1 Definition and basic properties
1.1 Probability density function (PDF) and parameters
The beta distribution with parameters \(\alpha>0\) and \(\beta>0\), written \(X\sim \mathrm{Beta}(\alpha,\beta)\), is a continuous distribution on the interval \([0,1]\) with probability density function \[ f(x)=\frac{1}{B(\alpha,\beta)}x^{\alpha-1}(1-x)^{\beta-1},\quad 0\le x\le 1, \] and \(f(x)=0\) outside \([0,1]\). The function \(B(\alpha,\beta)\) is the beta function, which serves as the normalizing constant.
1.2 Support, shape, and parameter interpretation (α, β)
The support of the beta distribution is exactly \([0,1]\). Shape is controlled by the pair \((\alpha,\beta)\): \(\alpha\) governs how the density behaves near \(0\), while \(\beta\) governs behavior near \(1\). When \(\alpha\) is small relative to \(\beta\), mass concentrates closer to \(0\); when \(\beta\) is small relative to \(\alpha\), mass concentrates closer to \(1\). When both parameters are large, the distribution becomes more concentrated around its central tendency.
1.3 Cumulative distribution function (CDF) and quantiles
The cumulative distribution function is \[ F(x)=\Pr(X\le x)=I_x(\alpha,\beta), \] where \(I_x(\alpha,\beta)\) denotes the regularized incomplete beta function. Quantiles, such as the \(p\)-th quantile \(q_p\) satisfying \(\Pr(X\le q_p)=p\), are computed as the inverse of this regularized incomplete beta function. In practice, numerical libraries provide stable routines for both \(F(x)\) and \(q_p\).
1.4 Mean, variance, and higher moments
For \(X\sim \mathrm{Beta}(\alpha,\beta)\), \[ \mathbb{E}[X]=\frac{\alpha}{\alpha+\beta}, \qquad \mathrm{Var}(X)=\frac{\alpha\beta}{(\alpha+\beta)^2(\alpha+\beta+1)}. \] Higher moments exist for all positive \(\alpha,\beta\). For integer \(k\ge 1\), \[ \mathbb{E}[X^k]=\frac{\alpha(\alpha+1)\cdots(\alpha+k-1)}{(\alpha+\beta)(\alpha+\beta+1)\cdots(\alpha+\beta+k-1)}. \] These formulas are useful for moment matching and for studying how tail behavior changes with parameters.
1.5 Mode, median, and limiting behavior
If \(\alpha>1\) and \(\beta>1\), the mode is \[ \mathrm{mode}=\frac{\alpha-1}{\alpha+\beta-2}. \] The median generally has no simple closed form; it is often approximated numerically or using analytic approximations. As parameters grow large with fixed ratio \(\alpha/(\alpha+\beta)\), the distribution concentrates around its mean. In contrast, when \(\alpha\) or \(\beta\) approaches values near \(0\) (while staying positive), the density can become extremely peaked at the corresponding endpoint.
2 Alternative characterizations
2.1 Relationship to gamma distributions
A common characterization uses gamma variables. Let \(Y_1\sim \mathrm{Gamma}(\alpha,1)\) and \(Y_2\sim \mathrm{Gamma}(\beta,1)\) be independent, with shape parameters \(\alpha\) and \(\beta\) and a shared scale (rate 1). Then \[ X=\frac{Y_1}{Y_1+Y_2} \] has the \(\mathrm{Beta}(\alpha,\beta)\) distribution. This representation explains both the flexibility of the beta family and its connection to other gamma-based constructions.
2.2 Transformation from Dirichlet distribution (marginals)
The beta distribution is the two-dimensional marginal of the Dirichlet distribution. If \((P_1,\dots,P_k)\sim \mathrm{Dirichlet}(\alpha_1,\dots,\alpha_k)\), then \(P_1\) alone has \[ P_1\sim \mathrm{Beta}(\alpha_1,\alpha_0-\alpha_1), \quad \text{where } \alpha_0=\sum_{i=1}^k \alpha_i. \] Thus, beta behavior can be viewed as a special case of multivariate probability mass allocation models.
2.3 Order statistics connections
In order statistics, the beta family appears naturally through the distribution of gaps and ranks for samples from uniform distributions. For example, if \(U_{(r)}\) denotes the \(r\)-th order statistic from \(n\) independent \(\mathrm{Uniform}(0,1)\) draws, then \[ U_{(r)}\sim \mathrm{Beta}(r, n-r+1). \] This connection is a standard route to computing distributional properties of quantiles and ranks.
2.4 Special/degenerate cases (e.g., α=1 or β=1)
When \(\alpha=1\), the density simplifies to \[ f(x)=\frac{1}{B(1,\beta)}(1-x)^{\beta-1}, \] which places relatively more weight near \(0\) and decreases toward \(1\). Similarly, when \(\beta=1\), the density behaves like a power of \(x\) and emphasizes values near \(1\). While \(\alpha,\beta\) are typically assumed strictly positive, parameter limits as \(\alpha\to 0^+\) or \(\beta\to 0^+\) lead to increasingly extreme mass near an endpoint (in a distributional limiting sense).
3 Shape behavior and intuition
3.1 Uniform case
The uniform distribution on \([0,1]\) corresponds to \(\mathrm{Beta}(1,1)\). Its constant density reflects maximal spread with no preference for any subinterval.
3.2 U-shaped, J-shaped, and bell-shaped cases
Different parameter regimes yield distinct shapes:
- Bell-shaped (unimodal): If \(\alpha>1\) and \(\beta>1\), the density has an interior peak.
- U-shaped: If \(0<\alpha<1\) and \(0<\beta<1\), the density blows up near both endpoints, creating high mass close to \(0\) and \(1\).
- J-shaped (endpoint-peaked): If \(\alpha<1\) but \(\beta>1\), mass concentrates near \(0\); if \(\alpha>1\) but \(\beta<1\), mass concentrates near \(1\).
- Monotone decreasing/increasing: Cases like \(\alpha=1\) with \(\beta>1\) produce monotone behavior toward one side.
3.3 Skewness and how α vs. β changes the distribution
Skewness depends on the relative size of \(\alpha\) and \(\beta\). Since \(\mathbb{E}[X]=\alpha/(\alpha+\beta)\), larger \(\alpha\) shifts the center toward higher values. For \(\alpha>\beta\), the distribution tends to be right-skewed (heavier tail toward \(0\)); for \(\beta>\alpha\), it tends to be left-skewed (heavier tail toward \(1\)). The mode (when defined) also shifts accordingly, tracking the endpoint preferences induced by the exponents \(\alpha-1\) and \(\beta-1\).
3.4 Tail behavior near 0 and 1
Near \(x=0\), the density behaves like \(x^{\alpha-1}\). If \(\alpha<1\), this factor diverges, producing strong mass at \(0\); if \(\alpha>1\), it goes to \(0\), suppressing values near \(0\). Analogously, near \(x=1\), the density behaves like \((1-x)^{\beta-1}\), with the same qualitative conclusions controlled by whether \(\beta\) is less than or greater than \(1\).
4 Bayesian modeling applications
4.1 Conjugacy with Bernoulli trials
4.1.1 Prior and posterior update rules
Consider a Bernoulli model with success probability \(\theta\). Place a beta prior \(\theta\sim \mathrm{Beta}(\alpha,\beta)\). After observing one Bernoulli trial, the posterior remains in the beta family:
- observing a success updates \(\alpha\mapsto \alpha+1\),
- observing a failure updates \(\beta\mapsto \beta+1\).
After \(s\) successes and \(f\) failures, the posterior is \[ \theta\mid \text{data}\sim \mathrm{Beta}(\alpha+s,\;\beta+f). \]
4.1.2 Posterior predictive distribution
The beta prior allows an analytic posterior predictive probability for the next Bernoulli outcome. The probability that the next trial is a success equals the posterior mean: \[ \Pr(\text{next success}\mid \text{data})=\frac{\alpha+s}{\alpha+\beta+s+f}. \] This result follows directly from the beta-binomial structure underlying future observations.
4.2 Conjugacy with binomial likelihood
4.2.1 Updating parameters from counts
For binomial observations \(Y\sim \mathrm{Binomial}(n,\theta)\), with prior \(\theta\sim \mathrm{Beta}(\alpha,\beta)\), conjugacy gives the posterior \[ \theta\mid Y=y \sim \mathrm{Beta}(\alpha+y,\;\beta+n-y). \] Thus, only cumulative counts are needed to update the belief about \(\theta\).
4.2.2 Predictive probabilities for future trials
Integrating out \(\theta\) yields the beta-binomial predictive distribution. For a future binomial count \(Y_{\text{new}}\) from \(m\) trials, the predictive distribution is \[ \Pr(Y_{\text{new}}=k\mid \alpha,\beta,s,f) \] with parameters effectively carried by the posterior beta distribution. This provides closed-form predictive probabilities without sampling \(\theta\) explicitly.
4.3 Credible intervals for probabilities
4.3.1 Computing intervals using beta quantiles
A Bayesian credible interval for \(\theta\) is commonly defined via beta quantiles. For a \(100(1-\gamma)\%\) central credible interval, \[ \left[q_{\gamma/2},\; q_{1-\gamma/2}\right], \] where \(q_p\) is the \(p\)-th quantile of \(\mathrm{Beta}(\alpha+s,\beta+f)\). Because the CDF is expressed through the regularized incomplete beta function, accurate quantile computation relies on numerical inversion.
5 Estimation and inference
5.1 Method of moments
Given sample data \(x_1,\dots,x_n\) with sample mean \(\bar{x}\) and sample variance \(s^2\), one can match these to the beta distribution moments: \[ \bar{x}=\frac{\alpha}{\alpha+\beta}, \qquad s^2=\frac{\alpha\beta}{(\alpha+\beta)^2(\alpha+\beta+1)}. \] Solving yields parameter estimates (when the sample moments are compatible with a beta model). This approach is straightforward but can be sensitive when data are near \(0\) or \(1\).
5.2 Maximum likelihood estimation (MLE) and practical considerations
MLE chooses \(\alpha,\beta\) to maximize the beta log-likelihood based on observed \(x_i\). The resulting score equations involve digamma functions and generally require iterative numerical solvers. Practical issues include:
- boundary observations: if data include exact \(0\) or \(1\), the likelihood can become problematic when \(\alpha\) or \(\beta\) are less than 1;
- initialization: good starting values improve convergence;
- identifiability: extreme shapes can lead to flat likelihood regions, making optimization sensitive.
5.3 Bayesian estimation with hyperpriors (overview)
One can place hyperpriors on \(\alpha\) and \(\beta\) (or on reparameterizations such as mean and concentration) and infer them from data. This yields a hierarchical model that captures uncertainty in prior strength rather than treating \(\alpha,\beta\) as fixed. Posterior computation typically uses Markov chain Monte Carlo or variational methods, since closed-form updates are not generally available.
5.4 Model checking and goodness-of-fit concepts
After fitting a beta model, model checking may use:
- residual-style diagnostics adapted to the beta setting (e.g., comparing empirical distribution to the fitted CDF);
- calibration checks for probabilistic forecasts;
- graphical methods such as overlaying fitted density curves on histograms or using probability plots.
Goodness-of-fit tests for continuous distributions can be applied to the beta CDF, provided appropriate care is taken with parameter estimation and boundary effects.
6 Computation and numerical aspects
6.1 Evaluating the Beta function and normalization
The normalization constant is \[ B(\alpha,\beta)=\frac{\Gamma(\alpha)\Gamma(\beta)}{\Gamma(\alpha+\beta)}. \] In numerical work, computing via log-gamma functions improves stability, especially for large parameters where direct evaluation of \(\Gamma(\cdot)\) can overflow.
6.2 Computing PDF/CDF values reliably
The PDF can underflow or overflow for extreme \(\alpha,\beta\) and values close to the boundaries. Robust implementations typically compute in log space and exponentiate when safe. The CDF requires evaluating the regularized incomplete beta function; reliable algorithms use continued fractions or power series expansions depending on parameter and argument regions.
6.3 Random variate generation
Sampling from \(\mathrm{Beta}(\alpha,\beta)\) is commonly done using the gamma representation: draw \(Y_1\sim\mathrm{Gamma}(\alpha,1)\) and \(Y_2\sim\mathrm{Gamma}(\beta,1)\), then output \(X=Y_1/(Y_1+Y_2)\). This method leverages well-tested gamma samplers and handles a wide range of shapes.
6.4 Parameter constraints and edge-case handling
Most computations assume \(\alpha>0\) and \(\beta>0\). Implementations must guard against invalid parameters, and special handling is often needed when parameters are close to \(1\) (where certain derivatives or approximations change behavior) or when values are extremely close to \(0\) or \(1\).
7 Connections and generalizations
7.1 Beta-binomial distribution
The beta-binomial distribution arises when a binomial likelihood has its success probability integrated over a beta prior. It produces overdispersion relative to a pure binomial model and is useful for modeling counts when success probability varies between groups or over time. Its parameters inherit the beta prior’s \(\alpha,\beta\) and the binomial trial count.
7.2 Beta prime distribution (comparison and transformation)
The beta prime distribution is supported on \((0,\infty)\) and can be obtained by a transformation involving the ratio of gamma variables similar to the beta construction. While the beta distribution describes proportions constrained to \([0,1]\), beta prime is often used for positive quantities without an upper bound. The two families are closely related by transformations such as \(X \mapsto X/(1+X)\) in appropriate settings.
7.3 Dirichlet distribution and multivariate extension
The Dirichlet distribution generalizes the beta distribution to the simplex in \(\mathbb{R}^k\), modeling vectors of proportions that sum to \(1\). Each component marginally follows a beta distribution, and dependence between components is governed by the shared concentration parameters. This makes Dirichlet models natural for categorical data and compositional analyses.
7.4 Mixtures involving beta components (overview)
Mixtures of beta distributions combine multiple beta-shaped components to approximate complex densities on \([0,1]\). Such mixtures can model multimodality or irregular shapes that a single beta cannot capture. Bayesian mixture models may place priors on mixture weights and component parameters, often using latent assignment variables for inference.
8 Worked examples and common tasks
8.1 Fitting a beta distribution to proportion data
Suppose observations \(x_i\) represent measured proportions in \([0,1]\). One can estimate \(\alpha,\beta\) via moment matching to obtain an initial fit, then refine using MLE if desired. After fitting, it is common to compare the empirical mean and variance to the theoretical beta mean and variance to verify that the chosen parameters reproduce key summary features.
8.2 Updating beliefs after observing successes and failures
Start with a prior \(\theta\sim \mathrm{Beta}(\alpha,\beta)\). After observing \(s\) successes and \(f\) failures in Bernoulli trials, update to \(\mathrm{Beta}(\alpha+s,\beta+f)\). This update rule is simple enough for sequential analysis: each new success increments \(\alpha\), each new failure increments \(\beta\). The resulting posterior mean provides an updated point estimate, while quantiles yield uncertainty bounds.
8.3 Estimating probabilities of thresholds (e.g., P(X>t))
To compute \(\Pr(X>t)\) for \(X\sim \mathrm{Beta}(\alpha,\beta)\), use the CDF: \[ \Pr(X>t)=1-F(t)=1-I_t(\alpha,\beta). \] This is frequently used to answer “exceedance” questions, such as the chance that a success probability is above a target threshold.
8.4 Visualizing parameter effects on shape
A standard exploratory task is to fix one parameter and vary the other to see how the density changes. For instance, increasing \(\alpha\) while holding \(\beta\) constant moves the mass toward larger \(x\) and tightens concentration if both parameters grow. Plotting multiple beta densities on the same axes helps develop intuition about how endpoint behavior and skewness depend on whether \(\alpha\) and \(\beta\) are below, near, or above \(1\).