See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/264168336

A Note on the Characteristic Function of Multivariate t Distribution

Article · January 2014

DOI: 10.5351/CSAM.2014.21.1.081

CITATIONS READS

3 327

3 authors, including:

Hyoung-Moon Kim
Konkuk University
39 PUBLICATIONS   321 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Skew-Normal distribution and its application View project

All content following this page was uploaded by Hyoung-Moon Kim on 16 January 2015.

The user has requested enhancement of the downloaded file.

Communications for Statistical Applications and Methods DOI: http://dx.doi.org/10.5351/CSAM.2014.21.1.081
2014, Vol. 21, No. 1, 81–91 ISSN 2287-7843

A Note on the Characteristic Function of Multivariate t

Distribution

Dae-Kun Songa , Hyoung-Jin Parkb , Hyoung-Moon Kim 1, b

a
Department of Statistics, Colorado State University, USA
b
Department of Applied Statistics, Konkuk University, Korea

Abstract
This study derives the characteristic functions of (multivariate/generalized) t distributions without contour
integration. We extended Hursts method (1995) to (multivariate/generalized) t distributions based on the principle
of randomization and mixtures. The derivation methods are relatively straightforward and are appropriate for
graduate level statistics theory courses.
Keywords: Randomization and mixtures, modified Bessel function of the third kind, contour inte-
gration, Laplace-Stieltjes transform.

1. Introduction
If X is a random variable, the characteristic function (cf) of X is defined by
[ ] ∫ ∞
ϕX (t) = E eitX = eitX dF X (x), t ∈ R,
−∞
√
where i = −1, F X (x) is the distribution function of X and eitX = cos(tX) + i sin(tX). The focus of
this note is on deriving the cf of the t distribution. Having heavier tails than the normal distribution,
t distribution is widely applied in many areas of statistics such as tests for two sample means, linear
regression analysis, and robust parametric modeling. Significant data is drawn from t distribution and
studying t distribution has become increasingly important in financial analysis.
It is natural to focus on the cf of t distribution because they uniquely determine distribution func-
tions. The cfs allow us to easily manipulate convolved probability density functions (pdfs) when they
represent linear combinations of independent random variables. With regard to asymptotics, cfs play
an important role in the Levy continuity theorem. Most importantly, the derivation of the central limit
theorem rests on an asymptotic expansion of cfs. These topics are covered in all standard graduate
level textbooks on probability (Durrett, 1996). Although cfs have many useful properties, most statis-
tics theory courses avoid rigorous derivations of these functions because statistics students are not
generally exposed to contour integration. A direct application of contour integration to develop the cf
of (multivariate/generalized) t distribution is burdensome since derivation methods require more steps
and more contour integration knowledge (Sutradhar, 1986) than the ones used in our study.
Students are typically advised to compute the moment generating function (when it is finite) of a
random variable and then substitute it for t in the moment generating function to obtain the cf, where
The corresponding author’s research was supported by Konkuk University in 2011.
1 Corresponding author: Professor, Department of Applied Statistics, Konkuk University, Seoul 143-701, Korea.
E-mail: hmkim@konkuk.ac.kr

Published online 31 January 2014 / journal homepage: http://csam.or.kr

⃝
c 2014 The Korean Statistical Society, Korean International Statistical Society. All rights reserved.
82 Dae-Kun Song, Hyoung-Jin Park, Hyoung-Moon Kim

√
i = −1. This approach works for many distributions; however, this method does not work when the
moment generating function is not finite (such as in t distribution with finite degrees of freedom). To
resolve this problem, Datta and Ghosh (2007) suggested the easy and creative derivations of cfs for
some basic probability density functions such as normal, double exponential, and Cauchy distribu-
tions. Their study did not include t distribution; however, the Cauchy distribution is a special case of
the t distribution.
Owing to the property of odd function, the cf of the univariate t distribution (with degrees of
freedom ν) is given by,
( ) ∫ ( √ )
Γ ν+1
2
∞ cos tx ν
ϕX (t) = ( ) √ dx.
Γ 2ν
v+1
π −∞ (1 + x2 ) 2

In 1956, Fisher and Healy tried to obtain a simple form of the cf of t distribution. However, they gave
it only for odd degrees of freedom, as follows:
√ ( √ )
ϕX (t) = e−|t ν| S ν t ν ,

where
ν−1 ( )
∑2 (ν − r − 1)! ν−1 r
2 !2 r
S ν (t) = ( ) t
ν−1
r=0 (ν − 1)! 2 − r !r!

with odd ν (ν = 1, 3, 5, . . .). That is,

S1 = 1
S3 = 1 + t
1
S 5 = 1 + t + t2
3
2 2 1 3
S7 = 1 + t + t + t
5 15
..
.

Since their work, many statisticians have derived the cf of t distribution using various methods of
their own. Ifram (1970) gave an expression of the cf for all degrees of freedom. Ifram derived the cf of
F distribution and found the cf of t distribution to be a special case of F distribution; however, Pestana
(1977) showed that the expression was incorrect. He indicated that Ifram (1970) made a mistake in
the contour integration when he used various quadratic substitutions. Hurst (1995) also derived the
cf of the univariate t distribution based on three facts: i) randomization and mixtures; ii) the cf of
symmetric generalized hyperbolic distribution in terms of Bessel function; and iii) some well known
approximations. A more recent derivation of the univariate case was presented by Dreier and Kotz
(2002), based on the theory of positive definite densities.
Sutradhar (1986, 1988) derived the first version of the cf of multivariate t distribution based on a
series representation composed of three types. The first is the case when the degree of freedom, ν, is
odd. The second case is when the degree of freedom, ν, is even. The third case deals with a fractional
value of ν. However, these representations are too complex. See Kotz and Nadarajah (2004) for a
general overview of the cfs of multivariate t distribution.
A Note on the Characteristic Function of Multivariate t Distribution 83

The rest of this paper is organized as follows. Section 2 presents some basic materials essential
to deriving the cf of t distribution. In Section 3, we suggest a simple method to derive the cf of
the t distribution for both the univariate and multivariate cases. The method of Hurst (1995) primar-
ily used the idea of Laplace-Stieltjes transform based on mixtures. Here, we extend his method to
(multivariate/generalized) t distribution. Finally, Section 4 concludes the paper.

2. Preliminaries
In this section, we present some basic materials essential to deriving the cf of t distribution. Let F be
a distribution function depending on a parameter θ, and u a probability density. Then
∫ ∞
W(x) = F(x, θ)u(θ)dθ
−∞

is a monotone function of x increasing from 0 to 1 and hence a distribution function, so we will denote
W(x) as F X (x) to emphasize the fact that it is a distribution function of X. If F has a continuous density
f , then W has a density w given by
∫ ∞
w(x) = f (x, θ)u(θ)dθ,
−∞

so we will denote w(x) as fX (x) to emphasize the fact that it is a density function of X. This process
may be described probabilistically as randomization and mixtures or simply mixtures (Feller, 1966).
One example of the mixtures
√ is called the variance-mean mixture (Barndorff-Nielsen et al., 1982).
Let X = µ + βU + σ UZ, where σ > 0, Z ∼ N(0, 1), and U follows a distribution function F on
[0, ∞). Further, assume that Z and U are independent. Then, X|U ∼ N(µ + βU, σ2 U) and we say
that the distribution of X is a normal variance-mean mixture with position µ, drift β, structure σ, and
mixing distribution F(u). Note that if β = 0, it is also called the scale mixtures of normal distributions.
Hence, the pdf of X is given by
{ }∫ ∞ { }
(x − µ)β 1 1 (x − µ)2 1 uβ2
exp √ exp − − dFU (u). (2.1)
σ2 0 2πσ2 u 2 σ2 u 2 σ2
It is well-known that the (multivariate/generalized) t distributions can be expressed as a normal
variance-mean mixture and as the scale mixtures of normal distributions. Let the random variable
√
X = UZ, (2.2)
where U ∼ IG(v/2, δ/2), Z ∼ N(0, 1), and U is independent of Z. The pdf of U ∼ IG(v/2, δ/2) is
given by
( ) 2ν  δ
δ 
 2
 

2 − 2ν −1
fU (u) = ( ) u exp 
 −  , u > 0.
Γ ν  u 
2

Then, the random variable X has a univariate generalized t distribution or Pearson type VII distribution
(see Fang et al., 1990, Section 3.3). We denote this as X ∼ gt(ν, δ) with the following pdf
( ) ( )− ν+1
Γ ν+1
2 x2 2
fX (x) = ( ) √ 1+ , x ∈ R, ν, δ > 0.
Γ ν δπ δ
2
84 Dae-Kun Song, Hyoung-Jin Park, Hyoung-Moon Kim

If ν = δ, then X follows the univariate t distribution; that is, X ∼ t(ν).

The multivariate extension is similar to the univariate case. Let the random vector
√
X = UZ, (2.3)

where U ∼ IG(ν/2, δ/2) independently of Z ∼ N p (0, I). Then, the random vector X follows the
multivariate generalized t distribution. We denote this as X ∼ gt p (ν, δ) with a pdf given by
( ) ( )− ν+p
Γ ν+p
2 x⊤ x 2
fX (x) = ( ) 1+ , x ∈ R p , ν, δ > 0.
Γ ν (δπ) 2
p
δ
2

If ν = δ, then X follows the multivariate t distribution; that is, X ∼ t p (ν). Deriving the univariate t
densities is straightforward using gamma integration, with µ = 0, β = 0, and σ = 1 in Equation (2.1).
The multivariate case is similar, but uses a density of N p (0, I) and gamma integration.
Now, we define the integral representation of the modified Bessel function of the third kind (Wat-
son, 1966, p.182) and examine some of its properties (Abramowitz and Stegun, 1972).

Definition 1. The integral representation of the modified Bessel function of the third kind is defined
by
∫ { ( )}
1 ∞ λ−1 1 1
Kλ (w) = x exp − w x + dx, w > 0.
2 0 2 x
Some properties of the modified Bessel function of the third kind are
I. Kλ (w) = K−λ (w), w > 0, λ ∈ R and
II. Kλ (w)  Γ(λ)2λ−1 w−λ as w → 0+ , λ > 0.

Proof: Because Abramowitz and Stegun (1972) skipped the proofs of these properties, we include
them here. We can prove property I using a transformation. Let y = 1/x, then Kλ (w) can be expressed
in terms of y, as follows:
∫ { ( )}
1 ∞ −λ−1 1 1
Kλ (w) = y exp − w + y dy
2 0 2 y
= K−λ (w).

To prove property II, we show that

Kλ (w)
lim+  1.
w→0 Γ(λ)2λ−1 w−λ
To do so, we first show
∫ ∞ ∫ ∞
lim+ fw (x)dx = lim fw (x)dx,
w→0 0 0 w→0+

where
( wx )λ−1 { wx } ( { w} )w
fw (x) = exp − exp − −1 .
2 2 2x 2
A Note on the Characteristic Function of Multivariate t Distribution 85

Now, let k = 1/w, then

( x )λ−1 { x }( {
1
} )
1
fw (x) = exp − exp − −1 .
2k 2k 2kx 2k

Note that limw→0+ fw (x) = limk→∞ fw (x) = 0 and | fw (x)| ≤ g(x), where g(x) = (wx/2)λ−1 exp {−wx/2}
w/2, since |e−w/(2x) − 1| ≤ 1. The application of the dominated convergence theorem ∫ ∞ allows us to
interchange the order of the limit and the integration since it is integrable. That is, 0 g(x)dx = Γ(λ).
∫∞ ∫∞
Thus, limw→0+ 0 fw (x)dx = 0 limw→0+ fw (x)dx. From these results, we have the following:

∫ ∞ 1 ( 2 )λ ( wx )λ−1 { } { }
Kλ (w) 0 2 w 2 exp − wx
2 exp − 2x 2 dx
w w
lim+ = lim+
w→0 Γ(λ)2λ−1 w−λ w→0 Γ(λ)2λ−1 w−λ
∫ (
∞ ) { wx } ( { w} )w
1 wx λ−1
= lim+ exp − 1 + exp − − 1 dx
w→0 Γ(λ) 0 2 2 2x 2
[ ∫ ∞ ( wx )λ−1 { }( { } ) ]
Γ(λ) + 0 2 exp − 2 exp − 2x − 1 2 dx
wx w w

= lim+
w→0 Γ(λ)
 ∫∞  ∫∞
 f (x)dx 

 limw→0+ fw (x)dx
= lim+ 1 + 0  = 1 + 0
w
= 1.
w→0 Γ(λ) Γ(λ)

Hence, we have proved property II. 

3. Extension of Hurst’s Method

Using Feller’s randomization and mixture, Hurst (1995) verified the relationship between the cf of
X|U = u and the√Laplace-Stieltjes transform of U. The relationship can be described briefly as
follows. Let X = UZ, where Z ∼ N(0, 1) and the non-negative random variable U are independent.
Then X|U ∼ N(0, U). Hence, the cf of X is given by

[ ] ∫ ∞ ∫ ∞ ∫ ∞
ϕX (t) = E eitX = eitx f (x)dx = eitx f (x|u) f (u)dudx
−∞ −∞
∫ ∞∫ ∞ ∫ ∞ 0

= eitx f (x|u)dx f (u)du = ϕX|u (t) f (u)du

0 −∞ 0
∫ ∞ ( )
t2
e− 2 t f (u)du = Lu
u 2
= , t ∈ R. (3.1)
0 2
∫∞
Note that the Laplace-Stieltjes transform Lu (t) of u is defined as E(e−tu ) = 0 e−tu dFU (u) when U is a
non-negative random variable (Feller, 1966). Hurst (1995) derived the cf of the symmetric generalized
hyperbolic distribution. He then found the cf of the univariate t distribution using this relationship and
the fact that t distribution is a special case of the generalized hyperbolic distribution.
We employ a similar idea, but use a simple transformation to derive the Laplace-Stieltjes trans-
form. We then extend this to the (multivariate/generalized) versions. The pdf of the generalized
inverse Gaussian distribution with parameter (λ, ν, ξ), X ∼ GIG(λ, ν, ξ), is given in Barndorff-Nielsen
86 Dae-Kun Song, Hyoung-Jin Park, Hyoung-Moon Kim

(1978) as
( ξ ) λ2 { }
ν λ−1 1( ν)
fX (x) = (√ )x exp − ξx + , x ∈ R+ ,
2Kλ νξ 2 x

where λ ∈ R, and ξ, ν ≥ 0.
Next, we express Lu (t) in terms of the modified Bessel function of the third kind. When U follows
GIG(λ, ν, ξ), the Laplace-Stieltjes transform is expressed as follows:
∫ ∞
Lu (t) = e−tu f (u)du
0
∫ ( ξ ) λ2 { }
∞
−tu ν uλ−1 1( ν)
= e ( √ ) exp − ξu + du
0 2Kλ νξ 2 u
( ξ ) λ2 ∫ { }
ν
∞
1( ν)
= (√ ) uλ−1 exp − (2t + ξ)u + du.
2Kλ νξ 0 2 u
√
If we now let y = (2t + ξ)/ν u, then Lu (t) becomes
( ξ ) λ2 (√ ) (√ )
Kλ ν(2t + ξ) ( ξ ) 2
λ
ν 2Kλ ν(2t + ξ)
(√ ) ( 2t+ξ ) λ2 = ( √ ) , t ∈ R+ . (3.2)
2Kλ νξ Kλ νξ 2t + ξ
ν

Hence, when λ ∈ R and ν = δ2 , ξ = α2 ≥ 0, the cf of the symmetric generalized hyperbolic

distribution can be obtained using (3.1) and (3.2), as follows:
( 2 ) K (δ √t2 + α2 ) ( ) λ2
t λ α2
ϕX (t) = Lu = , t ∈ R. (3.3)
2 Kλ (δα) t2 + α2

Note that this cf was derived in an alternative way in Barndorff-Nielsen and Blaesild (1981) using an
appropriate Fourier cosine transform. Hurst (1995) derived the cf using the Laplace-Stieltjes transform
of u.
Now we show that a generalized inverse Gaussian distribution with parameters, (−ν/2, ν, ξ → 0+ ),
is an inverse gamma distribution with parameters (ν/2, ν/2). Let λ = −ν/2, then the pdf of the
generalized inverse Gaussian distribution becomes
( ξ )− 4ν { }
ν − 2ν −1 1( ν)
fX (x) = (√ )x exp − ξx + .
2K− 2ν νξ 2 x

For ν > 0, using properties I and II of the modified Bessel function of the third kind, we have
( ξ )− 4ν ( ξ )− 4ν ( ) 4ν ( ) 2ν
ν ν
ν ν ξ 2
(√ ) = (√ )  ( ) ν ν
= ( )
2K− 2ν νξ 2K 2ν νξ 2Γ 2ν 2 2 −1 (νξ)− 4 Γ 2ν
A Note on the Characteristic Function of Multivariate t Distribution 87

as ξ → 0+ . The remaining part is

{ }
− 2ν −1 1( ν) ν ν/2
x exp − ξx +  x− 2 −1 e− x
2 x

as ξ → 0+ . Hence, fX (x) with λ = −ν/2 goes to

( ) 2ν
ν
ν ν/2
( ) x− 2 −1 e− x , x > 0
2
(3.4)
ν
Γ 2

as ξ → 0+ , which is the density of IG(ν/2, ν/2).

Therefore, we can now derive the cf of the t distribution, with degree of freedom ν. This is
explained in the next Section.

3.1. Univariate results

We first calculate the cf of the t distribution and then derive the cf of the generalized t distribution.
Result 1: Let X ∼ t(ν), then the cf of the t distribution is
( √ ) ( √ ) 2ν
K 2ν ν|t| ν|t|
ϕX (t) = ( ) ν , t ∈ R and ν > 0.
Γ 2ν 2 2 −1

Proof: Since the t distribution can be expressed as a normal variance-mean mixture using equation
(2.2), with ν = δ and because GIG(−ν/2, ν, ξ → 0+ ) becomes IG(ν/2, ν/2), we have
( 2 ) K ( √(t2 + ξ) ν) ( ) λ2
t λ ξ
Lu = (√ )  ϕX (t)
2 Kλ νξ t2 + ξ

as ξ → 0+ when λ = −ν/2. By properties I and II of the modified Bessel function of the third kind,
we obtain the cf of the t distribution as
( 2 ) K ν ( √(t2 + ξ) ν) ( )− 4ν
t ξ
 ( ) ν [ ]
2
Lu
Γ 2ν 2 2 −1 (νξ)− 4 t + ξ
ν 2
2
(√ )
K 2ν (t2 + ξ)ν ( [ ]) ν
= ( ) ν ν t2 + ξ 4
Γ 2ν 2 2 −1
( √ ) ( √ ) 2ν
K 2ν ν|t| ν|t|
= ( ) ν = ϕX (t)
Γ 2ν 2 2 −1

by substituting in ξ = 0. 

Extending this result to the generalized t distribution is straightforward. The generalized t dis-
tribution (Arellano-Valle and Bolfarine, 1995) retains its conditional distribution, but not in the t
distribution.
88 Dae-Kun Song, Hyoung-Jin Park, Hyoung-Moon Kim

Result 2: Let X ∼ gt(ν, δ), then the cf is

( √ ) ( √ ) 2ν
K 2ν δ|t| δ|t|
ϕX (t) = ( ) ν , t ∈ R and ν, δ > 0.
Γ 2ν 2 2 −1

Proof: Using the stochastic representation in (2.2) and that GIG(−ν/2, δ, ξ → 0+ ) becomes IG(ν/2,
δ/2) by a simple adjustment of previous result in (3.4), we have
( 2 ) K ν ( √(t2 + ξ) δ) ( )− 4ν
t ξ
 ( ) ν [2 ]
2
Lu
Γ 2ν 2 2 −1 (δξ)− 4 t + ξ
ν
2
( √( ) )
K 2ν t2 + ξ δ ( [ ]) ν
= ( ) ν δ t2 + ξ 4
Γ 2ν 2 2 −1
( √ ) ( √ ) 2ν
K 2ν δ|t| δ|t|
= ( ) ν = ϕX (t)
Γ 2ν 2 2 −1

by substituting in ξ = 0. 

3.2. Multivariate results

√
Let X = UZ, where Z ∼ N p (0, I) and a non-negative random variable U are independent. Then
X|U ∼ N p (0, UI). Hence, the cf of X is given by
[ ⊤ ] ∫ ⊤
∫
⊤
∫ ∞
ϕX (t) = E eit X = eit X fX (x)dx = eit X fX|U (x|u) fU (u)dudx
Rp Rp
∫ ∞∫ ∫ ∞ 0
⊤
= eit X fX|U (x|u)dx fU (u)du = ϕX|U (t) fU (u)du
Rp
0
∫ ∞ ( ) ( 0
)
1 t⊤ t
= exp − ut⊤ t fU (u)du = Lu F .cd
0 2 2

If the non-negative random variable U has GIG(λ, ν, ξ), then the cf of X is

√ ( ) λ2
Kλ ν (ξ + t⊤ t) ξ
ϕX (x) = ( √ ) [ ]
Kλ νξ ξ + t⊤ t

by (3.2), and substituting t⊤ t/2 for t. Using similar techniques for the univariate results, we can derive
the cf of the multivariate t distribution.
Result 3: Let X ∼ t p (ν). Then the cf of the multivariate t distribution is
( √ ) ( √ ) ν
K 2ν  νt  νt  2
ϕX (t) = ( ) ν , t ∈ R p and ν > 0,
Γ 2ν 2 2 −1
√
where ∥ t ∥= t⊤ t.
A Note on the Characteristic Function of Multivariate t Distribution 89

Proof: Since the multivariate t distribution can be expressed as a normal variance-mean mixture by
(2.3), with ν = δ, and GIG(−ν/2, ν, ξ → 0+ ) becomes IG(ν/2, ν/2), Lu (t⊤ t/2) converges to ϕX (t) as
ξ → 0+ when λ = −ν/2. Using properties I and II of the modified Bessel function of the third kind,
we obtain the cf of the multivariate t as follows:
( ⊤ ) √ ( )− 4ν
t t K 2ν (t⊤ t + ξ) ν ξ
Lu  ( ) ν [⊤ ]
Γ 2ν 2 2 −1 (νξ)− 4 t t + ξ
ν
2
√
K 2ν (t⊤ t + ξ) ν ( [ ⊤ ]) ν
= ( ) ν ν t t+ξ 4
ν −1
Γ 2 22
( √ ) ( √ ) ν
K 2ν  νt   νt  2
= ( ) ν = ϕX (t)
Γ 2ν 2 2 − 1

by substituting in ξ = 0. 

The cf of the multivariate generalized t distribution is obtained as follows.

Result 4: Let X ∼ gt p (ν, δ), then the cf of the multivariate generalized t distribution is
( √ ) ( √ ) ν
K 2ν  δt   δt  2
ϕX (t) = ( ) ν , t ∈ R p and ν, δ > 0,
Γ 2ν 2 2 −1
√
where ∥ t ∥= t⊤ t.

Proof: Using the stochastic representation (2.3) and that GIG(−ν/2, δ, ξ → 0+ ) becomes IG(ν/2, δ/2)
by simple adjustment of previous result in (3.4), we have
( ⊤ ) K ν ( √(t⊤ t + ξ)δ) ( )− 4ν
t t ξ
 ( ) ν [⊤ ]
2
Lu
Γ 2ν 2 2 −1 (δξ)− 4 t t + ξ
ν
2
(√ )
K 2ν (t⊤ t + ξ)δ ( [ ]) ν
= ( ) ν δ t⊤ t + ξ 4
ν −1
Γ 2 22
( √ ) ( √ ) 2ν
K 2ν  δt   δt
= ( ) ν = ϕX (t)
Γ 2ν 2 2 −1

by substituting in ξ = 0. 

3.3. Introducing location-scale parameters

Once we have the Results 1–4, the location-scale extensions are straightforward using the properties
of cf. Similar to normal distribution, for the univariate case, we have Y = µ + σX, where X ∼ gt(ν, δ)
and we have Y = µ+Σ1/2 X, where X ∼ gt p (ν, δ) for the multivariate case. Then, Y ∼ gt(µ, σ2 , ν, δ) and
Y ∼ gt p (µ, Σ, ν, δ), respectively. When ν = δ, we have the (multivariate) t distribution with location
and scale parameters.
90 Dae-Kun Song, Hyoung-Jin Park, Hyoung-Moon Kim

Since ϕY (t) = E[eitY ] = eitµ E[eitσX ] = eitµ ϕX (σt), we have the cf of Y as

(√ )(√ )ν
K 2ν δ|σt| δ|σt| 2
ϕY (t) = exp(itµ) ( ) ν , t ∈ R and ν, δ, σ > 0.
Γ 2ν 2 2 −1
⊤
Similarly ϕY (t) = eit µ ϕX (Σ1/2 t). Hence, we have the cf of Y as
( √ ) ( √ 1 ) 2ν
1 
ν  2 t 
( ) 2 
K δΣ   δΣ 2 t 
ϕY (t) = exp it⊤ µ ( ) ν ,
Γ 2ν 2 2 −1
√
t ∈ R p and ν, δ > 0, where ∥ t ∥= t⊤ t. Note that (Y − µ)/σ ∼ gt(ν, δ) and Σ−1/2 (Y − µ) ∼ gt p (ν, δ).

4. Conclusion
The t distribution has properties that make it a useful alternative to normal distribution. The additional
parameter of degrees of freedom governs the heaviness of the tails. This paper derived the cfs of
(multivariate/generalized) t distributions based on the principle of randomization and mixtures as well
as the cf of the symmetric generalized hyperbolic distribution. The expression of the cfs depend on the
integral representation of the modified Bessel function of the third kind. The derivation methods are
simple and independent of contour integration; consequently, the methods are suitable for graduate
students.

References
Abramowitz, M. and Stegun, I. A. (1972). Handbook of Mathematical Functions with Formulas,
Graphs, and Mathematical Tables, Dover Publications, INC, New York.
Arellano-Valle, R. B. and Bolfarine, H. (1995). On some characterizations of the t distribution, Statis-
tics & Probability Letters, 25, 79–85.
Barndorff-Nielsen, O. (1978). Hyperbolic distributions and distributions on hyperbola, Scandinavian
Journal of Statistics, 5, 151–157.
Barndorff-Nielsen, O. and Blaesild, P. (1981). Hyperbolic distributions and ramifications: Contribu-
tions to theory and application, In Taillie, C., Patil, G. P. and Baldessari, B. A., editors, Statistical
Distributions in Scientific Work, 4, 19–44, D. Reidel Publishing Company, Dordrecht.
Barndorff-Nielsen, O., Kent, J. and Sorensen, M. (1982). Normal variance-mean mixtures and z
distribution, International Statistical Review, 50, 145–159.
Datta, G. S. and Ghosh, M. (2007). Characteristic functions without contour integration, The Ameri-
can Statistician, 61, 67–70.
Dreier, I. and Kotz, S. (2002). A note on the characteristic function of the t-distribution, Statistics &
Probability Letters, 57, 221–224.
Durrett, R. (1996). Probability: Theory and Examples, 2nd edition, Duxbury Press, New York.
Fang, K. T., Kotz, S. and Ng, K. W. (1990). Symmetric Multivariate and Related Distributions,
Chapman and Hall, London.
Feller, W. (1966). An Introduction to Probability Theory and Its Applications, Volume II, Wiley, New
York.
Fisher, R. A. and Healy, M. J. R. (1956). New tables of Behrens’ test of significance, Journal of Royal
Statistical Society B, 18, 212–216.
 A Note on the Characteristic Function of Multivariate t Distribution 91

Hurst, S. (1995). The characteristic function of the student t distribution, Financial Mathematic Re-
search Report 006-95, Australian National University, Canberra ACT 0200, Australia.
Ifram, A. F. (1970). On the characteristic function of F and t distributions, Sankhya A, 32, 350–352.
Kotz, S. and Nadarajah, S. (2004). Multivariate t Distributions and Their Applications, Cambridge
University Press, Cambridge.
Pestana, D. (1977). Note on a paper by Ifram, Sankhya Series A, 39, 396–397.
Sutradhar, B. C. (1986). On the characteristic function of multivariate Student t-distribution, Cana-
dian Journal of Statistics, 14, 329–337.
Sutradhar, B. C. (1988). Author’s revision on the characteristic function of multivariate Student t-
distribution, Canadian Journal of Statistics, 16, 323.
Watson, G. N. (1966). A Treatise on the Theory of Bessel Functions, Cambridge University Press,
Cambridge.

Received October 31, 2013; Revised November 29, 2013; Accepted December 18, 2013

View publication stats