Skip to content

Distributions

We first introduce some notation for probability distributions:

Mathematical Description Programmatic
$p(\mathrm{d}x)$ A distribution. p:Distribution<X>
$p(x)$ Evaluate the probability density (mass) function associated with the probability distribution $p(\mathrm{d}x)$ at the point $x$.
$\log p(x)$ Evaluate the logarithm of the probability density (mass) function associated with the distribution $p(\mathrm{d}x)$ at the point $x$. x ~> p
$x\sim p(\mathrm{d}x)$ Simulate a variate $x$ from the distribution $p(\mathrm{d}x)$. x <~ p

This notation may be unfamiliar, particularly as many texts rely on context, rather than notation, to distinguish between a probability distribution $p(\mathrm{d}x)$ and its associated probability density (mass) function $p(x)$, using the notation $p(x)$ for both. In what follows, the distinction in notation becomes useful. In a probabilistic program we can perform many computations associated with the one probability distribution: simulate from it, evaluate its probability density (mass) function, evaluate its cumulative distribution function, compute its mean or variance, upper or lower bound, median or some other quantile. So we make the distinction in notation to use $p(\mathrm{d}x)$ to denote the distribution itself, and $p(x)$ for the particular computation of its probability density (mass) function.

Tip

You may recognize the notation $p(\mathrm{d}x)$ from measure theory. We will not adopt measure-theoretic terms otherwise, but find the notation useful.

In Birch code, a distribution is represented by an object of the Distribution class. This is a generic class: we use it as Distribution<X>, where X is the domain of the distribution, e.g. Distribution<Real> (over $\mathbb{R}$), Distribution<Integer> (over $\mathbb{Z}$), Real[_] (over $\mathbb{R}^D$), etc. However, we do not usually use Distribution<X> directly. Instead we use one of its derived classes, such as Gaussian, Gamma, Beta, Uniform. The idiom is to use a factory function for the particular distribution of interest in combination with a probabilistic operator. For example, we can simulate from a distribution with the simulate operator (<~):

x:Real;
x <~ Gaussian(0.0, 4.0);

The factory function Gaussian() creates an object of class Gaussian, which derives from class Distribution<Real>. The <~ operator then simulates a variate from it, and assigns the value of that variate to the variable x. We can instead use code such as the following:

x:Real;
p:Distribution<Real> <- Gaussian(0.0, 4.0);
x <~ p;
which works, and is perfectly correct, just not idiomatic.

We can observe a variate with the observe operator (~>). For example, to observe a variate of value 1.5823 from a Gaussian distribution with mean 0.0 and variance 4.0:

1.5823 ~> Gaussian(0.0, 4.0);
Or we could, of course, assign the value to a variable first and use that:
let x <- 1.5823;
x ~> Gaussian(0.0, 4.0);