The Dirichlet-Multinomial and Dirichlet-Categorical models for Bayesian inference

Size: px

Start display at page:

Download "The Dirichlet-Multinomial and Dirichlet-Categorical models for Bayesian inference"

Randall Horn
7 years ago
Views:

1 The Dirichlet-Multinomial and Dirichlet-Categorical models for Bayesian inference Stephen Tu 1 Introduction This document collects in one place various results for both the Dirichlet-multinomial and Dirichlet-categorical likelihood model. Both models, while simple, are actually a source of confusion because the terminology has been very sloppily overloaded on the internet. We ll clear the confusion in this writeup. 2 Preliminaries Let s standardize terminology and outline a few preliminaries. 2.1 Dirichlet distribution The Dirichlet distribution, which we denote Dir(α 1,..., α ), is parameterized by positive scalars α i > 0 for i1,...,, where 2. The support of the Dirichlet distribution is the ( 1)-dimensional simplex S ; that is, all dimensional vectors which form a valid probability distribution. The probability density of x (x 1,..., x ) when x S is f(x 1,..., x ; α 1,..., α ) Γ( i1 α i) i1 Γ(α i) i1 x α i 1 i There are some pedagogical concerns to be noted. First, the pdf f(x) is technically defined only in ( 1)-dimensional space and not ; this is because S has zero measure in dimensions 1. Second, the support is actually the open simplex, meaning all x i > 0 (which implies no x i 1). The Dirichlet distribution is a generalization of the Beta distribution, which is the conugate prior for coin flipping. 1 If you read the Wikipedia entry on the Dirichlet distribution, this is what the mathematical argon referring to the Lebesgue measure is saying. 1

Both models, while simple, are actually a source of confusion because the terminology has been very sloppily overloaded on the internet. We ll clear the confusion in this writeup.

2 2.2 Multinomial distribution The Multinomial distribution, which we denote Mult(p 1,..., p, n), is a discrete distribution over dimensional non-negative integer vectors x Z + where i1 x i n. Here, p (p 1,..., p ) is a element of S and n 1. The probability mass function is given as f(x 1,..., x ; p 1,..., p, n) i1 Γ(x i + 1) The interpretation is simple; a draw x from Mult(p 1,..., p ; n) can be intepreted as drawing n iid values from a Categorical distribution with pmf f(xi) p i. Each entry x i counts the number of times value i was drawn. The strange looking gamma functions in the pmf account for the combinatorics of a draw; it is simply the number of ways of placing n balls in bins. Indeed, since n! for non-negative integer n, i1 Γ(x i + 1) n! x 1! x! The Multinomial distribution is a generalization of the Binomial distribution. 2.3 Categorical distribution The Categorical distribution, which we denote as Cat(p 1,..., p ), is a discrete distribution with support {1,..., }. Once again, p (p 1,..., p ) S. It has a probability mass function given as f(xi) p i The source of confusion with the Multinomial distribution is because often the popular 1-of- encoding is used to encode a value drawn from the Categorical distribution, and in this case, we can actually see the Categorical distribution is ust a Multinomial with n 1. In the scalar form, the Categorical distribution is a generalization of the Bernoulli distribution (coin flipping). 2.4 Conugate priors Much ink has been spilled on conugate priors, so we won t attempt to provide any lengthy discussion. What we will mention now is that, when doing Bayesian inference, we are often interested in a few key distributions. Below, D refers to the dataset we have at hand, and we typically assume each y i D is drawn iid from the same distribution f(y; θ), where θ parameterizes the likelihood model. When people say they are being Bayesian, what this often means is they are treating θ as an unknown, but postulating that θ follows some prior distribution f(θ; α), where α parameterizes the prior distribution (and is often called the hyper-parameter). Of course you can keep playing this game and put a prior on α (called the hyper-prior), but we won t go there. i1 p x i i 2

.., p ; n) can be intepreted as drawing n iid values from a Categorical distribution with pmf f(xi) p i. Each entry x i counts the number of times value i was drawn.

3 While so far this seems reasonable, often Bayesian analysis is intractable since we have to integrate out the unknown θ, and for many problems the integral cannot be solved analytically (and one has to resort to numerical techniques). Conugate priors, however, are the (tiny 2 ) class of models where we can analytically compute distributions of interest. These distributions are: Posterior predictive. This is the distribution denoted notationally by f(y D), where y is a new datapoint of interest. By the iid assumption, we have f(y D) f(y, θ D) dθ f(y θ)f(θ D) dθ The distribution f(θ D), often called the posterior distribution, can be broken down further f(θ D) f(θ, D) f(d) f(θ, D) f(θ α) f(y i θ) Marginal distribution of the data. This is denoted f(d), and is derived by integrating out the model parameter f(d) f(d, θ) dθ f(θ α) f(y i θ) dθ Dirichlet distribution as a prior. It turns out (to further the confusion), that the Dirichlet distribution is the conugate prior for both the Categorical and Multinomial distributions! For the remainder of this document, we will list the results of both the posterior predictive and marginal distribution on both Dirichlet-Categorical and Dirichlet-Multinomial. 3 Dirichlet-Multinomial Model. p 1,..., p Dir(α 1,..., α ) y 1,...y Mult(p 1,..., p ) 2 This class is essentially ust the exponential family of distributions. 3

These distributions are: Posterior predictive. This is the distribution denoted notationally by f(y D), where y is a new datapoint of interest.

4 Posterior. f(θ D) f(θ, D) f(p 1,..., p α 1,..., α ) p α 1 p α 1+ y() i p y() i f(y i p 1,...p ) This density is exactly that of a Dirichlet distribution, except we have α α + y () i That is, f(θ D) Dir(α 1,..., α ). Posterior Predictive. f(y D) f(y θ)f(θ D) dθ f(y p 1,..., p )f(p 1,..., p D) Γ(y() + 1) p y() Γ( α ) Γ(y() + 1) Γ(α ) Γ( α ) Γ(α ) p y() +α 1 p α 1 Γ( α ) Γ(y() + α ) Γ(y() + 1) Γ(α ) Γ(n + α ) (1) where denotes integrating (p 1,..., p ) with respect to the ( 1) simplex. 4

Posterior Predictive. f(y D) f(y θ)f(θ D) dθ f(y p 1,..., p )f(p 1,.

5 Marignal. This derivation is almost the same as the posterior predictive. f(d) f(θ α) f(y i θ) dθ f(p 1,..., p α 1,..., α ) f(y i p 1,..., p ) Γ( α ) Γ(α ) Γ( α ) Γ(α ) Γ( α ) Γ(α ) [ [ p α 1 Γ(y() Γ(y() i + 1) ] i + 1) Γ(y() i + 1) p p y() i y() i +α 1 ] Γ( y() i + α ) Γ( D n + α ) (2) 4 Dirichlet-Categorical The derivations here are almost identical to before (with some minor syntatic differences). Model. Posterior. p 1,..., p Dir(α 1,..., α ) y Cat(p 1,..., p ) f(θ D) f(θ, D) f(p 1,..., p α 1,..., α ) p α 1 p 1{y i} p α 1+ 1{y i} f(y i p 1,...p ) This density is exactly that of a Dirichlet distribution, except we have α α + 1{y i } That is, f(θ D) Dir(α 1,..., α ). 5

Dirichlet-Categorical The derivations here are almost identical to before (with some minor syntatic differences). Model. Posterior. p 1,..., p Dir(α 1,..., α ) y Cat(p 1,.

6 Posterior Predictive. f(yx D) f(yx θ)f(θ D) dθ f(yx p 1,..., p )f(p 1,..., p D) Γ( p α ) x Γ(α ) Γ( α ) Γ(α ) p α 1 p 1{x}+α 1 Γ( α ) Γ(α ) Γ(1{x } + α ) Γ(1 + α ) α x α (3) where we used the fact that nγ(n) to simplify the second to last line. Marignal. f(d) f(θ α) f(y i θ) dθ f(p 1,..., p α 1,..., α ) f(y i p 1,..., p ) Γ( α ) Γ(α ) Γ( α ) Γ(α ) p α 1 p 1{y i} y p i D 1{y i}+α 1 Γ( α ) Γ( 1{y i } + α ) Γ(α ) Γ( D + α ) (4) 6

where we used the fact that nγ(n) to simplify the second to last line. Marignal.

Principle of Data Reduction

Chapter 6 Principle of Data Reduction 6.1 Introduction An experimenter uses the information in a sample X 1,..., X n to make inferences about an unknown parameter θ. If the sample size n is large, then