Matrix t-distribution

Matrix t
Notation Tn,p(ν,𝐌,𝜮,𝜴)
Parameters

𝐌 location (real n×p matrix)
𝜴 scale (positive-definite real n×n matrix)
𝜮 scale (positive-definite real p×p matrix)

ν>0 degrees of freedom (real)
Support 𝐗∈ℝn×p
PDF

Γp(ν+n+p−12)(π)np2Γp(ν+p−12)|𝜴|−n2|𝜮|−p2

×|𝐈p+𝜮−1(𝐗−𝐌)𝜴−1(𝐗−𝐌)T|−ν+n+p−12
CDF No analytic expression
Mean 𝐌 if ν>1, else undefined
Mode 𝐌
Variance cov(vec(𝐗))=𝜮⊗𝜴ν−2 if ν>2, else undefined
CF see below

In statistics, the matrix t-distribution (or matrix variate t-distribution) is the generalization of the multivariate t-distribution from vectors to matrices.

The matrix t-distribution shares the same relationship with the multivariate t-distribution that the matrix normal distribution shares with the multivariate normal distribution: If the matrix has only one row, or only one column, the distributions become equivalent to the corresponding (vector-)multivariate distribution. The matrix t-distribution is the compound distribution that results from an infinite mixture of a matrix normal distribution with an inverse Wishart distribution placed over either of its covariance matrices, and the multivariate t-distribution can be generated in a similar way.

In a Bayesian analysis of a multivariate linear regression model based on the matrix normal distribution, the matrix t-distribution is the posterior predictive distribution.

Definition

For a matrix t-distribution, the probability density function at the point 𝐗 of an n×p space is

f(𝐗;ν,𝐌,𝜮,𝜴)=K×|𝐈n+𝜮−1(𝐗−𝐌)𝜴−1(𝐗−𝐌)T|−ν+n+p−12,

where the constant of integration K is given by

K=Γp(ν+n+p−12)(π)np2Γp(ν+p−12)|𝜴|−n2|𝜮|−p2.

Here Γp is the multivariate gamma function.

Properties

If 𝐗∼𝒯n×p(ν,𝐌,𝜮,𝜴), then we have the following properties:

Expected values

The mean, or expected value is, if ν>1:

E[𝐗]=𝐌

and we have the following second-order expectations, if ν>2:

E[(𝐗−𝐌)(𝐗−𝐌)T]=𝜮tr⁡(𝜴)ν−2
E[(𝐗−𝐌)T(𝐗−𝐌)]=𝜴tr⁡(𝜮)ν−2

where tr denotes trace.

More generally, for appropriately dimensioned matrices A,B,C:

E[(𝐗−𝐌)𝐀(𝐗−𝐌)T]=𝜮tr⁡(𝐀T𝜴)ν−2E[(𝐗−𝐌)T𝐁(𝐗−𝐌)]=𝜴tr⁡(𝐁T𝜮)ν−2E[(𝐗−𝐌)𝐂(𝐗−𝐌)]=𝜮𝐂T𝜴ν−2

Transformation

Transpose transform:

𝐗T∼𝒯p×n(ν,𝐌T,𝜴,𝜮)

Linear transform: let A (r-by-n), be of full rank r ≤ n and B (p-by-s), be of full rank s ≤ p, then:

𝐀𝐗𝐁∼𝒯r×s(ν,𝐀𝐌𝐁,𝐀𝜮𝐀T,𝐁T𝜴𝐁)

The characteristic function and various other properties can be derived from the re-parameterised formulation (see below).

Re-parameterized matrix t-distribution

Re-parameterized matrix t
Notation Tn,p(α,β,𝐌,𝜮,𝜴)
Parameters

𝐌 location (real n×p matrix)
𝜴 scale (positive-definite real p×p matrix)
𝜮 scale (positive-definite real n×n matrix)
α>(p−1)/2 shape parameter

β>0 scale parameter
Support 𝐗∈ℝn×p
PDF

Γp(α+n/2)(2π/β)np2Γp(α)|𝜴|−n2|𝜮|−p2

×|𝐈n+β2𝜮−1(𝐗−𝐌)𝜴−1(𝐗−𝐌)T|−(α+n/2)
CDF No analytic expression
Mean 𝐌 if α>p/2, else undefined
Variance 2(𝜮⊗𝜴)β(2α−p−1) if α>(p+1)/2, else undefined
CF see below

An alternative parameterisation of the matrix t-distribution uses two parameters α and β in place of ν.

This formulation reduces to the standard matrix t-distribution with β=2,α=ν+p−12.

This formulation of the matrix t-distribution can be derived as the compound distribution that results from an infinite mixture of a matrix normal distribution with an inverse multivariate gamma distribution placed over either of its covariance matrices.

Properties

If 𝐗∼Tn,p(α,β,𝐌,𝜮,𝜴) then

𝐗T∼Tp,n(α,β,𝐌T,𝜴,𝜮).

The property above comes from Sylvester's determinant theorem:

det⁡(𝐈n+β2𝜮−1(𝐗−𝐌)𝜴−1(𝐗−𝐌)T)=
det⁡(𝐈p+β2𝜴−1(𝐗T−𝐌T)𝜮−1(𝐗T−𝐌T)T).

If 𝐗∼Tn,p(α,β,𝐌,𝜮,𝜴) and 𝐀(n×n) and 𝐁(p×p) are nonsingular matrices then

𝐀𝐗𝐁∼Tn,p(α,β,𝐀𝐌𝐁,𝐀𝜮𝐀T,𝐁T𝜴𝐁).

The characteristic function is

ϕT(𝐙)=exp⁡(tr(i𝐙′𝐌))|𝜴|αΓp(α)(2β)αp|𝐙′𝜮𝐙|αBα(12β𝐙′𝜮𝐙𝜴),

where

Bδ(𝐖𝐙)=|𝐖|−δ∫𝐒>0exp⁡(tr(−𝐒𝐖−𝐒−𝟏𝐙))|𝐒|−δ−12(p+1)d𝐒,

and where Bδ is the type-two Bessel function of Herz[clarification needed] of a matrix argument.

See also