diststudent {mstudentd}

Distance/Divergence between Centered Multivariate t Distributions

Description

Computes the distance or divergence (Renyi divergence, Bhattacharyya distance or Hellinger distance) between two random vectors distributed according to multivariate $t$ distributions (MTD) with zero mean vector.

Usage

diststudent(nu1, Sigma1, nu2, Sigma2,
                   dist = c("renyi", "bhattacharyya", "hellinger"),
                   bet = NULL, eps = 1e-06)

Arguments

nu1

numéric. The degrees of freedom of the first distribution.

Sigma1

symmetric, positive-definite matrix. The correlation matrix of the first distribution.

nu2

numéric. The degrees of freedom of the second distribution.

Sigma2

symmetric, positive-definite matrix. The correlation matrix of the second distribution.

dist

character. The distance or divergence used. One of "renyi" (default), "battacharyya" or "hellinger".

bet

numeric, positive and not equal to 1. Order of the Renyi divergence. Ignored if distance="bhattacharyya" or distance="hellinger".

eps

numeric. Precision for the computation of the partial derivative of the Lauricella D-hypergeometric function (see Details). Default: 1e-06.

Details

Given X_1, a random vector of R^p distributed according to the MTD with parameters (\nu_1, \mathbf{0}, \Sigma_1) and X_2, a random vector of R^p distributed according to the MTD with parameters (\nu_2, \mathbf{0}, \Sigma_2).

Let \delta_1 = \frac{\nu_1 + p}{2} \beta, \delta_2 = \frac{\nu_2 + p}{2} (1 - \beta) and \lambda_1, \dots, \lambda_p the eigenvalues of the square matrix \Sigma_1 \Sigma_2^{-1} sorted in increasing order:

\lambda_1 < \dots < \lambda_{p-1} < \lambda_p

The Renyi divergence between X_1 and X_2 is:

\begin{aligned} D_R^\beta(\mathbf{X}_1||\mathbf{X}_1) & = & \displaystyle{\frac{1}{\beta - 1} \bigg[ \beta \ln\left(\frac{\Gamma\left(\frac{\nu_1+p}{2}\right) \Gamma\left(\frac{\nu_2}{2}\right) \nu_2^{\frac{p}{2}}}{\Gamma\left(\frac{\nu_2+p}{2}\right) \Gamma\left(\frac{\nu_1}{2}\right) \nu_1^{\frac{p}{2}}}\right) + \ln\left(\frac{\Gamma\left(\frac{\nu_2+p}{2}\right)}{\Gamma\left(\frac{\nu_2}{2}\right)}\right) + \ln\left(\frac{\Gamma\left(\delta_1 + \delta_2 - \frac{p}{2}\right)}{\Gamma(\delta_1 + \delta_2)}\right) } \\ && \displaystyle{- \frac{\beta}{2} \sum_{i=1}^p{\ln\lambda_i} + \ln F_D \bigg]} \end{aligned}

with F_D given by:

• If \displaystyle{\frac{\nu_1}{\nu_2} \lambda_1 > 1}: \displaystyle{ F_D = F_D^{(p)}{\bigg( \delta_1, \underbrace{\frac{1}{2}, \dots, \frac{1}{2}}_p; \delta_1+\delta_2; 1-\frac{\nu_2}{\nu_1 \lambda_1}, \dots, 1-\frac{\nu_2}{\nu_1 \lambda_p} \bigg)} }

• If \displaystyle{\frac{\nu_1}{\nu_2} \lambda_p < 1}: \displaystyle{ F_D = \prod_{i=1}^p{\left(\frac{\nu_1}{\nu_2} \lambda_i\right)^{\frac{1}{2}}} F_D^{(p)}\bigg(\delta_2, \underbrace{\frac{1}{2}, \dots, \frac{1}{2}}_p; \delta_1+\delta_2; 1-\frac{\nu_1}{\nu_2}\lambda_1, \dots, 1-\frac{\nu_1}{\nu_2}\lambda_p\bigg) }

• If \displaystyle{\frac{\nu_1}{\nu_2} \lambda_1 < 1} and \displaystyle{\frac{\nu_1}{\nu_2} \lambda_p > 1}:

  \displaystyle{ F_D = \left(\frac{\nu_2}{\nu_1} \frac{1}{\lambda_p}\right)^{\delta_2} \prod_{i=1}^p\left(\frac{\nu_1}{\nu_2}\lambda_i\right)^\frac{1}{2} F_D^{(p)}\bigg(\delta_2, \underbrace{\frac{1}{2}, \dots, \frac{1}{2}}_p, \delta_1+\delta_2-\frac{p}{2}; \delta_1+\delta2; 1-\frac{\lambda_1}{\lambda_p}, \dots, 1-\frac{\lambda_{p-1}}{\lambda_p}, 1-\frac{\nu_2}{\nu_1}\frac{1}{\lambda_p}\bigg) }

where F_D^{(p)} is the Lauricella D-hypergeometric function defined for p variables:

\displaystyle{ F_D^{(p)}\left(a; b_1, ..., b_p; g; x_1, ..., x_p\right) = \sum\limits_{m_1 \geq 0} ... \sum\limits_{m_p \geq 0}{ \frac{ (a)_{m_1+...+m_p}(b_1)_{m_1} ... (b_p)_{m_p} }{ (g)_{m_1+...+m_p} } \frac{x_1^{m_1}}{m_1!} ... \frac{x_p^{m_p}}{m_p!} } }

Its computation uses the lauricella function.

The Bhattacharyya distance is given by:

D_B(\mathbf{X}_1||\mathbf{X}_2) = \frac{1}{2} D_R^{1/2}(\mathbf{X}_1||\mathbf{X}_2)

And the Hellinger distance is given by:

D_H(\mathbf{X}_1||\mathbf{X}_2) = 1 - \exp{\left(-\frac{1}{2} D_R^{1/2}(\mathbf{X}_1||\mathbf{X}_2)\right)}

Value

A numeric value: the Renyi divergence between the two distributions, with two attributes attr(, "epsilon") (precision of the result of the Lauricella D-hypergeometric function,see Details) and attr(, "k") (number of iterations).

Author(s)

Pierre Santagostini, Nizar Bouhlel

References

N. Bouhlel and D. Rousseau (2023), Exact Rényi and Kullback-Leibler Divergences Between Multivariate t-Distributions, IEEE Signal Processing Letters. doi:10.1109/LSP.2023.3324594

Examples

nu1 <- 2
Sigma1 <- matrix(c(2, 1.2, 0.4, 1.2, 2, 0.6, 0.4, 0.6, 2), nrow = 3)
nu2 <- 4
Sigma2 <- matrix(c(1, 0.3, 0.1, 0.3, 1, 0.4, 0.1, 0.4, 1), nrow = 3)

# Renyi divergence
diststudent(nu1, Sigma1, nu2, Sigma2, bet = 0.25)
diststudent(nu2, Sigma2, nu1, Sigma1, bet = 0.25)

# Bhattacharyya distance
diststudent(nu1, Sigma1, nu2, Sigma2, dist = "bhattacharyya")
diststudent(nu2, Sigma2, nu1, Sigma1, dist = "bhattacharyya")

# Hellinger distance
diststudent(nu1, Sigma1, nu2, Sigma2, dist = "hellinger")
diststudent(nu2, Sigma2, nu1, Sigma1, dist = "hellinger")