Is the Wishart distribution only for covariance matrices?

No. It can describe scatter and can also be used for positive-definite parameters such as precision.

Does the scale equal the mean?

Under the stated convention, the mean is degrees of freedom times scale.

Can a scatter matrix be singular?

Yes, especially when too few independent observations span the variable space.

AI Glossary / AI and machine learning / Advanced

Wishart distribution

Definition

The Wishart distribution is a probability distribution over real symmetric positive-semidefinite matrices, with the usual nonsingular form supported on positive-definite matrices. It describes scatter matrices formed from suitable Gaussian observations and is a standard tool for covariance estimation and multivariate Bayesian modeling.

What it means

For a single numerical variable, sums of squared Gaussian observations lead to a chi-squared distribution after appropriate scaling. The Wishart distribution extends that idea to several variables, replacing a single sum of squares with a matrix of sums of products.
The resulting matrix records how observations vary in each direction and together. Repeating the sampling process produces different matrices, and the distribution describes that variability.
The Wishart family also appears as a prior for a precision matrix in some Bayesian Gaussian models. Its role depends on the context: a random matrix representing observed scatter is conceptually different from a prior representing uncertainty about an unknown parameter.

How it works

Take nu independent zero-mean multivariate normal vectors with covariance Sigma, when nu is an integer. Sum their outer products to form S. Then S follows a Wishart distribution with degrees of freedom nu and scale Sigma under a common convention, and its expectation is nu Sigma.
The mathematical family extends degrees of freedom beyond this integer construction within its validity conditions. If observations are centered using an estimated mean, the usual normal-sample covariance result uses adjusted degrees of freedom. Dividing scatter by a normalizing factor rescales the matrix distribution. Check rank and parameter restrictions when using a software implementation, especially when the sample size is small relative to dimension.

Three practical examples

A statistician models sampling variation in a covariance matrix estimated from normal data.
A Bayesian analyst uses a Wishart prior for a multivariate normal model's precision matrix.
An engineering simulation generates valid random scatter matrices to assess a covariance-dependent calculation.

Think of it this way

Think of each observation contributing a small structured tile to a mosaic. Adding the tiles produces a summary of variation and co-variation across measurements. The Wishart model describes how the whole mosaic changes across repeated samples. The analogy captures addition but not the Gaussian assumptions or matrix constraints; arbitrary tiles do not produce a Wishart distribution.

Worked example

Consider two hand-chosen vectors x1=(1,2) and x2=(2,0). The first outer product has rows (1,2) and (2,4). The second has rows (4,0) and (0,0). Their sum S has rows (5,2) and (2,4).
The matrix is symmetric and positive definite, with determinant 5×4−2×2=16. Dividing it by 2 gives an average outer-product matrix with rows (2.5,1) and (1,2). It is not the same numerical object as scatter.
This arithmetic illustrates the construction, not a demonstration of a sampling distribution: the vectors were chosen rather than drawn from a specified Gaussian population. In a separate hypothetical model with nu=10 and scale equal to the two-dimensional identity matrix, the expected scatter is 10 times the identity. The scale itself is not the expected scatter under that convention.

Illustrative case study

Fictional case: Oak Reliability studies paired sensor errors and uses a Wishart model to simulate covariance-estimation uncertainty. An early implementation treats an estimated sample covariance as if it were an unnormalized scatter matrix, giving the simulated matrices an incorrect scale.
The team records the mean-handling and normalization conventions, verifies a small calculation and compares simulated summaries with the data. It also investigates serial dependence between successive measurements, since treating correlated records as independent can exaggerate the effective information. This case illustrates checking the model and implementation together; it does not claim that all sensor errors are Gaussian or that the Wishart assumption guarantees reliable forecasts.

Common misconceptions and mistakes

Scatter, sample covariance and population covariance are different quantities. Their relationships involve normalization and, sometimes, estimating the mean. A Wishart scale matrix is not generally the mean of the random matrix.
Positive diagonal entries alone do not establish positive definiteness. Limited sample rank can make a scatter matrix singular. Real and complex Wishart formulas differ. Finally, using a Wishart prior for precision does not mean the underlying measurements themselves have a Wishart distribution; those observations may instead be multivariate normal under the model.

Business use

The distribution is relevant to specialist work involving multivariate uncertainty, covariance estimates and Gaussian models. Business users may see its consequences in forecast ranges, simulation results or estimated dependence between operational measurements.
Ask which matrix the distribution represents, how its scale is defined and what assumptions justify treating observations as independent Gaussian vectors. Check whether sample count supports the chosen dimension and whether numerical calculations remain stable. Compare model-generated matrices or predictions with real data. A mathematically valid matrix distribution can still be a poor description of heavy-tailed, dependent or changing measurements, so its use should remain tied to evidence.

Check your understanding

A Wishart model uses degrees of freedom 6 and scale matrix Sigma. What is its expected matrix under the convention in this entry?

Answer: It is 6 Sigma. If a program expects a different parameterization, that relationship must be checked rather than assumed from the parameter name alone.

Related terms

Learn next

Learn outer products and sample covariance first. Construct a small scatter matrix by hand, then compare it with centered and normalized versions. Follow this with rank and positive-definiteness checks before using Wishart formulas in inference or simulation.

Scatter matrix

Sources

SciPy Wishart reference, density, scale convention and numerical-rank cautions consulted: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.wishart.html. Will Penny, Bayesian Inference for the Multivariate Normal, supplies a Gaussian outer-product construction using its own parameter convention: https://www.fil.ion.ucl.ac.uk/~wpenny/publications/bmn.pdf. The original scatter arithmetic and expected-matrix relationship were independently checked. The fictional implementation case describes incorrect scaling without asserting an unsupported direction.

Previous: Wisdom of the crowd Next: Wizard of Oz experiment