a Glance > **One-sentence summary:** The gamma distribution models positive magnitudes and accumulated waiting times, while the beta distribution models uncertain probabilities or continuous proportions between zero and one. | Distribution | Support | Parameters used here | Mean | Variance | |---|---|---|---|---| | Gamma | $x>0$ | shape $a>0$, rate $b>0$ | $a/b$ | $a/b^2$ | | Beta | $0<x<1$ | shapes $a>0$, $b>0$ | $a/(a+b)$ | $ab/\{(a+b)^2(a+b+1)\}$ | | Prerequisites | [Exponential and geometric distributions](exponential-and-geometric-distributions.qmd); [transformations of random variables](transformations-of-random-variables.qmd) | | | | The symbols $a$ and $b$ play different roles in the two families. For gamma distributions, software and texts also alternate between a **rate** $b$ and a **scale** $\theta=1/b$; stating the convention is essential. # The Gamma Distribution A positive random variable $X$ has a gamma distribution with shape $a>0$ and rate $b>0$ when $$ \boxed{ f_X(x)=\frac{b^a}{\Gamma(a)}x^{a-1}e^{-bx}, \qquad x>0, } $$ where the gamma function is $$ \Gamma(a)=\int_0^\infty t^{a-1}e^{-t}\,dt. $$ The substitution $t=bx$ shows that the density integrates to one. Integration by parts gives $\Gamma(a+1)=a\Gamma(a)$, so $\Gamma(n)=(n-1)!$ for positive integers $n$. ## Moments and shape For any real $r>-a$, $$ E(X^r)=\frac{\Gamma(a+r)}{\Gamma(a)b^r}. $$ Taking $r=1$ and $r=2$ yields $$ \boxed{E(X)=\frac ab,\qquad \operatorname{Var}(X)=\frac{a}{b^2}.} $$ The moment-generating function is $$ M_X(t)=\left(\frac{b}{b-t}\right)^a, \qquad t<b. $$ When $a<1$, the density decreases from an integrable singularity at zero. When $a=1$, it is exponential. When $a>1$, its mode is $(a-1)/b$. Increasing $a$ while holding the mean fixed reduces relative variation and makes the distribution less skewed. ## Waiting times and additivity If $T_1,\ldots,T_k$ are independent exponential waiting times with common rate $b$, then $$ T_1+\cdots+T_k\sim\operatorname{Gamma}(k,b). $$ Thus the time of the $k$th event in a homogeneous Poisson process is gamma distributed. More generally, the moment-generating function proves $$ X_i\sim\operatorname{Gamma}(a_i,b)\text{ independently} \quad\Longrightarrow\quad \sum_iX_i\sim\operatorname{Gamma}\left(\sum_i a_i,b\right). $$ The common rate is required. Adding gamma variables with different rates does not generally produce another gamma variable. ## Parameterization warning The shape--scale density is $$ f_X(x)=\frac{x^{a-1}e^{-x/\theta}}{\Gamma(a)\theta^a}, \qquad \theta>0, $$ with mean $a\theta$ and variance $a\theta^2$. It describes exactly the same family when $\theta=1/b$. Confusing rate and scale replaces a parameter by its reciprocal and can radically change simulated values or likelihood calculations. # The Beta Distribution A random variable $P$ has a beta distribution with shapes $a>0$ and $b>0$ when $$ \boxed{ f_P(p)=\frac{1}{B(a,b)}p^{a-1}(1-p)^{b-1}, \qquad 0<p<1, } $$ where $$ B(a,b)=\int_0^1p^{a-1}(1-p)^{b-1}\,dp =\frac{\Gamma(a)\Gamma(b)}{\Gamma(a+b)}. $$ Its principal moments are $$ \boxed{ E(P)=\frac{a}{a+b}, \qquad \operatorname{Var}(P)= \frac{ab}{(a+b)^2(a+b+1)}. } $$ For $a,b>1$, the mode is $(a-1)/(a+b-2)$. The uniform distribution on $(0,1)$ is $\operatorname{Beta}(1,1)$. Shapes below one put more density near boundaries; $a,b>1$ produce an interior mode; and unequal shapes produce skewness. ## Mean--concentration parameterization Writing $$ \mu=\frac{a}{a+b}, \qquad \kappa=a+b $$ gives $a=\mu\kappa$ and $b=(1-\mu)\kappa$. Then $$ E(P)=\mu, \qquad \operatorname{Var}(P)=\frac{\mu(1-\mu)}{\kappa+1}. $$ This separates location from concentration. At a fixed mean, larger $\kappa$ concentrates probability more tightly around $\mu$. Calling $\kappa$ a sample size is only an analogy unless the model was constructed from actual commensurate observations. # The Beta--Gamma Construction Let $$ U\sim\operatorname{Gamma}(a,1), \qquad V\sim\operatorname{Gamma}(b,1), \qquad U\perp\!\!\!\perp V. $$ Define $$ S=U+V, \qquad P=\frac{U}{U+V}. $$ The inverse transformation is $U=PS$, $V=(1-P)S$, with absolute Jacobian determinant $s$. Therefore $$ \begin{aligned} f_{P,S}(p,s) &=\frac{p^{a-1}(1-p)^{b-1}}{B(a,b)} \frac{s^{a+b-1}e^{-s}}{\Gamma(a+b)}. \end{aligned} $$ The factorization shows that $$ \boxed{P\sim\operatorname{Beta}(a,b),\qquad S\sim\operatorname{Gamma}(a+b,1),\qquad P\perp\!\!\!\perp S.} $$ The beta variable represents the fraction of two independent gamma contributions attributable to the first; the gamma variable represents their total. This construction generalizes to the [Dirichlet distribution](dirichlet-distribution.qmd). # Conjugate Roles ## Beta prior with a binomial likelihood If $$ Y\mid P=p\sim\operatorname{Binomial}(n,p), \qquad P\sim\operatorname{Beta}(a,b), $$ then multiplication of likelihood and prior gives $$ P\mid Y=y\sim\operatorname{Beta}(a+y,b+n-y). $$ The posterior updates the two shape parameters with successes and failures. This is the conjugate calculation introduced in [Bayes' theorem and Bayesian updating](bayes-theorem-and-bayesian-updating.qmd). It does not make beta distributions appropriate for every proportion: structural zeros or ones, dependence among trials, misclassification, and variation across experimental units require a richer observation model. ## Gamma prior with a Poisson likelihood If conditionally independent counts satisfy $$ Y_i\mid\Lambda=\lambda\sim\operatorname{Poisson}(t_i\lambda), \qquad \Lambda\sim\operatorname{Gamma}(a,b), $$ where $t_i$ are known exposures, then $$ \boxed{ \Lambda\mid y \sim\operatorname{Gamma}\left(a+\sum_i y_i, b+\sum_i t_i\right). } $$ The rate parameter accumulates exposure, while the shape accumulates events. Mixing a Poisson count over a gamma-distributed rate also yields a negative-binomial marginal distribution, explaining one important source of count overdispersion. # Worked Biological Example Suppose a sequencing calibration experiment observes $y=18$ correctly classified control molecules out of $n=20$ conditionally independent trials. With a $\operatorname{Beta}(2,2)$ prior for the assay's classification probability $p$, $$ p\mid y\sim\operatorname{Beta}(20,4), $$ whose posterior mean is $$ E(p\mid y)=\frac{20}{24}\approx0.833. $$ The raw observed fraction is $0.9$; the posterior mean is pulled toward the prior mean $0.5$. Statistically, the posterior describes $p$ under the binomial sampling model and chosen prior. Biologically, it supports generalization only to specimens and conditions represented by the calibration design. Molecules or reads from one library are technical units. Estimating performance across laboratories, organisms, or batches requires independently replicated biological or operational units and a hierarchical model if their probabilities differ. # Assumptions and Model Choice | Intended use | Crucial assumption | Common failure | Sensitivity or alternative | |---|---|---|---| | Gamma waiting-time model | positive duration generated by accumulated constant-rate waits | time-varying hazard, censoring, or competing events | survival model with appropriate hazard and observation process | | Gamma magnitude model | continuous positive support with gamma-shaped mean--variance relation | zeros, multimodality, or heavy tails | two-part, mixture, log-normal, or robust model | | Beta proportion model | a continuous latent quantity strictly inside $(0,1)$ | observed boundary values or discrete numerator/denominator data | zero/one-inflated beta or explicit binomial model | | Conjugate updating | stated conditional independence and common parameter | clustered trials or heterogeneous rates | hierarchical beta--binomial or gamma--Poisson model | # Common Pitfalls 1. **Failing to state rate versus scale:** the same numeric value implies reciprocal means under the two gamma conventions. 2. **Using a beta density for a count proportion without its denominator:** $3/5$ and $300/500$ have different measurement uncertainty; model the counts when available. 3. **Treating density at zero or one as point probability:** an ordinary beta law assigns zero probability to exact boundary values. 4. **Adding gamma variables with unequal rates as though closure still held:** common rate is part of the additivity result. 5. **Reading conjugacy as validation:** convenient algebra does not check sampling, dependence, or prior assumptions. 6. **Confusing technical observations with biological replication:** deeper measurement within one specimen does not estimate between-specimen variation. # Connections to the Knowledge Map - [Exponential and geometric distributions](exponential-and-geometric-distributions.qmd) gives the one-event waiting-time case of the gamma family. - [Poisson distribution and Poisson processes](poisson-distribution-and-poisson-processes.qmd) supplies counts and arrival times linked to gamma variables. - [Bernoulli and binomial distributions](bernoulli-and-binomial-distributions.qmd) provides the likelihood paired with a beta prior. - [Bayes' theorem and Bayesian updating](bayes-theorem-and-bayesian-updating.qmd) explains posterior normalization and prediction. - [Chi-squared, Student's t, and F distributions](chi-squared-students-t-and-f-distributions.qmd) shows that $\chi^2_\nu$ is $\operatorname{Gamma}(\nu/2,1/2)$ under the rate convention. - [Dirichlet distribution](dirichlet-distribution.qmd) extends normalized independent gamma variables to multiple proportions. # Take-home Messages 1. The gamma distribution models positive quantities; its rate and scale parameterizations must never be left implicit. 2. The beta distribution models a latent continuous probability or proportion on $(0,1)$, with shapes controlling its mean and concentration. 3. Normalizing two independent gamma variables produces a beta proportion independent of their gamma-distributed total. 4. Beta--binomial and gamma--Poisson conjugacy give transparent updates, but their sampling assumptions still require scrutiny. 5. Boundary values, heterogeneous units, dependence, and the biological replication structure often require extensions. # References - Johnson, N. L., Kotz, S., and Balakrishnan, N. *Continuous Univariate Distributions*, Volume 2, 2nd ed. Wiley, 1995. - Casella, G. and Berger, R. L. *Statistical Inference*, 2nd ed. Duxbury, 2002. - Gelman, A. et al. *Bayesian Data Analysis*, 3rd ed. CRC Press, 2013. - R Core Team. “The Gamma Distribution.” *R Statistical Functions Documentation*. [https://stat.ethz.ch/R-manual/R-devel/library/stats/html/GammaDist.html](https://stat.ethz.ch/R-manual/R-devel/library/stats/html/GammaDist.html). - R Core Team. “The Beta Distribution.” *R Statistical Functions Documentation*. [https://stat.ethz.ch/R-manual/R-devel/library/stats/html/Beta.html](https://stat.ethz.ch/R-manual/R-devel/library/stats/html/Beta.html).