liveBook · Manning

Dirichlet distribution

As explained in Grokking Bayes

The Dirichlet distribution is a continuous probability distribution over categorical probability vectors, parameterized by positive concentration parameters.

When an experiment has more than two possible outcomes, each outcome has its own unknown probability. The Dirichlet distribution represents beliefs about that whole collection of probabilities, extending the beta distribution beyond two outcomes. section 3.9

It is a conjugate prior for the multinomial likelihood. Its hyperparameters record prior information for each outcome, and observed counts update those values directly. section 3.9

Updating a three-outcome prior

Start with Dir(1, 2, 1) for three possible outcomes. After observing n₁, n₂, and n₃ instances, the posterior becomes Dir(n₁ + 1, n₂ + 2, n₃ + 1). Each prior hyperparameter increases by the count for its corresponding outcome. section 3.9

How Grokking Bayes covers it

How deep it goes
Explains the conjugate posterior rule and illustrates it with a three-outcome calculation.
Examples
Uses a Dir(1, 2, 1) prior and updates it with three observed outcome counts.
What it assumes
Assumes familiarity with Bayesian priors and likelihoods while extending the beta-distribution idea from two outcomes to many.
What you can do after
Evaluate a three-outcome Bayesian update by adding observed counts to matching Dirichlet hyperparameters.
How it connects
Connects categorical probabilities with multinomial likelihoods, beta distributions, conjugate priors, and Bayesian models.

Also explained in

Compare how 3 books explain Dirichlet distribution →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Grokking Bayes. The section links point to the parts of the book it draws on.