Short Notes: The Multivariate Gaussian Distribution With a Diagonal Covariance Matrix

When the off-diagonal entries of a covariance matrix are negligible, it’s often convenient to use an alternative representation of a multivariate Gaussian distribution. In such cases, we can assume the covariance matrix is diagonal, allowing the mean and variance to be estimated independently for each dimension. This simplification leads to a factorized form of the multivariate density function as a product of univariate Gaussians, as shown below.

Assume we are given a diagonal covariance matrix of the form:

\[\Sigma = \begin{pmatrix} \sigma_1^2 & 0 & \cdots & 0 \\ 0 & \sigma_2^2 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & \sigma_k^2 \\ \end{pmatrix}.\]

The density function of a multivariate Gaussian distribution is given by:

\[f_x(x_1, \ldots, x_k) = \frac{\exp\left(-\frac{1}{2} (\vec{x} - \vec{\mu})^T \Sigma^{-1} (\vec{x} - \vec{\mu})\right)}{\sqrt{(2\pi)^k |\Sigma|}}.\]

Let us define:

\[\vec{y} = \vec{x} - \vec{\mu}.\]

Then the expression becomes:

\[f_x(x_1, \ldots, x_k) = \frac{\exp\left(-\frac{1}{2} \vec{y}^T \Sigma^{-1} \vec{y}\right)}{\sqrt{(2\pi)^k |\Sigma|}}.\]

Recall that the inverse of a diagonal matrix is simply the reciprocal of its diagonal elements:

\[\Sigma^{-1} = \begin{pmatrix} \sigma_1^2 & 0 & \cdots & 0 \\ 0 & \sigma_2^2 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & \sigma_k^2 \\ \end{pmatrix}^{-1} = \begin{pmatrix} \frac{1}{\sigma_1^2} & 0 & \cdots & 0 \\ 0 & \frac{1}{\sigma_2^2} & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & \frac{1}{\sigma_k^2} \end{pmatrix}.\]

Substituting this back into the expression:

\[f_x(x_1, \ldots, x_k) = \frac{\exp\left(-\frac{1}{2} \left( \frac{y_1^2}{\sigma_1^2} + \frac{y_2^2}{\sigma_2^2} + \cdots + \frac{y_k^2}{\sigma_k^2} \right) \right)}{\sqrt{(2\pi)^k |\Sigma|}}.\]

The determinant of a diagonal matrix is the product of its diagonal entries:

\[|\Sigma| = \begin{vmatrix} \sigma_1^2 & 0 & \cdots & 0 \\ 0 & \sigma_2^2 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \cdots & \sigma_k^2 \\ \end{vmatrix} = \sigma_1^2 \cdot \sigma_2^2 \cdots \sigma_k^2.\]

Putting it all together:

\[\begin{align} f_x(x_1, \ldots, x_k) &= \frac{\exp\left(-\frac{1}{2} \left( \sum_{i=1}^k \frac{y_i^2}{\sigma_i^2} \right) \right)}{\sqrt{(2\pi)^k \prod_{i=1}^k \sigma_i^2}} \\ &= \prod_{i=1}^k \frac{\exp\left( -\frac{y_i^2}{2\sigma_i^2} \right)}{\sqrt{2\pi \sigma_i^2}} \\ &= \prod_{i=1}^k \frac{\exp\left( -\frac{(x_i - \mu_i)^2}{2\sigma_i^2} \right)}{\sqrt{2\pi \sigma_i^2}} \\ &= f_1(x_1) \cdot f_2(x_2) \cdots f_k(x_k). \end{align}\]

This shows that under the assumption of a diagonal covariance matrix, the multivariate Gaussian density factorizes into a product of independent univariate Gaussian densities.

Related posts: The full (non-diagonal) covariance matrix plays a central role in the Mahalanobis distance, which measures how far a point lies from a distribution — with applications to anomaly detection. For Python implementations and benchmarks, see Implementing the Mahalanobis Distance in Python.