Statistical Notation for Conditioning (|) and Parameterization (;)

[Wolf]

Wolf

February 16, 2026 at 1:31 pm #5461

Understanding Probability Notation for AI Learners

In the context of Variational Autoencoders (VAEs) and Bayesian Deep Learning, we often see expressions like:

$$q_{\phi}(\mathbf{z} | \mathbf{x}) = \mathcal{N}(\mathbf{z}; \boldsymbol{\mu}, \boldsymbol{\Sigma})$$

This equation describes how an encoder maps an input $\mathbf{x}$ to a latent distribution $q$. Here is the breakdown of the specific symbols used to separate variables.

---

1. The Conditioning Bar ($|$)

The vertical bar $|$ denotes conditional probability. It represents a relationship between two random variables.

• Meaning: $P(A | B)$ means “The probability of $A$ occurring, given that $B$ has already occurred or is known.”
• In our example: $q(\mathbf{z} | \mathbf{x})$ means we are looking for the probability of a latent code $\mathbf{z}$ specifically for a given input image or data point $\mathbf{x}$.
• AI Context: This represents the Encoder. The latent vector $\mathbf{z}$ is “conditioned” on the input data $\mathbf{x}$.

---

2. The Parameter Semicolon ($;$)

The semicolon $;$ is used to separate the random variables from the fixed parameters that define the shape of the distribution.

• Meaning: Everything to the left of the semicolon is a variable we are evaluating; everything to the right is a constant (parameter) that tells the function how to behave.
• In our example: $\mathcal{N}(\mathbf{z}; \boldsymbol{\mu}, \boldsymbol{\Sigma})$ says “Evaluate the Gaussian density at point $\mathbf{z}$, using a specific mean $\boldsymbol{\mu}$ and covariance $\boldsymbol{\Sigma}$.”
• AI Context: Parameters are the “weights” or “settings.” While $\mathbf{z}$ changes during sampling, $\boldsymbol{\mu}$ and $\boldsymbol{\Sigma}$ are the outputs produced by your neural network layers.

---

Comparison Summary

Symbol	Name	Relationship Type	Example Role
$|$	Vertical Bar	Variable to Variable	Links input $\mathbf{x}$ to latent $\mathbf{z}$.
$;$	Semicolon	Variable to Parameter	Links $\mathbf{z}$ to its distribution shape ($\boldsymbol{\mu}, \boldsymbol{\Sigma}$).
$_{\phi}$	Subscript	Function to Parameter	Links the model $q$ to its learnable weights $\phi$.

---

Mathematical Structure

In deep learning, $\boldsymbol{\mu}$ and $\boldsymbol{\Sigma}$ are typically vectors/matrices generated by a network. For a 2D latent space, they might look like this:

$$\boldsymbol{\mu} = \begin{pmatrix} \mu_1 \cr \mu_2 \end{pmatrix}, \boldsymbol{\Sigma} = \begin{pmatrix} \sigma_{11} & \sigma_{12} \cr \sigma_{21} & \sigma_{22} \end{pmatrix}$$

By using the semicolon, we clarify that we aren’t “conditioning” on $\boldsymbol{\mu}$ in a probabilistic sense (treating it as a random variable), but rather using it as a fixed coefficient for the calculation.

• This topic was modified 3 months ago by Wolf.

[Author]

Posts