Do the shares always sum to one?

Yes, in this construction the final remainder completes the total.

Can the final shares be independent?

They are constrained by their fixed sum and generally dependent.

Is the ordinary Dirichlet included?

Yes, suitable parameter relationships recover it as a special case.

AI Glossary / AI and machine learning / Advanced

Generalized Dirichlet distribution

Definition

The generalized Dirichlet distribution is a family of probability distributions for vectors of proportions that sum to one. In the Connor - Mosimann construction, independent beta-distributed fractions allocate successive shares of the remaining total, allowing more flexible dependence than an ordinary Dirichlet distribution.

What it means

A vector of proportions might describe how activity is divided among categories. Increasing one share leaves less total for the others, so modeling each share independently is usually inappropriate. The ordinary Dirichlet distribution is a common solution, but it imposes a particular pattern of dependence.
The generalized Dirichlet considered here extends that model through a sequential allocation construction. It has more parameters and can represent dependence patterns unavailable to the ordinary family, including some positive pairwise correlations.
Generalized does not mean universally better. The extra flexibility requires enough information to estimate parameters, and the order assigned to categories can become part of the model specification.

How it works

Start with a total of one. Draw a beta-distributed fraction and give that fraction to the first category. Draw another independent beta fraction and allocate that fraction of the remainder to the second category. Continue until the final category receives whatever remains.
For three categories, write the independent fractions as V1 and V2. The shares are X1=V1, X2=(1−V1)V2 and X3=(1−V1)(1−V2). They are nonnegative and sum to one. Appropriate restrictions on the beta parameters recover an ordinary Dirichlet distribution. Other settings provide additional flexibility. Independence applies to the allocation fractions, not generally to the resulting category shares.

Three practical examples

A researcher models proportions of several activity types when a standard Dirichlet model fits their dependence poorly.
A planning analyst represents uncertainty in a composition of resource allocations across categories.
A statistical modeler generates synthetic category-probability vectors to test how an analysis responds to different dependence assumptions.

Think of it this way

Think of cutting a ribbon: take a fraction for the first use, then a fraction of what remains for the second, and leave the rest for the third. Each cut uses the remaining ribbon as its reference. The analogy captures the construction, but the category order need not describe a real chronological decision process. It is a mathematical modeling choice that should be documented.

Worked example

Suppose one draw gives V1=0.20 and V2=0.50. The first category receives 0.20. The second receives half of the remaining 0.80, or 0.40. The final category receives the other 0.40. The resulting composition is therefore (0.20,0.40,0.40), not (0.20,0.50,0.30).
For a separate calculation, suppose V1 has beta parameters (2,3), giving mean 2/5, and V2 has parameters (3,2), giving mean 3/5. Independence of the fractions gives expected shares 0.40, (0.60)(0.60)=0.36 and (0.60)(0.40)=0.24.
Those expected shares sum to one. They do not imply that any particular draw equals the mean composition. Also, independence of V1 and V2 does not imply independence of X1 and X2, because the second share depends on how much the first allocation leaves.

Illustrative case study

Fictional case: Pine Research analyses the proportions of time participants spend on three activities. A standard Dirichlet model does not reproduce some observed relationships between activity shares. The team considers a generalized Dirichlet alternative.
Before adopting it, the analysts compare predictive checks, explain the selected activity order and repeat the analysis with a different ordering. They discover that interpretation depends on the sequential parameterization, so the reporting includes that choice rather than presenting the distribution as an automatic upgrade. The example illustrates model assessment. It does not establish that this family will fit every composition or remove bias from inaccurate time records.

Common misconceptions and mistakes

Do not allocate every beta draw as a fraction of the original total; later fractions apply to the remainder. Do not confuse independent allocation fractions with independent final proportions.
The generalized Dirichlet is not the same as every distribution bearing a similar generalized name. State the construction or reference being used. Extra parameters can overfit limited observations. Exact zero shares and structural absence of categories also need deliberate treatment, since a continuous interior distribution does not automatically model a separate probability mass at zero.

Business use

This is a specialist option for analyzing compositions or uncertain category probabilities when simpler assumptions are inadequate. The business question should concern relationships among shares, rather than unrelated numerical forecasts.
Teams should compare it with an ordinary Dirichlet and other compositional models using relevant predictive checks. Record category definitions, ordering, any treatment of zeros and the distinction between observed counts and underlying probabilities. If counts are modeled, specify the observation model as well as the distribution over proportions. A more flexible prior does not compensate for inconsistent categories, a shifting data-collection process or a sample that fails to represent future conditions.

Check your understanding

The first category receives 30% of a total, and the next beta fraction is 40%. What share of the original total goes to the second category?

Answer: Forty per cent of the remaining 70%, which is 28% of the original total. The final remainder in a three-category construction is 42%.

Related terms

Dirichlet distribution
Covariance
Multinomial distribution
Grouped Dirichlet distribution
beta distribution
compositional data
stick-breaking construction

Learn next

Learn the ordinary Dirichlet first, then simulate the sequential beta construction and compare means and correlations. Change the category order deliberately and inspect what changes. This makes the additional flexibility and the assumptions behind it easier to understand than reading a density formula alone.

Sources

Connor and Mosimann, Concepts of Independence for Proportions with a Generalization of the Dirichlet Distribution: https://www.tandfonline.com/doi/abs/10.1080/01621459.1969.10500963. Original full text was unavailable. Wakai and colleagues, Section 1.1, independently consulted for the sequential beta construction, order dependence, correlation flexibility and ordinary-Dirichlet special case: https://link.springer.com/article/10.1007/s41237-025-00273-6. The original numerical examples were independently checked.

Previous: Generalized canonical correlation Next: Generalized distributive law