Is it a Gaussian distribution?

It is a useful spherical analog, but its support is the sphere rather than unrestricted Euclidean space.

Can it work with hundreds of coordinates?

Yes, the mathematical family extends to higher-dimensional unit spheres. Numerical stability and suitable concentration estimates remain important.

Does mu equal the average of observed unit vectors?

The sample mean direction is obtained by normalizing the resultant vector when it is nonzero. The unnormalized average generally has a different length.

AI Glossary / Probabilistic machine learning / Advanced

Von Mises-Fisher distribution

Definition

The von Mises-Fisher distribution is a probability model for directions represented by unit-length vectors. It describes observations scattered around a preferred direction on a circle, sphere or higher-dimensional sphere, with a concentration parameter controlling how tightly they gather.

What it means

An ordinary numerical measurement can vary in size and direction. A directional measurement keeps only orientation. A camera's viewing direction, for example, can be represented by three coordinates whose squared values sum to one.

The model has a unit vector mu specifying its central direction and a nonnegative concentration kappa. Larger concentration means a tighter directional cluster. At zero concentration, the mathematical family has a uniform limiting distribution over the sphere. Particular software interfaces may require a strictly positive value.

Mu is a mean direction, not generally the ordinary expected vector. Averaging uncertain unit vectors usually produces a vector shorter than one.

How it works

For a unit observation x in d coordinates, the density is proportional to exp(kappa times mu dot x). The dot product equals the cosine of the angle between the observation and preferred direction. It equals one for agreement, zero for perpendicular directions and minus one for opposites.

A dimension-dependent normalizing constant makes the density integrate to one over the sphere's surface. This constant involves a modified Bessel function. Numerical libraries handle its evaluation; manual exponentiation can become unstable for large concentration.

A single component is rotationally symmetric around mu: directions at the same angle receive equal density. Fitting uses the directional observations to estimate a center and concentration. Mixtures can represent several directional groups, but require additional fitting and model-selection choices.

Three practical examples

A robotics team models repeated estimates of a landmark's bearing. The distribution describes uncertainty in orientation while distance is handled separately.
A search team explores clustering normalized document embeddings around several preferred directions. Whether the clusters capture useful topics must be tested against actual retrieval tasks.
A computer-graphics researcher samples incoming directions around a preferred orientation. The chosen directional model must match the physical or rendering approximation being made.

Think of it this way

Think of arrows painted from the center of a globe toward nearby points on its surface. The central arrow gives the preferred direction; concentration controls how closely the others gather around it.

The analogy concerns direction only. Arrow lengths do not encode varying speed, distance or confidence, and one circular patch cannot describe every possible pattern of directional uncertainty.

Worked example

Take the ordinary three-dimensional unit sphere, mu = (0, 0, 1), and kappa = ln(2). Consider north xN = (0, 0, 1), an equatorial point xE = (1, 0, 0), and south xS = (0, 0, -1).

Their dot products with mu are 1, 0 and -1. Their unnormalized densities are therefore 2, 1 and 1/2. North has twice the density of the equatorial point and four times the density of south. These are density ratios, not probabilities assigned to individual points.

For this three-dimensional sphere, write t = mu dot x. Surface rings give t a density proportional to exp(kappa t) for -1 <= t <= 1. Integrating from zero to one and dividing by the integral from minus one to one gives the northern hemisphere probability:

(exp(kappa) - 1) / (exp(kappa) - exp(-kappa)) = (2 - 1) / (2 - 1/2) = 2/3.

Thus a broad area has a meaningful probability, even though each exact point has probability zero under this continuous model. The hemisphere calculation is specific to the ordinary sphere; do not reuse it unchanged in every dimension.

Illustrative case study

Fictional case: a warehouse research team compares repeated directional estimates of fixed markers. It normalizes each nonzero bearing vector, records the coordinate convention and fits a directional distribution for each experimental condition.

The team notices one condition produces two separated clusters. Reporting a single center and concentration would hide that structure, so it investigates whether marker confusion or distinct operating conditions caused it. It also retains the original distance measurements in a separate table.

The project treats the model as an approximation to measured uncertainty. It does not infer reliable navigation or improved safety from a good directional fit alone.

Common misconceptions and mistakes

Concentration is an ordinary variance: increasing concentration tightens the distribution, and its relationship to dispersion depends on dimension.
Normalizing any vector is harmless: it discards magnitude, and a zero vector has no direction to normalize.
Opposite vectors always mean the same thing: they are distinct directions here. Data describing unoriented axes need a suitable axial model.
A high density is the probability of that exact direction: continuous probabilities come from integrating over regions.
One component can model arbitrary spread: rotational symmetry limits its shape.

Business use

This distribution offers a compact way to represent directional clusters and uncertainty. It can support embedding analysis, geometric perception and simulation when direction is the relevant measurement.

Before adopting it, establish why magnitude can be removed, how coordinates are defined and whether opposite directions are distinct. Inspect whether observed scatter is approximately symmetric or contains several modes. Evaluate held-out directional predictions and the downstream decision, rather than judging only the fitted concentration. A highly concentrated estimate can still point in the wrong direction if measurements are biased.

Check your understanding

Two unit vectors make the same angle with mu. Can this single-component model assign them different densities?

Answer: No. With fixed concentration, the density depends on their dot product with mu, so equal angles give equal densities.

Related terms

Unit vector
Cosine similarity
Normalization
Directional statistics
concentration parameter
spherical clustering
probability density
mixture model
axial data

Learn next

Practice converting nonzero vectors into directions and measuring their angles. Compare uniform spherical samples with concentrated samples. Then inspect a multimodal dataset and decide whether one rotationally symmetric component is a defensible description.

Sources

SciPy reference, vonmises_fisher, mathematical density, parameter constraints and implementation examples: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.vonmises_fisher.html

The directional comparison, hemisphere integration and fictional warehouse example are original teaching illustrations. The reference's mathematical definition was reviewed; the separate sampling-note PDF linked by it could not be retrieved during this check.

Previous: Von Kries adaptation Next: Voronoi cells