Statistical Models on Circles and Tori: von Mises, Wrapped Cauchy, and Kato‑Jones Distributions

What are Circular and Toroidal Data?

Circular data are observations measured on a one‑dimensional periodic domain, typically angles expressed in radians or degrees (e.g., wind direction, time of day). Toroidal data extend this concept to two independent periodic dimensions, forming a doughnut‑shaped space (e.g., longitude‑latitude on a sphere projected onto a torus, phase angles of two coupled oscillators).

Why Use Specialized Distributions?

Standard linear distributions (normal, exponential, etc.) assume an unbounded real line and ignore the wrap‑around nature of angles. Applying them to periodic data leads to biased estimates, incorrect inference, and misleading visualizations. Specialized circular/toroidal distributions respect periodicity, provide proper probability density on the unit circle or torus, and enable accurate parameter estimation.

von Mises Distribution

The von Mises distribution is the circular analogue of the normal distribution. It is defined by a mean direction μ and a concentration parameter κ that controls dispersion around μ.

• Parameters: μ (mean direction, 0 ≤ μ < 2π), κ ≥ 0 (concentration; κ = 0 yields a uniform distribution).
• Probability density: f(θ)=π^{-1}exp(κcos(θ‑μ))/I_0(κ), where I_0 is the modified Bessel function of order 0.
• Properties: Symmetric about μ, unimodal for κ>0, approaches a normal distribution on the real line as κ→∞.
• Typical uses: Wind direction analysis, animal movement bearings, time‑of‑day event modeling.

Wrapped Cauchy Distribution

The Wrapped Cauchy distribution results from wrapping the linear Cauchy distribution around the unit circle. It can model data with heavier tails than von Mises.

• Parameters: μ (location/mean direction), ρ (0 ≤ ρ < 1) controlling concentration; larger ρ yields tighter clustering.
• Probability density: f(θ)=π^{-1}(1‑ρ^2)/(1+ρ^2‑2ρcos(θ‑μ)).
• Properties: Symmetric, can represent multimodal behavior when combined in mixtures, heavier tails than von Mises.
• Typical uses: Modeling wind gust directions, phase noise in communication systems, any scenario where outliers are common.

Kato‑Jones Distribution

The Kato‑Jones distribution is a flexible family for toroidal data, extending the von Mises to two dimensions with separate concentration parameters and a correlation term.

• Parameters: μ₁, μ₂ (mean directions for each dimension), κ₁, κ₂ (concentrations), λ (dependence parameter, |λ| < √(κ₁κ₂)).
• Probability density: f(θ₁,θ₂)=C·exp[κ₁cos(θ₁‑μ₁)+κ₂cos(θ₂‑μ₂)+λcos(θ₁‑μ₁‑θ₂+μ₂)], where C is a normalizing constant.
• Properties: Captures both marginal concentration and inter‑dimensional correlation; reduces to independent von Mises when λ=0.
• Typical uses: Joint modeling of dihedral angles in protein structures, paired phase angles in signal processing, bivariate wind direction‑speed analyses.

How to Fit and Sample These Distributions

Modern statistical software provides built‑in functions for parameter estimation (maximum likelihood or Bayesian) and random sampling.

• Python (SciPy, statsmodels, pycircstat): Use scipy.stats.vonmises for von Mises, pycircstat.wrapped_cauchy for Wrapped Cauchy, and custom likelihood for Kato‑Jones.
• R (circular, Directional packages): Functions vonmisesMLE, wrappedcauchyMLE, and fit.katojones (available in the Directional package).
• Estimation steps: 1) Prepare angle data in radians; 2) Choose an initial guess (e.g., sample mean for μ); 3) Optimize the log‑likelihood using Newton‑Raphson or BFGS; 4) Assess fit with goodness‑of‑fit tests (Kuiper, Watson) and visual tools (circular histograms, contour plots).
• Sampling: After fitting, draw samples via the inverse‑transform method (von Mises) or rejection sampling (Wrapped Cauchy); for Kato‑Jones, use Metropolis‑Hastings or Hamiltonian Monte Carlo due to the correlation term.

Applications

Specialized circular and toroidal models are essential across many scientific and engineering domains.

• Environmental science – modeling wind and ocean current directions.
• Biology – analyzing animal movement paths, protein dihedral angles.
• Neuroscience – phase relationships between neural oscillations.
• Communications – phase noise and carrier recovery in modulated signals.
• Geoscience – orientation of geological features, fault line azimuths.