Von Mises--Fisher (vMF) Distribution

• The von Mises–Fisher (vMF) distribution is a rotationally symmetric probability distribution used for directional data on the unit hypersphere, characterized by mean direction and concentration parameter, such as in predicting angular movements or modeling directional statistical following an exponential family form.
• Researchers leverage vMF for applications such as directional modeling in deep learning, wireless-channel modeling, and positioning because it is analytically tractable and interpretable for directional data, enhances the Mixture Model's directional clustering.
• In practical applications, vMF excels in defining robust geometric structures on spheres, making it a valuable tool for studying directional data in high-dimensional spaces, emphasizing the ability to integrate into larger models and reducing bias throughout the parameters.

The von Mises–Fisher (vMF) distribution is a rotationally symmetric probability distribution for directional observations on the unit hypersphere. For $\mathbf{x}\in\mathbb{S}^{d-1}\subset\mathbb{R}^{d}$, it is parameterized by a mean direction $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$ and a nonnegative concentration parameter $\kappa$. Its density depends on the direction only through the cosine similarity $\boldsymbol{\mu}^{T}\mathbf{x}$, making vMF the standard spherical analogue of an isotropic Gaussian for data whose orientation matters but whose magnitude does not. It is used in directional statistics, mixture modelling, Bayesian inference, deep representation learning, positioning, spatial regression, wireless-channel modelling, orbital dynamics, reinforcement learning, and scientific analyses of directional physical processes.

1. Definition and geometry

The unit hypersphere is

$$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$

The vMF density with respect to surface-area measure on $\mathbb{S}^{d-1}$ is

$$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$

where

$$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$

and $I_\nu(\kappa)$ is the modified Bessel function of the first kind of order $\nu$. The constraints are $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$0 and $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$1.

Because both $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$2 and $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$3 have unit norm,

$\boldsymbol{\mu}\in\mathbb{S}^{d-1}$4

where $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$5 is their angular separation. Consequently,

$\boldsymbol{\mu}\in\mathbb{S}^{d-1}$6

The density is maximal at $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$7 and constant on hyperspherical level sets having the same angular distance from $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$8. The distribution is rotationally symmetric around its mean direction. For $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$9, the support is the unit circle and the distribution reduces to the ordinary von Mises distribution; for $\kappa$0, it is the Fisher distribution on the ordinary unit sphere.

The concentration parameter controls angular dispersion. At $\kappa$1, the distribution is uniform on the sphere. Increasing $\kappa$2 concentrates probability increasingly tightly around $\kappa$3, and the limit $\kappa$4 approaches a point mass at $\kappa$5. The vMF distribution is therefore not a Gaussian conditioned literally on a sphere; rather, it is a canonical rotationally symmetric exponential-family distribution on spherical support (Cao et al., 2020).

The Gaussian analogy follows from the identity

$\kappa$6

Thus, on the unit sphere,

$\kappa$7

This interpretation is extrinsic: it uses chordal Euclidean distance in the ambient space. An intrinsic spherical normal instead uses squared great-circle distance, and a spherical Laplace model uses unsquared geodesic distance (You, 2021, You et al., 2022).

The vMF family is an exponential family with natural parameter

$\kappa$8

Its sufficient statistic is $\kappa$9, and its exponential-family structure supports analytical moment, entropy, and divergence calculations. The natural-parameter representation is also useful when vMF distributions are employed as priors or quasi-posteriors on parameters constrained to have unit norm (Kitagawa et al., 2022).

2. Moments and concentration

Define the Bessel-function ratio

$\boldsymbol{\mu}^{T}\mathbf{x}$0

The first moment is

$\boldsymbol{\mu}^{T}\mathbf{x}$1

The vector expectation generally lies inside the unit ball, while $\boldsymbol{\mu}^{T}\mathbf{x}$2 is the direction of the mean resultant. The scalar $\boldsymbol{\mu}^{T}\mathbf{x}$3, also called the mean resultant length, lies between zero and one for positive $\boldsymbol{\mu}^{T}\mathbf{x}$4. It tends to zero as $\boldsymbol{\mu}^{T}\mathbf{x}$5 tends to zero and tends to one as concentration becomes large.

The covariance can be expressed as

$\boldsymbol{\mu}^{T}\mathbf{x}$6

For large $\boldsymbol{\mu}^{T}\mathbf{x}$7, $\boldsymbol{\mu}^{T}\mathbf{x}$8 is of order $\boldsymbol{\mu}^{T}\mathbf{x}$9, so angular variance decreases with concentration. Locally, a highly concentrated vMF distribution resembles an isotropic Gaussian in the tangent plane, with approximate tangent-space variance $$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$0 in each of the $$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$1 tangent directions (Kitagawa et al., 2022).

For observations $$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$2, the resultant vector and resultant length are

$$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$3

The resultant vector is the principal sufficient statistic for single-component estimation. The maximum-likelihood estimate of the mean direction is

$$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$4

The maximum-likelihood concentration solves

$$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$5

There is generally no elementary closed-form inverse of $$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$6. A commonly used approximation is

$$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$7

This approximation is used in vMF mixture estimation, deep learning, gradient-direction analysis, and hyperspherical hashing (Kasarapu et al., 2015, Hasnat et al., 2017, Lee et al., 2018, Kim, 2021).

Concentration estimation can be substantially biased for small samples and high dimensions. Minimum Message Length (MML) estimation modifies the maximum-likelihood equation by adding prior and Fisher-information precision costs. Conventional estimation solves $$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$8, whereas MML solves a penalized nonlinear equation. The resulting correction is particularly relevant for small effective sample sizes (Kasarapu et al., 2015).

3. Likelihood inference and mixture models

A finite vMF mixture with $$\mathbb{S}^{d-1}=\{\mathbf{x}\in\mathbb{R}^{d}:\|\mathbf{x}\|_2=1\}.$$9 components has density

$\mathbb{S}^{d-1}$0

where $\mathbb{S}^{d-1}$1, $\mathbb{S}^{d-1}$2, $\mathbb{S}^{d-1}$3, and $\mathbb{S}^{d-1}$4.

Introducing latent assignments $\mathbb{S}^{d-1}$5, the posterior responsibility of component $\mathbb{S}^{d-1}$6 for observation $\mathbb{S}^{d-1}$7 is

$\mathbb{S}^{d-1}$8

The effective membership and weighted resultant are

$\mathbb{S}^{d-1}$9

The EM updates for a fixed number of components include

$$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$0

$$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$1

and

$$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$2

The likelihood formulation has an important degeneracy. A component can place its mean direction exactly at an observation and let $$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$3 diverge. Since

$$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$4

the density at the aligned observation diverges polynomially, and the ordinary mixture likelihood is unbounded. Consequently, a global ordinary maximum-likelihood estimator need not exist for finite vMF mixtures (Ng, 2020).

Penalized maximum likelihood addresses this problem by adding a penalty depending on component concentrations. A computationally used linear penalty is

$$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$5

with $$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$6 of order $$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$7. The resulting penalized EM algorithm retains the standard responsibility and direction updates, while the concentration equation becomes

$$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$8

where $$f(\mathbf{x};\boldsymbol{\mu},\kappa) = C_d(\kappa) \exp\!\left(\kappa\boldsymbol{\mu}^{T}\mathbf{x}\right),$$9 is the responsibility-weighted resultant. Under the stated penalty conditions, the estimator is strongly consistent in the quotient parameter space that identifies mixtures differing only by component-label permutations (Ng, 2020).

MML provides a different form of regularization. It jointly encodes the model and the observations,

$$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$0

and selects the number of components by minimizing total message length. The message includes the number of components, mixture weights, component parameters, and data given the mixture. Split, delete, and merge operations search over model sizes, accepting a perturbation only when it decreases total message length. This approach differs from ordinary EM because it estimates parameters and mixture complexity jointly rather than assuming $$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$1 is known (Kasarapu et al., 2015).

Finite vMF mixtures are used for directional clustering of normalized text vectors, protein conformations, image embeddings, and other spherical data. In text clustering, the number of statistically useful components can exceed the number of externally supplied semantic categories, because MML selects components according to compression rather than label agreement. In protein-conformation modelling, smooth vMF mixtures assign small but nonzero probability to rare observations, avoiding the infinite code lengths that would result from a hard-truncated directional model (Kasarapu et al., 2015, Kim, 2021).

4. Estimation, computation, and alternatives

The principal computational difficulty is evaluation of the Bessel-function normalizer, especially when the dimension or concentration is large or when the Bessel argument is small. Direct floating-point evaluation can underflow, and ratios of separately evaluated Bessel functions can become numerically unstable. Implementations use logarithmic normalizers, stable Bessel-ratio routines, arbitrary-precision evaluation, or custom automatic-differentiation functions. A PyTorch implementation uses MPMath for arbitrary-precision Bessel evaluation and supports both full-batch EM and minibatch stochastic-gradient learning (Kim, 2021).

For mixture probabilities, log-domain calculations and logsumexp are used. Mixture weights can be parameterized by a softmax, concentrations by a positive parameterization such as softplus, and mean directions by normalizing unconstrained vectors. These parameterizations enforce

$$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$2

Exact vMF sampling is commonly based on a tangent-normal decomposition. If $$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$3, then

$$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$4

where $$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$5 is uniformly distributed on the orthogonal subsphere. The marginal density of $$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$6 has the form

$$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$7

The cumulative distribution of this marginal has no known analytic inverse, so standard sampling uses rejection sampling. This can be slow, inefficient in batched GPU execution, and difficult to differentiate through with the reparameterization trick. The Power Spherical distribution replaces the exponential-Bessel marginal with an affine-transformed Beta distribution, enabling direct differentiable sampling and improved numerical stability while retaining spherical support and rotational symmetry (Cao et al., 2020).

The vMF is also an extrinsic model. Its density is a function of chordal distance, while spherical-normal models use intrinsic great-circle distance,

$$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$8

The spherical normal has density proportional to

$$C_d(\kappa) = \frac{\kappa^{d/2-1}} {(2\pi)^{d/2}I_{d/2-1}(\kappa)},$$9

whose location parameter is an intrinsic Fréchet mean rather than the normalized Euclidean resultant. Its normalizer requires a one-dimensional integral, so it is less algebraically convenient than the vMF, but it may better represent models whose geometry is intrinsically geodesic (You, 2021).

The vMF also assumes rotational symmetry around a single direction. Fisher–Bingham, Kent, Bingham, and related distributions provide greater flexibility for anisotropic, axial, or asymmetric directional data. In particular, vMF treats $I_\nu(\kappa)$0 and $I_\nu(\kappa)$1 as distinct directions, making it unsuitable when observations represent unoriented axes rather than directed vectors (lan et al., 2022, Li et al., 16 Oct 2026).

5. Deep learning and directional representation learning

In deep learning, a network output $I_\nu(\kappa)$2 can be normalized to

$I_\nu(\kappa)$3

so that features lie on $I_\nu(\kappa)$4. A class can then be represented by a vMF component with mean direction $I_\nu(\kappa)$5 and concentration $I_\nu(\kappa)$6. The resulting class posterior is

$I_\nu(\kappa)$7

Under the equal-privilege assumptions $I_\nu(\kappa)$8 and $I_\nu(\kappa)$9, the posterior simplifies to

$\nu$0

Thus, vMF classification is a softmax over scaled cosine similarities. The concentration acts as a shared logit scale and determines the sharpness of the class posterior. The cross-entropy objective simultaneously encourages intra-class compactness and inter-class angular separation. This formulation provides a probabilistic foundation for feature normalization, normalized class weights, zero bias, and cosine-based logits (Hasnat et al., 2017).

The vMF mixture loss is mathematically a softmax cross-entropy, but the logits have a constrained geometric interpretation:

$\nu$1

Its connection to center loss follows from the unit-sphere identity

$\nu$2

Under unit normalization, maximizing cosine similarity is equivalent to minimizing squared chordal distance to the class direction. The relationship to L2-softmax, NormFace, weight normalization, and angular-softmax objectives holds under the corresponding unit-norm, zero-bias, and scale assumptions; these objectives are not identical in all unconstrained implementations (Hasnat et al., 2017).

The same geometry is used in supervised deep hashing. Time series are mapped to normalized $\nu$3-dimensional embeddings, each class is assigned its own vMF distribution, and the final binary code is produced by sign thresholding. Class-specific concentrations are estimated from the mean resultant length and can be downscaled during training to induce an implicit angular margin. The method seeks to make semantic clusters compact and separated before discretization, although it does not mathematically eliminate quantization loss or directly penalize coordinates near zero (Perez et al., 23 Sep 2025).

vMF models have also been used to analyse stochastic gradient descent. Normalized minibatch-gradient directions are treated as observations on a high-dimensional sphere. Their mean direction estimates the dominant update direction, while the estimated concentration summarizes directional consistency. Under local smoothness, small-step, approximately equal-norm, and well-conditioned-Hessian assumptions, the analysis predicts that SGD decreases directional concentration over training. This is interpreted as increasing directional uniformity and eventual directional cancellation, rather than merely the disappearance of gradient norms (Lee et al., 2018).

6. Applications and extensions

The vMF model is applicable whenever observations are directions, normalized embeddings, or parameters constrained to a sphere. In angle-of-arrival positioning, azimuth and elevation are converted into a three-dimensional unit vector,

$\nu$4

The vMF measurement model avoids azimuth wrap-around, the undefined azimuth at the poles, and coordinate-dependent uncertainty. Particle filters can use the exact vMF likelihood, while EKF and UKF implementations use a local Cartesian Gaussian approximation. In pole-heavy positioning scenarios, vMF-based filters outperform conventional Gaussian azimuth/elevation filters, especially for EKF and UKF methods (Nurminen et al., 2017).

Spatial vMF regression models covariate-dependent directional responses using a link from Cartesian coordinates to spherical angles, followed by logit transformations that produce unconstrained predictors. An unknown rotation matrix separates coordinate-system adjustment from the regression link, while autoregressive random effects model dependence along fiber streamlines. In diffusion-tensor imaging, the model has been applied to brain-fiber directions and used to study associations between clinical covariates, including APOE status, and spatially varying fiber orientations (lan et al., 2022).

In wireless propagation, vMF distributions model three-dimensional directions of arrival of multipath scatterers. The Doppler frequency is the projection of a random arrival direction onto the motion direction. For three-dimensional scattering, the Doppler density is

$\nu$5

for $\nu$6, where $\nu$7 is the angle between the mean scattering direction and the motion direction. Parallel motion produces an exponential spectrum, while perpendicular motion produces an approximately Gaussian-like spectrum at high concentration (Turbic et al., 2024).

The associated spatial correlation function for vMF scattering has a closed form,

$\nu$8

The isotropic limit is the three-dimensional sinc correlation. The concentrated limit gives slower decorrelation for displacements compatible with the mean scattering direction. The model extends to multiple scattering clusters through a power-weighted sum of component correlations (Turbic et al., 2024).

The first two Doppler moments are

$\nu$9

and

$\boldsymbol{\mu}\in\mathbb{S}^{d-1}$00

where $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$01. The level-crossing rate and average fade duration for Rayleigh fading depend on $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$02:

$\boldsymbol{\mu}\in\mathbb{S}^{d-1}$03

$\boldsymbol{\mu}\in\mathbb{S}^{d-1}$04

Isotropic scattering gives the largest level-crossing rate, whereas increasing concentration reduces Doppler spread and slows fading (Turbic et al., 6 Jun 2025).

Other physical applications include orbital-pole analysis and solar-wind magnetic-field directions. For Plutino orbital poles, the vMF model respects the physical spherical support of inclinations and avoids the truncation required by Gaussian or Rayleigh inclination models (Matheson et al., 2023). In the young solar wind, a two-component vMF mixture represents a broad weakly confined population associated with large rotations and switchbacks and a more concentrated field-aligned population associated with small transverse Alfvénic fluctuations. The broad component is exponential in pitch-angle cosine rather than polar angle, while the concentrated component has a Gaussian small-angle limit. Weak transverse anisotropy motivates a Bingham refinement of the vMF core (Li et al., 16 Oct 2026).

In reinforcement learning with very large action sets, normalized action embeddings are points on a hypersphere. vMF exploration samples a continuous direction around the current state embedding and maps it to a discrete action by nearest-neighbor retrieval. The probability of an action is the vMF probability mass of its spherical Voronoi cell, not merely the density evaluated at the action embedding. Under independent uniformly distributed action embeddings, this procedure is asymptotically equivalent to Boltzmann exploration while requiring one hyperspherical sample and an approximate nearest-neighbor query rather than an action-wide softmax (Bendada et al., 1 Jul 2025).

A recent geometry for comparing non-degenerate vMF distributions combines angular separation of mean directions with differences in tangent-space standard deviation. For $\boldsymbol{\mu}\in\mathbb{S}^{d-1}$05, the Wasserstein-like distance is

$\boldsymbol{\mu}\in\mathbb{S}^{d-1}$06

The first term is geodesic displacement and the second is a variance-like concentration discrepancy derived from a high-concentration tangent-Gaussian approximation. The construction supports barycenters and vMF mixture reduction, but it is not the exact Wasserstein distance between arbitrary vMF measures and is most justified in reasonably concentrated, non-degenerate regimes (You et al., 19 Apr 2025).

The principal limitations of vMF are its unit-norm requirement, loss of radial information, rotational symmetry, unimodality, isotropic angular dispersion, and conditional independence assumptions in mixture applications. More flexible directional distributions are required for anisotropic, asymmetric, multimodal, or axial data. Concentration estimation and density evaluation can be numerically delicate, particularly in high dimensions and at extreme concentration values. These limitations do not undermine the model’s central role: vMF provides a compact, interpretable, and computationally tractable probabilistic language for directional structure on hyperspherical spaces.