Normal distributions

Outline

Topics

  • Quick review of the normal distribution.
  • Different parameterizations.

Rationale

The normal distribution is often used in Bayesian analysis when one needs a prior over an unknown \(x\) such that \(x \in (-\infty, \infty) = \mathbb{R}\).

Definition

A probability density \(f\) is normal (or Gaussian) when its logarithm is a degree two polynomial:

\[f(x) \propto \exp(\text{polynomial of degree two in }x).\]

Examples of normal densities

Parameterizations

  • There are different conventions to measure the spread.
    • Standard deviation \(\sigma\).
    • Variance, \(\sigma^2\).
    • Precision, \(\tau = 1/\sigma^2\).
  • Keep that in mind as different languages will use different conventions!
  • Standard deviation is the most intuitive:
    • it is the width of the bell,
    • the only one that has the same units as \(x\) (e.g. if \(x\) is in meters, so is \(\sigma\)).

Density for variance \(\sigma^2\) and mean \(\mu\):

\[f(x) \propto \exp\left( - \frac{(x - \mu)^2}{2\sigma^2} \right).\]

Multivariate version

Exact same definition: A joint probability density \(f\) is multivariate normal when its logarithm is a degree two polynomial:

\[f(x_1, x_2, \dots, x_d) \propto \exp(\text{polynomial of degree two in }x_1, x_2, \dots, x_d).\]

In the multivariate case the most popular parametrization is the generalization of variance, the covariance matrix.

Warning: notation trap

Unfortunately, the covariance is almost always denoted \(\Sigma\), which is the capital letter for \(\sigma\), but really it is the generalization of \(\sigma^2\), not of \(\sigma\)! We will stick to the standard notation.

Density: where \(x = (x_1, x_2, \dots, x_d)\),

\[f(x) \propto \exp\left( - \frac{1}{2} (x - \mu)^{\text{T}} \Sigma^{-1} (x - \mu) \right).\]