rfayat / sphereproba

Fit and manipulate a few probability distribution functions on the unit S2 sphere.

SphereProba

Fit and manipulate a few probability distribution functions on the unit S2 sphere.

Installation

Clone the repository and use pip to install the package.

git clone https://github.com/rfayat/SphereProba.git
cd SphereProba
pip install .

Generating the example figures

Code for generating the figures shown here can be found in the generate_examples.py script. To generate the example figures, install seaborn and angle_visualization and run:

python -m examples.generate_examples

von Mises-Fisher distribution on the unit sphere

The von Mises-Fisher distribution in dimension 3 (or Fisher distribution) is an isotropic probability distribution on the unit sphere. It is defined by a mean direction mu and a concentration parameter kappa (respectively analogue to the mean and invert variance of a Gaussian).

SphereProba.distributions.VonMisesFisher allows to fit such a distribution on an input set of 3-dimensional arrays (which are normalized if needed). VonMisesFisher objects support a __call__ method returning the value of the probability distribution function of input vectors.

Initialization

A VonMisesFisher object can be created by providing an array of length 3 representing its mean direction mu and a strictly positive float kappa corresponding to its concentration parameter:

>>> import numpy as np
>>> from SphereProba.distributions import VonMisesFisher
>>> vmf = VonMisesFisher(np.array([0, 0, 1]), 10.)
>>> print(vmf)
vMF distribution with parameters:
        μ = [0 0 1]
        κ = 10.0

Alternatively, it can be fitted on a 2-dimensional array of shape (n_arrays, 3) containing the x, y and z coordinates of a set of 3-dimensional vectors:

>>> dummy_data = np.array([[0, 0, 1.], [0, 0.01, 1.01]])
>>> vmf = VonMisesFisher.fit(dummy_data)
>>> print(vmf)
vMF distribution with parameters:
        μ = [0.         0.00495031 0.99998775]
        κ = 81613.99993282309

Example for different values of kappa

A visualization of the impact of the concentration parameter kappa on the shape of the distribution is shown below:

Converting kappa to an equivalent angle

Taking advantage of the isotropy of the von Mises-Fisher distribution (cylindrical symmetry around mu), we can convert the concentration parameter kappa in an equivalent angle theta such that integrating over the spherical cap centered on mu and having polar angle theta yields a user-defined value.

For instance, the angle theta for which 95% of the distribution is contained within angle theta of mu can be computed as follows:

>>> vmf = VonMisesFisher(np.array([0, 0, 1]), 10.)
>>> vmf.kappa_to_thetamax(.95)  # result in degrees by default
45.53874561998862

The proof for the explicit formula for computing the value of theta given both kappa and the expected value of the integral on the corresponding spherical cap is defined in this short pdf.

Kent distribution on the unit sphere

The Kent distribution (or 5-parameter Fisher-Bingham distribution) extends the von Mises-Fisher by making it anisotropic (ellipsoid on the sphere). It has parameters gamma, an orthonormal matrix whose components define the mean direction of the distribution (gamma1) and directions of major and minor axes of the ellipsoid (respectively gamma2 and gamma3).

Similarly to the von Mises-Fisher distribution, a parameter kapparules the concentration of the distribution while beta defines the extent to which the ellipsoid is stretched.

Fit from a range of 3D vectors

Similarly to SphereProba.distributions.VonMisesFisher, SphereProba.distributions.Kent supports initialization from user-defined parameters and pdf value using a __call__ method. Fitting the distribution on a set of 3-dimensional arrays (which are normalized if needed) can be done as follows:

>>> dummy_data = np.array([[ 0.        , -0.4472136 ,  0.89442719],
                           [-0.81649658,  0.40824829,  0.40824829],
                           [-0.85811633, -0.19069252,  0.47673129],
                           [ 0.30151134, -0.90453403,  0.30151134],
                           [ 0.53452248, -0.80178373, -0.26726124],
                           [-0.53452248, -0.80178373, -0.26726124]])
>>> kent = Kent.fit(dummy_data)
>>> print(kent)
Kent distribution with parameters:
   γ1 = [-0.19689796 -0.64910057  0.73477864]
   γ2 = [-0.69428639  0.621472    0.36295864]
   γ3 = [-0.692241   -0.43868099 -0.57302826]
   κ = 3.6618129030645763
   β = 0.572062833756161

Example for different values of kappa and beta

A visualization of the impact of kappa and beta on the shape of the distribution is shown below: