3.5. The Multivariate Normal Distribution 201
Note thatΣ^1 /^2 is symmetric and positive semi-definite. SupposeΣis positive
definite; that is, all of its eigenvalues are strictly positive. Based on this, it is then
easy to show that (
Σ^1 /^2
)− 1
=Γ′Λ−^1 /^2 Γ; (3.5.11)
see Exercise 3.5.13. We write the left side of this equation asΣ−^1 /^2. These matrices
enjoy many additional properties of the law of exponents for numbers; see, for
example, Arnold (1981). Here, though, all we need are the properties given above.
SupposeZhas aNn( 0 ,In) distribution. LetΣbe a positive semi-definite,
symmetric matrix and letμbe ann×1 vector of constants. Define the random
vectorXby
X=Σ^1 /^2 Z+μ. (3.5.12)
By (3.5.6) and Theorem 2.6.3, we immediately have
E[X]=μand Cov[X]=Σ^1 /^2 Σ^1 /^2 =Σ. (3.5.13)
Further, the mgf ofXis given by
MX(t)=E[exp{t′X}]=E
[
exp{t′Σ^1 /^2 Z+t′μ}
]
=exp{t′μ}E
[
exp
{(
Σ^1 /^2 t
)′
Z
}]
=exp{t′μ}exp
{
(1/2)
(
Σ^1 /^2 t
)′
Σ^1 /^2 t
}
=exp{t′μ}exp{(1/2)t′Σt}. (3.5.14)
This leads to the following definition:
Definition 3.5.1(Multivariate Normal).We say ann-dimensional random vector
Xhas amultivariate normal distributionif its mgf is
MX(t)=exp{t′μ+(1/2)t′Σt}, for allt∈Rn. (3.5.15)
whereΣis a symmetric, positive semi-definite matrix andμ∈Rn. We abbreviate
this by saying thatXhas aNn(μ,Σ)distribution.
Note that our definition is for positive semi-definite matricesΣ. UsuallyΣis
positive definite, in which case we can further obtain the density ofX.IfΣis
positive definite, then so isΣ^1 /^2 and, as discussed above, its inverse is given by
expression (3.5.11). Thus the transformation betweenXandZ, (3.5.12), is one-to-
one with the inverse transformation
Z=Σ−^1 /^2 (X−μ)
and the Jacobian|Σ−^1 /^2 |=|Σ|−^1 /^2. Hence, upon simplification, the pdf ofXis
given by
fX(x)=
1
(2π)n/^2 |Σ|^1 /^2
exp
{
−
1
2
(x−μ)′Σ−^1 (x−μ)
}
, forx∈Rn. (3.5.16)