What does the Wishart part model?

The positive-definite precision matrix in this construction.

Are mean and precision independent?

Not generally; the conditional spread of the mean depends on precision.

Is normal-Wishart only another spelling of normal-inverse-Wishart?

No. They use related but different parameter representations that require a correct transformation.

AI Glossary / AI and machine learning / Advanced

Normal-Wishart distribution

Definition

The normal-Wishart distribution is a joint probability distribution for a mean vector and a precision matrix. It combines a Wishart distribution for precision with a multivariate normal distribution for the mean conditional on that precision, providing a conjugate prior for multivariate normal observations.

What it means

Precision is the matrix inverse of covariance. A covariance matrix describes spread and relationships among measurements; a precision matrix expresses related information in an inverse form useful in many statistical calculations. It is not obtained by taking the reciprocal of each covariance entry.
The normal-Wishart family represents uncertainty about the average vector and precision together. Conditional on a precision matrix, the mean has a normal distribution whose covariance depends on that matrix.
The family closely relates to normal-inverse-Wishart modeling, which uses covariance instead. Confusing the two representations can produce incorrect parameter settings even when the model names sound almost interchangeable.

How it works

Let Lambda denote precision. In a common scale convention, Lambda follows Wishart(nu,W). Conditional on Lambda, the mean mu follows a multivariate normal distribution with center m and covariance (kappa Lambda) inverse, with kappa positive.
Combine this prior with independent multivariate normal observations having mean mu and precision Lambda. Updating produces another normal-Wishart distribution, using the sample count, mean and scatter. The posterior preserves uncertainty rather than replacing the matrix with a single estimate. Different references may parameterize the Wishart using scale or inverse-scale quantities, so inspect the density or explicit definition before using an update formula.

Three practical examples

A Bayesian Gaussian model estimates a vector of average measurements and its precision matrix jointly.
A clustering research implementation uses normal-Wishart priors for Gaussian component parameters.
A simulation study compares inference expressed through covariance with an equivalent precision-based parameterization.

Think of it this way

Think of describing a cloud either by how widely it spreads or by how tightly it is concentrated in different directions. Covariance and precision offer those related perspectives. The normal-Wishart distribution also allows uncertainty about the cloud's center. The analogy is limited: in several dimensions, inversion mixes relationships across coordinates, so precision cannot be understood by independently reversing each visible spread value.

Worked example

Suppose a particular draw of the precision matrix Lambda is diagonal with entries 4 and 1, and kappa=2. The corresponding observation covariance Lambda inverse is diagonal with entries 1/4 and 1. The conditional covariance of the mean is (2 Lambda) inverse, giving diagonal entries 1/8 and 1/2.
These are different quantities: individual observations vary with covariance Lambda inverse, while the conditional mean uncertainty is smaller by the factor kappa. This example uses a diagonal matrix only to make the arithmetic transparent.
Now suppose Lambda is random under the Wishart part of the model. The inverse of its expected value generally differs from the expected value of its inverse. Therefore an analyst cannot simply invert a reported mean precision matrix and label the result the mean covariance without checking the required distributional calculation.

Illustrative case study

Fictional case: Meadow Analytics ports a Gaussian modeling routine between two libraries. One library's documentation describes covariance priors, while the other uses precision and a normal-Wishart parameterization. A developer initially reuses the same numerical matrix.
Reviewers compare simulated prior observations and discover that the implied spread is inconsistent. The team documents the matrix transformation and parameter conventions, then checks the posterior against a small independently calculated example. It also tests whether the Gaussian assumption fits the data. This case illustrates a translation risk; it does not suggest that one representation is inherently more accurate than the other.

Common misconceptions and mistakes

Precision is a matrix inverse, not an elementwise reciprocal. Wishart scale and inverse scale are not interchangeable. The mean vector and precision are linked through the conditional normal distribution rather than generally independent.
Do not confuse uncertainty about a mean with variation in individual observations. Conjugate updates are exact only for the stated probabilistic model. They do not establish normality, independence or causal relationships. Comparing covariance and precision summaries also requires care because taking expectations and taking matrix inverses do not generally commute.

Business use

Business teams usually encounter this distribution inside statistical software or a specialist modeling proposal. It can offer transparent, efficient updates when a multivariate normal model and its prior assumptions are appropriate.
Require analysts to describe the model in the units of the observed measurements, even if implementation uses precision matrices. Prior predictive simulations can reveal assumptions more clearly than a matrix of abstract parameters. Document any conversion between normal-Wishart and normal-inverse-Wishart forms, and compare outputs on a simple test case. If decisions are sensitive to the covariance structure, examine alternative priors and report that sensitivity instead of relying only on mathematical convenience.

Check your understanding

A precision matrix is diagonal with entries 2 and 8. What is the corresponding covariance matrix?

Answer: It is diagonal with entries 1/2 and 1/8. This simple reciprocal calculation works because the matrix is diagonal; a general matrix requires proper matrix inversion.

Related terms

Learn next

Learn covariance and precision matrices together, then compare equivalent Gaussian models written in each form. Simulate a Wishart precision matrix and a conditional mean to see the joint construction. Check scale conventions before attempting a posterior update from a textbook or software example.

Sources

Kevin P. Murphy, Conjugate Bayesian analysis of the Gaussian distribution, normal-Wishart section: https://www.cs.ubc.ca/~murphyk/Papers/bayesGauss.pdf. Will Penny, Bayesian Inference for the Multivariate Normal, Sections 3.1–3.2 consulted: https://www.fil.ion.ucl.ac.uk/~wpenny/publications/bmn.pdf. These sources use differing Wishart conventions, so the entry states its scale convention explicitly. Conditional covariance and diagonal examples were independently checked; inverse-of-expectation is not identified with expected inverse.

Previous: Normal matrix Next: Normality