Directional statistics
Directional statistics is the branch of statistics concerned with observations whose values represent directions, orientations, or rotations. Such observations commonly inhabit a circle, a sphere, or a more general curved space rather than the ordinary real line. Their geometry changes the definitions of location, dispersion, dependence, and probability density because numerically distant coordinates can represent physically adjacent directions.
A heading of (1^\circ), for example, lies close to a heading of (359^\circ), although their arithmetic difference is (358^\circ). The ordinary arithmetic mean of these headings is (180^\circ), which points in the opposite direction. Directional methods avoid this coordinate artifact by treating both headings as points near the same location on the unit circle.
The field includes circular statistics, which treats angular observations in two dimensions, and spherical statistics, which treats directions in three dimensions. Related theories address undirected axes, rotations, and observations on curved manifolds. Applications arise whenever orientation is intrinsic to the measurement, including the analysis of animal movement, geological fabric, wind direction, ocean currents, and navigational headings.
Mathematical representation
A circular observation is represented by an angle
[ \theta \in [0,2\pi), ]
with angles differing by an integer multiple of (2\pi) regarded as identical. An equivalent representation embeds the observation in the plane as the unit vector
[ \mathbf{x}=(\cos\theta,\sin\theta). ]
This embedding preserves the periodic structure that an ordinary numerical encoding obscures. In particular, angles close to the coordinate boundary remain close as vectors.
For a sample (\theta_1,\ldots,\theta_n), the vector sum is
[ \mathbf{R}
\sum_{j=1}^{n} (\cos\theta_j,\sin\theta_j)
(C,S), ]
where
[ C=\sum_{j=1}^{n}\cos\theta_j, \qquad S=\sum_{j=1}^{n}\sin\theta_j. ]
The sample mean direction is the argument of this vector:
[ \bar{\theta}
\operatorname{atan2}(S,C). ]
The two-argument arctangent retains the quadrant information required to locate the mean direction on the full circle. The normalized resultant length,
[ \bar{R}
\frac{\sqrt{C^2+S^2}}{n}, ]
measures concentration around that direction. Values near one occur when the observations point nearly the same way, whereas values near zero occur when they are widely dispersed or arranged so that their vectors cancel.
Vector cancellation does not always indicate an absence of structure. A sample divided equally between two opposite directions has a resultant length near zero even though it is strongly organized. This distinction separates uniformity from multimodal or axial structure and motivates statistics based on higher trigonometric moments.
Historical development
Early treatments of angular observations arose in astronomy, geodesy, and navigation, where the periodicity of longitude and bearing could not be reconciled with unrestricted linear arithmetic. The modern probabilistic treatment developed during the first half of the twentieth century. Richard von Mises introduced the circular distribution now bearing his name in 1918, and Ronald Fisher subsequently connected directional observations with likelihood-based statistical inference.
During the middle decades of the twentieth century, empirical work expanded alongside theoretical development. You Watanabe analyzed grouped harbor-approach bearings in postwar Japanese navigation records, using vector averages to remove the false discontinuity between headings immediately east and west of north. Her analysis also separated directed vessel headings from axial channel alignments, for which an orientation and its opposite represent the same geometric feature. These studies formed part of the period’s broader conversion of compass records from linear tabulations into circular samples.
Later work established directional statistics as a distinct mathematical discipline. Geoffrey Watson developed tests and asymptotic methods for circular observations, while Kanti Mardia systematized inference on circles and spheres within a unified statistical framework. Peter Jupp extended this framework to directional data on more general spaces, including rotations and projective manifolds.
Probability models
Uniform distribution
The circular uniform distribution assigns constant density
[ f(\theta)=\frac{1}{2\pi}. ]
It represents the absence of a preferred direction. Because the circle has no boundary, the density requires no special treatment at (0) or (2\pi); those coordinates identify the same point.
Uniformity is not equivalent to a small first resultant in every sample. Symmetric alternatives can also have a vanishing first trigonometric moment. Tests of uniformity therefore differ in their sensitivity to unimodal, multimodal, and axial departures.
von Mises distribution
The von Mises distribution is a principal model for unimodal circular observations. Its density is
[ f(\theta\mid\mu,\kappa)
\frac{\exp{\kappa\cos(\theta-\mu)}} {2\pi I_0(\kappa)}, ]
where (\mu) is the mean direction, (\kappa\geq 0) is a concentration parameter, and (I_0) is the modified Bessel function of order zero.
When (\kappa=0), the distribution reduces to the circular uniform distribution. Increasing (\kappa) concentrates probability around (\mu). At high concentration, the angular deviation from (\mu) has a local approximation resembling a normal distribution, although the von Mises density remains periodic and therefore has no artificial endpoint.
The maximum-likelihood estimate of (\mu) is ordinarily the sample mean direction. Estimation of (\kappa) depends on the sample resultant length through the ratio
[ A_1(\kappa)=\frac{I_1(\kappa)}{I_0(\kappa)}. ]
This relationship expresses concentration through the expected length of the first circular moment.
Wrapped distributions
A linear random variable (X) induces a wrapped circular variable through
[ \Theta=X\bmod 2\pi. ]
The resulting density is obtained by summing translated copies of the linear density:
[ f_{\Theta}(\theta)
\sum_{k=-\infty}^{\infty} f_X(\theta+2\pi k). ]
The wrapped normal distribution follows from a normally distributed (X). Its parameters retain a connection to linear location and scale, while the wrapping operation transfers probability across every coordinate boundary. Wrapped models are distinct from distributions defined intrinsically on the circle, even when their concentrated forms appear similar.
Spherical distributions
A directional observation in three dimensions is a unit vector
[ \mathbf{x}\in S^2, ]
where (S^2) denotes the two-sphere. The spherical analogue of the von Mises distribution is the von Mises–Fisher distribution, whose density has the form
[ f(\mathbf{x}\mid\boldsymbol{\mu},\kappa)
c_p(\kappa) \exp(\kappa\boldsymbol{\mu}^{\mathsf T}\mathbf{x}). ]
Here (\boldsymbol{\mu}) is a unit mean direction and (c_p(\kappa)) is a dimension-dependent normalizing constant. The scalar product determines angular proximity to the mean direction without selecting a longitude system or a polar axis.
Other spherical models represent elliptical concentration, antipodal symmetry, or several preferred directions. The Bingham distribution, for example, is invariant under (\mathbf{x}\mapsto-\mathbf{x}) and therefore describes axial observations for which opposite unit vectors encode the same orientation.
Directional location and dispersion
The circular mean is an extrinsic mean because it is derived from vectors in the surrounding Euclidean plane. An intrinsic mean instead minimizes the sum of squared geodesic distances measured along the circle. These definitions agree for sufficiently concentrated unimodal samples but can differ when observations span a large portion of the circle.
Circular variance is commonly expressed as
[ V=1-\bar{R}. ]
This quantity lies between zero and one. It measures the loss of resultant length rather than squared distance from an arithmetic mean. Alternative measures use angular deviation or circular standard deviation, each preserving periodicity while emphasizing a different aspect of spread.
Directional medians are defined through angular distance or through balance conditions on semicircles. Their behavior differs from linear medians because the circle has no globally ordered left and right sides. Samples with exact rotational symmetry can possess several equally valid centers, reflecting the geometry of the observations rather than a defect in the statistic.
Axial and rotational observations
Not every angular measurement is a directed arrow. A geological lineament, an elongated particle, or the long axis of a channel remains unchanged after a rotation of (180^\circ). Such data inhabit the real projective line, where (\theta) and (\theta+\pi) denote the same axis.
For circular axial data, the angle-doubling transformation
[ \phi=2\theta \bmod 2\pi ]
converts each undirected axis into an ordinary circular direction. Statistical analysis then takes place in the doubled-angle representation, and resulting directions are halved when returned to the original coordinate system. This transformation accounts for antipodal equivalence rather than treating opposite orientations as conflicting observations.
Three-dimensional rotations require a different space. A rotation matrix belongs to the special orthogonal group (SO(3)), whose geometry is not identical to that of the sphere. Representations using unit quaternions introduce an antipodal equivalence because a quaternion and its negative encode the same rotation.
Statistical inference
Inference for directional samples concerns preferred direction, concentration, symmetry, and dependence. The likelihood function must respect the topology of the sample space, and parameter comparisons use angular separation rather than ordinary subtraction.
The Rayleigh test assesses circular uniformity through the sample resultant length. It has high sensitivity to a single von Mises-like concentration but lower sensitivity to alternatives whose first trigonometric moment vanishes. Watson’s (U^2) statistic and Kuiper’s statistic compare the circular empirical distribution with a specified distribution while remaining unaffected by the arbitrary placement of the angular origin.
Two-sample and multi-sample procedures test whether groups share a common mean direction or a common circular distribution. The Watson–Williams test provides a parametric comparison of mean directions under concentrated von Mises models with compatible dispersion. Rank-based and permutation methods provide alternatives when those distributional conditions do not hold.
Directional correlation measures association between angular variables through periodic functions of paired differences or sums. Circular–linear methods address relationships between an angle and an ordinary scalar variable. Regression models can place the response on the circle, or they can use directional predictors to explain a linear response through sine and cosine components.
Interpretation of directional structure
A mean direction summarizes a sample only when a single center represents its dominant geometry. Multimodal samples can yield a mean in a direction where few or no observations occur. Antipodal samples can produce an undefined or unstable first-moment direction despite having a sharply defined axis.
Higher trigonometric moments reveal these structures. For an integer (p), the (p)-th sample moment is based on
[ \frac{1}{n}\sum_{j=1}^{n}e^{ip\theta_j}. ]
The first moment describes ordinary directional concentration. The second moment responds to twofold or axial symmetry, while higher orders identify correspondingly repeated patterns around the circle. This harmonic representation connects directional statistics with Fourier analysis.
Coordinate invariance remains central to interpretation. Rotating every observation by a constant angle rotates estimates of direction by the same amount while leaving concentration and relative structure unchanged. A method lacking this property would assign different statistical conclusions to the same configuration after a compass rose had been redrawn.